[FFmpeg-devel] FFmpeg 3.4.1

Fri Dec 8 14:25:16 EET 2017

On Fri, 8 Dec 2017, wm4 wrote:

> On Fri, 8 Dec 2017 06:52:20 +0100
> Michael Niedermayer <michael at niedermayer.cc> wrote:
>
>> On Fri, Dec 08, 2017 at 01:09:50AM -0300, James Almer wrote:
>> > On 12/8/2017 12:26 AM, wm4 wrote: 
>> > > On Thu, 7 Dec 2017 23:23:51 +0100
>> > > Michael Niedermayer <michael at niedermayer.cc> wrote:
>> > > 
>> > >> On Tue, Nov 21, 2017 at 07:58:18PM +0100, Michael Niedermayer wrote: 
>> > >>> On Sat, Nov 18, 2017 at 09:11:17PM +0100, Michael Niedermayer wrote: 
>> > >>>> On Sat, Nov 18, 2017 at 09:50:33AM +0100, Hendrik Leppkes wrote: 
>> > >>>>> On Sat, Nov 18, 2017 at 3:05 AM, Michael Niedermayer
>> > >>>>> <michael at niedermayer.cc> wrote: 
>> > >>>>>> On Fri, Nov 17, 2017 at 09:55:47AM +0100, Hendrik Leppkes wrote: 
>> > >>>>>>> On Fri, Nov 17, 2017 at 5:00 AM, Michael Niedermayer
>> > >>>>>>> <michael at niedermayer.cc> wrote: 
>> > >>>>>>>> On Thu, Nov 16, 2017 at 01:51:34PM +0100, Carl Eugen Hoyos wrote: 
>> > >>>>>>>>> 2017-11-16 13:44 GMT+01:00 Michael Niedermayer <michael at niedermayer.cc>: 
>> > >>>>>>>>>> On Thu, Nov 16, 2017 at 01:04:27AM +0100, Carl Eugen Hoyos wrote: 
>> > >>>>>>>>>>> 2017-11-15 13:34 GMT+01:00 Michael Niedermayer <michael at niedermayer.cc>: 
>> > >>>>>>>>>>>> Hi all
>> > >>>>>>>>>>>>
>> > >>>>>>>>>>>> I intend to make 3.4.1 very soon 
>> > >>>>>>>>>>>
>> > >>>>>>>>>>> Shouldn't we first decide on how to proceed with #6775? 
>> > >>>>>>>>>>
>> > >>>>>>>>>> This would be ideal.
>> > >>>>>>>>>>
>> > >>>>>>>>>> IIUC this is a regression from bddb2343b6e594e312dadb5d21b408702929ae04 
>> > >>>>>>>>>
>> > >>>>>>>>> This was confirmed by more than one developer, yes.
>> > >>>>>>>>> 
>> > >>>>>>>>>> I see a patch that is said to improve 6775, can that be applied and
>> > >>>>>>>>>> would that resolve this ? 
>> > >>>>>>>>>
>> > >>>>>>>>> Iiuc, it would not completely resolve the issue, see:
>> > >>>>>>>>> https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=881536
>> > >>>>>>>>> 
>> > >>>>>>>>>> If so why was it not applied yet ? 
>> > >>>>>>>>>
>> > >>>>>>>>> The patch did not get support here, see:
>> > >>>>>>>>> [FFmpeg-devel] [PATCH] lavc: reset codec on receiving packet after EOF
>> > >>>>>>>>> in compat_decode 
>> > >>>>>>>>
>> > >>>>>>>>
>> > >>>>>>>> Is someone working on fixing this better than with the available patch ?
>> > >>>>>>>> 
>> > >>>>>>>
>> > >>>>>>> I don't even agree this needs fixing. Those projects use the API wrong. 
>> > >>>>>>
>> > >>>>>> Had we documented the correct/wrong use precissely in the past when
>> > >>>>>> these wrong uses where written ?
>> > >>>>>>
>> > >>>>>> Because if it was documented then few should have made the mistake.
>> > >>>>>> But it seems this affects multiple projects, so i wonder if our API
>> > >>>>>> really excluded the use
>> > >>>>>> 
>> > >>>>>
>> > >>>>> Apparently not well enough, but I also don't even understand why you
>> > >>>>> would *want* to drain in the middle of decoding.
>> > >>>>>
>> > >>>>> The only mention of sending NULL/0 packets (in 3.0 docs, before the
>> > >>>>> new API was introduced) do include the "at the end", however.
>> > >>>>>
>> > >>>>> From CODEC_CAP_DELAY:
>> > >>>>>  * Decoders:
>> > >>>>>  * The decoder has a non-zero delay and needs to be fed with avpkt->data=NULL,
>> > >>>>>  * avpkt->size=0 at the end to get the delayed data until the decoder no longer
>> > >>>>>  * returns frames.
>> > >>>>>
>> > >>>>> From avcodec_decode_video2
>> > >>>>> * @note Codecs which have the AV_CODEC_CAP_DELAY capability set have a delay
>> > >>>>> * between input and output, these need to be fed with avpkt->data=NULL,
>> > >>>>> * avpkt->size=0 at the end to return the remaining frames.
>> > >>>>>
>> > >>>>> There is a few more mentions of the same concept, but as far as I can
>> > >>>>> see all include the key words "at the end".
>> > >>>>> 
>> > >>>> 
>> > >>>>> For the suggested patch, draining and flushing in the middle of a
>> > >>>>> bitstream is still going to result in problems, though, since it
>> > >>>>> removes all reference frames, for example.
>> > >>>>> The original behavior cannot really be stored, which was to just keep
>> > >>>>> feeding frames into the decoder after a drain without a flush.
>> > >>>>> However, some decoders actually crashed when you did this, so this was
>> > >>>>> a rather unsafe action to begin with (not an issue any longer, since
>> > >>>>> this pattern is now blocked). 
>> > >>>>
>> > >>>> Did the previous code really work if the frame after a flush was not a
>> > >>>> new keyframe or there was some use of previous references ? 
>> > >>>
>> > >>> ping
>> > >>>
>> > >>> so what is the status of this?
>> > >>>
>> > >>> Ticket 6775 is still open, neither a workaround was applied nor was
>> > >>> it closed as invalid. Only one workaround was proposed which was
>> > >>> claimed to be worse than the code before.
>> > >>> It seems the discussion died down.
>> > >>> If theres no activity on this in the next days then i intend to make
>> > >>> the release with whats in release/3.4 at the time. I dont think
>> > >>> blind waiting will do any good, id rather release early and often ...
>> > >>>
>> > >>> Also if someone wants to write some release notes about this issue,
>> > >>> that is IMO a good idea ... 
>> > >>
>> > >> So this code is completely unmaintained ?
>> > >> Noone cares about pushing a workaround ?
>> > >> Noone cares about closing the ticket as wont fix ?
>> > >> Noone cares about explaining why neither should be done ?
>> > >>
>> > >> I intend to make the release tomorrow or as soon as i have time, we
>> > >> have waited too long already. I can of course wait more if people want
>> > >> but then please have a plan on resolving the issue that the release is
>> > >> delayed for 
>> > > 
>> > > I didn't read all of this, but it's probably due to the API user
>> > > passing as null packet, which indeed signals EOF. What happens if you
>> > > send more packets after sending EOF has always been undefined behavior,
>> > > and some audio decoders could crash if you did that. (It's fine if
>> > > you flush the decoder, of course.) I think I brought it up in the past
>> > > (and nobody cared), so I added explicit workarounds to my own API usage
>> > > code to avoid crashes. Why are you making so much noise around it now? 
>> > 
>> > When the old decode API was turned into a wrapper for the new, some
>> > applications using said API this way started to experience
>> > issues/crashes that did not happen before. So basically, a change of
>> > behavior which, as you put it, was apparently undefined.
>> > 
>> > Some argue it should not crash if it didn't before, others argue that it
>> > was never meant to be used this way. What Michael wants to know in order
>> > to release 3.4.1 is, what should be done? Should it be addressed, or
>> > left as is? If the former, do we apply the patch someone proposed in
>> > another thread, or something else? If the latter, then the ticket should
>> > be closed and the discussion ends there. 
>> 
>> yes
>> also to add to this, 6775 is set to the highest priority (critical)
>
> Also definitely wrong API usage.

IMHO the documentation was not entirely clear. In any case, 
undefined behaviour was changed, which broke two major projects. As I 
mentioned in the other thread, I see no downside of applying the patch, 
therefore it should be applied.

Regards,
Marton