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_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd)
1402 {
1403         LASSERT(hdev->ibh_mrs != NULL);
1404
1405         if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1406             *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags)
1407                 return NULL;
1408
1409         return hdev->ibh_mrs;
1410 }
1411
1412 static void
1413 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
1414 {
1415         LASSERT (pool->fpo_map_count == 0);
1416
1417         if (pool->fpo_fmr_pool != NULL)
1418                 ib_destroy_fmr_pool(pool->fpo_fmr_pool);
1419
1420         if (pool->fpo_hdev != NULL)
1421                 kiblnd_hdev_decref(pool->fpo_hdev);
1422
1423         LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t));
1424 }
1425
1426 static void
1427 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1428 {
1429         kib_fmr_pool_t *pool;
1430
1431         while (!list_empty(head)) {
1432                 pool = list_entry(head->next, kib_fmr_pool_t, fpo_list);
1433                 list_del(&pool->fpo_list);
1434                 kiblnd_destroy_fmr_pool(pool);
1435         }
1436 }
1437
1438 static int kiblnd_fmr_pool_size(int ncpts)
1439 {
1440         int size = *kiblnd_tunables.kib_fmr_pool_size / ncpts;
1441
1442         return max(IBLND_FMR_POOL, size);
1443 }
1444
1445 static int kiblnd_fmr_flush_trigger(int ncpts)
1446 {
1447         int size = *kiblnd_tunables.kib_fmr_flush_trigger / ncpts;
1448
1449         return max(IBLND_FMR_POOL_FLUSH, size);
1450 }
1451
1452 static int
1453 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1454 {
1455         /* FMR pool for RDMA */
1456         kib_dev_t               *dev = fps->fps_net->ibn_dev;
1457         kib_fmr_pool_t          *fpo;
1458         struct ib_fmr_pool_param param = {
1459                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1460                 .page_shift        = PAGE_SHIFT,
1461                 .access            = (IB_ACCESS_LOCAL_WRITE |
1462                                       IB_ACCESS_REMOTE_WRITE),
1463                 .pool_size         = fps->fps_pool_size,
1464                 .dirty_watermark   = fps->fps_flush_trigger,
1465                 .flush_function    = NULL,
1466                 .flush_arg         = NULL,
1467                 .cache             = !!*kiblnd_tunables.kib_fmr_cache};
1468         int rc;
1469
1470         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1471         if (fpo == NULL)
1472                 return -ENOMEM;
1473
1474         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1475
1476         fpo->fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, &param);
1477         if (IS_ERR(fpo->fpo_fmr_pool)) {
1478                 rc = PTR_ERR(fpo->fpo_fmr_pool);
1479                 CERROR("Failed to create FMR pool: %d\n", rc);
1480
1481                 kiblnd_hdev_decref(fpo->fpo_hdev);
1482                 LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t));
1483                 return rc;
1484         }
1485
1486         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1487         fpo->fpo_owner    = fps;
1488         *pp_fpo = fpo;
1489
1490         return 0;
1491 }
1492
1493 static void
1494 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies)
1495 {
1496         if (fps->fps_net == NULL) /* intialized? */
1497                 return;
1498
1499         spin_lock(&fps->fps_lock);
1500
1501         while (!list_empty(&fps->fps_pool_list)) {
1502                 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1503                                                  kib_fmr_pool_t, fpo_list);
1504                 fpo->fpo_failed = 1;
1505                 list_del(&fpo->fpo_list);
1506                 if (fpo->fpo_map_count == 0)
1507                         list_add(&fpo->fpo_list, zombies);
1508                 else
1509                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1510         }
1511
1512         spin_unlock(&fps->fps_lock);
1513 }
1514
1515 static void
1516 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1517 {
1518         if (fps->fps_net != NULL) { /* initialized? */
1519                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1520                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1521         }
1522 }
1523
1524 static int
1525 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, kib_net_t *net,
1526                         int pool_size, int flush_trigger)
1527 {
1528         kib_fmr_pool_t *fpo;
1529         int             rc;
1530
1531         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1532
1533         fps->fps_net = net;
1534         fps->fps_cpt = cpt;
1535         fps->fps_pool_size = pool_size;
1536         fps->fps_flush_trigger = flush_trigger;
1537         spin_lock_init(&fps->fps_lock);
1538         INIT_LIST_HEAD(&fps->fps_pool_list);
1539         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1540
1541         rc = kiblnd_create_fmr_pool(fps, &fpo);
1542         if (rc == 0)
1543                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1544
1545         return rc;
1546 }
1547
1548 static int
1549 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1550 {
1551         if (fpo->fpo_map_count != 0) /* still in use */
1552                 return 0;
1553         if (fpo->fpo_failed)
1554                 return 1;
1555         return cfs_time_aftereq(now, fpo->fpo_deadline);
1556 }
1557
1558 void
1559 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1560 {
1561         struct list_head   zombies = LIST_HEAD_INIT(zombies);
1562         kib_fmr_pool_t    *fpo = fmr->fmr_pool;
1563         kib_fmr_poolset_t *fps = fpo->fpo_owner;
1564         cfs_time_t         now = cfs_time_current();
1565         kib_fmr_pool_t    *tmp;
1566         int                rc;
1567
1568         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1569         LASSERT(rc == 0);
1570
1571         if (status != 0) {
1572                 rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
1573                 LASSERT(rc == 0);
1574         }
1575
1576         fmr->fmr_pool = NULL;
1577         fmr->fmr_pfmr = NULL;
1578
1579         spin_lock(&fps->fps_lock);
1580         fpo->fpo_map_count--;   /* decref the pool */
1581
1582         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1583                 /* the first pool is persistent */
1584                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1585                         continue;
1586
1587                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1588                         list_move(&fpo->fpo_list, &zombies);
1589                         fps->fps_version++;
1590                 }
1591         }
1592         spin_unlock(&fps->fps_lock);
1593
1594         if (!list_empty(&zombies))
1595                 kiblnd_destroy_fmr_pool_list(&zombies);
1596 }
1597
1598 int
1599 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1600                     __u64 iov, kib_fmr_t *fmr)
1601 {
1602         struct ib_pool_fmr *pfmr;
1603         kib_fmr_pool_t     *fpo;
1604         __u64               version;
1605         int                 rc;
1606
1607 again:
1608         spin_lock(&fps->fps_lock);
1609         version = fps->fps_version;
1610         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1611                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1612                 fpo->fpo_map_count++;
1613                 spin_unlock(&fps->fps_lock);
1614
1615                 pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
1616                                             pages, npages, iov);
1617                 if (likely(!IS_ERR(pfmr))) {
1618                         fmr->fmr_pool = fpo;
1619                         fmr->fmr_pfmr = pfmr;
1620                         return 0;
1621                 }
1622
1623                 spin_lock(&fps->fps_lock);
1624                 fpo->fpo_map_count--;
1625                 if (PTR_ERR(pfmr) != -EAGAIN) {
1626                         spin_unlock(&fps->fps_lock);
1627                         return PTR_ERR(pfmr);
1628                 }
1629
1630                 /* EAGAIN and ... */
1631                 if (version != fps->fps_version) {
1632                         spin_unlock(&fps->fps_lock);
1633                         goto again;
1634                 }
1635         }
1636
1637         if (fps->fps_increasing) {
1638                 spin_unlock(&fps->fps_lock);
1639                 CDEBUG(D_NET, "Another thread is allocating new "
1640                        "FMR pool, waiting for her to complete\n");
1641                 schedule();
1642                 goto again;
1643
1644         }
1645
1646         if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1647                 /* someone failed recently */
1648                 spin_unlock(&fps->fps_lock);
1649                 return -EAGAIN;
1650         }
1651
1652         fps->fps_increasing = 1;
1653         spin_unlock(&fps->fps_lock);
1654
1655         CDEBUG(D_NET, "Allocate new FMR pool\n");
1656         rc = kiblnd_create_fmr_pool(fps, &fpo);
1657         spin_lock(&fps->fps_lock);
1658         fps->fps_increasing = 0;
1659         if (rc == 0) {
1660                 fps->fps_version++;
1661                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1662         } else {
1663                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1664         }
1665         spin_unlock(&fps->fps_lock);
1666
1667         goto again;
1668 }
1669
1670 static void
1671 kiblnd_fini_pool(kib_pool_t *pool)
1672 {
1673         LASSERT(list_empty(&pool->po_free_list));
1674         LASSERT(pool->po_allocated == 0);
1675
1676         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1677 }
1678
1679 static void
1680 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1681 {
1682         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1683
1684         memset(pool, 0, sizeof(kib_pool_t));
1685         INIT_LIST_HEAD(&pool->po_free_list);
1686         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1687         pool->po_owner    = ps;
1688         pool->po_size     = size;
1689 }
1690
1691 static void
1692 kiblnd_destroy_pool_list(struct list_head *head)
1693 {
1694         kib_pool_t *pool;
1695
1696         while (!list_empty(head)) {
1697                 pool = list_entry(head->next, kib_pool_t, po_list);
1698                 list_del(&pool->po_list);
1699
1700                 LASSERT(pool->po_owner != NULL);
1701                 pool->po_owner->ps_pool_destroy(pool);
1702         }
1703 }
1704
1705 static void
1706 kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
1707 {
1708         if (ps->ps_net == NULL) /* intialized? */
1709                 return;
1710
1711         spin_lock(&ps->ps_lock);
1712         while (!list_empty(&ps->ps_pool_list)) {
1713                 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
1714                                             kib_pool_t, po_list);
1715                 po->po_failed = 1;
1716                 list_del(&po->po_list);
1717                 if (po->po_allocated == 0)
1718                         list_add(&po->po_list, zombies);
1719                 else
1720                         list_add(&po->po_list, &ps->ps_failed_pool_list);
1721         }
1722         spin_unlock(&ps->ps_lock);
1723 }
1724
1725 static void
1726 kiblnd_fini_poolset(kib_poolset_t *ps)
1727 {
1728         if (ps->ps_net != NULL) { /* initialized? */
1729                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1730                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1731         }
1732 }
1733
1734 static int
1735 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
1736                     kib_net_t *net, char *name, int size,
1737                     kib_ps_pool_create_t po_create,
1738                     kib_ps_pool_destroy_t po_destroy,
1739                     kib_ps_node_init_t nd_init,
1740                     kib_ps_node_fini_t nd_fini)
1741 {
1742         kib_pool_t      *pool;
1743         int             rc;
1744
1745         memset(ps, 0, sizeof(kib_poolset_t));
1746
1747         ps->ps_cpt          = cpt;
1748         ps->ps_net          = net;
1749         ps->ps_pool_create  = po_create;
1750         ps->ps_pool_destroy = po_destroy;
1751         ps->ps_node_init    = nd_init;
1752         ps->ps_node_fini    = nd_fini;
1753         ps->ps_pool_size    = size;
1754         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
1755             >= sizeof(ps->ps_name))
1756                 return -E2BIG;
1757         spin_lock_init(&ps->ps_lock);
1758         INIT_LIST_HEAD(&ps->ps_pool_list);
1759         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1760
1761         rc = ps->ps_pool_create(ps, size, &pool);
1762         if (rc == 0)
1763                 list_add(&pool->po_list, &ps->ps_pool_list);
1764         else
1765                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1766
1767         return rc;
1768 }
1769
1770 static int
1771 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
1772 {
1773         if (pool->po_allocated != 0) /* still in use */
1774                 return 0;
1775         if (pool->po_failed)
1776                 return 1;
1777         return cfs_time_aftereq(now, pool->po_deadline);
1778 }
1779
1780 void
1781 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
1782 {
1783         struct list_head zombies = LIST_HEAD_INIT(zombies);
1784         kib_poolset_t   *ps = pool->po_owner;
1785         kib_pool_t      *tmp;
1786         cfs_time_t       now = cfs_time_current();
1787
1788         spin_lock(&ps->ps_lock);
1789
1790         if (ps->ps_node_fini != NULL)
1791                 ps->ps_node_fini(pool, node);
1792
1793         LASSERT(pool->po_allocated > 0);
1794         list_add(node, &pool->po_free_list);
1795         pool->po_allocated--;
1796
1797         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1798                 /* the first pool is persistent */
1799                 if (ps->ps_pool_list.next == &pool->po_list)
1800                         continue;
1801
1802                 if (kiblnd_pool_is_idle(pool, now))
1803                         list_move(&pool->po_list, &zombies);
1804         }
1805         spin_unlock(&ps->ps_lock);
1806
1807         if (!list_empty(&zombies))
1808                 kiblnd_destroy_pool_list(&zombies);
1809 }
1810
1811 struct list_head *
1812 kiblnd_pool_alloc_node(kib_poolset_t *ps)
1813 {
1814         struct list_head        *node;
1815         kib_pool_t              *pool;
1816         int                     rc;
1817
1818 again:
1819         spin_lock(&ps->ps_lock);
1820         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1821                 if (list_empty(&pool->po_free_list))
1822                         continue;
1823
1824                 pool->po_allocated++;
1825                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1826                 node = pool->po_free_list.next;
1827                 list_del(node);
1828
1829                 if (ps->ps_node_init != NULL) {
1830                         /* still hold the lock */
1831                         ps->ps_node_init(pool, node);
1832                 }
1833                 spin_unlock(&ps->ps_lock);
1834                 return node;
1835         }
1836
1837         /* no available tx pool and ... */
1838         if (ps->ps_increasing) {
1839                 /* another thread is allocating a new pool */
1840                 spin_unlock(&ps->ps_lock);
1841                 CDEBUG(D_NET, "Another thread is allocating new "
1842                        "%s pool, waiting for her to complete\n",
1843                        ps->ps_name);
1844                 schedule();
1845                 goto again;
1846         }
1847
1848         if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
1849                 /* someone failed recently */
1850                 spin_unlock(&ps->ps_lock);
1851                 return NULL;
1852         }
1853
1854         ps->ps_increasing = 1;
1855         spin_unlock(&ps->ps_lock);
1856
1857         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1858
1859         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1860
1861         spin_lock(&ps->ps_lock);
1862         ps->ps_increasing = 0;
1863         if (rc == 0) {
1864                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
1865         } else {
1866                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1867                 CERROR("Can't allocate new %s pool because out of memory\n",
1868                        ps->ps_name);
1869         }
1870         spin_unlock(&ps->ps_lock);
1871
1872         goto again;
1873 }
1874
1875 static void
1876 kiblnd_destroy_tx_pool(kib_pool_t *pool)
1877 {
1878         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
1879         int             i;
1880
1881         LASSERT (pool->po_allocated == 0);
1882
1883         if (tpo->tpo_tx_pages != NULL) {
1884                 kiblnd_unmap_tx_pool(tpo);
1885                 kiblnd_free_pages(tpo->tpo_tx_pages);
1886         }
1887
1888         if (tpo->tpo_tx_descs == NULL)
1889                 goto out;
1890
1891         for (i = 0; i < pool->po_size; i++) {
1892                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1893
1894                 list_del(&tx->tx_list);
1895                 if (tx->tx_pages != NULL)
1896                         LIBCFS_FREE(tx->tx_pages,
1897                                     LNET_MAX_IOV *
1898                                     sizeof(*tx->tx_pages));
1899                 if (tx->tx_frags != NULL)
1900                         LIBCFS_FREE(tx->tx_frags,
1901                                     IBLND_MAX_RDMA_FRAGS *
1902                                             sizeof(*tx->tx_frags));
1903                 if (tx->tx_wrq != NULL)
1904                         LIBCFS_FREE(tx->tx_wrq,
1905                                     (1 + IBLND_MAX_RDMA_FRAGS) *
1906                                     sizeof(*tx->tx_wrq));
1907                 if (tx->tx_sge != NULL)
1908                         LIBCFS_FREE(tx->tx_sge,
1909                                     (1 + IBLND_MAX_RDMA_FRAGS) *
1910                                     sizeof(*tx->tx_sge));
1911                 if (tx->tx_rd != NULL)
1912                         LIBCFS_FREE(tx->tx_rd,
1913                                     offsetof(kib_rdma_desc_t,
1914                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
1915         }
1916
1917         LIBCFS_FREE(tpo->tpo_tx_descs,
1918                     pool->po_size * sizeof(kib_tx_t));
1919 out:
1920         kiblnd_fini_pool(pool);
1921         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
1922 }
1923
1924 static int kiblnd_tx_pool_size(int ncpts)
1925 {
1926         int ntx = *kiblnd_tunables.kib_ntx / ncpts;
1927
1928         return max(IBLND_TX_POOL, ntx);
1929 }
1930
1931 static int
1932 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
1933 {
1934         int            i;
1935         int            npg;
1936         kib_pool_t    *pool;
1937         kib_tx_pool_t *tpo;
1938
1939         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
1940         if (tpo == NULL) {
1941                 CERROR("Failed to allocate TX pool\n");
1942                 return -ENOMEM;
1943         }
1944
1945         pool = &tpo->tpo_pool;
1946         kiblnd_init_pool(ps, pool, size);
1947         tpo->tpo_tx_descs = NULL;
1948         tpo->tpo_tx_pages = NULL;
1949
1950         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
1951         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
1952                 CERROR("Can't allocate tx pages: %d\n", npg);
1953                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
1954                 return -ENOMEM;
1955         }
1956
1957         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
1958                          size * sizeof(kib_tx_t));
1959         if (tpo->tpo_tx_descs == NULL) {
1960                 CERROR("Can't allocate %d tx descriptors\n", size);
1961                 ps->ps_pool_destroy(pool);
1962                 return -ENOMEM;
1963         }
1964
1965         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
1966
1967         for (i = 0; i < size; i++) {
1968                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1969
1970                 tx->tx_pool = tpo;
1971                 if (ps->ps_net->ibn_fmr_ps != NULL) {
1972                         LIBCFS_CPT_ALLOC(tx->tx_pages,
1973                                          lnet_cpt_table(), ps->ps_cpt,
1974                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
1975                         if (tx->tx_pages == NULL)
1976                                 break;
1977                 }
1978
1979                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
1980                                  IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags));
1981                 if (tx->tx_frags == NULL)
1982                         break;
1983
1984                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
1985
1986                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
1987                                  (1 + IBLND_MAX_RDMA_FRAGS) *
1988                                  sizeof(*tx->tx_wrq));
1989                 if (tx->tx_wrq == NULL)
1990                         break;
1991
1992                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
1993                                  (1 + IBLND_MAX_RDMA_FRAGS) *
1994                                  sizeof(*tx->tx_sge));
1995                 if (tx->tx_sge == NULL)
1996                         break;
1997
1998                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
1999                                  offsetof(kib_rdma_desc_t,
2000                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2001                 if (tx->tx_rd == NULL)
2002                         break;
2003         }
2004
2005         if (i == size) {
2006                 kiblnd_map_tx_pool(tpo);
2007                 *pp_po = pool;
2008                 return 0;
2009         }
2010
2011         ps->ps_pool_destroy(pool);
2012         return -ENOMEM;
2013 }
2014
2015 static void
2016 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2017 {
2018         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2019                                              tps_poolset);
2020         kib_tx_t         *tx  = list_entry(node, kib_tx_t, tx_list);
2021
2022         tx->tx_cookie = tps->tps_next_tx_cookie++;
2023 }
2024
2025 static void
2026 kiblnd_net_fini_pools(kib_net_t *net)
2027 {
2028         int     i;
2029
2030         cfs_cpt_for_each(i, lnet_cpt_table()) {
2031                 kib_tx_poolset_t        *tps;
2032                 kib_fmr_poolset_t       *fps;
2033
2034                 if (net->ibn_tx_ps != NULL) {
2035                         tps = net->ibn_tx_ps[i];
2036                         kiblnd_fini_poolset(&tps->tps_poolset);
2037                 }
2038
2039                 if (net->ibn_fmr_ps != NULL) {
2040                         fps = net->ibn_fmr_ps[i];
2041                         kiblnd_fini_fmr_poolset(fps);
2042                 }
2043         }
2044
2045         if (net->ibn_tx_ps != NULL) {
2046                 cfs_percpt_free(net->ibn_tx_ps);
2047                 net->ibn_tx_ps = NULL;
2048         }
2049
2050         if (net->ibn_fmr_ps != NULL) {
2051                 cfs_percpt_free(net->ibn_fmr_ps);
2052                 net->ibn_fmr_ps = NULL;
2053         }
2054 }
2055
2056 static int
2057 kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts)
2058 {
2059         unsigned long   flags;
2060         int             cpt;
2061         int             rc;
2062         int             i;
2063
2064         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2065         if (*kiblnd_tunables.kib_map_on_demand == 0) {
2066                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2067                                            flags);
2068                 goto create_tx_pool;
2069         }
2070
2071         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2072
2073         if (*kiblnd_tunables.kib_fmr_pool_size <
2074             *kiblnd_tunables.kib_ntx / 4) {
2075                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2076                        *kiblnd_tunables.kib_fmr_pool_size,
2077                        *kiblnd_tunables.kib_ntx / 4);
2078                 rc = -EINVAL;
2079                 goto failed;
2080         }
2081
2082         /* TX pool must be created later than FMR, see LU-2268
2083          * for details */
2084         LASSERT(net->ibn_tx_ps == NULL);
2085
2086         /* premapping can fail if ibd_nmr > 1, so we always create
2087          * FMR pool and map-on-demand if premapping failed */
2088
2089         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2090                                            sizeof(kib_fmr_poolset_t));
2091         if (net->ibn_fmr_ps == NULL) {
2092                 CERROR("Failed to allocate FMR pool array\n");
2093                 rc = -ENOMEM;
2094                 goto failed;
2095         }
2096
2097         for (i = 0; i < ncpts; i++) {
2098                 cpt = (cpts == NULL) ? i : cpts[i];
2099                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, net,
2100                                              kiblnd_fmr_pool_size(ncpts),
2101                                              kiblnd_fmr_flush_trigger(ncpts));
2102                 if (rc != 0) {
2103                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2104                                cpt, rc);
2105                         goto failed;
2106                 }
2107         }
2108
2109         if (i > 0)
2110                 LASSERT(i == ncpts);
2111
2112  create_tx_pool:
2113         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2114                                           sizeof(kib_tx_poolset_t));
2115         if (net->ibn_tx_ps == NULL) {
2116                 CERROR("Failed to allocate tx pool array\n");
2117                 rc = -ENOMEM;
2118                 goto failed;
2119         }
2120
2121         for (i = 0; i < ncpts; i++) {
2122                 cpt = (cpts == NULL) ? i : cpts[i];
2123                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2124                                          cpt, net, "TX",
2125                                          kiblnd_tx_pool_size(ncpts),
2126                                          kiblnd_create_tx_pool,
2127                                          kiblnd_destroy_tx_pool,
2128                                          kiblnd_tx_init, NULL);
2129                 if (rc != 0) {
2130                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2131                                cpt, rc);
2132                         goto failed;
2133                 }
2134         }
2135
2136         return 0;
2137  failed:
2138         kiblnd_net_fini_pools(net);
2139         LASSERT(rc != 0);
2140         return rc;
2141 }
2142
2143 static int
2144 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2145 {
2146         struct ib_device_attr *attr;
2147         int                    rc;
2148
2149         /* It's safe to assume a HCA can handle a page size
2150          * matching that of the native system */
2151         hdev->ibh_page_shift = PAGE_SHIFT;
2152         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2153         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2154
2155         LIBCFS_ALLOC(attr, sizeof(*attr));
2156         if (attr == NULL) {
2157                 CERROR("Out of memory\n");
2158                 return -ENOMEM;
2159         }
2160
2161         rc = ib_query_device(hdev->ibh_ibdev, attr);
2162         if (rc == 0)
2163                 hdev->ibh_mr_size = attr->max_mr_size;
2164
2165         LIBCFS_FREE(attr, sizeof(*attr));
2166
2167         if (rc != 0) {
2168                 CERROR("Failed to query IB device: %d\n", rc);
2169                 return rc;
2170         }
2171
2172         if (hdev->ibh_mr_size == ~0ULL) {
2173                 hdev->ibh_mr_shift = 64;
2174                 return 0;
2175         }
2176
2177         CERROR("Invalid mr size: "LPX64"\n", hdev->ibh_mr_size);
2178         return -EINVAL;
2179 }
2180
2181 static void
2182 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2183 {
2184         if (hdev->ibh_mrs == NULL)
2185                 return;
2186
2187         ib_dereg_mr(hdev->ibh_mrs);
2188
2189         hdev->ibh_mrs = NULL;
2190 }
2191
2192 void
2193 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2194 {
2195         kiblnd_hdev_cleanup_mrs(hdev);
2196
2197         if (hdev->ibh_pd != NULL)
2198                 ib_dealloc_pd(hdev->ibh_pd);
2199
2200         if (hdev->ibh_cmid != NULL)
2201                 rdma_destroy_id(hdev->ibh_cmid);
2202
2203         LIBCFS_FREE(hdev, sizeof(*hdev));
2204 }
2205
2206 static int
2207 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2208 {
2209         struct ib_mr *mr;
2210         int           rc;
2211         int           acflags = IB_ACCESS_LOCAL_WRITE |
2212                                 IB_ACCESS_REMOTE_WRITE;
2213
2214         rc = kiblnd_hdev_get_attr(hdev);
2215         if (rc != 0)
2216                 return rc;
2217
2218         mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2219         if (IS_ERR(mr)) {
2220                 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2221                 kiblnd_hdev_cleanup_mrs(hdev);
2222                 return PTR_ERR(mr);
2223         }
2224
2225         hdev->ibh_mrs = mr;
2226
2227         return 0;
2228 }
2229
2230 static int
2231 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2232 {       /* DUMMY */
2233         return 0;
2234 }
2235
2236 static int
2237 kiblnd_dev_need_failover(kib_dev_t *dev)
2238 {
2239         struct rdma_cm_id  *cmid;
2240         struct sockaddr_in  srcaddr;
2241         struct sockaddr_in  dstaddr;
2242         int                 rc;
2243
2244         if (dev->ibd_hdev == NULL || /* initializing */
2245             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2246             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2247                 return 1;
2248
2249         /* XXX: it's UGLY, but I don't have better way to find
2250          * ib-bonding HCA failover because:
2251          *
2252          * a. no reliable CM event for HCA failover...
2253          * b. no OFED API to get ib_device for current net_device...
2254          *
2255          * We have only two choices at this point:
2256          *
2257          * a. rdma_bind_addr(), it will conflict with listener cmid
2258          * b. rdma_resolve_addr() to zero addr */
2259         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2260                                      IB_QPT_RC);
2261         if (IS_ERR(cmid)) {
2262                 rc = PTR_ERR(cmid);
2263                 CERROR("Failed to create cmid for failover: %d\n", rc);
2264                 return rc;
2265         }
2266
2267         memset(&srcaddr, 0, sizeof(srcaddr));
2268         srcaddr.sin_family      = AF_INET;
2269         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2270
2271         memset(&dstaddr, 0, sizeof(dstaddr));
2272         dstaddr.sin_family = AF_INET;
2273         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2274                                (struct sockaddr *)&dstaddr, 1);
2275         if (rc != 0 || cmid->device == NULL) {
2276                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2277                        dev->ibd_ifname, &dev->ibd_ifip,
2278                        cmid->device, rc);
2279                 rdma_destroy_id(cmid);
2280                 return rc;
2281         }
2282
2283         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2284         rdma_destroy_id(cmid);
2285         return rc;
2286 }
2287
2288 int
2289 kiblnd_dev_failover(kib_dev_t *dev)
2290 {
2291         struct list_head    zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2292         struct list_head    zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2293         struct list_head    zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2294         struct rdma_cm_id  *cmid  = NULL;
2295         kib_hca_dev_t      *hdev  = NULL;
2296         kib_hca_dev_t      *old;
2297         struct ib_pd       *pd;
2298         kib_net_t          *net;
2299         struct sockaddr_in  addr;
2300         unsigned long       flags;
2301         int                 rc = 0;
2302         int                 i;
2303
2304         LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2305                  dev->ibd_can_failover ||
2306                  dev->ibd_hdev == NULL);
2307
2308         rc = kiblnd_dev_need_failover(dev);
2309         if (rc <= 0)
2310                 goto out;
2311
2312         if (dev->ibd_hdev != NULL &&
2313             dev->ibd_hdev->ibh_cmid != NULL) {
2314                 /* XXX it's not good to close old listener at here,
2315                  * because we can fail to create new listener.
2316                  * But we have to close it now, otherwise rdma_bind_addr
2317                  * will return EADDRINUSE... How crap! */
2318                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2319
2320                 cmid = dev->ibd_hdev->ibh_cmid;
2321                 /* make next schedule of kiblnd_dev_need_failover()
2322                  * return 1 for me */
2323                 dev->ibd_hdev->ibh_cmid  = NULL;
2324                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2325
2326                 rdma_destroy_id(cmid);
2327         }
2328
2329         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2330                                      IB_QPT_RC);
2331         if (IS_ERR(cmid)) {
2332                 rc = PTR_ERR(cmid);
2333                 CERROR("Failed to create cmid for failover: %d\n", rc);
2334                 goto out;
2335         }
2336
2337         memset(&addr, 0, sizeof(addr));
2338         addr.sin_family      = AF_INET;
2339         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2340         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2341
2342         /* Bind to failover device or port */
2343         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2344         if (rc != 0 || cmid->device == NULL) {
2345                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2346                        dev->ibd_ifname, &dev->ibd_ifip,
2347                        cmid->device, rc);
2348                 rdma_destroy_id(cmid);
2349                 goto out;
2350         }
2351
2352         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2353         if (hdev == NULL) {
2354                 CERROR("Failed to allocate kib_hca_dev\n");
2355                 rdma_destroy_id(cmid);
2356                 rc = -ENOMEM;
2357                 goto out;
2358         }
2359
2360         atomic_set(&hdev->ibh_ref, 1);
2361         hdev->ibh_dev   = dev;
2362         hdev->ibh_cmid  = cmid;
2363         hdev->ibh_ibdev = cmid->device;
2364
2365         pd = ib_alloc_pd(cmid->device);
2366         if (IS_ERR(pd)) {
2367                 rc = PTR_ERR(pd);
2368                 CERROR("Can't allocate PD: %d\n", rc);
2369                 goto out;
2370         }
2371
2372         hdev->ibh_pd = pd;
2373
2374         rc = rdma_listen(cmid, 0);
2375         if (rc != 0) {
2376                 CERROR("Can't start new listener: %d\n", rc);
2377                 goto out;
2378         }
2379
2380         rc = kiblnd_hdev_setup_mrs(hdev);
2381         if (rc != 0) {
2382                 CERROR("Can't setup device: %d\n", rc);
2383                 goto out;
2384         }
2385
2386         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2387
2388         old = dev->ibd_hdev;
2389         dev->ibd_hdev = hdev;   /* take over the refcount */
2390         hdev = old;
2391
2392         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2393                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2394                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2395                                             &zombie_tpo);
2396
2397                         if (net->ibn_fmr_ps != NULL)
2398                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2399                                                         &zombie_fpo);
2400                 }
2401         }
2402
2403         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2404  out:
2405         if (!list_empty(&zombie_tpo))
2406                 kiblnd_destroy_pool_list(&zombie_tpo);
2407         if (!list_empty(&zombie_ppo))
2408                 kiblnd_destroy_pool_list(&zombie_ppo);
2409         if (!list_empty(&zombie_fpo))
2410                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2411         if (hdev != NULL)
2412                 kiblnd_hdev_decref(hdev);
2413
2414         if (rc != 0)
2415                 dev->ibd_failed_failover++;
2416         else
2417                 dev->ibd_failed_failover = 0;
2418
2419         return rc;
2420 }
2421
2422 void
2423 kiblnd_destroy_dev (kib_dev_t *dev)
2424 {
2425         LASSERT (dev->ibd_nnets == 0);
2426         LASSERT(list_empty(&dev->ibd_nets));
2427
2428         list_del(&dev->ibd_fail_list);
2429         list_del(&dev->ibd_list);
2430
2431         if (dev->ibd_hdev != NULL)
2432                 kiblnd_hdev_decref(dev->ibd_hdev);
2433
2434         LIBCFS_FREE(dev, sizeof(*dev));
2435 }
2436
2437 static kib_dev_t *
2438 kiblnd_create_dev(char *ifname)
2439 {
2440         struct net_device *netdev;
2441         kib_dev_t         *dev;
2442         __u32              netmask;
2443         __u32              ip;
2444         int                up;
2445         int                rc;
2446
2447         rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2448         if (rc != 0) {
2449                 CERROR("Can't query IPoIB interface %s: %d\n",
2450                        ifname, rc);
2451                 return NULL;
2452         }
2453
2454         if (!up) {
2455                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2456                 return NULL;
2457         }
2458
2459         LIBCFS_ALLOC(dev, sizeof(*dev));
2460         if (dev == NULL)
2461                 return NULL;
2462
2463         memset(dev, 0, sizeof(*dev));
2464         netdev = dev_get_by_name(&init_net, ifname);
2465         if (netdev == NULL) {
2466                 dev->ibd_can_failover = 0;
2467         } else {
2468                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2469                 dev_put(netdev);
2470         }
2471
2472         INIT_LIST_HEAD(&dev->ibd_nets);
2473         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2474         INIT_LIST_HEAD(&dev->ibd_fail_list);
2475         dev->ibd_ifip = ip;
2476         strcpy(&dev->ibd_ifname[0], ifname);
2477
2478         /* initialize the device */
2479         rc = kiblnd_dev_failover(dev);
2480         if (rc != 0) {
2481                 CERROR("Can't initialize device: %d\n", rc);
2482                 LIBCFS_FREE(dev, sizeof(*dev));
2483                 return NULL;
2484         }
2485
2486         list_add_tail(&dev->ibd_list,
2487                           &kiblnd_data.kib_devs);
2488         return dev;
2489 }
2490
2491 static void
2492 kiblnd_base_shutdown(void)
2493 {
2494         struct kib_sched_info   *sched;
2495         int                     i;
2496
2497         LASSERT(list_empty(&kiblnd_data.kib_devs));
2498
2499         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2500                atomic_read(&libcfs_kmemory));
2501
2502         switch (kiblnd_data.kib_init) {
2503         default:
2504                 LBUG();
2505
2506         case IBLND_INIT_ALL:
2507         case IBLND_INIT_DATA:
2508                 LASSERT (kiblnd_data.kib_peers != NULL);
2509                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2510                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2511                 }
2512                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2513                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2514
2515                 /* flag threads to terminate; wake and wait for them to die */
2516                 kiblnd_data.kib_shutdown = 1;
2517
2518                 /* NB: we really want to stop scheduler threads net by net
2519                  * instead of the whole module, this should be improved
2520                  * with dynamic configuration LNet */
2521                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2522                         wake_up_all(&sched->ibs_waitq);
2523
2524                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2525                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2526
2527                 i = 2;
2528                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2529                         i++;
2530                         /* power of 2? */
2531                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2532                                "Waiting for %d threads to terminate\n",
2533                                atomic_read(&kiblnd_data.kib_nthreads));
2534                         set_current_state(TASK_UNINTERRUPTIBLE);
2535                         schedule_timeout(cfs_time_seconds(1));
2536                 }
2537
2538                 /* fall through */
2539
2540         case IBLND_INIT_NOTHING:
2541                 break;
2542         }
2543
2544         if (kiblnd_data.kib_peers != NULL) {
2545                 LIBCFS_FREE(kiblnd_data.kib_peers,
2546                             sizeof(struct list_head) *
2547                             kiblnd_data.kib_peer_hash_size);
2548         }
2549
2550         if (kiblnd_data.kib_scheds != NULL)
2551                 cfs_percpt_free(kiblnd_data.kib_scheds);
2552
2553         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2554                atomic_read(&libcfs_kmemory));
2555
2556         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2557         module_put(THIS_MODULE);
2558 }
2559
2560 static void
2561 kiblnd_shutdown (lnet_ni_t *ni)
2562 {
2563         kib_net_t        *net = ni->ni_data;
2564         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2565         int               i;
2566         unsigned long     flags;
2567
2568         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2569
2570         if (net == NULL)
2571                 goto out;
2572
2573         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2574                atomic_read(&libcfs_kmemory));
2575
2576         write_lock_irqsave(g_lock, flags);
2577         net->ibn_shutdown = 1;
2578         write_unlock_irqrestore(g_lock, flags);
2579
2580         switch (net->ibn_init) {
2581         default:
2582                 LBUG();
2583
2584         case IBLND_INIT_ALL:
2585                 /* nuke all existing peers within this net */
2586                 kiblnd_del_peer(ni, LNET_NID_ANY);
2587
2588                 /* Wait for all peer state to clean up */
2589                 i = 2;
2590                 while (atomic_read(&net->ibn_npeers) != 0) {
2591                         i++;
2592                         /* power of 2? */
2593                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2594                                "%s: waiting for %d peers to disconnect\n",
2595                                libcfs_nid2str(ni->ni_nid),
2596                                atomic_read(&net->ibn_npeers));
2597                         set_current_state(TASK_UNINTERRUPTIBLE);
2598                         schedule_timeout(cfs_time_seconds(1));
2599                 }
2600
2601                 kiblnd_net_fini_pools(net);
2602
2603                 write_lock_irqsave(g_lock, flags);
2604                 LASSERT(net->ibn_dev->ibd_nnets > 0);
2605                 net->ibn_dev->ibd_nnets--;
2606                 list_del(&net->ibn_list);
2607                 write_unlock_irqrestore(g_lock, flags);
2608
2609                 /* fall through */
2610
2611         case IBLND_INIT_NOTHING:
2612                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
2613
2614                 if (net->ibn_dev != NULL &&
2615                     net->ibn_dev->ibd_nnets == 0)
2616                         kiblnd_destroy_dev(net->ibn_dev);
2617
2618                 break;
2619         }
2620
2621         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2622                atomic_read(&libcfs_kmemory));
2623
2624         net->ibn_init = IBLND_INIT_NOTHING;
2625         ni->ni_data = NULL;
2626
2627         LIBCFS_FREE(net, sizeof(*net));
2628
2629 out:
2630         if (list_empty(&kiblnd_data.kib_devs))
2631                 kiblnd_base_shutdown();
2632         return;
2633 }
2634
2635 static int
2636 kiblnd_base_startup(void)
2637 {
2638         struct kib_sched_info   *sched;
2639         int                     rc;
2640         int                     i;
2641
2642         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2643
2644         try_module_get(THIS_MODULE);
2645         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2646
2647         rwlock_init(&kiblnd_data.kib_global_lock);
2648
2649         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2650         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2651
2652         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2653         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2654                      sizeof(struct list_head) *
2655                      kiblnd_data.kib_peer_hash_size);
2656         if (kiblnd_data.kib_peers == NULL)
2657                 goto failed;
2658
2659         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2660                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2661
2662         spin_lock_init(&kiblnd_data.kib_connd_lock);
2663         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2664         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2665         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2666         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
2667
2668         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2669                                                   sizeof(*sched));
2670         if (kiblnd_data.kib_scheds == NULL)
2671                 goto failed;
2672
2673         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2674                 int     nthrs;
2675
2676                 spin_lock_init(&sched->ibs_lock);
2677                 INIT_LIST_HEAD(&sched->ibs_conns);
2678                 init_waitqueue_head(&sched->ibs_waitq);
2679
2680                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2681                 if (*kiblnd_tunables.kib_nscheds > 0) {
2682                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
2683                 } else {
2684                         /* max to half of CPUs, another half is reserved for
2685                          * upper layer modules */
2686                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2687                 }
2688
2689                 sched->ibs_nthreads_max = nthrs;
2690                 sched->ibs_cpt = i;
2691         }
2692
2693         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2694
2695         /* lists/ptrs/locks initialised */
2696         kiblnd_data.kib_init = IBLND_INIT_DATA;
2697         /*****************************************************/
2698
2699         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
2700         if (rc != 0) {
2701                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2702                 goto failed;
2703         }
2704
2705         if (*kiblnd_tunables.kib_dev_failover != 0)
2706                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
2707                                          "kiblnd_failover");
2708
2709         if (rc != 0) {
2710                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
2711                 goto failed;
2712         }
2713
2714         /* flag everything initialised */
2715         kiblnd_data.kib_init = IBLND_INIT_ALL;
2716         /*****************************************************/
2717
2718         return 0;
2719
2720  failed:
2721         kiblnd_base_shutdown();
2722         return -ENETDOWN;
2723 }
2724
2725 static int
2726 kiblnd_start_schedulers(struct kib_sched_info *sched)
2727 {
2728         int     rc = 0;
2729         int     nthrs;
2730         int     i;
2731
2732         if (sched->ibs_nthreads == 0) {
2733                 if (*kiblnd_tunables.kib_nscheds > 0) {
2734                         nthrs = sched->ibs_nthreads_max;
2735                 } else {
2736                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
2737                                                sched->ibs_cpt);
2738                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2739                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
2740                 }
2741         } else {
2742                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
2743                 /* increase one thread if there is new interface */
2744                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
2745         }
2746
2747         for (i = 0; i < nthrs; i++) {
2748                 long    id;
2749                 char    name[20];
2750                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
2751                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
2752                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
2753                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
2754                 if (rc == 0)
2755                         continue;
2756
2757                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2758                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
2759                 break;
2760         }
2761
2762         sched->ibs_nthreads += i;
2763         return rc;
2764 }
2765
2766 static int
2767 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
2768 {
2769         int     cpt;
2770         int     rc;
2771         int     i;
2772
2773         for (i = 0; i < ncpts; i++) {
2774                 struct kib_sched_info *sched;
2775
2776                 cpt = (cpts == NULL) ? i : cpts[i];
2777                 sched = kiblnd_data.kib_scheds[cpt];
2778
2779                 if (!newdev && sched->ibs_nthreads > 0)
2780                         continue;
2781
2782                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
2783                 if (rc != 0) {
2784                         CERROR("Failed to start scheduler threads for %s\n",
2785                                dev->ibd_ifname);
2786                         return rc;
2787                 }
2788         }
2789         return 0;
2790 }
2791
2792 static kib_dev_t *
2793 kiblnd_dev_search(char *ifname)
2794 {
2795         kib_dev_t       *alias = NULL;
2796         kib_dev_t       *dev;
2797         char            *colon;
2798         char            *colon2;
2799
2800         colon = strchr(ifname, ':');
2801         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
2802                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
2803                         return dev;
2804
2805                 if (alias != NULL)
2806                         continue;
2807
2808                 colon2 = strchr(dev->ibd_ifname, ':');
2809                 if (colon != NULL)
2810                         *colon = 0;
2811                 if (colon2 != NULL)
2812                         *colon2 = 0;
2813
2814                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
2815                         alias = dev;
2816
2817                 if (colon != NULL)
2818                         *colon = ':';
2819                 if (colon2 != NULL)
2820                         *colon2 = ':';
2821         }
2822         return alias;
2823 }
2824
2825 static int
2826 kiblnd_startup (lnet_ni_t *ni)
2827 {
2828         char                     *ifname;
2829         kib_dev_t                *ibdev = NULL;
2830         kib_net_t                *net;
2831         struct timeval            tv;
2832         unsigned long             flags;
2833         int                       rc;
2834         int                       newdev;
2835
2836         LASSERT (ni->ni_lnd == &the_o2iblnd);
2837
2838         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
2839                 rc = kiblnd_base_startup();
2840                 if (rc != 0)
2841                         return rc;
2842         }
2843
2844         LIBCFS_ALLOC(net, sizeof(*net));
2845         ni->ni_data = net;
2846         if (net == NULL)
2847                 goto failed;
2848
2849         memset(net, 0, sizeof(*net));
2850
2851         do_gettimeofday(&tv);
2852         net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
2853
2854         ni->ni_peertimeout    = *kiblnd_tunables.kib_peertimeout;
2855         ni->ni_maxtxcredits   = *kiblnd_tunables.kib_credits;
2856         ni->ni_peertxcredits  = *kiblnd_tunables.kib_peertxcredits;
2857         ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
2858
2859         if (ni->ni_interfaces[0] != NULL) {
2860                 /* Use the IPoIB interface specified in 'networks=' */
2861
2862                 CLASSERT (LNET_MAX_INTERFACES > 1);
2863                 if (ni->ni_interfaces[1] != NULL) {
2864                         CERROR("Multiple interfaces not supported\n");
2865                         goto failed;
2866                 }
2867
2868                 ifname = ni->ni_interfaces[0];
2869         } else {
2870                 ifname = *kiblnd_tunables.kib_default_ipif;
2871         }
2872
2873         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
2874                 CERROR("IPoIB interface name too long: %s\n", ifname);
2875                 goto failed;
2876         }
2877
2878         ibdev = kiblnd_dev_search(ifname);
2879
2880         newdev = ibdev == NULL;
2881         /* hmm...create kib_dev even for alias */
2882         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
2883                 ibdev = kiblnd_create_dev(ifname);
2884
2885         if (ibdev == NULL)
2886                 goto failed;
2887
2888         net->ibn_dev = ibdev;
2889         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
2890
2891         rc = kiblnd_dev_start_threads(ibdev, newdev,
2892                                       ni->ni_cpts, ni->ni_ncpts);
2893         if (rc != 0)
2894                 goto failed;
2895
2896         rc = kiblnd_net_init_pools(net, ni->ni_cpts, ni->ni_ncpts);
2897         if (rc != 0) {
2898                 CERROR("Failed to initialize NI pools: %d\n", rc);
2899                 goto failed;
2900         }
2901
2902         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2903         ibdev->ibd_nnets++;
2904         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
2905         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2906
2907         net->ibn_init = IBLND_INIT_ALL;
2908
2909         return 0;
2910
2911 failed:
2912         if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
2913                 kiblnd_destroy_dev(ibdev);
2914
2915         kiblnd_shutdown(ni);
2916
2917         CDEBUG(D_NET, "kiblnd_startup failed\n");
2918         return -ENETDOWN;
2919 }
2920
2921 static lnd_t the_o2iblnd = {
2922         .lnd_type       = O2IBLND,
2923         .lnd_startup    = kiblnd_startup,
2924         .lnd_shutdown   = kiblnd_shutdown,
2925         .lnd_ctl        = kiblnd_ctl,
2926         .lnd_query      = kiblnd_query,
2927         .lnd_send       = kiblnd_send,
2928         .lnd_recv       = kiblnd_recv,
2929 };
2930
2931 static void __exit
2932 kiblnd_module_fini (void)
2933 {
2934         lnet_unregister_lnd(&the_o2iblnd);
2935         kiblnd_tunables_fini();
2936 }
2937
2938 static int __init
2939 kiblnd_module_init (void)
2940 {
2941         int    rc;
2942
2943         CLASSERT (sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
2944         CLASSERT (offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
2945                   <= IBLND_MSG_SIZE);
2946         CLASSERT (offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
2947                   <= IBLND_MSG_SIZE);
2948
2949         rc = kiblnd_tunables_init();
2950         if (rc != 0)
2951                 return rc;
2952
2953         lnet_register_lnd(&the_o2iblnd);
2954
2955         return 0;
2956 }
2957
2958 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
2959 MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v2.00");
2960 MODULE_LICENSE("GPL");
2961
2962 module_init(kiblnd_module_init);
2963 module_exit(kiblnd_module_fini);