[FFmpeg-devel] [PATCH] avcodec/libmp3lame: properly handle unaligned frame data

[wm4]

On Wed, 26 Apr 2017 17:16:05 +0700
Muhammad Faiz <mfcc64 at gmail.com> wrote:

> This should fix Ticket6349
> 
> Since 383057f8e744efeaaa3648a59bc577b25b055835, framequeue may
> generate unaligned frame data.
> 
> Signed-off-by: Muhammad Faiz <mfcc64 at gmail.com>
> ---
>  libavcodec/libmp3lame.c | 13 +++++++++----
>  1 file changed, 9 insertions(+), 4 deletions(-)
> 
> diff --git a/libavcodec/libmp3lame.c b/libavcodec/libmp3lame.c
> index 5e26743..cd505bb 100644
> --- a/libavcodec/libmp3lame.c
> +++ b/libavcodec/libmp3lame.c
> @@ -203,15 +203,20 @@ static int mp3lame_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
>              ENCODE_BUFFER(lame_encode_buffer_int, int32_t, frame->data);
>              break;
>          case AV_SAMPLE_FMT_FLTP:
> -            if (frame->linesize[0] < 4 * FFALIGN(frame->nb_samples, 8)) {
> -                av_log(avctx, AV_LOG_ERROR, "inadequate AVFrame plane padding\n");
> -                return AVERROR(EINVAL);
> -            }
>              for (ch = 0; ch < avctx->channels; ch++) {
> +                if (frame->linesize[0] < 4 * FFALIGN(frame->nb_samples, 8) || 0x1F & (intptr_t)(frame->data[ch])) {
> +                    float *src = (float *) frame->data[ch];
> +                    float *dst = s->samples_flt[ch];
> +                    int k;
> +
> +                    for (k = 0; k < frame->nb_samples; k++)
> +                        dst[k] = src[k] * 32768.0f;
> +                } else {
>                  s->fdsp->vector_fmul_scalar(s->samples_flt[ch],
>                                             (const float *)frame->data[ch],
>                                             32768.0f,
>                                             FFALIGN(frame->nb_samples, 8));
> +                }
>              }
>              ENCODE_BUFFER(lame_encode_buffer_float, float, s->samples_flt);
>              break;

Looks like the right solution is fixing framequeue.

I don't think encoders generally allow lax alignment, and this might
not be the only case.