On 7/14/2025 9:21 PM, Michael Niedermayer wrote:
> On Sun, Jul 13, 2025 at 01:42:28PM +0200, Timo Rothenpieler wrote:
>> On 7/13/2025 3:10 AM, Michael Niedermayer wrote:
>>> Fixes: Infinite loop
>>> Fixes: 427538726/clusterfuzz-testcase-minimized-ffmpeg_dem_FLV_fuzzer-6582567304495104
>>>
>>> Found-by: continuous fuzzing process https://github.com/google/oss-fuzz/tree/master/projects/ffmpeg
>>> Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
>>> ---
>>>    libavformat/flvdec.c | 3 +++
>>>    1 file changed, 3 insertions(+)
>>>
>>> diff --git a/libavformat/flvdec.c b/libavformat/flvdec.c
>>> index ac681954cb7..a4fa0157512 100644
>>> --- a/libavformat/flvdec.c
>>> +++ b/libavformat/flvdec.c
>>> @@ -1715,6 +1715,9 @@ retry_duration:
>>>                    av_log(s, AV_LOG_DEBUG, "Set channel data from MultiChannel info.\n");
>>> +                if (avio_feof(s->pb))
>>> +                    return AVERROR_EOF;
>>> +
>>>                    goto next_track;
>>>                }
>>>            } else if (stream_type == FLV_STREAM_TYPE_VIDEO) {
>>
>> I don't think just returning from here is correct.
>> The goto next_track right after it already checks for EOF.
> 
>> I do not see how between here and the eof check there there'd be any way to
>> infinite loop.
> 
> avio_skip() with a negative value will reset the EOF flag
> 
> 
>>
>> It returns FFERROR_REDO there, which is important to drain queued up
>> packages.
> 
> I think the state becomes corrupted once it reads into EOF
> that is the size accounting goes "oops" as the code keeps running
> things that read and keeps accounting for these reads but in reality
> nothing is read as its at EOF
> and then it seeks back all these "not reads" and thats where it hits the
> infinite loop as theres a mismatch what the code thinks it moved forward
> and what it actually moved forward.
> Thats how it looked to me at least, i have not verified every step of this
> 
> ill mail you the testcase, then you can check if my analysis is right
> and fix the code in a way that can recover queued packets in such truncated
> packet at EOF case.
> Also please make sure its not forgotten that whatever fix this gets is backported
I'm unable to reproduce any infinite loops, even with the sample.
But the code there definitely is sub-optimal, given the seek can go the 
wrong way, and even when going the right way can potentially reset the 
EOF flag.

Proposed patch is attached.