Re: [PATCH v5 07/17] arm64: add basic pointer authentication support

[Will Deacon]

[+Cyrill Gorcunov for CRIU stuff]

On Fri, Oct 19, 2018 at 12:15:43PM +0100, Catalin Marinas wrote:
> On Fri, Oct 05, 2018 at 09:47:44AM +0100, Kristina Martsenko wrote:
> > diff --git a/arch/arm64/include/asm/pointer_auth.h b/arch/arm64/include/asm/pointer_auth.h
> > new file mode 100644
> > index 000000000000..2aefedc31d9e
> > --- /dev/null
> > +++ b/arch/arm64/include/asm/pointer_auth.h
> > @@ -0,0 +1,63 @@
> > +// SPDX-License-Identifier: GPL-2.0
> > +#ifndef __ASM_POINTER_AUTH_H
> > +#define __ASM_POINTER_AUTH_H
> > +
> > +#include <linux/random.h>
> > +
> > +#include <asm/cpufeature.h>
> > +#include <asm/sysreg.h>
> > +
> > +#ifdef CONFIG_ARM64_PTR_AUTH
> > +/*
> > + * Each key is a 128-bit quantity which is split across a pair of 64-bit
> > + * registers (Lo and Hi).
> > + */
> > +struct ptrauth_key {
> > +	unsigned long lo, hi;
> > +};
> > +
> > +/*
> > + * We give each process its own instruction A key (APIAKey), which is shared by
> > + * all threads. This is inherited upon fork(), and reinitialised upon exec*().
> > + * All other keys are currently unused, with APIBKey, APDAKey, and APBAKey
> > + * instructions behaving as NOPs.
> > + */
> 
> I don't remember the past discussions but I assume the tools guys are ok
> with a single key shared by multiple threads. Ramana, could you ack this
> part, FTR?
> 
> (and it would help if someone from the Android and Chrome camps can
> confirm)

FWIW: I think we should be entertaining a prctl() interface to use a new
key on a per-thread basis. Obviously, this would need to be used with care
(e.g. you'd fork(); use the prctl() and then you'd better not return from
the calling function!).

Assuming we want this (Kees -- I was under the impression that everything in
Android would end up with the same key otherwise?), then the question is
do we want:

  - prctl() get/set operations for the key, or
  - prctl() set_random_key operation, or
  - both of the above?

Part of the answer to that may lie in the requirements of CRIU, where I
strongly suspect they need explicit get/set operations, although these
could be gated on CONFIG_CHECKPOINT_RESTORE=y.

Will