[RFC PATCH 3/4] arm64/sve: KVM: Ensure user SVE use traps after vcpu execution

[Christoffer Dall]

On Fri, Nov 17, 2017 at 04:38:54PM +0000, Dave Martin wrote:
> Currently, SVE use can remain untrapped if a KVM vcpu thread is
> preempted inside the kernel and we then switch back to some user
> thread.
> 
> This patch ensures that SVE traps for userspace are enabled before
> switching away from the vcpu thread.

I don't really understand why KVM is any different then any other thread
which could be using SVE that gets preempted?

> 
> In an attempt to preserve some clarity about why and when this is
> needed, kvm_fpsimd_flush_cpu_state() is used as a hook for doing
> this.  This means that this function needs to be called after
> exiting the vcpu instead of before entry: 

I don't understand why the former means the latter?

> this patch moves the call
> as appropriate.  As a side-effect, this will avoid the call if vcpu
> entry is shortcircuited by a signal etc.
> 
> Signed-off-by: Dave Martin <Dave.Martin at arm.com>
> ---
>  arch/arm64/kernel/fpsimd.c | 2 ++
>  virt/kvm/arm/arm.c         | 6 +++---
>  2 files changed, 5 insertions(+), 3 deletions(-)
> 
> diff --git a/arch/arm64/kernel/fpsimd.c b/arch/arm64/kernel/fpsimd.c
> index 3dc8058..3b135eb 100644
> --- a/arch/arm64/kernel/fpsimd.c
> +++ b/arch/arm64/kernel/fpsimd.c
> @@ -1083,6 +1083,8 @@ void sve_flush_cpu_state(void)
>  
>  	if (last->st && last->sve_in_use)
>  		fpsimd_flush_cpu_state();
> +
> +	sve_user_disable();
>  }
>  #endif /* CONFIG_ARM64_SVE */
>  
> diff --git a/virt/kvm/arm/arm.c b/virt/kvm/arm/arm.c
> index 772bf74..554b157 100644
> --- a/virt/kvm/arm/arm.c
> +++ b/virt/kvm/arm/arm.c
> @@ -651,9 +651,6 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu, struct kvm_run *run)
>  		 */
>  		preempt_disable();
>  
> -		/* Flush FP/SIMD state that can't survive guest entry/exit */
> -		kvm_fpsimd_flush_cpu_state();
> -
>  		kvm_pmu_flush_hwstate(vcpu);
>  
>  		local_irq_disable();
> @@ -754,6 +751,9 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu, struct kvm_run *run)
>  		guest_exit();
>  		trace_kvm_exit(ret, kvm_vcpu_trap_get_class(vcpu), *vcpu_pc(vcpu));
>  
> +		/* Flush FP/SIMD state that can't survive guest entry/exit */
> +		kvm_fpsimd_flush_cpu_state();
> +

Could this be done in kvm_arch_vcpu_put() instead?

>  		preempt_enable();
>  
>  		ret = handle_exit(vcpu, run, ret);
> -- 
> 2.1.4
> 

Thanks,
-Christoffer