4 * Copyright (C) 2013 Proxmox Server Solutions
7 * Dietmar Maurer (dietmar@proxmox.com)
9 * This work is licensed under the terms of the GNU GPL, version 2 or later.
10 * See the COPYING file in the top-level directory.
14 #include "qemu/osdep.h"
17 #include "block/block.h"
18 #include "block/block_int.h"
19 #include "block/blockjob_int.h"
20 #include "block/block_backup.h"
21 #include "qapi/error.h"
22 #include "qapi/qmp/qerror.h"
23 #include "qemu/ratelimit.h"
24 #include "qemu/cutils.h"
25 #include "sysemu/block-backend.h"
26 #include "qemu/bitmap.h"
27 #include "qemu/error-report.h"
29 #define BACKUP_CLUSTER_SIZE_DEFAULT (1 << 16)
31 typedef struct BackupBlockJob
{
34 /* bitmap for sync=incremental */
35 BdrvDirtyBitmap
*sync_bitmap
;
36 MirrorSyncMode sync_mode
;
37 BlockdevOnError on_source_error
;
38 BlockdevOnError on_target_error
;
39 CoRwlock flush_rwlock
;
44 NotifierWithReturn before_write
;
45 QLIST_HEAD(, CowRequest
) inflight_reqs
;
49 int64_t copy_range_size
;
51 bool serialize_target_writes
;
54 static const BlockJobDriver backup_job_driver
;
56 /* See if in-flight requests overlap and wait for them to complete */
57 static void coroutine_fn
wait_for_overlapping_requests(BackupBlockJob
*job
,
66 QLIST_FOREACH(req
, &job
->inflight_reqs
, list
) {
67 if (end
> req
->start_byte
&& start
< req
->end_byte
) {
68 qemu_co_queue_wait(&req
->wait_queue
, NULL
);
76 /* Keep track of an in-flight request */
77 static void cow_request_begin(CowRequest
*req
, BackupBlockJob
*job
,
78 int64_t start
, int64_t end
)
80 req
->start_byte
= start
;
82 qemu_co_queue_init(&req
->wait_queue
);
83 QLIST_INSERT_HEAD(&job
->inflight_reqs
, req
, list
);
86 /* Forget about a completed request */
87 static void cow_request_end(CowRequest
*req
)
89 QLIST_REMOVE(req
, list
);
90 qemu_co_queue_restart_all(&req
->wait_queue
);
93 /* Copy range to target with a bounce buffer and return the bytes copied. If
94 * error occured, return a negative error number */
95 static int coroutine_fn
backup_cow_with_bounce_buffer(BackupBlockJob
*job
,
98 bool is_write_notifier
,
100 void **bounce_buffer
)
105 BlockBackend
*blk
= job
->common
.blk
;
107 int read_flags
= is_write_notifier
? BDRV_REQ_NO_SERIALISING
: 0;
108 int write_flags
= job
->serialize_target_writes
? BDRV_REQ_SERIALISING
: 0;
110 hbitmap_reset(job
->copy_bitmap
, start
/ job
->cluster_size
, 1);
111 nbytes
= MIN(job
->cluster_size
, job
->len
- start
);
112 if (!*bounce_buffer
) {
113 *bounce_buffer
= blk_blockalign(blk
, job
->cluster_size
);
115 iov
.iov_base
= *bounce_buffer
;
116 iov
.iov_len
= nbytes
;
117 qemu_iovec_init_external(&qiov
, &iov
, 1);
119 ret
= blk_co_preadv(blk
, start
, qiov
.size
, &qiov
, read_flags
);
121 trace_backup_do_cow_read_fail(job
, start
, ret
);
123 *error_is_read
= true;
128 if (qemu_iovec_is_zero(&qiov
)) {
129 ret
= blk_co_pwrite_zeroes(job
->target
, start
,
130 qiov
.size
, write_flags
| BDRV_REQ_MAY_UNMAP
);
132 ret
= blk_co_pwritev(job
->target
, start
,
133 qiov
.size
, &qiov
, write_flags
|
134 (job
->compress
? BDRV_REQ_WRITE_COMPRESSED
: 0));
137 trace_backup_do_cow_write_fail(job
, start
, ret
);
139 *error_is_read
= false;
146 hbitmap_set(job
->copy_bitmap
, start
/ job
->cluster_size
, 1);
151 /* Copy range to target and return the bytes copied. If error occured, return a
152 * negative error number. */
153 static int coroutine_fn
backup_cow_with_offload(BackupBlockJob
*job
,
156 bool is_write_notifier
)
160 BlockBackend
*blk
= job
->common
.blk
;
162 int read_flags
= is_write_notifier
? BDRV_REQ_NO_SERIALISING
: 0;
163 int write_flags
= job
->serialize_target_writes
? BDRV_REQ_SERIALISING
: 0;
165 assert(QEMU_IS_ALIGNED(job
->copy_range_size
, job
->cluster_size
));
166 nbytes
= MIN(job
->copy_range_size
, end
- start
);
167 nr_clusters
= DIV_ROUND_UP(nbytes
, job
->cluster_size
);
168 hbitmap_reset(job
->copy_bitmap
, start
/ job
->cluster_size
,
170 ret
= blk_co_copy_range(blk
, start
, job
->target
, start
, nbytes
,
171 read_flags
, write_flags
);
173 trace_backup_do_cow_copy_range_fail(job
, start
, ret
);
174 hbitmap_set(job
->copy_bitmap
, start
/ job
->cluster_size
,
182 static int coroutine_fn
backup_do_cow(BackupBlockJob
*job
,
183 int64_t offset
, uint64_t bytes
,
185 bool is_write_notifier
)
187 CowRequest cow_request
;
189 int64_t start
, end
; /* bytes */
190 void *bounce_buffer
= NULL
;
192 qemu_co_rwlock_rdlock(&job
->flush_rwlock
);
194 start
= QEMU_ALIGN_DOWN(offset
, job
->cluster_size
);
195 end
= QEMU_ALIGN_UP(bytes
+ offset
, job
->cluster_size
);
197 trace_backup_do_cow_enter(job
, start
, offset
, bytes
);
199 wait_for_overlapping_requests(job
, start
, end
);
200 cow_request_begin(&cow_request
, job
, start
, end
);
202 while (start
< end
) {
203 if (!hbitmap_get(job
->copy_bitmap
, start
/ job
->cluster_size
)) {
204 trace_backup_do_cow_skip(job
, start
);
205 start
+= job
->cluster_size
;
206 continue; /* already copied */
209 trace_backup_do_cow_process(job
, start
);
211 if (job
->use_copy_range
) {
212 ret
= backup_cow_with_offload(job
, start
, end
, is_write_notifier
);
214 job
->use_copy_range
= false;
217 if (!job
->use_copy_range
) {
218 ret
= backup_cow_with_bounce_buffer(job
, start
, end
, is_write_notifier
,
219 error_is_read
, &bounce_buffer
);
225 /* Publish progress, guest I/O counts as progress too. Note that the
226 * offset field is an opaque progress value, it is not a disk offset.
229 job
->bytes_read
+= ret
;
230 job_progress_update(&job
->common
.job
, ret
);
235 qemu_vfree(bounce_buffer
);
238 cow_request_end(&cow_request
);
240 trace_backup_do_cow_return(job
, offset
, bytes
, ret
);
242 qemu_co_rwlock_unlock(&job
->flush_rwlock
);
247 static int coroutine_fn
backup_before_write_notify(
248 NotifierWithReturn
*notifier
,
251 BackupBlockJob
*job
= container_of(notifier
, BackupBlockJob
, before_write
);
252 BdrvTrackedRequest
*req
= opaque
;
254 assert(req
->bs
== blk_bs(job
->common
.blk
));
255 assert(QEMU_IS_ALIGNED(req
->offset
, BDRV_SECTOR_SIZE
));
256 assert(QEMU_IS_ALIGNED(req
->bytes
, BDRV_SECTOR_SIZE
));
258 return backup_do_cow(job
, req
->offset
, req
->bytes
, NULL
, true);
261 static void backup_cleanup_sync_bitmap(BackupBlockJob
*job
, int ret
)
264 BlockDriverState
*bs
= blk_bs(job
->common
.blk
);
267 /* Merge the successor back into the parent, delete nothing. */
268 bm
= bdrv_reclaim_dirty_bitmap(bs
, job
->sync_bitmap
, NULL
);
271 /* Everything is fine, delete this bitmap and install the backup. */
272 bm
= bdrv_dirty_bitmap_abdicate(bs
, job
->sync_bitmap
, NULL
);
277 static void backup_commit(Job
*job
)
279 BackupBlockJob
*s
= container_of(job
, BackupBlockJob
, common
.job
);
280 if (s
->sync_bitmap
) {
281 backup_cleanup_sync_bitmap(s
, 0);
285 static void backup_abort(Job
*job
)
287 BackupBlockJob
*s
= container_of(job
, BackupBlockJob
, common
.job
);
288 if (s
->sync_bitmap
) {
289 backup_cleanup_sync_bitmap(s
, -1);
293 static void backup_clean(Job
*job
)
295 BackupBlockJob
*s
= container_of(job
, BackupBlockJob
, common
.job
);
297 blk_unref(s
->target
);
301 static void backup_attached_aio_context(BlockJob
*job
, AioContext
*aio_context
)
303 BackupBlockJob
*s
= container_of(job
, BackupBlockJob
, common
);
305 blk_set_aio_context(s
->target
, aio_context
);
308 void backup_do_checkpoint(BlockJob
*job
, Error
**errp
)
310 BackupBlockJob
*backup_job
= container_of(job
, BackupBlockJob
, common
);
313 assert(block_job_driver(job
) == &backup_job_driver
);
315 if (backup_job
->sync_mode
!= MIRROR_SYNC_MODE_NONE
) {
316 error_setg(errp
, "The backup job only supports block checkpoint in"
321 len
= DIV_ROUND_UP(backup_job
->len
, backup_job
->cluster_size
);
322 hbitmap_set(backup_job
->copy_bitmap
, 0, len
);
325 void backup_wait_for_overlapping_requests(BlockJob
*job
, int64_t offset
,
328 BackupBlockJob
*backup_job
= container_of(job
, BackupBlockJob
, common
);
331 assert(block_job_driver(job
) == &backup_job_driver
);
333 start
= QEMU_ALIGN_DOWN(offset
, backup_job
->cluster_size
);
334 end
= QEMU_ALIGN_UP(offset
+ bytes
, backup_job
->cluster_size
);
335 wait_for_overlapping_requests(backup_job
, start
, end
);
338 void backup_cow_request_begin(CowRequest
*req
, BlockJob
*job
,
339 int64_t offset
, uint64_t bytes
)
341 BackupBlockJob
*backup_job
= container_of(job
, BackupBlockJob
, common
);
344 assert(block_job_driver(job
) == &backup_job_driver
);
346 start
= QEMU_ALIGN_DOWN(offset
, backup_job
->cluster_size
);
347 end
= QEMU_ALIGN_UP(offset
+ bytes
, backup_job
->cluster_size
);
348 cow_request_begin(req
, backup_job
, start
, end
);
351 void backup_cow_request_end(CowRequest
*req
)
353 cow_request_end(req
);
356 static void backup_drain(BlockJob
*job
)
358 BackupBlockJob
*s
= container_of(job
, BackupBlockJob
, common
);
360 /* Need to keep a reference in case blk_drain triggers execution
361 * of backup_complete...
364 BlockBackend
*target
= s
->target
;
371 static BlockErrorAction
backup_error_action(BackupBlockJob
*job
,
372 bool read
, int error
)
375 return block_job_error_action(&job
->common
, job
->on_source_error
,
378 return block_job_error_action(&job
->common
, job
->on_target_error
,
385 } BackupCompleteData
;
387 static void backup_complete(Job
*job
, void *opaque
)
389 BackupCompleteData
*data
= opaque
;
391 job_completed(job
, data
->ret
, NULL
);
395 static bool coroutine_fn
yield_and_check(BackupBlockJob
*job
)
399 if (job_is_cancelled(&job
->common
.job
)) {
403 /* We need to yield even for delay_ns = 0 so that bdrv_drain_all() can
404 * return. Without a yield, the VM would not reboot. */
405 delay_ns
= block_job_ratelimit_get_delay(&job
->common
, job
->bytes_read
);
407 job_sleep_ns(&job
->common
.job
, delay_ns
);
409 if (job_is_cancelled(&job
->common
.job
)) {
416 static int coroutine_fn
backup_run_incremental(BackupBlockJob
*job
)
423 hbitmap_iter_init(&hbi
, job
->copy_bitmap
, 0);
424 while ((cluster
= hbitmap_iter_next(&hbi
, true)) != -1) {
426 if (yield_and_check(job
)) {
429 ret
= backup_do_cow(job
, cluster
* job
->cluster_size
,
430 job
->cluster_size
, &error_is_read
, false);
431 if (ret
< 0 && backup_error_action(job
, error_is_read
, -ret
) ==
432 BLOCK_ERROR_ACTION_REPORT
)
442 /* init copy_bitmap from sync_bitmap */
443 static void backup_incremental_init_copy_bitmap(BackupBlockJob
*job
)
445 BdrvDirtyBitmapIter
*dbi
;
447 int64_t end
= DIV_ROUND_UP(bdrv_dirty_bitmap_size(job
->sync_bitmap
),
450 dbi
= bdrv_dirty_iter_new(job
->sync_bitmap
);
451 while ((offset
= bdrv_dirty_iter_next(dbi
)) != -1) {
452 int64_t cluster
= offset
/ job
->cluster_size
;
453 int64_t next_cluster
;
455 offset
+= bdrv_dirty_bitmap_granularity(job
->sync_bitmap
);
456 if (offset
>= bdrv_dirty_bitmap_size(job
->sync_bitmap
)) {
457 hbitmap_set(job
->copy_bitmap
, cluster
, end
- cluster
);
461 offset
= bdrv_dirty_bitmap_next_zero(job
->sync_bitmap
, offset
);
463 hbitmap_set(job
->copy_bitmap
, cluster
, end
- cluster
);
467 next_cluster
= DIV_ROUND_UP(offset
, job
->cluster_size
);
468 hbitmap_set(job
->copy_bitmap
, cluster
, next_cluster
- cluster
);
469 if (next_cluster
>= end
) {
473 bdrv_set_dirty_iter(dbi
, next_cluster
* job
->cluster_size
);
476 /* TODO job_progress_set_remaining() would make more sense */
477 job_progress_update(&job
->common
.job
,
478 job
->len
- hbitmap_count(job
->copy_bitmap
) * job
->cluster_size
);
480 bdrv_dirty_iter_free(dbi
);
483 static void coroutine_fn
backup_run(void *opaque
)
485 BackupBlockJob
*job
= opaque
;
486 BackupCompleteData
*data
;
487 BlockDriverState
*bs
= blk_bs(job
->common
.blk
);
488 int64_t offset
, nb_clusters
;
491 QLIST_INIT(&job
->inflight_reqs
);
492 qemu_co_rwlock_init(&job
->flush_rwlock
);
494 nb_clusters
= DIV_ROUND_UP(job
->len
, job
->cluster_size
);
495 job_progress_set_remaining(&job
->common
.job
, job
->len
);
497 job
->copy_bitmap
= hbitmap_alloc(nb_clusters
, 0);
498 if (job
->sync_mode
== MIRROR_SYNC_MODE_INCREMENTAL
) {
499 backup_incremental_init_copy_bitmap(job
);
501 hbitmap_set(job
->copy_bitmap
, 0, nb_clusters
);
505 job
->before_write
.notify
= backup_before_write_notify
;
506 bdrv_add_before_write_notifier(bs
, &job
->before_write
);
508 if (job
->sync_mode
== MIRROR_SYNC_MODE_NONE
) {
509 /* All bits are set in copy_bitmap to allow any cluster to be copied.
510 * This does not actually require them to be copied. */
511 while (!job_is_cancelled(&job
->common
.job
)) {
512 /* Yield until the job is cancelled. We just let our before_write
513 * notify callback service CoW requests. */
514 job_yield(&job
->common
.job
);
516 } else if (job
->sync_mode
== MIRROR_SYNC_MODE_INCREMENTAL
) {
517 ret
= backup_run_incremental(job
);
519 /* Both FULL and TOP SYNC_MODE's require copying.. */
520 for (offset
= 0; offset
< job
->len
;
521 offset
+= job
->cluster_size
) {
525 if (yield_and_check(job
)) {
529 if (job
->sync_mode
== MIRROR_SYNC_MODE_TOP
) {
533 /* Check to see if these blocks are already in the
536 for (i
= 0; i
< job
->cluster_size
;) {
537 /* bdrv_is_allocated() only returns true/false based
538 * on the first set of sectors it comes across that
539 * are are all in the same state.
540 * For that reason we must verify each sector in the
541 * backup cluster length. We end up copying more than
542 * needed but at some point that is always the case. */
544 bdrv_is_allocated(bs
, offset
+ i
,
545 job
->cluster_size
- i
, &n
);
548 if (alloced
|| n
== 0) {
553 /* If the above loop never found any sectors that are in
554 * the topmost image, skip this backup. */
559 /* FULL sync mode we copy the whole drive. */
563 ret
= backup_do_cow(job
, offset
, job
->cluster_size
,
564 &error_is_read
, false);
567 /* Depending on error action, fail now or retry cluster */
568 BlockErrorAction action
=
569 backup_error_action(job
, error_is_read
, -ret
);
570 if (action
== BLOCK_ERROR_ACTION_REPORT
) {
573 offset
-= job
->cluster_size
;
580 notifier_with_return_remove(&job
->before_write
);
582 /* wait until pending backup_do_cow() calls have completed */
583 qemu_co_rwlock_wrlock(&job
->flush_rwlock
);
584 qemu_co_rwlock_unlock(&job
->flush_rwlock
);
585 hbitmap_free(job
->copy_bitmap
);
587 data
= g_malloc(sizeof(*data
));
589 job_defer_to_main_loop(&job
->common
.job
, backup_complete
, data
);
592 static const BlockJobDriver backup_job_driver
= {
594 .instance_size
= sizeof(BackupBlockJob
),
595 .job_type
= JOB_TYPE_BACKUP
,
596 .free
= block_job_free
,
597 .user_resume
= block_job_user_resume
,
598 .drain
= block_job_drain
,
600 .commit
= backup_commit
,
601 .abort
= backup_abort
,
602 .clean
= backup_clean
,
604 .attached_aio_context
= backup_attached_aio_context
,
605 .drain
= backup_drain
,
608 BlockJob
*backup_job_create(const char *job_id
, BlockDriverState
*bs
,
609 BlockDriverState
*target
, int64_t speed
,
610 MirrorSyncMode sync_mode
, BdrvDirtyBitmap
*sync_bitmap
,
612 BlockdevOnError on_source_error
,
613 BlockdevOnError on_target_error
,
615 BlockCompletionFunc
*cb
, void *opaque
,
616 JobTxn
*txn
, Error
**errp
)
620 BackupBlockJob
*job
= NULL
;
627 error_setg(errp
, "Source and target cannot be the same");
631 if (!bdrv_is_inserted(bs
)) {
632 error_setg(errp
, "Device is not inserted: %s",
633 bdrv_get_device_name(bs
));
637 if (!bdrv_is_inserted(target
)) {
638 error_setg(errp
, "Device is not inserted: %s",
639 bdrv_get_device_name(target
));
643 if (compress
&& target
->drv
->bdrv_co_pwritev_compressed
== NULL
) {
644 error_setg(errp
, "Compression is not supported for this drive %s",
645 bdrv_get_device_name(target
));
649 if (bdrv_op_is_blocked(bs
, BLOCK_OP_TYPE_BACKUP_SOURCE
, errp
)) {
653 if (bdrv_op_is_blocked(target
, BLOCK_OP_TYPE_BACKUP_TARGET
, errp
)) {
657 if (sync_mode
== MIRROR_SYNC_MODE_INCREMENTAL
) {
659 error_setg(errp
, "must provide a valid bitmap name for "
660 "\"incremental\" sync mode");
664 /* Create a new bitmap, and freeze/disable this one. */
665 if (bdrv_dirty_bitmap_create_successor(bs
, sync_bitmap
, errp
) < 0) {
668 } else if (sync_bitmap
) {
670 "a sync_bitmap was provided to backup_run, "
671 "but received an incompatible sync_mode (%s)",
672 MirrorSyncMode_str(sync_mode
));
676 len
= bdrv_getlength(bs
);
678 error_setg_errno(errp
, -len
, "unable to get length for '%s'",
679 bdrv_get_device_name(bs
));
683 /* job->len is fixed, so we can't allow resize */
684 job
= block_job_create(job_id
, &backup_job_driver
, txn
, bs
,
685 BLK_PERM_CONSISTENT_READ
,
686 BLK_PERM_CONSISTENT_READ
| BLK_PERM_WRITE
|
687 BLK_PERM_WRITE_UNCHANGED
| BLK_PERM_GRAPH_MOD
,
688 speed
, creation_flags
, cb
, opaque
, errp
);
693 /* The target must match the source in size, so no resize here either */
694 job
->target
= blk_new(BLK_PERM_WRITE
,
695 BLK_PERM_CONSISTENT_READ
| BLK_PERM_WRITE
|
696 BLK_PERM_WRITE_UNCHANGED
| BLK_PERM_GRAPH_MOD
);
697 ret
= blk_insert_bs(job
->target
, target
, errp
);
702 job
->on_source_error
= on_source_error
;
703 job
->on_target_error
= on_target_error
;
704 job
->sync_mode
= sync_mode
;
705 job
->sync_bitmap
= sync_mode
== MIRROR_SYNC_MODE_INCREMENTAL
?
707 job
->compress
= compress
;
709 /* Detect image-fleecing (and similar) schemes */
710 job
->serialize_target_writes
= bdrv_chain_contains(target
, bs
);
712 /* If there is no backing file on the target, we cannot rely on COW if our
713 * backup cluster size is smaller than the target cluster size. Even for
714 * targets with a backing file, try to avoid COW if possible. */
715 ret
= bdrv_get_info(target
, &bdi
);
716 if (ret
== -ENOTSUP
&& !target
->backing
) {
717 /* Cluster size is not defined */
718 warn_report("The target block device doesn't provide "
719 "information about the block size and it doesn't have a "
720 "backing file. The default block size of %u bytes is "
721 "used. If the actual block size of the target exceeds "
722 "this default, the backup may be unusable",
723 BACKUP_CLUSTER_SIZE_DEFAULT
);
724 job
->cluster_size
= BACKUP_CLUSTER_SIZE_DEFAULT
;
725 } else if (ret
< 0 && !target
->backing
) {
726 error_setg_errno(errp
, -ret
,
727 "Couldn't determine the cluster size of the target image, "
728 "which has no backing file");
729 error_append_hint(errp
,
730 "Aborting, since this may create an unusable destination image\n");
732 } else if (ret
< 0 && target
->backing
) {
733 /* Not fatal; just trudge on ahead. */
734 job
->cluster_size
= BACKUP_CLUSTER_SIZE_DEFAULT
;
736 job
->cluster_size
= MAX(BACKUP_CLUSTER_SIZE_DEFAULT
, bdi
.cluster_size
);
738 job
->use_copy_range
= true;
739 job
->copy_range_size
= MIN_NON_ZERO(blk_get_max_transfer(job
->common
.blk
),
740 blk_get_max_transfer(job
->target
));
741 job
->copy_range_size
= MAX(job
->cluster_size
,
742 QEMU_ALIGN_UP(job
->copy_range_size
,
745 /* Required permissions are already taken with target's blk_new() */
746 block_job_add_bdrv(&job
->common
, "target", target
, 0, BLK_PERM_ALL
,
754 bdrv_reclaim_dirty_bitmap(bs
, sync_bitmap
, NULL
);
757 backup_clean(&job
->common
.job
);
758 job_early_fail(&job
->common
.job
);