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, 2015, 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 <asm/div64.h>
  42 #include <asm/page.h>
  43 #include "o2iblnd.h"
  44
  45 static lnd_t the_o2iblnd = {
  46         .lnd_type     = O2IBLND,
  47         .lnd_startup  = kiblnd_startup,
  48         .lnd_shutdown = kiblnd_shutdown,
  49         .lnd_ctl      = kiblnd_ctl,
  50         .lnd_query    = kiblnd_query,
  51         .lnd_send     = kiblnd_send,
  52         .lnd_recv     = kiblnd_recv,
  53 };
  54
  55 kib_data_t kiblnd_data;
  56
  57 static __u32 kiblnd_cksum(void *ptr, int nob)
  58 {
  59         char *c = ptr;
  60         __u32 sum = 0;
  61
  62         while (nob-- > 0)
  63                 sum = ((sum << 1) | (sum >> 31)) + *c++;
  64
  65         /* ensure I don't return 0 (== no checksum) */
  66         return !sum ? 1 : sum;
  67 }
  68
  69 static char *kiblnd_msgtype2str(int type)
  70 {
  71         switch (type) {
  72         case IBLND_MSG_CONNREQ:
  73                 return "CONNREQ";
  74
  75         case IBLND_MSG_CONNACK:
  76                 return "CONNACK";
  77
  78         case IBLND_MSG_NOOP:
  79                 return "NOOP";
  80
  81         case IBLND_MSG_IMMEDIATE:
  82                 return "IMMEDIATE";
  83
  84         case IBLND_MSG_PUT_REQ:
  85                 return "PUT_REQ";
  86
  87         case IBLND_MSG_PUT_NAK:
  88                 return "PUT_NAK";
  89
  90         case IBLND_MSG_PUT_ACK:
  91                 return "PUT_ACK";
  92
  93         case IBLND_MSG_PUT_DONE:
  94                 return "PUT_DONE";
  95
  96         case IBLND_MSG_GET_REQ:
  97                 return "GET_REQ";
  98
  99         case IBLND_MSG_GET_DONE:
 100                 return "GET_DONE";
 101
 102         default:
 103                 return "???";
 104         }
 105 }
 106
 107 static int kiblnd_msgtype2size(int type)
 108 {
 109         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 110
 111         switch (type) {
 112         case IBLND_MSG_CONNREQ:
 113         case IBLND_MSG_CONNACK:
 114                 return hdr_size + sizeof(kib_connparams_t);
 115
 116         case IBLND_MSG_NOOP:
 117                 return hdr_size;
 118
 119         case IBLND_MSG_IMMEDIATE:
 120                 return offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[0]);
 121
 122         case IBLND_MSG_PUT_REQ:
 123                 return hdr_size + sizeof(kib_putreq_msg_t);
 124
 125         case IBLND_MSG_PUT_ACK:
 126                 return hdr_size + sizeof(kib_putack_msg_t);
 127
 128         case IBLND_MSG_GET_REQ:
 129                 return hdr_size + sizeof(kib_get_msg_t);
 130
 131         case IBLND_MSG_PUT_NAK:
 132         case IBLND_MSG_PUT_DONE:
 133         case IBLND_MSG_GET_DONE:
 134                 return hdr_size + sizeof(kib_completion_msg_t);
 135         default:
 136                 return -1;
 137         }
 138 }
 139
 140 static int kiblnd_unpack_rd(kib_msg_t *msg, int flip)
 141 {
 142         kib_rdma_desc_t *rd;
 143         int nob;
 144         int n;
 145         int i;
 146
 147         LASSERT(msg->ibm_type == IBLND_MSG_GET_REQ ||
 148                 msg->ibm_type == IBLND_MSG_PUT_ACK);
 149
 150         rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
 151                               &msg->ibm_u.get.ibgm_rd :
 152                               &msg->ibm_u.putack.ibpam_rd;
 153
 154         if (flip) {
 155                 __swab32s(&rd->rd_key);
 156                 __swab32s(&rd->rd_nfrags);
 157         }
 158
 159         n = rd->rd_nfrags;
 160
 161         if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
 162                 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
 163                        n, IBLND_MAX_RDMA_FRAGS);
 164                 return 1;
 165         }
 166
 167         nob = offsetof(kib_msg_t, ibm_u) +
 168               kiblnd_rd_msg_size(rd, msg->ibm_type, n);
 169
 170         if (msg->ibm_nob < nob) {
 171                 CERROR("Short %s: %d(%d)\n",
 172                        kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
 173                 return 1;
 174         }
 175
 176         if (!flip)
 177                 return 0;
 178
 179         for (i = 0; i < n; i++) {
 180                 __swab32s(&rd->rd_frags[i].rf_nob);
 181                 __swab64s(&rd->rd_frags[i].rf_addr);
 182         }
 183
 184         return 0;
 185 }
 186
 187 void kiblnd_pack_msg(lnet_ni_t *ni, kib_msg_t *msg, int version,
 188                      int credits, lnet_nid_t dstnid, __u64 dststamp)
 189 {
 190         kib_net_t *net = ni->ni_data;
 191
 192         /*
 193          * CAVEAT EMPTOR! all message fields not set here should have been
 194          * initialised previously.
 195          */
 196         msg->ibm_magic    = IBLND_MSG_MAGIC;
 197         msg->ibm_version  = version;
 198         /*   ibm_type */
 199         msg->ibm_credits  = credits;
 200         /*   ibm_nob */
 201         msg->ibm_cksum    = 0;
 202         msg->ibm_srcnid   = ni->ni_nid;
 203         msg->ibm_srcstamp = net->ibn_incarnation;
 204         msg->ibm_dstnid   = dstnid;
 205         msg->ibm_dststamp = dststamp;
 206
 207         if (*kiblnd_tunables.kib_cksum) {
 208                 /* NB ibm_cksum zero while computing cksum */
 209                 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
 210         }
 211 }
 212
 213 int kiblnd_unpack_msg(kib_msg_t *msg, int nob)
 214 {
 215         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 216         __u32 msg_cksum;
 217         __u16 version;
 218         int msg_nob;
 219         int flip;
 220
 221         /* 6 bytes are enough to have received magic + version */
 222         if (nob < 6) {
 223                 CERROR("Short message: %d\n", nob);
 224                 return -EPROTO;
 225         }
 226
 227         if (msg->ibm_magic == IBLND_MSG_MAGIC) {
 228                 flip = 0;
 229         } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
 230                 flip = 1;
 231         } else {
 232                 CERROR("Bad magic: %08x\n", msg->ibm_magic);
 233                 return -EPROTO;
 234         }
 235
 236         version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
 237         if (version != IBLND_MSG_VERSION &&
 238             version != IBLND_MSG_VERSION_1) {
 239                 CERROR("Bad version: %x\n", version);
 240                 return -EPROTO;
 241         }
 242
 243         if (nob < hdr_size) {
 244                 CERROR("Short message: %d\n", nob);
 245                 return -EPROTO;
 246         }
 247
 248         msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
 249         if (msg_nob > nob) {
 250                 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
 251                 return -EPROTO;
 252         }
 253
 254         /*
 255          * checksum must be computed with ibm_cksum zero and BEFORE anything
 256          * gets flipped
 257          */
 258         msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
 259         msg->ibm_cksum = 0;
 260         if (msg_cksum &&
 261             msg_cksum != kiblnd_cksum(msg, msg_nob)) {
 262                 CERROR("Bad checksum\n");
 263                 return -EPROTO;
 264         }
 265
 266         msg->ibm_cksum = msg_cksum;
 267
 268         if (flip) {
 269                 /* leave magic unflipped as a clue to peer endianness */
 270                 msg->ibm_version = version;
 271                 CLASSERT(sizeof(msg->ibm_type) == 1);
 272                 CLASSERT(sizeof(msg->ibm_credits) == 1);
 273                 msg->ibm_nob     = msg_nob;
 274                 __swab64s(&msg->ibm_srcnid);
 275                 __swab64s(&msg->ibm_srcstamp);
 276                 __swab64s(&msg->ibm_dstnid);
 277                 __swab64s(&msg->ibm_dststamp);
 278         }
 279
 280         if (msg->ibm_srcnid == LNET_NID_ANY) {
 281                 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
 282                 return -EPROTO;
 283         }
 284
 285         if (msg_nob < kiblnd_msgtype2size(msg->ibm_type)) {
 286                 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg->ibm_type),
 287                        msg_nob, kiblnd_msgtype2size(msg->ibm_type));
 288                 return -EPROTO;
 289         }
 290
 291         switch (msg->ibm_type) {
 292         default:
 293                 CERROR("Unknown message type %x\n", msg->ibm_type);
 294                 return -EPROTO;
 295
 296         case IBLND_MSG_NOOP:
 297         case IBLND_MSG_IMMEDIATE:
 298         case IBLND_MSG_PUT_REQ:
 299                 break;
 300
 301         case IBLND_MSG_PUT_ACK:
 302         case IBLND_MSG_GET_REQ:
 303                 if (kiblnd_unpack_rd(msg, flip))
 304                         return -EPROTO;
 305                 break;
 306
 307         case IBLND_MSG_PUT_NAK:
 308         case IBLND_MSG_PUT_DONE:
 309         case IBLND_MSG_GET_DONE:
 310                 if (flip)
 311                         __swab32s(&msg->ibm_u.completion.ibcm_status);
 312                 break;
 313
 314         case IBLND_MSG_CONNREQ:
 315         case IBLND_MSG_CONNACK:
 316                 if (flip) {
 317                         __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
 318                         __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
 319                         __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
 320                 }
 321                 break;
 322         }
 323         return 0;
 324 }
 325
 326 int kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid)
 327 {
 328         kib_peer_t *peer;
 329         kib_net_t *net = ni->ni_data;
 330         int cpt = lnet_cpt_of_nid(nid);
 331         unsigned long flags;
 332
 333         LASSERT(net);
 334         LASSERT(nid != LNET_NID_ANY);
 335
 336         LIBCFS_CPT_ALLOC(peer, lnet_cpt_table(), cpt, sizeof(*peer));
 337         if (!peer) {
 338                 CERROR("Cannot allocate peer\n");
 339                 return -ENOMEM;
 340         }
 341
 342         peer->ibp_ni = ni;
 343         peer->ibp_nid = nid;
 344         peer->ibp_error = 0;
 345         peer->ibp_last_alive = 0;
 346         atomic_set(&peer->ibp_refcount, 1);  /* 1 ref for caller */
 347
 348         INIT_LIST_HEAD(&peer->ibp_list);     /* not in the peer table yet */
 349         INIT_LIST_HEAD(&peer->ibp_conns);
 350         INIT_LIST_HEAD(&peer->ibp_tx_queue);
 351
 352         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 353
 354         /* always called with a ref on ni, which prevents ni being shutdown */
 355         LASSERT(!net->ibn_shutdown);
 356
 357         /* npeers only grows with the global lock held */
 358         atomic_inc(&net->ibn_npeers);
 359
 360         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 361
 362         *peerp = peer;
 363         return 0;
 364 }
 365
 366 void kiblnd_destroy_peer(kib_peer_t *peer)
 367 {
 368         kib_net_t *net = peer->ibp_ni->ni_data;
 369
 370         LASSERT(net);
 371         LASSERT(!atomic_read(&peer->ibp_refcount));
 372         LASSERT(!kiblnd_peer_active(peer));
 373         LASSERT(!peer->ibp_connecting);
 374         LASSERT(!peer->ibp_accepting);
 375         LASSERT(list_empty(&peer->ibp_conns));
 376         LASSERT(list_empty(&peer->ibp_tx_queue));
 377
 378         LIBCFS_FREE(peer, sizeof(*peer));
 379
 380         /*
 381          * NB a peer's connections keep a reference on their peer until
 382          * they are destroyed, so we can be assured that _all_ state to do
 383          * with this peer has been cleaned up when its refcount drops to
 384          * zero.
 385          */
 386         atomic_dec(&net->ibn_npeers);
 387 }
 388
 389 kib_peer_t *kiblnd_find_peer_locked(lnet_nid_t nid)
 390 {
 391         /*
 392          * the caller is responsible for accounting the additional reference
 393          * that this creates
 394          */
 395         struct list_head *peer_list = kiblnd_nid2peerlist(nid);
 396         struct list_head *tmp;
 397         kib_peer_t *peer;
 398
 399         list_for_each(tmp, peer_list) {
 400                 peer = list_entry(tmp, kib_peer_t, ibp_list);
 401
 402                 LASSERT(peer->ibp_connecting > 0 || /* creating conns */
 403                          peer->ibp_accepting > 0 ||
 404                          !list_empty(&peer->ibp_conns));  /* active conn */
 405
 406                 if (peer->ibp_nid != nid)
 407                         continue;
 408
 409                 CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n",
 410                        peer, libcfs_nid2str(nid),
 411                        atomic_read(&peer->ibp_refcount),
 412                        peer->ibp_version);
 413                 return peer;
 414         }
 415         return NULL;
 416 }
 417
 418 void kiblnd_unlink_peer_locked(kib_peer_t *peer)
 419 {
 420         LASSERT(list_empty(&peer->ibp_conns));
 421
 422         LASSERT(kiblnd_peer_active(peer));
 423         list_del_init(&peer->ibp_list);
 424         /* lose peerlist's ref */
 425         kiblnd_peer_decref(peer);
 426 }
 427
 428 static int kiblnd_get_peer_info(lnet_ni_t *ni, int index,
 429                                 lnet_nid_t *nidp, int *count)
 430 {
 431         kib_peer_t *peer;
 432         struct list_head *ptmp;
 433         int i;
 434         unsigned long flags;
 435
 436         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 437
 438         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 439                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 440                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 441                         LASSERT(peer->ibp_connecting > 0 ||
 442                                 peer->ibp_accepting > 0 ||
 443                                 !list_empty(&peer->ibp_conns));
 444
 445                         if (peer->ibp_ni != ni)
 446                                 continue;
 447
 448                         if (index-- > 0)
 449                                 continue;
 450
 451                         *nidp = peer->ibp_nid;
 452                         *count = atomic_read(&peer->ibp_refcount);
 453
 454                         read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 455                                                flags);
 456                         return 0;
 457                 }
 458         }
 459
 460         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 461         return -ENOENT;
 462 }
 463
 464 static void kiblnd_del_peer_locked(kib_peer_t *peer)
 465 {
 466         struct list_head *ctmp;
 467         struct list_head *cnxt;
 468         kib_conn_t *conn;
 469
 470         if (list_empty(&peer->ibp_conns)) {
 471                 kiblnd_unlink_peer_locked(peer);
 472         } else {
 473                 list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 474                         conn = list_entry(ctmp, kib_conn_t, ibc_list);
 475
 476                         kiblnd_close_conn_locked(conn, 0);
 477                 }
 478                 /* NB closing peer's last conn unlinked it. */
 479         }
 480         /*
 481          * NB peer now unlinked; might even be freed if the peer table had the
 482          * last ref on it.
 483          */
 484 }
 485
 486 static int kiblnd_del_peer(lnet_ni_t *ni, lnet_nid_t nid)
 487 {
 488         LIST_HEAD(zombies);
 489         struct list_head *ptmp;
 490         struct list_head *pnxt;
 491         kib_peer_t *peer;
 492         int lo;
 493         int hi;
 494         int i;
 495         unsigned long flags;
 496         int rc = -ENOENT;
 497
 498         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 499
 500         if (nid != LNET_NID_ANY) {
 501                 lo = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 502                 hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 503         } else {
 504                 lo = 0;
 505                 hi = kiblnd_data.kib_peer_hash_size - 1;
 506         }
 507
 508         for (i = lo; i <= hi; i++) {
 509                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 510                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 511                         LASSERT(peer->ibp_connecting > 0 ||
 512                                 peer->ibp_accepting > 0 ||
 513                                 !list_empty(&peer->ibp_conns));
 514
 515                         if (peer->ibp_ni != ni)
 516                                 continue;
 517
 518                         if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
 519                                 continue;
 520
 521                         if (!list_empty(&peer->ibp_tx_queue)) {
 522                                 LASSERT(list_empty(&peer->ibp_conns));
 523
 524                                 list_splice_init(&peer->ibp_tx_queue,
 525                                                  &zombies);
 526                         }
 527
 528                         kiblnd_del_peer_locked(peer);
 529                         rc = 0;  /* matched something */
 530                 }
 531         }
 532
 533         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 534
 535         kiblnd_txlist_done(ni, &zombies, -EIO);
 536
 537         return rc;
 538 }
 539
 540 static kib_conn_t *kiblnd_get_conn_by_idx(lnet_ni_t *ni, int index)
 541 {
 542         kib_peer_t *peer;
 543         struct list_head *ptmp;
 544         kib_conn_t *conn;
 545         struct list_head *ctmp;
 546         int i;
 547         unsigned long flags;
 548
 549         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 550
 551         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 552                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 553                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 554                         LASSERT(peer->ibp_connecting > 0 ||
 555                                 peer->ibp_accepting > 0 ||
 556                                 !list_empty(&peer->ibp_conns));
 557
 558                         if (peer->ibp_ni != ni)
 559                                 continue;
 560
 561                         list_for_each(ctmp, &peer->ibp_conns) {
 562                                 if (index-- > 0)
 563                                         continue;
 564
 565                                 conn = list_entry(ctmp, kib_conn_t,
 566                                                   ibc_list);
 567                                 kiblnd_conn_addref(conn);
 568                                 read_unlock_irqrestore(
 569                                         &kiblnd_data.kib_global_lock,
 570                                         flags);
 571                                 return conn;
 572                         }
 573                 }
 574         }
 575
 576         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 577         return NULL;
 578 }
 579
 580 int kiblnd_translate_mtu(int value)
 581 {
 582         switch (value) {
 583         default:
 584                 return -1;
 585         case 0:
 586                 return 0;
 587         case 256:
 588                 return IB_MTU_256;
 589         case 512:
 590                 return IB_MTU_512;
 591         case 1024:
 592                 return IB_MTU_1024;
 593         case 2048:
 594                 return IB_MTU_2048;
 595         case 4096:
 596                 return IB_MTU_4096;
 597         }
 598 }
 599
 600 static void kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
 601 {
 602         int mtu;
 603
 604         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 605         if (!cmid->route.path_rec)
 606                 return;
 607
 608         mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
 609         LASSERT(mtu >= 0);
 610         if (mtu)
 611                 cmid->route.path_rec->mtu = mtu;
 612 }
 613
 614 static int kiblnd_get_completion_vector(kib_conn_t *conn, int cpt)
 615 {
 616         cpumask_t *mask;
 617         int vectors;
 618         int off;
 619         int i;
 620         lnet_nid_t nid = conn->ibc_peer->ibp_nid;
 621
 622         vectors = conn->ibc_cmid->device->num_comp_vectors;
 623         if (vectors <= 1)
 624                 return 0;
 625
 626         mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
 627         if (!mask)
 628                 return 0;
 629
 630         /* hash NID to CPU id in this partition... */
 631         off = do_div(nid, cpumask_weight(mask));
 632         for_each_cpu(i, mask) {
 633                 if (!off--)
 634                         return i % vectors;
 635         }
 636
 637         LBUG();
 638         return 1;
 639 }
 640
 641 kib_conn_t *kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
 642                                int state, int version)
 643 {
 644         /*
 645          * CAVEAT EMPTOR:
 646          * If the new conn is created successfully it takes over the caller's
 647          * ref on 'peer'.  It also "owns" 'cmid' and destroys it when it itself
 648          * is destroyed.  On failure, the caller's ref on 'peer' remains and
 649          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 650          * to destroy 'cmid' here since I'm called from the CM which still has
 651          * its ref on 'cmid').
 652          */
 653         rwlock_t *glock = &kiblnd_data.kib_global_lock;
 654         kib_net_t *net = peer->ibp_ni->ni_data;
 655         kib_dev_t *dev;
 656         struct ib_qp_init_attr *init_qp_attr;
 657         struct kib_sched_info *sched;
 658         struct ib_cq_init_attr cq_attr = {};
 659         kib_conn_t *conn;
 660         struct ib_cq *cq;
 661         unsigned long flags;
 662         int cpt;
 663         int rc;
 664         int i;
 665
 666         LASSERT(net);
 667         LASSERT(!in_interrupt());
 668
 669         dev = net->ibn_dev;
 670
 671         cpt = lnet_cpt_of_nid(peer->ibp_nid);
 672         sched = kiblnd_data.kib_scheds[cpt];
 673
 674         LASSERT(sched->ibs_nthreads > 0);
 675
 676         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 677                          sizeof(*init_qp_attr));
 678         if (!init_qp_attr) {
 679                 CERROR("Can't allocate qp_attr for %s\n",
 680                        libcfs_nid2str(peer->ibp_nid));
 681                 goto failed_0;
 682         }
 683
 684         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 685         if (!conn) {
 686                 CERROR("Can't allocate connection for %s\n",
 687                        libcfs_nid2str(peer->ibp_nid));
 688                 goto failed_1;
 689         }
 690
 691         conn->ibc_state = IBLND_CONN_INIT;
 692         conn->ibc_version = version;
 693         conn->ibc_peer = peer;            /* I take the caller's ref */
 694         cmid->context = conn;              /* for future CM callbacks */
 695         conn->ibc_cmid = cmid;
 696
 697         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 698         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 699         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 700         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 701         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 702         INIT_LIST_HEAD(&conn->ibc_active_txs);
 703         spin_lock_init(&conn->ibc_lock);
 704
 705         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 706                          sizeof(*conn->ibc_connvars));
 707         if (!conn->ibc_connvars) {
 708                 CERROR("Can't allocate in-progress connection state\n");
 709                 goto failed_2;
 710         }
 711
 712         write_lock_irqsave(glock, flags);
 713         if (dev->ibd_failover) {
 714                 write_unlock_irqrestore(glock, flags);
 715                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 716                 goto failed_2;
 717         }
 718
 719         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 720                 /* wakeup failover thread and teardown connection */
 721                 if (kiblnd_dev_can_failover(dev)) {
 722                         list_add_tail(&dev->ibd_fail_list,
 723                                       &kiblnd_data.kib_failed_devs);
 724                         wake_up(&kiblnd_data.kib_failover_waitq);
 725                 }
 726
 727                 write_unlock_irqrestore(glock, flags);
 728                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 729                        cmid->device->name, dev->ibd_ifname);
 730                 goto failed_2;
 731         }
 732
 733         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 734         conn->ibc_hdev = dev->ibd_hdev;
 735
 736         kiblnd_setup_mtu_locked(cmid);
 737
 738         write_unlock_irqrestore(glock, flags);
 739
 740         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 741                          IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
 742         if (!conn->ibc_rxs) {
 743                 CERROR("Cannot allocate RX buffers\n");
 744                 goto failed_2;
 745         }
 746
 747         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 748                                 IBLND_RX_MSG_PAGES(version));
 749         if (rc)
 750                 goto failed_2;
 751
 752         kiblnd_map_rx_descs(conn);
 753
 754         cq_attr.cqe = IBLND_CQ_ENTRIES(version);
 755         cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
 756         cq = ib_create_cq(cmid->device,
 757                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 758                           &cq_attr);
 759         if (IS_ERR(cq)) {
 760                 CERROR("Can't create CQ: %ld, cqe: %d\n",
 761                        PTR_ERR(cq), IBLND_CQ_ENTRIES(version));
 762                 goto failed_2;
 763         }
 764
 765         conn->ibc_cq = cq;
 766
 767         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 768         if (rc) {
 769                 CERROR("Can't request completion notificiation: %d\n", rc);
 770                 goto failed_2;
 771         }
 772
 773         init_qp_attr->event_handler = kiblnd_qp_event;
 774         init_qp_attr->qp_context = conn;
 775         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(version);
 776         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(version);
 777         init_qp_attr->cap.max_send_sge = 1;
 778         init_qp_attr->cap.max_recv_sge = 1;
 779         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 780         init_qp_attr->qp_type = IB_QPT_RC;
 781         init_qp_attr->send_cq = cq;
 782         init_qp_attr->recv_cq = cq;
 783
 784         conn->ibc_sched = sched;
 785
 786         rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 787         if (rc) {
 788                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
 789                        rc, init_qp_attr->cap.max_send_wr,
 790                        init_qp_attr->cap.max_recv_wr);
 791                 goto failed_2;
 792         }
 793
 794         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 795
 796         /* 1 ref for caller and each rxmsg */
 797         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(version));
 798         conn->ibc_nrx = IBLND_RX_MSGS(version);
 799
 800         /* post receives */
 801         for (i = 0; i < IBLND_RX_MSGS(version); i++) {
 802                 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
 803                                     IBLND_POSTRX_NO_CREDIT);
 804                 if (rc) {
 805                         CERROR("Can't post rxmsg: %d\n", rc);
 806
 807                         /* Make posted receives complete */
 808                         kiblnd_abort_receives(conn);
 809
 810                         /*
 811                          * correct # of posted buffers
 812                          * NB locking needed now I'm racing with completion
 813                          */
 814                         spin_lock_irqsave(&sched->ibs_lock, flags);
 815                         conn->ibc_nrx -= IBLND_RX_MSGS(version) - i;
 816                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 817
 818                         /*
 819                          * cmid will be destroyed by CM(ofed) after cm_callback
 820                          * returned, so we can't refer it anymore
 821                          * (by kiblnd_connd()->kiblnd_destroy_conn)
 822                          */
 823                         rdma_destroy_qp(conn->ibc_cmid);
 824                         conn->ibc_cmid = NULL;
 825
 826                         /* Drop my own and unused rxbuffer refcounts */
 827                         while (i++ <= IBLND_RX_MSGS(version))
 828                                 kiblnd_conn_decref(conn);
 829
 830                         return NULL;
 831                 }
 832         }
 833
 834         /* Init successful! */
 835         LASSERT(state == IBLND_CONN_ACTIVE_CONNECT ||
 836                 state == IBLND_CONN_PASSIVE_WAIT);
 837         conn->ibc_state = state;
 838
 839         /* 1 more conn */
 840         atomic_inc(&net->ibn_nconns);
 841         return conn;
 842
 843  failed_2:
 844         kiblnd_destroy_conn(conn);
 845  failed_1:
 846         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 847  failed_0:
 848         return NULL;
 849 }
 850
 851 void kiblnd_destroy_conn(kib_conn_t *conn)
 852 {
 853         struct rdma_cm_id *cmid = conn->ibc_cmid;
 854         kib_peer_t *peer = conn->ibc_peer;
 855         int rc;
 856
 857         LASSERT(!in_interrupt());
 858         LASSERT(!atomic_read(&conn->ibc_refcount));
 859         LASSERT(list_empty(&conn->ibc_early_rxs));
 860         LASSERT(list_empty(&conn->ibc_tx_noops));
 861         LASSERT(list_empty(&conn->ibc_tx_queue));
 862         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
 863         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
 864         LASSERT(list_empty(&conn->ibc_active_txs));
 865         LASSERT(!conn->ibc_noops_posted);
 866         LASSERT(!conn->ibc_nsends_posted);
 867
 868         switch (conn->ibc_state) {
 869         default:
 870                 /* conn must be completely disengaged from the network */
 871                 LBUG();
 872
 873         case IBLND_CONN_DISCONNECTED:
 874                 /* connvars should have been freed already */
 875                 LASSERT(!conn->ibc_connvars);
 876                 break;
 877
 878         case IBLND_CONN_INIT:
 879                 break;
 880         }
 881
 882         /* conn->ibc_cmid might be destroyed by CM already */
 883         if (cmid && cmid->qp)
 884                 rdma_destroy_qp(cmid);
 885
 886         if (conn->ibc_cq) {
 887                 rc = ib_destroy_cq(conn->ibc_cq);
 888                 if (rc)
 889                         CWARN("Error destroying CQ: %d\n", rc);
 890         }
 891
 892         if (conn->ibc_rx_pages)
 893                 kiblnd_unmap_rx_descs(conn);
 894
 895         if (conn->ibc_rxs) {
 896                 LIBCFS_FREE(conn->ibc_rxs,
 897                             IBLND_RX_MSGS(conn->ibc_version)
 898                               * sizeof(kib_rx_t));
 899         }
 900
 901         if (conn->ibc_connvars)
 902                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 903
 904         if (conn->ibc_hdev)
 905                 kiblnd_hdev_decref(conn->ibc_hdev);
 906
 907         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 908         if (conn->ibc_state != IBLND_CONN_INIT) {
 909                 kib_net_t *net = peer->ibp_ni->ni_data;
 910
 911                 kiblnd_peer_decref(peer);
 912                 rdma_destroy_id(cmid);
 913                 atomic_dec(&net->ibn_nconns);
 914         }
 915
 916         LIBCFS_FREE(conn, sizeof(*conn));
 917 }
 918
 919 int kiblnd_close_peer_conns_locked(kib_peer_t *peer, int why)
 920 {
 921         kib_conn_t *conn;
 922         struct list_head *ctmp;
 923         struct list_head *cnxt;
 924         int count = 0;
 925
 926         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 927                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 928
 929                 CDEBUG(D_NET, "Closing conn -> %s, version: %x, reason: %d\n",
 930                        libcfs_nid2str(peer->ibp_nid),
 931                        conn->ibc_version, why);
 932
 933                 kiblnd_close_conn_locked(conn, why);
 934                 count++;
 935         }
 936
 937         return count;
 938 }
 939
 940 int kiblnd_close_stale_conns_locked(kib_peer_t *peer,
 941                                     int version, __u64 incarnation)
 942 {
 943         kib_conn_t *conn;
 944         struct list_head *ctmp;
 945         struct list_head *cnxt;
 946         int count = 0;
 947
 948         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 949                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 950
 951                 if (conn->ibc_version     == version &&
 952                     conn->ibc_incarnation == incarnation)
 953                         continue;
 954
 955                 CDEBUG(D_NET,
 956                        "Closing stale conn -> %s version: %x, incarnation:%#llx(%x, %#llx)\n",
 957                        libcfs_nid2str(peer->ibp_nid),
 958                        conn->ibc_version, conn->ibc_incarnation,
 959                        version, incarnation);
 960
 961                 kiblnd_close_conn_locked(conn, -ESTALE);
 962                 count++;
 963         }
 964
 965         return count;
 966 }
 967
 968 static int kiblnd_close_matching_conns(lnet_ni_t *ni, lnet_nid_t nid)
 969 {
 970         kib_peer_t *peer;
 971         struct list_head *ptmp;
 972         struct list_head *pnxt;
 973         int lo;
 974         int hi;
 975         int i;
 976         unsigned long flags;
 977         int count = 0;
 978
 979         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 980
 981         if (nid != LNET_NID_ANY) {
 982                 lo = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 983                 hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 984         } else {
 985                 lo = 0;
 986                 hi = kiblnd_data.kib_peer_hash_size - 1;
 987         }
 988
 989         for (i = lo; i <= hi; i++) {
 990                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 991                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 992                         LASSERT(peer->ibp_connecting > 0 ||
 993                                 peer->ibp_accepting > 0 ||
 994                                 !list_empty(&peer->ibp_conns));
 995
 996                         if (peer->ibp_ni != ni)
 997                                 continue;
 998
 999                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1000                                 continue;
1001
1002                         count += kiblnd_close_peer_conns_locked(peer, 0);
1003                 }
1004         }
1005
1006         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1007
1008         /* wildcards always succeed */
1009         if (nid == LNET_NID_ANY)
1010                 return 0;
1011
1012         return !count ? -ENOENT : 0;
1013 }
1014
1015 int kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1016 {
1017         struct libcfs_ioctl_data *data = arg;
1018         int rc = -EINVAL;
1019
1020         switch (cmd) {
1021         case IOC_LIBCFS_GET_PEER: {
1022                 lnet_nid_t nid = 0;
1023                 int count = 0;
1024
1025                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1026                                           &nid, &count);
1027                 data->ioc_nid   = nid;
1028                 data->ioc_count = count;
1029                 break;
1030         }
1031
1032         case IOC_LIBCFS_DEL_PEER: {
1033                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1034                 break;
1035         }
1036         case IOC_LIBCFS_GET_CONN: {
1037                 kib_conn_t *conn;
1038
1039                 rc = 0;
1040                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1041                 if (!conn) {
1042                         rc = -ENOENT;
1043                         break;
1044                 }
1045
1046                 LASSERT(conn->ibc_cmid);
1047                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1048                 if (!conn->ibc_cmid->route.path_rec)
1049                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1050                 else
1051                         data->ioc_u32[0] =
1052                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1053                 kiblnd_conn_decref(conn);
1054                 break;
1055         }
1056         case IOC_LIBCFS_CLOSE_CONNECTION: {
1057                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1058                 break;
1059         }
1060
1061         default:
1062                 break;
1063         }
1064
1065         return rc;
1066 }
1067
1068 void kiblnd_query(lnet_ni_t *ni, lnet_nid_t nid, unsigned long *when)
1069 {
1070         unsigned long last_alive = 0;
1071         unsigned long now = cfs_time_current();
1072         rwlock_t *glock = &kiblnd_data.kib_global_lock;
1073         kib_peer_t *peer;
1074         unsigned long flags;
1075
1076         read_lock_irqsave(glock, flags);
1077
1078         peer = kiblnd_find_peer_locked(nid);
1079         if (peer) {
1080                 LASSERT(peer->ibp_connecting > 0 || /* creating conns */
1081                          peer->ibp_accepting > 0 ||
1082                          !list_empty(&peer->ibp_conns));  /* active conn */
1083                 last_alive = peer->ibp_last_alive;
1084         }
1085
1086         read_unlock_irqrestore(glock, flags);
1087
1088         if (last_alive)
1089                 *when = last_alive;
1090
1091         /*
1092          * peer is not persistent in hash, trigger peer creation
1093          * and connection establishment with a NULL tx
1094          */
1095         if (!peer)
1096                 kiblnd_launch_tx(ni, NULL, nid);
1097
1098         CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1099                libcfs_nid2str(nid), peer,
1100                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1101 }
1102
1103 void kiblnd_free_pages(kib_pages_t *p)
1104 {
1105         int npages = p->ibp_npages;
1106         int i;
1107
1108         for (i = 0; i < npages; i++) {
1109                 if (p->ibp_pages[i])
1110                         __free_page(p->ibp_pages[i]);
1111         }
1112
1113         LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1114 }
1115
1116 int kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1117 {
1118         kib_pages_t *p;
1119         int i;
1120
1121         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1122                          offsetof(kib_pages_t, ibp_pages[npages]));
1123         if (!p) {
1124                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1125                 return -ENOMEM;
1126         }
1127
1128         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1129         p->ibp_npages = npages;
1130
1131         for (i = 0; i < npages; i++) {
1132                 p->ibp_pages[i] = alloc_pages_node(
1133                                     cfs_cpt_spread_node(lnet_cpt_table(), cpt),
1134                                     GFP_NOFS, 0);
1135                 if (!p->ibp_pages[i]) {
1136                         CERROR("Can't allocate page %d of %d\n", i, npages);
1137                         kiblnd_free_pages(p);
1138                         return -ENOMEM;
1139                 }
1140         }
1141
1142         *pp = p;
1143         return 0;
1144 }
1145
1146 void kiblnd_unmap_rx_descs(kib_conn_t *conn)
1147 {
1148         kib_rx_t *rx;
1149         int i;
1150
1151         LASSERT(conn->ibc_rxs);
1152         LASSERT(conn->ibc_hdev);
1153
1154         for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1155                 rx = &conn->ibc_rxs[i];
1156
1157                 LASSERT(rx->rx_nob >= 0); /* not posted */
1158
1159                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1160                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1161                                                           rx->rx_msgaddr),
1162                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1163         }
1164
1165         kiblnd_free_pages(conn->ibc_rx_pages);
1166
1167         conn->ibc_rx_pages = NULL;
1168 }
1169
1170 void kiblnd_map_rx_descs(kib_conn_t *conn)
1171 {
1172         kib_rx_t *rx;
1173         struct page *pg;
1174         int pg_off;
1175         int ipg;
1176         int i;
1177
1178         for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1179                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1180                 rx = &conn->ibc_rxs[i];
1181
1182                 rx->rx_conn = conn;
1183                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1184
1185                 rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1186                                                        rx->rx_msg,
1187                                                        IBLND_MSG_SIZE,
1188                                                        DMA_FROM_DEVICE);
1189                 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1190                                                   rx->rx_msgaddr));
1191                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1192
1193                 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1194                        i, rx->rx_msg, rx->rx_msgaddr,
1195                        (__u64)(page_to_phys(pg) + pg_off));
1196
1197                 pg_off += IBLND_MSG_SIZE;
1198                 LASSERT(pg_off <= PAGE_SIZE);
1199
1200                 if (pg_off == PAGE_SIZE) {
1201                         pg_off = 0;
1202                         ipg++;
1203                         LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version));
1204                 }
1205         }
1206 }
1207
1208 static void kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1209 {
1210         kib_hca_dev_t *hdev = tpo->tpo_hdev;
1211         kib_tx_t *tx;
1212         int i;
1213
1214         LASSERT(!tpo->tpo_pool.po_allocated);
1215
1216         if (!hdev)
1217                 return;
1218
1219         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1220                 tx = &tpo->tpo_tx_descs[i];
1221                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1222                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1223                                                           tx->tx_msgaddr),
1224                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1225         }
1226
1227         kiblnd_hdev_decref(hdev);
1228         tpo->tpo_hdev = NULL;
1229 }
1230
1231 static kib_hca_dev_t *kiblnd_current_hdev(kib_dev_t *dev)
1232 {
1233         kib_hca_dev_t *hdev;
1234         unsigned long flags;
1235         int i = 0;
1236
1237         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1238         while (dev->ibd_failover) {
1239                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1240                 if (!(i++ % 50))
1241                         CDEBUG(D_NET, "%s: Wait for failover\n",
1242                                dev->ibd_ifname);
1243                 schedule_timeout(cfs_time_seconds(1) / 100);
1244
1245                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1246         }
1247
1248         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1249         hdev = dev->ibd_hdev;
1250
1251         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1252
1253         return hdev;
1254 }
1255
1256 static void kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1257 {
1258         kib_pages_t *txpgs = tpo->tpo_tx_pages;
1259         kib_pool_t *pool = &tpo->tpo_pool;
1260         kib_net_t *net = pool->po_owner->ps_net;
1261         kib_dev_t *dev;
1262         struct page *page;
1263         kib_tx_t *tx;
1264         int page_offset;
1265         int ipage;
1266         int i;
1267
1268         LASSERT(net);
1269
1270         dev = net->ibn_dev;
1271
1272         /* pre-mapped messages are not bigger than 1 page */
1273         CLASSERT(IBLND_MSG_SIZE <= PAGE_SIZE);
1274
1275         /* No fancy arithmetic when we do the buffer calculations */
1276         CLASSERT(!(PAGE_SIZE % IBLND_MSG_SIZE));
1277
1278         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1279
1280         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1281                 page = txpgs->ibp_pages[ipage];
1282                 tx = &tpo->tpo_tx_descs[i];
1283
1284                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1285                                            page_offset);
1286
1287                 tx->tx_msgaddr = kiblnd_dma_map_single(
1288                         tpo->tpo_hdev->ibh_ibdev, tx->tx_msg,
1289                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1290                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1291                                                   tx->tx_msgaddr));
1292                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1293
1294                 list_add(&tx->tx_list, &pool->po_free_list);
1295
1296                 page_offset += IBLND_MSG_SIZE;
1297                 LASSERT(page_offset <= PAGE_SIZE);
1298
1299                 if (page_offset == PAGE_SIZE) {
1300                         page_offset = 0;
1301                         ipage++;
1302                         LASSERT(ipage <= txpgs->ibp_npages);
1303                 }
1304         }
1305 }
1306
1307 struct ib_mr *kiblnd_find_dma_mr(kib_hca_dev_t *hdev, __u64 addr, __u64 size)
1308 {
1309         __u64 index;
1310
1311         LASSERT(hdev->ibh_mrs[0]);
1312
1313         if (hdev->ibh_nmrs == 1)
1314                 return hdev->ibh_mrs[0];
1315
1316         index = addr >> hdev->ibh_mr_shift;
1317
1318         if (index <  hdev->ibh_nmrs &&
1319             index == ((addr + size - 1) >> hdev->ibh_mr_shift))
1320                 return hdev->ibh_mrs[index];
1321
1322         return NULL;
1323 }
1324
1325 struct ib_mr *kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd)
1326 {
1327         struct ib_mr *prev_mr;
1328         struct ib_mr *mr;
1329         int i;
1330
1331         LASSERT(hdev->ibh_mrs[0]);
1332
1333         if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1334             *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags)
1335                 return NULL;
1336
1337         if (hdev->ibh_nmrs == 1)
1338                 return hdev->ibh_mrs[0];
1339
1340         for (i = 0, mr = prev_mr = NULL;
1341              i < rd->rd_nfrags; i++) {
1342                 mr = kiblnd_find_dma_mr(hdev,
1343                                         rd->rd_frags[i].rf_addr,
1344                                         rd->rd_frags[i].rf_nob);
1345                 if (!prev_mr)
1346                         prev_mr = mr;
1347
1348                 if (!mr || prev_mr != mr) {
1349                         /* Can't covered by one single MR */
1350                         mr = NULL;
1351                         break;
1352                 }
1353         }
1354
1355         return mr;
1356 }
1357
1358 static void kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
1359 {
1360         LASSERT(!pool->fpo_map_count);
1361
1362         if (pool->fpo_fmr_pool)
1363                 ib_destroy_fmr_pool(pool->fpo_fmr_pool);
1364
1365         if (pool->fpo_hdev)
1366                 kiblnd_hdev_decref(pool->fpo_hdev);
1367
1368         LIBCFS_FREE(pool, sizeof(*pool));
1369 }
1370
1371 static void kiblnd_destroy_fmr_pool_list(struct list_head *head)
1372 {
1373         kib_fmr_pool_t *pool;
1374
1375         while (!list_empty(head)) {
1376                 pool = list_entry(head->next, kib_fmr_pool_t, fpo_list);
1377                 list_del(&pool->fpo_list);
1378                 kiblnd_destroy_fmr_pool(pool);
1379         }
1380 }
1381
1382 static int kiblnd_fmr_pool_size(int ncpts)
1383 {
1384         int size = *kiblnd_tunables.kib_fmr_pool_size / ncpts;
1385
1386         return max(IBLND_FMR_POOL, size);
1387 }
1388
1389 static int kiblnd_fmr_flush_trigger(int ncpts)
1390 {
1391         int size = *kiblnd_tunables.kib_fmr_flush_trigger / ncpts;
1392
1393         return max(IBLND_FMR_POOL_FLUSH, size);
1394 }
1395
1396 static int kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps,
1397                                   kib_fmr_pool_t **pp_fpo)
1398 {
1399         /* FMR pool for RDMA */
1400         kib_dev_t *dev = fps->fps_net->ibn_dev;
1401         kib_fmr_pool_t *fpo;
1402         struct ib_fmr_pool_param param = {
1403                 .max_pages_per_fmr = LNET_MAX_PAYLOAD / PAGE_SIZE,
1404                 .page_shift        = PAGE_SHIFT,
1405                 .access            = (IB_ACCESS_LOCAL_WRITE |
1406                                       IB_ACCESS_REMOTE_WRITE),
1407                 .pool_size         = fps->fps_pool_size,
1408                 .dirty_watermark   = fps->fps_flush_trigger,
1409                 .flush_function    = NULL,
1410                 .flush_arg         = NULL,
1411                 .cache             = !!*kiblnd_tunables.kib_fmr_cache};
1412         int rc;
1413
1414         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1415         if (!fpo)
1416                 return -ENOMEM;
1417
1418         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1419
1420         fpo->fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, &param);
1421         if (IS_ERR(fpo->fpo_fmr_pool)) {
1422                 rc = PTR_ERR(fpo->fpo_fmr_pool);
1423                 CERROR("Failed to create FMR pool: %d\n", rc);
1424
1425                 kiblnd_hdev_decref(fpo->fpo_hdev);
1426                 LIBCFS_FREE(fpo, sizeof(*fpo));
1427                 return rc;
1428         }
1429
1430         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1431         fpo->fpo_owner    = fps;
1432         *pp_fpo = fpo;
1433
1434         return 0;
1435 }
1436
1437 static void kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps,
1438                                     struct list_head *zombies)
1439 {
1440         if (!fps->fps_net) /* intialized? */
1441                 return;
1442
1443         spin_lock(&fps->fps_lock);
1444
1445         while (!list_empty(&fps->fps_pool_list)) {
1446                 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1447                                                  kib_fmr_pool_t, fpo_list);
1448                 fpo->fpo_failed = 1;
1449                 list_del(&fpo->fpo_list);
1450                 if (!fpo->fpo_map_count)
1451                         list_add(&fpo->fpo_list, zombies);
1452                 else
1453                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1454         }
1455
1456         spin_unlock(&fps->fps_lock);
1457 }
1458
1459 static void kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1460 {
1461         if (fps->fps_net) { /* initialized? */
1462                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1463                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1464         }
1465 }
1466
1467 static int kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt,
1468                                    kib_net_t *net, int pool_size,
1469                                    int flush_trigger)
1470 {
1471         kib_fmr_pool_t *fpo;
1472         int rc;
1473
1474         memset(fps, 0, sizeof(*fps));
1475
1476         fps->fps_net = net;
1477         fps->fps_cpt = cpt;
1478         fps->fps_pool_size = pool_size;
1479         fps->fps_flush_trigger = flush_trigger;
1480         spin_lock_init(&fps->fps_lock);
1481         INIT_LIST_HEAD(&fps->fps_pool_list);
1482         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1483
1484         rc = kiblnd_create_fmr_pool(fps, &fpo);
1485         if (!rc)
1486                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1487
1488         return rc;
1489 }
1490
1491 static int kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, unsigned long now)
1492 {
1493         if (fpo->fpo_map_count) /* still in use */
1494                 return 0;
1495         if (fpo->fpo_failed)
1496                 return 1;
1497         return cfs_time_aftereq(now, fpo->fpo_deadline);
1498 }
1499
1500 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1501 {
1502         LIST_HEAD(zombies);
1503         kib_fmr_pool_t *fpo = fmr->fmr_pool;
1504         kib_fmr_poolset_t *fps = fpo->fpo_owner;
1505         unsigned long now = cfs_time_current();
1506         kib_fmr_pool_t *tmp;
1507         int rc;
1508
1509         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1510         LASSERT(!rc);
1511
1512         if (status) {
1513                 rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
1514                 LASSERT(!rc);
1515         }
1516
1517         fmr->fmr_pool = NULL;
1518         fmr->fmr_pfmr = NULL;
1519
1520         spin_lock(&fps->fps_lock);
1521         fpo->fpo_map_count--;  /* decref the pool */
1522
1523         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1524                 /* the first pool is persistent */
1525                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1526                         continue;
1527
1528                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1529                         list_move(&fpo->fpo_list, &zombies);
1530                         fps->fps_version++;
1531                 }
1532         }
1533         spin_unlock(&fps->fps_lock);
1534
1535         if (!list_empty(&zombies))
1536                 kiblnd_destroy_fmr_pool_list(&zombies);
1537 }
1538
1539 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1540                         __u64 iov, kib_fmr_t *fmr)
1541 {
1542         struct ib_pool_fmr *pfmr;
1543         kib_fmr_pool_t *fpo;
1544         __u64 version;
1545         int rc;
1546
1547  again:
1548         spin_lock(&fps->fps_lock);
1549         version = fps->fps_version;
1550         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1551                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1552                 fpo->fpo_map_count++;
1553                 spin_unlock(&fps->fps_lock);
1554
1555                 pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
1556                                             pages, npages, iov);
1557                 if (likely(!IS_ERR(pfmr))) {
1558                         fmr->fmr_pool = fpo;
1559                         fmr->fmr_pfmr = pfmr;
1560                         return 0;
1561                 }
1562
1563                 spin_lock(&fps->fps_lock);
1564                 fpo->fpo_map_count--;
1565                 if (PTR_ERR(pfmr) != -EAGAIN) {
1566                         spin_unlock(&fps->fps_lock);
1567                         return PTR_ERR(pfmr);
1568                 }
1569
1570                 /* EAGAIN and ... */
1571                 if (version != fps->fps_version) {
1572                         spin_unlock(&fps->fps_lock);
1573                         goto again;
1574                 }
1575         }
1576
1577         if (fps->fps_increasing) {
1578                 spin_unlock(&fps->fps_lock);
1579                 CDEBUG(D_NET, "Another thread is allocating new FMR pool, waiting for her to complete\n");
1580                 schedule();
1581                 goto again;
1582         }
1583
1584         if (time_before(cfs_time_current(), fps->fps_next_retry)) {
1585                 /* someone failed recently */
1586                 spin_unlock(&fps->fps_lock);
1587                 return -EAGAIN;
1588         }
1589
1590         fps->fps_increasing = 1;
1591         spin_unlock(&fps->fps_lock);
1592
1593         CDEBUG(D_NET, "Allocate new FMR pool\n");
1594         rc = kiblnd_create_fmr_pool(fps, &fpo);
1595         spin_lock(&fps->fps_lock);
1596         fps->fps_increasing = 0;
1597         if (!rc) {
1598                 fps->fps_version++;
1599                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1600         } else {
1601                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1602         }
1603         spin_unlock(&fps->fps_lock);
1604
1605         goto again;
1606 }
1607
1608 static void kiblnd_fini_pool(kib_pool_t *pool)
1609 {
1610         LASSERT(list_empty(&pool->po_free_list));
1611         LASSERT(!pool->po_allocated);
1612
1613         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1614 }
1615
1616 static void kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1617 {
1618         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1619
1620         memset(pool, 0, sizeof(*pool));
1621         INIT_LIST_HEAD(&pool->po_free_list);
1622         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1623         pool->po_owner    = ps;
1624         pool->po_size     = size;
1625 }
1626
1627 static void kiblnd_destroy_pool_list(struct list_head *head)
1628 {
1629         kib_pool_t *pool;
1630
1631         while (!list_empty(head)) {
1632                 pool = list_entry(head->next, kib_pool_t, po_list);
1633                 list_del(&pool->po_list);
1634
1635                 LASSERT(pool->po_owner);
1636                 pool->po_owner->ps_pool_destroy(pool);
1637         }
1638 }
1639
1640 static void kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
1641 {
1642         if (!ps->ps_net) /* intialized? */
1643                 return;
1644
1645         spin_lock(&ps->ps_lock);
1646         while (!list_empty(&ps->ps_pool_list)) {
1647                 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
1648                                             kib_pool_t, po_list);
1649                 po->po_failed = 1;
1650                 list_del(&po->po_list);
1651                 if (!po->po_allocated)
1652                         list_add(&po->po_list, zombies);
1653                 else
1654                         list_add(&po->po_list, &ps->ps_failed_pool_list);
1655         }
1656         spin_unlock(&ps->ps_lock);
1657 }
1658
1659 static void kiblnd_fini_poolset(kib_poolset_t *ps)
1660 {
1661         if (ps->ps_net) { /* initialized? */
1662                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1663                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1664         }
1665 }
1666
1667 static int kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
1668                                kib_net_t *net, char *name, int size,
1669                                kib_ps_pool_create_t po_create,
1670                                kib_ps_pool_destroy_t po_destroy,
1671                                kib_ps_node_init_t nd_init,
1672                                kib_ps_node_fini_t nd_fini)
1673 {
1674         kib_pool_t *pool;
1675         int rc;
1676
1677         memset(ps, 0, sizeof(*ps));
1678
1679         ps->ps_cpt          = cpt;
1680         ps->ps_net          = net;
1681         ps->ps_pool_create  = po_create;
1682         ps->ps_pool_destroy = po_destroy;
1683         ps->ps_node_init    = nd_init;
1684         ps->ps_node_fini    = nd_fini;
1685         ps->ps_pool_size    = size;
1686         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
1687             >= sizeof(ps->ps_name))
1688                 return -E2BIG;
1689         spin_lock_init(&ps->ps_lock);
1690         INIT_LIST_HEAD(&ps->ps_pool_list);
1691         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1692
1693         rc = ps->ps_pool_create(ps, size, &pool);
1694         if (!rc)
1695                 list_add(&pool->po_list, &ps->ps_pool_list);
1696         else
1697                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1698
1699         return rc;
1700 }
1701
1702 static int kiblnd_pool_is_idle(kib_pool_t *pool, unsigned long now)
1703 {
1704         if (pool->po_allocated) /* still in use */
1705                 return 0;
1706         if (pool->po_failed)
1707                 return 1;
1708         return cfs_time_aftereq(now, pool->po_deadline);
1709 }
1710
1711 void kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
1712 {
1713         LIST_HEAD(zombies);
1714         kib_poolset_t *ps = pool->po_owner;
1715         kib_pool_t *tmp;
1716         unsigned long now = cfs_time_current();
1717
1718         spin_lock(&ps->ps_lock);
1719
1720         if (ps->ps_node_fini)
1721                 ps->ps_node_fini(pool, node);
1722
1723         LASSERT(pool->po_allocated > 0);
1724         list_add(node, &pool->po_free_list);
1725         pool->po_allocated--;
1726
1727         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1728                 /* the first pool is persistent */
1729                 if (ps->ps_pool_list.next == &pool->po_list)
1730                         continue;
1731
1732                 if (kiblnd_pool_is_idle(pool, now))
1733                         list_move(&pool->po_list, &zombies);
1734         }
1735         spin_unlock(&ps->ps_lock);
1736
1737         if (!list_empty(&zombies))
1738                 kiblnd_destroy_pool_list(&zombies);
1739 }
1740
1741 struct list_head *kiblnd_pool_alloc_node(kib_poolset_t *ps)
1742 {
1743         struct list_head *node;
1744         kib_pool_t *pool;
1745         int rc;
1746
1747  again:
1748         spin_lock(&ps->ps_lock);
1749         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1750                 if (list_empty(&pool->po_free_list))
1751                         continue;
1752
1753                 pool->po_allocated++;
1754                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1755                 node = pool->po_free_list.next;
1756                 list_del(node);
1757
1758                 if (ps->ps_node_init) {
1759                         /* still hold the lock */
1760                         ps->ps_node_init(pool, node);
1761                 }
1762                 spin_unlock(&ps->ps_lock);
1763                 return node;
1764         }
1765
1766         /* no available tx pool and ... */
1767         if (ps->ps_increasing) {
1768                 /* another thread is allocating a new pool */
1769                 spin_unlock(&ps->ps_lock);
1770                 CDEBUG(D_NET, "Another thread is allocating new %s pool, waiting for her to complete\n",
1771                        ps->ps_name);
1772                 schedule();
1773                 goto again;
1774         }
1775
1776         if (time_before(cfs_time_current(), ps->ps_next_retry)) {
1777                 /* someone failed recently */
1778                 spin_unlock(&ps->ps_lock);
1779                 return NULL;
1780         }
1781
1782         ps->ps_increasing = 1;
1783         spin_unlock(&ps->ps_lock);
1784
1785         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1786
1787         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1788
1789         spin_lock(&ps->ps_lock);
1790         ps->ps_increasing = 0;
1791         if (!rc) {
1792                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
1793         } else {
1794                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1795                 CERROR("Can't allocate new %s pool because out of memory\n",
1796                        ps->ps_name);
1797         }
1798         spin_unlock(&ps->ps_lock);
1799
1800         goto again;
1801 }
1802
1803 static void kiblnd_destroy_tx_pool(kib_pool_t *pool)
1804 {
1805         kib_tx_pool_t *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
1806         int i;
1807
1808         LASSERT(!pool->po_allocated);
1809
1810         if (tpo->tpo_tx_pages) {
1811                 kiblnd_unmap_tx_pool(tpo);
1812                 kiblnd_free_pages(tpo->tpo_tx_pages);
1813         }
1814
1815         if (!tpo->tpo_tx_descs)
1816                 goto out;
1817
1818         for (i = 0; i < pool->po_size; i++) {
1819                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1820
1821                 list_del(&tx->tx_list);
1822                 if (tx->tx_pages)
1823                         LIBCFS_FREE(tx->tx_pages,
1824                                     LNET_MAX_IOV *
1825                                     sizeof(*tx->tx_pages));
1826                 if (tx->tx_frags)
1827                         LIBCFS_FREE(tx->tx_frags,
1828                                     IBLND_MAX_RDMA_FRAGS *
1829                                             sizeof(*tx->tx_frags));
1830                 if (tx->tx_wrq)
1831                         LIBCFS_FREE(tx->tx_wrq,
1832                                     (1 + IBLND_MAX_RDMA_FRAGS) *
1833                                     sizeof(*tx->tx_wrq));
1834                 if (tx->tx_sge)
1835                         LIBCFS_FREE(tx->tx_sge,
1836                                     (1 + IBLND_MAX_RDMA_FRAGS) *
1837                                     sizeof(*tx->tx_sge));
1838                 if (tx->tx_rd)
1839                         LIBCFS_FREE(tx->tx_rd,
1840                                     offsetof(kib_rdma_desc_t,
1841                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
1842         }
1843
1844         LIBCFS_FREE(tpo->tpo_tx_descs,
1845                     pool->po_size * sizeof(kib_tx_t));
1846 out:
1847         kiblnd_fini_pool(pool);
1848         LIBCFS_FREE(tpo, sizeof(*tpo));
1849 }
1850
1851 static int kiblnd_tx_pool_size(int ncpts)
1852 {
1853         int ntx = *kiblnd_tunables.kib_ntx / ncpts;
1854
1855         return max(IBLND_TX_POOL, ntx);
1856 }
1857
1858 static int kiblnd_create_tx_pool(kib_poolset_t *ps, int size,
1859                                  kib_pool_t **pp_po)
1860 {
1861         int i;
1862         int npg;
1863         kib_pool_t *pool;
1864         kib_tx_pool_t *tpo;
1865
1866         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
1867         if (!tpo) {
1868                 CERROR("Failed to allocate TX pool\n");
1869                 return -ENOMEM;
1870         }
1871
1872         pool = &tpo->tpo_pool;
1873         kiblnd_init_pool(ps, pool, size);
1874         tpo->tpo_tx_descs = NULL;
1875         tpo->tpo_tx_pages = NULL;
1876
1877         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
1878         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg)) {
1879                 CERROR("Can't allocate tx pages: %d\n", npg);
1880                 LIBCFS_FREE(tpo, sizeof(*tpo));
1881                 return -ENOMEM;
1882         }
1883
1884         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
1885                          size * sizeof(kib_tx_t));
1886         if (!tpo->tpo_tx_descs) {
1887                 CERROR("Can't allocate %d tx descriptors\n", size);
1888                 ps->ps_pool_destroy(pool);
1889                 return -ENOMEM;
1890         }
1891
1892         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
1893
1894         for (i = 0; i < size; i++) {
1895                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1896
1897                 tx->tx_pool = tpo;
1898                 if (ps->ps_net->ibn_fmr_ps) {
1899                         LIBCFS_CPT_ALLOC(tx->tx_pages,
1900                                          lnet_cpt_table(), ps->ps_cpt,
1901                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
1902                         if (!tx->tx_pages)
1903                                 break;
1904                 }
1905
1906                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
1907                                  IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags));
1908                 if (!tx->tx_frags)
1909                         break;
1910
1911                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
1912
1913                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
1914                                  (1 + IBLND_MAX_RDMA_FRAGS) *
1915                                  sizeof(*tx->tx_wrq));
1916                 if (!tx->tx_wrq)
1917                         break;
1918
1919                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
1920                                  (1 + IBLND_MAX_RDMA_FRAGS) *
1921                                  sizeof(*tx->tx_sge));
1922                 if (!tx->tx_sge)
1923                         break;
1924
1925                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
1926                                  offsetof(kib_rdma_desc_t,
1927                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
1928                 if (!tx->tx_rd)
1929                         break;
1930         }
1931
1932         if (i == size) {
1933                 kiblnd_map_tx_pool(tpo);
1934                 *pp_po = pool;
1935                 return 0;
1936         }
1937
1938         ps->ps_pool_destroy(pool);
1939         return -ENOMEM;
1940 }
1941
1942 static void kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
1943 {
1944         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
1945                                              tps_poolset);
1946         kib_tx_t *tx  = list_entry(node, kib_tx_t, tx_list);
1947
1948         tx->tx_cookie = tps->tps_next_tx_cookie++;
1949 }
1950
1951 static void kiblnd_net_fini_pools(kib_net_t *net)
1952 {
1953         int i;
1954
1955         cfs_cpt_for_each(i, lnet_cpt_table()) {
1956                 kib_tx_poolset_t *tps;
1957                 kib_fmr_poolset_t *fps;
1958
1959                 if (net->ibn_tx_ps) {
1960                         tps = net->ibn_tx_ps[i];
1961                         kiblnd_fini_poolset(&tps->tps_poolset);
1962                 }
1963
1964                 if (net->ibn_fmr_ps) {
1965                         fps = net->ibn_fmr_ps[i];
1966                         kiblnd_fini_fmr_poolset(fps);
1967                 }
1968         }
1969
1970         if (net->ibn_tx_ps) {
1971                 cfs_percpt_free(net->ibn_tx_ps);
1972                 net->ibn_tx_ps = NULL;
1973         }
1974
1975         if (net->ibn_fmr_ps) {
1976                 cfs_percpt_free(net->ibn_fmr_ps);
1977                 net->ibn_fmr_ps = NULL;
1978         }
1979 }
1980
1981 static int kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts)
1982 {
1983         unsigned long flags;
1984         int cpt;
1985         int             rc = 0;
1986         int i;
1987
1988         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1989         if (!*kiblnd_tunables.kib_map_on_demand &&
1990             net->ibn_dev->ibd_hdev->ibh_nmrs == 1) {
1991                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1992                 goto create_tx_pool;
1993         }
1994
1995         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1996
1997         if (*kiblnd_tunables.kib_fmr_pool_size <
1998             *kiblnd_tunables.kib_ntx / 4) {
1999                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2000                        *kiblnd_tunables.kib_fmr_pool_size,
2001                        *kiblnd_tunables.kib_ntx / 4);
2002                 rc = -EINVAL;
2003                 goto failed;
2004         }
2005
2006         /*
2007          * TX pool must be created later than FMR, see LU-2268
2008          * for details
2009          */
2010         LASSERT(!net->ibn_tx_ps);
2011
2012         /*
2013          * premapping can fail if ibd_nmr > 1, so we always create
2014          * FMR pool and map-on-demand if premapping failed
2015          */
2016
2017         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2018                                            sizeof(kib_fmr_poolset_t));
2019         if (!net->ibn_fmr_ps) {
2020                 CERROR("Failed to allocate FMR pool array\n");
2021                 rc = -ENOMEM;
2022                 goto failed;
2023         }
2024
2025         for (i = 0; i < ncpts; i++) {
2026                 cpt = !cpts ? i : cpts[i];
2027                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, net,
2028                                              kiblnd_fmr_pool_size(ncpts),
2029                                              kiblnd_fmr_flush_trigger(ncpts));
2030                 if (rc == -ENOSYS && !i) /* no FMR */
2031                         break;
2032
2033                 if (rc) { /* a real error */
2034                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2035                                cpt, rc);
2036                         goto failed;
2037                 }
2038         }
2039
2040         if (i > 0) {
2041                 LASSERT(i == ncpts);
2042                 goto create_tx_pool;
2043         }
2044
2045         cfs_percpt_free(net->ibn_fmr_ps);
2046         net->ibn_fmr_ps = NULL;
2047
2048         CWARN("Device does not support FMR\n");
2049                         goto failed;
2050
2051  create_tx_pool:
2052         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2053                                           sizeof(kib_tx_poolset_t));
2054         if (!net->ibn_tx_ps) {
2055                 CERROR("Failed to allocate tx pool array\n");
2056                 rc = -ENOMEM;
2057                 goto failed;
2058         }
2059
2060         for (i = 0; i < ncpts; i++) {
2061                 cpt = !cpts ? i : cpts[i];
2062                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2063                                          cpt, net, "TX",
2064                                          kiblnd_tx_pool_size(ncpts),
2065                                          kiblnd_create_tx_pool,
2066                                          kiblnd_destroy_tx_pool,
2067                                          kiblnd_tx_init, NULL);
2068                 if (rc) {
2069                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2070                                cpt, rc);
2071                         goto failed;
2072                 }
2073         }
2074
2075         return 0;
2076  failed:
2077         kiblnd_net_fini_pools(net);
2078         LASSERT(rc);
2079         return rc;
2080 }
2081
2082 static int kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2083 {
2084         /*
2085          * It's safe to assume a HCA can handle a page size
2086          * matching that of the native system
2087          */
2088         hdev->ibh_page_shift = PAGE_SHIFT;
2089         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2090         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2091
2092         hdev->ibh_mr_size = hdev->ibh_ibdev->attrs.max_mr_size;
2093         if (hdev->ibh_mr_size == ~0ULL) {
2094                 hdev->ibh_mr_shift = 64;
2095                 return 0;
2096         }
2097
2098         for (hdev->ibh_mr_shift = 0;
2099              hdev->ibh_mr_shift < 64; hdev->ibh_mr_shift++) {
2100                 if (hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) ||
2101                     hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) - 1)
2102                         return 0;
2103         }
2104
2105         CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2106         return -EINVAL;
2107 }
2108
2109 static void kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2110 {
2111         int i;
2112
2113         if (!hdev->ibh_nmrs || !hdev->ibh_mrs)
2114                 return;
2115
2116         for (i = 0; i < hdev->ibh_nmrs; i++) {
2117                 if (!hdev->ibh_mrs[i])
2118                         break;
2119
2120                 ib_dereg_mr(hdev->ibh_mrs[i]);
2121         }
2122
2123         LIBCFS_FREE(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2124         hdev->ibh_mrs  = NULL;
2125         hdev->ibh_nmrs = 0;
2126 }
2127
2128 void kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2129 {
2130         kiblnd_hdev_cleanup_mrs(hdev);
2131
2132         if (hdev->ibh_pd)
2133                 ib_dealloc_pd(hdev->ibh_pd);
2134
2135         if (hdev->ibh_cmid)
2136                 rdma_destroy_id(hdev->ibh_cmid);
2137
2138         LIBCFS_FREE(hdev, sizeof(*hdev));
2139 }
2140
2141 static int kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2142 {
2143         struct ib_mr *mr;
2144         int rc;
2145         int acflags = IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE;
2146
2147         rc = kiblnd_hdev_get_attr(hdev);
2148         if (rc)
2149                 return rc;
2150
2151         LIBCFS_ALLOC(hdev->ibh_mrs, 1 * sizeof(*hdev->ibh_mrs));
2152         if (!hdev->ibh_mrs) {
2153                 CERROR("Failed to allocate MRs table\n");
2154                 return -ENOMEM;
2155         }
2156
2157         hdev->ibh_mrs[0] = NULL;
2158         hdev->ibh_nmrs   = 1;
2159
2160         mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2161         if (IS_ERR(mr)) {
2162                 CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
2163                 kiblnd_hdev_cleanup_mrs(hdev);
2164                 return PTR_ERR(mr);
2165         }
2166
2167         hdev->ibh_mrs[0] = mr;
2168
2169         return 0;
2170 }
2171
2172 /* DUMMY */
2173 static int kiblnd_dummy_callback(struct rdma_cm_id *cmid,
2174                                  struct rdma_cm_event *event)
2175 {
2176         return 0;
2177 }
2178
2179 static int kiblnd_dev_need_failover(kib_dev_t *dev)
2180 {
2181         struct rdma_cm_id *cmid;
2182         struct sockaddr_in srcaddr;
2183         struct sockaddr_in dstaddr;
2184         int rc;
2185
2186         if (!dev->ibd_hdev || /* initializing */
2187             !dev->ibd_hdev->ibh_cmid || /* listener is dead */
2188             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2189                 return 1;
2190
2191         /*
2192          * XXX: it's UGLY, but I don't have better way to find
2193          * ib-bonding HCA failover because:
2194          *
2195          * a. no reliable CM event for HCA failover...
2196          * b. no OFED API to get ib_device for current net_device...
2197          *
2198          * We have only two choices at this point:
2199          *
2200          * a. rdma_bind_addr(), it will conflict with listener cmid
2201          * b. rdma_resolve_addr() to zero addr
2202          */
2203         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2204                                      IB_QPT_RC);
2205         if (IS_ERR(cmid)) {
2206                 rc = PTR_ERR(cmid);
2207                 CERROR("Failed to create cmid for failover: %d\n", rc);
2208                 return rc;
2209         }
2210
2211         memset(&srcaddr, 0, sizeof(srcaddr));
2212         srcaddr.sin_family = AF_INET;
2213         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2214
2215         memset(&dstaddr, 0, sizeof(dstaddr));
2216         dstaddr.sin_family = AF_INET;
2217         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2218                                (struct sockaddr *)&dstaddr, 1);
2219         if (rc || !cmid->device) {
2220                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2221                        dev->ibd_ifname, &dev->ibd_ifip,
2222                        cmid->device, rc);
2223                 rdma_destroy_id(cmid);
2224                 return rc;
2225         }
2226
2227         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2228         rdma_destroy_id(cmid);
2229
2230         return rc;
2231 }
2232
2233 int kiblnd_dev_failover(kib_dev_t *dev)
2234 {
2235         LIST_HEAD(zombie_tpo);
2236         LIST_HEAD(zombie_ppo);
2237         LIST_HEAD(zombie_fpo);
2238         struct rdma_cm_id *cmid  = NULL;
2239         kib_hca_dev_t *hdev  = NULL;
2240         struct ib_pd *pd;
2241         kib_net_t *net;
2242         struct sockaddr_in addr;
2243         unsigned long flags;
2244         int rc = 0;
2245         int i;
2246
2247         LASSERT(*kiblnd_tunables.kib_dev_failover > 1 ||
2248                 dev->ibd_can_failover || !dev->ibd_hdev);
2249
2250         rc = kiblnd_dev_need_failover(dev);
2251         if (rc <= 0)
2252                 goto out;
2253
2254         if (dev->ibd_hdev &&
2255             dev->ibd_hdev->ibh_cmid) {
2256                 /*
2257                  * XXX it's not good to close old listener at here,
2258                  * because we can fail to create new listener.
2259                  * But we have to close it now, otherwise rdma_bind_addr
2260                  * will return EADDRINUSE... How crap!
2261                  */
2262                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2263
2264                 cmid = dev->ibd_hdev->ibh_cmid;
2265                 /*
2266                  * make next schedule of kiblnd_dev_need_failover()
2267                  * return 1 for me
2268                  */
2269                 dev->ibd_hdev->ibh_cmid  = NULL;
2270                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2271
2272                 rdma_destroy_id(cmid);
2273         }
2274
2275         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2276                                      IB_QPT_RC);
2277         if (IS_ERR(cmid)) {
2278                 rc = PTR_ERR(cmid);
2279                 CERROR("Failed to create cmid for failover: %d\n", rc);
2280                 goto out;
2281         }
2282
2283         memset(&addr, 0, sizeof(addr));
2284         addr.sin_family      = AF_INET;
2285         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2286         addr.sin_port   = htons(*kiblnd_tunables.kib_service);
2287
2288         /* Bind to failover device or port */
2289         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2290         if (rc || !cmid->device) {
2291                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2292                        dev->ibd_ifname, &dev->ibd_ifip,
2293                        cmid->device, rc);
2294                 rdma_destroy_id(cmid);
2295                 goto out;
2296         }
2297
2298         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2299         if (!hdev) {
2300                 CERROR("Failed to allocate kib_hca_dev\n");
2301                 rdma_destroy_id(cmid);
2302                 rc = -ENOMEM;
2303                 goto out;
2304         }
2305
2306         atomic_set(&hdev->ibh_ref, 1);
2307         hdev->ibh_dev   = dev;
2308         hdev->ibh_cmid  = cmid;
2309         hdev->ibh_ibdev = cmid->device;
2310
2311         pd = ib_alloc_pd(cmid->device);
2312         if (IS_ERR(pd)) {
2313                 rc = PTR_ERR(pd);
2314                 CERROR("Can't allocate PD: %d\n", rc);
2315                 goto out;
2316         }
2317
2318         hdev->ibh_pd = pd;
2319
2320         rc = rdma_listen(cmid, 0);
2321         if (rc) {
2322                 CERROR("Can't start new listener: %d\n", rc);
2323                 goto out;
2324         }
2325
2326         rc = kiblnd_hdev_setup_mrs(hdev);
2327         if (rc) {
2328                 CERROR("Can't setup device: %d\n", rc);
2329                 goto out;
2330         }
2331
2332         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2333
2334         swap(dev->ibd_hdev, hdev); /* take over the refcount */
2335
2336         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2337                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2338                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2339                                             &zombie_tpo);
2340
2341                         if (net->ibn_fmr_ps)
2342                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2343                                                         &zombie_fpo);
2344                 }
2345         }
2346
2347         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2348  out:
2349         if (!list_empty(&zombie_tpo))
2350                 kiblnd_destroy_pool_list(&zombie_tpo);
2351         if (!list_empty(&zombie_ppo))
2352                 kiblnd_destroy_pool_list(&zombie_ppo);
2353         if (!list_empty(&zombie_fpo))
2354                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2355         if (hdev)
2356                 kiblnd_hdev_decref(hdev);
2357
2358         if (rc)
2359                 dev->ibd_failed_failover++;
2360         else
2361                 dev->ibd_failed_failover = 0;
2362
2363         return rc;
2364 }
2365
2366 void kiblnd_destroy_dev(kib_dev_t *dev)
2367 {
2368         LASSERT(!dev->ibd_nnets);
2369         LASSERT(list_empty(&dev->ibd_nets));
2370
2371         list_del(&dev->ibd_fail_list);
2372         list_del(&dev->ibd_list);
2373
2374         if (dev->ibd_hdev)
2375                 kiblnd_hdev_decref(dev->ibd_hdev);
2376
2377         LIBCFS_FREE(dev, sizeof(*dev));
2378 }
2379
2380 static kib_dev_t *kiblnd_create_dev(char *ifname)
2381 {
2382         struct net_device *netdev;
2383         kib_dev_t *dev;
2384         __u32 netmask;
2385         __u32 ip;
2386         int up;
2387         int rc;
2388
2389         rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2390         if (rc) {
2391                 CERROR("Can't query IPoIB interface %s: %d\n",
2392                        ifname, rc);
2393                 return NULL;
2394         }
2395
2396         if (!up) {
2397                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2398                 return NULL;
2399         }
2400
2401         LIBCFS_ALLOC(dev, sizeof(*dev));
2402         if (!dev)
2403                 return NULL;
2404
2405         netdev = dev_get_by_name(&init_net, ifname);
2406         if (!netdev) {
2407                 dev->ibd_can_failover = 0;
2408         } else {
2409                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2410                 dev_put(netdev);
2411         }
2412
2413         INIT_LIST_HEAD(&dev->ibd_nets);
2414         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2415         INIT_LIST_HEAD(&dev->ibd_fail_list);
2416         dev->ibd_ifip = ip;
2417         strcpy(&dev->ibd_ifname[0], ifname);
2418
2419         /* initialize the device */
2420         rc = kiblnd_dev_failover(dev);
2421         if (rc) {
2422                 CERROR("Can't initialize device: %d\n", rc);
2423                 LIBCFS_FREE(dev, sizeof(*dev));
2424                 return NULL;
2425         }
2426
2427         list_add_tail(&dev->ibd_list, &kiblnd_data.kib_devs);
2428         return dev;
2429 }
2430
2431 static void kiblnd_base_shutdown(void)
2432 {
2433         struct kib_sched_info *sched;
2434         int i;
2435
2436         LASSERT(list_empty(&kiblnd_data.kib_devs));
2437
2438         switch (kiblnd_data.kib_init) {
2439         default:
2440                 LBUG();
2441
2442         case IBLND_INIT_ALL:
2443         case IBLND_INIT_DATA:
2444                 LASSERT(kiblnd_data.kib_peers);
2445                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2446                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2447                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2448                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2449
2450                 /* flag threads to terminate; wake and wait for them to die */
2451                 kiblnd_data.kib_shutdown = 1;
2452
2453                 /*
2454                  * NB: we really want to stop scheduler threads net by net
2455                  * instead of the whole module, this should be improved
2456                  * with dynamic configuration LNet
2457                  */
2458                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2459                         wake_up_all(&sched->ibs_waitq);
2460
2461                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2462                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2463
2464                 i = 2;
2465                 while (atomic_read(&kiblnd_data.kib_nthreads)) {
2466                         i++;
2467                         /* power of 2 ? */
2468                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2469                                "Waiting for %d threads to terminate\n",
2470                                atomic_read(&kiblnd_data.kib_nthreads));
2471                         set_current_state(TASK_UNINTERRUPTIBLE);
2472                         schedule_timeout(cfs_time_seconds(1));
2473                 }
2474
2475                 /* fall through */
2476
2477         case IBLND_INIT_NOTHING:
2478                 break;
2479         }
2480
2481         if (kiblnd_data.kib_peers) {
2482                 LIBCFS_FREE(kiblnd_data.kib_peers,
2483                             sizeof(struct list_head) *
2484                             kiblnd_data.kib_peer_hash_size);
2485         }
2486
2487         if (kiblnd_data.kib_scheds)
2488                 cfs_percpt_free(kiblnd_data.kib_scheds);
2489
2490         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2491         module_put(THIS_MODULE);
2492 }
2493
2494 void kiblnd_shutdown(lnet_ni_t *ni)
2495 {
2496         kib_net_t *net = ni->ni_data;
2497         rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2498         int i;
2499         unsigned long flags;
2500
2501         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2502
2503         if (!net)
2504                 goto out;
2505
2506         write_lock_irqsave(g_lock, flags);
2507         net->ibn_shutdown = 1;
2508         write_unlock_irqrestore(g_lock, flags);
2509
2510         switch (net->ibn_init) {
2511         default:
2512                 LBUG();
2513
2514         case IBLND_INIT_ALL:
2515                 /* nuke all existing peers within this net */
2516                 kiblnd_del_peer(ni, LNET_NID_ANY);
2517
2518                 /* Wait for all peer state to clean up */
2519                 i = 2;
2520                 while (atomic_read(&net->ibn_npeers)) {
2521                         i++;
2522                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */
2523                                "%s: waiting for %d peers to disconnect\n",
2524                                libcfs_nid2str(ni->ni_nid),
2525                                atomic_read(&net->ibn_npeers));
2526                         set_current_state(TASK_UNINTERRUPTIBLE);
2527                         schedule_timeout(cfs_time_seconds(1));
2528                 }
2529
2530                 kiblnd_net_fini_pools(net);
2531
2532                 write_lock_irqsave(g_lock, flags);
2533                 LASSERT(net->ibn_dev->ibd_nnets > 0);
2534                 net->ibn_dev->ibd_nnets--;
2535                 list_del(&net->ibn_list);
2536                 write_unlock_irqrestore(g_lock, flags);
2537
2538                 /* fall through */
2539
2540         case IBLND_INIT_NOTHING:
2541                 LASSERT(!atomic_read(&net->ibn_nconns));
2542
2543                 if (net->ibn_dev && !net->ibn_dev->ibd_nnets)
2544                         kiblnd_destroy_dev(net->ibn_dev);
2545
2546                 break;
2547         }
2548
2549         net->ibn_init = IBLND_INIT_NOTHING;
2550         ni->ni_data = NULL;
2551
2552         LIBCFS_FREE(net, sizeof(*net));
2553
2554 out:
2555         if (list_empty(&kiblnd_data.kib_devs))
2556                 kiblnd_base_shutdown();
2557 }
2558
2559 static int kiblnd_base_startup(void)
2560 {
2561         struct kib_sched_info *sched;
2562         int rc;
2563         int i;
2564
2565         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2566
2567         try_module_get(THIS_MODULE);
2568         /* zero pointers, flags etc */
2569         memset(&kiblnd_data, 0, sizeof(kiblnd_data));
2570
2571         rwlock_init(&kiblnd_data.kib_global_lock);
2572
2573         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2574         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2575
2576         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2577         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2578                      sizeof(struct list_head) * kiblnd_data.kib_peer_hash_size);
2579         if (!kiblnd_data.kib_peers)
2580                 goto failed;
2581         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2582                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2583
2584         spin_lock_init(&kiblnd_data.kib_connd_lock);
2585         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2586         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2587         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2588         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
2589
2590         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2591                                                   sizeof(*sched));
2592         if (!kiblnd_data.kib_scheds)
2593                 goto failed;
2594
2595         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2596                 int nthrs;
2597
2598                 spin_lock_init(&sched->ibs_lock);
2599                 INIT_LIST_HEAD(&sched->ibs_conns);
2600                 init_waitqueue_head(&sched->ibs_waitq);
2601
2602                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2603                 if (*kiblnd_tunables.kib_nscheds > 0) {
2604                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
2605                 } else {
2606                         /*
2607                          * max to half of CPUs, another half is reserved for
2608                          * upper layer modules
2609                          */
2610                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2611                 }
2612
2613                 sched->ibs_nthreads_max = nthrs;
2614                 sched->ibs_cpt = i;
2615         }
2616
2617         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2618
2619         /* lists/ptrs/locks initialised */
2620         kiblnd_data.kib_init = IBLND_INIT_DATA;
2621         /*****************************************************/
2622
2623         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
2624         if (rc) {
2625                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2626                 goto failed;
2627         }
2628
2629         if (*kiblnd_tunables.kib_dev_failover)
2630                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
2631                                          "kiblnd_failover");
2632
2633         if (rc) {
2634                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
2635                 goto failed;
2636         }
2637
2638         /* flag everything initialised */
2639         kiblnd_data.kib_init = IBLND_INIT_ALL;
2640         /*****************************************************/
2641
2642         return 0;
2643
2644  failed:
2645         kiblnd_base_shutdown();
2646         return -ENETDOWN;
2647 }
2648
2649 static int kiblnd_start_schedulers(struct kib_sched_info *sched)
2650 {
2651         int rc = 0;
2652         int nthrs;
2653         int i;
2654
2655         if (!sched->ibs_nthreads) {
2656                 if (*kiblnd_tunables.kib_nscheds > 0) {
2657                         nthrs = sched->ibs_nthreads_max;
2658                 } else {
2659                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
2660                                                sched->ibs_cpt);
2661                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2662                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
2663                 }
2664         } else {
2665                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
2666                 /* increase one thread if there is new interface */
2667                 nthrs = sched->ibs_nthreads < sched->ibs_nthreads_max;
2668         }
2669
2670         for (i = 0; i < nthrs; i++) {
2671                 long id;
2672                 char name[20];
2673
2674                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
2675                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
2676                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
2677                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
2678                 if (!rc)
2679                         continue;
2680
2681                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2682                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
2683                 break;
2684         }
2685
2686         sched->ibs_nthreads += i;
2687         return rc;
2688 }
2689
2690 static int kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts,
2691                                     int ncpts)
2692 {
2693         int cpt;
2694         int rc;
2695         int i;
2696
2697         for (i = 0; i < ncpts; i++) {
2698                 struct kib_sched_info *sched;
2699
2700                 cpt = !cpts ? i : cpts[i];
2701                 sched = kiblnd_data.kib_scheds[cpt];
2702
2703                 if (!newdev && sched->ibs_nthreads > 0)
2704                         continue;
2705
2706                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
2707                 if (rc) {
2708                         CERROR("Failed to start scheduler threads for %s\n",
2709                                dev->ibd_ifname);
2710                         return rc;
2711                 }
2712         }
2713         return 0;
2714 }
2715
2716 static kib_dev_t *kiblnd_dev_search(char *ifname)
2717 {
2718         kib_dev_t *alias = NULL;
2719         kib_dev_t *dev;
2720         char *colon;
2721         char *colon2;
2722
2723         colon = strchr(ifname, ':');
2724         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
2725                 if (!strcmp(&dev->ibd_ifname[0], ifname))
2726                         return dev;
2727
2728                 if (alias)
2729                         continue;
2730
2731                 colon2 = strchr(dev->ibd_ifname, ':');
2732                 if (colon)
2733                         *colon = 0;
2734                 if (colon2)
2735                         *colon2 = 0;
2736
2737                 if (!strcmp(&dev->ibd_ifname[0], ifname))
2738                         alias = dev;
2739
2740                 if (colon)
2741                         *colon = ':';
2742                 if (colon2)
2743                         *colon2 = ':';
2744         }
2745         return alias;
2746 }
2747
2748 int kiblnd_startup(lnet_ni_t *ni)
2749 {
2750         char *ifname;
2751         kib_dev_t *ibdev = NULL;
2752         kib_net_t *net;
2753         struct timespec64 tv;
2754         unsigned long flags;
2755         int rc;
2756         int newdev;
2757
2758         LASSERT(ni->ni_lnd == &the_o2iblnd);
2759
2760         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
2761                 rc = kiblnd_base_startup();
2762                 if (rc)
2763                         return rc;
2764         }
2765
2766         LIBCFS_ALLOC(net, sizeof(*net));
2767         ni->ni_data = net;
2768         if (!net)
2769                 goto net_failed;
2770
2771         ktime_get_real_ts64(&tv);
2772         net->ibn_incarnation = tv.tv_sec * USEC_PER_SEC +
2773                                tv.tv_nsec / NSEC_PER_USEC;
2774
2775         ni->ni_peertimeout    = *kiblnd_tunables.kib_peertimeout;
2776         ni->ni_maxtxcredits   = *kiblnd_tunables.kib_credits;
2777         ni->ni_peertxcredits  = *kiblnd_tunables.kib_peertxcredits;
2778         ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
2779
2780         if (ni->ni_interfaces[0]) {
2781                 /* Use the IPoIB interface specified in 'networks=' */
2782
2783                 CLASSERT(LNET_MAX_INTERFACES > 1);
2784                 if (ni->ni_interfaces[1]) {
2785                         CERROR("Multiple interfaces not supported\n");
2786                         goto failed;
2787                 }
2788
2789                 ifname = ni->ni_interfaces[0];
2790         } else {
2791                 ifname = *kiblnd_tunables.kib_default_ipif;
2792         }
2793
2794         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
2795                 CERROR("IPoIB interface name too long: %s\n", ifname);
2796                 goto failed;
2797         }
2798
2799         ibdev = kiblnd_dev_search(ifname);
2800
2801         newdev = !ibdev;
2802         /* hmm...create kib_dev even for alias */
2803         if (!ibdev || strcmp(&ibdev->ibd_ifname[0], ifname))
2804                 ibdev = kiblnd_create_dev(ifname);
2805
2806         if (!ibdev)
2807                 goto failed;
2808
2809         net->ibn_dev = ibdev;
2810         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
2811
2812         rc = kiblnd_dev_start_threads(ibdev, newdev,
2813                                       ni->ni_cpts, ni->ni_ncpts);
2814         if (rc)
2815                 goto failed;
2816
2817         rc = kiblnd_net_init_pools(net, ni->ni_cpts, ni->ni_ncpts);
2818         if (rc) {
2819                 CERROR("Failed to initialize NI pools: %d\n", rc);
2820                 goto failed;
2821         }
2822
2823         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2824         ibdev->ibd_nnets++;
2825         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
2826         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2827
2828         net->ibn_init = IBLND_INIT_ALL;
2829
2830         return 0;
2831
2832 failed:
2833         if (!net->ibn_dev && ibdev)
2834                 kiblnd_destroy_dev(ibdev);
2835
2836 net_failed:
2837         kiblnd_shutdown(ni);
2838
2839         CDEBUG(D_NET, "kiblnd_startup failed\n");
2840         return -ENETDOWN;
2841 }
2842
2843 static void __exit ko2iblnd_exit(void)
2844 {
2845         lnet_unregister_lnd(&the_o2iblnd);
2846 }
2847
2848 static int __init ko2iblnd_init(void)
2849 {
2850         int rc;
2851
2852         CLASSERT(sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
2853         CLASSERT(offsetof(kib_msg_t,
2854                           ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
2855                           <= IBLND_MSG_SIZE);
2856         CLASSERT(offsetof(kib_msg_t,
2857                           ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
2858                           <= IBLND_MSG_SIZE);
2859
2860         rc = kiblnd_tunables_init();
2861         if (rc)
2862                 return rc;
2863
2864         lnet_register_lnd(&the_o2iblnd);
2865
2866         return 0;
2867 }
2868
2869 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2870 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
2871 MODULE_VERSION("2.7.0");
2872 MODULE_LICENSE("GPL");
2873
2874 module_init(ko2iblnd_init);
2875 module_exit(ko2iblnd_exit);