Re: [PATCH 00/22][RFC] Unionfs: Stackable Namespace Unification Filesystem

[Shaya Potter <spotter@cs.columbia.edu>]

On Tue, 2006-09-05 at 08:02 +0200, Jan Engelhardt wrote:
> >
> >I agree that unionfs shouldn't oops, it should handle that situation in
> >a more graceful manner, but once the "backing store" is modified
> >underneath it, all bets are off for either unionfs or ext2/3 behaving
> >"correctly" (where "correctly" doesn't just mean handle the error
> >gracefully).
> >
> >But are you also 100% sure that messing with the underlying backing
> >store wouldn't be considered an admin bug as opposed to an administrator
> >bug? I mean there's nothing that we can do to prevent an administrator
> >from FUBAR'ing their system by 
> >
> >dd if=/dev/random of=/dev/kmem.
> >
> >where does one draw the line?  I agree that stackable file systems make
> >this a more pressing issue, as the "backing store" can be visible within
> >the file system namespace as a regular file system that people are
> >generally accustomed to interacting with.
> 
> So here's an idea. When a branch is added, mount an empty space onto the
> branch. (From within the kernel, so it appears as a side-effect of mount(2))
> 
>   mount -t unionfs -o dirs=/a=rw:/b=ro none /union
> 
> should imply something like
> 
>   mount --bind /var/lib/empty /a
>   mount --bind /var/lib/empty /b
> 
> Or better, yet, make them read-only:
> 
>   mount --rbind -o ro /a /a
>   mount --rbind -o ro /b /b
> (hope this works as intended?)
> 
> So that no one can tinker with /a and /b while the union is mounted.

I thought about that, but that doesn't help you w/ the NFS as branch
case.