Re: [kernel-hardening] Re: [PATCH 4/4] refcount: Report failures through CHECK_DATA_CORRUPTION

[Mark Rutland <mark.rutland@arm.com>]

On Tue, Feb 07, 2017 at 01:36:30PM +0100, Peter Zijlstra wrote:
> On Tue, Feb 07, 2017 at 11:10:12AM +0000, Mark Rutland wrote:
> > On Tue, Feb 07, 2017 at 09:34:05AM +0100, Peter Zijlstra wrote:
> > > On Mon, Feb 06, 2017 at 08:54:38AM -0800, Kees Cook wrote:
> > > > >
> > > > > Like I wrote, ideally we'd end up using something like the x86 exception
> > > > > table with a custom handler. Just no idea how to pull that off without
> > > > > doing a full blown arch specific implementation, so I didn't go there
> > > > > quite yet.
> > > > 
> > > > I haven't spent much time looking at the extable stuff. (Though
> > > > coincidentally, I was poking at it for x86's test_nx stuff...) I
> > > > thought there was a way to build arch-agnostic extables already?
> > > > kernel/extable.c is unconditionally built-in, for example.
> > > 
> > > That doesn't seem to be of much use. It only contains section sort and
> > > search functions.
> > > 
> > > Another problem for generic code would be to figure out what register
> > > the relevant variable would live in at the time of exception. Here its
> > > 'obviously' EAX because that's what cmpxchg requires, but in generic
> > > you'd need a means of querying GCC's register allocator at the exception
> > > point and somehow using that information for the generation of the
> > > exception handler.
> > 
> > I think we only need two arch-specific primitives:
> > (a) mangle a GCC assigned register into an idx stored in the extable
> > (b) take said index, and grab the relevant register from pt_regs
> > 
> > Then you can have a BUG_VALUE(v, ...), where we use an input "r" (val),
> > and mangle that into the idx in the extable. In the common case, I'd
> > hope GCC would leave the register in-place from the cmpxchg.
> > 
> > ... or have I misundertood? :)
> 
> Right something along those lines. (a) will need GCC help, and (b) would
> be kernel-arch specific. So this isn't something we can quickly do.

I agree this isn't something that can be hacked together quickly, and
certainly shouldn't block these patches.

However, I don't think we need anything new from GCC, and I think we
already have a generic API for (b).

For (a) we don't need new GCC help if we do something like we did in
commit 72c5839515260dce to do the mangling. Prepend a prefix to the
register, e.g. changing 'x0' to '__pt_regs_offset_x0', which we arrange
to hold the correct value.

For (b) we already have regs_get_register().

Thanks,
Mark.