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"
16 #include "block/blockjob.h"
17 #include "block/block_int.h"
18 #include "sysemu/block-backend.h"
19 #include "qapi/error.h"
20 #include "qapi/qmp/qerror.h"
21 #include "qemu/ratelimit.h"
22 #include "qemu/bitmap.h"
23 #include "qemu/error-report.h"
25 #define SLICE_TIME 100000000ULL /* ns */
26 #define MAX_IN_FLIGHT 16
27 #define DEFAULT_MIRROR_BUF_SIZE (10 << 20)
29 /* The mirroring buffer is a list of granularity-sized chunks.
30 * Free chunks are organized in a list.
32 typedef struct MirrorBuffer
{
33 QSIMPLEQ_ENTRY(MirrorBuffer
) next
;
36 typedef struct MirrorBlockJob
{
39 BlockDriverState
*target
;
40 BlockDriverState
*base
;
41 /* The name of the graph node to replace */
43 /* The BDS to replace */
44 BlockDriverState
*to_replace
;
45 /* Used to block operations on the drive-mirror-replace target */
46 Error
*replace_blocker
;
48 BlockdevOnError on_source_error
, on_target_error
;
54 unsigned long *cow_bitmap
;
55 BdrvDirtyBitmap
*dirty_bitmap
;
58 QSIMPLEQ_HEAD(, MirrorBuffer
) buf_free
;
61 unsigned long *in_flight_bitmap
;
63 int sectors_in_flight
;
67 int target_cluster_sectors
;
71 typedef struct MirrorOp
{
78 static BlockErrorAction
mirror_error_action(MirrorBlockJob
*s
, bool read
,
83 return block_job_error_action(&s
->common
, s
->on_source_error
,
86 return block_job_error_action(&s
->common
, s
->on_target_error
,
91 static void mirror_iteration_done(MirrorOp
*op
, int ret
)
93 MirrorBlockJob
*s
= op
->s
;
96 int i
, nb_chunks
, sectors_per_chunk
;
98 trace_mirror_iteration_done(s
, op
->sector_num
, op
->nb_sectors
, ret
);
101 s
->sectors_in_flight
-= op
->nb_sectors
;
103 for (i
= 0; i
< op
->qiov
.niov
; i
++) {
104 MirrorBuffer
*buf
= (MirrorBuffer
*) iov
[i
].iov_base
;
105 QSIMPLEQ_INSERT_TAIL(&s
->buf_free
, buf
, next
);
109 sectors_per_chunk
= s
->granularity
>> BDRV_SECTOR_BITS
;
110 chunk_num
= op
->sector_num
/ sectors_per_chunk
;
111 nb_chunks
= DIV_ROUND_UP(op
->nb_sectors
, sectors_per_chunk
);
112 bitmap_clear(s
->in_flight_bitmap
, chunk_num
, nb_chunks
);
115 bitmap_set(s
->cow_bitmap
, chunk_num
, nb_chunks
);
117 s
->common
.offset
+= (uint64_t)op
->nb_sectors
* BDRV_SECTOR_SIZE
;
120 qemu_iovec_destroy(&op
->qiov
);
123 if (s
->waiting_for_io
) {
124 qemu_coroutine_enter(s
->common
.co
, NULL
);
128 static void mirror_write_complete(void *opaque
, int ret
)
130 MirrorOp
*op
= opaque
;
131 MirrorBlockJob
*s
= op
->s
;
133 BlockErrorAction action
;
135 bdrv_set_dirty_bitmap(s
->dirty_bitmap
, op
->sector_num
, op
->nb_sectors
);
136 action
= mirror_error_action(s
, false, -ret
);
137 if (action
== BLOCK_ERROR_ACTION_REPORT
&& s
->ret
>= 0) {
141 mirror_iteration_done(op
, ret
);
144 static void mirror_read_complete(void *opaque
, int ret
)
146 MirrorOp
*op
= opaque
;
147 MirrorBlockJob
*s
= op
->s
;
149 BlockErrorAction action
;
151 bdrv_set_dirty_bitmap(s
->dirty_bitmap
, op
->sector_num
, op
->nb_sectors
);
152 action
= mirror_error_action(s
, true, -ret
);
153 if (action
== BLOCK_ERROR_ACTION_REPORT
&& s
->ret
>= 0) {
157 mirror_iteration_done(op
, ret
);
160 bdrv_aio_writev(s
->target
, op
->sector_num
, &op
->qiov
, op
->nb_sectors
,
161 mirror_write_complete
, op
);
164 static inline void mirror_clip_sectors(MirrorBlockJob
*s
,
168 *nb_sectors
= MIN(*nb_sectors
,
169 s
->bdev_length
/ BDRV_SECTOR_SIZE
- sector_num
);
172 /* Round sector_num and/or nb_sectors to target cluster if COW is needed, and
173 * return the offset of the adjusted tail sector against original. */
174 static int mirror_cow_align(MirrorBlockJob
*s
,
180 int chunk_sectors
= s
->granularity
>> BDRV_SECTOR_BITS
;
181 int64_t align_sector_num
= *sector_num
;
182 int align_nb_sectors
= *nb_sectors
;
183 int max_sectors
= chunk_sectors
* s
->max_iov
;
185 need_cow
= !test_bit(*sector_num
/ chunk_sectors
, s
->cow_bitmap
);
186 need_cow
|= !test_bit((*sector_num
+ *nb_sectors
- 1) / chunk_sectors
,
189 bdrv_round_to_clusters(s
->target
, *sector_num
, *nb_sectors
,
190 &align_sector_num
, &align_nb_sectors
);
193 if (align_nb_sectors
> max_sectors
) {
194 align_nb_sectors
= max_sectors
;
196 align_nb_sectors
= QEMU_ALIGN_DOWN(align_nb_sectors
,
197 s
->target_cluster_sectors
);
200 /* Clipping may result in align_nb_sectors unaligned to chunk boundary, but
201 * that doesn't matter because it's already the end of source image. */
202 mirror_clip_sectors(s
, align_sector_num
, &align_nb_sectors
);
204 ret
= align_sector_num
+ align_nb_sectors
- (*sector_num
+ *nb_sectors
);
205 *sector_num
= align_sector_num
;
206 *nb_sectors
= align_nb_sectors
;
211 static inline void mirror_wait_for_io(MirrorBlockJob
*s
)
213 assert(!s
->waiting_for_io
);
214 s
->waiting_for_io
= true;
215 qemu_coroutine_yield();
216 s
->waiting_for_io
= false;
219 /* Submit async read while handling COW.
220 * Returns: nb_sectors if no alignment is necessary, or
221 * (new_end - sector_num) if tail is rounded up or down due to
222 * alignment or buffer limit.
224 static int mirror_do_read(MirrorBlockJob
*s
, int64_t sector_num
,
227 BlockDriverState
*source
= s
->common
.bs
;
228 int sectors_per_chunk
, nb_chunks
;
229 int ret
= nb_sectors
;
232 sectors_per_chunk
= s
->granularity
>> BDRV_SECTOR_BITS
;
234 /* We can only handle as much as buf_size at a time. */
235 nb_sectors
= MIN(s
->buf_size
>> BDRV_SECTOR_BITS
, nb_sectors
);
239 ret
+= mirror_cow_align(s
, §or_num
, &nb_sectors
);
241 assert(nb_sectors
<< BDRV_SECTOR_BITS
<= s
->buf_size
);
242 /* The sector range must meet granularity because:
243 * 1) Caller passes in aligned values;
244 * 2) mirror_cow_align is used only when target cluster is larger. */
245 assert(!(sector_num
% sectors_per_chunk
));
246 nb_chunks
= DIV_ROUND_UP(nb_sectors
, sectors_per_chunk
);
248 while (s
->buf_free_count
< nb_chunks
) {
249 trace_mirror_yield_in_flight(s
, sector_num
, s
->in_flight
);
250 mirror_wait_for_io(s
);
253 /* Allocate a MirrorOp that is used as an AIO callback. */
254 op
= g_new(MirrorOp
, 1);
256 op
->sector_num
= sector_num
;
257 op
->nb_sectors
= nb_sectors
;
259 /* Now make a QEMUIOVector taking enough granularity-sized chunks
262 qemu_iovec_init(&op
->qiov
, nb_chunks
);
263 while (nb_chunks
-- > 0) {
264 MirrorBuffer
*buf
= QSIMPLEQ_FIRST(&s
->buf_free
);
265 size_t remaining
= nb_sectors
* BDRV_SECTOR_SIZE
- op
->qiov
.size
;
267 QSIMPLEQ_REMOVE_HEAD(&s
->buf_free
, next
);
269 qemu_iovec_add(&op
->qiov
, buf
, MIN(s
->granularity
, remaining
));
272 /* Copy the dirty cluster. */
274 s
->sectors_in_flight
+= nb_sectors
;
275 trace_mirror_one_iteration(s
, sector_num
, nb_sectors
);
277 bdrv_aio_readv(source
, sector_num
, &op
->qiov
, nb_sectors
,
278 mirror_read_complete
, op
);
282 static void mirror_do_zero_or_discard(MirrorBlockJob
*s
,
289 /* Allocate a MirrorOp that is used as an AIO callback. The qiov is zeroed
290 * so the freeing in mirror_iteration_done is nop. */
291 op
= g_new0(MirrorOp
, 1);
293 op
->sector_num
= sector_num
;
294 op
->nb_sectors
= nb_sectors
;
297 s
->sectors_in_flight
+= nb_sectors
;
299 bdrv_aio_discard(s
->target
, sector_num
, op
->nb_sectors
,
300 mirror_write_complete
, op
);
302 bdrv_aio_write_zeroes(s
->target
, sector_num
, op
->nb_sectors
,
303 s
->unmap
? BDRV_REQ_MAY_UNMAP
: 0,
304 mirror_write_complete
, op
);
308 static uint64_t coroutine_fn
mirror_iteration(MirrorBlockJob
*s
)
310 BlockDriverState
*source
= s
->common
.bs
;
311 int64_t sector_num
, first_chunk
;
312 uint64_t delay_ns
= 0;
313 /* At least the first dirty chunk is mirrored in one iteration. */
315 int64_t end
= s
->bdev_length
/ BDRV_SECTOR_SIZE
;
316 int sectors_per_chunk
= s
->granularity
>> BDRV_SECTOR_BITS
;
318 sector_num
= hbitmap_iter_next(&s
->hbi
);
319 if (sector_num
< 0) {
320 bdrv_dirty_iter_init(s
->dirty_bitmap
, &s
->hbi
);
321 sector_num
= hbitmap_iter_next(&s
->hbi
);
322 trace_mirror_restart_iter(s
, bdrv_get_dirty_count(s
->dirty_bitmap
));
323 assert(sector_num
>= 0);
326 first_chunk
= sector_num
/ sectors_per_chunk
;
327 while (test_bit(first_chunk
, s
->in_flight_bitmap
)) {
328 trace_mirror_yield_in_flight(s
, first_chunk
, s
->in_flight
);
329 mirror_wait_for_io(s
);
332 /* Find the number of consective dirty chunks following the first dirty
333 * one, and wait for in flight requests in them. */
334 while (nb_chunks
* sectors_per_chunk
< (s
->buf_size
>> BDRV_SECTOR_BITS
)) {
335 int64_t hbitmap_next
;
336 int64_t next_sector
= sector_num
+ nb_chunks
* sectors_per_chunk
;
337 int64_t next_chunk
= next_sector
/ sectors_per_chunk
;
338 if (next_sector
>= end
||
339 !bdrv_get_dirty(source
, s
->dirty_bitmap
, next_sector
)) {
342 if (test_bit(next_chunk
, s
->in_flight_bitmap
)) {
346 hbitmap_next
= hbitmap_iter_next(&s
->hbi
);
347 if (hbitmap_next
> next_sector
|| hbitmap_next
< 0) {
348 /* The bitmap iterator's cache is stale, refresh it */
349 bdrv_set_dirty_iter(&s
->hbi
, next_sector
);
350 hbitmap_next
= hbitmap_iter_next(&s
->hbi
);
352 assert(hbitmap_next
== next_sector
);
356 /* Clear dirty bits before querying the block status, because
357 * calling bdrv_get_block_status_above could yield - if some blocks are
358 * marked dirty in this window, we need to know.
360 bdrv_reset_dirty_bitmap(s
->dirty_bitmap
, sector_num
,
361 nb_chunks
* sectors_per_chunk
);
362 bitmap_set(s
->in_flight_bitmap
, sector_num
/ sectors_per_chunk
, nb_chunks
);
363 while (nb_chunks
> 0 && sector_num
< end
) {
366 BlockDriverState
*file
;
370 MIRROR_METHOD_DISCARD
371 } mirror_method
= MIRROR_METHOD_COPY
;
373 assert(!(sector_num
% sectors_per_chunk
));
374 ret
= bdrv_get_block_status_above(source
, NULL
, sector_num
,
375 nb_chunks
* sectors_per_chunk
,
378 io_sectors
= nb_chunks
* sectors_per_chunk
;
381 io_sectors
-= io_sectors
% sectors_per_chunk
;
382 if (io_sectors
< sectors_per_chunk
) {
383 io_sectors
= sectors_per_chunk
;
384 } else if (ret
>= 0 && !(ret
& BDRV_BLOCK_DATA
)) {
385 int64_t target_sector_num
;
386 int target_nb_sectors
;
387 bdrv_round_to_clusters(s
->target
, sector_num
, io_sectors
,
388 &target_sector_num
, &target_nb_sectors
);
389 if (target_sector_num
== sector_num
&&
390 target_nb_sectors
== io_sectors
) {
391 mirror_method
= ret
& BDRV_BLOCK_ZERO
?
393 MIRROR_METHOD_DISCARD
;
397 mirror_clip_sectors(s
, sector_num
, &io_sectors
);
398 switch (mirror_method
) {
399 case MIRROR_METHOD_COPY
:
400 io_sectors
= mirror_do_read(s
, sector_num
, io_sectors
);
402 case MIRROR_METHOD_ZERO
:
403 mirror_do_zero_or_discard(s
, sector_num
, io_sectors
, false);
405 case MIRROR_METHOD_DISCARD
:
406 mirror_do_zero_or_discard(s
, sector_num
, io_sectors
, true);
412 sector_num
+= io_sectors
;
413 nb_chunks
-= DIV_ROUND_UP(io_sectors
, sectors_per_chunk
);
414 delay_ns
+= ratelimit_calculate_delay(&s
->limit
, io_sectors
);
419 static void mirror_free_init(MirrorBlockJob
*s
)
421 int granularity
= s
->granularity
;
422 size_t buf_size
= s
->buf_size
;
423 uint8_t *buf
= s
->buf
;
425 assert(s
->buf_free_count
== 0);
426 QSIMPLEQ_INIT(&s
->buf_free
);
427 while (buf_size
!= 0) {
428 MirrorBuffer
*cur
= (MirrorBuffer
*)buf
;
429 QSIMPLEQ_INSERT_TAIL(&s
->buf_free
, cur
, next
);
431 buf_size
-= granularity
;
436 static void mirror_drain(MirrorBlockJob
*s
)
438 while (s
->in_flight
> 0) {
439 mirror_wait_for_io(s
);
447 static void mirror_exit(BlockJob
*job
, void *opaque
)
449 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
450 MirrorExitData
*data
= opaque
;
451 AioContext
*replace_aio_context
= NULL
;
452 BlockDriverState
*src
= s
->common
.bs
;
454 /* Make sure that the source BDS doesn't go away before we called
455 * block_job_completed(). */
459 replace_aio_context
= bdrv_get_aio_context(s
->to_replace
);
460 aio_context_acquire(replace_aio_context
);
463 if (s
->should_complete
&& data
->ret
== 0) {
464 BlockDriverState
*to_replace
= s
->common
.bs
;
466 to_replace
= s
->to_replace
;
469 /* This was checked in mirror_start_job(), but meanwhile one of the
470 * nodes could have been newly attached to a BlockBackend. */
471 if (bdrv_has_blk(to_replace
) && bdrv_has_blk(s
->target
)) {
472 error_report("block job: Can't create node with two BlockBackends");
477 if (bdrv_get_flags(s
->target
) != bdrv_get_flags(to_replace
)) {
478 bdrv_reopen(s
->target
, bdrv_get_flags(to_replace
), NULL
);
480 bdrv_replace_in_backing_chain(to_replace
, s
->target
);
481 /* We just changed the BDS the job BB refers to */
482 blk_remove_bs(job
->blk
);
483 blk_insert_bs(job
->blk
, src
);
488 bdrv_op_unblock_all(s
->to_replace
, s
->replace_blocker
);
489 error_free(s
->replace_blocker
);
490 bdrv_unref(s
->to_replace
);
492 if (replace_aio_context
) {
493 aio_context_release(replace_aio_context
);
496 bdrv_op_unblock_all(s
->target
, s
->common
.blocker
);
497 bdrv_unref(s
->target
);
498 block_job_completed(&s
->common
, data
->ret
);
500 bdrv_drained_end(src
);
501 if (qemu_get_aio_context() == bdrv_get_aio_context(src
)) {
502 aio_enable_external(iohandler_get_aio_context());
507 static void coroutine_fn
mirror_run(void *opaque
)
509 MirrorBlockJob
*s
= opaque
;
510 MirrorExitData
*data
;
511 BlockDriverState
*bs
= s
->common
.bs
;
512 int64_t sector_num
, end
, length
;
513 uint64_t last_pause_ns
;
515 char backing_filename
[2]; /* we only need 2 characters because we are only
516 checking for a NULL string */
519 int target_cluster_size
= BDRV_SECTOR_SIZE
;
521 if (block_job_is_cancelled(&s
->common
)) {
525 s
->bdev_length
= bdrv_getlength(bs
);
526 if (s
->bdev_length
< 0) {
527 ret
= s
->bdev_length
;
529 } else if (s
->bdev_length
== 0) {
530 /* Report BLOCK_JOB_READY and wait for complete. */
531 block_job_event_ready(&s
->common
);
533 while (!block_job_is_cancelled(&s
->common
) && !s
->should_complete
) {
534 block_job_yield(&s
->common
);
536 s
->common
.cancelled
= false;
540 length
= DIV_ROUND_UP(s
->bdev_length
, s
->granularity
);
541 s
->in_flight_bitmap
= bitmap_new(length
);
543 /* If we have no backing file yet in the destination, we cannot let
544 * the destination do COW. Instead, we copy sectors around the
545 * dirty data if needed. We need a bitmap to do that.
547 bdrv_get_backing_filename(s
->target
, backing_filename
,
548 sizeof(backing_filename
));
549 if (!bdrv_get_info(s
->target
, &bdi
) && bdi
.cluster_size
) {
550 target_cluster_size
= bdi
.cluster_size
;
552 if (backing_filename
[0] && !s
->target
->backing
553 && s
->granularity
< target_cluster_size
) {
554 s
->buf_size
= MAX(s
->buf_size
, target_cluster_size
);
555 s
->cow_bitmap
= bitmap_new(length
);
557 s
->target_cluster_sectors
= target_cluster_size
>> BDRV_SECTOR_BITS
;
558 s
->max_iov
= MIN(s
->common
.bs
->bl
.max_iov
, s
->target
->bl
.max_iov
);
560 end
= s
->bdev_length
/ BDRV_SECTOR_SIZE
;
561 s
->buf
= qemu_try_blockalign(bs
, s
->buf_size
);
562 if (s
->buf
== NULL
) {
569 last_pause_ns
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
);
570 if (!s
->is_none_mode
) {
571 /* First part, loop on the sectors and initialize the dirty bitmap. */
572 BlockDriverState
*base
= s
->base
;
573 bool mark_all_dirty
= s
->base
== NULL
&& !bdrv_has_zero_init(s
->target
);
575 for (sector_num
= 0; sector_num
< end
; ) {
576 /* Just to make sure we are not exceeding int limit. */
577 int nb_sectors
= MIN(INT_MAX
>> BDRV_SECTOR_BITS
,
579 int64_t now
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
);
581 if (now
- last_pause_ns
> SLICE_TIME
) {
583 block_job_sleep_ns(&s
->common
, QEMU_CLOCK_REALTIME
, 0);
586 if (block_job_is_cancelled(&s
->common
)) {
590 ret
= bdrv_is_allocated_above(bs
, base
, sector_num
, nb_sectors
, &n
);
597 if (ret
== 1 || mark_all_dirty
) {
598 bdrv_set_dirty_bitmap(s
->dirty_bitmap
, sector_num
, n
);
604 bdrv_dirty_iter_init(s
->dirty_bitmap
, &s
->hbi
);
606 uint64_t delay_ns
= 0;
608 bool should_complete
;
615 cnt
= bdrv_get_dirty_count(s
->dirty_bitmap
);
616 /* s->common.offset contains the number of bytes already processed so
617 * far, cnt is the number of dirty sectors remaining and
618 * s->sectors_in_flight is the number of sectors currently being
619 * processed; together those are the current total operation length */
620 s
->common
.len
= s
->common
.offset
+
621 (cnt
+ s
->sectors_in_flight
) * BDRV_SECTOR_SIZE
;
623 /* Note that even when no rate limit is applied we need to yield
624 * periodically with no pending I/O so that bdrv_drain_all() returns.
625 * We do so every SLICE_TIME nanoseconds, or when there is an error,
626 * or when the source is clean, whichever comes first.
628 if (qemu_clock_get_ns(QEMU_CLOCK_REALTIME
) - last_pause_ns
< SLICE_TIME
&&
629 s
->common
.iostatus
== BLOCK_DEVICE_IO_STATUS_OK
) {
630 if (s
->in_flight
== MAX_IN_FLIGHT
|| s
->buf_free_count
== 0 ||
631 (cnt
== 0 && s
->in_flight
> 0)) {
632 trace_mirror_yield(s
, s
->in_flight
, s
->buf_free_count
, cnt
);
633 mirror_wait_for_io(s
);
635 } else if (cnt
!= 0) {
636 delay_ns
= mirror_iteration(s
);
640 should_complete
= false;
641 if (s
->in_flight
== 0 && cnt
== 0) {
642 trace_mirror_before_flush(s
);
643 ret
= bdrv_flush(s
->target
);
645 if (mirror_error_action(s
, false, -ret
) ==
646 BLOCK_ERROR_ACTION_REPORT
) {
650 /* We're out of the streaming phase. From now on, if the job
651 * is cancelled we will actually complete all pending I/O and
652 * report completion. This way, block-job-cancel will leave
653 * the target in a consistent state.
656 block_job_event_ready(&s
->common
);
660 should_complete
= s
->should_complete
||
661 block_job_is_cancelled(&s
->common
);
662 cnt
= bdrv_get_dirty_count(s
->dirty_bitmap
);
666 if (cnt
== 0 && should_complete
) {
667 /* The dirty bitmap is not updated while operations are pending.
668 * If we're about to exit, wait for pending operations before
669 * calling bdrv_get_dirty_count(bs), or we may exit while the
670 * source has dirty data to copy!
672 * Note that I/O can be submitted by the guest while
673 * mirror_populate runs.
675 trace_mirror_before_drain(s
, cnt
);
677 cnt
= bdrv_get_dirty_count(s
->dirty_bitmap
);
681 trace_mirror_before_sleep(s
, cnt
, s
->synced
, delay_ns
);
683 block_job_sleep_ns(&s
->common
, QEMU_CLOCK_REALTIME
, delay_ns
);
684 if (block_job_is_cancelled(&s
->common
)) {
687 } else if (!should_complete
) {
688 delay_ns
= (s
->in_flight
== 0 && cnt
== 0 ? SLICE_TIME
: 0);
689 block_job_sleep_ns(&s
->common
, QEMU_CLOCK_REALTIME
, delay_ns
);
690 } else if (cnt
== 0) {
691 /* The two disks are in sync. Exit and report successful
694 assert(QLIST_EMPTY(&bs
->tracked_requests
));
695 s
->common
.cancelled
= false;
698 last_pause_ns
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
);
702 if (s
->in_flight
> 0) {
703 /* We get here only if something went wrong. Either the job failed,
704 * or it was cancelled prematurely so that we do not guarantee that
705 * the target is a copy of the source.
707 assert(ret
< 0 || (!s
->synced
&& block_job_is_cancelled(&s
->common
)));
711 assert(s
->in_flight
== 0);
713 g_free(s
->cow_bitmap
);
714 g_free(s
->in_flight_bitmap
);
715 bdrv_release_dirty_bitmap(bs
, s
->dirty_bitmap
);
717 data
= g_malloc(sizeof(*data
));
719 /* Before we switch to target in mirror_exit, make sure data doesn't
721 bdrv_drained_begin(s
->common
.bs
);
722 if (qemu_get_aio_context() == bdrv_get_aio_context(bs
)) {
723 /* FIXME: virtio host notifiers run on iohandler_ctx, therefore the
724 * above bdrv_drained_end isn't enough to quiesce it. This is ugly, we
725 * need a block layer API change to achieve this. */
726 aio_disable_external(iohandler_get_aio_context());
728 block_job_defer_to_main_loop(&s
->common
, mirror_exit
, data
);
731 static void mirror_set_speed(BlockJob
*job
, int64_t speed
, Error
**errp
)
733 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
736 error_setg(errp
, QERR_INVALID_PARAMETER
, "speed");
739 ratelimit_set_speed(&s
->limit
, speed
/ BDRV_SECTOR_SIZE
, SLICE_TIME
);
742 static void mirror_complete(BlockJob
*job
, Error
**errp
)
744 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
745 Error
*local_err
= NULL
;
748 ret
= bdrv_open_backing_file(s
->target
, NULL
, "backing", &local_err
);
750 error_propagate(errp
, local_err
);
754 error_setg(errp
, QERR_BLOCK_JOB_NOT_READY
, job
->id
);
758 /* check the target bs is not blocked and block all operations on it */
760 AioContext
*replace_aio_context
;
762 s
->to_replace
= bdrv_find_node(s
->replaces
);
763 if (!s
->to_replace
) {
764 error_setg(errp
, "Node name '%s' not found", s
->replaces
);
768 replace_aio_context
= bdrv_get_aio_context(s
->to_replace
);
769 aio_context_acquire(replace_aio_context
);
771 error_setg(&s
->replace_blocker
,
772 "block device is in use by block-job-complete");
773 bdrv_op_block_all(s
->to_replace
, s
->replace_blocker
);
774 bdrv_ref(s
->to_replace
);
776 aio_context_release(replace_aio_context
);
779 s
->should_complete
= true;
780 block_job_enter(&s
->common
);
783 static const BlockJobDriver mirror_job_driver
= {
784 .instance_size
= sizeof(MirrorBlockJob
),
785 .job_type
= BLOCK_JOB_TYPE_MIRROR
,
786 .set_speed
= mirror_set_speed
,
787 .complete
= mirror_complete
,
790 static const BlockJobDriver commit_active_job_driver
= {
791 .instance_size
= sizeof(MirrorBlockJob
),
792 .job_type
= BLOCK_JOB_TYPE_COMMIT
,
793 .set_speed
= mirror_set_speed
,
794 .complete
= mirror_complete
,
797 static void mirror_start_job(BlockDriverState
*bs
, BlockDriverState
*target
,
798 const char *replaces
,
799 int64_t speed
, uint32_t granularity
,
801 BlockdevOnError on_source_error
,
802 BlockdevOnError on_target_error
,
804 BlockCompletionFunc
*cb
,
805 void *opaque
, Error
**errp
,
806 const BlockJobDriver
*driver
,
807 bool is_none_mode
, BlockDriverState
*base
)
810 BlockDriverState
*replaced_bs
;
812 if (granularity
== 0) {
813 granularity
= bdrv_get_default_bitmap_granularity(target
);
816 assert ((granularity
& (granularity
- 1)) == 0);
819 error_setg(errp
, "Invalid parameter 'buf-size'");
824 buf_size
= DEFAULT_MIRROR_BUF_SIZE
;
827 /* We can't support this case as long as the block layer can't handle
828 * multiple BlockBackends per BlockDriverState. */
830 replaced_bs
= bdrv_lookup_bs(replaces
, replaces
, errp
);
831 if (replaced_bs
== NULL
) {
837 if (bdrv_has_blk(replaced_bs
) && bdrv_has_blk(target
)) {
838 error_setg(errp
, "Can't create node with two BlockBackends");
842 s
= block_job_create(driver
, bs
, speed
, cb
, opaque
, errp
);
847 s
->replaces
= g_strdup(replaces
);
848 s
->on_source_error
= on_source_error
;
849 s
->on_target_error
= on_target_error
;
851 s
->is_none_mode
= is_none_mode
;
853 s
->granularity
= granularity
;
854 s
->buf_size
= ROUND_UP(buf_size
, granularity
);
857 s
->dirty_bitmap
= bdrv_create_dirty_bitmap(bs
, granularity
, NULL
, errp
);
858 if (!s
->dirty_bitmap
) {
860 block_job_unref(&s
->common
);
864 bdrv_op_block_all(s
->target
, s
->common
.blocker
);
866 s
->common
.co
= qemu_coroutine_create(mirror_run
);
867 trace_mirror_start(bs
, s
, s
->common
.co
, opaque
);
868 qemu_coroutine_enter(s
->common
.co
, s
);
871 void mirror_start(BlockDriverState
*bs
, BlockDriverState
*target
,
872 const char *replaces
,
873 int64_t speed
, uint32_t granularity
, int64_t buf_size
,
874 MirrorSyncMode mode
, BlockdevOnError on_source_error
,
875 BlockdevOnError on_target_error
,
877 BlockCompletionFunc
*cb
,
878 void *opaque
, Error
**errp
)
881 BlockDriverState
*base
;
883 if (mode
== MIRROR_SYNC_MODE_INCREMENTAL
) {
884 error_setg(errp
, "Sync mode 'incremental' not supported");
887 is_none_mode
= mode
== MIRROR_SYNC_MODE_NONE
;
888 base
= mode
== MIRROR_SYNC_MODE_TOP
? backing_bs(bs
) : NULL
;
889 mirror_start_job(bs
, target
, replaces
,
890 speed
, granularity
, buf_size
,
891 on_source_error
, on_target_error
, unmap
, cb
, opaque
, errp
,
892 &mirror_job_driver
, is_none_mode
, base
);
895 void commit_active_start(BlockDriverState
*bs
, BlockDriverState
*base
,
897 BlockdevOnError on_error
,
898 BlockCompletionFunc
*cb
,
899 void *opaque
, Error
**errp
)
901 int64_t length
, base_length
;
904 Error
*local_err
= NULL
;
906 orig_base_flags
= bdrv_get_flags(base
);
908 if (bdrv_reopen(base
, bs
->open_flags
, errp
)) {
912 length
= bdrv_getlength(bs
);
914 error_setg_errno(errp
, -length
,
915 "Unable to determine length of %s", bs
->filename
);
916 goto error_restore_flags
;
919 base_length
= bdrv_getlength(base
);
920 if (base_length
< 0) {
921 error_setg_errno(errp
, -base_length
,
922 "Unable to determine length of %s", base
->filename
);
923 goto error_restore_flags
;
926 if (length
> base_length
) {
927 ret
= bdrv_truncate(base
, length
);
929 error_setg_errno(errp
, -ret
,
930 "Top image %s is larger than base image %s, and "
931 "resize of base image failed",
932 bs
->filename
, base
->filename
);
933 goto error_restore_flags
;
938 mirror_start_job(bs
, base
, NULL
, speed
, 0, 0,
939 on_error
, on_error
, false, cb
, opaque
, &local_err
,
940 &commit_active_job_driver
, false, base
);
942 error_propagate(errp
, local_err
);
943 goto error_restore_flags
;
949 /* ignore error and errp for bdrv_reopen, because we want to propagate
950 * the original error */
951 bdrv_reopen(base
, orig_base_flags
, NULL
);