4 * Copyright (C) 2012-2014 Nodalink, EURL.
7 * BenoƮt Canet <benoit.canet@irqsave.net>
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).
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.
16 #include <gnutls/gnutls.h>
17 #include <gnutls/crypto.h>
18 #include "block/block_int.h"
19 #include "qapi/qmp/qjson.h"
21 #define HASH_LENGTH 32
23 #define QUORUM_OPT_VOTE_THRESHOLD "vote-threshold"
24 #define QUORUM_OPT_BLKVERIFY "blkverify"
26 /* This union holds a vote hash value */
27 typedef union QuorumVoteValue
{
28 char h
[HASH_LENGTH
]; /* SHA-256 hash */
29 int64_t l
; /* simpler 64 bits hash */
33 typedef struct QuorumVoteItem
{
35 QLIST_ENTRY(QuorumVoteItem
) next
;
38 /* this structure is a vote version. A version is the set of votes sharing the
40 * The set of votes will be tracked with the items field and its cardinality is
43 typedef struct QuorumVoteVersion
{
44 QuorumVoteValue value
;
47 QLIST_HEAD(, QuorumVoteItem
) items
;
48 QLIST_ENTRY(QuorumVoteVersion
) next
;
51 /* this structure holds a group of vote versions together */
52 typedef struct QuorumVotes
{
53 QLIST_HEAD(, QuorumVoteVersion
) vote_list
;
54 bool (*compare
)(QuorumVoteValue
*a
, QuorumVoteValue
*b
);
57 /* the following structure holds the state of one quorum instance */
58 typedef struct BDRVQuorumState
{
59 BlockDriverState
**bs
; /* children BlockDriverStates */
60 int num_children
; /* children count */
61 int threshold
; /* if less than threshold children reads gave the
62 * same result a quorum error occurs.
64 bool is_blkverify
; /* true if the driver is in blkverify mode
65 * Writes are mirrored on two children devices.
66 * On reads the two children devices' contents are
67 * compared and if a difference is spotted its
68 * location is printed and the code aborts.
69 * It is useful to debug other block drivers by
70 * comparing them with a reference one.
74 typedef struct QuorumAIOCB QuorumAIOCB
;
76 /* Quorum will create one instance of the following structure per operation it
77 * performs on its children.
78 * So for each read/write operation coming from the upper layer there will be
79 * $children_count QuorumChildRequest.
81 typedef struct QuorumChildRequest
{
82 BlockDriverAIOCB
*aiocb
;
89 /* Quorum will use the following structure to track progress of each read/write
90 * operation received by the upper layer.
91 * This structure hold pointers to the QuorumChildRequest structures instances
92 * used to do operations on each children and track overall progress.
95 BlockDriverAIOCB common
;
97 /* Request metadata */
101 QEMUIOVector
*qiov
; /* calling IOV */
103 QuorumChildRequest
*qcrs
; /* individual child requests */
104 int count
; /* number of completed AIOCB */
105 int success_count
; /* number of successfully completed AIOCB */
113 static void quorum_vote(QuorumAIOCB
*acb
);
115 static void quorum_aio_cancel(BlockDriverAIOCB
*blockacb
)
117 QuorumAIOCB
*acb
= container_of(blockacb
, QuorumAIOCB
, common
);
118 BDRVQuorumState
*s
= acb
->common
.bs
->opaque
;
121 /* cancel all callbacks */
122 for (i
= 0; i
< s
->num_children
; i
++) {
123 bdrv_aio_cancel(acb
->qcrs
[i
].aiocb
);
127 qemu_aio_release(acb
);
130 static AIOCBInfo quorum_aiocb_info
= {
131 .aiocb_size
= sizeof(QuorumAIOCB
),
132 .cancel
= quorum_aio_cancel
,
135 static void quorum_aio_finalize(QuorumAIOCB
*acb
)
137 BDRVQuorumState
*s
= acb
->common
.bs
->opaque
;
144 acb
->common
.cb(acb
->common
.opaque
, ret
);
147 for (i
= 0; i
< s
->num_children
; i
++) {
148 qemu_vfree(acb
->qcrs
[i
].buf
);
149 qemu_iovec_destroy(&acb
->qcrs
[i
].qiov
);
154 qemu_aio_release(acb
);
157 static bool quorum_sha256_compare(QuorumVoteValue
*a
, QuorumVoteValue
*b
)
159 return !memcmp(a
->h
, b
->h
, HASH_LENGTH
);
162 static bool quorum_64bits_compare(QuorumVoteValue
*a
, QuorumVoteValue
*b
)
167 static QuorumAIOCB
*quorum_aio_get(BDRVQuorumState
*s
,
168 BlockDriverState
*bs
,
172 BlockDriverCompletionFunc
*cb
,
175 QuorumAIOCB
*acb
= qemu_aio_get(&quorum_aiocb_info
, bs
, cb
, opaque
);
178 acb
->common
.bs
->opaque
= s
;
179 acb
->sector_num
= sector_num
;
180 acb
->nb_sectors
= nb_sectors
;
182 acb
->qcrs
= g_new0(QuorumChildRequest
, s
->num_children
);
184 acb
->success_count
= 0;
185 acb
->votes
.compare
= quorum_sha256_compare
;
186 QLIST_INIT(&acb
->votes
.vote_list
);
187 acb
->is_read
= false;
190 for (i
= 0; i
< s
->num_children
; i
++) {
191 acb
->qcrs
[i
].buf
= NULL
;
192 acb
->qcrs
[i
].ret
= 0;
193 acb
->qcrs
[i
].parent
= acb
;
199 static void quorum_report_bad(QuorumAIOCB
*acb
, char *node_name
, int ret
)
203 data
= qobject_from_jsonf("{ 'node-name': %s"
204 ", 'sector-num': %" PRId64
205 ", 'sectors-count': %d }",
206 node_name
, acb
->sector_num
, acb
->nb_sectors
);
208 QDict
*dict
= qobject_to_qdict(data
);
209 qdict_put(dict
, "error", qstring_from_str(strerror(-ret
)));
211 monitor_protocol_event(QEVENT_QUORUM_REPORT_BAD
, data
);
212 qobject_decref(data
);
215 static void quorum_report_failure(QuorumAIOCB
*acb
)
218 const char *reference
= acb
->common
.bs
->device_name
[0] ?
219 acb
->common
.bs
->device_name
:
220 acb
->common
.bs
->node_name
;
221 data
= qobject_from_jsonf("{ 'reference': %s"
222 ", 'sector-num': %" PRId64
223 ", 'sectors-count': %d }",
224 reference
, acb
->sector_num
, acb
->nb_sectors
);
225 monitor_protocol_event(QEVENT_QUORUM_FAILURE
, data
);
226 qobject_decref(data
);
229 static int quorum_vote_error(QuorumAIOCB
*acb
);
231 static bool quorum_has_too_much_io_failed(QuorumAIOCB
*acb
)
233 BDRVQuorumState
*s
= acb
->common
.bs
->opaque
;
235 if (acb
->success_count
< s
->threshold
) {
236 acb
->vote_ret
= quorum_vote_error(acb
);
237 quorum_report_failure(acb
);
244 static void quorum_aio_cb(void *opaque
, int ret
)
246 QuorumChildRequest
*sacb
= opaque
;
247 QuorumAIOCB
*acb
= sacb
->parent
;
248 BDRVQuorumState
*s
= acb
->common
.bs
->opaque
;
253 acb
->success_count
++;
255 quorum_report_bad(acb
, sacb
->aiocb
->bs
->node_name
, ret
);
257 assert(acb
->count
<= s
->num_children
);
258 assert(acb
->success_count
<= s
->num_children
);
259 if (acb
->count
< s
->num_children
) {
263 /* Do the vote on read */
267 quorum_has_too_much_io_failed(acb
);
270 quorum_aio_finalize(acb
);
273 static void quorum_report_bad_versions(BDRVQuorumState
*s
,
275 QuorumVoteValue
*value
)
277 QuorumVoteVersion
*version
;
278 QuorumVoteItem
*item
;
280 QLIST_FOREACH(version
, &acb
->votes
.vote_list
, next
) {
281 if (acb
->votes
.compare(&version
->value
, value
)) {
284 QLIST_FOREACH(item
, &version
->items
, next
) {
285 quorum_report_bad(acb
, s
->bs
[item
->index
]->node_name
, 0);
290 static void quorum_copy_qiov(QEMUIOVector
*dest
, QEMUIOVector
*source
)
293 assert(dest
->niov
== source
->niov
);
294 assert(dest
->size
== source
->size
);
295 for (i
= 0; i
< source
->niov
; i
++) {
296 assert(dest
->iov
[i
].iov_len
== source
->iov
[i
].iov_len
);
297 memcpy(dest
->iov
[i
].iov_base
,
298 source
->iov
[i
].iov_base
,
299 source
->iov
[i
].iov_len
);
303 static void quorum_count_vote(QuorumVotes
*votes
,
304 QuorumVoteValue
*value
,
307 QuorumVoteVersion
*v
= NULL
, *version
= NULL
;
308 QuorumVoteItem
*item
;
310 /* look if we have something with this hash */
311 QLIST_FOREACH(v
, &votes
->vote_list
, next
) {
312 if (votes
->compare(&v
->value
, value
)) {
318 /* It's a version not yet in the list add it */
320 version
= g_new0(QuorumVoteVersion
, 1);
321 QLIST_INIT(&version
->items
);
322 memcpy(&version
->value
, value
, sizeof(version
->value
));
323 version
->index
= index
;
324 version
->vote_count
= 0;
325 QLIST_INSERT_HEAD(&votes
->vote_list
, version
, next
);
328 version
->vote_count
++;
330 item
= g_new0(QuorumVoteItem
, 1);
332 QLIST_INSERT_HEAD(&version
->items
, item
, next
);
335 static void quorum_free_vote_list(QuorumVotes
*votes
)
337 QuorumVoteVersion
*version
, *next_version
;
338 QuorumVoteItem
*item
, *next_item
;
340 QLIST_FOREACH_SAFE(version
, &votes
->vote_list
, next
, next_version
) {
341 QLIST_REMOVE(version
, next
);
342 QLIST_FOREACH_SAFE(item
, &version
->items
, next
, next_item
) {
343 QLIST_REMOVE(item
, next
);
350 static int quorum_compute_hash(QuorumAIOCB
*acb
, int i
, QuorumVoteValue
*hash
)
353 gnutls_hash_hd_t dig
;
354 QEMUIOVector
*qiov
= &acb
->qcrs
[i
].qiov
;
356 ret
= gnutls_hash_init(&dig
, GNUTLS_DIG_SHA256
);
362 for (j
= 0; j
< qiov
->niov
; j
++) {
363 ret
= gnutls_hash(dig
, qiov
->iov
[j
].iov_base
, qiov
->iov
[j
].iov_len
);
369 gnutls_hash_deinit(dig
, (void *) hash
);
373 static QuorumVoteVersion
*quorum_get_vote_winner(QuorumVotes
*votes
)
376 QuorumVoteVersion
*candidate
, *winner
= NULL
;
378 QLIST_FOREACH(candidate
, &votes
->vote_list
, next
) {
379 if (candidate
->vote_count
> max
) {
380 max
= candidate
->vote_count
;
388 /* qemu_iovec_compare is handy for blkverify mode because it returns the first
389 * differing byte location. Yet it is handcoded to compare vectors one byte
390 * after another so it does not benefit from the libc SIMD optimizations.
391 * quorum_iovec_compare is written for speed and should be used in the non
392 * blkverify mode of quorum.
394 static bool quorum_iovec_compare(QEMUIOVector
*a
, QEMUIOVector
*b
)
399 assert(a
->niov
== b
->niov
);
400 for (i
= 0; i
< a
->niov
; i
++) {
401 assert(a
->iov
[i
].iov_len
== b
->iov
[i
].iov_len
);
402 result
= memcmp(a
->iov
[i
].iov_base
,
413 static void GCC_FMT_ATTR(2, 3) quorum_err(QuorumAIOCB
*acb
,
414 const char *fmt
, ...)
419 fprintf(stderr
, "quorum: sector_num=%" PRId64
" nb_sectors=%d ",
420 acb
->sector_num
, acb
->nb_sectors
);
421 vfprintf(stderr
, fmt
, ap
);
422 fprintf(stderr
, "\n");
427 static bool quorum_compare(QuorumAIOCB
*acb
,
431 BDRVQuorumState
*s
= acb
->common
.bs
->opaque
;
434 /* This driver will replace blkverify in this particular case */
435 if (s
->is_blkverify
) {
436 offset
= qemu_iovec_compare(a
, b
);
438 quorum_err(acb
, "contents mismatch in sector %" PRId64
,
440 (uint64_t)(offset
/ BDRV_SECTOR_SIZE
));
445 return quorum_iovec_compare(a
, b
);
448 /* Do a vote to get the error code */
449 static int quorum_vote_error(QuorumAIOCB
*acb
)
451 BDRVQuorumState
*s
= acb
->common
.bs
->opaque
;
452 QuorumVoteVersion
*winner
= NULL
;
453 QuorumVotes error_votes
;
454 QuorumVoteValue result_value
;
458 QLIST_INIT(&error_votes
.vote_list
);
459 error_votes
.compare
= quorum_64bits_compare
;
461 for (i
= 0; i
< s
->num_children
; i
++) {
462 ret
= acb
->qcrs
[i
].ret
;
465 result_value
.l
= ret
;
466 quorum_count_vote(&error_votes
, &result_value
, i
);
471 winner
= quorum_get_vote_winner(&error_votes
);
472 ret
= winner
->value
.l
;
475 quorum_free_vote_list(&error_votes
);
480 static void quorum_vote(QuorumAIOCB
*acb
)
484 QuorumVoteValue hash
;
485 BDRVQuorumState
*s
= acb
->common
.bs
->opaque
;
486 QuorumVoteVersion
*winner
;
488 if (quorum_has_too_much_io_failed(acb
)) {
492 /* get the index of the first successful read */
493 for (i
= 0; i
< s
->num_children
; i
++) {
494 if (!acb
->qcrs
[i
].ret
) {
499 assert(i
< s
->num_children
);
501 /* compare this read with all other successful reads stopping at quorum
504 for (j
= i
+ 1; j
< s
->num_children
; j
++) {
505 if (acb
->qcrs
[j
].ret
) {
508 quorum
= quorum_compare(acb
, &acb
->qcrs
[i
].qiov
, &acb
->qcrs
[j
].qiov
);
514 /* Every successful read agrees */
516 quorum_copy_qiov(acb
->qiov
, &acb
->qcrs
[i
].qiov
);
520 /* compute hashes for each successful read, also store indexes */
521 for (i
= 0; i
< s
->num_children
; i
++) {
522 if (acb
->qcrs
[i
].ret
) {
525 ret
= quorum_compute_hash(acb
, i
, &hash
);
526 /* if ever the hash computation failed */
531 quorum_count_vote(&acb
->votes
, &hash
, i
);
534 /* vote to select the most represented version */
535 winner
= quorum_get_vote_winner(&acb
->votes
);
537 /* if the winner count is smaller than threshold the read fails */
538 if (winner
->vote_count
< s
->threshold
) {
539 quorum_report_failure(acb
);
540 acb
->vote_ret
= -EIO
;
544 /* we have a winner: copy it */
545 quorum_copy_qiov(acb
->qiov
, &acb
->qcrs
[winner
->index
].qiov
);
547 /* some versions are bad print them */
548 quorum_report_bad_versions(s
, acb
, &winner
->value
);
552 quorum_free_vote_list(&acb
->votes
);
555 static BlockDriverAIOCB
*quorum_aio_readv(BlockDriverState
*bs
,
559 BlockDriverCompletionFunc
*cb
,
562 BDRVQuorumState
*s
= bs
->opaque
;
563 QuorumAIOCB
*acb
= quorum_aio_get(s
, bs
, qiov
, sector_num
,
564 nb_sectors
, cb
, opaque
);
569 for (i
= 0; i
< s
->num_children
; i
++) {
570 acb
->qcrs
[i
].buf
= qemu_blockalign(s
->bs
[i
], qiov
->size
);
571 qemu_iovec_init(&acb
->qcrs
[i
].qiov
, qiov
->niov
);
572 qemu_iovec_clone(&acb
->qcrs
[i
].qiov
, qiov
, acb
->qcrs
[i
].buf
);
575 for (i
= 0; i
< s
->num_children
; i
++) {
576 bdrv_aio_readv(s
->bs
[i
], sector_num
, &acb
->qcrs
[i
].qiov
, nb_sectors
,
577 quorum_aio_cb
, &acb
->qcrs
[i
]);
583 static BlockDriverAIOCB
*quorum_aio_writev(BlockDriverState
*bs
,
587 BlockDriverCompletionFunc
*cb
,
590 BDRVQuorumState
*s
= bs
->opaque
;
591 QuorumAIOCB
*acb
= quorum_aio_get(s
, bs
, qiov
, sector_num
, nb_sectors
,
595 for (i
= 0; i
< s
->num_children
; i
++) {
596 acb
->qcrs
[i
].aiocb
= bdrv_aio_writev(s
->bs
[i
], sector_num
, qiov
,
597 nb_sectors
, &quorum_aio_cb
,
604 static int64_t quorum_getlength(BlockDriverState
*bs
)
606 BDRVQuorumState
*s
= bs
->opaque
;
610 /* check that all file have the same length */
611 result
= bdrv_getlength(s
->bs
[0]);
615 for (i
= 1; i
< s
->num_children
; i
++) {
616 int64_t value
= bdrv_getlength(s
->bs
[i
]);
620 if (value
!= result
) {
628 static void quorum_invalidate_cache(BlockDriverState
*bs
, Error
**errp
)
630 BDRVQuorumState
*s
= bs
->opaque
;
631 Error
*local_err
= NULL
;
634 for (i
= 0; i
< s
->num_children
; i
++) {
635 bdrv_invalidate_cache(s
->bs
[i
], &local_err
);
637 error_propagate(errp
, local_err
);
643 static coroutine_fn
int quorum_co_flush(BlockDriverState
*bs
)
645 BDRVQuorumState
*s
= bs
->opaque
;
646 QuorumVoteVersion
*winner
= NULL
;
647 QuorumVotes error_votes
;
648 QuorumVoteValue result_value
;
652 QLIST_INIT(&error_votes
.vote_list
);
653 error_votes
.compare
= quorum_64bits_compare
;
655 for (i
= 0; i
< s
->num_children
; i
++) {
656 result
= bdrv_co_flush(s
->bs
[i
]);
657 result_value
.l
= result
;
658 quorum_count_vote(&error_votes
, &result_value
, i
);
661 winner
= quorum_get_vote_winner(&error_votes
);
662 result
= winner
->value
.l
;
664 quorum_free_vote_list(&error_votes
);
669 static bool quorum_recurse_is_first_non_filter(BlockDriverState
*bs
,
670 BlockDriverState
*candidate
)
672 BDRVQuorumState
*s
= bs
->opaque
;
675 for (i
= 0; i
< s
->num_children
; i
++) {
676 bool perm
= bdrv_recurse_is_first_non_filter(s
->bs
[i
],
686 static int quorum_valid_threshold(int threshold
, int num_children
, Error
**errp
)
690 error_set(errp
, QERR_INVALID_PARAMETER_VALUE
,
691 "vote-threshold", "value >= 1");
695 if (threshold
> num_children
) {
696 error_setg(errp
, "threshold may not exceed children count");
703 static QemuOptsList quorum_runtime_opts
= {
705 .head
= QTAILQ_HEAD_INITIALIZER(quorum_runtime_opts
.head
),
708 .name
= QUORUM_OPT_VOTE_THRESHOLD
,
709 .type
= QEMU_OPT_NUMBER
,
710 .help
= "The number of vote needed for reaching quorum",
713 .name
= QUORUM_OPT_BLKVERIFY
,
714 .type
= QEMU_OPT_BOOL
,
715 .help
= "Trigger block verify mode if set",
717 { /* end of list */ }
721 static int quorum_open(BlockDriverState
*bs
, QDict
*options
, int flags
,
724 BDRVQuorumState
*s
= bs
->opaque
;
725 Error
*local_err
= NULL
;
730 const QListEntry
*lentry
;
734 qdict_flatten(options
);
735 qdict_extract_subqdict(options
, &sub
, "children.");
736 qdict_array_split(sub
, &list
);
738 if (qdict_size(sub
)) {
739 error_setg(&local_err
, "Invalid option children.%s",
740 qdict_first(sub
)->key
);
745 /* count how many different children are present */
746 s
->num_children
= qlist_size(list
);
747 if (s
->num_children
< 2) {
748 error_setg(&local_err
,
749 "Number of provided children must be greater than 1");
754 opts
= qemu_opts_create(&quorum_runtime_opts
, NULL
, 0, &error_abort
);
755 qemu_opts_absorb_qdict(opts
, options
, &local_err
);
761 s
->threshold
= qemu_opt_get_number(opts
, QUORUM_OPT_VOTE_THRESHOLD
, 0);
763 /* and validate it against s->num_children */
764 ret
= quorum_valid_threshold(s
->threshold
, s
->num_children
, &local_err
);
769 /* is the driver in blkverify mode */
770 if (qemu_opt_get_bool(opts
, QUORUM_OPT_BLKVERIFY
, false) &&
771 s
->num_children
== 2 && s
->threshold
== 2) {
772 s
->is_blkverify
= true;
773 } else if (qemu_opt_get_bool(opts
, QUORUM_OPT_BLKVERIFY
, false)) {
774 fprintf(stderr
, "blkverify mode is set by setting blkverify=on "
775 "and using two files with vote_threshold=2\n");
778 /* allocate the children BlockDriverState array */
779 s
->bs
= g_new0(BlockDriverState
*, s
->num_children
);
780 opened
= g_new0(bool, s
->num_children
);
782 for (i
= 0, lentry
= qlist_first(list
); lentry
;
783 lentry
= qlist_next(lentry
), i
++) {
787 switch (qobject_type(lentry
->value
))
789 /* List of options */
791 d
= qobject_to_qdict(lentry
->value
);
793 ret
= bdrv_open(&s
->bs
[i
], NULL
, NULL
, d
, flags
, NULL
,
799 string
= qobject_to_qstring(lentry
->value
);
800 ret
= bdrv_open(&s
->bs
[i
], NULL
, qstring_get_str(string
), NULL
,
801 flags
, NULL
, &local_err
);
805 error_setg(&local_err
, "Specification of child block device %i "
820 /* cleanup on error */
821 for (i
= 0; i
< s
->num_children
; i
++) {
825 bdrv_unref(s
->bs
[i
]);
830 /* propagate error */
832 error_propagate(errp
, local_err
);
839 static void quorum_close(BlockDriverState
*bs
)
841 BDRVQuorumState
*s
= bs
->opaque
;
844 for (i
= 0; i
< s
->num_children
; i
++) {
845 bdrv_unref(s
->bs
[i
]);
851 static BlockDriver bdrv_quorum
= {
852 .format_name
= "quorum",
853 .protocol_name
= "quorum",
855 .instance_size
= sizeof(BDRVQuorumState
),
857 .bdrv_file_open
= quorum_open
,
858 .bdrv_close
= quorum_close
,
860 .bdrv_co_flush_to_disk
= quorum_co_flush
,
862 .bdrv_getlength
= quorum_getlength
,
864 .bdrv_aio_readv
= quorum_aio_readv
,
865 .bdrv_aio_writev
= quorum_aio_writev
,
866 .bdrv_invalidate_cache
= quorum_invalidate_cache
,
869 .bdrv_recurse_is_first_non_filter
= quorum_recurse_is_first_non_filter
,
872 static void bdrv_quorum_init(void)
874 bdrv_register(&bdrv_quorum
);
877 block_init(bdrv_quorum_init
);