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 "block/block_int.h"
  17 #include "qapi/qmp/qbool.h"
  18 #include "qapi/qmp/qdict.h"
  19 #include "qapi/qmp/qerror.h"
  20 #include "qapi/qmp/qint.h"
  21 #include "qapi/qmp/qjson.h"
  22 #include "qapi/qmp/qlist.h"
  23 #include "qapi/qmp/qstring.h"
  24 #include "qapi-event.h"
  25 #include "crypto/hash.h"
  26
  27 #define HASH_LENGTH 32
  28
  29 #define QUORUM_OPT_VOTE_THRESHOLD "vote-threshold"
  30 #define QUORUM_OPT_BLKVERIFY      "blkverify"
  31 #define QUORUM_OPT_REWRITE        "rewrite-corrupted"
  32 #define QUORUM_OPT_READ_PATTERN   "read-pattern"
  33
  34 /* This union holds a vote hash value */
  35 typedef union QuorumVoteValue {
  36     uint8_t h[HASH_LENGTH];    /* SHA-256 hash */
  37     int64_t l;                 /* simpler 64 bits hash */
  38 } QuorumVoteValue;
  39
  40 /* A vote item */
  41 typedef struct QuorumVoteItem {
  42     int index;
  43     QLIST_ENTRY(QuorumVoteItem) next;
  44 } QuorumVoteItem;
  45
  46 /* this structure is a vote version. A version is the set of votes sharing the
  47  * same vote value.
  48  * The set of votes will be tracked with the items field and its cardinality is
  49  * vote_count.
  50  */
  51 typedef struct QuorumVoteVersion {
  52     QuorumVoteValue value;
  53     int index;
  54     int vote_count;
  55     QLIST_HEAD(, QuorumVoteItem) items;
  56     QLIST_ENTRY(QuorumVoteVersion) next;
  57 } QuorumVoteVersion;
  58
  59 /* this structure holds a group of vote versions together */
  60 typedef struct QuorumVotes {
  61     QLIST_HEAD(, QuorumVoteVersion) vote_list;
  62     bool (*compare)(QuorumVoteValue *a, QuorumVoteValue *b);
  63 } QuorumVotes;
  64
  65 /* the following structure holds the state of one quorum instance */
  66 typedef struct BDRVQuorumState {
  67     BlockDriverState **bs; /* children BlockDriverStates */
  68     int num_children;      /* children count */
  69     int threshold;         /* if less than threshold children reads gave the
  70                             * same result a quorum error occurs.
  71                             */
  72     bool is_blkverify;     /* true if the driver is in blkverify mode
  73                             * Writes are mirrored on two children devices.
  74                             * On reads the two children devices' contents are
  75                             * compared and if a difference is spotted its
  76                             * location is printed and the code aborts.
  77                             * It is useful to debug other block drivers by
  78                             * comparing them with a reference one.
  79                             */
  80     bool rewrite_corrupted;/* true if the driver must rewrite-on-read corrupted
  81                             * block if Quorum is reached.
  82                             */
  83
  84     QuorumReadPattern read_pattern;
  85 } BDRVQuorumState;
  86
  87 typedef struct QuorumAIOCB QuorumAIOCB;
  88
  89 /* Quorum will create one instance of the following structure per operation it
  90  * performs on its children.
  91  * So for each read/write operation coming from the upper layer there will be
  92  * $children_count QuorumChildRequest.
  93  */
  94 typedef struct QuorumChildRequest {
  95     BlockAIOCB *aiocb;
  96     QEMUIOVector qiov;
  97     uint8_t *buf;
  98     int ret;
  99     QuorumAIOCB *parent;
 100 } QuorumChildRequest;
 101
 102 /* Quorum will use the following structure to track progress of each read/write
 103  * operation received by the upper layer.
 104  * This structure hold pointers to the QuorumChildRequest structures instances
 105  * used to do operations on each children and track overall progress.
 106  */
 107 struct QuorumAIOCB {
 108     BlockAIOCB common;
 109
 110     /* Request metadata */
 111     uint64_t sector_num;
 112     int nb_sectors;
 113
 114     QEMUIOVector *qiov;         /* calling IOV */
 115
 116     QuorumChildRequest *qcrs;   /* individual child requests */
 117     int count;                  /* number of completed AIOCB */
 118     int success_count;          /* number of successfully completed AIOCB */
 119
 120     int rewrite_count;          /* number of replica to rewrite: count down to
 121                                  * zero once writes are fired
 122                                  */
 123
 124     QuorumVotes votes;
 125
 126     bool is_read;
 127     int vote_ret;
 128     int child_iter;             /* which child to read in fifo pattern */
 129 };
 130
 131 static bool quorum_vote(QuorumAIOCB *acb);
 132
 133 static void quorum_aio_cancel(BlockAIOCB *blockacb)
 134 {
 135     QuorumAIOCB *acb = container_of(blockacb, QuorumAIOCB, common);
 136     BDRVQuorumState *s = acb->common.bs->opaque;
 137     int i;
 138
 139     /* cancel all callbacks */
 140     for (i = 0; i < s->num_children; i++) {
 141         if (acb->qcrs[i].aiocb) {
 142             bdrv_aio_cancel_async(acb->qcrs[i].aiocb);
 143         }
 144     }
 145 }
 146
 147 static AIOCBInfo quorum_aiocb_info = {
 148     .aiocb_size         = sizeof(QuorumAIOCB),
 149     .cancel_async       = quorum_aio_cancel,
 150 };
 151
 152 static void quorum_aio_finalize(QuorumAIOCB *acb)
 153 {
 154     int i, ret = 0;
 155
 156     if (acb->vote_ret) {
 157         ret = acb->vote_ret;
 158     }
 159
 160     acb->common.cb(acb->common.opaque, ret);
 161
 162     if (acb->is_read) {
 163         /* on the quorum case acb->child_iter == s->num_children - 1 */
 164         for (i = 0; i <= acb->child_iter; i++) {
 165             qemu_vfree(acb->qcrs[i].buf);
 166             qemu_iovec_destroy(&acb->qcrs[i].qiov);
 167         }
 168     }
 169
 170     g_free(acb->qcrs);
 171     qemu_aio_unref(acb);
 172 }
 173
 174 static bool quorum_sha256_compare(QuorumVoteValue *a, QuorumVoteValue *b)
 175 {
 176     return !memcmp(a->h, b->h, HASH_LENGTH);
 177 }
 178
 179 static bool quorum_64bits_compare(QuorumVoteValue *a, QuorumVoteValue *b)
 180 {
 181     return a->l == b->l;
 182 }
 183
 184 static QuorumAIOCB *quorum_aio_get(BDRVQuorumState *s,
 185                                    BlockDriverState *bs,
 186                                    QEMUIOVector *qiov,
 187                                    uint64_t sector_num,
 188                                    int nb_sectors,
 189                                    BlockCompletionFunc *cb,
 190                                    void *opaque)
 191 {
 192     QuorumAIOCB *acb = qemu_aio_get(&quorum_aiocb_info, bs, cb, opaque);
 193     int i;
 194
 195     acb->common.bs->opaque = s;
 196     acb->sector_num = sector_num;
 197     acb->nb_sectors = nb_sectors;
 198     acb->qiov = qiov;
 199     acb->qcrs = g_new0(QuorumChildRequest, s->num_children);
 200     acb->count = 0;
 201     acb->success_count = 0;
 202     acb->rewrite_count = 0;
 203     acb->votes.compare = quorum_sha256_compare;
 204     QLIST_INIT(&acb->votes.vote_list);
 205     acb->is_read = false;
 206     acb->vote_ret = 0;
 207
 208     for (i = 0; i < s->num_children; i++) {
 209         acb->qcrs[i].buf = NULL;
 210         acb->qcrs[i].ret = 0;
 211         acb->qcrs[i].parent = acb;
 212     }
 213
 214     return acb;
 215 }
 216
 217 static void quorum_report_bad(QuorumAIOCB *acb, char *node_name, int ret)
 218 {
 219     const char *msg = NULL;
 220     if (ret < 0) {
 221         msg = strerror(-ret);
 222     }
 223     qapi_event_send_quorum_report_bad(!!msg, msg, node_name,
 224                                       acb->sector_num, acb->nb_sectors, &error_abort);
 225 }
 226
 227 static void quorum_report_failure(QuorumAIOCB *acb)
 228 {
 229     const char *reference = bdrv_get_device_or_node_name(acb->common.bs);
 230     qapi_event_send_quorum_failure(reference, acb->sector_num,
 231                                    acb->nb_sectors, &error_abort);
 232 }
 233
 234 static int quorum_vote_error(QuorumAIOCB *acb);
 235
 236 static bool quorum_has_too_much_io_failed(QuorumAIOCB *acb)
 237 {
 238     BDRVQuorumState *s = acb->common.bs->opaque;
 239
 240     if (acb->success_count < s->threshold) {
 241         acb->vote_ret = quorum_vote_error(acb);
 242         quorum_report_failure(acb);
 243         return true;
 244     }
 245
 246     return false;
 247 }
 248
 249 static void quorum_rewrite_aio_cb(void *opaque, int ret)
 250 {
 251     QuorumAIOCB *acb = opaque;
 252
 253     /* one less rewrite to do */
 254     acb->rewrite_count--;
 255
 256     /* wait until all rewrite callbacks have completed */
 257     if (acb->rewrite_count) {
 258         return;
 259     }
 260
 261     quorum_aio_finalize(acb);
 262 }
 263
 264 static BlockAIOCB *read_fifo_child(QuorumAIOCB *acb);
 265
 266 static void quorum_copy_qiov(QEMUIOVector *dest, QEMUIOVector *source)
 267 {
 268     int i;
 269     assert(dest->niov == source->niov);
 270     assert(dest->size == source->size);
 271     for (i = 0; i < source->niov; i++) {
 272         assert(dest->iov[i].iov_len == source->iov[i].iov_len);
 273         memcpy(dest->iov[i].iov_base,
 274                source->iov[i].iov_base,
 275                source->iov[i].iov_len);
 276     }
 277 }
 278
 279 static void quorum_aio_cb(void *opaque, int ret)
 280 {
 281     QuorumChildRequest *sacb = opaque;
 282     QuorumAIOCB *acb = sacb->parent;
 283     BDRVQuorumState *s = acb->common.bs->opaque;
 284     bool rewrite = false;
 285
 286     if (acb->is_read && s->read_pattern == QUORUM_READ_PATTERN_FIFO) {
 287         /* We try to read next child in FIFO order if we fail to read */
 288         if (ret < 0 && ++acb->child_iter < s->num_children) {
 289             read_fifo_child(acb);
 290             return;
 291         }
 292
 293         if (ret == 0) {
 294             quorum_copy_qiov(acb->qiov, &acb->qcrs[acb->child_iter].qiov);
 295         }
 296         acb->vote_ret = ret;
 297         quorum_aio_finalize(acb);
 298         return;
 299     }
 300
 301     sacb->ret = ret;
 302     acb->count++;
 303     if (ret == 0) {
 304         acb->success_count++;
 305     } else {
 306         quorum_report_bad(acb, sacb->aiocb->bs->node_name, ret);
 307     }
 308     assert(acb->count <= s->num_children);
 309     assert(acb->success_count <= s->num_children);
 310     if (acb->count < s->num_children) {
 311         return;
 312     }
 313
 314     /* Do the vote on read */
 315     if (acb->is_read) {
 316         rewrite = quorum_vote(acb);
 317     } else {
 318         quorum_has_too_much_io_failed(acb);
 319     }
 320
 321     /* if no rewrite is done the code will finish right away */
 322     if (!rewrite) {
 323         quorum_aio_finalize(acb);
 324     }
 325 }
 326
 327 static void quorum_report_bad_versions(BDRVQuorumState *s,
 328                                        QuorumAIOCB *acb,
 329                                        QuorumVoteValue *value)
 330 {
 331     QuorumVoteVersion *version;
 332     QuorumVoteItem *item;
 333
 334     QLIST_FOREACH(version, &acb->votes.vote_list, next) {
 335         if (acb->votes.compare(&version->value, value)) {
 336             continue;
 337         }
 338         QLIST_FOREACH(item, &version->items, next) {
 339             quorum_report_bad(acb, s->bs[item->index]->node_name, 0);
 340         }
 341     }
 342 }
 343
 344 static bool quorum_rewrite_bad_versions(BDRVQuorumState *s, QuorumAIOCB *acb,
 345                                         QuorumVoteValue *value)
 346 {
 347     QuorumVoteVersion *version;
 348     QuorumVoteItem *item;
 349     int count = 0;
 350
 351     /* first count the number of bad versions: done first to avoid concurrency
 352      * issues.
 353      */
 354     QLIST_FOREACH(version, &acb->votes.vote_list, next) {
 355         if (acb->votes.compare(&version->value, value)) {
 356             continue;
 357         }
 358         QLIST_FOREACH(item, &version->items, next) {
 359             count++;
 360         }
 361     }
 362
 363     /* quorum_rewrite_aio_cb will count down this to zero */
 364     acb->rewrite_count = count;
 365
 366     /* now fire the correcting rewrites */
 367     QLIST_FOREACH(version, &acb->votes.vote_list, next) {
 368         if (acb->votes.compare(&version->value, value)) {
 369             continue;
 370         }
 371         QLIST_FOREACH(item, &version->items, next) {
 372             bdrv_aio_writev(s->bs[item->index], acb->sector_num, acb->qiov,
 373                             acb->nb_sectors, quorum_rewrite_aio_cb, acb);
 374         }
 375     }
 376
 377     /* return true if any rewrite is done else false */
 378     return count;
 379 }
 380
 381 static void quorum_count_vote(QuorumVotes *votes,
 382                               QuorumVoteValue *value,
 383                               int index)
 384 {
 385     QuorumVoteVersion *v = NULL, *version = NULL;
 386     QuorumVoteItem *item;
 387
 388     /* look if we have something with this hash */
 389     QLIST_FOREACH(v, &votes->vote_list, next) {
 390         if (votes->compare(&v->value, value)) {
 391             version = v;
 392             break;
 393         }
 394     }
 395
 396     /* It's a version not yet in the list add it */
 397     if (!version) {
 398         version = g_new0(QuorumVoteVersion, 1);
 399         QLIST_INIT(&version->items);
 400         memcpy(&version->value, value, sizeof(version->value));
 401         version->index = index;
 402         version->vote_count = 0;
 403         QLIST_INSERT_HEAD(&votes->vote_list, version, next);
 404     }
 405
 406     version->vote_count++;
 407
 408     item = g_new0(QuorumVoteItem, 1);
 409     item->index = index;
 410     QLIST_INSERT_HEAD(&version->items, item, next);
 411 }
 412
 413 static void quorum_free_vote_list(QuorumVotes *votes)
 414 {
 415     QuorumVoteVersion *version, *next_version;
 416     QuorumVoteItem *item, *next_item;
 417
 418     QLIST_FOREACH_SAFE(version, &votes->vote_list, next, next_version) {
 419         QLIST_REMOVE(version, next);
 420         QLIST_FOREACH_SAFE(item, &version->items, next, next_item) {
 421             QLIST_REMOVE(item, next);
 422             g_free(item);
 423         }
 424         g_free(version);
 425     }
 426 }
 427
 428 static int quorum_compute_hash(QuorumAIOCB *acb, int i, QuorumVoteValue *hash)
 429 {
 430     QEMUIOVector *qiov = &acb->qcrs[i].qiov;
 431     size_t len = sizeof(hash->h);
 432     uint8_t *data = hash->h;
 433
 434     /* XXX - would be nice if we could pass in the Error **
 435      * and propagate that back, but this quorum code is
 436      * restricted to just errno values currently */
 437     if (qcrypto_hash_bytesv(QCRYPTO_HASH_ALG_SHA256,
 438                             qiov->iov, qiov->niov,
 439                             &data, &len,
 440                             NULL) < 0) {
 441         return -EINVAL;
 442     }
 443
 444     return 0;
 445 }
 446
 447 static QuorumVoteVersion *quorum_get_vote_winner(QuorumVotes *votes)
 448 {
 449     int max = 0;
 450     QuorumVoteVersion *candidate, *winner = NULL;
 451
 452     QLIST_FOREACH(candidate, &votes->vote_list, next) {
 453         if (candidate->vote_count > max) {
 454             max = candidate->vote_count;
 455             winner = candidate;
 456         }
 457     }
 458
 459     return winner;
 460 }
 461
 462 /* qemu_iovec_compare is handy for blkverify mode because it returns the first
 463  * differing byte location. Yet it is handcoded to compare vectors one byte
 464  * after another so it does not benefit from the libc SIMD optimizations.
 465  * quorum_iovec_compare is written for speed and should be used in the non
 466  * blkverify mode of quorum.
 467  */
 468 static bool quorum_iovec_compare(QEMUIOVector *a, QEMUIOVector *b)
 469 {
 470     int i;
 471     int result;
 472
 473     assert(a->niov == b->niov);
 474     for (i = 0; i < a->niov; i++) {
 475         assert(a->iov[i].iov_len == b->iov[i].iov_len);
 476         result = memcmp(a->iov[i].iov_base,
 477                         b->iov[i].iov_base,
 478                         a->iov[i].iov_len);
 479         if (result) {
 480             return false;
 481         }
 482     }
 483
 484     return true;
 485 }
 486
 487 static void GCC_FMT_ATTR(2, 3) quorum_err(QuorumAIOCB *acb,
 488                                           const char *fmt, ...)
 489 {
 490     va_list ap;
 491
 492     va_start(ap, fmt);
 493     fprintf(stderr, "quorum: sector_num=%" PRId64 " nb_sectors=%d ",
 494             acb->sector_num, acb->nb_sectors);
 495     vfprintf(stderr, fmt, ap);
 496     fprintf(stderr, "\n");
 497     va_end(ap);
 498     exit(1);
 499 }
 500
 501 static bool quorum_compare(QuorumAIOCB *acb,
 502                            QEMUIOVector *a,
 503                            QEMUIOVector *b)
 504 {
 505     BDRVQuorumState *s = acb->common.bs->opaque;
 506     ssize_t offset;
 507
 508     /* This driver will replace blkverify in this particular case */
 509     if (s->is_blkverify) {
 510         offset = qemu_iovec_compare(a, b);
 511         if (offset != -1) {
 512             quorum_err(acb, "contents mismatch in sector %" PRId64,
 513                        acb->sector_num +
 514                        (uint64_t)(offset / BDRV_SECTOR_SIZE));
 515         }
 516         return true;
 517     }
 518
 519     return quorum_iovec_compare(a, b);
 520 }
 521
 522 /* Do a vote to get the error code */
 523 static int quorum_vote_error(QuorumAIOCB *acb)
 524 {
 525     BDRVQuorumState *s = acb->common.bs->opaque;
 526     QuorumVoteVersion *winner = NULL;
 527     QuorumVotes error_votes;
 528     QuorumVoteValue result_value;
 529     int i, ret = 0;
 530     bool error = false;
 531
 532     QLIST_INIT(&error_votes.vote_list);
 533     error_votes.compare = quorum_64bits_compare;
 534
 535     for (i = 0; i < s->num_children; i++) {
 536         ret = acb->qcrs[i].ret;
 537         if (ret) {
 538             error = true;
 539             result_value.l = ret;
 540             quorum_count_vote(&error_votes, &result_value, i);
 541         }
 542     }
 543
 544     if (error) {
 545         winner = quorum_get_vote_winner(&error_votes);
 546         ret = winner->value.l;
 547     }
 548
 549     quorum_free_vote_list(&error_votes);
 550
 551     return ret;
 552 }
 553
 554 static bool quorum_vote(QuorumAIOCB *acb)
 555 {
 556     bool quorum = true;
 557     bool rewrite = false;
 558     int i, j, ret;
 559     QuorumVoteValue hash;
 560     BDRVQuorumState *s = acb->common.bs->opaque;
 561     QuorumVoteVersion *winner;
 562
 563     if (quorum_has_too_much_io_failed(acb)) {
 564         return false;
 565     }
 566
 567     /* get the index of the first successful read */
 568     for (i = 0; i < s->num_children; i++) {
 569         if (!acb->qcrs[i].ret) {
 570             break;
 571         }
 572     }
 573
 574     assert(i < s->num_children);
 575
 576     /* compare this read with all other successful reads stopping at quorum
 577      * failure
 578      */
 579     for (j = i + 1; j < s->num_children; j++) {
 580         if (acb->qcrs[j].ret) {
 581             continue;
 582         }
 583         quorum = quorum_compare(acb, &acb->qcrs[i].qiov, &acb->qcrs[j].qiov);
 584         if (!quorum) {
 585             break;
 586        }
 587     }
 588
 589     /* Every successful read agrees */
 590     if (quorum) {
 591         quorum_copy_qiov(acb->qiov, &acb->qcrs[i].qiov);
 592         return false;
 593     }
 594
 595     /* compute hashes for each successful read, also store indexes */
 596     for (i = 0; i < s->num_children; i++) {
 597         if (acb->qcrs[i].ret) {
 598             continue;
 599         }
 600         ret = quorum_compute_hash(acb, i, &hash);
 601         /* if ever the hash computation failed */
 602         if (ret < 0) {
 603             acb->vote_ret = ret;
 604             goto free_exit;
 605         }
 606         quorum_count_vote(&acb->votes, &hash, i);
 607     }
 608
 609     /* vote to select the most represented version */
 610     winner = quorum_get_vote_winner(&acb->votes);
 611
 612     /* if the winner count is smaller than threshold the read fails */
 613     if (winner->vote_count < s->threshold) {
 614         quorum_report_failure(acb);
 615         acb->vote_ret = -EIO;
 616         goto free_exit;
 617     }
 618
 619     /* we have a winner: copy it */
 620     quorum_copy_qiov(acb->qiov, &acb->qcrs[winner->index].qiov);
 621
 622     /* some versions are bad print them */
 623     quorum_report_bad_versions(s, acb, &winner->value);
 624
 625     /* corruption correction is enabled */
 626     if (s->rewrite_corrupted) {
 627         rewrite = quorum_rewrite_bad_versions(s, acb, &winner->value);
 628     }
 629
 630 free_exit:
 631     /* free lists */
 632     quorum_free_vote_list(&acb->votes);
 633     return rewrite;
 634 }
 635
 636 static BlockAIOCB *read_quorum_children(QuorumAIOCB *acb)
 637 {
 638     BDRVQuorumState *s = acb->common.bs->opaque;
 639     int i;
 640
 641     for (i = 0; i < s->num_children; i++) {
 642         acb->qcrs[i].buf = qemu_blockalign(s->bs[i], acb->qiov->size);
 643         qemu_iovec_init(&acb->qcrs[i].qiov, acb->qiov->niov);
 644         qemu_iovec_clone(&acb->qcrs[i].qiov, acb->qiov, acb->qcrs[i].buf);
 645     }
 646
 647     for (i = 0; i < s->num_children; i++) {
 648         bdrv_aio_readv(s->bs[i], acb->sector_num, &acb->qcrs[i].qiov,
 649                        acb->nb_sectors, quorum_aio_cb, &acb->qcrs[i]);
 650     }
 651
 652     return &acb->common;
 653 }
 654
 655 static BlockAIOCB *read_fifo_child(QuorumAIOCB *acb)
 656 {
 657     BDRVQuorumState *s = acb->common.bs->opaque;
 658
 659     acb->qcrs[acb->child_iter].buf = qemu_blockalign(s->bs[acb->child_iter],
 660                                                      acb->qiov->size);
 661     qemu_iovec_init(&acb->qcrs[acb->child_iter].qiov, acb->qiov->niov);
 662     qemu_iovec_clone(&acb->qcrs[acb->child_iter].qiov, acb->qiov,
 663                      acb->qcrs[acb->child_iter].buf);
 664     bdrv_aio_readv(s->bs[acb->child_iter], acb->sector_num,
 665                    &acb->qcrs[acb->child_iter].qiov, acb->nb_sectors,
 666                    quorum_aio_cb, &acb->qcrs[acb->child_iter]);
 667
 668     return &acb->common;
 669 }
 670
 671 static BlockAIOCB *quorum_aio_readv(BlockDriverState *bs,
 672                                     int64_t sector_num,
 673                                     QEMUIOVector *qiov,
 674                                     int nb_sectors,
 675                                     BlockCompletionFunc *cb,
 676                                     void *opaque)
 677 {
 678     BDRVQuorumState *s = bs->opaque;
 679     QuorumAIOCB *acb = quorum_aio_get(s, bs, qiov, sector_num,
 680                                       nb_sectors, cb, opaque);
 681     acb->is_read = true;
 682
 683     if (s->read_pattern == QUORUM_READ_PATTERN_QUORUM) {
 684         acb->child_iter = s->num_children - 1;
 685         return read_quorum_children(acb);
 686     }
 687
 688     acb->child_iter = 0;
 689     return read_fifo_child(acb);
 690 }
 691
 692 static BlockAIOCB *quorum_aio_writev(BlockDriverState *bs,
 693                                      int64_t sector_num,
 694                                      QEMUIOVector *qiov,
 695                                      int nb_sectors,
 696                                      BlockCompletionFunc *cb,
 697                                      void *opaque)
 698 {
 699     BDRVQuorumState *s = bs->opaque;
 700     QuorumAIOCB *acb = quorum_aio_get(s, bs, qiov, sector_num, nb_sectors,
 701                                       cb, opaque);
 702     int i;
 703
 704     for (i = 0; i < s->num_children; i++) {
 705         acb->qcrs[i].aiocb = bdrv_aio_writev(s->bs[i], sector_num, qiov,
 706                                              nb_sectors, &quorum_aio_cb,
 707                                              &acb->qcrs[i]);
 708     }
 709
 710     return &acb->common;
 711 }
 712
 713 static int64_t quorum_getlength(BlockDriverState *bs)
 714 {
 715     BDRVQuorumState *s = bs->opaque;
 716     int64_t result;
 717     int i;
 718
 719     /* check that all file have the same length */
 720     result = bdrv_getlength(s->bs[0]);
 721     if (result < 0) {
 722         return result;
 723     }
 724     for (i = 1; i < s->num_children; i++) {
 725         int64_t value = bdrv_getlength(s->bs[i]);
 726         if (value < 0) {
 727             return value;
 728         }
 729         if (value != result) {
 730             return -EIO;
 731         }
 732     }
 733
 734     return result;
 735 }
 736
 737 static void quorum_invalidate_cache(BlockDriverState *bs, Error **errp)
 738 {
 739     BDRVQuorumState *s = bs->opaque;
 740     Error *local_err = NULL;
 741     int i;
 742
 743     for (i = 0; i < s->num_children; i++) {
 744         bdrv_invalidate_cache(s->bs[i], &local_err);
 745         if (local_err) {
 746             error_propagate(errp, local_err);
 747             return;
 748         }
 749     }
 750 }
 751
 752 static coroutine_fn int quorum_co_flush(BlockDriverState *bs)
 753 {
 754     BDRVQuorumState *s = bs->opaque;
 755     QuorumVoteVersion *winner = NULL;
 756     QuorumVotes error_votes;
 757     QuorumVoteValue result_value;
 758     int i;
 759     int result = 0;
 760
 761     QLIST_INIT(&error_votes.vote_list);
 762     error_votes.compare = quorum_64bits_compare;
 763
 764     for (i = 0; i < s->num_children; i++) {
 765         result = bdrv_co_flush(s->bs[i]);
 766         result_value.l = result;
 767         quorum_count_vote(&error_votes, &result_value, i);
 768     }
 769
 770     winner = quorum_get_vote_winner(&error_votes);
 771     result = winner->value.l;
 772
 773     quorum_free_vote_list(&error_votes);
 774
 775     return result;
 776 }
 777
 778 static bool quorum_recurse_is_first_non_filter(BlockDriverState *bs,
 779                                                BlockDriverState *candidate)
 780 {
 781     BDRVQuorumState *s = bs->opaque;
 782     int i;
 783
 784     for (i = 0; i < s->num_children; i++) {
 785         bool perm = bdrv_recurse_is_first_non_filter(s->bs[i],
 786                                                      candidate);
 787         if (perm) {
 788             return true;
 789         }
 790     }
 791
 792     return false;
 793 }
 794
 795 static int quorum_valid_threshold(int threshold, int num_children, Error **errp)
 796 {
 797
 798     if (threshold < 1) {
 799         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 800                    "vote-threshold", "value >= 1");
 801         return -ERANGE;
 802     }
 803
 804     if (threshold > num_children) {
 805         error_setg(errp, "threshold may not exceed children count");
 806         return -ERANGE;
 807     }
 808
 809     return 0;
 810 }
 811
 812 static QemuOptsList quorum_runtime_opts = {
 813     .name = "quorum",
 814     .head = QTAILQ_HEAD_INITIALIZER(quorum_runtime_opts.head),
 815     .desc = {
 816         {
 817             .name = QUORUM_OPT_VOTE_THRESHOLD,
 818             .type = QEMU_OPT_NUMBER,
 819             .help = "The number of vote needed for reaching quorum",
 820         },
 821         {
 822             .name = QUORUM_OPT_BLKVERIFY,
 823             .type = QEMU_OPT_BOOL,
 824             .help = "Trigger block verify mode if set",
 825         },
 826         {
 827             .name = QUORUM_OPT_REWRITE,
 828             .type = QEMU_OPT_BOOL,
 829             .help = "Rewrite corrupted block on read quorum",
 830         },
 831         {
 832             .name = QUORUM_OPT_READ_PATTERN,
 833             .type = QEMU_OPT_STRING,
 834             .help = "Allowed pattern: quorum, fifo. Quorum is default",
 835         },
 836         { /* end of list */ }
 837     },
 838 };
 839
 840 static int parse_read_pattern(const char *opt)
 841 {
 842     int i;
 843
 844     if (!opt) {
 845         /* Set quorum as default */
 846         return QUORUM_READ_PATTERN_QUORUM;
 847     }
 848
 849     for (i = 0; i < QUORUM_READ_PATTERN_MAX; i++) {
 850         if (!strcmp(opt, QuorumReadPattern_lookup[i])) {
 851             return i;
 852         }
 853     }
 854
 855     return -EINVAL;
 856 }
 857
 858 static int quorum_open(BlockDriverState *bs, QDict *options, int flags,
 859                        Error **errp)
 860 {
 861     BDRVQuorumState *s = bs->opaque;
 862     Error *local_err = NULL;
 863     QemuOpts *opts = NULL;
 864     bool *opened;
 865     int i;
 866     int ret = 0;
 867
 868     if (!qcrypto_hash_supports(QCRYPTO_HASH_ALG_SHA256)) {
 869         error_setg(errp,
 870                    "SHA256 hash support is required for quorum device");
 871         return -EINVAL;
 872     }
 873
 874     qdict_flatten(options);
 875
 876     /* count how many different children are present */
 877     s->num_children = qdict_array_entries(options, "children.");
 878     if (s->num_children < 0) {
 879         error_setg(&local_err, "Option children is not a valid array");
 880         ret = -EINVAL;
 881         goto exit;
 882     }
 883     if (s->num_children < 2) {
 884         error_setg(&local_err,
 885                    "Number of provided children must be greater than 1");
 886         ret = -EINVAL;
 887         goto exit;
 888     }
 889
 890     opts = qemu_opts_create(&quorum_runtime_opts, NULL, 0, &error_abort);
 891     qemu_opts_absorb_qdict(opts, options, &local_err);
 892     if (local_err) {
 893         ret = -EINVAL;
 894         goto exit;
 895     }
 896
 897     s->threshold = qemu_opt_get_number(opts, QUORUM_OPT_VOTE_THRESHOLD, 0);
 898     ret = parse_read_pattern(qemu_opt_get(opts, QUORUM_OPT_READ_PATTERN));
 899     if (ret < 0) {
 900         error_setg(&local_err, "Please set read-pattern as fifo or quorum");
 901         goto exit;
 902     }
 903     s->read_pattern = ret;
 904
 905     if (s->read_pattern == QUORUM_READ_PATTERN_QUORUM) {
 906         /* and validate it against s->num_children */
 907         ret = quorum_valid_threshold(s->threshold, s->num_children, &local_err);
 908         if (ret < 0) {
 909             goto exit;
 910         }
 911
 912         /* is the driver in blkverify mode */
 913         if (qemu_opt_get_bool(opts, QUORUM_OPT_BLKVERIFY, false) &&
 914             s->num_children == 2 && s->threshold == 2) {
 915             s->is_blkverify = true;
 916         } else if (qemu_opt_get_bool(opts, QUORUM_OPT_BLKVERIFY, false)) {
 917             fprintf(stderr, "blkverify mode is set by setting blkverify=on "
 918                     "and using two files with vote_threshold=2\n");
 919         }
 920
 921         s->rewrite_corrupted = qemu_opt_get_bool(opts, QUORUM_OPT_REWRITE,
 922                                                  false);
 923         if (s->rewrite_corrupted && s->is_blkverify) {
 924             error_setg(&local_err,
 925                        "rewrite-corrupted=on cannot be used with blkverify=on");
 926             ret = -EINVAL;
 927             goto exit;
 928         }
 929     }
 930
 931     /* allocate the children BlockDriverState array */
 932     s->bs = g_new0(BlockDriverState *, s->num_children);
 933     opened = g_new0(bool, s->num_children);
 934
 935     for (i = 0; i < s->num_children; i++) {
 936         char indexstr[32];
 937         ret = snprintf(indexstr, 32, "children.%d", i);
 938         assert(ret < 32);
 939
 940         ret = bdrv_open_image(&s->bs[i], NULL, options, indexstr, bs,
 941                               &child_format, false, &local_err);
 942         if (ret < 0) {
 943             goto close_exit;
 944         }
 945
 946         opened[i] = true;
 947     }
 948
 949     g_free(opened);
 950     goto exit;
 951
 952 close_exit:
 953     /* cleanup on error */
 954     for (i = 0; i < s->num_children; i++) {
 955         if (!opened[i]) {
 956             continue;
 957         }
 958         bdrv_unref(s->bs[i]);
 959     }
 960     g_free(s->bs);
 961     g_free(opened);
 962 exit:
 963     qemu_opts_del(opts);
 964     /* propagate error */
 965     if (local_err) {
 966         error_propagate(errp, local_err);
 967     }
 968     return ret;
 969 }
 970
 971 static void quorum_close(BlockDriverState *bs)
 972 {
 973     BDRVQuorumState *s = bs->opaque;
 974     int i;
 975
 976     for (i = 0; i < s->num_children; i++) {
 977         bdrv_unref(s->bs[i]);
 978     }
 979
 980     g_free(s->bs);
 981 }
 982
 983 static void quorum_detach_aio_context(BlockDriverState *bs)
 984 {
 985     BDRVQuorumState *s = bs->opaque;
 986     int i;
 987
 988     for (i = 0; i < s->num_children; i++) {
 989         bdrv_detach_aio_context(s->bs[i]);
 990     }
 991 }
 992
 993 static void quorum_attach_aio_context(BlockDriverState *bs,
 994                                       AioContext *new_context)
 995 {
 996     BDRVQuorumState *s = bs->opaque;
 997     int i;
 998
 999     for (i = 0; i < s->num_children; i++) {
1000         bdrv_attach_aio_context(s->bs[i], new_context);
1001     }
1002 }
1003
1004 static void quorum_refresh_filename(BlockDriverState *bs)
1005 {
1006     BDRVQuorumState *s = bs->opaque;
1007     QDict *opts;
1008     QList *children;
1009     int i;
1010
1011     for (i = 0; i < s->num_children; i++) {
1012         bdrv_refresh_filename(s->bs[i]);
1013         if (!s->bs[i]->full_open_options) {
1014             return;
1015         }
1016     }
1017
1018     children = qlist_new();
1019     for (i = 0; i < s->num_children; i++) {
1020         QINCREF(s->bs[i]->full_open_options);
1021         qlist_append_obj(children, QOBJECT(s->bs[i]->full_open_options));
1022     }
1023
1024     opts = qdict_new();
1025     qdict_put_obj(opts, "driver", QOBJECT(qstring_from_str("quorum")));
1026     qdict_put_obj(opts, QUORUM_OPT_VOTE_THRESHOLD,
1027                   QOBJECT(qint_from_int(s->threshold)));
1028     qdict_put_obj(opts, QUORUM_OPT_BLKVERIFY,
1029                   QOBJECT(qbool_from_bool(s->is_blkverify)));
1030     qdict_put_obj(opts, QUORUM_OPT_REWRITE,
1031                   QOBJECT(qbool_from_bool(s->rewrite_corrupted)));
1032     qdict_put_obj(opts, "children", QOBJECT(children));
1033
1034     bs->full_open_options = opts;
1035 }
1036
1037 static BlockDriver bdrv_quorum = {
1038     .format_name                        = "quorum",
1039     .protocol_name                      = "quorum",
1040
1041     .instance_size                      = sizeof(BDRVQuorumState),
1042
1043     .bdrv_file_open                     = quorum_open,
1044     .bdrv_close                         = quorum_close,
1045     .bdrv_refresh_filename              = quorum_refresh_filename,
1046
1047     .bdrv_co_flush_to_disk              = quorum_co_flush,
1048
1049     .bdrv_getlength                     = quorum_getlength,
1050
1051     .bdrv_aio_readv                     = quorum_aio_readv,
1052     .bdrv_aio_writev                    = quorum_aio_writev,
1053     .bdrv_invalidate_cache              = quorum_invalidate_cache,
1054
1055     .bdrv_detach_aio_context            = quorum_detach_aio_context,
1056     .bdrv_attach_aio_context            = quorum_attach_aio_context,
1057
1058     .is_filter                          = true,
1059     .bdrv_recurse_is_first_non_filter   = quorum_recurse_is_first_non_filter,
1060 };
1061
1062 static void bdrv_quorum_init(void)
1063 {
1064     bdrv_register(&bdrv_quorum);
1065 }
1066
1067 block_init(bdrv_quorum_init);