On Sat, Sep 11, 2010 at 09:57:02PM +0200, Iustin Pop wrote:
> On Sat, Sep 11, 2010 at 02:38:48PM -0500, Stan Hoeppner wrote:
> > Dave Chinner put forth on 9/11/2010 3:23 AM:
> > > On Sat, Sep 11, 2010 at 07:55:32AM +0100, John Lister wrote:
> > >> Stan Hoeppner wrote on 9/10/2010 14:00
> > >>>> On 10/09/2010 15:41, John Lister wrote:
> > >>> Try unmounting and remounting the filesystem, and see if the various
> > >>> tools all report the same thing afterwards. This solved the exact same
> > >>> problem for me very recently, though I'm on kernel 18.104.22.168 and xfsprogs
> > >>> 2.9.8.
> > >>
> > >> Cheers, that got rid of most of it, there is still a slight
> > >> discrepency (50 extra fragments) which I can live with.
> > >
> > > xfs_db used buffered IO on the block device, which is not coherent
> > > with the filesystem. If you are using it on an active filesystem,
> > > then running "echo 1 > /proc/sys/vm/drop_caches" before you run
> > > xfs_db should make it read from disk at least once....
> I wonder if xfs_db shouldn't use by default direct I/O, or at least take
> a flag to allow it to do direct I/O only against the blocks it needs.
> Dropping the entire caches on a big box is not nice :)
True - I need to look at whether posix_fadvise(POSIX_FADV_DONTNEED)
will clear the bdev pages, and if so that is the easiest solution.
Rewriting xfs_db to use direct IO is a pretty major undertaking...