xref: /openbmc/qemu/block/quorum.c (revision 7d87775f)
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/module.h"
19 #include "qemu/option.h"
20 #include "qemu/memalign.h"
21 #include "block/block_int.h"
22 #include "block/coroutines.h"
23 #include "block/qdict.h"
24 #include "qapi/error.h"
25 #include "qapi/qapi-events-block.h"
26 #include "qapi/qmp/qdict.h"
27 #include "qapi/qmp/qerror.h"
28 #include "qapi/qmp/qlist.h"
29 #include "qapi/qmp/qstring.h"
30 #include "crypto/hash.h"
31 
32 #define HASH_LENGTH 32
33 
34 #define INDEXSTR_LEN 32
35 
36 #define QUORUM_OPT_VOTE_THRESHOLD "vote-threshold"
37 #define QUORUM_OPT_BLKVERIFY      "blkverify"
38 #define QUORUM_OPT_REWRITE        "rewrite-corrupted"
39 #define QUORUM_OPT_READ_PATTERN   "read-pattern"
40 
41 /* This union holds a vote hash value */
42 typedef union QuorumVoteValue {
43     uint8_t h[HASH_LENGTH];    /* SHA-256 hash */
44     int64_t l;                 /* simpler 64 bits hash */
45 } QuorumVoteValue;
46 
47 /* A vote item */
48 typedef struct QuorumVoteItem {
49     int index;
50     QLIST_ENTRY(QuorumVoteItem) next;
51 } QuorumVoteItem;
52 
53 /* this structure is a vote version. A version is the set of votes sharing the
54  * same vote value.
55  * The set of votes will be tracked with the items field and its cardinality is
56  * vote_count.
57  */
58 typedef struct QuorumVoteVersion {
59     QuorumVoteValue value;
60     int index;
61     int vote_count;
62     QLIST_HEAD(, QuorumVoteItem) items;
63     QLIST_ENTRY(QuorumVoteVersion) next;
64 } QuorumVoteVersion;
65 
66 /* this structure holds a group of vote versions together */
67 typedef struct QuorumVotes {
68     QLIST_HEAD(, QuorumVoteVersion) vote_list;
69     bool (*compare)(QuorumVoteValue *a, QuorumVoteValue *b);
70 } QuorumVotes;
71 
72 /* the following structure holds the state of one quorum instance */
73 typedef struct BDRVQuorumState {
74     BdrvChild **children;  /* children BlockDriverStates */
75     int num_children;      /* children count */
76     unsigned next_child_index;  /* the index of the next child that should
77                                  * be added
78                                  */
79     int threshold;         /* if less than threshold children reads gave the
80                             * same result a quorum error occurs.
81                             */
82     bool is_blkverify;     /* true if the driver is in blkverify mode
83                             * Writes are mirrored on two children devices.
84                             * On reads the two children devices' contents are
85                             * compared and if a difference is spotted its
86                             * location is printed and the code aborts.
87                             * It is useful to debug other block drivers by
88                             * comparing them with a reference one.
89                             */
90     bool rewrite_corrupted;/* true if the driver must rewrite-on-read corrupted
91                             * block if Quorum is reached.
92                             */
93 
94     QuorumReadPattern read_pattern;
95 } BDRVQuorumState;
96 
97 typedef struct QuorumAIOCB QuorumAIOCB;
98 
99 /* Quorum will create one instance of the following structure per operation it
100  * performs on its children.
101  * So for each read/write operation coming from the upper layer there will be
102  * $children_count QuorumChildRequest.
103  */
104 typedef struct QuorumChildRequest {
105     BlockDriverState *bs;
106     QEMUIOVector qiov;
107     uint8_t *buf;
108     int ret;
109     QuorumAIOCB *parent;
110 } QuorumChildRequest;
111 
112 /* Quorum will use the following structure to track progress of each read/write
113  * operation received by the upper layer.
114  * This structure hold pointers to the QuorumChildRequest structures instances
115  * used to do operations on each children and track overall progress.
116  */
117 struct QuorumAIOCB {
118     BlockDriverState *bs;
119     Coroutine *co;
120 
121     /* Request metadata */
122     uint64_t offset;
123     uint64_t bytes;
124     int flags;
125 
126     QEMUIOVector *qiov;         /* calling IOV */
127 
128     QuorumChildRequest *qcrs;   /* individual child requests */
129     int count;                  /* number of completed AIOCB */
130     int success_count;          /* number of successfully completed AIOCB */
131 
132     int rewrite_count;          /* number of replica to rewrite: count down to
133                                  * zero once writes are fired
134                                  */
135 
136     QuorumVotes votes;
137 
138     bool is_read;
139     int vote_ret;
140     int children_read;          /* how many children have been read from */
141 };
142 
143 typedef struct QuorumCo {
144     QuorumAIOCB *acb;
145     int idx;
146 } QuorumCo;
147 
148 static void quorum_aio_finalize(QuorumAIOCB *acb)
149 {
150     g_free(acb->qcrs);
151     g_free(acb);
152 }
153 
154 static bool quorum_sha256_compare(QuorumVoteValue *a, QuorumVoteValue *b)
155 {
156     return !memcmp(a->h, b->h, HASH_LENGTH);
157 }
158 
159 static bool quorum_64bits_compare(QuorumVoteValue *a, QuorumVoteValue *b)
160 {
161     return a->l == b->l;
162 }
163 
164 static QuorumAIOCB *coroutine_fn quorum_aio_get(BlockDriverState *bs,
165                                                 QEMUIOVector *qiov,
166                                                 uint64_t offset, uint64_t bytes,
167                                                 int flags)
168 {
169     BDRVQuorumState *s = bs->opaque;
170     QuorumAIOCB *acb = g_new(QuorumAIOCB, 1);
171     int i;
172 
173     *acb = (QuorumAIOCB) {
174         .co                 = qemu_coroutine_self(),
175         .bs                 = bs,
176         .offset             = offset,
177         .bytes              = bytes,
178         .flags              = flags,
179         .qiov               = qiov,
180         .votes.compare      = quorum_sha256_compare,
181         .votes.vote_list    = QLIST_HEAD_INITIALIZER(acb.votes.vote_list),
182     };
183 
184     acb->qcrs = g_new0(QuorumChildRequest, s->num_children);
185     for (i = 0; i < s->num_children; i++) {
186         acb->qcrs[i].buf = NULL;
187         acb->qcrs[i].ret = 0;
188         acb->qcrs[i].parent = acb;
189     }
190 
191     return acb;
192 }
193 
194 static void quorum_report_bad(QuorumOpType type, uint64_t offset,
195                               uint64_t bytes, char *node_name, int ret)
196 {
197     const char *msg = NULL;
198     int64_t start_sector = offset / BDRV_SECTOR_SIZE;
199     int64_t end_sector = DIV_ROUND_UP(offset + bytes, BDRV_SECTOR_SIZE);
200 
201     if (ret < 0) {
202         msg = strerror(-ret);
203     }
204 
205     qapi_event_send_quorum_report_bad(type, msg, node_name, start_sector,
206                                       end_sector - start_sector);
207 }
208 
209 static void GRAPH_RDLOCK quorum_report_failure(QuorumAIOCB *acb)
210 {
211     const char *reference = bdrv_get_device_or_node_name(acb->bs);
212     int64_t start_sector = acb->offset / BDRV_SECTOR_SIZE;
213     int64_t end_sector = DIV_ROUND_UP(acb->offset + acb->bytes,
214                                       BDRV_SECTOR_SIZE);
215 
216     qapi_event_send_quorum_failure(reference, start_sector,
217                                    end_sector - start_sector);
218 }
219 
220 static int quorum_vote_error(QuorumAIOCB *acb);
221 
222 static bool GRAPH_RDLOCK quorum_has_too_much_io_failed(QuorumAIOCB *acb)
223 {
224     BDRVQuorumState *s = acb->bs->opaque;
225 
226     if (acb->success_count < s->threshold) {
227         acb->vote_ret = quorum_vote_error(acb);
228         quorum_report_failure(acb);
229         return true;
230     }
231 
232     return false;
233 }
234 
235 static void quorum_copy_qiov(QEMUIOVector *dest, QEMUIOVector *source)
236 {
237     int i;
238     assert(dest->niov == source->niov);
239     assert(dest->size == source->size);
240     for (i = 0; i < source->niov; i++) {
241         assert(dest->iov[i].iov_len == source->iov[i].iov_len);
242         memcpy(dest->iov[i].iov_base,
243                source->iov[i].iov_base,
244                source->iov[i].iov_len);
245     }
246 }
247 
248 static void quorum_report_bad_acb(QuorumChildRequest *sacb, int ret)
249 {
250     QuorumAIOCB *acb = sacb->parent;
251     QuorumOpType type = acb->is_read ? QUORUM_OP_TYPE_READ : QUORUM_OP_TYPE_WRITE;
252     quorum_report_bad(type, acb->offset, acb->bytes, sacb->bs->node_name, ret);
253 }
254 
255 static void quorum_report_bad_versions(BDRVQuorumState *s,
256                                        QuorumAIOCB *acb,
257                                        QuorumVoteValue *value)
258 {
259     QuorumVoteVersion *version;
260     QuorumVoteItem *item;
261 
262     QLIST_FOREACH(version, &acb->votes.vote_list, next) {
263         if (acb->votes.compare(&version->value, value)) {
264             continue;
265         }
266         QLIST_FOREACH(item, &version->items, next) {
267             quorum_report_bad(QUORUM_OP_TYPE_READ, acb->offset, acb->bytes,
268                               s->children[item->index]->bs->node_name, 0);
269         }
270     }
271 }
272 
273 /*
274  * This function can count as GRAPH_RDLOCK because read_quorum_children() holds
275  * the graph lock and keeps it until this coroutine has terminated.
276  */
277 static void coroutine_fn GRAPH_RDLOCK quorum_rewrite_entry(void *opaque)
278 {
279     QuorumCo *co = opaque;
280     QuorumAIOCB *acb = co->acb;
281     BDRVQuorumState *s = acb->bs->opaque;
282 
283     /* Ignore any errors, it's just a correction attempt for already
284      * corrupted data.
285      * Mask out BDRV_REQ_WRITE_UNCHANGED because this overwrites the
286      * area with different data from the other children. */
287     bdrv_co_pwritev(s->children[co->idx], acb->offset, acb->bytes,
288                     acb->qiov, acb->flags & ~BDRV_REQ_WRITE_UNCHANGED);
289 
290     /* Wake up the caller after the last rewrite */
291     acb->rewrite_count--;
292     if (!acb->rewrite_count) {
293         qemu_coroutine_enter_if_inactive(acb->co);
294     }
295 }
296 
297 static bool coroutine_fn GRAPH_RDLOCK
298 quorum_rewrite_bad_versions(QuorumAIOCB *acb, QuorumVoteValue *value)
299 {
300     QuorumVoteVersion *version;
301     QuorumVoteItem *item;
302     int count = 0;
303 
304     /* first count the number of bad versions: done first to avoid concurrency
305      * issues.
306      */
307     QLIST_FOREACH(version, &acb->votes.vote_list, next) {
308         if (acb->votes.compare(&version->value, value)) {
309             continue;
310         }
311         QLIST_FOREACH(item, &version->items, next) {
312             count++;
313         }
314     }
315 
316     /* quorum_rewrite_entry will count down this to zero */
317     acb->rewrite_count = count;
318 
319     /* now fire the correcting rewrites */
320     QLIST_FOREACH(version, &acb->votes.vote_list, next) {
321         if (acb->votes.compare(&version->value, value)) {
322             continue;
323         }
324         QLIST_FOREACH(item, &version->items, next) {
325             Coroutine *co;
326             QuorumCo data = {
327                 .acb = acb,
328                 .idx = item->index,
329             };
330 
331             co = qemu_coroutine_create(quorum_rewrite_entry, &data);
332             qemu_coroutine_enter(co);
333         }
334     }
335 
336     /* return true if any rewrite is done else false */
337     return count;
338 }
339 
340 static void quorum_count_vote(QuorumVotes *votes,
341                               QuorumVoteValue *value,
342                               int index)
343 {
344     QuorumVoteVersion *v = NULL, *version = NULL;
345     QuorumVoteItem *item;
346 
347     /* look if we have something with this hash */
348     QLIST_FOREACH(v, &votes->vote_list, next) {
349         if (votes->compare(&v->value, value)) {
350             version = v;
351             break;
352         }
353     }
354 
355     /* It's a version not yet in the list add it */
356     if (!version) {
357         version = g_new0(QuorumVoteVersion, 1);
358         QLIST_INIT(&version->items);
359         memcpy(&version->value, value, sizeof(version->value));
360         version->index = index;
361         version->vote_count = 0;
362         QLIST_INSERT_HEAD(&votes->vote_list, version, next);
363     }
364 
365     version->vote_count++;
366 
367     item = g_new0(QuorumVoteItem, 1);
368     item->index = index;
369     QLIST_INSERT_HEAD(&version->items, item, next);
370 }
371 
372 static void quorum_free_vote_list(QuorumVotes *votes)
373 {
374     QuorumVoteVersion *version, *next_version;
375     QuorumVoteItem *item, *next_item;
376 
377     QLIST_FOREACH_SAFE(version, &votes->vote_list, next, next_version) {
378         QLIST_REMOVE(version, next);
379         QLIST_FOREACH_SAFE(item, &version->items, next, next_item) {
380             QLIST_REMOVE(item, next);
381             g_free(item);
382         }
383         g_free(version);
384     }
385 }
386 
387 static int quorum_compute_hash(QuorumAIOCB *acb, int i, QuorumVoteValue *hash)
388 {
389     QEMUIOVector *qiov = &acb->qcrs[i].qiov;
390     size_t len = sizeof(hash->h);
391     uint8_t *data = hash->h;
392 
393     /* XXX - would be nice if we could pass in the Error **
394      * and propagate that back, but this quorum code is
395      * restricted to just errno values currently */
396     if (qcrypto_hash_bytesv(QCRYPTO_HASH_ALGO_SHA256,
397                             qiov->iov, qiov->niov,
398                             &data, &len,
399                             NULL) < 0) {
400         return -EINVAL;
401     }
402 
403     return 0;
404 }
405 
406 static QuorumVoteVersion *quorum_get_vote_winner(QuorumVotes *votes)
407 {
408     int max = 0;
409     QuorumVoteVersion *candidate, *winner = NULL;
410 
411     QLIST_FOREACH(candidate, &votes->vote_list, next) {
412         if (candidate->vote_count > max) {
413             max = candidate->vote_count;
414             winner = candidate;
415         }
416     }
417 
418     return winner;
419 }
420 
421 /* qemu_iovec_compare is handy for blkverify mode because it returns the first
422  * differing byte location. Yet it is handcoded to compare vectors one byte
423  * after another so it does not benefit from the libc SIMD optimizations.
424  * quorum_iovec_compare is written for speed and should be used in the non
425  * blkverify mode of quorum.
426  */
427 static bool quorum_iovec_compare(QEMUIOVector *a, QEMUIOVector *b)
428 {
429     int i;
430     int result;
431 
432     assert(a->niov == b->niov);
433     for (i = 0; i < a->niov; i++) {
434         assert(a->iov[i].iov_len == b->iov[i].iov_len);
435         result = memcmp(a->iov[i].iov_base,
436                         b->iov[i].iov_base,
437                         a->iov[i].iov_len);
438         if (result) {
439             return false;
440         }
441     }
442 
443     return true;
444 }
445 
446 static bool quorum_compare(QuorumAIOCB *acb, QEMUIOVector *a, QEMUIOVector *b)
447 {
448     BDRVQuorumState *s = acb->bs->opaque;
449     ssize_t offset;
450 
451     /* This driver will replace blkverify in this particular case */
452     if (s->is_blkverify) {
453         offset = qemu_iovec_compare(a, b);
454         if (offset != -1) {
455             fprintf(stderr, "quorum: offset=%" PRIu64 " bytes=%" PRIu64
456                     " contents mismatch at offset %" PRIu64 "\n",
457                     acb->offset, acb->bytes, acb->offset + offset);
458             exit(1);
459         }
460         return true;
461     }
462 
463     return quorum_iovec_compare(a, b);
464 }
465 
466 /* Do a vote to get the error code */
467 static int quorum_vote_error(QuorumAIOCB *acb)
468 {
469     BDRVQuorumState *s = acb->bs->opaque;
470     QuorumVoteVersion *winner = NULL;
471     QuorumVotes error_votes;
472     QuorumVoteValue result_value;
473     int i, ret = 0;
474     bool error = false;
475 
476     QLIST_INIT(&error_votes.vote_list);
477     error_votes.compare = quorum_64bits_compare;
478 
479     for (i = 0; i < s->num_children; i++) {
480         ret = acb->qcrs[i].ret;
481         if (ret) {
482             error = true;
483             result_value.l = ret;
484             quorum_count_vote(&error_votes, &result_value, i);
485         }
486     }
487 
488     if (error) {
489         winner = quorum_get_vote_winner(&error_votes);
490         ret = winner->value.l;
491     }
492 
493     quorum_free_vote_list(&error_votes);
494 
495     return ret;
496 }
497 
498 static void coroutine_fn GRAPH_RDLOCK quorum_vote(QuorumAIOCB *acb)
499 {
500     bool quorum = true;
501     int i, j, ret;
502     QuorumVoteValue hash;
503     BDRVQuorumState *s = acb->bs->opaque;
504     QuorumVoteVersion *winner;
505 
506     if (quorum_has_too_much_io_failed(acb)) {
507         return;
508     }
509 
510     /* get the index of the first successful read */
511     for (i = 0; i < s->num_children; i++) {
512         if (!acb->qcrs[i].ret) {
513             break;
514         }
515     }
516 
517     assert(i < s->num_children);
518 
519     /* compare this read with all other successful reads stopping at quorum
520      * failure
521      */
522     for (j = i + 1; j < s->num_children; j++) {
523         if (acb->qcrs[j].ret) {
524             continue;
525         }
526         quorum = quorum_compare(acb, &acb->qcrs[i].qiov, &acb->qcrs[j].qiov);
527         if (!quorum) {
528             break;
529        }
530     }
531 
532     /* Every successful read agrees */
533     if (quorum) {
534         quorum_copy_qiov(acb->qiov, &acb->qcrs[i].qiov);
535         return;
536     }
537 
538     /* compute hashes for each successful read, also store indexes */
539     for (i = 0; i < s->num_children; i++) {
540         if (acb->qcrs[i].ret) {
541             continue;
542         }
543         ret = quorum_compute_hash(acb, i, &hash);
544         /* if ever the hash computation failed */
545         if (ret < 0) {
546             acb->vote_ret = ret;
547             goto free_exit;
548         }
549         quorum_count_vote(&acb->votes, &hash, i);
550     }
551 
552     /* vote to select the most represented version */
553     winner = quorum_get_vote_winner(&acb->votes);
554 
555     /* if the winner count is smaller than threshold the read fails */
556     if (winner->vote_count < s->threshold) {
557         quorum_report_failure(acb);
558         acb->vote_ret = -EIO;
559         goto free_exit;
560     }
561 
562     /* we have a winner: copy it */
563     quorum_copy_qiov(acb->qiov, &acb->qcrs[winner->index].qiov);
564 
565     /* some versions are bad print them */
566     quorum_report_bad_versions(s, acb, &winner->value);
567 
568     /* corruption correction is enabled */
569     if (s->rewrite_corrupted) {
570         quorum_rewrite_bad_versions(acb, &winner->value);
571     }
572 
573 free_exit:
574     /* free lists */
575     quorum_free_vote_list(&acb->votes);
576 }
577 
578 /*
579  * This function can count as GRAPH_RDLOCK because read_quorum_children() holds
580  * the graph lock and keeps it until this coroutine has terminated.
581  */
582 static void coroutine_fn GRAPH_RDLOCK read_quorum_children_entry(void *opaque)
583 {
584     QuorumCo *co = opaque;
585     QuorumAIOCB *acb = co->acb;
586     BDRVQuorumState *s = acb->bs->opaque;
587     int i = co->idx;
588     QuorumChildRequest *sacb = &acb->qcrs[i];
589 
590     sacb->bs = s->children[i]->bs;
591     sacb->ret = bdrv_co_preadv(s->children[i], acb->offset, acb->bytes,
592                                &acb->qcrs[i].qiov, 0);
593 
594     if (sacb->ret == 0) {
595         acb->success_count++;
596     } else {
597         quorum_report_bad_acb(sacb, sacb->ret);
598     }
599 
600     acb->count++;
601     assert(acb->count <= s->num_children);
602     assert(acb->success_count <= s->num_children);
603 
604     /* Wake up the caller after the last read */
605     if (acb->count == s->num_children) {
606         qemu_coroutine_enter_if_inactive(acb->co);
607     }
608 }
609 
610 static int coroutine_fn GRAPH_RDLOCK read_quorum_children(QuorumAIOCB *acb)
611 {
612     BDRVQuorumState *s = acb->bs->opaque;
613     int i;
614 
615     acb->children_read = s->num_children;
616     for (i = 0; i < s->num_children; i++) {
617         acb->qcrs[i].buf = qemu_blockalign(s->children[i]->bs, acb->qiov->size);
618         qemu_iovec_init(&acb->qcrs[i].qiov, acb->qiov->niov);
619         qemu_iovec_clone(&acb->qcrs[i].qiov, acb->qiov, acb->qcrs[i].buf);
620     }
621 
622     for (i = 0; i < s->num_children; i++) {
623         Coroutine *co;
624         QuorumCo data = {
625             .acb = acb,
626             .idx = i,
627         };
628 
629         co = qemu_coroutine_create(read_quorum_children_entry, &data);
630         qemu_coroutine_enter(co);
631     }
632 
633     while (acb->count < s->num_children) {
634         qemu_coroutine_yield();
635     }
636 
637     /* Do the vote on read */
638     quorum_vote(acb);
639     for (i = 0; i < s->num_children; i++) {
640         qemu_vfree(acb->qcrs[i].buf);
641         qemu_iovec_destroy(&acb->qcrs[i].qiov);
642     }
643 
644     while (acb->rewrite_count) {
645         qemu_coroutine_yield();
646     }
647 
648     return acb->vote_ret;
649 }
650 
651 static int coroutine_fn GRAPH_RDLOCK read_fifo_child(QuorumAIOCB *acb)
652 {
653     BDRVQuorumState *s = acb->bs->opaque;
654     int n, ret;
655 
656     /* We try to read the next child in FIFO order if we failed to read */
657     do {
658         n = acb->children_read++;
659         acb->qcrs[n].bs = s->children[n]->bs;
660         ret = bdrv_co_preadv(s->children[n], acb->offset, acb->bytes,
661                              acb->qiov, 0);
662         if (ret < 0) {
663             quorum_report_bad_acb(&acb->qcrs[n], ret);
664         }
665     } while (ret < 0 && acb->children_read < s->num_children);
666 
667     /* FIXME: rewrite failed children if acb->children_read > 1? */
668 
669     return ret;
670 }
671 
672 static int coroutine_fn GRAPH_RDLOCK
673 quorum_co_preadv(BlockDriverState *bs, int64_t offset, int64_t bytes,
674                  QEMUIOVector *qiov, BdrvRequestFlags flags)
675 {
676     BDRVQuorumState *s = bs->opaque;
677     QuorumAIOCB *acb = quorum_aio_get(bs, qiov, offset, bytes, flags);
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 /*
694  * This function can count as GRAPH_RDLOCK because quorum_co_pwritev() holds the
695  * graph lock and keeps it until this coroutine has terminated.
696  */
697 static void coroutine_fn GRAPH_RDLOCK write_quorum_entry(void *opaque)
698 {
699     QuorumCo *co = opaque;
700     QuorumAIOCB *acb = co->acb;
701     BDRVQuorumState *s = acb->bs->opaque;
702     int i = co->idx;
703     QuorumChildRequest *sacb = &acb->qcrs[i];
704 
705     sacb->bs = s->children[i]->bs;
706     if (acb->flags & BDRV_REQ_ZERO_WRITE) {
707         sacb->ret = bdrv_co_pwrite_zeroes(s->children[i], acb->offset,
708                                           acb->bytes, acb->flags);
709     } else {
710         sacb->ret = bdrv_co_pwritev(s->children[i], acb->offset, acb->bytes,
711                                     acb->qiov, acb->flags);
712     }
713     if (sacb->ret == 0) {
714         acb->success_count++;
715     } else {
716         quorum_report_bad_acb(sacb, sacb->ret);
717     }
718     acb->count++;
719     assert(acb->count <= s->num_children);
720     assert(acb->success_count <= s->num_children);
721 
722     /* Wake up the caller after the last write */
723     if (acb->count == s->num_children) {
724         qemu_coroutine_enter_if_inactive(acb->co);
725     }
726 }
727 
728 static int coroutine_fn GRAPH_RDLOCK
729 quorum_co_pwritev(BlockDriverState *bs, int64_t offset, int64_t bytes,
730                   QEMUIOVector *qiov, BdrvRequestFlags flags)
731 {
732     BDRVQuorumState *s = bs->opaque;
733     QuorumAIOCB *acb = quorum_aio_get(bs, qiov, offset, bytes, flags);
734     int i, ret;
735 
736     for (i = 0; i < s->num_children; i++) {
737         Coroutine *co;
738         QuorumCo data = {
739             .acb = acb,
740             .idx = i,
741         };
742 
743         co = qemu_coroutine_create(write_quorum_entry, &data);
744         qemu_coroutine_enter(co);
745     }
746 
747     while (acb->count < s->num_children) {
748         qemu_coroutine_yield();
749     }
750 
751     quorum_has_too_much_io_failed(acb);
752 
753     ret = acb->vote_ret;
754     quorum_aio_finalize(acb);
755 
756     return ret;
757 }
758 
759 static int coroutine_fn GRAPH_RDLOCK
760 quorum_co_pwrite_zeroes(BlockDriverState *bs, int64_t offset, int64_t bytes,
761                         BdrvRequestFlags flags)
762 {
763     return quorum_co_pwritev(bs, offset, bytes, NULL,
764                              flags | BDRV_REQ_ZERO_WRITE);
765 }
766 
767 static int64_t coroutine_fn GRAPH_RDLOCK
768 quorum_co_getlength(BlockDriverState *bs)
769 {
770     BDRVQuorumState *s = bs->opaque;
771     int64_t result;
772     int i;
773 
774     /* check that all file have the same length */
775     result = bdrv_co_getlength(s->children[0]->bs);
776     if (result < 0) {
777         return result;
778     }
779     for (i = 1; i < s->num_children; i++) {
780         int64_t value = bdrv_co_getlength(s->children[i]->bs);
781         if (value < 0) {
782             return value;
783         }
784         if (value != result) {
785             return -EIO;
786         }
787     }
788 
789     return result;
790 }
791 
792 static coroutine_fn GRAPH_RDLOCK int quorum_co_flush(BlockDriverState *bs)
793 {
794     BDRVQuorumState *s = bs->opaque;
795     QuorumVoteVersion *winner = NULL;
796     QuorumVotes error_votes;
797     QuorumVoteValue result_value;
798     int i;
799     int result = 0;
800     int success_count = 0;
801 
802     QLIST_INIT(&error_votes.vote_list);
803     error_votes.compare = quorum_64bits_compare;
804 
805     for (i = 0; i < s->num_children; i++) {
806         result = bdrv_co_flush(s->children[i]->bs);
807         if (result) {
808             quorum_report_bad(QUORUM_OP_TYPE_FLUSH, 0, 0,
809                               s->children[i]->bs->node_name, result);
810             result_value.l = result;
811             quorum_count_vote(&error_votes, &result_value, i);
812         } else {
813             success_count++;
814         }
815     }
816 
817     if (success_count >= s->threshold) {
818         result = 0;
819     } else {
820         winner = quorum_get_vote_winner(&error_votes);
821         result = winner->value.l;
822     }
823     quorum_free_vote_list(&error_votes);
824 
825     return result;
826 }
827 
828 static bool GRAPH_RDLOCK
829 quorum_recurse_can_replace(BlockDriverState *bs, BlockDriverState *to_replace)
830 {
831     BDRVQuorumState *s = bs->opaque;
832     int i;
833 
834     for (i = 0; i < s->num_children; i++) {
835         /*
836          * We have no idea whether our children show the same data as
837          * this node (@bs).  It is actually highly likely that
838          * @to_replace does not, because replacing a broken child is
839          * one of the main use cases here.
840          *
841          * We do know that the new BDS will match @bs, so replacing
842          * any of our children by it will be safe.  It cannot change
843          * the data this quorum node presents to its parents.
844          *
845          * However, replacing @to_replace by @bs in any of our
846          * children's chains may change visible data somewhere in
847          * there.  We therefore cannot recurse down those chains with
848          * bdrv_recurse_can_replace().
849          * (More formally, bdrv_recurse_can_replace() requires that
850          * @to_replace will be replaced by something matching the @bs
851          * passed to it.  We cannot guarantee that.)
852          *
853          * Thus, we can only check whether any of our immediate
854          * children matches @to_replace.
855          *
856          * (In the future, we might add a function to recurse down a
857          * chain that checks that nothing there cares about a change
858          * in data from the respective child in question.  For
859          * example, most filters do not care when their child's data
860          * suddenly changes, as long as their parents do not care.)
861          */
862         if (s->children[i]->bs == to_replace) {
863             /*
864              * We now have to ensure that there is no other parent
865              * that cares about replacing this child by a node with
866              * potentially different data.
867              * We do so by checking whether there are any other parents
868              * at all, which is stricter than necessary, but also very
869              * simple.  (We may decide to implement something more
870              * complex and permissive when there is an actual need for
871              * it.)
872              */
873             return QLIST_FIRST(&to_replace->parents) == s->children[i] &&
874                 QLIST_NEXT(s->children[i], next_parent) == NULL;
875         }
876     }
877 
878     return false;
879 }
880 
881 static int quorum_valid_threshold(int threshold, int num_children, Error **errp)
882 {
883 
884     if (threshold < 1) {
885         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
886                    "vote-threshold", "a value >= 1");
887         return -ERANGE;
888     }
889 
890     if (threshold > num_children) {
891         error_setg(errp, "threshold may not exceed children count");
892         return -ERANGE;
893     }
894 
895     return 0;
896 }
897 
898 static QemuOptsList quorum_runtime_opts = {
899     .name = "quorum",
900     .head = QTAILQ_HEAD_INITIALIZER(quorum_runtime_opts.head),
901     .desc = {
902         {
903             .name = QUORUM_OPT_VOTE_THRESHOLD,
904             .type = QEMU_OPT_NUMBER,
905             .help = "The number of vote needed for reaching quorum",
906         },
907         {
908             .name = QUORUM_OPT_BLKVERIFY,
909             .type = QEMU_OPT_BOOL,
910             .help = "Trigger block verify mode if set",
911         },
912         {
913             .name = QUORUM_OPT_REWRITE,
914             .type = QEMU_OPT_BOOL,
915             .help = "Rewrite corrupted block on read quorum",
916         },
917         {
918             .name = QUORUM_OPT_READ_PATTERN,
919             .type = QEMU_OPT_STRING,
920             .help = "Allowed pattern: quorum, fifo. Quorum is default",
921         },
922         { /* end of list */ }
923     },
924 };
925 
926 static void quorum_refresh_flags(BlockDriverState *bs)
927 {
928     BDRVQuorumState *s = bs->opaque;
929     int i;
930 
931     bs->supported_zero_flags =
932         BDRV_REQ_FUA | BDRV_REQ_MAY_UNMAP | BDRV_REQ_NO_FALLBACK;
933 
934     for (i = 0; i < s->num_children; i++) {
935         bs->supported_zero_flags &= s->children[i]->bs->supported_zero_flags;
936     }
937 
938     bs->supported_zero_flags |= BDRV_REQ_WRITE_UNCHANGED;
939 }
940 
941 static int quorum_open(BlockDriverState *bs, QDict *options, int flags,
942                        Error **errp)
943 {
944     BDRVQuorumState *s = bs->opaque;
945     QemuOpts *opts = NULL;
946     const char *pattern_str;
947     bool *opened;
948     int i;
949     int ret = 0;
950 
951     qdict_flatten(options);
952 
953     /* count how many different children are present */
954     s->num_children = qdict_array_entries(options, "children.");
955     if (s->num_children < 0) {
956         error_setg(errp, "Option children is not a valid array");
957         ret = -EINVAL;
958         goto exit;
959     }
960     if (s->num_children < 1) {
961         error_setg(errp, "Number of provided children must be 1 or more");
962         ret = -EINVAL;
963         goto exit;
964     }
965 
966     opts = qemu_opts_create(&quorum_runtime_opts, NULL, 0, &error_abort);
967     if (!qemu_opts_absorb_qdict(opts, options, errp)) {
968         ret = -EINVAL;
969         goto exit;
970     }
971 
972     s->threshold = qemu_opt_get_number(opts, QUORUM_OPT_VOTE_THRESHOLD, 0);
973     /* and validate it against s->num_children */
974     ret = quorum_valid_threshold(s->threshold, s->num_children, errp);
975     if (ret < 0) {
976         goto exit;
977     }
978 
979     pattern_str = qemu_opt_get(opts, QUORUM_OPT_READ_PATTERN);
980     if (!pattern_str) {
981         ret = QUORUM_READ_PATTERN_QUORUM;
982     } else {
983         ret = qapi_enum_parse(&QuorumReadPattern_lookup, pattern_str,
984                               -EINVAL, NULL);
985     }
986     if (ret < 0) {
987         error_setg(errp, "Please set read-pattern as fifo or quorum");
988         goto exit;
989     }
990     s->read_pattern = ret;
991 
992     if (s->read_pattern == QUORUM_READ_PATTERN_QUORUM) {
993         s->is_blkverify = qemu_opt_get_bool(opts, QUORUM_OPT_BLKVERIFY, false);
994         if (s->is_blkverify && (s->num_children != 2 || s->threshold != 2)) {
995             error_setg(errp, "blkverify=on can only be set if there are "
996                        "exactly two files and vote-threshold is 2");
997             ret = -EINVAL;
998             goto exit;
999         }
1000 
1001         s->rewrite_corrupted = qemu_opt_get_bool(opts, QUORUM_OPT_REWRITE,
1002                                                  false);
1003         if (s->rewrite_corrupted && s->is_blkverify) {
1004             error_setg(errp,
1005                        "rewrite-corrupted=on cannot be used with blkverify=on");
1006             ret = -EINVAL;
1007             goto exit;
1008         }
1009     }
1010 
1011     /* allocate the children array */
1012     s->children = g_new0(BdrvChild *, s->num_children);
1013     opened = g_new0(bool, s->num_children);
1014 
1015     for (i = 0; i < s->num_children; i++) {
1016         char indexstr[INDEXSTR_LEN];
1017         ret = snprintf(indexstr, INDEXSTR_LEN, "children.%d", i);
1018         assert(ret < INDEXSTR_LEN);
1019 
1020         s->children[i] = bdrv_open_child(NULL, options, indexstr, bs,
1021                                          &child_of_bds, BDRV_CHILD_DATA, false,
1022                                          errp);
1023         if (!s->children[i]) {
1024             ret = -EINVAL;
1025             goto close_exit;
1026         }
1027 
1028         opened[i] = true;
1029     }
1030     s->next_child_index = s->num_children;
1031 
1032     bs->supported_write_flags = BDRV_REQ_WRITE_UNCHANGED;
1033     quorum_refresh_flags(bs);
1034 
1035     g_free(opened);
1036     goto exit;
1037 
1038 close_exit:
1039     /* cleanup on error */
1040     bdrv_graph_wrlock();
1041     for (i = 0; i < s->num_children; i++) {
1042         if (!opened[i]) {
1043             continue;
1044         }
1045         bdrv_unref_child(bs, s->children[i]);
1046     }
1047     bdrv_graph_wrunlock();
1048     g_free(s->children);
1049     g_free(opened);
1050 exit:
1051     qemu_opts_del(opts);
1052     return ret;
1053 }
1054 
1055 static void quorum_close(BlockDriverState *bs)
1056 {
1057     BDRVQuorumState *s = bs->opaque;
1058     int i;
1059 
1060     bdrv_graph_wrlock();
1061     for (i = 0; i < s->num_children; i++) {
1062         bdrv_unref_child(bs, s->children[i]);
1063     }
1064     bdrv_graph_wrunlock();
1065 
1066     g_free(s->children);
1067 }
1068 
1069 static void GRAPH_WRLOCK
1070 quorum_add_child(BlockDriverState *bs, BlockDriverState *child_bs, Error **errp)
1071 {
1072     BDRVQuorumState *s = bs->opaque;
1073     BdrvChild *child;
1074     char indexstr[INDEXSTR_LEN];
1075     int ret;
1076 
1077     if (s->is_blkverify) {
1078         error_setg(errp, "Cannot add a child to a quorum in blkverify mode");
1079         return;
1080     }
1081 
1082     assert(s->num_children <= INT_MAX / sizeof(BdrvChild *));
1083     if (s->num_children == INT_MAX / sizeof(BdrvChild *) ||
1084         s->next_child_index == UINT_MAX) {
1085         error_setg(errp, "Too many children");
1086         return;
1087     }
1088 
1089     ret = snprintf(indexstr, INDEXSTR_LEN, "children.%u", s->next_child_index);
1090     if (ret < 0 || ret >= INDEXSTR_LEN) {
1091         error_setg(errp, "cannot generate child name");
1092         return;
1093     }
1094     s->next_child_index++;
1095 
1096     /* We can safely add the child now */
1097     bdrv_ref(child_bs);
1098 
1099     child = bdrv_attach_child(bs, child_bs, indexstr, &child_of_bds,
1100                               BDRV_CHILD_DATA, errp);
1101     if (child == NULL) {
1102         s->next_child_index--;
1103         return;
1104     }
1105     s->children = g_renew(BdrvChild *, s->children, s->num_children + 1);
1106     s->children[s->num_children++] = child;
1107     quorum_refresh_flags(bs);
1108 }
1109 
1110 static void GRAPH_WRLOCK
1111 quorum_del_child(BlockDriverState *bs, BdrvChild *child, Error **errp)
1112 {
1113     BDRVQuorumState *s = bs->opaque;
1114     char indexstr[INDEXSTR_LEN];
1115     int i;
1116 
1117     for (i = 0; i < s->num_children; i++) {
1118         if (s->children[i] == child) {
1119             break;
1120         }
1121     }
1122 
1123     /* we have checked it in bdrv_del_child() */
1124     assert(i < s->num_children);
1125 
1126     if (s->num_children <= s->threshold) {
1127         error_setg(errp,
1128             "The number of children cannot be lower than the vote threshold %d",
1129             s->threshold);
1130         return;
1131     }
1132 
1133     /* We know now that num_children > threshold, so blkverify must be false */
1134     assert(!s->is_blkverify);
1135 
1136     snprintf(indexstr, INDEXSTR_LEN, "children.%u", s->next_child_index - 1);
1137     if (!strncmp(child->name, indexstr, INDEXSTR_LEN)) {
1138         s->next_child_index--;
1139     }
1140 
1141     /* We can safely remove this child now */
1142     memmove(&s->children[i], &s->children[i + 1],
1143             (s->num_children - i - 1) * sizeof(BdrvChild *));
1144     s->children = g_renew(BdrvChild *, s->children, --s->num_children);
1145 
1146     bdrv_unref_child(bs, child);
1147 
1148     quorum_refresh_flags(bs);
1149 }
1150 
1151 static void quorum_gather_child_options(BlockDriverState *bs, QDict *target,
1152                                         bool backing_overridden)
1153 {
1154     BDRVQuorumState *s = bs->opaque;
1155     QList *children_list;
1156     int i;
1157 
1158     /*
1159      * The generic implementation for gathering child options in
1160      * bdrv_refresh_filename() would use the names of the children
1161      * as specified for bdrv_open_child() or bdrv_attach_child(),
1162      * which is "children.%u" with %u being a value
1163      * (s->next_child_index) that is incremented each time a new child
1164      * is added (and never decremented).  Since children can be
1165      * deleted at runtime, there may be gaps in that enumeration.
1166      * When creating a new quorum BDS and specifying the children for
1167      * it through runtime options, the enumeration used there may not
1168      * have any gaps, though.
1169      *
1170      * Therefore, we have to create a new gap-less enumeration here
1171      * (which we can achieve by simply putting all of the children's
1172      * full_open_options into a QList).
1173      *
1174      * XXX: Note that there are issues with the current child option
1175      *      structure quorum uses (such as the fact that children do
1176      *      not really have unique permanent names).  Therefore, this
1177      *      is going to have to change in the future and ideally we
1178      *      want quorum to be covered by the generic implementation.
1179      */
1180 
1181     children_list = qlist_new();
1182     qdict_put(target, "children", children_list);
1183 
1184     for (i = 0; i < s->num_children; i++) {
1185         qlist_append(children_list,
1186                      qobject_ref(s->children[i]->bs->full_open_options));
1187     }
1188 }
1189 
1190 static char *quorum_dirname(BlockDriverState *bs, Error **errp)
1191 {
1192     /* In general, there are multiple BDSs with different dirnames below this
1193      * one; so there is no unique dirname we could return (unless all are equal
1194      * by chance, or there is only one). Therefore, to be consistent, just
1195      * always return NULL. */
1196     error_setg(errp, "Cannot generate a base directory for quorum nodes");
1197     return NULL;
1198 }
1199 
1200 static void quorum_child_perm(BlockDriverState *bs, BdrvChild *c,
1201                               BdrvChildRole role,
1202                               BlockReopenQueue *reopen_queue,
1203                               uint64_t perm, uint64_t shared,
1204                               uint64_t *nperm, uint64_t *nshared)
1205 {
1206     BDRVQuorumState *s = bs->opaque;
1207 
1208     *nperm = perm & DEFAULT_PERM_PASSTHROUGH;
1209     if (s->rewrite_corrupted) {
1210         *nperm |= BLK_PERM_WRITE;
1211     }
1212 
1213     /*
1214      * We cannot share RESIZE or WRITE, as this would make the
1215      * children differ from each other.
1216      */
1217     *nshared = (shared & (BLK_PERM_CONSISTENT_READ |
1218                           BLK_PERM_WRITE_UNCHANGED))
1219              | DEFAULT_PERM_UNCHANGED;
1220 }
1221 
1222 /*
1223  * Each one of the children can report different status flags even
1224  * when they contain the same data, so what this function does is
1225  * return BDRV_BLOCK_ZERO if *all* children agree that a certain
1226  * region contains zeroes, and BDRV_BLOCK_DATA otherwise.
1227  */
1228 static int coroutine_fn GRAPH_RDLOCK
1229 quorum_co_block_status(BlockDriverState *bs, bool want_zero,
1230                        int64_t offset, int64_t count,
1231                        int64_t *pnum, int64_t *map, BlockDriverState **file)
1232 {
1233     BDRVQuorumState *s = bs->opaque;
1234     int i, ret;
1235     int64_t pnum_zero = count;
1236     int64_t pnum_data = 0;
1237 
1238     for (i = 0; i < s->num_children; i++) {
1239         int64_t bytes;
1240         ret = bdrv_co_common_block_status_above(s->children[i]->bs, NULL, false,
1241                                                 want_zero, offset, count,
1242                                                 &bytes, NULL, NULL, NULL);
1243         if (ret < 0) {
1244             quorum_report_bad(QUORUM_OP_TYPE_READ, offset, count,
1245                               s->children[i]->bs->node_name, ret);
1246             pnum_data = count;
1247             break;
1248         }
1249         /*
1250          * Even if all children agree about whether there are zeroes
1251          * or not at @offset they might disagree on the size, so use
1252          * the smallest when reporting BDRV_BLOCK_ZERO and the largest
1253          * when reporting BDRV_BLOCK_DATA.
1254          */
1255         if (ret & BDRV_BLOCK_ZERO) {
1256             pnum_zero = MIN(pnum_zero, bytes);
1257         } else {
1258             pnum_data = MAX(pnum_data, bytes);
1259         }
1260     }
1261 
1262     if (pnum_data) {
1263         *pnum = pnum_data;
1264         return BDRV_BLOCK_DATA;
1265     } else {
1266         *pnum = pnum_zero;
1267         return BDRV_BLOCK_ZERO;
1268     }
1269 }
1270 
1271 static const char *const quorum_strong_runtime_opts[] = {
1272     QUORUM_OPT_VOTE_THRESHOLD,
1273     QUORUM_OPT_BLKVERIFY,
1274     QUORUM_OPT_REWRITE,
1275     QUORUM_OPT_READ_PATTERN,
1276 
1277     NULL
1278 };
1279 
1280 static BlockDriver bdrv_quorum = {
1281     .format_name                        = "quorum",
1282 
1283     .instance_size                      = sizeof(BDRVQuorumState),
1284 
1285     .bdrv_open                          = quorum_open,
1286     .bdrv_close                         = quorum_close,
1287     .bdrv_gather_child_options          = quorum_gather_child_options,
1288     .bdrv_dirname                       = quorum_dirname,
1289     .bdrv_co_block_status               = quorum_co_block_status,
1290 
1291     .bdrv_co_flush                      = quorum_co_flush,
1292 
1293     .bdrv_co_getlength                  = quorum_co_getlength,
1294 
1295     .bdrv_co_preadv                     = quorum_co_preadv,
1296     .bdrv_co_pwritev                    = quorum_co_pwritev,
1297     .bdrv_co_pwrite_zeroes              = quorum_co_pwrite_zeroes,
1298 
1299     .bdrv_add_child                     = quorum_add_child,
1300     .bdrv_del_child                     = quorum_del_child,
1301 
1302     .bdrv_child_perm                    = quorum_child_perm,
1303 
1304     .bdrv_recurse_can_replace           = quorum_recurse_can_replace,
1305 
1306     .strong_runtime_opts                = quorum_strong_runtime_opts,
1307 };
1308 
1309 static void bdrv_quorum_init(void)
1310 {
1311     if (!qcrypto_hash_supports(QCRYPTO_HASH_ALGO_SHA256)) {
1312         /* SHA256 hash support is required for quorum device */
1313         return;
1314     }
1315     bdrv_register(&bdrv_quorum);
1316 }
1317 
1318 block_init(bdrv_quorum_init);
1319