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 "qemu/ratelimit.h"
25 #include "qemu/bitmap.h"
26 #include "qemu/memalign.h"
28 #define MAX_IN_FLIGHT 16
29 #define MAX_IO_BYTES (1 << 20) /* 1 Mb */
30 #define DEFAULT_MIRROR_BUF_SIZE (MAX_IN_FLIGHT * MAX_IO_BYTES)
32 /* The mirroring buffer is a list of granularity-sized chunks.
33 * Free chunks are organized in a list.
35 typedef struct MirrorBuffer
{
36 QSIMPLEQ_ENTRY(MirrorBuffer
) next
;
39 typedef struct MirrorOp MirrorOp
;
41 typedef struct MirrorBlockJob
{
44 BlockDriverState
*mirror_top_bs
;
45 BlockDriverState
*base
;
46 BlockDriverState
*base_overlay
;
48 /* The name of the graph node to replace */
50 /* The BDS to replace */
51 BlockDriverState
*to_replace
;
52 /* Used to block operations on the drive-mirror-replace target */
53 Error
*replace_blocker
;
55 BlockMirrorBackingMode backing_mode
;
56 /* Whether the target image requires explicit zero-initialization */
58 MirrorCopyMode copy_mode
;
59 BlockdevOnError on_source_error
, on_target_error
;
60 /* Set when the target is synced (dirty bitmap is clean, nothing
61 * in flight) and the job is running in active mode */
67 unsigned long *cow_bitmap
;
68 BdrvDirtyBitmap
*dirty_bitmap
;
69 BdrvDirtyBitmapIter
*dbi
;
71 QSIMPLEQ_HEAD(, MirrorBuffer
) buf_free
;
74 uint64_t last_pause_ns
;
75 unsigned long *in_flight_bitmap
;
77 int64_t bytes_in_flight
;
78 QTAILQ_HEAD(, MirrorOp
) ops_in_flight
;
81 int target_cluster_size
;
83 bool initial_zeroing_ongoing
;
84 int in_active_write_counter
;
85 int64_t active_write_bytes_in_flight
;
90 typedef struct MirrorBDSOpaque
{
102 /* The pointee is set by mirror_co_read(), mirror_co_zero(), and
103 * mirror_co_discard() before yielding for the first time */
104 int64_t *bytes_handled
;
107 bool is_active_write
;
109 CoQueue waiting_requests
;
111 MirrorOp
*waiting_for_op
;
113 QTAILQ_ENTRY(MirrorOp
) next
;
116 typedef enum MirrorMethod
{
119 MIRROR_METHOD_DISCARD
,
122 static BlockErrorAction
mirror_error_action(MirrorBlockJob
*s
, bool read
,
125 s
->actively_synced
= false;
127 return block_job_error_action(&s
->common
, s
->on_source_error
,
130 return block_job_error_action(&s
->common
, s
->on_target_error
,
135 static void coroutine_fn
mirror_wait_on_conflicts(MirrorOp
*self
,
140 uint64_t self_start_chunk
= offset
/ s
->granularity
;
141 uint64_t self_end_chunk
= DIV_ROUND_UP(offset
+ bytes
, s
->granularity
);
142 uint64_t self_nb_chunks
= self_end_chunk
- self_start_chunk
;
144 while (find_next_bit(s
->in_flight_bitmap
, self_end_chunk
,
145 self_start_chunk
) < self_end_chunk
&&
150 QTAILQ_FOREACH(op
, &s
->ops_in_flight
, next
) {
151 uint64_t op_start_chunk
= op
->offset
/ s
->granularity
;
152 uint64_t op_nb_chunks
= DIV_ROUND_UP(op
->offset
+ op
->bytes
,
160 if (ranges_overlap(self_start_chunk
, self_nb_chunks
,
161 op_start_chunk
, op_nb_chunks
))
165 * If the operation is already (indirectly) waiting for us,
166 * or will wait for us as soon as it wakes up, then just go
167 * on (instead of producing a deadlock in the former case).
169 if (op
->waiting_for_op
) {
173 self
->waiting_for_op
= op
;
176 qemu_co_queue_wait(&op
->waiting_requests
, NULL
);
179 self
->waiting_for_op
= NULL
;
188 static void coroutine_fn
mirror_iteration_done(MirrorOp
*op
, int ret
)
190 MirrorBlockJob
*s
= op
->s
;
195 trace_mirror_iteration_done(s
, op
->offset
, op
->bytes
, ret
);
198 s
->bytes_in_flight
-= op
->bytes
;
200 for (i
= 0; i
< op
->qiov
.niov
; i
++) {
201 MirrorBuffer
*buf
= (MirrorBuffer
*) iov
[i
].iov_base
;
202 QSIMPLEQ_INSERT_TAIL(&s
->buf_free
, buf
, next
);
206 chunk_num
= op
->offset
/ s
->granularity
;
207 nb_chunks
= DIV_ROUND_UP(op
->bytes
, s
->granularity
);
209 bitmap_clear(s
->in_flight_bitmap
, chunk_num
, nb_chunks
);
210 QTAILQ_REMOVE(&s
->ops_in_flight
, op
, next
);
213 bitmap_set(s
->cow_bitmap
, chunk_num
, nb_chunks
);
215 if (!s
->initial_zeroing_ongoing
) {
216 job_progress_update(&s
->common
.job
, op
->bytes
);
219 qemu_iovec_destroy(&op
->qiov
);
221 qemu_co_queue_restart_all(&op
->waiting_requests
);
225 static void coroutine_fn
mirror_write_complete(MirrorOp
*op
, int ret
)
227 MirrorBlockJob
*s
= op
->s
;
230 BlockErrorAction action
;
232 bdrv_set_dirty_bitmap(s
->dirty_bitmap
, op
->offset
, op
->bytes
);
233 action
= mirror_error_action(s
, false, -ret
);
234 if (action
== BLOCK_ERROR_ACTION_REPORT
&& s
->ret
>= 0) {
239 mirror_iteration_done(op
, ret
);
242 static void coroutine_fn
mirror_read_complete(MirrorOp
*op
, int ret
)
244 MirrorBlockJob
*s
= op
->s
;
247 BlockErrorAction action
;
249 bdrv_set_dirty_bitmap(s
->dirty_bitmap
, op
->offset
, op
->bytes
);
250 action
= mirror_error_action(s
, true, -ret
);
251 if (action
== BLOCK_ERROR_ACTION_REPORT
&& s
->ret
>= 0) {
255 mirror_iteration_done(op
, ret
);
259 ret
= blk_co_pwritev(s
->target
, op
->offset
, op
->qiov
.size
, &op
->qiov
, 0);
260 mirror_write_complete(op
, ret
);
263 /* Clip bytes relative to offset to not exceed end-of-file */
264 static inline int64_t mirror_clip_bytes(MirrorBlockJob
*s
,
268 return MIN(bytes
, s
->bdev_length
- offset
);
271 /* Round offset and/or bytes to target cluster if COW is needed, and
272 * return the offset of the adjusted tail against original. */
273 static int coroutine_fn
mirror_cow_align(MirrorBlockJob
*s
, int64_t *offset
,
278 int64_t align_offset
= *offset
;
279 int64_t align_bytes
= *bytes
;
280 int max_bytes
= s
->granularity
* s
->max_iov
;
282 need_cow
= !test_bit(*offset
/ s
->granularity
, s
->cow_bitmap
);
283 need_cow
|= !test_bit((*offset
+ *bytes
- 1) / s
->granularity
,
286 bdrv_round_to_clusters(blk_bs(s
->target
), *offset
, *bytes
,
287 &align_offset
, &align_bytes
);
290 if (align_bytes
> max_bytes
) {
291 align_bytes
= max_bytes
;
293 align_bytes
= QEMU_ALIGN_DOWN(align_bytes
, s
->target_cluster_size
);
296 /* Clipping may result in align_bytes unaligned to chunk boundary, but
297 * that doesn't matter because it's already the end of source image. */
298 align_bytes
= mirror_clip_bytes(s
, align_offset
, align_bytes
);
300 ret
= align_offset
+ align_bytes
- (*offset
+ *bytes
);
301 *offset
= align_offset
;
302 *bytes
= align_bytes
;
307 static inline void coroutine_fn
308 mirror_wait_for_free_in_flight_slot(MirrorBlockJob
*s
)
312 QTAILQ_FOREACH(op
, &s
->ops_in_flight
, next
) {
314 * Do not wait on pseudo ops, because it may in turn wait on
315 * some other operation to start, which may in fact be the
316 * caller of this function. Since there is only one pseudo op
317 * at any given time, we will always find some real operation
319 * Also, do not wait on active operations, because they do not
320 * use up in-flight slots.
322 if (!op
->is_pseudo_op
&& op
->is_in_flight
&& !op
->is_active_write
) {
323 qemu_co_queue_wait(&op
->waiting_requests
, NULL
);
330 /* Perform a mirror copy operation.
332 * *op->bytes_handled is set to the number of bytes copied after and
333 * including offset, excluding any bytes copied prior to offset due
334 * to alignment. This will be op->bytes if no alignment is necessary,
335 * or (new_end - op->offset) if the tail is rounded up or down due to
336 * alignment or buffer limit.
338 static void coroutine_fn
mirror_co_read(void *opaque
)
340 MirrorOp
*op
= opaque
;
341 MirrorBlockJob
*s
= op
->s
;
346 max_bytes
= s
->granularity
* s
->max_iov
;
348 /* We can only handle as much as buf_size at a time. */
349 op
->bytes
= MIN(s
->buf_size
, MIN(max_bytes
, op
->bytes
));
351 assert(op
->bytes
< BDRV_REQUEST_MAX_BYTES
);
352 *op
->bytes_handled
= op
->bytes
;
355 *op
->bytes_handled
+= mirror_cow_align(s
, &op
->offset
, &op
->bytes
);
357 /* Cannot exceed BDRV_REQUEST_MAX_BYTES + INT_MAX */
358 assert(*op
->bytes_handled
<= UINT_MAX
);
359 assert(op
->bytes
<= s
->buf_size
);
360 /* The offset is granularity-aligned because:
361 * 1) Caller passes in aligned values;
362 * 2) mirror_cow_align is used only when target cluster is larger. */
363 assert(QEMU_IS_ALIGNED(op
->offset
, s
->granularity
));
364 /* The range is sector-aligned, since bdrv_getlength() rounds up. */
365 assert(QEMU_IS_ALIGNED(op
->bytes
, BDRV_SECTOR_SIZE
));
366 nb_chunks
= DIV_ROUND_UP(op
->bytes
, s
->granularity
);
368 while (s
->buf_free_count
< nb_chunks
) {
369 trace_mirror_yield_in_flight(s
, op
->offset
, s
->in_flight
);
370 mirror_wait_for_free_in_flight_slot(s
);
373 /* Now make a QEMUIOVector taking enough granularity-sized chunks
376 qemu_iovec_init(&op
->qiov
, nb_chunks
);
377 while (nb_chunks
-- > 0) {
378 MirrorBuffer
*buf
= QSIMPLEQ_FIRST(&s
->buf_free
);
379 size_t remaining
= op
->bytes
- op
->qiov
.size
;
381 QSIMPLEQ_REMOVE_HEAD(&s
->buf_free
, next
);
383 qemu_iovec_add(&op
->qiov
, buf
, MIN(s
->granularity
, remaining
));
386 /* Copy the dirty cluster. */
388 s
->bytes_in_flight
+= op
->bytes
;
389 op
->is_in_flight
= true;
390 trace_mirror_one_iteration(s
, op
->offset
, op
->bytes
);
392 WITH_GRAPH_RDLOCK_GUARD() {
393 ret
= bdrv_co_preadv(s
->mirror_top_bs
->backing
, op
->offset
, op
->bytes
,
396 mirror_read_complete(op
, ret
);
399 static void coroutine_fn
mirror_co_zero(void *opaque
)
401 MirrorOp
*op
= opaque
;
405 op
->s
->bytes_in_flight
+= op
->bytes
;
406 *op
->bytes_handled
= op
->bytes
;
407 op
->is_in_flight
= true;
409 ret
= blk_co_pwrite_zeroes(op
->s
->target
, op
->offset
, op
->bytes
,
410 op
->s
->unmap
? BDRV_REQ_MAY_UNMAP
: 0);
411 mirror_write_complete(op
, ret
);
414 static void coroutine_fn
mirror_co_discard(void *opaque
)
416 MirrorOp
*op
= opaque
;
420 op
->s
->bytes_in_flight
+= op
->bytes
;
421 *op
->bytes_handled
= op
->bytes
;
422 op
->is_in_flight
= true;
424 ret
= blk_co_pdiscard(op
->s
->target
, op
->offset
, op
->bytes
);
425 mirror_write_complete(op
, ret
);
428 static unsigned mirror_perform(MirrorBlockJob
*s
, int64_t offset
,
429 unsigned bytes
, MirrorMethod mirror_method
)
433 int64_t bytes_handled
= -1;
435 op
= g_new(MirrorOp
, 1);
440 .bytes_handled
= &bytes_handled
,
442 qemu_co_queue_init(&op
->waiting_requests
);
444 switch (mirror_method
) {
445 case MIRROR_METHOD_COPY
:
446 co
= qemu_coroutine_create(mirror_co_read
, op
);
448 case MIRROR_METHOD_ZERO
:
449 co
= qemu_coroutine_create(mirror_co_zero
, op
);
451 case MIRROR_METHOD_DISCARD
:
452 co
= qemu_coroutine_create(mirror_co_discard
, op
);
459 QTAILQ_INSERT_TAIL(&s
->ops_in_flight
, op
, next
);
460 qemu_coroutine_enter(co
);
461 /* At this point, ownership of op has been moved to the coroutine
462 * and the object may already be freed */
464 /* Assert that this value has been set */
465 assert(bytes_handled
>= 0);
467 /* Same assertion as in mirror_co_read() (and for mirror_co_read()
468 * and mirror_co_discard(), bytes_handled == op->bytes, which
469 * is the @bytes parameter given to this function) */
470 assert(bytes_handled
<= UINT_MAX
);
471 return bytes_handled
;
474 static uint64_t coroutine_fn
mirror_iteration(MirrorBlockJob
*s
)
476 BlockDriverState
*source
= s
->mirror_top_bs
->backing
->bs
;
479 uint64_t delay_ns
= 0, ret
= 0;
480 /* At least the first dirty chunk is mirrored in one iteration. */
482 bool write_zeroes_ok
= bdrv_can_write_zeroes_with_unmap(blk_bs(s
->target
));
483 int max_io_bytes
= MAX(s
->buf_size
/ MAX_IN_FLIGHT
, MAX_IO_BYTES
);
485 bdrv_dirty_bitmap_lock(s
->dirty_bitmap
);
486 offset
= bdrv_dirty_iter_next(s
->dbi
);
488 bdrv_set_dirty_iter(s
->dbi
, 0);
489 offset
= bdrv_dirty_iter_next(s
->dbi
);
490 trace_mirror_restart_iter(s
, bdrv_get_dirty_count(s
->dirty_bitmap
));
493 bdrv_dirty_bitmap_unlock(s
->dirty_bitmap
);
496 * Wait for concurrent requests to @offset. The next loop will limit the
497 * copied area based on in_flight_bitmap so we only copy an area that does
498 * not overlap with concurrent in-flight requests. Still, we would like to
499 * copy something, so wait until there are at least no more requests to the
500 * very beginning of the area.
502 mirror_wait_on_conflicts(NULL
, s
, offset
, 1);
504 job_pause_point(&s
->common
.job
);
506 /* Find the number of consective dirty chunks following the first dirty
507 * one, and wait for in flight requests in them. */
508 bdrv_dirty_bitmap_lock(s
->dirty_bitmap
);
509 while (nb_chunks
* s
->granularity
< s
->buf_size
) {
511 int64_t next_offset
= offset
+ nb_chunks
* s
->granularity
;
512 int64_t next_chunk
= next_offset
/ s
->granularity
;
513 if (next_offset
>= s
->bdev_length
||
514 !bdrv_dirty_bitmap_get_locked(s
->dirty_bitmap
, next_offset
)) {
517 if (test_bit(next_chunk
, s
->in_flight_bitmap
)) {
521 next_dirty
= bdrv_dirty_iter_next(s
->dbi
);
522 if (next_dirty
> next_offset
|| next_dirty
< 0) {
523 /* The bitmap iterator's cache is stale, refresh it */
524 bdrv_set_dirty_iter(s
->dbi
, next_offset
);
525 next_dirty
= bdrv_dirty_iter_next(s
->dbi
);
527 assert(next_dirty
== next_offset
);
531 /* Clear dirty bits before querying the block status, because
532 * calling bdrv_block_status_above could yield - if some blocks are
533 * marked dirty in this window, we need to know.
535 bdrv_reset_dirty_bitmap_locked(s
->dirty_bitmap
, offset
,
536 nb_chunks
* s
->granularity
);
537 bdrv_dirty_bitmap_unlock(s
->dirty_bitmap
);
539 /* Before claiming an area in the in-flight bitmap, we have to
540 * create a MirrorOp for it so that conflicting requests can wait
541 * for it. mirror_perform() will create the real MirrorOps later,
542 * for now we just create a pseudo operation that will wake up all
543 * conflicting requests once all real operations have been
545 pseudo_op
= g_new(MirrorOp
, 1);
546 *pseudo_op
= (MirrorOp
){
548 .bytes
= nb_chunks
* s
->granularity
,
549 .is_pseudo_op
= true,
551 qemu_co_queue_init(&pseudo_op
->waiting_requests
);
552 QTAILQ_INSERT_TAIL(&s
->ops_in_flight
, pseudo_op
, next
);
554 bitmap_set(s
->in_flight_bitmap
, offset
/ s
->granularity
, nb_chunks
);
555 while (nb_chunks
> 0 && offset
< s
->bdev_length
) {
558 int64_t io_bytes_acct
;
559 MirrorMethod mirror_method
= MIRROR_METHOD_COPY
;
561 assert(!(offset
% s
->granularity
));
562 WITH_GRAPH_RDLOCK_GUARD() {
563 ret
= bdrv_block_status_above(source
, NULL
, offset
,
564 nb_chunks
* s
->granularity
,
565 &io_bytes
, NULL
, NULL
);
568 io_bytes
= MIN(nb_chunks
* s
->granularity
, max_io_bytes
);
569 } else if (ret
& BDRV_BLOCK_DATA
) {
570 io_bytes
= MIN(io_bytes
, max_io_bytes
);
573 io_bytes
-= io_bytes
% s
->granularity
;
574 if (io_bytes
< s
->granularity
) {
575 io_bytes
= s
->granularity
;
576 } else if (ret
>= 0 && !(ret
& BDRV_BLOCK_DATA
)) {
577 int64_t target_offset
;
578 int64_t target_bytes
;
579 WITH_GRAPH_RDLOCK_GUARD() {
580 bdrv_round_to_clusters(blk_bs(s
->target
), offset
, io_bytes
,
581 &target_offset
, &target_bytes
);
583 if (target_offset
== offset
&&
584 target_bytes
== io_bytes
) {
585 mirror_method
= ret
& BDRV_BLOCK_ZERO
?
587 MIRROR_METHOD_DISCARD
;
591 while (s
->in_flight
>= MAX_IN_FLIGHT
) {
592 trace_mirror_yield_in_flight(s
, offset
, s
->in_flight
);
593 mirror_wait_for_free_in_flight_slot(s
);
601 io_bytes
= mirror_clip_bytes(s
, offset
, io_bytes
);
602 io_bytes
= mirror_perform(s
, offset
, io_bytes
, mirror_method
);
603 if (mirror_method
!= MIRROR_METHOD_COPY
&& write_zeroes_ok
) {
606 io_bytes_acct
= io_bytes
;
610 nb_chunks
-= DIV_ROUND_UP(io_bytes
, s
->granularity
);
611 delay_ns
= block_job_ratelimit_get_delay(&s
->common
, io_bytes_acct
);
616 QTAILQ_REMOVE(&s
->ops_in_flight
, pseudo_op
, next
);
617 qemu_co_queue_restart_all(&pseudo_op
->waiting_requests
);
623 static void mirror_free_init(MirrorBlockJob
*s
)
625 int granularity
= s
->granularity
;
626 size_t buf_size
= s
->buf_size
;
627 uint8_t *buf
= s
->buf
;
629 assert(s
->buf_free_count
== 0);
630 QSIMPLEQ_INIT(&s
->buf_free
);
631 while (buf_size
!= 0) {
632 MirrorBuffer
*cur
= (MirrorBuffer
*)buf
;
633 QSIMPLEQ_INSERT_TAIL(&s
->buf_free
, cur
, next
);
635 buf_size
-= granularity
;
640 /* This is also used for the .pause callback. There is no matching
641 * mirror_resume() because mirror_run() will begin iterating again
642 * when the job is resumed.
644 static void coroutine_fn
mirror_wait_for_all_io(MirrorBlockJob
*s
)
646 while (s
->in_flight
> 0) {
647 mirror_wait_for_free_in_flight_slot(s
);
652 * mirror_exit_common: handle both abort() and prepare() cases.
653 * for .prepare, returns 0 on success and -errno on failure.
654 * for .abort cases, denoted by abort = true, MUST return 0.
656 static int mirror_exit_common(Job
*job
)
658 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
.job
);
659 BlockJob
*bjob
= &s
->common
;
660 MirrorBDSOpaque
*bs_opaque
;
661 AioContext
*replace_aio_context
= NULL
;
662 BlockDriverState
*src
;
663 BlockDriverState
*target_bs
;
664 BlockDriverState
*mirror_top_bs
;
665 Error
*local_err
= NULL
;
666 bool abort
= job
->ret
< 0;
674 mirror_top_bs
= s
->mirror_top_bs
;
675 bs_opaque
= mirror_top_bs
->opaque
;
676 src
= mirror_top_bs
->backing
->bs
;
677 target_bs
= blk_bs(s
->target
);
679 if (bdrv_chain_contains(src
, target_bs
)) {
680 bdrv_unfreeze_backing_chain(mirror_top_bs
, target_bs
);
683 bdrv_release_dirty_bitmap(s
->dirty_bitmap
);
685 /* Make sure that the source BDS doesn't go away during bdrv_replace_node,
686 * before we can call bdrv_drained_end */
688 bdrv_ref(mirror_top_bs
);
692 * Remove target parent that still uses BLK_PERM_WRITE/RESIZE before
693 * inserting target_bs at s->to_replace, where we might not be able to get
696 blk_unref(s
->target
);
699 /* We don't access the source any more. Dropping any WRITE/RESIZE is
700 * required before it could become a backing file of target_bs. Not having
701 * these permissions any more means that we can't allow any new requests on
702 * mirror_top_bs from now on, so keep it drained. */
703 bdrv_drained_begin(mirror_top_bs
);
704 bs_opaque
->stop
= true;
705 bdrv_child_refresh_perms(mirror_top_bs
, mirror_top_bs
->backing
,
707 if (!abort
&& s
->backing_mode
== MIRROR_SOURCE_BACKING_CHAIN
) {
708 BlockDriverState
*backing
= s
->is_none_mode
? src
: s
->base
;
709 BlockDriverState
*unfiltered_target
= bdrv_skip_filters(target_bs
);
711 if (bdrv_cow_bs(unfiltered_target
) != backing
) {
712 bdrv_set_backing_hd(unfiltered_target
, backing
, &local_err
);
714 error_report_err(local_err
);
719 } else if (!abort
&& s
->backing_mode
== MIRROR_OPEN_BACKING_CHAIN
) {
720 assert(!bdrv_backing_chain_next(target_bs
));
721 ret
= bdrv_open_backing_file(bdrv_skip_filters(target_bs
), NULL
,
722 "backing", &local_err
);
724 error_report_err(local_err
);
730 replace_aio_context
= bdrv_get_aio_context(s
->to_replace
);
731 aio_context_acquire(replace_aio_context
);
734 if (s
->should_complete
&& !abort
) {
735 BlockDriverState
*to_replace
= s
->to_replace
?: src
;
736 bool ro
= bdrv_is_read_only(to_replace
);
738 if (ro
!= bdrv_is_read_only(target_bs
)) {
739 bdrv_reopen_set_read_only(target_bs
, ro
, NULL
);
742 /* The mirror job has no requests in flight any more, but we need to
743 * drain potential other users of the BDS before changing the graph. */
745 bdrv_drained_begin(target_bs
);
747 * Cannot use check_to_replace_node() here, because that would
748 * check for an op blocker on @to_replace, and we have our own
751 if (bdrv_recurse_can_replace(src
, to_replace
)) {
752 bdrv_replace_node(to_replace
, target_bs
, &local_err
);
754 error_setg(&local_err
, "Can no longer replace '%s' by '%s', "
755 "because it can no longer be guaranteed that doing so "
756 "would not lead to an abrupt change of visible data",
757 to_replace
->node_name
, target_bs
->node_name
);
759 bdrv_drained_end(target_bs
);
761 error_report_err(local_err
);
766 bdrv_op_unblock_all(s
->to_replace
, s
->replace_blocker
);
767 error_free(s
->replace_blocker
);
768 bdrv_unref(s
->to_replace
);
770 if (replace_aio_context
) {
771 aio_context_release(replace_aio_context
);
774 bdrv_unref(target_bs
);
777 * Remove the mirror filter driver from the graph. Before this, get rid of
778 * the blockers on the intermediate nodes so that the resulting state is
781 block_job_remove_all_bdrv(bjob
);
782 bdrv_replace_node(mirror_top_bs
, mirror_top_bs
->backing
->bs
, &error_abort
);
784 bs_opaque
->job
= NULL
;
786 bdrv_drained_end(src
);
787 bdrv_drained_end(mirror_top_bs
);
789 bdrv_unref(mirror_top_bs
);
795 static int mirror_prepare(Job
*job
)
797 return mirror_exit_common(job
);
800 static void mirror_abort(Job
*job
)
802 int ret
= mirror_exit_common(job
);
806 static void coroutine_fn
mirror_throttle(MirrorBlockJob
*s
)
808 int64_t now
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
);
810 if (now
- s
->last_pause_ns
> BLOCK_JOB_SLICE_TIME
) {
811 s
->last_pause_ns
= now
;
812 job_sleep_ns(&s
->common
.job
, 0);
814 job_pause_point(&s
->common
.job
);
818 static int coroutine_fn
mirror_dirty_init(MirrorBlockJob
*s
)
821 BlockDriverState
*bs
= s
->mirror_top_bs
->backing
->bs
;
822 BlockDriverState
*target_bs
= blk_bs(s
->target
);
826 if (s
->zero_target
) {
827 if (!bdrv_can_write_zeroes_with_unmap(target_bs
)) {
828 bdrv_set_dirty_bitmap(s
->dirty_bitmap
, 0, s
->bdev_length
);
832 s
->initial_zeroing_ongoing
= true;
833 for (offset
= 0; offset
< s
->bdev_length
; ) {
834 int bytes
= MIN(s
->bdev_length
- offset
,
835 QEMU_ALIGN_DOWN(INT_MAX
, s
->granularity
));
839 if (job_is_cancelled(&s
->common
.job
)) {
840 s
->initial_zeroing_ongoing
= false;
844 if (s
->in_flight
>= MAX_IN_FLIGHT
) {
845 trace_mirror_yield(s
, UINT64_MAX
, s
->buf_free_count
,
847 mirror_wait_for_free_in_flight_slot(s
);
851 mirror_perform(s
, offset
, bytes
, MIRROR_METHOD_ZERO
);
855 mirror_wait_for_all_io(s
);
856 s
->initial_zeroing_ongoing
= false;
859 /* First part, loop on the sectors and initialize the dirty bitmap. */
860 for (offset
= 0; offset
< s
->bdev_length
; ) {
861 /* Just to make sure we are not exceeding int limit. */
862 int bytes
= MIN(s
->bdev_length
- offset
,
863 QEMU_ALIGN_DOWN(INT_MAX
, s
->granularity
));
867 if (job_is_cancelled(&s
->common
.job
)) {
871 WITH_GRAPH_RDLOCK_GUARD() {
872 ret
= bdrv_is_allocated_above(bs
, s
->base_overlay
, true, offset
,
881 bdrv_set_dirty_bitmap(s
->dirty_bitmap
, offset
, count
);
888 /* Called when going out of the streaming phase to flush the bulk of the
889 * data to the medium, or just before completing.
891 static int coroutine_fn
mirror_flush(MirrorBlockJob
*s
)
893 int ret
= blk_co_flush(s
->target
);
895 if (mirror_error_action(s
, false, -ret
) == BLOCK_ERROR_ACTION_REPORT
) {
902 static int coroutine_fn
mirror_run(Job
*job
, Error
**errp
)
904 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
.job
);
905 BlockDriverState
*bs
= s
->mirror_top_bs
->backing
->bs
;
906 MirrorBDSOpaque
*mirror_top_opaque
= s
->mirror_top_bs
->opaque
;
907 BlockDriverState
*target_bs
= blk_bs(s
->target
);
908 bool need_drain
= true;
909 BlockDeviceIoStatus iostatus
;
911 int64_t target_length
;
913 char backing_filename
[2]; /* we only need 2 characters because we are only
914 checking for a NULL string */
917 if (job_is_cancelled(&s
->common
.job
)) {
921 bdrv_graph_co_rdlock();
922 s
->bdev_length
= bdrv_co_getlength(bs
);
923 bdrv_graph_co_rdunlock();
925 if (s
->bdev_length
< 0) {
926 ret
= s
->bdev_length
;
930 target_length
= blk_co_getlength(s
->target
);
931 if (target_length
< 0) {
936 /* Active commit must resize the base image if its size differs from the
938 if (s
->base
== blk_bs(s
->target
)) {
939 if (s
->bdev_length
> target_length
) {
940 ret
= blk_co_truncate(s
->target
, s
->bdev_length
, false,
941 PREALLOC_MODE_OFF
, 0, NULL
);
946 } else if (s
->bdev_length
!= target_length
) {
947 error_setg(errp
, "Source and target image have different sizes");
952 if (s
->bdev_length
== 0) {
953 /* Transition to the READY state and wait for complete. */
954 job_transition_to_ready(&s
->common
.job
);
955 s
->actively_synced
= true;
956 while (!job_cancel_requested(&s
->common
.job
) && !s
->should_complete
) {
957 job_yield(&s
->common
.job
);
962 length
= DIV_ROUND_UP(s
->bdev_length
, s
->granularity
);
963 s
->in_flight_bitmap
= bitmap_new(length
);
965 /* If we have no backing file yet in the destination, we cannot let
966 * the destination do COW. Instead, we copy sectors around the
967 * dirty data if needed. We need a bitmap to do that.
969 bdrv_get_backing_filename(target_bs
, backing_filename
,
970 sizeof(backing_filename
));
971 bdrv_graph_co_rdlock();
972 if (!bdrv_co_get_info(target_bs
, &bdi
) && bdi
.cluster_size
) {
973 s
->target_cluster_size
= bdi
.cluster_size
;
975 s
->target_cluster_size
= BDRV_SECTOR_SIZE
;
977 bdrv_graph_co_rdunlock();
978 if (backing_filename
[0] && !bdrv_backing_chain_next(target_bs
) &&
979 s
->granularity
< s
->target_cluster_size
) {
980 s
->buf_size
= MAX(s
->buf_size
, s
->target_cluster_size
);
981 s
->cow_bitmap
= bitmap_new(length
);
983 s
->max_iov
= MIN(bs
->bl
.max_iov
, target_bs
->bl
.max_iov
);
985 s
->buf
= qemu_try_blockalign(bs
, s
->buf_size
);
986 if (s
->buf
== NULL
) {
993 s
->last_pause_ns
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
);
994 if (!s
->is_none_mode
) {
995 ret
= mirror_dirty_init(s
);
996 if (ret
< 0 || job_is_cancelled(&s
->common
.job
)) {
1002 * Only now the job is fully initialised and mirror_top_bs should start
1005 mirror_top_opaque
->job
= s
;
1008 s
->dbi
= bdrv_dirty_iter_new(s
->dirty_bitmap
);
1010 uint64_t delay_ns
= 0;
1012 bool should_complete
;
1016 goto immediate_exit
;
1019 job_pause_point(&s
->common
.job
);
1021 if (job_is_cancelled(&s
->common
.job
)) {
1023 goto immediate_exit
;
1026 cnt
= bdrv_get_dirty_count(s
->dirty_bitmap
);
1027 /* cnt is the number of dirty bytes remaining and s->bytes_in_flight is
1028 * the number of bytes currently being processed; together those are
1029 * the current remaining operation length */
1030 job_progress_set_remaining(&s
->common
.job
,
1031 s
->bytes_in_flight
+ cnt
+
1032 s
->active_write_bytes_in_flight
);
1034 /* Note that even when no rate limit is applied we need to yield
1035 * periodically with no pending I/O so that bdrv_drain_all() returns.
1036 * We do so every BLKOCK_JOB_SLICE_TIME nanoseconds, or when there is
1037 * an error, or when the source is clean, whichever comes first. */
1038 delta
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
) - s
->last_pause_ns
;
1039 WITH_JOB_LOCK_GUARD() {
1040 iostatus
= s
->common
.iostatus
;
1042 if (delta
< BLOCK_JOB_SLICE_TIME
&&
1043 iostatus
== BLOCK_DEVICE_IO_STATUS_OK
) {
1044 if (s
->in_flight
>= MAX_IN_FLIGHT
|| s
->buf_free_count
== 0 ||
1045 (cnt
== 0 && s
->in_flight
> 0)) {
1046 trace_mirror_yield(s
, cnt
, s
->buf_free_count
, s
->in_flight
);
1047 mirror_wait_for_free_in_flight_slot(s
);
1049 } else if (cnt
!= 0) {
1050 delay_ns
= mirror_iteration(s
);
1054 should_complete
= false;
1055 if (s
->in_flight
== 0 && cnt
== 0) {
1056 trace_mirror_before_flush(s
);
1057 if (!job_is_ready(&s
->common
.job
)) {
1058 if (mirror_flush(s
) < 0) {
1059 /* Go check s->ret. */
1062 /* We're out of the streaming phase. From now on, if the job
1063 * is cancelled we will actually complete all pending I/O and
1064 * report completion. This way, block-job-cancel will leave
1065 * the target in a consistent state.
1067 job_transition_to_ready(&s
->common
.job
);
1068 if (s
->copy_mode
!= MIRROR_COPY_MODE_BACKGROUND
) {
1069 s
->actively_synced
= true;
1073 should_complete
= s
->should_complete
||
1074 job_cancel_requested(&s
->common
.job
);
1075 cnt
= bdrv_get_dirty_count(s
->dirty_bitmap
);
1078 if (cnt
== 0 && should_complete
) {
1079 /* The dirty bitmap is not updated while operations are pending.
1080 * If we're about to exit, wait for pending operations before
1081 * calling bdrv_get_dirty_count(bs), or we may exit while the
1082 * source has dirty data to copy!
1084 * Note that I/O can be submitted by the guest while
1085 * mirror_populate runs, so pause it now. Before deciding
1086 * whether to switch to target check one last time if I/O has
1087 * come in the meanwhile, and if not flush the data to disk.
1089 trace_mirror_before_drain(s
, cnt
);
1092 bdrv_drained_begin(bs
);
1094 /* Must be zero because we are drained */
1095 assert(s
->in_active_write_counter
== 0);
1097 cnt
= bdrv_get_dirty_count(s
->dirty_bitmap
);
1098 if (cnt
> 0 || mirror_flush(s
) < 0) {
1099 bdrv_drained_end(bs
);
1100 s
->in_drain
= false;
1104 /* The two disks are in sync. Exit and report successful
1107 assert(QLIST_EMPTY(&bs
->tracked_requests
));
1112 if (job_is_ready(&s
->common
.job
) && !should_complete
) {
1113 delay_ns
= (s
->in_flight
== 0 &&
1114 cnt
== 0 ? BLOCK_JOB_SLICE_TIME
: 0);
1116 trace_mirror_before_sleep(s
, cnt
, job_is_ready(&s
->common
.job
),
1118 job_sleep_ns(&s
->common
.job
, delay_ns
);
1119 s
->last_pause_ns
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
);
1123 if (s
->in_flight
> 0) {
1124 /* We get here only if something went wrong. Either the job failed,
1125 * or it was cancelled prematurely so that we do not guarantee that
1126 * the target is a copy of the source.
1128 assert(ret
< 0 || job_is_cancelled(&s
->common
.job
));
1130 mirror_wait_for_all_io(s
);
1133 assert(s
->in_flight
== 0);
1135 g_free(s
->cow_bitmap
);
1136 g_free(s
->in_flight_bitmap
);
1137 bdrv_dirty_iter_free(s
->dbi
);
1141 bdrv_drained_begin(bs
);
1147 static void mirror_complete(Job
*job
, Error
**errp
)
1149 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
.job
);
1151 if (!job_is_ready(job
)) {
1152 error_setg(errp
, "The active block job '%s' cannot be completed",
1157 /* block all operations on to_replace bs */
1159 AioContext
*replace_aio_context
;
1161 s
->to_replace
= bdrv_find_node(s
->replaces
);
1162 if (!s
->to_replace
) {
1163 error_setg(errp
, "Node name '%s' not found", s
->replaces
);
1167 replace_aio_context
= bdrv_get_aio_context(s
->to_replace
);
1168 aio_context_acquire(replace_aio_context
);
1170 /* TODO Translate this into child freeze system. */
1171 error_setg(&s
->replace_blocker
,
1172 "block device is in use by block-job-complete");
1173 bdrv_op_block_all(s
->to_replace
, s
->replace_blocker
);
1174 bdrv_ref(s
->to_replace
);
1176 aio_context_release(replace_aio_context
);
1179 s
->should_complete
= true;
1181 /* If the job is paused, it will be re-entered when it is resumed */
1182 WITH_JOB_LOCK_GUARD() {
1184 job_enter_cond_locked(job
, NULL
);
1189 static void coroutine_fn
mirror_pause(Job
*job
)
1191 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
.job
);
1193 mirror_wait_for_all_io(s
);
1196 static bool mirror_drained_poll(BlockJob
*job
)
1198 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
1200 /* If the job isn't paused nor cancelled, we can't be sure that it won't
1201 * issue more requests. We make an exception if we've reached this point
1202 * from one of our own drain sections, to avoid a deadlock waiting for
1205 WITH_JOB_LOCK_GUARD() {
1206 if (!s
->common
.job
.paused
&& !job_is_cancelled_locked(&job
->job
)
1212 return !!s
->in_flight
;
1215 static bool mirror_cancel(Job
*job
, bool force
)
1217 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
.job
);
1218 BlockDriverState
*target
= blk_bs(s
->target
);
1221 * Before the job is READY, we treat any cancellation like a
1222 * force-cancellation.
1224 force
= force
|| !job_is_ready(job
);
1227 bdrv_cancel_in_flight(target
);
1232 static bool commit_active_cancel(Job
*job
, bool force
)
1234 /* Same as above in mirror_cancel() */
1235 return force
|| !job_is_ready(job
);
1238 static const BlockJobDriver mirror_job_driver
= {
1240 .instance_size
= sizeof(MirrorBlockJob
),
1241 .job_type
= JOB_TYPE_MIRROR
,
1242 .free
= block_job_free
,
1243 .user_resume
= block_job_user_resume
,
1245 .prepare
= mirror_prepare
,
1246 .abort
= mirror_abort
,
1247 .pause
= mirror_pause
,
1248 .complete
= mirror_complete
,
1249 .cancel
= mirror_cancel
,
1251 .drained_poll
= mirror_drained_poll
,
1254 static const BlockJobDriver commit_active_job_driver
= {
1256 .instance_size
= sizeof(MirrorBlockJob
),
1257 .job_type
= JOB_TYPE_COMMIT
,
1258 .free
= block_job_free
,
1259 .user_resume
= block_job_user_resume
,
1261 .prepare
= mirror_prepare
,
1262 .abort
= mirror_abort
,
1263 .pause
= mirror_pause
,
1264 .complete
= mirror_complete
,
1265 .cancel
= commit_active_cancel
,
1267 .drained_poll
= mirror_drained_poll
,
1270 static void coroutine_fn
1271 do_sync_target_write(MirrorBlockJob
*job
, MirrorMethod method
,
1272 uint64_t offset
, uint64_t bytes
,
1273 QEMUIOVector
*qiov
, int flags
)
1276 size_t qiov_offset
= 0;
1277 int64_t bitmap_offset
, bitmap_end
;
1279 if (!QEMU_IS_ALIGNED(offset
, job
->granularity
) &&
1280 bdrv_dirty_bitmap_get(job
->dirty_bitmap
, offset
))
1283 * Dirty unaligned padding: ignore it.
1286 * 1. If we copy it, we can't reset corresponding bit in
1287 * dirty_bitmap as there may be some "dirty" bytes still not
1289 * 2. It's already dirty, so skipping it we don't diverge mirror
1292 * Note, that because of this, guest write may have no contribution
1293 * into mirror converge, but that's not bad, as we have background
1294 * process of mirroring. If under some bad circumstances (high guest
1295 * IO load) background process starve, we will not converge anyway,
1296 * even if each write will contribute, as guest is not guaranteed to
1297 * rewrite the whole disk.
1299 qiov_offset
= QEMU_ALIGN_UP(offset
, job
->granularity
) - offset
;
1300 if (bytes
<= qiov_offset
) {
1301 /* nothing to do after shrink */
1304 offset
+= qiov_offset
;
1305 bytes
-= qiov_offset
;
1308 if (!QEMU_IS_ALIGNED(offset
+ bytes
, job
->granularity
) &&
1309 bdrv_dirty_bitmap_get(job
->dirty_bitmap
, offset
+ bytes
- 1))
1311 uint64_t tail
= (offset
+ bytes
) % job
->granularity
;
1313 if (bytes
<= tail
) {
1314 /* nothing to do after shrink */
1321 * Tails are either clean or shrunk, so for bitmap resetting
1322 * we safely align the range down.
1324 bitmap_offset
= QEMU_ALIGN_UP(offset
, job
->granularity
);
1325 bitmap_end
= QEMU_ALIGN_DOWN(offset
+ bytes
, job
->granularity
);
1326 if (bitmap_offset
< bitmap_end
) {
1327 bdrv_reset_dirty_bitmap(job
->dirty_bitmap
, bitmap_offset
,
1328 bitmap_end
- bitmap_offset
);
1331 job_progress_increase_remaining(&job
->common
.job
, bytes
);
1332 job
->active_write_bytes_in_flight
+= bytes
;
1335 case MIRROR_METHOD_COPY
:
1336 ret
= blk_co_pwritev_part(job
->target
, offset
, bytes
,
1337 qiov
, qiov_offset
, flags
);
1340 case MIRROR_METHOD_ZERO
:
1342 ret
= blk_co_pwrite_zeroes(job
->target
, offset
, bytes
, flags
);
1345 case MIRROR_METHOD_DISCARD
:
1347 ret
= blk_co_pdiscard(job
->target
, offset
, bytes
);
1354 job
->active_write_bytes_in_flight
-= bytes
;
1356 job_progress_update(&job
->common
.job
, bytes
);
1358 BlockErrorAction action
;
1361 * We failed, so we should mark dirty the whole area, aligned up.
1362 * Note that we don't care about shrunk tails if any: they were dirty
1363 * at function start, and they must be still dirty, as we've locked
1364 * the region for in-flight op.
1366 bitmap_offset
= QEMU_ALIGN_DOWN(offset
, job
->granularity
);
1367 bitmap_end
= QEMU_ALIGN_UP(offset
+ bytes
, job
->granularity
);
1368 bdrv_set_dirty_bitmap(job
->dirty_bitmap
, bitmap_offset
,
1369 bitmap_end
- bitmap_offset
);
1370 job
->actively_synced
= false;
1372 action
= mirror_error_action(job
, false, -ret
);
1373 if (action
== BLOCK_ERROR_ACTION_REPORT
) {
1381 static MirrorOp
*coroutine_fn
active_write_prepare(MirrorBlockJob
*s
,
1386 uint64_t start_chunk
= offset
/ s
->granularity
;
1387 uint64_t end_chunk
= DIV_ROUND_UP(offset
+ bytes
, s
->granularity
);
1389 op
= g_new(MirrorOp
, 1);
1394 .is_active_write
= true,
1395 .is_in_flight
= true,
1396 .co
= qemu_coroutine_self(),
1398 qemu_co_queue_init(&op
->waiting_requests
);
1399 QTAILQ_INSERT_TAIL(&s
->ops_in_flight
, op
, next
);
1401 s
->in_active_write_counter
++;
1404 * Wait for concurrent requests affecting the area. If there are already
1405 * running requests that are copying off now-to-be stale data in the area,
1406 * we must wait for them to finish before we begin writing fresh data to the
1407 * target so that the write operations appear in the correct order.
1408 * Note that background requests (see mirror_iteration()) in contrast only
1409 * wait for conflicting requests at the start of the dirty area, and then
1410 * (based on the in_flight_bitmap) truncate the area to copy so it will not
1411 * conflict with any requests beyond that. For active writes, however, we
1412 * cannot truncate that area. The request from our parent must be blocked
1413 * until the area is copied in full. Therefore, we must wait for the whole
1414 * area to become free of concurrent requests.
1416 mirror_wait_on_conflicts(op
, s
, offset
, bytes
);
1418 bitmap_set(s
->in_flight_bitmap
, start_chunk
, end_chunk
- start_chunk
);
1423 static void coroutine_fn GRAPH_RDLOCK
active_write_settle(MirrorOp
*op
)
1425 uint64_t start_chunk
= op
->offset
/ op
->s
->granularity
;
1426 uint64_t end_chunk
= DIV_ROUND_UP(op
->offset
+ op
->bytes
,
1427 op
->s
->granularity
);
1429 if (!--op
->s
->in_active_write_counter
&& op
->s
->actively_synced
) {
1430 BdrvChild
*source
= op
->s
->mirror_top_bs
->backing
;
1432 if (QLIST_FIRST(&source
->bs
->parents
) == source
&&
1433 QLIST_NEXT(source
, next_parent
) == NULL
)
1435 /* Assert that we are back in sync once all active write
1436 * operations are settled.
1437 * Note that we can only assert this if the mirror node
1438 * is the source node's only parent. */
1439 assert(!bdrv_get_dirty_count(op
->s
->dirty_bitmap
));
1442 bitmap_clear(op
->s
->in_flight_bitmap
, start_chunk
, end_chunk
- start_chunk
);
1443 QTAILQ_REMOVE(&op
->s
->ops_in_flight
, op
, next
);
1444 qemu_co_queue_restart_all(&op
->waiting_requests
);
1448 static int coroutine_fn GRAPH_RDLOCK
1449 bdrv_mirror_top_preadv(BlockDriverState
*bs
, int64_t offset
, int64_t bytes
,
1450 QEMUIOVector
*qiov
, BdrvRequestFlags flags
)
1452 return bdrv_co_preadv(bs
->backing
, offset
, bytes
, qiov
, flags
);
1455 static int coroutine_fn GRAPH_RDLOCK
1456 bdrv_mirror_top_do_write(BlockDriverState
*bs
, MirrorMethod method
,
1457 uint64_t offset
, uint64_t bytes
, QEMUIOVector
*qiov
,
1460 MirrorOp
*op
= NULL
;
1461 MirrorBDSOpaque
*s
= bs
->opaque
;
1463 bool copy_to_target
= false;
1466 copy_to_target
= s
->job
->ret
>= 0 &&
1467 !job_is_cancelled(&s
->job
->common
.job
) &&
1468 s
->job
->copy_mode
== MIRROR_COPY_MODE_WRITE_BLOCKING
;
1471 if (copy_to_target
) {
1472 op
= active_write_prepare(s
->job
, offset
, bytes
);
1476 case MIRROR_METHOD_COPY
:
1477 ret
= bdrv_co_pwritev(bs
->backing
, offset
, bytes
, qiov
, flags
);
1480 case MIRROR_METHOD_ZERO
:
1481 ret
= bdrv_co_pwrite_zeroes(bs
->backing
, offset
, bytes
, flags
);
1484 case MIRROR_METHOD_DISCARD
:
1485 ret
= bdrv_co_pdiscard(bs
->backing
, offset
, bytes
);
1496 if (copy_to_target
) {
1497 do_sync_target_write(s
->job
, method
, offset
, bytes
, qiov
, flags
);
1501 if (copy_to_target
) {
1502 active_write_settle(op
);
1507 static int coroutine_fn GRAPH_RDLOCK
1508 bdrv_mirror_top_pwritev(BlockDriverState
*bs
, int64_t offset
, int64_t bytes
,
1509 QEMUIOVector
*qiov
, BdrvRequestFlags flags
)
1511 MirrorBDSOpaque
*s
= bs
->opaque
;
1512 QEMUIOVector bounce_qiov
;
1515 bool copy_to_target
= false;
1518 copy_to_target
= s
->job
->ret
>= 0 &&
1519 !job_is_cancelled(&s
->job
->common
.job
) &&
1520 s
->job
->copy_mode
== MIRROR_COPY_MODE_WRITE_BLOCKING
;
1523 if (copy_to_target
) {
1524 /* The guest might concurrently modify the data to write; but
1525 * the data on source and destination must match, so we have
1526 * to use a bounce buffer if we are going to write to the
1528 bounce_buf
= qemu_blockalign(bs
, bytes
);
1529 iov_to_buf_full(qiov
->iov
, qiov
->niov
, 0, bounce_buf
, bytes
);
1531 qemu_iovec_init(&bounce_qiov
, 1);
1532 qemu_iovec_add(&bounce_qiov
, bounce_buf
, bytes
);
1533 qiov
= &bounce_qiov
;
1535 flags
&= ~BDRV_REQ_REGISTERED_BUF
;
1538 ret
= bdrv_mirror_top_do_write(bs
, MIRROR_METHOD_COPY
, offset
, bytes
, qiov
,
1541 if (copy_to_target
) {
1542 qemu_iovec_destroy(&bounce_qiov
);
1543 qemu_vfree(bounce_buf
);
1549 static int coroutine_fn GRAPH_RDLOCK
bdrv_mirror_top_flush(BlockDriverState
*bs
)
1551 if (bs
->backing
== NULL
) {
1552 /* we can be here after failed bdrv_append in mirror_start_job */
1555 return bdrv_co_flush(bs
->backing
->bs
);
1558 static int coroutine_fn GRAPH_RDLOCK
1559 bdrv_mirror_top_pwrite_zeroes(BlockDriverState
*bs
, int64_t offset
,
1560 int64_t bytes
, BdrvRequestFlags flags
)
1562 return bdrv_mirror_top_do_write(bs
, MIRROR_METHOD_ZERO
, offset
, bytes
, NULL
,
1566 static int coroutine_fn GRAPH_RDLOCK
1567 bdrv_mirror_top_pdiscard(BlockDriverState
*bs
, int64_t offset
, int64_t bytes
)
1569 return bdrv_mirror_top_do_write(bs
, MIRROR_METHOD_DISCARD
, offset
, bytes
,
1573 static void bdrv_mirror_top_refresh_filename(BlockDriverState
*bs
)
1575 if (bs
->backing
== NULL
) {
1576 /* we can be here after failed bdrv_attach_child in
1577 * bdrv_set_backing_hd */
1580 pstrcpy(bs
->exact_filename
, sizeof(bs
->exact_filename
),
1581 bs
->backing
->bs
->filename
);
1584 static void bdrv_mirror_top_child_perm(BlockDriverState
*bs
, BdrvChild
*c
,
1586 BlockReopenQueue
*reopen_queue
,
1587 uint64_t perm
, uint64_t shared
,
1588 uint64_t *nperm
, uint64_t *nshared
)
1590 MirrorBDSOpaque
*s
= bs
->opaque
;
1594 * If the job is to be stopped, we do not need to forward
1595 * anything to the real image.
1598 *nshared
= BLK_PERM_ALL
;
1602 bdrv_default_perms(bs
, c
, role
, reopen_queue
,
1603 perm
, shared
, nperm
, nshared
);
1607 * For commit jobs, we cannot take CONSISTENT_READ, because
1608 * that permission is unshared for everything above the base
1609 * node (except for filters on the base node).
1610 * We also have to force-share the WRITE permission, or
1611 * otherwise we would block ourselves at the base node (if
1612 * writes are blocked for a node, they are also blocked for
1613 * its backing file).
1614 * (We could also share RESIZE, because it may be needed for
1615 * the target if its size is less than the top node's; but
1616 * bdrv_default_perms_for_cow() automatically shares RESIZE
1617 * for backing nodes if WRITE is shared, so there is no need
1620 *nperm
&= ~BLK_PERM_CONSISTENT_READ
;
1621 *nshared
|= BLK_PERM_WRITE
;
1625 /* Dummy node that provides consistent read to its users without requiring it
1626 * from its backing file and that allows writes on the backing file chain. */
1627 static BlockDriver bdrv_mirror_top
= {
1628 .format_name
= "mirror_top",
1629 .bdrv_co_preadv
= bdrv_mirror_top_preadv
,
1630 .bdrv_co_pwritev
= bdrv_mirror_top_pwritev
,
1631 .bdrv_co_pwrite_zeroes
= bdrv_mirror_top_pwrite_zeroes
,
1632 .bdrv_co_pdiscard
= bdrv_mirror_top_pdiscard
,
1633 .bdrv_co_flush
= bdrv_mirror_top_flush
,
1634 .bdrv_refresh_filename
= bdrv_mirror_top_refresh_filename
,
1635 .bdrv_child_perm
= bdrv_mirror_top_child_perm
,
1638 .filtered_child_is_backing
= true,
1641 static BlockJob
*mirror_start_job(
1642 const char *job_id
, BlockDriverState
*bs
,
1643 int creation_flags
, BlockDriverState
*target
,
1644 const char *replaces
, int64_t speed
,
1645 uint32_t granularity
, int64_t buf_size
,
1646 BlockMirrorBackingMode backing_mode
,
1648 BlockdevOnError on_source_error
,
1649 BlockdevOnError on_target_error
,
1651 BlockCompletionFunc
*cb
,
1653 const BlockJobDriver
*driver
,
1654 bool is_none_mode
, BlockDriverState
*base
,
1655 bool auto_complete
, const char *filter_node_name
,
1656 bool is_mirror
, MirrorCopyMode copy_mode
,
1660 MirrorBDSOpaque
*bs_opaque
;
1661 BlockDriverState
*mirror_top_bs
;
1662 bool target_is_backing
;
1663 uint64_t target_perms
, target_shared_perms
;
1666 if (granularity
== 0) {
1667 granularity
= bdrv_get_default_bitmap_granularity(target
);
1670 assert(is_power_of_2(granularity
));
1673 error_setg(errp
, "Invalid parameter 'buf-size'");
1677 if (buf_size
== 0) {
1678 buf_size
= DEFAULT_MIRROR_BUF_SIZE
;
1681 if (bdrv_skip_filters(bs
) == bdrv_skip_filters(target
)) {
1682 error_setg(errp
, "Can't mirror node into itself");
1686 target_is_backing
= bdrv_chain_contains(bs
, target
);
1688 /* In the case of active commit, add dummy driver to provide consistent
1689 * reads on the top, while disabling it in the intermediate nodes, and make
1690 * the backing chain writable. */
1691 mirror_top_bs
= bdrv_new_open_driver(&bdrv_mirror_top
, filter_node_name
,
1693 if (mirror_top_bs
== NULL
) {
1696 if (!filter_node_name
) {
1697 mirror_top_bs
->implicit
= true;
1700 /* So that we can always drop this node */
1701 mirror_top_bs
->never_freeze
= true;
1703 mirror_top_bs
->total_sectors
= bs
->total_sectors
;
1704 mirror_top_bs
->supported_write_flags
= BDRV_REQ_WRITE_UNCHANGED
;
1705 mirror_top_bs
->supported_zero_flags
= BDRV_REQ_WRITE_UNCHANGED
|
1706 BDRV_REQ_NO_FALLBACK
;
1707 bs_opaque
= g_new0(MirrorBDSOpaque
, 1);
1708 mirror_top_bs
->opaque
= bs_opaque
;
1710 bs_opaque
->is_commit
= target_is_backing
;
1712 bdrv_drained_begin(bs
);
1713 ret
= bdrv_append(mirror_top_bs
, bs
, errp
);
1714 bdrv_drained_end(bs
);
1717 bdrv_unref(mirror_top_bs
);
1721 /* Make sure that the source is not resized while the job is running */
1722 s
= block_job_create(job_id
, driver
, NULL
, mirror_top_bs
,
1723 BLK_PERM_CONSISTENT_READ
,
1724 BLK_PERM_CONSISTENT_READ
| BLK_PERM_WRITE_UNCHANGED
|
1725 BLK_PERM_WRITE
, speed
,
1726 creation_flags
, cb
, opaque
, errp
);
1731 /* The block job now has a reference to this node */
1732 bdrv_unref(mirror_top_bs
);
1734 s
->mirror_top_bs
= mirror_top_bs
;
1736 /* No resize for the target either; while the mirror is still running, a
1737 * consistent read isn't necessarily possible. We could possibly allow
1738 * writes and graph modifications, though it would likely defeat the
1739 * purpose of a mirror, so leave them blocked for now.
1741 * In the case of active commit, things look a bit different, though,
1742 * because the target is an already populated backing file in active use.
1743 * We can allow anything except resize there.*/
1745 target_perms
= BLK_PERM_WRITE
;
1746 target_shared_perms
= BLK_PERM_WRITE_UNCHANGED
;
1748 if (target_is_backing
) {
1749 int64_t bs_size
, target_size
;
1750 bs_size
= bdrv_getlength(bs
);
1752 error_setg_errno(errp
, -bs_size
,
1753 "Could not inquire top image size");
1757 target_size
= bdrv_getlength(target
);
1758 if (target_size
< 0) {
1759 error_setg_errno(errp
, -target_size
,
1760 "Could not inquire base image size");
1764 if (target_size
< bs_size
) {
1765 target_perms
|= BLK_PERM_RESIZE
;
1768 target_shared_perms
|= BLK_PERM_CONSISTENT_READ
| BLK_PERM_WRITE
;
1769 } else if (bdrv_chain_contains(bs
, bdrv_skip_filters(target
))) {
1771 * We may want to allow this in the future, but it would
1772 * require taking some extra care.
1774 error_setg(errp
, "Cannot mirror to a filter on top of a node in the "
1775 "source's backing chain");
1779 s
->target
= blk_new(s
->common
.job
.aio_context
,
1780 target_perms
, target_shared_perms
);
1781 ret
= blk_insert_bs(s
->target
, target
, errp
);
1786 /* XXX: Mirror target could be a NBD server of target QEMU in the case
1787 * of non-shared block migration. To allow migration completion, we
1788 * have to allow "inactivate" of the target BB. When that happens, we
1789 * know the job is drained, and the vcpus are stopped, so no write
1790 * operation will be performed. Block layer already has assertions to
1792 blk_set_force_allow_inactivate(s
->target
);
1794 blk_set_allow_aio_context_change(s
->target
, true);
1795 blk_set_disable_request_queuing(s
->target
, true);
1797 s
->replaces
= g_strdup(replaces
);
1798 s
->on_source_error
= on_source_error
;
1799 s
->on_target_error
= on_target_error
;
1800 s
->is_none_mode
= is_none_mode
;
1801 s
->backing_mode
= backing_mode
;
1802 s
->zero_target
= zero_target
;
1803 s
->copy_mode
= copy_mode
;
1805 s
->base_overlay
= bdrv_find_overlay(bs
, base
);
1806 s
->granularity
= granularity
;
1807 s
->buf_size
= ROUND_UP(buf_size
, granularity
);
1809 if (auto_complete
) {
1810 s
->should_complete
= true;
1813 s
->dirty_bitmap
= bdrv_create_dirty_bitmap(bs
, granularity
, NULL
, errp
);
1814 if (!s
->dirty_bitmap
) {
1817 if (s
->copy_mode
== MIRROR_COPY_MODE_WRITE_BLOCKING
) {
1818 bdrv_disable_dirty_bitmap(s
->dirty_bitmap
);
1821 ret
= block_job_add_bdrv(&s
->common
, "source", bs
, 0,
1822 BLK_PERM_WRITE_UNCHANGED
| BLK_PERM_WRITE
|
1823 BLK_PERM_CONSISTENT_READ
,
1829 /* Required permissions are already taken with blk_new() */
1830 block_job_add_bdrv(&s
->common
, "target", target
, 0, BLK_PERM_ALL
,
1833 /* In commit_active_start() all intermediate nodes disappear, so
1834 * any jobs in them must be blocked */
1835 if (target_is_backing
) {
1836 BlockDriverState
*iter
, *filtered_target
;
1837 uint64_t iter_shared_perms
;
1840 * The topmost node with
1841 * bdrv_skip_filters(filtered_target) == bdrv_skip_filters(target)
1843 filtered_target
= bdrv_cow_bs(bdrv_find_overlay(bs
, target
));
1845 assert(bdrv_skip_filters(filtered_target
) ==
1846 bdrv_skip_filters(target
));
1849 * XXX BLK_PERM_WRITE needs to be allowed so we don't block
1850 * ourselves at s->base (if writes are blocked for a node, they are
1851 * also blocked for its backing file). The other options would be a
1852 * second filter driver above s->base (== target).
1854 iter_shared_perms
= BLK_PERM_WRITE_UNCHANGED
| BLK_PERM_WRITE
;
1856 for (iter
= bdrv_filter_or_cow_bs(bs
); iter
!= target
;
1857 iter
= bdrv_filter_or_cow_bs(iter
))
1859 if (iter
== filtered_target
) {
1861 * From here on, all nodes are filters on the base.
1862 * This allows us to share BLK_PERM_CONSISTENT_READ.
1864 iter_shared_perms
|= BLK_PERM_CONSISTENT_READ
;
1867 ret
= block_job_add_bdrv(&s
->common
, "intermediate node", iter
, 0,
1868 iter_shared_perms
, errp
);
1874 if (bdrv_freeze_backing_chain(mirror_top_bs
, target
, errp
) < 0) {
1879 QTAILQ_INIT(&s
->ops_in_flight
);
1881 trace_mirror_start(bs
, s
, opaque
);
1882 job_start(&s
->common
.job
);
1888 /* Make sure this BDS does not go away until we have completed the graph
1890 bdrv_ref(mirror_top_bs
);
1892 g_free(s
->replaces
);
1893 blk_unref(s
->target
);
1894 bs_opaque
->job
= NULL
;
1895 if (s
->dirty_bitmap
) {
1896 bdrv_release_dirty_bitmap(s
->dirty_bitmap
);
1898 job_early_fail(&s
->common
.job
);
1901 bs_opaque
->stop
= true;
1902 bdrv_child_refresh_perms(mirror_top_bs
, mirror_top_bs
->backing
,
1904 bdrv_replace_node(mirror_top_bs
, mirror_top_bs
->backing
->bs
, &error_abort
);
1906 bdrv_unref(mirror_top_bs
);
1911 void mirror_start(const char *job_id
, BlockDriverState
*bs
,
1912 BlockDriverState
*target
, const char *replaces
,
1913 int creation_flags
, int64_t speed
,
1914 uint32_t granularity
, int64_t buf_size
,
1915 MirrorSyncMode mode
, BlockMirrorBackingMode backing_mode
,
1917 BlockdevOnError on_source_error
,
1918 BlockdevOnError on_target_error
,
1919 bool unmap
, const char *filter_node_name
,
1920 MirrorCopyMode copy_mode
, Error
**errp
)
1923 BlockDriverState
*base
;
1925 GLOBAL_STATE_CODE();
1927 if ((mode
== MIRROR_SYNC_MODE_INCREMENTAL
) ||
1928 (mode
== MIRROR_SYNC_MODE_BITMAP
)) {
1929 error_setg(errp
, "Sync mode '%s' not supported",
1930 MirrorSyncMode_str(mode
));
1933 is_none_mode
= mode
== MIRROR_SYNC_MODE_NONE
;
1934 base
= mode
== MIRROR_SYNC_MODE_TOP
? bdrv_backing_chain_next(bs
) : NULL
;
1935 mirror_start_job(job_id
, bs
, creation_flags
, target
, replaces
,
1936 speed
, granularity
, buf_size
, backing_mode
, zero_target
,
1937 on_source_error
, on_target_error
, unmap
, NULL
, NULL
,
1938 &mirror_job_driver
, is_none_mode
, base
, false,
1939 filter_node_name
, true, copy_mode
, errp
);
1942 BlockJob
*commit_active_start(const char *job_id
, BlockDriverState
*bs
,
1943 BlockDriverState
*base
, int creation_flags
,
1944 int64_t speed
, BlockdevOnError on_error
,
1945 const char *filter_node_name
,
1946 BlockCompletionFunc
*cb
, void *opaque
,
1947 bool auto_complete
, Error
**errp
)
1949 bool base_read_only
;
1952 GLOBAL_STATE_CODE();
1954 base_read_only
= bdrv_is_read_only(base
);
1956 if (base_read_only
) {
1957 if (bdrv_reopen_set_read_only(base
, false, errp
) < 0) {
1962 job
= mirror_start_job(
1963 job_id
, bs
, creation_flags
, base
, NULL
, speed
, 0, 0,
1964 MIRROR_LEAVE_BACKING_CHAIN
, false,
1965 on_error
, on_error
, true, cb
, opaque
,
1966 &commit_active_job_driver
, false, base
, auto_complete
,
1967 filter_node_name
, false, MIRROR_COPY_MODE_BACKGROUND
,
1970 goto error_restore_flags
;
1975 error_restore_flags
:
1976 /* ignore error and errp for bdrv_reopen, because we want to propagate
1977 * the original error */
1978 if (base_read_only
) {
1979 bdrv_reopen_set_read_only(base
, true, NULL
);