Re: [RFC PATCH v2 0/3] ceph: add support for snapshot names encryption

[Jeff Layton <jlayton@kernel.org>]

On Thu, 2022-03-17 at 11:11 +0000, Luís Henriques wrote:
> Xiubo Li <xiubli@redhat.com> writes:
> 
> > On 3/17/22 6:01 PM, Jeff Layton wrote:
> > > I'm not sure we want to worry about .snap directories here since they
> > > aren't "real". IIRC, snaps are inherited from parents too, so you could
> > > do something like
> > > 
> > >      mkdir dir1
> > >      mkdir dir1/.snap/snap1
> > >      mkdir dir1/dir2
> > >      fscrypt encrypt dir1/dir2
> > > 
> > > There should be nothing to prevent encrypting dir2, but I'm pretty sure
> > > dir2/.snap will not be empty at that point.
> > 
> > If we don't take care of this. Then we don't know which snapshots should do
> > encrypt/dencrypt and which shouldn't when building the path in lookup and when
> > reading the snapdir ?
> 
> In my patchset (which I plan to send a new revision later today, I think I
> still need to rebase it) this is handled by using the *real* snapshot
> parent inode.  If we're decrypting/encrypting a name for a snapshot that
> starts with a '_' character, we first find the parent inode for that
> snapshot and only do the operation if that parent is encrypted.
> 
> In the other email I suggested that we could prevent enabling encryption
> in a directory when there are snapshots above in the hierarchy.  But now
> that I think more about it, it won't solve any problem because you could
> create those snapshots later and then you would still need to handle these
> (non-encrypted) "_name_xxxx" snapshots anyway.
> 

Yeah, that sounds about right.

What happens if you don't have the snapshot parent's inode in cache?
That can happen if you (e.g.) are running NFS over ceph, or if you get
crafty with name_to_handle_at() and open_by_handle_at().

Do we have to do a LOOKUPINO in that case or does the trace contain that
info? If it doesn't then that could really suck in a big hierarchy if
there are a lot of different snapshot parent inodes to hunt down.

I think this is a case where the client just doesn't have complete
control over the dentry name. It may be better to just not encrypt them
if it's too ugly.

Another idea might be to just use the same parent inode (maybe the
root?) for all snapshot names. It's not as secure, but it's probably
better than nothing.
-- 
Jeff Layton <jlayton@kernel.org>