[Lustre-discuss] MDT Snapshot orphaned inodes

Andreas Dilger adilger at sun.com
Tue Jan 29 01:22:04 PST 2008


On Jan 28, 2008  14:44 -0500, Brian J. Murrell wrote:
> On Mon, 2008-01-28 at 14:34 -0500, Aaron Knister wrote:
> > I'm trying to snapshot my mdt using lvm and back it up.
> > 
> > After i take the snapshot and go to mount the volume (as ldiskfs) I  
> > end up with all kinds of warnings about orphaned inodes. Is this  
> > normal?

> It is.  You are trying to mount a (copy of a) filesystem that is in an
> "open" state rather than the state it would be in if you were mounting a
> filesystem that had had gone through an "unmount".

Partly true, partly not.  It is normal if you expect that applications
are keeping files open after unlinking them...

> > Is this a sign of potential problems with my MDT?
> 
> No.  But your backup will not truly be a snapshot in time at the exact
> moment you took it since any transactions which have not been flushed
> from the journal at the snapshot time will be backed out when you mount
> it.
> 
> Ideally, you could have ldiskfs close off the journal and quiesce the
> filesystem before you take your snapshot and you would not see those
> errors and would have a real snapshot it time copy.  I don't know that
> there is a way to do that though, and really the result is probably
> close enough if you are doing something like a daily backup or such.

Well, the LVM kernel code _does_ flush everything to disk (including the
journal) and block all device IO before taking the snapshot.  The orphan
inodes are the one thing it can't clean up before the snapshot, since
that would lose the open-unlinked files otherwise.

This isn't harmful or inconsistent for the snapshot, because those files
have actually been deleted already.

Cheers, Andreas
--
Andreas Dilger
Sr. Staff Engineer, Lustre Group
Sun Microsystems of Canada, Inc.




More information about the lustre-discuss mailing list