4 * Copyright Red Hat, Inc. 2012
7 * Paolo Bonzini <pbonzini@redhat.com>
9 * This work is licensed under the terms of the GNU LGPL, version 2 or later.
10 * See the COPYING.LIB file in the top-level directory.
14 #include "qemu/osdep.h"
15 #include "qemu/cutils.h"
17 #include "block/blockjob_int.h"
18 #include "block/block_int.h"
19 #include "sysemu/block-backend.h"
20 #include "qapi/error.h"
21 #include "qapi/qmp/qerror.h"
22 #include "qemu/ratelimit.h"
23 #include "qemu/bitmap.h"
25 #define SLICE_TIME 100000000ULL /* ns */
26 #define MAX_IN_FLIGHT 16
27 #define MAX_IO_BYTES (1 << 20) /* 1 Mb */
28 #define DEFAULT_MIRROR_BUF_SIZE (MAX_IN_FLIGHT * MAX_IO_BYTES)
30 /* The mirroring buffer is a list of granularity-sized chunks.
31 * Free chunks are organized in a list.
33 typedef struct MirrorBuffer
{
34 QSIMPLEQ_ENTRY(MirrorBuffer
) next
;
37 typedef struct MirrorBlockJob
{
41 BlockDriverState
*mirror_top_bs
;
42 BlockDriverState
*source
;
43 BlockDriverState
*base
;
45 /* The name of the graph node to replace */
47 /* The BDS to replace */
48 BlockDriverState
*to_replace
;
49 /* Used to block operations on the drive-mirror-replace target */
50 Error
*replace_blocker
;
52 BlockMirrorBackingMode backing_mode
;
53 BlockdevOnError on_source_error
, on_target_error
;
59 unsigned long *cow_bitmap
;
60 BdrvDirtyBitmap
*dirty_bitmap
;
61 BdrvDirtyBitmapIter
*dbi
;
63 QSIMPLEQ_HEAD(, MirrorBuffer
) buf_free
;
66 uint64_t last_pause_ns
;
67 unsigned long *in_flight_bitmap
;
69 int64_t bytes_in_flight
;
73 int target_cluster_size
;
75 bool initial_zeroing_ongoing
;
78 typedef struct MirrorOp
{
85 static BlockErrorAction
mirror_error_action(MirrorBlockJob
*s
, bool read
,
90 return block_job_error_action(&s
->common
, s
->on_source_error
,
93 return block_job_error_action(&s
->common
, s
->on_target_error
,
98 static void mirror_iteration_done(MirrorOp
*op
, int ret
)
100 MirrorBlockJob
*s
= op
->s
;
105 trace_mirror_iteration_done(s
, op
->offset
, op
->bytes
, ret
);
108 s
->bytes_in_flight
-= op
->bytes
;
110 for (i
= 0; i
< op
->qiov
.niov
; i
++) {
111 MirrorBuffer
*buf
= (MirrorBuffer
*) iov
[i
].iov_base
;
112 QSIMPLEQ_INSERT_TAIL(&s
->buf_free
, buf
, next
);
116 chunk_num
= op
->offset
/ s
->granularity
;
117 nb_chunks
= DIV_ROUND_UP(op
->bytes
, s
->granularity
);
118 bitmap_clear(s
->in_flight_bitmap
, chunk_num
, nb_chunks
);
121 bitmap_set(s
->cow_bitmap
, chunk_num
, nb_chunks
);
123 if (!s
->initial_zeroing_ongoing
) {
124 s
->common
.offset
+= op
->bytes
;
127 qemu_iovec_destroy(&op
->qiov
);
130 if (s
->waiting_for_io
) {
131 qemu_coroutine_enter(s
->common
.co
);
135 static void mirror_write_complete(void *opaque
, int ret
)
137 MirrorOp
*op
= opaque
;
138 MirrorBlockJob
*s
= op
->s
;
140 aio_context_acquire(blk_get_aio_context(s
->common
.blk
));
142 BlockErrorAction action
;
144 bdrv_set_dirty_bitmap(s
->dirty_bitmap
, op
->offset
>> BDRV_SECTOR_BITS
,
145 op
->bytes
>> BDRV_SECTOR_BITS
);
146 action
= mirror_error_action(s
, false, -ret
);
147 if (action
== BLOCK_ERROR_ACTION_REPORT
&& s
->ret
>= 0) {
151 mirror_iteration_done(op
, ret
);
152 aio_context_release(blk_get_aio_context(s
->common
.blk
));
155 static void mirror_read_complete(void *opaque
, int ret
)
157 MirrorOp
*op
= opaque
;
158 MirrorBlockJob
*s
= op
->s
;
160 aio_context_acquire(blk_get_aio_context(s
->common
.blk
));
162 BlockErrorAction action
;
164 bdrv_set_dirty_bitmap(s
->dirty_bitmap
, op
->offset
>> BDRV_SECTOR_BITS
,
165 op
->bytes
>> BDRV_SECTOR_BITS
);
166 action
= mirror_error_action(s
, true, -ret
);
167 if (action
== BLOCK_ERROR_ACTION_REPORT
&& s
->ret
>= 0) {
171 mirror_iteration_done(op
, ret
);
173 blk_aio_pwritev(s
->target
, op
->offset
, &op
->qiov
,
174 0, mirror_write_complete
, op
);
176 aio_context_release(blk_get_aio_context(s
->common
.blk
));
179 /* Clip bytes relative to offset to not exceed end-of-file */
180 static inline int64_t mirror_clip_bytes(MirrorBlockJob
*s
,
184 return MIN(bytes
, s
->bdev_length
- offset
);
187 /* Round offset and/or bytes to target cluster if COW is needed, and
188 * return the offset of the adjusted tail against original. */
189 static int mirror_cow_align(MirrorBlockJob
*s
, int64_t *offset
,
194 int64_t align_offset
= *offset
;
195 unsigned int align_bytes
= *bytes
;
196 int max_bytes
= s
->granularity
* s
->max_iov
;
198 assert(*bytes
< INT_MAX
);
199 need_cow
= !test_bit(*offset
/ s
->granularity
, s
->cow_bitmap
);
200 need_cow
|= !test_bit((*offset
+ *bytes
- 1) / s
->granularity
,
203 bdrv_round_to_clusters(blk_bs(s
->target
), *offset
, *bytes
,
204 &align_offset
, &align_bytes
);
207 if (align_bytes
> max_bytes
) {
208 align_bytes
= max_bytes
;
210 align_bytes
= QEMU_ALIGN_DOWN(align_bytes
, s
->target_cluster_size
);
213 /* Clipping may result in align_bytes unaligned to chunk boundary, but
214 * that doesn't matter because it's already the end of source image. */
215 align_bytes
= mirror_clip_bytes(s
, align_offset
, align_bytes
);
217 ret
= align_offset
+ align_bytes
- (*offset
+ *bytes
);
218 *offset
= align_offset
;
219 *bytes
= align_bytes
;
224 static inline void mirror_wait_for_io(MirrorBlockJob
*s
)
226 assert(!s
->waiting_for_io
);
227 s
->waiting_for_io
= true;
228 qemu_coroutine_yield();
229 s
->waiting_for_io
= false;
232 /* Submit async read while handling COW.
233 * Returns: The number of bytes copied after and including offset,
234 * excluding any bytes copied prior to offset due to alignment.
235 * This will be @bytes if no alignment is necessary, or
236 * (new_end - offset) if tail is rounded up or down due to
237 * alignment or buffer limit.
239 static uint64_t mirror_do_read(MirrorBlockJob
*s
, int64_t offset
,
242 BlockBackend
*source
= s
->common
.blk
;
248 max_bytes
= s
->granularity
* s
->max_iov
;
250 /* We can only handle as much as buf_size at a time. */
251 bytes
= MIN(s
->buf_size
, MIN(max_bytes
, bytes
));
253 assert(bytes
< BDRV_REQUEST_MAX_BYTES
);
257 ret
+= mirror_cow_align(s
, &offset
, &bytes
);
259 assert(bytes
<= s
->buf_size
);
260 /* The offset is granularity-aligned because:
261 * 1) Caller passes in aligned values;
262 * 2) mirror_cow_align is used only when target cluster is larger. */
263 assert(QEMU_IS_ALIGNED(offset
, s
->granularity
));
264 /* The range is sector-aligned, since bdrv_getlength() rounds up. */
265 assert(QEMU_IS_ALIGNED(bytes
, BDRV_SECTOR_SIZE
));
266 nb_chunks
= DIV_ROUND_UP(bytes
, s
->granularity
);
268 while (s
->buf_free_count
< nb_chunks
) {
269 trace_mirror_yield_in_flight(s
, offset
, s
->in_flight
);
270 mirror_wait_for_io(s
);
273 /* Allocate a MirrorOp that is used as an AIO callback. */
274 op
= g_new(MirrorOp
, 1);
279 /* Now make a QEMUIOVector taking enough granularity-sized chunks
282 qemu_iovec_init(&op
->qiov
, nb_chunks
);
283 while (nb_chunks
-- > 0) {
284 MirrorBuffer
*buf
= QSIMPLEQ_FIRST(&s
->buf_free
);
285 size_t remaining
= bytes
- op
->qiov
.size
;
287 QSIMPLEQ_REMOVE_HEAD(&s
->buf_free
, next
);
289 qemu_iovec_add(&op
->qiov
, buf
, MIN(s
->granularity
, remaining
));
292 /* Copy the dirty cluster. */
294 s
->bytes_in_flight
+= bytes
;
295 trace_mirror_one_iteration(s
, offset
, bytes
);
297 blk_aio_preadv(source
, offset
, &op
->qiov
, 0, mirror_read_complete
, op
);
301 static void mirror_do_zero_or_discard(MirrorBlockJob
*s
,
308 /* Allocate a MirrorOp that is used as an AIO callback. The qiov is zeroed
309 * so the freeing in mirror_iteration_done is nop. */
310 op
= g_new0(MirrorOp
, 1);
316 s
->bytes_in_flight
+= bytes
;
318 blk_aio_pdiscard(s
->target
, offset
,
319 op
->bytes
, mirror_write_complete
, op
);
321 blk_aio_pwrite_zeroes(s
->target
, offset
,
322 op
->bytes
, s
->unmap
? BDRV_REQ_MAY_UNMAP
: 0,
323 mirror_write_complete
, op
);
327 static uint64_t coroutine_fn
mirror_iteration(MirrorBlockJob
*s
)
329 BlockDriverState
*source
= s
->source
;
330 int64_t offset
, first_chunk
;
331 uint64_t delay_ns
= 0;
332 /* At least the first dirty chunk is mirrored in one iteration. */
334 int sectors_per_chunk
= s
->granularity
>> BDRV_SECTOR_BITS
;
335 bool write_zeroes_ok
= bdrv_can_write_zeroes_with_unmap(blk_bs(s
->target
));
336 int max_io_bytes
= MAX(s
->buf_size
/ MAX_IN_FLIGHT
, MAX_IO_BYTES
);
338 bdrv_dirty_bitmap_lock(s
->dirty_bitmap
);
339 offset
= bdrv_dirty_iter_next(s
->dbi
);
341 bdrv_set_dirty_iter(s
->dbi
, 0);
342 offset
= bdrv_dirty_iter_next(s
->dbi
);
343 trace_mirror_restart_iter(s
, bdrv_get_dirty_count(s
->dirty_bitmap
));
346 bdrv_dirty_bitmap_unlock(s
->dirty_bitmap
);
348 first_chunk
= offset
/ s
->granularity
;
349 while (test_bit(first_chunk
, s
->in_flight_bitmap
)) {
350 trace_mirror_yield_in_flight(s
, offset
, s
->in_flight
);
351 mirror_wait_for_io(s
);
354 block_job_pause_point(&s
->common
);
356 /* Find the number of consective dirty chunks following the first dirty
357 * one, and wait for in flight requests in them. */
358 bdrv_dirty_bitmap_lock(s
->dirty_bitmap
);
359 while (nb_chunks
* s
->granularity
< s
->buf_size
) {
361 int64_t next_offset
= offset
+ nb_chunks
* s
->granularity
;
362 int64_t next_chunk
= next_offset
/ s
->granularity
;
363 if (next_offset
>= s
->bdev_length
||
364 !bdrv_get_dirty_locked(source
, s
->dirty_bitmap
, next_offset
)) {
367 if (test_bit(next_chunk
, s
->in_flight_bitmap
)) {
371 next_dirty
= bdrv_dirty_iter_next(s
->dbi
);
372 if (next_dirty
> next_offset
|| next_dirty
< 0) {
373 /* The bitmap iterator's cache is stale, refresh it */
374 bdrv_set_dirty_iter(s
->dbi
, next_offset
);
375 next_dirty
= bdrv_dirty_iter_next(s
->dbi
);
377 assert(next_dirty
== next_offset
);
381 /* Clear dirty bits before querying the block status, because
382 * calling bdrv_get_block_status_above could yield - if some blocks are
383 * marked dirty in this window, we need to know.
385 bdrv_reset_dirty_bitmap_locked(s
->dirty_bitmap
, offset
>> BDRV_SECTOR_BITS
,
386 nb_chunks
* sectors_per_chunk
);
387 bdrv_dirty_bitmap_unlock(s
->dirty_bitmap
);
389 bitmap_set(s
->in_flight_bitmap
, offset
/ s
->granularity
, nb_chunks
);
390 while (nb_chunks
> 0 && offset
< s
->bdev_length
) {
393 unsigned int io_bytes
;
394 int64_t io_bytes_acct
;
395 BlockDriverState
*file
;
399 MIRROR_METHOD_DISCARD
400 } mirror_method
= MIRROR_METHOD_COPY
;
402 assert(!(offset
% s
->granularity
));
403 ret
= bdrv_get_block_status_above(source
, NULL
,
404 offset
>> BDRV_SECTOR_BITS
,
405 nb_chunks
* sectors_per_chunk
,
407 io_bytes
= io_sectors
* BDRV_SECTOR_SIZE
;
409 io_bytes
= MIN(nb_chunks
* s
->granularity
, max_io_bytes
);
410 } else if (ret
& BDRV_BLOCK_DATA
) {
411 io_bytes
= MIN(io_bytes
, max_io_bytes
);
414 io_bytes
-= io_bytes
% s
->granularity
;
415 if (io_bytes
< s
->granularity
) {
416 io_bytes
= s
->granularity
;
417 } else if (ret
>= 0 && !(ret
& BDRV_BLOCK_DATA
)) {
418 int64_t target_offset
;
419 unsigned int target_bytes
;
420 bdrv_round_to_clusters(blk_bs(s
->target
), offset
, io_bytes
,
421 &target_offset
, &target_bytes
);
422 if (target_offset
== offset
&&
423 target_bytes
== io_bytes
) {
424 mirror_method
= ret
& BDRV_BLOCK_ZERO
?
426 MIRROR_METHOD_DISCARD
;
430 while (s
->in_flight
>= MAX_IN_FLIGHT
) {
431 trace_mirror_yield_in_flight(s
, offset
, s
->in_flight
);
432 mirror_wait_for_io(s
);
439 io_bytes
= mirror_clip_bytes(s
, offset
, io_bytes
);
440 switch (mirror_method
) {
441 case MIRROR_METHOD_COPY
:
442 io_bytes
= io_bytes_acct
= mirror_do_read(s
, offset
, io_bytes
);
444 case MIRROR_METHOD_ZERO
:
445 case MIRROR_METHOD_DISCARD
:
446 mirror_do_zero_or_discard(s
, offset
, io_bytes
,
447 mirror_method
== MIRROR_METHOD_DISCARD
);
448 if (write_zeroes_ok
) {
451 io_bytes_acct
= io_bytes
;
459 nb_chunks
-= DIV_ROUND_UP(io_bytes
, s
->granularity
);
460 if (s
->common
.speed
) {
461 delay_ns
= ratelimit_calculate_delay(&s
->limit
, io_bytes_acct
);
467 static void mirror_free_init(MirrorBlockJob
*s
)
469 int granularity
= s
->granularity
;
470 size_t buf_size
= s
->buf_size
;
471 uint8_t *buf
= s
->buf
;
473 assert(s
->buf_free_count
== 0);
474 QSIMPLEQ_INIT(&s
->buf_free
);
475 while (buf_size
!= 0) {
476 MirrorBuffer
*cur
= (MirrorBuffer
*)buf
;
477 QSIMPLEQ_INSERT_TAIL(&s
->buf_free
, cur
, next
);
479 buf_size
-= granularity
;
484 /* This is also used for the .pause callback. There is no matching
485 * mirror_resume() because mirror_run() will begin iterating again
486 * when the job is resumed.
488 static void mirror_wait_for_all_io(MirrorBlockJob
*s
)
490 while (s
->in_flight
> 0) {
491 mirror_wait_for_io(s
);
499 static void mirror_exit(BlockJob
*job
, void *opaque
)
501 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
502 MirrorExitData
*data
= opaque
;
503 AioContext
*replace_aio_context
= NULL
;
504 BlockDriverState
*src
= s
->source
;
505 BlockDriverState
*target_bs
= blk_bs(s
->target
);
506 BlockDriverState
*mirror_top_bs
= s
->mirror_top_bs
;
507 Error
*local_err
= NULL
;
509 bdrv_release_dirty_bitmap(src
, s
->dirty_bitmap
);
511 /* Make sure that the source BDS doesn't go away before we called
512 * block_job_completed(). */
514 bdrv_ref(mirror_top_bs
);
517 /* Remove target parent that still uses BLK_PERM_WRITE/RESIZE before
518 * inserting target_bs at s->to_replace, where we might not be able to get
521 * Note that blk_unref() alone doesn't necessarily drop permissions because
522 * we might be running nested inside mirror_drain(), which takes an extra
523 * reference, so use an explicit blk_set_perm() first. */
524 blk_set_perm(s
->target
, 0, BLK_PERM_ALL
, &error_abort
);
525 blk_unref(s
->target
);
528 /* We don't access the source any more. Dropping any WRITE/RESIZE is
529 * required before it could become a backing file of target_bs. */
530 bdrv_child_try_set_perm(mirror_top_bs
->backing
, 0, BLK_PERM_ALL
,
532 if (s
->backing_mode
== MIRROR_SOURCE_BACKING_CHAIN
) {
533 BlockDriverState
*backing
= s
->is_none_mode
? src
: s
->base
;
534 if (backing_bs(target_bs
) != backing
) {
535 bdrv_set_backing_hd(target_bs
, backing
, &local_err
);
537 error_report_err(local_err
);
544 replace_aio_context
= bdrv_get_aio_context(s
->to_replace
);
545 aio_context_acquire(replace_aio_context
);
548 if (s
->should_complete
&& data
->ret
== 0) {
549 BlockDriverState
*to_replace
= src
;
551 to_replace
= s
->to_replace
;
554 if (bdrv_get_flags(target_bs
) != bdrv_get_flags(to_replace
)) {
555 bdrv_reopen(target_bs
, bdrv_get_flags(to_replace
), NULL
);
558 /* The mirror job has no requests in flight any more, but we need to
559 * drain potential other users of the BDS before changing the graph. */
560 bdrv_drained_begin(target_bs
);
561 bdrv_replace_node(to_replace
, target_bs
, &local_err
);
562 bdrv_drained_end(target_bs
);
564 error_report_err(local_err
);
569 bdrv_op_unblock_all(s
->to_replace
, s
->replace_blocker
);
570 error_free(s
->replace_blocker
);
571 bdrv_unref(s
->to_replace
);
573 if (replace_aio_context
) {
574 aio_context_release(replace_aio_context
);
577 bdrv_unref(target_bs
);
579 /* Remove the mirror filter driver from the graph. Before this, get rid of
580 * the blockers on the intermediate nodes so that the resulting state is
581 * valid. Also give up permissions on mirror_top_bs->backing, which might
582 * block the removal. */
583 block_job_remove_all_bdrv(job
);
584 bdrv_child_try_set_perm(mirror_top_bs
->backing
, 0, BLK_PERM_ALL
,
586 bdrv_replace_node(mirror_top_bs
, backing_bs(mirror_top_bs
), &error_abort
);
588 /* We just changed the BDS the job BB refers to (with either or both of the
589 * bdrv_replace_node() calls), so switch the BB back so the cleanup does
590 * the right thing. We don't need any permissions any more now. */
591 blk_remove_bs(job
->blk
);
592 blk_set_perm(job
->blk
, 0, BLK_PERM_ALL
, &error_abort
);
593 blk_insert_bs(job
->blk
, mirror_top_bs
, &error_abort
);
595 block_job_completed(&s
->common
, data
->ret
);
598 bdrv_drained_end(src
);
599 bdrv_unref(mirror_top_bs
);
603 static void mirror_throttle(MirrorBlockJob
*s
)
605 int64_t now
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
);
607 if (now
- s
->last_pause_ns
> SLICE_TIME
) {
608 s
->last_pause_ns
= now
;
609 block_job_sleep_ns(&s
->common
, QEMU_CLOCK_REALTIME
, 0);
611 block_job_pause_point(&s
->common
);
615 static int coroutine_fn
mirror_dirty_init(MirrorBlockJob
*s
)
617 int64_t sector_num
, end
;
618 BlockDriverState
*base
= s
->base
;
619 BlockDriverState
*bs
= s
->source
;
620 BlockDriverState
*target_bs
= blk_bs(s
->target
);
624 end
= s
->bdev_length
/ BDRV_SECTOR_SIZE
;
626 if (base
== NULL
&& !bdrv_has_zero_init(target_bs
)) {
627 if (!bdrv_can_write_zeroes_with_unmap(target_bs
)) {
628 bdrv_set_dirty_bitmap(s
->dirty_bitmap
, 0, end
);
632 s
->initial_zeroing_ongoing
= true;
633 for (sector_num
= 0; sector_num
< end
; ) {
634 int nb_sectors
= MIN(end
- sector_num
,
635 QEMU_ALIGN_DOWN(INT_MAX
, s
->granularity
) >> BDRV_SECTOR_BITS
);
639 if (block_job_is_cancelled(&s
->common
)) {
640 s
->initial_zeroing_ongoing
= false;
644 if (s
->in_flight
>= MAX_IN_FLIGHT
) {
645 trace_mirror_yield(s
, UINT64_MAX
, s
->buf_free_count
,
647 mirror_wait_for_io(s
);
651 mirror_do_zero_or_discard(s
, sector_num
* BDRV_SECTOR_SIZE
,
652 nb_sectors
* BDRV_SECTOR_SIZE
, false);
653 sector_num
+= nb_sectors
;
656 mirror_wait_for_all_io(s
);
657 s
->initial_zeroing_ongoing
= false;
660 /* First part, loop on the sectors and initialize the dirty bitmap. */
661 for (sector_num
= 0; sector_num
< end
; ) {
662 /* Just to make sure we are not exceeding int limit. */
663 int nb_sectors
= MIN(INT_MAX
>> BDRV_SECTOR_BITS
,
668 if (block_job_is_cancelled(&s
->common
)) {
672 ret
= bdrv_is_allocated_above(bs
, base
, sector_num
* BDRV_SECTOR_SIZE
,
673 nb_sectors
* BDRV_SECTOR_SIZE
, &count
);
678 /* TODO: Relax this once bdrv_is_allocated_above and dirty
679 * bitmaps no longer require sector alignment. */
680 assert(QEMU_IS_ALIGNED(count
, BDRV_SECTOR_SIZE
));
681 n
= count
>> BDRV_SECTOR_BITS
;
684 bdrv_set_dirty_bitmap(s
->dirty_bitmap
, sector_num
, n
);
691 /* Called when going out of the streaming phase to flush the bulk of the
692 * data to the medium, or just before completing.
694 static int mirror_flush(MirrorBlockJob
*s
)
696 int ret
= blk_flush(s
->target
);
698 if (mirror_error_action(s
, false, -ret
) == BLOCK_ERROR_ACTION_REPORT
) {
705 static void coroutine_fn
mirror_run(void *opaque
)
707 MirrorBlockJob
*s
= opaque
;
708 MirrorExitData
*data
;
709 BlockDriverState
*bs
= s
->source
;
710 BlockDriverState
*target_bs
= blk_bs(s
->target
);
711 bool need_drain
= true;
714 char backing_filename
[2]; /* we only need 2 characters because we are only
715 checking for a NULL string */
718 if (block_job_is_cancelled(&s
->common
)) {
722 s
->bdev_length
= bdrv_getlength(bs
);
723 if (s
->bdev_length
< 0) {
724 ret
= s
->bdev_length
;
728 /* Active commit must resize the base image if its size differs from the
730 if (s
->base
== blk_bs(s
->target
)) {
733 base_length
= blk_getlength(s
->target
);
734 if (base_length
< 0) {
739 if (s
->bdev_length
> base_length
) {
740 ret
= blk_truncate(s
->target
, s
->bdev_length
, PREALLOC_MODE_OFF
,
748 if (s
->bdev_length
== 0) {
749 /* Report BLOCK_JOB_READY and wait for complete. */
750 block_job_event_ready(&s
->common
);
752 while (!block_job_is_cancelled(&s
->common
) && !s
->should_complete
) {
753 block_job_yield(&s
->common
);
755 s
->common
.cancelled
= false;
759 length
= DIV_ROUND_UP(s
->bdev_length
, s
->granularity
);
760 s
->in_flight_bitmap
= bitmap_new(length
);
762 /* If we have no backing file yet in the destination, we cannot let
763 * the destination do COW. Instead, we copy sectors around the
764 * dirty data if needed. We need a bitmap to do that.
766 bdrv_get_backing_filename(target_bs
, backing_filename
,
767 sizeof(backing_filename
));
768 if (!bdrv_get_info(target_bs
, &bdi
) && bdi
.cluster_size
) {
769 s
->target_cluster_size
= bdi
.cluster_size
;
771 s
->target_cluster_size
= BDRV_SECTOR_SIZE
;
773 if (backing_filename
[0] && !target_bs
->backing
&&
774 s
->granularity
< s
->target_cluster_size
) {
775 s
->buf_size
= MAX(s
->buf_size
, s
->target_cluster_size
);
776 s
->cow_bitmap
= bitmap_new(length
);
778 s
->max_iov
= MIN(bs
->bl
.max_iov
, target_bs
->bl
.max_iov
);
780 s
->buf
= qemu_try_blockalign(bs
, s
->buf_size
);
781 if (s
->buf
== NULL
) {
788 s
->last_pause_ns
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
);
789 if (!s
->is_none_mode
) {
790 ret
= mirror_dirty_init(s
);
791 if (ret
< 0 || block_job_is_cancelled(&s
->common
)) {
797 s
->dbi
= bdrv_dirty_iter_new(s
->dirty_bitmap
);
799 uint64_t delay_ns
= 0;
801 bool should_complete
;
808 block_job_pause_point(&s
->common
);
810 cnt
= bdrv_get_dirty_count(s
->dirty_bitmap
);
811 /* s->common.offset contains the number of bytes already processed so
812 * far, cnt is the number of dirty bytes remaining and
813 * s->bytes_in_flight is the number of bytes currently being
814 * processed; together those are the current total operation length */
815 s
->common
.len
= s
->common
.offset
+ s
->bytes_in_flight
+ cnt
;
817 /* Note that even when no rate limit is applied we need to yield
818 * periodically with no pending I/O so that bdrv_drain_all() returns.
819 * We do so every SLICE_TIME nanoseconds, or when there is an error,
820 * or when the source is clean, whichever comes first.
822 delta
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
) - s
->last_pause_ns
;
823 if (delta
< SLICE_TIME
&&
824 s
->common
.iostatus
== BLOCK_DEVICE_IO_STATUS_OK
) {
825 if (s
->in_flight
>= MAX_IN_FLIGHT
|| s
->buf_free_count
== 0 ||
826 (cnt
== 0 && s
->in_flight
> 0)) {
827 trace_mirror_yield(s
, cnt
, s
->buf_free_count
, s
->in_flight
);
828 mirror_wait_for_io(s
);
830 } else if (cnt
!= 0) {
831 delay_ns
= mirror_iteration(s
);
835 should_complete
= false;
836 if (s
->in_flight
== 0 && cnt
== 0) {
837 trace_mirror_before_flush(s
);
839 if (mirror_flush(s
) < 0) {
840 /* Go check s->ret. */
843 /* We're out of the streaming phase. From now on, if the job
844 * is cancelled we will actually complete all pending I/O and
845 * report completion. This way, block-job-cancel will leave
846 * the target in a consistent state.
848 block_job_event_ready(&s
->common
);
852 should_complete
= s
->should_complete
||
853 block_job_is_cancelled(&s
->common
);
854 cnt
= bdrv_get_dirty_count(s
->dirty_bitmap
);
857 if (cnt
== 0 && should_complete
) {
858 /* The dirty bitmap is not updated while operations are pending.
859 * If we're about to exit, wait for pending operations before
860 * calling bdrv_get_dirty_count(bs), or we may exit while the
861 * source has dirty data to copy!
863 * Note that I/O can be submitted by the guest while
864 * mirror_populate runs, so pause it now. Before deciding
865 * whether to switch to target check one last time if I/O has
866 * come in the meanwhile, and if not flush the data to disk.
868 trace_mirror_before_drain(s
, cnt
);
870 bdrv_drained_begin(bs
);
871 cnt
= bdrv_get_dirty_count(s
->dirty_bitmap
);
872 if (cnt
> 0 || mirror_flush(s
) < 0) {
873 bdrv_drained_end(bs
);
877 /* The two disks are in sync. Exit and report successful
880 assert(QLIST_EMPTY(&bs
->tracked_requests
));
881 s
->common
.cancelled
= false;
887 trace_mirror_before_sleep(s
, cnt
, s
->synced
, delay_ns
);
889 block_job_sleep_ns(&s
->common
, QEMU_CLOCK_REALTIME
, delay_ns
);
890 if (block_job_is_cancelled(&s
->common
)) {
893 } else if (!should_complete
) {
894 delay_ns
= (s
->in_flight
== 0 && cnt
== 0 ? SLICE_TIME
: 0);
895 block_job_sleep_ns(&s
->common
, QEMU_CLOCK_REALTIME
, delay_ns
);
897 s
->last_pause_ns
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
);
901 if (s
->in_flight
> 0) {
902 /* We get here only if something went wrong. Either the job failed,
903 * or it was cancelled prematurely so that we do not guarantee that
904 * the target is a copy of the source.
906 assert(ret
< 0 || (!s
->synced
&& block_job_is_cancelled(&s
->common
)));
908 mirror_wait_for_all_io(s
);
911 assert(s
->in_flight
== 0);
913 g_free(s
->cow_bitmap
);
914 g_free(s
->in_flight_bitmap
);
915 bdrv_dirty_iter_free(s
->dbi
);
917 data
= g_malloc(sizeof(*data
));
921 bdrv_drained_begin(bs
);
923 block_job_defer_to_main_loop(&s
->common
, mirror_exit
, data
);
926 static void mirror_set_speed(BlockJob
*job
, int64_t speed
, Error
**errp
)
928 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
931 error_setg(errp
, QERR_INVALID_PARAMETER
, "speed");
934 ratelimit_set_speed(&s
->limit
, speed
, SLICE_TIME
);
937 static void mirror_complete(BlockJob
*job
, Error
**errp
)
939 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
940 BlockDriverState
*target
;
942 target
= blk_bs(s
->target
);
945 error_setg(errp
, "The active block job '%s' cannot be completed",
950 if (s
->backing_mode
== MIRROR_OPEN_BACKING_CHAIN
) {
953 assert(!target
->backing
);
954 ret
= bdrv_open_backing_file(target
, NULL
, "backing", errp
);
960 /* block all operations on to_replace bs */
962 AioContext
*replace_aio_context
;
964 s
->to_replace
= bdrv_find_node(s
->replaces
);
965 if (!s
->to_replace
) {
966 error_setg(errp
, "Node name '%s' not found", s
->replaces
);
970 replace_aio_context
= bdrv_get_aio_context(s
->to_replace
);
971 aio_context_acquire(replace_aio_context
);
973 /* TODO Translate this into permission system. Current definition of
974 * GRAPH_MOD would require to request it for the parents; they might
975 * not even be BlockDriverStates, however, so a BdrvChild can't address
976 * them. May need redefinition of GRAPH_MOD. */
977 error_setg(&s
->replace_blocker
,
978 "block device is in use by block-job-complete");
979 bdrv_op_block_all(s
->to_replace
, s
->replace_blocker
);
980 bdrv_ref(s
->to_replace
);
982 aio_context_release(replace_aio_context
);
985 s
->should_complete
= true;
986 block_job_enter(&s
->common
);
989 static void mirror_pause(BlockJob
*job
)
991 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
993 mirror_wait_for_all_io(s
);
996 static void mirror_attached_aio_context(BlockJob
*job
, AioContext
*new_context
)
998 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
1000 blk_set_aio_context(s
->target
, new_context
);
1003 static void mirror_drain(BlockJob
*job
)
1005 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
1007 /* Need to keep a reference in case blk_drain triggers execution
1008 * of mirror_complete...
1011 BlockBackend
*target
= s
->target
;
1018 static const BlockJobDriver mirror_job_driver
= {
1019 .instance_size
= sizeof(MirrorBlockJob
),
1020 .job_type
= BLOCK_JOB_TYPE_MIRROR
,
1021 .set_speed
= mirror_set_speed
,
1022 .start
= mirror_run
,
1023 .complete
= mirror_complete
,
1024 .pause
= mirror_pause
,
1025 .attached_aio_context
= mirror_attached_aio_context
,
1026 .drain
= mirror_drain
,
1029 static const BlockJobDriver commit_active_job_driver
= {
1030 .instance_size
= sizeof(MirrorBlockJob
),
1031 .job_type
= BLOCK_JOB_TYPE_COMMIT
,
1032 .set_speed
= mirror_set_speed
,
1033 .start
= mirror_run
,
1034 .complete
= mirror_complete
,
1035 .pause
= mirror_pause
,
1036 .attached_aio_context
= mirror_attached_aio_context
,
1037 .drain
= mirror_drain
,
1040 static int coroutine_fn
bdrv_mirror_top_preadv(BlockDriverState
*bs
,
1041 uint64_t offset
, uint64_t bytes
, QEMUIOVector
*qiov
, int flags
)
1043 return bdrv_co_preadv(bs
->backing
, offset
, bytes
, qiov
, flags
);
1046 static int coroutine_fn
bdrv_mirror_top_pwritev(BlockDriverState
*bs
,
1047 uint64_t offset
, uint64_t bytes
, QEMUIOVector
*qiov
, int flags
)
1049 return bdrv_co_pwritev(bs
->backing
, offset
, bytes
, qiov
, flags
);
1052 static int coroutine_fn
bdrv_mirror_top_flush(BlockDriverState
*bs
)
1054 return bdrv_co_flush(bs
->backing
->bs
);
1057 static int coroutine_fn
bdrv_mirror_top_pwrite_zeroes(BlockDriverState
*bs
,
1058 int64_t offset
, int bytes
, BdrvRequestFlags flags
)
1060 return bdrv_co_pwrite_zeroes(bs
->backing
, offset
, bytes
, flags
);
1063 static int coroutine_fn
bdrv_mirror_top_pdiscard(BlockDriverState
*bs
,
1064 int64_t offset
, int bytes
)
1066 return bdrv_co_pdiscard(bs
->backing
->bs
, offset
, bytes
);
1069 static void bdrv_mirror_top_refresh_filename(BlockDriverState
*bs
, QDict
*opts
)
1071 bdrv_refresh_filename(bs
->backing
->bs
);
1072 pstrcpy(bs
->exact_filename
, sizeof(bs
->exact_filename
),
1073 bs
->backing
->bs
->filename
);
1076 static void bdrv_mirror_top_close(BlockDriverState
*bs
)
1080 static void bdrv_mirror_top_child_perm(BlockDriverState
*bs
, BdrvChild
*c
,
1081 const BdrvChildRole
*role
,
1082 BlockReopenQueue
*reopen_queue
,
1083 uint64_t perm
, uint64_t shared
,
1084 uint64_t *nperm
, uint64_t *nshared
)
1086 /* Must be able to forward guest writes to the real image */
1088 if (perm
& BLK_PERM_WRITE
) {
1089 *nperm
|= BLK_PERM_WRITE
;
1092 *nshared
= BLK_PERM_ALL
;
1095 /* Dummy node that provides consistent read to its users without requiring it
1096 * from its backing file and that allows writes on the backing file chain. */
1097 static BlockDriver bdrv_mirror_top
= {
1098 .format_name
= "mirror_top",
1099 .bdrv_co_preadv
= bdrv_mirror_top_preadv
,
1100 .bdrv_co_pwritev
= bdrv_mirror_top_pwritev
,
1101 .bdrv_co_pwrite_zeroes
= bdrv_mirror_top_pwrite_zeroes
,
1102 .bdrv_co_pdiscard
= bdrv_mirror_top_pdiscard
,
1103 .bdrv_co_flush
= bdrv_mirror_top_flush
,
1104 .bdrv_co_get_block_status
= bdrv_co_get_block_status_from_backing
,
1105 .bdrv_refresh_filename
= bdrv_mirror_top_refresh_filename
,
1106 .bdrv_close
= bdrv_mirror_top_close
,
1107 .bdrv_child_perm
= bdrv_mirror_top_child_perm
,
1110 static void mirror_start_job(const char *job_id
, BlockDriverState
*bs
,
1111 int creation_flags
, BlockDriverState
*target
,
1112 const char *replaces
, int64_t speed
,
1113 uint32_t granularity
, int64_t buf_size
,
1114 BlockMirrorBackingMode backing_mode
,
1115 BlockdevOnError on_source_error
,
1116 BlockdevOnError on_target_error
,
1118 BlockCompletionFunc
*cb
,
1120 const BlockJobDriver
*driver
,
1121 bool is_none_mode
, BlockDriverState
*base
,
1122 bool auto_complete
, const char *filter_node_name
,
1127 BlockDriverState
*mirror_top_bs
;
1128 bool target_graph_mod
;
1129 bool target_is_backing
;
1130 Error
*local_err
= NULL
;
1133 if (granularity
== 0) {
1134 granularity
= bdrv_get_default_bitmap_granularity(target
);
1137 assert ((granularity
& (granularity
- 1)) == 0);
1138 /* Granularity must be large enough for sector-based dirty bitmap */
1139 assert(granularity
>= BDRV_SECTOR_SIZE
);
1142 error_setg(errp
, "Invalid parameter 'buf-size'");
1146 if (buf_size
== 0) {
1147 buf_size
= DEFAULT_MIRROR_BUF_SIZE
;
1150 /* In the case of active commit, add dummy driver to provide consistent
1151 * reads on the top, while disabling it in the intermediate nodes, and make
1152 * the backing chain writable. */
1153 mirror_top_bs
= bdrv_new_open_driver(&bdrv_mirror_top
, filter_node_name
,
1155 if (mirror_top_bs
== NULL
) {
1158 if (!filter_node_name
) {
1159 mirror_top_bs
->implicit
= true;
1161 mirror_top_bs
->total_sectors
= bs
->total_sectors
;
1162 bdrv_set_aio_context(mirror_top_bs
, bdrv_get_aio_context(bs
));
1164 /* bdrv_append takes ownership of the mirror_top_bs reference, need to keep
1165 * it alive until block_job_create() succeeds even if bs has no parent. */
1166 bdrv_ref(mirror_top_bs
);
1167 bdrv_drained_begin(bs
);
1168 bdrv_append(mirror_top_bs
, bs
, &local_err
);
1169 bdrv_drained_end(bs
);
1172 bdrv_unref(mirror_top_bs
);
1173 error_propagate(errp
, local_err
);
1177 /* Make sure that the source is not resized while the job is running */
1178 s
= block_job_create(job_id
, driver
, mirror_top_bs
,
1179 BLK_PERM_CONSISTENT_READ
,
1180 BLK_PERM_CONSISTENT_READ
| BLK_PERM_WRITE_UNCHANGED
|
1181 BLK_PERM_WRITE
| BLK_PERM_GRAPH_MOD
, speed
,
1182 creation_flags
, cb
, opaque
, errp
);
1186 /* The block job now has a reference to this node */
1187 bdrv_unref(mirror_top_bs
);
1190 s
->mirror_top_bs
= mirror_top_bs
;
1192 /* No resize for the target either; while the mirror is still running, a
1193 * consistent read isn't necessarily possible. We could possibly allow
1194 * writes and graph modifications, though it would likely defeat the
1195 * purpose of a mirror, so leave them blocked for now.
1197 * In the case of active commit, things look a bit different, though,
1198 * because the target is an already populated backing file in active use.
1199 * We can allow anything except resize there.*/
1200 target_is_backing
= bdrv_chain_contains(bs
, target
);
1201 target_graph_mod
= (backing_mode
!= MIRROR_LEAVE_BACKING_CHAIN
);
1202 s
->target
= blk_new(BLK_PERM_WRITE
| BLK_PERM_RESIZE
|
1203 (target_graph_mod
? BLK_PERM_GRAPH_MOD
: 0),
1204 BLK_PERM_WRITE_UNCHANGED
|
1205 (target_is_backing
? BLK_PERM_CONSISTENT_READ
|
1207 BLK_PERM_GRAPH_MOD
: 0));
1208 ret
= blk_insert_bs(s
->target
, target
, errp
);
1213 /* XXX: Mirror target could be a NBD server of target QEMU in the case
1214 * of non-shared block migration. To allow migration completion, we
1215 * have to allow "inactivate" of the target BB. When that happens, we
1216 * know the job is drained, and the vcpus are stopped, so no write
1217 * operation will be performed. Block layer already has assertions to
1219 blk_set_force_allow_inactivate(s
->target
);
1222 s
->replaces
= g_strdup(replaces
);
1223 s
->on_source_error
= on_source_error
;
1224 s
->on_target_error
= on_target_error
;
1225 s
->is_none_mode
= is_none_mode
;
1226 s
->backing_mode
= backing_mode
;
1228 s
->granularity
= granularity
;
1229 s
->buf_size
= ROUND_UP(buf_size
, granularity
);
1231 if (auto_complete
) {
1232 s
->should_complete
= true;
1235 s
->dirty_bitmap
= bdrv_create_dirty_bitmap(bs
, granularity
, NULL
, errp
);
1236 if (!s
->dirty_bitmap
) {
1240 /* Required permissions are already taken with blk_new() */
1241 block_job_add_bdrv(&s
->common
, "target", target
, 0, BLK_PERM_ALL
,
1244 /* In commit_active_start() all intermediate nodes disappear, so
1245 * any jobs in them must be blocked */
1246 if (target_is_backing
) {
1247 BlockDriverState
*iter
;
1248 for (iter
= backing_bs(bs
); iter
!= target
; iter
= backing_bs(iter
)) {
1249 /* XXX BLK_PERM_WRITE needs to be allowed so we don't block
1250 * ourselves at s->base (if writes are blocked for a node, they are
1251 * also blocked for its backing file). The other options would be a
1252 * second filter driver above s->base (== target). */
1253 ret
= block_job_add_bdrv(&s
->common
, "intermediate node", iter
, 0,
1254 BLK_PERM_WRITE_UNCHANGED
| BLK_PERM_WRITE
,
1262 trace_mirror_start(bs
, s
, opaque
);
1263 block_job_start(&s
->common
);
1268 /* Make sure this BDS does not go away until we have completed the graph
1270 bdrv_ref(mirror_top_bs
);
1272 g_free(s
->replaces
);
1273 blk_unref(s
->target
);
1274 block_job_early_fail(&s
->common
);
1277 bdrv_child_try_set_perm(mirror_top_bs
->backing
, 0, BLK_PERM_ALL
,
1279 bdrv_replace_node(mirror_top_bs
, backing_bs(mirror_top_bs
), &error_abort
);
1281 bdrv_unref(mirror_top_bs
);
1284 void mirror_start(const char *job_id
, BlockDriverState
*bs
,
1285 BlockDriverState
*target
, const char *replaces
,
1286 int64_t speed
, uint32_t granularity
, int64_t buf_size
,
1287 MirrorSyncMode mode
, BlockMirrorBackingMode backing_mode
,
1288 BlockdevOnError on_source_error
,
1289 BlockdevOnError on_target_error
,
1290 bool unmap
, const char *filter_node_name
, Error
**errp
)
1293 BlockDriverState
*base
;
1295 if (mode
== MIRROR_SYNC_MODE_INCREMENTAL
) {
1296 error_setg(errp
, "Sync mode 'incremental' not supported");
1299 is_none_mode
= mode
== MIRROR_SYNC_MODE_NONE
;
1300 base
= mode
== MIRROR_SYNC_MODE_TOP
? backing_bs(bs
) : NULL
;
1301 mirror_start_job(job_id
, bs
, BLOCK_JOB_DEFAULT
, target
, replaces
,
1302 speed
, granularity
, buf_size
, backing_mode
,
1303 on_source_error
, on_target_error
, unmap
, NULL
, NULL
,
1304 &mirror_job_driver
, is_none_mode
, base
, false,
1305 filter_node_name
, true, errp
);
1308 void commit_active_start(const char *job_id
, BlockDriverState
*bs
,
1309 BlockDriverState
*base
, int creation_flags
,
1310 int64_t speed
, BlockdevOnError on_error
,
1311 const char *filter_node_name
,
1312 BlockCompletionFunc
*cb
, void *opaque
,
1313 bool auto_complete
, Error
**errp
)
1315 int orig_base_flags
;
1316 Error
*local_err
= NULL
;
1318 orig_base_flags
= bdrv_get_flags(base
);
1320 if (bdrv_reopen(base
, bs
->open_flags
, errp
)) {
1324 mirror_start_job(job_id
, bs
, creation_flags
, base
, NULL
, speed
, 0, 0,
1325 MIRROR_LEAVE_BACKING_CHAIN
,
1326 on_error
, on_error
, true, cb
, opaque
,
1327 &commit_active_job_driver
, false, base
, auto_complete
,
1328 filter_node_name
, false, &local_err
);
1330 error_propagate(errp
, local_err
);
1331 goto error_restore_flags
;
1336 error_restore_flags
:
1337 /* ignore error and errp for bdrv_reopen, because we want to propagate
1338 * the original error */
1339 bdrv_reopen(base
, orig_base_flags
, NULL
);