lnet/klnds/o2iblnd/o2iblnd.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lnet/klnds/o2iblnd/o2iblnd.c
  37  *
  38  * Author: Eric Barton <eric@bartonsoftware.com>
  39  */
  40
  41 #include <asm/page.h>
  42 #include "o2iblnd.h"
  43
  44 static lnd_t the_o2iblnd;
  45
  46 kib_data_t              kiblnd_data;
  47
  48 static __u32
  49 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 == 0) ? 1 : sum;
  59 }
  60
  61 static char *
  62 kiblnd_msgtype2str(int type)
  63 {
  64         switch (type) {
  65         case IBLND_MSG_CONNREQ:
  66                 return "CONNREQ";
  67
  68         case IBLND_MSG_CONNACK:
  69                 return "CONNACK";
  70
  71         case IBLND_MSG_NOOP:
  72                 return "NOOP";
  73
  74         case IBLND_MSG_IMMEDIATE:
  75                 return "IMMEDIATE";
  76
  77         case IBLND_MSG_PUT_REQ:
  78                 return "PUT_REQ";
  79
  80         case IBLND_MSG_PUT_NAK:
  81                 return "PUT_NAK";
  82
  83         case IBLND_MSG_PUT_ACK:
  84                 return "PUT_ACK";
  85
  86         case IBLND_MSG_PUT_DONE:
  87                 return "PUT_DONE";
  88
  89         case IBLND_MSG_GET_REQ:
  90                 return "GET_REQ";
  91
  92         case IBLND_MSG_GET_DONE:
  93                 return "GET_DONE";
  94
  95         default:
  96                 return "???";
  97         }
  98 }
  99
 100 static int
 101 kiblnd_msgtype2size(int type)
 102 {
 103         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 104
 105         switch (type) {
 106         case IBLND_MSG_CONNREQ:
 107         case IBLND_MSG_CONNACK:
 108                 return hdr_size + sizeof(kib_connparams_t);
 109
 110         case IBLND_MSG_NOOP:
 111                 return hdr_size;
 112
 113         case IBLND_MSG_IMMEDIATE:
 114                 return offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[0]);
 115
 116         case IBLND_MSG_PUT_REQ:
 117                 return hdr_size + sizeof(kib_putreq_msg_t);
 118
 119         case IBLND_MSG_PUT_ACK:
 120                 return hdr_size + sizeof(kib_putack_msg_t);
 121
 122         case IBLND_MSG_GET_REQ:
 123                 return hdr_size + sizeof(kib_get_msg_t);
 124
 125         case IBLND_MSG_PUT_NAK:
 126         case IBLND_MSG_PUT_DONE:
 127         case IBLND_MSG_GET_DONE:
 128                 return hdr_size + sizeof(kib_completion_msg_t);
 129         default:
 130                 return -1;
 131         }
 132 }
 133
 134 static int
 135 kiblnd_unpack_rd(kib_msg_t *msg, int flip)
 136 {
 137         kib_rdma_desc_t   *rd;
 138         int                nob;
 139         int                n;
 140         int                i;
 141
 142         LASSERT (msg->ibm_type == IBLND_MSG_GET_REQ ||
 143                  msg->ibm_type == IBLND_MSG_PUT_ACK);
 144
 145         rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
 146                               &msg->ibm_u.get.ibgm_rd :
 147                               &msg->ibm_u.putack.ibpam_rd;
 148
 149         if (flip) {
 150                 __swab32s(&rd->rd_key);
 151                 __swab32s(&rd->rd_nfrags);
 152         }
 153
 154         n = rd->rd_nfrags;
 155
 156         if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
 157                 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
 158                        n, IBLND_MAX_RDMA_FRAGS);
 159                 return 1;
 160         }
 161
 162         nob = offsetof (kib_msg_t, ibm_u) +
 163               kiblnd_rd_msg_size(rd, msg->ibm_type, n);
 164
 165         if (msg->ibm_nob < nob) {
 166                 CERROR("Short %s: %d(%d)\n",
 167                        kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
 168                 return 1;
 169         }
 170
 171         if (!flip)
 172                 return 0;
 173
 174         for (i = 0; i < n; i++) {
 175                 __swab32s(&rd->rd_frags[i].rf_nob);
 176                 __swab64s(&rd->rd_frags[i].rf_addr);
 177         }
 178
 179         return 0;
 180 }
 181
 182 void
 183 kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
 184                  int credits, lnet_nid_t dstnid, __u64 dststamp)
 185 {
 186         kib_net_t *net = ni->ni_data;
 187
 188         /* CAVEAT EMPTOR! all message fields not set here should have been
 189          * initialised previously. */
 190         msg->ibm_magic    = IBLND_MSG_MAGIC;
 191         msg->ibm_version  = version;
 192         /*   ibm_type */
 193         msg->ibm_credits  = credits;
 194         /*   ibm_nob */
 195         msg->ibm_cksum    = 0;
 196         msg->ibm_srcnid   = ni->ni_nid;
 197         msg->ibm_srcstamp = net->ibn_incarnation;
 198         msg->ibm_dstnid   = dstnid;
 199         msg->ibm_dststamp = dststamp;
 200
 201         if (*kiblnd_tunables.kib_cksum) {
 202                 /* NB ibm_cksum zero while computing cksum */
 203                 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
 204         }
 205 }
 206
 207 int
 208 kiblnd_unpack_msg(kib_msg_t *msg, int nob)
 209 {
 210         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 211         __u32     msg_cksum;
 212         __u16     version;
 213         int       msg_nob;
 214         int       flip;
 215
 216         /* 6 bytes are enough to have received magic + version */
 217         if (nob < 6) {
 218                 CERROR("Short message: %d\n", nob);
 219                 return -EPROTO;
 220         }
 221
 222         if (msg->ibm_magic == IBLND_MSG_MAGIC) {
 223                 flip = 0;
 224         } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
 225                 flip = 1;
 226         } else {
 227                 CERROR("Bad magic: %08x\n", msg->ibm_magic);
 228                 return -EPROTO;
 229         }
 230
 231         version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
 232         if (version != IBLND_MSG_VERSION &&
 233             version != IBLND_MSG_VERSION_1) {
 234                 CERROR("Bad version: %x\n", version);
 235                 return -EPROTO;
 236         }
 237
 238         if (nob < hdr_size) {
 239                 CERROR("Short message: %d\n", nob);
 240                 return -EPROTO;
 241         }
 242
 243         msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
 244         if (msg_nob > nob) {
 245                 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
 246                 return -EPROTO;
 247         }
 248
 249         /* checksum must be computed with ibm_cksum zero and BEFORE anything
 250          * gets flipped */
 251         msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
 252         msg->ibm_cksum = 0;
 253         if (msg_cksum != 0 &&
 254             msg_cksum != kiblnd_cksum(msg, msg_nob)) {
 255                 CERROR("Bad checksum\n");
 256                 return -EPROTO;
 257         }
 258
 259         msg->ibm_cksum = msg_cksum;
 260
 261         if (flip) {
 262                 /* leave magic unflipped as a clue to peer endianness */
 263                 msg->ibm_version = version;
 264                 CLASSERT (sizeof(msg->ibm_type) == 1);
 265                 CLASSERT (sizeof(msg->ibm_credits) == 1);
 266                 msg->ibm_nob     = msg_nob;
 267                 __swab64s(&msg->ibm_srcnid);
 268                 __swab64s(&msg->ibm_srcstamp);
 269                 __swab64s(&msg->ibm_dstnid);
 270                 __swab64s(&msg->ibm_dststamp);
 271         }
 272
 273         if (msg->ibm_srcnid == LNET_NID_ANY) {
 274                 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
 275                 return -EPROTO;
 276         }
 277
 278         if (msg_nob < kiblnd_msgtype2size(msg->ibm_type)) {
 279                 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg->ibm_type),
 280                        msg_nob, kiblnd_msgtype2size(msg->ibm_type));
 281                 return -EPROTO;
 282         }
 283
 284         switch (msg->ibm_type) {
 285         default:
 286                 CERROR("Unknown message type %x\n", msg->ibm_type);
 287                 return -EPROTO;
 288
 289         case IBLND_MSG_NOOP:
 290         case IBLND_MSG_IMMEDIATE:
 291         case IBLND_MSG_PUT_REQ:
 292                 break;
 293
 294         case IBLND_MSG_PUT_ACK:
 295         case IBLND_MSG_GET_REQ:
 296                 if (kiblnd_unpack_rd(msg, flip))
 297                         return -EPROTO;
 298                 break;
 299
 300         case IBLND_MSG_PUT_NAK:
 301         case IBLND_MSG_PUT_DONE:
 302         case IBLND_MSG_GET_DONE:
 303                 if (flip)
 304                         __swab32s(&msg->ibm_u.completion.ibcm_status);
 305                 break;
 306
 307         case IBLND_MSG_CONNREQ:
 308         case IBLND_MSG_CONNACK:
 309                 if (flip) {
 310                         __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
 311                         __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
 312                         __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
 313                 }
 314                 break;
 315         }
 316         return 0;
 317 }
 318
 319 int
 320 kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid)
 321 {
 322         kib_peer_t      *peer;
 323         kib_net_t       *net = ni->ni_data;
 324         int             cpt = lnet_cpt_of_nid(nid);
 325         unsigned long   flags;
 326
 327         LASSERT(net != NULL);
 328         LASSERT(nid != LNET_NID_ANY);
 329
 330         LIBCFS_CPT_ALLOC(peer, lnet_cpt_table(), cpt, sizeof(*peer));
 331         if (peer == NULL) {
 332                 CERROR("Cannot allocate peer\n");
 333                 return -ENOMEM;
 334         }
 335
 336         peer->ibp_ni = ni;
 337         peer->ibp_nid = nid;
 338         peer->ibp_error = 0;
 339         peer->ibp_last_alive = 0;
 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 == 0);
 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
 361 kiblnd_destroy_peer (kib_peer_t *peer)
 362 {
 363         kib_net_t *net = peer->ibp_ni->ni_data;
 364
 365         LASSERT(net != NULL);
 366         LASSERT (atomic_read(&peer->ibp_refcount) == 0);
 367         LASSERT(!kiblnd_peer_active(peer));
 368         LASSERT(peer->ibp_connecting == 0);
 369         LASSERT(peer->ibp_accepting == 0);
 370         LASSERT(list_empty(&peer->ibp_conns));
 371         LASSERT(list_empty(&peer->ibp_tx_queue));
 372
 373         LIBCFS_FREE(peer, sizeof(*peer));
 374
 375         /* NB a peer's connections keep a reference on their peer until
 376          * they are destroyed, so we can be assured that _all_ state to do
 377          * with this peer has been cleaned up when its refcount drops to
 378          * zero. */
 379         atomic_dec(&net->ibn_npeers);
 380 }
 381
 382 kib_peer_t *
 383 kiblnd_find_peer_locked (lnet_nid_t nid)
 384 {
 385         /* the caller is responsible for accounting the additional reference
 386          * that this creates */
 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
 393                 peer = list_entry(tmp, kib_peer_t, ibp_list);
 394
 395                 LASSERT(peer->ibp_connecting > 0 ||     /* creating conns */
 396                         peer->ibp_accepting > 0 ||
 397                         !list_empty(&peer->ibp_conns)); /* active conn */
 398
 399                 if (peer->ibp_nid != nid)
 400                         continue;
 401
 402                 CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n",
 403                        peer, libcfs_nid2str(nid),
 404                        atomic_read(&peer->ibp_refcount),
 405                        peer->ibp_version);
 406                 return peer;
 407         }
 408         return NULL;
 409 }
 410
 411 void
 412 kiblnd_unlink_peer_locked (kib_peer_t *peer)
 413 {
 414         LASSERT(list_empty(&peer->ibp_conns));
 415
 416         LASSERT (kiblnd_peer_active(peer));
 417         list_del_init(&peer->ibp_list);
 418         /* lose peerlist's ref */
 419         kiblnd_peer_decref(peer);
 420 }
 421
 422 static int
 423 kiblnd_get_peer_info(lnet_ni_t *ni, int index,
 424                      lnet_nid_t *nidp, int *count)
 425 {
 426         kib_peer_t              *peer;
 427         struct list_head        *ptmp;
 428         int                      i;
 429         unsigned long            flags;
 430
 431         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 432
 433         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 434
 435                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 436
 437                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 438                         LASSERT(peer->ibp_connecting > 0 ||
 439                                 peer->ibp_accepting > 0 ||
 440                                 !list_empty(&peer->ibp_conns));
 441
 442                         if (peer->ibp_ni != ni)
 443                                 continue;
 444
 445                         if (index-- > 0)
 446                                 continue;
 447
 448                         *nidp = peer->ibp_nid;
 449                         *count = atomic_read(&peer->ibp_refcount);
 450
 451                         read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 452                                                flags);
 453                         return 0;
 454                 }
 455         }
 456
 457         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 458         return -ENOENT;
 459 }
 460
 461 static void
 462 kiblnd_del_peer_locked (kib_peer_t *peer)
 463 {
 464         struct list_head        *ctmp;
 465         struct list_head        *cnxt;
 466         kib_conn_t              *conn;
 467
 468         if (list_empty(&peer->ibp_conns)) {
 469                 kiblnd_unlink_peer_locked(peer);
 470         } else {
 471                 list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 472                         conn = list_entry(ctmp, kib_conn_t, ibc_list);
 473
 474                         kiblnd_close_conn_locked(conn, 0);
 475                 }
 476                 /* NB closing peer's last conn unlinked it. */
 477         }
 478         /* NB peer now unlinked; might even be freed if the peer table had the
 479          * last ref on it. */
 480 }
 481
 482 static int
 483 kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid)
 484 {
 485         struct list_head        zombies = LIST_HEAD_INIT(zombies);
 486         struct list_head        *ptmp;
 487         struct list_head        *pnxt;
 488         kib_peer_t              *peer;
 489         int                     lo;
 490         int                     hi;
 491         int                     i;
 492         unsigned long           flags;
 493         int                     rc = -ENOENT;
 494
 495         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 496
 497         if (nid != LNET_NID_ANY) {
 498                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 499         } else {
 500                 lo = 0;
 501                 hi = kiblnd_data.kib_peer_hash_size - 1;
 502         }
 503
 504         for (i = lo; i <= hi; i++) {
 505                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 506                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 507                         LASSERT(peer->ibp_connecting > 0 ||
 508                                 peer->ibp_accepting > 0 ||
 509                                 !list_empty(&peer->ibp_conns));
 510
 511                         if (peer->ibp_ni != ni)
 512                                 continue;
 513
 514                         if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
 515                                 continue;
 516
 517                         if (!list_empty(&peer->ibp_tx_queue)) {
 518                                 LASSERT(list_empty(&peer->ibp_conns));
 519
 520                                 list_splice_init(&peer->ibp_tx_queue,
 521                                                  &zombies);
 522                         }
 523
 524                         kiblnd_del_peer_locked(peer);
 525                         rc = 0;         /* matched something */
 526                 }
 527         }
 528
 529         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 530
 531         kiblnd_txlist_done(ni, &zombies, -EIO);
 532
 533         return rc;
 534 }
 535
 536 static kib_conn_t *
 537 kiblnd_get_conn_by_idx(lnet_ni_t *ni, int index)
 538 {
 539         kib_peer_t              *peer;
 540         struct list_head        *ptmp;
 541         kib_conn_t              *conn;
 542         struct list_head        *ctmp;
 543         int                     i;
 544         unsigned long           flags;
 545
 546         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 547
 548         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 549                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 550
 551                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 552                         LASSERT(peer->ibp_connecting > 0 ||
 553                                 peer->ibp_accepting > 0 ||
 554                                 !list_empty(&peer->ibp_conns));
 555
 556                         if (peer->ibp_ni != ni)
 557                                 continue;
 558
 559                         list_for_each(ctmp, &peer->ibp_conns) {
 560                                 if (index-- > 0)
 561                                         continue;
 562
 563                                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 564                                 kiblnd_conn_addref(conn);
 565                                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 566                                                        flags);
 567                                 return conn;
 568                         }
 569                 }
 570         }
 571
 572         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 573         return NULL;
 574 }
 575
 576 static void
 577 kiblnd_debug_rx (kib_rx_t *rx)
 578 {
 579         CDEBUG(D_CONSOLE, "      %p status %d msg_type %x cred %d\n",
 580                rx, rx->rx_status, rx->rx_msg->ibm_type,
 581                rx->rx_msg->ibm_credits);
 582 }
 583
 584 static void
 585 kiblnd_debug_tx (kib_tx_t *tx)
 586 {
 587         CDEBUG(D_CONSOLE, "      %p snd %d q %d w %d rc %d dl %lx "
 588                "cookie "LPX64" msg %s%s type %x cred %d\n",
 589                tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
 590                tx->tx_status, tx->tx_deadline, tx->tx_cookie,
 591                tx->tx_lntmsg[0] == NULL ? "-" : "!",
 592                tx->tx_lntmsg[1] == NULL ? "-" : "!",
 593                tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
 594 }
 595
 596 void
 597 kiblnd_debug_conn (kib_conn_t *conn)
 598 {
 599         struct list_head        *tmp;
 600         int                     i;
 601
 602         spin_lock(&conn->ibc_lock);
 603
 604         CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s:\n",
 605                atomic_read(&conn->ibc_refcount), conn,
 606                conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 607         CDEBUG(D_CONSOLE, "   state %d nposted %d/%d cred %d o_cred %d "
 608                " r_cred %d\n", conn->ibc_state, conn->ibc_noops_posted,
 609                conn->ibc_nsends_posted, conn->ibc_credits,
 610                conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
 611         CDEBUG(D_CONSOLE, "   comms_err %d\n", conn->ibc_comms_error);
 612
 613         CDEBUG(D_CONSOLE, "   early_rxs:\n");
 614         list_for_each(tmp, &conn->ibc_early_rxs)
 615                 kiblnd_debug_rx(list_entry(tmp, kib_rx_t, rx_list));
 616
 617         CDEBUG(D_CONSOLE, "   tx_noops:\n");
 618         list_for_each(tmp, &conn->ibc_tx_noops)
 619                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 620
 621         CDEBUG(D_CONSOLE, "   tx_queue_nocred:\n");
 622         list_for_each(tmp, &conn->ibc_tx_queue_nocred)
 623                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 624
 625         CDEBUG(D_CONSOLE, "   tx_queue_rsrvd:\n");
 626         list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
 627                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 628
 629         CDEBUG(D_CONSOLE, "   tx_queue:\n");
 630         list_for_each(tmp, &conn->ibc_tx_queue)
 631                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 632
 633         CDEBUG(D_CONSOLE, "   active_txs:\n");
 634         list_for_each(tmp, &conn->ibc_active_txs)
 635                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 636
 637         CDEBUG(D_CONSOLE, "   rxs:\n");
 638         for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++)
 639                 kiblnd_debug_rx(&conn->ibc_rxs[i]);
 640
 641         spin_unlock(&conn->ibc_lock);
 642 }
 643
 644 int
 645 kiblnd_translate_mtu(int value)
 646 {
 647         switch (value) {
 648         default:
 649                 return -1;
 650         case 0:
 651                 return 0;
 652         case 256:
 653                 return IB_MTU_256;
 654         case 512:
 655                 return IB_MTU_512;
 656         case 1024:
 657                 return IB_MTU_1024;
 658         case 2048:
 659                 return IB_MTU_2048;
 660         case 4096:
 661                 return IB_MTU_4096;
 662         }
 663 }
 664
 665 static void
 666 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
 667 {
 668         int           mtu;
 669
 670         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 671         if (cmid->route.path_rec == NULL)
 672                 return;
 673
 674         mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
 675         LASSERT (mtu >= 0);
 676         if (mtu != 0)
 677                 cmid->route.path_rec->mtu = mtu;
 678 }
 679
 680 static int
 681 kiblnd_get_completion_vector(kib_conn_t *conn, int cpt)
 682 {
 683         cpumask_t       *mask;
 684         int             vectors;
 685         int             off;
 686         int             i;
 687
 688         vectors = conn->ibc_cmid->device->num_comp_vectors;
 689         if (vectors <= 1)
 690                 return 0;
 691
 692         mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
 693
 694         /* hash NID to CPU id in this partition... */
 695         off = conn->ibc_peer->ibp_nid % cpumask_weight(mask);
 696         for_each_cpu(i, mask) {
 697                 if (off-- == 0)
 698                         return i % vectors;
 699         }
 700
 701         LBUG();
 702         return 1;
 703 }
 704
 705 kib_conn_t *
 706 kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
 707                    int state, int version)
 708 {
 709         /* CAVEAT EMPTOR:
 710          * If the new conn is created successfully it takes over the caller's
 711          * ref on 'peer'.  It also "owns" 'cmid' and destroys it when it itself
 712          * is destroyed.  On failure, the caller's ref on 'peer' remains and
 713          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 714          * to destroy 'cmid' here since I'm called from the CM which still has
 715          * its ref on 'cmid'). */
 716         rwlock_t               *glock = &kiblnd_data.kib_global_lock;
 717         kib_net_t              *net = peer->ibp_ni->ni_data;
 718         kib_dev_t              *dev;
 719         struct ib_qp_init_attr *init_qp_attr;
 720         struct kib_sched_info   *sched;
 721 #ifdef HAVE_IB_CQ_INIT_ATTR
 722         struct ib_cq_init_attr  cq_attr = {};
 723 #endif
 724         kib_conn_t              *conn;
 725         struct ib_cq            *cq;
 726         unsigned long           flags;
 727         int                     cpt;
 728         int                     rc;
 729         int                     i;
 730
 731         LASSERT(net != NULL);
 732         LASSERT(!in_interrupt());
 733
 734         dev = net->ibn_dev;
 735
 736         cpt = lnet_cpt_of_nid(peer->ibp_nid);
 737         sched = kiblnd_data.kib_scheds[cpt];
 738
 739         LASSERT(sched->ibs_nthreads > 0);
 740
 741         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 742                          sizeof(*init_qp_attr));
 743         if (init_qp_attr == NULL) {
 744                 CERROR("Can't allocate qp_attr for %s\n",
 745                        libcfs_nid2str(peer->ibp_nid));
 746                 goto failed_0;
 747         }
 748
 749         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 750         if (conn == NULL) {
 751                 CERROR("Can't allocate connection for %s\n",
 752                        libcfs_nid2str(peer->ibp_nid));
 753                 goto failed_1;
 754         }
 755
 756         conn->ibc_state = IBLND_CONN_INIT;
 757         conn->ibc_version = version;
 758         conn->ibc_peer = peer;                  /* I take the caller's ref */
 759         cmid->context = conn;                   /* for future CM callbacks */
 760         conn->ibc_cmid = cmid;
 761
 762         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 763         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 764         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 765         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 766         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 767         INIT_LIST_HEAD(&conn->ibc_active_txs);
 768         spin_lock_init(&conn->ibc_lock);
 769
 770         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 771                          sizeof(*conn->ibc_connvars));
 772         if (conn->ibc_connvars == NULL) {
 773                 CERROR("Can't allocate in-progress connection state\n");
 774                 goto failed_2;
 775         }
 776
 777         write_lock_irqsave(glock, flags);
 778         if (dev->ibd_failover) {
 779                 write_unlock_irqrestore(glock, flags);
 780                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 781                 goto failed_2;
 782         }
 783
 784         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 785                 /* wakeup failover thread and teardown connection */
 786                 if (kiblnd_dev_can_failover(dev)) {
 787                         list_add_tail(&dev->ibd_fail_list,
 788                                       &kiblnd_data.kib_failed_devs);
 789                         wake_up(&kiblnd_data.kib_failover_waitq);
 790                 }
 791
 792                 write_unlock_irqrestore(glock, flags);
 793                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 794                        cmid->device->name, dev->ibd_ifname);
 795                 goto failed_2;
 796         }
 797
 798         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 799         conn->ibc_hdev = dev->ibd_hdev;
 800
 801         kiblnd_setup_mtu_locked(cmid);
 802
 803         write_unlock_irqrestore(glock, flags);
 804
 805         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 806                          IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
 807         if (conn->ibc_rxs == NULL) {
 808                 CERROR("Cannot allocate RX buffers\n");
 809                 goto failed_2;
 810         }
 811
 812         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 813                                 IBLND_RX_MSG_PAGES(version));
 814         if (rc != 0)
 815                 goto failed_2;
 816
 817         kiblnd_map_rx_descs(conn);
 818
 819 #ifdef HAVE_IB_CQ_INIT_ATTR
 820         cq_attr.cqe = IBLND_CQ_ENTRIES(version);
 821         cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
 822         cq = ib_create_cq(cmid->device,
 823                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 824                           &cq_attr);
 825 #else
 826         cq = ib_create_cq(cmid->device,
 827                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 828                           IBLND_CQ_ENTRIES(version),
 829                           kiblnd_get_completion_vector(conn, cpt));
 830 #endif
 831         if (IS_ERR(cq)) {
 832                 CERROR("Can't create CQ: %ld, cqe: %d\n",
 833                        PTR_ERR(cq), IBLND_CQ_ENTRIES(version));
 834                 goto failed_2;
 835         }
 836
 837         conn->ibc_cq = cq;
 838
 839         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 840         if (rc != 0) {
 841                 CERROR("Can't request completion notification: %d\n", rc);
 842                 goto failed_2;
 843         }
 844
 845         init_qp_attr->event_handler = kiblnd_qp_event;
 846         init_qp_attr->qp_context = conn;
 847         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(version);
 848         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(version);
 849         init_qp_attr->cap.max_send_sge = 1;
 850         init_qp_attr->cap.max_recv_sge = 1;
 851         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 852         init_qp_attr->qp_type = IB_QPT_RC;
 853         init_qp_attr->send_cq = cq;
 854         init_qp_attr->recv_cq = cq;
 855
 856         conn->ibc_sched = sched;
 857
 858         rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 859         if (rc != 0) {
 860                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
 861                        rc, init_qp_attr->cap.max_send_wr,
 862                        init_qp_attr->cap.max_recv_wr);
 863                 goto failed_2;
 864         }
 865
 866         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 867
 868         /* 1 ref for caller and each rxmsg */
 869         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(version));
 870         conn->ibc_nrx = IBLND_RX_MSGS(version);
 871
 872         /* post receives */
 873         for (i = 0; i < IBLND_RX_MSGS(version); i++) {
 874                 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
 875                                     IBLND_POSTRX_NO_CREDIT);
 876                 if (rc != 0) {
 877                         CERROR("Can't post rxmsg: %d\n", rc);
 878
 879                         /* Make posted receives complete */
 880                         kiblnd_abort_receives(conn);
 881
 882                         /* correct # of posted buffers
 883                          * NB locking needed now I'm racing with completion */
 884                         spin_lock_irqsave(&sched->ibs_lock, flags);
 885                         conn->ibc_nrx -= IBLND_RX_MSGS(version) - i;
 886                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 887
 888                         /* cmid will be destroyed by CM(ofed) after cm_callback
 889                          * returned, so we can't refer it anymore
 890                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 891                         rdma_destroy_qp(conn->ibc_cmid);
 892                         conn->ibc_cmid = NULL;
 893
 894                         /* Drop my own and unused rxbuffer refcounts */
 895                         while (i++ <= IBLND_RX_MSGS(version))
 896                                 kiblnd_conn_decref(conn);
 897
 898                         return NULL;
 899                 }
 900         }
 901
 902         /* Init successful! */
 903         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 904                  state == IBLND_CONN_PASSIVE_WAIT);
 905         conn->ibc_state = state;
 906
 907         /* 1 more conn */
 908         atomic_inc(&net->ibn_nconns);
 909         return conn;
 910
 911  failed_2:
 912         kiblnd_destroy_conn(conn);
 913  failed_1:
 914         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 915  failed_0:
 916         return NULL;
 917 }
 918
 919 void
 920 kiblnd_destroy_conn (kib_conn_t *conn)
 921 {
 922         struct rdma_cm_id *cmid = conn->ibc_cmid;
 923         kib_peer_t        *peer = conn->ibc_peer;
 924         int                rc;
 925
 926         LASSERT (!in_interrupt());
 927         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
 928         LASSERT(list_empty(&conn->ibc_early_rxs));
 929         LASSERT(list_empty(&conn->ibc_tx_noops));
 930         LASSERT(list_empty(&conn->ibc_tx_queue));
 931         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
 932         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
 933         LASSERT(list_empty(&conn->ibc_active_txs));
 934         LASSERT (conn->ibc_noops_posted == 0);
 935         LASSERT (conn->ibc_nsends_posted == 0);
 936
 937         switch (conn->ibc_state) {
 938         default:
 939                 /* conn must be completely disengaged from the network */
 940                 LBUG();
 941
 942         case IBLND_CONN_DISCONNECTED:
 943                 /* connvars should have been freed already */
 944                 LASSERT (conn->ibc_connvars == NULL);
 945                 break;
 946
 947         case IBLND_CONN_INIT:
 948                 break;
 949         }
 950
 951         /* conn->ibc_cmid might be destroyed by CM already */
 952         if (cmid != NULL && cmid->qp != NULL)
 953                 rdma_destroy_qp(cmid);
 954
 955         if (conn->ibc_cq != NULL) {
 956                 rc = ib_destroy_cq(conn->ibc_cq);
 957                 if (rc != 0)
 958                         CWARN("Error destroying CQ: %d\n", rc);
 959         }
 960
 961         if (conn->ibc_rx_pages != NULL)
 962                 kiblnd_unmap_rx_descs(conn);
 963
 964         if (conn->ibc_rxs != NULL) {
 965                 LIBCFS_FREE(conn->ibc_rxs,
 966                             IBLND_RX_MSGS(conn->ibc_version) * sizeof(kib_rx_t));
 967         }
 968
 969         if (conn->ibc_connvars != NULL)
 970                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 971
 972         if (conn->ibc_hdev != NULL)
 973                 kiblnd_hdev_decref(conn->ibc_hdev);
 974
 975         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 976         if (conn->ibc_state != IBLND_CONN_INIT) {
 977                 kib_net_t *net = peer->ibp_ni->ni_data;
 978
 979                 rdma_destroy_id(cmid);
 980                 atomic_dec(&net->ibn_nconns);
 981                 if (conn->ibc_conn_race) {
 982                         if (peer->ibp_accepting == 0 &&
 983                             !list_empty(&peer->ibp_tx_queue)) {
 984                                 kiblnd_connect_peer(peer);
 985                         } else  {
 986                                 rwlock_t *glock = &kiblnd_data.kib_global_lock;
 987                                 unsigned long flags;
 988
 989                                 write_lock_irqsave(glock, flags);
 990                                 peer->ibp_connecting--;
 991                                 write_unlock_irqrestore(glock, flags);
 992                         }
 993                 }
 994                 kiblnd_peer_decref(peer);
 995         }
 996
 997         LIBCFS_FREE(conn, sizeof(*conn));
 998 }
 999
1000 int
1001 kiblnd_close_peer_conns_locked(kib_peer_t *peer, int why)
1002 {
1003         kib_conn_t              *conn;
1004         struct list_head        *ctmp;
1005         struct list_head        *cnxt;
1006         int                     count = 0;
1007
1008         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
1009                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1010
1011                 CDEBUG(D_NET, "Closing conn -> %s, "
1012                               "version: %x, reason: %d\n",
1013                        libcfs_nid2str(peer->ibp_nid),
1014                        conn->ibc_version, why);
1015
1016                 kiblnd_close_conn_locked(conn, why);
1017                 count++;
1018         }
1019
1020         return count;
1021 }
1022
1023 int
1024 kiblnd_close_stale_conns_locked(kib_peer_t *peer,
1025                                 int version, __u64 incarnation)
1026 {
1027         kib_conn_t              *conn;
1028         struct list_head        *ctmp;
1029         struct list_head        *cnxt;
1030         int                     count = 0;
1031
1032         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
1033                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1034
1035                 if (conn->ibc_version     == version &&
1036                     conn->ibc_incarnation == incarnation)
1037                         continue;
1038
1039                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1040                               "incarnation:"LPX64"(%x, "LPX64")\n",
1041                        libcfs_nid2str(peer->ibp_nid),
1042                        conn->ibc_version, conn->ibc_incarnation,
1043                        version, incarnation);
1044
1045                 kiblnd_close_conn_locked(conn, -ESTALE);
1046                 count++;
1047         }
1048
1049         return count;
1050 }
1051
1052 static int
1053 kiblnd_close_matching_conns(lnet_ni_t *ni, lnet_nid_t nid)
1054 {
1055         kib_peer_t              *peer;
1056         struct list_head        *ptmp;
1057         struct list_head        *pnxt;
1058         int                     lo;
1059         int                     hi;
1060         int                     i;
1061         unsigned long           flags;
1062         int                     count = 0;
1063
1064         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1065
1066         if (nid != LNET_NID_ANY)
1067                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1068         else {
1069                 lo = 0;
1070                 hi = kiblnd_data.kib_peer_hash_size - 1;
1071         }
1072
1073         for (i = lo; i <= hi; i++) {
1074                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1075
1076                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
1077                         LASSERT(peer->ibp_connecting > 0 ||
1078                                 peer->ibp_accepting > 0 ||
1079                                 !list_empty(&peer->ibp_conns));
1080
1081                         if (peer->ibp_ni != ni)
1082                                 continue;
1083
1084                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1085                                 continue;
1086
1087                         count += kiblnd_close_peer_conns_locked(peer, 0);
1088                 }
1089         }
1090
1091         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1092
1093         /* wildcards always succeed */
1094         if (nid == LNET_NID_ANY)
1095                 return 0;
1096
1097         return (count == 0) ? -ENOENT : 0;
1098 }
1099
1100 static int
1101 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1102 {
1103         struct libcfs_ioctl_data *data = arg;
1104         int                       rc = -EINVAL;
1105
1106         switch(cmd) {
1107         case IOC_LIBCFS_GET_PEER: {
1108                 lnet_nid_t   nid = 0;
1109                 int          count = 0;
1110
1111                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1112                                           &nid, &count);
1113                 data->ioc_nid    = nid;
1114                 data->ioc_count  = count;
1115                 break;
1116         }
1117
1118         case IOC_LIBCFS_DEL_PEER: {
1119                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1120                 break;
1121         }
1122         case IOC_LIBCFS_GET_CONN: {
1123                 kib_conn_t *conn;
1124
1125                 rc = 0;
1126                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1127                 if (conn == NULL) {
1128                         rc = -ENOENT;
1129                         break;
1130                 }
1131
1132                 LASSERT (conn->ibc_cmid != NULL);
1133                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1134                 if (conn->ibc_cmid->route.path_rec == NULL)
1135                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1136                 else
1137                         data->ioc_u32[0] =
1138                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1139                 kiblnd_conn_decref(conn);
1140                 break;
1141         }
1142         case IOC_LIBCFS_CLOSE_CONNECTION: {
1143                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1144                 break;
1145         }
1146
1147         default:
1148                 break;
1149         }
1150
1151         return rc;
1152 }
1153
1154 static void
1155 kiblnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1156 {
1157         cfs_time_t      last_alive = 0;
1158         cfs_time_t      now = cfs_time_current();
1159         rwlock_t        *glock = &kiblnd_data.kib_global_lock;
1160         kib_peer_t      *peer;
1161         unsigned long   flags;
1162
1163         read_lock_irqsave(glock, flags);
1164
1165         peer = kiblnd_find_peer_locked(nid);
1166         if (peer != NULL) {
1167                 LASSERT(peer->ibp_connecting > 0 || /* creating conns */
1168                         peer->ibp_accepting > 0 ||
1169                         !list_empty(&peer->ibp_conns));  /* active conn */
1170                 last_alive = peer->ibp_last_alive;
1171         }
1172
1173         read_unlock_irqrestore(glock, flags);
1174
1175         if (last_alive != 0)
1176                 *when = last_alive;
1177
1178         /* peer is not persistent in hash, trigger peer creation
1179          * and connection establishment with a NULL tx */
1180         if (peer == NULL)
1181                 kiblnd_launch_tx(ni, NULL, nid);
1182
1183         CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1184                libcfs_nid2str(nid), peer,
1185                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1186         return;
1187 }
1188
1189 static void
1190 kiblnd_free_pages(kib_pages_t *p)
1191 {
1192         int     npages = p->ibp_npages;
1193         int     i;
1194
1195         for (i = 0; i < npages; i++) {
1196                 if (p->ibp_pages[i] != NULL)
1197                         __free_page(p->ibp_pages[i]);
1198         }
1199
1200         LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1201 }
1202
1203 int
1204 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1205 {
1206         kib_pages_t     *p;
1207         int             i;
1208
1209         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1210                          offsetof(kib_pages_t, ibp_pages[npages]));
1211         if (p == NULL) {
1212                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1213                 return -ENOMEM;
1214         }
1215
1216         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1217         p->ibp_npages = npages;
1218
1219         for (i = 0; i < npages; i++) {
1220                 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1221                                                      GFP_NOFS);
1222                 if (p->ibp_pages[i] == NULL) {
1223                         CERROR("Can't allocate page %d of %d\n", i, npages);
1224                         kiblnd_free_pages(p);
1225                         return -ENOMEM;
1226                 }
1227         }
1228
1229         *pp = p;
1230         return 0;
1231 }
1232
1233 void
1234 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1235 {
1236         kib_rx_t *rx;
1237         int       i;
1238
1239         LASSERT (conn->ibc_rxs != NULL);
1240         LASSERT (conn->ibc_hdev != NULL);
1241
1242         for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1243                 rx = &conn->ibc_rxs[i];
1244
1245                 LASSERT (rx->rx_nob >= 0); /* not posted */
1246
1247                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1248                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1249                                                           rx->rx_msgaddr),
1250                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1251         }
1252
1253         kiblnd_free_pages(conn->ibc_rx_pages);
1254
1255         conn->ibc_rx_pages = NULL;
1256 }
1257
1258 void
1259 kiblnd_map_rx_descs(kib_conn_t *conn)
1260 {
1261         kib_rx_t       *rx;
1262         struct page    *pg;
1263         int             pg_off;
1264         int             ipg;
1265         int             i;
1266
1267         for (pg_off = ipg = i = 0;
1268              i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1269                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1270                 rx = &conn->ibc_rxs[i];
1271
1272                 rx->rx_conn = conn;
1273                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1274
1275                 rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1276                                                        rx->rx_msg, IBLND_MSG_SIZE,
1277                                                        DMA_FROM_DEVICE);
1278                 LASSERT (!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1279                                                    rx->rx_msgaddr));
1280                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1281
1282                 CDEBUG(D_NET, "rx %d: %p "LPX64"("LPX64")\n",
1283                        i, rx->rx_msg, rx->rx_msgaddr,
1284                        (__u64)(page_to_phys(pg) + pg_off));
1285
1286                 pg_off += IBLND_MSG_SIZE;
1287                 LASSERT (pg_off <= PAGE_SIZE);
1288
1289                 if (pg_off == PAGE_SIZE) {
1290                         pg_off = 0;
1291                         ipg++;
1292                         LASSERT (ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version));
1293                 }
1294         }
1295 }
1296
1297 static void
1298 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1299 {
1300         kib_hca_dev_t  *hdev = tpo->tpo_hdev;
1301         kib_tx_t       *tx;
1302         int             i;
1303
1304         LASSERT (tpo->tpo_pool.po_allocated == 0);
1305
1306         if (hdev == NULL)
1307                 return;
1308
1309         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1310                 tx = &tpo->tpo_tx_descs[i];
1311                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1312                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1313                                                           tx->tx_msgaddr),
1314                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1315         }
1316
1317         kiblnd_hdev_decref(hdev);
1318         tpo->tpo_hdev = NULL;
1319 }
1320
1321 static kib_hca_dev_t *
1322 kiblnd_current_hdev(kib_dev_t *dev)
1323 {
1324         kib_hca_dev_t *hdev;
1325         unsigned long  flags;
1326         int            i = 0;
1327
1328         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1329         while (dev->ibd_failover) {
1330                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1331                 if (i++ % 50 == 0)
1332                         CDEBUG(D_NET, "%s: Wait for failover\n",
1333                                dev->ibd_ifname);
1334                 schedule_timeout(cfs_time_seconds(1) / 100);
1335
1336                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1337         }
1338
1339         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1340         hdev = dev->ibd_hdev;
1341
1342         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1343
1344         return hdev;
1345 }
1346
1347 static void
1348 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1349 {
1350         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1351         kib_pool_t     *pool  = &tpo->tpo_pool;
1352         kib_net_t      *net   = pool->po_owner->ps_net;
1353         kib_dev_t      *dev;
1354         struct page    *page;
1355         kib_tx_t       *tx;
1356         int             page_offset;
1357         int             ipage;
1358         int             i;
1359
1360         LASSERT (net != NULL);
1361
1362         dev = net->ibn_dev;
1363
1364         /* pre-mapped messages are not bigger than 1 page */
1365         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1366
1367         /* No fancy arithmetic when we do the buffer calculations */
1368         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1369
1370         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1371
1372         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1373                 page = txpgs->ibp_pages[ipage];
1374                 tx = &tpo->tpo_tx_descs[i];
1375
1376                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1377                                            page_offset);
1378
1379                 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1380                                                        tx->tx_msg,
1381                                                        IBLND_MSG_SIZE,
1382                                                        DMA_TO_DEVICE);
1383                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1384                                                   tx->tx_msgaddr));
1385                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1386
1387                 list_add(&tx->tx_list, &pool->po_free_list);
1388
1389                 page_offset += IBLND_MSG_SIZE;
1390                 LASSERT(page_offset <= PAGE_SIZE);
1391
1392                 if (page_offset == PAGE_SIZE) {
1393                         page_offset = 0;
1394                         ipage++;
1395                         LASSERT(ipage <= txpgs->ibp_npages);
1396                 }
1397         }
1398 }
1399
1400 struct ib_mr *
1401 kiblnd_find_dma_mr(kib_hca_dev_t *hdev, __u64 addr, __u64 size)
1402 {
1403         __u64   index;
1404
1405         LASSERT (hdev->ibh_mrs[0] != NULL);
1406
1407         if (hdev->ibh_nmrs == 1)
1408                 return hdev->ibh_mrs[0];
1409
1410         index = addr >> hdev->ibh_mr_shift;
1411
1412         if (index <  hdev->ibh_nmrs &&
1413             index == ((addr + size - 1) >> hdev->ibh_mr_shift))
1414                 return hdev->ibh_mrs[index];
1415
1416         return NULL;
1417 }
1418
1419 struct ib_mr *
1420 kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd)
1421 {
1422         struct ib_mr *prev_mr;
1423         struct ib_mr *mr;
1424         int           i;
1425
1426         LASSERT (hdev->ibh_mrs[0] != NULL);
1427
1428         if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1429             *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags)
1430                 return NULL;
1431
1432         if (hdev->ibh_nmrs == 1)
1433                 return hdev->ibh_mrs[0];
1434
1435         for (i = 0, mr = prev_mr = NULL;
1436              i < rd->rd_nfrags; i++) {
1437                 mr = kiblnd_find_dma_mr(hdev,
1438                                         rd->rd_frags[i].rf_addr,
1439                                         rd->rd_frags[i].rf_nob);
1440                 if (prev_mr == NULL)
1441                         prev_mr = mr;
1442
1443                 if (mr == NULL || prev_mr != mr) {
1444                         /* Can't covered by one single MR */
1445                         mr = NULL;
1446                         break;
1447                 }
1448         }
1449
1450         return mr;
1451 }
1452
1453 static void
1454 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
1455 {
1456         LASSERT (pool->fpo_map_count == 0);
1457
1458         if (pool->fpo_fmr_pool != NULL)
1459                 ib_destroy_fmr_pool(pool->fpo_fmr_pool);
1460
1461         if (pool->fpo_hdev != NULL)
1462                 kiblnd_hdev_decref(pool->fpo_hdev);
1463
1464         LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t));
1465 }
1466
1467 static void
1468 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1469 {
1470         kib_fmr_pool_t *pool;
1471
1472         while (!list_empty(head)) {
1473                 pool = list_entry(head->next, kib_fmr_pool_t, fpo_list);
1474                 list_del(&pool->fpo_list);
1475                 kiblnd_destroy_fmr_pool(pool);
1476         }
1477 }
1478
1479 static int kiblnd_fmr_pool_size(int ncpts)
1480 {
1481         int size = *kiblnd_tunables.kib_fmr_pool_size / ncpts;
1482
1483         return max(IBLND_FMR_POOL, size);
1484 }
1485
1486 static int kiblnd_fmr_flush_trigger(int ncpts)
1487 {
1488         int size = *kiblnd_tunables.kib_fmr_flush_trigger / ncpts;
1489
1490         return max(IBLND_FMR_POOL_FLUSH, size);
1491 }
1492
1493 static int
1494 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1495 {
1496         /* FMR pool for RDMA */
1497         kib_dev_t               *dev = fps->fps_net->ibn_dev;
1498         kib_fmr_pool_t          *fpo;
1499         struct ib_fmr_pool_param param = {
1500                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1501                 .page_shift        = PAGE_SHIFT,
1502                 .access            = (IB_ACCESS_LOCAL_WRITE |
1503                                       IB_ACCESS_REMOTE_WRITE),
1504                 .pool_size         = fps->fps_pool_size,
1505                 .dirty_watermark   = fps->fps_flush_trigger,
1506                 .flush_function    = NULL,
1507                 .flush_arg         = NULL,
1508                 .cache             = !!*kiblnd_tunables.kib_fmr_cache};
1509         int rc;
1510
1511         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1512         if (fpo == NULL)
1513                 return -ENOMEM;
1514
1515         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1516
1517         fpo->fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, &param);
1518         if (IS_ERR(fpo->fpo_fmr_pool)) {
1519                 rc = PTR_ERR(fpo->fpo_fmr_pool);
1520                 CERROR("Failed to create FMR pool: %d\n", rc);
1521
1522                 kiblnd_hdev_decref(fpo->fpo_hdev);
1523                 LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t));
1524                 return rc;
1525         }
1526
1527         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1528         fpo->fpo_owner    = fps;
1529         *pp_fpo = fpo;
1530
1531         return 0;
1532 }
1533
1534 static void
1535 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies)
1536 {
1537         if (fps->fps_net == NULL) /* intialized? */
1538                 return;
1539
1540         spin_lock(&fps->fps_lock);
1541
1542         while (!list_empty(&fps->fps_pool_list)) {
1543                 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1544                                                  kib_fmr_pool_t, fpo_list);
1545                 fpo->fpo_failed = 1;
1546                 list_del(&fpo->fpo_list);
1547                 if (fpo->fpo_map_count == 0)
1548                         list_add(&fpo->fpo_list, zombies);
1549                 else
1550                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1551         }
1552
1553         spin_unlock(&fps->fps_lock);
1554 }
1555
1556 static void
1557 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1558 {
1559         if (fps->fps_net != NULL) { /* initialized? */
1560                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1561                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1562         }
1563 }
1564
1565 static int
1566 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, kib_net_t *net,
1567                         int pool_size, int flush_trigger)
1568 {
1569         kib_fmr_pool_t *fpo;
1570         int             rc;
1571
1572         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1573
1574         fps->fps_net = net;
1575         fps->fps_cpt = cpt;
1576         fps->fps_pool_size = pool_size;
1577         fps->fps_flush_trigger = flush_trigger;
1578         spin_lock_init(&fps->fps_lock);
1579         INIT_LIST_HEAD(&fps->fps_pool_list);
1580         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1581
1582         rc = kiblnd_create_fmr_pool(fps, &fpo);
1583         if (rc == 0)
1584                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1585
1586         return rc;
1587 }
1588
1589 static int
1590 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1591 {
1592         if (fpo->fpo_map_count != 0) /* still in use */
1593                 return 0;
1594         if (fpo->fpo_failed)
1595                 return 1;
1596         return cfs_time_aftereq(now, fpo->fpo_deadline);
1597 }
1598
1599 void
1600 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1601 {
1602         struct list_head   zombies = LIST_HEAD_INIT(zombies);
1603         kib_fmr_pool_t    *fpo = fmr->fmr_pool;
1604         kib_fmr_poolset_t *fps = fpo->fpo_owner;
1605         cfs_time_t         now = cfs_time_current();
1606         kib_fmr_pool_t    *tmp;
1607         int                rc;
1608
1609         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1610         LASSERT(rc == 0);
1611
1612         if (status != 0) {
1613                 rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
1614                 LASSERT(rc == 0);
1615         }
1616
1617         fmr->fmr_pool = NULL;
1618         fmr->fmr_pfmr = NULL;
1619
1620         spin_lock(&fps->fps_lock);
1621         fpo->fpo_map_count--;   /* decref the pool */
1622
1623         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1624                 /* the first pool is persistent */
1625                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1626                         continue;
1627
1628                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1629                         list_move(&fpo->fpo_list, &zombies);
1630                         fps->fps_version++;
1631                 }
1632         }
1633         spin_unlock(&fps->fps_lock);
1634
1635         if (!list_empty(&zombies))
1636                 kiblnd_destroy_fmr_pool_list(&zombies);
1637 }
1638
1639 int
1640 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1641                     __u64 iov, kib_fmr_t *fmr)
1642 {
1643         struct ib_pool_fmr *pfmr;
1644         kib_fmr_pool_t     *fpo;
1645         __u64               version;
1646         int                 rc;
1647
1648 again:
1649         spin_lock(&fps->fps_lock);
1650         version = fps->fps_version;
1651         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1652                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1653                 fpo->fpo_map_count++;
1654                 spin_unlock(&fps->fps_lock);
1655
1656                 pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
1657                                             pages, npages, iov);
1658                 if (likely(!IS_ERR(pfmr))) {
1659                         fmr->fmr_pool = fpo;
1660                         fmr->fmr_pfmr = pfmr;
1661                         return 0;
1662                 }
1663
1664                 spin_lock(&fps->fps_lock);
1665                 fpo->fpo_map_count--;
1666                 if (PTR_ERR(pfmr) != -EAGAIN) {
1667                         spin_unlock(&fps->fps_lock);
1668                         return PTR_ERR(pfmr);
1669                 }
1670
1671                 /* EAGAIN and ... */
1672                 if (version != fps->fps_version) {
1673                         spin_unlock(&fps->fps_lock);
1674                         goto again;
1675                 }
1676         }
1677
1678         if (fps->fps_increasing) {
1679                 spin_unlock(&fps->fps_lock);
1680                 CDEBUG(D_NET, "Another thread is allocating new "
1681                        "FMR pool, waiting for her to complete\n");
1682                 schedule();
1683                 goto again;
1684
1685         }
1686
1687         if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1688                 /* someone failed recently */
1689                 spin_unlock(&fps->fps_lock);
1690                 return -EAGAIN;
1691         }
1692
1693         fps->fps_increasing = 1;
1694         spin_unlock(&fps->fps_lock);
1695
1696         CDEBUG(D_NET, "Allocate new FMR pool\n");
1697         rc = kiblnd_create_fmr_pool(fps, &fpo);
1698         spin_lock(&fps->fps_lock);
1699         fps->fps_increasing = 0;
1700         if (rc == 0) {
1701                 fps->fps_version++;
1702                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1703         } else {
1704                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1705         }
1706         spin_unlock(&fps->fps_lock);
1707
1708         goto again;
1709 }
1710
1711 static void
1712 kiblnd_fini_pool(kib_pool_t *pool)
1713 {
1714         LASSERT(list_empty(&pool->po_free_list));
1715         LASSERT(pool->po_allocated == 0);
1716
1717         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1718 }
1719
1720 static void
1721 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1722 {
1723         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1724
1725         memset(pool, 0, sizeof(kib_pool_t));
1726         INIT_LIST_HEAD(&pool->po_free_list);
1727         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1728         pool->po_owner    = ps;
1729         pool->po_size     = size;
1730 }
1731
1732 static void
1733 kiblnd_destroy_pool_list(struct list_head *head)
1734 {
1735         kib_pool_t *pool;
1736
1737         while (!list_empty(head)) {
1738                 pool = list_entry(head->next, kib_pool_t, po_list);
1739                 list_del(&pool->po_list);
1740
1741                 LASSERT(pool->po_owner != NULL);
1742                 pool->po_owner->ps_pool_destroy(pool);
1743         }
1744 }
1745
1746 static void
1747 kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
1748 {
1749         if (ps->ps_net == NULL) /* intialized? */
1750                 return;
1751
1752         spin_lock(&ps->ps_lock);
1753         while (!list_empty(&ps->ps_pool_list)) {
1754                 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
1755                                             kib_pool_t, po_list);
1756                 po->po_failed = 1;
1757                 list_del(&po->po_list);
1758                 if (po->po_allocated == 0)
1759                         list_add(&po->po_list, zombies);
1760                 else
1761                         list_add(&po->po_list, &ps->ps_failed_pool_list);
1762         }
1763         spin_unlock(&ps->ps_lock);
1764 }
1765
1766 static void
1767 kiblnd_fini_poolset(kib_poolset_t *ps)
1768 {
1769         if (ps->ps_net != NULL) { /* initialized? */
1770                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1771                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1772         }
1773 }
1774
1775 static int
1776 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
1777                     kib_net_t *net, char *name, int size,
1778                     kib_ps_pool_create_t po_create,
1779                     kib_ps_pool_destroy_t po_destroy,
1780                     kib_ps_node_init_t nd_init,
1781                     kib_ps_node_fini_t nd_fini)
1782 {
1783         kib_pool_t      *pool;
1784         int             rc;
1785
1786         memset(ps, 0, sizeof(kib_poolset_t));
1787
1788         ps->ps_cpt          = cpt;
1789         ps->ps_net          = net;
1790         ps->ps_pool_create  = po_create;
1791         ps->ps_pool_destroy = po_destroy;
1792         ps->ps_node_init    = nd_init;
1793         ps->ps_node_fini    = nd_fini;
1794         ps->ps_pool_size    = size;
1795         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
1796             >= sizeof(ps->ps_name))
1797                 return -E2BIG;
1798         spin_lock_init(&ps->ps_lock);
1799         INIT_LIST_HEAD(&ps->ps_pool_list);
1800         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1801
1802         rc = ps->ps_pool_create(ps, size, &pool);
1803         if (rc == 0)
1804                 list_add(&pool->po_list, &ps->ps_pool_list);
1805         else
1806                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1807
1808         return rc;
1809 }
1810
1811 static int
1812 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
1813 {
1814         if (pool->po_allocated != 0) /* still in use */
1815                 return 0;
1816         if (pool->po_failed)
1817                 return 1;
1818         return cfs_time_aftereq(now, pool->po_deadline);
1819 }
1820
1821 void
1822 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
1823 {
1824         struct list_head zombies = LIST_HEAD_INIT(zombies);
1825         kib_poolset_t   *ps = pool->po_owner;
1826         kib_pool_t      *tmp;
1827         cfs_time_t       now = cfs_time_current();
1828
1829         spin_lock(&ps->ps_lock);
1830
1831         if (ps->ps_node_fini != NULL)
1832                 ps->ps_node_fini(pool, node);
1833
1834         LASSERT(pool->po_allocated > 0);
1835         list_add(node, &pool->po_free_list);
1836         pool->po_allocated--;
1837
1838         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1839                 /* the first pool is persistent */
1840                 if (ps->ps_pool_list.next == &pool->po_list)
1841                         continue;
1842
1843                 if (kiblnd_pool_is_idle(pool, now))
1844                         list_move(&pool->po_list, &zombies);
1845         }
1846         spin_unlock(&ps->ps_lock);
1847
1848         if (!list_empty(&zombies))
1849                 kiblnd_destroy_pool_list(&zombies);
1850 }
1851
1852 struct list_head *
1853 kiblnd_pool_alloc_node(kib_poolset_t *ps)
1854 {
1855         struct list_head        *node;
1856         kib_pool_t              *pool;
1857         int                     rc;
1858
1859 again:
1860         spin_lock(&ps->ps_lock);
1861         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1862                 if (list_empty(&pool->po_free_list))
1863                         continue;
1864
1865                 pool->po_allocated++;
1866                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1867                 node = pool->po_free_list.next;
1868                 list_del(node);
1869
1870                 if (ps->ps_node_init != NULL) {
1871                         /* still hold the lock */
1872                         ps->ps_node_init(pool, node);
1873                 }
1874                 spin_unlock(&ps->ps_lock);
1875                 return node;
1876         }
1877
1878         /* no available tx pool and ... */
1879         if (ps->ps_increasing) {
1880                 /* another thread is allocating a new pool */
1881                 spin_unlock(&ps->ps_lock);
1882                 CDEBUG(D_NET, "Another thread is allocating new "
1883                        "%s pool, waiting for her to complete\n",
1884                        ps->ps_name);
1885                 schedule();
1886                 goto again;
1887         }
1888
1889         if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
1890                 /* someone failed recently */
1891                 spin_unlock(&ps->ps_lock);
1892                 return NULL;
1893         }
1894
1895         ps->ps_increasing = 1;
1896         spin_unlock(&ps->ps_lock);
1897
1898         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1899
1900         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1901
1902         spin_lock(&ps->ps_lock);
1903         ps->ps_increasing = 0;
1904         if (rc == 0) {
1905                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
1906         } else {
1907                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1908                 CERROR("Can't allocate new %s pool because out of memory\n",
1909                        ps->ps_name);
1910         }
1911         spin_unlock(&ps->ps_lock);
1912
1913         goto again;
1914 }
1915
1916 void
1917 kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr)
1918 {
1919         kib_pmr_pool_t      *ppo = pmr->pmr_pool;
1920         struct ib_mr        *mr  = pmr->pmr_mr;
1921
1922         pmr->pmr_mr = NULL;
1923         kiblnd_pool_free_node(&ppo->ppo_pool, &pmr->pmr_list);
1924         if (mr != NULL)
1925                 ib_dereg_mr(mr);
1926 }
1927
1928 int
1929 kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
1930                     kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr)
1931 {
1932         kib_phys_mr_t    *pmr;
1933         struct list_head *node;
1934         int               rc;
1935         int               i;
1936
1937         node = kiblnd_pool_alloc_node(&pps->pps_poolset);
1938         if (node == NULL) {
1939                 CERROR("Failed to allocate PMR descriptor\n");
1940                 return -ENOMEM;
1941         }
1942
1943         pmr = container_of(node, kib_phys_mr_t, pmr_list);
1944         if (pmr->pmr_pool->ppo_hdev != hdev) {
1945                 kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1946                 return -EAGAIN;
1947         }
1948
1949         for (i = 0; i < rd->rd_nfrags; i ++) {
1950                 pmr->pmr_ipb[i].addr = rd->rd_frags[i].rf_addr;
1951                 pmr->pmr_ipb[i].size = rd->rd_frags[i].rf_nob;
1952         }
1953
1954         pmr->pmr_mr = ib_reg_phys_mr(hdev->ibh_pd,
1955                                      pmr->pmr_ipb, rd->rd_nfrags,
1956                                      IB_ACCESS_LOCAL_WRITE |
1957                                      IB_ACCESS_REMOTE_WRITE,
1958                                      iova);
1959         if (!IS_ERR(pmr->pmr_mr)) {
1960                 pmr->pmr_iova = *iova;
1961                 *pp_pmr = pmr;
1962                 return 0;
1963         }
1964
1965         rc = PTR_ERR(pmr->pmr_mr);
1966         CERROR("Failed ib_reg_phys_mr: %d\n", rc);
1967
1968         pmr->pmr_mr = NULL;
1969         kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1970
1971         return rc;
1972 }
1973
1974 static void
1975 kiblnd_destroy_pmr_pool(kib_pool_t *pool)
1976 {
1977         kib_pmr_pool_t *ppo = container_of(pool, kib_pmr_pool_t, ppo_pool);
1978         kib_phys_mr_t  *pmr;
1979
1980         LASSERT (pool->po_allocated == 0);
1981
1982         while (!list_empty(&pool->po_free_list)) {
1983                 pmr = list_entry(pool->po_free_list.next,
1984                                      kib_phys_mr_t, pmr_list);
1985
1986                 LASSERT (pmr->pmr_mr == NULL);
1987                 list_del(&pmr->pmr_list);
1988
1989                 if (pmr->pmr_ipb != NULL) {
1990                         LIBCFS_FREE(pmr->pmr_ipb,
1991                                     IBLND_MAX_RDMA_FRAGS *
1992                                     sizeof(struct ib_phys_buf));
1993                 }
1994
1995                 LIBCFS_FREE(pmr, sizeof(kib_phys_mr_t));
1996         }
1997
1998         kiblnd_fini_pool(pool);
1999         if (ppo->ppo_hdev != NULL)
2000                 kiblnd_hdev_decref(ppo->ppo_hdev);
2001
2002         LIBCFS_FREE(ppo, sizeof(kib_pmr_pool_t));
2003 }
2004
2005 static inline int kiblnd_pmr_pool_size(int ncpts)
2006 {
2007         int size = *kiblnd_tunables.kib_pmr_pool_size / ncpts;
2008
2009         return max(IBLND_PMR_POOL, size);
2010 }
2011
2012 static int
2013 kiblnd_create_pmr_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2014 {
2015         struct kib_pmr_pool     *ppo;
2016         struct kib_pool         *pool;
2017         kib_phys_mr_t           *pmr;
2018         int                     i;
2019
2020         LIBCFS_CPT_ALLOC(ppo, lnet_cpt_table(),
2021                          ps->ps_cpt, sizeof(kib_pmr_pool_t));
2022         if (ppo == NULL) {
2023                 CERROR("Failed to allocate PMR pool\n");
2024                 return -ENOMEM;
2025         }
2026
2027         pool = &ppo->ppo_pool;
2028         kiblnd_init_pool(ps, pool, size);
2029
2030         for (i = 0; i < size; i++) {
2031                 LIBCFS_CPT_ALLOC(pmr, lnet_cpt_table(),
2032                                  ps->ps_cpt, sizeof(kib_phys_mr_t));
2033                 if (pmr == NULL)
2034                         break;
2035
2036                 pmr->pmr_pool = ppo;
2037                 LIBCFS_CPT_ALLOC(pmr->pmr_ipb, lnet_cpt_table(), ps->ps_cpt,
2038                                  IBLND_MAX_RDMA_FRAGS * sizeof(*pmr->pmr_ipb));
2039                 if (pmr->pmr_ipb == NULL)
2040                         break;
2041
2042                 list_add(&pmr->pmr_list, &pool->po_free_list);
2043         }
2044
2045         if (i < size) {
2046                 ps->ps_pool_destroy(pool);
2047                 return -ENOMEM;
2048         }
2049
2050         ppo->ppo_hdev = kiblnd_current_hdev(ps->ps_net->ibn_dev);
2051         *pp_po = pool;
2052         return 0;
2053 }
2054
2055 static void
2056 kiblnd_destroy_tx_pool(kib_pool_t *pool)
2057 {
2058         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
2059         int             i;
2060
2061         LASSERT (pool->po_allocated == 0);
2062
2063         if (tpo->tpo_tx_pages != NULL) {
2064                 kiblnd_unmap_tx_pool(tpo);
2065                 kiblnd_free_pages(tpo->tpo_tx_pages);
2066         }
2067
2068         if (tpo->tpo_tx_descs == NULL)
2069                 goto out;
2070
2071         for (i = 0; i < pool->po_size; i++) {
2072                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2073
2074                 list_del(&tx->tx_list);
2075                 if (tx->tx_pages != NULL)
2076                         LIBCFS_FREE(tx->tx_pages,
2077                                     LNET_MAX_IOV *
2078                                     sizeof(*tx->tx_pages));
2079                 if (tx->tx_frags != NULL)
2080                         LIBCFS_FREE(tx->tx_frags,
2081                                     IBLND_MAX_RDMA_FRAGS *
2082                                             sizeof(*tx->tx_frags));
2083                 if (tx->tx_wrq != NULL)
2084                         LIBCFS_FREE(tx->tx_wrq,
2085                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2086                                     sizeof(*tx->tx_wrq));
2087                 if (tx->tx_sge != NULL)
2088                         LIBCFS_FREE(tx->tx_sge,
2089                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2090                                     sizeof(*tx->tx_sge));
2091                 if (tx->tx_rd != NULL)
2092                         LIBCFS_FREE(tx->tx_rd,
2093                                     offsetof(kib_rdma_desc_t,
2094                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2095         }
2096
2097         LIBCFS_FREE(tpo->tpo_tx_descs,
2098                     pool->po_size * sizeof(kib_tx_t));
2099 out:
2100         kiblnd_fini_pool(pool);
2101         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2102 }
2103
2104 static int kiblnd_tx_pool_size(int ncpts)
2105 {
2106         int ntx = *kiblnd_tunables.kib_ntx / ncpts;
2107
2108         return max(IBLND_TX_POOL, ntx);
2109 }
2110
2111 static int
2112 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2113 {
2114         int            i;
2115         int            npg;
2116         kib_pool_t    *pool;
2117         kib_tx_pool_t *tpo;
2118
2119         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2120         if (tpo == NULL) {
2121                 CERROR("Failed to allocate TX pool\n");
2122                 return -ENOMEM;
2123         }
2124
2125         pool = &tpo->tpo_pool;
2126         kiblnd_init_pool(ps, pool, size);
2127         tpo->tpo_tx_descs = NULL;
2128         tpo->tpo_tx_pages = NULL;
2129
2130         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2131         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2132                 CERROR("Can't allocate tx pages: %d\n", npg);
2133                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2134                 return -ENOMEM;
2135         }
2136
2137         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2138                          size * sizeof(kib_tx_t));
2139         if (tpo->tpo_tx_descs == NULL) {
2140                 CERROR("Can't allocate %d tx descriptors\n", size);
2141                 ps->ps_pool_destroy(pool);
2142                 return -ENOMEM;
2143         }
2144
2145         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2146
2147         for (i = 0; i < size; i++) {
2148                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2149
2150                 tx->tx_pool = tpo;
2151                 if (ps->ps_net->ibn_fmr_ps != NULL) {
2152                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2153                                          lnet_cpt_table(), ps->ps_cpt,
2154                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2155                         if (tx->tx_pages == NULL)
2156                                 break;
2157                 }
2158
2159                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2160                                  IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags));
2161                 if (tx->tx_frags == NULL)
2162                         break;
2163
2164                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
2165
2166                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2167                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2168                                  sizeof(*tx->tx_wrq));
2169                 if (tx->tx_wrq == NULL)
2170                         break;
2171
2172                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2173                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2174                                  sizeof(*tx->tx_sge));
2175                 if (tx->tx_sge == NULL)
2176                         break;
2177
2178                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2179                                  offsetof(kib_rdma_desc_t,
2180                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2181                 if (tx->tx_rd == NULL)
2182                         break;
2183         }
2184
2185         if (i == size) {
2186                 kiblnd_map_tx_pool(tpo);
2187                 *pp_po = pool;
2188                 return 0;
2189         }
2190
2191         ps->ps_pool_destroy(pool);
2192         return -ENOMEM;
2193 }
2194
2195 static void
2196 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2197 {
2198         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2199                                              tps_poolset);
2200         kib_tx_t         *tx  = list_entry(node, kib_tx_t, tx_list);
2201
2202         tx->tx_cookie = tps->tps_next_tx_cookie++;
2203 }
2204
2205 static void
2206 kiblnd_net_fini_pools(kib_net_t *net)
2207 {
2208         int     i;
2209
2210         cfs_cpt_for_each(i, lnet_cpt_table()) {
2211                 kib_tx_poolset_t        *tps;
2212                 kib_fmr_poolset_t       *fps;
2213                 kib_pmr_poolset_t       *pps;
2214
2215                 if (net->ibn_tx_ps != NULL) {
2216                         tps = net->ibn_tx_ps[i];
2217                         kiblnd_fini_poolset(&tps->tps_poolset);
2218                 }
2219
2220                 if (net->ibn_fmr_ps != NULL) {
2221                         fps = net->ibn_fmr_ps[i];
2222                         kiblnd_fini_fmr_poolset(fps);
2223                 }
2224
2225                 if (net->ibn_pmr_ps != NULL) {
2226                         pps = net->ibn_pmr_ps[i];
2227                         kiblnd_fini_poolset(&pps->pps_poolset);
2228                 }
2229         }
2230
2231         if (net->ibn_tx_ps != NULL) {
2232                 cfs_percpt_free(net->ibn_tx_ps);
2233                 net->ibn_tx_ps = NULL;
2234         }
2235
2236         if (net->ibn_fmr_ps != NULL) {
2237                 cfs_percpt_free(net->ibn_fmr_ps);
2238                 net->ibn_fmr_ps = NULL;
2239         }
2240
2241         if (net->ibn_pmr_ps != NULL) {
2242                 cfs_percpt_free(net->ibn_pmr_ps);
2243                 net->ibn_pmr_ps = NULL;
2244         }
2245 }
2246
2247 static int
2248 kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts)
2249 {
2250         unsigned long   flags;
2251         int             cpt;
2252         int             rc;
2253         int             i;
2254
2255         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2256         if (*kiblnd_tunables.kib_map_on_demand == 0 &&
2257             net->ibn_dev->ibd_hdev->ibh_nmrs == 1) {
2258                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2259                                            flags);
2260                 goto create_tx_pool;
2261         }
2262
2263         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2264
2265         if (*kiblnd_tunables.kib_fmr_pool_size <
2266             *kiblnd_tunables.kib_ntx / 4) {
2267                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2268                        *kiblnd_tunables.kib_fmr_pool_size,
2269                        *kiblnd_tunables.kib_ntx / 4);
2270                 rc = -EINVAL;
2271                 goto failed;
2272         }
2273
2274         /* TX pool must be created later than FMR/PMR, see LU-2268
2275          * for details */
2276         LASSERT(net->ibn_tx_ps == NULL);
2277
2278         /* premapping can fail if ibd_nmr > 1, so we always create
2279          * FMR/PMR pool and map-on-demand if premapping failed */
2280
2281         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2282                                            sizeof(kib_fmr_poolset_t));
2283         if (net->ibn_fmr_ps == NULL) {
2284                 CERROR("Failed to allocate FMR pool array\n");
2285                 rc = -ENOMEM;
2286                 goto failed;
2287         }
2288
2289         for (i = 0; i < ncpts; i++) {
2290                 cpt = (cpts == NULL) ? i : cpts[i];
2291                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, net,
2292                                              kiblnd_fmr_pool_size(ncpts),
2293                                              kiblnd_fmr_flush_trigger(ncpts));
2294                 if (rc == -ENOSYS && i == 0) /* no FMR */
2295                         break; /* create PMR pool */
2296
2297                 if (rc != 0) { /* a real error */
2298                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2299                                cpt, rc);
2300                         goto failed;
2301                 }
2302         }
2303
2304         if (i > 0) {
2305                 LASSERT(i == ncpts);
2306                 goto create_tx_pool;
2307         }
2308
2309         cfs_percpt_free(net->ibn_fmr_ps);
2310         net->ibn_fmr_ps = NULL;
2311
2312         CWARN("Device does not support FMR, failing back to PMR\n");
2313
2314         if (*kiblnd_tunables.kib_pmr_pool_size <
2315             *kiblnd_tunables.kib_ntx / 4) {
2316                 CERROR("Can't set pmr pool size (%d) < ntx / 4(%d)\n",
2317                        *kiblnd_tunables.kib_pmr_pool_size,
2318                        *kiblnd_tunables.kib_ntx / 4);
2319                 rc = -EINVAL;
2320                 goto failed;
2321         }
2322
2323         net->ibn_pmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2324                                            sizeof(kib_pmr_poolset_t));
2325         if (net->ibn_pmr_ps == NULL) {
2326                 CERROR("Failed to allocate PMR pool array\n");
2327                 rc = -ENOMEM;
2328                 goto failed;
2329         }
2330
2331         for (i = 0; i < ncpts; i++) {
2332                 cpt = (cpts == NULL) ? i : cpts[i];
2333                 rc = kiblnd_init_poolset(&net->ibn_pmr_ps[cpt]->pps_poolset,
2334                                          cpt, net, "PMR",
2335                                          kiblnd_pmr_pool_size(ncpts),
2336                                          kiblnd_create_pmr_pool,
2337                                          kiblnd_destroy_pmr_pool, NULL, NULL);
2338                 if (rc != 0) {
2339                         CERROR("Can't initialize PMR pool for CPT %d: %d\n",
2340                                cpt, rc);
2341                         goto failed;
2342                 }
2343         }
2344
2345  create_tx_pool:
2346         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2347                                           sizeof(kib_tx_poolset_t));
2348         if (net->ibn_tx_ps == NULL) {
2349                 CERROR("Failed to allocate tx pool array\n");
2350                 rc = -ENOMEM;
2351                 goto failed;
2352         }
2353
2354         for (i = 0; i < ncpts; i++) {
2355                 cpt = (cpts == NULL) ? i : cpts[i];
2356                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2357                                          cpt, net, "TX",
2358                                          kiblnd_tx_pool_size(ncpts),
2359                                          kiblnd_create_tx_pool,
2360                                          kiblnd_destroy_tx_pool,
2361                                          kiblnd_tx_init, NULL);
2362                 if (rc != 0) {
2363                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2364                                cpt, rc);
2365                         goto failed;
2366                 }
2367         }
2368
2369         return 0;
2370  failed:
2371         kiblnd_net_fini_pools(net);
2372         LASSERT(rc != 0);
2373         return rc;
2374 }
2375
2376 static int
2377 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2378 {
2379         struct ib_device_attr *attr;
2380         int                    rc;
2381
2382         /* It's safe to assume a HCA can handle a page size
2383          * matching that of the native system */
2384         hdev->ibh_page_shift = PAGE_SHIFT;
2385         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2386         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2387
2388         LIBCFS_ALLOC(attr, sizeof(*attr));
2389         if (attr == NULL) {
2390                 CERROR("Out of memory\n");
2391                 return -ENOMEM;
2392         }
2393
2394         rc = ib_query_device(hdev->ibh_ibdev, attr);
2395         if (rc == 0)
2396                 hdev->ibh_mr_size = attr->max_mr_size;
2397
2398         LIBCFS_FREE(attr, sizeof(*attr));
2399
2400         if (rc != 0) {
2401                 CERROR("Failed to query IB device: %d\n", rc);
2402                 return rc;
2403         }
2404
2405         if (hdev->ibh_mr_size == ~0ULL) {
2406                 hdev->ibh_mr_shift = 64;
2407                 return 0;
2408         }
2409
2410         for (hdev->ibh_mr_shift = 0;
2411              hdev->ibh_mr_shift < 64; hdev->ibh_mr_shift ++) {
2412                 if (hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) ||
2413                     hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) - 1)
2414                         return 0;
2415         }
2416
2417         CERROR("Invalid mr size: "LPX64"\n", hdev->ibh_mr_size);
2418         return -EINVAL;
2419 }
2420
2421 static void
2422 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2423 {
2424         int     i;
2425
2426         if (hdev->ibh_nmrs == 0 || hdev->ibh_mrs == NULL)
2427                 return;
2428
2429         for (i = 0; i < hdev->ibh_nmrs; i++) {
2430                 if (hdev->ibh_mrs[i] == NULL)
2431                         break;
2432
2433                 ib_dereg_mr(hdev->ibh_mrs[i]);
2434         }
2435
2436         LIBCFS_FREE(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2437         hdev->ibh_mrs  = NULL;
2438         hdev->ibh_nmrs = 0;
2439 }
2440
2441 void
2442 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2443 {
2444         kiblnd_hdev_cleanup_mrs(hdev);
2445
2446         if (hdev->ibh_pd != NULL)
2447                 ib_dealloc_pd(hdev->ibh_pd);
2448
2449         if (hdev->ibh_cmid != NULL)
2450                 rdma_destroy_id(hdev->ibh_cmid);
2451
2452         LIBCFS_FREE(hdev, sizeof(*hdev));
2453 }
2454
2455 static int
2456 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2457 {
2458         struct ib_mr *mr;
2459         int           i;
2460         int           rc;
2461         __u64         mm_size;
2462         __u64         mr_size;
2463         int           acflags = IB_ACCESS_LOCAL_WRITE |
2464                                 IB_ACCESS_REMOTE_WRITE;
2465
2466         rc = kiblnd_hdev_get_attr(hdev);
2467         if (rc != 0)
2468                 return rc;
2469
2470         if (hdev->ibh_mr_shift == 64) {
2471                 LIBCFS_ALLOC(hdev->ibh_mrs, 1 * sizeof(*hdev->ibh_mrs));
2472                 if (hdev->ibh_mrs == NULL) {
2473                         CERROR("Failed to allocate MRs table\n");
2474                         return -ENOMEM;
2475                 }
2476
2477                 hdev->ibh_mrs[0] = NULL;
2478                 hdev->ibh_nmrs   = 1;
2479
2480                 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2481                 if (IS_ERR(mr)) {
2482                         CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
2483                         kiblnd_hdev_cleanup_mrs(hdev);
2484                         return PTR_ERR(mr);
2485                 }
2486
2487                 hdev->ibh_mrs[0] = mr;
2488
2489                 goto out;
2490         }
2491
2492         mr_size = (1ULL << hdev->ibh_mr_shift);
2493         mm_size = (unsigned long)high_memory - PAGE_OFFSET;
2494
2495         hdev->ibh_nmrs = (int)((mm_size + mr_size - 1) >> hdev->ibh_mr_shift);
2496
2497         if (hdev->ibh_mr_shift < 32 || hdev->ibh_nmrs > 1024) {
2498                 /* it's 4T..., assume we will re-code at that time */
2499                 CERROR("Can't support memory size: x"LPX64
2500                        " with MR size: x"LPX64"\n", mm_size, mr_size);
2501                 return -EINVAL;
2502         }
2503
2504         /* create an array of MRs to cover all memory */
2505         LIBCFS_ALLOC(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2506         if (hdev->ibh_mrs == NULL) {
2507                 CERROR("Failed to allocate MRs' table\n");
2508                 return -ENOMEM;
2509         }
2510
2511         memset(hdev->ibh_mrs, 0, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2512
2513         for (i = 0; i < hdev->ibh_nmrs; i++) {
2514                 struct ib_phys_buf ipb;
2515                 __u64              iova;
2516
2517                 ipb.size = hdev->ibh_mr_size;
2518                 ipb.addr = i * mr_size;
2519                 iova     = ipb.addr;
2520
2521                 mr = ib_reg_phys_mr(hdev->ibh_pd, &ipb, 1, acflags, &iova);
2522                 if (IS_ERR(mr)) {
2523                         CERROR("Failed ib_reg_phys_mr addr "LPX64
2524                                " size "LPX64" : %ld\n",
2525                                ipb.addr, ipb.size, PTR_ERR(mr));
2526                         kiblnd_hdev_cleanup_mrs(hdev);
2527                         return PTR_ERR(mr);
2528                 }
2529
2530                 LASSERT (iova == ipb.addr);
2531
2532                 hdev->ibh_mrs[i] = mr;
2533         }
2534
2535 out:
2536         if (hdev->ibh_mr_size != ~0ULL || hdev->ibh_nmrs != 1)
2537                 LCONSOLE_INFO("Register global MR array, MR size: "
2538                               LPX64", array size: %d\n",
2539                               hdev->ibh_mr_size, hdev->ibh_nmrs);
2540         return 0;
2541 }
2542
2543 static int
2544 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2545 {       /* DUMMY */
2546         return 0;
2547 }
2548
2549 static int
2550 kiblnd_dev_need_failover(kib_dev_t *dev)
2551 {
2552         struct rdma_cm_id  *cmid;
2553         struct sockaddr_in  srcaddr;
2554         struct sockaddr_in  dstaddr;
2555         int                 rc;
2556
2557         if (dev->ibd_hdev == NULL || /* initializing */
2558             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2559             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2560                 return 1;
2561
2562         /* XXX: it's UGLY, but I don't have better way to find
2563          * ib-bonding HCA failover because:
2564          *
2565          * a. no reliable CM event for HCA failover...
2566          * b. no OFED API to get ib_device for current net_device...
2567          *
2568          * We have only two choices at this point:
2569          *
2570          * a. rdma_bind_addr(), it will conflict with listener cmid
2571          * b. rdma_resolve_addr() to zero addr */
2572         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2573                                      IB_QPT_RC);
2574         if (IS_ERR(cmid)) {
2575                 rc = PTR_ERR(cmid);
2576                 CERROR("Failed to create cmid for failover: %d\n", rc);
2577                 return rc;
2578         }
2579
2580         memset(&srcaddr, 0, sizeof(srcaddr));
2581         srcaddr.sin_family      = AF_INET;
2582         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2583
2584         memset(&dstaddr, 0, sizeof(dstaddr));
2585         dstaddr.sin_family = AF_INET;
2586         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2587                                (struct sockaddr *)&dstaddr, 1);
2588         if (rc != 0 || cmid->device == NULL) {
2589                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2590                        dev->ibd_ifname, &dev->ibd_ifip,
2591                        cmid->device, rc);
2592                 rdma_destroy_id(cmid);
2593                 return rc;
2594         }
2595
2596         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2597         rdma_destroy_id(cmid);
2598         return rc;
2599 }
2600
2601 int
2602 kiblnd_dev_failover(kib_dev_t *dev)
2603 {
2604         struct list_head    zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2605         struct list_head    zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2606         struct list_head    zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2607         struct rdma_cm_id  *cmid  = NULL;
2608         kib_hca_dev_t      *hdev  = NULL;
2609         kib_hca_dev_t      *old;
2610         struct ib_pd       *pd;
2611         kib_net_t          *net;
2612         struct sockaddr_in  addr;
2613         unsigned long       flags;
2614         int                 rc = 0;
2615         int                 i;
2616
2617         LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2618                  dev->ibd_can_failover ||
2619                  dev->ibd_hdev == NULL);
2620
2621         rc = kiblnd_dev_need_failover(dev);
2622         if (rc <= 0)
2623                 goto out;
2624
2625         if (dev->ibd_hdev != NULL &&
2626             dev->ibd_hdev->ibh_cmid != NULL) {
2627                 /* XXX it's not good to close old listener at here,
2628                  * because we can fail to create new listener.
2629                  * But we have to close it now, otherwise rdma_bind_addr
2630                  * will return EADDRINUSE... How crap! */
2631                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2632
2633                 cmid = dev->ibd_hdev->ibh_cmid;
2634                 /* make next schedule of kiblnd_dev_need_failover()
2635                  * return 1 for me */
2636                 dev->ibd_hdev->ibh_cmid  = NULL;
2637                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2638
2639                 rdma_destroy_id(cmid);
2640         }
2641
2642         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2643                                      IB_QPT_RC);
2644         if (IS_ERR(cmid)) {
2645                 rc = PTR_ERR(cmid);
2646                 CERROR("Failed to create cmid for failover: %d\n", rc);
2647                 goto out;
2648         }
2649
2650         memset(&addr, 0, sizeof(addr));
2651         addr.sin_family      = AF_INET;
2652         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2653         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2654
2655         /* Bind to failover device or port */
2656         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2657         if (rc != 0 || cmid->device == NULL) {
2658                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2659                        dev->ibd_ifname, &dev->ibd_ifip,
2660                        cmid->device, rc);
2661                 rdma_destroy_id(cmid);
2662                 goto out;
2663         }
2664
2665         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2666         if (hdev == NULL) {
2667                 CERROR("Failed to allocate kib_hca_dev\n");
2668                 rdma_destroy_id(cmid);
2669                 rc = -ENOMEM;
2670                 goto out;
2671         }
2672
2673         atomic_set(&hdev->ibh_ref, 1);
2674         hdev->ibh_dev   = dev;
2675         hdev->ibh_cmid  = cmid;
2676         hdev->ibh_ibdev = cmid->device;
2677
2678         pd = ib_alloc_pd(cmid->device);
2679         if (IS_ERR(pd)) {
2680                 rc = PTR_ERR(pd);
2681                 CERROR("Can't allocate PD: %d\n", rc);
2682                 goto out;
2683         }
2684
2685         hdev->ibh_pd = pd;
2686
2687         rc = rdma_listen(cmid, 0);
2688         if (rc != 0) {
2689                 CERROR("Can't start new listener: %d\n", rc);
2690                 goto out;
2691         }
2692
2693         rc = kiblnd_hdev_setup_mrs(hdev);
2694         if (rc != 0) {
2695                 CERROR("Can't setup device: %d\n", rc);
2696                 goto out;
2697         }
2698
2699         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2700
2701         old = dev->ibd_hdev;
2702         dev->ibd_hdev = hdev;   /* take over the refcount */
2703         hdev = old;
2704
2705         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2706                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2707                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2708                                             &zombie_tpo);
2709
2710                         if (net->ibn_fmr_ps != NULL) {
2711                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2712                                                         &zombie_fpo);
2713
2714                         } else if (net->ibn_pmr_ps != NULL) {
2715                                 kiblnd_fail_poolset(&net->ibn_pmr_ps[i]->
2716                                                     pps_poolset, &zombie_ppo);
2717                         }
2718                 }
2719         }
2720
2721         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2722  out:
2723         if (!list_empty(&zombie_tpo))
2724                 kiblnd_destroy_pool_list(&zombie_tpo);
2725         if (!list_empty(&zombie_ppo))
2726                 kiblnd_destroy_pool_list(&zombie_ppo);
2727         if (!list_empty(&zombie_fpo))
2728                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2729         if (hdev != NULL)
2730                 kiblnd_hdev_decref(hdev);
2731
2732         if (rc != 0)
2733                 dev->ibd_failed_failover++;
2734         else
2735                 dev->ibd_failed_failover = 0;
2736
2737         return rc;
2738 }
2739
2740 void
2741 kiblnd_destroy_dev (kib_dev_t *dev)
2742 {
2743         LASSERT (dev->ibd_nnets == 0);
2744         LASSERT(list_empty(&dev->ibd_nets));
2745
2746         list_del(&dev->ibd_fail_list);
2747         list_del(&dev->ibd_list);
2748
2749         if (dev->ibd_hdev != NULL)
2750                 kiblnd_hdev_decref(dev->ibd_hdev);
2751
2752         LIBCFS_FREE(dev, sizeof(*dev));
2753 }
2754
2755 static kib_dev_t *
2756 kiblnd_create_dev(char *ifname)
2757 {
2758         struct net_device *netdev;
2759         kib_dev_t         *dev;
2760         __u32              netmask;
2761         __u32              ip;
2762         int                up;
2763         int                rc;
2764
2765         rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2766         if (rc != 0) {
2767                 CERROR("Can't query IPoIB interface %s: %d\n",
2768                        ifname, rc);
2769                 return NULL;
2770         }
2771
2772         if (!up) {
2773                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2774                 return NULL;
2775         }
2776
2777         LIBCFS_ALLOC(dev, sizeof(*dev));
2778         if (dev == NULL)
2779                 return NULL;
2780
2781         memset(dev, 0, sizeof(*dev));
2782         netdev = dev_get_by_name(&init_net, ifname);
2783         if (netdev == NULL) {
2784                 dev->ibd_can_failover = 0;
2785         } else {
2786                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2787                 dev_put(netdev);
2788         }
2789
2790         INIT_LIST_HEAD(&dev->ibd_nets);
2791         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2792         INIT_LIST_HEAD(&dev->ibd_fail_list);
2793         dev->ibd_ifip = ip;
2794         strcpy(&dev->ibd_ifname[0], ifname);
2795
2796         /* initialize the device */
2797         rc = kiblnd_dev_failover(dev);
2798         if (rc != 0) {
2799                 CERROR("Can't initialize device: %d\n", rc);
2800                 LIBCFS_FREE(dev, sizeof(*dev));
2801                 return NULL;
2802         }
2803
2804         list_add_tail(&dev->ibd_list,
2805                           &kiblnd_data.kib_devs);
2806         return dev;
2807 }
2808
2809 static void
2810 kiblnd_base_shutdown(void)
2811 {
2812         struct kib_sched_info   *sched;
2813         int                     i;
2814
2815         LASSERT(list_empty(&kiblnd_data.kib_devs));
2816
2817         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2818                atomic_read(&libcfs_kmemory));
2819
2820         switch (kiblnd_data.kib_init) {
2821         default:
2822                 LBUG();
2823
2824         case IBLND_INIT_ALL:
2825         case IBLND_INIT_DATA:
2826                 LASSERT (kiblnd_data.kib_peers != NULL);
2827                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2828                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2829                 }
2830                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2831                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2832
2833                 /* flag threads to terminate; wake and wait for them to die */
2834                 kiblnd_data.kib_shutdown = 1;
2835
2836                 /* NB: we really want to stop scheduler threads net by net
2837                  * instead of the whole module, this should be improved
2838                  * with dynamic configuration LNet */
2839                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2840                         wake_up_all(&sched->ibs_waitq);
2841
2842                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2843                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2844
2845                 i = 2;
2846                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2847                         i++;
2848                         /* power of 2? */
2849                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2850                                "Waiting for %d threads to terminate\n",
2851                                atomic_read(&kiblnd_data.kib_nthreads));
2852                         set_current_state(TASK_UNINTERRUPTIBLE);
2853                         schedule_timeout(cfs_time_seconds(1));
2854                 }
2855
2856                 /* fall through */
2857
2858         case IBLND_INIT_NOTHING:
2859                 break;
2860         }
2861
2862         if (kiblnd_data.kib_peers != NULL) {
2863                 LIBCFS_FREE(kiblnd_data.kib_peers,
2864                             sizeof(struct list_head) *
2865                             kiblnd_data.kib_peer_hash_size);
2866         }
2867
2868         if (kiblnd_data.kib_scheds != NULL)
2869                 cfs_percpt_free(kiblnd_data.kib_scheds);
2870
2871         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2872                atomic_read(&libcfs_kmemory));
2873
2874         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2875         module_put(THIS_MODULE);
2876 }
2877
2878 static void
2879 kiblnd_shutdown (lnet_ni_t *ni)
2880 {
2881         kib_net_t        *net = ni->ni_data;
2882         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2883         int               i;
2884         unsigned long     flags;
2885
2886         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2887
2888         if (net == NULL)
2889                 goto out;
2890
2891         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2892                atomic_read(&libcfs_kmemory));
2893
2894         write_lock_irqsave(g_lock, flags);
2895         net->ibn_shutdown = 1;
2896         write_unlock_irqrestore(g_lock, flags);
2897
2898         switch (net->ibn_init) {
2899         default:
2900                 LBUG();
2901
2902         case IBLND_INIT_ALL:
2903                 /* nuke all existing peers within this net */
2904                 kiblnd_del_peer(ni, LNET_NID_ANY);
2905
2906                 /* Wait for all peer state to clean up */
2907                 i = 2;
2908                 while (atomic_read(&net->ibn_npeers) != 0) {
2909                         i++;
2910                         /* power of 2? */
2911                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2912                                "%s: waiting for %d peers to disconnect\n",
2913                                libcfs_nid2str(ni->ni_nid),
2914                                atomic_read(&net->ibn_npeers));
2915                         set_current_state(TASK_UNINTERRUPTIBLE);
2916                         schedule_timeout(cfs_time_seconds(1));
2917                 }
2918
2919                 kiblnd_net_fini_pools(net);
2920
2921                 write_lock_irqsave(g_lock, flags);
2922                 LASSERT(net->ibn_dev->ibd_nnets > 0);
2923                 net->ibn_dev->ibd_nnets--;
2924                 list_del(&net->ibn_list);
2925                 write_unlock_irqrestore(g_lock, flags);
2926
2927                 /* fall through */
2928
2929         case IBLND_INIT_NOTHING:
2930                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
2931
2932                 if (net->ibn_dev != NULL &&
2933                     net->ibn_dev->ibd_nnets == 0)
2934                         kiblnd_destroy_dev(net->ibn_dev);
2935
2936                 break;
2937         }
2938
2939         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2940                atomic_read(&libcfs_kmemory));
2941
2942         net->ibn_init = IBLND_INIT_NOTHING;
2943         ni->ni_data = NULL;
2944
2945         LIBCFS_FREE(net, sizeof(*net));
2946
2947 out:
2948         if (list_empty(&kiblnd_data.kib_devs))
2949                 kiblnd_base_shutdown();
2950         return;
2951 }
2952
2953 static int
2954 kiblnd_base_startup(void)
2955 {
2956         struct kib_sched_info   *sched;
2957         int                     rc;
2958         int                     i;
2959
2960         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2961
2962         try_module_get(THIS_MODULE);
2963         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2964
2965         rwlock_init(&kiblnd_data.kib_global_lock);
2966
2967         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2968         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2969
2970         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2971         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2972                      sizeof(struct list_head) *
2973                      kiblnd_data.kib_peer_hash_size);
2974         if (kiblnd_data.kib_peers == NULL)
2975                 goto failed;
2976
2977         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2978                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2979
2980         spin_lock_init(&kiblnd_data.kib_connd_lock);
2981         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2982         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2983         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2984         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
2985
2986         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2987                                                   sizeof(*sched));
2988         if (kiblnd_data.kib_scheds == NULL)
2989                 goto failed;
2990
2991         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2992                 int     nthrs;
2993
2994                 spin_lock_init(&sched->ibs_lock);
2995                 INIT_LIST_HEAD(&sched->ibs_conns);
2996                 init_waitqueue_head(&sched->ibs_waitq);
2997
2998                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2999                 if (*kiblnd_tunables.kib_nscheds > 0) {
3000                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
3001                 } else {
3002                         /* max to half of CPUs, another half is reserved for
3003                          * upper layer modules */
3004                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3005                 }
3006
3007                 sched->ibs_nthreads_max = nthrs;
3008                 sched->ibs_cpt = i;
3009         }
3010
3011         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3012
3013         /* lists/ptrs/locks initialised */
3014         kiblnd_data.kib_init = IBLND_INIT_DATA;
3015         /*****************************************************/
3016
3017         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3018         if (rc != 0) {
3019                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3020                 goto failed;
3021         }
3022
3023         if (*kiblnd_tunables.kib_dev_failover != 0)
3024                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
3025                                          "kiblnd_failover");
3026
3027         if (rc != 0) {
3028                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3029                 goto failed;
3030         }
3031
3032         /* flag everything initialised */
3033         kiblnd_data.kib_init = IBLND_INIT_ALL;
3034         /*****************************************************/
3035
3036         return 0;
3037
3038  failed:
3039         kiblnd_base_shutdown();
3040         return -ENETDOWN;
3041 }
3042
3043 static int
3044 kiblnd_start_schedulers(struct kib_sched_info *sched)
3045 {
3046         int     rc = 0;
3047         int     nthrs;
3048         int     i;
3049
3050         if (sched->ibs_nthreads == 0) {
3051                 if (*kiblnd_tunables.kib_nscheds > 0) {
3052                         nthrs = sched->ibs_nthreads_max;
3053                 } else {
3054                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
3055                                                sched->ibs_cpt);
3056                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3057                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3058                 }
3059         } else {
3060                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3061                 /* increase one thread if there is new interface */
3062                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3063         }
3064
3065         for (i = 0; i < nthrs; i++) {
3066                 long    id;
3067                 char    name[20];
3068                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3069                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
3070                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3071                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
3072                 if (rc == 0)
3073                         continue;
3074
3075                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3076                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3077                 break;
3078         }
3079
3080         sched->ibs_nthreads += i;
3081         return rc;
3082 }
3083
3084 static int
3085 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
3086 {
3087         int     cpt;
3088         int     rc;
3089         int     i;
3090
3091         for (i = 0; i < ncpts; i++) {
3092                 struct kib_sched_info *sched;
3093
3094                 cpt = (cpts == NULL) ? i : cpts[i];
3095                 sched = kiblnd_data.kib_scheds[cpt];
3096
3097                 if (!newdev && sched->ibs_nthreads > 0)
3098                         continue;
3099
3100                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3101                 if (rc != 0) {
3102                         CERROR("Failed to start scheduler threads for %s\n",
3103                                dev->ibd_ifname);
3104                         return rc;
3105                 }
3106         }
3107         return 0;
3108 }
3109
3110 static kib_dev_t *
3111 kiblnd_dev_search(char *ifname)
3112 {
3113         kib_dev_t       *alias = NULL;
3114         kib_dev_t       *dev;
3115         char            *colon;
3116         char            *colon2;
3117
3118         colon = strchr(ifname, ':');
3119         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3120                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3121                         return dev;
3122
3123                 if (alias != NULL)
3124                         continue;
3125
3126                 colon2 = strchr(dev->ibd_ifname, ':');
3127                 if (colon != NULL)
3128                         *colon = 0;
3129                 if (colon2 != NULL)
3130                         *colon2 = 0;
3131
3132                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3133                         alias = dev;
3134
3135                 if (colon != NULL)
3136                         *colon = ':';
3137                 if (colon2 != NULL)
3138                         *colon2 = ':';
3139         }
3140         return alias;
3141 }
3142
3143 static int
3144 kiblnd_startup (lnet_ni_t *ni)
3145 {
3146         char                     *ifname;
3147         kib_dev_t                *ibdev = NULL;
3148         kib_net_t                *net;
3149         struct timeval            tv;
3150         unsigned long             flags;
3151         int                       rc;
3152         int                       newdev;
3153
3154         LASSERT (ni->ni_lnd == &the_o2iblnd);
3155
3156         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3157                 rc = kiblnd_base_startup();
3158                 if (rc != 0)
3159                         return rc;
3160         }
3161
3162         LIBCFS_ALLOC(net, sizeof(*net));
3163         ni->ni_data = net;
3164         if (net == NULL)
3165                 goto failed;
3166
3167         memset(net, 0, sizeof(*net));
3168
3169         do_gettimeofday(&tv);
3170         net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
3171
3172         ni->ni_peertimeout    = *kiblnd_tunables.kib_peertimeout;
3173         ni->ni_maxtxcredits   = *kiblnd_tunables.kib_credits;
3174         ni->ni_peertxcredits  = *kiblnd_tunables.kib_peertxcredits;
3175         ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
3176
3177         if (ni->ni_interfaces[0] != NULL) {
3178                 /* Use the IPoIB interface specified in 'networks=' */
3179
3180                 CLASSERT (LNET_MAX_INTERFACES > 1);
3181                 if (ni->ni_interfaces[1] != NULL) {
3182                         CERROR("Multiple interfaces not supported\n");
3183                         goto failed;
3184                 }
3185
3186                 ifname = ni->ni_interfaces[0];
3187         } else {
3188                 ifname = *kiblnd_tunables.kib_default_ipif;
3189         }
3190
3191         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3192                 CERROR("IPoIB interface name too long: %s\n", ifname);
3193                 goto failed;
3194         }
3195
3196         ibdev = kiblnd_dev_search(ifname);
3197
3198         newdev = ibdev == NULL;
3199         /* hmm...create kib_dev even for alias */
3200         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3201                 ibdev = kiblnd_create_dev(ifname);
3202
3203         if (ibdev == NULL)
3204                 goto failed;
3205
3206         net->ibn_dev = ibdev;
3207         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3208
3209         rc = kiblnd_dev_start_threads(ibdev, newdev,
3210                                       ni->ni_cpts, ni->ni_ncpts);
3211         if (rc != 0)
3212                 goto failed;
3213
3214         rc = kiblnd_net_init_pools(net, ni->ni_cpts, ni->ni_ncpts);
3215         if (rc != 0) {
3216                 CERROR("Failed to initialize NI pools: %d\n", rc);
3217                 goto failed;
3218         }
3219
3220         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3221         ibdev->ibd_nnets++;
3222         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3223         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3224
3225         net->ibn_init = IBLND_INIT_ALL;
3226
3227         return 0;
3228
3229 failed:
3230         if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3231                 kiblnd_destroy_dev(ibdev);
3232
3233         kiblnd_shutdown(ni);
3234
3235         CDEBUG(D_NET, "kiblnd_startup failed\n");
3236         return -ENETDOWN;
3237 }
3238
3239 static lnd_t the_o2iblnd = {
3240         .lnd_type       = O2IBLND,
3241         .lnd_startup    = kiblnd_startup,
3242         .lnd_shutdown   = kiblnd_shutdown,
3243         .lnd_ctl        = kiblnd_ctl,
3244         .lnd_query      = kiblnd_query,
3245         .lnd_send       = kiblnd_send,
3246         .lnd_recv       = kiblnd_recv,
3247 };
3248
3249 static void __exit
3250 kiblnd_module_fini (void)
3251 {
3252         lnet_unregister_lnd(&the_o2iblnd);
3253         kiblnd_tunables_fini();
3254 }
3255
3256 static int __init
3257 kiblnd_module_init (void)
3258 {
3259         int    rc;
3260
3261         CLASSERT (sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3262         CLASSERT (offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3263                   <= IBLND_MSG_SIZE);
3264         CLASSERT (offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3265                   <= IBLND_MSG_SIZE);
3266
3267         rc = kiblnd_tunables_init();
3268         if (rc != 0)
3269                 return rc;
3270
3271         lnet_register_lnd(&the_o2iblnd);
3272
3273         return 0;
3274 }
3275
3276 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3277 MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v2.00");
3278 MODULE_LICENSE("GPL");
3279
3280 module_init(kiblnd_module_init);
3281 module_exit(kiblnd_module_fini);