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"
16 #include "qemu/coroutine.h"
17 #include "qemu/range.h"
19 #include "block/blockjob_int.h"
20 #include "block/block_int.h"
21 #include "block/dirty-bitmap.h"
22 #include "sysemu/block-backend.h"
23 #include "qapi/error.h"
24 #include "qapi/qmp/qerror.h"
25 #include "qemu/ratelimit.h"
26 #include "qemu/bitmap.h"
27 #include "qemu/memalign.h"
29 #define MAX_IN_FLIGHT 16
30 #define MAX_IO_BYTES (1 << 20) /* 1 Mb */
31 #define DEFAULT_MIRROR_BUF_SIZE (MAX_IN_FLIGHT * MAX_IO_BYTES)
33 /* The mirroring buffer is a list of granularity-sized chunks.
34 * Free chunks are organized in a list.
36 typedef struct MirrorBuffer
{
37 QSIMPLEQ_ENTRY(MirrorBuffer
) next
;
40 typedef struct MirrorOp MirrorOp
;
42 typedef struct MirrorBlockJob
{
45 BlockDriverState
*mirror_top_bs
;
46 BlockDriverState
*base
;
47 BlockDriverState
*base_overlay
;
49 /* The name of the graph node to replace */
51 /* The BDS to replace */
52 BlockDriverState
*to_replace
;
53 /* Used to block operations on the drive-mirror-replace target */
54 Error
*replace_blocker
;
56 BlockMirrorBackingMode backing_mode
;
57 /* Whether the target image requires explicit zero-initialization */
59 MirrorCopyMode copy_mode
;
60 BlockdevOnError on_source_error
, on_target_error
;
61 /* Set when the target is synced (dirty bitmap is clean, nothing
62 * in flight) and the job is running in active mode */
68 unsigned long *cow_bitmap
;
69 BdrvDirtyBitmap
*dirty_bitmap
;
70 BdrvDirtyBitmapIter
*dbi
;
72 QSIMPLEQ_HEAD(, MirrorBuffer
) buf_free
;
75 uint64_t last_pause_ns
;
76 unsigned long *in_flight_bitmap
;
78 int64_t bytes_in_flight
;
79 QTAILQ_HEAD(, MirrorOp
) ops_in_flight
;
82 int target_cluster_size
;
84 bool initial_zeroing_ongoing
;
85 int in_active_write_counter
;
86 int64_t active_write_bytes_in_flight
;
91 typedef struct MirrorBDSOpaque
{
103 /* The pointee is set by mirror_co_read(), mirror_co_zero(), and
104 * mirror_co_discard() before yielding for the first time */
105 int64_t *bytes_handled
;
108 bool is_active_write
;
110 CoQueue waiting_requests
;
112 MirrorOp
*waiting_for_op
;
114 QTAILQ_ENTRY(MirrorOp
) next
;
117 typedef enum MirrorMethod
{
120 MIRROR_METHOD_DISCARD
,
123 static BlockErrorAction
mirror_error_action(MirrorBlockJob
*s
, bool read
,
126 s
->actively_synced
= false;
128 return block_job_error_action(&s
->common
, s
->on_source_error
,
131 return block_job_error_action(&s
->common
, s
->on_target_error
,
136 static void coroutine_fn
mirror_wait_on_conflicts(MirrorOp
*self
,
141 uint64_t self_start_chunk
= offset
/ s
->granularity
;
142 uint64_t self_end_chunk
= DIV_ROUND_UP(offset
+ bytes
, s
->granularity
);
143 uint64_t self_nb_chunks
= self_end_chunk
- self_start_chunk
;
145 while (find_next_bit(s
->in_flight_bitmap
, self_end_chunk
,
146 self_start_chunk
) < self_end_chunk
&&
151 QTAILQ_FOREACH(op
, &s
->ops_in_flight
, next
) {
152 uint64_t op_start_chunk
= op
->offset
/ s
->granularity
;
153 uint64_t op_nb_chunks
= DIV_ROUND_UP(op
->offset
+ op
->bytes
,
161 if (ranges_overlap(self_start_chunk
, self_nb_chunks
,
162 op_start_chunk
, op_nb_chunks
))
166 * If the operation is already (indirectly) waiting for us,
167 * or will wait for us as soon as it wakes up, then just go
168 * on (instead of producing a deadlock in the former case).
170 if (op
->waiting_for_op
) {
174 self
->waiting_for_op
= op
;
177 qemu_co_queue_wait(&op
->waiting_requests
, NULL
);
180 self
->waiting_for_op
= NULL
;
189 static void coroutine_fn
mirror_iteration_done(MirrorOp
*op
, int ret
)
191 MirrorBlockJob
*s
= op
->s
;
196 trace_mirror_iteration_done(s
, op
->offset
, op
->bytes
, ret
);
199 s
->bytes_in_flight
-= op
->bytes
;
201 for (i
= 0; i
< op
->qiov
.niov
; i
++) {
202 MirrorBuffer
*buf
= (MirrorBuffer
*) iov
[i
].iov_base
;
203 QSIMPLEQ_INSERT_TAIL(&s
->buf_free
, buf
, next
);
207 chunk_num
= op
->offset
/ s
->granularity
;
208 nb_chunks
= DIV_ROUND_UP(op
->bytes
, s
->granularity
);
210 bitmap_clear(s
->in_flight_bitmap
, chunk_num
, nb_chunks
);
211 QTAILQ_REMOVE(&s
->ops_in_flight
, op
, next
);
214 bitmap_set(s
->cow_bitmap
, chunk_num
, nb_chunks
);
216 if (!s
->initial_zeroing_ongoing
) {
217 job_progress_update(&s
->common
.job
, op
->bytes
);
220 qemu_iovec_destroy(&op
->qiov
);
222 qemu_co_queue_restart_all(&op
->waiting_requests
);
226 static void coroutine_fn
mirror_write_complete(MirrorOp
*op
, int ret
)
228 MirrorBlockJob
*s
= op
->s
;
231 BlockErrorAction action
;
233 bdrv_set_dirty_bitmap(s
->dirty_bitmap
, op
->offset
, op
->bytes
);
234 action
= mirror_error_action(s
, false, -ret
);
235 if (action
== BLOCK_ERROR_ACTION_REPORT
&& s
->ret
>= 0) {
240 mirror_iteration_done(op
, ret
);
243 static void coroutine_fn
mirror_read_complete(MirrorOp
*op
, int ret
)
245 MirrorBlockJob
*s
= op
->s
;
248 BlockErrorAction action
;
250 bdrv_set_dirty_bitmap(s
->dirty_bitmap
, op
->offset
, op
->bytes
);
251 action
= mirror_error_action(s
, true, -ret
);
252 if (action
== BLOCK_ERROR_ACTION_REPORT
&& s
->ret
>= 0) {
256 mirror_iteration_done(op
, ret
);
260 ret
= blk_co_pwritev(s
->target
, op
->offset
, op
->qiov
.size
, &op
->qiov
, 0);
261 mirror_write_complete(op
, ret
);
264 /* Clip bytes relative to offset to not exceed end-of-file */
265 static inline int64_t mirror_clip_bytes(MirrorBlockJob
*s
,
269 return MIN(bytes
, s
->bdev_length
- offset
);
272 /* Round offset and/or bytes to target cluster if COW is needed, and
273 * return the offset of the adjusted tail against original. */
274 static int mirror_cow_align(MirrorBlockJob
*s
, int64_t *offset
,
279 int64_t align_offset
= *offset
;
280 int64_t align_bytes
= *bytes
;
281 int max_bytes
= s
->granularity
* s
->max_iov
;
283 need_cow
= !test_bit(*offset
/ s
->granularity
, s
->cow_bitmap
);
284 need_cow
|= !test_bit((*offset
+ *bytes
- 1) / s
->granularity
,
287 bdrv_round_to_clusters(blk_bs(s
->target
), *offset
, *bytes
,
288 &align_offset
, &align_bytes
);
291 if (align_bytes
> max_bytes
) {
292 align_bytes
= max_bytes
;
294 align_bytes
= QEMU_ALIGN_DOWN(align_bytes
, s
->target_cluster_size
);
297 /* Clipping may result in align_bytes unaligned to chunk boundary, but
298 * that doesn't matter because it's already the end of source image. */
299 align_bytes
= mirror_clip_bytes(s
, align_offset
, align_bytes
);
301 ret
= align_offset
+ align_bytes
- (*offset
+ *bytes
);
302 *offset
= align_offset
;
303 *bytes
= align_bytes
;
308 static inline void coroutine_fn
309 mirror_wait_for_free_in_flight_slot(MirrorBlockJob
*s
)
313 QTAILQ_FOREACH(op
, &s
->ops_in_flight
, next
) {
315 * Do not wait on pseudo ops, because it may in turn wait on
316 * some other operation to start, which may in fact be the
317 * caller of this function. Since there is only one pseudo op
318 * at any given time, we will always find some real operation
320 * Also, do not wait on active operations, because they do not
321 * use up in-flight slots.
323 if (!op
->is_pseudo_op
&& op
->is_in_flight
&& !op
->is_active_write
) {
324 qemu_co_queue_wait(&op
->waiting_requests
, NULL
);
331 /* Perform a mirror copy operation.
333 * *op->bytes_handled is set to the number of bytes copied after and
334 * including offset, excluding any bytes copied prior to offset due
335 * to alignment. This will be op->bytes if no alignment is necessary,
336 * or (new_end - op->offset) if the tail is rounded up or down due to
337 * alignment or buffer limit.
339 static void coroutine_fn
mirror_co_read(void *opaque
)
341 MirrorOp
*op
= opaque
;
342 MirrorBlockJob
*s
= op
->s
;
347 max_bytes
= s
->granularity
* s
->max_iov
;
349 /* We can only handle as much as buf_size at a time. */
350 op
->bytes
= MIN(s
->buf_size
, MIN(max_bytes
, op
->bytes
));
352 assert(op
->bytes
< BDRV_REQUEST_MAX_BYTES
);
353 *op
->bytes_handled
= op
->bytes
;
356 *op
->bytes_handled
+= mirror_cow_align(s
, &op
->offset
, &op
->bytes
);
358 /* Cannot exceed BDRV_REQUEST_MAX_BYTES + INT_MAX */
359 assert(*op
->bytes_handled
<= UINT_MAX
);
360 assert(op
->bytes
<= s
->buf_size
);
361 /* The offset is granularity-aligned because:
362 * 1) Caller passes in aligned values;
363 * 2) mirror_cow_align is used only when target cluster is larger. */
364 assert(QEMU_IS_ALIGNED(op
->offset
, s
->granularity
));
365 /* The range is sector-aligned, since bdrv_getlength() rounds up. */
366 assert(QEMU_IS_ALIGNED(op
->bytes
, BDRV_SECTOR_SIZE
));
367 nb_chunks
= DIV_ROUND_UP(op
->bytes
, s
->granularity
);
369 while (s
->buf_free_count
< nb_chunks
) {
370 trace_mirror_yield_in_flight(s
, op
->offset
, s
->in_flight
);
371 mirror_wait_for_free_in_flight_slot(s
);
374 /* Now make a QEMUIOVector taking enough granularity-sized chunks
377 qemu_iovec_init(&op
->qiov
, nb_chunks
);
378 while (nb_chunks
-- > 0) {
379 MirrorBuffer
*buf
= QSIMPLEQ_FIRST(&s
->buf_free
);
380 size_t remaining
= op
->bytes
- op
->qiov
.size
;
382 QSIMPLEQ_REMOVE_HEAD(&s
->buf_free
, next
);
384 qemu_iovec_add(&op
->qiov
, buf
, MIN(s
->granularity
, remaining
));
387 /* Copy the dirty cluster. */
389 s
->bytes_in_flight
+= op
->bytes
;
390 op
->is_in_flight
= true;
391 trace_mirror_one_iteration(s
, op
->offset
, op
->bytes
);
393 ret
= bdrv_co_preadv(s
->mirror_top_bs
->backing
, op
->offset
, op
->bytes
,
395 mirror_read_complete(op
, ret
);
398 static void coroutine_fn
mirror_co_zero(void *opaque
)
400 MirrorOp
*op
= opaque
;
404 op
->s
->bytes_in_flight
+= op
->bytes
;
405 *op
->bytes_handled
= op
->bytes
;
406 op
->is_in_flight
= true;
408 ret
= blk_co_pwrite_zeroes(op
->s
->target
, op
->offset
, op
->bytes
,
409 op
->s
->unmap
? BDRV_REQ_MAY_UNMAP
: 0);
410 mirror_write_complete(op
, ret
);
413 static void coroutine_fn
mirror_co_discard(void *opaque
)
415 MirrorOp
*op
= opaque
;
419 op
->s
->bytes_in_flight
+= op
->bytes
;
420 *op
->bytes_handled
= op
->bytes
;
421 op
->is_in_flight
= true;
423 ret
= blk_co_pdiscard(op
->s
->target
, op
->offset
, op
->bytes
);
424 mirror_write_complete(op
, ret
);
427 static unsigned mirror_perform(MirrorBlockJob
*s
, int64_t offset
,
428 unsigned bytes
, MirrorMethod mirror_method
)
432 int64_t bytes_handled
= -1;
434 op
= g_new(MirrorOp
, 1);
439 .bytes_handled
= &bytes_handled
,
441 qemu_co_queue_init(&op
->waiting_requests
);
443 switch (mirror_method
) {
444 case MIRROR_METHOD_COPY
:
445 co
= qemu_coroutine_create(mirror_co_read
, op
);
447 case MIRROR_METHOD_ZERO
:
448 co
= qemu_coroutine_create(mirror_co_zero
, op
);
450 case MIRROR_METHOD_DISCARD
:
451 co
= qemu_coroutine_create(mirror_co_discard
, op
);
458 QTAILQ_INSERT_TAIL(&s
->ops_in_flight
, op
, next
);
459 qemu_coroutine_enter(co
);
460 /* At this point, ownership of op has been moved to the coroutine
461 * and the object may already be freed */
463 /* Assert that this value has been set */
464 assert(bytes_handled
>= 0);
466 /* Same assertion as in mirror_co_read() (and for mirror_co_read()
467 * and mirror_co_discard(), bytes_handled == op->bytes, which
468 * is the @bytes parameter given to this function) */
469 assert(bytes_handled
<= UINT_MAX
);
470 return bytes_handled
;
473 static uint64_t coroutine_fn
mirror_iteration(MirrorBlockJob
*s
)
475 BlockDriverState
*source
= s
->mirror_top_bs
->backing
->bs
;
478 uint64_t delay_ns
= 0, ret
= 0;
479 /* At least the first dirty chunk is mirrored in one iteration. */
481 bool write_zeroes_ok
= bdrv_can_write_zeroes_with_unmap(blk_bs(s
->target
));
482 int max_io_bytes
= MAX(s
->buf_size
/ MAX_IN_FLIGHT
, MAX_IO_BYTES
);
484 bdrv_dirty_bitmap_lock(s
->dirty_bitmap
);
485 offset
= bdrv_dirty_iter_next(s
->dbi
);
487 bdrv_set_dirty_iter(s
->dbi
, 0);
488 offset
= bdrv_dirty_iter_next(s
->dbi
);
489 trace_mirror_restart_iter(s
, bdrv_get_dirty_count(s
->dirty_bitmap
));
492 bdrv_dirty_bitmap_unlock(s
->dirty_bitmap
);
495 * Wait for concurrent requests to @offset. The next loop will limit the
496 * copied area based on in_flight_bitmap so we only copy an area that does
497 * not overlap with concurrent in-flight requests. Still, we would like to
498 * copy something, so wait until there are at least no more requests to the
499 * very beginning of the area.
501 mirror_wait_on_conflicts(NULL
, s
, offset
, 1);
503 job_pause_point(&s
->common
.job
);
505 /* Find the number of consective dirty chunks following the first dirty
506 * one, and wait for in flight requests in them. */
507 bdrv_dirty_bitmap_lock(s
->dirty_bitmap
);
508 while (nb_chunks
* s
->granularity
< s
->buf_size
) {
510 int64_t next_offset
= offset
+ nb_chunks
* s
->granularity
;
511 int64_t next_chunk
= next_offset
/ s
->granularity
;
512 if (next_offset
>= s
->bdev_length
||
513 !bdrv_dirty_bitmap_get_locked(s
->dirty_bitmap
, next_offset
)) {
516 if (test_bit(next_chunk
, s
->in_flight_bitmap
)) {
520 next_dirty
= bdrv_dirty_iter_next(s
->dbi
);
521 if (next_dirty
> next_offset
|| next_dirty
< 0) {
522 /* The bitmap iterator's cache is stale, refresh it */
523 bdrv_set_dirty_iter(s
->dbi
, next_offset
);
524 next_dirty
= bdrv_dirty_iter_next(s
->dbi
);
526 assert(next_dirty
== next_offset
);
530 /* Clear dirty bits before querying the block status, because
531 * calling bdrv_block_status_above could yield - if some blocks are
532 * marked dirty in this window, we need to know.
534 bdrv_reset_dirty_bitmap_locked(s
->dirty_bitmap
, offset
,
535 nb_chunks
* s
->granularity
);
536 bdrv_dirty_bitmap_unlock(s
->dirty_bitmap
);
538 /* Before claiming an area in the in-flight bitmap, we have to
539 * create a MirrorOp for it so that conflicting requests can wait
540 * for it. mirror_perform() will create the real MirrorOps later,
541 * for now we just create a pseudo operation that will wake up all
542 * conflicting requests once all real operations have been
544 pseudo_op
= g_new(MirrorOp
, 1);
545 *pseudo_op
= (MirrorOp
){
547 .bytes
= nb_chunks
* s
->granularity
,
548 .is_pseudo_op
= true,
550 qemu_co_queue_init(&pseudo_op
->waiting_requests
);
551 QTAILQ_INSERT_TAIL(&s
->ops_in_flight
, pseudo_op
, next
);
553 bitmap_set(s
->in_flight_bitmap
, offset
/ s
->granularity
, nb_chunks
);
554 while (nb_chunks
> 0 && offset
< s
->bdev_length
) {
557 int64_t io_bytes_acct
;
558 MirrorMethod mirror_method
= MIRROR_METHOD_COPY
;
560 assert(!(offset
% s
->granularity
));
561 WITH_GRAPH_RDLOCK_GUARD() {
562 ret
= bdrv_block_status_above(source
, NULL
, offset
,
563 nb_chunks
* s
->granularity
,
564 &io_bytes
, NULL
, NULL
);
567 io_bytes
= MIN(nb_chunks
* s
->granularity
, max_io_bytes
);
568 } else if (ret
& BDRV_BLOCK_DATA
) {
569 io_bytes
= MIN(io_bytes
, max_io_bytes
);
572 io_bytes
-= io_bytes
% s
->granularity
;
573 if (io_bytes
< s
->granularity
) {
574 io_bytes
= s
->granularity
;
575 } else if (ret
>= 0 && !(ret
& BDRV_BLOCK_DATA
)) {
576 int64_t target_offset
;
577 int64_t target_bytes
;
578 bdrv_round_to_clusters(blk_bs(s
->target
), offset
, io_bytes
,
579 &target_offset
, &target_bytes
);
580 if (target_offset
== offset
&&
581 target_bytes
== io_bytes
) {
582 mirror_method
= ret
& BDRV_BLOCK_ZERO
?
584 MIRROR_METHOD_DISCARD
;
588 while (s
->in_flight
>= MAX_IN_FLIGHT
) {
589 trace_mirror_yield_in_flight(s
, offset
, s
->in_flight
);
590 mirror_wait_for_free_in_flight_slot(s
);
598 io_bytes
= mirror_clip_bytes(s
, offset
, io_bytes
);
599 io_bytes
= mirror_perform(s
, offset
, io_bytes
, mirror_method
);
600 if (mirror_method
!= MIRROR_METHOD_COPY
&& write_zeroes_ok
) {
603 io_bytes_acct
= io_bytes
;
607 nb_chunks
-= DIV_ROUND_UP(io_bytes
, s
->granularity
);
608 delay_ns
= block_job_ratelimit_get_delay(&s
->common
, io_bytes_acct
);
613 QTAILQ_REMOVE(&s
->ops_in_flight
, pseudo_op
, next
);
614 qemu_co_queue_restart_all(&pseudo_op
->waiting_requests
);
620 static void mirror_free_init(MirrorBlockJob
*s
)
622 int granularity
= s
->granularity
;
623 size_t buf_size
= s
->buf_size
;
624 uint8_t *buf
= s
->buf
;
626 assert(s
->buf_free_count
== 0);
627 QSIMPLEQ_INIT(&s
->buf_free
);
628 while (buf_size
!= 0) {
629 MirrorBuffer
*cur
= (MirrorBuffer
*)buf
;
630 QSIMPLEQ_INSERT_TAIL(&s
->buf_free
, cur
, next
);
632 buf_size
-= granularity
;
637 /* This is also used for the .pause callback. There is no matching
638 * mirror_resume() because mirror_run() will begin iterating again
639 * when the job is resumed.
641 static void coroutine_fn
mirror_wait_for_all_io(MirrorBlockJob
*s
)
643 while (s
->in_flight
> 0) {
644 mirror_wait_for_free_in_flight_slot(s
);
649 * mirror_exit_common: handle both abort() and prepare() cases.
650 * for .prepare, returns 0 on success and -errno on failure.
651 * for .abort cases, denoted by abort = true, MUST return 0.
653 static int mirror_exit_common(Job
*job
)
655 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
.job
);
656 BlockJob
*bjob
= &s
->common
;
657 MirrorBDSOpaque
*bs_opaque
;
658 AioContext
*replace_aio_context
= NULL
;
659 BlockDriverState
*src
;
660 BlockDriverState
*target_bs
;
661 BlockDriverState
*mirror_top_bs
;
662 Error
*local_err
= NULL
;
663 bool abort
= job
->ret
< 0;
671 mirror_top_bs
= s
->mirror_top_bs
;
672 bs_opaque
= mirror_top_bs
->opaque
;
673 src
= mirror_top_bs
->backing
->bs
;
674 target_bs
= blk_bs(s
->target
);
676 if (bdrv_chain_contains(src
, target_bs
)) {
677 bdrv_unfreeze_backing_chain(mirror_top_bs
, target_bs
);
680 bdrv_release_dirty_bitmap(s
->dirty_bitmap
);
682 /* Make sure that the source BDS doesn't go away during bdrv_replace_node,
683 * before we can call bdrv_drained_end */
685 bdrv_ref(mirror_top_bs
);
689 * Remove target parent that still uses BLK_PERM_WRITE/RESIZE before
690 * inserting target_bs at s->to_replace, where we might not be able to get
693 blk_unref(s
->target
);
696 /* We don't access the source any more. Dropping any WRITE/RESIZE is
697 * required before it could become a backing file of target_bs. Not having
698 * these permissions any more means that we can't allow any new requests on
699 * mirror_top_bs from now on, so keep it drained. */
700 bdrv_drained_begin(mirror_top_bs
);
701 bs_opaque
->stop
= true;
702 bdrv_child_refresh_perms(mirror_top_bs
, mirror_top_bs
->backing
,
704 if (!abort
&& s
->backing_mode
== MIRROR_SOURCE_BACKING_CHAIN
) {
705 BlockDriverState
*backing
= s
->is_none_mode
? src
: s
->base
;
706 BlockDriverState
*unfiltered_target
= bdrv_skip_filters(target_bs
);
708 if (bdrv_cow_bs(unfiltered_target
) != backing
) {
709 bdrv_set_backing_hd(unfiltered_target
, backing
, &local_err
);
711 error_report_err(local_err
);
716 } else if (!abort
&& s
->backing_mode
== MIRROR_OPEN_BACKING_CHAIN
) {
717 assert(!bdrv_backing_chain_next(target_bs
));
718 ret
= bdrv_open_backing_file(bdrv_skip_filters(target_bs
), NULL
,
719 "backing", &local_err
);
721 error_report_err(local_err
);
727 replace_aio_context
= bdrv_get_aio_context(s
->to_replace
);
728 aio_context_acquire(replace_aio_context
);
731 if (s
->should_complete
&& !abort
) {
732 BlockDriverState
*to_replace
= s
->to_replace
?: src
;
733 bool ro
= bdrv_is_read_only(to_replace
);
735 if (ro
!= bdrv_is_read_only(target_bs
)) {
736 bdrv_reopen_set_read_only(target_bs
, ro
, NULL
);
739 /* The mirror job has no requests in flight any more, but we need to
740 * drain potential other users of the BDS before changing the graph. */
742 bdrv_drained_begin(target_bs
);
744 * Cannot use check_to_replace_node() here, because that would
745 * check for an op blocker on @to_replace, and we have our own
748 if (bdrv_recurse_can_replace(src
, to_replace
)) {
749 bdrv_replace_node(to_replace
, target_bs
, &local_err
);
751 error_setg(&local_err
, "Can no longer replace '%s' by '%s', "
752 "because it can no longer be guaranteed that doing so "
753 "would not lead to an abrupt change of visible data",
754 to_replace
->node_name
, target_bs
->node_name
);
756 bdrv_drained_end(target_bs
);
758 error_report_err(local_err
);
763 bdrv_op_unblock_all(s
->to_replace
, s
->replace_blocker
);
764 error_free(s
->replace_blocker
);
765 bdrv_unref(s
->to_replace
);
767 if (replace_aio_context
) {
768 aio_context_release(replace_aio_context
);
771 bdrv_unref(target_bs
);
774 * Remove the mirror filter driver from the graph. Before this, get rid of
775 * the blockers on the intermediate nodes so that the resulting state is
778 block_job_remove_all_bdrv(bjob
);
779 bdrv_replace_node(mirror_top_bs
, mirror_top_bs
->backing
->bs
, &error_abort
);
781 bs_opaque
->job
= NULL
;
783 bdrv_drained_end(src
);
784 bdrv_drained_end(mirror_top_bs
);
786 bdrv_unref(mirror_top_bs
);
792 static int mirror_prepare(Job
*job
)
794 return mirror_exit_common(job
);
797 static void mirror_abort(Job
*job
)
799 int ret
= mirror_exit_common(job
);
803 static void coroutine_fn
mirror_throttle(MirrorBlockJob
*s
)
805 int64_t now
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
);
807 if (now
- s
->last_pause_ns
> BLOCK_JOB_SLICE_TIME
) {
808 s
->last_pause_ns
= now
;
809 job_sleep_ns(&s
->common
.job
, 0);
811 job_pause_point(&s
->common
.job
);
815 static int coroutine_fn
mirror_dirty_init(MirrorBlockJob
*s
)
818 BlockDriverState
*bs
= s
->mirror_top_bs
->backing
->bs
;
819 BlockDriverState
*target_bs
= blk_bs(s
->target
);
823 if (s
->zero_target
) {
824 if (!bdrv_can_write_zeroes_with_unmap(target_bs
)) {
825 bdrv_set_dirty_bitmap(s
->dirty_bitmap
, 0, s
->bdev_length
);
829 s
->initial_zeroing_ongoing
= true;
830 for (offset
= 0; offset
< s
->bdev_length
; ) {
831 int bytes
= MIN(s
->bdev_length
- offset
,
832 QEMU_ALIGN_DOWN(INT_MAX
, s
->granularity
));
836 if (job_is_cancelled(&s
->common
.job
)) {
837 s
->initial_zeroing_ongoing
= false;
841 if (s
->in_flight
>= MAX_IN_FLIGHT
) {
842 trace_mirror_yield(s
, UINT64_MAX
, s
->buf_free_count
,
844 mirror_wait_for_free_in_flight_slot(s
);
848 mirror_perform(s
, offset
, bytes
, MIRROR_METHOD_ZERO
);
852 mirror_wait_for_all_io(s
);
853 s
->initial_zeroing_ongoing
= false;
856 /* First part, loop on the sectors and initialize the dirty bitmap. */
857 for (offset
= 0; offset
< s
->bdev_length
; ) {
858 /* Just to make sure we are not exceeding int limit. */
859 int bytes
= MIN(s
->bdev_length
- offset
,
860 QEMU_ALIGN_DOWN(INT_MAX
, s
->granularity
));
864 if (job_is_cancelled(&s
->common
.job
)) {
868 WITH_GRAPH_RDLOCK_GUARD() {
869 ret
= bdrv_is_allocated_above(bs
, s
->base_overlay
, true, offset
,
878 bdrv_set_dirty_bitmap(s
->dirty_bitmap
, offset
, count
);
885 /* Called when going out of the streaming phase to flush the bulk of the
886 * data to the medium, or just before completing.
888 static int mirror_flush(MirrorBlockJob
*s
)
890 int ret
= blk_flush(s
->target
);
892 if (mirror_error_action(s
, false, -ret
) == BLOCK_ERROR_ACTION_REPORT
) {
899 static int coroutine_fn
mirror_run(Job
*job
, Error
**errp
)
901 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
.job
);
902 BlockDriverState
*bs
= s
->mirror_top_bs
->backing
->bs
;
903 MirrorBDSOpaque
*mirror_top_opaque
= s
->mirror_top_bs
->opaque
;
904 BlockDriverState
*target_bs
= blk_bs(s
->target
);
905 bool need_drain
= true;
906 BlockDeviceIoStatus iostatus
;
908 int64_t target_length
;
910 char backing_filename
[2]; /* we only need 2 characters because we are only
911 checking for a NULL string */
914 if (job_is_cancelled(&s
->common
.job
)) {
918 s
->bdev_length
= bdrv_co_getlength(bs
);
919 if (s
->bdev_length
< 0) {
920 ret
= s
->bdev_length
;
924 target_length
= blk_co_getlength(s
->target
);
925 if (target_length
< 0) {
930 /* Active commit must resize the base image if its size differs from the
932 if (s
->base
== blk_bs(s
->target
)) {
933 if (s
->bdev_length
> target_length
) {
934 ret
= blk_co_truncate(s
->target
, s
->bdev_length
, false,
935 PREALLOC_MODE_OFF
, 0, NULL
);
940 } else if (s
->bdev_length
!= target_length
) {
941 error_setg(errp
, "Source and target image have different sizes");
946 if (s
->bdev_length
== 0) {
947 /* Transition to the READY state and wait for complete. */
948 job_transition_to_ready(&s
->common
.job
);
949 s
->actively_synced
= true;
950 while (!job_cancel_requested(&s
->common
.job
) && !s
->should_complete
) {
951 job_yield(&s
->common
.job
);
956 length
= DIV_ROUND_UP(s
->bdev_length
, s
->granularity
);
957 s
->in_flight_bitmap
= bitmap_new(length
);
959 /* If we have no backing file yet in the destination, we cannot let
960 * the destination do COW. Instead, we copy sectors around the
961 * dirty data if needed. We need a bitmap to do that.
963 bdrv_get_backing_filename(target_bs
, backing_filename
,
964 sizeof(backing_filename
));
965 if (!bdrv_co_get_info(target_bs
, &bdi
) && bdi
.cluster_size
) {
966 s
->target_cluster_size
= bdi
.cluster_size
;
968 s
->target_cluster_size
= BDRV_SECTOR_SIZE
;
970 if (backing_filename
[0] && !bdrv_backing_chain_next(target_bs
) &&
971 s
->granularity
< s
->target_cluster_size
) {
972 s
->buf_size
= MAX(s
->buf_size
, s
->target_cluster_size
);
973 s
->cow_bitmap
= bitmap_new(length
);
975 s
->max_iov
= MIN(bs
->bl
.max_iov
, target_bs
->bl
.max_iov
);
977 s
->buf
= qemu_try_blockalign(bs
, s
->buf_size
);
978 if (s
->buf
== NULL
) {
985 s
->last_pause_ns
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
);
986 if (!s
->is_none_mode
) {
987 ret
= mirror_dirty_init(s
);
988 if (ret
< 0 || job_is_cancelled(&s
->common
.job
)) {
994 * Only now the job is fully initialised and mirror_top_bs should start
997 mirror_top_opaque
->job
= s
;
1000 s
->dbi
= bdrv_dirty_iter_new(s
->dirty_bitmap
);
1002 uint64_t delay_ns
= 0;
1004 bool should_complete
;
1008 goto immediate_exit
;
1011 job_pause_point(&s
->common
.job
);
1013 if (job_is_cancelled(&s
->common
.job
)) {
1015 goto immediate_exit
;
1018 cnt
= bdrv_get_dirty_count(s
->dirty_bitmap
);
1019 /* cnt is the number of dirty bytes remaining and s->bytes_in_flight is
1020 * the number of bytes currently being processed; together those are
1021 * the current remaining operation length */
1022 job_progress_set_remaining(&s
->common
.job
,
1023 s
->bytes_in_flight
+ cnt
+
1024 s
->active_write_bytes_in_flight
);
1026 /* Note that even when no rate limit is applied we need to yield
1027 * periodically with no pending I/O so that bdrv_drain_all() returns.
1028 * We do so every BLKOCK_JOB_SLICE_TIME nanoseconds, or when there is
1029 * an error, or when the source is clean, whichever comes first. */
1030 delta
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
) - s
->last_pause_ns
;
1031 WITH_JOB_LOCK_GUARD() {
1032 iostatus
= s
->common
.iostatus
;
1034 if (delta
< BLOCK_JOB_SLICE_TIME
&&
1035 iostatus
== BLOCK_DEVICE_IO_STATUS_OK
) {
1036 if (s
->in_flight
>= MAX_IN_FLIGHT
|| s
->buf_free_count
== 0 ||
1037 (cnt
== 0 && s
->in_flight
> 0)) {
1038 trace_mirror_yield(s
, cnt
, s
->buf_free_count
, s
->in_flight
);
1039 mirror_wait_for_free_in_flight_slot(s
);
1041 } else if (cnt
!= 0) {
1042 delay_ns
= mirror_iteration(s
);
1046 should_complete
= false;
1047 if (s
->in_flight
== 0 && cnt
== 0) {
1048 trace_mirror_before_flush(s
);
1049 if (!job_is_ready(&s
->common
.job
)) {
1050 if (mirror_flush(s
) < 0) {
1051 /* Go check s->ret. */
1054 /* We're out of the streaming phase. From now on, if the job
1055 * is cancelled we will actually complete all pending I/O and
1056 * report completion. This way, block-job-cancel will leave
1057 * the target in a consistent state.
1059 job_transition_to_ready(&s
->common
.job
);
1060 if (s
->copy_mode
!= MIRROR_COPY_MODE_BACKGROUND
) {
1061 s
->actively_synced
= true;
1065 should_complete
= s
->should_complete
||
1066 job_cancel_requested(&s
->common
.job
);
1067 cnt
= bdrv_get_dirty_count(s
->dirty_bitmap
);
1070 if (cnt
== 0 && should_complete
) {
1071 /* The dirty bitmap is not updated while operations are pending.
1072 * If we're about to exit, wait for pending operations before
1073 * calling bdrv_get_dirty_count(bs), or we may exit while the
1074 * source has dirty data to copy!
1076 * Note that I/O can be submitted by the guest while
1077 * mirror_populate runs, so pause it now. Before deciding
1078 * whether to switch to target check one last time if I/O has
1079 * come in the meanwhile, and if not flush the data to disk.
1081 trace_mirror_before_drain(s
, cnt
);
1084 bdrv_drained_begin(bs
);
1086 /* Must be zero because we are drained */
1087 assert(s
->in_active_write_counter
== 0);
1089 cnt
= bdrv_get_dirty_count(s
->dirty_bitmap
);
1090 if (cnt
> 0 || mirror_flush(s
) < 0) {
1091 bdrv_drained_end(bs
);
1092 s
->in_drain
= false;
1096 /* The two disks are in sync. Exit and report successful
1099 assert(QLIST_EMPTY(&bs
->tracked_requests
));
1104 if (job_is_ready(&s
->common
.job
) && !should_complete
) {
1105 delay_ns
= (s
->in_flight
== 0 &&
1106 cnt
== 0 ? BLOCK_JOB_SLICE_TIME
: 0);
1108 trace_mirror_before_sleep(s
, cnt
, job_is_ready(&s
->common
.job
),
1110 job_sleep_ns(&s
->common
.job
, delay_ns
);
1111 s
->last_pause_ns
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
);
1115 if (s
->in_flight
> 0) {
1116 /* We get here only if something went wrong. Either the job failed,
1117 * or it was cancelled prematurely so that we do not guarantee that
1118 * the target is a copy of the source.
1120 assert(ret
< 0 || job_is_cancelled(&s
->common
.job
));
1122 mirror_wait_for_all_io(s
);
1125 assert(s
->in_flight
== 0);
1127 g_free(s
->cow_bitmap
);
1128 g_free(s
->in_flight_bitmap
);
1129 bdrv_dirty_iter_free(s
->dbi
);
1133 bdrv_drained_begin(bs
);
1139 static void mirror_complete(Job
*job
, Error
**errp
)
1141 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
.job
);
1143 if (!job_is_ready(job
)) {
1144 error_setg(errp
, "The active block job '%s' cannot be completed",
1149 /* block all operations on to_replace bs */
1151 AioContext
*replace_aio_context
;
1153 s
->to_replace
= bdrv_find_node(s
->replaces
);
1154 if (!s
->to_replace
) {
1155 error_setg(errp
, "Node name '%s' not found", s
->replaces
);
1159 replace_aio_context
= bdrv_get_aio_context(s
->to_replace
);
1160 aio_context_acquire(replace_aio_context
);
1162 /* TODO Translate this into child freeze system. */
1163 error_setg(&s
->replace_blocker
,
1164 "block device is in use by block-job-complete");
1165 bdrv_op_block_all(s
->to_replace
, s
->replace_blocker
);
1166 bdrv_ref(s
->to_replace
);
1168 aio_context_release(replace_aio_context
);
1171 s
->should_complete
= true;
1173 /* If the job is paused, it will be re-entered when it is resumed */
1174 WITH_JOB_LOCK_GUARD() {
1176 job_enter_cond_locked(job
, NULL
);
1181 static void coroutine_fn
mirror_pause(Job
*job
)
1183 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
.job
);
1185 mirror_wait_for_all_io(s
);
1188 static bool mirror_drained_poll(BlockJob
*job
)
1190 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
1192 /* If the job isn't paused nor cancelled, we can't be sure that it won't
1193 * issue more requests. We make an exception if we've reached this point
1194 * from one of our own drain sections, to avoid a deadlock waiting for
1197 WITH_JOB_LOCK_GUARD() {
1198 if (!s
->common
.job
.paused
&& !job_is_cancelled_locked(&job
->job
)
1204 return !!s
->in_flight
;
1207 static bool mirror_cancel(Job
*job
, bool force
)
1209 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
.job
);
1210 BlockDriverState
*target
= blk_bs(s
->target
);
1213 * Before the job is READY, we treat any cancellation like a
1214 * force-cancellation.
1216 force
= force
|| !job_is_ready(job
);
1219 bdrv_cancel_in_flight(target
);
1224 static bool commit_active_cancel(Job
*job
, bool force
)
1226 /* Same as above in mirror_cancel() */
1227 return force
|| !job_is_ready(job
);
1230 static const BlockJobDriver mirror_job_driver
= {
1232 .instance_size
= sizeof(MirrorBlockJob
),
1233 .job_type
= JOB_TYPE_MIRROR
,
1234 .free
= block_job_free
,
1235 .user_resume
= block_job_user_resume
,
1237 .prepare
= mirror_prepare
,
1238 .abort
= mirror_abort
,
1239 .pause
= mirror_pause
,
1240 .complete
= mirror_complete
,
1241 .cancel
= mirror_cancel
,
1243 .drained_poll
= mirror_drained_poll
,
1246 static const BlockJobDriver commit_active_job_driver
= {
1248 .instance_size
= sizeof(MirrorBlockJob
),
1249 .job_type
= JOB_TYPE_COMMIT
,
1250 .free
= block_job_free
,
1251 .user_resume
= block_job_user_resume
,
1253 .prepare
= mirror_prepare
,
1254 .abort
= mirror_abort
,
1255 .pause
= mirror_pause
,
1256 .complete
= mirror_complete
,
1257 .cancel
= commit_active_cancel
,
1259 .drained_poll
= mirror_drained_poll
,
1262 static void coroutine_fn
1263 do_sync_target_write(MirrorBlockJob
*job
, MirrorMethod method
,
1264 uint64_t offset
, uint64_t bytes
,
1265 QEMUIOVector
*qiov
, int flags
)
1268 size_t qiov_offset
= 0;
1269 int64_t bitmap_offset
, bitmap_end
;
1271 if (!QEMU_IS_ALIGNED(offset
, job
->granularity
) &&
1272 bdrv_dirty_bitmap_get(job
->dirty_bitmap
, offset
))
1275 * Dirty unaligned padding: ignore it.
1278 * 1. If we copy it, we can't reset corresponding bit in
1279 * dirty_bitmap as there may be some "dirty" bytes still not
1281 * 2. It's already dirty, so skipping it we don't diverge mirror
1284 * Note, that because of this, guest write may have no contribution
1285 * into mirror converge, but that's not bad, as we have background
1286 * process of mirroring. If under some bad circumstances (high guest
1287 * IO load) background process starve, we will not converge anyway,
1288 * even if each write will contribute, as guest is not guaranteed to
1289 * rewrite the whole disk.
1291 qiov_offset
= QEMU_ALIGN_UP(offset
, job
->granularity
) - offset
;
1292 if (bytes
<= qiov_offset
) {
1293 /* nothing to do after shrink */
1296 offset
+= qiov_offset
;
1297 bytes
-= qiov_offset
;
1300 if (!QEMU_IS_ALIGNED(offset
+ bytes
, job
->granularity
) &&
1301 bdrv_dirty_bitmap_get(job
->dirty_bitmap
, offset
+ bytes
- 1))
1303 uint64_t tail
= (offset
+ bytes
) % job
->granularity
;
1305 if (bytes
<= tail
) {
1306 /* nothing to do after shrink */
1313 * Tails are either clean or shrunk, so for bitmap resetting
1314 * we safely align the range down.
1316 bitmap_offset
= QEMU_ALIGN_UP(offset
, job
->granularity
);
1317 bitmap_end
= QEMU_ALIGN_DOWN(offset
+ bytes
, job
->granularity
);
1318 if (bitmap_offset
< bitmap_end
) {
1319 bdrv_reset_dirty_bitmap(job
->dirty_bitmap
, bitmap_offset
,
1320 bitmap_end
- bitmap_offset
);
1323 job_progress_increase_remaining(&job
->common
.job
, bytes
);
1324 job
->active_write_bytes_in_flight
+= bytes
;
1327 case MIRROR_METHOD_COPY
:
1328 ret
= blk_co_pwritev_part(job
->target
, offset
, bytes
,
1329 qiov
, qiov_offset
, flags
);
1332 case MIRROR_METHOD_ZERO
:
1334 ret
= blk_co_pwrite_zeroes(job
->target
, offset
, bytes
, flags
);
1337 case MIRROR_METHOD_DISCARD
:
1339 ret
= blk_co_pdiscard(job
->target
, offset
, bytes
);
1346 job
->active_write_bytes_in_flight
-= bytes
;
1348 job_progress_update(&job
->common
.job
, bytes
);
1350 BlockErrorAction action
;
1353 * We failed, so we should mark dirty the whole area, aligned up.
1354 * Note that we don't care about shrunk tails if any: they were dirty
1355 * at function start, and they must be still dirty, as we've locked
1356 * the region for in-flight op.
1358 bitmap_offset
= QEMU_ALIGN_DOWN(offset
, job
->granularity
);
1359 bitmap_end
= QEMU_ALIGN_UP(offset
+ bytes
, job
->granularity
);
1360 bdrv_set_dirty_bitmap(job
->dirty_bitmap
, bitmap_offset
,
1361 bitmap_end
- bitmap_offset
);
1362 job
->actively_synced
= false;
1364 action
= mirror_error_action(job
, false, -ret
);
1365 if (action
== BLOCK_ERROR_ACTION_REPORT
) {
1373 static MirrorOp
*coroutine_fn
active_write_prepare(MirrorBlockJob
*s
,
1378 uint64_t start_chunk
= offset
/ s
->granularity
;
1379 uint64_t end_chunk
= DIV_ROUND_UP(offset
+ bytes
, s
->granularity
);
1381 op
= g_new(MirrorOp
, 1);
1386 .is_active_write
= true,
1387 .is_in_flight
= true,
1388 .co
= qemu_coroutine_self(),
1390 qemu_co_queue_init(&op
->waiting_requests
);
1391 QTAILQ_INSERT_TAIL(&s
->ops_in_flight
, op
, next
);
1393 s
->in_active_write_counter
++;
1396 * Wait for concurrent requests affecting the area. If there are already
1397 * running requests that are copying off now-to-be stale data in the area,
1398 * we must wait for them to finish before we begin writing fresh data to the
1399 * target so that the write operations appear in the correct order.
1400 * Note that background requests (see mirror_iteration()) in contrast only
1401 * wait for conflicting requests at the start of the dirty area, and then
1402 * (based on the in_flight_bitmap) truncate the area to copy so it will not
1403 * conflict with any requests beyond that. For active writes, however, we
1404 * cannot truncate that area. The request from our parent must be blocked
1405 * until the area is copied in full. Therefore, we must wait for the whole
1406 * area to become free of concurrent requests.
1408 mirror_wait_on_conflicts(op
, s
, offset
, bytes
);
1410 bitmap_set(s
->in_flight_bitmap
, start_chunk
, end_chunk
- start_chunk
);
1415 static void coroutine_fn
active_write_settle(MirrorOp
*op
)
1417 uint64_t start_chunk
= op
->offset
/ op
->s
->granularity
;
1418 uint64_t end_chunk
= DIV_ROUND_UP(op
->offset
+ op
->bytes
,
1419 op
->s
->granularity
);
1421 if (!--op
->s
->in_active_write_counter
&& op
->s
->actively_synced
) {
1422 BdrvChild
*source
= op
->s
->mirror_top_bs
->backing
;
1424 if (QLIST_FIRST(&source
->bs
->parents
) == source
&&
1425 QLIST_NEXT(source
, next_parent
) == NULL
)
1427 /* Assert that we are back in sync once all active write
1428 * operations are settled.
1429 * Note that we can only assert this if the mirror node
1430 * is the source node's only parent. */
1431 assert(!bdrv_get_dirty_count(op
->s
->dirty_bitmap
));
1434 bitmap_clear(op
->s
->in_flight_bitmap
, start_chunk
, end_chunk
- start_chunk
);
1435 QTAILQ_REMOVE(&op
->s
->ops_in_flight
, op
, next
);
1436 qemu_co_queue_restart_all(&op
->waiting_requests
);
1440 static int coroutine_fn
bdrv_mirror_top_preadv(BlockDriverState
*bs
,
1441 int64_t offset
, int64_t bytes
, QEMUIOVector
*qiov
, BdrvRequestFlags flags
)
1443 return bdrv_co_preadv(bs
->backing
, offset
, bytes
, qiov
, flags
);
1446 static int coroutine_fn GRAPH_RDLOCK
1447 bdrv_mirror_top_do_write(BlockDriverState
*bs
, MirrorMethod method
,
1448 uint64_t offset
, uint64_t bytes
, QEMUIOVector
*qiov
,
1451 MirrorOp
*op
= NULL
;
1452 MirrorBDSOpaque
*s
= bs
->opaque
;
1454 bool copy_to_target
= false;
1457 copy_to_target
= s
->job
->ret
>= 0 &&
1458 !job_is_cancelled(&s
->job
->common
.job
) &&
1459 s
->job
->copy_mode
== MIRROR_COPY_MODE_WRITE_BLOCKING
;
1462 if (copy_to_target
) {
1463 op
= active_write_prepare(s
->job
, offset
, bytes
);
1467 case MIRROR_METHOD_COPY
:
1468 ret
= bdrv_co_pwritev(bs
->backing
, offset
, bytes
, qiov
, flags
);
1471 case MIRROR_METHOD_ZERO
:
1472 ret
= bdrv_co_pwrite_zeroes(bs
->backing
, offset
, bytes
, flags
);
1475 case MIRROR_METHOD_DISCARD
:
1476 ret
= bdrv_co_pdiscard(bs
->backing
, offset
, bytes
);
1487 if (copy_to_target
) {
1488 do_sync_target_write(s
->job
, method
, offset
, bytes
, qiov
, flags
);
1492 if (copy_to_target
) {
1493 active_write_settle(op
);
1498 static int coroutine_fn
bdrv_mirror_top_pwritev(BlockDriverState
*bs
,
1499 int64_t offset
, int64_t bytes
, QEMUIOVector
*qiov
, BdrvRequestFlags flags
)
1501 MirrorBDSOpaque
*s
= bs
->opaque
;
1502 QEMUIOVector bounce_qiov
;
1505 bool copy_to_target
= false;
1507 assume_graph_lock(); /* FIXME */
1510 copy_to_target
= s
->job
->ret
>= 0 &&
1511 !job_is_cancelled(&s
->job
->common
.job
) &&
1512 s
->job
->copy_mode
== MIRROR_COPY_MODE_WRITE_BLOCKING
;
1515 if (copy_to_target
) {
1516 /* The guest might concurrently modify the data to write; but
1517 * the data on source and destination must match, so we have
1518 * to use a bounce buffer if we are going to write to the
1520 bounce_buf
= qemu_blockalign(bs
, bytes
);
1521 iov_to_buf_full(qiov
->iov
, qiov
->niov
, 0, bounce_buf
, bytes
);
1523 qemu_iovec_init(&bounce_qiov
, 1);
1524 qemu_iovec_add(&bounce_qiov
, bounce_buf
, bytes
);
1525 qiov
= &bounce_qiov
;
1527 flags
&= ~BDRV_REQ_REGISTERED_BUF
;
1530 ret
= bdrv_mirror_top_do_write(bs
, MIRROR_METHOD_COPY
, offset
, bytes
, qiov
,
1533 if (copy_to_target
) {
1534 qemu_iovec_destroy(&bounce_qiov
);
1535 qemu_vfree(bounce_buf
);
1541 static int coroutine_fn GRAPH_RDLOCK
bdrv_mirror_top_flush(BlockDriverState
*bs
)
1543 if (bs
->backing
== NULL
) {
1544 /* we can be here after failed bdrv_append in mirror_start_job */
1547 return bdrv_co_flush(bs
->backing
->bs
);
1550 static int coroutine_fn GRAPH_RDLOCK
1551 bdrv_mirror_top_pwrite_zeroes(BlockDriverState
*bs
, int64_t offset
,
1552 int64_t bytes
, BdrvRequestFlags flags
)
1554 return bdrv_mirror_top_do_write(bs
, MIRROR_METHOD_ZERO
, offset
, bytes
, NULL
,
1558 static int coroutine_fn GRAPH_RDLOCK
1559 bdrv_mirror_top_pdiscard(BlockDriverState
*bs
, int64_t offset
, int64_t bytes
)
1561 return bdrv_mirror_top_do_write(bs
, MIRROR_METHOD_DISCARD
, offset
, bytes
,
1565 static void bdrv_mirror_top_refresh_filename(BlockDriverState
*bs
)
1567 if (bs
->backing
== NULL
) {
1568 /* we can be here after failed bdrv_attach_child in
1569 * bdrv_set_backing_hd */
1572 pstrcpy(bs
->exact_filename
, sizeof(bs
->exact_filename
),
1573 bs
->backing
->bs
->filename
);
1576 static void bdrv_mirror_top_child_perm(BlockDriverState
*bs
, BdrvChild
*c
,
1578 BlockReopenQueue
*reopen_queue
,
1579 uint64_t perm
, uint64_t shared
,
1580 uint64_t *nperm
, uint64_t *nshared
)
1582 MirrorBDSOpaque
*s
= bs
->opaque
;
1586 * If the job is to be stopped, we do not need to forward
1587 * anything to the real image.
1590 *nshared
= BLK_PERM_ALL
;
1594 bdrv_default_perms(bs
, c
, role
, reopen_queue
,
1595 perm
, shared
, nperm
, nshared
);
1599 * For commit jobs, we cannot take CONSISTENT_READ, because
1600 * that permission is unshared for everything above the base
1601 * node (except for filters on the base node).
1602 * We also have to force-share the WRITE permission, or
1603 * otherwise we would block ourselves at the base node (if
1604 * writes are blocked for a node, they are also blocked for
1605 * its backing file).
1606 * (We could also share RESIZE, because it may be needed for
1607 * the target if its size is less than the top node's; but
1608 * bdrv_default_perms_for_cow() automatically shares RESIZE
1609 * for backing nodes if WRITE is shared, so there is no need
1612 *nperm
&= ~BLK_PERM_CONSISTENT_READ
;
1613 *nshared
|= BLK_PERM_WRITE
;
1617 /* Dummy node that provides consistent read to its users without requiring it
1618 * from its backing file and that allows writes on the backing file chain. */
1619 static BlockDriver bdrv_mirror_top
= {
1620 .format_name
= "mirror_top",
1621 .bdrv_co_preadv
= bdrv_mirror_top_preadv
,
1622 .bdrv_co_pwritev
= bdrv_mirror_top_pwritev
,
1623 .bdrv_co_pwrite_zeroes
= bdrv_mirror_top_pwrite_zeroes
,
1624 .bdrv_co_pdiscard
= bdrv_mirror_top_pdiscard
,
1625 .bdrv_co_flush
= bdrv_mirror_top_flush
,
1626 .bdrv_refresh_filename
= bdrv_mirror_top_refresh_filename
,
1627 .bdrv_child_perm
= bdrv_mirror_top_child_perm
,
1630 .filtered_child_is_backing
= true,
1633 static BlockJob
*mirror_start_job(
1634 const char *job_id
, BlockDriverState
*bs
,
1635 int creation_flags
, BlockDriverState
*target
,
1636 const char *replaces
, int64_t speed
,
1637 uint32_t granularity
, int64_t buf_size
,
1638 BlockMirrorBackingMode backing_mode
,
1640 BlockdevOnError on_source_error
,
1641 BlockdevOnError on_target_error
,
1643 BlockCompletionFunc
*cb
,
1645 const BlockJobDriver
*driver
,
1646 bool is_none_mode
, BlockDriverState
*base
,
1647 bool auto_complete
, const char *filter_node_name
,
1648 bool is_mirror
, MirrorCopyMode copy_mode
,
1652 MirrorBDSOpaque
*bs_opaque
;
1653 BlockDriverState
*mirror_top_bs
;
1654 bool target_is_backing
;
1655 uint64_t target_perms
, target_shared_perms
;
1658 if (granularity
== 0) {
1659 granularity
= bdrv_get_default_bitmap_granularity(target
);
1662 assert(is_power_of_2(granularity
));
1665 error_setg(errp
, "Invalid parameter 'buf-size'");
1669 if (buf_size
== 0) {
1670 buf_size
= DEFAULT_MIRROR_BUF_SIZE
;
1673 if (bdrv_skip_filters(bs
) == bdrv_skip_filters(target
)) {
1674 error_setg(errp
, "Can't mirror node into itself");
1678 target_is_backing
= bdrv_chain_contains(bs
, target
);
1680 /* In the case of active commit, add dummy driver to provide consistent
1681 * reads on the top, while disabling it in the intermediate nodes, and make
1682 * the backing chain writable. */
1683 mirror_top_bs
= bdrv_new_open_driver(&bdrv_mirror_top
, filter_node_name
,
1685 if (mirror_top_bs
== NULL
) {
1688 if (!filter_node_name
) {
1689 mirror_top_bs
->implicit
= true;
1692 /* So that we can always drop this node */
1693 mirror_top_bs
->never_freeze
= true;
1695 mirror_top_bs
->total_sectors
= bs
->total_sectors
;
1696 mirror_top_bs
->supported_write_flags
= BDRV_REQ_WRITE_UNCHANGED
;
1697 mirror_top_bs
->supported_zero_flags
= BDRV_REQ_WRITE_UNCHANGED
|
1698 BDRV_REQ_NO_FALLBACK
;
1699 bs_opaque
= g_new0(MirrorBDSOpaque
, 1);
1700 mirror_top_bs
->opaque
= bs_opaque
;
1702 bs_opaque
->is_commit
= target_is_backing
;
1704 bdrv_drained_begin(bs
);
1705 ret
= bdrv_append(mirror_top_bs
, bs
, errp
);
1706 bdrv_drained_end(bs
);
1709 bdrv_unref(mirror_top_bs
);
1713 /* Make sure that the source is not resized while the job is running */
1714 s
= block_job_create(job_id
, driver
, NULL
, mirror_top_bs
,
1715 BLK_PERM_CONSISTENT_READ
,
1716 BLK_PERM_CONSISTENT_READ
| BLK_PERM_WRITE_UNCHANGED
|
1717 BLK_PERM_WRITE
, speed
,
1718 creation_flags
, cb
, opaque
, errp
);
1723 /* The block job now has a reference to this node */
1724 bdrv_unref(mirror_top_bs
);
1726 s
->mirror_top_bs
= mirror_top_bs
;
1728 /* No resize for the target either; while the mirror is still running, a
1729 * consistent read isn't necessarily possible. We could possibly allow
1730 * writes and graph modifications, though it would likely defeat the
1731 * purpose of a mirror, so leave them blocked for now.
1733 * In the case of active commit, things look a bit different, though,
1734 * because the target is an already populated backing file in active use.
1735 * We can allow anything except resize there.*/
1737 target_perms
= BLK_PERM_WRITE
;
1738 target_shared_perms
= BLK_PERM_WRITE_UNCHANGED
;
1740 if (target_is_backing
) {
1741 int64_t bs_size
, target_size
;
1742 bs_size
= bdrv_getlength(bs
);
1744 error_setg_errno(errp
, -bs_size
,
1745 "Could not inquire top image size");
1749 target_size
= bdrv_getlength(target
);
1750 if (target_size
< 0) {
1751 error_setg_errno(errp
, -target_size
,
1752 "Could not inquire base image size");
1756 if (target_size
< bs_size
) {
1757 target_perms
|= BLK_PERM_RESIZE
;
1760 target_shared_perms
|= BLK_PERM_CONSISTENT_READ
| BLK_PERM_WRITE
;
1761 } else if (bdrv_chain_contains(bs
, bdrv_skip_filters(target
))) {
1763 * We may want to allow this in the future, but it would
1764 * require taking some extra care.
1766 error_setg(errp
, "Cannot mirror to a filter on top of a node in the "
1767 "source's backing chain");
1771 s
->target
= blk_new(s
->common
.job
.aio_context
,
1772 target_perms
, target_shared_perms
);
1773 ret
= blk_insert_bs(s
->target
, target
, errp
);
1778 /* XXX: Mirror target could be a NBD server of target QEMU in the case
1779 * of non-shared block migration. To allow migration completion, we
1780 * have to allow "inactivate" of the target BB. When that happens, we
1781 * know the job is drained, and the vcpus are stopped, so no write
1782 * operation will be performed. Block layer already has assertions to
1784 blk_set_force_allow_inactivate(s
->target
);
1786 blk_set_allow_aio_context_change(s
->target
, true);
1787 blk_set_disable_request_queuing(s
->target
, true);
1789 s
->replaces
= g_strdup(replaces
);
1790 s
->on_source_error
= on_source_error
;
1791 s
->on_target_error
= on_target_error
;
1792 s
->is_none_mode
= is_none_mode
;
1793 s
->backing_mode
= backing_mode
;
1794 s
->zero_target
= zero_target
;
1795 s
->copy_mode
= copy_mode
;
1797 s
->base_overlay
= bdrv_find_overlay(bs
, base
);
1798 s
->granularity
= granularity
;
1799 s
->buf_size
= ROUND_UP(buf_size
, granularity
);
1801 if (auto_complete
) {
1802 s
->should_complete
= true;
1805 s
->dirty_bitmap
= bdrv_create_dirty_bitmap(bs
, granularity
, NULL
, errp
);
1806 if (!s
->dirty_bitmap
) {
1809 if (s
->copy_mode
== MIRROR_COPY_MODE_WRITE_BLOCKING
) {
1810 bdrv_disable_dirty_bitmap(s
->dirty_bitmap
);
1813 ret
= block_job_add_bdrv(&s
->common
, "source", bs
, 0,
1814 BLK_PERM_WRITE_UNCHANGED
| BLK_PERM_WRITE
|
1815 BLK_PERM_CONSISTENT_READ
,
1821 /* Required permissions are already taken with blk_new() */
1822 block_job_add_bdrv(&s
->common
, "target", target
, 0, BLK_PERM_ALL
,
1825 /* In commit_active_start() all intermediate nodes disappear, so
1826 * any jobs in them must be blocked */
1827 if (target_is_backing
) {
1828 BlockDriverState
*iter
, *filtered_target
;
1829 uint64_t iter_shared_perms
;
1832 * The topmost node with
1833 * bdrv_skip_filters(filtered_target) == bdrv_skip_filters(target)
1835 filtered_target
= bdrv_cow_bs(bdrv_find_overlay(bs
, target
));
1837 assert(bdrv_skip_filters(filtered_target
) ==
1838 bdrv_skip_filters(target
));
1841 * XXX BLK_PERM_WRITE needs to be allowed so we don't block
1842 * ourselves at s->base (if writes are blocked for a node, they are
1843 * also blocked for its backing file). The other options would be a
1844 * second filter driver above s->base (== target).
1846 iter_shared_perms
= BLK_PERM_WRITE_UNCHANGED
| BLK_PERM_WRITE
;
1848 for (iter
= bdrv_filter_or_cow_bs(bs
); iter
!= target
;
1849 iter
= bdrv_filter_or_cow_bs(iter
))
1851 if (iter
== filtered_target
) {
1853 * From here on, all nodes are filters on the base.
1854 * This allows us to share BLK_PERM_CONSISTENT_READ.
1856 iter_shared_perms
|= BLK_PERM_CONSISTENT_READ
;
1859 ret
= block_job_add_bdrv(&s
->common
, "intermediate node", iter
, 0,
1860 iter_shared_perms
, errp
);
1866 if (bdrv_freeze_backing_chain(mirror_top_bs
, target
, errp
) < 0) {
1871 QTAILQ_INIT(&s
->ops_in_flight
);
1873 trace_mirror_start(bs
, s
, opaque
);
1874 job_start(&s
->common
.job
);
1880 /* Make sure this BDS does not go away until we have completed the graph
1882 bdrv_ref(mirror_top_bs
);
1884 g_free(s
->replaces
);
1885 blk_unref(s
->target
);
1886 bs_opaque
->job
= NULL
;
1887 if (s
->dirty_bitmap
) {
1888 bdrv_release_dirty_bitmap(s
->dirty_bitmap
);
1890 job_early_fail(&s
->common
.job
);
1893 bs_opaque
->stop
= true;
1894 bdrv_child_refresh_perms(mirror_top_bs
, mirror_top_bs
->backing
,
1896 bdrv_replace_node(mirror_top_bs
, mirror_top_bs
->backing
->bs
, &error_abort
);
1898 bdrv_unref(mirror_top_bs
);
1903 void mirror_start(const char *job_id
, BlockDriverState
*bs
,
1904 BlockDriverState
*target
, const char *replaces
,
1905 int creation_flags
, int64_t speed
,
1906 uint32_t granularity
, int64_t buf_size
,
1907 MirrorSyncMode mode
, BlockMirrorBackingMode backing_mode
,
1909 BlockdevOnError on_source_error
,
1910 BlockdevOnError on_target_error
,
1911 bool unmap
, const char *filter_node_name
,
1912 MirrorCopyMode copy_mode
, Error
**errp
)
1915 BlockDriverState
*base
;
1917 GLOBAL_STATE_CODE();
1919 if ((mode
== MIRROR_SYNC_MODE_INCREMENTAL
) ||
1920 (mode
== MIRROR_SYNC_MODE_BITMAP
)) {
1921 error_setg(errp
, "Sync mode '%s' not supported",
1922 MirrorSyncMode_str(mode
));
1925 is_none_mode
= mode
== MIRROR_SYNC_MODE_NONE
;
1926 base
= mode
== MIRROR_SYNC_MODE_TOP
? bdrv_backing_chain_next(bs
) : NULL
;
1927 mirror_start_job(job_id
, bs
, creation_flags
, target
, replaces
,
1928 speed
, granularity
, buf_size
, backing_mode
, zero_target
,
1929 on_source_error
, on_target_error
, unmap
, NULL
, NULL
,
1930 &mirror_job_driver
, is_none_mode
, base
, false,
1931 filter_node_name
, true, copy_mode
, errp
);
1934 BlockJob
*commit_active_start(const char *job_id
, BlockDriverState
*bs
,
1935 BlockDriverState
*base
, int creation_flags
,
1936 int64_t speed
, BlockdevOnError on_error
,
1937 const char *filter_node_name
,
1938 BlockCompletionFunc
*cb
, void *opaque
,
1939 bool auto_complete
, Error
**errp
)
1941 bool base_read_only
;
1944 GLOBAL_STATE_CODE();
1946 base_read_only
= bdrv_is_read_only(base
);
1948 if (base_read_only
) {
1949 if (bdrv_reopen_set_read_only(base
, false, errp
) < 0) {
1954 job
= mirror_start_job(
1955 job_id
, bs
, creation_flags
, base
, NULL
, speed
, 0, 0,
1956 MIRROR_LEAVE_BACKING_CHAIN
, false,
1957 on_error
, on_error
, true, cb
, opaque
,
1958 &commit_active_job_driver
, false, base
, auto_complete
,
1959 filter_node_name
, false, MIRROR_COPY_MODE_BACKGROUND
,
1962 goto error_restore_flags
;
1967 error_restore_flags
:
1968 /* ignore error and errp for bdrv_reopen, because we want to propagate
1969 * the original error */
1970 if (base_read_only
) {
1971 bdrv_reopen_set_read_only(base
, true, NULL
);