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/qbool.h"
  20 #include "qapi/qmp/qdict.h"
  21 #include "qapi/qmp/qerror.h"
  22 #include "qapi/qmp/qint.h"
  23 #include "qapi/qmp/qjson.h"
  24 #include "qapi/qmp/qlist.h"
  25 #include "qapi/qmp/qstring.h"
  26 #include "qapi-event.h"
  27
  28 #define HASH_LENGTH 32
  29
  30 #define QUORUM_OPT_VOTE_THRESHOLD "vote-threshold"
  31 #define QUORUM_OPT_BLKVERIFY      "blkverify"
  32 #define QUORUM_OPT_REWRITE        "rewrite-corrupted"
  33 #define QUORUM_OPT_READ_PATTERN   "read-pattern"
  34
  35 /* This union holds a vote hash value */
  36 typedef union QuorumVoteValue {
  37     char h[HASH_LENGTH];       /* SHA-256 hash */
  38     int64_t l;                 /* simpler 64 bits hash */
  39 } QuorumVoteValue;
  40
  41 /* A vote item */
  42 typedef struct QuorumVoteItem {
  43     int index;
  44     QLIST_ENTRY(QuorumVoteItem) next;
  45 } QuorumVoteItem;
  46
  47 /* this structure is a vote version. A version is the set of votes sharing the
  48  * same vote value.
  49  * The set of votes will be tracked with the items field and its cardinality is
  50  * vote_count.
  51  */
  52 typedef struct QuorumVoteVersion {
  53     QuorumVoteValue value;
  54     int index;
  55     int vote_count;
  56     QLIST_HEAD(, QuorumVoteItem) items;
  57     QLIST_ENTRY(QuorumVoteVersion) next;
  58 } QuorumVoteVersion;
  59
  60 /* this structure holds a group of vote versions together */
  61 typedef struct QuorumVotes {
  62     QLIST_HEAD(, QuorumVoteVersion) vote_list;
  63     bool (*compare)(QuorumVoteValue *a, QuorumVoteValue *b);
  64 } QuorumVotes;
  65
  66 /* the following structure holds the state of one quorum instance */
  67 typedef struct BDRVQuorumState {
  68     BlockDriverState **bs; /* children BlockDriverStates */
  69     int num_children;      /* children count */
  70     int threshold;         /* if less than threshold children reads gave the
  71                             * same result a quorum error occurs.
  72                             */
  73     bool is_blkverify;     /* true if the driver is in blkverify mode
  74                             * Writes are mirrored on two children devices.
  75                             * On reads the two children devices' contents are
  76                             * compared and if a difference is spotted its
  77                             * location is printed and the code aborts.
  78                             * It is useful to debug other block drivers by
  79                             * comparing them with a reference one.
  80                             */
  81     bool rewrite_corrupted;/* true if the driver must rewrite-on-read corrupted
  82                             * block if Quorum is reached.
  83                             */
  84
  85     QuorumReadPattern read_pattern;
  86 } BDRVQuorumState;
  87
  88 typedef struct QuorumAIOCB QuorumAIOCB;
  89
  90 /* Quorum will create one instance of the following structure per operation it
  91  * performs on its children.
  92  * So for each read/write operation coming from the upper layer there will be
  93  * $children_count QuorumChildRequest.
  94  */
  95 typedef struct QuorumChildRequest {
  96     BlockAIOCB *aiocb;
  97     QEMUIOVector qiov;
  98     uint8_t *buf;
  99     int ret;
 100     QuorumAIOCB *parent;
 101 } QuorumChildRequest;
 102
 103 /* Quorum will use the following structure to track progress of each read/write
 104  * operation received by the upper layer.
 105  * This structure hold pointers to the QuorumChildRequest structures instances
 106  * used to do operations on each children and track overall progress.
 107  */
 108 struct QuorumAIOCB {
 109     BlockAIOCB common;
 110
 111     /* Request metadata */
 112     uint64_t sector_num;
 113     int nb_sectors;
 114
 115     QEMUIOVector *qiov;         /* calling IOV */
 116
 117     QuorumChildRequest *qcrs;   /* individual child requests */
 118     int count;                  /* number of completed AIOCB */
 119     int success_count;          /* number of successfully completed AIOCB */
 120
 121     int rewrite_count;          /* number of replica to rewrite: count down to
 122                                  * zero once writes are fired
 123                                  */
 124
 125     QuorumVotes votes;
 126
 127     bool is_read;
 128     int vote_ret;
 129     int child_iter;             /* which child to read in fifo pattern */
 130 };
 131
 132 static bool quorum_vote(QuorumAIOCB *acb);
 133
 134 static void quorum_aio_cancel(BlockAIOCB *blockacb)
 135 {
 136     QuorumAIOCB *acb = container_of(blockacb, QuorumAIOCB, common);
 137     BDRVQuorumState *s = acb->common.bs->opaque;
 138     int i;
 139
 140     /* cancel all callbacks */
 141     for (i = 0; i < s->num_children; i++) {
 142         if (acb->qcrs[i].aiocb) {
 143             bdrv_aio_cancel_async(acb->qcrs[i].aiocb);
 144         }
 145     }
 146 }
 147
 148 static AIOCBInfo quorum_aiocb_info = {
 149     .aiocb_size         = sizeof(QuorumAIOCB),
 150     .cancel_async       = quorum_aio_cancel,
 151 };
 152
 153 static void quorum_aio_finalize(QuorumAIOCB *acb)
 154 {
 155     int i, ret = 0;
 156
 157     if (acb->vote_ret) {
 158         ret = acb->vote_ret;
 159     }
 160
 161     acb->common.cb(acb->common.opaque, ret);
 162
 163     if (acb->is_read) {
 164         /* on the quorum case acb->child_iter == s->num_children - 1 */
 165         for (i = 0; i <= acb->child_iter; i++) {
 166             qemu_vfree(acb->qcrs[i].buf);
 167             qemu_iovec_destroy(&acb->qcrs[i].qiov);
 168         }
 169     }
 170
 171     g_free(acb->qcrs);
 172     qemu_aio_unref(acb);
 173 }
 174
 175 static bool quorum_sha256_compare(QuorumVoteValue *a, QuorumVoteValue *b)
 176 {
 177     return !memcmp(a->h, b->h, HASH_LENGTH);
 178 }
 179
 180 static bool quorum_64bits_compare(QuorumVoteValue *a, QuorumVoteValue *b)
 181 {
 182     return a->l == b->l;
 183 }
 184
 185 static QuorumAIOCB *quorum_aio_get(BDRVQuorumState *s,
 186                                    BlockDriverState *bs,
 187                                    QEMUIOVector *qiov,
 188                                    uint64_t sector_num,
 189                                    int nb_sectors,
 190                                    BlockCompletionFunc *cb,
 191                                    void *opaque)
 192 {
 193     QuorumAIOCB *acb = qemu_aio_get(&quorum_aiocb_info, bs, cb, opaque);
 194     int i;
 195
 196     acb->common.bs->opaque = s;
 197     acb->sector_num = sector_num;
 198     acb->nb_sectors = nb_sectors;
 199     acb->qiov = qiov;
 200     acb->qcrs = g_new0(QuorumChildRequest, s->num_children);
 201     acb->count = 0;
 202     acb->success_count = 0;
 203     acb->rewrite_count = 0;
 204     acb->votes.compare = quorum_sha256_compare;
 205     QLIST_INIT(&acb->votes.vote_list);
 206     acb->is_read = false;
 207     acb->vote_ret = 0;
 208
 209     for (i = 0; i < s->num_children; i++) {
 210         acb->qcrs[i].buf = NULL;
 211         acb->qcrs[i].ret = 0;
 212         acb->qcrs[i].parent = acb;
 213     }
 214
 215     return acb;
 216 }
 217
 218 static void quorum_report_bad(QuorumAIOCB *acb, char *node_name, int ret)
 219 {
 220     const char *msg = NULL;
 221     if (ret < 0) {
 222         msg = strerror(-ret);
 223     }
 224     qapi_event_send_quorum_report_bad(!!msg, msg, node_name,
 225                                       acb->sector_num, acb->nb_sectors, &error_abort);
 226 }
 227
 228 static void quorum_report_failure(QuorumAIOCB *acb)
 229 {
 230     const char *reference = bdrv_get_device_or_node_name(acb->common.bs);
 231     qapi_event_send_quorum_failure(reference, acb->sector_num,
 232                                    acb->nb_sectors, &error_abort);
 233 }
 234
 235 static int quorum_vote_error(QuorumAIOCB *acb);
 236
 237 static bool quorum_has_too_much_io_failed(QuorumAIOCB *acb)
 238 {
 239     BDRVQuorumState *s = acb->common.bs->opaque;
 240
 241     if (acb->success_count < s->threshold) {
 242         acb->vote_ret = quorum_vote_error(acb);
 243         quorum_report_failure(acb);
 244         return true;
 245     }
 246
 247     return false;
 248 }
 249
 250 static void quorum_rewrite_aio_cb(void *opaque, int ret)
 251 {
 252     QuorumAIOCB *acb = opaque;
 253
 254     /* one less rewrite to do */
 255     acb->rewrite_count--;
 256
 257     /* wait until all rewrite callbacks have completed */
 258     if (acb->rewrite_count) {
 259         return;
 260     }
 261
 262     quorum_aio_finalize(acb);
 263 }
 264
 265 static BlockAIOCB *read_fifo_child(QuorumAIOCB *acb);
 266
 267 static void quorum_copy_qiov(QEMUIOVector *dest, QEMUIOVector *source)
 268 {
 269     int i;
 270     assert(dest->niov == source->niov);
 271     assert(dest->size == source->size);
 272     for (i = 0; i < source->niov; i++) {
 273         assert(dest->iov[i].iov_len == source->iov[i].iov_len);
 274         memcpy(dest->iov[i].iov_base,
 275                source->iov[i].iov_base,
 276                source->iov[i].iov_len);
 277     }
 278 }
 279
 280 static void quorum_aio_cb(void *opaque, int ret)
 281 {
 282     QuorumChildRequest *sacb = opaque;
 283     QuorumAIOCB *acb = sacb->parent;
 284     BDRVQuorumState *s = acb->common.bs->opaque;
 285     bool rewrite = false;
 286
 287     if (acb->is_read && s->read_pattern == QUORUM_READ_PATTERN_FIFO) {
 288         /* We try to read next child in FIFO order if we fail to read */
 289         if (ret < 0 && ++acb->child_iter < s->num_children) {
 290             read_fifo_child(acb);
 291             return;
 292         }
 293
 294         if (ret == 0) {
 295             quorum_copy_qiov(acb->qiov, &acb->qcrs[acb->child_iter].qiov);
 296         }
 297         acb->vote_ret = ret;
 298         quorum_aio_finalize(acb);
 299         return;
 300     }
 301
 302     sacb->ret = ret;
 303     acb->count++;
 304     if (ret == 0) {
 305         acb->success_count++;
 306     } else {
 307         quorum_report_bad(acb, sacb->aiocb->bs->node_name, ret);
 308     }
 309     assert(acb->count <= s->num_children);
 310     assert(acb->success_count <= s->num_children);
 311     if (acb->count < s->num_children) {
 312         return;
 313     }
 314
 315     /* Do the vote on read */
 316     if (acb->is_read) {
 317         rewrite = quorum_vote(acb);
 318     } else {
 319         quorum_has_too_much_io_failed(acb);
 320     }
 321
 322     /* if no rewrite is done the code will finish right away */
 323     if (!rewrite) {
 324         quorum_aio_finalize(acb);
 325     }
 326 }
 327
 328 static void quorum_report_bad_versions(BDRVQuorumState *s,
 329                                        QuorumAIOCB *acb,
 330                                        QuorumVoteValue *value)
 331 {
 332     QuorumVoteVersion *version;
 333     QuorumVoteItem *item;
 334
 335     QLIST_FOREACH(version, &acb->votes.vote_list, next) {
 336         if (acb->votes.compare(&version->value, value)) {
 337             continue;
 338         }
 339         QLIST_FOREACH(item, &version->items, next) {
 340             quorum_report_bad(acb, s->bs[item->index]->node_name, 0);
 341         }
 342     }
 343 }
 344
 345 static bool quorum_rewrite_bad_versions(BDRVQuorumState *s, QuorumAIOCB *acb,
 346                                         QuorumVoteValue *value)
 347 {
 348     QuorumVoteVersion *version;
 349     QuorumVoteItem *item;
 350     int count = 0;
 351
 352     /* first count the number of bad versions: done first to avoid concurrency
 353      * issues.
 354      */
 355     QLIST_FOREACH(version, &acb->votes.vote_list, next) {
 356         if (acb->votes.compare(&version->value, value)) {
 357             continue;
 358         }
 359         QLIST_FOREACH(item, &version->items, next) {
 360             count++;
 361         }
 362     }
 363
 364     /* quorum_rewrite_aio_cb will count down this to zero */
 365     acb->rewrite_count = count;
 366
 367     /* now fire the correcting rewrites */
 368     QLIST_FOREACH(version, &acb->votes.vote_list, next) {
 369         if (acb->votes.compare(&version->value, value)) {
 370             continue;
 371         }
 372         QLIST_FOREACH(item, &version->items, next) {
 373             bdrv_aio_writev(s->bs[item->index], acb->sector_num, acb->qiov,
 374                             acb->nb_sectors, quorum_rewrite_aio_cb, acb);
 375         }
 376     }
 377
 378     /* return true if any rewrite is done else false */
 379     return count;
 380 }
 381
 382 static void quorum_count_vote(QuorumVotes *votes,
 383                               QuorumVoteValue *value,
 384                               int index)
 385 {
 386     QuorumVoteVersion *v = NULL, *version = NULL;
 387     QuorumVoteItem *item;
 388
 389     /* look if we have something with this hash */
 390     QLIST_FOREACH(v, &votes->vote_list, next) {
 391         if (votes->compare(&v->value, value)) {
 392             version = v;
 393             break;
 394         }
 395     }
 396
 397     /* It's a version not yet in the list add it */
 398     if (!version) {
 399         version = g_new0(QuorumVoteVersion, 1);
 400         QLIST_INIT(&version->items);
 401         memcpy(&version->value, value, sizeof(version->value));
 402         version->index = index;
 403         version->vote_count = 0;
 404         QLIST_INSERT_HEAD(&votes->vote_list, version, next);
 405     }
 406
 407     version->vote_count++;
 408
 409     item = g_new0(QuorumVoteItem, 1);
 410     item->index = index;
 411     QLIST_INSERT_HEAD(&version->items, item, next);
 412 }
 413
 414 static void quorum_free_vote_list(QuorumVotes *votes)
 415 {
 416     QuorumVoteVersion *version, *next_version;
 417     QuorumVoteItem *item, *next_item;
 418
 419     QLIST_FOREACH_SAFE(version, &votes->vote_list, next, next_version) {
 420         QLIST_REMOVE(version, next);
 421         QLIST_FOREACH_SAFE(item, &version->items, next, next_item) {
 422             QLIST_REMOVE(item, next);
 423             g_free(item);
 424         }
 425         g_free(version);
 426     }
 427 }
 428
 429 static int quorum_compute_hash(QuorumAIOCB *acb, int i, QuorumVoteValue *hash)
 430 {
 431     int j, ret;
 432     gnutls_hash_hd_t dig;
 433     QEMUIOVector *qiov = &acb->qcrs[i].qiov;
 434
 435     ret = gnutls_hash_init(&dig, GNUTLS_DIG_SHA256);
 436
 437     if (ret < 0) {
 438         return ret;
 439     }
 440
 441     for (j = 0; j < qiov->niov; j++) {
 442         ret = gnutls_hash(dig, qiov->iov[j].iov_base, qiov->iov[j].iov_len);
 443         if (ret < 0) {
 444             break;
 445         }
 446     }
 447
 448     gnutls_hash_deinit(dig, (void *) hash);
 449     return ret;
 450 }
 451
 452 static QuorumVoteVersion *quorum_get_vote_winner(QuorumVotes *votes)
 453 {
 454     int max = 0;
 455     QuorumVoteVersion *candidate, *winner = NULL;
 456
 457     QLIST_FOREACH(candidate, &votes->vote_list, next) {
 458         if (candidate->vote_count > max) {
 459             max = candidate->vote_count;
 460             winner = candidate;
 461         }
 462     }
 463
 464     return winner;
 465 }
 466
 467 /* qemu_iovec_compare is handy for blkverify mode because it returns the first
 468  * differing byte location. Yet it is handcoded to compare vectors one byte
 469  * after another so it does not benefit from the libc SIMD optimizations.
 470  * quorum_iovec_compare is written for speed and should be used in the non
 471  * blkverify mode of quorum.
 472  */
 473 static bool quorum_iovec_compare(QEMUIOVector *a, QEMUIOVector *b)
 474 {
 475     int i;
 476     int result;
 477
 478     assert(a->niov == b->niov);
 479     for (i = 0; i < a->niov; i++) {
 480         assert(a->iov[i].iov_len == b->iov[i].iov_len);
 481         result = memcmp(a->iov[i].iov_base,
 482                         b->iov[i].iov_base,
 483                         a->iov[i].iov_len);
 484         if (result) {
 485             return false;
 486         }
 487     }
 488
 489     return true;
 490 }
 491
 492 static void GCC_FMT_ATTR(2, 3) quorum_err(QuorumAIOCB *acb,
 493                                           const char *fmt, ...)
 494 {
 495     va_list ap;
 496
 497     va_start(ap, fmt);
 498     fprintf(stderr, "quorum: sector_num=%" PRId64 " nb_sectors=%d ",
 499             acb->sector_num, acb->nb_sectors);
 500     vfprintf(stderr, fmt, ap);
 501     fprintf(stderr, "\n");
 502     va_end(ap);
 503     exit(1);
 504 }
 505
 506 static bool quorum_compare(QuorumAIOCB *acb,
 507                            QEMUIOVector *a,
 508                            QEMUIOVector *b)
 509 {
 510     BDRVQuorumState *s = acb->common.bs->opaque;
 511     ssize_t offset;
 512
 513     /* This driver will replace blkverify in this particular case */
 514     if (s->is_blkverify) {
 515         offset = qemu_iovec_compare(a, b);
 516         if (offset != -1) {
 517             quorum_err(acb, "contents mismatch in sector %" PRId64,
 518                        acb->sector_num +
 519                        (uint64_t)(offset / BDRV_SECTOR_SIZE));
 520         }
 521         return true;
 522     }
 523
 524     return quorum_iovec_compare(a, b);
 525 }
 526
 527 /* Do a vote to get the error code */
 528 static int quorum_vote_error(QuorumAIOCB *acb)
 529 {
 530     BDRVQuorumState *s = acb->common.bs->opaque;
 531     QuorumVoteVersion *winner = NULL;
 532     QuorumVotes error_votes;
 533     QuorumVoteValue result_value;
 534     int i, ret = 0;
 535     bool error = false;
 536
 537     QLIST_INIT(&error_votes.vote_list);
 538     error_votes.compare = quorum_64bits_compare;
 539
 540     for (i = 0; i < s->num_children; i++) {
 541         ret = acb->qcrs[i].ret;
 542         if (ret) {
 543             error = true;
 544             result_value.l = ret;
 545             quorum_count_vote(&error_votes, &result_value, i);
 546         }
 547     }
 548
 549     if (error) {
 550         winner = quorum_get_vote_winner(&error_votes);
 551         ret = winner->value.l;
 552     }
 553
 554     quorum_free_vote_list(&error_votes);
 555
 556     return ret;
 557 }
 558
 559 static bool quorum_vote(QuorumAIOCB *acb)
 560 {
 561     bool quorum = true;
 562     bool rewrite = false;
 563     int i, j, ret;
 564     QuorumVoteValue hash;
 565     BDRVQuorumState *s = acb->common.bs->opaque;
 566     QuorumVoteVersion *winner;
 567
 568     if (quorum_has_too_much_io_failed(acb)) {
 569         return false;
 570     }
 571
 572     /* get the index of the first successful read */
 573     for (i = 0; i < s->num_children; i++) {
 574         if (!acb->qcrs[i].ret) {
 575             break;
 576         }
 577     }
 578
 579     assert(i < s->num_children);
 580
 581     /* compare this read with all other successful reads stopping at quorum
 582      * failure
 583      */
 584     for (j = i + 1; j < s->num_children; j++) {
 585         if (acb->qcrs[j].ret) {
 586             continue;
 587         }
 588         quorum = quorum_compare(acb, &acb->qcrs[i].qiov, &acb->qcrs[j].qiov);
 589         if (!quorum) {
 590             break;
 591        }
 592     }
 593
 594     /* Every successful read agrees */
 595     if (quorum) {
 596         quorum_copy_qiov(acb->qiov, &acb->qcrs[i].qiov);
 597         return false;
 598     }
 599
 600     /* compute hashes for each successful read, also store indexes */
 601     for (i = 0; i < s->num_children; i++) {
 602         if (acb->qcrs[i].ret) {
 603             continue;
 604         }
 605         ret = quorum_compute_hash(acb, i, &hash);
 606         /* if ever the hash computation failed */
 607         if (ret < 0) {
 608             acb->vote_ret = ret;
 609             goto free_exit;
 610         }
 611         quorum_count_vote(&acb->votes, &hash, i);
 612     }
 613
 614     /* vote to select the most represented version */
 615     winner = quorum_get_vote_winner(&acb->votes);
 616
 617     /* if the winner count is smaller than threshold the read fails */
 618     if (winner->vote_count < s->threshold) {
 619         quorum_report_failure(acb);
 620         acb->vote_ret = -EIO;
 621         goto free_exit;
 622     }
 623
 624     /* we have a winner: copy it */
 625     quorum_copy_qiov(acb->qiov, &acb->qcrs[winner->index].qiov);
 626
 627     /* some versions are bad print them */
 628     quorum_report_bad_versions(s, acb, &winner->value);
 629
 630     /* corruption correction is enabled */
 631     if (s->rewrite_corrupted) {
 632         rewrite = quorum_rewrite_bad_versions(s, acb, &winner->value);
 633     }
 634
 635 free_exit:
 636     /* free lists */
 637     quorum_free_vote_list(&acb->votes);
 638     return rewrite;
 639 }
 640
 641 static BlockAIOCB *read_quorum_children(QuorumAIOCB *acb)
 642 {
 643     BDRVQuorumState *s = acb->common.bs->opaque;
 644     int i;
 645
 646     for (i = 0; i < s->num_children; i++) {
 647         acb->qcrs[i].buf = qemu_blockalign(s->bs[i], acb->qiov->size);
 648         qemu_iovec_init(&acb->qcrs[i].qiov, acb->qiov->niov);
 649         qemu_iovec_clone(&acb->qcrs[i].qiov, acb->qiov, acb->qcrs[i].buf);
 650     }
 651
 652     for (i = 0; i < s->num_children; i++) {
 653         bdrv_aio_readv(s->bs[i], acb->sector_num, &acb->qcrs[i].qiov,
 654                        acb->nb_sectors, quorum_aio_cb, &acb->qcrs[i]);
 655     }
 656
 657     return &acb->common;
 658 }
 659
 660 static BlockAIOCB *read_fifo_child(QuorumAIOCB *acb)
 661 {
 662     BDRVQuorumState *s = acb->common.bs->opaque;
 663
 664     acb->qcrs[acb->child_iter].buf = qemu_blockalign(s->bs[acb->child_iter],
 665                                                      acb->qiov->size);
 666     qemu_iovec_init(&acb->qcrs[acb->child_iter].qiov, acb->qiov->niov);
 667     qemu_iovec_clone(&acb->qcrs[acb->child_iter].qiov, acb->qiov,
 668                      acb->qcrs[acb->child_iter].buf);
 669     bdrv_aio_readv(s->bs[acb->child_iter], acb->sector_num,
 670                    &acb->qcrs[acb->child_iter].qiov, acb->nb_sectors,
 671                    quorum_aio_cb, &acb->qcrs[acb->child_iter]);
 672
 673     return &acb->common;
 674 }
 675
 676 static BlockAIOCB *quorum_aio_readv(BlockDriverState *bs,
 677                                     int64_t sector_num,
 678                                     QEMUIOVector *qiov,
 679                                     int nb_sectors,
 680                                     BlockCompletionFunc *cb,
 681                                     void *opaque)
 682 {
 683     BDRVQuorumState *s = bs->opaque;
 684     QuorumAIOCB *acb = quorum_aio_get(s, bs, qiov, sector_num,
 685                                       nb_sectors, cb, opaque);
 686     acb->is_read = true;
 687
 688     if (s->read_pattern == QUORUM_READ_PATTERN_QUORUM) {
 689         acb->child_iter = s->num_children - 1;
 690         return read_quorum_children(acb);
 691     }
 692
 693     acb->child_iter = 0;
 694     return read_fifo_child(acb);
 695 }
 696
 697 static BlockAIOCB *quorum_aio_writev(BlockDriverState *bs,
 698                                      int64_t sector_num,
 699                                      QEMUIOVector *qiov,
 700                                      int nb_sectors,
 701                                      BlockCompletionFunc *cb,
 702                                      void *opaque)
 703 {
 704     BDRVQuorumState *s = bs->opaque;
 705     QuorumAIOCB *acb = quorum_aio_get(s, bs, qiov, sector_num, nb_sectors,
 706                                       cb, opaque);
 707     int i;
 708
 709     for (i = 0; i < s->num_children; i++) {
 710         acb->qcrs[i].aiocb = bdrv_aio_writev(s->bs[i], sector_num, qiov,
 711                                              nb_sectors, &quorum_aio_cb,
 712                                              &acb->qcrs[i]);
 713     }
 714
 715     return &acb->common;
 716 }
 717
 718 static int64_t quorum_getlength(BlockDriverState *bs)
 719 {
 720     BDRVQuorumState *s = bs->opaque;
 721     int64_t result;
 722     int i;
 723
 724     /* check that all file have the same length */
 725     result = bdrv_getlength(s->bs[0]);
 726     if (result < 0) {
 727         return result;
 728     }
 729     for (i = 1; i < s->num_children; i++) {
 730         int64_t value = bdrv_getlength(s->bs[i]);
 731         if (value < 0) {
 732             return value;
 733         }
 734         if (value != result) {
 735             return -EIO;
 736         }
 737     }
 738
 739     return result;
 740 }
 741
 742 static void quorum_invalidate_cache(BlockDriverState *bs, Error **errp)
 743 {
 744     BDRVQuorumState *s = bs->opaque;
 745     Error *local_err = NULL;
 746     int i;
 747
 748     for (i = 0; i < s->num_children; i++) {
 749         bdrv_invalidate_cache(s->bs[i], &local_err);
 750         if (local_err) {
 751             error_propagate(errp, local_err);
 752             return;
 753         }
 754     }
 755 }
 756
 757 static coroutine_fn int quorum_co_flush(BlockDriverState *bs)
 758 {
 759     BDRVQuorumState *s = bs->opaque;
 760     QuorumVoteVersion *winner = NULL;
 761     QuorumVotes error_votes;
 762     QuorumVoteValue result_value;
 763     int i;
 764     int result = 0;
 765
 766     QLIST_INIT(&error_votes.vote_list);
 767     error_votes.compare = quorum_64bits_compare;
 768
 769     for (i = 0; i < s->num_children; i++) {
 770         result = bdrv_co_flush(s->bs[i]);
 771         result_value.l = result;
 772         quorum_count_vote(&error_votes, &result_value, i);
 773     }
 774
 775     winner = quorum_get_vote_winner(&error_votes);
 776     result = winner->value.l;
 777
 778     quorum_free_vote_list(&error_votes);
 779
 780     return result;
 781 }
 782
 783 static bool quorum_recurse_is_first_non_filter(BlockDriverState *bs,
 784                                                BlockDriverState *candidate)
 785 {
 786     BDRVQuorumState *s = bs->opaque;
 787     int i;
 788
 789     for (i = 0; i < s->num_children; i++) {
 790         bool perm = bdrv_recurse_is_first_non_filter(s->bs[i],
 791                                                      candidate);
 792         if (perm) {
 793             return true;
 794         }
 795     }
 796
 797     return false;
 798 }
 799
 800 static int quorum_valid_threshold(int threshold, int num_children, Error **errp)
 801 {
 802
 803     if (threshold < 1) {
 804         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 805                    "vote-threshold", "value >= 1");
 806         return -ERANGE;
 807     }
 808
 809     if (threshold > num_children) {
 810         error_setg(errp, "threshold may not exceed children count");
 811         return -ERANGE;
 812     }
 813
 814     return 0;
 815 }
 816
 817 static QemuOptsList quorum_runtime_opts = {
 818     .name = "quorum",
 819     .head = QTAILQ_HEAD_INITIALIZER(quorum_runtime_opts.head),
 820     .desc = {
 821         {
 822             .name = QUORUM_OPT_VOTE_THRESHOLD,
 823             .type = QEMU_OPT_NUMBER,
 824             .help = "The number of vote needed for reaching quorum",
 825         },
 826         {
 827             .name = QUORUM_OPT_BLKVERIFY,
 828             .type = QEMU_OPT_BOOL,
 829             .help = "Trigger block verify mode if set",
 830         },
 831         {
 832             .name = QUORUM_OPT_REWRITE,
 833             .type = QEMU_OPT_BOOL,
 834             .help = "Rewrite corrupted block on read quorum",
 835         },
 836         {
 837             .name = QUORUM_OPT_READ_PATTERN,
 838             .type = QEMU_OPT_STRING,
 839             .help = "Allowed pattern: quorum, fifo. Quorum is default",
 840         },
 841         { /* end of list */ }
 842     },
 843 };
 844
 845 static int parse_read_pattern(const char *opt)
 846 {
 847     int i;
 848
 849     if (!opt) {
 850         /* Set quorum as default */
 851         return QUORUM_READ_PATTERN_QUORUM;
 852     }
 853
 854     for (i = 0; i < QUORUM_READ_PATTERN_MAX; i++) {
 855         if (!strcmp(opt, QuorumReadPattern_lookup[i])) {
 856             return i;
 857         }
 858     }
 859
 860     return -EINVAL;
 861 }
 862
 863 static int quorum_open(BlockDriverState *bs, QDict *options, int flags,
 864                        Error **errp)
 865 {
 866     BDRVQuorumState *s = bs->opaque;
 867     Error *local_err = NULL;
 868     QemuOpts *opts = NULL;
 869     bool *opened;
 870     int i;
 871     int ret = 0;
 872
 873     qdict_flatten(options);
 874
 875     /* count how many different children are present */
 876     s->num_children = qdict_array_entries(options, "children.");
 877     if (s->num_children < 0) {
 878         error_setg(&local_err, "Option children is not a valid array");
 879         ret = -EINVAL;
 880         goto exit;
 881     }
 882     if (s->num_children < 2) {
 883         error_setg(&local_err,
 884                    "Number of provided children must be greater than 1");
 885         ret = -EINVAL;
 886         goto exit;
 887     }
 888
 889     opts = qemu_opts_create(&quorum_runtime_opts, NULL, 0, &error_abort);
 890     qemu_opts_absorb_qdict(opts, options, &local_err);
 891     if (local_err) {
 892         ret = -EINVAL;
 893         goto exit;
 894     }
 895
 896     s->threshold = qemu_opt_get_number(opts, QUORUM_OPT_VOTE_THRESHOLD, 0);
 897     ret = parse_read_pattern(qemu_opt_get(opts, QUORUM_OPT_READ_PATTERN));
 898     if (ret < 0) {
 899         error_setg(&local_err, "Please set read-pattern as fifo or quorum");
 900         goto exit;
 901     }
 902     s->read_pattern = ret;
 903
 904     if (s->read_pattern == QUORUM_READ_PATTERN_QUORUM) {
 905         /* and validate it against s->num_children */
 906         ret = quorum_valid_threshold(s->threshold, s->num_children, &local_err);
 907         if (ret < 0) {
 908             goto exit;
 909         }
 910
 911         /* is the driver in blkverify mode */
 912         if (qemu_opt_get_bool(opts, QUORUM_OPT_BLKVERIFY, false) &&
 913             s->num_children == 2 && s->threshold == 2) {
 914             s->is_blkverify = true;
 915         } else if (qemu_opt_get_bool(opts, QUORUM_OPT_BLKVERIFY, false)) {
 916             fprintf(stderr, "blkverify mode is set by setting blkverify=on "
 917                     "and using two files with vote_threshold=2\n");
 918         }
 919
 920         s->rewrite_corrupted = qemu_opt_get_bool(opts, QUORUM_OPT_REWRITE,
 921                                                  false);
 922         if (s->rewrite_corrupted && s->is_blkverify) {
 923             error_setg(&local_err,
 924                        "rewrite-corrupted=on cannot be used with blkverify=on");
 925             ret = -EINVAL;
 926             goto exit;
 927         }
 928     }
 929
 930     /* allocate the children BlockDriverState array */
 931     s->bs = g_new0(BlockDriverState *, s->num_children);
 932     opened = g_new0(bool, s->num_children);
 933
 934     for (i = 0; i < s->num_children; i++) {
 935         char indexstr[32];
 936         ret = snprintf(indexstr, 32, "children.%d", i);
 937         assert(ret < 32);
 938
 939         ret = bdrv_open_image(&s->bs[i], NULL, options, indexstr, bs,
 940                               &child_format, false, &local_err);
 941         if (ret < 0) {
 942             goto close_exit;
 943         }
 944
 945         opened[i] = true;
 946     }
 947
 948     g_free(opened);
 949     goto exit;
 950
 951 close_exit:
 952     /* cleanup on error */
 953     for (i = 0; i < s->num_children; i++) {
 954         if (!opened[i]) {
 955             continue;
 956         }
 957         bdrv_unref(s->bs[i]);
 958     }
 959     g_free(s->bs);
 960     g_free(opened);
 961 exit:
 962     qemu_opts_del(opts);
 963     /* propagate error */
 964     if (local_err) {
 965         error_propagate(errp, local_err);
 966     }
 967     return ret;
 968 }
 969
 970 static void quorum_close(BlockDriverState *bs)
 971 {
 972     BDRVQuorumState *s = bs->opaque;
 973     int i;
 974
 975     for (i = 0; i < s->num_children; i++) {
 976         bdrv_unref(s->bs[i]);
 977     }
 978
 979     g_free(s->bs);
 980 }
 981
 982 static void quorum_detach_aio_context(BlockDriverState *bs)
 983 {
 984     BDRVQuorumState *s = bs->opaque;
 985     int i;
 986
 987     for (i = 0; i < s->num_children; i++) {
 988         bdrv_detach_aio_context(s->bs[i]);
 989     }
 990 }
 991
 992 static void quorum_attach_aio_context(BlockDriverState *bs,
 993                                       AioContext *new_context)
 994 {
 995     BDRVQuorumState *s = bs->opaque;
 996     int i;
 997
 998     for (i = 0; i < s->num_children; i++) {
 999         bdrv_attach_aio_context(s->bs[i], new_context);
1000     }
1001 }
1002
1003 static void quorum_refresh_filename(BlockDriverState *bs)
1004 {
1005     BDRVQuorumState *s = bs->opaque;
1006     QDict *opts;
1007     QList *children;
1008     int i;
1009
1010     for (i = 0; i < s->num_children; i++) {
1011         bdrv_refresh_filename(s->bs[i]);
1012         if (!s->bs[i]->full_open_options) {
1013             return;
1014         }
1015     }
1016
1017     children = qlist_new();
1018     for (i = 0; i < s->num_children; i++) {
1019         QINCREF(s->bs[i]->full_open_options);
1020         qlist_append_obj(children, QOBJECT(s->bs[i]->full_open_options));
1021     }
1022
1023     opts = qdict_new();
1024     qdict_put_obj(opts, "driver", QOBJECT(qstring_from_str("quorum")));
1025     qdict_put_obj(opts, QUORUM_OPT_VOTE_THRESHOLD,
1026                   QOBJECT(qint_from_int(s->threshold)));
1027     qdict_put_obj(opts, QUORUM_OPT_BLKVERIFY,
1028                   QOBJECT(qbool_from_bool(s->is_blkverify)));
1029     qdict_put_obj(opts, QUORUM_OPT_REWRITE,
1030                   QOBJECT(qbool_from_bool(s->rewrite_corrupted)));
1031     qdict_put_obj(opts, "children", QOBJECT(children));
1032
1033     bs->full_open_options = opts;
1034 }
1035
1036 static BlockDriver bdrv_quorum = {
1037     .format_name                        = "quorum",
1038     .protocol_name                      = "quorum",
1039
1040     .instance_size                      = sizeof(BDRVQuorumState),
1041
1042     .bdrv_file_open                     = quorum_open,
1043     .bdrv_close                         = quorum_close,
1044     .bdrv_refresh_filename              = quorum_refresh_filename,
1045
1046     .bdrv_co_flush_to_disk              = quorum_co_flush,
1047
1048     .bdrv_getlength                     = quorum_getlength,
1049
1050     .bdrv_aio_readv                     = quorum_aio_readv,
1051     .bdrv_aio_writev                    = quorum_aio_writev,
1052     .bdrv_invalidate_cache              = quorum_invalidate_cache,
1053
1054     .bdrv_detach_aio_context            = quorum_detach_aio_context,
1055     .bdrv_attach_aio_context            = quorum_attach_aio_context,
1056
1057     .is_filter                          = true,
1058     .bdrv_recurse_is_first_non_filter   = quorum_recurse_is_first_non_filter,
1059 };
1060
1061 static void bdrv_quorum_init(void)
1062 {
1063     bdrv_register(&bdrv_quorum);
1064 }
1065
1066 block_init(bdrv_quorum_init);