[FFmpeg-devel] [PATCH] avformat/mov: (v4) fix get_eia608_packet

[Pavel Koshevoy]

On Thu, Feb 13, 2025, 22:04 Andreas Rheinhardt <
andreas.rheinhardt at outlook.com> wrote:

> Pavel Koshevoy:
> > The problem is reproducible with "Test for Quicktime 608 CC file.mov"
> > from https://samples.ffmpeg.org/MPEG2/subcc/
> >
> > ffmpeg -i "Test for Quicktime 608 CC file.mov" -map 0 -c copy -y
> remuxed.mov
> >
> > Prior to the fix QuickTime Player playback of remuxed.mov would
> > render garbage text for "English CC" subtitles.
>
> Is remuxing necessary for there being garbage?
>
> > ---
> >  libavformat/mov.c | 70 +++++++++++++++++++++++++++++++++++++++--------
> >  1 file changed, 59 insertions(+), 11 deletions(-)
> >
> > diff --git a/libavformat/mov.c b/libavformat/mov.c
> > index 85aef33b19..5a91ef5b8c 100644
> > --- a/libavformat/mov.c
> > +++ b/libavformat/mov.c
> > @@ -10788,25 +10788,73 @@ static int mov_change_extradata(AVStream *st,
> AVPacket *pkt)
> >      return 0;
> >  }
> >
> > -static int get_eia608_packet(AVIOContext *pb, AVPacket *pkt, int size)
> > +static int get_eia608_packet(AVIOContext *pb, AVPacket *pkt, int
> src_size)
> >  {
> > -    int new_size, ret;
> > +    /* We can't make assumptions about the structure of the payload,
> > +       because it may include multiple cdat and cdt2 samples. */
> > +    const uint32_t cdat = AV_RB32("cdat");
> > +    const uint32_t cdt2 = AV_RB32("cdt2");
>
> I don't think that using (non-variable) variables for these improves
> clarity (e.g. it means that the definition of the actual values used for
> the comparisons below is now further away from its use). Why not simply
> use MKBETAG('c','d','a','t') below?
>
> > +    int ret, out_size = 0;
> >
> > -    if (size <= 8)
> > +    /* a valid payload must have size, 4cc, and at least 1 byte pair: */
> > +    if (src_size < 10)
> >          return AVERROR_INVALIDDATA;
> > -    new_size = ((size - 8) / 2) * 3;
> > -    ret = av_new_packet(pkt, new_size);
> > +
> > +    /* avoid an int overflow: */
> > +    if ((src_size - 8) / 2 >= INT_MAX / 3)
> > +        return AVERROR_INVALIDDATA;
> > +
> > +    ret = av_new_packet(pkt, ((src_size - 8) / 2) * 3);
> >      if (ret < 0)
> >          return ret;
> >
> > -    avio_skip(pb, 8);
> > -    for (int j = 0; j < new_size; j += 3) {
> > -        pkt->data[j] = 0xFC;
> > -        pkt->data[j+1] = avio_r8(pb);
> > -        pkt->data[j+2] = avio_r8(pb);
> > +    /* parse and re-format the c608 payload in one pass. */
> > +    while (src_size >= 10) {
> > +        const uint32_t atom_size = avio_rb32(pb);
> > +        const uint32_t atom_type = avio_rb32(pb);
> > +        const uint32_t data_size = atom_size - 8;
>
> This may wrap around (if atom_size is < 8). If int is 32 bits, then the
> data_size > src_size check will catch this, but in case of 64 bit ints
> it may not. Relying on (unsigned, defined) integer wraparound should be
> avoided unless it is advantageous to use it; in this case, this is just
> not true: Just compare atom_size to 10 below.
>
> > +        const uint8_t cc_field =
> > +            atom_type == cdat ? 1 :
> > +            atom_type == cdt2 ? 2 :
> > +            0;
> > +
> > +        /* account for bytes consumed for atom size and type. */
> > +        src_size -= 8;
> > +
> > +        /* make sure the data size stays within the buffer boundaries.
> */
> > +        if (data_size < 2 || data_size > src_size) {
> > +            ret = AVERROR_INVALIDDATA;
> > +            break;
> > +        }
> > +
> > +        /* make sure the data size is consistent with N byte pairs. */
> > +        if (data_size % 2 != 0) {
>
> We typically try to avoid redundant "!= 0".
>
> > +            ret = AVERROR_INVALIDDATA;
> > +            break;
> > +        }
> > +
> > +        if (!cc_field) {
> > +            /* neither cdat or cdt2 ... skip it */
> > +            avio_skip(pb, data_size);
> > +            src_size -= data_size;
> > +            continue;
> > +        }
> > +
> > +        for (int32_t i = 0; i < data_size; i += 2) {
>
> int32_t? Why signed? (And why use a separate loop counter at all? Simply
> decrement data_size by 2 in each iteration.
>
> > +            pkt->data[out_size] = (0x1F << 3) | (1 << 2) | (cc_field -
> 1);
> > +            pkt->data[out_size + 1] = avio_r8(pb);
> > +            pkt->data[out_size + 2] = avio_r8(pb);
> > +            out_size += 3;
> > +            src_size -= 2;
> > +        }
> >      }
> >
> > -    return 0;
> > +    if (src_size > 0)
> > +        /* skip any remaining unread portion of the input payload */
> > +        avio_skip(pb, src_size);
> > +
> > +    av_shrink_packet(pkt, out_size);
> > +    return ret;
> >  }
> >
> >  static int mov_finalize_packet(AVFormatContext *s, AVStream *st,
> AVIndexEntry *sample,
>
> Generally, I believe that reading the input into pkt->data[size / 2]
> would be advantageous: It would make it simple to check for EOF and I/O
> errors (notice that the avio_r* reads above are unchecked) and would
> read the data in one go, avoiding all the avio_skip().
>
> - Andreas
>

Then perhaps you would find v2 of the patch more agreeable to your taste,
could you review that instead?

This function has been corrupting closed captions since 2020.  There was a
different fix posted in 2023 (mentioned by Devin in the 1st version of this
patch), perhaps that should be merged instead, as it also solves the
problem.

Pavel.