[FFmpeg-devel] [PATCHv2] vf_colorspace: Forbid odd dimensions

[Vittorio Giovara]

On Thu, Nov 17, 2016 at 4:11 PM, Ronald S. Bultje <rsbultje at gmail.com> wrote:
> Hi,
>
> On Thu, Nov 17, 2016 at 4:09 PM, Vittorio Giovara
> <vittorio.giovara at gmail.com> wrote:
>>
>> This prevents writing past bounds.
>>
>> Signed-off-by: Vittorio Giovara <vittorio.giovara at gmail.com>
>> ---
>> Updated according review.
>> Vittorio
>>
>>  doc/filters.texi            | 1 +
>>  libavfilter/vf_colorspace.c | 9 ++++++++-
>>  2 files changed, 9 insertions(+), 1 deletion(-)
>>
>> diff --git a/doc/filters.texi b/doc/filters.texi
>> index 208be42..aee43a8 100644
>> --- a/doc/filters.texi
>> +++ b/doc/filters.texi
>> @@ -5278,6 +5278,7 @@ colormatrix=bt601:smpte240m
>>  @section colorspace
>>
>>  Convert colorspace, transfer characteristics or color primaries.
>> +Input video needs to have an even size.
>>
>>  The filter accepts the following options:
>>
>> diff --git a/libavfilter/vf_colorspace.c b/libavfilter/vf_colorspace.c
>> index c7a2286..0024505 100644
>> --- a/libavfilter/vf_colorspace.c
>> +++ b/libavfilter/vf_colorspace.c
>> @@ -168,7 +168,7 @@ typedef struct ColorSpaceContext {
>>      int did_warn_range;
>>  } ColorSpaceContext;
>>
>> -// FIXME deal with odd width/heights (or just forbid it)
>> +// FIXME deal with odd width/heights
>>  // FIXME faster linearize/delinearize implementation (integer pow)
>>  // FIXME bt2020cl support (linearization between yuv/rgb step instead of
>> between rgb/xyz)
>>  // FIXME test that the values in (de)lin_lut don't exceed their container
>> storage
>> @@ -1031,8 +1031,15 @@ static int query_formats(AVFilterContext *ctx)
>>
>>  static int config_props(AVFilterLink *outlink)
>>  {
>> +    AVFilterContext *ctx = outlink->dst;
>>      AVFilterLink *inlink = outlink->src->inputs[0];
>>
>> +    if (inlink->w % 2 || inlink->h % 2) {
>> +        av_log(ctx, AV_LOG_ERROR, "Invalid odd size (%dx%d)\n",
>> +               inlink->w, inlink->h);
>> +        return AVERROR_PATCHWELCOME;
>> +    }
>
>
> If 422 && w%2 or 420 && (w%2||h%2), or some variant thereof. Odd sizes for
> 444 or odd vertical sizes for 422 should be OK, right?

well in theory, but i actually got a segfault on 422

==4064== Invalid write of size 2
==4064==    at 0x4ED6637: yuv2rgb_422p10_c (colorspacedsp_template.c:89)
==4064==    by 0x4F0444C: convert (vf_colorspace.c:504)

and on 444 (with testsrc).
-- 
Vittorio