blob: 46be65a95f10b9e095027f4d986f60b406adb8ce [file] [log] [blame]
Benoît Canet27cec152014-02-21 22:21:10 +01001/*
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
Peter Maydell80c71a22016-01-18 18:01:42 +000016#include "qemu/osdep.h"
Wen Congyang98292c62016-05-10 15:36:38 +080017#include "qemu/cutils.h"
Markus Armbruster0b8fa322019-05-23 16:35:07 +020018#include "qemu/module.h"
Markus Armbruster922a01a2018-02-01 12:18:46 +010019#include "qemu/option.h"
Peter Maydell5df022c2022-02-26 18:07:23 +000020#include "qemu/memalign.h"
Benoît Canet27cec152014-02-21 22:21:10 +010021#include "block/block_int.h"
Alberto Garciaef9bba12020-11-13 17:52:31 +010022#include "block/coroutines.h"
Max Reitz609f45e2018-06-14 21:14:28 +020023#include "block/qdict.h"
Markus Armbrustere688df62018-02-01 12:18:31 +010024#include "qapi/error.h"
Markus Armbruster9af23982018-02-11 10:36:01 +010025#include "qapi/qapi-events-block.h"
Max Reitzfafcfe22014-07-18 20:25:00 +020026#include "qapi/qmp/qdict.h"
Markus Armbrustercc7a8ea2015-03-17 17:22:46 +010027#include "qapi/qmp/qerror.h"
Max Reitzfafcfe22014-07-18 20:25:00 +020028#include "qapi/qmp/qlist.h"
29#include "qapi/qmp/qstring.h"
Daniel P. Berrange488981a42015-07-01 18:10:35 +010030#include "crypto/hash.h"
Benoît Canet95c6bff2014-02-21 22:21:15 +010031
32#define HASH_LENGTH 32
33
Lukas Straub5eb9a3c2020-08-04 12:46:42 +020034#define INDEXSTR_LEN 32
35
Benoît Canetc88a1de2014-02-21 22:21:20 +010036#define QUORUM_OPT_VOTE_THRESHOLD "vote-threshold"
37#define QUORUM_OPT_BLKVERIFY "blkverify"
Benoît Canetcf29a572014-06-11 15:24:10 +020038#define QUORUM_OPT_REWRITE "rewrite-corrupted"
Liu Yuana9db86b2014-08-18 17:41:05 +080039#define QUORUM_OPT_READ_PATTERN "read-pattern"
Benoît Canetc88a1de2014-02-21 22:21:20 +010040
Benoît Canet95c6bff2014-02-21 22:21:15 +010041/* This union holds a vote hash value */
42typedef union QuorumVoteValue {
Daniel P. Berrange488981a42015-07-01 18:10:35 +010043 uint8_t h[HASH_LENGTH]; /* SHA-256 hash */
Benoît Canet95c6bff2014-02-21 22:21:15 +010044 int64_t l; /* simpler 64 bits hash */
45} QuorumVoteValue;
46
47/* A vote item */
48typedef struct QuorumVoteItem {
49 int index;
50 QLIST_ENTRY(QuorumVoteItem) next;
51} QuorumVoteItem;
52
53/* this structure is a vote version. A version is the set of votes sharing the
54 * same vote value.
55 * The set of votes will be tracked with the items field and its cardinality is
56 * vote_count.
57 */
58typedef struct QuorumVoteVersion {
59 QuorumVoteValue value;
60 int index;
61 int vote_count;
62 QLIST_HEAD(, QuorumVoteItem) items;
63 QLIST_ENTRY(QuorumVoteVersion) next;
64} QuorumVoteVersion;
65
66/* this structure holds a group of vote versions together */
67typedef struct QuorumVotes {
68 QLIST_HEAD(, QuorumVoteVersion) vote_list;
69 bool (*compare)(QuorumVoteValue *a, QuorumVoteValue *b);
70} QuorumVotes;
Benoît Canet27cec152014-02-21 22:21:10 +010071
Benoît Canetcadebd72014-02-21 22:21:11 +010072/* the following structure holds the state of one quorum instance */
73typedef struct BDRVQuorumState {
Kevin Wolf0bd6e912015-06-16 11:29:22 +020074 BdrvChild **children; /* children BlockDriverStates */
Benoît Canetcadebd72014-02-21 22:21:11 +010075 int num_children; /* children count */
Wen Congyang98292c62016-05-10 15:36:38 +080076 unsigned next_child_index; /* the index of the next child that should
77 * be added
78 */
Benoît Canetcadebd72014-02-21 22:21:11 +010079 int threshold; /* if less than threshold children reads gave the
80 * same result a quorum error occurs.
81 */
82 bool is_blkverify; /* true if the driver is in blkverify mode
83 * Writes are mirrored on two children devices.
84 * On reads the two children devices' contents are
85 * compared and if a difference is spotted its
86 * location is printed and the code aborts.
87 * It is useful to debug other block drivers by
88 * comparing them with a reference one.
89 */
Benoît Canetcf29a572014-06-11 15:24:10 +020090 bool rewrite_corrupted;/* true if the driver must rewrite-on-read corrupted
91 * block if Quorum is reached.
92 */
Liu Yuana9db86b2014-08-18 17:41:05 +080093
94 QuorumReadPattern read_pattern;
Benoît Canetcadebd72014-02-21 22:21:11 +010095} BDRVQuorumState;
96
Benoît Canet27cec152014-02-21 22:21:10 +010097typedef struct QuorumAIOCB QuorumAIOCB;
98
99/* Quorum will create one instance of the following structure per operation it
100 * performs on its children.
101 * So for each read/write operation coming from the upper layer there will be
102 * $children_count QuorumChildRequest.
103 */
104typedef struct QuorumChildRequest {
Kevin Wolfce15dc02016-11-08 11:10:14 +0100105 BlockDriverState *bs;
Benoît Canet27cec152014-02-21 22:21:10 +0100106 QEMUIOVector qiov;
107 uint8_t *buf;
108 int ret;
109 QuorumAIOCB *parent;
110} QuorumChildRequest;
111
112/* Quorum will use the following structure to track progress of each read/write
113 * operation received by the upper layer.
114 * This structure hold pointers to the QuorumChildRequest structures instances
115 * used to do operations on each children and track overall progress.
116 */
117struct QuorumAIOCB {
Kevin Wolfce15dc02016-11-08 11:10:14 +0100118 BlockDriverState *bs;
119 Coroutine *co;
Benoît Canet27cec152014-02-21 22:21:10 +0100120
121 /* Request metadata */
Kevin Wolf6847da32016-11-10 17:22:07 +0100122 uint64_t offset;
123 uint64_t bytes;
Max Reitz1b1a9202018-04-21 15:29:25 +0200124 int flags;
Benoît Canet27cec152014-02-21 22:21:10 +0100125
126 QEMUIOVector *qiov; /* calling IOV */
127
128 QuorumChildRequest *qcrs; /* individual child requests */
129 int count; /* number of completed AIOCB */
130 int success_count; /* number of successfully completed AIOCB */
131
Benoît Canetcf29a572014-06-11 15:24:10 +0200132 int rewrite_count; /* number of replica to rewrite: count down to
133 * zero once writes are fired
134 */
135
Benoît Canet95c6bff2014-02-21 22:21:15 +0100136 QuorumVotes votes;
137
Benoît Canet27cec152014-02-21 22:21:10 +0100138 bool is_read;
139 int vote_ret;
Paolo Bonzini86ec2522016-10-05 18:35:26 +0200140 int children_read; /* how many children have been read from */
Benoît Canet27cec152014-02-21 22:21:10 +0100141};
Benoît Canetcadebd72014-02-21 22:21:11 +0100142
Kevin Wolfce15dc02016-11-08 11:10:14 +0100143typedef struct QuorumCo {
144 QuorumAIOCB *acb;
145 int idx;
146} QuorumCo;
147
Benoît Canet13e79562014-02-21 22:21:12 +0100148static void quorum_aio_finalize(QuorumAIOCB *acb)
149{
Benoît Canet13e79562014-02-21 22:21:12 +0100150 g_free(acb->qcrs);
Kevin Wolf0f319772016-11-10 14:24:27 +0100151 g_free(acb);
Benoît Canet13e79562014-02-21 22:21:12 +0100152}
153
Benoît Canet95c6bff2014-02-21 22:21:15 +0100154static bool quorum_sha256_compare(QuorumVoteValue *a, QuorumVoteValue *b)
155{
156 return !memcmp(a->h, b->h, HASH_LENGTH);
157}
158
159static bool quorum_64bits_compare(QuorumVoteValue *a, QuorumVoteValue *b)
160{
161 return a->l == b->l;
162}
163
Paolo Bonzini2987ae72022-09-22 10:49:16 +0200164static QuorumAIOCB *coroutine_fn quorum_aio_get(BlockDriverState *bs,
165 QEMUIOVector *qiov,
166 uint64_t offset, uint64_t bytes,
167 int flags)
Benoît Canet13e79562014-02-21 22:21:12 +0100168{
Kevin Wolf10c85512016-11-07 18:00:29 +0100169 BDRVQuorumState *s = bs->opaque;
Kevin Wolfce15dc02016-11-08 11:10:14 +0100170 QuorumAIOCB *acb = g_new(QuorumAIOCB, 1);
Benoît Canet13e79562014-02-21 22:21:12 +0100171 int i;
172
Kevin Wolf7c37f942016-11-22 12:49:49 +0100173 *acb = (QuorumAIOCB) {
174 .co = qemu_coroutine_self(),
175 .bs = bs,
176 .offset = offset,
177 .bytes = bytes,
Max Reitz1b1a9202018-04-21 15:29:25 +0200178 .flags = flags,
Kevin Wolf7c37f942016-11-22 12:49:49 +0100179 .qiov = qiov,
180 .votes.compare = quorum_sha256_compare,
181 .votes.vote_list = QLIST_HEAD_INITIALIZER(acb.votes.vote_list),
182 };
Benoît Canet13e79562014-02-21 22:21:12 +0100183
Kevin Wolf7c37f942016-11-22 12:49:49 +0100184 acb->qcrs = g_new0(QuorumChildRequest, s->num_children);
Benoît Canet13e79562014-02-21 22:21:12 +0100185 for (i = 0; i < s->num_children; i++) {
186 acb->qcrs[i].buf = NULL;
187 acb->qcrs[i].ret = 0;
188 acb->qcrs[i].parent = acb;
189 }
190
191 return acb;
192}
193
Kevin Wolf6847da32016-11-10 17:22:07 +0100194static void quorum_report_bad(QuorumOpType type, uint64_t offset,
195 uint64_t bytes, char *node_name, int ret)
Benoît Canet95c6bff2014-02-21 22:21:15 +0100196{
Wenchao Xiafe069d92014-06-18 08:43:53 +0200197 const char *msg = NULL;
Kevin Wolf6847da32016-11-10 17:22:07 +0100198 int64_t start_sector = offset / BDRV_SECTOR_SIZE;
199 int64_t end_sector = DIV_ROUND_UP(offset + bytes, BDRV_SECTOR_SIZE);
200
Benoît Canet0c762732014-02-22 18:43:41 +0100201 if (ret < 0) {
Wenchao Xiafe069d92014-06-18 08:43:53 +0200202 msg = strerror(-ret);
Benoît Canet0c762732014-02-22 18:43:41 +0100203 }
Changlong Xie0ae053b2016-02-26 09:39:01 +0800204
Markus Armbruster54fde4f2022-11-04 17:06:52 +0100205 qapi_event_send_quorum_report_bad(type, msg, node_name, start_sector,
Peter Xu3ab72382018-08-15 21:37:37 +0800206 end_sector - start_sector);
Benoît Canet95c6bff2014-02-21 22:21:15 +0100207}
208
Kevin Wolf4026f1c2023-09-29 16:51:47 +0200209static void GRAPH_RDLOCK quorum_report_failure(QuorumAIOCB *acb)
Benoît Canet95c6bff2014-02-21 22:21:15 +0100210{
Kevin Wolfce15dc02016-11-08 11:10:14 +0100211 const char *reference = bdrv_get_device_or_node_name(acb->bs);
Kevin Wolf6847da32016-11-10 17:22:07 +0100212 int64_t start_sector = acb->offset / BDRV_SECTOR_SIZE;
213 int64_t end_sector = DIV_ROUND_UP(acb->offset + acb->bytes,
214 BDRV_SECTOR_SIZE);
215
216 qapi_event_send_quorum_failure(reference, start_sector,
Peter Xu3ab72382018-08-15 21:37:37 +0800217 end_sector - start_sector);
Benoît Canet95c6bff2014-02-21 22:21:15 +0100218}
219
220static int quorum_vote_error(QuorumAIOCB *acb);
221
Kevin Wolf4026f1c2023-09-29 16:51:47 +0200222static bool GRAPH_RDLOCK quorum_has_too_much_io_failed(QuorumAIOCB *acb)
Benoît Canet95c6bff2014-02-21 22:21:15 +0100223{
Kevin Wolfce15dc02016-11-08 11:10:14 +0100224 BDRVQuorumState *s = acb->bs->opaque;
Benoît Canet95c6bff2014-02-21 22:21:15 +0100225
226 if (acb->success_count < s->threshold) {
227 acb->vote_ret = quorum_vote_error(acb);
228 quorum_report_failure(acb);
229 return true;
230 }
231
232 return false;
233}
234
Liu Yuana9db86b2014-08-18 17:41:05 +0800235static void quorum_copy_qiov(QEMUIOVector *dest, QEMUIOVector *source)
236{
237 int i;
238 assert(dest->niov == source->niov);
239 assert(dest->size == source->size);
240 for (i = 0; i < source->niov; i++) {
241 assert(dest->iov[i].iov_len == source->iov[i].iov_len);
242 memcpy(dest->iov[i].iov_base,
243 source->iov[i].iov_base,
244 source->iov[i].iov_len);
245 }
246}
247
Paolo Bonzini1ba7e152016-10-05 18:35:27 +0200248static void quorum_report_bad_acb(QuorumChildRequest *sacb, int ret)
249{
250 QuorumAIOCB *acb = sacb->parent;
251 QuorumOpType type = acb->is_read ? QUORUM_OP_TYPE_READ : QUORUM_OP_TYPE_WRITE;
Kevin Wolf6847da32016-11-10 17:22:07 +0100252 quorum_report_bad(type, acb->offset, acb->bytes, sacb->bs->node_name, ret);
Paolo Bonzini1ba7e152016-10-05 18:35:27 +0200253}
254
Benoît Canet95c6bff2014-02-21 22:21:15 +0100255static void quorum_report_bad_versions(BDRVQuorumState *s,
256 QuorumAIOCB *acb,
257 QuorumVoteValue *value)
258{
259 QuorumVoteVersion *version;
260 QuorumVoteItem *item;
261
262 QLIST_FOREACH(version, &acb->votes.vote_list, next) {
263 if (acb->votes.compare(&version->value, value)) {
264 continue;
265 }
266 QLIST_FOREACH(item, &version->items, next) {
Kevin Wolf6847da32016-11-10 17:22:07 +0100267 quorum_report_bad(QUORUM_OP_TYPE_READ, acb->offset, acb->bytes,
Changlong Xie0ae053b2016-02-26 09:39:01 +0800268 s->children[item->index]->bs->node_name, 0);
Benoît Canet95c6bff2014-02-21 22:21:15 +0100269 }
270 }
271}
272
Kevin Wolfb9b10c32023-02-03 16:21:50 +0100273/*
274 * This function can count as GRAPH_RDLOCK because read_quorum_children() holds
275 * the graph lock and keeps it until this coroutine has terminated.
276 */
277static void coroutine_fn GRAPH_RDLOCK quorum_rewrite_entry(void *opaque)
Kevin Wolfdee66e22016-11-10 16:50:16 +0100278{
279 QuorumCo *co = opaque;
280 QuorumAIOCB *acb = co->acb;
281 BDRVQuorumState *s = acb->bs->opaque;
282
283 /* Ignore any errors, it's just a correction attempt for already
Max Reitz1b1a9202018-04-21 15:29:25 +0200284 * corrupted data.
285 * Mask out BDRV_REQ_WRITE_UNCHANGED because this overwrites the
286 * area with different data from the other children. */
Kevin Wolf6847da32016-11-10 17:22:07 +0100287 bdrv_co_pwritev(s->children[co->idx], acb->offset, acb->bytes,
Max Reitz1b1a9202018-04-21 15:29:25 +0200288 acb->qiov, acb->flags & ~BDRV_REQ_WRITE_UNCHANGED);
Kevin Wolfdee66e22016-11-10 16:50:16 +0100289
290 /* Wake up the caller after the last rewrite */
291 acb->rewrite_count--;
292 if (!acb->rewrite_count) {
293 qemu_coroutine_enter_if_inactive(acb->co);
294 }
295}
296
Kevin Wolfb9b10c32023-02-03 16:21:50 +0100297static bool coroutine_fn GRAPH_RDLOCK
298quorum_rewrite_bad_versions(QuorumAIOCB *acb, QuorumVoteValue *value)
Benoît Canetcf29a572014-06-11 15:24:10 +0200299{
300 QuorumVoteVersion *version;
301 QuorumVoteItem *item;
302 int count = 0;
303
304 /* first count the number of bad versions: done first to avoid concurrency
305 * issues.
306 */
307 QLIST_FOREACH(version, &acb->votes.vote_list, next) {
308 if (acb->votes.compare(&version->value, value)) {
309 continue;
310 }
311 QLIST_FOREACH(item, &version->items, next) {
312 count++;
313 }
314 }
315
Kevin Wolfdee66e22016-11-10 16:50:16 +0100316 /* quorum_rewrite_entry will count down this to zero */
Benoît Canetcf29a572014-06-11 15:24:10 +0200317 acb->rewrite_count = count;
318
319 /* now fire the correcting rewrites */
320 QLIST_FOREACH(version, &acb->votes.vote_list, next) {
321 if (acb->votes.compare(&version->value, value)) {
322 continue;
323 }
324 QLIST_FOREACH(item, &version->items, next) {
Kevin Wolfdee66e22016-11-10 16:50:16 +0100325 Coroutine *co;
326 QuorumCo data = {
327 .acb = acb,
328 .idx = item->index,
329 };
330
331 co = qemu_coroutine_create(quorum_rewrite_entry, &data);
332 qemu_coroutine_enter(co);
Benoît Canetcf29a572014-06-11 15:24:10 +0200333 }
334 }
335
336 /* return true if any rewrite is done else false */
337 return count;
338}
339
Benoît Canet95c6bff2014-02-21 22:21:15 +0100340static void quorum_count_vote(QuorumVotes *votes,
341 QuorumVoteValue *value,
342 int index)
343{
344 QuorumVoteVersion *v = NULL, *version = NULL;
345 QuorumVoteItem *item;
346
347 /* look if we have something with this hash */
348 QLIST_FOREACH(v, &votes->vote_list, next) {
349 if (votes->compare(&v->value, value)) {
350 version = v;
351 break;
352 }
353 }
354
355 /* It's a version not yet in the list add it */
356 if (!version) {
357 version = g_new0(QuorumVoteVersion, 1);
358 QLIST_INIT(&version->items);
359 memcpy(&version->value, value, sizeof(version->value));
360 version->index = index;
361 version->vote_count = 0;
362 QLIST_INSERT_HEAD(&votes->vote_list, version, next);
363 }
364
365 version->vote_count++;
366
367 item = g_new0(QuorumVoteItem, 1);
368 item->index = index;
369 QLIST_INSERT_HEAD(&version->items, item, next);
370}
371
372static void quorum_free_vote_list(QuorumVotes *votes)
373{
374 QuorumVoteVersion *version, *next_version;
375 QuorumVoteItem *item, *next_item;
376
377 QLIST_FOREACH_SAFE(version, &votes->vote_list, next, next_version) {
378 QLIST_REMOVE(version, next);
379 QLIST_FOREACH_SAFE(item, &version->items, next, next_item) {
380 QLIST_REMOVE(item, next);
381 g_free(item);
382 }
383 g_free(version);
384 }
385}
386
387static int quorum_compute_hash(QuorumAIOCB *acb, int i, QuorumVoteValue *hash)
388{
Benoît Canet95c6bff2014-02-21 22:21:15 +0100389 QEMUIOVector *qiov = &acb->qcrs[i].qiov;
Daniel P. Berrange488981a42015-07-01 18:10:35 +0100390 size_t len = sizeof(hash->h);
391 uint8_t *data = hash->h;
Benoît Canet95c6bff2014-02-21 22:21:15 +0100392
Daniel P. Berrange488981a42015-07-01 18:10:35 +0100393 /* XXX - would be nice if we could pass in the Error **
394 * and propagate that back, but this quorum code is
395 * restricted to just errno values currently */
Markus Armbrusteref834aa2024-09-04 13:18:28 +0200396 if (qcrypto_hash_bytesv(QCRYPTO_HASH_ALGO_SHA256,
Daniel P. Berrange488981a42015-07-01 18:10:35 +0100397 qiov->iov, qiov->niov,
398 &data, &len,
399 NULL) < 0) {
400 return -EINVAL;
Benoît Canet95c6bff2014-02-21 22:21:15 +0100401 }
402
Daniel P. Berrange488981a42015-07-01 18:10:35 +0100403 return 0;
Benoît Canet95c6bff2014-02-21 22:21:15 +0100404}
405
406static QuorumVoteVersion *quorum_get_vote_winner(QuorumVotes *votes)
407{
408 int max = 0;
409 QuorumVoteVersion *candidate, *winner = NULL;
410
411 QLIST_FOREACH(candidate, &votes->vote_list, next) {
412 if (candidate->vote_count > max) {
413 max = candidate->vote_count;
414 winner = candidate;
415 }
416 }
417
418 return winner;
419}
420
421/* qemu_iovec_compare is handy for blkverify mode because it returns the first
422 * differing byte location. Yet it is handcoded to compare vectors one byte
423 * after another so it does not benefit from the libc SIMD optimizations.
424 * quorum_iovec_compare is written for speed and should be used in the non
425 * blkverify mode of quorum.
426 */
427static bool quorum_iovec_compare(QEMUIOVector *a, QEMUIOVector *b)
428{
429 int i;
430 int result;
431
432 assert(a->niov == b->niov);
433 for (i = 0; i < a->niov; i++) {
434 assert(a->iov[i].iov_len == b->iov[i].iov_len);
435 result = memcmp(a->iov[i].iov_base,
436 b->iov[i].iov_base,
437 a->iov[i].iov_len);
438 if (result) {
439 return false;
440 }
441 }
442
443 return true;
444}
445
Alberto Garcia6840e8d2018-10-17 17:33:49 +0300446static bool quorum_compare(QuorumAIOCB *acb, QEMUIOVector *a, QEMUIOVector *b)
Benoît Canet95c6bff2014-02-21 22:21:15 +0100447{
Kevin Wolfce15dc02016-11-08 11:10:14 +0100448 BDRVQuorumState *s = acb->bs->opaque;
Benoît Canet95c6bff2014-02-21 22:21:15 +0100449 ssize_t offset;
450
451 /* This driver will replace blkverify in this particular case */
452 if (s->is_blkverify) {
453 offset = qemu_iovec_compare(a, b);
454 if (offset != -1) {
Alberto Garcia6840e8d2018-10-17 17:33:49 +0300455 fprintf(stderr, "quorum: offset=%" PRIu64 " bytes=%" PRIu64
456 " contents mismatch at offset %" PRIu64 "\n",
457 acb->offset, acb->bytes, acb->offset + offset);
458 exit(1);
Benoît Canet95c6bff2014-02-21 22:21:15 +0100459 }
460 return true;
461 }
462
463 return quorum_iovec_compare(a, b);
464}
465
466/* Do a vote to get the error code */
467static int quorum_vote_error(QuorumAIOCB *acb)
468{
Kevin Wolfce15dc02016-11-08 11:10:14 +0100469 BDRVQuorumState *s = acb->bs->opaque;
Benoît Canet95c6bff2014-02-21 22:21:15 +0100470 QuorumVoteVersion *winner = NULL;
471 QuorumVotes error_votes;
472 QuorumVoteValue result_value;
473 int i, ret = 0;
474 bool error = false;
475
476 QLIST_INIT(&error_votes.vote_list);
477 error_votes.compare = quorum_64bits_compare;
478
479 for (i = 0; i < s->num_children; i++) {
480 ret = acb->qcrs[i].ret;
481 if (ret) {
482 error = true;
483 result_value.l = ret;
484 quorum_count_vote(&error_votes, &result_value, i);
485 }
486 }
487
488 if (error) {
489 winner = quorum_get_vote_winner(&error_votes);
490 ret = winner->value.l;
491 }
492
493 quorum_free_vote_list(&error_votes);
494
495 return ret;
496}
497
Kevin Wolfb9b10c32023-02-03 16:21:50 +0100498static void coroutine_fn GRAPH_RDLOCK quorum_vote(QuorumAIOCB *acb)
Benoît Canet95c6bff2014-02-21 22:21:15 +0100499{
500 bool quorum = true;
501 int i, j, ret;
502 QuorumVoteValue hash;
Kevin Wolfce15dc02016-11-08 11:10:14 +0100503 BDRVQuorumState *s = acb->bs->opaque;
Benoît Canet95c6bff2014-02-21 22:21:15 +0100504 QuorumVoteVersion *winner;
505
506 if (quorum_has_too_much_io_failed(acb)) {
Kevin Wolf7cd9b392016-11-10 16:13:15 +0100507 return;
Benoît Canet95c6bff2014-02-21 22:21:15 +0100508 }
509
510 /* get the index of the first successful read */
511 for (i = 0; i < s->num_children; i++) {
512 if (!acb->qcrs[i].ret) {
513 break;
514 }
515 }
516
517 assert(i < s->num_children);
518
519 /* compare this read with all other successful reads stopping at quorum
520 * failure
521 */
522 for (j = i + 1; j < s->num_children; j++) {
523 if (acb->qcrs[j].ret) {
524 continue;
525 }
526 quorum = quorum_compare(acb, &acb->qcrs[i].qiov, &acb->qcrs[j].qiov);
527 if (!quorum) {
528 break;
529 }
530 }
531
532 /* Every successful read agrees */
533 if (quorum) {
534 quorum_copy_qiov(acb->qiov, &acb->qcrs[i].qiov);
Kevin Wolf7cd9b392016-11-10 16:13:15 +0100535 return;
Benoît Canet95c6bff2014-02-21 22:21:15 +0100536 }
537
538 /* compute hashes for each successful read, also store indexes */
539 for (i = 0; i < s->num_children; i++) {
540 if (acb->qcrs[i].ret) {
541 continue;
542 }
543 ret = quorum_compute_hash(acb, i, &hash);
544 /* if ever the hash computation failed */
545 if (ret < 0) {
546 acb->vote_ret = ret;
547 goto free_exit;
548 }
549 quorum_count_vote(&acb->votes, &hash, i);
550 }
551
552 /* vote to select the most represented version */
553 winner = quorum_get_vote_winner(&acb->votes);
554
555 /* if the winner count is smaller than threshold the read fails */
556 if (winner->vote_count < s->threshold) {
557 quorum_report_failure(acb);
558 acb->vote_ret = -EIO;
559 goto free_exit;
560 }
561
562 /* we have a winner: copy it */
563 quorum_copy_qiov(acb->qiov, &acb->qcrs[winner->index].qiov);
564
565 /* some versions are bad print them */
566 quorum_report_bad_versions(s, acb, &winner->value);
567
Benoît Canetcf29a572014-06-11 15:24:10 +0200568 /* corruption correction is enabled */
569 if (s->rewrite_corrupted) {
Kevin Wolfdee66e22016-11-10 16:50:16 +0100570 quorum_rewrite_bad_versions(acb, &winner->value);
Benoît Canetcf29a572014-06-11 15:24:10 +0200571 }
572
Benoît Canet95c6bff2014-02-21 22:21:15 +0100573free_exit:
574 /* free lists */
575 quorum_free_vote_list(&acb->votes);
576}
577
Kevin Wolfb9b10c32023-02-03 16:21:50 +0100578/*
579 * This function can count as GRAPH_RDLOCK because read_quorum_children() holds
580 * the graph lock and keeps it until this coroutine has terminated.
581 */
582static void coroutine_fn GRAPH_RDLOCK read_quorum_children_entry(void *opaque)
Liu Yuana9db86b2014-08-18 17:41:05 +0800583{
Kevin Wolfce15dc02016-11-08 11:10:14 +0100584 QuorumCo *co = opaque;
585 QuorumAIOCB *acb = co->acb;
586 BDRVQuorumState *s = acb->bs->opaque;
587 int i = co->idx;
Kevin Wolf7cd9b392016-11-10 16:13:15 +0100588 QuorumChildRequest *sacb = &acb->qcrs[i];
Kevin Wolfce15dc02016-11-08 11:10:14 +0100589
Kevin Wolf7cd9b392016-11-10 16:13:15 +0100590 sacb->bs = s->children[i]->bs;
Kevin Wolf6847da32016-11-10 17:22:07 +0100591 sacb->ret = bdrv_co_preadv(s->children[i], acb->offset, acb->bytes,
Kevin Wolf7cd9b392016-11-10 16:13:15 +0100592 &acb->qcrs[i].qiov, 0);
593
594 if (sacb->ret == 0) {
595 acb->success_count++;
596 } else {
597 quorum_report_bad_acb(sacb, sacb->ret);
598 }
599
600 acb->count++;
601 assert(acb->count <= s->num_children);
602 assert(acb->success_count <= s->num_children);
603
604 /* Wake up the caller after the last read */
605 if (acb->count == s->num_children) {
606 qemu_coroutine_enter_if_inactive(acb->co);
607 }
Kevin Wolfce15dc02016-11-08 11:10:14 +0100608}
609
Kevin Wolfb9b10c32023-02-03 16:21:50 +0100610static int coroutine_fn GRAPH_RDLOCK read_quorum_children(QuorumAIOCB *acb)
Kevin Wolfce15dc02016-11-08 11:10:14 +0100611{
612 BDRVQuorumState *s = acb->bs->opaque;
Laurent Vivier4a4ff4c2018-03-23 15:32:02 +0100613 int i;
Liu Yuana9db86b2014-08-18 17:41:05 +0800614
Paolo Bonzini86ec2522016-10-05 18:35:26 +0200615 acb->children_read = s->num_children;
Liu Yuana9db86b2014-08-18 17:41:05 +0800616 for (i = 0; i < s->num_children; i++) {
Kevin Wolf0bd6e912015-06-16 11:29:22 +0200617 acb->qcrs[i].buf = qemu_blockalign(s->children[i]->bs, acb->qiov->size);
Liu Yuana9db86b2014-08-18 17:41:05 +0800618 qemu_iovec_init(&acb->qcrs[i].qiov, acb->qiov->niov);
619 qemu_iovec_clone(&acb->qcrs[i].qiov, acb->qiov, acb->qcrs[i].buf);
620 }
621
622 for (i = 0; i < s->num_children; i++) {
Kevin Wolfce15dc02016-11-08 11:10:14 +0100623 Coroutine *co;
624 QuorumCo data = {
625 .acb = acb,
626 .idx = i,
627 };
628
629 co = qemu_coroutine_create(read_quorum_children_entry, &data);
630 qemu_coroutine_enter(co);
Liu Yuana9db86b2014-08-18 17:41:05 +0800631 }
632
Kevin Wolf7cd9b392016-11-10 16:13:15 +0100633 while (acb->count < s->num_children) {
634 qemu_coroutine_yield();
635 }
636
637 /* Do the vote on read */
638 quorum_vote(acb);
639 for (i = 0; i < s->num_children; i++) {
640 qemu_vfree(acb->qcrs[i].buf);
641 qemu_iovec_destroy(&acb->qcrs[i].qiov);
642 }
643
644 while (acb->rewrite_count) {
Kevin Wolfce15dc02016-11-08 11:10:14 +0100645 qemu_coroutine_yield();
646 }
647
Laurent Vivier4a4ff4c2018-03-23 15:32:02 +0100648 return acb->vote_ret;
Liu Yuana9db86b2014-08-18 17:41:05 +0800649}
650
Kevin Wolfb9b10c32023-02-03 16:21:50 +0100651static int coroutine_fn GRAPH_RDLOCK read_fifo_child(QuorumAIOCB *acb)
Liu Yuana9db86b2014-08-18 17:41:05 +0800652{
Kevin Wolfce15dc02016-11-08 11:10:14 +0100653 BDRVQuorumState *s = acb->bs->opaque;
Kevin Wolfa7e15902016-11-10 17:40:34 +0100654 int n, ret;
Liu Yuana9db86b2014-08-18 17:41:05 +0800655
Kevin Wolfa7e15902016-11-10 17:40:34 +0100656 /* We try to read the next child in FIFO order if we failed to read */
657 do {
658 n = acb->children_read++;
659 acb->qcrs[n].bs = s->children[n]->bs;
660 ret = bdrv_co_preadv(s->children[n], acb->offset, acb->bytes,
661 acb->qiov, 0);
662 if (ret < 0) {
663 quorum_report_bad_acb(&acb->qcrs[n], ret);
664 }
665 } while (ret < 0 && acb->children_read < s->num_children);
666
667 /* FIXME: rewrite failed children if acb->children_read > 1? */
Liu Yuana9db86b2014-08-18 17:41:05 +0800668
Kevin Wolfce15dc02016-11-08 11:10:14 +0100669 return ret;
Liu Yuana9db86b2014-08-18 17:41:05 +0800670}
671
Kevin Wolfb9b10c32023-02-03 16:21:50 +0100672static int coroutine_fn GRAPH_RDLOCK
673quorum_co_preadv(BlockDriverState *bs, int64_t offset, int64_t bytes,
674 QEMUIOVector *qiov, BdrvRequestFlags flags)
Benoît Canet7db69822014-02-21 22:21:14 +0100675{
676 BDRVQuorumState *s = bs->opaque;
Max Reitz1b1a9202018-04-21 15:29:25 +0200677 QuorumAIOCB *acb = quorum_aio_get(bs, qiov, offset, bytes, flags);
Kevin Wolfce15dc02016-11-08 11:10:14 +0100678 int ret;
679
Benoît Canet7db69822014-02-21 22:21:14 +0100680 acb->is_read = true;
Paolo Bonzini86ec2522016-10-05 18:35:26 +0200681 acb->children_read = 0;
Benoît Canet7db69822014-02-21 22:21:14 +0100682
Liu Yuana9db86b2014-08-18 17:41:05 +0800683 if (s->read_pattern == QUORUM_READ_PATTERN_QUORUM) {
Kevin Wolfce15dc02016-11-08 11:10:14 +0100684 ret = read_quorum_children(acb);
685 } else {
686 ret = read_fifo_child(acb);
Benoît Canet7db69822014-02-21 22:21:14 +0100687 }
Kevin Wolf0f319772016-11-10 14:24:27 +0100688 quorum_aio_finalize(acb);
689
Kevin Wolfce15dc02016-11-08 11:10:14 +0100690 return ret;
Benoît Canet7db69822014-02-21 22:21:14 +0100691}
692
Kevin Wolfabaf8b72023-02-03 16:21:48 +0100693/*
694 * This function can count as GRAPH_RDLOCK because quorum_co_pwritev() holds the
695 * graph lock and keeps it until this coroutine has terminated.
696 */
697static void coroutine_fn GRAPH_RDLOCK write_quorum_entry(void *opaque)
Kevin Wolfce15dc02016-11-08 11:10:14 +0100698{
699 QuorumCo *co = opaque;
700 QuorumAIOCB *acb = co->acb;
701 BDRVQuorumState *s = acb->bs->opaque;
702 int i = co->idx;
Kevin Wolf7cd9b392016-11-10 16:13:15 +0100703 QuorumChildRequest *sacb = &acb->qcrs[i];
Kevin Wolfce15dc02016-11-08 11:10:14 +0100704
Kevin Wolf7cd9b392016-11-10 16:13:15 +0100705 sacb->bs = s->children[i]->bs;
Alberto Garcia5cddb2e2020-11-13 17:52:32 +0100706 if (acb->flags & BDRV_REQ_ZERO_WRITE) {
707 sacb->ret = bdrv_co_pwrite_zeroes(s->children[i], acb->offset,
708 acb->bytes, acb->flags);
709 } else {
710 sacb->ret = bdrv_co_pwritev(s->children[i], acb->offset, acb->bytes,
711 acb->qiov, acb->flags);
712 }
Kevin Wolf7cd9b392016-11-10 16:13:15 +0100713 if (sacb->ret == 0) {
714 acb->success_count++;
715 } else {
716 quorum_report_bad_acb(sacb, sacb->ret);
717 }
718 acb->count++;
719 assert(acb->count <= s->num_children);
720 assert(acb->success_count <= s->num_children);
721
722 /* Wake up the caller after the last write */
723 if (acb->count == s->num_children) {
724 qemu_coroutine_enter_if_inactive(acb->co);
725 }
Kevin Wolfce15dc02016-11-08 11:10:14 +0100726}
727
Kevin Wolf7b1fb722023-02-03 16:21:49 +0100728static int coroutine_fn GRAPH_RDLOCK
729quorum_co_pwritev(BlockDriverState *bs, int64_t offset, int64_t bytes,
730 QEMUIOVector *qiov, BdrvRequestFlags flags)
Benoît Canet13e79562014-02-21 22:21:12 +0100731{
732 BDRVQuorumState *s = bs->opaque;
Max Reitz1b1a9202018-04-21 15:29:25 +0200733 QuorumAIOCB *acb = quorum_aio_get(bs, qiov, offset, bytes, flags);
Kevin Wolfce15dc02016-11-08 11:10:14 +0100734 int i, ret;
Benoît Canet13e79562014-02-21 22:21:12 +0100735
736 for (i = 0; i < s->num_children; i++) {
Kevin Wolfce15dc02016-11-08 11:10:14 +0100737 Coroutine *co;
738 QuorumCo data = {
739 .acb = acb,
740 .idx = i,
741 };
742
743 co = qemu_coroutine_create(write_quorum_entry, &data);
744 qemu_coroutine_enter(co);
Benoît Canet13e79562014-02-21 22:21:12 +0100745 }
746
Kevin Wolf7cd9b392016-11-10 16:13:15 +0100747 while (acb->count < s->num_children) {
Kevin Wolfce15dc02016-11-08 11:10:14 +0100748 qemu_coroutine_yield();
749 }
750
Kevin Wolf7cd9b392016-11-10 16:13:15 +0100751 quorum_has_too_much_io_failed(acb);
752
Kevin Wolfce15dc02016-11-08 11:10:14 +0100753 ret = acb->vote_ret;
Kevin Wolf0f319772016-11-10 14:24:27 +0100754 quorum_aio_finalize(acb);
Kevin Wolfce15dc02016-11-08 11:10:14 +0100755
756 return ret;
Benoît Canet13e79562014-02-21 22:21:12 +0100757}
758
Kevin Wolfabaf8b72023-02-03 16:21:48 +0100759static int coroutine_fn GRAPH_RDLOCK
760quorum_co_pwrite_zeroes(BlockDriverState *bs, int64_t offset, int64_t bytes,
761 BdrvRequestFlags flags)
Alberto Garcia5cddb2e2020-11-13 17:52:32 +0100762{
763 return quorum_co_pwritev(bs, offset, bytes, NULL,
764 flags | BDRV_REQ_ZERO_WRITE);
765}
766
Kevin Wolf8ab81402023-02-03 16:22:02 +0100767static int64_t coroutine_fn GRAPH_RDLOCK
768quorum_co_getlength(BlockDriverState *bs)
Benoît Canetd55dee22014-02-21 22:21:16 +0100769{
770 BDRVQuorumState *s = bs->opaque;
771 int64_t result;
772 int i;
773
774 /* check that all file have the same length */
Emanuele Giuseppe Espositoc86422c2023-01-13 21:42:04 +0100775 result = bdrv_co_getlength(s->children[0]->bs);
Benoît Canetd55dee22014-02-21 22:21:16 +0100776 if (result < 0) {
777 return result;
778 }
779 for (i = 1; i < s->num_children; i++) {
Emanuele Giuseppe Espositoc86422c2023-01-13 21:42:04 +0100780 int64_t value = bdrv_co_getlength(s->children[i]->bs);
Benoît Canetd55dee22014-02-21 22:21:16 +0100781 if (value < 0) {
782 return value;
783 }
784 if (value != result) {
785 return -EIO;
786 }
787 }
788
789 return result;
790}
791
Emanuele Giuseppe Esposito88095342023-02-03 16:21:46 +0100792static coroutine_fn GRAPH_RDLOCK int quorum_co_flush(BlockDriverState *bs)
Benoît Canet1c508d12014-02-21 22:21:18 +0100793{
794 BDRVQuorumState *s = bs->opaque;
795 QuorumVoteVersion *winner = NULL;
796 QuorumVotes error_votes;
797 QuorumVoteValue result_value;
798 int i;
799 int result = 0;
Changlong Xie924e8a22016-02-26 09:39:02 +0800800 int success_count = 0;
Benoît Canet1c508d12014-02-21 22:21:18 +0100801
802 QLIST_INIT(&error_votes.vote_list);
803 error_votes.compare = quorum_64bits_compare;
804
805 for (i = 0; i < s->num_children; i++) {
Kevin Wolf0bd6e912015-06-16 11:29:22 +0200806 result = bdrv_co_flush(s->children[i]->bs);
Changlong Xie924e8a22016-02-26 09:39:02 +0800807 if (result) {
Alberto Garcia795be062017-08-07 15:36:58 +0300808 quorum_report_bad(QUORUM_OP_TYPE_FLUSH, 0, 0,
Changlong Xie924e8a22016-02-26 09:39:02 +0800809 s->children[i]->bs->node_name, result);
810 result_value.l = result;
811 quorum_count_vote(&error_votes, &result_value, i);
812 } else {
813 success_count++;
814 }
Benoît Canet1c508d12014-02-21 22:21:18 +0100815 }
816
Changlong Xie924e8a22016-02-26 09:39:02 +0800817 if (success_count >= s->threshold) {
818 result = 0;
819 } else {
820 winner = quorum_get_vote_winner(&error_votes);
821 result = winner->value.l;
822 }
Benoît Canet1c508d12014-02-21 22:21:18 +0100823 quorum_free_vote_list(&error_votes);
824
825 return result;
826}
827
Kevin Wolf533c6e42023-05-04 13:57:49 +0200828static bool GRAPH_RDLOCK
829quorum_recurse_can_replace(BlockDriverState *bs, BlockDriverState *to_replace)
Max Reitza3ed7942020-02-18 11:34:43 +0100830{
831 BDRVQuorumState *s = bs->opaque;
832 int i;
833
834 for (i = 0; i < s->num_children; i++) {
835 /*
836 * We have no idea whether our children show the same data as
837 * this node (@bs). It is actually highly likely that
838 * @to_replace does not, because replacing a broken child is
839 * one of the main use cases here.
840 *
841 * We do know that the new BDS will match @bs, so replacing
842 * any of our children by it will be safe. It cannot change
843 * the data this quorum node presents to its parents.
844 *
845 * However, replacing @to_replace by @bs in any of our
846 * children's chains may change visible data somewhere in
847 * there. We therefore cannot recurse down those chains with
848 * bdrv_recurse_can_replace().
849 * (More formally, bdrv_recurse_can_replace() requires that
850 * @to_replace will be replaced by something matching the @bs
851 * passed to it. We cannot guarantee that.)
852 *
853 * Thus, we can only check whether any of our immediate
854 * children matches @to_replace.
855 *
856 * (In the future, we might add a function to recurse down a
857 * chain that checks that nothing there cares about a change
858 * in data from the respective child in question. For
859 * example, most filters do not care when their child's data
860 * suddenly changes, as long as their parents do not care.)
861 */
862 if (s->children[i]->bs == to_replace) {
863 /*
864 * We now have to ensure that there is no other parent
865 * that cares about replacing this child by a node with
866 * potentially different data.
867 * We do so by checking whether there are any other parents
868 * at all, which is stricter than necessary, but also very
869 * simple. (We may decide to implement something more
870 * complex and permissive when there is an actual need for
871 * it.)
872 */
873 return QLIST_FIRST(&to_replace->parents) == s->children[i] &&
874 QLIST_NEXT(s->children[i], next_parent) == NULL;
875 }
876 }
877
878 return false;
879}
880
Benoît Canetc88a1de2014-02-21 22:21:20 +0100881static int quorum_valid_threshold(int threshold, int num_children, Error **errp)
882{
883
884 if (threshold < 1) {
Markus Armbrusterc6bd8c72015-03-17 11:54:50 +0100885 error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
Markus Armbruster6cc06672020-11-13 09:26:26 +0100886 "vote-threshold", "a value >= 1");
Benoît Canetc88a1de2014-02-21 22:21:20 +0100887 return -ERANGE;
888 }
889
890 if (threshold > num_children) {
891 error_setg(errp, "threshold may not exceed children count");
892 return -ERANGE;
893 }
894
895 return 0;
896}
897
898static QemuOptsList quorum_runtime_opts = {
899 .name = "quorum",
900 .head = QTAILQ_HEAD_INITIALIZER(quorum_runtime_opts.head),
901 .desc = {
902 {
903 .name = QUORUM_OPT_VOTE_THRESHOLD,
904 .type = QEMU_OPT_NUMBER,
905 .help = "The number of vote needed for reaching quorum",
906 },
907 {
908 .name = QUORUM_OPT_BLKVERIFY,
909 .type = QEMU_OPT_BOOL,
910 .help = "Trigger block verify mode if set",
911 },
Benoît Canetcf29a572014-06-11 15:24:10 +0200912 {
913 .name = QUORUM_OPT_REWRITE,
914 .type = QEMU_OPT_BOOL,
915 .help = "Rewrite corrupted block on read quorum",
916 },
Liu Yuana9db86b2014-08-18 17:41:05 +0800917 {
918 .name = QUORUM_OPT_READ_PATTERN,
919 .type = QEMU_OPT_STRING,
920 .help = "Allowed pattern: quorum, fifo. Quorum is default",
921 },
Benoît Canetc88a1de2014-02-21 22:21:20 +0100922 { /* end of list */ }
923 },
924};
925
Alberto Garcia5cddb2e2020-11-13 17:52:32 +0100926static void quorum_refresh_flags(BlockDriverState *bs)
927{
928 BDRVQuorumState *s = bs->opaque;
929 int i;
930
931 bs->supported_zero_flags =
932 BDRV_REQ_FUA | BDRV_REQ_MAY_UNMAP | BDRV_REQ_NO_FALLBACK;
933
934 for (i = 0; i < s->num_children; i++) {
935 bs->supported_zero_flags &= s->children[i]->bs->supported_zero_flags;
936 }
937
938 bs->supported_zero_flags |= BDRV_REQ_WRITE_UNCHANGED;
939}
940
Benoît Canetc88a1de2014-02-21 22:21:20 +0100941static int quorum_open(BlockDriverState *bs, QDict *options, int flags,
942 Error **errp)
943{
944 BDRVQuorumState *s = bs->opaque;
Fam Zheng8df3abf2014-08-28 13:56:12 +0800945 QemuOpts *opts = NULL;
Marc-André Lureau8d5fb192017-08-24 10:46:03 +0200946 const char *pattern_str;
Benoît Canetc88a1de2014-02-21 22:21:20 +0100947 bool *opened;
Benoît Canetc88a1de2014-02-21 22:21:20 +0100948 int i;
949 int ret = 0;
950
951 qdict_flatten(options);
Benoît Canetc88a1de2014-02-21 22:21:20 +0100952
Kevin Wolfea6828d2015-01-21 18:49:28 +0100953 /* count how many different children are present */
954 s->num_children = qdict_array_entries(options, "children.");
955 if (s->num_children < 0) {
Markus Armbrusterdcfe4802020-07-07 18:06:01 +0200956 error_setg(errp, "Option children is not a valid array");
Max Reitz8a87f3d2014-02-21 22:30:37 +0100957 ret = -EINVAL;
958 goto exit;
959 }
Wen Congyang98292c62016-05-10 15:36:38 +0800960 if (s->num_children < 1) {
Markus Armbrusterdcfe4802020-07-07 18:06:01 +0200961 error_setg(errp, "Number of provided children must be 1 or more");
Benoît Canetc88a1de2014-02-21 22:21:20 +0100962 ret = -EINVAL;
963 goto exit;
964 }
965
966 opts = qemu_opts_create(&quorum_runtime_opts, NULL, 0, &error_abort);
Markus Armbrustera5f9b9d2020-07-07 18:06:05 +0200967 if (!qemu_opts_absorb_qdict(opts, options, errp)) {
Benoît Canetc88a1de2014-02-21 22:21:20 +0100968 ret = -EINVAL;
969 goto exit;
970 }
971
972 s->threshold = qemu_opt_get_number(opts, QUORUM_OPT_VOTE_THRESHOLD, 0);
Wen Congyang834cb2a2015-07-03 14:45:06 +0800973 /* and validate it against s->num_children */
Markus Armbrusterdcfe4802020-07-07 18:06:01 +0200974 ret = quorum_valid_threshold(s->threshold, s->num_children, errp);
Wen Congyang834cb2a2015-07-03 14:45:06 +0800975 if (ret < 0) {
976 goto exit;
977 }
978
Marc-André Lureau8d5fb192017-08-24 10:46:03 +0200979 pattern_str = qemu_opt_get(opts, QUORUM_OPT_READ_PATTERN);
980 if (!pattern_str) {
981 ret = QUORUM_READ_PATTERN_QUORUM;
982 } else {
Marc-André Lureauf7abe0e2017-08-24 10:46:10 +0200983 ret = qapi_enum_parse(&QuorumReadPattern_lookup, pattern_str,
Marc-André Lureau8d5fb192017-08-24 10:46:03 +0200984 -EINVAL, NULL);
985 }
Benoît Canetc88a1de2014-02-21 22:21:20 +0100986 if (ret < 0) {
Markus Armbrusterdcfe4802020-07-07 18:06:01 +0200987 error_setg(errp, "Please set read-pattern as fifo or quorum");
Benoît Canetc88a1de2014-02-21 22:21:20 +0100988 goto exit;
989 }
Liu Yuana9db86b2014-08-18 17:41:05 +0800990 s->read_pattern = ret;
Benoît Canetc88a1de2014-02-21 22:21:20 +0100991
Liu Yuana9db86b2014-08-18 17:41:05 +0800992 if (s->read_pattern == QUORUM_READ_PATTERN_QUORUM) {
Alberto Garcia83aedca2018-10-17 17:33:50 +0300993 s->is_blkverify = qemu_opt_get_bool(opts, QUORUM_OPT_BLKVERIFY, false);
994 if (s->is_blkverify && (s->num_children != 2 || s->threshold != 2)) {
Markus Armbrusterdcfe4802020-07-07 18:06:01 +0200995 error_setg(errp, "blkverify=on can only be set if there are "
Alberto Garcia83aedca2018-10-17 17:33:50 +0300996 "exactly two files and vote-threshold is 2");
997 ret = -EINVAL;
998 goto exit;
Liu Yuana9db86b2014-08-18 17:41:05 +0800999 }
1000
1001 s->rewrite_corrupted = qemu_opt_get_bool(opts, QUORUM_OPT_REWRITE,
1002 false);
1003 if (s->rewrite_corrupted && s->is_blkverify) {
Markus Armbrusterdcfe4802020-07-07 18:06:01 +02001004 error_setg(errp,
Liu Yuana9db86b2014-08-18 17:41:05 +08001005 "rewrite-corrupted=on cannot be used with blkverify=on");
1006 ret = -EINVAL;
1007 goto exit;
1008 }
Benoît Canetcf29a572014-06-11 15:24:10 +02001009 }
1010
Kevin Wolf0bd6e912015-06-16 11:29:22 +02001011 /* allocate the children array */
1012 s->children = g_new0(BdrvChild *, s->num_children);
Benoît Canetc88a1de2014-02-21 22:21:20 +01001013 opened = g_new0(bool, s->num_children);
1014
Kevin Wolfea6828d2015-01-21 18:49:28 +01001015 for (i = 0; i < s->num_children; i++) {
Lukas Straub5eb9a3c2020-08-04 12:46:42 +02001016 char indexstr[INDEXSTR_LEN];
1017 ret = snprintf(indexstr, INDEXSTR_LEN, "children.%d", i);
1018 assert(ret < INDEXSTR_LEN);
Max Reitz8a87f3d2014-02-21 22:30:37 +01001019
Kevin Wolf0bd6e912015-06-16 11:29:22 +02001020 s->children[i] = bdrv_open_child(NULL, options, indexstr, bs,
Max Reitz36ee58d2020-05-13 13:05:31 +02001021 &child_of_bds, BDRV_CHILD_DATA, false,
Vladimir Sementsov-Ogievskiybc520242021-02-02 15:49:45 +03001022 errp);
1023 if (!s->children[i]) {
Kevin Wolf0bd6e912015-06-16 11:29:22 +02001024 ret = -EINVAL;
Max Reitz8a87f3d2014-02-21 22:30:37 +01001025 goto close_exit;
Benoît Canetc88a1de2014-02-21 22:21:20 +01001026 }
Kevin Wolfea6828d2015-01-21 18:49:28 +01001027
Max Reitz8a87f3d2014-02-21 22:30:37 +01001028 opened[i] = true;
Benoît Canetc88a1de2014-02-21 22:21:20 +01001029 }
Wen Congyang98292c62016-05-10 15:36:38 +08001030 s->next_child_index = s->num_children;
Benoît Canetc88a1de2014-02-21 22:21:20 +01001031
Max Reitz1b1a9202018-04-21 15:29:25 +02001032 bs->supported_write_flags = BDRV_REQ_WRITE_UNCHANGED;
Alberto Garcia5cddb2e2020-11-13 17:52:32 +01001033 quorum_refresh_flags(bs);
Max Reitz1b1a9202018-04-21 15:29:25 +02001034
Benoît Canetc88a1de2014-02-21 22:21:20 +01001035 g_free(opened);
1036 goto exit;
1037
1038close_exit:
1039 /* cleanup on error */
Stefan Hajnoczi6bc30f12023-12-05 13:20:02 -05001040 bdrv_graph_wrlock();
Benoît Canetc88a1de2014-02-21 22:21:20 +01001041 for (i = 0; i < s->num_children; i++) {
1042 if (!opened[i]) {
1043 continue;
1044 }
Kevin Wolf0bd6e912015-06-16 11:29:22 +02001045 bdrv_unref_child(bs, s->children[i]);
Benoît Canetc88a1de2014-02-21 22:21:20 +01001046 }
Stefan Hajnoczi6bc30f12023-12-05 13:20:02 -05001047 bdrv_graph_wrunlock();
Kevin Wolf0bd6e912015-06-16 11:29:22 +02001048 g_free(s->children);
Benoît Canetc88a1de2014-02-21 22:21:20 +01001049 g_free(opened);
1050exit:
Fam Zheng8df3abf2014-08-28 13:56:12 +08001051 qemu_opts_del(opts);
Benoît Canetc88a1de2014-02-21 22:21:20 +01001052 return ret;
1053}
1054
1055static void quorum_close(BlockDriverState *bs)
1056{
1057 BDRVQuorumState *s = bs->opaque;
1058 int i;
1059
Stefan Hajnoczi6bc30f12023-12-05 13:20:02 -05001060 bdrv_graph_wrlock();
Benoît Canetc88a1de2014-02-21 22:21:20 +01001061 for (i = 0; i < s->num_children; i++) {
Kevin Wolf0bd6e912015-06-16 11:29:22 +02001062 bdrv_unref_child(bs, s->children[i]);
Benoît Canetc88a1de2014-02-21 22:21:20 +01001063 }
Stefan Hajnoczi6bc30f12023-12-05 13:20:02 -05001064 bdrv_graph_wrunlock();
Benoît Canetc88a1de2014-02-21 22:21:20 +01001065
Kevin Wolf0bd6e912015-06-16 11:29:22 +02001066 g_free(s->children);
Benoît Canetc88a1de2014-02-21 22:21:20 +01001067}
1068
Kevin Wolf9def6082023-09-11 11:46:20 +02001069static void GRAPH_WRLOCK
1070quorum_add_child(BlockDriverState *bs, BlockDriverState *child_bs, Error **errp)
Wen Congyang98292c62016-05-10 15:36:38 +08001071{
1072 BDRVQuorumState *s = bs->opaque;
1073 BdrvChild *child;
Lukas Straub5eb9a3c2020-08-04 12:46:42 +02001074 char indexstr[INDEXSTR_LEN];
Wen Congyang98292c62016-05-10 15:36:38 +08001075 int ret;
1076
Alberto Garcia808b27d2018-10-18 11:59:03 +03001077 if (s->is_blkverify) {
1078 error_setg(errp, "Cannot add a child to a quorum in blkverify mode");
1079 return;
1080 }
1081
Wen Congyang98292c62016-05-10 15:36:38 +08001082 assert(s->num_children <= INT_MAX / sizeof(BdrvChild *));
1083 if (s->num_children == INT_MAX / sizeof(BdrvChild *) ||
1084 s->next_child_index == UINT_MAX) {
1085 error_setg(errp, "Too many children");
1086 return;
1087 }
1088
Lukas Straub5eb9a3c2020-08-04 12:46:42 +02001089 ret = snprintf(indexstr, INDEXSTR_LEN, "children.%u", s->next_child_index);
1090 if (ret < 0 || ret >= INDEXSTR_LEN) {
Wen Congyang98292c62016-05-10 15:36:38 +08001091 error_setg(errp, "cannot generate child name");
1092 return;
1093 }
1094 s->next_child_index++;
1095
Wen Congyang98292c62016-05-10 15:36:38 +08001096 /* We can safely add the child now */
1097 bdrv_ref(child_bs);
Kevin Wolf8b2ff522016-12-20 22:21:17 +01001098
Max Reitz36ee58d2020-05-13 13:05:31 +02001099 child = bdrv_attach_child(bs, child_bs, indexstr, &child_of_bds,
1100 BDRV_CHILD_DATA, errp);
Kevin Wolf8b2ff522016-12-20 22:21:17 +01001101 if (child == NULL) {
1102 s->next_child_index--;
Kevin Wolf9def6082023-09-11 11:46:20 +02001103 return;
Kevin Wolf8b2ff522016-12-20 22:21:17 +01001104 }
Wen Congyang98292c62016-05-10 15:36:38 +08001105 s->children = g_renew(BdrvChild *, s->children, s->num_children + 1);
1106 s->children[s->num_children++] = child;
Alberto Garcia5cddb2e2020-11-13 17:52:32 +01001107 quorum_refresh_flags(bs);
Wen Congyang98292c62016-05-10 15:36:38 +08001108}
1109
Kevin Wolf9def6082023-09-11 11:46:20 +02001110static void GRAPH_WRLOCK
1111quorum_del_child(BlockDriverState *bs, BdrvChild *child, Error **errp)
Wen Congyang98292c62016-05-10 15:36:38 +08001112{
1113 BDRVQuorumState *s = bs->opaque;
Lukas Straub5eb9a3c2020-08-04 12:46:42 +02001114 char indexstr[INDEXSTR_LEN];
Wen Congyang98292c62016-05-10 15:36:38 +08001115 int i;
1116
1117 for (i = 0; i < s->num_children; i++) {
1118 if (s->children[i] == child) {
1119 break;
1120 }
1121 }
1122
1123 /* we have checked it in bdrv_del_child() */
1124 assert(i < s->num_children);
1125
1126 if (s->num_children <= s->threshold) {
1127 error_setg(errp,
1128 "The number of children cannot be lower than the vote threshold %d",
1129 s->threshold);
1130 return;
1131 }
1132
Alberto Garcia808b27d2018-10-18 11:59:03 +03001133 /* We know now that num_children > threshold, so blkverify must be false */
1134 assert(!s->is_blkverify);
1135
Lukas Straub5eb9a3c2020-08-04 12:46:42 +02001136 snprintf(indexstr, INDEXSTR_LEN, "children.%u", s->next_child_index - 1);
1137 if (!strncmp(child->name, indexstr, INDEXSTR_LEN)) {
1138 s->next_child_index--;
1139 }
1140
Wen Congyang98292c62016-05-10 15:36:38 +08001141 /* We can safely remove this child now */
1142 memmove(&s->children[i], &s->children[i + 1],
1143 (s->num_children - i - 1) * sizeof(BdrvChild *));
1144 s->children = g_renew(BdrvChild *, s->children, --s->num_children);
Kevin Wolf9def6082023-09-11 11:46:20 +02001145
Wen Congyang98292c62016-05-10 15:36:38 +08001146 bdrv_unref_child(bs, child);
1147
Alberto Garcia5cddb2e2020-11-13 17:52:32 +01001148 quorum_refresh_flags(bs);
Wen Congyang98292c62016-05-10 15:36:38 +08001149}
1150
Max Reitzabc521a2019-02-01 20:29:26 +01001151static void quorum_gather_child_options(BlockDriverState *bs, QDict *target,
1152 bool backing_overridden)
1153{
1154 BDRVQuorumState *s = bs->opaque;
1155 QList *children_list;
1156 int i;
1157
1158 /*
1159 * The generic implementation for gathering child options in
1160 * bdrv_refresh_filename() would use the names of the children
1161 * as specified for bdrv_open_child() or bdrv_attach_child(),
1162 * which is "children.%u" with %u being a value
1163 * (s->next_child_index) that is incremented each time a new child
1164 * is added (and never decremented). Since children can be
1165 * deleted at runtime, there may be gaps in that enumeration.
1166 * When creating a new quorum BDS and specifying the children for
1167 * it through runtime options, the enumeration used there may not
1168 * have any gaps, though.
1169 *
1170 * Therefore, we have to create a new gap-less enumeration here
1171 * (which we can achieve by simply putting all of the children's
1172 * full_open_options into a QList).
1173 *
1174 * XXX: Note that there are issues with the current child option
1175 * structure quorum uses (such as the fact that children do
1176 * not really have unique permanent names). Therefore, this
1177 * is going to have to change in the future and ideally we
1178 * want quorum to be covered by the generic implementation.
1179 */
1180
1181 children_list = qlist_new();
1182 qdict_put(target, "children", children_list);
1183
1184 for (i = 0; i < s->num_children; i++) {
1185 qlist_append(children_list,
1186 qobject_ref(s->children[i]->bs->full_open_options));
1187 }
1188}
1189
Max Reitzf3037bd2019-02-01 20:29:20 +01001190static char *quorum_dirname(BlockDriverState *bs, Error **errp)
1191{
1192 /* In general, there are multiple BDSs with different dirnames below this
1193 * one; so there is no unique dirname we could return (unless all are equal
1194 * by chance, or there is only one). Therefore, to be consistent, just
1195 * always return NULL. */
1196 error_setg(errp, "Cannot generate a base directory for quorum nodes");
1197 return NULL;
1198}
1199
Max Reitz37a37912020-02-18 11:34:40 +01001200static void quorum_child_perm(BlockDriverState *bs, BdrvChild *c,
Max Reitzbf8e9252020-05-13 13:05:16 +02001201 BdrvChildRole role,
Max Reitz37a37912020-02-18 11:34:40 +01001202 BlockReopenQueue *reopen_queue,
1203 uint64_t perm, uint64_t shared,
1204 uint64_t *nperm, uint64_t *nshared)
1205{
Max Reitz9ca5b0e2020-11-13 22:17:16 +01001206 BDRVQuorumState *s = bs->opaque;
1207
Max Reitz37a37912020-02-18 11:34:40 +01001208 *nperm = perm & DEFAULT_PERM_PASSTHROUGH;
Max Reitz9ca5b0e2020-11-13 22:17:16 +01001209 if (s->rewrite_corrupted) {
1210 *nperm |= BLK_PERM_WRITE;
1211 }
Max Reitz37a37912020-02-18 11:34:40 +01001212
1213 /*
1214 * We cannot share RESIZE or WRITE, as this would make the
1215 * children differ from each other.
1216 */
1217 *nshared = (shared & (BLK_PERM_CONSISTENT_READ |
1218 BLK_PERM_WRITE_UNCHANGED))
1219 | DEFAULT_PERM_UNCHANGED;
1220}
1221
Alberto Garciaef9bba12020-11-13 17:52:31 +01001222/*
1223 * Each one of the children can report different status flags even
1224 * when they contain the same data, so what this function does is
1225 * return BDRV_BLOCK_ZERO if *all* children agree that a certain
1226 * region contains zeroes, and BDRV_BLOCK_DATA otherwise.
1227 */
Kevin Wolf7ff95792023-02-03 16:21:43 +01001228static int coroutine_fn GRAPH_RDLOCK
1229quorum_co_block_status(BlockDriverState *bs, bool want_zero,
1230 int64_t offset, int64_t count,
1231 int64_t *pnum, int64_t *map, BlockDriverState **file)
Alberto Garciaef9bba12020-11-13 17:52:31 +01001232{
1233 BDRVQuorumState *s = bs->opaque;
1234 int i, ret;
1235 int64_t pnum_zero = count;
1236 int64_t pnum_data = 0;
1237
1238 for (i = 0; i < s->num_children; i++) {
1239 int64_t bytes;
1240 ret = bdrv_co_common_block_status_above(s->children[i]->bs, NULL, false,
1241 want_zero, offset, count,
1242 &bytes, NULL, NULL, NULL);
1243 if (ret < 0) {
1244 quorum_report_bad(QUORUM_OP_TYPE_READ, offset, count,
1245 s->children[i]->bs->node_name, ret);
1246 pnum_data = count;
1247 break;
1248 }
1249 /*
1250 * Even if all children agree about whether there are zeroes
1251 * or not at @offset they might disagree on the size, so use
1252 * the smallest when reporting BDRV_BLOCK_ZERO and the largest
1253 * when reporting BDRV_BLOCK_DATA.
1254 */
1255 if (ret & BDRV_BLOCK_ZERO) {
1256 pnum_zero = MIN(pnum_zero, bytes);
1257 } else {
1258 pnum_data = MAX(pnum_data, bytes);
1259 }
1260 }
1261
1262 if (pnum_data) {
1263 *pnum = pnum_data;
1264 return BDRV_BLOCK_DATA;
1265 } else {
1266 *pnum = pnum_zero;
1267 return BDRV_BLOCK_ZERO;
1268 }
1269}
1270
Max Reitz26542672019-02-01 20:29:25 +01001271static const char *const quorum_strong_runtime_opts[] = {
1272 QUORUM_OPT_VOTE_THRESHOLD,
1273 QUORUM_OPT_BLKVERIFY,
1274 QUORUM_OPT_REWRITE,
1275 QUORUM_OPT_READ_PATTERN,
1276
1277 NULL
1278};
1279
Benoît Canetcadebd72014-02-21 22:21:11 +01001280static BlockDriver bdrv_quorum = {
Stefan Hajnoczie3625d32014-05-08 16:34:46 +02001281 .format_name = "quorum",
Benoît Canetcadebd72014-02-21 22:21:11 +01001282
Stefan Hajnoczie3625d32014-05-08 16:34:46 +02001283 .instance_size = sizeof(BDRVQuorumState),
Benoît Canet13e79562014-02-21 22:21:12 +01001284
Fabiano Rosas65d2c3e2018-03-12 19:07:50 -03001285 .bdrv_open = quorum_open,
Stefan Hajnoczie3625d32014-05-08 16:34:46 +02001286 .bdrv_close = quorum_close,
Max Reitzabc521a2019-02-01 20:29:26 +01001287 .bdrv_gather_child_options = quorum_gather_child_options,
Max Reitzf3037bd2019-02-01 20:29:20 +01001288 .bdrv_dirname = quorum_dirname,
Alberto Garciaef9bba12020-11-13 17:52:31 +01001289 .bdrv_co_block_status = quorum_co_block_status,
Benoît Canetc88a1de2014-02-21 22:21:20 +01001290
Lukas Straub5529b022021-05-18 13:42:14 +02001291 .bdrv_co_flush = quorum_co_flush,
Benoît Canet1c508d12014-02-21 22:21:18 +01001292
Emanuele Giuseppe Espositoc86422c2023-01-13 21:42:04 +01001293 .bdrv_co_getlength = quorum_co_getlength,
Benoît Canetd55dee22014-02-21 22:21:16 +01001294
Kevin Wolf6847da32016-11-10 17:22:07 +01001295 .bdrv_co_preadv = quorum_co_preadv,
1296 .bdrv_co_pwritev = quorum_co_pwritev,
Alberto Garcia5cddb2e2020-11-13 17:52:32 +01001297 .bdrv_co_pwrite_zeroes = quorum_co_pwrite_zeroes,
Benoît Canet98a7a382014-02-21 22:21:19 +01001298
Wen Congyang98292c62016-05-10 15:36:38 +08001299 .bdrv_add_child = quorum_add_child,
1300 .bdrv_del_child = quorum_del_child,
1301
Max Reitz37a37912020-02-18 11:34:40 +01001302 .bdrv_child_perm = quorum_child_perm,
Kevin Wolfd7010df2016-12-15 12:28:58 +01001303
Max Reitza3ed7942020-02-18 11:34:43 +01001304 .bdrv_recurse_can_replace = quorum_recurse_can_replace,
Max Reitz26542672019-02-01 20:29:25 +01001305
1306 .strong_runtime_opts = quorum_strong_runtime_opts,
Benoît Canetcadebd72014-02-21 22:21:11 +01001307};
1308
1309static void bdrv_quorum_init(void)
1310{
Markus Armbrusteref834aa2024-09-04 13:18:28 +02001311 if (!qcrypto_hash_supports(QCRYPTO_HASH_ALGO_SHA256)) {
Sascha Silbee94867e2015-08-04 16:48:25 +02001312 /* SHA256 hash support is required for quorum device */
1313 return;
1314 }
Benoît Canetcadebd72014-02-21 22:21:11 +01001315 bdrv_register(&bdrv_quorum);
1316}
1317
1318block_init(bdrv_quorum_init);