Re: NFS/credentials leak in 2.6.29-rc1

[David Howells <dhowells@redhat.com>]

J. Bruce Fields <bfields@fieldses.org> wrote:

> 	- Finally, we put_cred(override_creds(new)).  That modifies
> 	  current->cred again, putting the old value and getting the
> 	  new.
> 
> Hm.  But that last part's not OK; aren't we still holding our own
> reference to new, in addition to the one that override_creds() just
> took?  So I think we need the following?

Yes, you're right.  override_creds() takes an extra ref on the argument it is
passed, thus leaving the caller with their original reference intact.

So really, you don't want to call override_creds() as that will cost you an
extra atomic_inc() and atomic_dec_and_test().  I recommend you replace:

        put_cred(override_creds(new));

with:

	revert_creds(new);

I think that should do the right thing.  It may look a bit odd, but it'll be
quicker.  If you object to using revert_creds)( because of the name, we can
come up with an alternative name.

> Looking through nfsd_setuser(), one obvious bug: in the (flags &
> NFSEXP_ALLSQUASH) case, we never check the return value from the
> groups_alloc(0).  If it returns NULL, we dereference it anyway.

Since a zero-length groups list must be copied before writing, can I recommend
that we make groups_alloc(0) a special case that returns pointer to a
statically allocated groups list (after inc'ing the refcount) that represents
a zero-length list, thus meaning groups_alloc(0) will never fail?

David