On Thu, May 06, 2004 at 09:40:42PM +0100, Christoph Hellwig wrote:
> On Thu, May 06, 2004 at 09:26:38AM -0700, Jeffrey W. Baker wrote:
> > XFS just decided out of the blue it was hosed. This has been reported
> > on this list a number of times, normally with NFS involved. But this
> > filesystem was not exported with NFS, it was simply running bonnie,
> > memtest.sh, and dd md0 all at once.
>
> Don't touch the underlying device if XFS is active. This is know to
> cause data corruption. I have an experimental fix for 2.6 here, I just
> need to get rid of a occasional deadlock in there that I'm hitting.
>
> While it's conceptually backportable to 2.4 I doubt marcelo will take
> it.
>
> Nathan, Eric: any good idea how to prominently warn about this?
Not really, its a tough one. It really needs to be in the kernel
to be effective ... would it have to be a complete 2.6 backport,
or could there be some simplified infrastructure we could use?
Maybe just something setting a flag when we claim the device(s)
(while holding bd_sem), and preventing blkdev opens on those in
the block_dev.c::do_open (and vice-versa, of course, preventing
a mount on an already open device). There does seem to be some
infrastructure we might use (bd_sem/bd_openers)?
Probably we should claim the device excl in 2.6 too until its
sorted out too (but if the fix is close at hand, better to focus
on that I geuss).
thanks.
--
Nathan
|