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/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
  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     char 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     int j, ret;
 431     gnutls_hash_hd_t dig;
 432     QEMUIOVector *qiov = &acb->qcrs[i].qiov;
 433
 434     ret = gnutls_hash_init(&dig, GNUTLS_DIG_SHA256);
 435
 436     if (ret < 0) {
 437         return ret;
 438     }
 439
 440     for (j = 0; j < qiov->niov; j++) {
 441         ret = gnutls_hash(dig, qiov->iov[j].iov_base, qiov->iov[j].iov_len);
 442         if (ret < 0) {
 443             break;
 444         }
 445     }
 446
 447     gnutls_hash_deinit(dig, (void *) hash);
 448     return ret;
 449 }
 450
 451 static QuorumVoteVersion *quorum_get_vote_winner(QuorumVotes *votes)
 452 {
 453     int max = 0;
 454     QuorumVoteVersion *candidate, *winner = NULL;
 455
 456     QLIST_FOREACH(candidate, &votes->vote_list, next) {
 457         if (candidate->vote_count > max) {
 458             max = candidate->vote_count;
 459             winner = candidate;
 460         }
 461     }
 462
 463     return winner;
 464 }
 465
 466 /* qemu_iovec_compare is handy for blkverify mode because it returns the first
 467  * differing byte location. Yet it is handcoded to compare vectors one byte
 468  * after another so it does not benefit from the libc SIMD optimizations.
 469  * quorum_iovec_compare is written for speed and should be used in the non
 470  * blkverify mode of quorum.
 471  */
 472 static bool quorum_iovec_compare(QEMUIOVector *a, QEMUIOVector *b)
 473 {
 474     int i;
 475     int result;
 476
 477     assert(a->niov == b->niov);
 478     for (i = 0; i < a->niov; i++) {
 479         assert(a->iov[i].iov_len == b->iov[i].iov_len);
 480         result = memcmp(a->iov[i].iov_base,
 481                         b->iov[i].iov_base,
 482                         a->iov[i].iov_len);
 483         if (result) {
 484             return false;
 485         }
 486     }
 487
 488     return true;
 489 }
 490
 491 static void GCC_FMT_ATTR(2, 3) quorum_err(QuorumAIOCB *acb,
 492                                           const char *fmt, ...)
 493 {
 494     va_list ap;
 495
 496     va_start(ap, fmt);
 497     fprintf(stderr, "quorum: sector_num=%" PRId64 " nb_sectors=%d ",
 498             acb->sector_num, acb->nb_sectors);
 499     vfprintf(stderr, fmt, ap);
 500     fprintf(stderr, "\n");
 501     va_end(ap);
 502     exit(1);
 503 }
 504
 505 static bool quorum_compare(QuorumAIOCB *acb,
 506                            QEMUIOVector *a,
 507                            QEMUIOVector *b)
 508 {
 509     BDRVQuorumState *s = acb->common.bs->opaque;
 510     ssize_t offset;
 511
 512     /* This driver will replace blkverify in this particular case */
 513     if (s->is_blkverify) {
 514         offset = qemu_iovec_compare(a, b);
 515         if (offset != -1) {
 516             quorum_err(acb, "contents mismatch in sector %" PRId64,
 517                        acb->sector_num +
 518                        (uint64_t)(offset / BDRV_SECTOR_SIZE));
 519         }
 520         return true;
 521     }
 522
 523     return quorum_iovec_compare(a, b);
 524 }
 525
 526 /* Do a vote to get the error code */
 527 static int quorum_vote_error(QuorumAIOCB *acb)
 528 {
 529     BDRVQuorumState *s = acb->common.bs->opaque;
 530     QuorumVoteVersion *winner = NULL;
 531     QuorumVotes error_votes;
 532     QuorumVoteValue result_value;
 533     int i, ret = 0;
 534     bool error = false;
 535
 536     QLIST_INIT(&error_votes.vote_list);
 537     error_votes.compare = quorum_64bits_compare;
 538
 539     for (i = 0; i < s->num_children; i++) {
 540         ret = acb->qcrs[i].ret;
 541         if (ret) {
 542             error = true;
 543             result_value.l = ret;
 544             quorum_count_vote(&error_votes, &result_value, i);
 545         }
 546     }
 547
 548     if (error) {
 549         winner = quorum_get_vote_winner(&error_votes);
 550         ret = winner->value.l;
 551     }
 552
 553     quorum_free_vote_list(&error_votes);
 554
 555     return ret;
 556 }
 557
 558 static bool quorum_vote(QuorumAIOCB *acb)
 559 {
 560     bool quorum = true;
 561     bool rewrite = false;
 562     int i, j, ret;
 563     QuorumVoteValue hash;
 564     BDRVQuorumState *s = acb->common.bs->opaque;
 565     QuorumVoteVersion *winner;
 566
 567     if (quorum_has_too_much_io_failed(acb)) {
 568         return false;
 569     }
 570
 571     /* get the index of the first successful read */
 572     for (i = 0; i < s->num_children; i++) {
 573         if (!acb->qcrs[i].ret) {
 574             break;
 575         }
 576     }
 577
 578     assert(i < s->num_children);
 579
 580     /* compare this read with all other successful reads stopping at quorum
 581      * failure
 582      */
 583     for (j = i + 1; j < s->num_children; j++) {
 584         if (acb->qcrs[j].ret) {
 585             continue;
 586         }
 587         quorum = quorum_compare(acb, &acb->qcrs[i].qiov, &acb->qcrs[j].qiov);
 588         if (!quorum) {
 589             break;
 590        }
 591     }
 592
 593     /* Every successful read agrees */
 594     if (quorum) {
 595         quorum_copy_qiov(acb->qiov, &acb->qcrs[i].qiov);
 596         return false;
 597     }
 598
 599     /* compute hashes for each successful read, also store indexes */
 600     for (i = 0; i < s->num_children; i++) {
 601         if (acb->qcrs[i].ret) {
 602             continue;
 603         }
 604         ret = quorum_compute_hash(acb, i, &hash);
 605         /* if ever the hash computation failed */
 606         if (ret < 0) {
 607             acb->vote_ret = ret;
 608             goto free_exit;
 609         }
 610         quorum_count_vote(&acb->votes, &hash, i);
 611     }
 612
 613     /* vote to select the most represented version */
 614     winner = quorum_get_vote_winner(&acb->votes);
 615
 616     /* if the winner count is smaller than threshold the read fails */
 617     if (winner->vote_count < s->threshold) {
 618         quorum_report_failure(acb);
 619         acb->vote_ret = -EIO;
 620         goto free_exit;
 621     }
 622
 623     /* we have a winner: copy it */
 624     quorum_copy_qiov(acb->qiov, &acb->qcrs[winner->index].qiov);
 625
 626     /* some versions are bad print them */
 627     quorum_report_bad_versions(s, acb, &winner->value);
 628
 629     /* corruption correction is enabled */
 630     if (s->rewrite_corrupted) {
 631         rewrite = quorum_rewrite_bad_versions(s, acb, &winner->value);
 632     }
 633
 634 free_exit:
 635     /* free lists */
 636     quorum_free_vote_list(&acb->votes);
 637     return rewrite;
 638 }
 639
 640 static BlockAIOCB *read_quorum_children(QuorumAIOCB *acb)
 641 {
 642     BDRVQuorumState *s = acb->common.bs->opaque;
 643     int i;
 644
 645     for (i = 0; i < s->num_children; i++) {
 646         acb->qcrs[i].buf = qemu_blockalign(s->bs[i], acb->qiov->size);
 647         qemu_iovec_init(&acb->qcrs[i].qiov, acb->qiov->niov);
 648         qemu_iovec_clone(&acb->qcrs[i].qiov, acb->qiov, acb->qcrs[i].buf);
 649     }
 650
 651     for (i = 0; i < s->num_children; i++) {
 652         bdrv_aio_readv(s->bs[i], acb->sector_num, &acb->qcrs[i].qiov,
 653                        acb->nb_sectors, quorum_aio_cb, &acb->qcrs[i]);
 654     }
 655
 656     return &acb->common;
 657 }
 658
 659 static BlockAIOCB *read_fifo_child(QuorumAIOCB *acb)
 660 {
 661     BDRVQuorumState *s = acb->common.bs->opaque;
 662
 663     acb->qcrs[acb->child_iter].buf = qemu_blockalign(s->bs[acb->child_iter],
 664                                                      acb->qiov->size);
 665     qemu_iovec_init(&acb->qcrs[acb->child_iter].qiov, acb->qiov->niov);
 666     qemu_iovec_clone(&acb->qcrs[acb->child_iter].qiov, acb->qiov,
 667                      acb->qcrs[acb->child_iter].buf);
 668     bdrv_aio_readv(s->bs[acb->child_iter], acb->sector_num,
 669                    &acb->qcrs[acb->child_iter].qiov, acb->nb_sectors,
 670                    quorum_aio_cb, &acb->qcrs[acb->child_iter]);
 671
 672     return &acb->common;
 673 }
 674
 675 static BlockAIOCB *quorum_aio_readv(BlockDriverState *bs,
 676                                     int64_t sector_num,
 677                                     QEMUIOVector *qiov,
 678                                     int nb_sectors,
 679                                     BlockCompletionFunc *cb,
 680                                     void *opaque)
 681 {
 682     BDRVQuorumState *s = bs->opaque;
 683     QuorumAIOCB *acb = quorum_aio_get(s, bs, qiov, sector_num,
 684                                       nb_sectors, cb, opaque);
 685     acb->is_read = true;
 686
 687     if (s->read_pattern == QUORUM_READ_PATTERN_QUORUM) {
 688         acb->child_iter = s->num_children - 1;
 689         return read_quorum_children(acb);
 690     }
 691
 692     acb->child_iter = 0;
 693     return read_fifo_child(acb);
 694 }
 695
 696 static BlockAIOCB *quorum_aio_writev(BlockDriverState *bs,
 697                                      int64_t sector_num,
 698                                      QEMUIOVector *qiov,
 699                                      int nb_sectors,
 700                                      BlockCompletionFunc *cb,
 701                                      void *opaque)
 702 {
 703     BDRVQuorumState *s = bs->opaque;
 704     QuorumAIOCB *acb = quorum_aio_get(s, bs, qiov, sector_num, nb_sectors,
 705                                       cb, opaque);
 706     int i;
 707
 708     for (i = 0; i < s->num_children; i++) {
 709         acb->qcrs[i].aiocb = bdrv_aio_writev(s->bs[i], sector_num, qiov,
 710                                              nb_sectors, &quorum_aio_cb,
 711                                              &acb->qcrs[i]);
 712     }
 713
 714     return &acb->common;
 715 }
 716
 717 static int64_t quorum_getlength(BlockDriverState *bs)
 718 {
 719     BDRVQuorumState *s = bs->opaque;
 720     int64_t result;
 721     int i;
 722
 723     /* check that all file have the same length */
 724     result = bdrv_getlength(s->bs[0]);
 725     if (result < 0) {
 726         return result;
 727     }
 728     for (i = 1; i < s->num_children; i++) {
 729         int64_t value = bdrv_getlength(s->bs[i]);
 730         if (value < 0) {
 731             return value;
 732         }
 733         if (value != result) {
 734             return -EIO;
 735         }
 736     }
 737
 738     return result;
 739 }
 740
 741 static void quorum_invalidate_cache(BlockDriverState *bs, Error **errp)
 742 {
 743     BDRVQuorumState *s = bs->opaque;
 744     Error *local_err = NULL;
 745     int i;
 746
 747     for (i = 0; i < s->num_children; i++) {
 748         bdrv_invalidate_cache(s->bs[i], &local_err);
 749         if (local_err) {
 750             error_propagate(errp, local_err);
 751             return;
 752         }
 753     }
 754 }
 755
 756 static coroutine_fn int quorum_co_flush(BlockDriverState *bs)
 757 {
 758     BDRVQuorumState *s = bs->opaque;
 759     QuorumVoteVersion *winner = NULL;
 760     QuorumVotes error_votes;
 761     QuorumVoteValue result_value;
 762     int i;
 763     int result = 0;
 764
 765     QLIST_INIT(&error_votes.vote_list);
 766     error_votes.compare = quorum_64bits_compare;
 767
 768     for (i = 0; i < s->num_children; i++) {
 769         result = bdrv_co_flush(s->bs[i]);
 770         result_value.l = result;
 771         quorum_count_vote(&error_votes, &result_value, i);
 772     }
 773
 774     winner = quorum_get_vote_winner(&error_votes);
 775     result = winner->value.l;
 776
 777     quorum_free_vote_list(&error_votes);
 778
 779     return result;
 780 }
 781
 782 static bool quorum_recurse_is_first_non_filter(BlockDriverState *bs,
 783                                                BlockDriverState *candidate)
 784 {
 785     BDRVQuorumState *s = bs->opaque;
 786     int i;
 787
 788     for (i = 0; i < s->num_children; i++) {
 789         bool perm = bdrv_recurse_is_first_non_filter(s->bs[i],
 790                                                      candidate);
 791         if (perm) {
 792             return true;
 793         }
 794     }
 795
 796     return false;
 797 }
 798
 799 static int quorum_valid_threshold(int threshold, int num_children, Error **errp)
 800 {
 801
 802     if (threshold < 1) {
 803         error_set(errp, QERR_INVALID_PARAMETER_VALUE,
 804                   "vote-threshold", "value >= 1");
 805         return -ERANGE;
 806     }
 807
 808     if (threshold > num_children) {
 809         error_setg(errp, "threshold may not exceed children count");
 810         return -ERANGE;
 811     }
 812
 813     return 0;
 814 }
 815
 816 static QemuOptsList quorum_runtime_opts = {
 817     .name = "quorum",
 818     .head = QTAILQ_HEAD_INITIALIZER(quorum_runtime_opts.head),
 819     .desc = {
 820         {
 821             .name = QUORUM_OPT_VOTE_THRESHOLD,
 822             .type = QEMU_OPT_NUMBER,
 823             .help = "The number of vote needed for reaching quorum",
 824         },
 825         {
 826             .name = QUORUM_OPT_BLKVERIFY,
 827             .type = QEMU_OPT_BOOL,
 828             .help = "Trigger block verify mode if set",
 829         },
 830         {
 831             .name = QUORUM_OPT_REWRITE,
 832             .type = QEMU_OPT_BOOL,
 833             .help = "Rewrite corrupted block on read quorum",
 834         },
 835         {
 836             .name = QUORUM_OPT_READ_PATTERN,
 837             .type = QEMU_OPT_STRING,
 838             .help = "Allowed pattern: quorum, fifo. Quorum is default",
 839         },
 840         { /* end of list */ }
 841     },
 842 };
 843
 844 static int parse_read_pattern(const char *opt)
 845 {
 846     int i;
 847
 848     if (!opt) {
 849         /* Set quorum as default */
 850         return QUORUM_READ_PATTERN_QUORUM;
 851     }
 852
 853     for (i = 0; i < QUORUM_READ_PATTERN_MAX; i++) {
 854         if (!strcmp(opt, QuorumReadPattern_lookup[i])) {
 855             return i;
 856         }
 857     }
 858
 859     return -EINVAL;
 860 }
 861
 862 static int quorum_open(BlockDriverState *bs, QDict *options, int flags,
 863                        Error **errp)
 864 {
 865     BDRVQuorumState *s = bs->opaque;
 866     Error *local_err = NULL;
 867     QemuOpts *opts = NULL;
 868     bool *opened;
 869     QDict *sub = NULL;
 870     QList *list = NULL;
 871     const QListEntry *lentry;
 872     int i;
 873     int ret = 0;
 874
 875     qdict_flatten(options);
 876     qdict_extract_subqdict(options, &sub, "children.");
 877     qdict_array_split(sub, &list);
 878
 879     if (qdict_size(sub)) {
 880         error_setg(&local_err, "Invalid option children.%s",
 881                    qdict_first(sub)->key);
 882         ret = -EINVAL;
 883         goto exit;
 884     }
 885
 886     /* count how many different children are present */
 887     s->num_children = qlist_size(list);
 888     if (s->num_children < 2) {
 889         error_setg(&local_err,
 890                    "Number of provided children must be greater than 1");
 891         ret = -EINVAL;
 892         goto exit;
 893     }
 894
 895     opts = qemu_opts_create(&quorum_runtime_opts, NULL, 0, &error_abort);
 896     qemu_opts_absorb_qdict(opts, options, &local_err);
 897     if (local_err) {
 898         ret = -EINVAL;
 899         goto exit;
 900     }
 901
 902     s->threshold = qemu_opt_get_number(opts, QUORUM_OPT_VOTE_THRESHOLD, 0);
 903     ret = parse_read_pattern(qemu_opt_get(opts, QUORUM_OPT_READ_PATTERN));
 904     if (ret < 0) {
 905         error_setg(&local_err, "Please set read-pattern as fifo or quorum");
 906         goto exit;
 907     }
 908     s->read_pattern = ret;
 909
 910     if (s->read_pattern == QUORUM_READ_PATTERN_QUORUM) {
 911         /* and validate it against s->num_children */
 912         ret = quorum_valid_threshold(s->threshold, s->num_children, &local_err);
 913         if (ret < 0) {
 914             goto exit;
 915         }
 916
 917         /* is the driver in blkverify mode */
 918         if (qemu_opt_get_bool(opts, QUORUM_OPT_BLKVERIFY, false) &&
 919             s->num_children == 2 && s->threshold == 2) {
 920             s->is_blkverify = true;
 921         } else if (qemu_opt_get_bool(opts, QUORUM_OPT_BLKVERIFY, false)) {
 922             fprintf(stderr, "blkverify mode is set by setting blkverify=on "
 923                     "and using two files with vote_threshold=2\n");
 924         }
 925
 926         s->rewrite_corrupted = qemu_opt_get_bool(opts, QUORUM_OPT_REWRITE,
 927                                                  false);
 928         if (s->rewrite_corrupted && s->is_blkverify) {
 929             error_setg(&local_err,
 930                        "rewrite-corrupted=on cannot be used with blkverify=on");
 931             ret = -EINVAL;
 932             goto exit;
 933         }
 934     }
 935
 936     /* allocate the children BlockDriverState array */
 937     s->bs = g_new0(BlockDriverState *, s->num_children);
 938     opened = g_new0(bool, s->num_children);
 939
 940     for (i = 0, lentry = qlist_first(list); lentry;
 941          lentry = qlist_next(lentry), i++) {
 942         QDict *d;
 943         QString *string;
 944
 945         switch (qobject_type(lentry->value))
 946         {
 947             /* List of options */
 948             case QTYPE_QDICT:
 949                 d = qobject_to_qdict(lentry->value);
 950                 QINCREF(d);
 951                 ret = bdrv_open(&s->bs[i], NULL, NULL, d, flags, NULL,
 952                                 &local_err);
 953                 break;
 954
 955             /* QMP reference */
 956             case QTYPE_QSTRING:
 957                 string = qobject_to_qstring(lentry->value);
 958                 ret = bdrv_open(&s->bs[i], NULL, qstring_get_str(string), NULL,
 959                                 flags, NULL, &local_err);
 960                 break;
 961
 962             default:
 963                 error_setg(&local_err, "Specification of child block device %i "
 964                            "is invalid", i);
 965                 ret = -EINVAL;
 966         }
 967
 968         if (ret < 0) {
 969             goto close_exit;
 970         }
 971         opened[i] = true;
 972     }
 973
 974     g_free(opened);
 975     goto exit;
 976
 977 close_exit:
 978     /* cleanup on error */
 979     for (i = 0; i < s->num_children; i++) {
 980         if (!opened[i]) {
 981             continue;
 982         }
 983         bdrv_unref(s->bs[i]);
 984     }
 985     g_free(s->bs);
 986     g_free(opened);
 987 exit:
 988     qemu_opts_del(opts);
 989     /* propagate error */
 990     if (local_err) {
 991         error_propagate(errp, local_err);
 992     }
 993     QDECREF(list);
 994     QDECREF(sub);
 995     return ret;
 996 }
 997
 998 static void quorum_close(BlockDriverState *bs)
 999 {
1000     BDRVQuorumState *s = bs->opaque;
1001     int i;
1002
1003     for (i = 0; i < s->num_children; i++) {
1004         bdrv_unref(s->bs[i]);
1005     }
1006
1007     g_free(s->bs);
1008 }
1009
1010 static void quorum_detach_aio_context(BlockDriverState *bs)
1011 {
1012     BDRVQuorumState *s = bs->opaque;
1013     int i;
1014
1015     for (i = 0; i < s->num_children; i++) {
1016         bdrv_detach_aio_context(s->bs[i]);
1017     }
1018 }
1019
1020 static void quorum_attach_aio_context(BlockDriverState *bs,
1021                                       AioContext *new_context)
1022 {
1023     BDRVQuorumState *s = bs->opaque;
1024     int i;
1025
1026     for (i = 0; i < s->num_children; i++) {
1027         bdrv_attach_aio_context(s->bs[i], new_context);
1028     }
1029 }
1030
1031 static void quorum_refresh_filename(BlockDriverState *bs)
1032 {
1033     BDRVQuorumState *s = bs->opaque;
1034     QDict *opts;
1035     QList *children;
1036     int i;
1037
1038     for (i = 0; i < s->num_children; i++) {
1039         bdrv_refresh_filename(s->bs[i]);
1040         if (!s->bs[i]->full_open_options) {
1041             return;
1042         }
1043     }
1044
1045     children = qlist_new();
1046     for (i = 0; i < s->num_children; i++) {
1047         QINCREF(s->bs[i]->full_open_options);
1048         qlist_append_obj(children, QOBJECT(s->bs[i]->full_open_options));
1049     }
1050
1051     opts = qdict_new();
1052     qdict_put_obj(opts, "driver", QOBJECT(qstring_from_str("quorum")));
1053     qdict_put_obj(opts, QUORUM_OPT_VOTE_THRESHOLD,
1054                   QOBJECT(qint_from_int(s->threshold)));
1055     qdict_put_obj(opts, QUORUM_OPT_BLKVERIFY,
1056                   QOBJECT(qbool_from_int(s->is_blkverify)));
1057     qdict_put_obj(opts, QUORUM_OPT_REWRITE,
1058                   QOBJECT(qbool_from_int(s->rewrite_corrupted)));
1059     qdict_put_obj(opts, "children", QOBJECT(children));
1060
1061     bs->full_open_options = opts;
1062 }
1063
1064 static BlockDriver bdrv_quorum = {
1065     .format_name                        = "quorum",
1066     .protocol_name                      = "quorum",
1067
1068     .instance_size                      = sizeof(BDRVQuorumState),
1069
1070     .bdrv_file_open                     = quorum_open,
1071     .bdrv_close                         = quorum_close,
1072     .bdrv_refresh_filename              = quorum_refresh_filename,
1073
1074     .bdrv_co_flush_to_disk              = quorum_co_flush,
1075
1076     .bdrv_getlength                     = quorum_getlength,
1077
1078     .bdrv_aio_readv                     = quorum_aio_readv,
1079     .bdrv_aio_writev                    = quorum_aio_writev,
1080     .bdrv_invalidate_cache              = quorum_invalidate_cache,
1081
1082     .bdrv_detach_aio_context            = quorum_detach_aio_context,
1083     .bdrv_attach_aio_context            = quorum_attach_aio_context,
1084
1085     .is_filter                          = true,
1086     .bdrv_recurse_is_first_non_filter   = quorum_recurse_is_first_non_filter,
1087 };
1088
1089 static void bdrv_quorum_init(void)
1090 {
1091     bdrv_register(&bdrv_quorum);
1092 }
1093
1094 block_init(bdrv_quorum_init);