[FFmpeg-devel] [RFC][PATCH] return packets for known codecs earlier when codec probing is done

[Michael Niedermayer]

On Thu, May 28, 2009 at 12:25:22PM -0700, Baptiste Coudurier wrote:
> Hi Michael,
> 
> On 5/28/2009 4:01 AM, Michael Niedermayer wrote:
> > On Wed, May 27, 2009 at 06:12:37PM -0700, Baptiste Coudurier wrote:
> >> On 5/27/2009 5:48 PM, Michael Niedermayer wrote:
> >>> On Wed, May 27, 2009 at 05:47:22PM -0700, Baptiste Coudurier wrote:
> >>>> On 5/27/2009 5:10 PM, Michael Niedermayer wrote:
> >>>>> On Wed, May 27, 2009 at 04:40:49PM -0700, Baptiste Coudurier wrote:
> >>>>>> On 5/27/2009 3:00 PM, Michael Niedermayer wrote:
> >>>>>>> On Wed, May 27, 2009 at 12:54:27AM -0700, Baptiste Coudurier wrote:
> >>>>>>>> Hi Michael,
> >>>>>>>>
> >>>>>>>> Michael Niedermayer wrote:
> >>>>>>>>> [...]
> >>>>>>>>>
> >>>>>>>>>> @@ -537,10 +537,18 @@ int av_read_packet(AVFormatContext *s, AVPacket *p
> >>>>>>>>>>  
> >>>>>>>>>>          av_init_packet(pkt);
> >>>>>>>>>>          ret= s->iformat->read_packet(s, pkt);
> >>>>>>>>>> -        if (ret < 0)
> >>>>>>>>>> -            return ret;
> >>>>>>>>>> +        if (ret < 0) {
> >>>>>>>>>> +            if (!pktl)
> >>>>>>>>>> +                return ret;
> >>>>>>>>>> +            for (i = 0; i < s->nb_streams; i++)
> >>>>>>>>>> +                if (s->streams[i]->codec->codec_id == CODEC_ID_PROBE)
> >>>>>>>>>> +                    s->streams[i]->codec->codec_id = CODEC_ID_NONE;
> >>>>>>>>>> +            continue;
> >>>>>>>>>> +        }
> >>>>>>>>> i dont like this
> >>>>>>>>>
> >>>>>>>> Well, we have a problem if probe_packets is not 0 when eof is reached.
> >>>>>>>> Do you see another way to avoid this ?
> >>>>>>> is this special case really relevant?
> >>>>>>> it just looses data on streams when all of the following is true
> >>>>>>> * stream is set to CODEC_ID_PROBE
> >>>>>>> * stream contains less than 100 packets in the whole file
> >>>>>>> * probing fails even with all packets in the whole file
> >>>>>> I think it is:
> >>>>>> code will call av_read_packet until st->probe_packets is 0, which will
> >>>>>> provoke an infinite loop.
> >>>>> hmm, i must be missing something, as i thought the packets would just never
> >>>>> be output and the player would end at EOF in whichever way it did before the
> >>>>> patch
> >>>> Before the patch, ie currently, if EOF is reach, packets are not outputed.
> >>>> This is a bug IMHO.
> >>> i agree its a bug, the patch without the controversal part would reduce it
> >>> to the case where less than 100 packets are in a whole file && probing
> >>> fails.
> >>> Your suggested fix, fixes it but introduces a new bug, namely that seeking
> >>> to EOF causes probing to fail for all streams permanently, even if the
> >>> user seeked back afterwards. 
> >> Well, probing should be executed by av_find_stream_info basically, and
> >> seeking should not occur during this process.
> > 
> > thats the ideal world case, yes but with mpeg-ts streams can start later
> > if iam not mistaken (like 2 movies which each happen to use different 
> > codecs or have subtitles in more or less languages)
> 
> Ok, thanks for clarifying.
> 
> >> But I don't really get what you mean by seeking to EOF and probing.
> >> Why would an user seek to EOF ?
> > 
> > because he clicked the wrong button ;)
> > users are silly, but ffmpeg should still not misbehave, especially not in
> > such a obscure way, i mean a mistaken seek to EOF shortly after a transition
> > between 2movies could lead to missing audio, maybe its a 1 in 100000 chance
> > or even less likely but its a bug from the users point of view
> 
> Ok, let's try to handle this the correct way.
> If the user seeks to EOF and a new stream (pid) was detected, codec will
> either be set to the right codec according to the pmt or if unknown
> codec probing will be called. If EOF is reached we must output packets

> buffered but it's not reasonable to set codecs to CODEC_ID_PROBE.

i assume you meant CODEC_ID_NONE here


> 
> So during seek, probe packets should be reset to MAX_PROBE_PACKETS, and
> eof must set probe_packets to 0.
> 
> Patch attached.

looks good

[...]
-- 
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

There will always be a question for which you do not know the correct awnser.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: Digital signature
URL: <http://lists.mplayerhq.hu/pipermail/ffmpeg-devel/attachments/20090531/8485d252/attachment.pgp>