Re: [PATCH 0/5] hybrid union filesystem prototype

[Valerie Aurora <vaurora@redhat.com>]

On Tue, Aug 31, 2010 at 11:05:18AM -0400, Kyle Moffett wrote:
> On Tue, Aug 31, 2010 at 07:24, Neil Brown <neilb@suse.de> wrote:
> > On Tue, 31 Aug 2010 13:00:45 +0200 Miklos Szeredi <miklos@szeredi.hu> wrote:
> >> No, I don't think this design will do that. ??So it might be enough
> >> just to document that online modification of upper or lower
> >> filesystems results in undefined behavior.
> >>
> >> But to prevent accidental damage, it's prudent (at least by default)
> >> to enforce the no-modification policy.
> >>
> >> Why do you think this feature of allowing modification is important?
> >> Lets take some typical use cases:
> >>
> >> ??- live cd: lower layer is hard r/o, upper layer makes no sense to
> >> ?? ??modify online
> >>
> >> ??- thin client: lower layer is static except upgrades, which need
> >> ?? ??special tools to support and is done offline, upper layer makes no
> >> ?? ??sense to modify online
> >>
> >> Do you have some cases in mind where it makes at least a little sense
> >> to allow online modification of the underlying filesystems?
> >
> > No, I don't have a particular use case in mind that would take advantage of
> > the layers being directly modifiable. ??But I know that sys-admins can be very
> > ingenious and may well come up with something clever.
> >
> > My point is more that I don't think that is it *possible* to prevent changes
> > to the underlying filesystem (NFS being the prime example) so if there are
> > easy steps we can take to make the behaviour of overlayfs more predictable in
> > those cases, we should.
> 
> There's certainly already weird behaviors you can cause by regular
> filesystem over-mounts on NFS.  For example, I have an NFS server that
> exports a "/srv/git" directory; if I was to do the following actions
> on a client:
> 
> # mkdir /srv/git
> # mount -t nfs myserver:/srv/git /srv/git
> # mkdir /srv/git/mnt
> # mount -t ext3 /dev/sda3 /srv/git/mnt
> 
> And then from the server I were to:
> # rmdir /srv/git/mnt
> 
> Terrible terrible things would happen... by which I mean I can no
> longer access or unmount that filesystem from the client.  That use
> case in particular seems to be much worse than your regular unionfs
> example even, and it's easily possible today (even by accident).

While this definitely sucks, the concern in this case with unioning
file systems is a deadlock or kernel panic, not just "weird" behavior
or inability to unmount a file system.  Although in general I like the
standard for union behavior as "not as bad as NFS." :)

-VAL