Eyevinn / mp4ff / 36253177160

83%

push

github

fix(hevc,vvc,mp4): reject invalid and mismatched NALU length sizes

ISO/IEC 14496-15 allows lengthSizeMinusOne 0, 1, or 3 (1-, 2-, or 4-byte NALU length fields). The hvcC and avcC decoders reject everything but 3, but the lhvC and vvcC decoders accepted 2 as well, and the hvcC, lhvC, and vvcC encoders wrote the field unchecked: 2 gave an invalid record, and 4 or more spilled into TemporalIDNested/NumTemporalLayers (hvcC, lhvC) or into reserved bits, reading back as another size (vvcC). The decoders now reject 2 and the encoders anything but 0, 1, or 3, with the new hevc.ErrInvalidLengthSize and vvc.ErrInvalidLengthSize.

14496-15 9.5.3 also requires the HEVC and L-HEVC configurations in a sample entry to have the same lengthSizeMinusOne, since the NALUs of all layers are in the same samples and share their length fields. A visual sample entry with an hvcC and an lhvC that disagree now fails to decode, and Encode and EncodeSW refuse to write one, so mp4ff no longer produces a file it then refuses to read.

Since the vvcC decoder accepts 1- and 2-byte lengths, mp4ff-nallister read such VVC tracks with 4-byte length fields. It now reports them as unsupported, as the avcC and hvcC decoders already make it do for AVC and HEVC.

hevc.SplitNalusByLayerID compared pos+int(naluLength) against the sample length, which wraps on 32-bit platforms for a length field of 2^31 or more and then panics on the slice. The check is now done in uint64.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

45 of 48 new or added lines in 5 files covered. (93.75%)

22245 of 27055 relevant lines covered (82.22%)

0.82 hits per line