4 * Copyright Red Hat, Inc. 2012
7 * Paolo Bonzini <pbonzini@redhat.com>
9 * This work is licensed under the terms of the GNU LGPL, version 2 or later.
10 * See the COPYING.LIB file in the top-level directory.
14 #include "qemu/osdep.h"
15 #include "qemu/cutils.h"
17 #include "block/blockjob_int.h"
18 #include "block/block_int.h"
19 #include "sysemu/block-backend.h"
20 #include "qapi/error.h"
21 #include "qapi/qmp/qerror.h"
22 #include "qemu/ratelimit.h"
23 #include "qemu/bitmap.h"
25 #define SLICE_TIME 100000000ULL /* ns */
26 #define MAX_IN_FLIGHT 16
27 #define MAX_IO_BYTES (1 << 20) /* 1 Mb */
28 #define DEFAULT_MIRROR_BUF_SIZE (MAX_IN_FLIGHT * MAX_IO_BYTES)
30 /* The mirroring buffer is a list of granularity-sized chunks.
31 * Free chunks are organized in a list.
33 typedef struct MirrorBuffer
{
34 QSIMPLEQ_ENTRY(MirrorBuffer
) next
;
37 typedef struct MirrorBlockJob
{
41 BlockDriverState
*mirror_top_bs
;
42 BlockDriverState
*source
;
43 BlockDriverState
*base
;
45 /* The name of the graph node to replace */
47 /* The BDS to replace */
48 BlockDriverState
*to_replace
;
49 /* Used to block operations on the drive-mirror-replace target */
50 Error
*replace_blocker
;
52 BlockMirrorBackingMode backing_mode
;
53 BlockdevOnError on_source_error
, on_target_error
;
59 unsigned long *cow_bitmap
;
60 BdrvDirtyBitmap
*dirty_bitmap
;
61 BdrvDirtyBitmapIter
*dbi
;
63 QSIMPLEQ_HEAD(, MirrorBuffer
) buf_free
;
66 uint64_t last_pause_ns
;
67 unsigned long *in_flight_bitmap
;
69 int64_t bytes_in_flight
;
73 int target_cluster_size
;
75 bool initial_zeroing_ongoing
;
78 typedef struct MirrorOp
{
85 static BlockErrorAction
mirror_error_action(MirrorBlockJob
*s
, bool read
,
90 return block_job_error_action(&s
->common
, s
->on_source_error
,
93 return block_job_error_action(&s
->common
, s
->on_target_error
,
98 static void mirror_iteration_done(MirrorOp
*op
, int ret
)
100 MirrorBlockJob
*s
= op
->s
;
105 trace_mirror_iteration_done(s
, op
->offset
, op
->bytes
, ret
);
108 s
->bytes_in_flight
-= op
->bytes
;
110 for (i
= 0; i
< op
->qiov
.niov
; i
++) {
111 MirrorBuffer
*buf
= (MirrorBuffer
*) iov
[i
].iov_base
;
112 QSIMPLEQ_INSERT_TAIL(&s
->buf_free
, buf
, next
);
116 chunk_num
= op
->offset
/ s
->granularity
;
117 nb_chunks
= DIV_ROUND_UP(op
->bytes
, s
->granularity
);
118 bitmap_clear(s
->in_flight_bitmap
, chunk_num
, nb_chunks
);
121 bitmap_set(s
->cow_bitmap
, chunk_num
, nb_chunks
);
123 if (!s
->initial_zeroing_ongoing
) {
124 s
->common
.offset
+= op
->bytes
;
127 qemu_iovec_destroy(&op
->qiov
);
130 if (s
->waiting_for_io
) {
131 qemu_coroutine_enter(s
->common
.co
);
135 static void mirror_write_complete(void *opaque
, int ret
)
137 MirrorOp
*op
= opaque
;
138 MirrorBlockJob
*s
= op
->s
;
140 aio_context_acquire(blk_get_aio_context(s
->common
.blk
));
142 BlockErrorAction action
;
144 bdrv_set_dirty_bitmap(s
->dirty_bitmap
, op
->offset
, op
->bytes
);
145 action
= mirror_error_action(s
, false, -ret
);
146 if (action
== BLOCK_ERROR_ACTION_REPORT
&& s
->ret
>= 0) {
150 mirror_iteration_done(op
, ret
);
151 aio_context_release(blk_get_aio_context(s
->common
.blk
));
154 static void mirror_read_complete(void *opaque
, int ret
)
156 MirrorOp
*op
= opaque
;
157 MirrorBlockJob
*s
= op
->s
;
159 aio_context_acquire(blk_get_aio_context(s
->common
.blk
));
161 BlockErrorAction action
;
163 bdrv_set_dirty_bitmap(s
->dirty_bitmap
, op
->offset
, op
->bytes
);
164 action
= mirror_error_action(s
, true, -ret
);
165 if (action
== BLOCK_ERROR_ACTION_REPORT
&& s
->ret
>= 0) {
169 mirror_iteration_done(op
, ret
);
171 blk_aio_pwritev(s
->target
, op
->offset
, &op
->qiov
,
172 0, mirror_write_complete
, op
);
174 aio_context_release(blk_get_aio_context(s
->common
.blk
));
177 /* Clip bytes relative to offset to not exceed end-of-file */
178 static inline int64_t mirror_clip_bytes(MirrorBlockJob
*s
,
182 return MIN(bytes
, s
->bdev_length
- offset
);
185 /* Round offset and/or bytes to target cluster if COW is needed, and
186 * return the offset of the adjusted tail against original. */
187 static int mirror_cow_align(MirrorBlockJob
*s
, int64_t *offset
,
192 int64_t align_offset
= *offset
;
193 int64_t align_bytes
= *bytes
;
194 int max_bytes
= s
->granularity
* s
->max_iov
;
196 need_cow
= !test_bit(*offset
/ s
->granularity
, s
->cow_bitmap
);
197 need_cow
|= !test_bit((*offset
+ *bytes
- 1) / s
->granularity
,
200 bdrv_round_to_clusters(blk_bs(s
->target
), *offset
, *bytes
,
201 &align_offset
, &align_bytes
);
204 if (align_bytes
> max_bytes
) {
205 align_bytes
= max_bytes
;
207 align_bytes
= QEMU_ALIGN_DOWN(align_bytes
, s
->target_cluster_size
);
210 /* Clipping may result in align_bytes unaligned to chunk boundary, but
211 * that doesn't matter because it's already the end of source image. */
212 align_bytes
= mirror_clip_bytes(s
, align_offset
, align_bytes
);
214 ret
= align_offset
+ align_bytes
- (*offset
+ *bytes
);
215 *offset
= align_offset
;
216 *bytes
= align_bytes
;
221 static inline void mirror_wait_for_io(MirrorBlockJob
*s
)
223 assert(!s
->waiting_for_io
);
224 s
->waiting_for_io
= true;
225 qemu_coroutine_yield();
226 s
->waiting_for_io
= false;
229 /* Submit async read while handling COW.
230 * Returns: The number of bytes copied after and including offset,
231 * excluding any bytes copied prior to offset due to alignment.
232 * This will be @bytes if no alignment is necessary, or
233 * (new_end - offset) if tail is rounded up or down due to
234 * alignment or buffer limit.
236 static uint64_t mirror_do_read(MirrorBlockJob
*s
, int64_t offset
,
239 BlockBackend
*source
= s
->common
.blk
;
245 max_bytes
= s
->granularity
* s
->max_iov
;
247 /* We can only handle as much as buf_size at a time. */
248 bytes
= MIN(s
->buf_size
, MIN(max_bytes
, bytes
));
250 assert(bytes
< BDRV_REQUEST_MAX_BYTES
);
254 ret
+= mirror_cow_align(s
, &offset
, &bytes
);
256 assert(bytes
<= s
->buf_size
);
257 /* The offset is granularity-aligned because:
258 * 1) Caller passes in aligned values;
259 * 2) mirror_cow_align is used only when target cluster is larger. */
260 assert(QEMU_IS_ALIGNED(offset
, s
->granularity
));
261 /* The range is sector-aligned, since bdrv_getlength() rounds up. */
262 assert(QEMU_IS_ALIGNED(bytes
, BDRV_SECTOR_SIZE
));
263 nb_chunks
= DIV_ROUND_UP(bytes
, s
->granularity
);
265 while (s
->buf_free_count
< nb_chunks
) {
266 trace_mirror_yield_in_flight(s
, offset
, s
->in_flight
);
267 mirror_wait_for_io(s
);
270 /* Allocate a MirrorOp that is used as an AIO callback. */
271 op
= g_new(MirrorOp
, 1);
276 /* Now make a QEMUIOVector taking enough granularity-sized chunks
279 qemu_iovec_init(&op
->qiov
, nb_chunks
);
280 while (nb_chunks
-- > 0) {
281 MirrorBuffer
*buf
= QSIMPLEQ_FIRST(&s
->buf_free
);
282 size_t remaining
= bytes
- op
->qiov
.size
;
284 QSIMPLEQ_REMOVE_HEAD(&s
->buf_free
, next
);
286 qemu_iovec_add(&op
->qiov
, buf
, MIN(s
->granularity
, remaining
));
289 /* Copy the dirty cluster. */
291 s
->bytes_in_flight
+= bytes
;
292 trace_mirror_one_iteration(s
, offset
, bytes
);
294 blk_aio_preadv(source
, offset
, &op
->qiov
, 0, mirror_read_complete
, op
);
298 static void mirror_do_zero_or_discard(MirrorBlockJob
*s
,
305 /* Allocate a MirrorOp that is used as an AIO callback. The qiov is zeroed
306 * so the freeing in mirror_iteration_done is nop. */
307 op
= g_new0(MirrorOp
, 1);
313 s
->bytes_in_flight
+= bytes
;
315 blk_aio_pdiscard(s
->target
, offset
,
316 op
->bytes
, mirror_write_complete
, op
);
318 blk_aio_pwrite_zeroes(s
->target
, offset
,
319 op
->bytes
, s
->unmap
? BDRV_REQ_MAY_UNMAP
: 0,
320 mirror_write_complete
, op
);
324 static uint64_t coroutine_fn
mirror_iteration(MirrorBlockJob
*s
)
326 BlockDriverState
*source
= s
->source
;
327 int64_t offset
, first_chunk
;
328 uint64_t delay_ns
= 0;
329 /* At least the first dirty chunk is mirrored in one iteration. */
331 int sectors_per_chunk
= s
->granularity
>> BDRV_SECTOR_BITS
;
332 bool write_zeroes_ok
= bdrv_can_write_zeroes_with_unmap(blk_bs(s
->target
));
333 int max_io_bytes
= MAX(s
->buf_size
/ MAX_IN_FLIGHT
, MAX_IO_BYTES
);
335 bdrv_dirty_bitmap_lock(s
->dirty_bitmap
);
336 offset
= bdrv_dirty_iter_next(s
->dbi
);
338 bdrv_set_dirty_iter(s
->dbi
, 0);
339 offset
= bdrv_dirty_iter_next(s
->dbi
);
340 trace_mirror_restart_iter(s
, bdrv_get_dirty_count(s
->dirty_bitmap
));
343 bdrv_dirty_bitmap_unlock(s
->dirty_bitmap
);
345 first_chunk
= offset
/ s
->granularity
;
346 while (test_bit(first_chunk
, s
->in_flight_bitmap
)) {
347 trace_mirror_yield_in_flight(s
, offset
, s
->in_flight
);
348 mirror_wait_for_io(s
);
351 block_job_pause_point(&s
->common
);
353 /* Find the number of consective dirty chunks following the first dirty
354 * one, and wait for in flight requests in them. */
355 bdrv_dirty_bitmap_lock(s
->dirty_bitmap
);
356 while (nb_chunks
* s
->granularity
< s
->buf_size
) {
358 int64_t next_offset
= offset
+ nb_chunks
* s
->granularity
;
359 int64_t next_chunk
= next_offset
/ s
->granularity
;
360 if (next_offset
>= s
->bdev_length
||
361 !bdrv_get_dirty_locked(source
, s
->dirty_bitmap
, next_offset
)) {
364 if (test_bit(next_chunk
, s
->in_flight_bitmap
)) {
368 next_dirty
= bdrv_dirty_iter_next(s
->dbi
);
369 if (next_dirty
> next_offset
|| next_dirty
< 0) {
370 /* The bitmap iterator's cache is stale, refresh it */
371 bdrv_set_dirty_iter(s
->dbi
, next_offset
);
372 next_dirty
= bdrv_dirty_iter_next(s
->dbi
);
374 assert(next_dirty
== next_offset
);
378 /* Clear dirty bits before querying the block status, because
379 * calling bdrv_get_block_status_above could yield - if some blocks are
380 * marked dirty in this window, we need to know.
382 bdrv_reset_dirty_bitmap_locked(s
->dirty_bitmap
, offset
,
383 nb_chunks
* s
->granularity
);
384 bdrv_dirty_bitmap_unlock(s
->dirty_bitmap
);
386 bitmap_set(s
->in_flight_bitmap
, offset
/ s
->granularity
, nb_chunks
);
387 while (nb_chunks
> 0 && offset
< s
->bdev_length
) {
391 int64_t io_bytes_acct
;
395 MIRROR_METHOD_DISCARD
396 } mirror_method
= MIRROR_METHOD_COPY
;
398 assert(!(offset
% s
->granularity
));
399 ret
= bdrv_get_block_status_above(source
, NULL
,
400 offset
>> BDRV_SECTOR_BITS
,
401 nb_chunks
* sectors_per_chunk
,
403 io_bytes
= io_sectors
* BDRV_SECTOR_SIZE
;
405 io_bytes
= MIN(nb_chunks
* s
->granularity
, max_io_bytes
);
406 } else if (ret
& BDRV_BLOCK_DATA
) {
407 io_bytes
= MIN(io_bytes
, max_io_bytes
);
410 io_bytes
-= io_bytes
% s
->granularity
;
411 if (io_bytes
< s
->granularity
) {
412 io_bytes
= s
->granularity
;
413 } else if (ret
>= 0 && !(ret
& BDRV_BLOCK_DATA
)) {
414 int64_t target_offset
;
415 int64_t target_bytes
;
416 bdrv_round_to_clusters(blk_bs(s
->target
), offset
, io_bytes
,
417 &target_offset
, &target_bytes
);
418 if (target_offset
== offset
&&
419 target_bytes
== io_bytes
) {
420 mirror_method
= ret
& BDRV_BLOCK_ZERO
?
422 MIRROR_METHOD_DISCARD
;
426 while (s
->in_flight
>= MAX_IN_FLIGHT
) {
427 trace_mirror_yield_in_flight(s
, offset
, s
->in_flight
);
428 mirror_wait_for_io(s
);
435 io_bytes
= mirror_clip_bytes(s
, offset
, io_bytes
);
436 switch (mirror_method
) {
437 case MIRROR_METHOD_COPY
:
438 io_bytes
= io_bytes_acct
= mirror_do_read(s
, offset
, io_bytes
);
440 case MIRROR_METHOD_ZERO
:
441 case MIRROR_METHOD_DISCARD
:
442 mirror_do_zero_or_discard(s
, offset
, io_bytes
,
443 mirror_method
== MIRROR_METHOD_DISCARD
);
444 if (write_zeroes_ok
) {
447 io_bytes_acct
= io_bytes
;
455 nb_chunks
-= DIV_ROUND_UP(io_bytes
, s
->granularity
);
456 if (s
->common
.speed
) {
457 delay_ns
= ratelimit_calculate_delay(&s
->limit
, io_bytes_acct
);
463 static void mirror_free_init(MirrorBlockJob
*s
)
465 int granularity
= s
->granularity
;
466 size_t buf_size
= s
->buf_size
;
467 uint8_t *buf
= s
->buf
;
469 assert(s
->buf_free_count
== 0);
470 QSIMPLEQ_INIT(&s
->buf_free
);
471 while (buf_size
!= 0) {
472 MirrorBuffer
*cur
= (MirrorBuffer
*)buf
;
473 QSIMPLEQ_INSERT_TAIL(&s
->buf_free
, cur
, next
);
475 buf_size
-= granularity
;
480 /* This is also used for the .pause callback. There is no matching
481 * mirror_resume() because mirror_run() will begin iterating again
482 * when the job is resumed.
484 static void mirror_wait_for_all_io(MirrorBlockJob
*s
)
486 while (s
->in_flight
> 0) {
487 mirror_wait_for_io(s
);
495 static void mirror_exit(BlockJob
*job
, void *opaque
)
497 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
498 MirrorExitData
*data
= opaque
;
499 AioContext
*replace_aio_context
= NULL
;
500 BlockDriverState
*src
= s
->source
;
501 BlockDriverState
*target_bs
= blk_bs(s
->target
);
502 BlockDriverState
*mirror_top_bs
= s
->mirror_top_bs
;
503 Error
*local_err
= NULL
;
505 bdrv_release_dirty_bitmap(src
, s
->dirty_bitmap
);
507 /* Make sure that the source BDS doesn't go away before we called
508 * block_job_completed(). */
510 bdrv_ref(mirror_top_bs
);
513 /* Remove target parent that still uses BLK_PERM_WRITE/RESIZE before
514 * inserting target_bs at s->to_replace, where we might not be able to get
517 * Note that blk_unref() alone doesn't necessarily drop permissions because
518 * we might be running nested inside mirror_drain(), which takes an extra
519 * reference, so use an explicit blk_set_perm() first. */
520 blk_set_perm(s
->target
, 0, BLK_PERM_ALL
, &error_abort
);
521 blk_unref(s
->target
);
524 /* We don't access the source any more. Dropping any WRITE/RESIZE is
525 * required before it could become a backing file of target_bs. */
526 bdrv_child_try_set_perm(mirror_top_bs
->backing
, 0, BLK_PERM_ALL
,
528 if (s
->backing_mode
== MIRROR_SOURCE_BACKING_CHAIN
) {
529 BlockDriverState
*backing
= s
->is_none_mode
? src
: s
->base
;
530 if (backing_bs(target_bs
) != backing
) {
531 bdrv_set_backing_hd(target_bs
, backing
, &local_err
);
533 error_report_err(local_err
);
540 replace_aio_context
= bdrv_get_aio_context(s
->to_replace
);
541 aio_context_acquire(replace_aio_context
);
544 if (s
->should_complete
&& data
->ret
== 0) {
545 BlockDriverState
*to_replace
= src
;
547 to_replace
= s
->to_replace
;
550 if (bdrv_get_flags(target_bs
) != bdrv_get_flags(to_replace
)) {
551 bdrv_reopen(target_bs
, bdrv_get_flags(to_replace
), NULL
);
554 /* The mirror job has no requests in flight any more, but we need to
555 * drain potential other users of the BDS before changing the graph. */
556 bdrv_drained_begin(target_bs
);
557 bdrv_replace_node(to_replace
, target_bs
, &local_err
);
558 bdrv_drained_end(target_bs
);
560 error_report_err(local_err
);
565 bdrv_op_unblock_all(s
->to_replace
, s
->replace_blocker
);
566 error_free(s
->replace_blocker
);
567 bdrv_unref(s
->to_replace
);
569 if (replace_aio_context
) {
570 aio_context_release(replace_aio_context
);
573 bdrv_unref(target_bs
);
575 /* Remove the mirror filter driver from the graph. Before this, get rid of
576 * the blockers on the intermediate nodes so that the resulting state is
577 * valid. Also give up permissions on mirror_top_bs->backing, which might
578 * block the removal. */
579 block_job_remove_all_bdrv(job
);
580 bdrv_child_try_set_perm(mirror_top_bs
->backing
, 0, BLK_PERM_ALL
,
582 bdrv_replace_node(mirror_top_bs
, backing_bs(mirror_top_bs
), &error_abort
);
584 /* We just changed the BDS the job BB refers to (with either or both of the
585 * bdrv_replace_node() calls), so switch the BB back so the cleanup does
586 * the right thing. We don't need any permissions any more now. */
587 blk_remove_bs(job
->blk
);
588 blk_set_perm(job
->blk
, 0, BLK_PERM_ALL
, &error_abort
);
589 blk_insert_bs(job
->blk
, mirror_top_bs
, &error_abort
);
591 block_job_completed(&s
->common
, data
->ret
);
594 bdrv_drained_end(src
);
595 bdrv_unref(mirror_top_bs
);
599 static void mirror_throttle(MirrorBlockJob
*s
)
601 int64_t now
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
);
603 if (now
- s
->last_pause_ns
> SLICE_TIME
) {
604 s
->last_pause_ns
= now
;
605 block_job_sleep_ns(&s
->common
, QEMU_CLOCK_REALTIME
, 0);
607 block_job_pause_point(&s
->common
);
611 static int coroutine_fn
mirror_dirty_init(MirrorBlockJob
*s
)
614 BlockDriverState
*base
= s
->base
;
615 BlockDriverState
*bs
= s
->source
;
616 BlockDriverState
*target_bs
= blk_bs(s
->target
);
620 if (base
== NULL
&& !bdrv_has_zero_init(target_bs
)) {
621 if (!bdrv_can_write_zeroes_with_unmap(target_bs
)) {
622 bdrv_set_dirty_bitmap(s
->dirty_bitmap
, 0, s
->bdev_length
);
626 s
->initial_zeroing_ongoing
= true;
627 for (offset
= 0; offset
< s
->bdev_length
; ) {
628 int bytes
= MIN(s
->bdev_length
- offset
,
629 QEMU_ALIGN_DOWN(INT_MAX
, s
->granularity
));
633 if (block_job_is_cancelled(&s
->common
)) {
634 s
->initial_zeroing_ongoing
= false;
638 if (s
->in_flight
>= MAX_IN_FLIGHT
) {
639 trace_mirror_yield(s
, UINT64_MAX
, s
->buf_free_count
,
641 mirror_wait_for_io(s
);
645 mirror_do_zero_or_discard(s
, offset
, bytes
, false);
649 mirror_wait_for_all_io(s
);
650 s
->initial_zeroing_ongoing
= false;
653 /* First part, loop on the sectors and initialize the dirty bitmap. */
654 for (offset
= 0; offset
< s
->bdev_length
; ) {
655 /* Just to make sure we are not exceeding int limit. */
656 int bytes
= MIN(s
->bdev_length
- offset
,
657 QEMU_ALIGN_DOWN(INT_MAX
, s
->granularity
));
661 if (block_job_is_cancelled(&s
->common
)) {
665 ret
= bdrv_is_allocated_above(bs
, base
, offset
, bytes
, &count
);
672 bdrv_set_dirty_bitmap(s
->dirty_bitmap
, offset
, count
);
679 /* Called when going out of the streaming phase to flush the bulk of the
680 * data to the medium, or just before completing.
682 static int mirror_flush(MirrorBlockJob
*s
)
684 int ret
= blk_flush(s
->target
);
686 if (mirror_error_action(s
, false, -ret
) == BLOCK_ERROR_ACTION_REPORT
) {
693 static void coroutine_fn
mirror_run(void *opaque
)
695 MirrorBlockJob
*s
= opaque
;
696 MirrorExitData
*data
;
697 BlockDriverState
*bs
= s
->source
;
698 BlockDriverState
*target_bs
= blk_bs(s
->target
);
699 bool need_drain
= true;
702 char backing_filename
[2]; /* we only need 2 characters because we are only
703 checking for a NULL string */
706 if (block_job_is_cancelled(&s
->common
)) {
710 s
->bdev_length
= bdrv_getlength(bs
);
711 if (s
->bdev_length
< 0) {
712 ret
= s
->bdev_length
;
716 /* Active commit must resize the base image if its size differs from the
718 if (s
->base
== blk_bs(s
->target
)) {
721 base_length
= blk_getlength(s
->target
);
722 if (base_length
< 0) {
727 if (s
->bdev_length
> base_length
) {
728 ret
= blk_truncate(s
->target
, s
->bdev_length
, PREALLOC_MODE_OFF
,
736 if (s
->bdev_length
== 0) {
737 /* Report BLOCK_JOB_READY and wait for complete. */
738 block_job_event_ready(&s
->common
);
740 while (!block_job_is_cancelled(&s
->common
) && !s
->should_complete
) {
741 block_job_yield(&s
->common
);
743 s
->common
.cancelled
= false;
747 length
= DIV_ROUND_UP(s
->bdev_length
, s
->granularity
);
748 s
->in_flight_bitmap
= bitmap_new(length
);
750 /* If we have no backing file yet in the destination, we cannot let
751 * the destination do COW. Instead, we copy sectors around the
752 * dirty data if needed. We need a bitmap to do that.
754 bdrv_get_backing_filename(target_bs
, backing_filename
,
755 sizeof(backing_filename
));
756 if (!bdrv_get_info(target_bs
, &bdi
) && bdi
.cluster_size
) {
757 s
->target_cluster_size
= bdi
.cluster_size
;
759 s
->target_cluster_size
= BDRV_SECTOR_SIZE
;
761 if (backing_filename
[0] && !target_bs
->backing
&&
762 s
->granularity
< s
->target_cluster_size
) {
763 s
->buf_size
= MAX(s
->buf_size
, s
->target_cluster_size
);
764 s
->cow_bitmap
= bitmap_new(length
);
766 s
->max_iov
= MIN(bs
->bl
.max_iov
, target_bs
->bl
.max_iov
);
768 s
->buf
= qemu_try_blockalign(bs
, s
->buf_size
);
769 if (s
->buf
== NULL
) {
776 s
->last_pause_ns
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
);
777 if (!s
->is_none_mode
) {
778 ret
= mirror_dirty_init(s
);
779 if (ret
< 0 || block_job_is_cancelled(&s
->common
)) {
785 s
->dbi
= bdrv_dirty_iter_new(s
->dirty_bitmap
);
787 uint64_t delay_ns
= 0;
789 bool should_complete
;
796 block_job_pause_point(&s
->common
);
798 cnt
= bdrv_get_dirty_count(s
->dirty_bitmap
);
799 /* s->common.offset contains the number of bytes already processed so
800 * far, cnt is the number of dirty bytes remaining and
801 * s->bytes_in_flight is the number of bytes currently being
802 * processed; together those are the current total operation length */
803 s
->common
.len
= s
->common
.offset
+ s
->bytes_in_flight
+ cnt
;
805 /* Note that even when no rate limit is applied we need to yield
806 * periodically with no pending I/O so that bdrv_drain_all() returns.
807 * We do so every SLICE_TIME nanoseconds, or when there is an error,
808 * or when the source is clean, whichever comes first.
810 delta
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
) - s
->last_pause_ns
;
811 if (delta
< SLICE_TIME
&&
812 s
->common
.iostatus
== BLOCK_DEVICE_IO_STATUS_OK
) {
813 if (s
->in_flight
>= MAX_IN_FLIGHT
|| s
->buf_free_count
== 0 ||
814 (cnt
== 0 && s
->in_flight
> 0)) {
815 trace_mirror_yield(s
, cnt
, s
->buf_free_count
, s
->in_flight
);
816 mirror_wait_for_io(s
);
818 } else if (cnt
!= 0) {
819 delay_ns
= mirror_iteration(s
);
823 should_complete
= false;
824 if (s
->in_flight
== 0 && cnt
== 0) {
825 trace_mirror_before_flush(s
);
827 if (mirror_flush(s
) < 0) {
828 /* Go check s->ret. */
831 /* We're out of the streaming phase. From now on, if the job
832 * is cancelled we will actually complete all pending I/O and
833 * report completion. This way, block-job-cancel will leave
834 * the target in a consistent state.
836 block_job_event_ready(&s
->common
);
840 should_complete
= s
->should_complete
||
841 block_job_is_cancelled(&s
->common
);
842 cnt
= bdrv_get_dirty_count(s
->dirty_bitmap
);
845 if (cnt
== 0 && should_complete
) {
846 /* The dirty bitmap is not updated while operations are pending.
847 * If we're about to exit, wait for pending operations before
848 * calling bdrv_get_dirty_count(bs), or we may exit while the
849 * source has dirty data to copy!
851 * Note that I/O can be submitted by the guest while
852 * mirror_populate runs, so pause it now. Before deciding
853 * whether to switch to target check one last time if I/O has
854 * come in the meanwhile, and if not flush the data to disk.
856 trace_mirror_before_drain(s
, cnt
);
858 bdrv_drained_begin(bs
);
859 cnt
= bdrv_get_dirty_count(s
->dirty_bitmap
);
860 if (cnt
> 0 || mirror_flush(s
) < 0) {
861 bdrv_drained_end(bs
);
865 /* The two disks are in sync. Exit and report successful
868 assert(QLIST_EMPTY(&bs
->tracked_requests
));
869 s
->common
.cancelled
= false;
875 trace_mirror_before_sleep(s
, cnt
, s
->synced
, delay_ns
);
877 block_job_sleep_ns(&s
->common
, QEMU_CLOCK_REALTIME
, delay_ns
);
878 if (block_job_is_cancelled(&s
->common
)) {
881 } else if (!should_complete
) {
882 delay_ns
= (s
->in_flight
== 0 && cnt
== 0 ? SLICE_TIME
: 0);
883 block_job_sleep_ns(&s
->common
, QEMU_CLOCK_REALTIME
, delay_ns
);
885 s
->last_pause_ns
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
);
889 if (s
->in_flight
> 0) {
890 /* We get here only if something went wrong. Either the job failed,
891 * or it was cancelled prematurely so that we do not guarantee that
892 * the target is a copy of the source.
894 assert(ret
< 0 || (!s
->synced
&& block_job_is_cancelled(&s
->common
)));
896 mirror_wait_for_all_io(s
);
899 assert(s
->in_flight
== 0);
901 g_free(s
->cow_bitmap
);
902 g_free(s
->in_flight_bitmap
);
903 bdrv_dirty_iter_free(s
->dbi
);
905 data
= g_malloc(sizeof(*data
));
909 bdrv_drained_begin(bs
);
911 block_job_defer_to_main_loop(&s
->common
, mirror_exit
, data
);
914 static void mirror_set_speed(BlockJob
*job
, int64_t speed
, Error
**errp
)
916 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
919 error_setg(errp
, QERR_INVALID_PARAMETER
, "speed");
922 ratelimit_set_speed(&s
->limit
, speed
, SLICE_TIME
);
925 static void mirror_complete(BlockJob
*job
, Error
**errp
)
927 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
928 BlockDriverState
*target
;
930 target
= blk_bs(s
->target
);
933 error_setg(errp
, "The active block job '%s' cannot be completed",
938 if (s
->backing_mode
== MIRROR_OPEN_BACKING_CHAIN
) {
941 assert(!target
->backing
);
942 ret
= bdrv_open_backing_file(target
, NULL
, "backing", errp
);
948 /* block all operations on to_replace bs */
950 AioContext
*replace_aio_context
;
952 s
->to_replace
= bdrv_find_node(s
->replaces
);
953 if (!s
->to_replace
) {
954 error_setg(errp
, "Node name '%s' not found", s
->replaces
);
958 replace_aio_context
= bdrv_get_aio_context(s
->to_replace
);
959 aio_context_acquire(replace_aio_context
);
961 /* TODO Translate this into permission system. Current definition of
962 * GRAPH_MOD would require to request it for the parents; they might
963 * not even be BlockDriverStates, however, so a BdrvChild can't address
964 * them. May need redefinition of GRAPH_MOD. */
965 error_setg(&s
->replace_blocker
,
966 "block device is in use by block-job-complete");
967 bdrv_op_block_all(s
->to_replace
, s
->replace_blocker
);
968 bdrv_ref(s
->to_replace
);
970 aio_context_release(replace_aio_context
);
973 s
->should_complete
= true;
974 block_job_enter(&s
->common
);
977 static void mirror_pause(BlockJob
*job
)
979 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
981 mirror_wait_for_all_io(s
);
984 static void mirror_attached_aio_context(BlockJob
*job
, AioContext
*new_context
)
986 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
988 blk_set_aio_context(s
->target
, new_context
);
991 static void mirror_drain(BlockJob
*job
)
993 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
995 /* Need to keep a reference in case blk_drain triggers execution
996 * of mirror_complete...
999 BlockBackend
*target
= s
->target
;
1006 static const BlockJobDriver mirror_job_driver
= {
1007 .instance_size
= sizeof(MirrorBlockJob
),
1008 .job_type
= BLOCK_JOB_TYPE_MIRROR
,
1009 .set_speed
= mirror_set_speed
,
1010 .start
= mirror_run
,
1011 .complete
= mirror_complete
,
1012 .pause
= mirror_pause
,
1013 .attached_aio_context
= mirror_attached_aio_context
,
1014 .drain
= mirror_drain
,
1017 static const BlockJobDriver commit_active_job_driver
= {
1018 .instance_size
= sizeof(MirrorBlockJob
),
1019 .job_type
= BLOCK_JOB_TYPE_COMMIT
,
1020 .set_speed
= mirror_set_speed
,
1021 .start
= mirror_run
,
1022 .complete
= mirror_complete
,
1023 .pause
= mirror_pause
,
1024 .attached_aio_context
= mirror_attached_aio_context
,
1025 .drain
= mirror_drain
,
1028 static int coroutine_fn
bdrv_mirror_top_preadv(BlockDriverState
*bs
,
1029 uint64_t offset
, uint64_t bytes
, QEMUIOVector
*qiov
, int flags
)
1031 return bdrv_co_preadv(bs
->backing
, offset
, bytes
, qiov
, flags
);
1034 static int coroutine_fn
bdrv_mirror_top_pwritev(BlockDriverState
*bs
,
1035 uint64_t offset
, uint64_t bytes
, QEMUIOVector
*qiov
, int flags
)
1037 return bdrv_co_pwritev(bs
->backing
, offset
, bytes
, qiov
, flags
);
1040 static int coroutine_fn
bdrv_mirror_top_flush(BlockDriverState
*bs
)
1042 if (bs
->backing
== NULL
) {
1043 /* we can be here after failed bdrv_append in mirror_start_job */
1046 return bdrv_co_flush(bs
->backing
->bs
);
1049 static int coroutine_fn
bdrv_mirror_top_pwrite_zeroes(BlockDriverState
*bs
,
1050 int64_t offset
, int bytes
, BdrvRequestFlags flags
)
1052 return bdrv_co_pwrite_zeroes(bs
->backing
, offset
, bytes
, flags
);
1055 static int coroutine_fn
bdrv_mirror_top_pdiscard(BlockDriverState
*bs
,
1056 int64_t offset
, int bytes
)
1058 return bdrv_co_pdiscard(bs
->backing
->bs
, offset
, bytes
);
1061 static void bdrv_mirror_top_refresh_filename(BlockDriverState
*bs
, QDict
*opts
)
1063 if (bs
->backing
== NULL
) {
1064 /* we can be here after failed bdrv_attach_child in
1065 * bdrv_set_backing_hd */
1068 bdrv_refresh_filename(bs
->backing
->bs
);
1069 pstrcpy(bs
->exact_filename
, sizeof(bs
->exact_filename
),
1070 bs
->backing
->bs
->filename
);
1073 static void bdrv_mirror_top_close(BlockDriverState
*bs
)
1077 static void bdrv_mirror_top_child_perm(BlockDriverState
*bs
, BdrvChild
*c
,
1078 const BdrvChildRole
*role
,
1079 BlockReopenQueue
*reopen_queue
,
1080 uint64_t perm
, uint64_t shared
,
1081 uint64_t *nperm
, uint64_t *nshared
)
1083 /* Must be able to forward guest writes to the real image */
1085 if (perm
& BLK_PERM_WRITE
) {
1086 *nperm
|= BLK_PERM_WRITE
;
1089 *nshared
= BLK_PERM_ALL
;
1092 /* Dummy node that provides consistent read to its users without requiring it
1093 * from its backing file and that allows writes on the backing file chain. */
1094 static BlockDriver bdrv_mirror_top
= {
1095 .format_name
= "mirror_top",
1096 .bdrv_co_preadv
= bdrv_mirror_top_preadv
,
1097 .bdrv_co_pwritev
= bdrv_mirror_top_pwritev
,
1098 .bdrv_co_pwrite_zeroes
= bdrv_mirror_top_pwrite_zeroes
,
1099 .bdrv_co_pdiscard
= bdrv_mirror_top_pdiscard
,
1100 .bdrv_co_flush
= bdrv_mirror_top_flush
,
1101 .bdrv_co_get_block_status
= bdrv_co_get_block_status_from_backing
,
1102 .bdrv_refresh_filename
= bdrv_mirror_top_refresh_filename
,
1103 .bdrv_close
= bdrv_mirror_top_close
,
1104 .bdrv_child_perm
= bdrv_mirror_top_child_perm
,
1107 static void mirror_start_job(const char *job_id
, BlockDriverState
*bs
,
1108 int creation_flags
, BlockDriverState
*target
,
1109 const char *replaces
, int64_t speed
,
1110 uint32_t granularity
, int64_t buf_size
,
1111 BlockMirrorBackingMode backing_mode
,
1112 BlockdevOnError on_source_error
,
1113 BlockdevOnError on_target_error
,
1115 BlockCompletionFunc
*cb
,
1117 const BlockJobDriver
*driver
,
1118 bool is_none_mode
, BlockDriverState
*base
,
1119 bool auto_complete
, const char *filter_node_name
,
1124 BlockDriverState
*mirror_top_bs
;
1125 bool target_graph_mod
;
1126 bool target_is_backing
;
1127 Error
*local_err
= NULL
;
1130 if (granularity
== 0) {
1131 granularity
= bdrv_get_default_bitmap_granularity(target
);
1134 assert ((granularity
& (granularity
- 1)) == 0);
1135 /* Granularity must be large enough for sector-based dirty bitmap */
1136 assert(granularity
>= BDRV_SECTOR_SIZE
);
1139 error_setg(errp
, "Invalid parameter 'buf-size'");
1143 if (buf_size
== 0) {
1144 buf_size
= DEFAULT_MIRROR_BUF_SIZE
;
1147 /* In the case of active commit, add dummy driver to provide consistent
1148 * reads on the top, while disabling it in the intermediate nodes, and make
1149 * the backing chain writable. */
1150 mirror_top_bs
= bdrv_new_open_driver(&bdrv_mirror_top
, filter_node_name
,
1152 if (mirror_top_bs
== NULL
) {
1155 if (!filter_node_name
) {
1156 mirror_top_bs
->implicit
= true;
1158 mirror_top_bs
->total_sectors
= bs
->total_sectors
;
1159 bdrv_set_aio_context(mirror_top_bs
, bdrv_get_aio_context(bs
));
1161 /* bdrv_append takes ownership of the mirror_top_bs reference, need to keep
1162 * it alive until block_job_create() succeeds even if bs has no parent. */
1163 bdrv_ref(mirror_top_bs
);
1164 bdrv_drained_begin(bs
);
1165 bdrv_append(mirror_top_bs
, bs
, &local_err
);
1166 bdrv_drained_end(bs
);
1169 bdrv_unref(mirror_top_bs
);
1170 error_propagate(errp
, local_err
);
1174 /* Make sure that the source is not resized while the job is running */
1175 s
= block_job_create(job_id
, driver
, mirror_top_bs
,
1176 BLK_PERM_CONSISTENT_READ
,
1177 BLK_PERM_CONSISTENT_READ
| BLK_PERM_WRITE_UNCHANGED
|
1178 BLK_PERM_WRITE
| BLK_PERM_GRAPH_MOD
, speed
,
1179 creation_flags
, cb
, opaque
, errp
);
1183 /* The block job now has a reference to this node */
1184 bdrv_unref(mirror_top_bs
);
1187 s
->mirror_top_bs
= mirror_top_bs
;
1189 /* No resize for the target either; while the mirror is still running, a
1190 * consistent read isn't necessarily possible. We could possibly allow
1191 * writes and graph modifications, though it would likely defeat the
1192 * purpose of a mirror, so leave them blocked for now.
1194 * In the case of active commit, things look a bit different, though,
1195 * because the target is an already populated backing file in active use.
1196 * We can allow anything except resize there.*/
1197 target_is_backing
= bdrv_chain_contains(bs
, target
);
1198 target_graph_mod
= (backing_mode
!= MIRROR_LEAVE_BACKING_CHAIN
);
1199 s
->target
= blk_new(BLK_PERM_WRITE
| BLK_PERM_RESIZE
|
1200 (target_graph_mod
? BLK_PERM_GRAPH_MOD
: 0),
1201 BLK_PERM_WRITE_UNCHANGED
|
1202 (target_is_backing
? BLK_PERM_CONSISTENT_READ
|
1204 BLK_PERM_GRAPH_MOD
: 0));
1205 ret
= blk_insert_bs(s
->target
, target
, errp
);
1210 /* XXX: Mirror target could be a NBD server of target QEMU in the case
1211 * of non-shared block migration. To allow migration completion, we
1212 * have to allow "inactivate" of the target BB. When that happens, we
1213 * know the job is drained, and the vcpus are stopped, so no write
1214 * operation will be performed. Block layer already has assertions to
1216 blk_set_force_allow_inactivate(s
->target
);
1219 s
->replaces
= g_strdup(replaces
);
1220 s
->on_source_error
= on_source_error
;
1221 s
->on_target_error
= on_target_error
;
1222 s
->is_none_mode
= is_none_mode
;
1223 s
->backing_mode
= backing_mode
;
1225 s
->granularity
= granularity
;
1226 s
->buf_size
= ROUND_UP(buf_size
, granularity
);
1228 if (auto_complete
) {
1229 s
->should_complete
= true;
1232 s
->dirty_bitmap
= bdrv_create_dirty_bitmap(bs
, granularity
, NULL
, errp
);
1233 if (!s
->dirty_bitmap
) {
1237 /* Required permissions are already taken with blk_new() */
1238 block_job_add_bdrv(&s
->common
, "target", target
, 0, BLK_PERM_ALL
,
1241 /* In commit_active_start() all intermediate nodes disappear, so
1242 * any jobs in them must be blocked */
1243 if (target_is_backing
) {
1244 BlockDriverState
*iter
;
1245 for (iter
= backing_bs(bs
); iter
!= target
; iter
= backing_bs(iter
)) {
1246 /* XXX BLK_PERM_WRITE needs to be allowed so we don't block
1247 * ourselves at s->base (if writes are blocked for a node, they are
1248 * also blocked for its backing file). The other options would be a
1249 * second filter driver above s->base (== target). */
1250 ret
= block_job_add_bdrv(&s
->common
, "intermediate node", iter
, 0,
1251 BLK_PERM_WRITE_UNCHANGED
| BLK_PERM_WRITE
,
1259 trace_mirror_start(bs
, s
, opaque
);
1260 block_job_start(&s
->common
);
1265 /* Make sure this BDS does not go away until we have completed the graph
1267 bdrv_ref(mirror_top_bs
);
1269 g_free(s
->replaces
);
1270 blk_unref(s
->target
);
1271 block_job_early_fail(&s
->common
);
1274 bdrv_child_try_set_perm(mirror_top_bs
->backing
, 0, BLK_PERM_ALL
,
1276 bdrv_replace_node(mirror_top_bs
, backing_bs(mirror_top_bs
), &error_abort
);
1278 bdrv_unref(mirror_top_bs
);
1281 void mirror_start(const char *job_id
, BlockDriverState
*bs
,
1282 BlockDriverState
*target
, const char *replaces
,
1283 int64_t speed
, uint32_t granularity
, int64_t buf_size
,
1284 MirrorSyncMode mode
, BlockMirrorBackingMode backing_mode
,
1285 BlockdevOnError on_source_error
,
1286 BlockdevOnError on_target_error
,
1287 bool unmap
, const char *filter_node_name
, Error
**errp
)
1290 BlockDriverState
*base
;
1292 if (mode
== MIRROR_SYNC_MODE_INCREMENTAL
) {
1293 error_setg(errp
, "Sync mode 'incremental' not supported");
1296 is_none_mode
= mode
== MIRROR_SYNC_MODE_NONE
;
1297 base
= mode
== MIRROR_SYNC_MODE_TOP
? backing_bs(bs
) : NULL
;
1298 mirror_start_job(job_id
, bs
, BLOCK_JOB_DEFAULT
, target
, replaces
,
1299 speed
, granularity
, buf_size
, backing_mode
,
1300 on_source_error
, on_target_error
, unmap
, NULL
, NULL
,
1301 &mirror_job_driver
, is_none_mode
, base
, false,
1302 filter_node_name
, true, errp
);
1305 void commit_active_start(const char *job_id
, BlockDriverState
*bs
,
1306 BlockDriverState
*base
, int creation_flags
,
1307 int64_t speed
, BlockdevOnError on_error
,
1308 const char *filter_node_name
,
1309 BlockCompletionFunc
*cb
, void *opaque
,
1310 bool auto_complete
, Error
**errp
)
1312 int orig_base_flags
;
1313 Error
*local_err
= NULL
;
1315 orig_base_flags
= bdrv_get_flags(base
);
1317 if (bdrv_reopen(base
, bs
->open_flags
, errp
)) {
1321 mirror_start_job(job_id
, bs
, creation_flags
, base
, NULL
, speed
, 0, 0,
1322 MIRROR_LEAVE_BACKING_CHAIN
,
1323 on_error
, on_error
, true, cb
, opaque
,
1324 &commit_active_job_driver
, false, base
, auto_complete
,
1325 filter_node_name
, false, &local_err
);
1327 error_propagate(errp
, local_err
);
1328 goto error_restore_flags
;
1333 error_restore_flags
:
1334 /* ignore error and errp for bdrv_reopen, because we want to propagate
1335 * the original error */
1336 bdrv_reopen(base
, orig_base_flags
, NULL
);