4 * Copyright (C) 2013 Proxmox Server Solutions
5 * Copyright (c) 2019 Virtuozzo International GmbH.
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
29 #define BLOCK_COPY_SLICE_TIME 100000000ULL /* ns */
31 static coroutine_fn int block_copy_task_entry(AioTask *task);
33 typedef struct BlockCopyCallState {
34 /* IN parameters. Initialized in block_copy_async() and never changed. */
40 bool ignore_ratelimit;
41 BlockCopyAsyncCallbackFunc cb;
44 /* Coroutine where async block-copy is running */
47 /* To reference all call states from BlockCopyState */
48 QLIST_ENTRY(BlockCopyCallState) list;
53 QemuCoSleepState *sleep_state;
60 typedef struct BlockCopyTask {
64 BlockCopyCallState *call_state;
68 QLIST_ENTRY(BlockCopyTask) list;
69 CoQueue wait_queue; /* coroutines blocked on this task */
72 static int64_t task_end(BlockCopyTask *task)
74 return task->offset + task->bytes;
77 typedef struct BlockCopyState {
79 * BdrvChild objects are not owned or managed by block-copy. They are
80 * provided by block-copy user and user is responsible for appropriate
81 * permissions on these children.
85 BdrvDirtyBitmap *copy_bitmap;
86 int64_t in_flight_bytes;
91 QLIST_HEAD(, BlockCopyTask) tasks; /* All tasks from all block-copy calls */
92 QLIST_HEAD(, BlockCopyCallState) calls;
94 BdrvRequestFlags write_flags;
99 * Used by sync=top jobs, which first scan the source node for unallocated
100 * areas and clear them in the copy_bitmap. During this process, the bitmap
101 * is thus not fully initialized: It may still have bits set for areas that
102 * are unallocated and should actually not be copied.
104 * This is indicated by skip_unallocated.
106 * In this case, block_copy() will query the source’s allocation status,
107 * skip unallocated regions, clear them in the copy_bitmap, and invoke
108 * block_copy_reset_unallocated() every time it does.
110 bool skip_unallocated;
112 ProgressMeter *progress;
117 RateLimit rate_limit;
120 static BlockCopyTask *find_conflicting_task(BlockCopyState *s,
121 int64_t offset, int64_t bytes)
125 QLIST_FOREACH(t, &s->tasks, list) {
126 if (offset + bytes > t->offset && offset < t->offset + t->bytes) {
135 * If there are no intersecting tasks return false. Otherwise, wait for the
136 * first found intersecting tasks to finish and return true.
138 static bool coroutine_fn block_copy_wait_one(BlockCopyState *s, int64_t offset,
141 BlockCopyTask *task = find_conflicting_task(s, offset, bytes);
147 qemu_co_queue_wait(&task->wait_queue, NULL);
153 * Search for the first dirty area in offset/bytes range and create task at
154 * the beginning of it.
156 static BlockCopyTask *block_copy_task_create(BlockCopyState *s,
157 BlockCopyCallState *call_state,
158 int64_t offset, int64_t bytes)
161 int64_t max_chunk = MIN_NON_ZERO(s->copy_size, call_state->max_chunk);
163 if (!bdrv_dirty_bitmap_next_dirty_area(s->copy_bitmap,
164 offset, offset + bytes,
165 max_chunk, &offset, &bytes))
170 assert(QEMU_IS_ALIGNED(offset, s->cluster_size));
171 bytes = QEMU_ALIGN_UP(bytes, s->cluster_size);
173 /* region is dirty, so no existent tasks possible in it */
174 assert(!find_conflicting_task(s, offset, bytes));
176 bdrv_reset_dirty_bitmap(s->copy_bitmap, offset, bytes);
177 s->in_flight_bytes += bytes;
179 task = g_new(BlockCopyTask, 1);
180 *task = (BlockCopyTask) {
181 .task.func = block_copy_task_entry,
183 .call_state = call_state,
187 qemu_co_queue_init(&task->wait_queue);
188 QLIST_INSERT_HEAD(&s->tasks, task, list);
194 * block_copy_task_shrink
196 * Drop the tail of the task to be handled later. Set dirty bits back and
197 * wake up all tasks waiting for us (may be some of them are not intersecting
200 static void coroutine_fn block_copy_task_shrink(BlockCopyTask *task,
203 if (new_bytes == task->bytes) {
207 assert(new_bytes > 0 && new_bytes < task->bytes);
209 task->s->in_flight_bytes -= task->bytes - new_bytes;
210 bdrv_set_dirty_bitmap(task->s->copy_bitmap,
211 task->offset + new_bytes, task->bytes - new_bytes);
213 task->bytes = new_bytes;
214 qemu_co_queue_restart_all(&task->wait_queue);
217 static void coroutine_fn block_copy_task_end(BlockCopyTask *task, int ret)
219 task->s->in_flight_bytes -= task->bytes;
221 bdrv_set_dirty_bitmap(task->s->copy_bitmap, task->offset, task->bytes);
223 QLIST_REMOVE(task, list);
224 qemu_co_queue_restart_all(&task->wait_queue);
227 void block_copy_state_free(BlockCopyState *s)
233 bdrv_release_dirty_bitmap(s->copy_bitmap);
234 shres_destroy(s->mem);
238 static uint32_t block_copy_max_transfer(BdrvChild *source, BdrvChild *target)
240 return MIN_NON_ZERO(INT_MAX,
241 MIN_NON_ZERO(source->bs->bl.max_transfer,
242 target->bs->bl.max_transfer));
245 BlockCopyState *block_copy_state_new(BdrvChild *source, BdrvChild *target,
246 int64_t cluster_size, bool use_copy_range,
247 BdrvRequestFlags write_flags, Error **errp)
250 BdrvDirtyBitmap *copy_bitmap;
252 copy_bitmap = bdrv_create_dirty_bitmap(source->bs, cluster_size, NULL,
257 bdrv_disable_dirty_bitmap(copy_bitmap);
259 s = g_new(BlockCopyState, 1);
260 *s = (BlockCopyState) {
263 .copy_bitmap = copy_bitmap,
264 .cluster_size = cluster_size,
265 .len = bdrv_dirty_bitmap_size(copy_bitmap),
266 .write_flags = write_flags,
267 .mem = shres_create(BLOCK_COPY_MAX_MEM),
270 if (block_copy_max_transfer(source, target) < cluster_size) {
272 * copy_range does not respect max_transfer. We don't want to bother
273 * with requests smaller than block-copy cluster size, so fallback to
274 * buffered copying (read and write respect max_transfer on their
277 s->use_copy_range = false;
278 s->copy_size = cluster_size;
279 } else if (write_flags & BDRV_REQ_WRITE_COMPRESSED) {
280 /* Compression supports only cluster-size writes and no copy-range. */
281 s->use_copy_range = false;
282 s->copy_size = cluster_size;
285 * We enable copy-range, but keep small copy_size, until first
286 * successful copy_range (look at block_copy_do_copy).
288 s->use_copy_range = use_copy_range;
289 s->copy_size = MAX(s->cluster_size, BLOCK_COPY_MAX_BUFFER);
292 QLIST_INIT(&s->tasks);
293 QLIST_INIT(&s->calls);
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);
446 co_put_to_shres(t->s->mem, t->bytes);
447 block_copy_task_end(t, ret);
452 static int block_copy_block_status(BlockCopyState *s, int64_t offset,
453 int64_t bytes, int64_t *pnum)
456 BlockDriverState *base;
459 if (s->skip_unallocated) {
460 base = bdrv_backing_chain_next(s->source->bs);
465 ret = bdrv_block_status_above(s->source->bs, base, offset, bytes, &num,
467 if (ret < 0 || num < s->cluster_size) {
469 * On error or if failed to obtain large enough chunk just fallback to
472 num = s->cluster_size;
473 ret = BDRV_BLOCK_ALLOCATED | BDRV_BLOCK_DATA;
474 } else if (offset + num == s->len) {
475 num = QEMU_ALIGN_UP(num, s->cluster_size);
477 num = QEMU_ALIGN_DOWN(num, s->cluster_size);
485 * Check if the cluster starting at offset is allocated or not.
486 * return via pnum the number of contiguous clusters sharing this allocation.
488 static int block_copy_is_cluster_allocated(BlockCopyState *s, int64_t offset,
491 BlockDriverState *bs = s->source->bs;
492 int64_t count, total_count = 0;
493 int64_t bytes = s->len - offset;
496 assert(QEMU_IS_ALIGNED(offset, s->cluster_size));
499 ret = bdrv_is_allocated(bs, offset, bytes, &count);
504 total_count += count;
506 if (ret || count == 0) {
508 * ret: partial segment(s) are considered allocated.
509 * otherwise: unallocated tail is treated as an entire segment.
511 *pnum = DIV_ROUND_UP(total_count, s->cluster_size);
515 /* Unallocated segment(s) with uncertain following segment(s) */
516 if (total_count >= s->cluster_size) {
517 *pnum = total_count / s->cluster_size;
527 * Reset bits in copy_bitmap starting at offset if they represent unallocated
528 * data in the image. May reset subsequent contiguous bits.
529 * @return 0 when the cluster at @offset was unallocated,
530 * 1 otherwise, and -ret on error.
532 int64_t block_copy_reset_unallocated(BlockCopyState *s,
533 int64_t offset, int64_t *count)
536 int64_t clusters, bytes;
538 ret = block_copy_is_cluster_allocated(s, offset, &clusters);
543 bytes = clusters * s->cluster_size;
546 bdrv_reset_dirty_bitmap(s->copy_bitmap, offset, bytes);
547 progress_set_remaining(s->progress,
548 bdrv_get_dirty_count(s->copy_bitmap) +
557 * block_copy_dirty_clusters
559 * Copy dirty clusters in @offset/@bytes range.
560 * Returns 1 if dirty clusters found and successfully copied, 0 if no dirty
561 * clusters found and -errno on failure.
563 static int coroutine_fn
564 block_copy_dirty_clusters(BlockCopyCallState *call_state)
566 BlockCopyState *s = call_state->s;
567 int64_t offset = call_state->offset;
568 int64_t bytes = call_state->bytes;
571 bool found_dirty = false;
572 int64_t end = offset + bytes;
573 AioTaskPool *aio = NULL;
576 * block_copy() user is responsible for keeping source and target in same
579 assert(bdrv_get_aio_context(s->source->bs) ==
580 bdrv_get_aio_context(s->target->bs));
582 assert(QEMU_IS_ALIGNED(offset, s->cluster_size));
583 assert(QEMU_IS_ALIGNED(bytes, s->cluster_size));
585 while (bytes && aio_task_pool_status(aio) == 0 && !call_state->cancelled) {
587 int64_t status_bytes;
589 task = block_copy_task_create(s, call_state, offset, bytes);
591 /* No more dirty bits in the bitmap */
592 trace_block_copy_skip_range(s, offset, bytes);
595 if (task->offset > offset) {
596 trace_block_copy_skip_range(s, offset, task->offset - offset);
601 ret = block_copy_block_status(s, task->offset, task->bytes,
603 assert(ret >= 0); /* never fail */
604 if (status_bytes < task->bytes) {
605 block_copy_task_shrink(task, status_bytes);
607 if (s->skip_unallocated && !(ret & BDRV_BLOCK_ALLOCATED)) {
608 block_copy_task_end(task, 0);
609 progress_set_remaining(s->progress,
610 bdrv_get_dirty_count(s->copy_bitmap) +
612 trace_block_copy_skip_range(s, task->offset, task->bytes);
613 offset = task_end(task);
614 bytes = end - offset;
618 task->zeroes = ret & BDRV_BLOCK_ZERO;
621 if (!call_state->ignore_ratelimit) {
622 uint64_t ns = ratelimit_calculate_delay(&s->rate_limit, 0);
624 block_copy_task_end(task, -EAGAIN);
626 qemu_co_sleep_ns_wakeable(QEMU_CLOCK_REALTIME, ns,
627 &call_state->sleep_state);
632 ratelimit_calculate_delay(&s->rate_limit, task->bytes);
635 trace_block_copy_process(s, task->offset);
637 co_get_from_shres(s->mem, task->bytes);
639 offset = task_end(task);
640 bytes = end - offset;
643 aio = aio_task_pool_new(call_state->max_workers);
646 ret = block_copy_task_run(aio, task);
654 aio_task_pool_wait_all(aio);
657 * We are not really interested in -ECANCELED returned from
658 * block_copy_task_run. If it fails, it means some task already failed
659 * for real reason, let's return first failure.
660 * Still, assert that we don't rewrite failure by success.
662 * Note: ret may be positive here because of block-status result.
664 assert(ret >= 0 || aio_task_pool_status(aio) < 0);
665 ret = aio_task_pool_status(aio);
667 aio_task_pool_free(aio);
670 return ret < 0 ? ret : found_dirty;
673 void block_copy_kick(BlockCopyCallState *call_state)
675 if (call_state->sleep_state) {
676 qemu_co_sleep_wake(call_state->sleep_state);
683 * Copy requested region, accordingly to dirty bitmap.
684 * Collaborate with parallel block_copy requests: if they succeed it will help
685 * us. If they fail, we will retry not-copied regions. So, if we return error,
686 * it means that some I/O operation failed in context of _this_ block_copy call,
687 * not some parallel operation.
689 static int coroutine_fn block_copy_common(BlockCopyCallState *call_state)
693 QLIST_INSERT_HEAD(&call_state->s->calls, call_state, list);
696 ret = block_copy_dirty_clusters(call_state);
698 if (ret == 0 && !call_state->cancelled) {
699 ret = block_copy_wait_one(call_state->s, call_state->offset,
704 * We retry in two cases:
705 * 1. Some progress done
706 * Something was copied, which means that there were yield points
707 * and some new dirty bits may have appeared (due to failed parallel
708 * block-copy requests).
709 * 2. We have waited for some intersecting block-copy request
710 * It may have failed and produced new dirty bits.
712 } while (ret > 0 && !call_state->cancelled);
714 call_state->finished = true;
716 if (call_state->cb) {
717 call_state->cb(call_state->cb_opaque);
720 QLIST_REMOVE(call_state, list);
725 int coroutine_fn block_copy(BlockCopyState *s, int64_t start, int64_t bytes,
726 bool ignore_ratelimit, bool *error_is_read)
728 BlockCopyCallState call_state = {
732 .ignore_ratelimit = ignore_ratelimit,
733 .max_workers = BLOCK_COPY_MAX_WORKERS,
736 int ret = block_copy_common(&call_state);
738 if (error_is_read && ret < 0) {
739 *error_is_read = call_state.error_is_read;
745 static void coroutine_fn block_copy_async_co_entry(void *opaque)
747 block_copy_common(opaque);
750 BlockCopyCallState *block_copy_async(BlockCopyState *s,
751 int64_t offset, int64_t bytes,
752 int max_workers, int64_t max_chunk,
753 BlockCopyAsyncCallbackFunc cb,
756 BlockCopyCallState *call_state = g_new(BlockCopyCallState, 1);
758 *call_state = (BlockCopyCallState) {
762 .max_workers = max_workers,
763 .max_chunk = max_chunk,
765 .cb_opaque = cb_opaque,
767 .co = qemu_coroutine_create(block_copy_async_co_entry, call_state),
770 qemu_coroutine_enter(call_state->co);
775 void block_copy_call_free(BlockCopyCallState *call_state)
781 assert(call_state->finished);
785 bool block_copy_call_finished(BlockCopyCallState *call_state)
787 return call_state->finished;
790 bool block_copy_call_succeeded(BlockCopyCallState *call_state)
792 return call_state->finished && !call_state->cancelled &&
793 call_state->ret == 0;
796 bool block_copy_call_failed(BlockCopyCallState *call_state)
798 return call_state->finished && !call_state->cancelled &&
802 bool block_copy_call_cancelled(BlockCopyCallState *call_state)
804 return call_state->cancelled;
807 int block_copy_call_status(BlockCopyCallState *call_state, bool *error_is_read)
809 assert(call_state->finished);
811 *error_is_read = call_state->error_is_read;
813 return call_state->ret;
816 void block_copy_call_cancel(BlockCopyCallState *call_state)
818 call_state->cancelled = true;
819 block_copy_kick(call_state);
822 BdrvDirtyBitmap *block_copy_dirty_bitmap(BlockCopyState *s)
824 return s->copy_bitmap;
827 void block_copy_set_skip_unallocated(BlockCopyState *s, bool skip)
829 s->skip_unallocated = skip;
832 void block_copy_set_speed(BlockCopyState *s, uint64_t speed)
836 ratelimit_set_speed(&s->rate_limit, speed, BLOCK_COPY_SLICE_TIME);
840 * Note: it's good to kick all call states from here, but it should be done
841 * only from a coroutine, to not crash if s->calls list changed while
842 * entering one call. So for now, the only user of this function kicks its
843 * only one call_state by hand.