Re: [PATCH 5/6] block/block-copy: add memory limit

[Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>]

07.10.2019 18:27, Max Reitz wrote:
> On 03.10.19 19:15, Vladimir Sementsov-Ogievskiy wrote:
>> Currently total allocation for parallel requests to block-copy instance
>> is unlimited. Let's limit it to 128 MiB.
>>
>> For now block-copy is used only in backup, so actually we limit total
>> allocation for backup job.
>>
>> Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
>> ---
>>   include/block/block-copy.h | 3 +++
>>   block/block-copy.c         | 5 +++++
>>   2 files changed, 8 insertions(+)
>>
>> diff --git a/include/block/block-copy.h b/include/block/block-copy.h
>> index e2e135ff1b..bb666e7068 100644
>> --- a/include/block/block-copy.h
>> +++ b/include/block/block-copy.h
>> @@ -16,6 +16,7 @@
>>   #define BLOCK_COPY_H
>>   
>>   #include "block/block.h"
>> +#include "qemu/co-shared-amount.h"
>>   
>>   typedef struct BlockCopyInFlightReq {
>>       int64_t start_byte;
>> @@ -69,6 +70,8 @@ typedef struct BlockCopyState {
>>        */
>>       ProgressResetCallbackFunc progress_reset_callback;
>>       void *progress_opaque;
>> +
>> +    QemuCoSharedAmount *mem;
>>   } BlockCopyState;
>>   
>>   BlockCopyState *block_copy_state_new(BdrvChild *source, BdrvChild *target,
>> diff --git a/block/block-copy.c b/block/block-copy.c
>> index cc49d2345d..e700c20d0f 100644
>> --- a/block/block-copy.c
>> +++ b/block/block-copy.c
>> @@ -21,6 +21,7 @@
>>   #include "qemu/units.h"
>>   
>>   #define BLOCK_COPY_MAX_COPY_RANGE (16 * MiB)
>> +#define BLOCK_COPY_MAX_MEM (128 * MiB)
>>   
>>   static void coroutine_fn block_copy_wait_inflight_reqs(BlockCopyState *s,
>>                                                          int64_t start,
>> @@ -64,6 +65,7 @@ void block_copy_state_free(BlockCopyState *s)
>>       }
>>   
>>       bdrv_release_dirty_bitmap(s->source->bs, s->copy_bitmap);
>> +    qemu_co_shared_amount_free(s->mem);
>>       g_free(s);
>>   }
>>   
>> @@ -95,6 +97,7 @@ BlockCopyState *block_copy_state_new(BdrvChild *source, BdrvChild *target,
>>           .cluster_size = cluster_size,
>>           .len = bdrv_dirty_bitmap_size(copy_bitmap),
>>           .write_flags = write_flags,
>> +        .mem = qemu_co_shared_amount_new(BLOCK_COPY_MAX_MEM),
>>       };
>>   
>>       s->copy_range_size = QEMU_ALIGN_DOWN(max_transfer, cluster_size),
>> @@ -316,7 +319,9 @@ int coroutine_fn block_copy(BlockCopyState *s,
>>   
>>           bdrv_reset_dirty_bitmap(s->copy_bitmap, start, chunk_end - start);
>>   
>> +        qemu_co_get_amount(s->mem, chunk_end - start);
> 
> Now that I see it like this, maybe the name is too short.  This sounds
> like it was trying to get some amount of coroutines.
> 
> Would “qemu_co_get_from_shared_amount” be too long?  (Something like
> qemu_co_sham_alloc() would be funny, but maybe not.  :-)  Or maybe
> exactly because it”s funny.)
> 

hmm sham may be interpreted as shared memory, not only like shame..

And if we call it _alloc, the opposite should be _free, but how to
distinguish it from freeing the whole object? Hmm, use create/destroy for
the whole object maybe.

May be, drop "qemu_" ? It's not very informative. Or may be drop "co_"?.

I don't like shaming my shared amount :)

May be, we should imagine, what are we allocating? May be balls?

struct BallAllocator

ball_allocator_create
ball_allocator_destroy

co_try_alloc_balls
co_alloc_balls
co_free_balls

Or bars? Or which thing may be used for funny naming and to not intersect
with existing concepts like memory?

> 
>>           ret = block_copy_do_copy(s, start, chunk_end, error_is_read);
>> +        qemu_co_put_amount(s->mem, chunk_end - start);
>>           if (ret < 0) {
>>               bdrv_set_dirty_bitmap(s->copy_bitmap, start, chunk_end - start);
>>               break;
>>
> 
> 


-- 
Best regards,
Vladimir