Re: [PATCH v2 5/5] KVM: arm64: Exclude FP ownership from kvm_vcpu_arch

From: Marc Zyngier <maz@kernel.org>
To: Mark Brown <broonie@kernel.org>
Cc: kvmarm@lists.linux.dev, linux-arm-kernel@lists.infradead.org,
	kvm@vger.kernel.org, James Morse <james.morse@arm.com>,
	Suzuki K Poulose <suzuki.poulose@arm.com>,
	Oliver Upton <oliver.upton@linux.dev>,
	Zenghui Yu <yuzenghui@huawei.com>,
	James Clark <james.clark@arm.com>,
	Anshuman Khandual <anshuman.khandual@arm.com>,
	Dongli Zhang <dongli.zhang@oracle.com>
Subject: Re: [PATCH v2 5/5] KVM: arm64: Exclude FP ownership from kvm_vcpu_arch
Date: Wed, 27 Mar 2024 09:04:42 +0000	[thread overview]
Message-ID: <87a5mkrqol.wl-maz@kernel.org> (raw)
In-Reply-To: <55f2f1f7-6f23-45f2-ae6c-a1111e3271db@sirena.org.uk>

On Mon, 25 Mar 2024 14:57:27 +0000,
Mark Brown <broonie@kernel.org> wrote:
> 
> [1  <text/plain; us-ascii (quoted-printable)>]
> On Mon, Mar 25, 2024 at 09:23:18AM +0000, Marc Zyngier wrote:
> > Mark Brown <broonie@kernel.org> wrote:
> 
> > > This was referring to the fact that currently when SVE is enabled access
> > > to the V registers as V registers via _CORE_REG() is blocked and they
> > > can only be accessed as a subset of the Z registers (see the check at
> > > the end of core_reg_size_from_offset() in guest.c).
> 
> > But what behaviour do you expect from allowing such a write? Insert in
> > place? Or zero the upper bits of the vector, as per R_WKYLB? One is
> > wrong, and the other wrecks havoc on unsuspecting userspace.
> 
> It would have to be the former due to the ABI issue I think.

No, that's an architecture violation.

> > My take on this is that when a VM is S*E aware, only the writes to the
> > largest *enabled* registers should take place. This is similar to what
> > we do for FP/SIMD: we only allow writes to the V registers, and not to
> > Q, D, S, H or B, although that happens by construction. For S*E,
> > dropping the write on the floor (or return some error that userspace
> > will understand as benign) is the least bad option.
> 
> OK, this does mean that in the case of a SME only guest we'll end up
> with registers not just changing size but appearing and disappearing
> depending on SVCR.SM.  It wasn't clear to me that this was a good idea
> from an ABI point of view, it's a level up beyond the size changing
> thing and there's a tradeoff with the "model what the architecture does"
> model.

The registers don't have to disappear, they just have to become WI,
just like it is the case today with SVE. Yes, the ABI becomes
contextual, but we're past the point where we can treat the various
register files as a bag of bits that can be save/restored without any
ordering.

We already have similar requirements for complex parts of KVM where
ordering is required, such as GICv3 and the ITS, and we follow what
the architecture requires. The same thing applies for the CPU.

	M.

-- 
Without deviation from the norm, progress is not possible.