Re: [PATCH v17 05/20] KVM: arm64: Add vcpu load/put call backs for flavors

On 13/09/2026 11:26, Marc Zyngier wrote:

On Tue, 08 Sep 2026 17:22:08 +0100,
Suzuki K Poulose [off-list ref] wrote:

quoted

Add VM flavor specific handlers for VCPU load/put, in an effort to make it
easier to follow the code.

Based on a patch by Marc Zyngier

Suggested-by: Marc Zyngier <maz@kernel.org>
Signed-off-by: Suzuki K Poulose <suzuki.poulose@arm.com>
---
  arch/arm64/include/asm/kvm_host.h |   6 ++
  arch/arm64/kvm/arm.c              | 156 ++++++++++++++++++++++--------
  2 files changed, 123 insertions(+), 39 deletions(-)

diff --git a/arch/arm64/include/asm/kvm_host.h b/arch/arm64/include/asm/kvm_host.h
index d0dccc9ad6aa8..b2e99c5cb1cd3 100644
--- a/arch/arm64/include/asm/kvm_host.h
+++ b/arch/arm64/include/asm/kvm_host.h

@@ -150,6 +150,11 @@ struct kvm_vmid {
  	atomic64_t id;

...

quoted

diff --git a/arch/arm64/kvm/arm.c b/arch/arm64/kvm/arm.c
index 51fc651267157..9af3bbb2f8c24 100644
--- a/arch/arm64/kvm/arm.c
+++ b/arch/arm64/kvm/arm.c

@@ -74,6 +74,8 @@ struct kvm_ioctl_cap_map {
  	long ext;
  };

...

quoted

+
+static void nvhe_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
+{
+	vcpu_prepare_mmu(vcpu);
+	/*
+	 * The timer must be loaded before the vgic to correctly set up physical
+	 * interrupt deactivation in nested state (e.g. timer interrupt).
+	 */

This comment makes no sense here -- it is strictly for VHE, which is
the only mode to implement NV. Same thing for the pKVM vcpu_load().

Ack, removed everywhere except for VHE.

quoted

+	kvm_timer_vcpu_load(vcpu);
+	kvm_vgic_load(vcpu);
+	kvm_vcpu_load_debug(vcpu);
+	kvm_vcpu_load_fgt(vcpu);
+	kvm_arch_vcpu_load_fp(vcpu);
+	kvm_vcpu_pmu_restore_guest(vcpu);
+
+	vcpu_load_pvtime(vcpu);
+	vcpu_set_wfx_traps(vcpu);
+	vcpu_set_pauth_traps(vcpu);
+}
+
+static void pkvm_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
+{
+	/*
+	 * The timer must be loaded before the vgic to correctly set up physical
+	 * interrupt deactivation in nested state (e.g. timer interrupt).
+	 */
+	kvm_timer_vcpu_load(vcpu);
+	kvm_vgic_load(vcpu);
+	kvm_vcpu_load_debug(vcpu);
+	kvm_vcpu_load_fgt(vcpu);
+	kvm_arch_vcpu_load_fp(vcpu);
+	kvm_vcpu_pmu_restore_guest(vcpu);
+
+	vcpu_load_pvtime(vcpu);
+	vcpu_set_wfx_traps(vcpu);
+
+	kvm_call_hyp_nvhe(__pkvm_vcpu_load,
+			  vcpu->kvm->arch.pkvm.handle,
+			  vcpu->vcpu_idx, vcpu->arch.hcr_el2);
+	kvm_call_hyp(__vgic_v3_restore_vmcr_aprs,
+		     &vcpu->arch.vgic_cpu.vgic_v3);

This can also be turned into a kvm_call_hyp_nvhe().

Ack

quoted

+}
+

...

quoted

+
+static void pkvm_vcpu_put(struct kvm_vcpu *vcpu)
+{
+	kvm_call_hyp(__vgic_v3_save_aprs, &vcpu->arch.vgic_cpu.vgic_v3);

Same thing here about kvm_call_hyp_nvhe().

Ack

quoted

+	kvm_call_hyp_nvhe(__pkvm_vcpu_put);
+
+	/* __pkvm_vcpu_put implies a sync of the state */
+	if (!kvm_vm_is_protected(vcpu->kvm))
+		vcpu_set_flag(vcpu, PKVM_HOST_STATE_DIRTY);
+
+	nvhe_vcpu_put(vcpu);

I'm not overly fond of this. Yes, that was in my original patch. But
for example, we end-up calling kvm_arm_vmid_clear_active() for pKVM.
This is harmless, but conceptually wrong.

I'd rather you expand the whole thing.

Done

quoted

+}
+
+void kvm_arch_vcpu_put(struct kvm_vcpu *vcpu)
+{
+	vcpu->arch.vcpu_ops->vcpu_put(vcpu);
  	vcpu_clear_on_unsupported_cpu(vcpu);
  	vcpu->cpu = -1;
  }

@@ -2136,6 +2192,28 @@ int kvm_arch_vm_ioctl(struct file *filp, unsigned int ioctl, unsigned long arg)
  	}
  }

quoted

+
+static const struct kvm_vcpu_ops *arm64_vcpu_ops[] = {
+	[VM_VHE] = &vhe_vcpu_ops,
+	[VM_NVHE] = &nvhe_vcpu_ops,
+	[VM_PKVM] = &pkvm_vcpu_ops,
+	[VM_PROTECTED_PKVM] = &pkvm_vcpu_ops,

nit: my OCD-self wants to align all the '=' signs vertically...

I have switched to a macro initialiser, consistent with vm_s2_mmu_ops. So it looks much cleaner

Cheers
Suzuki