Re: [PATCH 2/2] Add a new sysctl knob: unprivileged_userfaultfd_user_mode_only

[Kees Cook <keescook@chromium.org>]

On Wed, May 20, 2020 at 12:59:38AM -0400, Andrea Arcangeli wrote:
> Hello everyone,
> 
> On Fri, May 08, 2020 at 12:54:03PM -0400, Michael S. Tsirkin wrote:
> > On Fri, May 08, 2020 at 12:52:34PM -0400, Michael S. Tsirkin wrote:
> > > On Wed, Apr 22, 2020 at 05:26:32PM -0700, Daniel Colascione wrote:
> > > > This sysctl can be set to either zero or one. When zero (the default)
> > > > the system lets all users call userfaultfd with or without
> > > > UFFD_USER_MODE_ONLY, modulo other access controls. When
> > > > unprivileged_userfaultfd_user_mode_only is set to one, users without
> > > > CAP_SYS_PTRACE must pass UFFD_USER_MODE_ONLY to userfaultfd or the API
> > > > will fail with EPERM. This facility allows administrators to reduce
> > > > the likelihood that an attacker with access to userfaultfd can delay
> > > > faulting kernel code to widen timing windows for other exploits.
> > > > 
> > > > Signed-off-by: Daniel Colascione <dancol@google.com>
> > > 
> > > The approach taken looks like a hard-coded security policy.
> > > For example, it won't be possible to set the sysctl knob
> > > in question on any sytem running kvm. So this is
> > > no good for any general purpose system.

Not all systems run unprivileged KVM. :)

> > > What's wrong with using a security policy for this instead?
> > 
> > In fact I see the original thread already mentions selinux,
> > so it's just a question of making this controllable by
> > selinux.
> 
> I agree it'd be preferable if it was not hardcoded, but then this
> patchset is also much simpler than the previous controlling it through
> selinux..
> 
> I was thinking, an alternative policy that could control it without
> hard-coding it, is a seccomp-bpf filter, then you can drop 2/2 as
> well, not just 1/6-4/6.

Err, did I miss a separate 6-patch series? I can't find anything on lore.

> 
> If you keep only 1/2, can't seccomp-bpf enforce userfaultfd to be
> always called with flags==0x1 without requiring extra modifications in
> the kernel?

Please no. This is way too much overhead for something that a system
owner wants to enforce globally. A sysctl is the correct option here,
IMO. If it needs to be a per-userns sysctl, that would be fine too.

> Can't you get the feature party with the CAP_SYS_PTRACE capability
> too, if you don't wrap those tasks with the ptrace capability under
> that seccomp filter?
> 
> As far as I can tell, it's unprecedented to create a flag for a
> syscall API, with the only purpose of implementing a seccomp-bpf
> filter verifying such flag is set, but then if you want to control it
> with LSM it's even more complex than doing it with seccomp-bpf, and it
> requires more kernel code too. We could always add 2/2 later, such
> possibility won't disappear, in fact we could also add 1/6-4/6 later
> too if that is not enough.
> 
> If we could begin by merging only 1/2 from this new series and be done
> with the kernel changes, because we offload the rest of the work to
> the kernel eBPF JIT, I think it'd be ideal.

I'd agree that patch 1 should land, as it appears to be required for any
further policy considerations. I'm still a big fan of a sysctl since
this is the kind of thing I would absolutely turn on globally for all my
systems.

-- 
Kees Cook