Re: [RFC PATCH 1/2] selinux: Don't call avc_compute_av() from RCU path walk

From: Paul E. McKenney
Date: Wed Nov 20 2019 - 14:07:46 EST


On Wed, Nov 20, 2019 at 10:28:31AM -0500, Stephen Smalley wrote:
> On 11/20/19 8:12 AM, Will Deacon wrote:
> > Hi Stephen,
> >
> > Thanks for the quick reply.
> >
> > On Tue, Nov 19, 2019 at 01:59:40PM -0500, Stephen Smalley wrote:
> > > On 11/19/19 1:40 PM, Will Deacon wrote:
> > > > 'avc_compute_av()' can block, so we carefully exit the RCU read-side
> > > > critical section before calling it in 'avc_has_perm_noaudit()'.
> > > > Unfortunately, if we're calling from the VFS layer on the RCU path walk
> > > > via 'selinux_inode_permission()' then we're still actually in an RCU
> > > > read-side critical section and must not block.
> > >
> > > avc_compute_av() should never block AFAIK. The blocking concern was with
> > > slow_avc_audit(), and even that appears dubious to me. That seems to be more
> > > about misuse of d_find_alias in dump_common_audit_data() than anything.
> >
> > Apologies, I lost track of GFP_ATOMIC when I reading the code and didn't
> > think it was propagated down to all of the potential allocations and
> > string functions. Having looked at it again, I can't see where it blocks.
> >
> > Might be worth a comment in avc_compute_av(), because the temporary
> > dropping of rcu_read_lock() looks really dodgy when we could be running
> > on the RCU path walk path anyway.
>
> I don't think that's a problem but I'll defer to the fsdevel and rcu folks.
> The use of RCU within the SELinux AVC long predates the introduction of RCU
> path walk, and the rcu_read_lock()/unlock() pairs inside the AVC are not
> related in any way to RCU path walk. Hopefully they don't break it. The
> SELinux security server (i.e. security_compute_av() and the rest of
> security/selinux/ss/*) internally has its own locking for its data
> structures, primarily the policy rwlock. There was also a patch long ago to
> convert use of that policy rwlock to RCU but it didn't seem justified at the
> time. We are interested in revisiting that however. That would introduce
> its own set of rcu_read_lock/unlock pairs inside of security_compute_av() as
> well.

RCU readers nest, so it should be fine. (Famous last words...)

Thanx, Paul