Re: [PATCH v6 23/49] KVM: arm64: gic-v5: Set IRICHPPIDIS based on IRS enable state

[sashiko-bot@kernel.org]

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [High] Skipping the doorbell request when the IRS is disabled breaks vSGI wakeups and causes a KVM blocked-state deadlock when the IRS is re-enabled.
--

commit 614bfb31d4f0af06eb90b62b9d08cfb10894ae32
Author: Sascha Bischoff <Sascha.Bischoff@arm.com>

KVM: arm64: gic-v5: Set IRICHPPIDIS based on IRS enable state

This commit uses the GICv5 ICH_CONTEXTR_EL2 IRICHPPIDIS field to emulate the
guest enabling and disabling the IRS. It programs IRICHPPIDIS when loading a
VPE and skips calculating the doorbell priority mask and requesting a doorbell
when a VPE enters WFI with a disabled IRS.

> diff --git a/arch/arm64/kvm/vgic/vgic-v5.c b/arch/arm64/kvm/vgic/vgic-v5.c
> index bdf06dbff4231..5a5abf51d4ddd 100644
> --- a/arch/arm64/kvm/vgic/vgic-v5.c
> +++ b/arch/arm64/kvm/vgic/vgic-v5.c

[ ... ]

> @@ -1101,7 +1103,10 @@ void vgic_v5_put(struct kvm_vcpu *vcpu)
>  	kvm_call_hyp(__vgic_v5_save_apr, cpu_if);
>  
>  	cpu_if->vgic_contextr = 0;
> -	if (vcpu_get_flag(vcpu, IN_WFI)) {
> +
> +	/* Request a doorbell if entering WFI, unless the IRS is disabled */
> +	if (vcpu_get_flag(vcpu, IN_WFI) &&
> +	    READ_ONCE(vcpu->kvm->arch.vgic.enabled)) {

[Severity: High]
Does skipping the doorbell request when the IRS is disabled break vSGI wakeups
and cause a KVM blocked-state deadlock when the IRS is re-enabled?

If a guest disables the IRS (setting vgic.enabled = false) and a VCPU
executes WFI, vgic_v5_put() will skip requesting the hardware doorbell.
If another VCPU sends a vSGI to this sleeping VCPU, the hardware cannot
assert a doorbell to wake the physical CPU, so the vSGI is ignored.

Furthermore, if another VCPU later re-enables the IRS, KVM sets
KVM_REQ_RELOAD_GICv5 and kicks the sleeping VCPU. The VCPU thread wakes up in
kvm_vcpu_block(), but because KVM_REQ_RELOAD_GICv5 does not set
KVM_REQ_UNBLOCK and does not make kvm_arch_vcpu_runnable() return true
(as db_fired is false), kvm_vcpu_check_block() returns 0. The VCPU thread
goes right back to sleep without processing the request or re-requesting a
doorbell, resulting in a permanent hang.

>  		u32 priority_mask;
>  		int dbpm;
>

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20260904113404.4051341-1-sascha.bischoff@arm.com?part=23