Re: [PATCH 4/4] ovl: alllow remote upper

[Amir Goldstein <amir73il@gmail.com>]

On Tue, Feb 4, 2020 at 8:42 PM Vivek Goyal <vgoyal@redhat.com> wrote:
>
> On Tue, Feb 04, 2020 at 07:02:05PM +0200, Amir Goldstein wrote:
> > On Tue, Feb 4, 2020 at 6:17 PM Miklos Szeredi <miklos@szeredi.hu> wrote:
> > >
> > > On Tue, Feb 4, 2020 at 3:59 PM Vivek Goyal <vgoyal@redhat.com> wrote:
> > > >
> > > > On Fri, Jan 31, 2020 at 12:50:04PM +0100, Miklos Szeredi wrote:
> > > > > No reason to prevent upper layer being a remote filesystem.  Do the
> > > > > revalidation in that case, just as we already do for lower layers.
> > > > >
> > > > > This lets virtiofs be used as upper layer, which appears to be a real use
> > > > > case.
> > > >
> > > > Hi Miklos,
> > > >
> > > > I have couple of very basic questions.
> > > >
> > > > - So with this change, we will allow NFS to be upper layer also?
> > >
> > > I haven't tested, but I think it will fail on the d_type test.
> >
> > But we do not fail mount on no d_type support...
> > Besides, I though you were going to add the RENAME_WHITEOUT
> > test to avert untested network fs as upper.
> >
> > >
> > > > - What does revalidation on lower/upper mean? Does that mean that
> > > >   lower/upper can now change underneath overlayfs and overlayfs will
> > > >   cope with it.
> > >
> > > No, that's a more complicated thing.  Especially with redirected
> > > layers (i.e. revalidating a redirect actually means revalidating all
> > > the path components of that redirect).
> > >
> > > > If we still expect underlying layers not to change, then
> > > >   what's the point of calling ->revalidate().
> > >
> > > That's a good question; I guess because that's what the filesystem
> > > expects.  OTOH, it's probably unnecessary in most cases, since the
> > > path could come from an open file descriptor, in which case the vfs
> > > will not do any revalidation on that path.
> > >
> >
> > Note that ovl_dentry_revalidate() never returns 0 and therefore, vfs
> > will never actually redo the lookup, but instead return -ESTALE
> > to userspace. Right? This makes some sense considering that underlying
> > layers are not expected to change.
> >
> > The question is, with virtiofs, can we know that the server/host will not
> > invalidate entries?
>
> I don't think virtiofs will ensure that server/host will not invalidate
> entries. It will be more of a configuration thing where we will expect
> users to configure things in such a way that shared directory does not
> change.
>
> So that means, if user does not configure it properly and things change
> unexpectedly, then overlayfs should be able to detect it and flag error
> to user space?
>
> > And if it does, should it cause a permanent error
> > in overlayfs or a transient error? If we do not want a permanent error,
> > then ->revalidate() needs to be called to invalidate the overlay dentry. No?
>
> So as of now user space will get -ESTALE and that will get cleared when
> user space retries after corresponding ovl dentry has been dropped from
> cache (either dentry is evicted, cache is cleared forcibly or overlayfs
> is remounted)? If yes, that kind of makes sense. Overlay does not expect
> underlying layers to change and if a change it detected it is flagged
> to user space (and overlayfs does not try to fix it)?
>

I looks like it. I don't really understand why overlayfs shouldn't drop
the dentry on failure to revalidate. Maybe I am missing something.

Thanks,
Amir.