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
>> BDRV_SECTOR_BITS
,
145 op
->bytes
>> BDRV_SECTOR_BITS
);
146 action
= mirror_error_action(s
, false, -ret
);
147 if (action
== BLOCK_ERROR_ACTION_REPORT
&& s
->ret
>= 0) {
151 mirror_iteration_done(op
, ret
);
152 aio_context_release(blk_get_aio_context(s
->common
.blk
));
155 static void mirror_read_complete(void *opaque
, int ret
)
157 MirrorOp
*op
= opaque
;
158 MirrorBlockJob
*s
= op
->s
;
160 aio_context_acquire(blk_get_aio_context(s
->common
.blk
));
162 BlockErrorAction action
;
164 bdrv_set_dirty_bitmap(s
->dirty_bitmap
, op
->offset
>> BDRV_SECTOR_BITS
,
165 op
->bytes
>> BDRV_SECTOR_BITS
);
166 action
= mirror_error_action(s
, true, -ret
);
167 if (action
== BLOCK_ERROR_ACTION_REPORT
&& s
->ret
>= 0) {
171 mirror_iteration_done(op
, ret
);
173 blk_aio_pwritev(s
->target
, op
->offset
, &op
->qiov
,
174 0, mirror_write_complete
, op
);
176 aio_context_release(blk_get_aio_context(s
->common
.blk
));
179 /* Clip bytes relative to offset to not exceed end-of-file */
180 static inline int64_t mirror_clip_bytes(MirrorBlockJob
*s
,
184 return MIN(bytes
, s
->bdev_length
- offset
);
187 /* Clip nb_sectors relative to sector_num to not exceed end-of-file */
188 static inline int mirror_clip_sectors(MirrorBlockJob
*s
,
192 return MIN(nb_sectors
,
193 s
->bdev_length
/ BDRV_SECTOR_SIZE
- sector_num
);
196 /* Round offset and/or bytes to target cluster if COW is needed, and
197 * return the offset of the adjusted tail against original. */
198 static int mirror_cow_align(MirrorBlockJob
*s
, int64_t *offset
,
203 int64_t align_offset
= *offset
;
204 unsigned int align_bytes
= *bytes
;
205 int max_bytes
= s
->granularity
* s
->max_iov
;
207 assert(*bytes
< INT_MAX
);
208 need_cow
= !test_bit(*offset
/ s
->granularity
, s
->cow_bitmap
);
209 need_cow
|= !test_bit((*offset
+ *bytes
- 1) / s
->granularity
,
212 bdrv_round_to_clusters(blk_bs(s
->target
), *offset
, *bytes
,
213 &align_offset
, &align_bytes
);
216 if (align_bytes
> max_bytes
) {
217 align_bytes
= max_bytes
;
219 align_bytes
= QEMU_ALIGN_DOWN(align_bytes
, s
->target_cluster_size
);
222 /* Clipping may result in align_bytes unaligned to chunk boundary, but
223 * that doesn't matter because it's already the end of source image. */
224 align_bytes
= mirror_clip_bytes(s
, align_offset
, align_bytes
);
226 ret
= align_offset
+ align_bytes
- (*offset
+ *bytes
);
227 *offset
= align_offset
;
228 *bytes
= align_bytes
;
233 static inline void mirror_wait_for_io(MirrorBlockJob
*s
)
235 assert(!s
->waiting_for_io
);
236 s
->waiting_for_io
= true;
237 qemu_coroutine_yield();
238 s
->waiting_for_io
= false;
241 /* Submit async read while handling COW.
242 * Returns: The number of bytes copied after and including offset,
243 * excluding any bytes copied prior to offset due to alignment.
244 * This will be @bytes if no alignment is necessary, or
245 * (new_end - offset) if tail is rounded up or down due to
246 * alignment or buffer limit.
248 static uint64_t mirror_do_read(MirrorBlockJob
*s
, int64_t offset
,
251 BlockBackend
*source
= s
->common
.blk
;
257 max_bytes
= s
->granularity
* s
->max_iov
;
259 /* We can only handle as much as buf_size at a time. */
260 bytes
= MIN(s
->buf_size
, MIN(max_bytes
, bytes
));
262 assert(bytes
< BDRV_REQUEST_MAX_BYTES
);
266 ret
+= mirror_cow_align(s
, &offset
, &bytes
);
268 assert(bytes
<= s
->buf_size
);
269 /* The offset is granularity-aligned because:
270 * 1) Caller passes in aligned values;
271 * 2) mirror_cow_align is used only when target cluster is larger. */
272 assert(QEMU_IS_ALIGNED(offset
, s
->granularity
));
273 /* The range is sector-aligned, since bdrv_getlength() rounds up. */
274 assert(QEMU_IS_ALIGNED(bytes
, BDRV_SECTOR_SIZE
));
275 nb_chunks
= DIV_ROUND_UP(bytes
, s
->granularity
);
277 while (s
->buf_free_count
< nb_chunks
) {
278 trace_mirror_yield_in_flight(s
, offset
, s
->in_flight
);
279 mirror_wait_for_io(s
);
282 /* Allocate a MirrorOp that is used as an AIO callback. */
283 op
= g_new(MirrorOp
, 1);
288 /* Now make a QEMUIOVector taking enough granularity-sized chunks
291 qemu_iovec_init(&op
->qiov
, nb_chunks
);
292 while (nb_chunks
-- > 0) {
293 MirrorBuffer
*buf
= QSIMPLEQ_FIRST(&s
->buf_free
);
294 size_t remaining
= bytes
- op
->qiov
.size
;
296 QSIMPLEQ_REMOVE_HEAD(&s
->buf_free
, next
);
298 qemu_iovec_add(&op
->qiov
, buf
, MIN(s
->granularity
, remaining
));
301 /* Copy the dirty cluster. */
303 s
->bytes_in_flight
+= bytes
;
304 trace_mirror_one_iteration(s
, offset
, bytes
);
306 blk_aio_preadv(source
, offset
, &op
->qiov
, 0, mirror_read_complete
, op
);
310 static void mirror_do_zero_or_discard(MirrorBlockJob
*s
,
317 /* Allocate a MirrorOp that is used as an AIO callback. The qiov is zeroed
318 * so the freeing in mirror_iteration_done is nop. */
319 op
= g_new0(MirrorOp
, 1);
325 s
->bytes_in_flight
+= bytes
;
327 blk_aio_pdiscard(s
->target
, offset
,
328 op
->bytes
, mirror_write_complete
, op
);
330 blk_aio_pwrite_zeroes(s
->target
, offset
,
331 op
->bytes
, s
->unmap
? BDRV_REQ_MAY_UNMAP
: 0,
332 mirror_write_complete
, op
);
336 static uint64_t coroutine_fn
mirror_iteration(MirrorBlockJob
*s
)
338 BlockDriverState
*source
= s
->source
;
339 int64_t sector_num
, first_chunk
;
340 uint64_t delay_ns
= 0;
341 /* At least the first dirty chunk is mirrored in one iteration. */
343 int64_t end
= s
->bdev_length
/ BDRV_SECTOR_SIZE
;
344 int sectors_per_chunk
= s
->granularity
>> BDRV_SECTOR_BITS
;
345 bool write_zeroes_ok
= bdrv_can_write_zeroes_with_unmap(blk_bs(s
->target
));
346 int max_io_bytes
= MAX(s
->buf_size
/ MAX_IN_FLIGHT
, MAX_IO_BYTES
);
348 bdrv_dirty_bitmap_lock(s
->dirty_bitmap
);
349 sector_num
= bdrv_dirty_iter_next(s
->dbi
);
350 if (sector_num
< 0) {
351 bdrv_set_dirty_iter(s
->dbi
, 0);
352 sector_num
= bdrv_dirty_iter_next(s
->dbi
);
353 trace_mirror_restart_iter(s
, bdrv_get_dirty_count(s
->dirty_bitmap
) *
355 assert(sector_num
>= 0);
357 bdrv_dirty_bitmap_unlock(s
->dirty_bitmap
);
359 first_chunk
= sector_num
/ sectors_per_chunk
;
360 while (test_bit(first_chunk
, s
->in_flight_bitmap
)) {
361 trace_mirror_yield_in_flight(s
, sector_num
* BDRV_SECTOR_SIZE
,
363 mirror_wait_for_io(s
);
366 block_job_pause_point(&s
->common
);
368 /* Find the number of consective dirty chunks following the first dirty
369 * one, and wait for in flight requests in them. */
370 bdrv_dirty_bitmap_lock(s
->dirty_bitmap
);
371 while (nb_chunks
* sectors_per_chunk
< (s
->buf_size
>> BDRV_SECTOR_BITS
)) {
373 int64_t next_sector
= sector_num
+ nb_chunks
* sectors_per_chunk
;
374 int64_t next_chunk
= next_sector
/ sectors_per_chunk
;
375 if (next_sector
>= end
||
376 !bdrv_get_dirty_locked(source
, s
->dirty_bitmap
, next_sector
)) {
379 if (test_bit(next_chunk
, s
->in_flight_bitmap
)) {
383 next_dirty
= bdrv_dirty_iter_next(s
->dbi
);
384 if (next_dirty
> next_sector
|| next_dirty
< 0) {
385 /* The bitmap iterator's cache is stale, refresh it */
386 bdrv_set_dirty_iter(s
->dbi
, next_sector
);
387 next_dirty
= bdrv_dirty_iter_next(s
->dbi
);
389 assert(next_dirty
== next_sector
);
393 /* Clear dirty bits before querying the block status, because
394 * calling bdrv_get_block_status_above could yield - if some blocks are
395 * marked dirty in this window, we need to know.
397 bdrv_reset_dirty_bitmap_locked(s
->dirty_bitmap
, sector_num
,
398 nb_chunks
* sectors_per_chunk
);
399 bdrv_dirty_bitmap_unlock(s
->dirty_bitmap
);
401 bitmap_set(s
->in_flight_bitmap
, sector_num
/ sectors_per_chunk
, nb_chunks
);
402 while (nb_chunks
> 0 && sector_num
< end
) {
405 int64_t io_bytes_acct
;
406 BlockDriverState
*file
;
410 MIRROR_METHOD_DISCARD
411 } mirror_method
= MIRROR_METHOD_COPY
;
413 assert(!(sector_num
% sectors_per_chunk
));
414 ret
= bdrv_get_block_status_above(source
, NULL
, sector_num
,
415 nb_chunks
* sectors_per_chunk
,
418 io_sectors
= MIN(nb_chunks
* sectors_per_chunk
,
419 max_io_bytes
>> BDRV_SECTOR_BITS
);
420 } else if (ret
& BDRV_BLOCK_DATA
) {
421 io_sectors
= MIN(io_sectors
, max_io_bytes
>> BDRV_SECTOR_BITS
);
424 io_sectors
-= io_sectors
% sectors_per_chunk
;
425 if (io_sectors
< sectors_per_chunk
) {
426 io_sectors
= sectors_per_chunk
;
427 } else if (ret
>= 0 && !(ret
& BDRV_BLOCK_DATA
)) {
428 int64_t target_sector_num
;
429 int target_nb_sectors
;
430 bdrv_round_sectors_to_clusters(blk_bs(s
->target
), sector_num
,
431 io_sectors
, &target_sector_num
,
433 if (target_sector_num
== sector_num
&&
434 target_nb_sectors
== io_sectors
) {
435 mirror_method
= ret
& BDRV_BLOCK_ZERO
?
437 MIRROR_METHOD_DISCARD
;
441 while (s
->in_flight
>= MAX_IN_FLIGHT
) {
442 trace_mirror_yield_in_flight(s
, sector_num
* BDRV_SECTOR_SIZE
,
444 mirror_wait_for_io(s
);
451 io_sectors
= mirror_clip_sectors(s
, sector_num
, io_sectors
);
452 switch (mirror_method
) {
453 case MIRROR_METHOD_COPY
:
454 io_bytes_acct
= mirror_do_read(s
, sector_num
* BDRV_SECTOR_SIZE
,
455 io_sectors
* BDRV_SECTOR_SIZE
);
456 io_sectors
= io_bytes_acct
/ BDRV_SECTOR_SIZE
;
458 case MIRROR_METHOD_ZERO
:
459 case MIRROR_METHOD_DISCARD
:
460 mirror_do_zero_or_discard(s
, sector_num
* BDRV_SECTOR_SIZE
,
461 io_sectors
* BDRV_SECTOR_SIZE
,
462 mirror_method
== MIRROR_METHOD_DISCARD
);
463 if (write_zeroes_ok
) {
466 io_bytes_acct
= io_sectors
* BDRV_SECTOR_SIZE
;
473 sector_num
+= io_sectors
;
474 nb_chunks
-= DIV_ROUND_UP(io_sectors
, sectors_per_chunk
);
475 if (s
->common
.speed
) {
476 delay_ns
= ratelimit_calculate_delay(&s
->limit
, io_bytes_acct
);
482 static void mirror_free_init(MirrorBlockJob
*s
)
484 int granularity
= s
->granularity
;
485 size_t buf_size
= s
->buf_size
;
486 uint8_t *buf
= s
->buf
;
488 assert(s
->buf_free_count
== 0);
489 QSIMPLEQ_INIT(&s
->buf_free
);
490 while (buf_size
!= 0) {
491 MirrorBuffer
*cur
= (MirrorBuffer
*)buf
;
492 QSIMPLEQ_INSERT_TAIL(&s
->buf_free
, cur
, next
);
494 buf_size
-= granularity
;
499 /* This is also used for the .pause callback. There is no matching
500 * mirror_resume() because mirror_run() will begin iterating again
501 * when the job is resumed.
503 static void mirror_wait_for_all_io(MirrorBlockJob
*s
)
505 while (s
->in_flight
> 0) {
506 mirror_wait_for_io(s
);
514 static void mirror_exit(BlockJob
*job
, void *opaque
)
516 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
517 MirrorExitData
*data
= opaque
;
518 AioContext
*replace_aio_context
= NULL
;
519 BlockDriverState
*src
= s
->source
;
520 BlockDriverState
*target_bs
= blk_bs(s
->target
);
521 BlockDriverState
*mirror_top_bs
= s
->mirror_top_bs
;
522 Error
*local_err
= NULL
;
524 bdrv_release_dirty_bitmap(src
, s
->dirty_bitmap
);
526 /* Make sure that the source BDS doesn't go away before we called
527 * block_job_completed(). */
529 bdrv_ref(mirror_top_bs
);
532 /* Remove target parent that still uses BLK_PERM_WRITE/RESIZE before
533 * inserting target_bs at s->to_replace, where we might not be able to get
536 * Note that blk_unref() alone doesn't necessarily drop permissions because
537 * we might be running nested inside mirror_drain(), which takes an extra
538 * reference, so use an explicit blk_set_perm() first. */
539 blk_set_perm(s
->target
, 0, BLK_PERM_ALL
, &error_abort
);
540 blk_unref(s
->target
);
543 /* We don't access the source any more. Dropping any WRITE/RESIZE is
544 * required before it could become a backing file of target_bs. */
545 bdrv_child_try_set_perm(mirror_top_bs
->backing
, 0, BLK_PERM_ALL
,
547 if (s
->backing_mode
== MIRROR_SOURCE_BACKING_CHAIN
) {
548 BlockDriverState
*backing
= s
->is_none_mode
? src
: s
->base
;
549 if (backing_bs(target_bs
) != backing
) {
550 bdrv_set_backing_hd(target_bs
, backing
, &local_err
);
552 error_report_err(local_err
);
559 replace_aio_context
= bdrv_get_aio_context(s
->to_replace
);
560 aio_context_acquire(replace_aio_context
);
563 if (s
->should_complete
&& data
->ret
== 0) {
564 BlockDriverState
*to_replace
= src
;
566 to_replace
= s
->to_replace
;
569 if (bdrv_get_flags(target_bs
) != bdrv_get_flags(to_replace
)) {
570 bdrv_reopen(target_bs
, bdrv_get_flags(to_replace
), NULL
);
573 /* The mirror job has no requests in flight any more, but we need to
574 * drain potential other users of the BDS before changing the graph. */
575 bdrv_drained_begin(target_bs
);
576 bdrv_replace_node(to_replace
, target_bs
, &local_err
);
577 bdrv_drained_end(target_bs
);
579 error_report_err(local_err
);
584 bdrv_op_unblock_all(s
->to_replace
, s
->replace_blocker
);
585 error_free(s
->replace_blocker
);
586 bdrv_unref(s
->to_replace
);
588 if (replace_aio_context
) {
589 aio_context_release(replace_aio_context
);
592 bdrv_unref(target_bs
);
594 /* Remove the mirror filter driver from the graph. Before this, get rid of
595 * the blockers on the intermediate nodes so that the resulting state is
596 * valid. Also give up permissions on mirror_top_bs->backing, which might
597 * block the removal. */
598 block_job_remove_all_bdrv(job
);
599 bdrv_child_try_set_perm(mirror_top_bs
->backing
, 0, BLK_PERM_ALL
,
601 bdrv_replace_node(mirror_top_bs
, backing_bs(mirror_top_bs
), &error_abort
);
603 /* We just changed the BDS the job BB refers to (with either or both of the
604 * bdrv_replace_node() calls), so switch the BB back so the cleanup does
605 * the right thing. We don't need any permissions any more now. */
606 blk_remove_bs(job
->blk
);
607 blk_set_perm(job
->blk
, 0, BLK_PERM_ALL
, &error_abort
);
608 blk_insert_bs(job
->blk
, mirror_top_bs
, &error_abort
);
610 block_job_completed(&s
->common
, data
->ret
);
613 bdrv_drained_end(src
);
614 bdrv_unref(mirror_top_bs
);
618 static void mirror_throttle(MirrorBlockJob
*s
)
620 int64_t now
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
);
622 if (now
- s
->last_pause_ns
> SLICE_TIME
) {
623 s
->last_pause_ns
= now
;
624 block_job_sleep_ns(&s
->common
, QEMU_CLOCK_REALTIME
, 0);
626 block_job_pause_point(&s
->common
);
630 static int coroutine_fn
mirror_dirty_init(MirrorBlockJob
*s
)
632 int64_t sector_num
, end
;
633 BlockDriverState
*base
= s
->base
;
634 BlockDriverState
*bs
= s
->source
;
635 BlockDriverState
*target_bs
= blk_bs(s
->target
);
638 end
= s
->bdev_length
/ BDRV_SECTOR_SIZE
;
640 if (base
== NULL
&& !bdrv_has_zero_init(target_bs
)) {
641 if (!bdrv_can_write_zeroes_with_unmap(target_bs
)) {
642 bdrv_set_dirty_bitmap(s
->dirty_bitmap
, 0, end
);
646 s
->initial_zeroing_ongoing
= true;
647 for (sector_num
= 0; sector_num
< end
; ) {
648 int nb_sectors
= MIN(end
- sector_num
,
649 QEMU_ALIGN_DOWN(INT_MAX
, s
->granularity
) >> BDRV_SECTOR_BITS
);
653 if (block_job_is_cancelled(&s
->common
)) {
654 s
->initial_zeroing_ongoing
= false;
658 if (s
->in_flight
>= MAX_IN_FLIGHT
) {
659 trace_mirror_yield(s
, UINT64_MAX
, s
->buf_free_count
,
661 mirror_wait_for_io(s
);
665 mirror_do_zero_or_discard(s
, sector_num
* BDRV_SECTOR_SIZE
,
666 nb_sectors
* BDRV_SECTOR_SIZE
, false);
667 sector_num
+= nb_sectors
;
670 mirror_wait_for_all_io(s
);
671 s
->initial_zeroing_ongoing
= false;
674 /* First part, loop on the sectors and initialize the dirty bitmap. */
675 for (sector_num
= 0; sector_num
< end
; ) {
676 /* Just to make sure we are not exceeding int limit. */
677 int nb_sectors
= MIN(INT_MAX
>> BDRV_SECTOR_BITS
,
682 if (block_job_is_cancelled(&s
->common
)) {
686 ret
= bdrv_is_allocated_above(bs
, base
, sector_num
, nb_sectors
, &n
);
693 bdrv_set_dirty_bitmap(s
->dirty_bitmap
, sector_num
, n
);
700 /* Called when going out of the streaming phase to flush the bulk of the
701 * data to the medium, or just before completing.
703 static int mirror_flush(MirrorBlockJob
*s
)
705 int ret
= blk_flush(s
->target
);
707 if (mirror_error_action(s
, false, -ret
) == BLOCK_ERROR_ACTION_REPORT
) {
714 static void coroutine_fn
mirror_run(void *opaque
)
716 MirrorBlockJob
*s
= opaque
;
717 MirrorExitData
*data
;
718 BlockDriverState
*bs
= s
->source
;
719 BlockDriverState
*target_bs
= blk_bs(s
->target
);
720 bool need_drain
= true;
723 char backing_filename
[2]; /* we only need 2 characters because we are only
724 checking for a NULL string */
727 if (block_job_is_cancelled(&s
->common
)) {
731 s
->bdev_length
= bdrv_getlength(bs
);
732 if (s
->bdev_length
< 0) {
733 ret
= s
->bdev_length
;
737 /* Active commit must resize the base image if its size differs from the
739 if (s
->base
== blk_bs(s
->target
)) {
742 base_length
= blk_getlength(s
->target
);
743 if (base_length
< 0) {
748 if (s
->bdev_length
> base_length
) {
749 ret
= blk_truncate(s
->target
, s
->bdev_length
, NULL
);
756 if (s
->bdev_length
== 0) {
757 /* Report BLOCK_JOB_READY and wait for complete. */
758 block_job_event_ready(&s
->common
);
760 while (!block_job_is_cancelled(&s
->common
) && !s
->should_complete
) {
761 block_job_yield(&s
->common
);
763 s
->common
.cancelled
= false;
767 length
= DIV_ROUND_UP(s
->bdev_length
, s
->granularity
);
768 s
->in_flight_bitmap
= bitmap_new(length
);
770 /* If we have no backing file yet in the destination, we cannot let
771 * the destination do COW. Instead, we copy sectors around the
772 * dirty data if needed. We need a bitmap to do that.
774 bdrv_get_backing_filename(target_bs
, backing_filename
,
775 sizeof(backing_filename
));
776 if (!bdrv_get_info(target_bs
, &bdi
) && bdi
.cluster_size
) {
777 s
->target_cluster_size
= bdi
.cluster_size
;
779 s
->target_cluster_size
= BDRV_SECTOR_SIZE
;
781 if (backing_filename
[0] && !target_bs
->backing
&&
782 s
->granularity
< s
->target_cluster_size
) {
783 s
->buf_size
= MAX(s
->buf_size
, s
->target_cluster_size
);
784 s
->cow_bitmap
= bitmap_new(length
);
786 s
->max_iov
= MIN(bs
->bl
.max_iov
, target_bs
->bl
.max_iov
);
788 s
->buf
= qemu_try_blockalign(bs
, s
->buf_size
);
789 if (s
->buf
== NULL
) {
796 s
->last_pause_ns
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
);
797 if (!s
->is_none_mode
) {
798 ret
= mirror_dirty_init(s
);
799 if (ret
< 0 || block_job_is_cancelled(&s
->common
)) {
805 s
->dbi
= bdrv_dirty_iter_new(s
->dirty_bitmap
, 0);
807 uint64_t delay_ns
= 0;
809 bool should_complete
;
816 block_job_pause_point(&s
->common
);
818 cnt
= bdrv_get_dirty_count(s
->dirty_bitmap
);
819 /* s->common.offset contains the number of bytes already processed so
820 * far, cnt is the number of dirty sectors remaining and
821 * s->bytes_in_flight is the number of bytes currently being
822 * processed; together those are the current total operation length */
823 s
->common
.len
= s
->common
.offset
+ s
->bytes_in_flight
+
824 cnt
* BDRV_SECTOR_SIZE
;
826 /* Note that even when no rate limit is applied we need to yield
827 * periodically with no pending I/O so that bdrv_drain_all() returns.
828 * We do so every SLICE_TIME nanoseconds, or when there is an error,
829 * or when the source is clean, whichever comes first.
831 delta
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
) - s
->last_pause_ns
;
832 if (delta
< SLICE_TIME
&&
833 s
->common
.iostatus
== BLOCK_DEVICE_IO_STATUS_OK
) {
834 if (s
->in_flight
>= MAX_IN_FLIGHT
|| s
->buf_free_count
== 0 ||
835 (cnt
== 0 && s
->in_flight
> 0)) {
836 trace_mirror_yield(s
, cnt
* BDRV_SECTOR_SIZE
,
837 s
->buf_free_count
, s
->in_flight
);
838 mirror_wait_for_io(s
);
840 } else if (cnt
!= 0) {
841 delay_ns
= mirror_iteration(s
);
845 should_complete
= false;
846 if (s
->in_flight
== 0 && cnt
== 0) {
847 trace_mirror_before_flush(s
);
849 if (mirror_flush(s
) < 0) {
850 /* Go check s->ret. */
853 /* We're out of the streaming phase. From now on, if the job
854 * is cancelled we will actually complete all pending I/O and
855 * report completion. This way, block-job-cancel will leave
856 * the target in a consistent state.
858 block_job_event_ready(&s
->common
);
862 should_complete
= s
->should_complete
||
863 block_job_is_cancelled(&s
->common
);
864 cnt
= bdrv_get_dirty_count(s
->dirty_bitmap
);
867 if (cnt
== 0 && should_complete
) {
868 /* The dirty bitmap is not updated while operations are pending.
869 * If we're about to exit, wait for pending operations before
870 * calling bdrv_get_dirty_count(bs), or we may exit while the
871 * source has dirty data to copy!
873 * Note that I/O can be submitted by the guest while
874 * mirror_populate runs, so pause it now. Before deciding
875 * whether to switch to target check one last time if I/O has
876 * come in the meanwhile, and if not flush the data to disk.
878 trace_mirror_before_drain(s
, cnt
* BDRV_SECTOR_SIZE
);
880 bdrv_drained_begin(bs
);
881 cnt
= bdrv_get_dirty_count(s
->dirty_bitmap
);
882 if (cnt
> 0 || mirror_flush(s
) < 0) {
883 bdrv_drained_end(bs
);
887 /* The two disks are in sync. Exit and report successful
890 assert(QLIST_EMPTY(&bs
->tracked_requests
));
891 s
->common
.cancelled
= false;
897 trace_mirror_before_sleep(s
, cnt
* BDRV_SECTOR_SIZE
,
898 s
->synced
, delay_ns
);
900 block_job_sleep_ns(&s
->common
, QEMU_CLOCK_REALTIME
, delay_ns
);
901 if (block_job_is_cancelled(&s
->common
)) {
904 } else if (!should_complete
) {
905 delay_ns
= (s
->in_flight
== 0 && cnt
== 0 ? SLICE_TIME
: 0);
906 block_job_sleep_ns(&s
->common
, QEMU_CLOCK_REALTIME
, delay_ns
);
908 s
->last_pause_ns
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
);
912 if (s
->in_flight
> 0) {
913 /* We get here only if something went wrong. Either the job failed,
914 * or it was cancelled prematurely so that we do not guarantee that
915 * the target is a copy of the source.
917 assert(ret
< 0 || (!s
->synced
&& block_job_is_cancelled(&s
->common
)));
919 mirror_wait_for_all_io(s
);
922 assert(s
->in_flight
== 0);
924 g_free(s
->cow_bitmap
);
925 g_free(s
->in_flight_bitmap
);
926 bdrv_dirty_iter_free(s
->dbi
);
928 data
= g_malloc(sizeof(*data
));
932 bdrv_drained_begin(bs
);
934 block_job_defer_to_main_loop(&s
->common
, mirror_exit
, data
);
937 static void mirror_set_speed(BlockJob
*job
, int64_t speed
, Error
**errp
)
939 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
942 error_setg(errp
, QERR_INVALID_PARAMETER
, "speed");
945 ratelimit_set_speed(&s
->limit
, speed
, SLICE_TIME
);
948 static void mirror_complete(BlockJob
*job
, Error
**errp
)
950 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
951 BlockDriverState
*target
;
953 target
= blk_bs(s
->target
);
956 error_setg(errp
, "The active block job '%s' cannot be completed",
961 if (s
->backing_mode
== MIRROR_OPEN_BACKING_CHAIN
) {
964 assert(!target
->backing
);
965 ret
= bdrv_open_backing_file(target
, NULL
, "backing", errp
);
971 /* block all operations on to_replace bs */
973 AioContext
*replace_aio_context
;
975 s
->to_replace
= bdrv_find_node(s
->replaces
);
976 if (!s
->to_replace
) {
977 error_setg(errp
, "Node name '%s' not found", s
->replaces
);
981 replace_aio_context
= bdrv_get_aio_context(s
->to_replace
);
982 aio_context_acquire(replace_aio_context
);
984 /* TODO Translate this into permission system. Current definition of
985 * GRAPH_MOD would require to request it for the parents; they might
986 * not even be BlockDriverStates, however, so a BdrvChild can't address
987 * them. May need redefinition of GRAPH_MOD. */
988 error_setg(&s
->replace_blocker
,
989 "block device is in use by block-job-complete");
990 bdrv_op_block_all(s
->to_replace
, s
->replace_blocker
);
991 bdrv_ref(s
->to_replace
);
993 aio_context_release(replace_aio_context
);
996 s
->should_complete
= true;
997 block_job_enter(&s
->common
);
1000 static void mirror_pause(BlockJob
*job
)
1002 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
1004 mirror_wait_for_all_io(s
);
1007 static void mirror_attached_aio_context(BlockJob
*job
, AioContext
*new_context
)
1009 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
1011 blk_set_aio_context(s
->target
, new_context
);
1014 static void mirror_drain(BlockJob
*job
)
1016 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
1018 /* Need to keep a reference in case blk_drain triggers execution
1019 * of mirror_complete...
1022 BlockBackend
*target
= s
->target
;
1029 static const BlockJobDriver mirror_job_driver
= {
1030 .instance_size
= sizeof(MirrorBlockJob
),
1031 .job_type
= BLOCK_JOB_TYPE_MIRROR
,
1032 .set_speed
= mirror_set_speed
,
1033 .start
= mirror_run
,
1034 .complete
= mirror_complete
,
1035 .pause
= mirror_pause
,
1036 .attached_aio_context
= mirror_attached_aio_context
,
1037 .drain
= mirror_drain
,
1040 static const BlockJobDriver commit_active_job_driver
= {
1041 .instance_size
= sizeof(MirrorBlockJob
),
1042 .job_type
= BLOCK_JOB_TYPE_COMMIT
,
1043 .set_speed
= mirror_set_speed
,
1044 .start
= mirror_run
,
1045 .complete
= mirror_complete
,
1046 .pause
= mirror_pause
,
1047 .attached_aio_context
= mirror_attached_aio_context
,
1048 .drain
= mirror_drain
,
1051 static int coroutine_fn
bdrv_mirror_top_preadv(BlockDriverState
*bs
,
1052 uint64_t offset
, uint64_t bytes
, QEMUIOVector
*qiov
, int flags
)
1054 return bdrv_co_preadv(bs
->backing
, offset
, bytes
, qiov
, flags
);
1057 static int coroutine_fn
bdrv_mirror_top_pwritev(BlockDriverState
*bs
,
1058 uint64_t offset
, uint64_t bytes
, QEMUIOVector
*qiov
, int flags
)
1060 return bdrv_co_pwritev(bs
->backing
, offset
, bytes
, qiov
, flags
);
1063 static int coroutine_fn
bdrv_mirror_top_flush(BlockDriverState
*bs
)
1065 return bdrv_co_flush(bs
->backing
->bs
);
1068 static int64_t coroutine_fn
bdrv_mirror_top_get_block_status(
1069 BlockDriverState
*bs
, int64_t sector_num
, int nb_sectors
, int *pnum
,
1070 BlockDriverState
**file
)
1073 *file
= bs
->backing
->bs
;
1074 return BDRV_BLOCK_RAW
| BDRV_BLOCK_OFFSET_VALID
|
1075 (sector_num
<< BDRV_SECTOR_BITS
);
1078 static int coroutine_fn
bdrv_mirror_top_pwrite_zeroes(BlockDriverState
*bs
,
1079 int64_t offset
, int bytes
, BdrvRequestFlags flags
)
1081 return bdrv_co_pwrite_zeroes(bs
->backing
, offset
, bytes
, flags
);
1084 static int coroutine_fn
bdrv_mirror_top_pdiscard(BlockDriverState
*bs
,
1085 int64_t offset
, int bytes
)
1087 return bdrv_co_pdiscard(bs
->backing
->bs
, offset
, bytes
);
1090 static void bdrv_mirror_top_refresh_filename(BlockDriverState
*bs
, QDict
*opts
)
1092 bdrv_refresh_filename(bs
->backing
->bs
);
1093 pstrcpy(bs
->exact_filename
, sizeof(bs
->exact_filename
),
1094 bs
->backing
->bs
->filename
);
1097 static void bdrv_mirror_top_close(BlockDriverState
*bs
)
1101 static void bdrv_mirror_top_child_perm(BlockDriverState
*bs
, BdrvChild
*c
,
1102 const BdrvChildRole
*role
,
1103 uint64_t perm
, uint64_t shared
,
1104 uint64_t *nperm
, uint64_t *nshared
)
1106 /* Must be able to forward guest writes to the real image */
1108 if (perm
& BLK_PERM_WRITE
) {
1109 *nperm
|= BLK_PERM_WRITE
;
1112 *nshared
= BLK_PERM_ALL
;
1115 /* Dummy node that provides consistent read to its users without requiring it
1116 * from its backing file and that allows writes on the backing file chain. */
1117 static BlockDriver bdrv_mirror_top
= {
1118 .format_name
= "mirror_top",
1119 .bdrv_co_preadv
= bdrv_mirror_top_preadv
,
1120 .bdrv_co_pwritev
= bdrv_mirror_top_pwritev
,
1121 .bdrv_co_pwrite_zeroes
= bdrv_mirror_top_pwrite_zeroes
,
1122 .bdrv_co_pdiscard
= bdrv_mirror_top_pdiscard
,
1123 .bdrv_co_flush
= bdrv_mirror_top_flush
,
1124 .bdrv_co_get_block_status
= bdrv_mirror_top_get_block_status
,
1125 .bdrv_refresh_filename
= bdrv_mirror_top_refresh_filename
,
1126 .bdrv_close
= bdrv_mirror_top_close
,
1127 .bdrv_child_perm
= bdrv_mirror_top_child_perm
,
1130 static void mirror_start_job(const char *job_id
, BlockDriverState
*bs
,
1131 int creation_flags
, BlockDriverState
*target
,
1132 const char *replaces
, int64_t speed
,
1133 uint32_t granularity
, int64_t buf_size
,
1134 BlockMirrorBackingMode backing_mode
,
1135 BlockdevOnError on_source_error
,
1136 BlockdevOnError on_target_error
,
1138 BlockCompletionFunc
*cb
,
1140 const BlockJobDriver
*driver
,
1141 bool is_none_mode
, BlockDriverState
*base
,
1142 bool auto_complete
, const char *filter_node_name
,
1146 BlockDriverState
*mirror_top_bs
;
1147 bool target_graph_mod
;
1148 bool target_is_backing
;
1149 Error
*local_err
= NULL
;
1152 if (granularity
== 0) {
1153 granularity
= bdrv_get_default_bitmap_granularity(target
);
1156 assert ((granularity
& (granularity
- 1)) == 0);
1157 /* Granularity must be large enough for sector-based dirty bitmap */
1158 assert(granularity
>= BDRV_SECTOR_SIZE
);
1161 error_setg(errp
, "Invalid parameter 'buf-size'");
1165 if (buf_size
== 0) {
1166 buf_size
= DEFAULT_MIRROR_BUF_SIZE
;
1169 /* In the case of active commit, add dummy driver to provide consistent
1170 * reads on the top, while disabling it in the intermediate nodes, and make
1171 * the backing chain writable. */
1172 mirror_top_bs
= bdrv_new_open_driver(&bdrv_mirror_top
, filter_node_name
,
1174 if (mirror_top_bs
== NULL
) {
1177 mirror_top_bs
->total_sectors
= bs
->total_sectors
;
1178 bdrv_set_aio_context(mirror_top_bs
, bdrv_get_aio_context(bs
));
1180 /* bdrv_append takes ownership of the mirror_top_bs reference, need to keep
1181 * it alive until block_job_create() succeeds even if bs has no parent. */
1182 bdrv_ref(mirror_top_bs
);
1183 bdrv_drained_begin(bs
);
1184 bdrv_append(mirror_top_bs
, bs
, &local_err
);
1185 bdrv_drained_end(bs
);
1188 bdrv_unref(mirror_top_bs
);
1189 error_propagate(errp
, local_err
);
1193 /* Make sure that the source is not resized while the job is running */
1194 s
= block_job_create(job_id
, driver
, mirror_top_bs
,
1195 BLK_PERM_CONSISTENT_READ
,
1196 BLK_PERM_CONSISTENT_READ
| BLK_PERM_WRITE_UNCHANGED
|
1197 BLK_PERM_WRITE
| BLK_PERM_GRAPH_MOD
, speed
,
1198 creation_flags
, cb
, opaque
, errp
);
1202 /* The block job now has a reference to this node */
1203 bdrv_unref(mirror_top_bs
);
1206 s
->mirror_top_bs
= mirror_top_bs
;
1208 /* No resize for the target either; while the mirror is still running, a
1209 * consistent read isn't necessarily possible. We could possibly allow
1210 * writes and graph modifications, though it would likely defeat the
1211 * purpose of a mirror, so leave them blocked for now.
1213 * In the case of active commit, things look a bit different, though,
1214 * because the target is an already populated backing file in active use.
1215 * We can allow anything except resize there.*/
1216 target_is_backing
= bdrv_chain_contains(bs
, target
);
1217 target_graph_mod
= (backing_mode
!= MIRROR_LEAVE_BACKING_CHAIN
);
1218 s
->target
= blk_new(BLK_PERM_WRITE
| BLK_PERM_RESIZE
|
1219 (target_graph_mod
? BLK_PERM_GRAPH_MOD
: 0),
1220 BLK_PERM_WRITE_UNCHANGED
|
1221 (target_is_backing
? BLK_PERM_CONSISTENT_READ
|
1223 BLK_PERM_GRAPH_MOD
: 0));
1224 ret
= blk_insert_bs(s
->target
, target
, errp
);
1229 s
->replaces
= g_strdup(replaces
);
1230 s
->on_source_error
= on_source_error
;
1231 s
->on_target_error
= on_target_error
;
1232 s
->is_none_mode
= is_none_mode
;
1233 s
->backing_mode
= backing_mode
;
1235 s
->granularity
= granularity
;
1236 s
->buf_size
= ROUND_UP(buf_size
, granularity
);
1238 if (auto_complete
) {
1239 s
->should_complete
= true;
1242 s
->dirty_bitmap
= bdrv_create_dirty_bitmap(bs
, granularity
, NULL
, errp
);
1243 if (!s
->dirty_bitmap
) {
1247 /* Required permissions are already taken with blk_new() */
1248 block_job_add_bdrv(&s
->common
, "target", target
, 0, BLK_PERM_ALL
,
1251 /* In commit_active_start() all intermediate nodes disappear, so
1252 * any jobs in them must be blocked */
1253 if (target_is_backing
) {
1254 BlockDriverState
*iter
;
1255 for (iter
= backing_bs(bs
); iter
!= target
; iter
= backing_bs(iter
)) {
1256 /* XXX BLK_PERM_WRITE needs to be allowed so we don't block
1257 * ourselves at s->base (if writes are blocked for a node, they are
1258 * also blocked for its backing file). The other options would be a
1259 * second filter driver above s->base (== target). */
1260 ret
= block_job_add_bdrv(&s
->common
, "intermediate node", iter
, 0,
1261 BLK_PERM_WRITE_UNCHANGED
| BLK_PERM_WRITE
,
1269 trace_mirror_start(bs
, s
, opaque
);
1270 block_job_start(&s
->common
);
1275 /* Make sure this BDS does not go away until we have completed the graph
1277 bdrv_ref(mirror_top_bs
);
1279 g_free(s
->replaces
);
1280 blk_unref(s
->target
);
1281 block_job_early_fail(&s
->common
);
1284 bdrv_child_try_set_perm(mirror_top_bs
->backing
, 0, BLK_PERM_ALL
,
1286 bdrv_replace_node(mirror_top_bs
, backing_bs(mirror_top_bs
), &error_abort
);
1288 bdrv_unref(mirror_top_bs
);
1291 void mirror_start(const char *job_id
, BlockDriverState
*bs
,
1292 BlockDriverState
*target
, const char *replaces
,
1293 int64_t speed
, uint32_t granularity
, int64_t buf_size
,
1294 MirrorSyncMode mode
, BlockMirrorBackingMode backing_mode
,
1295 BlockdevOnError on_source_error
,
1296 BlockdevOnError on_target_error
,
1297 bool unmap
, const char *filter_node_name
, Error
**errp
)
1300 BlockDriverState
*base
;
1302 if (mode
== MIRROR_SYNC_MODE_INCREMENTAL
) {
1303 error_setg(errp
, "Sync mode 'incremental' not supported");
1306 is_none_mode
= mode
== MIRROR_SYNC_MODE_NONE
;
1307 base
= mode
== MIRROR_SYNC_MODE_TOP
? backing_bs(bs
) : NULL
;
1308 mirror_start_job(job_id
, bs
, BLOCK_JOB_DEFAULT
, target
, replaces
,
1309 speed
, granularity
, buf_size
, backing_mode
,
1310 on_source_error
, on_target_error
, unmap
, NULL
, NULL
,
1311 &mirror_job_driver
, is_none_mode
, base
, false,
1312 filter_node_name
, errp
);
1315 void commit_active_start(const char *job_id
, BlockDriverState
*bs
,
1316 BlockDriverState
*base
, int creation_flags
,
1317 int64_t speed
, BlockdevOnError on_error
,
1318 const char *filter_node_name
,
1319 BlockCompletionFunc
*cb
, void *opaque
,
1320 bool auto_complete
, Error
**errp
)
1322 int orig_base_flags
;
1323 Error
*local_err
= NULL
;
1325 orig_base_flags
= bdrv_get_flags(base
);
1327 if (bdrv_reopen(base
, bs
->open_flags
, errp
)) {
1331 mirror_start_job(job_id
, bs
, creation_flags
, base
, NULL
, speed
, 0, 0,
1332 MIRROR_LEAVE_BACKING_CHAIN
,
1333 on_error
, on_error
, true, cb
, opaque
,
1334 &commit_active_job_driver
, false, base
, auto_complete
,
1335 filter_node_name
, &local_err
);
1337 error_propagate(errp
, local_err
);
1338 goto error_restore_flags
;
1343 error_restore_flags
:
1344 /* ignore error and errp for bdrv_reopen, because we want to propagate
1345 * the original error */
1346 bdrv_reopen(base
, orig_base_flags
, NULL
);