xref: /qemu/block/quorum.c (revision 65a109ab)
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 "qemu/option.h"
19 #include "block/block_int.h"
20 #include "block/qdict.h"
21 #include "qapi/error.h"
22 #include "qapi/qapi-events-block.h"
23 #include "qapi/qmp/qdict.h"
24 #include "qapi/qmp/qerror.h"
25 #include "qapi/qmp/qlist.h"
26 #include "qapi/qmp/qstring.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     int flags;
120 
121     QEMUIOVector *qiov;         /* calling IOV */
122 
123     QuorumChildRequest *qcrs;   /* individual child requests */
124     int count;                  /* number of completed AIOCB */
125     int success_count;          /* number of successfully completed AIOCB */
126 
127     int rewrite_count;          /* number of replica to rewrite: count down to
128                                  * zero once writes are fired
129                                  */
130 
131     QuorumVotes votes;
132 
133     bool is_read;
134     int vote_ret;
135     int children_read;          /* how many children have been read from */
136 };
137 
138 typedef struct QuorumCo {
139     QuorumAIOCB *acb;
140     int idx;
141 } QuorumCo;
142 
143 static void quorum_aio_finalize(QuorumAIOCB *acb)
144 {
145     g_free(acb->qcrs);
146     g_free(acb);
147 }
148 
149 static bool quorum_sha256_compare(QuorumVoteValue *a, QuorumVoteValue *b)
150 {
151     return !memcmp(a->h, b->h, HASH_LENGTH);
152 }
153 
154 static bool quorum_64bits_compare(QuorumVoteValue *a, QuorumVoteValue *b)
155 {
156     return a->l == b->l;
157 }
158 
159 static QuorumAIOCB *quorum_aio_get(BlockDriverState *bs,
160                                    QEMUIOVector *qiov,
161                                    uint64_t offset,
162                                    uint64_t bytes,
163                                    int flags)
164 {
165     BDRVQuorumState *s = bs->opaque;
166     QuorumAIOCB *acb = g_new(QuorumAIOCB, 1);
167     int i;
168 
169     *acb = (QuorumAIOCB) {
170         .co                 = qemu_coroutine_self(),
171         .bs                 = bs,
172         .offset             = offset,
173         .bytes              = bytes,
174         .flags              = flags,
175         .qiov               = qiov,
176         .votes.compare      = quorum_sha256_compare,
177         .votes.vote_list    = QLIST_HEAD_INITIALIZER(acb.votes.vote_list),
178     };
179 
180     acb->qcrs = g_new0(QuorumChildRequest, s->num_children);
181     for (i = 0; i < s->num_children; i++) {
182         acb->qcrs[i].buf = NULL;
183         acb->qcrs[i].ret = 0;
184         acb->qcrs[i].parent = acb;
185     }
186 
187     return acb;
188 }
189 
190 static void quorum_report_bad(QuorumOpType type, uint64_t offset,
191                               uint64_t bytes, char *node_name, int ret)
192 {
193     const char *msg = NULL;
194     int64_t start_sector = offset / BDRV_SECTOR_SIZE;
195     int64_t end_sector = DIV_ROUND_UP(offset + bytes, BDRV_SECTOR_SIZE);
196 
197     if (ret < 0) {
198         msg = strerror(-ret);
199     }
200 
201     qapi_event_send_quorum_report_bad(type, !!msg, msg, node_name, start_sector,
202                                       end_sector - start_sector);
203 }
204 
205 static void quorum_report_failure(QuorumAIOCB *acb)
206 {
207     const char *reference = bdrv_get_device_or_node_name(acb->bs);
208     int64_t start_sector = acb->offset / BDRV_SECTOR_SIZE;
209     int64_t end_sector = DIV_ROUND_UP(acb->offset + acb->bytes,
210                                       BDRV_SECTOR_SIZE);
211 
212     qapi_event_send_quorum_failure(reference, start_sector,
213                                    end_sector - start_sector);
214 }
215 
216 static int quorum_vote_error(QuorumAIOCB *acb);
217 
218 static bool quorum_has_too_much_io_failed(QuorumAIOCB *acb)
219 {
220     BDRVQuorumState *s = acb->bs->opaque;
221 
222     if (acb->success_count < s->threshold) {
223         acb->vote_ret = quorum_vote_error(acb);
224         quorum_report_failure(acb);
225         return true;
226     }
227 
228     return false;
229 }
230 
231 static int read_fifo_child(QuorumAIOCB *acb);
232 
233 static void quorum_copy_qiov(QEMUIOVector *dest, QEMUIOVector *source)
234 {
235     int i;
236     assert(dest->niov == source->niov);
237     assert(dest->size == source->size);
238     for (i = 0; i < source->niov; i++) {
239         assert(dest->iov[i].iov_len == source->iov[i].iov_len);
240         memcpy(dest->iov[i].iov_base,
241                source->iov[i].iov_base,
242                source->iov[i].iov_len);
243     }
244 }
245 
246 static void quorum_report_bad_acb(QuorumChildRequest *sacb, int ret)
247 {
248     QuorumAIOCB *acb = sacb->parent;
249     QuorumOpType type = acb->is_read ? QUORUM_OP_TYPE_READ : QUORUM_OP_TYPE_WRITE;
250     quorum_report_bad(type, acb->offset, acb->bytes, sacb->bs->node_name, ret);
251 }
252 
253 static void quorum_report_bad_versions(BDRVQuorumState *s,
254                                        QuorumAIOCB *acb,
255                                        QuorumVoteValue *value)
256 {
257     QuorumVoteVersion *version;
258     QuorumVoteItem *item;
259 
260     QLIST_FOREACH(version, &acb->votes.vote_list, next) {
261         if (acb->votes.compare(&version->value, value)) {
262             continue;
263         }
264         QLIST_FOREACH(item, &version->items, next) {
265             quorum_report_bad(QUORUM_OP_TYPE_READ, acb->offset, acb->bytes,
266                               s->children[item->index]->bs->node_name, 0);
267         }
268     }
269 }
270 
271 static void quorum_rewrite_entry(void *opaque)
272 {
273     QuorumCo *co = opaque;
274     QuorumAIOCB *acb = co->acb;
275     BDRVQuorumState *s = acb->bs->opaque;
276 
277     /* Ignore any errors, it's just a correction attempt for already
278      * corrupted data.
279      * Mask out BDRV_REQ_WRITE_UNCHANGED because this overwrites the
280      * area with different data from the other children. */
281     bdrv_co_pwritev(s->children[co->idx], acb->offset, acb->bytes,
282                     acb->qiov, acb->flags & ~BDRV_REQ_WRITE_UNCHANGED);
283 
284     /* Wake up the caller after the last rewrite */
285     acb->rewrite_count--;
286     if (!acb->rewrite_count) {
287         qemu_coroutine_enter_if_inactive(acb->co);
288     }
289 }
290 
291 static bool quorum_rewrite_bad_versions(QuorumAIOCB *acb,
292                                         QuorumVoteValue *value)
293 {
294     QuorumVoteVersion *version;
295     QuorumVoteItem *item;
296     int count = 0;
297 
298     /* first count the number of bad versions: done first to avoid concurrency
299      * issues.
300      */
301     QLIST_FOREACH(version, &acb->votes.vote_list, next) {
302         if (acb->votes.compare(&version->value, value)) {
303             continue;
304         }
305         QLIST_FOREACH(item, &version->items, next) {
306             count++;
307         }
308     }
309 
310     /* quorum_rewrite_entry will count down this to zero */
311     acb->rewrite_count = count;
312 
313     /* now fire the correcting rewrites */
314     QLIST_FOREACH(version, &acb->votes.vote_list, next) {
315         if (acb->votes.compare(&version->value, value)) {
316             continue;
317         }
318         QLIST_FOREACH(item, &version->items, next) {
319             Coroutine *co;
320             QuorumCo data = {
321                 .acb = acb,
322                 .idx = item->index,
323             };
324 
325             co = qemu_coroutine_create(quorum_rewrite_entry, &data);
326             qemu_coroutine_enter(co);
327         }
328     }
329 
330     /* return true if any rewrite is done else false */
331     return count;
332 }
333 
334 static void quorum_count_vote(QuorumVotes *votes,
335                               QuorumVoteValue *value,
336                               int index)
337 {
338     QuorumVoteVersion *v = NULL, *version = NULL;
339     QuorumVoteItem *item;
340 
341     /* look if we have something with this hash */
342     QLIST_FOREACH(v, &votes->vote_list, next) {
343         if (votes->compare(&v->value, value)) {
344             version = v;
345             break;
346         }
347     }
348 
349     /* It's a version not yet in the list add it */
350     if (!version) {
351         version = g_new0(QuorumVoteVersion, 1);
352         QLIST_INIT(&version->items);
353         memcpy(&version->value, value, sizeof(version->value));
354         version->index = index;
355         version->vote_count = 0;
356         QLIST_INSERT_HEAD(&votes->vote_list, version, next);
357     }
358 
359     version->vote_count++;
360 
361     item = g_new0(QuorumVoteItem, 1);
362     item->index = index;
363     QLIST_INSERT_HEAD(&version->items, item, next);
364 }
365 
366 static void quorum_free_vote_list(QuorumVotes *votes)
367 {
368     QuorumVoteVersion *version, *next_version;
369     QuorumVoteItem *item, *next_item;
370 
371     QLIST_FOREACH_SAFE(version, &votes->vote_list, next, next_version) {
372         QLIST_REMOVE(version, next);
373         QLIST_FOREACH_SAFE(item, &version->items, next, next_item) {
374             QLIST_REMOVE(item, next);
375             g_free(item);
376         }
377         g_free(version);
378     }
379 }
380 
381 static int quorum_compute_hash(QuorumAIOCB *acb, int i, QuorumVoteValue *hash)
382 {
383     QEMUIOVector *qiov = &acb->qcrs[i].qiov;
384     size_t len = sizeof(hash->h);
385     uint8_t *data = hash->h;
386 
387     /* XXX - would be nice if we could pass in the Error **
388      * and propagate that back, but this quorum code is
389      * restricted to just errno values currently */
390     if (qcrypto_hash_bytesv(QCRYPTO_HASH_ALG_SHA256,
391                             qiov->iov, qiov->niov,
392                             &data, &len,
393                             NULL) < 0) {
394         return -EINVAL;
395     }
396 
397     return 0;
398 }
399 
400 static QuorumVoteVersion *quorum_get_vote_winner(QuorumVotes *votes)
401 {
402     int max = 0;
403     QuorumVoteVersion *candidate, *winner = NULL;
404 
405     QLIST_FOREACH(candidate, &votes->vote_list, next) {
406         if (candidate->vote_count > max) {
407             max = candidate->vote_count;
408             winner = candidate;
409         }
410     }
411 
412     return winner;
413 }
414 
415 /* qemu_iovec_compare is handy for blkverify mode because it returns the first
416  * differing byte location. Yet it is handcoded to compare vectors one byte
417  * after another so it does not benefit from the libc SIMD optimizations.
418  * quorum_iovec_compare is written for speed and should be used in the non
419  * blkverify mode of quorum.
420  */
421 static bool quorum_iovec_compare(QEMUIOVector *a, QEMUIOVector *b)
422 {
423     int i;
424     int result;
425 
426     assert(a->niov == b->niov);
427     for (i = 0; i < a->niov; i++) {
428         assert(a->iov[i].iov_len == b->iov[i].iov_len);
429         result = memcmp(a->iov[i].iov_base,
430                         b->iov[i].iov_base,
431                         a->iov[i].iov_len);
432         if (result) {
433             return false;
434         }
435     }
436 
437     return true;
438 }
439 
440 static bool quorum_compare(QuorumAIOCB *acb, QEMUIOVector *a, QEMUIOVector *b)
441 {
442     BDRVQuorumState *s = acb->bs->opaque;
443     ssize_t offset;
444 
445     /* This driver will replace blkverify in this particular case */
446     if (s->is_blkverify) {
447         offset = qemu_iovec_compare(a, b);
448         if (offset != -1) {
449             fprintf(stderr, "quorum: offset=%" PRIu64 " bytes=%" PRIu64
450                     " contents mismatch at offset %" PRIu64 "\n",
451                     acb->offset, acb->bytes, acb->offset + offset);
452             exit(1);
453         }
454         return true;
455     }
456 
457     return quorum_iovec_compare(a, b);
458 }
459 
460 /* Do a vote to get the error code */
461 static int quorum_vote_error(QuorumAIOCB *acb)
462 {
463     BDRVQuorumState *s = acb->bs->opaque;
464     QuorumVoteVersion *winner = NULL;
465     QuorumVotes error_votes;
466     QuorumVoteValue result_value;
467     int i, ret = 0;
468     bool error = false;
469 
470     QLIST_INIT(&error_votes.vote_list);
471     error_votes.compare = quorum_64bits_compare;
472 
473     for (i = 0; i < s->num_children; i++) {
474         ret = acb->qcrs[i].ret;
475         if (ret) {
476             error = true;
477             result_value.l = ret;
478             quorum_count_vote(&error_votes, &result_value, i);
479         }
480     }
481 
482     if (error) {
483         winner = quorum_get_vote_winner(&error_votes);
484         ret = winner->value.l;
485     }
486 
487     quorum_free_vote_list(&error_votes);
488 
489     return ret;
490 }
491 
492 static void quorum_vote(QuorumAIOCB *acb)
493 {
494     bool quorum = true;
495     int i, j, ret;
496     QuorumVoteValue hash;
497     BDRVQuorumState *s = acb->bs->opaque;
498     QuorumVoteVersion *winner;
499 
500     if (quorum_has_too_much_io_failed(acb)) {
501         return;
502     }
503 
504     /* get the index of the first successful read */
505     for (i = 0; i < s->num_children; i++) {
506         if (!acb->qcrs[i].ret) {
507             break;
508         }
509     }
510 
511     assert(i < s->num_children);
512 
513     /* compare this read with all other successful reads stopping at quorum
514      * failure
515      */
516     for (j = i + 1; j < s->num_children; j++) {
517         if (acb->qcrs[j].ret) {
518             continue;
519         }
520         quorum = quorum_compare(acb, &acb->qcrs[i].qiov, &acb->qcrs[j].qiov);
521         if (!quorum) {
522             break;
523        }
524     }
525 
526     /* Every successful read agrees */
527     if (quorum) {
528         quorum_copy_qiov(acb->qiov, &acb->qcrs[i].qiov);
529         return;
530     }
531 
532     /* compute hashes for each successful read, also store indexes */
533     for (i = 0; i < s->num_children; i++) {
534         if (acb->qcrs[i].ret) {
535             continue;
536         }
537         ret = quorum_compute_hash(acb, i, &hash);
538         /* if ever the hash computation failed */
539         if (ret < 0) {
540             acb->vote_ret = ret;
541             goto free_exit;
542         }
543         quorum_count_vote(&acb->votes, &hash, i);
544     }
545 
546     /* vote to select the most represented version */
547     winner = quorum_get_vote_winner(&acb->votes);
548 
549     /* if the winner count is smaller than threshold the read fails */
550     if (winner->vote_count < s->threshold) {
551         quorum_report_failure(acb);
552         acb->vote_ret = -EIO;
553         goto free_exit;
554     }
555 
556     /* we have a winner: copy it */
557     quorum_copy_qiov(acb->qiov, &acb->qcrs[winner->index].qiov);
558 
559     /* some versions are bad print them */
560     quorum_report_bad_versions(s, acb, &winner->value);
561 
562     /* corruption correction is enabled */
563     if (s->rewrite_corrupted) {
564         quorum_rewrite_bad_versions(acb, &winner->value);
565     }
566 
567 free_exit:
568     /* free lists */
569     quorum_free_vote_list(&acb->votes);
570 }
571 
572 static void read_quorum_children_entry(void *opaque)
573 {
574     QuorumCo *co = opaque;
575     QuorumAIOCB *acb = co->acb;
576     BDRVQuorumState *s = acb->bs->opaque;
577     int i = co->idx;
578     QuorumChildRequest *sacb = &acb->qcrs[i];
579 
580     sacb->bs = s->children[i]->bs;
581     sacb->ret = bdrv_co_preadv(s->children[i], acb->offset, acb->bytes,
582                                &acb->qcrs[i].qiov, 0);
583 
584     if (sacb->ret == 0) {
585         acb->success_count++;
586     } else {
587         quorum_report_bad_acb(sacb, sacb->ret);
588     }
589 
590     acb->count++;
591     assert(acb->count <= s->num_children);
592     assert(acb->success_count <= s->num_children);
593 
594     /* Wake up the caller after the last read */
595     if (acb->count == s->num_children) {
596         qemu_coroutine_enter_if_inactive(acb->co);
597     }
598 }
599 
600 static int read_quorum_children(QuorumAIOCB *acb)
601 {
602     BDRVQuorumState *s = acb->bs->opaque;
603     int i;
604 
605     acb->children_read = s->num_children;
606     for (i = 0; i < s->num_children; i++) {
607         acb->qcrs[i].buf = qemu_blockalign(s->children[i]->bs, acb->qiov->size);
608         qemu_iovec_init(&acb->qcrs[i].qiov, acb->qiov->niov);
609         qemu_iovec_clone(&acb->qcrs[i].qiov, acb->qiov, acb->qcrs[i].buf);
610     }
611 
612     for (i = 0; i < s->num_children; i++) {
613         Coroutine *co;
614         QuorumCo data = {
615             .acb = acb,
616             .idx = i,
617         };
618 
619         co = qemu_coroutine_create(read_quorum_children_entry, &data);
620         qemu_coroutine_enter(co);
621     }
622 
623     while (acb->count < s->num_children) {
624         qemu_coroutine_yield();
625     }
626 
627     /* Do the vote on read */
628     quorum_vote(acb);
629     for (i = 0; i < s->num_children; i++) {
630         qemu_vfree(acb->qcrs[i].buf);
631         qemu_iovec_destroy(&acb->qcrs[i].qiov);
632     }
633 
634     while (acb->rewrite_count) {
635         qemu_coroutine_yield();
636     }
637 
638     return acb->vote_ret;
639 }
640 
641 static int read_fifo_child(QuorumAIOCB *acb)
642 {
643     BDRVQuorumState *s = acb->bs->opaque;
644     int n, ret;
645 
646     /* We try to read the next child in FIFO order if we failed to read */
647     do {
648         n = acb->children_read++;
649         acb->qcrs[n].bs = s->children[n]->bs;
650         ret = bdrv_co_preadv(s->children[n], acb->offset, acb->bytes,
651                              acb->qiov, 0);
652         if (ret < 0) {
653             quorum_report_bad_acb(&acb->qcrs[n], ret);
654         }
655     } while (ret < 0 && acb->children_read < s->num_children);
656 
657     /* FIXME: rewrite failed children if acb->children_read > 1? */
658 
659     return ret;
660 }
661 
662 static int quorum_co_preadv(BlockDriverState *bs, uint64_t offset,
663                             uint64_t bytes, QEMUIOVector *qiov, int flags)
664 {
665     BDRVQuorumState *s = bs->opaque;
666     QuorumAIOCB *acb = quorum_aio_get(bs, qiov, offset, bytes, flags);
667     int ret;
668 
669     acb->is_read = true;
670     acb->children_read = 0;
671 
672     if (s->read_pattern == QUORUM_READ_PATTERN_QUORUM) {
673         ret = read_quorum_children(acb);
674     } else {
675         ret = read_fifo_child(acb);
676     }
677     quorum_aio_finalize(acb);
678 
679     return ret;
680 }
681 
682 static void write_quorum_entry(void *opaque)
683 {
684     QuorumCo *co = opaque;
685     QuorumAIOCB *acb = co->acb;
686     BDRVQuorumState *s = acb->bs->opaque;
687     int i = co->idx;
688     QuorumChildRequest *sacb = &acb->qcrs[i];
689 
690     sacb->bs = s->children[i]->bs;
691     sacb->ret = bdrv_co_pwritev(s->children[i], acb->offset, acb->bytes,
692                                 acb->qiov, acb->flags);
693     if (sacb->ret == 0) {
694         acb->success_count++;
695     } else {
696         quorum_report_bad_acb(sacb, sacb->ret);
697     }
698     acb->count++;
699     assert(acb->count <= s->num_children);
700     assert(acb->success_count <= s->num_children);
701 
702     /* Wake up the caller after the last write */
703     if (acb->count == s->num_children) {
704         qemu_coroutine_enter_if_inactive(acb->co);
705     }
706 }
707 
708 static int quorum_co_pwritev(BlockDriverState *bs, uint64_t offset,
709                              uint64_t bytes, QEMUIOVector *qiov, int flags)
710 {
711     BDRVQuorumState *s = bs->opaque;
712     QuorumAIOCB *acb = quorum_aio_get(bs, qiov, offset, bytes, flags);
713     int i, ret;
714 
715     for (i = 0; i < s->num_children; i++) {
716         Coroutine *co;
717         QuorumCo data = {
718             .acb = acb,
719             .idx = i,
720         };
721 
722         co = qemu_coroutine_create(write_quorum_entry, &data);
723         qemu_coroutine_enter(co);
724     }
725 
726     while (acb->count < s->num_children) {
727         qemu_coroutine_yield();
728     }
729 
730     quorum_has_too_much_io_failed(acb);
731 
732     ret = acb->vote_ret;
733     quorum_aio_finalize(acb);
734 
735     return ret;
736 }
737 
738 static int64_t quorum_getlength(BlockDriverState *bs)
739 {
740     BDRVQuorumState *s = bs->opaque;
741     int64_t result;
742     int i;
743 
744     /* check that all file have the same length */
745     result = bdrv_getlength(s->children[0]->bs);
746     if (result < 0) {
747         return result;
748     }
749     for (i = 1; i < s->num_children; i++) {
750         int64_t value = bdrv_getlength(s->children[i]->bs);
751         if (value < 0) {
752             return value;
753         }
754         if (value != result) {
755             return -EIO;
756         }
757     }
758 
759     return result;
760 }
761 
762 static coroutine_fn int quorum_co_flush(BlockDriverState *bs)
763 {
764     BDRVQuorumState *s = bs->opaque;
765     QuorumVoteVersion *winner = NULL;
766     QuorumVotes error_votes;
767     QuorumVoteValue result_value;
768     int i;
769     int result = 0;
770     int success_count = 0;
771 
772     QLIST_INIT(&error_votes.vote_list);
773     error_votes.compare = quorum_64bits_compare;
774 
775     for (i = 0; i < s->num_children; i++) {
776         result = bdrv_co_flush(s->children[i]->bs);
777         if (result) {
778             quorum_report_bad(QUORUM_OP_TYPE_FLUSH, 0, 0,
779                               s->children[i]->bs->node_name, result);
780             result_value.l = result;
781             quorum_count_vote(&error_votes, &result_value, i);
782         } else {
783             success_count++;
784         }
785     }
786 
787     if (success_count >= s->threshold) {
788         result = 0;
789     } else {
790         winner = quorum_get_vote_winner(&error_votes);
791         result = winner->value.l;
792     }
793     quorum_free_vote_list(&error_votes);
794 
795     return result;
796 }
797 
798 static bool quorum_recurse_is_first_non_filter(BlockDriverState *bs,
799                                                BlockDriverState *candidate)
800 {
801     BDRVQuorumState *s = bs->opaque;
802     int i;
803 
804     for (i = 0; i < s->num_children; i++) {
805         bool perm = bdrv_recurse_is_first_non_filter(s->children[i]->bs,
806                                                      candidate);
807         if (perm) {
808             return true;
809         }
810     }
811 
812     return false;
813 }
814 
815 static int quorum_valid_threshold(int threshold, int num_children, Error **errp)
816 {
817 
818     if (threshold < 1) {
819         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
820                    "vote-threshold", "value >= 1");
821         return -ERANGE;
822     }
823 
824     if (threshold > num_children) {
825         error_setg(errp, "threshold may not exceed children count");
826         return -ERANGE;
827     }
828 
829     return 0;
830 }
831 
832 static QemuOptsList quorum_runtime_opts = {
833     .name = "quorum",
834     .head = QTAILQ_HEAD_INITIALIZER(quorum_runtime_opts.head),
835     .desc = {
836         {
837             .name = QUORUM_OPT_VOTE_THRESHOLD,
838             .type = QEMU_OPT_NUMBER,
839             .help = "The number of vote needed for reaching quorum",
840         },
841         {
842             .name = QUORUM_OPT_BLKVERIFY,
843             .type = QEMU_OPT_BOOL,
844             .help = "Trigger block verify mode if set",
845         },
846         {
847             .name = QUORUM_OPT_REWRITE,
848             .type = QEMU_OPT_BOOL,
849             .help = "Rewrite corrupted block on read quorum",
850         },
851         {
852             .name = QUORUM_OPT_READ_PATTERN,
853             .type = QEMU_OPT_STRING,
854             .help = "Allowed pattern: quorum, fifo. Quorum is default",
855         },
856         { /* end of list */ }
857     },
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     const char *pattern_str;
867     bool *opened;
868     int i;
869     int ret = 0;
870 
871     qdict_flatten(options);
872 
873     /* count how many different children are present */
874     s->num_children = qdict_array_entries(options, "children.");
875     if (s->num_children < 0) {
876         error_setg(&local_err, "Option children is not a valid array");
877         ret = -EINVAL;
878         goto exit;
879     }
880     if (s->num_children < 1) {
881         error_setg(&local_err,
882                    "Number of provided children must be 1 or more");
883         ret = -EINVAL;
884         goto exit;
885     }
886 
887     opts = qemu_opts_create(&quorum_runtime_opts, NULL, 0, &error_abort);
888     qemu_opts_absorb_qdict(opts, options, &local_err);
889     if (local_err) {
890         ret = -EINVAL;
891         goto exit;
892     }
893 
894     s->threshold = qemu_opt_get_number(opts, QUORUM_OPT_VOTE_THRESHOLD, 0);
895     /* and validate it against s->num_children */
896     ret = quorum_valid_threshold(s->threshold, s->num_children, &local_err);
897     if (ret < 0) {
898         goto exit;
899     }
900 
901     pattern_str = qemu_opt_get(opts, QUORUM_OPT_READ_PATTERN);
902     if (!pattern_str) {
903         ret = QUORUM_READ_PATTERN_QUORUM;
904     } else {
905         ret = qapi_enum_parse(&QuorumReadPattern_lookup, pattern_str,
906                               -EINVAL, NULL);
907     }
908     if (ret < 0) {
909         error_setg(&local_err, "Please set read-pattern as fifo or quorum");
910         goto exit;
911     }
912     s->read_pattern = ret;
913 
914     if (s->read_pattern == QUORUM_READ_PATTERN_QUORUM) {
915         s->is_blkverify = qemu_opt_get_bool(opts, QUORUM_OPT_BLKVERIFY, false);
916         if (s->is_blkverify && (s->num_children != 2 || s->threshold != 2)) {
917             error_setg(&local_err, "blkverify=on can only be set if there are "
918                        "exactly two files and vote-threshold is 2");
919             ret = -EINVAL;
920             goto exit;
921         }
922 
923         s->rewrite_corrupted = qemu_opt_get_bool(opts, QUORUM_OPT_REWRITE,
924                                                  false);
925         if (s->rewrite_corrupted && s->is_blkverify) {
926             error_setg(&local_err,
927                        "rewrite-corrupted=on cannot be used with blkverify=on");
928             ret = -EINVAL;
929             goto exit;
930         }
931     }
932 
933     /* allocate the children array */
934     s->children = g_new0(BdrvChild *, s->num_children);
935     opened = g_new0(bool, s->num_children);
936 
937     for (i = 0; i < s->num_children; i++) {
938         char indexstr[32];
939         ret = snprintf(indexstr, 32, "children.%d", i);
940         assert(ret < 32);
941 
942         s->children[i] = bdrv_open_child(NULL, options, indexstr, bs,
943                                          &child_format, false, &local_err);
944         if (local_err) {
945             ret = -EINVAL;
946             goto close_exit;
947         }
948 
949         opened[i] = true;
950     }
951     s->next_child_index = s->num_children;
952 
953     bs->supported_write_flags = BDRV_REQ_WRITE_UNCHANGED;
954 
955     g_free(opened);
956     goto exit;
957 
958 close_exit:
959     /* cleanup on error */
960     for (i = 0; i < s->num_children; i++) {
961         if (!opened[i]) {
962             continue;
963         }
964         bdrv_unref_child(bs, s->children[i]);
965     }
966     g_free(s->children);
967     g_free(opened);
968 exit:
969     qemu_opts_del(opts);
970     /* propagate error */
971     error_propagate(errp, local_err);
972     return ret;
973 }
974 
975 static void quorum_close(BlockDriverState *bs)
976 {
977     BDRVQuorumState *s = bs->opaque;
978     int i;
979 
980     for (i = 0; i < s->num_children; i++) {
981         bdrv_unref_child(bs, s->children[i]);
982     }
983 
984     g_free(s->children);
985 }
986 
987 static void quorum_add_child(BlockDriverState *bs, BlockDriverState *child_bs,
988                              Error **errp)
989 {
990     BDRVQuorumState *s = bs->opaque;
991     BdrvChild *child;
992     char indexstr[32];
993     int ret;
994 
995     if (s->is_blkverify) {
996         error_setg(errp, "Cannot add a child to a quorum in blkverify mode");
997         return;
998     }
999 
1000     assert(s->num_children <= INT_MAX / sizeof(BdrvChild *));
1001     if (s->num_children == INT_MAX / sizeof(BdrvChild *) ||
1002         s->next_child_index == UINT_MAX) {
1003         error_setg(errp, "Too many children");
1004         return;
1005     }
1006 
1007     ret = snprintf(indexstr, 32, "children.%u", s->next_child_index);
1008     if (ret < 0 || ret >= 32) {
1009         error_setg(errp, "cannot generate child name");
1010         return;
1011     }
1012     s->next_child_index++;
1013 
1014     bdrv_drained_begin(bs);
1015 
1016     /* We can safely add the child now */
1017     bdrv_ref(child_bs);
1018 
1019     child = bdrv_attach_child(bs, child_bs, indexstr, &child_format, errp);
1020     if (child == NULL) {
1021         s->next_child_index--;
1022         bdrv_unref(child_bs);
1023         goto out;
1024     }
1025     s->children = g_renew(BdrvChild *, s->children, s->num_children + 1);
1026     s->children[s->num_children++] = child;
1027 
1028 out:
1029     bdrv_drained_end(bs);
1030 }
1031 
1032 static void quorum_del_child(BlockDriverState *bs, BdrvChild *child,
1033                              Error **errp)
1034 {
1035     BDRVQuorumState *s = bs->opaque;
1036     int i;
1037 
1038     for (i = 0; i < s->num_children; i++) {
1039         if (s->children[i] == child) {
1040             break;
1041         }
1042     }
1043 
1044     /* we have checked it in bdrv_del_child() */
1045     assert(i < s->num_children);
1046 
1047     if (s->num_children <= s->threshold) {
1048         error_setg(errp,
1049             "The number of children cannot be lower than the vote threshold %d",
1050             s->threshold);
1051         return;
1052     }
1053 
1054     /* We know now that num_children > threshold, so blkverify must be false */
1055     assert(!s->is_blkverify);
1056 
1057     bdrv_drained_begin(bs);
1058 
1059     /* We can safely remove this child now */
1060     memmove(&s->children[i], &s->children[i + 1],
1061             (s->num_children - i - 1) * sizeof(BdrvChild *));
1062     s->children = g_renew(BdrvChild *, s->children, --s->num_children);
1063     bdrv_unref_child(bs, child);
1064 
1065     bdrv_drained_end(bs);
1066 }
1067 
1068 static void quorum_gather_child_options(BlockDriverState *bs, QDict *target,
1069                                         bool backing_overridden)
1070 {
1071     BDRVQuorumState *s = bs->opaque;
1072     QList *children_list;
1073     int i;
1074 
1075     /*
1076      * The generic implementation for gathering child options in
1077      * bdrv_refresh_filename() would use the names of the children
1078      * as specified for bdrv_open_child() or bdrv_attach_child(),
1079      * which is "children.%u" with %u being a value
1080      * (s->next_child_index) that is incremented each time a new child
1081      * is added (and never decremented).  Since children can be
1082      * deleted at runtime, there may be gaps in that enumeration.
1083      * When creating a new quorum BDS and specifying the children for
1084      * it through runtime options, the enumeration used there may not
1085      * have any gaps, though.
1086      *
1087      * Therefore, we have to create a new gap-less enumeration here
1088      * (which we can achieve by simply putting all of the children's
1089      * full_open_options into a QList).
1090      *
1091      * XXX: Note that there are issues with the current child option
1092      *      structure quorum uses (such as the fact that children do
1093      *      not really have unique permanent names).  Therefore, this
1094      *      is going to have to change in the future and ideally we
1095      *      want quorum to be covered by the generic implementation.
1096      */
1097 
1098     children_list = qlist_new();
1099     qdict_put(target, "children", children_list);
1100 
1101     for (i = 0; i < s->num_children; i++) {
1102         qlist_append(children_list,
1103                      qobject_ref(s->children[i]->bs->full_open_options));
1104     }
1105 }
1106 
1107 static char *quorum_dirname(BlockDriverState *bs, Error **errp)
1108 {
1109     /* In general, there are multiple BDSs with different dirnames below this
1110      * one; so there is no unique dirname we could return (unless all are equal
1111      * by chance, or there is only one). Therefore, to be consistent, just
1112      * always return NULL. */
1113     error_setg(errp, "Cannot generate a base directory for quorum nodes");
1114     return NULL;
1115 }
1116 
1117 static const char *const quorum_strong_runtime_opts[] = {
1118     QUORUM_OPT_VOTE_THRESHOLD,
1119     QUORUM_OPT_BLKVERIFY,
1120     QUORUM_OPT_REWRITE,
1121     QUORUM_OPT_READ_PATTERN,
1122 
1123     NULL
1124 };
1125 
1126 static BlockDriver bdrv_quorum = {
1127     .format_name                        = "quorum",
1128 
1129     .instance_size                      = sizeof(BDRVQuorumState),
1130 
1131     .bdrv_open                          = quorum_open,
1132     .bdrv_close                         = quorum_close,
1133     .bdrv_gather_child_options          = quorum_gather_child_options,
1134     .bdrv_dirname                       = quorum_dirname,
1135 
1136     .bdrv_co_flush_to_disk              = quorum_co_flush,
1137 
1138     .bdrv_getlength                     = quorum_getlength,
1139 
1140     .bdrv_co_preadv                     = quorum_co_preadv,
1141     .bdrv_co_pwritev                    = quorum_co_pwritev,
1142 
1143     .bdrv_add_child                     = quorum_add_child,
1144     .bdrv_del_child                     = quorum_del_child,
1145 
1146     .bdrv_child_perm                    = bdrv_filter_default_perms,
1147 
1148     .is_filter                          = true,
1149     .bdrv_recurse_is_first_non_filter   = quorum_recurse_is_first_non_filter,
1150 
1151     .strong_runtime_opts                = quorum_strong_runtime_opts,
1152 };
1153 
1154 static void bdrv_quorum_init(void)
1155 {
1156     if (!qcrypto_hash_supports(QCRYPTO_HASH_ALG_SHA256)) {
1157         /* SHA256 hash support is required for quorum device */
1158         return;
1159     }
1160     bdrv_register(&bdrv_quorum);
1161 }
1162 
1163 block_init(bdrv_quorum_init);
1164