xref: /qemu/block/quorum.c (revision a81df1b6)
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 "block/block_int.h"
21 #include "block/qdict.h"
22 #include "qapi/error.h"
23 #include "qapi/qapi-events-block.h"
24 #include "qapi/qmp/qdict.h"
25 #include "qapi/qmp/qerror.h"
26 #include "qapi/qmp/qlist.h"
27 #include "qapi/qmp/qstring.h"
28 #include "crypto/hash.h"
29 
30 #define HASH_LENGTH 32
31 
32 #define QUORUM_OPT_VOTE_THRESHOLD "vote-threshold"
33 #define QUORUM_OPT_BLKVERIFY      "blkverify"
34 #define QUORUM_OPT_REWRITE        "rewrite-corrupted"
35 #define QUORUM_OPT_READ_PATTERN   "read-pattern"
36 
37 /* This union holds a vote hash value */
38 typedef union QuorumVoteValue {
39     uint8_t h[HASH_LENGTH];    /* SHA-256 hash */
40     int64_t l;                 /* simpler 64 bits hash */
41 } QuorumVoteValue;
42 
43 /* A vote item */
44 typedef struct QuorumVoteItem {
45     int index;
46     QLIST_ENTRY(QuorumVoteItem) next;
47 } QuorumVoteItem;
48 
49 /* this structure is a vote version. A version is the set of votes sharing the
50  * same vote value.
51  * The set of votes will be tracked with the items field and its cardinality is
52  * vote_count.
53  */
54 typedef struct QuorumVoteVersion {
55     QuorumVoteValue value;
56     int index;
57     int vote_count;
58     QLIST_HEAD(, QuorumVoteItem) items;
59     QLIST_ENTRY(QuorumVoteVersion) next;
60 } QuorumVoteVersion;
61 
62 /* this structure holds a group of vote versions together */
63 typedef struct QuorumVotes {
64     QLIST_HEAD(, QuorumVoteVersion) vote_list;
65     bool (*compare)(QuorumVoteValue *a, QuorumVoteValue *b);
66 } QuorumVotes;
67 
68 /* the following structure holds the state of one quorum instance */
69 typedef struct BDRVQuorumState {
70     BdrvChild **children;  /* children BlockDriverStates */
71     int num_children;      /* children count */
72     unsigned next_child_index;  /* the index of the next child that should
73                                  * be added
74                                  */
75     int threshold;         /* if less than threshold children reads gave the
76                             * same result a quorum error occurs.
77                             */
78     bool is_blkverify;     /* true if the driver is in blkverify mode
79                             * Writes are mirrored on two children devices.
80                             * On reads the two children devices' contents are
81                             * compared and if a difference is spotted its
82                             * location is printed and the code aborts.
83                             * It is useful to debug other block drivers by
84                             * comparing them with a reference one.
85                             */
86     bool rewrite_corrupted;/* true if the driver must rewrite-on-read corrupted
87                             * block if Quorum is reached.
88                             */
89 
90     QuorumReadPattern read_pattern;
91 } BDRVQuorumState;
92 
93 typedef struct QuorumAIOCB QuorumAIOCB;
94 
95 /* Quorum will create one instance of the following structure per operation it
96  * performs on its children.
97  * So for each read/write operation coming from the upper layer there will be
98  * $children_count QuorumChildRequest.
99  */
100 typedef struct QuorumChildRequest {
101     BlockDriverState *bs;
102     QEMUIOVector qiov;
103     uint8_t *buf;
104     int ret;
105     QuorumAIOCB *parent;
106 } QuorumChildRequest;
107 
108 /* Quorum will use the following structure to track progress of each read/write
109  * operation received by the upper layer.
110  * This structure hold pointers to the QuorumChildRequest structures instances
111  * used to do operations on each children and track overall progress.
112  */
113 struct QuorumAIOCB {
114     BlockDriverState *bs;
115     Coroutine *co;
116 
117     /* Request metadata */
118     uint64_t offset;
119     uint64_t bytes;
120     int flags;
121 
122     QEMUIOVector *qiov;         /* calling IOV */
123 
124     QuorumChildRequest *qcrs;   /* individual child requests */
125     int count;                  /* number of completed AIOCB */
126     int success_count;          /* number of successfully completed AIOCB */
127 
128     int rewrite_count;          /* number of replica to rewrite: count down to
129                                  * zero once writes are fired
130                                  */
131 
132     QuorumVotes votes;
133 
134     bool is_read;
135     int vote_ret;
136     int children_read;          /* how many children have been read from */
137 };
138 
139 typedef struct QuorumCo {
140     QuorumAIOCB *acb;
141     int idx;
142 } QuorumCo;
143 
144 static void quorum_aio_finalize(QuorumAIOCB *acb)
145 {
146     g_free(acb->qcrs);
147     g_free(acb);
148 }
149 
150 static bool quorum_sha256_compare(QuorumVoteValue *a, QuorumVoteValue *b)
151 {
152     return !memcmp(a->h, b->h, HASH_LENGTH);
153 }
154 
155 static bool quorum_64bits_compare(QuorumVoteValue *a, QuorumVoteValue *b)
156 {
157     return a->l == b->l;
158 }
159 
160 static QuorumAIOCB *quorum_aio_get(BlockDriverState *bs,
161                                    QEMUIOVector *qiov,
162                                    uint64_t offset,
163                                    uint64_t bytes,
164                                    int flags)
165 {
166     BDRVQuorumState *s = bs->opaque;
167     QuorumAIOCB *acb = g_new(QuorumAIOCB, 1);
168     int i;
169 
170     *acb = (QuorumAIOCB) {
171         .co                 = qemu_coroutine_self(),
172         .bs                 = bs,
173         .offset             = offset,
174         .bytes              = bytes,
175         .flags              = flags,
176         .qiov               = qiov,
177         .votes.compare      = quorum_sha256_compare,
178         .votes.vote_list    = QLIST_HEAD_INITIALIZER(acb.votes.vote_list),
179     };
180 
181     acb->qcrs = g_new0(QuorumChildRequest, s->num_children);
182     for (i = 0; i < s->num_children; i++) {
183         acb->qcrs[i].buf = NULL;
184         acb->qcrs[i].ret = 0;
185         acb->qcrs[i].parent = acb;
186     }
187 
188     return acb;
189 }
190 
191 static void quorum_report_bad(QuorumOpType type, uint64_t offset,
192                               uint64_t bytes, char *node_name, int ret)
193 {
194     const char *msg = NULL;
195     int64_t start_sector = offset / BDRV_SECTOR_SIZE;
196     int64_t end_sector = DIV_ROUND_UP(offset + bytes, BDRV_SECTOR_SIZE);
197 
198     if (ret < 0) {
199         msg = strerror(-ret);
200     }
201 
202     qapi_event_send_quorum_report_bad(type, !!msg, msg, node_name, start_sector,
203                                       end_sector - start_sector);
204 }
205 
206 static void quorum_report_failure(QuorumAIOCB *acb)
207 {
208     const char *reference = bdrv_get_device_or_node_name(acb->bs);
209     int64_t start_sector = acb->offset / BDRV_SECTOR_SIZE;
210     int64_t end_sector = DIV_ROUND_UP(acb->offset + acb->bytes,
211                                       BDRV_SECTOR_SIZE);
212 
213     qapi_event_send_quorum_failure(reference, start_sector,
214                                    end_sector - start_sector);
215 }
216 
217 static int quorum_vote_error(QuorumAIOCB *acb);
218 
219 static bool quorum_has_too_much_io_failed(QuorumAIOCB *acb)
220 {
221     BDRVQuorumState *s = acb->bs->opaque;
222 
223     if (acb->success_count < s->threshold) {
224         acb->vote_ret = quorum_vote_error(acb);
225         quorum_report_failure(acb);
226         return true;
227     }
228 
229     return false;
230 }
231 
232 static int read_fifo_child(QuorumAIOCB *acb);
233 
234 static void quorum_copy_qiov(QEMUIOVector *dest, QEMUIOVector *source)
235 {
236     int i;
237     assert(dest->niov == source->niov);
238     assert(dest->size == source->size);
239     for (i = 0; i < source->niov; i++) {
240         assert(dest->iov[i].iov_len == source->iov[i].iov_len);
241         memcpy(dest->iov[i].iov_base,
242                source->iov[i].iov_base,
243                source->iov[i].iov_len);
244     }
245 }
246 
247 static void quorum_report_bad_acb(QuorumChildRequest *sacb, int ret)
248 {
249     QuorumAIOCB *acb = sacb->parent;
250     QuorumOpType type = acb->is_read ? QUORUM_OP_TYPE_READ : QUORUM_OP_TYPE_WRITE;
251     quorum_report_bad(type, acb->offset, acb->bytes, sacb->bs->node_name, ret);
252 }
253 
254 static void quorum_report_bad_versions(BDRVQuorumState *s,
255                                        QuorumAIOCB *acb,
256                                        QuorumVoteValue *value)
257 {
258     QuorumVoteVersion *version;
259     QuorumVoteItem *item;
260 
261     QLIST_FOREACH(version, &acb->votes.vote_list, next) {
262         if (acb->votes.compare(&version->value, value)) {
263             continue;
264         }
265         QLIST_FOREACH(item, &version->items, next) {
266             quorum_report_bad(QUORUM_OP_TYPE_READ, acb->offset, acb->bytes,
267                               s->children[item->index]->bs->node_name, 0);
268         }
269     }
270 }
271 
272 static void quorum_rewrite_entry(void *opaque)
273 {
274     QuorumCo *co = opaque;
275     QuorumAIOCB *acb = co->acb;
276     BDRVQuorumState *s = acb->bs->opaque;
277 
278     /* Ignore any errors, it's just a correction attempt for already
279      * corrupted data.
280      * Mask out BDRV_REQ_WRITE_UNCHANGED because this overwrites the
281      * area with different data from the other children. */
282     bdrv_co_pwritev(s->children[co->idx], acb->offset, acb->bytes,
283                     acb->qiov, acb->flags & ~BDRV_REQ_WRITE_UNCHANGED);
284 
285     /* Wake up the caller after the last rewrite */
286     acb->rewrite_count--;
287     if (!acb->rewrite_count) {
288         qemu_coroutine_enter_if_inactive(acb->co);
289     }
290 }
291 
292 static bool quorum_rewrite_bad_versions(QuorumAIOCB *acb,
293                                         QuorumVoteValue *value)
294 {
295     QuorumVoteVersion *version;
296     QuorumVoteItem *item;
297     int count = 0;
298 
299     /* first count the number of bad versions: done first to avoid concurrency
300      * issues.
301      */
302     QLIST_FOREACH(version, &acb->votes.vote_list, next) {
303         if (acb->votes.compare(&version->value, value)) {
304             continue;
305         }
306         QLIST_FOREACH(item, &version->items, next) {
307             count++;
308         }
309     }
310 
311     /* quorum_rewrite_entry will count down this to zero */
312     acb->rewrite_count = count;
313 
314     /* now fire the correcting rewrites */
315     QLIST_FOREACH(version, &acb->votes.vote_list, next) {
316         if (acb->votes.compare(&version->value, value)) {
317             continue;
318         }
319         QLIST_FOREACH(item, &version->items, next) {
320             Coroutine *co;
321             QuorumCo data = {
322                 .acb = acb,
323                 .idx = item->index,
324             };
325 
326             co = qemu_coroutine_create(quorum_rewrite_entry, &data);
327             qemu_coroutine_enter(co);
328         }
329     }
330 
331     /* return true if any rewrite is done else false */
332     return count;
333 }
334 
335 static void quorum_count_vote(QuorumVotes *votes,
336                               QuorumVoteValue *value,
337                               int index)
338 {
339     QuorumVoteVersion *v = NULL, *version = NULL;
340     QuorumVoteItem *item;
341 
342     /* look if we have something with this hash */
343     QLIST_FOREACH(v, &votes->vote_list, next) {
344         if (votes->compare(&v->value, value)) {
345             version = v;
346             break;
347         }
348     }
349 
350     /* It's a version not yet in the list add it */
351     if (!version) {
352         version = g_new0(QuorumVoteVersion, 1);
353         QLIST_INIT(&version->items);
354         memcpy(&version->value, value, sizeof(version->value));
355         version->index = index;
356         version->vote_count = 0;
357         QLIST_INSERT_HEAD(&votes->vote_list, version, next);
358     }
359 
360     version->vote_count++;
361 
362     item = g_new0(QuorumVoteItem, 1);
363     item->index = index;
364     QLIST_INSERT_HEAD(&version->items, item, next);
365 }
366 
367 static void quorum_free_vote_list(QuorumVotes *votes)
368 {
369     QuorumVoteVersion *version, *next_version;
370     QuorumVoteItem *item, *next_item;
371 
372     QLIST_FOREACH_SAFE(version, &votes->vote_list, next, next_version) {
373         QLIST_REMOVE(version, next);
374         QLIST_FOREACH_SAFE(item, &version->items, next, next_item) {
375             QLIST_REMOVE(item, next);
376             g_free(item);
377         }
378         g_free(version);
379     }
380 }
381 
382 static int quorum_compute_hash(QuorumAIOCB *acb, int i, QuorumVoteValue *hash)
383 {
384     QEMUIOVector *qiov = &acb->qcrs[i].qiov;
385     size_t len = sizeof(hash->h);
386     uint8_t *data = hash->h;
387 
388     /* XXX - would be nice if we could pass in the Error **
389      * and propagate that back, but this quorum code is
390      * restricted to just errno values currently */
391     if (qcrypto_hash_bytesv(QCRYPTO_HASH_ALG_SHA256,
392                             qiov->iov, qiov->niov,
393                             &data, &len,
394                             NULL) < 0) {
395         return -EINVAL;
396     }
397 
398     return 0;
399 }
400 
401 static QuorumVoteVersion *quorum_get_vote_winner(QuorumVotes *votes)
402 {
403     int max = 0;
404     QuorumVoteVersion *candidate, *winner = NULL;
405 
406     QLIST_FOREACH(candidate, &votes->vote_list, next) {
407         if (candidate->vote_count > max) {
408             max = candidate->vote_count;
409             winner = candidate;
410         }
411     }
412 
413     return winner;
414 }
415 
416 /* qemu_iovec_compare is handy for blkverify mode because it returns the first
417  * differing byte location. Yet it is handcoded to compare vectors one byte
418  * after another so it does not benefit from the libc SIMD optimizations.
419  * quorum_iovec_compare is written for speed and should be used in the non
420  * blkverify mode of quorum.
421  */
422 static bool quorum_iovec_compare(QEMUIOVector *a, QEMUIOVector *b)
423 {
424     int i;
425     int result;
426 
427     assert(a->niov == b->niov);
428     for (i = 0; i < a->niov; i++) {
429         assert(a->iov[i].iov_len == b->iov[i].iov_len);
430         result = memcmp(a->iov[i].iov_base,
431                         b->iov[i].iov_base,
432                         a->iov[i].iov_len);
433         if (result) {
434             return false;
435         }
436     }
437 
438     return true;
439 }
440 
441 static bool quorum_compare(QuorumAIOCB *acb, QEMUIOVector *a, QEMUIOVector *b)
442 {
443     BDRVQuorumState *s = acb->bs->opaque;
444     ssize_t offset;
445 
446     /* This driver will replace blkverify in this particular case */
447     if (s->is_blkverify) {
448         offset = qemu_iovec_compare(a, b);
449         if (offset != -1) {
450             fprintf(stderr, "quorum: offset=%" PRIu64 " bytes=%" PRIu64
451                     " contents mismatch at offset %" PRIu64 "\n",
452                     acb->offset, acb->bytes, acb->offset + offset);
453             exit(1);
454         }
455         return true;
456     }
457 
458     return quorum_iovec_compare(a, b);
459 }
460 
461 /* Do a vote to get the error code */
462 static int quorum_vote_error(QuorumAIOCB *acb)
463 {
464     BDRVQuorumState *s = acb->bs->opaque;
465     QuorumVoteVersion *winner = NULL;
466     QuorumVotes error_votes;
467     QuorumVoteValue result_value;
468     int i, ret = 0;
469     bool error = false;
470 
471     QLIST_INIT(&error_votes.vote_list);
472     error_votes.compare = quorum_64bits_compare;
473 
474     for (i = 0; i < s->num_children; i++) {
475         ret = acb->qcrs[i].ret;
476         if (ret) {
477             error = true;
478             result_value.l = ret;
479             quorum_count_vote(&error_votes, &result_value, i);
480         }
481     }
482 
483     if (error) {
484         winner = quorum_get_vote_winner(&error_votes);
485         ret = winner->value.l;
486     }
487 
488     quorum_free_vote_list(&error_votes);
489 
490     return ret;
491 }
492 
493 static void quorum_vote(QuorumAIOCB *acb)
494 {
495     bool quorum = true;
496     int i, j, ret;
497     QuorumVoteValue hash;
498     BDRVQuorumState *s = acb->bs->opaque;
499     QuorumVoteVersion *winner;
500 
501     if (quorum_has_too_much_io_failed(acb)) {
502         return;
503     }
504 
505     /* get the index of the first successful read */
506     for (i = 0; i < s->num_children; i++) {
507         if (!acb->qcrs[i].ret) {
508             break;
509         }
510     }
511 
512     assert(i < s->num_children);
513 
514     /* compare this read with all other successful reads stopping at quorum
515      * failure
516      */
517     for (j = i + 1; j < s->num_children; j++) {
518         if (acb->qcrs[j].ret) {
519             continue;
520         }
521         quorum = quorum_compare(acb, &acb->qcrs[i].qiov, &acb->qcrs[j].qiov);
522         if (!quorum) {
523             break;
524        }
525     }
526 
527     /* Every successful read agrees */
528     if (quorum) {
529         quorum_copy_qiov(acb->qiov, &acb->qcrs[i].qiov);
530         return;
531     }
532 
533     /* compute hashes for each successful read, also store indexes */
534     for (i = 0; i < s->num_children; i++) {
535         if (acb->qcrs[i].ret) {
536             continue;
537         }
538         ret = quorum_compute_hash(acb, i, &hash);
539         /* if ever the hash computation failed */
540         if (ret < 0) {
541             acb->vote_ret = ret;
542             goto free_exit;
543         }
544         quorum_count_vote(&acb->votes, &hash, i);
545     }
546 
547     /* vote to select the most represented version */
548     winner = quorum_get_vote_winner(&acb->votes);
549 
550     /* if the winner count is smaller than threshold the read fails */
551     if (winner->vote_count < s->threshold) {
552         quorum_report_failure(acb);
553         acb->vote_ret = -EIO;
554         goto free_exit;
555     }
556 
557     /* we have a winner: copy it */
558     quorum_copy_qiov(acb->qiov, &acb->qcrs[winner->index].qiov);
559 
560     /* some versions are bad print them */
561     quorum_report_bad_versions(s, acb, &winner->value);
562 
563     /* corruption correction is enabled */
564     if (s->rewrite_corrupted) {
565         quorum_rewrite_bad_versions(acb, &winner->value);
566     }
567 
568 free_exit:
569     /* free lists */
570     quorum_free_vote_list(&acb->votes);
571 }
572 
573 static void read_quorum_children_entry(void *opaque)
574 {
575     QuorumCo *co = opaque;
576     QuorumAIOCB *acb = co->acb;
577     BDRVQuorumState *s = acb->bs->opaque;
578     int i = co->idx;
579     QuorumChildRequest *sacb = &acb->qcrs[i];
580 
581     sacb->bs = s->children[i]->bs;
582     sacb->ret = bdrv_co_preadv(s->children[i], acb->offset, acb->bytes,
583                                &acb->qcrs[i].qiov, 0);
584 
585     if (sacb->ret == 0) {
586         acb->success_count++;
587     } else {
588         quorum_report_bad_acb(sacb, sacb->ret);
589     }
590 
591     acb->count++;
592     assert(acb->count <= s->num_children);
593     assert(acb->success_count <= s->num_children);
594 
595     /* Wake up the caller after the last read */
596     if (acb->count == s->num_children) {
597         qemu_coroutine_enter_if_inactive(acb->co);
598     }
599 }
600 
601 static int read_quorum_children(QuorumAIOCB *acb)
602 {
603     BDRVQuorumState *s = acb->bs->opaque;
604     int i;
605 
606     acb->children_read = s->num_children;
607     for (i = 0; i < s->num_children; i++) {
608         acb->qcrs[i].buf = qemu_blockalign(s->children[i]->bs, acb->qiov->size);
609         qemu_iovec_init(&acb->qcrs[i].qiov, acb->qiov->niov);
610         qemu_iovec_clone(&acb->qcrs[i].qiov, acb->qiov, acb->qcrs[i].buf);
611     }
612 
613     for (i = 0; i < s->num_children; i++) {
614         Coroutine *co;
615         QuorumCo data = {
616             .acb = acb,
617             .idx = i,
618         };
619 
620         co = qemu_coroutine_create(read_quorum_children_entry, &data);
621         qemu_coroutine_enter(co);
622     }
623 
624     while (acb->count < s->num_children) {
625         qemu_coroutine_yield();
626     }
627 
628     /* Do the vote on read */
629     quorum_vote(acb);
630     for (i = 0; i < s->num_children; i++) {
631         qemu_vfree(acb->qcrs[i].buf);
632         qemu_iovec_destroy(&acb->qcrs[i].qiov);
633     }
634 
635     while (acb->rewrite_count) {
636         qemu_coroutine_yield();
637     }
638 
639     return acb->vote_ret;
640 }
641 
642 static int read_fifo_child(QuorumAIOCB *acb)
643 {
644     BDRVQuorumState *s = acb->bs->opaque;
645     int n, ret;
646 
647     /* We try to read the next child in FIFO order if we failed to read */
648     do {
649         n = acb->children_read++;
650         acb->qcrs[n].bs = s->children[n]->bs;
651         ret = bdrv_co_preadv(s->children[n], acb->offset, acb->bytes,
652                              acb->qiov, 0);
653         if (ret < 0) {
654             quorum_report_bad_acb(&acb->qcrs[n], ret);
655         }
656     } while (ret < 0 && acb->children_read < s->num_children);
657 
658     /* FIXME: rewrite failed children if acb->children_read > 1? */
659 
660     return ret;
661 }
662 
663 static int quorum_co_preadv(BlockDriverState *bs, uint64_t offset,
664                             uint64_t bytes, QEMUIOVector *qiov, int flags)
665 {
666     BDRVQuorumState *s = bs->opaque;
667     QuorumAIOCB *acb = quorum_aio_get(bs, qiov, offset, bytes, flags);
668     int ret;
669 
670     acb->is_read = true;
671     acb->children_read = 0;
672 
673     if (s->read_pattern == QUORUM_READ_PATTERN_QUORUM) {
674         ret = read_quorum_children(acb);
675     } else {
676         ret = read_fifo_child(acb);
677     }
678     quorum_aio_finalize(acb);
679 
680     return ret;
681 }
682 
683 static void write_quorum_entry(void *opaque)
684 {
685     QuorumCo *co = opaque;
686     QuorumAIOCB *acb = co->acb;
687     BDRVQuorumState *s = acb->bs->opaque;
688     int i = co->idx;
689     QuorumChildRequest *sacb = &acb->qcrs[i];
690 
691     sacb->bs = s->children[i]->bs;
692     sacb->ret = bdrv_co_pwritev(s->children[i], acb->offset, acb->bytes,
693                                 acb->qiov, acb->flags);
694     if (sacb->ret == 0) {
695         acb->success_count++;
696     } else {
697         quorum_report_bad_acb(sacb, sacb->ret);
698     }
699     acb->count++;
700     assert(acb->count <= s->num_children);
701     assert(acb->success_count <= s->num_children);
702 
703     /* Wake up the caller after the last write */
704     if (acb->count == s->num_children) {
705         qemu_coroutine_enter_if_inactive(acb->co);
706     }
707 }
708 
709 static int quorum_co_pwritev(BlockDriverState *bs, uint64_t offset,
710                              uint64_t bytes, QEMUIOVector *qiov, int flags)
711 {
712     BDRVQuorumState *s = bs->opaque;
713     QuorumAIOCB *acb = quorum_aio_get(bs, qiov, offset, bytes, flags);
714     int i, ret;
715 
716     for (i = 0; i < s->num_children; i++) {
717         Coroutine *co;
718         QuorumCo data = {
719             .acb = acb,
720             .idx = i,
721         };
722 
723         co = qemu_coroutine_create(write_quorum_entry, &data);
724         qemu_coroutine_enter(co);
725     }
726 
727     while (acb->count < s->num_children) {
728         qemu_coroutine_yield();
729     }
730 
731     quorum_has_too_much_io_failed(acb);
732 
733     ret = acb->vote_ret;
734     quorum_aio_finalize(acb);
735 
736     return ret;
737 }
738 
739 static int64_t quorum_getlength(BlockDriverState *bs)
740 {
741     BDRVQuorumState *s = bs->opaque;
742     int64_t result;
743     int i;
744 
745     /* check that all file have the same length */
746     result = bdrv_getlength(s->children[0]->bs);
747     if (result < 0) {
748         return result;
749     }
750     for (i = 1; i < s->num_children; i++) {
751         int64_t value = bdrv_getlength(s->children[i]->bs);
752         if (value < 0) {
753             return value;
754         }
755         if (value != result) {
756             return -EIO;
757         }
758     }
759 
760     return result;
761 }
762 
763 static coroutine_fn int quorum_co_flush(BlockDriverState *bs)
764 {
765     BDRVQuorumState *s = bs->opaque;
766     QuorumVoteVersion *winner = NULL;
767     QuorumVotes error_votes;
768     QuorumVoteValue result_value;
769     int i;
770     int result = 0;
771     int success_count = 0;
772 
773     QLIST_INIT(&error_votes.vote_list);
774     error_votes.compare = quorum_64bits_compare;
775 
776     for (i = 0; i < s->num_children; i++) {
777         result = bdrv_co_flush(s->children[i]->bs);
778         if (result) {
779             quorum_report_bad(QUORUM_OP_TYPE_FLUSH, 0, 0,
780                               s->children[i]->bs->node_name, result);
781             result_value.l = result;
782             quorum_count_vote(&error_votes, &result_value, i);
783         } else {
784             success_count++;
785         }
786     }
787 
788     if (success_count >= s->threshold) {
789         result = 0;
790     } else {
791         winner = quorum_get_vote_winner(&error_votes);
792         result = winner->value.l;
793     }
794     quorum_free_vote_list(&error_votes);
795 
796     return result;
797 }
798 
799 static bool quorum_recurse_can_replace(BlockDriverState *bs,
800                                        BlockDriverState *to_replace)
801 {
802     BDRVQuorumState *s = bs->opaque;
803     int i;
804 
805     for (i = 0; i < s->num_children; i++) {
806         /*
807          * We have no idea whether our children show the same data as
808          * this node (@bs).  It is actually highly likely that
809          * @to_replace does not, because replacing a broken child is
810          * one of the main use cases here.
811          *
812          * We do know that the new BDS will match @bs, so replacing
813          * any of our children by it will be safe.  It cannot change
814          * the data this quorum node presents to its parents.
815          *
816          * However, replacing @to_replace by @bs in any of our
817          * children's chains may change visible data somewhere in
818          * there.  We therefore cannot recurse down those chains with
819          * bdrv_recurse_can_replace().
820          * (More formally, bdrv_recurse_can_replace() requires that
821          * @to_replace will be replaced by something matching the @bs
822          * passed to it.  We cannot guarantee that.)
823          *
824          * Thus, we can only check whether any of our immediate
825          * children matches @to_replace.
826          *
827          * (In the future, we might add a function to recurse down a
828          * chain that checks that nothing there cares about a change
829          * in data from the respective child in question.  For
830          * example, most filters do not care when their child's data
831          * suddenly changes, as long as their parents do not care.)
832          */
833         if (s->children[i]->bs == to_replace) {
834             /*
835              * We now have to ensure that there is no other parent
836              * that cares about replacing this child by a node with
837              * potentially different data.
838              * We do so by checking whether there are any other parents
839              * at all, which is stricter than necessary, but also very
840              * simple.  (We may decide to implement something more
841              * complex and permissive when there is an actual need for
842              * it.)
843              */
844             return QLIST_FIRST(&to_replace->parents) == s->children[i] &&
845                 QLIST_NEXT(s->children[i], next_parent) == NULL;
846         }
847     }
848 
849     return false;
850 }
851 
852 static int quorum_valid_threshold(int threshold, int num_children, Error **errp)
853 {
854 
855     if (threshold < 1) {
856         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
857                    "vote-threshold", "value >= 1");
858         return -ERANGE;
859     }
860 
861     if (threshold > num_children) {
862         error_setg(errp, "threshold may not exceed children count");
863         return -ERANGE;
864     }
865 
866     return 0;
867 }
868 
869 static QemuOptsList quorum_runtime_opts = {
870     .name = "quorum",
871     .head = QTAILQ_HEAD_INITIALIZER(quorum_runtime_opts.head),
872     .desc = {
873         {
874             .name = QUORUM_OPT_VOTE_THRESHOLD,
875             .type = QEMU_OPT_NUMBER,
876             .help = "The number of vote needed for reaching quorum",
877         },
878         {
879             .name = QUORUM_OPT_BLKVERIFY,
880             .type = QEMU_OPT_BOOL,
881             .help = "Trigger block verify mode if set",
882         },
883         {
884             .name = QUORUM_OPT_REWRITE,
885             .type = QEMU_OPT_BOOL,
886             .help = "Rewrite corrupted block on read quorum",
887         },
888         {
889             .name = QUORUM_OPT_READ_PATTERN,
890             .type = QEMU_OPT_STRING,
891             .help = "Allowed pattern: quorum, fifo. Quorum is default",
892         },
893         { /* end of list */ }
894     },
895 };
896 
897 static int quorum_open(BlockDriverState *bs, QDict *options, int flags,
898                        Error **errp)
899 {
900     BDRVQuorumState *s = bs->opaque;
901     Error *local_err = NULL;
902     QemuOpts *opts = NULL;
903     const char *pattern_str;
904     bool *opened;
905     int i;
906     int ret = 0;
907 
908     qdict_flatten(options);
909 
910     /* count how many different children are present */
911     s->num_children = qdict_array_entries(options, "children.");
912     if (s->num_children < 0) {
913         error_setg(errp, "Option children is not a valid array");
914         ret = -EINVAL;
915         goto exit;
916     }
917     if (s->num_children < 1) {
918         error_setg(errp, "Number of provided children must be 1 or more");
919         ret = -EINVAL;
920         goto exit;
921     }
922 
923     opts = qemu_opts_create(&quorum_runtime_opts, NULL, 0, &error_abort);
924     if (!qemu_opts_absorb_qdict(opts, options, errp)) {
925         ret = -EINVAL;
926         goto exit;
927     }
928 
929     s->threshold = qemu_opt_get_number(opts, QUORUM_OPT_VOTE_THRESHOLD, 0);
930     /* and validate it against s->num_children */
931     ret = quorum_valid_threshold(s->threshold, s->num_children, errp);
932     if (ret < 0) {
933         goto exit;
934     }
935 
936     pattern_str = qemu_opt_get(opts, QUORUM_OPT_READ_PATTERN);
937     if (!pattern_str) {
938         ret = QUORUM_READ_PATTERN_QUORUM;
939     } else {
940         ret = qapi_enum_parse(&QuorumReadPattern_lookup, pattern_str,
941                               -EINVAL, NULL);
942     }
943     if (ret < 0) {
944         error_setg(errp, "Please set read-pattern as fifo or quorum");
945         goto exit;
946     }
947     s->read_pattern = ret;
948 
949     if (s->read_pattern == QUORUM_READ_PATTERN_QUORUM) {
950         s->is_blkverify = qemu_opt_get_bool(opts, QUORUM_OPT_BLKVERIFY, false);
951         if (s->is_blkverify && (s->num_children != 2 || s->threshold != 2)) {
952             error_setg(errp, "blkverify=on can only be set if there are "
953                        "exactly two files and vote-threshold is 2");
954             ret = -EINVAL;
955             goto exit;
956         }
957 
958         s->rewrite_corrupted = qemu_opt_get_bool(opts, QUORUM_OPT_REWRITE,
959                                                  false);
960         if (s->rewrite_corrupted && s->is_blkverify) {
961             error_setg(errp,
962                        "rewrite-corrupted=on cannot be used with blkverify=on");
963             ret = -EINVAL;
964             goto exit;
965         }
966     }
967 
968     /* allocate the children array */
969     s->children = g_new0(BdrvChild *, s->num_children);
970     opened = g_new0(bool, s->num_children);
971 
972     for (i = 0; i < s->num_children; i++) {
973         char indexstr[32];
974         ret = snprintf(indexstr, 32, "children.%d", i);
975         assert(ret < 32);
976 
977         s->children[i] = bdrv_open_child(NULL, options, indexstr, bs,
978                                          &child_of_bds, BDRV_CHILD_DATA, false,
979                                          &local_err);
980         if (local_err) {
981             error_propagate(errp, local_err);
982             ret = -EINVAL;
983             goto close_exit;
984         }
985 
986         opened[i] = true;
987     }
988     s->next_child_index = s->num_children;
989 
990     bs->supported_write_flags = BDRV_REQ_WRITE_UNCHANGED;
991 
992     g_free(opened);
993     goto exit;
994 
995 close_exit:
996     /* cleanup on error */
997     for (i = 0; i < s->num_children; i++) {
998         if (!opened[i]) {
999             continue;
1000         }
1001         bdrv_unref_child(bs, s->children[i]);
1002     }
1003     g_free(s->children);
1004     g_free(opened);
1005 exit:
1006     qemu_opts_del(opts);
1007     return ret;
1008 }
1009 
1010 static void quorum_close(BlockDriverState *bs)
1011 {
1012     BDRVQuorumState *s = bs->opaque;
1013     int i;
1014 
1015     for (i = 0; i < s->num_children; i++) {
1016         bdrv_unref_child(bs, s->children[i]);
1017     }
1018 
1019     g_free(s->children);
1020 }
1021 
1022 static void quorum_add_child(BlockDriverState *bs, BlockDriverState *child_bs,
1023                              Error **errp)
1024 {
1025     BDRVQuorumState *s = bs->opaque;
1026     BdrvChild *child;
1027     char indexstr[32];
1028     int ret;
1029 
1030     if (s->is_blkverify) {
1031         error_setg(errp, "Cannot add a child to a quorum in blkverify mode");
1032         return;
1033     }
1034 
1035     assert(s->num_children <= INT_MAX / sizeof(BdrvChild *));
1036     if (s->num_children == INT_MAX / sizeof(BdrvChild *) ||
1037         s->next_child_index == UINT_MAX) {
1038         error_setg(errp, "Too many children");
1039         return;
1040     }
1041 
1042     ret = snprintf(indexstr, 32, "children.%u", s->next_child_index);
1043     if (ret < 0 || ret >= 32) {
1044         error_setg(errp, "cannot generate child name");
1045         return;
1046     }
1047     s->next_child_index++;
1048 
1049     bdrv_drained_begin(bs);
1050 
1051     /* We can safely add the child now */
1052     bdrv_ref(child_bs);
1053 
1054     child = bdrv_attach_child(bs, child_bs, indexstr, &child_of_bds,
1055                               BDRV_CHILD_DATA, errp);
1056     if (child == NULL) {
1057         s->next_child_index--;
1058         goto out;
1059     }
1060     s->children = g_renew(BdrvChild *, s->children, s->num_children + 1);
1061     s->children[s->num_children++] = child;
1062 
1063 out:
1064     bdrv_drained_end(bs);
1065 }
1066 
1067 static void quorum_del_child(BlockDriverState *bs, BdrvChild *child,
1068                              Error **errp)
1069 {
1070     BDRVQuorumState *s = bs->opaque;
1071     int i;
1072 
1073     for (i = 0; i < s->num_children; i++) {
1074         if (s->children[i] == child) {
1075             break;
1076         }
1077     }
1078 
1079     /* we have checked it in bdrv_del_child() */
1080     assert(i < s->num_children);
1081 
1082     if (s->num_children <= s->threshold) {
1083         error_setg(errp,
1084             "The number of children cannot be lower than the vote threshold %d",
1085             s->threshold);
1086         return;
1087     }
1088 
1089     /* We know now that num_children > threshold, so blkverify must be false */
1090     assert(!s->is_blkverify);
1091 
1092     bdrv_drained_begin(bs);
1093 
1094     /* We can safely remove this child now */
1095     memmove(&s->children[i], &s->children[i + 1],
1096             (s->num_children - i - 1) * sizeof(BdrvChild *));
1097     s->children = g_renew(BdrvChild *, s->children, --s->num_children);
1098     bdrv_unref_child(bs, child);
1099 
1100     bdrv_drained_end(bs);
1101 }
1102 
1103 static void quorum_gather_child_options(BlockDriverState *bs, QDict *target,
1104                                         bool backing_overridden)
1105 {
1106     BDRVQuorumState *s = bs->opaque;
1107     QList *children_list;
1108     int i;
1109 
1110     /*
1111      * The generic implementation for gathering child options in
1112      * bdrv_refresh_filename() would use the names of the children
1113      * as specified for bdrv_open_child() or bdrv_attach_child(),
1114      * which is "children.%u" with %u being a value
1115      * (s->next_child_index) that is incremented each time a new child
1116      * is added (and never decremented).  Since children can be
1117      * deleted at runtime, there may be gaps in that enumeration.
1118      * When creating a new quorum BDS and specifying the children for
1119      * it through runtime options, the enumeration used there may not
1120      * have any gaps, though.
1121      *
1122      * Therefore, we have to create a new gap-less enumeration here
1123      * (which we can achieve by simply putting all of the children's
1124      * full_open_options into a QList).
1125      *
1126      * XXX: Note that there are issues with the current child option
1127      *      structure quorum uses (such as the fact that children do
1128      *      not really have unique permanent names).  Therefore, this
1129      *      is going to have to change in the future and ideally we
1130      *      want quorum to be covered by the generic implementation.
1131      */
1132 
1133     children_list = qlist_new();
1134     qdict_put(target, "children", children_list);
1135 
1136     for (i = 0; i < s->num_children; i++) {
1137         qlist_append(children_list,
1138                      qobject_ref(s->children[i]->bs->full_open_options));
1139     }
1140 }
1141 
1142 static char *quorum_dirname(BlockDriverState *bs, Error **errp)
1143 {
1144     /* In general, there are multiple BDSs with different dirnames below this
1145      * one; so there is no unique dirname we could return (unless all are equal
1146      * by chance, or there is only one). Therefore, to be consistent, just
1147      * always return NULL. */
1148     error_setg(errp, "Cannot generate a base directory for quorum nodes");
1149     return NULL;
1150 }
1151 
1152 static void quorum_child_perm(BlockDriverState *bs, BdrvChild *c,
1153                               BdrvChildRole role,
1154                               BlockReopenQueue *reopen_queue,
1155                               uint64_t perm, uint64_t shared,
1156                               uint64_t *nperm, uint64_t *nshared)
1157 {
1158     *nperm = perm & DEFAULT_PERM_PASSTHROUGH;
1159 
1160     /*
1161      * We cannot share RESIZE or WRITE, as this would make the
1162      * children differ from each other.
1163      */
1164     *nshared = (shared & (BLK_PERM_CONSISTENT_READ |
1165                           BLK_PERM_WRITE_UNCHANGED))
1166              | DEFAULT_PERM_UNCHANGED;
1167 }
1168 
1169 static const char *const quorum_strong_runtime_opts[] = {
1170     QUORUM_OPT_VOTE_THRESHOLD,
1171     QUORUM_OPT_BLKVERIFY,
1172     QUORUM_OPT_REWRITE,
1173     QUORUM_OPT_READ_PATTERN,
1174 
1175     NULL
1176 };
1177 
1178 static BlockDriver bdrv_quorum = {
1179     .format_name                        = "quorum",
1180 
1181     .instance_size                      = sizeof(BDRVQuorumState),
1182 
1183     .bdrv_open                          = quorum_open,
1184     .bdrv_close                         = quorum_close,
1185     .bdrv_gather_child_options          = quorum_gather_child_options,
1186     .bdrv_dirname                       = quorum_dirname,
1187 
1188     .bdrv_co_flush_to_disk              = quorum_co_flush,
1189 
1190     .bdrv_getlength                     = quorum_getlength,
1191 
1192     .bdrv_co_preadv                     = quorum_co_preadv,
1193     .bdrv_co_pwritev                    = quorum_co_pwritev,
1194 
1195     .bdrv_add_child                     = quorum_add_child,
1196     .bdrv_del_child                     = quorum_del_child,
1197 
1198     .bdrv_child_perm                    = quorum_child_perm,
1199 
1200     .bdrv_recurse_can_replace           = quorum_recurse_can_replace,
1201 
1202     .strong_runtime_opts                = quorum_strong_runtime_opts,
1203 };
1204 
1205 static void bdrv_quorum_init(void)
1206 {
1207     if (!qcrypto_hash_supports(QCRYPTO_HASH_ALG_SHA256)) {
1208         /* SHA256 hash support is required for quorum device */
1209         return;
1210     }
1211     bdrv_register(&bdrv_quorum);
1212 }
1213 
1214 block_init(bdrv_quorum_init);
1215