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 = kiblnd_cfg_rdma_frags(peer->ibp_ni);
 339         peer->ibp_queue_depth = ni->ni_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(kiblnd_peer_idle(peer));
 368         LASSERT(list_empty(&peer->ibp_tx_queue));
 369
 370         LIBCFS_FREE(peer, sizeof(*peer));
 371
 372         /*
 373          * NB a peer's connections keep a reference on their peer until
 374          * they are destroyed, so we can be assured that _all_ state to do
 375          * with this peer has been cleaned up when its refcount drops to
 376          * zero.
 377          */
 378         atomic_dec(&net->ibn_npeers);
 379 }
 380
 381 kib_peer_t *kiblnd_find_peer_locked(lnet_nid_t nid)
 382 {
 383         /*
 384          * the caller is responsible for accounting the additional reference
 385          * that this creates
 386          */
 387         struct list_head *peer_list = kiblnd_nid2peerlist(nid);
 388         struct list_head *tmp;
 389         kib_peer_t *peer;
 390
 391         list_for_each(tmp, peer_list) {
 392                 peer = list_entry(tmp, kib_peer_t, ibp_list);
 393                 LASSERT(!kiblnd_peer_idle(peer));
 394
 395                 if (peer->ibp_nid != nid)
 396                         continue;
 397
 398                 CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n",
 399                        peer, libcfs_nid2str(nid),
 400                        atomic_read(&peer->ibp_refcount),
 401                        peer->ibp_version);
 402                 return peer;
 403         }
 404         return NULL;
 405 }
 406
 407 void kiblnd_unlink_peer_locked(kib_peer_t *peer)
 408 {
 409         LASSERT(list_empty(&peer->ibp_conns));
 410
 411         LASSERT(kiblnd_peer_active(peer));
 412         list_del_init(&peer->ibp_list);
 413         /* lose peerlist's ref */
 414         kiblnd_peer_decref(peer);
 415 }
 416
 417 static int kiblnd_get_peer_info(lnet_ni_t *ni, int index,
 418                                 lnet_nid_t *nidp, int *count)
 419 {
 420         kib_peer_t *peer;
 421         struct list_head *ptmp;
 422         int i;
 423         unsigned long flags;
 424
 425         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 426
 427         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 428                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 429                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 430                         LASSERT(!kiblnd_peer_idle(peer));
 431
 432                         if (peer->ibp_ni != ni)
 433                                 continue;
 434
 435                         if (index-- > 0)
 436                                 continue;
 437
 438                         *nidp = peer->ibp_nid;
 439                         *count = atomic_read(&peer->ibp_refcount);
 440
 441                         read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 442                                                flags);
 443                         return 0;
 444                 }
 445         }
 446
 447         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 448         return -ENOENT;
 449 }
 450
 451 static void kiblnd_del_peer_locked(kib_peer_t *peer)
 452 {
 453         struct list_head *ctmp;
 454         struct list_head *cnxt;
 455         kib_conn_t *conn;
 456
 457         if (list_empty(&peer->ibp_conns)) {
 458                 kiblnd_unlink_peer_locked(peer);
 459         } else {
 460                 list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 461                         conn = list_entry(ctmp, kib_conn_t, ibc_list);
 462
 463                         kiblnd_close_conn_locked(conn, 0);
 464                 }
 465                 /* NB closing peer's last conn unlinked it. */
 466         }
 467         /*
 468          * NB peer now unlinked; might even be freed if the peer table had the
 469          * last ref on it.
 470          */
 471 }
 472
 473 static int kiblnd_del_peer(lnet_ni_t *ni, lnet_nid_t nid)
 474 {
 475         LIST_HEAD(zombies);
 476         struct list_head *ptmp;
 477         struct list_head *pnxt;
 478         kib_peer_t *peer;
 479         int lo;
 480         int hi;
 481         int i;
 482         unsigned long flags;
 483         int rc = -ENOENT;
 484
 485         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 486
 487         if (nid != LNET_NID_ANY) {
 488                 lo = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 489                 hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 490         } else {
 491                 lo = 0;
 492                 hi = kiblnd_data.kib_peer_hash_size - 1;
 493         }
 494
 495         for (i = lo; i <= hi; i++) {
 496                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 497                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 498                         LASSERT(!kiblnd_peer_idle(peer));
 499
 500                         if (peer->ibp_ni != ni)
 501                                 continue;
 502
 503                         if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
 504                                 continue;
 505
 506                         if (!list_empty(&peer->ibp_tx_queue)) {
 507                                 LASSERT(list_empty(&peer->ibp_conns));
 508
 509                                 list_splice_init(&peer->ibp_tx_queue,
 510                                                  &zombies);
 511                         }
 512
 513                         kiblnd_del_peer_locked(peer);
 514                         rc = 0;  /* matched something */
 515                 }
 516         }
 517
 518         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 519
 520         kiblnd_txlist_done(ni, &zombies, -EIO);
 521
 522         return rc;
 523 }
 524
 525 static kib_conn_t *kiblnd_get_conn_by_idx(lnet_ni_t *ni, int index)
 526 {
 527         kib_peer_t *peer;
 528         struct list_head *ptmp;
 529         kib_conn_t *conn;
 530         struct list_head *ctmp;
 531         int i;
 532         unsigned long flags;
 533
 534         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 535
 536         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 537                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 538                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 539                         LASSERT(!kiblnd_peer_idle(peer));
 540
 541                         if (peer->ibp_ni != ni)
 542                                 continue;
 543
 544                         list_for_each(ctmp, &peer->ibp_conns) {
 545                                 if (index-- > 0)
 546                                         continue;
 547
 548                                 conn = list_entry(ctmp, kib_conn_t,
 549                                                   ibc_list);
 550                                 kiblnd_conn_addref(conn);
 551                                 read_unlock_irqrestore(
 552                                         &kiblnd_data.kib_global_lock,
 553                                         flags);
 554                                 return conn;
 555                         }
 556                 }
 557         }
 558
 559         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 560         return NULL;
 561 }
 562
 563 int kiblnd_translate_mtu(int value)
 564 {
 565         switch (value) {
 566         default:
 567                 return -1;
 568         case 0:
 569                 return 0;
 570         case 256:
 571                 return IB_MTU_256;
 572         case 512:
 573                 return IB_MTU_512;
 574         case 1024:
 575                 return IB_MTU_1024;
 576         case 2048:
 577                 return IB_MTU_2048;
 578         case 4096:
 579                 return IB_MTU_4096;
 580         }
 581 }
 582
 583 static void kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
 584 {
 585         int mtu;
 586
 587         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 588         if (!cmid->route.path_rec)
 589                 return;
 590
 591         mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
 592         LASSERT(mtu >= 0);
 593         if (mtu)
 594                 cmid->route.path_rec->mtu = mtu;
 595 }
 596
 597 static int kiblnd_get_completion_vector(kib_conn_t *conn, int cpt)
 598 {
 599         cpumask_t *mask;
 600         int vectors;
 601         int off;
 602         int i;
 603         lnet_nid_t nid = conn->ibc_peer->ibp_nid;
 604
 605         vectors = conn->ibc_cmid->device->num_comp_vectors;
 606         if (vectors <= 1)
 607                 return 0;
 608
 609         mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
 610         if (!mask)
 611                 return 0;
 612
 613         /* hash NID to CPU id in this partition... */
 614         off = do_div(nid, cpumask_weight(mask));
 615         for_each_cpu(i, mask) {
 616                 if (!off--)
 617                         return i % vectors;
 618         }
 619
 620         LBUG();
 621         return 1;
 622 }
 623
 624 kib_conn_t *kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
 625                                int state, int version)
 626 {
 627         /*
 628          * CAVEAT EMPTOR:
 629          * If the new conn is created successfully it takes over the caller's
 630          * ref on 'peer'.  It also "owns" 'cmid' and destroys it when it itself
 631          * is destroyed.  On failure, the caller's ref on 'peer' remains and
 632          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 633          * to destroy 'cmid' here since I'm called from the CM which still has
 634          * its ref on 'cmid').
 635          */
 636         rwlock_t *glock = &kiblnd_data.kib_global_lock;
 637         kib_net_t *net = peer->ibp_ni->ni_data;
 638         kib_dev_t *dev;
 639         struct ib_qp_init_attr *init_qp_attr;
 640         struct kib_sched_info *sched;
 641         struct ib_cq_init_attr cq_attr = {};
 642         kib_conn_t *conn;
 643         struct ib_cq *cq;
 644         unsigned long flags;
 645         int cpt;
 646         int rc;
 647         int i;
 648
 649         LASSERT(net);
 650         LASSERT(!in_interrupt());
 651
 652         dev = net->ibn_dev;
 653
 654         cpt = lnet_cpt_of_nid(peer->ibp_nid);
 655         sched = kiblnd_data.kib_scheds[cpt];
 656
 657         LASSERT(sched->ibs_nthreads > 0);
 658
 659         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 660                          sizeof(*init_qp_attr));
 661         if (!init_qp_attr) {
 662                 CERROR("Can't allocate qp_attr for %s\n",
 663                        libcfs_nid2str(peer->ibp_nid));
 664                 goto failed_0;
 665         }
 666
 667         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 668         if (!conn) {
 669                 CERROR("Can't allocate connection for %s\n",
 670                        libcfs_nid2str(peer->ibp_nid));
 671                 goto failed_1;
 672         }
 673
 674         conn->ibc_state = IBLND_CONN_INIT;
 675         conn->ibc_version = version;
 676         conn->ibc_peer = peer;            /* I take the caller's ref */
 677         cmid->context = conn;              /* for future CM callbacks */
 678         conn->ibc_cmid = cmid;
 679         conn->ibc_max_frags = peer->ibp_max_frags;
 680         conn->ibc_queue_depth = peer->ibp_queue_depth;
 681
 682         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 683         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 684         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 685         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 686         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 687         INIT_LIST_HEAD(&conn->ibc_active_txs);
 688         spin_lock_init(&conn->ibc_lock);
 689
 690         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 691                          sizeof(*conn->ibc_connvars));
 692         if (!conn->ibc_connvars) {
 693                 CERROR("Can't allocate in-progress connection state\n");
 694                 goto failed_2;
 695         }
 696
 697         write_lock_irqsave(glock, flags);
 698         if (dev->ibd_failover) {
 699                 write_unlock_irqrestore(glock, flags);
 700                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 701                 goto failed_2;
 702         }
 703
 704         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 705                 /* wakeup failover thread and teardown connection */
 706                 if (kiblnd_dev_can_failover(dev)) {
 707                         list_add_tail(&dev->ibd_fail_list,
 708                                       &kiblnd_data.kib_failed_devs);
 709                         wake_up(&kiblnd_data.kib_failover_waitq);
 710                 }
 711
 712                 write_unlock_irqrestore(glock, flags);
 713                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 714                        cmid->device->name, dev->ibd_ifname);
 715                 goto failed_2;
 716         }
 717
 718         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 719         conn->ibc_hdev = dev->ibd_hdev;
 720
 721         kiblnd_setup_mtu_locked(cmid);
 722
 723         write_unlock_irqrestore(glock, flags);
 724
 725         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 726                          IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
 727         if (!conn->ibc_rxs) {
 728                 CERROR("Cannot allocate RX buffers\n");
 729                 goto failed_2;
 730         }
 731
 732         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 733                                 IBLND_RX_MSG_PAGES(conn));
 734         if (rc)
 735                 goto failed_2;
 736
 737         kiblnd_map_rx_descs(conn);
 738
 739         cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
 740         cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
 741         cq = ib_create_cq(cmid->device,
 742                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 743                           &cq_attr);
 744         if (IS_ERR(cq)) {
 745                 CERROR("Failed to create CQ with %d CQEs: %ld\n",
 746                        IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
 747                 goto failed_2;
 748         }
 749
 750         conn->ibc_cq = cq;
 751
 752         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 753         if (rc) {
 754                 CERROR("Can't request completion notification: %d\n", rc);
 755                 goto failed_2;
 756         }
 757
 758         init_qp_attr->event_handler = kiblnd_qp_event;
 759         init_qp_attr->qp_context = conn;
 760         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(conn);
 761         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(conn);
 762         init_qp_attr->cap.max_send_sge = 1;
 763         init_qp_attr->cap.max_recv_sge = 1;
 764         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 765         init_qp_attr->qp_type = IB_QPT_RC;
 766         init_qp_attr->send_cq = cq;
 767         init_qp_attr->recv_cq = cq;
 768
 769         conn->ibc_sched = sched;
 770
 771         rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 772         if (rc) {
 773                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
 774                        rc, init_qp_attr->cap.max_send_wr,
 775                        init_qp_attr->cap.max_recv_wr);
 776                 goto failed_2;
 777         }
 778
 779         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 780
 781         /* 1 ref for caller and each rxmsg */
 782         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
 783         conn->ibc_nrx = IBLND_RX_MSGS(conn);
 784
 785         /* post receives */
 786         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
 787                 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
 788                                     IBLND_POSTRX_NO_CREDIT);
 789                 if (rc) {
 790                         CERROR("Can't post rxmsg: %d\n", rc);
 791
 792                         /* Make posted receives complete */
 793                         kiblnd_abort_receives(conn);
 794
 795                         /*
 796                          * correct # of posted buffers
 797                          * NB locking needed now I'm racing with completion
 798                          */
 799                         spin_lock_irqsave(&sched->ibs_lock, flags);
 800                         conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
 801                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 802
 803                         /*
 804                          * cmid will be destroyed by CM(ofed) after cm_callback
 805                          * returned, so we can't refer it anymore
 806                          * (by kiblnd_connd()->kiblnd_destroy_conn)
 807                          */
 808                         rdma_destroy_qp(conn->ibc_cmid);
 809                         conn->ibc_cmid = NULL;
 810
 811                         /* Drop my own and unused rxbuffer refcounts */
 812                         while (i++ <= IBLND_RX_MSGS(conn))
 813                                 kiblnd_conn_decref(conn);
 814
 815                         return NULL;
 816                 }
 817         }
 818
 819         /* Init successful! */
 820         LASSERT(state == IBLND_CONN_ACTIVE_CONNECT ||
 821                 state == IBLND_CONN_PASSIVE_WAIT);
 822         conn->ibc_state = state;
 823
 824         /* 1 more conn */
 825         atomic_inc(&net->ibn_nconns);
 826         return conn;
 827
 828  failed_2:
 829         kiblnd_destroy_conn(conn, true);
 830  failed_1:
 831         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 832  failed_0:
 833         return NULL;
 834 }
 835
 836 void kiblnd_destroy_conn(kib_conn_t *conn, bool free_conn)
 837 {
 838         struct rdma_cm_id *cmid = conn->ibc_cmid;
 839         kib_peer_t *peer = conn->ibc_peer;
 840         int rc;
 841
 842         LASSERT(!in_interrupt());
 843         LASSERT(!atomic_read(&conn->ibc_refcount));
 844         LASSERT(list_empty(&conn->ibc_early_rxs));
 845         LASSERT(list_empty(&conn->ibc_tx_noops));
 846         LASSERT(list_empty(&conn->ibc_tx_queue));
 847         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
 848         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
 849         LASSERT(list_empty(&conn->ibc_active_txs));
 850         LASSERT(!conn->ibc_noops_posted);
 851         LASSERT(!conn->ibc_nsends_posted);
 852
 853         switch (conn->ibc_state) {
 854         default:
 855                 /* conn must be completely disengaged from the network */
 856                 LBUG();
 857
 858         case IBLND_CONN_DISCONNECTED:
 859                 /* connvars should have been freed already */
 860                 LASSERT(!conn->ibc_connvars);
 861                 break;
 862
 863         case IBLND_CONN_INIT:
 864                 break;
 865         }
 866
 867         /* conn->ibc_cmid might be destroyed by CM already */
 868         if (cmid && cmid->qp)
 869                 rdma_destroy_qp(cmid);
 870
 871         if (conn->ibc_cq) {
 872                 rc = ib_destroy_cq(conn->ibc_cq);
 873                 if (rc)
 874                         CWARN("Error destroying CQ: %d\n", rc);
 875         }
 876
 877         if (conn->ibc_rx_pages)
 878                 kiblnd_unmap_rx_descs(conn);
 879
 880         if (conn->ibc_rxs) {
 881                 LIBCFS_FREE(conn->ibc_rxs,
 882                             IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
 883         }
 884
 885         if (conn->ibc_connvars)
 886                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 887
 888         if (conn->ibc_hdev)
 889                 kiblnd_hdev_decref(conn->ibc_hdev);
 890
 891         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 892         if (conn->ibc_state != IBLND_CONN_INIT) {
 893                 kib_net_t *net = peer->ibp_ni->ni_data;
 894
 895                 kiblnd_peer_decref(peer);
 896                 rdma_destroy_id(cmid);
 897                 atomic_dec(&net->ibn_nconns);
 898         }
 899
 900         LIBCFS_FREE(conn, sizeof(*conn));
 901 }
 902
 903 int kiblnd_close_peer_conns_locked(kib_peer_t *peer, int why)
 904 {
 905         kib_conn_t *conn;
 906         struct list_head *ctmp;
 907         struct list_head *cnxt;
 908         int count = 0;
 909
 910         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 911                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 912
 913                 CDEBUG(D_NET, "Closing conn -> %s, version: %x, reason: %d\n",
 914                        libcfs_nid2str(peer->ibp_nid),
 915                        conn->ibc_version, why);
 916
 917                 kiblnd_close_conn_locked(conn, why);
 918                 count++;
 919         }
 920
 921         return count;
 922 }
 923
 924 int kiblnd_close_stale_conns_locked(kib_peer_t *peer,
 925                                     int version, __u64 incarnation)
 926 {
 927         kib_conn_t *conn;
 928         struct list_head *ctmp;
 929         struct list_head *cnxt;
 930         int count = 0;
 931
 932         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 933                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 934
 935                 if (conn->ibc_version     == version &&
 936                     conn->ibc_incarnation == incarnation)
 937                         continue;
 938
 939                 CDEBUG(D_NET,
 940                        "Closing stale conn -> %s version: %x, incarnation:%#llx(%x, %#llx)\n",
 941                        libcfs_nid2str(peer->ibp_nid),
 942                        conn->ibc_version, conn->ibc_incarnation,
 943                        version, incarnation);
 944
 945                 kiblnd_close_conn_locked(conn, -ESTALE);
 946                 count++;
 947         }
 948
 949         return count;
 950 }
 951
 952 static int kiblnd_close_matching_conns(lnet_ni_t *ni, lnet_nid_t nid)
 953 {
 954         kib_peer_t *peer;
 955         struct list_head *ptmp;
 956         struct list_head *pnxt;
 957         int lo;
 958         int hi;
 959         int i;
 960         unsigned long flags;
 961         int count = 0;
 962
 963         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 964
 965         if (nid != LNET_NID_ANY) {
 966                 lo = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 967                 hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 968         } else {
 969                 lo = 0;
 970                 hi = kiblnd_data.kib_peer_hash_size - 1;
 971         }
 972
 973         for (i = lo; i <= hi; i++) {
 974                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 975                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 976                         LASSERT(!kiblnd_peer_idle(peer));
 977
 978                         if (peer->ibp_ni != ni)
 979                                 continue;
 980
 981                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
 982                                 continue;
 983
 984                         count += kiblnd_close_peer_conns_locked(peer, 0);
 985                 }
 986         }
 987
 988         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 989
 990         /* wildcards always succeed */
 991         if (nid == LNET_NID_ANY)
 992                 return 0;
 993
 994         return !count ? -ENOENT : 0;
 995 }
 996
 997 static int kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
 998 {
 999         struct libcfs_ioctl_data *data = arg;
1000         int rc = -EINVAL;
1001
1002         switch (cmd) {
1003         case IOC_LIBCFS_GET_PEER: {
1004                 lnet_nid_t nid = 0;
1005                 int count = 0;
1006
1007                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1008                                           &nid, &count);
1009                 data->ioc_nid   = nid;
1010                 data->ioc_count = count;
1011                 break;
1012         }
1013
1014         case IOC_LIBCFS_DEL_PEER: {
1015                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1016                 break;
1017         }
1018         case IOC_LIBCFS_GET_CONN: {
1019                 kib_conn_t *conn;
1020
1021                 rc = 0;
1022                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1023                 if (!conn) {
1024                         rc = -ENOENT;
1025                         break;
1026                 }
1027
1028                 LASSERT(conn->ibc_cmid);
1029                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1030                 if (!conn->ibc_cmid->route.path_rec)
1031                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1032                 else
1033                         data->ioc_u32[0] =
1034                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1035                 kiblnd_conn_decref(conn);
1036                 break;
1037         }
1038         case IOC_LIBCFS_CLOSE_CONNECTION: {
1039                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1040                 break;
1041         }
1042
1043         default:
1044                 break;
1045         }
1046
1047         return rc;
1048 }
1049
1050 static void kiblnd_query(lnet_ni_t *ni, lnet_nid_t nid, unsigned long *when)
1051 {
1052         unsigned long last_alive = 0;
1053         unsigned long now = cfs_time_current();
1054         rwlock_t *glock = &kiblnd_data.kib_global_lock;
1055         kib_peer_t *peer;
1056         unsigned long flags;
1057
1058         read_lock_irqsave(glock, flags);
1059
1060         peer = kiblnd_find_peer_locked(nid);
1061         if (peer)
1062                 last_alive = peer->ibp_last_alive;
1063
1064         read_unlock_irqrestore(glock, flags);
1065
1066         if (last_alive)
1067                 *when = last_alive;
1068
1069         /*
1070          * peer is not persistent in hash, trigger peer creation
1071          * and connection establishment with a NULL tx
1072          */
1073         if (!peer)
1074                 kiblnd_launch_tx(ni, NULL, nid);
1075
1076         CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1077                libcfs_nid2str(nid), peer,
1078                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1079 }
1080
1081 static void kiblnd_free_pages(kib_pages_t *p)
1082 {
1083         int npages = p->ibp_npages;
1084         int i;
1085
1086         for (i = 0; i < npages; i++) {
1087                 if (p->ibp_pages[i])
1088                         __free_page(p->ibp_pages[i]);
1089         }
1090
1091         LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1092 }
1093
1094 int kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1095 {
1096         kib_pages_t *p;
1097         int i;
1098
1099         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1100                          offsetof(kib_pages_t, ibp_pages[npages]));
1101         if (!p) {
1102                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1103                 return -ENOMEM;
1104         }
1105
1106         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1107         p->ibp_npages = npages;
1108
1109         for (i = 0; i < npages; i++) {
1110                 p->ibp_pages[i] = alloc_pages_node(
1111                                     cfs_cpt_spread_node(lnet_cpt_table(), cpt),
1112                                     GFP_NOFS, 0);
1113                 if (!p->ibp_pages[i]) {
1114                         CERROR("Can't allocate page %d of %d\n", i, npages);
1115                         kiblnd_free_pages(p);
1116                         return -ENOMEM;
1117                 }
1118         }
1119
1120         *pp = p;
1121         return 0;
1122 }
1123
1124 void kiblnd_unmap_rx_descs(kib_conn_t *conn)
1125 {
1126         kib_rx_t *rx;
1127         int i;
1128
1129         LASSERT(conn->ibc_rxs);
1130         LASSERT(conn->ibc_hdev);
1131
1132         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1133                 rx = &conn->ibc_rxs[i];
1134
1135                 LASSERT(rx->rx_nob >= 0); /* not posted */
1136
1137                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1138                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1139                                                           rx->rx_msgaddr),
1140                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1141         }
1142
1143         kiblnd_free_pages(conn->ibc_rx_pages);
1144
1145         conn->ibc_rx_pages = NULL;
1146 }
1147
1148 void kiblnd_map_rx_descs(kib_conn_t *conn)
1149 {
1150         kib_rx_t *rx;
1151         struct page *pg;
1152         int pg_off;
1153         int ipg;
1154         int i;
1155
1156         for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1157                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1158                 rx = &conn->ibc_rxs[i];
1159
1160                 rx->rx_conn = conn;
1161                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1162
1163                 rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1164                                                        rx->rx_msg,
1165                                                        IBLND_MSG_SIZE,
1166                                                        DMA_FROM_DEVICE);
1167                 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1168                                                   rx->rx_msgaddr));
1169                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1170
1171                 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1172                        i, rx->rx_msg, rx->rx_msgaddr,
1173                        (__u64)(page_to_phys(pg) + pg_off));
1174
1175                 pg_off += IBLND_MSG_SIZE;
1176                 LASSERT(pg_off <= PAGE_SIZE);
1177
1178                 if (pg_off == PAGE_SIZE) {
1179                         pg_off = 0;
1180                         ipg++;
1181                         LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1182                 }
1183         }
1184 }
1185
1186 static void kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1187 {
1188         kib_hca_dev_t *hdev = tpo->tpo_hdev;
1189         kib_tx_t *tx;
1190         int i;
1191
1192         LASSERT(!tpo->tpo_pool.po_allocated);
1193
1194         if (!hdev)
1195                 return;
1196
1197         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1198                 tx = &tpo->tpo_tx_descs[i];
1199                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1200                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1201                                                           tx->tx_msgaddr),
1202                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1203         }
1204
1205         kiblnd_hdev_decref(hdev);
1206         tpo->tpo_hdev = NULL;
1207 }
1208
1209 static kib_hca_dev_t *kiblnd_current_hdev(kib_dev_t *dev)
1210 {
1211         kib_hca_dev_t *hdev;
1212         unsigned long flags;
1213         int i = 0;
1214
1215         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1216         while (dev->ibd_failover) {
1217                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1218                 if (!(i++ % 50))
1219                         CDEBUG(D_NET, "%s: Wait for failover\n",
1220                                dev->ibd_ifname);
1221                 set_current_state(TASK_INTERRUPTIBLE);
1222                 schedule_timeout(cfs_time_seconds(1) / 100);
1223
1224                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1225         }
1226
1227         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1228         hdev = dev->ibd_hdev;
1229
1230         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1231
1232         return hdev;
1233 }
1234
1235 static void kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1236 {
1237         kib_pages_t *txpgs = tpo->tpo_tx_pages;
1238         kib_pool_t *pool = &tpo->tpo_pool;
1239         kib_net_t *net = pool->po_owner->ps_net;
1240         kib_dev_t *dev;
1241         struct page *page;
1242         kib_tx_t *tx;
1243         int page_offset;
1244         int ipage;
1245         int i;
1246
1247         LASSERT(net);
1248
1249         dev = net->ibn_dev;
1250
1251         /* pre-mapped messages are not bigger than 1 page */
1252         CLASSERT(IBLND_MSG_SIZE <= PAGE_SIZE);
1253
1254         /* No fancy arithmetic when we do the buffer calculations */
1255         CLASSERT(!(PAGE_SIZE % IBLND_MSG_SIZE));
1256
1257         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1258
1259         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1260                 page = txpgs->ibp_pages[ipage];
1261                 tx = &tpo->tpo_tx_descs[i];
1262
1263                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1264                                            page_offset);
1265
1266                 tx->tx_msgaddr = kiblnd_dma_map_single(
1267                         tpo->tpo_hdev->ibh_ibdev, tx->tx_msg,
1268                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1269                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1270                                                   tx->tx_msgaddr));
1271                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1272
1273                 list_add(&tx->tx_list, &pool->po_free_list);
1274
1275                 page_offset += IBLND_MSG_SIZE;
1276                 LASSERT(page_offset <= PAGE_SIZE);
1277
1278                 if (page_offset == PAGE_SIZE) {
1279                         page_offset = 0;
1280                         ipage++;
1281                         LASSERT(ipage <= txpgs->ibp_npages);
1282                 }
1283         }
1284 }
1285
1286 struct ib_mr *kiblnd_find_rd_dma_mr(struct lnet_ni *ni, kib_rdma_desc_t *rd,
1287                                     int negotiated_nfrags)
1288 {
1289         kib_net_t *net = ni->ni_data;
1290         kib_hca_dev_t *hdev = net->ibn_dev->ibd_hdev;
1291         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
1292         __u16 nfrags;
1293         int mod;
1294
1295         tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
1296         mod = tunables->lnd_map_on_demand;
1297         nfrags = (negotiated_nfrags != -1) ? negotiated_nfrags : mod;
1298
1299         LASSERT(hdev->ibh_mrs);
1300
1301         if (mod > 0 && nfrags <= rd->rd_nfrags)
1302                 return NULL;
1303
1304         return hdev->ibh_mrs;
1305 }
1306
1307 static void kiblnd_destroy_fmr_pool(kib_fmr_pool_t *fpo)
1308 {
1309         LASSERT(!fpo->fpo_map_count);
1310
1311         if (fpo->fpo_is_fmr) {
1312                 if (fpo->fmr.fpo_fmr_pool)
1313                         ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1314         } else {
1315                 struct kib_fast_reg_descriptor *frd, *tmp;
1316                 int i = 0;
1317
1318                 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1319                                          frd_list) {
1320                         list_del(&frd->frd_list);
1321                         ib_dereg_mr(frd->frd_mr);
1322                         LIBCFS_FREE(frd, sizeof(*frd));
1323                         i++;
1324                 }
1325                 if (i < fpo->fast_reg.fpo_pool_size)
1326                         CERROR("FastReg pool still has %d regions registered\n",
1327                                fpo->fast_reg.fpo_pool_size - i);
1328         }
1329
1330         if (fpo->fpo_hdev)
1331                 kiblnd_hdev_decref(fpo->fpo_hdev);
1332
1333         LIBCFS_FREE(fpo, sizeof(*fpo));
1334 }
1335
1336 static void kiblnd_destroy_fmr_pool_list(struct list_head *head)
1337 {
1338         kib_fmr_pool_t *fpo, *tmp;
1339
1340         list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1341                 list_del(&fpo->fpo_list);
1342                 kiblnd_destroy_fmr_pool(fpo);
1343         }
1344 }
1345
1346 static int
1347 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1348                      int ncpts)
1349 {
1350         int size = tunables->lnd_fmr_pool_size / ncpts;
1351
1352         return max(IBLND_FMR_POOL, size);
1353 }
1354
1355 static int
1356 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1357                          int ncpts)
1358 {
1359         int size = tunables->lnd_fmr_flush_trigger / ncpts;
1360
1361         return max(IBLND_FMR_POOL_FLUSH, size);
1362 }
1363
1364 static int kiblnd_alloc_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
1365 {
1366         struct ib_fmr_pool_param param = {
1367                 .max_pages_per_fmr = LNET_MAX_PAYLOAD / PAGE_SIZE,
1368                 .page_shift        = PAGE_SHIFT,
1369                 .access            = (IB_ACCESS_LOCAL_WRITE |
1370                                       IB_ACCESS_REMOTE_WRITE),
1371                 .pool_size         = fps->fps_pool_size,
1372                 .dirty_watermark   = fps->fps_flush_trigger,
1373                 .flush_function    = NULL,
1374                 .flush_arg         = NULL,
1375                 .cache             = !!fps->fps_cache };
1376         int rc = 0;
1377
1378         fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1379                                                    &param);
1380         if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1381                 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1382                 if (rc != -ENOSYS)
1383                         CERROR("Failed to create FMR pool: %d\n", rc);
1384                 else
1385                         CERROR("FMRs are not supported\n");
1386         }
1387
1388         return rc;
1389 }
1390
1391 static int kiblnd_alloc_freg_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
1392 {
1393         struct kib_fast_reg_descriptor *frd, *tmp;
1394         int i, rc;
1395
1396         INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1397         fpo->fast_reg.fpo_pool_size = 0;
1398         for (i = 0; i < fps->fps_pool_size; i++) {
1399                 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1400                                  sizeof(*frd));
1401                 if (!frd) {
1402                         CERROR("Failed to allocate a new fast_reg descriptor\n");
1403                         rc = -ENOMEM;
1404                         goto out;
1405                 }
1406
1407                 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1408                                           IB_MR_TYPE_MEM_REG,
1409                                           LNET_MAX_PAYLOAD / PAGE_SIZE);
1410                 if (IS_ERR(frd->frd_mr)) {
1411                         rc = PTR_ERR(frd->frd_mr);
1412                         CERROR("Failed to allocate ib_alloc_mr: %d\n", rc);
1413                         frd->frd_mr = NULL;
1414                         goto out_middle;
1415                 }
1416
1417                 frd->frd_valid = true;
1418
1419                 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1420                 fpo->fast_reg.fpo_pool_size++;
1421         }
1422
1423         return 0;
1424
1425 out_middle:
1426         if (frd->frd_mr)
1427                 ib_dereg_mr(frd->frd_mr);
1428         LIBCFS_FREE(frd, sizeof(*frd));
1429
1430 out:
1431         list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1432                                  frd_list) {
1433                 list_del(&frd->frd_list);
1434                 ib_dereg_mr(frd->frd_mr);
1435                 LIBCFS_FREE(frd, sizeof(*frd));
1436         }
1437
1438         return rc;
1439 }
1440
1441 static int kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps,
1442                                   kib_fmr_pool_t **pp_fpo)
1443 {
1444         kib_dev_t *dev = fps->fps_net->ibn_dev;
1445         struct ib_device_attr *dev_attr;
1446         kib_fmr_pool_t *fpo;
1447         int rc;
1448
1449         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1450         if (!fpo)
1451                 return -ENOMEM;
1452
1453         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1454         dev_attr = &fpo->fpo_hdev->ibh_ibdev->attrs;
1455
1456         /* Check for FMR or FastReg support */
1457         fpo->fpo_is_fmr = 0;
1458         if (fpo->fpo_hdev->ibh_ibdev->alloc_fmr &&
1459             fpo->fpo_hdev->ibh_ibdev->dealloc_fmr &&
1460             fpo->fpo_hdev->ibh_ibdev->map_phys_fmr &&
1461             fpo->fpo_hdev->ibh_ibdev->unmap_fmr) {
1462                 LCONSOLE_INFO("Using FMR for registration\n");
1463                 fpo->fpo_is_fmr = 1;
1464         } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
1465                 LCONSOLE_INFO("Using FastReg for registration\n");
1466         } else {
1467                 rc = -ENOSYS;
1468                 LCONSOLE_ERROR_MSG(rc, "IB device does not support FMRs nor FastRegs, can't register memory\n");
1469                 goto out_fpo;
1470         }
1471
1472         if (fpo->fpo_is_fmr)
1473                 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1474         else
1475                 rc = kiblnd_alloc_freg_pool(fps, fpo);
1476         if (rc)
1477                 goto out_fpo;
1478
1479         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1480         fpo->fpo_owner = fps;
1481         *pp_fpo = fpo;
1482
1483         return 0;
1484
1485 out_fpo:
1486         kiblnd_hdev_decref(fpo->fpo_hdev);
1487         LIBCFS_FREE(fpo, sizeof(*fpo));
1488         return rc;
1489 }
1490
1491 static void kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps,
1492                                     struct list_head *zombies)
1493 {
1494         if (!fps->fps_net) /* intialized? */
1495                 return;
1496
1497         spin_lock(&fps->fps_lock);
1498
1499         while (!list_empty(&fps->fps_pool_list)) {
1500                 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1501                                                  kib_fmr_pool_t, fpo_list);
1502                 fpo->fpo_failed = 1;
1503                 list_del(&fpo->fpo_list);
1504                 if (!fpo->fpo_map_count)
1505                         list_add(&fpo->fpo_list, zombies);
1506                 else
1507                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1508         }
1509
1510         spin_unlock(&fps->fps_lock);
1511 }
1512
1513 static void kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1514 {
1515         if (fps->fps_net) { /* initialized? */
1516                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1517                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1518         }
1519 }
1520
1521 static int
1522 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, int ncpts,
1523                         kib_net_t *net,
1524                         struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1525 {
1526         kib_fmr_pool_t *fpo;
1527         int rc;
1528
1529         memset(fps, 0, sizeof(*fps));
1530
1531         fps->fps_net = net;
1532         fps->fps_cpt = cpt;
1533
1534         fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1535         fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1536         fps->fps_cache = tunables->lnd_fmr_cache;
1537
1538         spin_lock_init(&fps->fps_lock);
1539         INIT_LIST_HEAD(&fps->fps_pool_list);
1540         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1541
1542         rc = kiblnd_create_fmr_pool(fps, &fpo);
1543         if (!rc)
1544                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1545
1546         return rc;
1547 }
1548
1549 static int kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, unsigned long now)
1550 {
1551         if (fpo->fpo_map_count) /* still in use */
1552                 return 0;
1553         if (fpo->fpo_failed)
1554                 return 1;
1555         return cfs_time_aftereq(now, fpo->fpo_deadline);
1556 }
1557
1558 static int
1559 kiblnd_map_tx_pages(kib_tx_t *tx, kib_rdma_desc_t *rd)
1560 {
1561         __u64 *pages = tx->tx_pages;
1562         kib_hca_dev_t *hdev;
1563         int npages;
1564         int size;
1565         int i;
1566
1567         hdev = tx->tx_pool->tpo_hdev;
1568
1569         for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1570                 for (size = 0; size <  rd->rd_frags[i].rf_nob;
1571                      size += hdev->ibh_page_size) {
1572                         pages[npages++] = (rd->rd_frags[i].rf_addr &
1573                                            hdev->ibh_page_mask) + size;
1574                 }
1575         }
1576
1577         return npages;
1578 }
1579
1580 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1581 {
1582         LIST_HEAD(zombies);
1583         kib_fmr_pool_t *fpo = fmr->fmr_pool;
1584         kib_fmr_poolset_t *fps;
1585         unsigned long now = cfs_time_current();
1586         kib_fmr_pool_t *tmp;
1587         int rc;
1588
1589         if (!fpo)
1590                 return;
1591
1592         fps = fpo->fpo_owner;
1593         if (fpo->fpo_is_fmr) {
1594                 if (fmr->fmr_pfmr) {
1595                         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1596                         LASSERT(!rc);
1597                         fmr->fmr_pfmr = NULL;
1598                 }
1599
1600                 if (status) {
1601                         rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1602                         LASSERT(!rc);
1603                 }
1604         } else {
1605                 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1606
1607                 if (frd) {
1608                         frd->frd_valid = false;
1609                         spin_lock(&fps->fps_lock);
1610                         list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1611                         spin_unlock(&fps->fps_lock);
1612                         fmr->fmr_frd = NULL;
1613                 }
1614         }
1615         fmr->fmr_pool = NULL;
1616
1617         spin_lock(&fps->fps_lock);
1618         fpo->fpo_map_count--;  /* decref the pool */
1619
1620         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1621                 /* the first pool is persistent */
1622                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1623                         continue;
1624
1625                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1626                         list_move(&fpo->fpo_list, &zombies);
1627                         fps->fps_version++;
1628                 }
1629         }
1630         spin_unlock(&fps->fps_lock);
1631
1632         if (!list_empty(&zombies))
1633                 kiblnd_destroy_fmr_pool_list(&zombies);
1634 }
1635
1636 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, kib_tx_t *tx,
1637                         kib_rdma_desc_t *rd, __u32 nob, __u64 iov,
1638                         kib_fmr_t *fmr)
1639 {
1640         __u64 *pages = tx->tx_pages;
1641         bool is_rx = (rd != tx->tx_rd);
1642         bool tx_pages_mapped = 0;
1643         kib_fmr_pool_t *fpo;
1644         int npages = 0;
1645         __u64 version;
1646         int rc;
1647
1648  again:
1649         spin_lock(&fps->fps_lock);
1650         version = fps->fps_version;
1651         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1652                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1653                 fpo->fpo_map_count++;
1654
1655                 if (fpo->fpo_is_fmr) {
1656                         struct ib_pool_fmr *pfmr;
1657
1658                         spin_unlock(&fps->fps_lock);
1659
1660                         if (!tx_pages_mapped) {
1661                                 npages = kiblnd_map_tx_pages(tx, rd);
1662                                 tx_pages_mapped = 1;
1663                         }
1664
1665                         pfmr = ib_fmr_pool_map_phys(fpo->fmr.fpo_fmr_pool,
1666                                                     pages, npages, iov);
1667                         if (likely(!IS_ERR(pfmr))) {
1668                                 fmr->fmr_key = is_rx ? pfmr->fmr->rkey :
1669                                                        pfmr->fmr->lkey;
1670                                 fmr->fmr_frd = NULL;
1671                                 fmr->fmr_pfmr = pfmr;
1672                                 fmr->fmr_pool = fpo;
1673                                 return 0;
1674                         }
1675                         rc = PTR_ERR(pfmr);
1676                 } else {
1677                         if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1678                                 struct kib_fast_reg_descriptor *frd;
1679                                 struct ib_reg_wr *wr;
1680                                 struct ib_mr *mr;
1681                                 int n;
1682
1683                                 frd = list_first_entry(&fpo->fast_reg.fpo_pool_list,
1684                                                        struct kib_fast_reg_descriptor,
1685                                                        frd_list);
1686                                 list_del(&frd->frd_list);
1687                                 spin_unlock(&fps->fps_lock);
1688
1689                                 mr = frd->frd_mr;
1690
1691                                 if (!frd->frd_valid) {
1692                                         __u32 key = is_rx ? mr->rkey : mr->lkey;
1693                                         struct ib_send_wr *inv_wr;
1694
1695                                         inv_wr = &frd->frd_inv_wr;
1696                                         memset(inv_wr, 0, sizeof(*inv_wr));
1697                                         inv_wr->opcode = IB_WR_LOCAL_INV;
1698                                         inv_wr->wr_id = IBLND_WID_MR;
1699                                         inv_wr->ex.invalidate_rkey = key;
1700
1701                                         /* Bump the key */
1702                                         key = ib_inc_rkey(key);
1703                                         ib_update_fast_reg_key(mr, key);
1704                                 }
1705
1706                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1707                                                  tx->tx_nfrags, NULL, PAGE_SIZE);
1708                                 if (unlikely(n != tx->tx_nfrags)) {
1709                                         CERROR("Failed to map mr %d/%d elements\n",
1710                                                n, tx->tx_nfrags);
1711                                         return n < 0 ? n : -EINVAL;
1712                                 }
1713
1714                                 mr->iova = iov;
1715
1716                                 /* Prepare FastReg WR */
1717                                 wr = &frd->frd_fastreg_wr;
1718                                 memset(wr, 0, sizeof(*wr));
1719                                 wr->wr.opcode = IB_WR_REG_MR;
1720                                 wr->wr.wr_id = IBLND_WID_MR;
1721                                 wr->wr.num_sge = 0;
1722                                 wr->wr.send_flags = 0;
1723                                 wr->mr = mr;
1724                                 wr->key = is_rx ? mr->rkey : mr->lkey;
1725                                 wr->access = (IB_ACCESS_LOCAL_WRITE |
1726                                               IB_ACCESS_REMOTE_WRITE);
1727
1728                                 fmr->fmr_key = is_rx ? mr->rkey : mr->lkey;
1729                                 fmr->fmr_frd = frd;
1730                                 fmr->fmr_pfmr = NULL;
1731                                 fmr->fmr_pool = fpo;
1732                                 return 0;
1733                         }
1734                         spin_unlock(&fps->fps_lock);
1735                         rc = -EBUSY;
1736                 }
1737
1738                 spin_lock(&fps->fps_lock);
1739                 fpo->fpo_map_count--;
1740                 if (rc != -EAGAIN) {
1741                         spin_unlock(&fps->fps_lock);
1742                         return rc;
1743                 }
1744
1745                 /* EAGAIN and ... */
1746                 if (version != fps->fps_version) {
1747                         spin_unlock(&fps->fps_lock);
1748                         goto again;
1749                 }
1750         }
1751
1752         if (fps->fps_increasing) {
1753                 spin_unlock(&fps->fps_lock);
1754                 CDEBUG(D_NET, "Another thread is allocating new FMR pool, waiting for her to complete\n");
1755                 schedule();
1756                 goto again;
1757         }
1758
1759         if (time_before(cfs_time_current(), fps->fps_next_retry)) {
1760                 /* someone failed recently */
1761                 spin_unlock(&fps->fps_lock);
1762                 return -EAGAIN;
1763         }
1764
1765         fps->fps_increasing = 1;
1766         spin_unlock(&fps->fps_lock);
1767
1768         CDEBUG(D_NET, "Allocate new FMR pool\n");
1769         rc = kiblnd_create_fmr_pool(fps, &fpo);
1770         spin_lock(&fps->fps_lock);
1771         fps->fps_increasing = 0;
1772         if (!rc) {
1773                 fps->fps_version++;
1774                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1775         } else {
1776                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1777         }
1778         spin_unlock(&fps->fps_lock);
1779
1780         goto again;
1781 }
1782
1783 static void kiblnd_fini_pool(kib_pool_t *pool)
1784 {
1785         LASSERT(list_empty(&pool->po_free_list));
1786         LASSERT(!pool->po_allocated);
1787
1788         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1789 }
1790
1791 static void kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1792 {
1793         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1794
1795         memset(pool, 0, sizeof(*pool));
1796         INIT_LIST_HEAD(&pool->po_free_list);
1797         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1798         pool->po_owner    = ps;
1799         pool->po_size     = size;
1800 }
1801
1802 static void kiblnd_destroy_pool_list(struct list_head *head)
1803 {
1804         kib_pool_t *pool;
1805
1806         while (!list_empty(head)) {
1807                 pool = list_entry(head->next, kib_pool_t, po_list);
1808                 list_del(&pool->po_list);
1809
1810                 LASSERT(pool->po_owner);
1811                 pool->po_owner->ps_pool_destroy(pool);
1812         }
1813 }
1814
1815 static void kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
1816 {
1817         if (!ps->ps_net) /* intialized? */
1818                 return;
1819
1820         spin_lock(&ps->ps_lock);
1821         while (!list_empty(&ps->ps_pool_list)) {
1822                 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
1823                                             kib_pool_t, po_list);
1824                 po->po_failed = 1;
1825                 list_del(&po->po_list);
1826                 if (!po->po_allocated)
1827                         list_add(&po->po_list, zombies);
1828                 else
1829                         list_add(&po->po_list, &ps->ps_failed_pool_list);
1830         }
1831         spin_unlock(&ps->ps_lock);
1832 }
1833
1834 static void kiblnd_fini_poolset(kib_poolset_t *ps)
1835 {
1836         if (ps->ps_net) { /* initialized? */
1837                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1838                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1839         }
1840 }
1841
1842 static int kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
1843                                kib_net_t *net, char *name, int size,
1844                                kib_ps_pool_create_t po_create,
1845                                kib_ps_pool_destroy_t po_destroy,
1846                                kib_ps_node_init_t nd_init,
1847                                kib_ps_node_fini_t nd_fini)
1848 {
1849         kib_pool_t *pool;
1850         int rc;
1851
1852         memset(ps, 0, sizeof(*ps));
1853
1854         ps->ps_cpt          = cpt;
1855         ps->ps_net          = net;
1856         ps->ps_pool_create  = po_create;
1857         ps->ps_pool_destroy = po_destroy;
1858         ps->ps_node_init    = nd_init;
1859         ps->ps_node_fini    = nd_fini;
1860         ps->ps_pool_size    = size;
1861         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
1862             >= sizeof(ps->ps_name))
1863                 return -E2BIG;
1864         spin_lock_init(&ps->ps_lock);
1865         INIT_LIST_HEAD(&ps->ps_pool_list);
1866         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1867
1868         rc = ps->ps_pool_create(ps, size, &pool);
1869         if (!rc)
1870                 list_add(&pool->po_list, &ps->ps_pool_list);
1871         else
1872                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1873
1874         return rc;
1875 }
1876
1877 static int kiblnd_pool_is_idle(kib_pool_t *pool, unsigned long now)
1878 {
1879         if (pool->po_allocated) /* still in use */
1880                 return 0;
1881         if (pool->po_failed)
1882                 return 1;
1883         return cfs_time_aftereq(now, pool->po_deadline);
1884 }
1885
1886 void kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
1887 {
1888         LIST_HEAD(zombies);
1889         kib_poolset_t *ps = pool->po_owner;
1890         kib_pool_t *tmp;
1891         unsigned long now = cfs_time_current();
1892
1893         spin_lock(&ps->ps_lock);
1894
1895         if (ps->ps_node_fini)
1896                 ps->ps_node_fini(pool, node);
1897
1898         LASSERT(pool->po_allocated > 0);
1899         list_add(node, &pool->po_free_list);
1900         pool->po_allocated--;
1901
1902         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1903                 /* the first pool is persistent */
1904                 if (ps->ps_pool_list.next == &pool->po_list)
1905                         continue;
1906
1907                 if (kiblnd_pool_is_idle(pool, now))
1908                         list_move(&pool->po_list, &zombies);
1909         }
1910         spin_unlock(&ps->ps_lock);
1911
1912         if (!list_empty(&zombies))
1913                 kiblnd_destroy_pool_list(&zombies);
1914 }
1915
1916 struct list_head *kiblnd_pool_alloc_node(kib_poolset_t *ps)
1917 {
1918         struct list_head *node;
1919         kib_pool_t *pool;
1920         unsigned int interval = 1;
1921         unsigned long time_before;
1922         unsigned int trips = 0;
1923         int rc;
1924
1925  again:
1926         spin_lock(&ps->ps_lock);
1927         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1928                 if (list_empty(&pool->po_free_list))
1929                         continue;
1930
1931                 pool->po_allocated++;
1932                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1933                 node = pool->po_free_list.next;
1934                 list_del(node);
1935
1936                 if (ps->ps_node_init) {
1937                         /* still hold the lock */
1938                         ps->ps_node_init(pool, node);
1939                 }
1940                 spin_unlock(&ps->ps_lock);
1941                 return node;
1942         }
1943
1944         /* no available tx pool and ... */
1945         if (ps->ps_increasing) {
1946                 /* another thread is allocating a new pool */
1947                 spin_unlock(&ps->ps_lock);
1948                 trips++;
1949                 CDEBUG(D_NET, "Another thread is allocating new %s pool, waiting %d HZs for her to complete. trips = %d\n",
1950                        ps->ps_name, interval, trips);
1951
1952                 set_current_state(TASK_INTERRUPTIBLE);
1953                 schedule_timeout(interval);
1954                 if (interval < cfs_time_seconds(1))
1955                         interval *= 2;
1956
1957                 goto again;
1958         }
1959
1960         if (time_before(cfs_time_current(), ps->ps_next_retry)) {
1961                 /* someone failed recently */
1962                 spin_unlock(&ps->ps_lock);
1963                 return NULL;
1964         }
1965
1966         ps->ps_increasing = 1;
1967         spin_unlock(&ps->ps_lock);
1968
1969         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1970         time_before = cfs_time_current();
1971         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1972         CDEBUG(D_NET, "ps_pool_create took %lu HZ to complete",
1973                cfs_time_current() - time_before);
1974
1975         spin_lock(&ps->ps_lock);
1976         ps->ps_increasing = 0;
1977         if (!rc) {
1978                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
1979         } else {
1980                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1981                 CERROR("Can't allocate new %s pool because out of memory\n",
1982                        ps->ps_name);
1983         }
1984         spin_unlock(&ps->ps_lock);
1985
1986         goto again;
1987 }
1988
1989 static void kiblnd_destroy_tx_pool(kib_pool_t *pool)
1990 {
1991         kib_tx_pool_t *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
1992         int i;
1993
1994         LASSERT(!pool->po_allocated);
1995
1996         if (tpo->tpo_tx_pages) {
1997                 kiblnd_unmap_tx_pool(tpo);
1998                 kiblnd_free_pages(tpo->tpo_tx_pages);
1999         }
2000
2001         if (!tpo->tpo_tx_descs)
2002                 goto out;
2003
2004         for (i = 0; i < pool->po_size; i++) {
2005                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2006
2007                 list_del(&tx->tx_list);
2008                 if (tx->tx_pages)
2009                         LIBCFS_FREE(tx->tx_pages,
2010                                     LNET_MAX_IOV *
2011                                     sizeof(*tx->tx_pages));
2012                 if (tx->tx_frags)
2013                         LIBCFS_FREE(tx->tx_frags,
2014                                     IBLND_MAX_RDMA_FRAGS *
2015                                             sizeof(*tx->tx_frags));
2016                 if (tx->tx_wrq)
2017                         LIBCFS_FREE(tx->tx_wrq,
2018                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2019                                     sizeof(*tx->tx_wrq));
2020                 if (tx->tx_sge)
2021                         LIBCFS_FREE(tx->tx_sge,
2022                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2023                                     sizeof(*tx->tx_sge));
2024                 if (tx->tx_rd)
2025                         LIBCFS_FREE(tx->tx_rd,
2026                                     offsetof(kib_rdma_desc_t,
2027                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2028         }
2029
2030         LIBCFS_FREE(tpo->tpo_tx_descs,
2031                     pool->po_size * sizeof(kib_tx_t));
2032 out:
2033         kiblnd_fini_pool(pool);
2034         LIBCFS_FREE(tpo, sizeof(*tpo));
2035 }
2036
2037 static int kiblnd_tx_pool_size(int ncpts)
2038 {
2039         int ntx = *kiblnd_tunables.kib_ntx / ncpts;
2040
2041         return max(IBLND_TX_POOL, ntx);
2042 }
2043
2044 static int kiblnd_create_tx_pool(kib_poolset_t *ps, int size,
2045                                  kib_pool_t **pp_po)
2046 {
2047         int i;
2048         int npg;
2049         kib_pool_t *pool;
2050         kib_tx_pool_t *tpo;
2051
2052         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2053         if (!tpo) {
2054                 CERROR("Failed to allocate TX pool\n");
2055                 return -ENOMEM;
2056         }
2057
2058         pool = &tpo->tpo_pool;
2059         kiblnd_init_pool(ps, pool, size);
2060         tpo->tpo_tx_descs = NULL;
2061         tpo->tpo_tx_pages = NULL;
2062
2063         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2064         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg)) {
2065                 CERROR("Can't allocate tx pages: %d\n", npg);
2066                 LIBCFS_FREE(tpo, sizeof(*tpo));
2067                 return -ENOMEM;
2068         }
2069
2070         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2071                          size * sizeof(kib_tx_t));
2072         if (!tpo->tpo_tx_descs) {
2073                 CERROR("Can't allocate %d tx descriptors\n", size);
2074                 ps->ps_pool_destroy(pool);
2075                 return -ENOMEM;
2076         }
2077
2078         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2079
2080         for (i = 0; i < size; i++) {
2081                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2082
2083                 tx->tx_pool = tpo;
2084                 if (ps->ps_net->ibn_fmr_ps) {
2085                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2086                                          lnet_cpt_table(), ps->ps_cpt,
2087                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2088                         if (!tx->tx_pages)
2089                                 break;
2090                 }
2091
2092                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2093                                  IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags));
2094                 if (!tx->tx_frags)
2095                         break;
2096
2097                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
2098
2099                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2100                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2101                                  sizeof(*tx->tx_wrq));
2102                 if (!tx->tx_wrq)
2103                         break;
2104
2105                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2106                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2107                                  sizeof(*tx->tx_sge));
2108                 if (!tx->tx_sge)
2109                         break;
2110
2111                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2112                                  offsetof(kib_rdma_desc_t,
2113                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2114                 if (!tx->tx_rd)
2115                         break;
2116         }
2117
2118         if (i == size) {
2119                 kiblnd_map_tx_pool(tpo);
2120                 *pp_po = pool;
2121                 return 0;
2122         }
2123
2124         ps->ps_pool_destroy(pool);
2125         return -ENOMEM;
2126 }
2127
2128 static void kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2129 {
2130         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2131                                              tps_poolset);
2132         kib_tx_t *tx  = list_entry(node, kib_tx_t, tx_list);
2133
2134         tx->tx_cookie = tps->tps_next_tx_cookie++;
2135 }
2136
2137 static void kiblnd_net_fini_pools(kib_net_t *net)
2138 {
2139         int i;
2140
2141         cfs_cpt_for_each(i, lnet_cpt_table()) {
2142                 kib_tx_poolset_t *tps;
2143                 kib_fmr_poolset_t *fps;
2144
2145                 if (net->ibn_tx_ps) {
2146                         tps = net->ibn_tx_ps[i];
2147                         kiblnd_fini_poolset(&tps->tps_poolset);
2148                 }
2149
2150                 if (net->ibn_fmr_ps) {
2151                         fps = net->ibn_fmr_ps[i];
2152                         kiblnd_fini_fmr_poolset(fps);
2153                 }
2154         }
2155
2156         if (net->ibn_tx_ps) {
2157                 cfs_percpt_free(net->ibn_tx_ps);
2158                 net->ibn_tx_ps = NULL;
2159         }
2160
2161         if (net->ibn_fmr_ps) {
2162                 cfs_percpt_free(net->ibn_fmr_ps);
2163                 net->ibn_fmr_ps = NULL;
2164         }
2165 }
2166
2167 static int kiblnd_net_init_pools(kib_net_t *net, lnet_ni_t *ni, __u32 *cpts,
2168                                  int ncpts)
2169 {
2170         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2171         unsigned long flags;
2172         int cpt;
2173         int rc;
2174         int i;
2175
2176         tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
2177
2178         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2179         if (!tunables->lnd_map_on_demand) {
2180                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2181                 goto create_tx_pool;
2182         }
2183
2184         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2185
2186         if (tunables->lnd_fmr_pool_size < *kiblnd_tunables.kib_ntx / 4) {
2187                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2188                        tunables->lnd_fmr_pool_size,
2189                        *kiblnd_tunables.kib_ntx / 4);
2190                 rc = -EINVAL;
2191                 goto failed;
2192         }
2193
2194         /*
2195          * TX pool must be created later than FMR, see LU-2268
2196          * for details
2197          */
2198         LASSERT(!net->ibn_tx_ps);
2199
2200         /*
2201          * premapping can fail if ibd_nmr > 1, so we always create
2202          * FMR pool and map-on-demand if premapping failed
2203          *
2204          * cfs_precpt_alloc is creating an array of struct kib_fmr_poolset
2205          * The number of struct kib_fmr_poolsets create is equal to the
2206          * number of CPTs that exist, i.e net->ibn_fmr_ps[cpt].
2207          */
2208         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2209                                            sizeof(kib_fmr_poolset_t));
2210         if (!net->ibn_fmr_ps) {
2211                 CERROR("Failed to allocate FMR pool array\n");
2212                 rc = -ENOMEM;
2213                 goto failed;
2214         }
2215
2216         for (i = 0; i < ncpts; i++) {
2217                 cpt = !cpts ? i : cpts[i];
2218                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2219                                              net, tunables);
2220                 if (rc) {
2221                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2222                                cpt, rc);
2223                         goto failed;
2224                 }
2225         }
2226
2227         if (i > 0)
2228                 LASSERT(i == ncpts);
2229
2230  create_tx_pool:
2231         /*
2232          * cfs_precpt_alloc is creating an array of struct kib_tx_poolset
2233          * The number of struct kib_tx_poolsets create is equal to the
2234          * number of CPTs that exist, i.e net->ibn_tx_ps[cpt].
2235          */
2236         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2237                                           sizeof(kib_tx_poolset_t));
2238         if (!net->ibn_tx_ps) {
2239                 CERROR("Failed to allocate tx pool array\n");
2240                 rc = -ENOMEM;
2241                 goto failed;
2242         }
2243
2244         for (i = 0; i < ncpts; i++) {
2245                 cpt = !cpts ? i : cpts[i];
2246                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2247                                          cpt, net, "TX",
2248                                          kiblnd_tx_pool_size(ncpts),
2249                                          kiblnd_create_tx_pool,
2250                                          kiblnd_destroy_tx_pool,
2251                                          kiblnd_tx_init, NULL);
2252                 if (rc) {
2253                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2254                                cpt, rc);
2255                         goto failed;
2256                 }
2257         }
2258
2259         return 0;
2260  failed:
2261         kiblnd_net_fini_pools(net);
2262         LASSERT(rc);
2263         return rc;
2264 }
2265
2266 static int kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2267 {
2268         /*
2269          * It's safe to assume a HCA can handle a page size
2270          * matching that of the native system
2271          */
2272         hdev->ibh_page_shift = PAGE_SHIFT;
2273         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2274         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2275
2276         hdev->ibh_mr_size = hdev->ibh_ibdev->attrs.max_mr_size;
2277         if (hdev->ibh_mr_size == ~0ULL) {
2278                 hdev->ibh_mr_shift = 64;
2279                 return 0;
2280         }
2281
2282         CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2283         return -EINVAL;
2284 }
2285
2286 static void kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2287 {
2288         if (!hdev->ibh_mrs)
2289                 return;
2290
2291         ib_dereg_mr(hdev->ibh_mrs);
2292
2293         hdev->ibh_mrs = NULL;
2294 }
2295
2296 void kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2297 {
2298         kiblnd_hdev_cleanup_mrs(hdev);
2299
2300         if (hdev->ibh_pd)
2301                 ib_dealloc_pd(hdev->ibh_pd);
2302
2303         if (hdev->ibh_cmid)
2304                 rdma_destroy_id(hdev->ibh_cmid);
2305
2306         LIBCFS_FREE(hdev, sizeof(*hdev));
2307 }
2308
2309 static int kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2310 {
2311         struct ib_mr *mr;
2312         int rc;
2313         int acflags = IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE;
2314
2315         rc = kiblnd_hdev_get_attr(hdev);
2316         if (rc)
2317                 return rc;
2318
2319         mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2320         if (IS_ERR(mr)) {
2321                 CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
2322                 kiblnd_hdev_cleanup_mrs(hdev);
2323                 return PTR_ERR(mr);
2324         }
2325
2326         hdev->ibh_mrs = mr;
2327
2328         return 0;
2329 }
2330
2331 /* DUMMY */
2332 static int kiblnd_dummy_callback(struct rdma_cm_id *cmid,
2333                                  struct rdma_cm_event *event)
2334 {
2335         return 0;
2336 }
2337
2338 static int kiblnd_dev_need_failover(kib_dev_t *dev)
2339 {
2340         struct rdma_cm_id *cmid;
2341         struct sockaddr_in srcaddr;
2342         struct sockaddr_in dstaddr;
2343         int rc;
2344
2345         if (!dev->ibd_hdev || /* initializing */
2346             !dev->ibd_hdev->ibh_cmid || /* listener is dead */
2347             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2348                 return 1;
2349
2350         /*
2351          * XXX: it's UGLY, but I don't have better way to find
2352          * ib-bonding HCA failover because:
2353          *
2354          * a. no reliable CM event for HCA failover...
2355          * b. no OFED API to get ib_device for current net_device...
2356          *
2357          * We have only two choices at this point:
2358          *
2359          * a. rdma_bind_addr(), it will conflict with listener cmid
2360          * b. rdma_resolve_addr() to zero addr
2361          */
2362         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2363                                      IB_QPT_RC);
2364         if (IS_ERR(cmid)) {
2365                 rc = PTR_ERR(cmid);
2366                 CERROR("Failed to create cmid for failover: %d\n", rc);
2367                 return rc;
2368         }
2369
2370         memset(&srcaddr, 0, sizeof(srcaddr));
2371         srcaddr.sin_family = AF_INET;
2372         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2373
2374         memset(&dstaddr, 0, sizeof(dstaddr));
2375         dstaddr.sin_family = AF_INET;
2376         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2377                                (struct sockaddr *)&dstaddr, 1);
2378         if (rc || !cmid->device) {
2379                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2380                        dev->ibd_ifname, &dev->ibd_ifip,
2381                        cmid->device, rc);
2382                 rdma_destroy_id(cmid);
2383                 return rc;
2384         }
2385
2386         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2387         rdma_destroy_id(cmid);
2388
2389         return rc;
2390 }
2391
2392 int kiblnd_dev_failover(kib_dev_t *dev)
2393 {
2394         LIST_HEAD(zombie_tpo);
2395         LIST_HEAD(zombie_ppo);
2396         LIST_HEAD(zombie_fpo);
2397         struct rdma_cm_id *cmid  = NULL;
2398         kib_hca_dev_t *hdev  = NULL;
2399         struct ib_pd *pd;
2400         kib_net_t *net;
2401         struct sockaddr_in addr;
2402         unsigned long flags;
2403         int rc = 0;
2404         int i;
2405
2406         LASSERT(*kiblnd_tunables.kib_dev_failover > 1 ||
2407                 dev->ibd_can_failover || !dev->ibd_hdev);
2408
2409         rc = kiblnd_dev_need_failover(dev);
2410         if (rc <= 0)
2411                 goto out;
2412
2413         if (dev->ibd_hdev &&
2414             dev->ibd_hdev->ibh_cmid) {
2415                 /*
2416                  * XXX it's not good to close old listener at here,
2417                  * because we can fail to create new listener.
2418                  * But we have to close it now, otherwise rdma_bind_addr
2419                  * will return EADDRINUSE... How crap!
2420                  */
2421                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2422
2423                 cmid = dev->ibd_hdev->ibh_cmid;
2424                 /*
2425                  * make next schedule of kiblnd_dev_need_failover()
2426                  * return 1 for me
2427                  */
2428                 dev->ibd_hdev->ibh_cmid  = NULL;
2429                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2430
2431                 rdma_destroy_id(cmid);
2432         }
2433
2434         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2435                                      IB_QPT_RC);
2436         if (IS_ERR(cmid)) {
2437                 rc = PTR_ERR(cmid);
2438                 CERROR("Failed to create cmid for failover: %d\n", rc);
2439                 goto out;
2440         }
2441
2442         memset(&addr, 0, sizeof(addr));
2443         addr.sin_family      = AF_INET;
2444         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2445         addr.sin_port   = htons(*kiblnd_tunables.kib_service);
2446
2447         /* Bind to failover device or port */
2448         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2449         if (rc || !cmid->device) {
2450                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2451                        dev->ibd_ifname, &dev->ibd_ifip,
2452                        cmid->device, rc);
2453                 rdma_destroy_id(cmid);
2454                 goto out;
2455         }
2456
2457         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2458         if (!hdev) {
2459                 CERROR("Failed to allocate kib_hca_dev\n");
2460                 rdma_destroy_id(cmid);
2461                 rc = -ENOMEM;
2462                 goto out;
2463         }
2464
2465         atomic_set(&hdev->ibh_ref, 1);
2466         hdev->ibh_dev   = dev;
2467         hdev->ibh_cmid  = cmid;
2468         hdev->ibh_ibdev = cmid->device;
2469
2470         pd = ib_alloc_pd(cmid->device);
2471         if (IS_ERR(pd)) {
2472                 rc = PTR_ERR(pd);
2473                 CERROR("Can't allocate PD: %d\n", rc);
2474                 goto out;
2475         }
2476
2477         hdev->ibh_pd = pd;
2478
2479         rc = rdma_listen(cmid, 0);
2480         if (rc) {
2481                 CERROR("Can't start new listener: %d\n", rc);
2482                 goto out;
2483         }
2484
2485         rc = kiblnd_hdev_setup_mrs(hdev);
2486         if (rc) {
2487                 CERROR("Can't setup device: %d\n", rc);
2488                 goto out;
2489         }
2490
2491         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2492
2493         swap(dev->ibd_hdev, hdev); /* take over the refcount */
2494
2495         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2496                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2497                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2498                                             &zombie_tpo);
2499
2500                         if (net->ibn_fmr_ps)
2501                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2502                                                         &zombie_fpo);
2503                 }
2504         }
2505
2506         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2507  out:
2508         if (!list_empty(&zombie_tpo))
2509                 kiblnd_destroy_pool_list(&zombie_tpo);
2510         if (!list_empty(&zombie_ppo))
2511                 kiblnd_destroy_pool_list(&zombie_ppo);
2512         if (!list_empty(&zombie_fpo))
2513                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2514         if (hdev)
2515                 kiblnd_hdev_decref(hdev);
2516
2517         if (rc)
2518                 dev->ibd_failed_failover++;
2519         else
2520                 dev->ibd_failed_failover = 0;
2521
2522         return rc;
2523 }
2524
2525 void kiblnd_destroy_dev(kib_dev_t *dev)
2526 {
2527         LASSERT(!dev->ibd_nnets);
2528         LASSERT(list_empty(&dev->ibd_nets));
2529
2530         list_del(&dev->ibd_fail_list);
2531         list_del(&dev->ibd_list);
2532
2533         if (dev->ibd_hdev)
2534                 kiblnd_hdev_decref(dev->ibd_hdev);
2535
2536         LIBCFS_FREE(dev, sizeof(*dev));
2537 }
2538
2539 static kib_dev_t *kiblnd_create_dev(char *ifname)
2540 {
2541         struct net_device *netdev;
2542         kib_dev_t *dev;
2543         __u32 netmask;
2544         __u32 ip;
2545         int up;
2546         int rc;
2547
2548         rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2549         if (rc) {
2550                 CERROR("Can't query IPoIB interface %s: %d\n",
2551                        ifname, rc);
2552                 return NULL;
2553         }
2554
2555         if (!up) {
2556                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2557                 return NULL;
2558         }
2559
2560         LIBCFS_ALLOC(dev, sizeof(*dev));
2561         if (!dev)
2562                 return NULL;
2563
2564         netdev = dev_get_by_name(&init_net, ifname);
2565         if (!netdev) {
2566                 dev->ibd_can_failover = 0;
2567         } else {
2568                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2569                 dev_put(netdev);
2570         }
2571
2572         INIT_LIST_HEAD(&dev->ibd_nets);
2573         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2574         INIT_LIST_HEAD(&dev->ibd_fail_list);
2575         dev->ibd_ifip = ip;
2576         strcpy(&dev->ibd_ifname[0], ifname);
2577
2578         /* initialize the device */
2579         rc = kiblnd_dev_failover(dev);
2580         if (rc) {
2581                 CERROR("Can't initialize device: %d\n", rc);
2582                 LIBCFS_FREE(dev, sizeof(*dev));
2583                 return NULL;
2584         }
2585
2586         list_add_tail(&dev->ibd_list, &kiblnd_data.kib_devs);
2587         return dev;
2588 }
2589
2590 static void kiblnd_base_shutdown(void)
2591 {
2592         struct kib_sched_info *sched;
2593         int i;
2594
2595         LASSERT(list_empty(&kiblnd_data.kib_devs));
2596
2597         switch (kiblnd_data.kib_init) {
2598         default:
2599                 LBUG();
2600
2601         case IBLND_INIT_ALL:
2602         case IBLND_INIT_DATA:
2603                 LASSERT(kiblnd_data.kib_peers);
2604                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2605                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2606                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2607                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2608                 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2609                 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2610
2611                 /* flag threads to terminate; wake and wait for them to die */
2612                 kiblnd_data.kib_shutdown = 1;
2613
2614                 /*
2615                  * NB: we really want to stop scheduler threads net by net
2616                  * instead of the whole module, this should be improved
2617                  * with dynamic configuration LNet
2618                  */
2619                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2620                         wake_up_all(&sched->ibs_waitq);
2621
2622                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2623                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2624
2625                 i = 2;
2626                 while (atomic_read(&kiblnd_data.kib_nthreads)) {
2627                         i++;
2628                         /* power of 2 ? */
2629                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2630                                "Waiting for %d threads to terminate\n",
2631                                atomic_read(&kiblnd_data.kib_nthreads));
2632                         set_current_state(TASK_UNINTERRUPTIBLE);
2633                         schedule_timeout(cfs_time_seconds(1));
2634                 }
2635
2636                 /* fall through */
2637
2638         case IBLND_INIT_NOTHING:
2639                 break;
2640         }
2641
2642         if (kiblnd_data.kib_peers) {
2643                 LIBCFS_FREE(kiblnd_data.kib_peers,
2644                             sizeof(struct list_head) *
2645                             kiblnd_data.kib_peer_hash_size);
2646         }
2647
2648         if (kiblnd_data.kib_scheds)
2649                 cfs_percpt_free(kiblnd_data.kib_scheds);
2650
2651         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2652         module_put(THIS_MODULE);
2653 }
2654
2655 static void kiblnd_shutdown(lnet_ni_t *ni)
2656 {
2657         kib_net_t *net = ni->ni_data;
2658         rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2659         int i;
2660         unsigned long flags;
2661
2662         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2663
2664         if (!net)
2665                 goto out;
2666
2667         write_lock_irqsave(g_lock, flags);
2668         net->ibn_shutdown = 1;
2669         write_unlock_irqrestore(g_lock, flags);
2670
2671         switch (net->ibn_init) {
2672         default:
2673                 LBUG();
2674
2675         case IBLND_INIT_ALL:
2676                 /* nuke all existing peers within this net */
2677                 kiblnd_del_peer(ni, LNET_NID_ANY);
2678
2679                 /* Wait for all peer state to clean up */
2680                 i = 2;
2681                 while (atomic_read(&net->ibn_npeers)) {
2682                         i++;
2683                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */
2684                                "%s: waiting for %d peers to disconnect\n",
2685                                libcfs_nid2str(ni->ni_nid),
2686                                atomic_read(&net->ibn_npeers));
2687                         set_current_state(TASK_UNINTERRUPTIBLE);
2688                         schedule_timeout(cfs_time_seconds(1));
2689                 }
2690
2691                 kiblnd_net_fini_pools(net);
2692
2693                 write_lock_irqsave(g_lock, flags);
2694                 LASSERT(net->ibn_dev->ibd_nnets > 0);
2695                 net->ibn_dev->ibd_nnets--;
2696                 list_del(&net->ibn_list);
2697                 write_unlock_irqrestore(g_lock, flags);
2698
2699                 /* fall through */
2700
2701         case IBLND_INIT_NOTHING:
2702                 LASSERT(!atomic_read(&net->ibn_nconns));
2703
2704                 if (net->ibn_dev && !net->ibn_dev->ibd_nnets)
2705                         kiblnd_destroy_dev(net->ibn_dev);
2706
2707                 break;
2708         }
2709
2710         net->ibn_init = IBLND_INIT_NOTHING;
2711         ni->ni_data = NULL;
2712
2713         LIBCFS_FREE(net, sizeof(*net));
2714
2715 out:
2716         if (list_empty(&kiblnd_data.kib_devs))
2717                 kiblnd_base_shutdown();
2718 }
2719
2720 static int kiblnd_base_startup(void)
2721 {
2722         struct kib_sched_info *sched;
2723         int rc;
2724         int i;
2725
2726         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2727
2728         try_module_get(THIS_MODULE);
2729         /* zero pointers, flags etc */
2730         memset(&kiblnd_data, 0, sizeof(kiblnd_data));
2731
2732         rwlock_init(&kiblnd_data.kib_global_lock);
2733
2734         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2735         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2736
2737         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2738         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2739                      sizeof(struct list_head) * kiblnd_data.kib_peer_hash_size);
2740         if (!kiblnd_data.kib_peers)
2741                 goto failed;
2742         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2743                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2744
2745         spin_lock_init(&kiblnd_data.kib_connd_lock);
2746         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2747         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2748         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
2749         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
2750
2751         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2752         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
2753
2754         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2755                                                   sizeof(*sched));
2756         if (!kiblnd_data.kib_scheds)
2757                 goto failed;
2758
2759         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2760                 int nthrs;
2761
2762                 spin_lock_init(&sched->ibs_lock);
2763                 INIT_LIST_HEAD(&sched->ibs_conns);
2764                 init_waitqueue_head(&sched->ibs_waitq);
2765
2766                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2767                 if (*kiblnd_tunables.kib_nscheds > 0) {
2768                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
2769                 } else {
2770                         /*
2771                          * max to half of CPUs, another half is reserved for
2772                          * upper layer modules
2773                          */
2774                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2775                 }
2776
2777                 sched->ibs_nthreads_max = nthrs;
2778                 sched->ibs_cpt = i;
2779         }
2780
2781         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2782
2783         /* lists/ptrs/locks initialised */
2784         kiblnd_data.kib_init = IBLND_INIT_DATA;
2785         /*****************************************************/
2786
2787         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
2788         if (rc) {
2789                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2790                 goto failed;
2791         }
2792
2793         if (*kiblnd_tunables.kib_dev_failover)
2794                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
2795                                          "kiblnd_failover");
2796
2797         if (rc) {
2798                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
2799                 goto failed;
2800         }
2801
2802         /* flag everything initialised */
2803         kiblnd_data.kib_init = IBLND_INIT_ALL;
2804         /*****************************************************/
2805
2806         return 0;
2807
2808  failed:
2809         kiblnd_base_shutdown();
2810         return -ENETDOWN;
2811 }
2812
2813 static int kiblnd_start_schedulers(struct kib_sched_info *sched)
2814 {
2815         int rc = 0;
2816         int nthrs;
2817         int i;
2818
2819         if (!sched->ibs_nthreads) {
2820                 if (*kiblnd_tunables.kib_nscheds > 0) {
2821                         nthrs = sched->ibs_nthreads_max;
2822                 } else {
2823                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
2824                                                sched->ibs_cpt);
2825                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2826                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
2827                 }
2828         } else {
2829                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
2830                 /* increase one thread if there is new interface */
2831                 nthrs = sched->ibs_nthreads < sched->ibs_nthreads_max;
2832         }
2833
2834         for (i = 0; i < nthrs; i++) {
2835                 long id;
2836                 char name[20];
2837
2838                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
2839                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
2840                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
2841                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
2842                 if (!rc)
2843                         continue;
2844
2845                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2846                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
2847                 break;
2848         }
2849
2850         sched->ibs_nthreads += i;
2851         return rc;
2852 }
2853
2854 static int kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts,
2855                                     int ncpts)
2856 {
2857         int cpt;
2858         int rc;
2859         int i;
2860
2861         for (i = 0; i < ncpts; i++) {
2862                 struct kib_sched_info *sched;
2863
2864                 cpt = !cpts ? i : cpts[i];
2865                 sched = kiblnd_data.kib_scheds[cpt];
2866
2867                 if (!newdev && sched->ibs_nthreads > 0)
2868                         continue;
2869
2870                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
2871                 if (rc) {
2872                         CERROR("Failed to start scheduler threads for %s\n",
2873                                dev->ibd_ifname);
2874                         return rc;
2875                 }
2876         }
2877         return 0;
2878 }
2879
2880 static kib_dev_t *kiblnd_dev_search(char *ifname)
2881 {
2882         kib_dev_t *alias = NULL;
2883         kib_dev_t *dev;
2884         char *colon;
2885         char *colon2;
2886
2887         colon = strchr(ifname, ':');
2888         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
2889                 if (!strcmp(&dev->ibd_ifname[0], ifname))
2890                         return dev;
2891
2892                 if (alias)
2893                         continue;
2894
2895                 colon2 = strchr(dev->ibd_ifname, ':');
2896                 if (colon)
2897                         *colon = 0;
2898                 if (colon2)
2899                         *colon2 = 0;
2900
2901                 if (!strcmp(&dev->ibd_ifname[0], ifname))
2902                         alias = dev;
2903
2904                 if (colon)
2905                         *colon = ':';
2906                 if (colon2)
2907                         *colon2 = ':';
2908         }
2909         return alias;
2910 }
2911
2912 static int kiblnd_startup(lnet_ni_t *ni)
2913 {
2914         char *ifname;
2915         kib_dev_t *ibdev = NULL;
2916         kib_net_t *net;
2917         struct timespec64 tv;
2918         unsigned long flags;
2919         int rc;
2920         int newdev;
2921
2922         LASSERT(ni->ni_lnd == &the_o2iblnd);
2923
2924         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
2925                 rc = kiblnd_base_startup();
2926                 if (rc)
2927                         return rc;
2928         }
2929
2930         LIBCFS_ALLOC(net, sizeof(*net));
2931         ni->ni_data = net;
2932         if (!net)
2933                 goto net_failed;
2934
2935         ktime_get_real_ts64(&tv);
2936         net->ibn_incarnation = tv.tv_sec * USEC_PER_SEC +
2937                                tv.tv_nsec / NSEC_PER_USEC;
2938
2939         rc = kiblnd_tunables_setup(ni);
2940         if (rc)
2941                 goto net_failed;
2942
2943         if (ni->ni_interfaces[0]) {
2944                 /* Use the IPoIB interface specified in 'networks=' */
2945
2946                 CLASSERT(LNET_MAX_INTERFACES > 1);
2947                 if (ni->ni_interfaces[1]) {
2948                         CERROR("Multiple interfaces not supported\n");
2949                         goto failed;
2950                 }
2951
2952                 ifname = ni->ni_interfaces[0];
2953         } else {
2954                 ifname = *kiblnd_tunables.kib_default_ipif;
2955         }
2956
2957         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
2958                 CERROR("IPoIB interface name too long: %s\n", ifname);
2959                 goto failed;
2960         }
2961
2962         ibdev = kiblnd_dev_search(ifname);
2963
2964         newdev = !ibdev;
2965         /* hmm...create kib_dev even for alias */
2966         if (!ibdev || strcmp(&ibdev->ibd_ifname[0], ifname))
2967                 ibdev = kiblnd_create_dev(ifname);
2968
2969         if (!ibdev)
2970                 goto failed;
2971
2972         net->ibn_dev = ibdev;
2973         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
2974
2975         rc = kiblnd_dev_start_threads(ibdev, newdev,
2976                                       ni->ni_cpts, ni->ni_ncpts);
2977         if (rc)
2978                 goto failed;
2979
2980         rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
2981         if (rc) {
2982                 CERROR("Failed to initialize NI pools: %d\n", rc);
2983                 goto failed;
2984         }
2985
2986         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2987         ibdev->ibd_nnets++;
2988         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
2989         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2990
2991         net->ibn_init = IBLND_INIT_ALL;
2992
2993         return 0;
2994
2995 failed:
2996         if (!net->ibn_dev && ibdev)
2997                 kiblnd_destroy_dev(ibdev);
2998
2999 net_failed:
3000         kiblnd_shutdown(ni);
3001
3002         CDEBUG(D_NET, "kiblnd_startup failed\n");
3003         return -ENETDOWN;
3004 }
3005
3006 static lnd_t the_o2iblnd = {
3007         .lnd_type       = O2IBLND,
3008         .lnd_startup    = kiblnd_startup,
3009         .lnd_shutdown   = kiblnd_shutdown,
3010         .lnd_ctl        = kiblnd_ctl,
3011         .lnd_query      = kiblnd_query,
3012         .lnd_send       = kiblnd_send,
3013         .lnd_recv       = kiblnd_recv,
3014 };
3015
3016 static void __exit ko2iblnd_exit(void)
3017 {
3018         lnet_unregister_lnd(&the_o2iblnd);
3019 }
3020
3021 static int __init ko2iblnd_init(void)
3022 {
3023         CLASSERT(sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3024         CLASSERT(offsetof(kib_msg_t,
3025                           ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3026                           <= IBLND_MSG_SIZE);
3027         CLASSERT(offsetof(kib_msg_t,
3028                           ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3029                           <= IBLND_MSG_SIZE);
3030
3031         kiblnd_tunables_init();
3032
3033         lnet_register_lnd(&the_o2iblnd);
3034
3035         return 0;
3036 }
3037
3038 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3039 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3040 MODULE_VERSION("2.7.0");
3041 MODULE_LICENSE("GPL");
3042
3043 module_init(ko2iblnd_init);
3044 module_exit(ko2iblnd_exit);