[FFmpeg-devel] [PATCH] avformat/dhav: fix backward scanning for get_duration and optimize seeking

[Andreas Rheinhardt]

Derek Buitenhuis:
> From: Justin Ruggles <justinr at vimeo.com>
> 
> The backwards scanning done for incomplete final packets should not
> assume a specific alignment at the end of the file. Truncated files
> result in hundreds of thousands of seeks if the final packet does not
> fall on a specific byte boundary, which can be extremely slow.
> For example, with HTTP, each backwards seek results in a separate
> HTTP request.
> 
> This changes the scanning to check for the end tag 1 byte at a time
> and buffers the last 1 MiB using ffio_ensure_seekback to avoid additional
> seek operations.
> 
> Signed-off-by: Justin Ruggles <justinr at vimeo.com>
> Signed-off-by: Derek Buitenhuis <derek.buitenhuis at gmail.com>
> ---
>  libavformat/dhav.c | 38 +++++++++++++++++++++++++++++---------
>  1 file changed, 29 insertions(+), 9 deletions(-)
> 
> diff --git a/libavformat/dhav.c b/libavformat/dhav.c
> index b2ead99609..15a75b5033 100644
> --- a/libavformat/dhav.c
> +++ b/libavformat/dhav.c
> @@ -232,34 +232,54 @@ static void get_timeinfo(unsigned date, struct tm *timeinfo)
>      timeinfo->tm_sec  = sec;
>  }
>  
> +#define MAX_DURATION_BUFFER_SIZE (1024*1024)
> +
>  static int64_t get_duration(AVFormatContext *s)
>  {
>      DHAVContext *dhav = s->priv_data;
>      int64_t start_pos = avio_tell(s->pb);
> +    int64_t end_pos = -1;
>      int64_t start = 0, end = 0;
>      struct tm timeinfo;
> -    int max_interations = 100000;
> +    uint8_t *end_buffer;
> +    int64_t end_buffer_size;
> +    int64_t end_buffer_pos;
>  
>      if (!s->pb->seekable)
>          return 0;
>  
> +    end_buffer_size = FFMIN(MAX_DURATION_BUFFER_SIZE, avio_size(s->pb));
> +    end_buffer = av_malloc(end_buffer_size);
> +    if (!end_buffer)
> +        return 0;
> +    end_buffer_pos = avio_size(s->pb) - end_buffer_size;
> +    avio_seek(s->pb, end_buffer_pos, SEEK_SET);
> +    ffio_ensure_seekback(s->pb, end_buffer_size);

This can fail.

> +    avio_read(s->pb, end_buffer, end_buffer_size);
> +    av_freep(&end_buffer);

You read into a buffer just to ignore the read data? Why don't you just
rewrite the loop below to inspect the data directly?

> +
>      avio_seek(s->pb, avio_size(s->pb) - 8, SEEK_SET);
> -    while (avio_tell(s->pb) > 12 && max_interations--) {
> +    while (avio_tell(s->pb) > end_buffer_pos) {
>          if (avio_rl32(s->pb) == MKTAG('d','h','a','v')) {
>              int64_t seek_back = avio_rl32(s->pb);
> -
> -            avio_seek(s->pb, -seek_back, SEEK_CUR);
> -            read_chunk(s);
> -            get_timeinfo(dhav->date, &timeinfo);
> -            end = av_timegm(&timeinfo) * 1000LL;
> +            end_pos = avio_tell(s->pb) - seek_back;
>              break;
>          } else {
> -            avio_seek(s->pb, -12, SEEK_CUR);
> +            avio_seek(s->pb, -5, SEEK_CUR);
>          }
>      }
>  
> -    avio_seek(s->pb, start_pos, SEEK_SET);
> +    if (end_pos < 0) {
> +        avio_seek(s->pb, start_pos, SEEK_SET);
> +        return 0;
> +    }
> +
> +    avio_seek(s->pb, end_pos, SEEK_SET);
> +    read_chunk(s);
> +    get_timeinfo(dhav->date, &timeinfo);
> +    end = av_timegm(&timeinfo) * 1000LL;
>  
> +    avio_seek(s->pb, start_pos, SEEK_SET);
>      read_chunk(s);
>      get_timeinfo(dhav->date, &timeinfo);
>      start = av_timegm(&timeinfo) * 1000LL;