Hi,
On Fri, 2004-04-02 at 05:47, Chris Wedgwood wrote:
> On Thu, Apr 01, 2004 at 09:05:51PM -0600, Dmitry Nikiforov wrote:
>
> > Then there's no point in using XFS at all, is there? :)
>
> This has nothing to do with XFS. All filesystems journalling and
> otherwise are affected by this.
No, it's entirely a matter of the fs implementation. Ext3 by default
does not suffer from the problem, and it does _not_ have to journal its
data to avoid it. It simply arranges to flush data to disk first, then
metadata, when writing. Chris Mason also implemented such an "ordered"
mode for Reiserfs, and I believe SuSE use that as the default in their
kernels.
And there are designs which are fundamentally immune to the problem.
Both log-structured filesystems (such as jffs) and the tux2 experimental
phase-tree-based filesystem that Daniel Phillips did a while back are
guaranteed to hit disk in strict time order for both data and metadata.
It's a compromise, though: phase-tree and log-structure filesystems are
known to have serious fragmentation problems under some workloads. You
wouldn't necessarily want to use them for your Oracle database. :-) But
it's misleading to claim that all filesystems are necessarily affected
by this sort of thing.
Cheers,
Stephen
|