[FFmpeg-devel] [PATCH] lavf/mxfdec: fix seeking before the first keyframe

[Marton Balint]

On Mon, 7 Dec 2015, Tomas Härdin wrote:

> On Sat, 2015-11-28 at 02:56 +0100, Marton Balint wrote:
>> Regression since 53f2ef2c4afb1d49a679dea9163cb0e4671f3117.
>> Fixes ticket #5017.
>>
>> Signed-off-by: Marton Balint <cus at passwd.hu>
>> ---
>>  libavformat/mxfdec.c | 10 ++++++++++
>>  1 file changed, 10 insertions(+)
>>
>> diff --git a/libavformat/mxfdec.c b/libavformat/mxfdec.c
>> index 429f46a..926d2a3 100644
>> --- a/libavformat/mxfdec.c
>> +++ b/libavformat/mxfdec.c
>> @@ -3181,6 +3181,16 @@ static int mxf_read_seek(AVFormatContext *s, int stream_index, int64_t sample_ti
>>          sample_time = FFMAX(sample_time, 0);
>>
>>          if (t->fake_index) {
>> +            /* The first frames may not be keyframes in presentation order, so
>> +             * we have to advance the target to be able to find the first
>> +             * keyframe backwards... */
>> +            if (!(flags & AVSEEK_FLAG_ANY) &&
>> +                (flags & AVSEEK_FLAG_BACKWARD) &&
>> +                t->ptses[0] != AV_NOPTS_VALUE &&
>
> Can the size of t->ptses ever be zero here?

No, "paranoid check" in mxf_compute_ptses_fake_index should ensure that.

>
>
>> +                sample_time < t->ptses[0] &&
>> +                (t->fake_index[t->ptses[0]].flags & AVINDEX_KEYFRAME))
>
> Do t->ptses always point inside t->fake_index?

Yes, if it is not AV_NOPTS_VALUE, that is made sure when setting the 
ptses in mxf_compute_fake_index.

>
>> +                sample_time = t->ptses[0];
>> +
>
> Should be OK otherwise, since it only affects seeking and fixes a
> regression.

Ok, applied, thanks.

> Ideally we should be doing something better for seeking, I
> think..

Well, yes...

Regards,
Marton