[PATCH] KVM: arm64: Register irq_bypass consumers only with GICv4

[PATCH] KVM: arm64: Register irq_bypass consumers only with GICv4

[Jinqian Yang]

kvm_arch_has_irq_bypass() on arm64 returns true unconditionally, so
every irqfd assignment registers an irq_bypass consumer even when the
hardware cannot perform GICv4 direct MSI forwarding. In that case
kvm_vgic_v4_set_forwarding() bails out at its very first check, since
vgic_supports_direct_msis() is false without has_gicv4, and the
registration becomes pure overhead: every KVM_IRQFD ioctl takes the
global irq_bypass mutex, which becomes a contention point when many
VMs boot concurrently.

Gate the registration on kvm_vgic_global_state.has_gicv4, the way x86
gates on enable_device_posted_irqs. Functionally this is a no-op:
without GICv4 the consumer registration never did anything beyond
acquiring the lock.  On a GICv3-only host booting 100 microVMs
concurrently this cuts the average boot time by ~5 ms, matching the
gain observed with CONFIG_HAVE_KVM_IRQ_BYPASS=n.

Signed-off-by: Jinqian Yang <yangjinqian1@huawei.com>
---
 arch/arm64/include/asm/kvm_host.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/arch/arm64/include/asm/kvm_host.h b/arch/arm64/include/asm/kvm_host.h
index 97332ffcb130..785d57388032 100644
--- a/arch/arm64/include/asm/kvm_host.h
+++ b/arch/arm64/include/asm/kvm_host.h
@@ -1744,7 +1744,7 @@ void kvm_set_vm_id_reg(struct kvm *kvm, u32 reg, u64 val);
 
 static inline bool kvm_arch_has_irq_bypass(void)
 {
-	return true;
+	return kvm_vgic_global_state.has_gicv4;
 }
 
 void kvm_idregs_enforce_dependencies(struct kvm *kvm);
-- 
2.33.0

Re: [PATCH] KVM: arm64: Register irq_bypass consumers only with GICv4

[Oliver Upton]

Hi,

On Mon, Nov 09, 2026 at 07:20:18PM +0800, Jinqian Yang wrote:
> kvm_arch_has_irq_bypass() on arm64 returns true unconditionally, so
> every irqfd assignment registers an irq_bypass consumer even when the
> hardware cannot perform GICv4 direct MSI forwarding. In that case
> kvm_vgic_v4_set_forwarding() bails out at its very first check, since
> vgic_supports_direct_msis() is false without has_gicv4, and the
> registration becomes pure overhead: every KVM_IRQFD ioctl takes the
> global irq_bypass mutex, which becomes a contention point when many
> VMs boot concurrently.
> 
> Gate the registration on kvm_vgic_global_state.has_gicv4, the way x86
> gates on enable_device_posted_irqs. Functionally this is a no-op:
> without GICv4 the consumer registration never did anything beyond
> acquiring the lock.  On a GICv3-only host booting 100 microVMs
> concurrently this cuts the average boot time by ~5 ms, matching the
> gain observed with CONFIG_HAVE_KVM_IRQ_BYPASS=n.
> 
> Signed-off-by: Jinqian Yang <yangjinqian1@huawei.com>
> ---
>  arch/arm64/include/asm/kvm_host.h | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/arch/arm64/include/asm/kvm_host.h b/arch/arm64/include/asm/kvm_host.h
> index 97332ffcb130..785d57388032 100644
> --- a/arch/arm64/include/asm/kvm_host.h
> +++ b/arch/arm64/include/asm/kvm_host.h
> @@ -1744,7 +1744,7 @@ void kvm_set_vm_id_reg(struct kvm *kvm, u32 reg, u64 val);
>  
>  static inline bool kvm_arch_has_irq_bypass(void)
>  {
> -	return true;
> +	return kvm_vgic_global_state.has_gicv4;
>  }

This should just take a pointer on struct kvm then use
vgic_supports_direct_msis() since having an initialized irqchip is a
precondition of kvm_irqfd_assign().

Thanks,
Oliver