On Mon, Jun 06, 2022 at 12:28:32PM +0100, Marc Zyngier wrote:

Mark Brown [off-list ref] wrote:

quoted

On Sat, May 28, 2022 at 12:38:11PM +0100, Marc Zyngier wrote:

quoted

quoted

We probably never saw the issue because no VMM uses SVE, but
that's still pretty bad. Unconditionally clearing the flag
on vcpu load addresses the issue.

quoted

Unless I'm missing something since we currently always disable
SVE on syscall even if the VMM were using SVE for some reason
(SVE memcpy()?) we should already have disabled SVE for EL0 in
sve_user_discard() during kernel entry so EL0 access to SVE
should be disabled in the system register by the time we get
here.

Indeed. And this begs the question: what is this code actually doing?
Is there any way we can end-up running a guest with any valid host SVE
state?

I remember being >this< close to removing that code some time ago, and
only stopped because I vaguely remembered Dave Martin convincing me at
some point that it was necessary. I'm unable to piece the argument
together again though.

I've stared at that code a few times as well, I think I'd ended up
assuming it was some path to do with preempting and context switching
but in that case I've never been clear why there'd be anything left that
we'd need to preserve, or if we do why we don't just force a
fpsimd_save().  It's possible this was from some earlier stage in review
where the ABI didn't allow us to discard the SVE register state, or that
it's there as defensive programming so for future work where we don't
just disable on entry.

Conicidentally I am going to post some patches later today or tomorrow
which leave SVE enabled on syscall, they still have the hook for
disabling it when entering KVM though so we'd still not need to save the
EL0 state and the above should still apply.