[lxc-devel] [PATCH] utils: remove unnecessary check of mystat.st_dev

[Serge Hallyn]

Quoting Luka Perkov (luka.perkov at sartura.hr):
> Hi Serge,
> 
> On Wed, Oct 29, 2014 at 08:30:16PM +0000, Serge Hallyn wrote:
> > Quoting Luka Perkov (luka.perkov at sartura.hr):
> > > The check is not needed and it breaks lxc-destroy when container is installed
> > > on top of overlayfs. More information why this is a problem on overlayfs can be
> > > found here:
> > > 
> > > https://kernel.googlesource.com/pub/scm/linux/kernel/git/mszeredi/vfs/+/overlayfs.current/Documentation/filesystems/overlayfs.txt
> > 
> > Hm, not sure how I feel about this.  You say "it's not needed",
> > but removing it invalidates the "_onedev" part of the function
> > name.
> > 
> > I realize that in most cases the is-one-device part isn't
> > needed.  But in those rare cases where admin has messed up and
> > left a pointer to host's / or /etc in the container
> > somehow, it'll be missed.
> 
> I'm not sure I understand the above. Lets say you have container called
> foo, then:
> 
> /var/lib/lxc/foo/rootfs/mylink -> /etc
> 
> The link called mylink points to /etc. With the patch proposed the link
> will be removed but /etc on the host will not. The directory
> /var/lib/lxc/foo/rootfs/etc will be removed (if it exists of course). I
> thought that this was desired behavior.

What if it is a bind mount?

> > Really I don't know that I'd recommend running containers on top
> > of overlayfs.  This is not the usual overlayfs-container use-case,
> > and it overlayfs has enough niggles that it'd worry me.  Still,
> > if we can fix this in a safe way then I'm open to it.
> > 
> > Maybe we should detect whether the underlying fs is overlayfs,
> > and ignore the device if it is?
> 
> I'll see if there is an elegant way to do this ;)
> 
> Luka