Re: [PATCH v2] arm64: cpufeature: Export matrix and other features to userspace

[Mark Rutland]

On Wed, Jan 15, 2020 at 09:49:17AM +0000, Will Deacon wrote:
> On Mon, Dec 16, 2019 at 11:33:37AM +0000, Steven Price wrote:
> > Export the features introduced as part of ARMv8.6 exposed in the
> > ID_AA64ISAR1_EL1 and ID_AA64ZFR0_EL1 registers. This introduces the
> > Matrix features (ARMv8.2-I8MM, ARMv8.2-F64MM and ARMv8.2-F32MM) along
> > with BFloat16 (Armv8.2-BF16), speculation invalidation (SPECRES) and
> > Data Gathering Hint (ARMv8.0-DGH).
> > 
> > Signed-off-by: Julien Grall <julien.grall@xxxxxxx>
> > [Added other features in those registers]
> > Signed-off-by: Steven Price <steven.price@xxxxxxx>
> > ---
> > This is a v2 of Julien's patch[1] extended to export all the new
> > features contained within the ID_AA64ISAR1_EL1 and ID_AA64ZFR0_EL1
> > registers.
> > 
> > [1] https://lore.kernel.org/linux-arm-kernel/20191025171056.30641-1-julien.grall@xxxxxxx/
> > 
> >  Documentation/arm64/cpu-feature-registers.rst | 16 ++++++++++
> >  Documentation/arm64/elf_hwcaps.rst            | 31 +++++++++++++++++++
> >  arch/arm64/include/asm/hwcap.h                |  8 +++++
> >  arch/arm64/include/asm/sysreg.h               | 12 +++++++
> >  arch/arm64/include/uapi/asm/hwcap.h           |  8 +++++
> >  arch/arm64/kernel/cpufeature.c                | 20 ++++++++++++
> >  arch/arm64/kernel/cpuinfo.c                   |  8 +++++
> >  7 files changed, 103 insertions(+)
> > 
> > diff --git a/Documentation/arm64/cpu-feature-registers.rst b/Documentation/arm64/cpu-feature-registers.rst
> > index b6e44884e3ad..5382981533f8 100644
> > --- a/Documentation/arm64/cpu-feature-registers.rst
> > +++ b/Documentation/arm64/cpu-feature-registers.rst
> > @@ -200,6 +200,14 @@ infrastructure:
> >       +------------------------------+---------+---------+
> >       | Name                         |  bits   | visible |
> >       +------------------------------+---------+---------+
> > +     | I8MM                         | [55-52] |    y    |
> > +     +------------------------------+---------+---------+
> > +     | DGH                          | [51-48] |    y    |
> > +     +------------------------------+---------+---------+
> > +     | BF16                         | [47-44] |    y    |
> > +     +------------------------------+---------+---------+
> > +     | SPECRES                      | [43-40] |    y    |
> > +     +------------------------------+---------+---------+
> 
> I applied this for CI testing last night, but actually I think it's broken.
> AFAICT, the instructions introduced by SPECRES are behind an SCTLR_EL1
> enable (EnRCTX) which defaults to disabled, so we should either be enabling
> them before setting the HWCAP or not exposing them at all.
> 
> Given that the instructions are not broadcast and are likely to be very
> expensive, I don't think that exposing them to EL0 is a good idea.
> 
> In other words, I'll drop the SPECRES parts from this patch. Sound ok?

I completely agree. We deliberately avoided adding SPECRES for those
reasons previously:

  https://lore.kernel.org/r/20191212144633.GE46910@xxxxxxxxxxxxxxxxxxxxxxxxx

Thanks,
Mark.