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