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 "qemu/osdep.h"
  17 #include "block/block_int.h"
  18 #include "qapi/qmp/qbool.h"
  19 #include "qapi/qmp/qdict.h"
  20 #include "qapi/qmp/qerror.h"
  21 #include "qapi/qmp/qint.h"
  22 #include "qapi/qmp/qjson.h"
  23 #include "qapi/qmp/qlist.h"
  24 #include "qapi/qmp/qstring.h"
  25 #include "qapi-event.h"
  26 #include "crypto/hash.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     uint8_t 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     BdrvChild **children;  /* 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->children[item->index]->bs->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->children[item->index]->bs, acb->sector_num,
 374                             acb->qiov, acb->nb_sectors, quorum_rewrite_aio_cb,
 375                             acb);
 376         }
 377     }
 378
 379     /* return true if any rewrite is done else false */
 380     return count;
 381 }
 382
 383 static void quorum_count_vote(QuorumVotes *votes,
 384                               QuorumVoteValue *value,
 385                               int index)
 386 {
 387     QuorumVoteVersion *v = NULL, *version = NULL;
 388     QuorumVoteItem *item;
 389
 390     /* look if we have something with this hash */
 391     QLIST_FOREACH(v, &votes->vote_list, next) {
 392         if (votes->compare(&v->value, value)) {
 393             version = v;
 394             break;
 395         }
 396     }
 397
 398     /* It's a version not yet in the list add it */
 399     if (!version) {
 400         version = g_new0(QuorumVoteVersion, 1);
 401         QLIST_INIT(&version->items);
 402         memcpy(&version->value, value, sizeof(version->value));
 403         version->index = index;
 404         version->vote_count = 0;
 405         QLIST_INSERT_HEAD(&votes->vote_list, version, next);
 406     }
 407
 408     version->vote_count++;
 409
 410     item = g_new0(QuorumVoteItem, 1);
 411     item->index = index;
 412     QLIST_INSERT_HEAD(&version->items, item, next);
 413 }
 414
 415 static void quorum_free_vote_list(QuorumVotes *votes)
 416 {
 417     QuorumVoteVersion *version, *next_version;
 418     QuorumVoteItem *item, *next_item;
 419
 420     QLIST_FOREACH_SAFE(version, &votes->vote_list, next, next_version) {
 421         QLIST_REMOVE(version, next);
 422         QLIST_FOREACH_SAFE(item, &version->items, next, next_item) {
 423             QLIST_REMOVE(item, next);
 424             g_free(item);
 425         }
 426         g_free(version);
 427     }
 428 }
 429
 430 static int quorum_compute_hash(QuorumAIOCB *acb, int i, QuorumVoteValue *hash)
 431 {
 432     QEMUIOVector *qiov = &acb->qcrs[i].qiov;
 433     size_t len = sizeof(hash->h);
 434     uint8_t *data = hash->h;
 435
 436     /* XXX - would be nice if we could pass in the Error **
 437      * and propagate that back, but this quorum code is
 438      * restricted to just errno values currently */
 439     if (qcrypto_hash_bytesv(QCRYPTO_HASH_ALG_SHA256,
 440                             qiov->iov, qiov->niov,
 441                             &data, &len,
 442                             NULL) < 0) {
 443         return -EINVAL;
 444     }
 445
 446     return 0;
 447 }
 448
 449 static QuorumVoteVersion *quorum_get_vote_winner(QuorumVotes *votes)
 450 {
 451     int max = 0;
 452     QuorumVoteVersion *candidate, *winner = NULL;
 453
 454     QLIST_FOREACH(candidate, &votes->vote_list, next) {
 455         if (candidate->vote_count > max) {
 456             max = candidate->vote_count;
 457             winner = candidate;
 458         }
 459     }
 460
 461     return winner;
 462 }
 463
 464 /* qemu_iovec_compare is handy for blkverify mode because it returns the first
 465  * differing byte location. Yet it is handcoded to compare vectors one byte
 466  * after another so it does not benefit from the libc SIMD optimizations.
 467  * quorum_iovec_compare is written for speed and should be used in the non
 468  * blkverify mode of quorum.
 469  */
 470 static bool quorum_iovec_compare(QEMUIOVector *a, QEMUIOVector *b)
 471 {
 472     int i;
 473     int result;
 474
 475     assert(a->niov == b->niov);
 476     for (i = 0; i < a->niov; i++) {
 477         assert(a->iov[i].iov_len == b->iov[i].iov_len);
 478         result = memcmp(a->iov[i].iov_base,
 479                         b->iov[i].iov_base,
 480                         a->iov[i].iov_len);
 481         if (result) {
 482             return false;
 483         }
 484     }
 485
 486     return true;
 487 }
 488
 489 static void GCC_FMT_ATTR(2, 3) quorum_err(QuorumAIOCB *acb,
 490                                           const char *fmt, ...)
 491 {
 492     va_list ap;
 493
 494     va_start(ap, fmt);
 495     fprintf(stderr, "quorum: sector_num=%" PRId64 " nb_sectors=%d ",
 496             acb->sector_num, acb->nb_sectors);
 497     vfprintf(stderr, fmt, ap);
 498     fprintf(stderr, "\n");
 499     va_end(ap);
 500     exit(1);
 501 }
 502
 503 static bool quorum_compare(QuorumAIOCB *acb,
 504                            QEMUIOVector *a,
 505                            QEMUIOVector *b)
 506 {
 507     BDRVQuorumState *s = acb->common.bs->opaque;
 508     ssize_t offset;
 509
 510     /* This driver will replace blkverify in this particular case */
 511     if (s->is_blkverify) {
 512         offset = qemu_iovec_compare(a, b);
 513         if (offset != -1) {
 514             quorum_err(acb, "contents mismatch in sector %" PRId64,
 515                        acb->sector_num +
 516                        (uint64_t)(offset / BDRV_SECTOR_SIZE));
 517         }
 518         return true;
 519     }
 520
 521     return quorum_iovec_compare(a, b);
 522 }
 523
 524 /* Do a vote to get the error code */
 525 static int quorum_vote_error(QuorumAIOCB *acb)
 526 {
 527     BDRVQuorumState *s = acb->common.bs->opaque;
 528     QuorumVoteVersion *winner = NULL;
 529     QuorumVotes error_votes;
 530     QuorumVoteValue result_value;
 531     int i, ret = 0;
 532     bool error = false;
 533
 534     QLIST_INIT(&error_votes.vote_list);
 535     error_votes.compare = quorum_64bits_compare;
 536
 537     for (i = 0; i < s->num_children; i++) {
 538         ret = acb->qcrs[i].ret;
 539         if (ret) {
 540             error = true;
 541             result_value.l = ret;
 542             quorum_count_vote(&error_votes, &result_value, i);
 543         }
 544     }
 545
 546     if (error) {
 547         winner = quorum_get_vote_winner(&error_votes);
 548         ret = winner->value.l;
 549     }
 550
 551     quorum_free_vote_list(&error_votes);
 552
 553     return ret;
 554 }
 555
 556 static bool quorum_vote(QuorumAIOCB *acb)
 557 {
 558     bool quorum = true;
 559     bool rewrite = false;
 560     int i, j, ret;
 561     QuorumVoteValue hash;
 562     BDRVQuorumState *s = acb->common.bs->opaque;
 563     QuorumVoteVersion *winner;
 564
 565     if (quorum_has_too_much_io_failed(acb)) {
 566         return false;
 567     }
 568
 569     /* get the index of the first successful read */
 570     for (i = 0; i < s->num_children; i++) {
 571         if (!acb->qcrs[i].ret) {
 572             break;
 573         }
 574     }
 575
 576     assert(i < s->num_children);
 577
 578     /* compare this read with all other successful reads stopping at quorum
 579      * failure
 580      */
 581     for (j = i + 1; j < s->num_children; j++) {
 582         if (acb->qcrs[j].ret) {
 583             continue;
 584         }
 585         quorum = quorum_compare(acb, &acb->qcrs[i].qiov, &acb->qcrs[j].qiov);
 586         if (!quorum) {
 587             break;
 588        }
 589     }
 590
 591     /* Every successful read agrees */
 592     if (quorum) {
 593         quorum_copy_qiov(acb->qiov, &acb->qcrs[i].qiov);
 594         return false;
 595     }
 596
 597     /* compute hashes for each successful read, also store indexes */
 598     for (i = 0; i < s->num_children; i++) {
 599         if (acb->qcrs[i].ret) {
 600             continue;
 601         }
 602         ret = quorum_compute_hash(acb, i, &hash);
 603         /* if ever the hash computation failed */
 604         if (ret < 0) {
 605             acb->vote_ret = ret;
 606             goto free_exit;
 607         }
 608         quorum_count_vote(&acb->votes, &hash, i);
 609     }
 610
 611     /* vote to select the most represented version */
 612     winner = quorum_get_vote_winner(&acb->votes);
 613
 614     /* if the winner count is smaller than threshold the read fails */
 615     if (winner->vote_count < s->threshold) {
 616         quorum_report_failure(acb);
 617         acb->vote_ret = -EIO;
 618         goto free_exit;
 619     }
 620
 621     /* we have a winner: copy it */
 622     quorum_copy_qiov(acb->qiov, &acb->qcrs[winner->index].qiov);
 623
 624     /* some versions are bad print them */
 625     quorum_report_bad_versions(s, acb, &winner->value);
 626
 627     /* corruption correction is enabled */
 628     if (s->rewrite_corrupted) {
 629         rewrite = quorum_rewrite_bad_versions(s, acb, &winner->value);
 630     }
 631
 632 free_exit:
 633     /* free lists */
 634     quorum_free_vote_list(&acb->votes);
 635     return rewrite;
 636 }
 637
 638 static BlockAIOCB *read_quorum_children(QuorumAIOCB *acb)
 639 {
 640     BDRVQuorumState *s = acb->common.bs->opaque;
 641     int i;
 642
 643     for (i = 0; i < s->num_children; i++) {
 644         acb->qcrs[i].buf = qemu_blockalign(s->children[i]->bs, acb->qiov->size);
 645         qemu_iovec_init(&acb->qcrs[i].qiov, acb->qiov->niov);
 646         qemu_iovec_clone(&acb->qcrs[i].qiov, acb->qiov, acb->qcrs[i].buf);
 647     }
 648
 649     for (i = 0; i < s->num_children; i++) {
 650         bdrv_aio_readv(s->children[i]->bs, acb->sector_num, &acb->qcrs[i].qiov,
 651                        acb->nb_sectors, quorum_aio_cb, &acb->qcrs[i]);
 652     }
 653
 654     return &acb->common;
 655 }
 656
 657 static BlockAIOCB *read_fifo_child(QuorumAIOCB *acb)
 658 {
 659     BDRVQuorumState *s = acb->common.bs->opaque;
 660
 661     acb->qcrs[acb->child_iter].buf =
 662         qemu_blockalign(s->children[acb->child_iter]->bs, acb->qiov->size);
 663     qemu_iovec_init(&acb->qcrs[acb->child_iter].qiov, acb->qiov->niov);
 664     qemu_iovec_clone(&acb->qcrs[acb->child_iter].qiov, acb->qiov,
 665                      acb->qcrs[acb->child_iter].buf);
 666     bdrv_aio_readv(s->children[acb->child_iter]->bs, acb->sector_num,
 667                    &acb->qcrs[acb->child_iter].qiov, acb->nb_sectors,
 668                    quorum_aio_cb, &acb->qcrs[acb->child_iter]);
 669
 670     return &acb->common;
 671 }
 672
 673 static BlockAIOCB *quorum_aio_readv(BlockDriverState *bs,
 674                                     int64_t sector_num,
 675                                     QEMUIOVector *qiov,
 676                                     int nb_sectors,
 677                                     BlockCompletionFunc *cb,
 678                                     void *opaque)
 679 {
 680     BDRVQuorumState *s = bs->opaque;
 681     QuorumAIOCB *acb = quorum_aio_get(s, bs, qiov, sector_num,
 682                                       nb_sectors, cb, opaque);
 683     acb->is_read = true;
 684
 685     if (s->read_pattern == QUORUM_READ_PATTERN_QUORUM) {
 686         acb->child_iter = s->num_children - 1;
 687         return read_quorum_children(acb);
 688     }
 689
 690     acb->child_iter = 0;
 691     return read_fifo_child(acb);
 692 }
 693
 694 static BlockAIOCB *quorum_aio_writev(BlockDriverState *bs,
 695                                      int64_t sector_num,
 696                                      QEMUIOVector *qiov,
 697                                      int nb_sectors,
 698                                      BlockCompletionFunc *cb,
 699                                      void *opaque)
 700 {
 701     BDRVQuorumState *s = bs->opaque;
 702     QuorumAIOCB *acb = quorum_aio_get(s, bs, qiov, sector_num, nb_sectors,
 703                                       cb, opaque);
 704     int i;
 705
 706     for (i = 0; i < s->num_children; i++) {
 707         acb->qcrs[i].aiocb = bdrv_aio_writev(s->children[i]->bs, sector_num,
 708                                              qiov, nb_sectors, &quorum_aio_cb,
 709                                              &acb->qcrs[i]);
 710     }
 711
 712     return &acb->common;
 713 }
 714
 715 static int64_t quorum_getlength(BlockDriverState *bs)
 716 {
 717     BDRVQuorumState *s = bs->opaque;
 718     int64_t result;
 719     int i;
 720
 721     /* check that all file have the same length */
 722     result = bdrv_getlength(s->children[0]->bs);
 723     if (result < 0) {
 724         return result;
 725     }
 726     for (i = 1; i < s->num_children; i++) {
 727         int64_t value = bdrv_getlength(s->children[i]->bs);
 728         if (value < 0) {
 729             return value;
 730         }
 731         if (value != result) {
 732             return -EIO;
 733         }
 734     }
 735
 736     return result;
 737 }
 738
 739 static void quorum_invalidate_cache(BlockDriverState *bs, Error **errp)
 740 {
 741     BDRVQuorumState *s = bs->opaque;
 742     Error *local_err = NULL;
 743     int i;
 744
 745     for (i = 0; i < s->num_children; i++) {
 746         bdrv_invalidate_cache(s->children[i]->bs, &local_err);
 747         if (local_err) {
 748             error_propagate(errp, local_err);
 749             return;
 750         }
 751     }
 752 }
 753
 754 static coroutine_fn int quorum_co_flush(BlockDriverState *bs)
 755 {
 756     BDRVQuorumState *s = bs->opaque;
 757     QuorumVoteVersion *winner = NULL;
 758     QuorumVotes error_votes;
 759     QuorumVoteValue result_value;
 760     int i;
 761     int result = 0;
 762
 763     QLIST_INIT(&error_votes.vote_list);
 764     error_votes.compare = quorum_64bits_compare;
 765
 766     for (i = 0; i < s->num_children; i++) {
 767         result = bdrv_co_flush(s->children[i]->bs);
 768         result_value.l = result;
 769         quorum_count_vote(&error_votes, &result_value, i);
 770     }
 771
 772     winner = quorum_get_vote_winner(&error_votes);
 773     result = winner->value.l;
 774
 775     quorum_free_vote_list(&error_votes);
 776
 777     return result;
 778 }
 779
 780 static bool quorum_recurse_is_first_non_filter(BlockDriverState *bs,
 781                                                BlockDriverState *candidate)
 782 {
 783     BDRVQuorumState *s = bs->opaque;
 784     int i;
 785
 786     for (i = 0; i < s->num_children; i++) {
 787         bool perm = bdrv_recurse_is_first_non_filter(s->children[i]->bs,
 788                                                      candidate);
 789         if (perm) {
 790             return true;
 791         }
 792     }
 793
 794     return false;
 795 }
 796
 797 static int quorum_valid_threshold(int threshold, int num_children, Error **errp)
 798 {
 799
 800     if (threshold < 1) {
 801         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
 802                    "vote-threshold", "value >= 1");
 803         return -ERANGE;
 804     }
 805
 806     if (threshold > num_children) {
 807         error_setg(errp, "threshold may not exceed children count");
 808         return -ERANGE;
 809     }
 810
 811     return 0;
 812 }
 813
 814 static QemuOptsList quorum_runtime_opts = {
 815     .name = "quorum",
 816     .head = QTAILQ_HEAD_INITIALIZER(quorum_runtime_opts.head),
 817     .desc = {
 818         {
 819             .name = QUORUM_OPT_VOTE_THRESHOLD,
 820             .type = QEMU_OPT_NUMBER,
 821             .help = "The number of vote needed for reaching quorum",
 822         },
 823         {
 824             .name = QUORUM_OPT_BLKVERIFY,
 825             .type = QEMU_OPT_BOOL,
 826             .help = "Trigger block verify mode if set",
 827         },
 828         {
 829             .name = QUORUM_OPT_REWRITE,
 830             .type = QEMU_OPT_BOOL,
 831             .help = "Rewrite corrupted block on read quorum",
 832         },
 833         {
 834             .name = QUORUM_OPT_READ_PATTERN,
 835             .type = QEMU_OPT_STRING,
 836             .help = "Allowed pattern: quorum, fifo. Quorum is default",
 837         },
 838         { /* end of list */ }
 839     },
 840 };
 841
 842 static int parse_read_pattern(const char *opt)
 843 {
 844     int i;
 845
 846     if (!opt) {
 847         /* Set quorum as default */
 848         return QUORUM_READ_PATTERN_QUORUM;
 849     }
 850
 851     for (i = 0; i < QUORUM_READ_PATTERN__MAX; i++) {
 852         if (!strcmp(opt, QuorumReadPattern_lookup[i])) {
 853             return i;
 854         }
 855     }
 856
 857     return -EINVAL;
 858 }
 859
 860 static int quorum_open(BlockDriverState *bs, QDict *options, int flags,
 861                        Error **errp)
 862 {
 863     BDRVQuorumState *s = bs->opaque;
 864     Error *local_err = NULL;
 865     QemuOpts *opts = NULL;
 866     bool *opened;
 867     int i;
 868     int ret = 0;
 869
 870     qdict_flatten(options);
 871
 872     /* count how many different children are present */
 873     s->num_children = qdict_array_entries(options, "children.");
 874     if (s->num_children < 0) {
 875         error_setg(&local_err, "Option children is not a valid array");
 876         ret = -EINVAL;
 877         goto exit;
 878     }
 879     if (s->num_children < 2) {
 880         error_setg(&local_err,
 881                    "Number of provided children must be greater than 1");
 882         ret = -EINVAL;
 883         goto exit;
 884     }
 885
 886     opts = qemu_opts_create(&quorum_runtime_opts, NULL, 0, &error_abort);
 887     qemu_opts_absorb_qdict(opts, options, &local_err);
 888     if (local_err) {
 889         ret = -EINVAL;
 890         goto exit;
 891     }
 892
 893     s->threshold = qemu_opt_get_number(opts, QUORUM_OPT_VOTE_THRESHOLD, 0);
 894     /* and validate it against s->num_children */
 895     ret = quorum_valid_threshold(s->threshold, s->num_children, &local_err);
 896     if (ret < 0) {
 897         goto exit;
 898     }
 899
 900     ret = parse_read_pattern(qemu_opt_get(opts, QUORUM_OPT_READ_PATTERN));
 901     if (ret < 0) {
 902         error_setg(&local_err, "Please set read-pattern as fifo or quorum");
 903         goto exit;
 904     }
 905     s->read_pattern = ret;
 906
 907     if (s->read_pattern == QUORUM_READ_PATTERN_QUORUM) {
 908         /* is the driver in blkverify mode */
 909         if (qemu_opt_get_bool(opts, QUORUM_OPT_BLKVERIFY, false) &&
 910             s->num_children == 2 && s->threshold == 2) {
 911             s->is_blkverify = true;
 912         } else if (qemu_opt_get_bool(opts, QUORUM_OPT_BLKVERIFY, false)) {
 913             fprintf(stderr, "blkverify mode is set by setting blkverify=on "
 914                     "and using two files with vote_threshold=2\n");
 915         }
 916
 917         s->rewrite_corrupted = qemu_opt_get_bool(opts, QUORUM_OPT_REWRITE,
 918                                                  false);
 919         if (s->rewrite_corrupted && s->is_blkverify) {
 920             error_setg(&local_err,
 921                        "rewrite-corrupted=on cannot be used with blkverify=on");
 922             ret = -EINVAL;
 923             goto exit;
 924         }
 925     }
 926
 927     /* allocate the children array */
 928     s->children = g_new0(BdrvChild *, s->num_children);
 929     opened = g_new0(bool, s->num_children);
 930
 931     for (i = 0; i < s->num_children; i++) {
 932         char indexstr[32];
 933         ret = snprintf(indexstr, 32, "children.%d", i);
 934         assert(ret < 32);
 935
 936         s->children[i] = bdrv_open_child(NULL, options, indexstr, bs,
 937                                          &child_format, false, &local_err);
 938         if (local_err) {
 939             ret = -EINVAL;
 940             goto close_exit;
 941         }
 942
 943         opened[i] = true;
 944     }
 945
 946     g_free(opened);
 947     goto exit;
 948
 949 close_exit:
 950     /* cleanup on error */
 951     for (i = 0; i < s->num_children; i++) {
 952         if (!opened[i]) {
 953             continue;
 954         }
 955         bdrv_unref_child(bs, s->children[i]);
 956     }
 957     g_free(s->children);
 958     g_free(opened);
 959 exit:
 960     qemu_opts_del(opts);
 961     /* propagate error */
 962     if (local_err) {
 963         error_propagate(errp, local_err);
 964     }
 965     return ret;
 966 }
 967
 968 static void quorum_close(BlockDriverState *bs)
 969 {
 970     BDRVQuorumState *s = bs->opaque;
 971     int i;
 972
 973     for (i = 0; i < s->num_children; i++) {
 974         bdrv_unref_child(bs, s->children[i]);
 975     }
 976
 977     g_free(s->children);
 978 }
 979
 980 static void quorum_detach_aio_context(BlockDriverState *bs)
 981 {
 982     BDRVQuorumState *s = bs->opaque;
 983     int i;
 984
 985     for (i = 0; i < s->num_children; i++) {
 986         bdrv_detach_aio_context(s->children[i]->bs);
 987     }
 988 }
 989
 990 static void quorum_attach_aio_context(BlockDriverState *bs,
 991                                       AioContext *new_context)
 992 {
 993     BDRVQuorumState *s = bs->opaque;
 994     int i;
 995
 996     for (i = 0; i < s->num_children; i++) {
 997         bdrv_attach_aio_context(s->children[i]->bs, new_context);
 998     }
 999 }
1000
1001 static void quorum_refresh_filename(BlockDriverState *bs, QDict *options)
1002 {
1003     BDRVQuorumState *s = bs->opaque;
1004     QDict *opts;
1005     QList *children;
1006     int i;
1007
1008     for (i = 0; i < s->num_children; i++) {
1009         bdrv_refresh_filename(s->children[i]->bs);
1010         if (!s->children[i]->bs->full_open_options) {
1011             return;
1012         }
1013     }
1014
1015     children = qlist_new();
1016     for (i = 0; i < s->num_children; i++) {
1017         QINCREF(s->children[i]->bs->full_open_options);
1018         qlist_append_obj(children,
1019                          QOBJECT(s->children[i]->bs->full_open_options));
1020     }
1021
1022     opts = qdict_new();
1023     qdict_put_obj(opts, "driver", QOBJECT(qstring_from_str("quorum")));
1024     qdict_put_obj(opts, QUORUM_OPT_VOTE_THRESHOLD,
1025                   QOBJECT(qint_from_int(s->threshold)));
1026     qdict_put_obj(opts, QUORUM_OPT_BLKVERIFY,
1027                   QOBJECT(qbool_from_bool(s->is_blkverify)));
1028     qdict_put_obj(opts, QUORUM_OPT_REWRITE,
1029                   QOBJECT(qbool_from_bool(s->rewrite_corrupted)));
1030     qdict_put_obj(opts, "children", QOBJECT(children));
1031
1032     bs->full_open_options = opts;
1033 }
1034
1035 static BlockDriver bdrv_quorum = {
1036     .format_name                        = "quorum",
1037     .protocol_name                      = "quorum",
1038
1039     .instance_size                      = sizeof(BDRVQuorumState),
1040
1041     .bdrv_file_open                     = quorum_open,
1042     .bdrv_close                         = quorum_close,
1043     .bdrv_refresh_filename              = quorum_refresh_filename,
1044
1045     .bdrv_co_flush_to_disk              = quorum_co_flush,
1046
1047     .bdrv_getlength                     = quorum_getlength,
1048
1049     .bdrv_aio_readv                     = quorum_aio_readv,
1050     .bdrv_aio_writev                    = quorum_aio_writev,
1051     .bdrv_invalidate_cache              = quorum_invalidate_cache,
1052
1053     .bdrv_detach_aio_context            = quorum_detach_aio_context,
1054     .bdrv_attach_aio_context            = quorum_attach_aio_context,
1055
1056     .is_filter                          = true,
1057     .bdrv_recurse_is_first_non_filter   = quorum_recurse_is_first_non_filter,
1058 };
1059
1060 static void bdrv_quorum_init(void)
1061 {
1062     if (!qcrypto_hash_supports(QCRYPTO_HASH_ALG_SHA256)) {
1063         /* SHA256 hash support is required for quorum device */
1064         return;
1065     }
1066     bdrv_register(&bdrv_quorum);
1067 }
1068
1069 block_init(bdrv_quorum_init);