[FFmpeg-devel] [PATCH 1/8] avcodec/cinepak: Require 1 bit per 4x4 block as minimum input

[Michael Niedermayer]

On Sun, Aug 18, 2019 at 01:40:01PM +0200, Tomas Härdin wrote:
> sön 2019-08-18 klockan 12:19 +0200 skrev Michael Niedermayer:
> > On Sun, Aug 18, 2019 at 12:00:45PM +0200, Paul B Mahol wrote:
> > > On Sun, Aug 18, 2019 at 11:44 AM Michael Niedermayer <michael at niedermayer.cc>
> > > wrote:
> > > 
> > > > On Sun, Aug 18, 2019 at 10:47:26AM +0200, Tomas Härdin wrote:
> > > > > sön 2019-08-18 klockan 02:35 +0200 skrev Tomas Härdin:
> > > > > 
> > > > > I did some investigation, it is indeed ff_reget_buffer(). It copies the
> > > > > frame data for some reason. The fix is simple in this case: just call
> > > > > ff_get_buffer() once in cinepak_decode_init() and keep overwriting the
> > > > > same frame.
> > > > > 
> > > > > > As I said on IRC, this class of problems will exist for every codec.
> > > > > > Cinepak is easy to decode, even at these resolutions. Just imagine what
> > > > > > will happens when someone feeds in a 65535x209 av1 stream..
> > > > > 
> > > > > And related to this, ff_reget_buffer() is used for a lot of these
> > > > > codecs which only overwrite pixels in the old frame. flicvideo, gifdec,
> > > > > msrle, roqvideodec and others probably have the same flaw.
> > > > 
> > > > not calling any form of *get_buffer per frame breaks decoding into
> > > > user supplied buffers.
> > > > 
> > > > If you check the documentation of the get_buffer2 callback
> > > > 
> > > > " This callback is called at the beginning of each frame to get data
> > > > buffer(s) for it."
> > > > 
> > > > That would not be possible if its just called once in init
> 
> Sorry, I'm a bit rusty on lavc internals.
> 
> > > > and yes i too wish there was a magic fix but i think most things that
> > > > look like magic fixes have a fatal flaw. But maybe iam missing something
> > > > in fact i hope that iam missing something and that there is a magic fix
> > > > 
> > > 
> > > Magic fix is enabling reference counted frames in fuzzer.
> > 
> > That is covered by the part below which you maybe did not read
> > 
> > > > PS: if you think of changing the API, i dont think its the API.
> > > > I mean every user application will read the frames it receives, so
> > > > even if inside the decoder we just return the same frame with 2 pixels
> > > > different the user doesnt know this and has to read the whole frame.
> > > > The problem is moved around but its still there.
> 
> Copying is still slower than not copying. Enabling refcounting fixes
> the timeout issue here, and will likely silence a whole bunch of false
> positives for this class of files.

it makes probably sense to enable ref counting but we should
immedeatly in the next or a previous commit make the fuzzer read the frames
from the  decoder. Thats what basically every user app would do.
Otherwise we would have a bunch of issues closed and then reopened
later.

an alternative viewpoint to this would be to set the refcounting flag
from the input so the fuzzer itself has control over it and we test
both codepathes. This would improve coverage


> 
> It would be beneficial to have a consistent way of signalling that a
> frame didn't change, since a bunch of codecs have that property.
> Currently it's a mix of discontinuous timestamps, longer frame
> durations and repeating identical frames.

yes, i strongly agree.


> 
> > so feeding really crazy resolutions into a decoder requires some
> > small but proportional amout of input bytes.
> > double the width and the minimum input bytes double sort of.
> 
> Lavc is already very lenient with what it accepts. How do you detect
> the difference between "this packet is too small to decode to an entire
> frame" from "this packet is too small but we could still get a few MBs
> out of it"?

In reality this is actually not hard because a frame that is smaller
than the minimum valid size is generally for many codecs so small 
it really wont contain anything usefull to decode.
And we have discard_damaged_percentage where the user can tune it too.
Patches using discard_damaged_percentage are sometimes objected too
though so it is not consistently used.

Thanks

[...]

-- 
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

The worst form of inequality is to try to make unequal things equal.
-- Aristotle
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 181 bytes
Desc: not available
URL: <http://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20190818/c05d04f5/attachment.sig>