xref: /qemu/block/quorum.c (revision a942d8fa)
1 /*
2  * Quorum Block filter
3  *
4  * Copyright (C) 2012-2014 Nodalink, EURL.
5  *
6  * Author:
7  *   Benoît Canet <benoit.canet@irqsave.net>
8  *
9  * Based on the design and code of blkverify.c (Copyright (C) 2010 IBM, Corp)
10  * and blkmirror.c (Copyright (C) 2011 Red Hat, Inc).
11  *
12  * This work is licensed under the terms of the GNU GPL, version 2 or later.
13  * See the COPYING file in the top-level directory.
14  */
15 
16 #include "qemu/osdep.h"
17 #include "qemu/cutils.h"
18 #include "block/block_int.h"
19 #include "qapi/qmp/qbool.h"
20 #include "qapi/qmp/qdict.h"
21 #include "qapi/qmp/qerror.h"
22 #include "qapi/qmp/qint.h"
23 #include "qapi/qmp/qjson.h"
24 #include "qapi/qmp/qlist.h"
25 #include "qapi/qmp/qstring.h"
26 #include "qapi-event.h"
27 #include "crypto/hash.h"
28 
29 #define HASH_LENGTH 32
30 
31 #define QUORUM_OPT_VOTE_THRESHOLD "vote-threshold"
32 #define QUORUM_OPT_BLKVERIFY      "blkverify"
33 #define QUORUM_OPT_REWRITE        "rewrite-corrupted"
34 #define QUORUM_OPT_READ_PATTERN   "read-pattern"
35 
36 /* This union holds a vote hash value */
37 typedef union QuorumVoteValue {
38     uint8_t h[HASH_LENGTH];    /* SHA-256 hash */
39     int64_t l;                 /* simpler 64 bits hash */
40 } QuorumVoteValue;
41 
42 /* A vote item */
43 typedef struct QuorumVoteItem {
44     int index;
45     QLIST_ENTRY(QuorumVoteItem) next;
46 } QuorumVoteItem;
47 
48 /* this structure is a vote version. A version is the set of votes sharing the
49  * same vote value.
50  * The set of votes will be tracked with the items field and its cardinality is
51  * vote_count.
52  */
53 typedef struct QuorumVoteVersion {
54     QuorumVoteValue value;
55     int index;
56     int vote_count;
57     QLIST_HEAD(, QuorumVoteItem) items;
58     QLIST_ENTRY(QuorumVoteVersion) next;
59 } QuorumVoteVersion;
60 
61 /* this structure holds a group of vote versions together */
62 typedef struct QuorumVotes {
63     QLIST_HEAD(, QuorumVoteVersion) vote_list;
64     bool (*compare)(QuorumVoteValue *a, QuorumVoteValue *b);
65 } QuorumVotes;
66 
67 /* the following structure holds the state of one quorum instance */
68 typedef struct BDRVQuorumState {
69     BdrvChild **children;  /* children BlockDriverStates */
70     int num_children;      /* children count */
71     unsigned next_child_index;  /* the index of the next child that should
72                                  * be added
73                                  */
74     int threshold;         /* if less than threshold children reads gave the
75                             * same result a quorum error occurs.
76                             */
77     bool is_blkverify;     /* true if the driver is in blkverify mode
78                             * Writes are mirrored on two children devices.
79                             * On reads the two children devices' contents are
80                             * compared and if a difference is spotted its
81                             * location is printed and the code aborts.
82                             * It is useful to debug other block drivers by
83                             * comparing them with a reference one.
84                             */
85     bool rewrite_corrupted;/* true if the driver must rewrite-on-read corrupted
86                             * block if Quorum is reached.
87                             */
88 
89     QuorumReadPattern read_pattern;
90 } BDRVQuorumState;
91 
92 typedef struct QuorumAIOCB QuorumAIOCB;
93 
94 /* Quorum will create one instance of the following structure per operation it
95  * performs on its children.
96  * So for each read/write operation coming from the upper layer there will be
97  * $children_count QuorumChildRequest.
98  */
99 typedef struct QuorumChildRequest {
100     BlockAIOCB *aiocb;
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     BlockAIOCB common;
114 
115     /* Request metadata */
116     uint64_t sector_num;
117     int nb_sectors;
118 
119     QEMUIOVector *qiov;         /* calling IOV */
120 
121     QuorumChildRequest *qcrs;   /* individual child requests */
122     int count;                  /* number of completed AIOCB */
123     int success_count;          /* number of successfully completed AIOCB */
124 
125     int rewrite_count;          /* number of replica to rewrite: count down to
126                                  * zero once writes are fired
127                                  */
128 
129     QuorumVotes votes;
130 
131     bool is_read;
132     int vote_ret;
133     int child_iter;             /* which child to read in fifo pattern */
134 };
135 
136 static bool quorum_vote(QuorumAIOCB *acb);
137 
138 static void quorum_aio_cancel(BlockAIOCB *blockacb)
139 {
140     QuorumAIOCB *acb = container_of(blockacb, QuorumAIOCB, common);
141     BDRVQuorumState *s = acb->common.bs->opaque;
142     int i;
143 
144     /* cancel all callbacks */
145     for (i = 0; i < s->num_children; i++) {
146         if (acb->qcrs[i].aiocb) {
147             bdrv_aio_cancel_async(acb->qcrs[i].aiocb);
148         }
149     }
150 }
151 
152 static AIOCBInfo quorum_aiocb_info = {
153     .aiocb_size         = sizeof(QuorumAIOCB),
154     .cancel_async       = quorum_aio_cancel,
155 };
156 
157 static void quorum_aio_finalize(QuorumAIOCB *acb)
158 {
159     int i, ret = 0;
160 
161     if (acb->vote_ret) {
162         ret = acb->vote_ret;
163     }
164 
165     acb->common.cb(acb->common.opaque, ret);
166 
167     if (acb->is_read) {
168         /* on the quorum case acb->child_iter == s->num_children - 1 */
169         for (i = 0; i <= acb->child_iter; i++) {
170             qemu_vfree(acb->qcrs[i].buf);
171             qemu_iovec_destroy(&acb->qcrs[i].qiov);
172         }
173     }
174 
175     g_free(acb->qcrs);
176     qemu_aio_unref(acb);
177 }
178 
179 static bool quorum_sha256_compare(QuorumVoteValue *a, QuorumVoteValue *b)
180 {
181     return !memcmp(a->h, b->h, HASH_LENGTH);
182 }
183 
184 static bool quorum_64bits_compare(QuorumVoteValue *a, QuorumVoteValue *b)
185 {
186     return a->l == b->l;
187 }
188 
189 static QuorumAIOCB *quorum_aio_get(BDRVQuorumState *s,
190                                    BlockDriverState *bs,
191                                    QEMUIOVector *qiov,
192                                    uint64_t sector_num,
193                                    int nb_sectors,
194                                    BlockCompletionFunc *cb,
195                                    void *opaque)
196 {
197     QuorumAIOCB *acb = qemu_aio_get(&quorum_aiocb_info, bs, cb, opaque);
198     int i;
199 
200     acb->common.bs->opaque = s;
201     acb->sector_num = sector_num;
202     acb->nb_sectors = nb_sectors;
203     acb->qiov = qiov;
204     acb->qcrs = g_new0(QuorumChildRequest, s->num_children);
205     acb->count = 0;
206     acb->success_count = 0;
207     acb->rewrite_count = 0;
208     acb->votes.compare = quorum_sha256_compare;
209     QLIST_INIT(&acb->votes.vote_list);
210     acb->is_read = false;
211     acb->vote_ret = 0;
212 
213     for (i = 0; i < s->num_children; i++) {
214         acb->qcrs[i].buf = NULL;
215         acb->qcrs[i].ret = 0;
216         acb->qcrs[i].parent = acb;
217     }
218 
219     return acb;
220 }
221 
222 static void quorum_report_bad(QuorumOpType type, uint64_t sector_num,
223                               int nb_sectors, char *node_name, int ret)
224 {
225     const char *msg = NULL;
226     if (ret < 0) {
227         msg = strerror(-ret);
228     }
229 
230     qapi_event_send_quorum_report_bad(type, !!msg, msg, node_name,
231                                       sector_num, nb_sectors, &error_abort);
232 }
233 
234 static void quorum_report_failure(QuorumAIOCB *acb)
235 {
236     const char *reference = bdrv_get_device_or_node_name(acb->common.bs);
237     qapi_event_send_quorum_failure(reference, acb->sector_num,
238                                    acb->nb_sectors, &error_abort);
239 }
240 
241 static int quorum_vote_error(QuorumAIOCB *acb);
242 
243 static bool quorum_has_too_much_io_failed(QuorumAIOCB *acb)
244 {
245     BDRVQuorumState *s = acb->common.bs->opaque;
246 
247     if (acb->success_count < s->threshold) {
248         acb->vote_ret = quorum_vote_error(acb);
249         quorum_report_failure(acb);
250         return true;
251     }
252 
253     return false;
254 }
255 
256 static void quorum_rewrite_aio_cb(void *opaque, int ret)
257 {
258     QuorumAIOCB *acb = opaque;
259 
260     /* one less rewrite to do */
261     acb->rewrite_count--;
262 
263     /* wait until all rewrite callbacks have completed */
264     if (acb->rewrite_count) {
265         return;
266     }
267 
268     quorum_aio_finalize(acb);
269 }
270 
271 static BlockAIOCB *read_fifo_child(QuorumAIOCB *acb);
272 
273 static void quorum_copy_qiov(QEMUIOVector *dest, QEMUIOVector *source)
274 {
275     int i;
276     assert(dest->niov == source->niov);
277     assert(dest->size == source->size);
278     for (i = 0; i < source->niov; i++) {
279         assert(dest->iov[i].iov_len == source->iov[i].iov_len);
280         memcpy(dest->iov[i].iov_base,
281                source->iov[i].iov_base,
282                source->iov[i].iov_len);
283     }
284 }
285 
286 static void quorum_aio_cb(void *opaque, int ret)
287 {
288     QuorumChildRequest *sacb = opaque;
289     QuorumAIOCB *acb = sacb->parent;
290     BDRVQuorumState *s = acb->common.bs->opaque;
291     bool rewrite = false;
292 
293     if (ret == 0) {
294         acb->success_count++;
295     } else {
296         QuorumOpType type;
297         type = acb->is_read ? QUORUM_OP_TYPE_READ : QUORUM_OP_TYPE_WRITE;
298         quorum_report_bad(type, acb->sector_num, acb->nb_sectors,
299                           sacb->aiocb->bs->node_name, ret);
300     }
301 
302     if (acb->is_read && s->read_pattern == QUORUM_READ_PATTERN_FIFO) {
303         /* We try to read next child in FIFO order if we fail to read */
304         if (ret < 0 && (acb->child_iter + 1) < s->num_children) {
305             acb->child_iter++;
306             read_fifo_child(acb);
307             return;
308         }
309 
310         if (ret == 0) {
311             quorum_copy_qiov(acb->qiov, &acb->qcrs[acb->child_iter].qiov);
312         }
313         acb->vote_ret = ret;
314         quorum_aio_finalize(acb);
315         return;
316     }
317 
318     sacb->ret = ret;
319     acb->count++;
320     assert(acb->count <= s->num_children);
321     assert(acb->success_count <= s->num_children);
322     if (acb->count < s->num_children) {
323         return;
324     }
325 
326     /* Do the vote on read */
327     if (acb->is_read) {
328         rewrite = quorum_vote(acb);
329     } else {
330         quorum_has_too_much_io_failed(acb);
331     }
332 
333     /* if no rewrite is done the code will finish right away */
334     if (!rewrite) {
335         quorum_aio_finalize(acb);
336     }
337 }
338 
339 static void quorum_report_bad_versions(BDRVQuorumState *s,
340                                        QuorumAIOCB *acb,
341                                        QuorumVoteValue *value)
342 {
343     QuorumVoteVersion *version;
344     QuorumVoteItem *item;
345 
346     QLIST_FOREACH(version, &acb->votes.vote_list, next) {
347         if (acb->votes.compare(&version->value, value)) {
348             continue;
349         }
350         QLIST_FOREACH(item, &version->items, next) {
351             quorum_report_bad(QUORUM_OP_TYPE_READ, acb->sector_num,
352                               acb->nb_sectors,
353                               s->children[item->index]->bs->node_name, 0);
354         }
355     }
356 }
357 
358 static bool quorum_rewrite_bad_versions(BDRVQuorumState *s, QuorumAIOCB *acb,
359                                         QuorumVoteValue *value)
360 {
361     QuorumVoteVersion *version;
362     QuorumVoteItem *item;
363     int count = 0;
364 
365     /* first count the number of bad versions: done first to avoid concurrency
366      * issues.
367      */
368     QLIST_FOREACH(version, &acb->votes.vote_list, next) {
369         if (acb->votes.compare(&version->value, value)) {
370             continue;
371         }
372         QLIST_FOREACH(item, &version->items, next) {
373             count++;
374         }
375     }
376 
377     /* quorum_rewrite_aio_cb will count down this to zero */
378     acb->rewrite_count = count;
379 
380     /* now fire the correcting rewrites */
381     QLIST_FOREACH(version, &acb->votes.vote_list, next) {
382         if (acb->votes.compare(&version->value, value)) {
383             continue;
384         }
385         QLIST_FOREACH(item, &version->items, next) {
386             bdrv_aio_writev(s->children[item->index], acb->sector_num,
387                             acb->qiov, acb->nb_sectors, quorum_rewrite_aio_cb,
388                             acb);
389         }
390     }
391 
392     /* return true if any rewrite is done else false */
393     return count;
394 }
395 
396 static void quorum_count_vote(QuorumVotes *votes,
397                               QuorumVoteValue *value,
398                               int index)
399 {
400     QuorumVoteVersion *v = NULL, *version = NULL;
401     QuorumVoteItem *item;
402 
403     /* look if we have something with this hash */
404     QLIST_FOREACH(v, &votes->vote_list, next) {
405         if (votes->compare(&v->value, value)) {
406             version = v;
407             break;
408         }
409     }
410 
411     /* It's a version not yet in the list add it */
412     if (!version) {
413         version = g_new0(QuorumVoteVersion, 1);
414         QLIST_INIT(&version->items);
415         memcpy(&version->value, value, sizeof(version->value));
416         version->index = index;
417         version->vote_count = 0;
418         QLIST_INSERT_HEAD(&votes->vote_list, version, next);
419     }
420 
421     version->vote_count++;
422 
423     item = g_new0(QuorumVoteItem, 1);
424     item->index = index;
425     QLIST_INSERT_HEAD(&version->items, item, next);
426 }
427 
428 static void quorum_free_vote_list(QuorumVotes *votes)
429 {
430     QuorumVoteVersion *version, *next_version;
431     QuorumVoteItem *item, *next_item;
432 
433     QLIST_FOREACH_SAFE(version, &votes->vote_list, next, next_version) {
434         QLIST_REMOVE(version, next);
435         QLIST_FOREACH_SAFE(item, &version->items, next, next_item) {
436             QLIST_REMOVE(item, next);
437             g_free(item);
438         }
439         g_free(version);
440     }
441 }
442 
443 static int quorum_compute_hash(QuorumAIOCB *acb, int i, QuorumVoteValue *hash)
444 {
445     QEMUIOVector *qiov = &acb->qcrs[i].qiov;
446     size_t len = sizeof(hash->h);
447     uint8_t *data = hash->h;
448 
449     /* XXX - would be nice if we could pass in the Error **
450      * and propagate that back, but this quorum code is
451      * restricted to just errno values currently */
452     if (qcrypto_hash_bytesv(QCRYPTO_HASH_ALG_SHA256,
453                             qiov->iov, qiov->niov,
454                             &data, &len,
455                             NULL) < 0) {
456         return -EINVAL;
457     }
458 
459     return 0;
460 }
461 
462 static QuorumVoteVersion *quorum_get_vote_winner(QuorumVotes *votes)
463 {
464     int max = 0;
465     QuorumVoteVersion *candidate, *winner = NULL;
466 
467     QLIST_FOREACH(candidate, &votes->vote_list, next) {
468         if (candidate->vote_count > max) {
469             max = candidate->vote_count;
470             winner = candidate;
471         }
472     }
473 
474     return winner;
475 }
476 
477 /* qemu_iovec_compare is handy for blkverify mode because it returns the first
478  * differing byte location. Yet it is handcoded to compare vectors one byte
479  * after another so it does not benefit from the libc SIMD optimizations.
480  * quorum_iovec_compare is written for speed and should be used in the non
481  * blkverify mode of quorum.
482  */
483 static bool quorum_iovec_compare(QEMUIOVector *a, QEMUIOVector *b)
484 {
485     int i;
486     int result;
487 
488     assert(a->niov == b->niov);
489     for (i = 0; i < a->niov; i++) {
490         assert(a->iov[i].iov_len == b->iov[i].iov_len);
491         result = memcmp(a->iov[i].iov_base,
492                         b->iov[i].iov_base,
493                         a->iov[i].iov_len);
494         if (result) {
495             return false;
496         }
497     }
498 
499     return true;
500 }
501 
502 static void GCC_FMT_ATTR(2, 3) quorum_err(QuorumAIOCB *acb,
503                                           const char *fmt, ...)
504 {
505     va_list ap;
506 
507     va_start(ap, fmt);
508     fprintf(stderr, "quorum: sector_num=%" PRId64 " nb_sectors=%d ",
509             acb->sector_num, acb->nb_sectors);
510     vfprintf(stderr, fmt, ap);
511     fprintf(stderr, "\n");
512     va_end(ap);
513     exit(1);
514 }
515 
516 static bool quorum_compare(QuorumAIOCB *acb,
517                            QEMUIOVector *a,
518                            QEMUIOVector *b)
519 {
520     BDRVQuorumState *s = acb->common.bs->opaque;
521     ssize_t offset;
522 
523     /* This driver will replace blkverify in this particular case */
524     if (s->is_blkverify) {
525         offset = qemu_iovec_compare(a, b);
526         if (offset != -1) {
527             quorum_err(acb, "contents mismatch in sector %" PRId64,
528                        acb->sector_num +
529                        (uint64_t)(offset / BDRV_SECTOR_SIZE));
530         }
531         return true;
532     }
533 
534     return quorum_iovec_compare(a, b);
535 }
536 
537 /* Do a vote to get the error code */
538 static int quorum_vote_error(QuorumAIOCB *acb)
539 {
540     BDRVQuorumState *s = acb->common.bs->opaque;
541     QuorumVoteVersion *winner = NULL;
542     QuorumVotes error_votes;
543     QuorumVoteValue result_value;
544     int i, ret = 0;
545     bool error = false;
546 
547     QLIST_INIT(&error_votes.vote_list);
548     error_votes.compare = quorum_64bits_compare;
549 
550     for (i = 0; i < s->num_children; i++) {
551         ret = acb->qcrs[i].ret;
552         if (ret) {
553             error = true;
554             result_value.l = ret;
555             quorum_count_vote(&error_votes, &result_value, i);
556         }
557     }
558 
559     if (error) {
560         winner = quorum_get_vote_winner(&error_votes);
561         ret = winner->value.l;
562     }
563 
564     quorum_free_vote_list(&error_votes);
565 
566     return ret;
567 }
568 
569 static bool quorum_vote(QuorumAIOCB *acb)
570 {
571     bool quorum = true;
572     bool rewrite = false;
573     int i, j, ret;
574     QuorumVoteValue hash;
575     BDRVQuorumState *s = acb->common.bs->opaque;
576     QuorumVoteVersion *winner;
577 
578     if (quorum_has_too_much_io_failed(acb)) {
579         return false;
580     }
581 
582     /* get the index of the first successful read */
583     for (i = 0; i < s->num_children; i++) {
584         if (!acb->qcrs[i].ret) {
585             break;
586         }
587     }
588 
589     assert(i < s->num_children);
590 
591     /* compare this read with all other successful reads stopping at quorum
592      * failure
593      */
594     for (j = i + 1; j < s->num_children; j++) {
595         if (acb->qcrs[j].ret) {
596             continue;
597         }
598         quorum = quorum_compare(acb, &acb->qcrs[i].qiov, &acb->qcrs[j].qiov);
599         if (!quorum) {
600             break;
601        }
602     }
603 
604     /* Every successful read agrees */
605     if (quorum) {
606         quorum_copy_qiov(acb->qiov, &acb->qcrs[i].qiov);
607         return false;
608     }
609 
610     /* compute hashes for each successful read, also store indexes */
611     for (i = 0; i < s->num_children; i++) {
612         if (acb->qcrs[i].ret) {
613             continue;
614         }
615         ret = quorum_compute_hash(acb, i, &hash);
616         /* if ever the hash computation failed */
617         if (ret < 0) {
618             acb->vote_ret = ret;
619             goto free_exit;
620         }
621         quorum_count_vote(&acb->votes, &hash, i);
622     }
623 
624     /* vote to select the most represented version */
625     winner = quorum_get_vote_winner(&acb->votes);
626 
627     /* if the winner count is smaller than threshold the read fails */
628     if (winner->vote_count < s->threshold) {
629         quorum_report_failure(acb);
630         acb->vote_ret = -EIO;
631         goto free_exit;
632     }
633 
634     /* we have a winner: copy it */
635     quorum_copy_qiov(acb->qiov, &acb->qcrs[winner->index].qiov);
636 
637     /* some versions are bad print them */
638     quorum_report_bad_versions(s, acb, &winner->value);
639 
640     /* corruption correction is enabled */
641     if (s->rewrite_corrupted) {
642         rewrite = quorum_rewrite_bad_versions(s, acb, &winner->value);
643     }
644 
645 free_exit:
646     /* free lists */
647     quorum_free_vote_list(&acb->votes);
648     return rewrite;
649 }
650 
651 static BlockAIOCB *read_quorum_children(QuorumAIOCB *acb)
652 {
653     BDRVQuorumState *s = acb->common.bs->opaque;
654     int i;
655 
656     for (i = 0; i < s->num_children; i++) {
657         acb->qcrs[i].buf = qemu_blockalign(s->children[i]->bs, acb->qiov->size);
658         qemu_iovec_init(&acb->qcrs[i].qiov, acb->qiov->niov);
659         qemu_iovec_clone(&acb->qcrs[i].qiov, acb->qiov, acb->qcrs[i].buf);
660     }
661 
662     for (i = 0; i < s->num_children; i++) {
663         acb->qcrs[i].aiocb = bdrv_aio_readv(s->children[i], acb->sector_num,
664                                             &acb->qcrs[i].qiov, acb->nb_sectors,
665                                             quorum_aio_cb, &acb->qcrs[i]);
666     }
667 
668     return &acb->common;
669 }
670 
671 static BlockAIOCB *read_fifo_child(QuorumAIOCB *acb)
672 {
673     BDRVQuorumState *s = acb->common.bs->opaque;
674 
675     acb->qcrs[acb->child_iter].buf =
676         qemu_blockalign(s->children[acb->child_iter]->bs, acb->qiov->size);
677     qemu_iovec_init(&acb->qcrs[acb->child_iter].qiov, acb->qiov->niov);
678     qemu_iovec_clone(&acb->qcrs[acb->child_iter].qiov, acb->qiov,
679                      acb->qcrs[acb->child_iter].buf);
680     acb->qcrs[acb->child_iter].aiocb =
681         bdrv_aio_readv(s->children[acb->child_iter], acb->sector_num,
682                        &acb->qcrs[acb->child_iter].qiov, acb->nb_sectors,
683                        quorum_aio_cb, &acb->qcrs[acb->child_iter]);
684 
685     return &acb->common;
686 }
687 
688 static BlockAIOCB *quorum_aio_readv(BlockDriverState *bs,
689                                     int64_t sector_num,
690                                     QEMUIOVector *qiov,
691                                     int nb_sectors,
692                                     BlockCompletionFunc *cb,
693                                     void *opaque)
694 {
695     BDRVQuorumState *s = bs->opaque;
696     QuorumAIOCB *acb = quorum_aio_get(s, bs, qiov, sector_num,
697                                       nb_sectors, cb, opaque);
698     acb->is_read = true;
699 
700     if (s->read_pattern == QUORUM_READ_PATTERN_QUORUM) {
701         acb->child_iter = s->num_children - 1;
702         return read_quorum_children(acb);
703     }
704 
705     acb->child_iter = 0;
706     return read_fifo_child(acb);
707 }
708 
709 static BlockAIOCB *quorum_aio_writev(BlockDriverState *bs,
710                                      int64_t sector_num,
711                                      QEMUIOVector *qiov,
712                                      int nb_sectors,
713                                      BlockCompletionFunc *cb,
714                                      void *opaque)
715 {
716     BDRVQuorumState *s = bs->opaque;
717     QuorumAIOCB *acb = quorum_aio_get(s, bs, qiov, sector_num, nb_sectors,
718                                       cb, opaque);
719     int i;
720 
721     for (i = 0; i < s->num_children; i++) {
722         acb->qcrs[i].aiocb = bdrv_aio_writev(s->children[i], sector_num,
723                                              qiov, nb_sectors, &quorum_aio_cb,
724                                              &acb->qcrs[i]);
725     }
726 
727     return &acb->common;
728 }
729 
730 static int64_t quorum_getlength(BlockDriverState *bs)
731 {
732     BDRVQuorumState *s = bs->opaque;
733     int64_t result;
734     int i;
735 
736     /* check that all file have the same length */
737     result = bdrv_getlength(s->children[0]->bs);
738     if (result < 0) {
739         return result;
740     }
741     for (i = 1; i < s->num_children; i++) {
742         int64_t value = bdrv_getlength(s->children[i]->bs);
743         if (value < 0) {
744             return value;
745         }
746         if (value != result) {
747             return -EIO;
748         }
749     }
750 
751     return result;
752 }
753 
754 static coroutine_fn int quorum_co_flush(BlockDriverState *bs)
755 {
756     BDRVQuorumState *s = bs->opaque;
757     QuorumVoteVersion *winner = NULL;
758     QuorumVotes error_votes;
759     QuorumVoteValue result_value;
760     int i;
761     int result = 0;
762     int success_count = 0;
763 
764     QLIST_INIT(&error_votes.vote_list);
765     error_votes.compare = quorum_64bits_compare;
766 
767     for (i = 0; i < s->num_children; i++) {
768         result = bdrv_co_flush(s->children[i]->bs);
769         if (result) {
770             quorum_report_bad(QUORUM_OP_TYPE_FLUSH, 0,
771                               bdrv_nb_sectors(s->children[i]->bs),
772                               s->children[i]->bs->node_name, result);
773             result_value.l = result;
774             quorum_count_vote(&error_votes, &result_value, i);
775         } else {
776             success_count++;
777         }
778     }
779 
780     if (success_count >= s->threshold) {
781         result = 0;
782     } else {
783         winner = quorum_get_vote_winner(&error_votes);
784         result = winner->value.l;
785     }
786     quorum_free_vote_list(&error_votes);
787 
788     return result;
789 }
790 
791 static bool quorum_recurse_is_first_non_filter(BlockDriverState *bs,
792                                                BlockDriverState *candidate)
793 {
794     BDRVQuorumState *s = bs->opaque;
795     int i;
796 
797     for (i = 0; i < s->num_children; i++) {
798         bool perm = bdrv_recurse_is_first_non_filter(s->children[i]->bs,
799                                                      candidate);
800         if (perm) {
801             return true;
802         }
803     }
804 
805     return false;
806 }
807 
808 static int quorum_valid_threshold(int threshold, int num_children, Error **errp)
809 {
810 
811     if (threshold < 1) {
812         error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
813                    "vote-threshold", "value >= 1");
814         return -ERANGE;
815     }
816 
817     if (threshold > num_children) {
818         error_setg(errp, "threshold may not exceed children count");
819         return -ERANGE;
820     }
821 
822     return 0;
823 }
824 
825 static QemuOptsList quorum_runtime_opts = {
826     .name = "quorum",
827     .head = QTAILQ_HEAD_INITIALIZER(quorum_runtime_opts.head),
828     .desc = {
829         {
830             .name = QUORUM_OPT_VOTE_THRESHOLD,
831             .type = QEMU_OPT_NUMBER,
832             .help = "The number of vote needed for reaching quorum",
833         },
834         {
835             .name = QUORUM_OPT_BLKVERIFY,
836             .type = QEMU_OPT_BOOL,
837             .help = "Trigger block verify mode if set",
838         },
839         {
840             .name = QUORUM_OPT_REWRITE,
841             .type = QEMU_OPT_BOOL,
842             .help = "Rewrite corrupted block on read quorum",
843         },
844         {
845             .name = QUORUM_OPT_READ_PATTERN,
846             .type = QEMU_OPT_STRING,
847             .help = "Allowed pattern: quorum, fifo. Quorum is default",
848         },
849         { /* end of list */ }
850     },
851 };
852 
853 static int parse_read_pattern(const char *opt)
854 {
855     int i;
856 
857     if (!opt) {
858         /* Set quorum as default */
859         return QUORUM_READ_PATTERN_QUORUM;
860     }
861 
862     for (i = 0; i < QUORUM_READ_PATTERN__MAX; i++) {
863         if (!strcmp(opt, QuorumReadPattern_lookup[i])) {
864             return i;
865         }
866     }
867 
868     return -EINVAL;
869 }
870 
871 static int quorum_open(BlockDriverState *bs, QDict *options, int flags,
872                        Error **errp)
873 {
874     BDRVQuorumState *s = bs->opaque;
875     Error *local_err = NULL;
876     QemuOpts *opts = NULL;
877     bool *opened;
878     int i;
879     int ret = 0;
880 
881     qdict_flatten(options);
882 
883     /* count how many different children are present */
884     s->num_children = qdict_array_entries(options, "children.");
885     if (s->num_children < 0) {
886         error_setg(&local_err, "Option children is not a valid array");
887         ret = -EINVAL;
888         goto exit;
889     }
890     if (s->num_children < 1) {
891         error_setg(&local_err,
892                    "Number of provided children must be 1 or more");
893         ret = -EINVAL;
894         goto exit;
895     }
896 
897     opts = qemu_opts_create(&quorum_runtime_opts, NULL, 0, &error_abort);
898     qemu_opts_absorb_qdict(opts, options, &local_err);
899     if (local_err) {
900         ret = -EINVAL;
901         goto exit;
902     }
903 
904     s->threshold = qemu_opt_get_number(opts, QUORUM_OPT_VOTE_THRESHOLD, 0);
905     /* and validate it against s->num_children */
906     ret = quorum_valid_threshold(s->threshold, s->num_children, &local_err);
907     if (ret < 0) {
908         goto exit;
909     }
910 
911     ret = parse_read_pattern(qemu_opt_get(opts, QUORUM_OPT_READ_PATTERN));
912     if (ret < 0) {
913         error_setg(&local_err, "Please set read-pattern as fifo or quorum");
914         goto exit;
915     }
916     s->read_pattern = ret;
917 
918     if (s->read_pattern == QUORUM_READ_PATTERN_QUORUM) {
919         /* is the driver in blkverify mode */
920         if (qemu_opt_get_bool(opts, QUORUM_OPT_BLKVERIFY, false) &&
921             s->num_children == 2 && s->threshold == 2) {
922             s->is_blkverify = true;
923         } else if (qemu_opt_get_bool(opts, QUORUM_OPT_BLKVERIFY, false)) {
924             fprintf(stderr, "blkverify mode is set by setting blkverify=on "
925                     "and using two files with vote_threshold=2\n");
926         }
927 
928         s->rewrite_corrupted = qemu_opt_get_bool(opts, QUORUM_OPT_REWRITE,
929                                                  false);
930         if (s->rewrite_corrupted && s->is_blkverify) {
931             error_setg(&local_err,
932                        "rewrite-corrupted=on cannot be used with blkverify=on");
933             ret = -EINVAL;
934             goto exit;
935         }
936     }
937 
938     /* allocate the children array */
939     s->children = g_new0(BdrvChild *, s->num_children);
940     opened = g_new0(bool, s->num_children);
941 
942     for (i = 0; i < s->num_children; i++) {
943         char indexstr[32];
944         ret = snprintf(indexstr, 32, "children.%d", i);
945         assert(ret < 32);
946 
947         s->children[i] = bdrv_open_child(NULL, options, indexstr, bs,
948                                          &child_format, false, &local_err);
949         if (local_err) {
950             ret = -EINVAL;
951             goto close_exit;
952         }
953 
954         opened[i] = true;
955     }
956     s->next_child_index = s->num_children;
957 
958     g_free(opened);
959     goto exit;
960 
961 close_exit:
962     /* cleanup on error */
963     for (i = 0; i < s->num_children; i++) {
964         if (!opened[i]) {
965             continue;
966         }
967         bdrv_unref_child(bs, s->children[i]);
968     }
969     g_free(s->children);
970     g_free(opened);
971 exit:
972     qemu_opts_del(opts);
973     /* propagate error */
974     error_propagate(errp, local_err);
975     return ret;
976 }
977 
978 static void quorum_close(BlockDriverState *bs)
979 {
980     BDRVQuorumState *s = bs->opaque;
981     int i;
982 
983     for (i = 0; i < s->num_children; i++) {
984         bdrv_unref_child(bs, s->children[i]);
985     }
986 
987     g_free(s->children);
988 }
989 
990 static void quorum_add_child(BlockDriverState *bs, BlockDriverState *child_bs,
991                              Error **errp)
992 {
993     BDRVQuorumState *s = bs->opaque;
994     BdrvChild *child;
995     char indexstr[32];
996     int ret;
997 
998     assert(s->num_children <= INT_MAX / sizeof(BdrvChild *));
999     if (s->num_children == INT_MAX / sizeof(BdrvChild *) ||
1000         s->next_child_index == UINT_MAX) {
1001         error_setg(errp, "Too many children");
1002         return;
1003     }
1004 
1005     ret = snprintf(indexstr, 32, "children.%u", s->next_child_index);
1006     if (ret < 0 || ret >= 32) {
1007         error_setg(errp, "cannot generate child name");
1008         return;
1009     }
1010     s->next_child_index++;
1011 
1012     bdrv_drained_begin(bs);
1013 
1014     /* We can safely add the child now */
1015     bdrv_ref(child_bs);
1016     child = bdrv_attach_child(bs, child_bs, indexstr, &child_format);
1017     s->children = g_renew(BdrvChild *, s->children, s->num_children + 1);
1018     s->children[s->num_children++] = child;
1019 
1020     bdrv_drained_end(bs);
1021 }
1022 
1023 static void quorum_del_child(BlockDriverState *bs, BdrvChild *child,
1024                              Error **errp)
1025 {
1026     BDRVQuorumState *s = bs->opaque;
1027     int i;
1028 
1029     for (i = 0; i < s->num_children; i++) {
1030         if (s->children[i] == child) {
1031             break;
1032         }
1033     }
1034 
1035     /* we have checked it in bdrv_del_child() */
1036     assert(i < s->num_children);
1037 
1038     if (s->num_children <= s->threshold) {
1039         error_setg(errp,
1040             "The number of children cannot be lower than the vote threshold %d",
1041             s->threshold);
1042         return;
1043     }
1044 
1045     bdrv_drained_begin(bs);
1046 
1047     /* We can safely remove this child now */
1048     memmove(&s->children[i], &s->children[i + 1],
1049             (s->num_children - i - 1) * sizeof(BdrvChild *));
1050     s->children = g_renew(BdrvChild *, s->children, --s->num_children);
1051     bdrv_unref_child(bs, child);
1052 
1053     bdrv_drained_end(bs);
1054 }
1055 
1056 static void quorum_refresh_filename(BlockDriverState *bs, QDict *options)
1057 {
1058     BDRVQuorumState *s = bs->opaque;
1059     QDict *opts;
1060     QList *children;
1061     int i;
1062 
1063     for (i = 0; i < s->num_children; i++) {
1064         bdrv_refresh_filename(s->children[i]->bs);
1065         if (!s->children[i]->bs->full_open_options) {
1066             return;
1067         }
1068     }
1069 
1070     children = qlist_new();
1071     for (i = 0; i < s->num_children; i++) {
1072         QINCREF(s->children[i]->bs->full_open_options);
1073         qlist_append_obj(children,
1074                          QOBJECT(s->children[i]->bs->full_open_options));
1075     }
1076 
1077     opts = qdict_new();
1078     qdict_put_obj(opts, "driver", QOBJECT(qstring_from_str("quorum")));
1079     qdict_put_obj(opts, QUORUM_OPT_VOTE_THRESHOLD,
1080                   QOBJECT(qint_from_int(s->threshold)));
1081     qdict_put_obj(opts, QUORUM_OPT_BLKVERIFY,
1082                   QOBJECT(qbool_from_bool(s->is_blkverify)));
1083     qdict_put_obj(opts, QUORUM_OPT_REWRITE,
1084                   QOBJECT(qbool_from_bool(s->rewrite_corrupted)));
1085     qdict_put_obj(opts, "children", QOBJECT(children));
1086 
1087     bs->full_open_options = opts;
1088 }
1089 
1090 static BlockDriver bdrv_quorum = {
1091     .format_name                        = "quorum",
1092     .protocol_name                      = "quorum",
1093 
1094     .instance_size                      = sizeof(BDRVQuorumState),
1095 
1096     .bdrv_file_open                     = quorum_open,
1097     .bdrv_close                         = quorum_close,
1098     .bdrv_refresh_filename              = quorum_refresh_filename,
1099 
1100     .bdrv_co_flush_to_disk              = quorum_co_flush,
1101 
1102     .bdrv_getlength                     = quorum_getlength,
1103 
1104     .bdrv_aio_readv                     = quorum_aio_readv,
1105     .bdrv_aio_writev                    = quorum_aio_writev,
1106 
1107     .bdrv_add_child                     = quorum_add_child,
1108     .bdrv_del_child                     = quorum_del_child,
1109 
1110     .is_filter                          = true,
1111     .bdrv_recurse_is_first_non_filter   = quorum_recurse_is_first_non_filter,
1112 };
1113 
1114 static void bdrv_quorum_init(void)
1115 {
1116     if (!qcrypto_hash_supports(QCRYPTO_HASH_ALG_SHA256)) {
1117         /* SHA256 hash support is required for quorum device */
1118         return;
1119     }
1120     bdrv_register(&bdrv_quorum);
1121 }
1122 
1123 block_init(bdrv_quorum_init);
1124