[FFmpeg-devel] [PATCH] avfilter/framepool: fix adjustment that can crash filtering

James Almer jamrial at gmail.com
Sun Feb 20 19:03:14 EET 2022


On 2/20/2022 1:15 PM, Anton Khirnov wrote:
> Quoting Paul B Mahol (2022-02-18 11:46:05)
>> Fixes #9551.
>>
>> Signed-off-by: Paul B Mahol <onemda at gmail.com>
>> ---
>>   libavfilter/framepool.c | 2 +-
>>   1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/libavfilter/framepool.c b/libavfilter/framepool.c
>> index 7c63807df3..aab408d355 100644
>> --- a/libavfilter/framepool.c
>> +++ b/libavfilter/framepool.c
>> @@ -96,7 +96,7 @@ FFFramePool *ff_frame_pool_video_init(AVBufferRef* (*alloc)(size_t size),
>>           if (i == 1 || i == 2)
>>               h = AV_CEIL_RSHIFT(h, desc->log2_chroma_h);
>>   
>> -        pool->pools[i] = av_buffer_pool_init(pool->linesize[i] * h + 16 + 16 - 1,
>> +        pool->pools[i] = av_buffer_pool_init(pool->linesize[i] * h + 16 + 16 - 0,
>>                                                alloc);
> 
> all these magic constants are extremely non-obvious, why are they there
> and why does removing that 1 crash anything?

They are probably cargo culting from lavu's av_frame_get_buffer() and in 
turn from lavc's avcodec_default_get_buffer2().
In the latter, the padding is 16 + STRIDE_ALIGN - 1, where STRIDE_ALIGN 
is the size of the highest simd register enabled at configure time (16 
for sse/neon/altivec, 32 for avx, 64 for avx512).
In the former, the padding is 16 + 16 - 1, with a comment in one of 
those 16 that it's meant to be STRIDE_ALIGN, a lavc internal define, 
which means it's outdated and out of sync.

What the first 16 or the -1 are, i have no idea.


More information about the ffmpeg-devel mailing list