[FFmpeg-devel] [PATCH] avcodec/cbs_h2645: Check NAL space

[James Almer]

On 4/22/2024 6:01 PM, Michael Niedermayer wrote:
> On Mon, Apr 22, 2024 at 05:46:10PM -0300, James Almer wrote:
>> On 4/22/2024 5:40 PM, Mark Thompson wrote:
>>> On 22/04/2024 02:31, Michael Niedermayer wrote:
>>>> Found-by-reviewing: CID1419833 Untrusted loop bound
>>>>
>>>> Sponsored-by: Sovereign Tech Fund
>>>> Signed-off-by: Michael Niedermayer <michael at niedermayer.cc>
>>>> ---
>>>>    libavcodec/cbs_h2645.c | 4 ++++
>>>>    1 file changed, 4 insertions(+)
>>>>
>>>> diff --git a/libavcodec/cbs_h2645.c b/libavcodec/cbs_h2645.c
>>>> index fe2e383ff33..1a45d424bae 100644
>>>> --- a/libavcodec/cbs_h2645.c
>>>> +++ b/libavcodec/cbs_h2645.c
>>>> @@ -709,7 +709,11 @@ static int cbs_h2645_split_fragment(CodedBitstreamContext *ctx,
>>>>                start = bytestream2_tell(&gbc);
>>>>                for(i = 0; i < num_nalus; i++) {
>>>> +                if (bytestream2_get_bytes_left(&gbc) < 2)
>>>> +                    return AVERROR_INVALIDDATA;
>>>>                    size = bytestream2_get_be16(&gbc);
>>>> +                if (bytestream2_get_bytes_left(&gbc) < size)
>>>> +                    return AVERROR_INVALIDDATA;
>>>>                    bytestream2_skip(&gbc, size);
>>>>                }
>>>>                end = bytestream2_tell(&gbc);
>>>
>>> Seems fair.
>>>
>>> The problem looks more general with missing bounds checks in all the H.266 code around this, though?  Compare with H.26[45], which have checks on all the reads - seems like H.266 should be doing that.
>>>
>>> Thanks,
>>
>> Not against this approach, but since the bytestream2_get_* functions return
>> 0, never overread the buffer or move the internal pointer, wouldn't it be
>> enough to just ensure end > start?
>> Particularly in ff_h2645_packet_split(), we can return an error if length
>> (in this case being set to end - start) is < 4.
> 
> The patch adds the same kind of check as are used in the AV_CODEC_ID_HEVC
> case earlier in the file already
> 
> I think what is done should approximately stay in sync

fwiw, better not add an initial length check to ff_h2645_packet_split() 
like i suggested, as it could potentially break samples that would 
otherwise decode just fine. It setting nb_nals to 0 should be enough.

So this patch is fine. Further bytestream2_get_bytes_left() checks can 
be added too, namely one that checks the buffer is at least the smallest 
size a vvcC box can be.