[PATCH v1 11/11] KVM: arm64: Implement lazy vCPU state sync for non-protected guests

[Vincent Donnefort]

On Fri, Jun 12, 2026 at 07:59:25AM +0100, tabba at google.com wrote:
> pKVM copies a non-protected guest's register context between the host
> and the hypervisor on every world switch, even when the host never
> inspects it. Defer the copy: on entry, flush the host context into the
> hyp vCPU only when the host marked it dirty (PKVM_HOST_STATE_DIRTY); on
> exit, leave it in the hyp vCPU and copy it back only when the host needs
> it, via a __pkvm_vcpu_sync_state hypercall on trap handling or at vcpu
> put. A protected guest's context is copied as before, since lazy sync
> only helps where the host is trusted to see the guest's registers.
> 
> The PC is the exception: it is copied back on every exit so the
> kvm_exit tracepoint reports the guest's real exit PC rather than the
> value left by the previous sync.
> 
> Signed-off-by: Fuad Tabba <tabba at google.com>
> ---
>  arch/arm64/include/asm/kvm_asm.h   |  1 +
>  arch/arm64/include/asm/kvm_host.h  |  2 +
>  arch/arm64/kvm/arm.c               |  7 +++
>  arch/arm64/kvm/handle_exit.c       | 22 ++++++++
>  arch/arm64/kvm/hyp/nvhe/hyp-main.c | 88 ++++++++++++++++++++++++++++--
>  5 files changed, 115 insertions(+), 5 deletions(-)
> 
> diff --git a/arch/arm64/include/asm/kvm_asm.h b/arch/arm64/include/asm/kvm_asm.h
> index 043495f7fc78..6e1135b3ded4 100644
> --- a/arch/arm64/include/asm/kvm_asm.h
> +++ b/arch/arm64/include/asm/kvm_asm.h
> @@ -113,6 +113,7 @@ enum __kvm_host_smccc_func {
>  	__KVM_HOST_SMCCC_FUNC___pkvm_finalize_teardown_vm,
>  	__KVM_HOST_SMCCC_FUNC___pkvm_vcpu_load,
>  	__KVM_HOST_SMCCC_FUNC___pkvm_vcpu_put,
> +	__KVM_HOST_SMCCC_FUNC___pkvm_vcpu_sync_state,
>  	__KVM_HOST_SMCCC_FUNC___pkvm_tlb_flush_vmid,
>  
>  	MARKER(__KVM_HOST_SMCCC_FUNC_MAX)
> diff --git a/arch/arm64/include/asm/kvm_host.h b/arch/arm64/include/asm/kvm_host.h
> index a49042bfa801..1ef660774adc 100644
> --- a/arch/arm64/include/asm/kvm_host.h
> +++ b/arch/arm64/include/asm/kvm_host.h
> @@ -1113,6 +1113,8 @@ struct kvm_vcpu_arch {
>  /* SError pending for nested guest */
>  #define NESTED_SERROR_PENDING	__vcpu_single_flag(sflags, BIT(8))
>  
> +/* pKVM host vcpu state is dirty, needs resync (nVHE-only) */

nit: with hVHE, I guess we can just drop that nVHE-only? 

> +#define PKVM_HOST_STATE_DIRTY	__vcpu_single_flag(iflags, BIT(4))
>  
>  /* Pointer to the vcpu's SVE FFR for sve_{save,load}_state() */
>  #define vcpu_sve_pffr(vcpu) (kern_hyp_va((vcpu)->arch.sve_state) +	\
> diff --git a/arch/arm64/kvm/arm.c b/arch/arm64/kvm/arm.c
> index c9f36932c980..a5c54e37778b 100644
> --- a/arch/arm64/kvm/arm.c
> +++ b/arch/arm64/kvm/arm.c
> @@ -734,6 +734,10 @@ void kvm_arch_vcpu_put(struct kvm_vcpu *vcpu)
>  	if (is_protected_kvm_enabled()) {
>  		kvm_call_hyp(__vgic_v3_save_aprs, &vcpu->arch.vgic_cpu.vgic_v3);
>  		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);
>  	}
>  
>  	kvm_vcpu_put_debug(vcpu);
> @@ -961,6 +965,9 @@ int kvm_arch_vcpu_run_pid_change(struct kvm_vcpu *vcpu)
>  		return ret;
>  
>  	if (is_protected_kvm_enabled()) {
> +		/* Start with the vcpu in a dirty state */
> +		if (!kvm_vm_is_protected(vcpu->kvm))
> +			vcpu_set_flag(vcpu, PKVM_HOST_STATE_DIRTY);
>  		ret = pkvm_create_hyp_vm(kvm);
>  		if (ret)
>  			return ret;
> diff --git a/arch/arm64/kvm/handle_exit.c b/arch/arm64/kvm/handle_exit.c
> index 54aedf93c78b..dccc3786548b 100644
> --- a/arch/arm64/kvm/handle_exit.c
> +++ b/arch/arm64/kvm/handle_exit.c
> @@ -422,6 +422,21 @@ static int handle_trap_exceptions(struct kvm_vcpu *vcpu)
>  {
>  	int handled;
>  
> +	/*
> +	 * If we run a non-protected VM when protection is enabled
> +	 * system-wide, resync the state from the hypervisor and mark
> +	 * it as dirty on the host side if it wasn't dirty already
> +	 * (which could happen if preemption has taken place).
> +	 */
> +	if (is_protected_kvm_enabled() && !kvm_vm_is_protected(vcpu->kvm)) {
> +		preempt_disable();

nit: since we are introducing guard() with that series, this one could be
guard(preempt)().

> +		if (!(vcpu_get_flag(vcpu, PKVM_HOST_STATE_DIRTY))) {
> +			kvm_call_hyp_nvhe(__pkvm_vcpu_sync_state);
> +			vcpu_set_flag(vcpu, PKVM_HOST_STATE_DIRTY);
> +		}
> +		preempt_enable();
> +	}
> +
>  	/*
>  	 * See ARM ARM B1.14.1: "Hyp traps on instructions
>  	 * that fail their condition code check"
> @@ -489,6 +504,13 @@ int handle_exit(struct kvm_vcpu *vcpu, int exception_index)
>  /* For exit types that need handling before we can be preempted */
>  void handle_exit_early(struct kvm_vcpu *vcpu, int exception_index)
>  {
> +	/*
> +	 * We just exited, so the state is clean from a hypervisor
> +	 * perspective.
> +	 */
> +	if (is_protected_kvm_enabled())
> +		vcpu_clear_flag(vcpu, PKVM_HOST_STATE_DIRTY);
> +
>  	if (ARM_SERROR_PENDING(exception_index)) {
>  		if (this_cpu_has_cap(ARM64_HAS_RAS_EXTN)) {
>  			u64 disr = kvm_vcpu_get_disr(vcpu);
> diff --git a/arch/arm64/kvm/hyp/nvhe/hyp-main.c b/arch/arm64/kvm/hyp/nvhe/hyp-main.c
> index 23e644c24a03..02383b372258 100644
> --- a/arch/arm64/kvm/hyp/nvhe/hyp-main.c
> +++ b/arch/arm64/kvm/hyp/nvhe/hyp-main.c
> @@ -139,6 +139,49 @@ static void sync_hyp_vgic_state(struct pkvm_hyp_vcpu *hyp_vcpu)
>  		host_cpu_if->vgic_lr[i] = hyp_cpu_if->vgic_lr[i];
>  }
>  
> +
> +static void __copy_vcpu_state(const struct kvm_vcpu *from_vcpu,
> +			      struct kvm_vcpu *to_vcpu)
> +{
> +	int i;
> +
> +	to_vcpu->arch.ctxt.regs		= from_vcpu->arch.ctxt.regs;
> +	to_vcpu->arch.ctxt.spsr_abt	= from_vcpu->arch.ctxt.spsr_abt;
> +	to_vcpu->arch.ctxt.spsr_und	= from_vcpu->arch.ctxt.spsr_und;
> +	to_vcpu->arch.ctxt.spsr_irq	= from_vcpu->arch.ctxt.spsr_irq;
> +	to_vcpu->arch.ctxt.spsr_fiq	= from_vcpu->arch.ctxt.spsr_fiq;
> +	to_vcpu->arch.ctxt.fp_regs	= from_vcpu->arch.ctxt.fp_regs;
> +
> +	/*
> +	 * Copy the sysregs, but don't mess with the timer state which
> +	 * is directly handled by EL1 and is expected to be preserved.
> +	 * enum vcpu_sysreg is sparse: VNCR-mapped registers take values
> +	 * derived from their VNCR page offset, so the timer registers do
> +	 * not form a contiguous numeric range and must be skipped by name.
> +	 */
> +	for (i = 1; i < NR_SYS_REGS; i++) {
> +		switch (i) {
> +		case CNTVOFF_EL2:
> +		case CNTV_CVAL_EL0:
> +		case CNTV_CTL_EL0:
> +		case CNTP_CVAL_EL0:
> +		case CNTP_CTL_EL0:
> +			continue;
> +		}
> +		to_vcpu->arch.ctxt.sys_regs[i] = from_vcpu->arch.ctxt.sys_regs[i];
> +	}
> +}
> +
> +static void __sync_hyp_vcpu(struct pkvm_hyp_vcpu *hyp_vcpu)
> +{
> +	__copy_vcpu_state(&hyp_vcpu->vcpu, hyp_vcpu->host_vcpu);
> +}
> +
> +static void __flush_hyp_vcpu(struct pkvm_hyp_vcpu *hyp_vcpu)
> +{
> +	__copy_vcpu_state(hyp_vcpu->host_vcpu, &hyp_vcpu->vcpu);
> +}

nit: Could that be flush/sync_hyp_vcpu_state? as everything this is called
"state" and we already have flush_debug_state() below ?

> +
>  static void flush_debug_state(struct pkvm_hyp_vcpu *hyp_vcpu)
>  {
>  	struct kvm_vcpu *host_vcpu = hyp_vcpu->host_vcpu;
> @@ -168,7 +211,17 @@ static void flush_hyp_vcpu(struct pkvm_hyp_vcpu *hyp_vcpu)
>  	fpsimd_sve_flush();
>  	flush_debug_state(hyp_vcpu);
>  
> -	hyp_vcpu->vcpu.arch.ctxt	= host_vcpu->arch.ctxt;
> +	/*
> +	 * If we deal with a non-protected guest and the state is potentially
> +	 * dirty (from a host perspective), copy the state back into the hyp
> +	 * vcpu.
> +	 */
> +	if (!pkvm_hyp_vcpu_is_protected(hyp_vcpu)) {
> +		if (vcpu_get_flag(host_vcpu, PKVM_HOST_STATE_DIRTY))
> +			__flush_hyp_vcpu(hyp_vcpu);
> +	} else {
> +		hyp_vcpu->vcpu.arch.ctxt = host_vcpu->arch.ctxt;
> +	}
>  
>  	hyp_vcpu->vcpu.arch.mdcr_el2	= host_vcpu->arch.mdcr_el2;
>  	hyp_vcpu->vcpu.arch.hcr_el2 &= ~(HCR_TWI | HCR_TWE);
> @@ -191,9 +244,11 @@ static void sync_hyp_vcpu(struct pkvm_hyp_vcpu *hyp_vcpu)
>  	fpsimd_sve_sync(&hyp_vcpu->vcpu);
>  	sync_debug_state(hyp_vcpu);
>  
> -	host_vcpu->arch.ctxt		= hyp_vcpu->vcpu.arch.ctxt;
> -
> -	host_vcpu->arch.hcr_el2		= hyp_vcpu->vcpu.arch.hcr_el2;
> +	if (pkvm_hyp_vcpu_is_protected(hyp_vcpu))
> +		host_vcpu->arch.ctxt = hyp_vcpu->vcpu.arch.ctxt;
> +	else
> +		/* Keep the PC current for the kvm_exit tracepoint (lazy ctxt sync). */
> +		host_vcpu->arch.ctxt.regs.pc = hyp_vcpu->vcpu.arch.ctxt.regs.pc;
>  
>  	host_vcpu->arch.fault		= hyp_vcpu->vcpu.arch.fault;
>  
> @@ -227,8 +282,30 @@ static void handle___pkvm_vcpu_put(struct kvm_cpu_context *host_ctxt)
>  {
>  	struct pkvm_hyp_vcpu *hyp_vcpu = pkvm_get_loaded_hyp_vcpu();
>  
> -	if (hyp_vcpu)
> +	if (hyp_vcpu) {
> +		struct kvm_vcpu *host_vcpu = hyp_vcpu->host_vcpu;
> +
> +		if (!pkvm_hyp_vcpu_is_protected(hyp_vcpu) &&
> +		    !vcpu_get_flag(host_vcpu, PKVM_HOST_STATE_DIRTY)) {
> +			__sync_hyp_vcpu(hyp_vcpu);
> +		}
> +
>  		pkvm_put_hyp_vcpu(hyp_vcpu);
> +	}
> +}
> +
> +static void handle___pkvm_vcpu_sync_state(struct kvm_cpu_context *host_ctxt)
> +{
> +	struct pkvm_hyp_vcpu *hyp_vcpu;
> +
> +	if (!is_protected_kvm_enabled())
> +		return;

Since "KVM: arm64: Remove is_protected_kvm_enabled() checks from hypercalls" we
got rid of those is_protected_kvm_enabled() for pKVM-only HVCs. (also, it is
declared in the pKVM-only section of the HVCs)

> +
> +	hyp_vcpu = pkvm_get_loaded_hyp_vcpu();
> +	if (!hyp_vcpu || pkvm_hyp_vcpu_is_protected(hyp_vcpu))
> +		return;
> +
> +	__sync_hyp_vcpu(hyp_vcpu);
>  }
>  
>  static struct kvm_vcpu *__get_host_hyp_vcpus(struct kvm_vcpu *arg,
> @@ -859,6 +936,7 @@ static const hcall_t host_hcall[] = {
>  	HANDLE_FUNC(__pkvm_finalize_teardown_vm),
>  	HANDLE_FUNC(__pkvm_vcpu_load),
>  	HANDLE_FUNC(__pkvm_vcpu_put),
> +	HANDLE_FUNC(__pkvm_vcpu_sync_state),
>  	HANDLE_FUNC(__pkvm_tlb_flush_vmid),
>  };
>  
> -- 
> 2.54.0.1136.gdb2ca164c4-goog
>