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_int.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"
24 #define SLICE_TIME 100000000ULL /* ns */
25 #define MAX_IN_FLIGHT 16
26 #define MAX_IO_SECTORS ((1 << 20) >> BDRV_SECTOR_BITS) /* 1 Mb */
27 #define DEFAULT_MIRROR_BUF_SIZE \
28 (MAX_IN_FLIGHT * MAX_IO_SECTORS * BDRV_SECTOR_SIZE)
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
*base
;
42 /* The name of the graph node to replace */
44 /* The BDS to replace */
45 BlockDriverState
*to_replace
;
46 /* Used to block operations on the drive-mirror-replace target */
47 Error
*replace_blocker
;
49 BlockMirrorBackingMode backing_mode
;
50 BlockdevOnError on_source_error
, on_target_error
;
56 unsigned long *cow_bitmap
;
57 BdrvDirtyBitmap
*dirty_bitmap
;
58 BdrvDirtyBitmapIter
*dbi
;
60 QSIMPLEQ_HEAD(, MirrorBuffer
) buf_free
;
63 uint64_t last_pause_ns
;
64 unsigned long *in_flight_bitmap
;
66 int64_t sectors_in_flight
;
70 int target_cluster_sectors
;
72 bool initial_zeroing_ongoing
;
75 typedef struct MirrorOp
{
82 static BlockErrorAction
mirror_error_action(MirrorBlockJob
*s
, bool read
,
87 return block_job_error_action(&s
->common
, s
->on_source_error
,
90 return block_job_error_action(&s
->common
, s
->on_target_error
,
95 static void mirror_iteration_done(MirrorOp
*op
, int ret
)
97 MirrorBlockJob
*s
= op
->s
;
100 int i
, nb_chunks
, sectors_per_chunk
;
102 trace_mirror_iteration_done(s
, op
->sector_num
, op
->nb_sectors
, ret
);
105 s
->sectors_in_flight
-= op
->nb_sectors
;
107 for (i
= 0; i
< op
->qiov
.niov
; i
++) {
108 MirrorBuffer
*buf
= (MirrorBuffer
*) iov
[i
].iov_base
;
109 QSIMPLEQ_INSERT_TAIL(&s
->buf_free
, buf
, next
);
113 sectors_per_chunk
= s
->granularity
>> BDRV_SECTOR_BITS
;
114 chunk_num
= op
->sector_num
/ sectors_per_chunk
;
115 nb_chunks
= DIV_ROUND_UP(op
->nb_sectors
, sectors_per_chunk
);
116 bitmap_clear(s
->in_flight_bitmap
, chunk_num
, nb_chunks
);
119 bitmap_set(s
->cow_bitmap
, chunk_num
, nb_chunks
);
121 if (!s
->initial_zeroing_ongoing
) {
122 s
->common
.offset
+= (uint64_t)op
->nb_sectors
* BDRV_SECTOR_SIZE
;
125 qemu_iovec_destroy(&op
->qiov
);
128 if (s
->waiting_for_io
) {
129 qemu_coroutine_enter(s
->common
.co
);
133 static void mirror_write_complete(void *opaque
, int ret
)
135 MirrorOp
*op
= opaque
;
136 MirrorBlockJob
*s
= op
->s
;
138 aio_context_acquire(blk_get_aio_context(s
->common
.blk
));
140 BlockErrorAction action
;
142 bdrv_set_dirty_bitmap(s
->dirty_bitmap
, op
->sector_num
, op
->nb_sectors
);
143 action
= mirror_error_action(s
, false, -ret
);
144 if (action
== BLOCK_ERROR_ACTION_REPORT
&& s
->ret
>= 0) {
148 mirror_iteration_done(op
, ret
);
149 aio_context_release(blk_get_aio_context(s
->common
.blk
));
152 static void mirror_read_complete(void *opaque
, int ret
)
154 MirrorOp
*op
= opaque
;
155 MirrorBlockJob
*s
= op
->s
;
157 aio_context_acquire(blk_get_aio_context(s
->common
.blk
));
159 BlockErrorAction action
;
161 bdrv_set_dirty_bitmap(s
->dirty_bitmap
, op
->sector_num
, op
->nb_sectors
);
162 action
= mirror_error_action(s
, true, -ret
);
163 if (action
== BLOCK_ERROR_ACTION_REPORT
&& s
->ret
>= 0) {
167 mirror_iteration_done(op
, ret
);
169 blk_aio_pwritev(s
->target
, op
->sector_num
* BDRV_SECTOR_SIZE
, &op
->qiov
,
170 0, mirror_write_complete
, op
);
172 aio_context_release(blk_get_aio_context(s
->common
.blk
));
175 static inline void mirror_clip_sectors(MirrorBlockJob
*s
,
179 *nb_sectors
= MIN(*nb_sectors
,
180 s
->bdev_length
/ BDRV_SECTOR_SIZE
- sector_num
);
183 /* Round sector_num and/or nb_sectors to target cluster if COW is needed, and
184 * return the offset of the adjusted tail sector against original. */
185 static int mirror_cow_align(MirrorBlockJob
*s
,
191 int chunk_sectors
= s
->granularity
>> BDRV_SECTOR_BITS
;
192 int64_t align_sector_num
= *sector_num
;
193 int align_nb_sectors
= *nb_sectors
;
194 int max_sectors
= chunk_sectors
* s
->max_iov
;
196 need_cow
= !test_bit(*sector_num
/ chunk_sectors
, s
->cow_bitmap
);
197 need_cow
|= !test_bit((*sector_num
+ *nb_sectors
- 1) / chunk_sectors
,
200 bdrv_round_sectors_to_clusters(blk_bs(s
->target
), *sector_num
,
201 *nb_sectors
, &align_sector_num
,
205 if (align_nb_sectors
> max_sectors
) {
206 align_nb_sectors
= max_sectors
;
208 align_nb_sectors
= QEMU_ALIGN_DOWN(align_nb_sectors
,
209 s
->target_cluster_sectors
);
212 /* Clipping may result in align_nb_sectors unaligned to chunk boundary, but
213 * that doesn't matter because it's already the end of source image. */
214 mirror_clip_sectors(s
, align_sector_num
, &align_nb_sectors
);
216 ret
= align_sector_num
+ align_nb_sectors
- (*sector_num
+ *nb_sectors
);
217 *sector_num
= align_sector_num
;
218 *nb_sectors
= align_nb_sectors
;
223 static inline void mirror_wait_for_io(MirrorBlockJob
*s
)
225 assert(!s
->waiting_for_io
);
226 s
->waiting_for_io
= true;
227 qemu_coroutine_yield();
228 s
->waiting_for_io
= false;
231 /* Submit async read while handling COW.
232 * Returns: The number of sectors copied after and including sector_num,
233 * excluding any sectors copied prior to sector_num due to alignment.
234 * This will be nb_sectors if no alignment is necessary, or
235 * (new_end - sector_num) if tail is rounded up or down due to
236 * alignment or buffer limit.
238 static int mirror_do_read(MirrorBlockJob
*s
, int64_t sector_num
,
241 BlockBackend
*source
= s
->common
.blk
;
242 int sectors_per_chunk
, nb_chunks
;
247 sectors_per_chunk
= s
->granularity
>> BDRV_SECTOR_BITS
;
248 max_sectors
= sectors_per_chunk
* s
->max_iov
;
250 /* We can only handle as much as buf_size at a time. */
251 nb_sectors
= MIN(s
->buf_size
>> BDRV_SECTOR_BITS
, nb_sectors
);
252 nb_sectors
= MIN(max_sectors
, nb_sectors
);
257 ret
+= mirror_cow_align(s
, §or_num
, &nb_sectors
);
259 assert(nb_sectors
<< BDRV_SECTOR_BITS
<= s
->buf_size
);
260 /* The sector range must meet granularity because:
261 * 1) Caller passes in aligned values;
262 * 2) mirror_cow_align is used only when target cluster is larger. */
263 assert(!(sector_num
% sectors_per_chunk
));
264 nb_chunks
= DIV_ROUND_UP(nb_sectors
, sectors_per_chunk
);
266 while (s
->buf_free_count
< nb_chunks
) {
267 trace_mirror_yield_in_flight(s
, sector_num
, s
->in_flight
);
268 mirror_wait_for_io(s
);
271 /* Allocate a MirrorOp that is used as an AIO callback. */
272 op
= g_new(MirrorOp
, 1);
274 op
->sector_num
= sector_num
;
275 op
->nb_sectors
= nb_sectors
;
277 /* Now make a QEMUIOVector taking enough granularity-sized chunks
280 qemu_iovec_init(&op
->qiov
, nb_chunks
);
281 while (nb_chunks
-- > 0) {
282 MirrorBuffer
*buf
= QSIMPLEQ_FIRST(&s
->buf_free
);
283 size_t remaining
= nb_sectors
* BDRV_SECTOR_SIZE
- op
->qiov
.size
;
285 QSIMPLEQ_REMOVE_HEAD(&s
->buf_free
, next
);
287 qemu_iovec_add(&op
->qiov
, buf
, MIN(s
->granularity
, remaining
));
290 /* Copy the dirty cluster. */
292 s
->sectors_in_flight
+= nb_sectors
;
293 trace_mirror_one_iteration(s
, sector_num
, nb_sectors
);
295 blk_aio_preadv(source
, sector_num
* BDRV_SECTOR_SIZE
, &op
->qiov
, 0,
296 mirror_read_complete
, op
);
300 static void mirror_do_zero_or_discard(MirrorBlockJob
*s
,
307 /* Allocate a MirrorOp that is used as an AIO callback. The qiov is zeroed
308 * so the freeing in mirror_iteration_done is nop. */
309 op
= g_new0(MirrorOp
, 1);
311 op
->sector_num
= sector_num
;
312 op
->nb_sectors
= nb_sectors
;
315 s
->sectors_in_flight
+= nb_sectors
;
317 blk_aio_pdiscard(s
->target
, sector_num
<< BDRV_SECTOR_BITS
,
318 op
->nb_sectors
<< BDRV_SECTOR_BITS
,
319 mirror_write_complete
, op
);
321 blk_aio_pwrite_zeroes(s
->target
, sector_num
* BDRV_SECTOR_SIZE
,
322 op
->nb_sectors
* BDRV_SECTOR_SIZE
,
323 s
->unmap
? BDRV_REQ_MAY_UNMAP
: 0,
324 mirror_write_complete
, op
);
328 static uint64_t coroutine_fn
mirror_iteration(MirrorBlockJob
*s
)
330 BlockDriverState
*source
= blk_bs(s
->common
.blk
);
331 int64_t sector_num
, first_chunk
;
332 uint64_t delay_ns
= 0;
333 /* At least the first dirty chunk is mirrored in one iteration. */
335 int64_t end
= s
->bdev_length
/ BDRV_SECTOR_SIZE
;
336 int sectors_per_chunk
= s
->granularity
>> BDRV_SECTOR_BITS
;
337 bool write_zeroes_ok
= bdrv_can_write_zeroes_with_unmap(blk_bs(s
->target
));
338 int max_io_sectors
= MAX((s
->buf_size
>> BDRV_SECTOR_BITS
) / MAX_IN_FLIGHT
,
341 sector_num
= bdrv_dirty_iter_next(s
->dbi
);
342 if (sector_num
< 0) {
343 bdrv_set_dirty_iter(s
->dbi
, 0);
344 sector_num
= bdrv_dirty_iter_next(s
->dbi
);
345 trace_mirror_restart_iter(s
, bdrv_get_dirty_count(s
->dirty_bitmap
));
346 assert(sector_num
>= 0);
349 first_chunk
= sector_num
/ sectors_per_chunk
;
350 while (test_bit(first_chunk
, s
->in_flight_bitmap
)) {
351 trace_mirror_yield_in_flight(s
, sector_num
, s
->in_flight
);
352 mirror_wait_for_io(s
);
355 block_job_pause_point(&s
->common
);
357 /* Find the number of consective dirty chunks following the first dirty
358 * one, and wait for in flight requests in them. */
359 while (nb_chunks
* sectors_per_chunk
< (s
->buf_size
>> BDRV_SECTOR_BITS
)) {
361 int64_t next_sector
= sector_num
+ nb_chunks
* sectors_per_chunk
;
362 int64_t next_chunk
= next_sector
/ sectors_per_chunk
;
363 if (next_sector
>= end
||
364 !bdrv_get_dirty(source
, s
->dirty_bitmap
, next_sector
)) {
367 if (test_bit(next_chunk
, s
->in_flight_bitmap
)) {
371 next_dirty
= bdrv_dirty_iter_next(s
->dbi
);
372 if (next_dirty
> next_sector
|| next_dirty
< 0) {
373 /* The bitmap iterator's cache is stale, refresh it */
374 bdrv_set_dirty_iter(s
->dbi
, next_sector
);
375 next_dirty
= bdrv_dirty_iter_next(s
->dbi
);
377 assert(next_dirty
== next_sector
);
381 /* Clear dirty bits before querying the block status, because
382 * calling bdrv_get_block_status_above could yield - if some blocks are
383 * marked dirty in this window, we need to know.
385 bdrv_reset_dirty_bitmap(s
->dirty_bitmap
, sector_num
,
386 nb_chunks
* sectors_per_chunk
);
387 bitmap_set(s
->in_flight_bitmap
, sector_num
/ sectors_per_chunk
, nb_chunks
);
388 while (nb_chunks
> 0 && sector_num
< end
) {
390 int io_sectors
, io_sectors_acct
;
391 BlockDriverState
*file
;
395 MIRROR_METHOD_DISCARD
396 } mirror_method
= MIRROR_METHOD_COPY
;
398 assert(!(sector_num
% sectors_per_chunk
));
399 ret
= bdrv_get_block_status_above(source
, NULL
, sector_num
,
400 nb_chunks
* sectors_per_chunk
,
403 io_sectors
= MIN(nb_chunks
* sectors_per_chunk
, max_io_sectors
);
404 } else if (ret
& BDRV_BLOCK_DATA
) {
405 io_sectors
= MIN(io_sectors
, max_io_sectors
);
408 io_sectors
-= io_sectors
% sectors_per_chunk
;
409 if (io_sectors
< sectors_per_chunk
) {
410 io_sectors
= sectors_per_chunk
;
411 } else if (ret
>= 0 && !(ret
& BDRV_BLOCK_DATA
)) {
412 int64_t target_sector_num
;
413 int target_nb_sectors
;
414 bdrv_round_sectors_to_clusters(blk_bs(s
->target
), sector_num
,
415 io_sectors
, &target_sector_num
,
417 if (target_sector_num
== sector_num
&&
418 target_nb_sectors
== io_sectors
) {
419 mirror_method
= ret
& BDRV_BLOCK_ZERO
?
421 MIRROR_METHOD_DISCARD
;
425 while (s
->in_flight
>= MAX_IN_FLIGHT
) {
426 trace_mirror_yield_in_flight(s
, sector_num
, s
->in_flight
);
427 mirror_wait_for_io(s
);
434 mirror_clip_sectors(s
, sector_num
, &io_sectors
);
435 switch (mirror_method
) {
436 case MIRROR_METHOD_COPY
:
437 io_sectors
= mirror_do_read(s
, sector_num
, io_sectors
);
438 io_sectors_acct
= io_sectors
;
440 case MIRROR_METHOD_ZERO
:
441 case MIRROR_METHOD_DISCARD
:
442 mirror_do_zero_or_discard(s
, sector_num
, io_sectors
,
443 mirror_method
== MIRROR_METHOD_DISCARD
);
444 if (write_zeroes_ok
) {
447 io_sectors_acct
= io_sectors
;
454 sector_num
+= io_sectors
;
455 nb_chunks
-= DIV_ROUND_UP(io_sectors
, sectors_per_chunk
);
456 if (s
->common
.speed
) {
457 delay_ns
= ratelimit_calculate_delay(&s
->limit
, io_sectors_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
= blk_bs(s
->common
.blk
);
501 BlockDriverState
*target_bs
= blk_bs(s
->target
);
503 /* Make sure that the source BDS doesn't go away before we called
504 * block_job_completed(). */
508 replace_aio_context
= bdrv_get_aio_context(s
->to_replace
);
509 aio_context_acquire(replace_aio_context
);
512 if (s
->should_complete
&& data
->ret
== 0) {
513 BlockDriverState
*to_replace
= src
;
515 to_replace
= s
->to_replace
;
518 if (bdrv_get_flags(target_bs
) != bdrv_get_flags(to_replace
)) {
519 bdrv_reopen(target_bs
, bdrv_get_flags(to_replace
), NULL
);
522 /* The mirror job has no requests in flight any more, but we need to
523 * drain potential other users of the BDS before changing the graph. */
524 bdrv_drained_begin(target_bs
);
525 bdrv_replace_in_backing_chain(to_replace
, target_bs
);
526 bdrv_drained_end(target_bs
);
528 /* We just changed the BDS the job BB refers to, so switch the BB back
529 * so the cleanup does the right thing. We don't need any permissions
531 blk_remove_bs(job
->blk
);
532 blk_set_perm(job
->blk
, 0, BLK_PERM_ALL
, &error_abort
);
533 blk_insert_bs(job
->blk
, src
, &error_abort
);
536 bdrv_op_unblock_all(s
->to_replace
, s
->replace_blocker
);
537 error_free(s
->replace_blocker
);
538 bdrv_unref(s
->to_replace
);
540 if (replace_aio_context
) {
541 aio_context_release(replace_aio_context
);
544 blk_unref(s
->target
);
546 block_job_completed(&s
->common
, data
->ret
);
548 bdrv_drained_end(src
);
552 static void mirror_throttle(MirrorBlockJob
*s
)
554 int64_t now
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
);
556 if (now
- s
->last_pause_ns
> SLICE_TIME
) {
557 s
->last_pause_ns
= now
;
558 block_job_sleep_ns(&s
->common
, QEMU_CLOCK_REALTIME
, 0);
560 block_job_pause_point(&s
->common
);
564 static int coroutine_fn
mirror_dirty_init(MirrorBlockJob
*s
)
566 int64_t sector_num
, end
;
567 BlockDriverState
*base
= s
->base
;
568 BlockDriverState
*bs
= blk_bs(s
->common
.blk
);
569 BlockDriverState
*target_bs
= blk_bs(s
->target
);
572 end
= s
->bdev_length
/ BDRV_SECTOR_SIZE
;
574 if (base
== NULL
&& !bdrv_has_zero_init(target_bs
)) {
575 if (!bdrv_can_write_zeroes_with_unmap(target_bs
)) {
576 bdrv_set_dirty_bitmap(s
->dirty_bitmap
, 0, end
);
580 s
->initial_zeroing_ongoing
= true;
581 for (sector_num
= 0; sector_num
< end
; ) {
582 int nb_sectors
= MIN(end
- sector_num
,
583 QEMU_ALIGN_DOWN(INT_MAX
, s
->granularity
) >> BDRV_SECTOR_BITS
);
587 if (block_job_is_cancelled(&s
->common
)) {
588 s
->initial_zeroing_ongoing
= false;
592 if (s
->in_flight
>= MAX_IN_FLIGHT
) {
593 trace_mirror_yield(s
, s
->in_flight
, s
->buf_free_count
, -1);
594 mirror_wait_for_io(s
);
598 mirror_do_zero_or_discard(s
, sector_num
, nb_sectors
, false);
599 sector_num
+= nb_sectors
;
602 mirror_wait_for_all_io(s
);
603 s
->initial_zeroing_ongoing
= false;
606 /* First part, loop on the sectors and initialize the dirty bitmap. */
607 for (sector_num
= 0; sector_num
< end
; ) {
608 /* Just to make sure we are not exceeding int limit. */
609 int nb_sectors
= MIN(INT_MAX
>> BDRV_SECTOR_BITS
,
614 if (block_job_is_cancelled(&s
->common
)) {
618 ret
= bdrv_is_allocated_above(bs
, base
, sector_num
, nb_sectors
, &n
);
625 bdrv_set_dirty_bitmap(s
->dirty_bitmap
, sector_num
, n
);
632 /* Called when going out of the streaming phase to flush the bulk of the
633 * data to the medium, or just before completing.
635 static int mirror_flush(MirrorBlockJob
*s
)
637 int ret
= blk_flush(s
->target
);
639 if (mirror_error_action(s
, false, -ret
) == BLOCK_ERROR_ACTION_REPORT
) {
646 static void coroutine_fn
mirror_run(void *opaque
)
648 MirrorBlockJob
*s
= opaque
;
649 MirrorExitData
*data
;
650 BlockDriverState
*bs
= blk_bs(s
->common
.blk
);
651 BlockDriverState
*target_bs
= blk_bs(s
->target
);
652 bool need_drain
= true;
655 char backing_filename
[2]; /* we only need 2 characters because we are only
656 checking for a NULL string */
658 int target_cluster_size
= BDRV_SECTOR_SIZE
;
660 if (block_job_is_cancelled(&s
->common
)) {
664 s
->bdev_length
= bdrv_getlength(bs
);
665 if (s
->bdev_length
< 0) {
666 ret
= s
->bdev_length
;
670 /* Active commit must resize the base image if its size differs from the
672 if (s
->base
== blk_bs(s
->target
)) {
675 base_length
= blk_getlength(s
->target
);
676 if (base_length
< 0) {
681 if (s
->bdev_length
> base_length
) {
682 ret
= blk_truncate(s
->target
, s
->bdev_length
);
689 if (s
->bdev_length
== 0) {
690 /* Report BLOCK_JOB_READY and wait for complete. */
691 block_job_event_ready(&s
->common
);
693 while (!block_job_is_cancelled(&s
->common
) && !s
->should_complete
) {
694 block_job_yield(&s
->common
);
696 s
->common
.cancelled
= false;
700 length
= DIV_ROUND_UP(s
->bdev_length
, s
->granularity
);
701 s
->in_flight_bitmap
= bitmap_new(length
);
703 /* If we have no backing file yet in the destination, we cannot let
704 * the destination do COW. Instead, we copy sectors around the
705 * dirty data if needed. We need a bitmap to do that.
707 bdrv_get_backing_filename(target_bs
, backing_filename
,
708 sizeof(backing_filename
));
709 if (!bdrv_get_info(target_bs
, &bdi
) && bdi
.cluster_size
) {
710 target_cluster_size
= bdi
.cluster_size
;
712 if (backing_filename
[0] && !target_bs
->backing
713 && s
->granularity
< target_cluster_size
) {
714 s
->buf_size
= MAX(s
->buf_size
, target_cluster_size
);
715 s
->cow_bitmap
= bitmap_new(length
);
717 s
->target_cluster_sectors
= target_cluster_size
>> BDRV_SECTOR_BITS
;
718 s
->max_iov
= MIN(bs
->bl
.max_iov
, target_bs
->bl
.max_iov
);
720 s
->buf
= qemu_try_blockalign(bs
, s
->buf_size
);
721 if (s
->buf
== NULL
) {
728 s
->last_pause_ns
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
);
729 if (!s
->is_none_mode
) {
730 ret
= mirror_dirty_init(s
);
731 if (ret
< 0 || block_job_is_cancelled(&s
->common
)) {
737 s
->dbi
= bdrv_dirty_iter_new(s
->dirty_bitmap
, 0);
739 uint64_t delay_ns
= 0;
741 bool should_complete
;
748 block_job_pause_point(&s
->common
);
750 cnt
= bdrv_get_dirty_count(s
->dirty_bitmap
);
751 /* s->common.offset contains the number of bytes already processed so
752 * far, cnt is the number of dirty sectors remaining and
753 * s->sectors_in_flight is the number of sectors currently being
754 * processed; together those are the current total operation length */
755 s
->common
.len
= s
->common
.offset
+
756 (cnt
+ s
->sectors_in_flight
) * BDRV_SECTOR_SIZE
;
758 /* Note that even when no rate limit is applied we need to yield
759 * periodically with no pending I/O so that bdrv_drain_all() returns.
760 * We do so every SLICE_TIME nanoseconds, or when there is an error,
761 * or when the source is clean, whichever comes first.
763 delta
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
) - s
->last_pause_ns
;
764 if (delta
< SLICE_TIME
&&
765 s
->common
.iostatus
== BLOCK_DEVICE_IO_STATUS_OK
) {
766 if (s
->in_flight
>= MAX_IN_FLIGHT
|| s
->buf_free_count
== 0 ||
767 (cnt
== 0 && s
->in_flight
> 0)) {
768 trace_mirror_yield(s
, s
->in_flight
, s
->buf_free_count
, cnt
);
769 mirror_wait_for_io(s
);
771 } else if (cnt
!= 0) {
772 delay_ns
= mirror_iteration(s
);
776 should_complete
= false;
777 if (s
->in_flight
== 0 && cnt
== 0) {
778 trace_mirror_before_flush(s
);
780 if (mirror_flush(s
) < 0) {
781 /* Go check s->ret. */
784 /* We're out of the streaming phase. From now on, if the job
785 * is cancelled we will actually complete all pending I/O and
786 * report completion. This way, block-job-cancel will leave
787 * the target in a consistent state.
789 block_job_event_ready(&s
->common
);
793 should_complete
= s
->should_complete
||
794 block_job_is_cancelled(&s
->common
);
795 cnt
= bdrv_get_dirty_count(s
->dirty_bitmap
);
798 if (cnt
== 0 && should_complete
) {
799 /* The dirty bitmap is not updated while operations are pending.
800 * If we're about to exit, wait for pending operations before
801 * calling bdrv_get_dirty_count(bs), or we may exit while the
802 * source has dirty data to copy!
804 * Note that I/O can be submitted by the guest while
805 * mirror_populate runs, so pause it now. Before deciding
806 * whether to switch to target check one last time if I/O has
807 * come in the meanwhile, and if not flush the data to disk.
809 trace_mirror_before_drain(s
, cnt
);
811 bdrv_drained_begin(bs
);
812 cnt
= bdrv_get_dirty_count(s
->dirty_bitmap
);
813 if (cnt
> 0 || mirror_flush(s
) < 0) {
814 bdrv_drained_end(bs
);
818 /* The two disks are in sync. Exit and report successful
821 assert(QLIST_EMPTY(&bs
->tracked_requests
));
822 s
->common
.cancelled
= false;
828 trace_mirror_before_sleep(s
, cnt
, s
->synced
, delay_ns
);
830 block_job_sleep_ns(&s
->common
, QEMU_CLOCK_REALTIME
, delay_ns
);
831 if (block_job_is_cancelled(&s
->common
)) {
834 } else if (!should_complete
) {
835 delay_ns
= (s
->in_flight
== 0 && cnt
== 0 ? SLICE_TIME
: 0);
836 block_job_sleep_ns(&s
->common
, QEMU_CLOCK_REALTIME
, delay_ns
);
838 s
->last_pause_ns
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
);
842 if (s
->in_flight
> 0) {
843 /* We get here only if something went wrong. Either the job failed,
844 * or it was cancelled prematurely so that we do not guarantee that
845 * the target is a copy of the source.
847 assert(ret
< 0 || (!s
->synced
&& block_job_is_cancelled(&s
->common
)));
849 mirror_wait_for_all_io(s
);
852 assert(s
->in_flight
== 0);
854 g_free(s
->cow_bitmap
);
855 g_free(s
->in_flight_bitmap
);
856 bdrv_dirty_iter_free(s
->dbi
);
857 bdrv_release_dirty_bitmap(bs
, s
->dirty_bitmap
);
859 data
= g_malloc(sizeof(*data
));
863 bdrv_drained_begin(bs
);
865 block_job_defer_to_main_loop(&s
->common
, mirror_exit
, data
);
868 static void mirror_set_speed(BlockJob
*job
, int64_t speed
, Error
**errp
)
870 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
873 error_setg(errp
, QERR_INVALID_PARAMETER
, "speed");
876 ratelimit_set_speed(&s
->limit
, speed
/ BDRV_SECTOR_SIZE
, SLICE_TIME
);
879 static void mirror_complete(BlockJob
*job
, Error
**errp
)
881 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
882 BlockDriverState
*src
, *target
;
884 src
= blk_bs(job
->blk
);
885 target
= blk_bs(s
->target
);
888 error_setg(errp
, "The active block job '%s' cannot be completed",
893 if (s
->backing_mode
== MIRROR_OPEN_BACKING_CHAIN
) {
896 assert(!target
->backing
);
897 ret
= bdrv_open_backing_file(target
, NULL
, "backing", errp
);
903 /* block all operations on to_replace bs */
905 AioContext
*replace_aio_context
;
907 s
->to_replace
= bdrv_find_node(s
->replaces
);
908 if (!s
->to_replace
) {
909 error_setg(errp
, "Node name '%s' not found", s
->replaces
);
913 replace_aio_context
= bdrv_get_aio_context(s
->to_replace
);
914 aio_context_acquire(replace_aio_context
);
916 error_setg(&s
->replace_blocker
,
917 "block device is in use by block-job-complete");
918 bdrv_op_block_all(s
->to_replace
, s
->replace_blocker
);
919 bdrv_ref(s
->to_replace
);
921 aio_context_release(replace_aio_context
);
924 if (s
->backing_mode
== MIRROR_SOURCE_BACKING_CHAIN
) {
925 BlockDriverState
*backing
= s
->is_none_mode
? src
: s
->base
;
926 if (backing_bs(target
) != backing
) {
927 bdrv_set_backing_hd(target
, backing
);
931 s
->should_complete
= true;
932 block_job_enter(&s
->common
);
935 static void mirror_pause(BlockJob
*job
)
937 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
939 mirror_wait_for_all_io(s
);
942 static void mirror_attached_aio_context(BlockJob
*job
, AioContext
*new_context
)
944 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
946 blk_set_aio_context(s
->target
, new_context
);
949 static void mirror_drain(BlockJob
*job
)
951 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
953 /* Need to keep a reference in case blk_drain triggers execution
954 * of mirror_complete...
957 BlockBackend
*target
= s
->target
;
964 static const BlockJobDriver mirror_job_driver
= {
965 .instance_size
= sizeof(MirrorBlockJob
),
966 .job_type
= BLOCK_JOB_TYPE_MIRROR
,
967 .set_speed
= mirror_set_speed
,
969 .complete
= mirror_complete
,
970 .pause
= mirror_pause
,
971 .attached_aio_context
= mirror_attached_aio_context
,
972 .drain
= mirror_drain
,
975 static const BlockJobDriver commit_active_job_driver
= {
976 .instance_size
= sizeof(MirrorBlockJob
),
977 .job_type
= BLOCK_JOB_TYPE_COMMIT
,
978 .set_speed
= mirror_set_speed
,
980 .complete
= mirror_complete
,
981 .pause
= mirror_pause
,
982 .attached_aio_context
= mirror_attached_aio_context
,
983 .drain
= mirror_drain
,
986 static void mirror_start_job(const char *job_id
, BlockDriverState
*bs
,
987 int creation_flags
, BlockDriverState
*target
,
988 const char *replaces
, int64_t speed
,
989 uint32_t granularity
, int64_t buf_size
,
990 BlockMirrorBackingMode backing_mode
,
991 BlockdevOnError on_source_error
,
992 BlockdevOnError on_target_error
,
994 BlockCompletionFunc
*cb
,
995 void *opaque
, Error
**errp
,
996 const BlockJobDriver
*driver
,
997 bool is_none_mode
, BlockDriverState
*base
,
1003 if (granularity
== 0) {
1004 granularity
= bdrv_get_default_bitmap_granularity(target
);
1007 assert ((granularity
& (granularity
- 1)) == 0);
1010 error_setg(errp
, "Invalid parameter 'buf-size'");
1014 if (buf_size
== 0) {
1015 buf_size
= DEFAULT_MIRROR_BUF_SIZE
;
1018 /* FIXME Use real permissions */
1019 s
= block_job_create(job_id
, driver
, bs
, 0, BLK_PERM_ALL
, speed
,
1020 creation_flags
, cb
, opaque
, errp
);
1025 /* FIXME Use real permissions */
1026 s
->target
= blk_new(0, BLK_PERM_ALL
);
1027 ret
= blk_insert_bs(s
->target
, target
, errp
);
1029 blk_unref(s
->target
);
1030 block_job_unref(&s
->common
);
1034 s
->replaces
= g_strdup(replaces
);
1035 s
->on_source_error
= on_source_error
;
1036 s
->on_target_error
= on_target_error
;
1037 s
->is_none_mode
= is_none_mode
;
1038 s
->backing_mode
= backing_mode
;
1040 s
->granularity
= granularity
;
1041 s
->buf_size
= ROUND_UP(buf_size
, granularity
);
1043 if (auto_complete
) {
1044 s
->should_complete
= true;
1047 s
->dirty_bitmap
= bdrv_create_dirty_bitmap(bs
, granularity
, NULL
, errp
);
1048 if (!s
->dirty_bitmap
) {
1049 g_free(s
->replaces
);
1050 blk_unref(s
->target
);
1051 block_job_unref(&s
->common
);
1055 /* FIXME Use real permissions */
1056 block_job_add_bdrv(&s
->common
, "target", target
, 0, BLK_PERM_ALL
,
1059 /* In commit_active_start() all intermediate nodes disappear, so
1060 * any jobs in them must be blocked */
1061 if (bdrv_chain_contains(bs
, target
)) {
1062 BlockDriverState
*iter
;
1063 for (iter
= backing_bs(bs
); iter
!= target
; iter
= backing_bs(iter
)) {
1064 /* FIXME Use real permissions */
1065 block_job_add_bdrv(&s
->common
, "intermediate node", iter
, 0,
1066 BLK_PERM_ALL
, &error_abort
);
1070 trace_mirror_start(bs
, s
, opaque
);
1071 block_job_start(&s
->common
);
1074 void mirror_start(const char *job_id
, BlockDriverState
*bs
,
1075 BlockDriverState
*target
, const char *replaces
,
1076 int64_t speed
, uint32_t granularity
, int64_t buf_size
,
1077 MirrorSyncMode mode
, BlockMirrorBackingMode backing_mode
,
1078 BlockdevOnError on_source_error
,
1079 BlockdevOnError on_target_error
,
1080 bool unmap
, Error
**errp
)
1083 BlockDriverState
*base
;
1085 if (mode
== MIRROR_SYNC_MODE_INCREMENTAL
) {
1086 error_setg(errp
, "Sync mode 'incremental' not supported");
1089 is_none_mode
= mode
== MIRROR_SYNC_MODE_NONE
;
1090 base
= mode
== MIRROR_SYNC_MODE_TOP
? backing_bs(bs
) : NULL
;
1091 mirror_start_job(job_id
, bs
, BLOCK_JOB_DEFAULT
, target
, replaces
,
1092 speed
, granularity
, buf_size
, backing_mode
,
1093 on_source_error
, on_target_error
, unmap
, NULL
, NULL
, errp
,
1094 &mirror_job_driver
, is_none_mode
, base
, false);
1097 void commit_active_start(const char *job_id
, BlockDriverState
*bs
,
1098 BlockDriverState
*base
, int creation_flags
,
1099 int64_t speed
, BlockdevOnError on_error
,
1100 BlockCompletionFunc
*cb
, void *opaque
, Error
**errp
,
1103 int orig_base_flags
;
1104 Error
*local_err
= NULL
;
1106 orig_base_flags
= bdrv_get_flags(base
);
1108 if (bdrv_reopen(base
, bs
->open_flags
, errp
)) {
1112 mirror_start_job(job_id
, bs
, creation_flags
, base
, NULL
, speed
, 0, 0,
1113 MIRROR_LEAVE_BACKING_CHAIN
,
1114 on_error
, on_error
, true, cb
, opaque
, &local_err
,
1115 &commit_active_job_driver
, false, base
, auto_complete
);
1117 error_propagate(errp
, local_err
);
1118 goto error_restore_flags
;
1123 error_restore_flags
:
1124 /* ignore error and errp for bdrv_reopen, because we want to propagate
1125 * the original error */
1126 bdrv_reopen(base
, orig_base_flags
, NULL
);