drivers/staging/lustre/lnet/klnds/o2iblnd/o2iblnd.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2015, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lnet/klnds/o2iblnd/o2iblnd.c
  37  *
  38  * Author: Eric Barton <eric@bartonsoftware.com>
  39  */
  40
  41 #include <asm/div64.h>
  42 #include <asm/page.h>
  43 #include "o2iblnd.h"
  44
  45 static lnd_t the_o2iblnd;
  46
  47 kib_data_t kiblnd_data;
  48
  49 static __u32 kiblnd_cksum(void *ptr, int nob)
  50 {
  51         char *c = ptr;
  52         __u32 sum = 0;
  53
  54         while (nob-- > 0)
  55                 sum = ((sum << 1) | (sum >> 31)) + *c++;
  56
  57         /* ensure I don't return 0 (== no checksum) */
  58         return !sum ? 1 : sum;
  59 }
  60
  61 static char *kiblnd_msgtype2str(int type)
  62 {
  63         switch (type) {
  64         case IBLND_MSG_CONNREQ:
  65                 return "CONNREQ";
  66
  67         case IBLND_MSG_CONNACK:
  68                 return "CONNACK";
  69
  70         case IBLND_MSG_NOOP:
  71                 return "NOOP";
  72
  73         case IBLND_MSG_IMMEDIATE:
  74                 return "IMMEDIATE";
  75
  76         case IBLND_MSG_PUT_REQ:
  77                 return "PUT_REQ";
  78
  79         case IBLND_MSG_PUT_NAK:
  80                 return "PUT_NAK";
  81
  82         case IBLND_MSG_PUT_ACK:
  83                 return "PUT_ACK";
  84
  85         case IBLND_MSG_PUT_DONE:
  86                 return "PUT_DONE";
  87
  88         case IBLND_MSG_GET_REQ:
  89                 return "GET_REQ";
  90
  91         case IBLND_MSG_GET_DONE:
  92                 return "GET_DONE";
  93
  94         default:
  95                 return "???";
  96         }
  97 }
  98
  99 static int kiblnd_msgtype2size(int type)
 100 {
 101         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 102
 103         switch (type) {
 104         case IBLND_MSG_CONNREQ:
 105         case IBLND_MSG_CONNACK:
 106                 return hdr_size + sizeof(kib_connparams_t);
 107
 108         case IBLND_MSG_NOOP:
 109                 return hdr_size;
 110
 111         case IBLND_MSG_IMMEDIATE:
 112                 return offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[0]);
 113
 114         case IBLND_MSG_PUT_REQ:
 115                 return hdr_size + sizeof(kib_putreq_msg_t);
 116
 117         case IBLND_MSG_PUT_ACK:
 118                 return hdr_size + sizeof(kib_putack_msg_t);
 119
 120         case IBLND_MSG_GET_REQ:
 121                 return hdr_size + sizeof(kib_get_msg_t);
 122
 123         case IBLND_MSG_PUT_NAK:
 124         case IBLND_MSG_PUT_DONE:
 125         case IBLND_MSG_GET_DONE:
 126                 return hdr_size + sizeof(kib_completion_msg_t);
 127         default:
 128                 return -1;
 129         }
 130 }
 131
 132 static int kiblnd_unpack_rd(kib_msg_t *msg, int flip)
 133 {
 134         kib_rdma_desc_t *rd;
 135         int nob;
 136         int n;
 137         int i;
 138
 139         LASSERT(msg->ibm_type == IBLND_MSG_GET_REQ ||
 140                 msg->ibm_type == IBLND_MSG_PUT_ACK);
 141
 142         rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
 143                               &msg->ibm_u.get.ibgm_rd :
 144                               &msg->ibm_u.putack.ibpam_rd;
 145
 146         if (flip) {
 147                 __swab32s(&rd->rd_key);
 148                 __swab32s(&rd->rd_nfrags);
 149         }
 150
 151         n = rd->rd_nfrags;
 152
 153         if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
 154                 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
 155                        n, IBLND_MAX_RDMA_FRAGS);
 156                 return 1;
 157         }
 158
 159         nob = offsetof(kib_msg_t, ibm_u) +
 160               kiblnd_rd_msg_size(rd, msg->ibm_type, n);
 161
 162         if (msg->ibm_nob < nob) {
 163                 CERROR("Short %s: %d(%d)\n",
 164                        kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
 165                 return 1;
 166         }
 167
 168         if (!flip)
 169                 return 0;
 170
 171         for (i = 0; i < n; i++) {
 172                 __swab32s(&rd->rd_frags[i].rf_nob);
 173                 __swab64s(&rd->rd_frags[i].rf_addr);
 174         }
 175
 176         return 0;
 177 }
 178
 179 void kiblnd_pack_msg(lnet_ni_t *ni, kib_msg_t *msg, int version,
 180                      int credits, lnet_nid_t dstnid, __u64 dststamp)
 181 {
 182         kib_net_t *net = ni->ni_data;
 183
 184         /*
 185          * CAVEAT EMPTOR! all message fields not set here should have been
 186          * initialised previously.
 187          */
 188         msg->ibm_magic    = IBLND_MSG_MAGIC;
 189         msg->ibm_version  = version;
 190         /*   ibm_type */
 191         msg->ibm_credits  = credits;
 192         /*   ibm_nob */
 193         msg->ibm_cksum    = 0;
 194         msg->ibm_srcnid   = ni->ni_nid;
 195         msg->ibm_srcstamp = net->ibn_incarnation;
 196         msg->ibm_dstnid   = dstnid;
 197         msg->ibm_dststamp = dststamp;
 198
 199         if (*kiblnd_tunables.kib_cksum) {
 200                 /* NB ibm_cksum zero while computing cksum */
 201                 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
 202         }
 203 }
 204
 205 int kiblnd_unpack_msg(kib_msg_t *msg, int nob)
 206 {
 207         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 208         __u32 msg_cksum;
 209         __u16 version;
 210         int msg_nob;
 211         int flip;
 212
 213         /* 6 bytes are enough to have received magic + version */
 214         if (nob < 6) {
 215                 CERROR("Short message: %d\n", nob);
 216                 return -EPROTO;
 217         }
 218
 219         if (msg->ibm_magic == IBLND_MSG_MAGIC) {
 220                 flip = 0;
 221         } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
 222                 flip = 1;
 223         } else {
 224                 CERROR("Bad magic: %08x\n", msg->ibm_magic);
 225                 return -EPROTO;
 226         }
 227
 228         version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
 229         if (version != IBLND_MSG_VERSION &&
 230             version != IBLND_MSG_VERSION_1) {
 231                 CERROR("Bad version: %x\n", version);
 232                 return -EPROTO;
 233         }
 234
 235         if (nob < hdr_size) {
 236                 CERROR("Short message: %d\n", nob);
 237                 return -EPROTO;
 238         }
 239
 240         msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
 241         if (msg_nob > nob) {
 242                 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
 243                 return -EPROTO;
 244         }
 245
 246         /*
 247          * checksum must be computed with ibm_cksum zero and BEFORE anything
 248          * gets flipped
 249          */
 250         msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
 251         msg->ibm_cksum = 0;
 252         if (msg_cksum &&
 253             msg_cksum != kiblnd_cksum(msg, msg_nob)) {
 254                 CERROR("Bad checksum\n");
 255                 return -EPROTO;
 256         }
 257
 258         msg->ibm_cksum = msg_cksum;
 259
 260         if (flip) {
 261                 /* leave magic unflipped as a clue to peer endianness */
 262                 msg->ibm_version = version;
 263                 CLASSERT(sizeof(msg->ibm_type) == 1);
 264                 CLASSERT(sizeof(msg->ibm_credits) == 1);
 265                 msg->ibm_nob     = msg_nob;
 266                 __swab64s(&msg->ibm_srcnid);
 267                 __swab64s(&msg->ibm_srcstamp);
 268                 __swab64s(&msg->ibm_dstnid);
 269                 __swab64s(&msg->ibm_dststamp);
 270         }
 271
 272         if (msg->ibm_srcnid == LNET_NID_ANY) {
 273                 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
 274                 return -EPROTO;
 275         }
 276
 277         if (msg_nob < kiblnd_msgtype2size(msg->ibm_type)) {
 278                 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg->ibm_type),
 279                        msg_nob, kiblnd_msgtype2size(msg->ibm_type));
 280                 return -EPROTO;
 281         }
 282
 283         switch (msg->ibm_type) {
 284         default:
 285                 CERROR("Unknown message type %x\n", msg->ibm_type);
 286                 return -EPROTO;
 287
 288         case IBLND_MSG_NOOP:
 289         case IBLND_MSG_IMMEDIATE:
 290         case IBLND_MSG_PUT_REQ:
 291                 break;
 292
 293         case IBLND_MSG_PUT_ACK:
 294         case IBLND_MSG_GET_REQ:
 295                 if (kiblnd_unpack_rd(msg, flip))
 296                         return -EPROTO;
 297                 break;
 298
 299         case IBLND_MSG_PUT_NAK:
 300         case IBLND_MSG_PUT_DONE:
 301         case IBLND_MSG_GET_DONE:
 302                 if (flip)
 303                         __swab32s(&msg->ibm_u.completion.ibcm_status);
 304                 break;
 305
 306         case IBLND_MSG_CONNREQ:
 307         case IBLND_MSG_CONNACK:
 308                 if (flip) {
 309                         __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
 310                         __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
 311                         __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
 312                 }
 313                 break;
 314         }
 315         return 0;
 316 }
 317
 318 int kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid)
 319 {
 320         kib_peer_t *peer;
 321         kib_net_t *net = ni->ni_data;
 322         int cpt = lnet_cpt_of_nid(nid);
 323         unsigned long flags;
 324
 325         LASSERT(net);
 326         LASSERT(nid != LNET_NID_ANY);
 327
 328         LIBCFS_CPT_ALLOC(peer, lnet_cpt_table(), cpt, sizeof(*peer));
 329         if (!peer) {
 330                 CERROR("Cannot allocate peer\n");
 331                 return -ENOMEM;
 332         }
 333
 334         peer->ibp_ni = ni;
 335         peer->ibp_nid = nid;
 336         peer->ibp_error = 0;
 337         peer->ibp_last_alive = 0;
 338         peer->ibp_max_frags = IBLND_CFG_RDMA_FRAGS;
 339         peer->ibp_queue_depth = *kiblnd_tunables.kib_peertxcredits;
 340         atomic_set(&peer->ibp_refcount, 1);  /* 1 ref for caller */
 341
 342         INIT_LIST_HEAD(&peer->ibp_list);     /* not in the peer table yet */
 343         INIT_LIST_HEAD(&peer->ibp_conns);
 344         INIT_LIST_HEAD(&peer->ibp_tx_queue);
 345
 346         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 347
 348         /* always called with a ref on ni, which prevents ni being shutdown */
 349         LASSERT(!net->ibn_shutdown);
 350
 351         /* npeers only grows with the global lock held */
 352         atomic_inc(&net->ibn_npeers);
 353
 354         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 355
 356         *peerp = peer;
 357         return 0;
 358 }
 359
 360 void kiblnd_destroy_peer(kib_peer_t *peer)
 361 {
 362         kib_net_t *net = peer->ibp_ni->ni_data;
 363
 364         LASSERT(net);
 365         LASSERT(!atomic_read(&peer->ibp_refcount));
 366         LASSERT(!kiblnd_peer_active(peer));
 367         LASSERT(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(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd,
1287                                     int negotiated_nfrags)
1288 {
1289         __u16 nfrags = (negotiated_nfrags != -1) ?
1290                         negotiated_nfrags : *kiblnd_tunables.kib_map_on_demand;
1291
1292         LASSERT(hdev->ibh_mrs);
1293
1294         if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1295             nfrags <= rd->rd_nfrags)
1296                 return NULL;
1297
1298         return hdev->ibh_mrs;
1299 }
1300
1301 static void kiblnd_destroy_fmr_pool(kib_fmr_pool_t *fpo)
1302 {
1303         LASSERT(!fpo->fpo_map_count);
1304
1305         if (fpo->fpo_fmr_pool)
1306                 ib_destroy_fmr_pool(fpo->fpo_fmr_pool);
1307
1308         if (fpo->fpo_hdev)
1309                 kiblnd_hdev_decref(fpo->fpo_hdev);
1310
1311         LIBCFS_FREE(fpo, sizeof(*fpo));
1312 }
1313
1314 static void kiblnd_destroy_fmr_pool_list(struct list_head *head)
1315 {
1316         kib_fmr_pool_t *fpo, *tmp;
1317
1318         list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1319                 list_del(&fpo->fpo_list);
1320                 kiblnd_destroy_fmr_pool(fpo);
1321         }
1322 }
1323
1324 static int kiblnd_fmr_pool_size(int ncpts)
1325 {
1326         int size = *kiblnd_tunables.kib_fmr_pool_size / ncpts;
1327
1328         return max(IBLND_FMR_POOL, size);
1329 }
1330
1331 static int kiblnd_fmr_flush_trigger(int ncpts)
1332 {
1333         int size = *kiblnd_tunables.kib_fmr_flush_trigger / ncpts;
1334
1335         return max(IBLND_FMR_POOL_FLUSH, size);
1336 }
1337
1338 static int kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps,
1339                                   kib_fmr_pool_t **pp_fpo)
1340 {
1341         /* FMR pool for RDMA */
1342         kib_dev_t *dev = fps->fps_net->ibn_dev;
1343         kib_fmr_pool_t *fpo;
1344         struct ib_fmr_pool_param param = {
1345                 .max_pages_per_fmr = LNET_MAX_PAYLOAD / PAGE_SIZE,
1346                 .page_shift        = PAGE_SHIFT,
1347                 .access            = (IB_ACCESS_LOCAL_WRITE |
1348                                       IB_ACCESS_REMOTE_WRITE),
1349                 .pool_size         = fps->fps_pool_size,
1350                 .dirty_watermark   = fps->fps_flush_trigger,
1351                 .flush_function    = NULL,
1352                 .flush_arg         = NULL,
1353                 .cache             = !!*kiblnd_tunables.kib_fmr_cache};
1354         int rc;
1355
1356         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1357         if (!fpo)
1358                 return -ENOMEM;
1359
1360         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1361
1362         fpo->fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, &param);
1363         if (IS_ERR(fpo->fpo_fmr_pool)) {
1364                 rc = PTR_ERR(fpo->fpo_fmr_pool);
1365                 CERROR("Failed to create FMR pool: %d\n", rc);
1366
1367                 kiblnd_hdev_decref(fpo->fpo_hdev);
1368                 LIBCFS_FREE(fpo, sizeof(*fpo));
1369                 return rc;
1370         }
1371
1372         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1373         fpo->fpo_owner    = fps;
1374         *pp_fpo = fpo;
1375
1376         return 0;
1377 }
1378
1379 static void kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps,
1380                                     struct list_head *zombies)
1381 {
1382         if (!fps->fps_net) /* intialized? */
1383                 return;
1384
1385         spin_lock(&fps->fps_lock);
1386
1387         while (!list_empty(&fps->fps_pool_list)) {
1388                 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1389                                                  kib_fmr_pool_t, fpo_list);
1390                 fpo->fpo_failed = 1;
1391                 list_del(&fpo->fpo_list);
1392                 if (!fpo->fpo_map_count)
1393                         list_add(&fpo->fpo_list, zombies);
1394                 else
1395                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1396         }
1397
1398         spin_unlock(&fps->fps_lock);
1399 }
1400
1401 static void kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1402 {
1403         if (fps->fps_net) { /* initialized? */
1404                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1405                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1406         }
1407 }
1408
1409 static int kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt,
1410                                    kib_net_t *net, int pool_size,
1411                                    int flush_trigger)
1412 {
1413         kib_fmr_pool_t *fpo;
1414         int rc;
1415
1416         memset(fps, 0, sizeof(*fps));
1417
1418         fps->fps_net = net;
1419         fps->fps_cpt = cpt;
1420         fps->fps_pool_size = pool_size;
1421         fps->fps_flush_trigger = flush_trigger;
1422         spin_lock_init(&fps->fps_lock);
1423         INIT_LIST_HEAD(&fps->fps_pool_list);
1424         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1425
1426         rc = kiblnd_create_fmr_pool(fps, &fpo);
1427         if (!rc)
1428                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1429
1430         return rc;
1431 }
1432
1433 static int kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, unsigned long now)
1434 {
1435         if (fpo->fpo_map_count) /* still in use */
1436                 return 0;
1437         if (fpo->fpo_failed)
1438                 return 1;
1439         return cfs_time_aftereq(now, fpo->fpo_deadline);
1440 }
1441
1442 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1443 {
1444         LIST_HEAD(zombies);
1445         kib_fmr_pool_t *fpo = fmr->fmr_pool;
1446         kib_fmr_poolset_t *fps = fpo->fpo_owner;
1447         unsigned long now = cfs_time_current();
1448         kib_fmr_pool_t *tmp;
1449         int rc;
1450
1451         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1452         LASSERT(!rc);
1453
1454         if (status) {
1455                 rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
1456                 LASSERT(!rc);
1457         }
1458
1459         fmr->fmr_pool = NULL;
1460         fmr->fmr_pfmr = NULL;
1461
1462         spin_lock(&fps->fps_lock);
1463         fpo->fpo_map_count--;  /* decref the pool */
1464
1465         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1466                 /* the first pool is persistent */
1467                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1468                         continue;
1469
1470                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1471                         list_move(&fpo->fpo_list, &zombies);
1472                         fps->fps_version++;
1473                 }
1474         }
1475         spin_unlock(&fps->fps_lock);
1476
1477         if (!list_empty(&zombies))
1478                 kiblnd_destroy_fmr_pool_list(&zombies);
1479 }
1480
1481 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1482                         __u64 iov, kib_fmr_t *fmr)
1483 {
1484         struct ib_pool_fmr *pfmr;
1485         kib_fmr_pool_t *fpo;
1486         __u64 version;
1487         int rc;
1488
1489  again:
1490         spin_lock(&fps->fps_lock);
1491         version = fps->fps_version;
1492         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1493                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1494                 fpo->fpo_map_count++;
1495                 spin_unlock(&fps->fps_lock);
1496
1497                 pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
1498                                             pages, npages, iov);
1499                 if (likely(!IS_ERR(pfmr))) {
1500                         fmr->fmr_pool = fpo;
1501                         fmr->fmr_pfmr = pfmr;
1502                         return 0;
1503                 }
1504
1505                 spin_lock(&fps->fps_lock);
1506                 fpo->fpo_map_count--;
1507                 if (PTR_ERR(pfmr) != -EAGAIN) {
1508                         spin_unlock(&fps->fps_lock);
1509                         return PTR_ERR(pfmr);
1510                 }
1511
1512                 /* EAGAIN and ... */
1513                 if (version != fps->fps_version) {
1514                         spin_unlock(&fps->fps_lock);
1515                         goto again;
1516                 }
1517         }
1518
1519         if (fps->fps_increasing) {
1520                 spin_unlock(&fps->fps_lock);
1521                 CDEBUG(D_NET, "Another thread is allocating new FMR pool, waiting for her to complete\n");
1522                 schedule();
1523                 goto again;
1524         }
1525
1526         if (time_before(cfs_time_current(), fps->fps_next_retry)) {
1527                 /* someone failed recently */
1528                 spin_unlock(&fps->fps_lock);
1529                 return -EAGAIN;
1530         }
1531
1532         fps->fps_increasing = 1;
1533         spin_unlock(&fps->fps_lock);
1534
1535         CDEBUG(D_NET, "Allocate new FMR pool\n");
1536         rc = kiblnd_create_fmr_pool(fps, &fpo);
1537         spin_lock(&fps->fps_lock);
1538         fps->fps_increasing = 0;
1539         if (!rc) {
1540                 fps->fps_version++;
1541                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1542         } else {
1543                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1544         }
1545         spin_unlock(&fps->fps_lock);
1546
1547         goto again;
1548 }
1549
1550 static void kiblnd_fini_pool(kib_pool_t *pool)
1551 {
1552         LASSERT(list_empty(&pool->po_free_list));
1553         LASSERT(!pool->po_allocated);
1554
1555         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1556 }
1557
1558 static void kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1559 {
1560         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1561
1562         memset(pool, 0, sizeof(*pool));
1563         INIT_LIST_HEAD(&pool->po_free_list);
1564         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1565         pool->po_owner    = ps;
1566         pool->po_size     = size;
1567 }
1568
1569 static void kiblnd_destroy_pool_list(struct list_head *head)
1570 {
1571         kib_pool_t *pool;
1572
1573         while (!list_empty(head)) {
1574                 pool = list_entry(head->next, kib_pool_t, po_list);
1575                 list_del(&pool->po_list);
1576
1577                 LASSERT(pool->po_owner);
1578                 pool->po_owner->ps_pool_destroy(pool);
1579         }
1580 }
1581
1582 static void kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
1583 {
1584         if (!ps->ps_net) /* intialized? */
1585                 return;
1586
1587         spin_lock(&ps->ps_lock);
1588         while (!list_empty(&ps->ps_pool_list)) {
1589                 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
1590                                             kib_pool_t, po_list);
1591                 po->po_failed = 1;
1592                 list_del(&po->po_list);
1593                 if (!po->po_allocated)
1594                         list_add(&po->po_list, zombies);
1595                 else
1596                         list_add(&po->po_list, &ps->ps_failed_pool_list);
1597         }
1598         spin_unlock(&ps->ps_lock);
1599 }
1600
1601 static void kiblnd_fini_poolset(kib_poolset_t *ps)
1602 {
1603         if (ps->ps_net) { /* initialized? */
1604                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1605                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1606         }
1607 }
1608
1609 static int kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
1610                                kib_net_t *net, char *name, int size,
1611                                kib_ps_pool_create_t po_create,
1612                                kib_ps_pool_destroy_t po_destroy,
1613                                kib_ps_node_init_t nd_init,
1614                                kib_ps_node_fini_t nd_fini)
1615 {
1616         kib_pool_t *pool;
1617         int rc;
1618
1619         memset(ps, 0, sizeof(*ps));
1620
1621         ps->ps_cpt          = cpt;
1622         ps->ps_net          = net;
1623         ps->ps_pool_create  = po_create;
1624         ps->ps_pool_destroy = po_destroy;
1625         ps->ps_node_init    = nd_init;
1626         ps->ps_node_fini    = nd_fini;
1627         ps->ps_pool_size    = size;
1628         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
1629             >= sizeof(ps->ps_name))
1630                 return -E2BIG;
1631         spin_lock_init(&ps->ps_lock);
1632         INIT_LIST_HEAD(&ps->ps_pool_list);
1633         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1634
1635         rc = ps->ps_pool_create(ps, size, &pool);
1636         if (!rc)
1637                 list_add(&pool->po_list, &ps->ps_pool_list);
1638         else
1639                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1640
1641         return rc;
1642 }
1643
1644 static int kiblnd_pool_is_idle(kib_pool_t *pool, unsigned long now)
1645 {
1646         if (pool->po_allocated) /* still in use */
1647                 return 0;
1648         if (pool->po_failed)
1649                 return 1;
1650         return cfs_time_aftereq(now, pool->po_deadline);
1651 }
1652
1653 void kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
1654 {
1655         LIST_HEAD(zombies);
1656         kib_poolset_t *ps = pool->po_owner;
1657         kib_pool_t *tmp;
1658         unsigned long now = cfs_time_current();
1659
1660         spin_lock(&ps->ps_lock);
1661
1662         if (ps->ps_node_fini)
1663                 ps->ps_node_fini(pool, node);
1664
1665         LASSERT(pool->po_allocated > 0);
1666         list_add(node, &pool->po_free_list);
1667         pool->po_allocated--;
1668
1669         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1670                 /* the first pool is persistent */
1671                 if (ps->ps_pool_list.next == &pool->po_list)
1672                         continue;
1673
1674                 if (kiblnd_pool_is_idle(pool, now))
1675                         list_move(&pool->po_list, &zombies);
1676         }
1677         spin_unlock(&ps->ps_lock);
1678
1679         if (!list_empty(&zombies))
1680                 kiblnd_destroy_pool_list(&zombies);
1681 }
1682
1683 struct list_head *kiblnd_pool_alloc_node(kib_poolset_t *ps)
1684 {
1685         struct list_head *node;
1686         kib_pool_t *pool;
1687         unsigned int interval = 1;
1688         unsigned long time_before;
1689         unsigned int trips = 0;
1690         int rc;
1691
1692  again:
1693         spin_lock(&ps->ps_lock);
1694         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1695                 if (list_empty(&pool->po_free_list))
1696                         continue;
1697
1698                 pool->po_allocated++;
1699                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1700                 node = pool->po_free_list.next;
1701                 list_del(node);
1702
1703                 if (ps->ps_node_init) {
1704                         /* still hold the lock */
1705                         ps->ps_node_init(pool, node);
1706                 }
1707                 spin_unlock(&ps->ps_lock);
1708                 return node;
1709         }
1710
1711         /* no available tx pool and ... */
1712         if (ps->ps_increasing) {
1713                 /* another thread is allocating a new pool */
1714                 spin_unlock(&ps->ps_lock);
1715                 trips++;
1716                 CDEBUG(D_NET, "Another thread is allocating new %s pool, waiting %d HZs for her to complete. trips = %d\n",
1717                        ps->ps_name, interval, trips);
1718
1719                 set_current_state(TASK_INTERRUPTIBLE);
1720                 schedule_timeout(interval);
1721                 if (interval < cfs_time_seconds(1))
1722                         interval *= 2;
1723
1724                 goto again;
1725         }
1726
1727         if (time_before(cfs_time_current(), ps->ps_next_retry)) {
1728                 /* someone failed recently */
1729                 spin_unlock(&ps->ps_lock);
1730                 return NULL;
1731         }
1732
1733         ps->ps_increasing = 1;
1734         spin_unlock(&ps->ps_lock);
1735
1736         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1737         time_before = cfs_time_current();
1738         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1739         CDEBUG(D_NET, "ps_pool_create took %lu HZ to complete",
1740                cfs_time_current() - time_before);
1741
1742         spin_lock(&ps->ps_lock);
1743         ps->ps_increasing = 0;
1744         if (!rc) {
1745                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
1746         } else {
1747                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1748                 CERROR("Can't allocate new %s pool because out of memory\n",
1749                        ps->ps_name);
1750         }
1751         spin_unlock(&ps->ps_lock);
1752
1753         goto again;
1754 }
1755
1756 static void kiblnd_destroy_tx_pool(kib_pool_t *pool)
1757 {
1758         kib_tx_pool_t *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
1759         int i;
1760
1761         LASSERT(!pool->po_allocated);
1762
1763         if (tpo->tpo_tx_pages) {
1764                 kiblnd_unmap_tx_pool(tpo);
1765                 kiblnd_free_pages(tpo->tpo_tx_pages);
1766         }
1767
1768         if (!tpo->tpo_tx_descs)
1769                 goto out;
1770
1771         for (i = 0; i < pool->po_size; i++) {
1772                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1773
1774                 list_del(&tx->tx_list);
1775                 if (tx->tx_pages)
1776                         LIBCFS_FREE(tx->tx_pages,
1777                                     LNET_MAX_IOV *
1778                                     sizeof(*tx->tx_pages));
1779                 if (tx->tx_frags)
1780                         LIBCFS_FREE(tx->tx_frags,
1781                                     IBLND_MAX_RDMA_FRAGS *
1782                                             sizeof(*tx->tx_frags));
1783                 if (tx->tx_wrq)
1784                         LIBCFS_FREE(tx->tx_wrq,
1785                                     (1 + IBLND_MAX_RDMA_FRAGS) *
1786                                     sizeof(*tx->tx_wrq));
1787                 if (tx->tx_sge)
1788                         LIBCFS_FREE(tx->tx_sge,
1789                                     (1 + IBLND_MAX_RDMA_FRAGS) *
1790                                     sizeof(*tx->tx_sge));
1791                 if (tx->tx_rd)
1792                         LIBCFS_FREE(tx->tx_rd,
1793                                     offsetof(kib_rdma_desc_t,
1794                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
1795         }
1796
1797         LIBCFS_FREE(tpo->tpo_tx_descs,
1798                     pool->po_size * sizeof(kib_tx_t));
1799 out:
1800         kiblnd_fini_pool(pool);
1801         LIBCFS_FREE(tpo, sizeof(*tpo));
1802 }
1803
1804 static int kiblnd_tx_pool_size(int ncpts)
1805 {
1806         int ntx = *kiblnd_tunables.kib_ntx / ncpts;
1807
1808         return max(IBLND_TX_POOL, ntx);
1809 }
1810
1811 static int kiblnd_create_tx_pool(kib_poolset_t *ps, int size,
1812                                  kib_pool_t **pp_po)
1813 {
1814         int i;
1815         int npg;
1816         kib_pool_t *pool;
1817         kib_tx_pool_t *tpo;
1818
1819         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
1820         if (!tpo) {
1821                 CERROR("Failed to allocate TX pool\n");
1822                 return -ENOMEM;
1823         }
1824
1825         pool = &tpo->tpo_pool;
1826         kiblnd_init_pool(ps, pool, size);
1827         tpo->tpo_tx_descs = NULL;
1828         tpo->tpo_tx_pages = NULL;
1829
1830         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
1831         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg)) {
1832                 CERROR("Can't allocate tx pages: %d\n", npg);
1833                 LIBCFS_FREE(tpo, sizeof(*tpo));
1834                 return -ENOMEM;
1835         }
1836
1837         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
1838                          size * sizeof(kib_tx_t));
1839         if (!tpo->tpo_tx_descs) {
1840                 CERROR("Can't allocate %d tx descriptors\n", size);
1841                 ps->ps_pool_destroy(pool);
1842                 return -ENOMEM;
1843         }
1844
1845         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
1846
1847         for (i = 0; i < size; i++) {
1848                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1849
1850                 tx->tx_pool = tpo;
1851                 if (ps->ps_net->ibn_fmr_ps) {
1852                         LIBCFS_CPT_ALLOC(tx->tx_pages,
1853                                          lnet_cpt_table(), ps->ps_cpt,
1854                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
1855                         if (!tx->tx_pages)
1856                                 break;
1857                 }
1858
1859                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
1860                                  IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags));
1861                 if (!tx->tx_frags)
1862                         break;
1863
1864                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
1865
1866                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
1867                                  (1 + IBLND_MAX_RDMA_FRAGS) *
1868                                  sizeof(*tx->tx_wrq));
1869                 if (!tx->tx_wrq)
1870                         break;
1871
1872                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
1873                                  (1 + IBLND_MAX_RDMA_FRAGS) *
1874                                  sizeof(*tx->tx_sge));
1875                 if (!tx->tx_sge)
1876                         break;
1877
1878                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
1879                                  offsetof(kib_rdma_desc_t,
1880                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
1881                 if (!tx->tx_rd)
1882                         break;
1883         }
1884
1885         if (i == size) {
1886                 kiblnd_map_tx_pool(tpo);
1887                 *pp_po = pool;
1888                 return 0;
1889         }
1890
1891         ps->ps_pool_destroy(pool);
1892         return -ENOMEM;
1893 }
1894
1895 static void kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
1896 {
1897         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
1898                                              tps_poolset);
1899         kib_tx_t *tx  = list_entry(node, kib_tx_t, tx_list);
1900
1901         tx->tx_cookie = tps->tps_next_tx_cookie++;
1902 }
1903
1904 static void kiblnd_net_fini_pools(kib_net_t *net)
1905 {
1906         int i;
1907
1908         cfs_cpt_for_each(i, lnet_cpt_table()) {
1909                 kib_tx_poolset_t *tps;
1910                 kib_fmr_poolset_t *fps;
1911
1912                 if (net->ibn_tx_ps) {
1913                         tps = net->ibn_tx_ps[i];
1914                         kiblnd_fini_poolset(&tps->tps_poolset);
1915                 }
1916
1917                 if (net->ibn_fmr_ps) {
1918                         fps = net->ibn_fmr_ps[i];
1919                         kiblnd_fini_fmr_poolset(fps);
1920                 }
1921         }
1922
1923         if (net->ibn_tx_ps) {
1924                 cfs_percpt_free(net->ibn_tx_ps);
1925                 net->ibn_tx_ps = NULL;
1926         }
1927
1928         if (net->ibn_fmr_ps) {
1929                 cfs_percpt_free(net->ibn_fmr_ps);
1930                 net->ibn_fmr_ps = NULL;
1931         }
1932 }
1933
1934 static int kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts)
1935 {
1936         unsigned long flags;
1937         int cpt;
1938         int             rc = 0;
1939         int i;
1940
1941         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1942         if (!*kiblnd_tunables.kib_map_on_demand) {
1943                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1944                 goto create_tx_pool;
1945         }
1946
1947         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1948
1949         if (*kiblnd_tunables.kib_fmr_pool_size <
1950             *kiblnd_tunables.kib_ntx / 4) {
1951                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
1952                        *kiblnd_tunables.kib_fmr_pool_size,
1953                        *kiblnd_tunables.kib_ntx / 4);
1954                 rc = -EINVAL;
1955                 goto failed;
1956         }
1957
1958         /*
1959          * TX pool must be created later than FMR, see LU-2268
1960          * for details
1961          */
1962         LASSERT(!net->ibn_tx_ps);
1963
1964         /*
1965          * premapping can fail if ibd_nmr > 1, so we always create
1966          * FMR pool and map-on-demand if premapping failed
1967          *
1968          * cfs_precpt_alloc is creating an array of struct kib_fmr_poolset
1969          * The number of struct kib_fmr_poolsets create is equal to the
1970          * number of CPTs that exist, i.e net->ibn_fmr_ps[cpt].
1971          */
1972         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
1973                                            sizeof(kib_fmr_poolset_t));
1974         if (!net->ibn_fmr_ps) {
1975                 CERROR("Failed to allocate FMR pool array\n");
1976                 rc = -ENOMEM;
1977                 goto failed;
1978         }
1979
1980         for (i = 0; i < ncpts; i++) {
1981                 cpt = !cpts ? i : cpts[i];
1982                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, net,
1983                                              kiblnd_fmr_pool_size(ncpts),
1984                                              kiblnd_fmr_flush_trigger(ncpts));
1985                 if (rc) {
1986                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
1987                                cpt, rc);
1988                         goto failed;
1989                 }
1990         }
1991
1992         if (i > 0)
1993                 LASSERT(i == ncpts);
1994
1995  create_tx_pool:
1996         /*
1997          * cfs_precpt_alloc is creating an array of struct kib_tx_poolset
1998          * The number of struct kib_tx_poolsets create is equal to the
1999          * number of CPTs that exist, i.e net->ibn_tx_ps[cpt].
2000          */
2001         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2002                                           sizeof(kib_tx_poolset_t));
2003         if (!net->ibn_tx_ps) {
2004                 CERROR("Failed to allocate tx pool array\n");
2005                 rc = -ENOMEM;
2006                 goto failed;
2007         }
2008
2009         for (i = 0; i < ncpts; i++) {
2010                 cpt = !cpts ? i : cpts[i];
2011                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2012                                          cpt, net, "TX",
2013                                          kiblnd_tx_pool_size(ncpts),
2014                                          kiblnd_create_tx_pool,
2015                                          kiblnd_destroy_tx_pool,
2016                                          kiblnd_tx_init, NULL);
2017                 if (rc) {
2018                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2019                                cpt, rc);
2020                         goto failed;
2021                 }
2022         }
2023
2024         return 0;
2025  failed:
2026         kiblnd_net_fini_pools(net);
2027         LASSERT(rc);
2028         return rc;
2029 }
2030
2031 static int kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2032 {
2033         /*
2034          * It's safe to assume a HCA can handle a page size
2035          * matching that of the native system
2036          */
2037         hdev->ibh_page_shift = PAGE_SHIFT;
2038         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2039         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2040
2041         hdev->ibh_mr_size = hdev->ibh_ibdev->attrs.max_mr_size;
2042         if (hdev->ibh_mr_size == ~0ULL) {
2043                 hdev->ibh_mr_shift = 64;
2044                 return 0;
2045         }
2046
2047         CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2048         return -EINVAL;
2049 }
2050
2051 static void kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2052 {
2053         if (!hdev->ibh_mrs)
2054                 return;
2055
2056         ib_dereg_mr(hdev->ibh_mrs);
2057
2058         hdev->ibh_mrs = NULL;
2059 }
2060
2061 void kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2062 {
2063         kiblnd_hdev_cleanup_mrs(hdev);
2064
2065         if (hdev->ibh_pd)
2066                 ib_dealloc_pd(hdev->ibh_pd);
2067
2068         if (hdev->ibh_cmid)
2069                 rdma_destroy_id(hdev->ibh_cmid);
2070
2071         LIBCFS_FREE(hdev, sizeof(*hdev));
2072 }
2073
2074 static int kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2075 {
2076         struct ib_mr *mr;
2077         int rc;
2078         int acflags = IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE;
2079
2080         rc = kiblnd_hdev_get_attr(hdev);
2081         if (rc)
2082                 return rc;
2083
2084         mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2085         if (IS_ERR(mr)) {
2086                 CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
2087                 kiblnd_hdev_cleanup_mrs(hdev);
2088                 return PTR_ERR(mr);
2089         }
2090
2091         hdev->ibh_mrs = mr;
2092
2093         return 0;
2094 }
2095
2096 /* DUMMY */
2097 static int kiblnd_dummy_callback(struct rdma_cm_id *cmid,
2098                                  struct rdma_cm_event *event)
2099 {
2100         return 0;
2101 }
2102
2103 static int kiblnd_dev_need_failover(kib_dev_t *dev)
2104 {
2105         struct rdma_cm_id *cmid;
2106         struct sockaddr_in srcaddr;
2107         struct sockaddr_in dstaddr;
2108         int rc;
2109
2110         if (!dev->ibd_hdev || /* initializing */
2111             !dev->ibd_hdev->ibh_cmid || /* listener is dead */
2112             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2113                 return 1;
2114
2115         /*
2116          * XXX: it's UGLY, but I don't have better way to find
2117          * ib-bonding HCA failover because:
2118          *
2119          * a. no reliable CM event for HCA failover...
2120          * b. no OFED API to get ib_device for current net_device...
2121          *
2122          * We have only two choices at this point:
2123          *
2124          * a. rdma_bind_addr(), it will conflict with listener cmid
2125          * b. rdma_resolve_addr() to zero addr
2126          */
2127         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2128                                      IB_QPT_RC);
2129         if (IS_ERR(cmid)) {
2130                 rc = PTR_ERR(cmid);
2131                 CERROR("Failed to create cmid for failover: %d\n", rc);
2132                 return rc;
2133         }
2134
2135         memset(&srcaddr, 0, sizeof(srcaddr));
2136         srcaddr.sin_family = AF_INET;
2137         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2138
2139         memset(&dstaddr, 0, sizeof(dstaddr));
2140         dstaddr.sin_family = AF_INET;
2141         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2142                                (struct sockaddr *)&dstaddr, 1);
2143         if (rc || !cmid->device) {
2144                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2145                        dev->ibd_ifname, &dev->ibd_ifip,
2146                        cmid->device, rc);
2147                 rdma_destroy_id(cmid);
2148                 return rc;
2149         }
2150
2151         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2152         rdma_destroy_id(cmid);
2153
2154         return rc;
2155 }
2156
2157 int kiblnd_dev_failover(kib_dev_t *dev)
2158 {
2159         LIST_HEAD(zombie_tpo);
2160         LIST_HEAD(zombie_ppo);
2161         LIST_HEAD(zombie_fpo);
2162         struct rdma_cm_id *cmid  = NULL;
2163         kib_hca_dev_t *hdev  = NULL;
2164         struct ib_pd *pd;
2165         kib_net_t *net;
2166         struct sockaddr_in addr;
2167         unsigned long flags;
2168         int rc = 0;
2169         int i;
2170
2171         LASSERT(*kiblnd_tunables.kib_dev_failover > 1 ||
2172                 dev->ibd_can_failover || !dev->ibd_hdev);
2173
2174         rc = kiblnd_dev_need_failover(dev);
2175         if (rc <= 0)
2176                 goto out;
2177
2178         if (dev->ibd_hdev &&
2179             dev->ibd_hdev->ibh_cmid) {
2180                 /*
2181                  * XXX it's not good to close old listener at here,
2182                  * because we can fail to create new listener.
2183                  * But we have to close it now, otherwise rdma_bind_addr
2184                  * will return EADDRINUSE... How crap!
2185                  */
2186                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2187
2188                 cmid = dev->ibd_hdev->ibh_cmid;
2189                 /*
2190                  * make next schedule of kiblnd_dev_need_failover()
2191                  * return 1 for me
2192                  */
2193                 dev->ibd_hdev->ibh_cmid  = NULL;
2194                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2195
2196                 rdma_destroy_id(cmid);
2197         }
2198
2199         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2200                                      IB_QPT_RC);
2201         if (IS_ERR(cmid)) {
2202                 rc = PTR_ERR(cmid);
2203                 CERROR("Failed to create cmid for failover: %d\n", rc);
2204                 goto out;
2205         }
2206
2207         memset(&addr, 0, sizeof(addr));
2208         addr.sin_family      = AF_INET;
2209         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2210         addr.sin_port   = htons(*kiblnd_tunables.kib_service);
2211
2212         /* Bind to failover device or port */
2213         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2214         if (rc || !cmid->device) {
2215                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2216                        dev->ibd_ifname, &dev->ibd_ifip,
2217                        cmid->device, rc);
2218                 rdma_destroy_id(cmid);
2219                 goto out;
2220         }
2221
2222         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2223         if (!hdev) {
2224                 CERROR("Failed to allocate kib_hca_dev\n");
2225                 rdma_destroy_id(cmid);
2226                 rc = -ENOMEM;
2227                 goto out;
2228         }
2229
2230         atomic_set(&hdev->ibh_ref, 1);
2231         hdev->ibh_dev   = dev;
2232         hdev->ibh_cmid  = cmid;
2233         hdev->ibh_ibdev = cmid->device;
2234
2235         pd = ib_alloc_pd(cmid->device);
2236         if (IS_ERR(pd)) {
2237                 rc = PTR_ERR(pd);
2238                 CERROR("Can't allocate PD: %d\n", rc);
2239                 goto out;
2240         }
2241
2242         hdev->ibh_pd = pd;
2243
2244         rc = rdma_listen(cmid, 0);
2245         if (rc) {
2246                 CERROR("Can't start new listener: %d\n", rc);
2247                 goto out;
2248         }
2249
2250         rc = kiblnd_hdev_setup_mrs(hdev);
2251         if (rc) {
2252                 CERROR("Can't setup device: %d\n", rc);
2253                 goto out;
2254         }
2255
2256         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2257
2258         swap(dev->ibd_hdev, hdev); /* take over the refcount */
2259
2260         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2261                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2262                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2263                                             &zombie_tpo);
2264
2265                         if (net->ibn_fmr_ps)
2266                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2267                                                         &zombie_fpo);
2268                 }
2269         }
2270
2271         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2272  out:
2273         if (!list_empty(&zombie_tpo))
2274                 kiblnd_destroy_pool_list(&zombie_tpo);
2275         if (!list_empty(&zombie_ppo))
2276                 kiblnd_destroy_pool_list(&zombie_ppo);
2277         if (!list_empty(&zombie_fpo))
2278                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2279         if (hdev)
2280                 kiblnd_hdev_decref(hdev);
2281
2282         if (rc)
2283                 dev->ibd_failed_failover++;
2284         else
2285                 dev->ibd_failed_failover = 0;
2286
2287         return rc;
2288 }
2289
2290 void kiblnd_destroy_dev(kib_dev_t *dev)
2291 {
2292         LASSERT(!dev->ibd_nnets);
2293         LASSERT(list_empty(&dev->ibd_nets));
2294
2295         list_del(&dev->ibd_fail_list);
2296         list_del(&dev->ibd_list);
2297
2298         if (dev->ibd_hdev)
2299                 kiblnd_hdev_decref(dev->ibd_hdev);
2300
2301         LIBCFS_FREE(dev, sizeof(*dev));
2302 }
2303
2304 static kib_dev_t *kiblnd_create_dev(char *ifname)
2305 {
2306         struct net_device *netdev;
2307         kib_dev_t *dev;
2308         __u32 netmask;
2309         __u32 ip;
2310         int up;
2311         int rc;
2312
2313         rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2314         if (rc) {
2315                 CERROR("Can't query IPoIB interface %s: %d\n",
2316                        ifname, rc);
2317                 return NULL;
2318         }
2319
2320         if (!up) {
2321                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2322                 return NULL;
2323         }
2324
2325         LIBCFS_ALLOC(dev, sizeof(*dev));
2326         if (!dev)
2327                 return NULL;
2328
2329         netdev = dev_get_by_name(&init_net, ifname);
2330         if (!netdev) {
2331                 dev->ibd_can_failover = 0;
2332         } else {
2333                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2334                 dev_put(netdev);
2335         }
2336
2337         INIT_LIST_HEAD(&dev->ibd_nets);
2338         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2339         INIT_LIST_HEAD(&dev->ibd_fail_list);
2340         dev->ibd_ifip = ip;
2341         strcpy(&dev->ibd_ifname[0], ifname);
2342
2343         /* initialize the device */
2344         rc = kiblnd_dev_failover(dev);
2345         if (rc) {
2346                 CERROR("Can't initialize device: %d\n", rc);
2347                 LIBCFS_FREE(dev, sizeof(*dev));
2348                 return NULL;
2349         }
2350
2351         list_add_tail(&dev->ibd_list, &kiblnd_data.kib_devs);
2352         return dev;
2353 }
2354
2355 static void kiblnd_base_shutdown(void)
2356 {
2357         struct kib_sched_info *sched;
2358         int i;
2359
2360         LASSERT(list_empty(&kiblnd_data.kib_devs));
2361
2362         switch (kiblnd_data.kib_init) {
2363         default:
2364                 LBUG();
2365
2366         case IBLND_INIT_ALL:
2367         case IBLND_INIT_DATA:
2368                 LASSERT(kiblnd_data.kib_peers);
2369                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2370                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2371                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2372                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2373                 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2374                 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2375
2376                 /* flag threads to terminate; wake and wait for them to die */
2377                 kiblnd_data.kib_shutdown = 1;
2378
2379                 /*
2380                  * NB: we really want to stop scheduler threads net by net
2381                  * instead of the whole module, this should be improved
2382                  * with dynamic configuration LNet
2383                  */
2384                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2385                         wake_up_all(&sched->ibs_waitq);
2386
2387                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2388                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2389
2390                 i = 2;
2391                 while (atomic_read(&kiblnd_data.kib_nthreads)) {
2392                         i++;
2393                         /* power of 2 ? */
2394                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2395                                "Waiting for %d threads to terminate\n",
2396                                atomic_read(&kiblnd_data.kib_nthreads));
2397                         set_current_state(TASK_UNINTERRUPTIBLE);
2398                         schedule_timeout(cfs_time_seconds(1));
2399                 }
2400
2401                 /* fall through */
2402
2403         case IBLND_INIT_NOTHING:
2404                 break;
2405         }
2406
2407         if (kiblnd_data.kib_peers) {
2408                 LIBCFS_FREE(kiblnd_data.kib_peers,
2409                             sizeof(struct list_head) *
2410                             kiblnd_data.kib_peer_hash_size);
2411         }
2412
2413         if (kiblnd_data.kib_scheds)
2414                 cfs_percpt_free(kiblnd_data.kib_scheds);
2415
2416         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2417         module_put(THIS_MODULE);
2418 }
2419
2420 static void kiblnd_shutdown(lnet_ni_t *ni)
2421 {
2422         kib_net_t *net = ni->ni_data;
2423         rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2424         int i;
2425         unsigned long flags;
2426
2427         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2428
2429         if (!net)
2430                 goto out;
2431
2432         write_lock_irqsave(g_lock, flags);
2433         net->ibn_shutdown = 1;
2434         write_unlock_irqrestore(g_lock, flags);
2435
2436         switch (net->ibn_init) {
2437         default:
2438                 LBUG();
2439
2440         case IBLND_INIT_ALL:
2441                 /* nuke all existing peers within this net */
2442                 kiblnd_del_peer(ni, LNET_NID_ANY);
2443
2444                 /* Wait for all peer state to clean up */
2445                 i = 2;
2446                 while (atomic_read(&net->ibn_npeers)) {
2447                         i++;
2448                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */
2449                                "%s: waiting for %d peers to disconnect\n",
2450                                libcfs_nid2str(ni->ni_nid),
2451                                atomic_read(&net->ibn_npeers));
2452                         set_current_state(TASK_UNINTERRUPTIBLE);
2453                         schedule_timeout(cfs_time_seconds(1));
2454                 }
2455
2456                 kiblnd_net_fini_pools(net);
2457
2458                 write_lock_irqsave(g_lock, flags);
2459                 LASSERT(net->ibn_dev->ibd_nnets > 0);
2460                 net->ibn_dev->ibd_nnets--;
2461                 list_del(&net->ibn_list);
2462                 write_unlock_irqrestore(g_lock, flags);
2463
2464                 /* fall through */
2465
2466         case IBLND_INIT_NOTHING:
2467                 LASSERT(!atomic_read(&net->ibn_nconns));
2468
2469                 if (net->ibn_dev && !net->ibn_dev->ibd_nnets)
2470                         kiblnd_destroy_dev(net->ibn_dev);
2471
2472                 break;
2473         }
2474
2475         net->ibn_init = IBLND_INIT_NOTHING;
2476         ni->ni_data = NULL;
2477
2478         LIBCFS_FREE(net, sizeof(*net));
2479
2480 out:
2481         if (list_empty(&kiblnd_data.kib_devs))
2482                 kiblnd_base_shutdown();
2483 }
2484
2485 static int kiblnd_base_startup(void)
2486 {
2487         struct kib_sched_info *sched;
2488         int rc;
2489         int i;
2490
2491         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2492
2493         try_module_get(THIS_MODULE);
2494         /* zero pointers, flags etc */
2495         memset(&kiblnd_data, 0, sizeof(kiblnd_data));
2496
2497         rwlock_init(&kiblnd_data.kib_global_lock);
2498
2499         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2500         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2501
2502         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2503         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2504                      sizeof(struct list_head) * kiblnd_data.kib_peer_hash_size);
2505         if (!kiblnd_data.kib_peers)
2506                 goto failed;
2507         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2508                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2509
2510         spin_lock_init(&kiblnd_data.kib_connd_lock);
2511         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2512         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2513         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
2514         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
2515
2516         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2517         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
2518
2519         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2520                                                   sizeof(*sched));
2521         if (!kiblnd_data.kib_scheds)
2522                 goto failed;
2523
2524         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2525                 int nthrs;
2526
2527                 spin_lock_init(&sched->ibs_lock);
2528                 INIT_LIST_HEAD(&sched->ibs_conns);
2529                 init_waitqueue_head(&sched->ibs_waitq);
2530
2531                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2532                 if (*kiblnd_tunables.kib_nscheds > 0) {
2533                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
2534                 } else {
2535                         /*
2536                          * max to half of CPUs, another half is reserved for
2537                          * upper layer modules
2538                          */
2539                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2540                 }
2541
2542                 sched->ibs_nthreads_max = nthrs;
2543                 sched->ibs_cpt = i;
2544         }
2545
2546         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2547
2548         /* lists/ptrs/locks initialised */
2549         kiblnd_data.kib_init = IBLND_INIT_DATA;
2550         /*****************************************************/
2551
2552         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
2553         if (rc) {
2554                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2555                 goto failed;
2556         }
2557
2558         if (*kiblnd_tunables.kib_dev_failover)
2559                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
2560                                          "kiblnd_failover");
2561
2562         if (rc) {
2563                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
2564                 goto failed;
2565         }
2566
2567         /* flag everything initialised */
2568         kiblnd_data.kib_init = IBLND_INIT_ALL;
2569         /*****************************************************/
2570
2571         return 0;
2572
2573  failed:
2574         kiblnd_base_shutdown();
2575         return -ENETDOWN;
2576 }
2577
2578 static int kiblnd_start_schedulers(struct kib_sched_info *sched)
2579 {
2580         int rc = 0;
2581         int nthrs;
2582         int i;
2583
2584         if (!sched->ibs_nthreads) {
2585                 if (*kiblnd_tunables.kib_nscheds > 0) {
2586                         nthrs = sched->ibs_nthreads_max;
2587                 } else {
2588                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
2589                                                sched->ibs_cpt);
2590                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2591                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
2592                 }
2593         } else {
2594                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
2595                 /* increase one thread if there is new interface */
2596                 nthrs = sched->ibs_nthreads < sched->ibs_nthreads_max;
2597         }
2598
2599         for (i = 0; i < nthrs; i++) {
2600                 long id;
2601                 char name[20];
2602
2603                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
2604                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
2605                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
2606                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
2607                 if (!rc)
2608                         continue;
2609
2610                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2611                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
2612                 break;
2613         }
2614
2615         sched->ibs_nthreads += i;
2616         return rc;
2617 }
2618
2619 static int kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts,
2620                                     int ncpts)
2621 {
2622         int cpt;
2623         int rc;
2624         int i;
2625
2626         for (i = 0; i < ncpts; i++) {
2627                 struct kib_sched_info *sched;
2628
2629                 cpt = !cpts ? i : cpts[i];
2630                 sched = kiblnd_data.kib_scheds[cpt];
2631
2632                 if (!newdev && sched->ibs_nthreads > 0)
2633                         continue;
2634
2635                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
2636                 if (rc) {
2637                         CERROR("Failed to start scheduler threads for %s\n",
2638                                dev->ibd_ifname);
2639                         return rc;
2640                 }
2641         }
2642         return 0;
2643 }
2644
2645 static kib_dev_t *kiblnd_dev_search(char *ifname)
2646 {
2647         kib_dev_t *alias = NULL;
2648         kib_dev_t *dev;
2649         char *colon;
2650         char *colon2;
2651
2652         colon = strchr(ifname, ':');
2653         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
2654                 if (!strcmp(&dev->ibd_ifname[0], ifname))
2655                         return dev;
2656
2657                 if (alias)
2658                         continue;
2659
2660                 colon2 = strchr(dev->ibd_ifname, ':');
2661                 if (colon)
2662                         *colon = 0;
2663                 if (colon2)
2664                         *colon2 = 0;
2665
2666                 if (!strcmp(&dev->ibd_ifname[0], ifname))
2667                         alias = dev;
2668
2669                 if (colon)
2670                         *colon = ':';
2671                 if (colon2)
2672                         *colon2 = ':';
2673         }
2674         return alias;
2675 }
2676
2677 static int kiblnd_startup(lnet_ni_t *ni)
2678 {
2679         char *ifname;
2680         kib_dev_t *ibdev = NULL;
2681         kib_net_t *net;
2682         struct timespec64 tv;
2683         unsigned long flags;
2684         int rc;
2685         int newdev;
2686
2687         LASSERT(ni->ni_lnd == &the_o2iblnd);
2688
2689         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
2690                 rc = kiblnd_base_startup();
2691                 if (rc)
2692                         return rc;
2693         }
2694
2695         LIBCFS_ALLOC(net, sizeof(*net));
2696         ni->ni_data = net;
2697         if (!net)
2698                 goto net_failed;
2699
2700         ktime_get_real_ts64(&tv);
2701         net->ibn_incarnation = tv.tv_sec * USEC_PER_SEC +
2702                                tv.tv_nsec / NSEC_PER_USEC;
2703
2704         ni->ni_peertimeout    = *kiblnd_tunables.kib_peertimeout;
2705         ni->ni_maxtxcredits   = *kiblnd_tunables.kib_credits;
2706         ni->ni_peertxcredits  = *kiblnd_tunables.kib_peertxcredits;
2707         ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
2708
2709         if (ni->ni_interfaces[0]) {
2710                 /* Use the IPoIB interface specified in 'networks=' */
2711
2712                 CLASSERT(LNET_MAX_INTERFACES > 1);
2713                 if (ni->ni_interfaces[1]) {
2714                         CERROR("Multiple interfaces not supported\n");
2715                         goto failed;
2716                 }
2717
2718                 ifname = ni->ni_interfaces[0];
2719         } else {
2720                 ifname = *kiblnd_tunables.kib_default_ipif;
2721         }
2722
2723         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
2724                 CERROR("IPoIB interface name too long: %s\n", ifname);
2725                 goto failed;
2726         }
2727
2728         ibdev = kiblnd_dev_search(ifname);
2729
2730         newdev = !ibdev;
2731         /* hmm...create kib_dev even for alias */
2732         if (!ibdev || strcmp(&ibdev->ibd_ifname[0], ifname))
2733                 ibdev = kiblnd_create_dev(ifname);
2734
2735         if (!ibdev)
2736                 goto failed;
2737
2738         net->ibn_dev = ibdev;
2739         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
2740
2741         rc = kiblnd_dev_start_threads(ibdev, newdev,
2742                                       ni->ni_cpts, ni->ni_ncpts);
2743         if (rc)
2744                 goto failed;
2745
2746         rc = kiblnd_net_init_pools(net, ni->ni_cpts, ni->ni_ncpts);
2747         if (rc) {
2748                 CERROR("Failed to initialize NI pools: %d\n", rc);
2749                 goto failed;
2750         }
2751
2752         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2753         ibdev->ibd_nnets++;
2754         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
2755         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2756
2757         net->ibn_init = IBLND_INIT_ALL;
2758
2759         return 0;
2760
2761 failed:
2762         if (!net->ibn_dev && ibdev)
2763                 kiblnd_destroy_dev(ibdev);
2764
2765 net_failed:
2766         kiblnd_shutdown(ni);
2767
2768         CDEBUG(D_NET, "kiblnd_startup failed\n");
2769         return -ENETDOWN;
2770 }
2771
2772 static lnd_t the_o2iblnd = {
2773         .lnd_type       = O2IBLND,
2774         .lnd_startup    = kiblnd_startup,
2775         .lnd_shutdown   = kiblnd_shutdown,
2776         .lnd_ctl        = kiblnd_ctl,
2777         .lnd_query      = kiblnd_query,
2778         .lnd_send       = kiblnd_send,
2779         .lnd_recv       = kiblnd_recv,
2780 };
2781
2782 static void __exit ko2iblnd_exit(void)
2783 {
2784         lnet_unregister_lnd(&the_o2iblnd);
2785 }
2786
2787 static int __init ko2iblnd_init(void)
2788 {
2789         int rc;
2790
2791         CLASSERT(sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
2792         CLASSERT(offsetof(kib_msg_t,
2793                           ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
2794                           <= IBLND_MSG_SIZE);
2795         CLASSERT(offsetof(kib_msg_t,
2796                           ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
2797                           <= IBLND_MSG_SIZE);
2798
2799         rc = kiblnd_tunables_init();
2800         if (rc)
2801                 return rc;
2802
2803         lnet_register_lnd(&the_o2iblnd);
2804
2805         return 0;
2806 }
2807
2808 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2809 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
2810 MODULE_VERSION("2.7.0");
2811 MODULE_LICENSE("GPL");
2812
2813 module_init(ko2iblnd_init);
2814 module_exit(ko2iblnd_exit);