[FFmpeg-devel] [PATCH v3] avformat/avisynth: fix segfault when also compiled with libvmaf

[Daniel Bermond]

On 7/27/25 21:34, Kacper Michajlow wrote:
> On Mon, 28 Jul 2025 at 02:04, Daniel Bermond <danielbermond at gmail.com> wrote:
>>
>> When FFmpeg is compiled with support for both avisynth and libvmaf,
>> a segmentation fault occurs when using avisynth (.avs) input.
>>
>> This happens because both avisynthplus and vmaf have the exactly
>> same C++ symbol 'Cache::~Cache()'[1][2], which is a C++ destructor.
>> When using avisynth input, this destructor is also called in the
>> vmaf side, although no vmaf object was created, thus causing a
>> segmentation fault.
>>
>> Without changing this conflicting symbol in avisynthplus and/or
>> vmaf upstream code, a solution is to switch the 'RTLD_LOCAL'[3]
> 
> I'm not opposed to a workaround to the issues as an interim solution,
> but has this issue been reported to upstream? Classes with such
> generic names should be put in a namespace in the respective projects.

Thank you for the reply.

I have not reported this upstream. Both projects does not follow
such rules as encapsulating classes in a unique namespaces and/or
uniquely naming the classes. They simply use generic class names
in the top namespace of the public API. This is their code
design/style/standards, present the entire source code.
The programmer/manager is supposedly to be aware of such issues
in the code design stage, and yet they decided to be like this.

> Additionally it would be good for those projects to be built with
> -fvisibility=hidden to reduce the amount of exposed symbols and only
> expose API entry points.

Compiling avisynthplus with '-fvisibility=hidden' makes ffmpeg to
give an error when trying to open an .avs input:

'Error opening input: Unknown error occurred'

No more details when using the ffmpeg '-verbose' option.

But compiling vmaf with this flag works and avoids the segfault.
Not sure if this is the best solution, as this would be needed in
every GNU/Linux distribution out there that uses avisynth + vmaf
in ffmpeg, but surely it is another good solution. Thanks for the
suggestion.

> 
> - Kacper
> 
>> flag for 'RTLD_DEEPBIND'[4] when calling 'dlopen()' on the
>> avisynthplus library. This will make the avisynthplus symbols
>> loaded by 'dlopen()' to be ahead of the symbols loaded in the
>> global scope, and consequently avoid the segmentation fault by
>> preventing the conflicting vmaf C++ destructor to be called.
>>
>> More details can be seen in the discussion of the Arch Linux bug
>> report[5] about this issue.
>>
>> [1] https://github.com/AviSynth/AviSynthPlus/blob/v3.7.5/avs_core/core/cache.h#L62
>> [2] https://github.com/Netflix/vmaf/blob/v3.0.0/libvmaf/src/svm.cpp#L75
>> [3] https://man.archlinux.org/man/dlopen.3#RTLD_LOCAL
>> [4] https://man.archlinux.org/man/dlopen.3#RTLD_DEEPBIND
>> [5] https://gitlab.archlinux.org/archlinux/packaging/packages/ffmpeg/-/issues/19
>>
>> Signed-off-by: Daniel Bermond <danielbermond at gmail.com>
>> ---
>>  libavformat/avisynth.c | 2 +-
>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/libavformat/avisynth.c b/libavformat/avisynth.c
>> index cb2be10925..bef61b4a60 100644
>> --- a/libavformat/avisynth.c
>> +++ b/libavformat/avisynth.c
>> @@ -120,7 +120,7 @@ typedef struct AviSynthContext {
>>
>>  static av_cold int avisynth_load_library(AviSynthContext *avs)
>>  {
>> -    avs->avs_library.library = dlopen(AVISYNTH_LIB, RTLD_NOW | RTLD_LOCAL);
>> +    avs->avs_library.library = dlopen(AVISYNTH_LIB, RTLD_NOW | RTLD_DEEPBIND);
>>      if (!avs->avs_library.library)
>>          return AVERROR_UNKNOWN;
>>
>> --
>> 2.50.1
>> _______________________________________________
>> ffmpeg-devel mailing list
>> ffmpeg-devel at ffmpeg.org
>> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>>
>> To unsubscribe, visit link above, or email
>> ffmpeg-devel-request at ffmpeg.org with subject "unsubscribe".
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel at ffmpeg.org
> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
> 
> To unsubscribe, visit link above, or email
> ffmpeg-devel-request at ffmpeg.org with subject "unsubscribe".

-- 
Best regards,
Daniel Bermond