Run block_copy iterations in parallel in aio tasks.
Changes:
- BlockCopyTask becomes aio task structure. Add zeroes field to pass
it to block_copy_do_copy
- add call state - it's a state of one call of block_copy(), shared
between parallel tasks. For now used only to keep information about
first error: is it read or not.
- convert block_copy_dirty_clusters to aio-task loop.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
---
block/block-copy.c | 104 +++++++++++++++++++++++++++++++++++++++------
1 file changed, 91 insertions(+), 13 deletions(-)
diff --git a/block/block-copy.c b/block/block-copy.c
index 5cf032c4d8..f5ef91f292 100644
--- a/block/block-copy.c
+++ b/block/block-copy.c
@@ -19,15 +19,29 @@
#include "block/block-copy.h"
#include "sysemu/block-backend.h"
#include "qemu/units.h"
+#include "qemu/coroutine.h"
+#include "block/aio_task.h"
#define BLOCK_COPY_MAX_COPY_RANGE (16 * MiB)
#define BLOCK_COPY_MAX_BUFFER (1 * MiB)
#define BLOCK_COPY_MAX_MEM (128 * MiB)
+#define BLOCK_COPY_MAX_WORKERS 64
+
+static coroutine_fn int block_copy_task_entry(AioTask *task);
+
+typedef struct BlockCopyCallState {
+ bool failed;
+ bool error_is_read;
+} BlockCopyCallState;
typedef struct BlockCopyTask {
+ AioTask task;
+
BlockCopyState *s;
+ BlockCopyCallState *call_state;
int64_t offset;
int64_t bytes;
+ bool zeroes;
QLIST_ENTRY(BlockCopyTask) list;
CoQueue wait_queue; /* coroutines blocked on this task */
} BlockCopyTask;
@@ -116,6 +130,7 @@ static bool coroutine_fn block_copy_wait_one(BlockCopyState *s, int64_t offset,
* the beginning of it.
*/
static BlockCopyTask *block_copy_task_create(BlockCopyState *s,
+ BlockCopyCallState *call_state,
int64_t offset, int64_t bytes)
{
if (!bdrv_dirty_bitmap_next_dirty_area(s->copy_bitmap,
@@ -133,7 +148,9 @@ static BlockCopyTask *block_copy_task_create(BlockCopyState *s,
BlockCopyTask *task = g_new(BlockCopyTask, 1);
*task = (BlockCopyTask) {
+ .task.func = block_copy_task_entry,
.s = s,
+ .call_state = call_state,
.offset = offset,
.bytes = bytes,
};
@@ -261,6 +278,30 @@ void block_copy_set_progress_meter(BlockCopyState *s, ProgressMeter *pm)
s->progress = pm;
}
+/* Takes ownership on @task */
+static coroutine_fn int block_copy_task_run(AioTaskPool *pool,
+ BlockCopyTask *task)
+{
+ if (!pool) {
+ int ret = task->task.func(&task->task);
+
+ g_free(task);
+ return ret;
+ }
+
+ aio_task_pool_wait_slot(pool);
+ if (aio_task_pool_status(pool) < 0) {
+ co_put_to_shres(task->s->mem, task->bytes);
+ block_copy_task_end(task, -EAGAIN);
+ g_free(task);
+ return aio_task_pool_status(pool);
+ }
+
+ aio_task_pool_start_task(pool, &task->task);
+
+ return 0;
+}
+
/*
* block_copy_do_copy
*
@@ -364,6 +405,27 @@ out:
return ret;
}
+static coroutine_fn int block_copy_task_entry(AioTask *task)
+{
+ BlockCopyTask *t = container_of(task, BlockCopyTask, task);
+ bool error_is_read;
+ int ret;
+
+ ret = block_copy_do_copy(t->s, t->offset, t->bytes, t->zeroes,
+ &error_is_read);
+ if (ret < 0 && !t->call_state->failed) {
+ t->call_state->failed = true;
+ t->call_state->error_is_read = error_is_read;
+ } else {
+ progress_work_done(t->s->progress, t->bytes);
+ t->s->progress_bytes_callback(t->bytes, t->s->progress_opaque);
+ }
+ co_put_to_shres(t->s->mem, t->bytes);
+ block_copy_task_end(t, ret);
+
+ return ret;
+}
+
static int block_copy_block_status(BlockCopyState *s, int64_t offset,
int64_t bytes, int64_t *pnum)
{
@@ -482,6 +544,8 @@ static int coroutine_fn block_copy_dirty_clusters(BlockCopyState *s,
int ret = 0;
bool found_dirty = false;
int64_t end = offset + bytes;
+ AioTaskPool *aio = NULL;
+ BlockCopyCallState call_state = {false, false};
/*
* block_copy() user is responsible for keeping source and target in same
@@ -493,11 +557,11 @@ static int coroutine_fn block_copy_dirty_clusters(BlockCopyState *s,
assert(QEMU_IS_ALIGNED(offset, s->cluster_size));
assert(QEMU_IS_ALIGNED(bytes, s->cluster_size));
- while (bytes) {
- g_autofree BlockCopyTask *task = NULL;
+ while (bytes && aio_task_pool_status(aio) == 0) {
+ BlockCopyTask *task;
int64_t status_bytes;
- task = block_copy_task_create(s, offset, bytes);
+ task = block_copy_task_create(s, &call_state, offset, bytes);
if (!task) {
/* No more dirty bits in the bitmap */
trace_block_copy_skip_range(s, offset, bytes);
@@ -517,6 +581,7 @@ static int coroutine_fn block_copy_dirty_clusters(BlockCopyState *s,
}
if (s->skip_unallocated && !(ret & BDRV_BLOCK_ALLOCATED)) {
block_copy_task_end(task, 0);
+ g_free(task);
progress_set_remaining(s->progress,
bdrv_get_dirty_count(s->copy_bitmap) +
s->in_flight_bytes);
@@ -525,25 +590,38 @@ static int coroutine_fn block_copy_dirty_clusters(BlockCopyState *s,
bytes = end - offset;
continue;
}
+ task->zeroes = ret & BDRV_BLOCK_ZERO;
trace_block_copy_process(s, task->offset);
co_get_from_shres(s->mem, task->bytes);
- ret = block_copy_do_copy(s, task->offset, task->bytes,
- ret & BDRV_BLOCK_ZERO, error_is_read);
- co_put_to_shres(s->mem, task->bytes);
- block_copy_task_end(task, ret);
- if (ret < 0) {
- return ret;
- }
- progress_work_done(s->progress, task->bytes);
- s->progress_bytes_callback(task->bytes, s->progress_opaque);
offset = task_end(task);
bytes = end - offset;
+
+ if (!aio && bytes) {
+ aio = aio_task_pool_new(BLOCK_COPY_MAX_WORKERS);
+ }
+
+ ret = block_copy_task_run(aio, task);
+ if (ret < 0) {
+ goto out;
+ }
+ }
+
+out:
+ if (aio) {
+ aio_task_pool_wait_all(aio);
+ if (ret == 0) {
+ ret = aio_task_pool_status(aio);
+ }
+ g_free(aio);
+ }
+ if (error_is_read && ret < 0) {
+ *error_is_read = call_state.error_is_read;
}
- return found_dirty;
+ return ret < 0 ? ret : found_dirty;
}
/*
--
2.21.0
On 29.04.20 08:10, Vladimir Sementsov-Ogievskiy wrote:
> Run block_copy iterations in parallel in aio tasks.
>
> Changes:
> - BlockCopyTask becomes aio task structure. Add zeroes field to pass
> it to block_copy_do_copy
> - add call state - it's a state of one call of block_copy(), shared
> between parallel tasks. For now used only to keep information about
> first error: is it read or not.
> - convert block_copy_dirty_clusters to aio-task loop.
>
> Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
> ---
> block/block-copy.c | 104 +++++++++++++++++++++++++++++++++++++++------
> 1 file changed, 91 insertions(+), 13 deletions(-)
>
> diff --git a/block/block-copy.c b/block/block-copy.c
> index 5cf032c4d8..f5ef91f292 100644
> --- a/block/block-copy.c
> +++ b/block/block-copy.c
[...]
> @@ -261,6 +278,30 @@ void block_copy_set_progress_meter(BlockCopyState *s, ProgressMeter *pm)
> s->progress = pm;
> }
>
> +/* Takes ownership on @task */
Still *of
> +static coroutine_fn int block_copy_task_run(AioTaskPool *pool,
> + BlockCopyTask *task)
> +{
> + if (!pool) {
> + int ret = task->task.func(&task->task);
> +
> + g_free(task);
> + return ret;
> + }
> +
> + aio_task_pool_wait_slot(pool);
> + if (aio_task_pool_status(pool) < 0) {
> + co_put_to_shres(task->s->mem, task->bytes);
> + block_copy_task_end(task, -EAGAIN);
It looks like you may have missed my nit picks on v2 regarding this
patch, so I’m going to ask again whether -ECANCELED might be better here
(even though it still doesn’t really matter).
> + g_free(task);
> + return aio_task_pool_status(pool);
And whether it may be better to return a constant like -ECANCELED here,
because how a previous task failed shouldn’t really concern this task
(or its error code).
> + }
> +
> + aio_task_pool_start_task(pool, &task->task);
> +
> + return 0;
> +}
> +
> /*
> * block_copy_do_copy
> *
[...]
> @@ -525,25 +590,38 @@ static int coroutine_fn block_copy_dirty_clusters(BlockCopyState *s,
[...]
> +out:
> + if (aio) {
> + aio_task_pool_wait_all(aio);
> + if (ret == 0) {
> + ret = aio_task_pool_status(aio);
> + }
> + g_free(aio);
I’d still prefer aio_task_pool_free().
Max
29.04.2020 14:55, Max Reitz wrote:
> On 29.04.20 08:10, Vladimir Sementsov-Ogievskiy wrote:
>> Run block_copy iterations in parallel in aio tasks.
>>
>> Changes:
>> - BlockCopyTask becomes aio task structure. Add zeroes field to pass
>> it to block_copy_do_copy
>> - add call state - it's a state of one call of block_copy(), shared
>> between parallel tasks. For now used only to keep information about
>> first error: is it read or not.
>> - convert block_copy_dirty_clusters to aio-task loop.
>>
>> Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
>> ---
>> block/block-copy.c | 104 +++++++++++++++++++++++++++++++++++++++------
>> 1 file changed, 91 insertions(+), 13 deletions(-)
>>
>> diff --git a/block/block-copy.c b/block/block-copy.c
>> index 5cf032c4d8..f5ef91f292 100644
>> --- a/block/block-copy.c
>> +++ b/block/block-copy.c
>
> [...]
>
>> @@ -261,6 +278,30 @@ void block_copy_set_progress_meter(BlockCopyState *s, ProgressMeter *pm)
>> s->progress = pm;
>> }
>>
>> +/* Takes ownership on @task */
>
> Still *of
Ohhh, very sorry for it, I really forget to update the patch :(
>
>> +static coroutine_fn int block_copy_task_run(AioTaskPool *pool,
>> + BlockCopyTask *task)
>> +{
>> + if (!pool) {
>> + int ret = task->task.func(&task->task);
>> +
>> + g_free(task);
>> + return ret;
>> + }
>> +
>> + aio_task_pool_wait_slot(pool);
>> + if (aio_task_pool_status(pool) < 0) {
>> + co_put_to_shres(task->s->mem, task->bytes);
>> + block_copy_task_end(task, -EAGAIN);
>
> It looks like you may have missed my nit picks on v2 regarding this
> patch, so I’m going to ask again whether -ECANCELED might be better here
> (even though it still doesn’t really matter).
Hmm yes, sounds better in the context. And I don't see any specific usage of it, and don't remember why I've chosen EAGAIN :) Let's use ECANCELED.
>
>> + g_free(task);
>> + return aio_task_pool_status(pool);
>
> And whether it may be better to return a constant like -ECANCELED here,
> because how a previous task failed shouldn’t really concern this task
> (or its error code).
Looks correct, will change.
>
>> + }
>> +
>> + aio_task_pool_start_task(pool, &task->task);
>> +
>> + return 0;
>> +}
>> +
>> /*
>> * block_copy_do_copy
>> *
>
> [...]
>
>> @@ -525,25 +590,38 @@ static int coroutine_fn block_copy_dirty_clusters(BlockCopyState *s,
>
> [...]
>
>> +out:
>> + if (aio) {
>> + aio_task_pool_wait_all(aio);
>> + if (ret == 0) {
>> + ret = aio_task_pool_status(aio);
>> + }
>> + g_free(aio);
>
> I’d still prefer aio_task_pool_free().
>
a thousand apologies :(
--
Best regards,
Vladimir
© 2016 - 2026 Red Hat, Inc.