Re: gcc-8: arm64/kvm/pauth.: Error: unknown architectural extension `pauth'

[Mark Rutland]

On Mon, Apr 22, 2024 at 02:11:05PM +0200, Arnd Bergmann wrote:
> On Mon, Apr 22, 2024, at 11:40, Mark Rutland wrote:
> > On Mon, Apr 22, 2024 at 11:25:25AM +0200, Arnd Bergmann wrote:
> >> On Mon, Apr 22, 2024, at 11:13, Mark Rutland wrote:
> >> > On Mon, Apr 22, 2024 at 02:04:43PM +0530, Naresh Kamboju wrote:
> >> > Given the minimum supported toolchain comes with an assembler that doesn't
> >> > necessarily support ARMv8.3, I reckon we'll either have to make NV pauth
> >> > support depend upon AS_HAS_ARMV8_3, or manually assemble the PACGA instruction.
> >> >
> >> > I suspect the latter is the better option.
> >> 
> >> The .config linked from the report shows
> >> 
> >> CONFIG_AS_VERSION=23101
> >> CONFIG_ARM64_PTR_AUTH_KERNEL=y
> >> CONFIG_AS_HAS_ARMV8_3=y
> >> 
> >> So it gets detected as supporting ARMv8.3. Is this the wrong
> >> conditional to check, or does it get misdetected for an unsupported
> >> assembler?
> >
> > I suspect that means the 'pauth' arch extension was added after armv8.3
> > support, and the assembler supports `-march=armv8.3-a` but does not support
> > `.arch_extension pauth`. So for this code, it'd be wrong to check for
> > AS_HAS_ARMV8_3, unless we used `.march armv8.3-a`, but even then that'd still
> > mean configurations where we couldn't support this code.
> >
> > I reckon manually assembing the PACGA is the best thing to do; that sidesteps
> > the need for either `.arch_extension pauth` or `.march armv8.3-a`, and aligns
> > with what we do for CONFIG_ARM64_PTR_AUTH=y generally.
> >
> > Elsewhere in the kernel where we check for CONFIG_AS_HAS_ARMV8_3, we rely on
> > ARM64_ASM_PREAMBLE containing `.arch armv8.3-a` or a later version that implies
> > the presence of ARMv8.3-A instructions, and so pauth usage elsewhere is fine.
> 
> I tested with the old binutils versions I have here and found
> that anything that supports v8.3 also understands pacga, but
> '.arch_extension pauth' only works in binutils-2.35 and higher,
> presumably because it started out as a v8.3+ feature but was
> later turned into an optional extension for all versions.
> 
> Since there is a Kconfig check for armv8.3-a support already, I think
> it's safe to just drop the .arch_extension pauth.

That'll be safe, but it does mean that we'd need to *not* support pointer auth
for nested virt when we have a toolchain for which CONFIG_AS_HAS_ARMV8_3=n,
unless our minimum supported AS supports ARMv8.3.

If our minimum supported AS *doesn't* support ARMv8.3, then we'd either need a
new Kconfig symbol for NV_PAUTH support, or make CONFIG_ARM64_PTR_AUTH depend
upon CONFIG_AS_HAS_ARMV8_3.

AFAICT our options are:

(a) Manually assembly PACGA

(b) Change CONFIG_ARM64_PTR_AUTH to depend upon CONFIG_AS_HAS_ARMV8_3=y

(c) Add and use new Kconfig symbol for NV PAUTH, dependent upon
    CONFIG_AS_HAS_ARMV8_3=y

(d) Bump the minimum supported version of AS so that we can depend upon ARMv8.3
    support, and just open-code the ".arch armv8.3-a" in the NV pauth code.

.. and maybe some variations on that.

Mark.