Re: [kernel-hardening] Re: [RFC v4 PATCH 00/13] HARDENED_ATOMIC

[David Windsor]

On Thu, Nov 10, 2016 at 3:56 PM, Kees Cook <keescook@xxxxxxxxxxxx> wrote:
> On Thu, Nov 10, 2016 at 12:37 PM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>> On Thu, Nov 10, 2016 at 10:24:35PM +0200, Elena Reshetova wrote:
>>> This series brings the PaX/Grsecurity PAX_REFCOUNT
>>> feature support to the upstream kernel. All credit for the
>>> feature goes to the feature authors.
>>>
>>> The name of the upstream feature is HARDENED_ATOMIC
>>> and it is configured using CONFIG_HARDENED_ATOMIC and
>>> HAVE_ARCH_HARDENED_ATOMIC.
>>>
>>> This series only adds x86 support; other architectures are expected
>>> to add similar support gradually.
>>>
>>> More information about the feature can be found in the following
>>> commit messages.
>>
>> No, this should be here. As it stands this is completely without
>> content.
>>
>> In any case, NAK on this approach. Its the wrong way around.
>>
>> _IF_ you want to do a non-wrapping variant, it must not be the default.
>
> Unfortunately, we have to do it this way because there are so many
> misuses of atomic_t, and they just keep appearing. We can't do opt-in
> protections for the kernel -- we need to protect atomic_t and opt OUT
> of the protection where it's not needed.
>
> We must change the kernel culture to making things secure-by-default.
> Without this, we're wasting our time and continuing to leave people
> vulnerable every time some new driver lands that refcounts with
> atomic_t. Since education is proven to not work, we have to harden the
> _infrastructure_ of the kernel, of which atomic_t is a part.
>
>> Since you need to audit every single atomic_t user in the kernel anyway,
>> it doesn't matter. But changing atomic_t to non-wrap by default is not
>> robust, if you forgot one, you can then trivially dos the kernel.
>
> Correct: everything must be audited in either case. However, making a
> mistake using opt-out means a DoS. Making a mistake using opt-in means
> an exploitable kernel escalation. We must have the courage to
> recognize this distinction. Right now, every refcount mistake is an
> exploitable kernel flaw. Reducing this to a DoS is a giant
> improvement.
>

Agreed.  And once this DoS happens only once and gets reported, it's
solved forever.

What about the approach of changing the behavior or the overflow
response mechanism (hardened_atomic_overflow()) to only log the
overflow, rather than killing the offending process?  This way, it
could bake in distros for a while and all necessary conversions to
atomic_wrap_t can occur.  Then, we can turn back on the behavior of
killing the offending process.  While behavior after an overflow
occurs is undefined, this was happening before anyway.

>> That said, I still don't much like this.
>>
>> I would much rather you make kref useful and use that. It still means
>> you get to audit all refcounts in the kernel, but hey, you had to do
>> that anyway.
>
> This has already been suggested in the past, and suffers from the same
> opt-in problem. I'll let Greg comment on it, though, as he's agreed
> with going opt-out in the past when reviewing this work.
>
> -Kees
>
> --
> Kees Cook
> Nexus Security