[RFC PATCH untested] arm64: kernel: implement fast refcount checking

On Tue, Jul 25, 2017 at 10:20 AM, Ard Biesheuvel
[off-list ref] wrote:

On 25 July 2017 at 18:13, Kees Cook [off-list ref] wrote:

quoted

On Tue, Jul 25, 2017 at 4:49 AM, Ard Biesheuvel
[off-list ref] wrote:

quoted

Hi all,

I had a stab at porting the fast refcount checks to arm64. It is slightly
less straight-forward than x86 given that we need to support both LSE and
LL/SC, and fallback to the latter if running a kernel built with support
for the former on hardware that does not support it.

It is build tested with and without LSE support, and boots fine on non-LSE
hardware in both cases.

Ah! Very cool. Hopefully you and Li can compare notes; I think they've
been working on an implementation too.

I wasn't aware of that.

quoted

quoted

Suggestions welcome as to how to test and/or benchmark this,

I'll post a patch for LKDTM that I've been using. It's more
comprehensive than the existing ATOMIC checks (which predated the
refcount-only protection).

OK. One thing I couldn't figure out: is refcount_t signed or not? The
saturate tests set the initial value to UINT_MAX - 1, but this is
interpreted as a negative value and so the refcount manipulations that
are expected to succeed also fail in my case.

refcount_t under REFCOUNT_FULL is unsigned. Under the x86 fast
refcount, it's signed to gain the CPU flag detection for overflow. The
understanding is basically "omg, if you've got INT_MAX-many references
to something you already DoSed your machine".

I'll have the full LKDTM tests up in a moment here, just doing another
pass on them now...

-Kees

-- 
Kees Cook
Pixel Security