On Tue, Feb 04, 2025 at 06:00:24PM +0000, Marc Zyngier wrote:
> On Tue, 04 Feb 2025 15:21:00 +0000,
> Mark Rutland <mark.rutland@xxxxxxx> wrote:
> > +static inline void fpsimd_lazy_switch_to_guest(struct kvm_vcpu *vcpu)
> > +{
> > + u64 zcr_el1, zcr_el2;
> > +
> > + if (!guest_owns_fp_regs())
> > + return;
> > +
> > + if (vcpu_has_sve(vcpu)) {
> > + /* A guest hypervisor may restrict the effective max VL. */
> > + if (vcpu_has_nv(vcpu) && !is_hyp_ctxt(vcpu))
> > + zcr_el2 = __vcpu_sys_reg(vcpu, ZCR_EL2);
> > + else
> > + zcr_el2 = vcpu_sve_max_vq(vcpu) - 1;
> > +
> > + sve_cond_update_zcr_vq(zcr_el2, SYS_ZCR_EL2);
>
> Not a big deal, but I though I'd mention it here: Using ZCR_EL2 (or
> any other register using the _EL2 suffix) is a source of expensive
> traps with NV. We're much better off using the _EL1 accessor if we are
> running VHE, as this will involve no trap at all.
>
> nVHE will of course trap, but using nVHE with SVE under NV is not
> something I'm prepared to give a damn about.
Ah, sorry. I had forgotten that wrinkle.
Given the compiler warnings reported by Mark Brown [1] and the kernel
test robot [2], I'll go spin a v2 with that cleaned up.
I'll use write_sysreg_el2() here, i.e.
if (vcpu_has_sve(vcpu)) {
/* A guest hypervisor may restrict the effective max VL. */
if (vcpu_has_nv(vcpu) && !is_hyp_ctxt(vcpu))
zcr_el2 = __vcpu_sys_reg(vcpu, ZCR_EL2);
else
zcr_el2 = vcpu_sve_max_vq(vcpu) - 1;
write_sysreg_el2(scr_el2, SYS_ZCR);
}
That'll use the preferred alias automatically, and it matches the style
used to write to ZCR_EL{1,12} immediately after.
Likewise for the other instances.
Mark.
[1] https://lore.kernel.org/linux-arm-kernel/b76803b7-c1b3-426b-a375-0c01b98142c9@xxxxxxxxxxxxx/
[2] https://lore.kernel.org/oe-kbuild-all/202502061341.FvsCMKEH-lkp@xxxxxxxxx/
