Re: Security issues with local filesystem caching

[David Howells <dhowells@redhat.com>]

Stephen Smalley <sds@tycho.nsa.gov> wrote:

> > I was also wondering if I could generalise it to handle all cache types,
> > but the permissions checks are probably going to be quite different for
> > each type.  For instance, CacheFiles uses files on a mounted fs, whilst
> > CacheFS uses a block device.
> 
> So in the latter case, the daemon supplies the path of a block device
> node?

No.  In the latter case, there is no userspace daemon.  As there are no
dentries, filenames and paths in CacheFS, keeping track of the cull table
consumes a less space than for CacheFiles.

You start the cache by mounting it:

	mount -t cachefs /dev/hdx9 /cachefs

Then it's online.  However, you might want to check that whoever's calling
mount has permission to bring a cache online...

Actually, I think the permission to bring a cache online applies in all cases,
and is probably separate from checking that CacheFiles(d) is permitted to
mangle the filesystem it's using for a cache.  With CacheFS, we could do the
equivalent and do a MAC check to make sure we're permitted to read and write
the blockdev, as you suggest in the next bit:

> I suppose the hook could internally check the type of inode to decide what
> checks to apply, using the checks I previously sketched when it is a
> directory and using a different set of checks for the block device
> (substituting a write check against the block device for the
> directory-specific checks).  The hook interface itself would look the same
> IIUC, i.e. providing the (mnt, dentry) pair to which the path resolved and
> the secid to which the context resolved.

So, to summarise, is it worth having two checks:

 (1) Permission to bring a cache online or to take a cache offline.

 (2) Permission for the process bringing the cache online (cachefilesd or
     mount) to access the backing store, be it a set of files and directories,
     or be it a blockdev.

David