Re: [PATCH] use posix_fadvise to increase readahead

[Eric Wong <normalperson@yhbt.net>]

Eric Wong <normalperson@yhbt.net> wrote:
> All relevant audio file formats store data sequentially, so
> give a hint to the kernel to perform more readahead.  In current
> Linux, the readahead hint doubles readahead pages and can help
> with playback on slow devices.

Btw, I've been running this for a few years, too; but pretty
much all on FLAC.

I don't know if there's audio formats we support which aren't
sequential.  Maybe there's some wacky audio format which
requires random read all over the place...

> ---
>   If you prefer "git pull" instead of "git am":
> 
>   The following changes since commit 7e74b254b2a7c963be0bfce751fc5911fe681c12:
> 
>     Remove hepler script. It's mostly unmaintained, I don't know if anyone but me ever used it. In any case, those who want a custom Debian package should be capable of updating the debian/changelog entry on their own. (2015-02-26 22:48:40 -0500)
> 
>   are available in the git repository at:
> 
>     git://bogomips.org/sox fadvise

Nowadays moved to:

     https://yhbt.net/sox.git ew/fadvise

The commit SHA-1 remains unchanged

>   for you to fetch changes up to 1c47376a04218838447e7bd49fb42156e9cb13b6:
> 
>     use posix_fadvise to increase readahead (2015-09-07 00:39:24 +0000)
> 
>   ----------------------------------------------------------------
>   Eric Wong (1):
>         use posix_fadvise to increase readahead
> 
>    configure.ac  |  1 +
>    src/formats.c | 18 ++++++++++++++++--
>    src/sox_i.h   |  4 ++++
>    3 files changed, 21 insertions(+), 2 deletions(-)
> 
> diff --git a/configure.ac b/configure.ac
> index 23138a9..d9ae1ba 100644
> --- a/configure.ac
> +++ b/configure.ac
> @@ -208,6 +208,7 @@ AC_CHECK_HEADERS(fcntl.h unistd.h byteswap.h sys/stat.h sys/time.h sys/timeb.h s
>  
>  dnl Checks for library functions.
>  AC_CHECK_FUNCS(strcasecmp strdup popen vsnprintf gettimeofday mkstemp fmemopen)
> +AC_CHECK_FUNCS(posix_fadvise)
>  
>  dnl Check if math library is needed.
>  AC_SEARCH_LIBS([pow], [m])
> diff --git a/src/formats.c b/src/formats.c
> index 724a4cd..340b867 100644
> --- a/src/formats.c
> +++ b/src/formats.c
> @@ -329,12 +329,26 @@ static void set_endiannesses(sox_format_t * ft)
>  static sox_bool is_seekable(sox_format_t const * ft)
>  {
>    struct stat st;
> +  int fd, seekable;
>  
>    assert(ft);
>    if (!ft->fp)
>      return sox_false;
> -  fstat(fileno((FILE*)ft->fp), &st);
> -  return ((st.st_mode & S_IFMT) == S_IFREG);
> +  fd = fileno((FILE*)ft->fp);
> +  if (fd < 0)
> +     return 0;
> +  fstat(fd, &st);
> +  seekable = ((st.st_mode & S_IFMT) == S_IFREG);
> +#if defined HAVE_POSIX_FADVISE
> +  if (seekable) {
> +    /*
> +     * POSIX_FADV_NOREUSE can potentially be beneficial, too,
> +     * but is a no-op as of Linux 4.2.  Not sure about other kernels.

Still true as of Linux 5.8 :<

> +     */
> +    (void)posix_fadvise(fd, (off_t)0, st.st_size, POSIX_FADV_SEQUENTIAL);

Also, Linux has always ignored offset and size for
POSIX_FADV_SEQUENTIAL, so this won't do anything strange when
using the trim-as-first-effect to force seeking.

> +  }
> +#endif
> +  return seekable;
>  }
>  
>  /* check that all settings have been given */
> diff --git a/src/sox_i.h b/src/sox_i.h
> index 04bbe8c..2eec9d3 100644
> --- a/src/sox_i.h
> +++ b/src/sox_i.h
> @@ -20,6 +20,10 @@
>  #define _GNU_SOURCE
>  #endif
>  
> +#if defined HAVE_POSIX_FADVISE
> +#define _XOPEN_SOURCE 600
> +#endif
> +

I think _DEFAULT_SOURCE is needed, too, nowadays.

>  #include <errno.h>
>  #include <stdio.h>
>  #include <stdlib.h>


_______________________________________________
SoX-devel mailing list
SoX-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/sox-devel