xref: /qemu/block/quorum.c (revision e688df6b)
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 "qemu/cutils.h"
18 #include "block/block_int.h"
19 #include "qapi/error.h"
20 #include "qapi/qmp/qbool.h"
21 #include "qapi/qmp/qdict.h"
22 #include "qapi/qmp/qerror.h"
23 #include "qapi/qmp/qjson.h"
24 #include "qapi/qmp/qlist.h"
25 #include "qapi/qmp/qstring.h"
26 #include "qapi-event.h"
27 #include "crypto/hash.h"
28 
29 #define HASH_LENGTH 32
30 
31 #define QUORUM_OPT_VOTE_THRESHOLD "vote-threshold"
32 #define QUORUM_OPT_BLKVERIFY      "blkverify"
33 #define QUORUM_OPT_REWRITE        "rewrite-corrupted"
34 #define QUORUM_OPT_READ_PATTERN   "read-pattern"
35 
36 /* This union holds a vote hash value */
37 typedef union QuorumVoteValue {
38     uint8_t h[HASH_LENGTH];    /* SHA-256 hash */
39     int64_t l;                 /* simpler 64 bits hash */
40 } QuorumVoteValue;
41 
42 /* A vote item */
43 typedef struct QuorumVoteItem {
44     int index;
45     QLIST_ENTRY(QuorumVoteItem) next;
46 } QuorumVoteItem;
47 
48 /* this structure is a vote version. A version is the set of votes sharing the
49  * same vote value.
50  * The set of votes will be tracked with the items field and its cardinality is
51  * vote_count.
52  */
53 typedef struct QuorumVoteVersion {
54     QuorumVoteValue value;
55     int index;
56     int vote_count;
57     QLIST_HEAD(, QuorumVoteItem) items;
58     QLIST_ENTRY(QuorumVoteVersion) next;
59 } QuorumVoteVersion;
60 
61 /* this structure holds a group of vote versions together */
62 typedef struct QuorumVotes {
63     QLIST_HEAD(, QuorumVoteVersion) vote_list;
64     bool (*compare)(QuorumVoteValue *a, QuorumVoteValue *b);
65 } QuorumVotes;
66 
67 /* the following structure holds the state of one quorum instance */
68 typedef struct BDRVQuorumState {
69     BdrvChild **children;  /* children BlockDriverStates */
70     int num_children;      /* children count */
71     unsigned next_child_index;  /* the index of the next child that should
72                                  * be added
73                                  */
74     int threshold;         /* if less than threshold children reads gave the
75                             * same result a quorum error occurs.
76                             */
77     bool is_blkverify;     /* true if the driver is in blkverify mode
78                             * Writes are mirrored on two children devices.
79                             * On reads the two children devices' contents are
80                             * compared and if a difference is spotted its
81                             * location is printed and the code aborts.
82                             * It is useful to debug other block drivers by
83                             * comparing them with a reference one.
84                             */
85     bool rewrite_corrupted;/* true if the driver must rewrite-on-read corrupted
86                             * block if Quorum is reached.
87                             */
88 
89     QuorumReadPattern read_pattern;
90 } BDRVQuorumState;
91 
92 typedef struct QuorumAIOCB QuorumAIOCB;
93 
94 /* Quorum will create one instance of the following structure per operation it
95  * performs on its children.
96  * So for each read/write operation coming from the upper layer there will be
97  * $children_count QuorumChildRequest.
98  */
99 typedef struct QuorumChildRequest {
100     BlockDriverState *bs;
101     QEMUIOVector qiov;
102     uint8_t *buf;
103     int ret;
104     QuorumAIOCB *parent;
105 } QuorumChildRequest;
106 
107 /* Quorum will use the following structure to track progress of each read/write
108  * operation received by the upper layer.
109  * This structure hold pointers to the QuorumChildRequest structures instances
110  * used to do operations on each children and track overall progress.
111  */
112 struct QuorumAIOCB {
113     BlockDriverState *bs;
114     Coroutine *co;
115 
116     /* Request metadata */
117     uint64_t offset;
118     uint64_t bytes;
119 
120     QEMUIOVector *qiov;         /* calling IOV */
121 
122     QuorumChildRequest *qcrs;   /* individual child requests */
123     int count;                  /* number of completed AIOCB */
124     int success_count;          /* number of successfully completed AIOCB */
125 
126     int rewrite_count;          /* number of replica to rewrite: count down to
127                                  * zero once writes are fired
128                                  */
129 
130     QuorumVotes votes;
131 
132     bool is_read;
133     int vote_ret;
134     int children_read;          /* how many children have been read from */
135 };
136 
137 typedef struct QuorumCo {
138     QuorumAIOCB *acb;
139     int idx;
140 } QuorumCo;
141 
142 static void quorum_aio_finalize(QuorumAIOCB *acb)
143 {
144     g_free(acb->qcrs);
145     g_free(acb);
146 }
147 
148 static bool quorum_sha256_compare(QuorumVoteValue *a, QuorumVoteValue *b)
149 {
150     return !memcmp(a->h, b->h, HASH_LENGTH);
151 }
152 
153 static bool quorum_64bits_compare(QuorumVoteValue *a, QuorumVoteValue *b)
154 {
155     return a->l == b->l;
156 }
157 
158 static QuorumAIOCB *quorum_aio_get(BlockDriverState *bs,
159                                    QEMUIOVector *qiov,
160                                    uint64_t offset,
161                                    uint64_t bytes)
162 {
163     BDRVQuorumState *s = bs->opaque;
164     QuorumAIOCB *acb = g_new(QuorumAIOCB, 1);
165     int i;
166 
167     *acb = (QuorumAIOCB) {
168         .co                 = qemu_coroutine_self(),
169         .bs                 = bs,
170         .offset             = offset,
171         .bytes              = bytes,
172         .qiov               = qiov,
173         .votes.compare      = quorum_sha256_compare,
174         .votes.vote_list    = QLIST_HEAD_INITIALIZER(acb.votes.vote_list),
175     };
176 
177     acb->qcrs = g_new0(QuorumChildRequest, s->num_children);
178     for (i = 0; i < s->num_children; i++) {
179         acb->qcrs[i].buf = NULL;
180         acb->qcrs[i].ret = 0;
181         acb->qcrs[i].parent = acb;
182     }
183 
184     return acb;
185 }
186 
187 static void quorum_report_bad(QuorumOpType type, uint64_t offset,
188                               uint64_t bytes, char *node_name, int ret)
189 {
190     const char *msg = NULL;
191     int64_t start_sector = offset / BDRV_SECTOR_SIZE;
192     int64_t end_sector = DIV_ROUND_UP(offset + bytes, BDRV_SECTOR_SIZE);
193 
194     if (ret < 0) {
195         msg = strerror(-ret);
196     }
197 
198     qapi_event_send_quorum_report_bad(type, !!msg, msg, node_name, start_sector,
199                                       end_sector - start_sector, &error_abort);
200 }
201 
202 static void quorum_report_failure(QuorumAIOCB *acb)
203 {
204     const char *reference = bdrv_get_device_or_node_name(acb->bs);
205     int64_t start_sector = acb->offset / BDRV_SECTOR_SIZE;
206     int64_t end_sector = DIV_ROUND_UP(acb->offset + acb->bytes,
207                                       BDRV_SECTOR_SIZE);
208 
209     qapi_event_send_quorum_failure(reference, start_sector,
210                                    end_sector - start_sector, &error_abort);
211 }
212 
213 static int quorum_vote_error(QuorumAIOCB *acb);
214 
215 static bool quorum_has_too_much_io_failed(QuorumAIOCB *acb)
216 {
217     BDRVQuorumState *s = acb->bs->opaque;
218 
219     if (acb->success_count < s->threshold) {
220         acb->vote_ret = quorum_vote_error(acb);
221         quorum_report_failure(acb);
222         return true;
223     }
224 
225     return false;
226 }
227 
228 static int read_fifo_child(QuorumAIOCB *acb);
229 
230 static void quorum_copy_qiov(QEMUIOVector *dest, QEMUIOVector *source)
231 {
232     int i;
233     assert(dest->niov == source->niov);
234     assert(dest->size == source->size);
235     for (i = 0; i < source->niov; i++) {
236         assert(dest->iov[i].iov_len == source->iov[i].iov_len);
237         memcpy(dest->iov[i].iov_base,
238                source->iov[i].iov_base,
239                source->iov[i].iov_len);
240     }
241 }
242 
243 static void quorum_report_bad_acb(QuorumChildRequest *sacb, int ret)
244 {
245     QuorumAIOCB *acb = sacb->parent;
246     QuorumOpType type = acb->is_read ? QUORUM_OP_TYPE_READ : QUORUM_OP_TYPE_WRITE;
247     quorum_report_bad(type, acb->offset, acb->bytes, sacb->bs->node_name, ret);
248 }
249 
250 static void quorum_report_bad_versions(BDRVQuorumState *s,
251                                        QuorumAIOCB *acb,
252                                        QuorumVoteValue *value)
253 {
254     QuorumVoteVersion *version;
255     QuorumVoteItem *item;
256 
257     QLIST_FOREACH(version, &acb->votes.vote_list, next) {
258         if (acb->votes.compare(&version->value, value)) {
259             continue;
260         }
261         QLIST_FOREACH(item, &version->items, next) {
262             quorum_report_bad(QUORUM_OP_TYPE_READ, acb->offset, acb->bytes,
263                               s->children[item->index]->bs->node_name, 0);
264         }
265     }
266 }
267 
268 static void quorum_rewrite_entry(void *opaque)
269 {
270     QuorumCo *co = opaque;
271     QuorumAIOCB *acb = co->acb;
272     BDRVQuorumState *s = acb->bs->opaque;
273 
274     /* Ignore any errors, it's just a correction attempt for already
275      * corrupted data. */
276     bdrv_co_pwritev(s->children[co->idx], acb->offset, acb->bytes,
277                     acb->qiov, 0);
278 
279     /* Wake up the caller after the last rewrite */
280     acb->rewrite_count--;
281     if (!acb->rewrite_count) {
282         qemu_coroutine_enter_if_inactive(acb->co);
283     }
284 }
285 
286 static bool quorum_rewrite_bad_versions(QuorumAIOCB *acb,
287                                         QuorumVoteValue *value)
288 {
289     QuorumVoteVersion *version;
290     QuorumVoteItem *item;
291     int count = 0;
292 
293     /* first count the number of bad versions: done first to avoid concurrency
294      * issues.
295      */
296     QLIST_FOREACH(version, &acb->votes.vote_list, next) {
297         if (acb->votes.compare(&version->value, value)) {
298             continue;
299         }
300         QLIST_FOREACH(item, &version->items, next) {
301             count++;
302         }
303     }
304 
305     /* quorum_rewrite_entry will count down this to zero */
306     acb->rewrite_count = count;
307 
308     /* now fire the correcting rewrites */
309     QLIST_FOREACH(version, &acb->votes.vote_list, next) {
310         if (acb->votes.compare(&version->value, value)) {
311             continue;
312         }
313         QLIST_FOREACH(item, &version->items, next) {
314             Coroutine *co;
315             QuorumCo data = {
316                 .acb = acb,
317                 .idx = item->index,
318             };
319 
320             co = qemu_coroutine_create(quorum_rewrite_entry, &data);
321             qemu_coroutine_enter(co);
322         }
323     }
324 
325     /* return true if any rewrite is done else false */
326     return count;
327 }
328 
329 static void quorum_count_vote(QuorumVotes *votes,
330                               QuorumVoteValue *value,
331                               int index)
332 {
333     QuorumVoteVersion *v = NULL, *version = NULL;
334     QuorumVoteItem *item;
335 
336     /* look if we have something with this hash */
337     QLIST_FOREACH(v, &votes->vote_list, next) {
338         if (votes->compare(&v->value, value)) {
339             version = v;
340             break;
341         }
342     }
343 
344     /* It's a version not yet in the list add it */
345     if (!version) {
346         version = g_new0(QuorumVoteVersion, 1);
347         QLIST_INIT(&version->items);
348         memcpy(&version->value, value, sizeof(version->value));
349         version->index = index;
350         version->vote_count = 0;
351         QLIST_INSERT_HEAD(&votes->vote_list, version, next);
352     }
353 
354     version->vote_count++;
355 
356     item = g_new0(QuorumVoteItem, 1);
357     item->index = index;
358     QLIST_INSERT_HEAD(&version->items, item, next);
359 }
360 
361 static void quorum_free_vote_list(QuorumVotes *votes)
362 {
363     QuorumVoteVersion *version, *next_version;
364     QuorumVoteItem *item, *next_item;
365 
366     QLIST_FOREACH_SAFE(version, &votes->vote_list, next, next_version) {
367         QLIST_REMOVE(version, next);
368         QLIST_FOREACH_SAFE(item, &version->items, next, next_item) {
369             QLIST_REMOVE(item, next);
370             g_free(item);
371         }
372         g_free(version);
373     }
374 }
375 
376 static int quorum_compute_hash(QuorumAIOCB *acb, int i, QuorumVoteValue *hash)
377 {
378     QEMUIOVector *qiov = &acb->qcrs[i].qiov;
379     size_t len = sizeof(hash->h);
380     uint8_t *data = hash->h;
381 
382     /* XXX - would be nice if we could pass in the Error **
383      * and propagate that back, but this quorum code is
384      * restricted to just errno values currently */
385     if (qcrypto_hash_bytesv(QCRYPTO_HASH_ALG_SHA256,
386                             qiov->iov, qiov->niov,
387                             &data, &len,
388                             NULL) < 0) {
389         return -EINVAL;
390     }
391 
392     return 0;
393 }
394 
395 static QuorumVoteVersion *quorum_get_vote_winner(QuorumVotes *votes)
396 {
397     int max = 0;
398     QuorumVoteVersion *candidate, *winner = NULL;
399 
400     QLIST_FOREACH(candidate, &votes->vote_list, next) {
401         if (candidate->vote_count > max) {
402             max = candidate->vote_count;
403             winner = candidate;
404         }
405     }
406 
407     return winner;
408 }
409 
410 /* qemu_iovec_compare is handy for blkverify mode because it returns the first
411  * differing byte location. Yet it is handcoded to compare vectors one byte
412  * after another so it does not benefit from the libc SIMD optimizations.
413  * quorum_iovec_compare is written for speed and should be used in the non
414  * blkverify mode of quorum.
415  */
416 static bool quorum_iovec_compare(QEMUIOVector *a, QEMUIOVector *b)
417 {
418     int i;
419     int result;
420 
421     assert(a->niov == b->niov);
422     for (i = 0; i < a->niov; i++) {
423         assert(a->iov[i].iov_len == b->iov[i].iov_len);
424         result = memcmp(a->iov[i].iov_base,
425                         b->iov[i].iov_base,
426                         a->iov[i].iov_len);
427         if (result) {
428             return false;
429         }
430     }
431 
432     return true;
433 }
434 
435 static void GCC_FMT_ATTR(2, 3) quorum_err(QuorumAIOCB *acb,
436                                           const char *fmt, ...)
437 {
438     va_list ap;
439 
440     va_start(ap, fmt);
441     fprintf(stderr, "quorum: offset=%" PRIu64 " bytes=%" PRIu64 " ",
442             acb->offset, acb->bytes);
443     vfprintf(stderr, fmt, ap);
444     fprintf(stderr, "\n");
445     va_end(ap);
446     exit(1);
447 }
448 
449 static bool quorum_compare(QuorumAIOCB *acb,
450                            QEMUIOVector *a,
451                            QEMUIOVector *b)
452 {
453     BDRVQuorumState *s = acb->bs->opaque;
454     ssize_t offset;
455 
456     /* This driver will replace blkverify in this particular case */
457     if (s->is_blkverify) {
458         offset = qemu_iovec_compare(a, b);
459         if (offset != -1) {
460             quorum_err(acb, "contents mismatch at offset %" PRIu64,
461                        acb->offset + offset);
462         }
463         return true;
464     }
465 
466     return quorum_iovec_compare(a, b);
467 }
468 
469 /* Do a vote to get the error code */
470 static int quorum_vote_error(QuorumAIOCB *acb)
471 {
472     BDRVQuorumState *s = acb->bs->opaque;
473     QuorumVoteVersion *winner = NULL;
474     QuorumVotes error_votes;
475     QuorumVoteValue result_value;
476     int i, ret = 0;
477     bool error = false;
478 
479     QLIST_INIT(&error_votes.vote_list);
480     error_votes.compare = quorum_64bits_compare;
481 
482     for (i = 0; i < s->num_children; i++) {
483         ret = acb->qcrs[i].ret;
484         if (ret) {
485             error = true;
486             result_value.l = ret;
487             quorum_count_vote(&error_votes, &result_value, i);
488         }
489     }
490 
491     if (error) {
492         winner = quorum_get_vote_winner(&error_votes);
493         ret = winner->value.l;
494     }
495 
496     quorum_free_vote_list(&error_votes);
497 
498     return ret;
499 }
500 
501 static void quorum_vote(QuorumAIOCB *acb)
502 {
503     bool quorum = true;
504     int i, j, ret;
505     QuorumVoteValue hash;
506     BDRVQuorumState *s = acb->bs->opaque;
507     QuorumVoteVersion *winner;
508 
509     if (quorum_has_too_much_io_failed(acb)) {
510         return;
511     }
512 
513     /* get the index of the first successful read */
514     for (i = 0; i < s->num_children; i++) {
515         if (!acb->qcrs[i].ret) {
516             break;
517         }
518     }
519 
520     assert(i < s->num_children);
521 
522     /* compare this read with all other successful reads stopping at quorum
523      * failure
524      */
525     for (j = i + 1; j < s->num_children; j++) {
526         if (acb->qcrs[j].ret) {
527             continue;
528         }
529         quorum = quorum_compare(acb, &acb->qcrs[i].qiov, &acb->qcrs[j].qiov);
530         if (!quorum) {
531             break;
532        }
533     }
534 
535     /* Every successful read agrees */
536     if (quorum) {
537         quorum_copy_qiov(acb->qiov, &acb->qcrs[i].qiov);
538         return;
539     }
540 
541     /* compute hashes for each successful read, also store indexes */
542     for (i = 0; i < s->num_children; i++) {
543         if (acb->qcrs[i].ret) {
544             continue;
545         }
546         ret = quorum_compute_hash(acb, i, &hash);
547         /* if ever the hash computation failed */
548         if (ret < 0) {
549             acb->vote_ret = ret;
550             goto free_exit;
551         }
552         quorum_count_vote(&acb->votes, &hash, i);
553     }
554 
555     /* vote to select the most represented version */
556     winner = quorum_get_vote_winner(&acb->votes);
557 
558     /* if the winner count is smaller than threshold the read fails */
559     if (winner->vote_count < s->threshold) {
560         quorum_report_failure(acb);
561         acb->vote_ret = -EIO;
562         goto free_exit;
563     }
564 
565     /* we have a winner: copy it */
566     quorum_copy_qiov(acb->qiov, &acb->qcrs[winner->index].qiov);
567 
568     /* some versions are bad print them */
569     quorum_report_bad_versions(s, acb, &winner->value);
570 
571     /* corruption correction is enabled */
572     if (s->rewrite_corrupted) {
573         quorum_rewrite_bad_versions(acb, &winner->value);
574     }
575 
576 free_exit:
577     /* free lists */
578     quorum_free_vote_list(&acb->votes);
579 }
580 
581 static void read_quorum_children_entry(void *opaque)
582 {
583     QuorumCo *co = opaque;
584     QuorumAIOCB *acb = co->acb;
585     BDRVQuorumState *s = acb->bs->opaque;
586     int i = co->idx;
587     QuorumChildRequest *sacb = &acb->qcrs[i];
588 
589     sacb->bs = s->children[i]->bs;
590     sacb->ret = bdrv_co_preadv(s->children[i], acb->offset, acb->bytes,
591                                &acb->qcrs[i].qiov, 0);
592 
593     if (sacb->ret == 0) {
594         acb->success_count++;
595     } else {
596         quorum_report_bad_acb(sacb, sacb->ret);
597     }
598 
599     acb->count++;
600     assert(acb->count <= s->num_children);
601     assert(acb->success_count <= s->num_children);
602 
603     /* Wake up the caller after the last read */
604     if (acb->count == s->num_children) {
605         qemu_coroutine_enter_if_inactive(acb->co);
606     }
607 }
608 
609 static int read_quorum_children(QuorumAIOCB *acb)
610 {
611     BDRVQuorumState *s = acb->bs->opaque;
612     int i, ret;
613 
614     acb->children_read = s->num_children;
615     for (i = 0; i < s->num_children; i++) {
616         acb->qcrs[i].buf = qemu_blockalign(s->children[i]->bs, acb->qiov->size);
617         qemu_iovec_init(&acb->qcrs[i].qiov, acb->qiov->niov);
618         qemu_iovec_clone(&acb->qcrs[i].qiov, acb->qiov, acb->qcrs[i].buf);
619     }
620 
621     for (i = 0; i < s->num_children; i++) {
622         Coroutine *co;
623         QuorumCo data = {
624             .acb = acb,
625             .idx = i,
626         };
627 
628         co = qemu_coroutine_create(read_quorum_children_entry, &data);
629         qemu_coroutine_enter(co);
630     }
631 
632     while (acb->count < s->num_children) {
633         qemu_coroutine_yield();
634     }
635 
636     /* Do the vote on read */
637     quorum_vote(acb);
638     for (i = 0; i < s->num_children; i++) {
639         qemu_vfree(acb->qcrs[i].buf);
640         qemu_iovec_destroy(&acb->qcrs[i].qiov);
641     }
642 
643     while (acb->rewrite_count) {
644         qemu_coroutine_yield();
645     }
646 
647     ret = acb->vote_ret;
648 
649     return ret;
650 }
651 
652 static int read_fifo_child(QuorumAIOCB *acb)
653 {
654     BDRVQuorumState *s = acb->bs->opaque;
655     int n, ret;
656 
657     /* We try to read the next child in FIFO order if we failed to read */
658     do {
659         n = acb->children_read++;
660         acb->qcrs[n].bs = s->children[n]->bs;
661         ret = bdrv_co_preadv(s->children[n], acb->offset, acb->bytes,
662                              acb->qiov, 0);
663         if (ret < 0) {
664             quorum_report_bad_acb(&acb->qcrs[n], ret);
665         }
666     } while (ret < 0 && acb->children_read < s->num_children);
667 
668     /* FIXME: rewrite failed children if acb->children_read > 1? */
669 
670     return ret;
671 }
672 
673 static int quorum_co_preadv(BlockDriverState *bs, uint64_t offset,
674                             uint64_t bytes, QEMUIOVector *qiov, int flags)
675 {
676     BDRVQuorumState *s = bs->opaque;
677     QuorumAIOCB *acb = quorum_aio_get(bs, qiov, offset, bytes);
678     int ret;
679 
680     acb->is_read = true;
681     acb->children_read = 0;
682 
683     if (s->read_pattern == QUORUM_READ_PATTERN_QUORUM) {
684         ret = read_quorum_children(acb);
685     } else {
686         ret = read_fifo_child(acb);
687     }
688     quorum_aio_finalize(acb);
689 
690     return ret;
691 }
692 
693 static void write_quorum_entry(void *opaque)
694 {
695     QuorumCo *co = opaque;
696     QuorumAIOCB *acb = co->acb;
697     BDRVQuorumState *s = acb->bs->opaque;
698     int i = co->idx;
699     QuorumChildRequest *sacb = &acb->qcrs[i];
700 
701     sacb->bs = s->children[i]->bs;
702     sacb->ret = bdrv_co_pwritev(s->children[i], acb->offset, acb->bytes,
703                                 acb->qiov, 0);
704     if (sacb->ret == 0) {
705         acb->success_count++;
706     } else {
707         quorum_report_bad_acb(sacb, sacb->ret);
708     }
709     acb->count++;
710     assert(acb->count <= s->num_children);
711     assert(acb->success_count <= s->num_children);
712 
713     /* Wake up the caller after the last write */
714     if (acb->count == s->num_children) {
715         qemu_coroutine_enter_if_inactive(acb->co);
716     }
717 }
718 
719 static int quorum_co_pwritev(BlockDriverState *bs, uint64_t offset,
720                              uint64_t bytes, QEMUIOVector *qiov, int flags)
721 {
722     BDRVQuorumState *s = bs->opaque;
723     QuorumAIOCB *acb = quorum_aio_get(bs, qiov, offset, bytes);
724     int i, ret;
725 
726     for (i = 0; i < s->num_children; i++) {
727         Coroutine *co;
728         QuorumCo data = {
729             .acb = acb,
730             .idx = i,
731         };
732 
733         co = qemu_coroutine_create(write_quorum_entry, &data);
734         qemu_coroutine_enter(co);
735     }
736 
737     while (acb->count < s->num_children) {
738         qemu_coroutine_yield();
739     }
740 
741     quorum_has_too_much_io_failed(acb);
742 
743     ret = acb->vote_ret;
744     quorum_aio_finalize(acb);
745 
746     return ret;
747 }
748 
749 static int64_t quorum_getlength(BlockDriverState *bs)
750 {
751     BDRVQuorumState *s = bs->opaque;
752     int64_t result;
753     int i;
754 
755     /* check that all file have the same length */
756     result = bdrv_getlength(s->children[0]->bs);
757     if (result < 0) {
758         return result;
759     }
760     for (i = 1; i < s->num_children; i++) {
761         int64_t value = bdrv_getlength(s->children[i]->bs);
762         if (value < 0) {
763             return value;
764         }
765         if (value != result) {
766             return -EIO;
767         }
768     }
769 
770     return result;
771 }
772 
773 static coroutine_fn int quorum_co_flush(BlockDriverState *bs)
774 {
775     BDRVQuorumState *s = bs->opaque;
776     QuorumVoteVersion *winner = NULL;
777     QuorumVotes error_votes;
778     QuorumVoteValue result_value;
779     int i;
780     int result = 0;
781     int success_count = 0;
782 
783     QLIST_INIT(&error_votes.vote_list);
784     error_votes.compare = quorum_64bits_compare;
785 
786     for (i = 0; i < s->num_children; i++) {
787         result = bdrv_co_flush(s->children[i]->bs);
788         if (result) {
789             quorum_report_bad(QUORUM_OP_TYPE_FLUSH, 0, 0,
790                               s->children[i]->bs->node_name, result);
791             result_value.l = result;
792             quorum_count_vote(&error_votes, &result_value, i);
793         } else {
794             success_count++;
795         }
796     }
797 
798     if (success_count >= s->threshold) {
799         result = 0;
800     } else {
801         winner = quorum_get_vote_winner(&error_votes);
802         result = winner->value.l;
803     }
804     quorum_free_vote_list(&error_votes);
805 
806     return result;
807 }
808 
809 static bool quorum_recurse_is_first_non_filter(BlockDriverState *bs,
810                                                BlockDriverState *candidate)
811 {
812     BDRVQuorumState *s = bs->opaque;
813     int i;
814 
815     for (i = 0; i < s->num_children; i++) {
816         bool perm = bdrv_recurse_is_first_non_filter(s->children[i]->bs,
817                                                      candidate);
818         if (perm) {
819             return true;
820         }
821     }
822 
823     return false;
824 }
825 
826 static int quorum_valid_threshold(int threshold, int num_children, Error **errp)
827 {
828 
829     if (threshold < 1) {
830         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
831                    "vote-threshold", "value >= 1");
832         return -ERANGE;
833     }
834 
835     if (threshold > num_children) {
836         error_setg(errp, "threshold may not exceed children count");
837         return -ERANGE;
838     }
839 
840     return 0;
841 }
842 
843 static QemuOptsList quorum_runtime_opts = {
844     .name = "quorum",
845     .head = QTAILQ_HEAD_INITIALIZER(quorum_runtime_opts.head),
846     .desc = {
847         {
848             .name = QUORUM_OPT_VOTE_THRESHOLD,
849             .type = QEMU_OPT_NUMBER,
850             .help = "The number of vote needed for reaching quorum",
851         },
852         {
853             .name = QUORUM_OPT_BLKVERIFY,
854             .type = QEMU_OPT_BOOL,
855             .help = "Trigger block verify mode if set",
856         },
857         {
858             .name = QUORUM_OPT_REWRITE,
859             .type = QEMU_OPT_BOOL,
860             .help = "Rewrite corrupted block on read quorum",
861         },
862         {
863             .name = QUORUM_OPT_READ_PATTERN,
864             .type = QEMU_OPT_STRING,
865             .help = "Allowed pattern: quorum, fifo. Quorum is default",
866         },
867         { /* end of list */ }
868     },
869 };
870 
871 static int quorum_open(BlockDriverState *bs, QDict *options, int flags,
872                        Error **errp)
873 {
874     BDRVQuorumState *s = bs->opaque;
875     Error *local_err = NULL;
876     QemuOpts *opts = NULL;
877     const char *pattern_str;
878     bool *opened;
879     int i;
880     int ret = 0;
881 
882     qdict_flatten(options);
883 
884     /* count how many different children are present */
885     s->num_children = qdict_array_entries(options, "children.");
886     if (s->num_children < 0) {
887         error_setg(&local_err, "Option children is not a valid array");
888         ret = -EINVAL;
889         goto exit;
890     }
891     if (s->num_children < 1) {
892         error_setg(&local_err,
893                    "Number of provided children must be 1 or more");
894         ret = -EINVAL;
895         goto exit;
896     }
897 
898     opts = qemu_opts_create(&quorum_runtime_opts, NULL, 0, &error_abort);
899     qemu_opts_absorb_qdict(opts, options, &local_err);
900     if (local_err) {
901         ret = -EINVAL;
902         goto exit;
903     }
904 
905     s->threshold = qemu_opt_get_number(opts, QUORUM_OPT_VOTE_THRESHOLD, 0);
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     pattern_str = qemu_opt_get(opts, QUORUM_OPT_READ_PATTERN);
913     if (!pattern_str) {
914         ret = QUORUM_READ_PATTERN_QUORUM;
915     } else {
916         ret = qapi_enum_parse(&QuorumReadPattern_lookup, pattern_str,
917                               -EINVAL, NULL);
918     }
919     if (ret < 0) {
920         error_setg(&local_err, "Please set read-pattern as fifo or quorum");
921         goto exit;
922     }
923     s->read_pattern = ret;
924 
925     if (s->read_pattern == QUORUM_READ_PATTERN_QUORUM) {
926         /* is the driver in blkverify mode */
927         if (qemu_opt_get_bool(opts, QUORUM_OPT_BLKVERIFY, false) &&
928             s->num_children == 2 && s->threshold == 2) {
929             s->is_blkverify = true;
930         } else if (qemu_opt_get_bool(opts, QUORUM_OPT_BLKVERIFY, false)) {
931             fprintf(stderr, "blkverify mode is set by setting blkverify=on "
932                     "and using two files with vote_threshold=2\n");
933         }
934 
935         s->rewrite_corrupted = qemu_opt_get_bool(opts, QUORUM_OPT_REWRITE,
936                                                  false);
937         if (s->rewrite_corrupted && s->is_blkverify) {
938             error_setg(&local_err,
939                        "rewrite-corrupted=on cannot be used with blkverify=on");
940             ret = -EINVAL;
941             goto exit;
942         }
943     }
944 
945     /* allocate the children array */
946     s->children = g_new0(BdrvChild *, s->num_children);
947     opened = g_new0(bool, s->num_children);
948 
949     for (i = 0; i < s->num_children; i++) {
950         char indexstr[32];
951         ret = snprintf(indexstr, 32, "children.%d", i);
952         assert(ret < 32);
953 
954         s->children[i] = bdrv_open_child(NULL, options, indexstr, bs,
955                                          &child_format, false, &local_err);
956         if (local_err) {
957             ret = -EINVAL;
958             goto close_exit;
959         }
960 
961         opened[i] = true;
962     }
963     s->next_child_index = s->num_children;
964 
965     g_free(opened);
966     goto exit;
967 
968 close_exit:
969     /* cleanup on error */
970     for (i = 0; i < s->num_children; i++) {
971         if (!opened[i]) {
972             continue;
973         }
974         bdrv_unref_child(bs, s->children[i]);
975     }
976     g_free(s->children);
977     g_free(opened);
978 exit:
979     qemu_opts_del(opts);
980     /* propagate error */
981     error_propagate(errp, local_err);
982     return ret;
983 }
984 
985 static void quorum_close(BlockDriverState *bs)
986 {
987     BDRVQuorumState *s = bs->opaque;
988     int i;
989 
990     for (i = 0; i < s->num_children; i++) {
991         bdrv_unref_child(bs, s->children[i]);
992     }
993 
994     g_free(s->children);
995 }
996 
997 static void quorum_add_child(BlockDriverState *bs, BlockDriverState *child_bs,
998                              Error **errp)
999 {
1000     BDRVQuorumState *s = bs->opaque;
1001     BdrvChild *child;
1002     char indexstr[32];
1003     int ret;
1004 
1005     assert(s->num_children <= INT_MAX / sizeof(BdrvChild *));
1006     if (s->num_children == INT_MAX / sizeof(BdrvChild *) ||
1007         s->next_child_index == UINT_MAX) {
1008         error_setg(errp, "Too many children");
1009         return;
1010     }
1011 
1012     ret = snprintf(indexstr, 32, "children.%u", s->next_child_index);
1013     if (ret < 0 || ret >= 32) {
1014         error_setg(errp, "cannot generate child name");
1015         return;
1016     }
1017     s->next_child_index++;
1018 
1019     bdrv_drained_begin(bs);
1020 
1021     /* We can safely add the child now */
1022     bdrv_ref(child_bs);
1023 
1024     child = bdrv_attach_child(bs, child_bs, indexstr, &child_format, errp);
1025     if (child == NULL) {
1026         s->next_child_index--;
1027         bdrv_unref(child_bs);
1028         goto out;
1029     }
1030     s->children = g_renew(BdrvChild *, s->children, s->num_children + 1);
1031     s->children[s->num_children++] = child;
1032 
1033 out:
1034     bdrv_drained_end(bs);
1035 }
1036 
1037 static void quorum_del_child(BlockDriverState *bs, BdrvChild *child,
1038                              Error **errp)
1039 {
1040     BDRVQuorumState *s = bs->opaque;
1041     int i;
1042 
1043     for (i = 0; i < s->num_children; i++) {
1044         if (s->children[i] == child) {
1045             break;
1046         }
1047     }
1048 
1049     /* we have checked it in bdrv_del_child() */
1050     assert(i < s->num_children);
1051 
1052     if (s->num_children <= s->threshold) {
1053         error_setg(errp,
1054             "The number of children cannot be lower than the vote threshold %d",
1055             s->threshold);
1056         return;
1057     }
1058 
1059     bdrv_drained_begin(bs);
1060 
1061     /* We can safely remove this child now */
1062     memmove(&s->children[i], &s->children[i + 1],
1063             (s->num_children - i - 1) * sizeof(BdrvChild *));
1064     s->children = g_renew(BdrvChild *, s->children, --s->num_children);
1065     bdrv_unref_child(bs, child);
1066 
1067     bdrv_drained_end(bs);
1068 }
1069 
1070 static void quorum_refresh_filename(BlockDriverState *bs, QDict *options)
1071 {
1072     BDRVQuorumState *s = bs->opaque;
1073     QDict *opts;
1074     QList *children;
1075     int i;
1076 
1077     for (i = 0; i < s->num_children; i++) {
1078         bdrv_refresh_filename(s->children[i]->bs);
1079         if (!s->children[i]->bs->full_open_options) {
1080             return;
1081         }
1082     }
1083 
1084     children = qlist_new();
1085     for (i = 0; i < s->num_children; i++) {
1086         QINCREF(s->children[i]->bs->full_open_options);
1087         qlist_append(children, s->children[i]->bs->full_open_options);
1088     }
1089 
1090     opts = qdict_new();
1091     qdict_put_str(opts, "driver", "quorum");
1092     qdict_put_int(opts, QUORUM_OPT_VOTE_THRESHOLD, s->threshold);
1093     qdict_put_bool(opts, QUORUM_OPT_BLKVERIFY, s->is_blkverify);
1094     qdict_put_bool(opts, QUORUM_OPT_REWRITE, s->rewrite_corrupted);
1095     qdict_put(opts, "children", children);
1096 
1097     bs->full_open_options = opts;
1098 }
1099 
1100 static BlockDriver bdrv_quorum = {
1101     .format_name                        = "quorum",
1102     .protocol_name                      = "quorum",
1103 
1104     .instance_size                      = sizeof(BDRVQuorumState),
1105 
1106     .bdrv_file_open                     = quorum_open,
1107     .bdrv_close                         = quorum_close,
1108     .bdrv_refresh_filename              = quorum_refresh_filename,
1109 
1110     .bdrv_co_flush_to_disk              = quorum_co_flush,
1111 
1112     .bdrv_getlength                     = quorum_getlength,
1113 
1114     .bdrv_co_preadv                     = quorum_co_preadv,
1115     .bdrv_co_pwritev                    = quorum_co_pwritev,
1116 
1117     .bdrv_add_child                     = quorum_add_child,
1118     .bdrv_del_child                     = quorum_del_child,
1119 
1120     .bdrv_child_perm                    = bdrv_filter_default_perms,
1121 
1122     .is_filter                          = true,
1123     .bdrv_recurse_is_first_non_filter   = quorum_recurse_is_first_non_filter,
1124 };
1125 
1126 static void bdrv_quorum_init(void)
1127 {
1128     if (!qcrypto_hash_supports(QCRYPTO_HASH_ALG_SHA256)) {
1129         /* SHA256 hash support is required for quorum device */
1130         return;
1131     }
1132     bdrv_register(&bdrv_quorum);
1133 }
1134 
1135 block_init(bdrv_quorum_init);
1136