4 * Copyright (C) 2013 Proxmox Server Solutions
5 * Copyright (c) 2019 Virtuozzo International GmbH.
8 * Dietmar Maurer (dietmar@proxmox.com)
9 * Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
11 * This work is licensed under the terms of the GNU GPL, version 2 or later.
12 * See the COPYING file in the top-level directory.
15 #include "qemu/osdep.h"
18 #include "qapi/error.h"
19 #include "block/block-copy.h"
20 #include "sysemu/block-backend.h"
21 #include "qemu/units.h"
22 #include "qemu/coroutine.h"
23 #include "block/aio_task.h"
25 #define BLOCK_COPY_MAX_COPY_RANGE (16 * MiB)
26 #define BLOCK_COPY_MAX_BUFFER (1 * MiB)
27 #define BLOCK_COPY_MAX_MEM (128 * MiB)
28 #define BLOCK_COPY_MAX_WORKERS 64
30 static coroutine_fn
int block_copy_task_entry(AioTask
*task
);
32 typedef struct BlockCopyCallState
{
33 /* IN parameters. Initialized in block_copy_async() and never changed. */
39 BlockCopyAsyncCallbackFunc cb
;
42 /* Coroutine where async block-copy is running */
53 typedef struct BlockCopyTask
{
57 BlockCopyCallState
*call_state
;
61 QLIST_ENTRY(BlockCopyTask
) list
;
62 CoQueue wait_queue
; /* coroutines blocked on this task */
65 static int64_t task_end(BlockCopyTask
*task
)
67 return task
->offset
+ task
->bytes
;
70 typedef struct BlockCopyState
{
72 * BdrvChild objects are not owned or managed by block-copy. They are
73 * provided by block-copy user and user is responsible for appropriate
74 * permissions on these children.
78 BdrvDirtyBitmap
*copy_bitmap
;
79 int64_t in_flight_bytes
;
84 QLIST_HEAD(, BlockCopyTask
) tasks
;
86 BdrvRequestFlags write_flags
;
91 * Used by sync=top jobs, which first scan the source node for unallocated
92 * areas and clear them in the copy_bitmap. During this process, the bitmap
93 * is thus not fully initialized: It may still have bits set for areas that
94 * are unallocated and should actually not be copied.
96 * This is indicated by skip_unallocated.
98 * In this case, block_copy() will query the source’s allocation status,
99 * skip unallocated regions, clear them in the copy_bitmap, and invoke
100 * block_copy_reset_unallocated() every time it does.
102 bool skip_unallocated
;
104 ProgressMeter
*progress
;
105 /* progress_bytes_callback: called when some copying progress is done. */
106 ProgressBytesCallbackFunc progress_bytes_callback
;
107 void *progress_opaque
;
112 static BlockCopyTask
*find_conflicting_task(BlockCopyState
*s
,
113 int64_t offset
, int64_t bytes
)
117 QLIST_FOREACH(t
, &s
->tasks
, list
) {
118 if (offset
+ bytes
> t
->offset
&& offset
< t
->offset
+ t
->bytes
) {
127 * If there are no intersecting tasks return false. Otherwise, wait for the
128 * first found intersecting tasks to finish and return true.
130 static bool coroutine_fn
block_copy_wait_one(BlockCopyState
*s
, int64_t offset
,
133 BlockCopyTask
*task
= find_conflicting_task(s
, offset
, bytes
);
139 qemu_co_queue_wait(&task
->wait_queue
, NULL
);
145 * Search for the first dirty area in offset/bytes range and create task at
146 * the beginning of it.
148 static BlockCopyTask
*block_copy_task_create(BlockCopyState
*s
,
149 BlockCopyCallState
*call_state
,
150 int64_t offset
, int64_t bytes
)
153 int64_t max_chunk
= MIN_NON_ZERO(s
->copy_size
, call_state
->max_chunk
);
155 if (!bdrv_dirty_bitmap_next_dirty_area(s
->copy_bitmap
,
156 offset
, offset
+ bytes
,
157 max_chunk
, &offset
, &bytes
))
162 assert(QEMU_IS_ALIGNED(offset
, s
->cluster_size
));
163 bytes
= QEMU_ALIGN_UP(bytes
, s
->cluster_size
);
165 /* region is dirty, so no existent tasks possible in it */
166 assert(!find_conflicting_task(s
, offset
, bytes
));
168 bdrv_reset_dirty_bitmap(s
->copy_bitmap
, offset
, bytes
);
169 s
->in_flight_bytes
+= bytes
;
171 task
= g_new(BlockCopyTask
, 1);
172 *task
= (BlockCopyTask
) {
173 .task
.func
= block_copy_task_entry
,
175 .call_state
= call_state
,
179 qemu_co_queue_init(&task
->wait_queue
);
180 QLIST_INSERT_HEAD(&s
->tasks
, task
, list
);
186 * block_copy_task_shrink
188 * Drop the tail of the task to be handled later. Set dirty bits back and
189 * wake up all tasks waiting for us (may be some of them are not intersecting
192 static void coroutine_fn
block_copy_task_shrink(BlockCopyTask
*task
,
195 if (new_bytes
== task
->bytes
) {
199 assert(new_bytes
> 0 && new_bytes
< task
->bytes
);
201 task
->s
->in_flight_bytes
-= task
->bytes
- new_bytes
;
202 bdrv_set_dirty_bitmap(task
->s
->copy_bitmap
,
203 task
->offset
+ new_bytes
, task
->bytes
- new_bytes
);
205 task
->bytes
= new_bytes
;
206 qemu_co_queue_restart_all(&task
->wait_queue
);
209 static void coroutine_fn
block_copy_task_end(BlockCopyTask
*task
, int ret
)
211 task
->s
->in_flight_bytes
-= task
->bytes
;
213 bdrv_set_dirty_bitmap(task
->s
->copy_bitmap
, task
->offset
, task
->bytes
);
215 QLIST_REMOVE(task
, list
);
216 qemu_co_queue_restart_all(&task
->wait_queue
);
219 void block_copy_state_free(BlockCopyState
*s
)
225 bdrv_release_dirty_bitmap(s
->copy_bitmap
);
226 shres_destroy(s
->mem
);
230 static uint32_t block_copy_max_transfer(BdrvChild
*source
, BdrvChild
*target
)
232 return MIN_NON_ZERO(INT_MAX
,
233 MIN_NON_ZERO(source
->bs
->bl
.max_transfer
,
234 target
->bs
->bl
.max_transfer
));
237 BlockCopyState
*block_copy_state_new(BdrvChild
*source
, BdrvChild
*target
,
238 int64_t cluster_size
, bool use_copy_range
,
239 BdrvRequestFlags write_flags
, Error
**errp
)
242 BdrvDirtyBitmap
*copy_bitmap
;
244 copy_bitmap
= bdrv_create_dirty_bitmap(source
->bs
, cluster_size
, NULL
,
249 bdrv_disable_dirty_bitmap(copy_bitmap
);
251 s
= g_new(BlockCopyState
, 1);
252 *s
= (BlockCopyState
) {
255 .copy_bitmap
= copy_bitmap
,
256 .cluster_size
= cluster_size
,
257 .len
= bdrv_dirty_bitmap_size(copy_bitmap
),
258 .write_flags
= write_flags
,
259 .mem
= shres_create(BLOCK_COPY_MAX_MEM
),
262 if (block_copy_max_transfer(source
, target
) < cluster_size
) {
264 * copy_range does not respect max_transfer. We don't want to bother
265 * with requests smaller than block-copy cluster size, so fallback to
266 * buffered copying (read and write respect max_transfer on their
269 s
->use_copy_range
= false;
270 s
->copy_size
= cluster_size
;
271 } else if (write_flags
& BDRV_REQ_WRITE_COMPRESSED
) {
272 /* Compression supports only cluster-size writes and no copy-range. */
273 s
->use_copy_range
= false;
274 s
->copy_size
= cluster_size
;
277 * We enable copy-range, but keep small copy_size, until first
278 * successful copy_range (look at block_copy_do_copy).
280 s
->use_copy_range
= use_copy_range
;
281 s
->copy_size
= MAX(s
->cluster_size
, BLOCK_COPY_MAX_BUFFER
);
284 QLIST_INIT(&s
->tasks
);
289 void block_copy_set_progress_callback(
291 ProgressBytesCallbackFunc progress_bytes_callback
,
292 void *progress_opaque
)
294 s
->progress_bytes_callback
= progress_bytes_callback
;
295 s
->progress_opaque
= progress_opaque
;
298 void block_copy_set_progress_meter(BlockCopyState
*s
, ProgressMeter
*pm
)
304 * Takes ownership of @task
306 * If pool is NULL directly run the task, otherwise schedule it into the pool.
308 * Returns: task.func return code if pool is NULL
309 * otherwise -ECANCELED if pool status is bad
310 * otherwise 0 (successfully scheduled)
312 static coroutine_fn
int block_copy_task_run(AioTaskPool
*pool
,
316 int ret
= task
->task
.func(&task
->task
);
322 aio_task_pool_wait_slot(pool
);
323 if (aio_task_pool_status(pool
) < 0) {
324 co_put_to_shres(task
->s
->mem
, task
->bytes
);
325 block_copy_task_end(task
, -ECANCELED
);
330 aio_task_pool_start_task(pool
, &task
->task
);
338 * Do copy of cluster-aligned chunk. Requested region is allowed to exceed
339 * s->len only to cover last cluster when s->len is not aligned to clusters.
341 * No sync here: nor bitmap neighter intersecting requests handling, only copy.
343 * Returns 0 on success.
345 static int coroutine_fn
block_copy_do_copy(BlockCopyState
*s
,
346 int64_t offset
, int64_t bytes
,
347 bool zeroes
, bool *error_is_read
)
350 int64_t nbytes
= MIN(offset
+ bytes
, s
->len
) - offset
;
351 void *bounce_buffer
= NULL
;
353 assert(offset
>= 0 && bytes
> 0 && INT64_MAX
- offset
>= bytes
);
354 assert(QEMU_IS_ALIGNED(offset
, s
->cluster_size
));
355 assert(QEMU_IS_ALIGNED(bytes
, s
->cluster_size
));
356 assert(offset
< s
->len
);
357 assert(offset
+ bytes
<= s
->len
||
358 offset
+ bytes
== QEMU_ALIGN_UP(s
->len
, s
->cluster_size
));
359 assert(nbytes
< INT_MAX
);
362 ret
= bdrv_co_pwrite_zeroes(s
->target
, offset
, nbytes
, s
->write_flags
&
363 ~BDRV_REQ_WRITE_COMPRESSED
);
365 trace_block_copy_write_zeroes_fail(s
, offset
, ret
);
366 *error_is_read
= false;
371 if (s
->use_copy_range
) {
372 ret
= bdrv_co_copy_range(s
->source
, offset
, s
->target
, offset
, nbytes
,
375 trace_block_copy_copy_range_fail(s
, offset
, ret
);
376 s
->use_copy_range
= false;
377 s
->copy_size
= MAX(s
->cluster_size
, BLOCK_COPY_MAX_BUFFER
);
378 /* Fallback to read+write with allocated buffer */
380 if (s
->use_copy_range
) {
382 * Successful copy-range. Now increase copy_size. copy_range
383 * does not respect max_transfer (it's a TODO), so we factor
386 * Note: we double-check s->use_copy_range for the case when
387 * parallel block-copy request unsets it during previous
388 * bdrv_co_copy_range call.
391 MIN(MAX(s
->cluster_size
, BLOCK_COPY_MAX_COPY_RANGE
),
392 QEMU_ALIGN_DOWN(block_copy_max_transfer(s
->source
,
401 * In case of failed copy_range request above, we may proceed with buffered
402 * request larger than BLOCK_COPY_MAX_BUFFER. Still, further requests will
403 * be properly limited, so don't care too much. Moreover the most likely
404 * case (copy_range is unsupported for the configuration, so the very first
405 * copy_range request fails) is handled by setting large copy_size only
406 * after first successful copy_range.
409 bounce_buffer
= qemu_blockalign(s
->source
->bs
, nbytes
);
411 ret
= bdrv_co_pread(s
->source
, offset
, nbytes
, bounce_buffer
, 0);
413 trace_block_copy_read_fail(s
, offset
, ret
);
414 *error_is_read
= true;
418 ret
= bdrv_co_pwrite(s
->target
, offset
, nbytes
, bounce_buffer
,
421 trace_block_copy_write_fail(s
, offset
, ret
);
422 *error_is_read
= false;
427 qemu_vfree(bounce_buffer
);
432 static coroutine_fn
int block_copy_task_entry(AioTask
*task
)
434 BlockCopyTask
*t
= container_of(task
, BlockCopyTask
, task
);
435 bool error_is_read
= false;
438 ret
= block_copy_do_copy(t
->s
, t
->offset
, t
->bytes
, t
->zeroes
,
440 if (ret
< 0 && !t
->call_state
->ret
) {
441 t
->call_state
->ret
= ret
;
442 t
->call_state
->error_is_read
= error_is_read
;
444 progress_work_done(t
->s
->progress
, t
->bytes
);
445 t
->s
->progress_bytes_callback(t
->bytes
, t
->s
->progress_opaque
);
447 co_put_to_shres(t
->s
->mem
, t
->bytes
);
448 block_copy_task_end(t
, ret
);
453 static int block_copy_block_status(BlockCopyState
*s
, int64_t offset
,
454 int64_t bytes
, int64_t *pnum
)
457 BlockDriverState
*base
;
460 if (s
->skip_unallocated
) {
461 base
= bdrv_backing_chain_next(s
->source
->bs
);
466 ret
= bdrv_block_status_above(s
->source
->bs
, base
, offset
, bytes
, &num
,
468 if (ret
< 0 || num
< s
->cluster_size
) {
470 * On error or if failed to obtain large enough chunk just fallback to
473 num
= s
->cluster_size
;
474 ret
= BDRV_BLOCK_ALLOCATED
| BDRV_BLOCK_DATA
;
475 } else if (offset
+ num
== s
->len
) {
476 num
= QEMU_ALIGN_UP(num
, s
->cluster_size
);
478 num
= QEMU_ALIGN_DOWN(num
, s
->cluster_size
);
486 * Check if the cluster starting at offset is allocated or not.
487 * return via pnum the number of contiguous clusters sharing this allocation.
489 static int block_copy_is_cluster_allocated(BlockCopyState
*s
, int64_t offset
,
492 BlockDriverState
*bs
= s
->source
->bs
;
493 int64_t count
, total_count
= 0;
494 int64_t bytes
= s
->len
- offset
;
497 assert(QEMU_IS_ALIGNED(offset
, s
->cluster_size
));
500 ret
= bdrv_is_allocated(bs
, offset
, bytes
, &count
);
505 total_count
+= count
;
507 if (ret
|| count
== 0) {
509 * ret: partial segment(s) are considered allocated.
510 * otherwise: unallocated tail is treated as an entire segment.
512 *pnum
= DIV_ROUND_UP(total_count
, s
->cluster_size
);
516 /* Unallocated segment(s) with uncertain following segment(s) */
517 if (total_count
>= s
->cluster_size
) {
518 *pnum
= total_count
/ s
->cluster_size
;
528 * Reset bits in copy_bitmap starting at offset if they represent unallocated
529 * data in the image. May reset subsequent contiguous bits.
530 * @return 0 when the cluster at @offset was unallocated,
531 * 1 otherwise, and -ret on error.
533 int64_t block_copy_reset_unallocated(BlockCopyState
*s
,
534 int64_t offset
, int64_t *count
)
537 int64_t clusters
, bytes
;
539 ret
= block_copy_is_cluster_allocated(s
, offset
, &clusters
);
544 bytes
= clusters
* s
->cluster_size
;
547 bdrv_reset_dirty_bitmap(s
->copy_bitmap
, offset
, bytes
);
548 progress_set_remaining(s
->progress
,
549 bdrv_get_dirty_count(s
->copy_bitmap
) +
558 * block_copy_dirty_clusters
560 * Copy dirty clusters in @offset/@bytes range.
561 * Returns 1 if dirty clusters found and successfully copied, 0 if no dirty
562 * clusters found and -errno on failure.
564 static int coroutine_fn
565 block_copy_dirty_clusters(BlockCopyCallState
*call_state
)
567 BlockCopyState
*s
= call_state
->s
;
568 int64_t offset
= call_state
->offset
;
569 int64_t bytes
= call_state
->bytes
;
572 bool found_dirty
= false;
573 int64_t end
= offset
+ bytes
;
574 AioTaskPool
*aio
= NULL
;
577 * block_copy() user is responsible for keeping source and target in same
580 assert(bdrv_get_aio_context(s
->source
->bs
) ==
581 bdrv_get_aio_context(s
->target
->bs
));
583 assert(QEMU_IS_ALIGNED(offset
, s
->cluster_size
));
584 assert(QEMU_IS_ALIGNED(bytes
, s
->cluster_size
));
586 while (bytes
&& aio_task_pool_status(aio
) == 0) {
588 int64_t status_bytes
;
590 task
= block_copy_task_create(s
, call_state
, offset
, bytes
);
592 /* No more dirty bits in the bitmap */
593 trace_block_copy_skip_range(s
, offset
, bytes
);
596 if (task
->offset
> offset
) {
597 trace_block_copy_skip_range(s
, offset
, task
->offset
- offset
);
602 ret
= block_copy_block_status(s
, task
->offset
, task
->bytes
,
604 assert(ret
>= 0); /* never fail */
605 if (status_bytes
< task
->bytes
) {
606 block_copy_task_shrink(task
, status_bytes
);
608 if (s
->skip_unallocated
&& !(ret
& BDRV_BLOCK_ALLOCATED
)) {
609 block_copy_task_end(task
, 0);
610 progress_set_remaining(s
->progress
,
611 bdrv_get_dirty_count(s
->copy_bitmap
) +
613 trace_block_copy_skip_range(s
, task
->offset
, task
->bytes
);
614 offset
= task_end(task
);
615 bytes
= end
- offset
;
619 task
->zeroes
= ret
& BDRV_BLOCK_ZERO
;
621 trace_block_copy_process(s
, task
->offset
);
623 co_get_from_shres(s
->mem
, task
->bytes
);
625 offset
= task_end(task
);
626 bytes
= end
- offset
;
629 aio
= aio_task_pool_new(call_state
->max_workers
);
632 ret
= block_copy_task_run(aio
, task
);
640 aio_task_pool_wait_all(aio
);
643 * We are not really interested in -ECANCELED returned from
644 * block_copy_task_run. If it fails, it means some task already failed
645 * for real reason, let's return first failure.
646 * Still, assert that we don't rewrite failure by success.
648 * Note: ret may be positive here because of block-status result.
650 assert(ret
>= 0 || aio_task_pool_status(aio
) < 0);
651 ret
= aio_task_pool_status(aio
);
653 aio_task_pool_free(aio
);
656 return ret
< 0 ? ret
: found_dirty
;
662 * Copy requested region, accordingly to dirty bitmap.
663 * Collaborate with parallel block_copy requests: if they succeed it will help
664 * us. If they fail, we will retry not-copied regions. So, if we return error,
665 * it means that some I/O operation failed in context of _this_ block_copy call,
666 * not some parallel operation.
668 static int coroutine_fn
block_copy_common(BlockCopyCallState
*call_state
)
673 ret
= block_copy_dirty_clusters(call_state
);
676 ret
= block_copy_wait_one(call_state
->s
, call_state
->offset
,
681 * We retry in two cases:
682 * 1. Some progress done
683 * Something was copied, which means that there were yield points
684 * and some new dirty bits may have appeared (due to failed parallel
685 * block-copy requests).
686 * 2. We have waited for some intersecting block-copy request
687 * It may have failed and produced new dirty bits.
691 call_state
->finished
= true;
693 if (call_state
->cb
) {
694 call_state
->cb(call_state
->cb_opaque
);
700 int coroutine_fn
block_copy(BlockCopyState
*s
, int64_t start
, int64_t bytes
,
703 BlockCopyCallState call_state
= {
707 .max_workers
= BLOCK_COPY_MAX_WORKERS
,
710 int ret
= block_copy_common(&call_state
);
712 if (error_is_read
&& ret
< 0) {
713 *error_is_read
= call_state
.error_is_read
;
719 static void coroutine_fn
block_copy_async_co_entry(void *opaque
)
721 block_copy_common(opaque
);
724 BlockCopyCallState
*block_copy_async(BlockCopyState
*s
,
725 int64_t offset
, int64_t bytes
,
726 int max_workers
, int64_t max_chunk
,
727 BlockCopyAsyncCallbackFunc cb
,
730 BlockCopyCallState
*call_state
= g_new(BlockCopyCallState
, 1);
732 *call_state
= (BlockCopyCallState
) {
736 .max_workers
= max_workers
,
737 .max_chunk
= max_chunk
,
739 .cb_opaque
= cb_opaque
,
741 .co
= qemu_coroutine_create(block_copy_async_co_entry
, call_state
),
744 qemu_coroutine_enter(call_state
->co
);
749 void block_copy_call_free(BlockCopyCallState
*call_state
)
755 assert(call_state
->finished
);
759 bool block_copy_call_finished(BlockCopyCallState
*call_state
)
761 return call_state
->finished
;
764 bool block_copy_call_succeeded(BlockCopyCallState
*call_state
)
766 return call_state
->finished
&& call_state
->ret
== 0;
769 bool block_copy_call_failed(BlockCopyCallState
*call_state
)
771 return call_state
->finished
&& call_state
->ret
< 0;
774 int block_copy_call_status(BlockCopyCallState
*call_state
, bool *error_is_read
)
776 assert(call_state
->finished
);
778 *error_is_read
= call_state
->error_is_read
;
780 return call_state
->ret
;
783 BdrvDirtyBitmap
*block_copy_dirty_bitmap(BlockCopyState
*s
)
785 return s
->copy_bitmap
;
788 void block_copy_set_skip_unallocated(BlockCopyState
*s
, bool skip
)
790 s
->skip_unallocated
= skip
;