[PATCH v2] KVM: arm64: Mask SErrors in a protected vCPU's host copy at first run

[PATCH v2] KVM: arm64: Mask SErrors in a protected vCPU's host copy at first run

[Fuad Tabba]

The host's copy of a protected vCPU has SErrors masked from reset, so a
host-injected SError is pended through HCR_EL2.VSE and the guest takes
it when it unmasks SErrors. However, the VMM can unmask SErrors in that
copy before the first run, through PSTATE.A or SCTLR2_EL1.NMEA. KVM
then emulates the SError's entry on the host copy, and the guest never
takes it as an SError. Exits that copy PSTATE out set PSTATE.A again,
but nothing clears NMEA. With NMEA set, an SError injected after a
KVM_RUN that completes an MMIO access but returns before entering the
guest also trips WARN_ON(INCREMENT_PC).

Mask SErrors in the host copy when the first run creates the hyp vCPU,
after which KVM_SET_ONE_REG is rejected. An SError injected before then
is still emulated, but only writes registers the VMM can set itself.

Fixes: 872383bd12e11 ("KVM: arm64: Add per-EC entry/exit state marshalling for protected guests")
Reported-by: Sashiko <sashiko-bot@kernel.org>
Closes: https://lore.kernel.org/all/20261001142109.794CA1F000FF@smtp.kernel.org/
Signed-off-by: Fuad Tabba <fuad.tabba@linux.dev>
---
v2:
 - Set PSTATE.A and clear SCTLR2_EL1.NMEA in the host copy when the hyp
   vCPU is created, instead of testing vcpu_is_protected() in
   kvm_inject_serror_esr() (Marc).

Applies on kvmarm/next. A follow-up to "KVM: arm64: Confine protected VM
vCPU state to EL2" [1], from Sashiko's review of its v4 patch 12.

v1: https://lore.kernel.org/r/20261005050352.836980-1-fuad.tabba@linux.dev/
[1] https://lore.kernel.org/all/20261001135711.1640520-1-fuad.tabba@linux.dev/

 arch/arm64/kvm/pkvm.c | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/arch/arm64/kvm/pkvm.c b/arch/arm64/kvm/pkvm.c
index d4822d8bb16b8..49973c689789f 100644
--- a/arch/arm64/kvm/pkvm.c
+++ b/arch/arm64/kvm/pkvm.c
@@ -191,12 +191,18 @@ static int __pkvm_create_hyp_vcpu(struct kvm_vcpu *vcpu)
 	/*
 	 * Mirror EL2's seeding of power_state from mp_state. The hyp vCPU is
 	 * published, so take mp_state_lock against kvm_psci_vcpu_on().
+	 *
+	 * EL2 never reads the VMM's PSTATE or SCTLR2_EL1: undo any unmasking
+	 * so that a host SError is pended through HCR_EL2.VSE.
 	 */
 	if (vcpu_is_protected(vcpu)) {
 		spin_lock(&vcpu->arch.mp_state_lock);
 		if (kvm_arm_vcpu_stopped(vcpu))
 			WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_UNINITIALIZED);
 		spin_unlock(&vcpu->arch.mp_state_lock);
+
+		*vcpu_cpsr(vcpu) |= PSR_A_BIT;
+		__vcpu_rmw_sys_reg(vcpu, SCTLR2_EL1, &=, ~SCTLR2_EL1_NMEA);
 	}
 
 	vcpu_set_flag(vcpu, VCPU_PKVM_FINALIZED);

base-commit: fa22cd9947fc245d71bc40482f7f78eb0d5a4af0
-- 
2.39.5

Re: [PATCH v2] KVM: arm64: Mask SErrors in a protected vCPU's host copy at first run

[Oliver Upton]

Hi Fuad,

On Tue, Oct 06, 2026 at 10:28:26AM +0100, Fuad Tabba wrote:
> The host's copy of a protected vCPU has SErrors masked from reset, so a
> host-injected SError is pended through HCR_EL2.VSE and the guest takes
> it when it unmasks SErrors. However, the VMM can unmask SErrors in that
> copy before the first run, through PSTATE.A or SCTLR2_EL1.NMEA. KVM
> then emulates the SError's entry on the host copy, and the guest never
> takes it as an SError. Exits that copy PSTATE out set PSTATE.A again,
> but nothing clears NMEA. With NMEA set, an SError injected after a
> KVM_RUN that completes an MMIO access but returns before entering the
> guest also trips WARN_ON(INCREMENT_PC).
> 
> Mask SErrors in the host copy when the first run creates the hyp vCPU,
> after which KVM_SET_ONE_REG is rejected. An SError injected before then
> is still emulated, but only writes registers the VMM can set itself.
> 
> Fixes: 872383bd12e11 ("KVM: arm64: Add per-EC entry/exit state marshalling for protected guests")
> Reported-by: Sashiko <sashiko-bot@kernel.org>
> Closes: https://lore.kernel.org/all/20261001142109.794CA1F000FF@smtp.kernel.org/
> Signed-off-by: Fuad Tabba <fuad.tabba@linux.dev>
> ---
> v2:
>  - Set PSTATE.A and clear SCTLR2_EL1.NMEA in the host copy when the hyp
>    vCPU is created, instead of testing vcpu_is_protected() in
>    kvm_inject_serror_esr() (Marc).
> 
> Applies on kvmarm/next. A follow-up to "KVM: arm64: Confine protected VM
> vCPU state to EL2" [1], from Sashiko's review of its v4 patch 12.
> 
> v1: https://lore.kernel.org/r/20261005050352.836980-1-fuad.tabba@linux.dev/
> [1] https://lore.kernel.org/all/20261001135711.1640520-1-fuad.tabba@linux.dev/
> 
>  arch/arm64/kvm/pkvm.c | 6 ++++++
>  1 file changed, 6 insertions(+)
> 
> diff --git a/arch/arm64/kvm/pkvm.c b/arch/arm64/kvm/pkvm.c
> index d4822d8bb16b8..49973c689789f 100644
> --- a/arch/arm64/kvm/pkvm.c
> +++ b/arch/arm64/kvm/pkvm.c
> @@ -191,12 +191,18 @@ static int __pkvm_create_hyp_vcpu(struct kvm_vcpu *vcpu)
>  	/*
>  	 * Mirror EL2's seeding of power_state from mp_state. The hyp vCPU is
>  	 * published, so take mp_state_lock against kvm_psci_vcpu_on().
> +	 *
> +	 * EL2 never reads the VMM's PSTATE or SCTLR2_EL1: undo any unmasking
> +	 * so that a host SError is pended through HCR_EL2.VSE.
>  	 */
>  	if (vcpu_is_protected(vcpu)) {
>  		spin_lock(&vcpu->arch.mp_state_lock);
>  		if (kvm_arm_vcpu_stopped(vcpu))
>  			WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_UNINITIALIZED);
>  		spin_unlock(&vcpu->arch.mp_state_lock);
> +
> +		*vcpu_cpsr(vcpu) |= PSR_A_BIT;
> +		__vcpu_rmw_sys_reg(vcpu, SCTLR2_EL1, &=, ~SCTLR2_EL1_NMEA);
>  	}

Urgh, I hadn't realized that PSTATE.A is still modifiable by userspace
prior to KVM_RUN. In that case, I would prefer the vcpu_has_nv()
approach that I had recommended as a backup.

FWIW, since pVMs do not implement FEAT_SCTLR2 it should not be able to
write to the register at all.

Thanks,
Oliver

Re: [PATCH v2] KVM: arm64: Mask SErrors in a protected vCPU's host copy at first run

[Fuad Tabba]

Hi Oliver,

On Tue, 6 Oct 2026 at 15:15, Oliver Upton <oupton@kernel.org> wrote:
[...]
> > +
> > +             *vcpu_cpsr(vcpu) |= PSR_A_BIT;
> > +             __vcpu_rmw_sys_reg(vcpu, SCTLR2_EL1, &=, ~SCTLR2_EL1_NMEA);
> >       }
>
> Urgh, I hadn't realized that PSTATE.A is still modifiable by userspace
> prior to KVM_RUN. In that case, I would prefer the vcpu_has_nv()
> approach that I had recommended as a backup.

Sure. I'll respin with that.

Cheers,
/fuad

> FWIW, since pVMs do not implement FEAT_SCTLR2 it should not be able to
> write to the register at all.
>
> Thanks,
> Oliver