Re: File System Performance

[Andrew Morton <akpm@zip.com.au>]

Linus Torvalds wrote:
> 
> On Mon, 12 Nov 2001, Lionel Bouton wrote:
> >
> > Seems not the case with gnu tar : write isn't even called once on the fd
> > returned by open("/dev/null",...). In fact a "grep write" on the strace
> > output is empty in the "tar cf /dev/null" case. Every file in the tar-ed
> > tree is stat-ed but no-one is read-ed.
> 
> Wow. What a sleazy optimization - it can't be anything but a special case.
> 
> How do they do it anyway? By matching on the name Or by knowing what the
> minor/major numbers of /dev/null are supposed to be on that particular
> operating system?

  /* Detect if outputting to "/dev/null".  */
  {
    static char const dev_null[] = "/dev/null";
    struct stat dev_null_stat;

    dev_null_output =
      (strcmp (archive_name_array[0], dev_null) == 0
       || (! _isrmt (archive)
           && stat (dev_null, &dev_null_stat) == 0
           && S_ISCHR (archive_stat.st_mode)
           && archive_stat.st_rdev == dev_null_stat.st_rdev));
  }

> And what's the _point_ of the optimization? I've never heard of a "tar
> benchmark"..

Ask Al.  He understands those GNU folks.

It's actually a bug.  I may _want_ an ISREG /dev/null...

-