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_SECTORS ((1 << 20) >> BDRV_SECTOR_BITS) /* 1 Mb */
28 #define DEFAULT_MIRROR_BUF_SIZE \
29 (MAX_IN_FLIGHT * MAX_IO_SECTORS * BDRV_SECTOR_SIZE)
31 /* The mirroring buffer is a list of granularity-sized chunks.
32 * Free chunks are organized in a list.
34 typedef struct MirrorBuffer
{
35 QSIMPLEQ_ENTRY(MirrorBuffer
) next
;
38 typedef struct MirrorBlockJob
{
42 BlockDriverState
*mirror_top_bs
;
43 BlockDriverState
*source
;
44 BlockDriverState
*base
;
46 /* The name of the graph node to replace */
48 /* The BDS to replace */
49 BlockDriverState
*to_replace
;
50 /* Used to block operations on the drive-mirror-replace target */
51 Error
*replace_blocker
;
53 BlockMirrorBackingMode backing_mode
;
54 BlockdevOnError on_source_error
, on_target_error
;
60 unsigned long *cow_bitmap
;
61 BdrvDirtyBitmap
*dirty_bitmap
;
62 BdrvDirtyBitmapIter
*dbi
;
64 QSIMPLEQ_HEAD(, MirrorBuffer
) buf_free
;
67 uint64_t last_pause_ns
;
68 unsigned long *in_flight_bitmap
;
70 int64_t sectors_in_flight
;
74 int target_cluster_sectors
;
76 bool initial_zeroing_ongoing
;
79 typedef struct MirrorOp
{
86 static BlockErrorAction
mirror_error_action(MirrorBlockJob
*s
, bool read
,
91 return block_job_error_action(&s
->common
, s
->on_source_error
,
94 return block_job_error_action(&s
->common
, s
->on_target_error
,
99 static void mirror_iteration_done(MirrorOp
*op
, int ret
)
101 MirrorBlockJob
*s
= op
->s
;
104 int i
, nb_chunks
, sectors_per_chunk
;
106 trace_mirror_iteration_done(s
, op
->sector_num
, op
->nb_sectors
, ret
);
109 s
->sectors_in_flight
-= op
->nb_sectors
;
111 for (i
= 0; i
< op
->qiov
.niov
; i
++) {
112 MirrorBuffer
*buf
= (MirrorBuffer
*) iov
[i
].iov_base
;
113 QSIMPLEQ_INSERT_TAIL(&s
->buf_free
, buf
, next
);
117 sectors_per_chunk
= s
->granularity
>> BDRV_SECTOR_BITS
;
118 chunk_num
= op
->sector_num
/ sectors_per_chunk
;
119 nb_chunks
= DIV_ROUND_UP(op
->nb_sectors
, sectors_per_chunk
);
120 bitmap_clear(s
->in_flight_bitmap
, chunk_num
, nb_chunks
);
123 bitmap_set(s
->cow_bitmap
, chunk_num
, nb_chunks
);
125 if (!s
->initial_zeroing_ongoing
) {
126 s
->common
.offset
+= (uint64_t)op
->nb_sectors
* BDRV_SECTOR_SIZE
;
129 qemu_iovec_destroy(&op
->qiov
);
132 if (s
->waiting_for_io
) {
133 qemu_coroutine_enter(s
->common
.co
);
137 static void mirror_write_complete(void *opaque
, int ret
)
139 MirrorOp
*op
= opaque
;
140 MirrorBlockJob
*s
= op
->s
;
142 aio_context_acquire(blk_get_aio_context(s
->common
.blk
));
144 BlockErrorAction action
;
146 bdrv_set_dirty_bitmap(s
->dirty_bitmap
, op
->sector_num
, op
->nb_sectors
);
147 action
= mirror_error_action(s
, false, -ret
);
148 if (action
== BLOCK_ERROR_ACTION_REPORT
&& s
->ret
>= 0) {
152 mirror_iteration_done(op
, ret
);
153 aio_context_release(blk_get_aio_context(s
->common
.blk
));
156 static void mirror_read_complete(void *opaque
, int ret
)
158 MirrorOp
*op
= opaque
;
159 MirrorBlockJob
*s
= op
->s
;
161 aio_context_acquire(blk_get_aio_context(s
->common
.blk
));
163 BlockErrorAction action
;
165 bdrv_set_dirty_bitmap(s
->dirty_bitmap
, op
->sector_num
, op
->nb_sectors
);
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
->sector_num
* BDRV_SECTOR_SIZE
, &op
->qiov
,
174 0, mirror_write_complete
, op
);
176 aio_context_release(blk_get_aio_context(s
->common
.blk
));
179 static inline void mirror_clip_sectors(MirrorBlockJob
*s
,
183 *nb_sectors
= MIN(*nb_sectors
,
184 s
->bdev_length
/ BDRV_SECTOR_SIZE
- sector_num
);
187 /* Round sector_num and/or nb_sectors to target cluster if COW is needed, and
188 * return the offset of the adjusted tail sector against original. */
189 static int mirror_cow_align(MirrorBlockJob
*s
,
195 int chunk_sectors
= s
->granularity
>> BDRV_SECTOR_BITS
;
196 int64_t align_sector_num
= *sector_num
;
197 int align_nb_sectors
= *nb_sectors
;
198 int max_sectors
= chunk_sectors
* s
->max_iov
;
200 need_cow
= !test_bit(*sector_num
/ chunk_sectors
, s
->cow_bitmap
);
201 need_cow
|= !test_bit((*sector_num
+ *nb_sectors
- 1) / chunk_sectors
,
204 bdrv_round_sectors_to_clusters(blk_bs(s
->target
), *sector_num
,
205 *nb_sectors
, &align_sector_num
,
209 if (align_nb_sectors
> max_sectors
) {
210 align_nb_sectors
= max_sectors
;
212 align_nb_sectors
= QEMU_ALIGN_DOWN(align_nb_sectors
,
213 s
->target_cluster_sectors
);
216 /* Clipping may result in align_nb_sectors unaligned to chunk boundary, but
217 * that doesn't matter because it's already the end of source image. */
218 mirror_clip_sectors(s
, align_sector_num
, &align_nb_sectors
);
220 ret
= align_sector_num
+ align_nb_sectors
- (*sector_num
+ *nb_sectors
);
221 *sector_num
= align_sector_num
;
222 *nb_sectors
= align_nb_sectors
;
227 static inline void mirror_wait_for_io(MirrorBlockJob
*s
)
229 assert(!s
->waiting_for_io
);
230 s
->waiting_for_io
= true;
231 qemu_coroutine_yield();
232 s
->waiting_for_io
= false;
235 /* Submit async read while handling COW.
236 * Returns: The number of sectors copied after and including sector_num,
237 * excluding any sectors copied prior to sector_num due to alignment.
238 * This will be nb_sectors if no alignment is necessary, or
239 * (new_end - sector_num) if tail is rounded up or down due to
240 * alignment or buffer limit.
242 static int mirror_do_read(MirrorBlockJob
*s
, int64_t sector_num
,
245 BlockBackend
*source
= s
->common
.blk
;
246 int sectors_per_chunk
, nb_chunks
;
251 sectors_per_chunk
= s
->granularity
>> BDRV_SECTOR_BITS
;
252 max_sectors
= sectors_per_chunk
* s
->max_iov
;
254 /* We can only handle as much as buf_size at a time. */
255 nb_sectors
= MIN(s
->buf_size
>> BDRV_SECTOR_BITS
, nb_sectors
);
256 nb_sectors
= MIN(max_sectors
, nb_sectors
);
261 ret
+= mirror_cow_align(s
, §or_num
, &nb_sectors
);
263 assert(nb_sectors
<< BDRV_SECTOR_BITS
<= s
->buf_size
);
264 /* The sector range must meet granularity because:
265 * 1) Caller passes in aligned values;
266 * 2) mirror_cow_align is used only when target cluster is larger. */
267 assert(!(sector_num
% sectors_per_chunk
));
268 nb_chunks
= DIV_ROUND_UP(nb_sectors
, sectors_per_chunk
);
270 while (s
->buf_free_count
< nb_chunks
) {
271 trace_mirror_yield_in_flight(s
, sector_num
, s
->in_flight
);
272 mirror_wait_for_io(s
);
275 /* Allocate a MirrorOp that is used as an AIO callback. */
276 op
= g_new(MirrorOp
, 1);
278 op
->sector_num
= sector_num
;
279 op
->nb_sectors
= nb_sectors
;
281 /* Now make a QEMUIOVector taking enough granularity-sized chunks
284 qemu_iovec_init(&op
->qiov
, nb_chunks
);
285 while (nb_chunks
-- > 0) {
286 MirrorBuffer
*buf
= QSIMPLEQ_FIRST(&s
->buf_free
);
287 size_t remaining
= nb_sectors
* BDRV_SECTOR_SIZE
- op
->qiov
.size
;
289 QSIMPLEQ_REMOVE_HEAD(&s
->buf_free
, next
);
291 qemu_iovec_add(&op
->qiov
, buf
, MIN(s
->granularity
, remaining
));
294 /* Copy the dirty cluster. */
296 s
->sectors_in_flight
+= nb_sectors
;
297 trace_mirror_one_iteration(s
, sector_num
, nb_sectors
);
299 blk_aio_preadv(source
, sector_num
* BDRV_SECTOR_SIZE
, &op
->qiov
, 0,
300 mirror_read_complete
, op
);
304 static void mirror_do_zero_or_discard(MirrorBlockJob
*s
,
311 /* Allocate a MirrorOp that is used as an AIO callback. The qiov is zeroed
312 * so the freeing in mirror_iteration_done is nop. */
313 op
= g_new0(MirrorOp
, 1);
315 op
->sector_num
= sector_num
;
316 op
->nb_sectors
= nb_sectors
;
319 s
->sectors_in_flight
+= nb_sectors
;
321 blk_aio_pdiscard(s
->target
, sector_num
<< BDRV_SECTOR_BITS
,
322 op
->nb_sectors
<< BDRV_SECTOR_BITS
,
323 mirror_write_complete
, op
);
325 blk_aio_pwrite_zeroes(s
->target
, sector_num
* BDRV_SECTOR_SIZE
,
326 op
->nb_sectors
* BDRV_SECTOR_SIZE
,
327 s
->unmap
? BDRV_REQ_MAY_UNMAP
: 0,
328 mirror_write_complete
, op
);
332 static uint64_t coroutine_fn
mirror_iteration(MirrorBlockJob
*s
)
334 BlockDriverState
*source
= s
->source
;
335 int64_t sector_num
, first_chunk
;
336 uint64_t delay_ns
= 0;
337 /* At least the first dirty chunk is mirrored in one iteration. */
339 int64_t end
= s
->bdev_length
/ BDRV_SECTOR_SIZE
;
340 int sectors_per_chunk
= s
->granularity
>> BDRV_SECTOR_BITS
;
341 bool write_zeroes_ok
= bdrv_can_write_zeroes_with_unmap(blk_bs(s
->target
));
342 int max_io_sectors
= MAX((s
->buf_size
>> BDRV_SECTOR_BITS
) / MAX_IN_FLIGHT
,
345 sector_num
= bdrv_dirty_iter_next(s
->dbi
);
346 if (sector_num
< 0) {
347 bdrv_set_dirty_iter(s
->dbi
, 0);
348 sector_num
= bdrv_dirty_iter_next(s
->dbi
);
349 trace_mirror_restart_iter(s
, bdrv_get_dirty_count(s
->dirty_bitmap
));
350 assert(sector_num
>= 0);
353 first_chunk
= sector_num
/ sectors_per_chunk
;
354 while (test_bit(first_chunk
, s
->in_flight_bitmap
)) {
355 trace_mirror_yield_in_flight(s
, sector_num
, s
->in_flight
);
356 mirror_wait_for_io(s
);
359 block_job_pause_point(&s
->common
);
361 /* Find the number of consective dirty chunks following the first dirty
362 * one, and wait for in flight requests in them. */
363 while (nb_chunks
* sectors_per_chunk
< (s
->buf_size
>> BDRV_SECTOR_BITS
)) {
365 int64_t next_sector
= sector_num
+ nb_chunks
* sectors_per_chunk
;
366 int64_t next_chunk
= next_sector
/ sectors_per_chunk
;
367 if (next_sector
>= end
||
368 !bdrv_get_dirty(source
, s
->dirty_bitmap
, next_sector
)) {
371 if (test_bit(next_chunk
, s
->in_flight_bitmap
)) {
375 next_dirty
= bdrv_dirty_iter_next(s
->dbi
);
376 if (next_dirty
> next_sector
|| next_dirty
< 0) {
377 /* The bitmap iterator's cache is stale, refresh it */
378 bdrv_set_dirty_iter(s
->dbi
, next_sector
);
379 next_dirty
= bdrv_dirty_iter_next(s
->dbi
);
381 assert(next_dirty
== next_sector
);
385 /* Clear dirty bits before querying the block status, because
386 * calling bdrv_get_block_status_above could yield - if some blocks are
387 * marked dirty in this window, we need to know.
389 bdrv_reset_dirty_bitmap(s
->dirty_bitmap
, sector_num
,
390 nb_chunks
* sectors_per_chunk
);
391 bitmap_set(s
->in_flight_bitmap
, sector_num
/ sectors_per_chunk
, nb_chunks
);
392 while (nb_chunks
> 0 && sector_num
< end
) {
394 int io_sectors
, io_sectors_acct
;
395 BlockDriverState
*file
;
399 MIRROR_METHOD_DISCARD
400 } mirror_method
= MIRROR_METHOD_COPY
;
402 assert(!(sector_num
% sectors_per_chunk
));
403 ret
= bdrv_get_block_status_above(source
, NULL
, sector_num
,
404 nb_chunks
* sectors_per_chunk
,
407 io_sectors
= MIN(nb_chunks
* sectors_per_chunk
, max_io_sectors
);
408 } else if (ret
& BDRV_BLOCK_DATA
) {
409 io_sectors
= MIN(io_sectors
, max_io_sectors
);
412 io_sectors
-= io_sectors
% sectors_per_chunk
;
413 if (io_sectors
< sectors_per_chunk
) {
414 io_sectors
= sectors_per_chunk
;
415 } else if (ret
>= 0 && !(ret
& BDRV_BLOCK_DATA
)) {
416 int64_t target_sector_num
;
417 int target_nb_sectors
;
418 bdrv_round_sectors_to_clusters(blk_bs(s
->target
), sector_num
,
419 io_sectors
, &target_sector_num
,
421 if (target_sector_num
== sector_num
&&
422 target_nb_sectors
== io_sectors
) {
423 mirror_method
= ret
& BDRV_BLOCK_ZERO
?
425 MIRROR_METHOD_DISCARD
;
429 while (s
->in_flight
>= MAX_IN_FLIGHT
) {
430 trace_mirror_yield_in_flight(s
, sector_num
, s
->in_flight
);
431 mirror_wait_for_io(s
);
438 mirror_clip_sectors(s
, sector_num
, &io_sectors
);
439 switch (mirror_method
) {
440 case MIRROR_METHOD_COPY
:
441 io_sectors
= mirror_do_read(s
, sector_num
, io_sectors
);
442 io_sectors_acct
= io_sectors
;
444 case MIRROR_METHOD_ZERO
:
445 case MIRROR_METHOD_DISCARD
:
446 mirror_do_zero_or_discard(s
, sector_num
, io_sectors
,
447 mirror_method
== MIRROR_METHOD_DISCARD
);
448 if (write_zeroes_ok
) {
451 io_sectors_acct
= io_sectors
;
458 sector_num
+= io_sectors
;
459 nb_chunks
-= DIV_ROUND_UP(io_sectors
, sectors_per_chunk
);
460 if (s
->common
.speed
) {
461 delay_ns
= ratelimit_calculate_delay(&s
->limit
, io_sectors_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 /* Make sure that the source BDS doesn't go away before we called
510 * block_job_completed(). */
512 bdrv_ref(mirror_top_bs
);
515 /* Remove target parent that still uses BLK_PERM_WRITE/RESIZE before
516 * inserting target_bs at s->to_replace, where we might not be able to get
517 * these permissions. */
518 blk_unref(s
->target
);
521 /* We don't access the source any more. Dropping any WRITE/RESIZE is
522 * required before it could become a backing file of target_bs. */
523 bdrv_child_try_set_perm(mirror_top_bs
->backing
, 0, BLK_PERM_ALL
,
525 if (s
->backing_mode
== MIRROR_SOURCE_BACKING_CHAIN
) {
526 BlockDriverState
*backing
= s
->is_none_mode
? src
: s
->base
;
527 if (backing_bs(target_bs
) != backing
) {
528 bdrv_set_backing_hd(target_bs
, backing
, &local_err
);
530 error_report_err(local_err
);
537 replace_aio_context
= bdrv_get_aio_context(s
->to_replace
);
538 aio_context_acquire(replace_aio_context
);
541 if (s
->should_complete
&& data
->ret
== 0) {
542 BlockDriverState
*to_replace
= src
;
544 to_replace
= s
->to_replace
;
547 if (bdrv_get_flags(target_bs
) != bdrv_get_flags(to_replace
)) {
548 bdrv_reopen(target_bs
, bdrv_get_flags(to_replace
), NULL
);
551 /* The mirror job has no requests in flight any more, but we need to
552 * drain potential other users of the BDS before changing the graph. */
553 bdrv_drained_begin(target_bs
);
554 bdrv_replace_node(to_replace
, target_bs
, &local_err
);
555 bdrv_drained_end(target_bs
);
557 error_report_err(local_err
);
562 bdrv_op_unblock_all(s
->to_replace
, s
->replace_blocker
);
563 error_free(s
->replace_blocker
);
564 bdrv_unref(s
->to_replace
);
566 if (replace_aio_context
) {
567 aio_context_release(replace_aio_context
);
570 bdrv_unref(target_bs
);
572 /* Remove the mirror filter driver from the graph. Before this, get rid of
573 * the blockers on the intermediate nodes so that the resulting state is
574 * valid. Also give up permissions on mirror_top_bs->backing, which might
575 * block the removal. */
576 block_job_remove_all_bdrv(job
);
577 bdrv_child_try_set_perm(mirror_top_bs
->backing
, 0, BLK_PERM_ALL
,
579 bdrv_replace_node(mirror_top_bs
, backing_bs(mirror_top_bs
), &error_abort
);
581 /* We just changed the BDS the job BB refers to (with either or both of the
582 * bdrv_replace_node() calls), so switch the BB back so the cleanup does
583 * the right thing. We don't need any permissions any more now. */
584 blk_remove_bs(job
->blk
);
585 blk_set_perm(job
->blk
, 0, BLK_PERM_ALL
, &error_abort
);
586 blk_insert_bs(job
->blk
, mirror_top_bs
, &error_abort
);
588 block_job_completed(&s
->common
, data
->ret
);
591 bdrv_drained_end(src
);
592 bdrv_unref(mirror_top_bs
);
596 static void mirror_throttle(MirrorBlockJob
*s
)
598 int64_t now
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
);
600 if (now
- s
->last_pause_ns
> SLICE_TIME
) {
601 s
->last_pause_ns
= now
;
602 block_job_sleep_ns(&s
->common
, QEMU_CLOCK_REALTIME
, 0);
604 block_job_pause_point(&s
->common
);
608 static int coroutine_fn
mirror_dirty_init(MirrorBlockJob
*s
)
610 int64_t sector_num
, end
;
611 BlockDriverState
*base
= s
->base
;
612 BlockDriverState
*bs
= s
->source
;
613 BlockDriverState
*target_bs
= blk_bs(s
->target
);
616 end
= s
->bdev_length
/ BDRV_SECTOR_SIZE
;
618 if (base
== NULL
&& !bdrv_has_zero_init(target_bs
)) {
619 if (!bdrv_can_write_zeroes_with_unmap(target_bs
)) {
620 bdrv_set_dirty_bitmap(s
->dirty_bitmap
, 0, end
);
624 s
->initial_zeroing_ongoing
= true;
625 for (sector_num
= 0; sector_num
< end
; ) {
626 int nb_sectors
= MIN(end
- sector_num
,
627 QEMU_ALIGN_DOWN(INT_MAX
, s
->granularity
) >> BDRV_SECTOR_BITS
);
631 if (block_job_is_cancelled(&s
->common
)) {
632 s
->initial_zeroing_ongoing
= false;
636 if (s
->in_flight
>= MAX_IN_FLIGHT
) {
637 trace_mirror_yield(s
, s
->in_flight
, s
->buf_free_count
, -1);
638 mirror_wait_for_io(s
);
642 mirror_do_zero_or_discard(s
, sector_num
, nb_sectors
, false);
643 sector_num
+= nb_sectors
;
646 mirror_wait_for_all_io(s
);
647 s
->initial_zeroing_ongoing
= false;
650 /* First part, loop on the sectors and initialize the dirty bitmap. */
651 for (sector_num
= 0; sector_num
< end
; ) {
652 /* Just to make sure we are not exceeding int limit. */
653 int nb_sectors
= MIN(INT_MAX
>> BDRV_SECTOR_BITS
,
658 if (block_job_is_cancelled(&s
->common
)) {
662 ret
= bdrv_is_allocated_above(bs
, base
, sector_num
, nb_sectors
, &n
);
669 bdrv_set_dirty_bitmap(s
->dirty_bitmap
, sector_num
, n
);
676 /* Called when going out of the streaming phase to flush the bulk of the
677 * data to the medium, or just before completing.
679 static int mirror_flush(MirrorBlockJob
*s
)
681 int ret
= blk_flush(s
->target
);
683 if (mirror_error_action(s
, false, -ret
) == BLOCK_ERROR_ACTION_REPORT
) {
690 static void coroutine_fn
mirror_run(void *opaque
)
692 MirrorBlockJob
*s
= opaque
;
693 MirrorExitData
*data
;
694 BlockDriverState
*bs
= s
->source
;
695 BlockDriverState
*target_bs
= blk_bs(s
->target
);
696 bool need_drain
= true;
699 char backing_filename
[2]; /* we only need 2 characters because we are only
700 checking for a NULL string */
702 int target_cluster_size
= BDRV_SECTOR_SIZE
;
704 if (block_job_is_cancelled(&s
->common
)) {
708 s
->bdev_length
= bdrv_getlength(bs
);
709 if (s
->bdev_length
< 0) {
710 ret
= s
->bdev_length
;
714 /* Active commit must resize the base image if its size differs from the
716 if (s
->base
== blk_bs(s
->target
)) {
719 base_length
= blk_getlength(s
->target
);
720 if (base_length
< 0) {
725 if (s
->bdev_length
> base_length
) {
726 ret
= blk_truncate(s
->target
, s
->bdev_length
);
733 if (s
->bdev_length
== 0) {
734 /* Report BLOCK_JOB_READY and wait for complete. */
735 block_job_event_ready(&s
->common
);
737 while (!block_job_is_cancelled(&s
->common
) && !s
->should_complete
) {
738 block_job_yield(&s
->common
);
740 s
->common
.cancelled
= false;
744 length
= DIV_ROUND_UP(s
->bdev_length
, s
->granularity
);
745 s
->in_flight_bitmap
= bitmap_new(length
);
747 /* If we have no backing file yet in the destination, we cannot let
748 * the destination do COW. Instead, we copy sectors around the
749 * dirty data if needed. We need a bitmap to do that.
751 bdrv_get_backing_filename(target_bs
, backing_filename
,
752 sizeof(backing_filename
));
753 if (!bdrv_get_info(target_bs
, &bdi
) && bdi
.cluster_size
) {
754 target_cluster_size
= bdi
.cluster_size
;
756 if (backing_filename
[0] && !target_bs
->backing
757 && s
->granularity
< target_cluster_size
) {
758 s
->buf_size
= MAX(s
->buf_size
, target_cluster_size
);
759 s
->cow_bitmap
= bitmap_new(length
);
761 s
->target_cluster_sectors
= target_cluster_size
>> BDRV_SECTOR_BITS
;
762 s
->max_iov
= MIN(bs
->bl
.max_iov
, target_bs
->bl
.max_iov
);
764 s
->buf
= qemu_try_blockalign(bs
, s
->buf_size
);
765 if (s
->buf
== NULL
) {
772 s
->last_pause_ns
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
);
773 if (!s
->is_none_mode
) {
774 ret
= mirror_dirty_init(s
);
775 if (ret
< 0 || block_job_is_cancelled(&s
->common
)) {
781 s
->dbi
= bdrv_dirty_iter_new(s
->dirty_bitmap
, 0);
783 uint64_t delay_ns
= 0;
785 bool should_complete
;
792 block_job_pause_point(&s
->common
);
794 cnt
= bdrv_get_dirty_count(s
->dirty_bitmap
);
795 /* s->common.offset contains the number of bytes already processed so
796 * far, cnt is the number of dirty sectors remaining and
797 * s->sectors_in_flight is the number of sectors currently being
798 * processed; together those are the current total operation length */
799 s
->common
.len
= s
->common
.offset
+
800 (cnt
+ s
->sectors_in_flight
) * BDRV_SECTOR_SIZE
;
802 /* Note that even when no rate limit is applied we need to yield
803 * periodically with no pending I/O so that bdrv_drain_all() returns.
804 * We do so every SLICE_TIME nanoseconds, or when there is an error,
805 * or when the source is clean, whichever comes first.
807 delta
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
) - s
->last_pause_ns
;
808 if (delta
< SLICE_TIME
&&
809 s
->common
.iostatus
== BLOCK_DEVICE_IO_STATUS_OK
) {
810 if (s
->in_flight
>= MAX_IN_FLIGHT
|| s
->buf_free_count
== 0 ||
811 (cnt
== 0 && s
->in_flight
> 0)) {
812 trace_mirror_yield(s
, s
->in_flight
, s
->buf_free_count
, cnt
);
813 mirror_wait_for_io(s
);
815 } else if (cnt
!= 0) {
816 delay_ns
= mirror_iteration(s
);
820 should_complete
= false;
821 if (s
->in_flight
== 0 && cnt
== 0) {
822 trace_mirror_before_flush(s
);
824 if (mirror_flush(s
) < 0) {
825 /* Go check s->ret. */
828 /* We're out of the streaming phase. From now on, if the job
829 * is cancelled we will actually complete all pending I/O and
830 * report completion. This way, block-job-cancel will leave
831 * the target in a consistent state.
833 block_job_event_ready(&s
->common
);
837 should_complete
= s
->should_complete
||
838 block_job_is_cancelled(&s
->common
);
839 cnt
= bdrv_get_dirty_count(s
->dirty_bitmap
);
842 if (cnt
== 0 && should_complete
) {
843 /* The dirty bitmap is not updated while operations are pending.
844 * If we're about to exit, wait for pending operations before
845 * calling bdrv_get_dirty_count(bs), or we may exit while the
846 * source has dirty data to copy!
848 * Note that I/O can be submitted by the guest while
849 * mirror_populate runs, so pause it now. Before deciding
850 * whether to switch to target check one last time if I/O has
851 * come in the meanwhile, and if not flush the data to disk.
853 trace_mirror_before_drain(s
, cnt
);
855 bdrv_drained_begin(bs
);
856 cnt
= bdrv_get_dirty_count(s
->dirty_bitmap
);
857 if (cnt
> 0 || mirror_flush(s
) < 0) {
858 bdrv_drained_end(bs
);
862 /* The two disks are in sync. Exit and report successful
865 assert(QLIST_EMPTY(&bs
->tracked_requests
));
866 s
->common
.cancelled
= false;
872 trace_mirror_before_sleep(s
, cnt
, s
->synced
, delay_ns
);
874 block_job_sleep_ns(&s
->common
, QEMU_CLOCK_REALTIME
, delay_ns
);
875 if (block_job_is_cancelled(&s
->common
)) {
878 } else if (!should_complete
) {
879 delay_ns
= (s
->in_flight
== 0 && cnt
== 0 ? SLICE_TIME
: 0);
880 block_job_sleep_ns(&s
->common
, QEMU_CLOCK_REALTIME
, delay_ns
);
882 s
->last_pause_ns
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
);
886 if (s
->in_flight
> 0) {
887 /* We get here only if something went wrong. Either the job failed,
888 * or it was cancelled prematurely so that we do not guarantee that
889 * the target is a copy of the source.
891 assert(ret
< 0 || (!s
->synced
&& block_job_is_cancelled(&s
->common
)));
893 mirror_wait_for_all_io(s
);
896 assert(s
->in_flight
== 0);
898 g_free(s
->cow_bitmap
);
899 g_free(s
->in_flight_bitmap
);
900 bdrv_dirty_iter_free(s
->dbi
);
901 bdrv_release_dirty_bitmap(bs
, s
->dirty_bitmap
);
903 data
= g_malloc(sizeof(*data
));
907 bdrv_drained_begin(bs
);
909 block_job_defer_to_main_loop(&s
->common
, mirror_exit
, data
);
912 static void mirror_set_speed(BlockJob
*job
, int64_t speed
, Error
**errp
)
914 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
917 error_setg(errp
, QERR_INVALID_PARAMETER
, "speed");
920 ratelimit_set_speed(&s
->limit
, speed
/ BDRV_SECTOR_SIZE
, SLICE_TIME
);
923 static void mirror_complete(BlockJob
*job
, Error
**errp
)
925 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
926 BlockDriverState
*target
;
928 target
= blk_bs(s
->target
);
931 error_setg(errp
, "The active block job '%s' cannot be completed",
936 if (s
->backing_mode
== MIRROR_OPEN_BACKING_CHAIN
) {
939 assert(!target
->backing
);
940 ret
= bdrv_open_backing_file(target
, NULL
, "backing", errp
);
946 /* block all operations on to_replace bs */
948 AioContext
*replace_aio_context
;
950 s
->to_replace
= bdrv_find_node(s
->replaces
);
951 if (!s
->to_replace
) {
952 error_setg(errp
, "Node name '%s' not found", s
->replaces
);
956 replace_aio_context
= bdrv_get_aio_context(s
->to_replace
);
957 aio_context_acquire(replace_aio_context
);
959 /* TODO Translate this into permission system. Current definition of
960 * GRAPH_MOD would require to request it for the parents; they might
961 * not even be BlockDriverStates, however, so a BdrvChild can't address
962 * them. May need redefinition of GRAPH_MOD. */
963 error_setg(&s
->replace_blocker
,
964 "block device is in use by block-job-complete");
965 bdrv_op_block_all(s
->to_replace
, s
->replace_blocker
);
966 bdrv_ref(s
->to_replace
);
968 aio_context_release(replace_aio_context
);
971 s
->should_complete
= true;
972 block_job_enter(&s
->common
);
975 static void mirror_pause(BlockJob
*job
)
977 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
979 mirror_wait_for_all_io(s
);
982 static void mirror_attached_aio_context(BlockJob
*job
, AioContext
*new_context
)
984 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
986 blk_set_aio_context(s
->target
, new_context
);
989 static void mirror_drain(BlockJob
*job
)
991 MirrorBlockJob
*s
= container_of(job
, MirrorBlockJob
, common
);
993 /* Need to keep a reference in case blk_drain triggers execution
994 * of mirror_complete...
997 BlockBackend
*target
= s
->target
;
1004 static const BlockJobDriver mirror_job_driver
= {
1005 .instance_size
= sizeof(MirrorBlockJob
),
1006 .job_type
= BLOCK_JOB_TYPE_MIRROR
,
1007 .set_speed
= mirror_set_speed
,
1008 .start
= mirror_run
,
1009 .complete
= mirror_complete
,
1010 .pause
= mirror_pause
,
1011 .attached_aio_context
= mirror_attached_aio_context
,
1012 .drain
= mirror_drain
,
1015 static const BlockJobDriver commit_active_job_driver
= {
1016 .instance_size
= sizeof(MirrorBlockJob
),
1017 .job_type
= BLOCK_JOB_TYPE_COMMIT
,
1018 .set_speed
= mirror_set_speed
,
1019 .start
= mirror_run
,
1020 .complete
= mirror_complete
,
1021 .pause
= mirror_pause
,
1022 .attached_aio_context
= mirror_attached_aio_context
,
1023 .drain
= mirror_drain
,
1026 static int coroutine_fn
bdrv_mirror_top_preadv(BlockDriverState
*bs
,
1027 uint64_t offset
, uint64_t bytes
, QEMUIOVector
*qiov
, int flags
)
1029 return bdrv_co_preadv(bs
->backing
, offset
, bytes
, qiov
, flags
);
1032 static int coroutine_fn
bdrv_mirror_top_pwritev(BlockDriverState
*bs
,
1033 uint64_t offset
, uint64_t bytes
, QEMUIOVector
*qiov
, int flags
)
1035 return bdrv_co_pwritev(bs
->backing
, offset
, bytes
, qiov
, flags
);
1038 static int coroutine_fn
bdrv_mirror_top_flush(BlockDriverState
*bs
)
1040 return bdrv_co_flush(bs
->backing
->bs
);
1043 static int64_t coroutine_fn
bdrv_mirror_top_get_block_status(
1044 BlockDriverState
*bs
, int64_t sector_num
, int nb_sectors
, int *pnum
,
1045 BlockDriverState
**file
)
1048 *file
= bs
->backing
->bs
;
1049 return BDRV_BLOCK_RAW
| BDRV_BLOCK_OFFSET_VALID
| BDRV_BLOCK_DATA
|
1050 (sector_num
<< BDRV_SECTOR_BITS
);
1053 static int coroutine_fn
bdrv_mirror_top_pwrite_zeroes(BlockDriverState
*bs
,
1054 int64_t offset
, int count
, BdrvRequestFlags flags
)
1056 return bdrv_co_pwrite_zeroes(bs
->backing
, offset
, count
, flags
);
1059 static int coroutine_fn
bdrv_mirror_top_pdiscard(BlockDriverState
*bs
,
1060 int64_t offset
, int count
)
1062 return bdrv_co_pdiscard(bs
->backing
->bs
, offset
, count
);
1065 static void bdrv_mirror_top_refresh_filename(BlockDriverState
*bs
, QDict
*opts
)
1067 bdrv_refresh_filename(bs
->backing
->bs
);
1068 pstrcpy(bs
->exact_filename
, sizeof(bs
->exact_filename
),
1069 bs
->backing
->bs
->filename
);
1072 static void bdrv_mirror_top_close(BlockDriverState
*bs
)
1076 static void bdrv_mirror_top_child_perm(BlockDriverState
*bs
, BdrvChild
*c
,
1077 const BdrvChildRole
*role
,
1078 uint64_t perm
, uint64_t shared
,
1079 uint64_t *nperm
, uint64_t *nshared
)
1081 /* Must be able to forward guest writes to the real image */
1083 if (perm
& BLK_PERM_WRITE
) {
1084 *nperm
|= BLK_PERM_WRITE
;
1087 *nshared
= BLK_PERM_ALL
;
1090 /* Dummy node that provides consistent read to its users without requiring it
1091 * from its backing file and that allows writes on the backing file chain. */
1092 static BlockDriver bdrv_mirror_top
= {
1093 .format_name
= "mirror_top",
1094 .bdrv_co_preadv
= bdrv_mirror_top_preadv
,
1095 .bdrv_co_pwritev
= bdrv_mirror_top_pwritev
,
1096 .bdrv_co_pwrite_zeroes
= bdrv_mirror_top_pwrite_zeroes
,
1097 .bdrv_co_pdiscard
= bdrv_mirror_top_pdiscard
,
1098 .bdrv_co_flush
= bdrv_mirror_top_flush
,
1099 .bdrv_co_get_block_status
= bdrv_mirror_top_get_block_status
,
1100 .bdrv_refresh_filename
= bdrv_mirror_top_refresh_filename
,
1101 .bdrv_close
= bdrv_mirror_top_close
,
1102 .bdrv_child_perm
= bdrv_mirror_top_child_perm
,
1105 static void mirror_start_job(const char *job_id
, BlockDriverState
*bs
,
1106 int creation_flags
, BlockDriverState
*target
,
1107 const char *replaces
, int64_t speed
,
1108 uint32_t granularity
, int64_t buf_size
,
1109 BlockMirrorBackingMode backing_mode
,
1110 BlockdevOnError on_source_error
,
1111 BlockdevOnError on_target_error
,
1113 BlockCompletionFunc
*cb
,
1114 void *opaque
, Error
**errp
,
1115 const BlockJobDriver
*driver
,
1116 bool is_none_mode
, BlockDriverState
*base
,
1117 bool auto_complete
, const char *filter_node_name
)
1120 BlockDriverState
*mirror_top_bs
;
1121 bool target_graph_mod
;
1122 bool target_is_backing
;
1123 Error
*local_err
= NULL
;
1126 if (granularity
== 0) {
1127 granularity
= bdrv_get_default_bitmap_granularity(target
);
1130 assert ((granularity
& (granularity
- 1)) == 0);
1133 error_setg(errp
, "Invalid parameter 'buf-size'");
1137 if (buf_size
== 0) {
1138 buf_size
= DEFAULT_MIRROR_BUF_SIZE
;
1141 /* In the case of active commit, add dummy driver to provide consistent
1142 * reads on the top, while disabling it in the intermediate nodes, and make
1143 * the backing chain writable. */
1144 mirror_top_bs
= bdrv_new_open_driver(&bdrv_mirror_top
, filter_node_name
,
1146 if (mirror_top_bs
== NULL
) {
1149 mirror_top_bs
->total_sectors
= bs
->total_sectors
;
1151 /* bdrv_append takes ownership of the mirror_top_bs reference, need to keep
1152 * it alive until block_job_create() even if bs has no parent. */
1153 bdrv_ref(mirror_top_bs
);
1154 bdrv_drained_begin(bs
);
1155 bdrv_append(mirror_top_bs
, bs
, &local_err
);
1156 bdrv_drained_end(bs
);
1159 bdrv_unref(mirror_top_bs
);
1160 error_propagate(errp
, local_err
);
1164 /* Make sure that the source is not resized while the job is running */
1165 s
= block_job_create(job_id
, driver
, mirror_top_bs
,
1166 BLK_PERM_CONSISTENT_READ
,
1167 BLK_PERM_CONSISTENT_READ
| BLK_PERM_WRITE_UNCHANGED
|
1168 BLK_PERM_WRITE
| BLK_PERM_GRAPH_MOD
, speed
,
1169 creation_flags
, cb
, opaque
, errp
);
1170 bdrv_unref(mirror_top_bs
);
1175 s
->mirror_top_bs
= mirror_top_bs
;
1177 /* No resize for the target either; while the mirror is still running, a
1178 * consistent read isn't necessarily possible. We could possibly allow
1179 * writes and graph modifications, though it would likely defeat the
1180 * purpose of a mirror, so leave them blocked for now.
1182 * In the case of active commit, things look a bit different, though,
1183 * because the target is an already populated backing file in active use.
1184 * We can allow anything except resize there.*/
1185 target_is_backing
= bdrv_chain_contains(bs
, target
);
1186 target_graph_mod
= (backing_mode
!= MIRROR_LEAVE_BACKING_CHAIN
);
1187 s
->target
= blk_new(BLK_PERM_WRITE
| BLK_PERM_RESIZE
|
1188 (target_graph_mod
? BLK_PERM_GRAPH_MOD
: 0),
1189 BLK_PERM_WRITE_UNCHANGED
|
1190 (target_is_backing
? BLK_PERM_CONSISTENT_READ
|
1192 BLK_PERM_GRAPH_MOD
: 0));
1193 ret
= blk_insert_bs(s
->target
, target
, errp
);
1198 s
->replaces
= g_strdup(replaces
);
1199 s
->on_source_error
= on_source_error
;
1200 s
->on_target_error
= on_target_error
;
1201 s
->is_none_mode
= is_none_mode
;
1202 s
->backing_mode
= backing_mode
;
1204 s
->granularity
= granularity
;
1205 s
->buf_size
= ROUND_UP(buf_size
, granularity
);
1207 if (auto_complete
) {
1208 s
->should_complete
= true;
1211 s
->dirty_bitmap
= bdrv_create_dirty_bitmap(bs
, granularity
, NULL
, errp
);
1212 if (!s
->dirty_bitmap
) {
1216 /* Required permissions are already taken with blk_new() */
1217 block_job_add_bdrv(&s
->common
, "target", target
, 0, BLK_PERM_ALL
,
1220 /* In commit_active_start() all intermediate nodes disappear, so
1221 * any jobs in them must be blocked */
1222 if (target_is_backing
) {
1223 BlockDriverState
*iter
;
1224 for (iter
= backing_bs(bs
); iter
!= target
; iter
= backing_bs(iter
)) {
1225 /* XXX BLK_PERM_WRITE needs to be allowed so we don't block
1226 * ourselves at s->base (if writes are blocked for a node, they are
1227 * also blocked for its backing file). The other options would be a
1228 * second filter driver above s->base (== target). */
1229 ret
= block_job_add_bdrv(&s
->common
, "intermediate node", iter
, 0,
1230 BLK_PERM_WRITE_UNCHANGED
| BLK_PERM_WRITE
,
1238 trace_mirror_start(bs
, s
, opaque
);
1239 block_job_start(&s
->common
);
1244 g_free(s
->replaces
);
1245 blk_unref(s
->target
);
1246 block_job_unref(&s
->common
);
1249 bdrv_child_try_set_perm(mirror_top_bs
->backing
, 0, BLK_PERM_ALL
,
1251 bdrv_replace_node(mirror_top_bs
, backing_bs(mirror_top_bs
), &error_abort
);
1254 void mirror_start(const char *job_id
, BlockDriverState
*bs
,
1255 BlockDriverState
*target
, const char *replaces
,
1256 int64_t speed
, uint32_t granularity
, int64_t buf_size
,
1257 MirrorSyncMode mode
, BlockMirrorBackingMode backing_mode
,
1258 BlockdevOnError on_source_error
,
1259 BlockdevOnError on_target_error
,
1260 bool unmap
, const char *filter_node_name
, Error
**errp
)
1263 BlockDriverState
*base
;
1265 if (mode
== MIRROR_SYNC_MODE_INCREMENTAL
) {
1266 error_setg(errp
, "Sync mode 'incremental' not supported");
1269 is_none_mode
= mode
== MIRROR_SYNC_MODE_NONE
;
1270 base
= mode
== MIRROR_SYNC_MODE_TOP
? backing_bs(bs
) : NULL
;
1271 mirror_start_job(job_id
, bs
, BLOCK_JOB_DEFAULT
, target
, replaces
,
1272 speed
, granularity
, buf_size
, backing_mode
,
1273 on_source_error
, on_target_error
, unmap
, NULL
, NULL
, errp
,
1274 &mirror_job_driver
, is_none_mode
, base
, false,
1278 void commit_active_start(const char *job_id
, BlockDriverState
*bs
,
1279 BlockDriverState
*base
, int creation_flags
,
1280 int64_t speed
, BlockdevOnError on_error
,
1281 const char *filter_node_name
,
1282 BlockCompletionFunc
*cb
, void *opaque
, Error
**errp
,
1285 int orig_base_flags
;
1286 Error
*local_err
= NULL
;
1288 orig_base_flags
= bdrv_get_flags(base
);
1290 if (bdrv_reopen(base
, bs
->open_flags
, errp
)) {
1294 mirror_start_job(job_id
, bs
, creation_flags
, base
, NULL
, speed
, 0, 0,
1295 MIRROR_LEAVE_BACKING_CHAIN
,
1296 on_error
, on_error
, true, cb
, opaque
, &local_err
,
1297 &commit_active_job_driver
, false, base
, auto_complete
,
1300 error_propagate(errp
, local_err
);
1301 goto error_restore_flags
;
1306 error_restore_flags
:
1307 /* ignore error and errp for bdrv_reopen, because we want to propagate
1308 * the original error */
1309 bdrv_reopen(base
, orig_base_flags
, NULL
);