[FFmpeg-devel] [PATCH v3] avformat/dhav: fix backward scanning for get_duration and optimize seeking

[Kacper Michajlow]

On Wed, 21 May 2025 at 15:23, Derek Buitenhuis
<derek.buitenhuis at gmail.com> wrote:
>
> 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.
>
> Co-authored-by: Derek Buitenhuis <derek.buitenhuis at gmail.com>
> Signed-off-by: Justin Ruggles <justinr at vimeo.com>
> Signed-off-by: Derek Buitenhuis <derek.buitenhuis at gmail.com>
> ---
>  libavformat/dhav.c | 54 +++++++++++++++++++++++++++++++++-------------
>  1 file changed, 39 insertions(+), 15 deletions(-)
>
> diff --git a/libavformat/dhav.c b/libavformat/dhav.c
> index b2ead99609..d9db775802 100644
> --- a/libavformat/dhav.c
> +++ b/libavformat/dhav.c
> @@ -22,6 +22,7 @@
>
>  #include <time.h>
>
> +#include "libavutil/intreadwrite.h"
>  #include "libavutil/mem.h"
>  #include "libavutil/parseutils.h"
>  #include "avio_internal.h"
> @@ -232,37 +233,60 @@ 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;
> +    int64_t offset;
> +    unsigned date;
>
>      if (!s->pb->seekable)
>          return 0;
>
> -    avio_seek(s->pb, avio_size(s->pb) - 8, SEEK_SET);
> -    while (avio_tell(s->pb) > 12 && max_interations--) {
> -        if (avio_rl32(s->pb) == MKTAG('d','h','a','v')) {
> -            int64_t seek_back = avio_rl32(s->pb);
> +    if (start_pos + 16 > avio_size(s->pb))
> +        return 0;
>
> -            avio_seek(s->pb, -seek_back, SEEK_CUR);
> -            read_chunk(s);
> -            get_timeinfo(dhav->date, &timeinfo);
> -            end = av_timegm(&timeinfo) * 1000LL;
> +    avio_skip(s->pb, 16);
> +    date = avio_rl32(s->pb);
> +    get_timeinfo(date, &timeinfo);
> +    start = av_timegm(&timeinfo) * 1000LL;
> +
> +    end_buffer_size = FFMIN(MAX_DURATION_BUFFER_SIZE, avio_size(s->pb));
> +    end_buffer = av_malloc(end_buffer_size);

This allocation is never freed.

- Kacper