block/quorum.c

   1 /*
   2  * Quorum Block filter
   3  *
   4  * Copyright (C) 2012-2014 Nodalink, EURL.
   5  *
   6  * Author:
   7  *   Benoît Canet <benoit.canet@irqsave.net>
   8  *
   9  * Based on the design and code of blkverify.c (Copyright (C) 2010 IBM, Corp)
  10  * and blkmirror.c (Copyright (C) 2011 Red Hat, Inc).
  11  *
  12  * This work is licensed under the terms of the GNU GPL, version 2 or later.
  13  * See the COPYING file in the top-level directory.
  14  */
  15
  16 #include <gnutls/gnutls.h>
  17 #include <gnutls/crypto.h>
  18 #include "block/block_int.h"
  19 #include "qapi/qmp/qjson.h"
  20
  21 #define HASH_LENGTH 32
  22
  23 #define QUORUM_OPT_VOTE_THRESHOLD "vote-threshold"
  24 #define QUORUM_OPT_BLKVERIFY      "blkverify"
  25
  26 /* This union holds a vote hash value */
  27 typedef union QuorumVoteValue {
  28     char h[HASH_LENGTH];       /* SHA-256 hash */
  29     int64_t l;                 /* simpler 64 bits hash */
  30 } QuorumVoteValue;
  31
  32 /* A vote item */
  33 typedef struct QuorumVoteItem {
  34     int index;
  35     QLIST_ENTRY(QuorumVoteItem) next;
  36 } QuorumVoteItem;
  37
  38 /* this structure is a vote version. A version is the set of votes sharing the
  39  * same vote value.
  40  * The set of votes will be tracked with the items field and its cardinality is
  41  * vote_count.
  42  */
  43 typedef struct QuorumVoteVersion {
  44     QuorumVoteValue value;
  45     int index;
  46     int vote_count;
  47     QLIST_HEAD(, QuorumVoteItem) items;
  48     QLIST_ENTRY(QuorumVoteVersion) next;
  49 } QuorumVoteVersion;
  50
  51 /* this structure holds a group of vote versions together */
  52 typedef struct QuorumVotes {
  53     QLIST_HEAD(, QuorumVoteVersion) vote_list;
  54     bool (*compare)(QuorumVoteValue *a, QuorumVoteValue *b);
  55 } QuorumVotes;
  56
  57 /* the following structure holds the state of one quorum instance */
  58 typedef struct BDRVQuorumState {
  59     BlockDriverState **bs; /* children BlockDriverStates */
  60     int num_children;      /* children count */
  61     int threshold;         /* if less than threshold children reads gave the
  62                             * same result a quorum error occurs.
  63                             */
  64     bool is_blkverify;     /* true if the driver is in blkverify mode
  65                             * Writes are mirrored on two children devices.
  66                             * On reads the two children devices' contents are
  67                             * compared and if a difference is spotted its
  68                             * location is printed and the code aborts.
  69                             * It is useful to debug other block drivers by
  70                             * comparing them with a reference one.
  71                             */
  72 } BDRVQuorumState;
  73
  74 typedef struct QuorumAIOCB QuorumAIOCB;
  75
  76 /* Quorum will create one instance of the following structure per operation it
  77  * performs on its children.
  78  * So for each read/write operation coming from the upper layer there will be
  79  * $children_count QuorumChildRequest.
  80  */
  81 typedef struct QuorumChildRequest {
  82     BlockDriverAIOCB *aiocb;
  83     QEMUIOVector qiov;
  84     uint8_t *buf;
  85     int ret;
  86     QuorumAIOCB *parent;
  87 } QuorumChildRequest;
  88
  89 /* Quorum will use the following structure to track progress of each read/write
  90  * operation received by the upper layer.
  91  * This structure hold pointers to the QuorumChildRequest structures instances
  92  * used to do operations on each children and track overall progress.
  93  */
  94 struct QuorumAIOCB {
  95     BlockDriverAIOCB common;
  96
  97     /* Request metadata */
  98     uint64_t sector_num;
  99     int nb_sectors;
 100
 101     QEMUIOVector *qiov;         /* calling IOV */
 102
 103     QuorumChildRequest *qcrs;   /* individual child requests */
 104     int count;                  /* number of completed AIOCB */
 105     int success_count;          /* number of successfully completed AIOCB */
 106
 107     QuorumVotes votes;
 108
 109     bool is_read;
 110     int vote_ret;
 111 };
 112
 113 static void quorum_vote(QuorumAIOCB *acb);
 114
 115 static void quorum_aio_cancel(BlockDriverAIOCB *blockacb)
 116 {
 117     QuorumAIOCB *acb = container_of(blockacb, QuorumAIOCB, common);
 118     BDRVQuorumState *s = acb->common.bs->opaque;
 119     int i;
 120
 121     /* cancel all callbacks */
 122     for (i = 0; i < s->num_children; i++) {
 123         bdrv_aio_cancel(acb->qcrs[i].aiocb);
 124     }
 125
 126     g_free(acb->qcrs);
 127     qemu_aio_release(acb);
 128 }
 129
 130 static AIOCBInfo quorum_aiocb_info = {
 131     .aiocb_size         = sizeof(QuorumAIOCB),
 132     .cancel             = quorum_aio_cancel,
 133 };
 134
 135 static void quorum_aio_finalize(QuorumAIOCB *acb)
 136 {
 137     BDRVQuorumState *s = acb->common.bs->opaque;
 138     int i, ret = 0;
 139
 140     if (acb->vote_ret) {
 141         ret = acb->vote_ret;
 142     }
 143
 144     acb->common.cb(acb->common.opaque, ret);
 145
 146     if (acb->is_read) {
 147         for (i = 0; i < s->num_children; i++) {
 148             qemu_vfree(acb->qcrs[i].buf);
 149             qemu_iovec_destroy(&acb->qcrs[i].qiov);
 150         }
 151     }
 152
 153     g_free(acb->qcrs);
 154     qemu_aio_release(acb);
 155 }
 156
 157 static bool quorum_sha256_compare(QuorumVoteValue *a, QuorumVoteValue *b)
 158 {
 159     return !memcmp(a->h, b->h, HASH_LENGTH);
 160 }
 161
 162 static bool quorum_64bits_compare(QuorumVoteValue *a, QuorumVoteValue *b)
 163 {
 164     return a->l == b->l;
 165 }
 166
 167 static QuorumAIOCB *quorum_aio_get(BDRVQuorumState *s,
 168                                    BlockDriverState *bs,
 169                                    QEMUIOVector *qiov,
 170                                    uint64_t sector_num,
 171                                    int nb_sectors,
 172                                    BlockDriverCompletionFunc *cb,
 173                                    void *opaque)
 174 {
 175     QuorumAIOCB *acb = qemu_aio_get(&quorum_aiocb_info, bs, cb, opaque);
 176     int i;
 177
 178     acb->common.bs->opaque = s;
 179     acb->sector_num = sector_num;
 180     acb->nb_sectors = nb_sectors;
 181     acb->qiov = qiov;
 182     acb->qcrs = g_new0(QuorumChildRequest, s->num_children);
 183     acb->count = 0;
 184     acb->success_count = 0;
 185     acb->votes.compare = quorum_sha256_compare;
 186     QLIST_INIT(&acb->votes.vote_list);
 187     acb->is_read = false;
 188     acb->vote_ret = 0;
 189
 190     for (i = 0; i < s->num_children; i++) {
 191         acb->qcrs[i].buf = NULL;
 192         acb->qcrs[i].ret = 0;
 193         acb->qcrs[i].parent = acb;
 194     }
 195
 196     return acb;
 197 }
 198
 199 static void quorum_report_bad(QuorumAIOCB *acb, char *node_name, int ret)
 200 {
 201     QObject *data;
 202     assert(node_name);
 203     data = qobject_from_jsonf("{ 'ret': %d"
 204                               ", 'node-name': %s"
 205                               ", 'sector-num': %" PRId64
 206                               ", 'sectors-count': %d }",
 207                               ret, node_name, acb->sector_num, acb->nb_sectors);
 208     monitor_protocol_event(QEVENT_QUORUM_REPORT_BAD, data);
 209     qobject_decref(data);
 210 }
 211
 212 static void quorum_report_failure(QuorumAIOCB *acb)
 213 {
 214     QObject *data;
 215     const char *reference = acb->common.bs->device_name[0] ?
 216                             acb->common.bs->device_name :
 217                             acb->common.bs->node_name;
 218     data = qobject_from_jsonf("{ 'reference': %s"
 219                               ", 'sector-num': %" PRId64
 220                               ", 'sectors-count': %d }",
 221                               reference, acb->sector_num, acb->nb_sectors);
 222     monitor_protocol_event(QEVENT_QUORUM_FAILURE, data);
 223     qobject_decref(data);
 224 }
 225
 226 static int quorum_vote_error(QuorumAIOCB *acb);
 227
 228 static bool quorum_has_too_much_io_failed(QuorumAIOCB *acb)
 229 {
 230     BDRVQuorumState *s = acb->common.bs->opaque;
 231
 232     if (acb->success_count < s->threshold) {
 233         acb->vote_ret = quorum_vote_error(acb);
 234         quorum_report_failure(acb);
 235         return true;
 236     }
 237
 238     return false;
 239 }
 240
 241 static void quorum_aio_cb(void *opaque, int ret)
 242 {
 243     QuorumChildRequest *sacb = opaque;
 244     QuorumAIOCB *acb = sacb->parent;
 245     BDRVQuorumState *s = acb->common.bs->opaque;
 246
 247     sacb->ret = ret;
 248     acb->count++;
 249     if (ret == 0) {
 250         acb->success_count++;
 251     } else {
 252         quorum_report_bad(acb, sacb->aiocb->bs->node_name, ret);
 253     }
 254     assert(acb->count <= s->num_children);
 255     assert(acb->success_count <= s->num_children);
 256     if (acb->count < s->num_children) {
 257         return;
 258     }
 259
 260     /* Do the vote on read */
 261     if (acb->is_read) {
 262         quorum_vote(acb);
 263     } else {
 264         quorum_has_too_much_io_failed(acb);
 265     }
 266
 267     quorum_aio_finalize(acb);
 268 }
 269
 270 static void quorum_report_bad_versions(BDRVQuorumState *s,
 271                                        QuorumAIOCB *acb,
 272                                        QuorumVoteValue *value)
 273 {
 274     QuorumVoteVersion *version;
 275     QuorumVoteItem *item;
 276
 277     QLIST_FOREACH(version, &acb->votes.vote_list, next) {
 278         if (acb->votes.compare(&version->value, value)) {
 279             continue;
 280         }
 281         QLIST_FOREACH(item, &version->items, next) {
 282             quorum_report_bad(acb, s->bs[item->index]->node_name, 0);
 283         }
 284     }
 285 }
 286
 287 static void quorum_copy_qiov(QEMUIOVector *dest, QEMUIOVector *source)
 288 {
 289     int i;
 290     assert(dest->niov == source->niov);
 291     assert(dest->size == source->size);
 292     for (i = 0; i < source->niov; i++) {
 293         assert(dest->iov[i].iov_len == source->iov[i].iov_len);
 294         memcpy(dest->iov[i].iov_base,
 295                source->iov[i].iov_base,
 296                source->iov[i].iov_len);
 297     }
 298 }
 299
 300 static void quorum_count_vote(QuorumVotes *votes,
 301                               QuorumVoteValue *value,
 302                               int index)
 303 {
 304     QuorumVoteVersion *v = NULL, *version = NULL;
 305     QuorumVoteItem *item;
 306
 307     /* look if we have something with this hash */
 308     QLIST_FOREACH(v, &votes->vote_list, next) {
 309         if (votes->compare(&v->value, value)) {
 310             version = v;
 311             break;
 312         }
 313     }
 314
 315     /* It's a version not yet in the list add it */
 316     if (!version) {
 317         version = g_new0(QuorumVoteVersion, 1);
 318         QLIST_INIT(&version->items);
 319         memcpy(&version->value, value, sizeof(version->value));
 320         version->index = index;
 321         version->vote_count = 0;
 322         QLIST_INSERT_HEAD(&votes->vote_list, version, next);
 323     }
 324
 325     version->vote_count++;
 326
 327     item = g_new0(QuorumVoteItem, 1);
 328     item->index = index;
 329     QLIST_INSERT_HEAD(&version->items, item, next);
 330 }
 331
 332 static void quorum_free_vote_list(QuorumVotes *votes)
 333 {
 334     QuorumVoteVersion *version, *next_version;
 335     QuorumVoteItem *item, *next_item;
 336
 337     QLIST_FOREACH_SAFE(version, &votes->vote_list, next, next_version) {
 338         QLIST_REMOVE(version, next);
 339         QLIST_FOREACH_SAFE(item, &version->items, next, next_item) {
 340             QLIST_REMOVE(item, next);
 341             g_free(item);
 342         }
 343         g_free(version);
 344     }
 345 }
 346
 347 static int quorum_compute_hash(QuorumAIOCB *acb, int i, QuorumVoteValue *hash)
 348 {
 349     int j, ret;
 350     gnutls_hash_hd_t dig;
 351     QEMUIOVector *qiov = &acb->qcrs[i].qiov;
 352
 353     ret = gnutls_hash_init(&dig, GNUTLS_DIG_SHA256);
 354
 355     if (ret < 0) {
 356         return ret;
 357     }
 358
 359     for (j = 0; j < qiov->niov; j++) {
 360         ret = gnutls_hash(dig, qiov->iov[j].iov_base, qiov->iov[j].iov_len);
 361         if (ret < 0) {
 362             break;
 363         }
 364     }
 365
 366     gnutls_hash_deinit(dig, (void *) hash);
 367     return ret;
 368 }
 369
 370 static QuorumVoteVersion *quorum_get_vote_winner(QuorumVotes *votes)
 371 {
 372     int max = 0;
 373     QuorumVoteVersion *candidate, *winner = NULL;
 374
 375     QLIST_FOREACH(candidate, &votes->vote_list, next) {
 376         if (candidate->vote_count > max) {
 377             max = candidate->vote_count;
 378             winner = candidate;
 379         }
 380     }
 381
 382     return winner;
 383 }
 384
 385 /* qemu_iovec_compare is handy for blkverify mode because it returns the first
 386  * differing byte location. Yet it is handcoded to compare vectors one byte
 387  * after another so it does not benefit from the libc SIMD optimizations.
 388  * quorum_iovec_compare is written for speed and should be used in the non
 389  * blkverify mode of quorum.
 390  */
 391 static bool quorum_iovec_compare(QEMUIOVector *a, QEMUIOVector *b)
 392 {
 393     int i;
 394     int result;
 395
 396     assert(a->niov == b->niov);
 397     for (i = 0; i < a->niov; i++) {
 398         assert(a->iov[i].iov_len == b->iov[i].iov_len);
 399         result = memcmp(a->iov[i].iov_base,
 400                         b->iov[i].iov_base,
 401                         a->iov[i].iov_len);
 402         if (result) {
 403             return false;
 404         }
 405     }
 406
 407     return true;
 408 }
 409
 410 static void GCC_FMT_ATTR(2, 3) quorum_err(QuorumAIOCB *acb,
 411                                           const char *fmt, ...)
 412 {
 413     va_list ap;
 414
 415     va_start(ap, fmt);
 416     fprintf(stderr, "quorum: sector_num=%" PRId64 " nb_sectors=%d ",
 417             acb->sector_num, acb->nb_sectors);
 418     vfprintf(stderr, fmt, ap);
 419     fprintf(stderr, "\n");
 420     va_end(ap);
 421     exit(1);
 422 }
 423
 424 static bool quorum_compare(QuorumAIOCB *acb,
 425                            QEMUIOVector *a,
 426                            QEMUIOVector *b)
 427 {
 428     BDRVQuorumState *s = acb->common.bs->opaque;
 429     ssize_t offset;
 430
 431     /* This driver will replace blkverify in this particular case */
 432     if (s->is_blkverify) {
 433         offset = qemu_iovec_compare(a, b);
 434         if (offset != -1) {
 435             quorum_err(acb, "contents mismatch in sector %" PRId64,
 436                        acb->sector_num +
 437                        (uint64_t)(offset / BDRV_SECTOR_SIZE));
 438         }
 439         return true;
 440     }
 441
 442     return quorum_iovec_compare(a, b);
 443 }
 444
 445 /* Do a vote to get the error code */
 446 static int quorum_vote_error(QuorumAIOCB *acb)
 447 {
 448     BDRVQuorumState *s = acb->common.bs->opaque;
 449     QuorumVoteVersion *winner = NULL;
 450     QuorumVotes error_votes;
 451     QuorumVoteValue result_value;
 452     int i, ret = 0;
 453     bool error = false;
 454
 455     QLIST_INIT(&error_votes.vote_list);
 456     error_votes.compare = quorum_64bits_compare;
 457
 458     for (i = 0; i < s->num_children; i++) {
 459         ret = acb->qcrs[i].ret;
 460         if (ret) {
 461             error = true;
 462             result_value.l = ret;
 463             quorum_count_vote(&error_votes, &result_value, i);
 464         }
 465     }
 466
 467     if (error) {
 468         winner = quorum_get_vote_winner(&error_votes);
 469         ret = winner->value.l;
 470     }
 471
 472     quorum_free_vote_list(&error_votes);
 473
 474     return ret;
 475 }
 476
 477 static void quorum_vote(QuorumAIOCB *acb)
 478 {
 479     bool quorum = true;
 480     int i, j, ret;
 481     QuorumVoteValue hash;
 482     BDRVQuorumState *s = acb->common.bs->opaque;
 483     QuorumVoteVersion *winner;
 484
 485     if (quorum_has_too_much_io_failed(acb)) {
 486         return;
 487     }
 488
 489     /* get the index of the first successful read */
 490     for (i = 0; i < s->num_children; i++) {
 491         if (!acb->qcrs[i].ret) {
 492             break;
 493         }
 494     }
 495
 496     assert(i < s->num_children);
 497
 498     /* compare this read with all other successful reads stopping at quorum
 499      * failure
 500      */
 501     for (j = i + 1; j < s->num_children; j++) {
 502         if (acb->qcrs[j].ret) {
 503             continue;
 504         }
 505         quorum = quorum_compare(acb, &acb->qcrs[i].qiov, &acb->qcrs[j].qiov);
 506         if (!quorum) {
 507             break;
 508        }
 509     }
 510
 511     /* Every successful read agrees */
 512     if (quorum) {
 513         quorum_copy_qiov(acb->qiov, &acb->qcrs[i].qiov);
 514         return;
 515     }
 516
 517     /* compute hashes for each successful read, also store indexes */
 518     for (i = 0; i < s->num_children; i++) {
 519         if (acb->qcrs[i].ret) {
 520             continue;
 521         }
 522         ret = quorum_compute_hash(acb, i, &hash);
 523         /* if ever the hash computation failed */
 524         if (ret < 0) {
 525             acb->vote_ret = ret;
 526             goto free_exit;
 527         }
 528         quorum_count_vote(&acb->votes, &hash, i);
 529     }
 530
 531     /* vote to select the most represented version */
 532     winner = quorum_get_vote_winner(&acb->votes);
 533
 534     /* if the winner count is smaller than threshold the read fails */
 535     if (winner->vote_count < s->threshold) {
 536         quorum_report_failure(acb);
 537         acb->vote_ret = -EIO;
 538         goto free_exit;
 539     }
 540
 541     /* we have a winner: copy it */
 542     quorum_copy_qiov(acb->qiov, &acb->qcrs[winner->index].qiov);
 543
 544     /* some versions are bad print them */
 545     quorum_report_bad_versions(s, acb, &winner->value);
 546
 547 free_exit:
 548     /* free lists */
 549     quorum_free_vote_list(&acb->votes);
 550 }
 551
 552 static BlockDriverAIOCB *quorum_aio_readv(BlockDriverState *bs,
 553                                          int64_t sector_num,
 554                                          QEMUIOVector *qiov,
 555                                          int nb_sectors,
 556                                          BlockDriverCompletionFunc *cb,
 557                                          void *opaque)
 558 {
 559     BDRVQuorumState *s = bs->opaque;
 560     QuorumAIOCB *acb = quorum_aio_get(s, bs, qiov, sector_num,
 561                                       nb_sectors, cb, opaque);
 562     int i;
 563
 564     acb->is_read = true;
 565
 566     for (i = 0; i < s->num_children; i++) {
 567         acb->qcrs[i].buf = qemu_blockalign(s->bs[i], qiov->size);
 568         qemu_iovec_init(&acb->qcrs[i].qiov, qiov->niov);
 569         qemu_iovec_clone(&acb->qcrs[i].qiov, qiov, acb->qcrs[i].buf);
 570     }
 571
 572     for (i = 0; i < s->num_children; i++) {
 573         bdrv_aio_readv(s->bs[i], sector_num, &acb->qcrs[i].qiov, nb_sectors,
 574                        quorum_aio_cb, &acb->qcrs[i]);
 575     }
 576
 577     return &acb->common;
 578 }
 579
 580 static BlockDriverAIOCB *quorum_aio_writev(BlockDriverState *bs,
 581                                           int64_t sector_num,
 582                                           QEMUIOVector *qiov,
 583                                           int nb_sectors,
 584                                           BlockDriverCompletionFunc *cb,
 585                                           void *opaque)
 586 {
 587     BDRVQuorumState *s = bs->opaque;
 588     QuorumAIOCB *acb = quorum_aio_get(s, bs, qiov, sector_num, nb_sectors,
 589                                       cb, opaque);
 590     int i;
 591
 592     for (i = 0; i < s->num_children; i++) {
 593         acb->qcrs[i].aiocb = bdrv_aio_writev(s->bs[i], sector_num, qiov,
 594                                              nb_sectors, &quorum_aio_cb,
 595                                              &acb->qcrs[i]);
 596     }
 597
 598     return &acb->common;
 599 }
 600
 601 static int64_t quorum_getlength(BlockDriverState *bs)
 602 {
 603     BDRVQuorumState *s = bs->opaque;
 604     int64_t result;
 605     int i;
 606
 607     /* check that all file have the same length */
 608     result = bdrv_getlength(s->bs[0]);
 609     if (result < 0) {
 610         return result;
 611     }
 612     for (i = 1; i < s->num_children; i++) {
 613         int64_t value = bdrv_getlength(s->bs[i]);
 614         if (value < 0) {
 615             return value;
 616         }
 617         if (value != result) {
 618             return -EIO;
 619         }
 620     }
 621
 622     return result;
 623 }
 624
 625 static void quorum_invalidate_cache(BlockDriverState *bs)
 626 {
 627     BDRVQuorumState *s = bs->opaque;
 628     int i;
 629
 630     for (i = 0; i < s->num_children; i++) {
 631         bdrv_invalidate_cache(s->bs[i]);
 632     }
 633 }
 634
 635 static coroutine_fn int quorum_co_flush(BlockDriverState *bs)
 636 {
 637     BDRVQuorumState *s = bs->opaque;
 638     QuorumVoteVersion *winner = NULL;
 639     QuorumVotes error_votes;
 640     QuorumVoteValue result_value;
 641     int i;
 642     int result = 0;
 643
 644     QLIST_INIT(&error_votes.vote_list);
 645     error_votes.compare = quorum_64bits_compare;
 646
 647     for (i = 0; i < s->num_children; i++) {
 648         result = bdrv_co_flush(s->bs[i]);
 649         result_value.l = result;
 650         quorum_count_vote(&error_votes, &result_value, i);
 651     }
 652
 653     winner = quorum_get_vote_winner(&error_votes);
 654     result = winner->value.l;
 655
 656     quorum_free_vote_list(&error_votes);
 657
 658     return result;
 659 }
 660
 661 static bool quorum_recurse_is_first_non_filter(BlockDriverState *bs,
 662                                                BlockDriverState *candidate)
 663 {
 664     BDRVQuorumState *s = bs->opaque;
 665     int i;
 666
 667     for (i = 0; i < s->num_children; i++) {
 668         bool perm = bdrv_recurse_is_first_non_filter(s->bs[i],
 669                                                      candidate);
 670         if (perm) {
 671             return true;
 672         }
 673     }
 674
 675     return false;
 676 }
 677
 678 static int quorum_valid_threshold(int threshold, int num_children, Error **errp)
 679 {
 680
 681     if (threshold < 1) {
 682         error_set(errp, QERR_INVALID_PARAMETER_VALUE,
 683                   "vote-threshold", "value >= 1");
 684         return -ERANGE;
 685     }
 686
 687     if (threshold > num_children) {
 688         error_setg(errp, "threshold may not exceed children count");
 689         return -ERANGE;
 690     }
 691
 692     return 0;
 693 }
 694
 695 static QemuOptsList quorum_runtime_opts = {
 696     .name = "quorum",
 697     .head = QTAILQ_HEAD_INITIALIZER(quorum_runtime_opts.head),
 698     .desc = {
 699         {
 700             .name = QUORUM_OPT_VOTE_THRESHOLD,
 701             .type = QEMU_OPT_NUMBER,
 702             .help = "The number of vote needed for reaching quorum",
 703         },
 704         {
 705             .name = QUORUM_OPT_BLKVERIFY,
 706             .type = QEMU_OPT_BOOL,
 707             .help = "Trigger block verify mode if set",
 708         },
 709         { /* end of list */ }
 710     },
 711 };
 712
 713 static int quorum_open(BlockDriverState *bs, QDict *options, int flags,
 714                        Error **errp)
 715 {
 716     BDRVQuorumState *s = bs->opaque;
 717     Error *local_err = NULL;
 718     QemuOpts *opts;
 719     bool *opened;
 720     QDict *sub = NULL;
 721     QList *list = NULL;
 722     const QListEntry *lentry;
 723     int i;
 724     int ret = 0;
 725
 726     qdict_flatten(options);
 727     qdict_extract_subqdict(options, &sub, "children.");
 728     qdict_array_split(sub, &list);
 729
 730     if (qdict_size(sub)) {
 731         error_setg(&local_err, "Invalid option children.%s",
 732                    qdict_first(sub)->key);
 733         ret = -EINVAL;
 734         goto exit;
 735     }
 736
 737     /* count how many different children are present */
 738     s->num_children = qlist_size(list);
 739     if (s->num_children < 2) {
 740         error_setg(&local_err,
 741                    "Number of provided children must be greater than 1");
 742         ret = -EINVAL;
 743         goto exit;
 744     }
 745
 746     opts = qemu_opts_create(&quorum_runtime_opts, NULL, 0, &error_abort);
 747     qemu_opts_absorb_qdict(opts, options, &local_err);
 748     if (error_is_set(&local_err)) {
 749         ret = -EINVAL;
 750         goto exit;
 751     }
 752
 753     s->threshold = qemu_opt_get_number(opts, QUORUM_OPT_VOTE_THRESHOLD, 0);
 754
 755     /* and validate it against s->num_children */
 756     ret = quorum_valid_threshold(s->threshold, s->num_children, &local_err);
 757     if (ret < 0) {
 758         goto exit;
 759     }
 760
 761     /* is the driver in blkverify mode */
 762     if (qemu_opt_get_bool(opts, QUORUM_OPT_BLKVERIFY, false) &&
 763         s->num_children == 2 && s->threshold == 2) {
 764         s->is_blkverify = true;
 765     } else if (qemu_opt_get_bool(opts, QUORUM_OPT_BLKVERIFY, false)) {
 766         fprintf(stderr, "blkverify mode is set by setting blkverify=on "
 767                 "and using two files with vote_threshold=2\n");
 768     }
 769
 770     /* allocate the children BlockDriverState array */
 771     s->bs = g_new0(BlockDriverState *, s->num_children);
 772     opened = g_new0(bool, s->num_children);
 773
 774     for (i = 0, lentry = qlist_first(list); lentry;
 775          lentry = qlist_next(lentry), i++) {
 776         QDict *d;
 777         QString *string;
 778
 779         switch (qobject_type(lentry->value))
 780         {
 781             /* List of options */
 782             case QTYPE_QDICT:
 783                 d = qobject_to_qdict(lentry->value);
 784                 QINCREF(d);
 785                 ret = bdrv_open(&s->bs[i], NULL, NULL, d, flags, NULL,
 786                                 &local_err);
 787                 break;
 788
 789             /* QMP reference */
 790             case QTYPE_QSTRING:
 791                 string = qobject_to_qstring(lentry->value);
 792                 ret = bdrv_open(&s->bs[i], NULL, qstring_get_str(string), NULL,
 793                                 flags, NULL, &local_err);
 794                 break;
 795
 796             default:
 797                 error_setg(&local_err, "Specification of child block device %i "
 798                            "is invalid", i);
 799                 ret = -EINVAL;
 800         }
 801
 802         if (ret < 0) {
 803             goto close_exit;
 804         }
 805         opened[i] = true;
 806     }
 807
 808     g_free(opened);
 809     goto exit;
 810
 811 close_exit:
 812     /* cleanup on error */
 813     for (i = 0; i < s->num_children; i++) {
 814         if (!opened[i]) {
 815             continue;
 816         }
 817         bdrv_unref(s->bs[i]);
 818     }
 819     g_free(s->bs);
 820     g_free(opened);
 821 exit:
 822     /* propagate error */
 823     if (error_is_set(&local_err)) {
 824         error_propagate(errp, local_err);
 825     }
 826     QDECREF(list);
 827     QDECREF(sub);
 828     return ret;
 829 }
 830
 831 static void quorum_close(BlockDriverState *bs)
 832 {
 833     BDRVQuorumState *s = bs->opaque;
 834     int i;
 835
 836     for (i = 0; i < s->num_children; i++) {
 837         bdrv_unref(s->bs[i]);
 838     }
 839
 840     g_free(s->bs);
 841 }
 842
 843 static BlockDriver bdrv_quorum = {
 844     .format_name        = "quorum",
 845     .protocol_name      = "quorum",
 846
 847     .instance_size      = sizeof(BDRVQuorumState),
 848
 849     .bdrv_file_open     = quorum_open,
 850     .bdrv_close         = quorum_close,
 851
 852     .authorizations     = { true, true },
 853
 854     .bdrv_co_flush_to_disk = quorum_co_flush,
 855
 856     .bdrv_getlength     = quorum_getlength,
 857
 858     .bdrv_aio_readv     = quorum_aio_readv,
 859     .bdrv_aio_writev    = quorum_aio_writev,
 860     .bdrv_invalidate_cache = quorum_invalidate_cache,
 861
 862     .bdrv_recurse_is_first_non_filter = quorum_recurse_is_first_non_filter,
 863 };
 864
 865 static void bdrv_quorum_init(void)
 866 {
 867     bdrv_register(&bdrv_quorum);
 868 }
 869
 870 block_init(bdrv_quorum_init);