[FFmpeg-devel] [PATCH 1/5] avutil/buffer: add av_buffer_pool_flush()

[James Almer]

On 12/9/2020 8:06 PM, Lynne wrote:
> Dec 9, 2020, 23:42 by jonas at kwiboo.se:
> 
>> On 2020-12-09 23:09, Lynne wrote:
>>
>>> Dec 9, 2020, 21:25 by jonas at kwiboo.se:
>>>
>>>> Signed-off-by: Jonas Karlman <jonas at kwiboo.se>
>>>> ---
>>>>   doc/APIchanges      |  3 +++
>>>>   libavutil/buffer.c  | 13 +++++++++++++
>>>>   libavutil/buffer.h  |  5 +++++
>>>>   libavutil/version.h |  2 +-
>>>>   4 files changed, 22 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/doc/APIchanges b/doc/APIchanges
>>>> index 3fb9e12525..4a739ce453 100644
>>>> --- a/doc/APIchanges
>>>> +++ b/doc/APIchanges
>>>> @@ -15,6 +15,9 @@ libavutil:     2017-10-21
>>>>   
>>>>   API changes, most recent first:
>>>>   
>>>> +2020-xx-xx - xxxxxxxxxx - lavu 56.63.100 - buffer.h
>>>> +  Add av_buffer_pool_flush().
>>>> +
>>>>   2020-12-03 - xxxxxxxxxx - lavu 56.62.100 - timecode.h
>>>>   Add av_timecode_init_from_components.
>>>>   
>>>> diff --git a/libavutil/buffer.c b/libavutil/buffer.c
>>>> index d67b4bbdaf..a0683664cf 100644
>>>> --- a/libavutil/buffer.c
>>>> +++ b/libavutil/buffer.c
>>>> @@ -300,6 +300,19 @@ static void buffer_pool_free(AVBufferPool *pool)
>>>>   av_freep(&pool);
>>>>   }
>>>>   
>>>> +void av_buffer_pool_flush(AVBufferPool *pool)
>>>> +{
>>>> +    ff_mutex_lock(&pool->mutex);
>>>> +    while (pool->pool) {
>>>> +        BufferPoolEntry *buf = pool->pool;
>>>> +        pool->pool = buf->next;
>>>> +
>>>> +        buf->free(buf->opaque, buf->data);
>>>> +        av_freep(&buf);
>>>> +    }
>>>> +    ff_mutex_unlock(&pool->mutex);
>>>> +}
>>>>
>>>
>>> This frees all buffers from the pool, which an API user probably has references to,
>>> leading to very nasty crashes. I can't see how this is even useful nor needed.
>>>
>>
>> This function should only free buffers that has been returned to the pool and
>> is once again part of the pool->pool linked list.
>>
>> Main use-case is to flush and free all returned buffers in a pool used by a
>> hwaccel ctx while a new hwaccel ctx is being initialized, i.e. during seek
>> of H.264 video and a SPS change triggers initialization of a new hwaccel ctx.
>>
>> For devices with memory constraints keeping all previously allocated buffers
>> for the old hwaccel ctx around while trying to allocate new buffers for a new
>> hwaccel ctx tend to cause to much memory usage, especially for 2160p video.
>>
>> This function works around this limitation by freeing all buffers that has
>> already been returned to the pool during hwaccel uninit, before the last
>> buffer is returned, the last buffer is most of the time being displayed and
>> is not returned to the pool until next frame has been decoded and displayed
>> from next hwaccel ctx.
>> Without this the pool buffers is only freed once the last buffer is returned.
>>
>> Please note that I may have mixed up hwaccel and hw frames ctx above :-)
>>
>> This function is being called from ff_v4l2_request_uninit and
>> v4l2_request_hwframe_ctx_free in next patch.
>>
> 
> Why can't av_buffer_pool_uninit() free all the buffers currently returned in the pool?
> You only seem to be calling it the function during uninitialization.

av_buffer_pool_uninit() frees them only after *all* the buffers have 
been returned. In the example Jonas gave above, there may be one or two 
frames created from buffers allocated by the pool still active and yet 
to be unref'd by the time another hwcontext is created and new buffers 
from a new pool start being allocated.

His intention is to free all the buffers currently held by the pool (as 
they will not be reused anymore) to free as much memory as possible. The 
remaining buffers will be freed once the few remaining frames using them 
are unref'd.