[PATCH 3/3] KVM/arm64: enable enhanced armv8 fp/simd lazy switch

[Christoffer Dall]

On Fri, Oct 30, 2015 at 02:56:33PM -0700, Mario Smarduch wrote:
> This patch enables arm64 lazy fp/simd switch, similar to arm described in
> second patch. Change from previous version - restore function is moved to
> host. 
> 
> Signed-off-by: Mario Smarduch <m.smarduch at samsung.com>
> ---
>  arch/arm64/include/asm/kvm_host.h |  2 +-
>  arch/arm64/kernel/asm-offsets.c   |  1 +
>  arch/arm64/kvm/hyp.S              | 37 +++++++++++++++++++++++++++++++------
>  3 files changed, 33 insertions(+), 7 deletions(-)
> 
> diff --git a/arch/arm64/include/asm/kvm_host.h b/arch/arm64/include/asm/kvm_host.h
> index 26a2347..dcecf92 100644
> --- a/arch/arm64/include/asm/kvm_host.h
> +++ b/arch/arm64/include/asm/kvm_host.h
> @@ -251,11 +251,11 @@ static inline void kvm_arch_hardware_unsetup(void) {}
>  static inline void kvm_arch_sync_events(struct kvm *kvm) {}
>  static inline void kvm_arch_vcpu_uninit(struct kvm_vcpu *vcpu) {}
>  static inline void kvm_arch_sched_in(struct kvm_vcpu *vcpu, int cpu) {}
> -static inline void kvm_restore_host_vfp_state(struct kvm_vcpu *vcpu) {}
>  
>  void kvm_arm_init_debug(void);
>  void kvm_arm_setup_debug(struct kvm_vcpu *vcpu);
>  void kvm_arm_clear_debug(struct kvm_vcpu *vcpu);
>  void kvm_arm_reset_debug_ptr(struct kvm_vcpu *vcpu);
> +void kvm_restore_host_vfp_state(struct kvm_vcpu *vcpu);
>  
>  #endif /* __ARM64_KVM_HOST_H__ */
> diff --git a/arch/arm64/kernel/asm-offsets.c b/arch/arm64/kernel/asm-offsets.c
> index 8d89cf8..c9c5242 100644
> --- a/arch/arm64/kernel/asm-offsets.c
> +++ b/arch/arm64/kernel/asm-offsets.c
> @@ -124,6 +124,7 @@ int main(void)
>    DEFINE(VCPU_HCR_EL2,		offsetof(struct kvm_vcpu, arch.hcr_el2));
>    DEFINE(VCPU_MDCR_EL2,	offsetof(struct kvm_vcpu, arch.mdcr_el2));
>    DEFINE(VCPU_IRQ_LINES,	offsetof(struct kvm_vcpu, arch.irq_lines));
> +  DEFINE(VCPU_VFP_DIRTY,	offsetof(struct kvm_vcpu, arch.vfp_dirty));
>    DEFINE(VCPU_HOST_CONTEXT,	offsetof(struct kvm_vcpu, arch.host_cpu_context));
>    DEFINE(VCPU_HOST_DEBUG_STATE, offsetof(struct kvm_vcpu, arch.host_debug_state));
>    DEFINE(VCPU_TIMER_CNTV_CTL,	offsetof(struct kvm_vcpu, arch.timer_cpu.cntv_ctl));
> diff --git a/arch/arm64/kvm/hyp.S b/arch/arm64/kvm/hyp.S
> index e583613..ed2c4cf 100644
> --- a/arch/arm64/kvm/hyp.S
> +++ b/arch/arm64/kvm/hyp.S
> @@ -36,6 +36,28 @@
>  #define CPU_SYSREG_OFFSET(x)	(CPU_SYSREGS + 8*x)
>  
>  	.text
> +
> +/**
> + * void kvm_restore_host_vfp_state(struct vcpu *vcpu) - Executes lazy
> + *	fp/simd switch, saves the guest, restores host. Called from host
> + *	mode, placed outside of hyp section.

same comments on style as previous patch

> + */
> +ENTRY(kvm_restore_host_vfp_state)
> +	push    xzr, lr
> +
> +	add     x2, x0, #VCPU_CONTEXT
> +	mov     w3, #0
> +	strb    w3, [x0, #VCPU_VFP_DIRTY]

I've been discussing with myself if it would make more sense to clear
the dirty flag in the C-code...

> +
> +	bl __save_fpsimd
> +
> +	ldr     x2, [x0, #VCPU_HOST_CONTEXT]
> +	bl __restore_fpsimd
> +
> +	pop     xzr, lr
> +	ret
> +ENDPROC(kvm_restore_host_vfp_state)
> +
>  	.pushsection	.hyp.text, "ax"
>  	.align	PAGE_SHIFT
>  
> @@ -482,7 +504,11 @@
>  99:
>  	msr     hcr_el2, x2
>  	mov	x2, #CPTR_EL2_TTA
> +
> +	ldrb	w3, [x0, #VCPU_VFP_DIRTY]
> +	tbnz    w3, #0, 98f
>  	orr     x2, x2, #CPTR_EL2_TFP
> +98:

mmm, don't you need to only set the fpexc32 when you're actually going
to trap the guest accesses?

also, you can consider only setting this in vcpu_load (jumping quickly
to EL2 to do so) if we're running a 32-bit guest.  Probably worth
measuring the difference between the extra EL2 jump on vcpu_load
compared to hitting this register on every entry/exit.

Code-wise, it will be nicer to do it on vcpu_load.

>  	msr	cptr_el2, x2
>  
>  	mov	x2, #(1 << 15)	// Trap CP15 Cr=15
> @@ -669,14 +695,12 @@ __restore_debug:
>  	ret
>  
>  __save_fpsimd:
> -	skip_fpsimd_state x3, 1f
>  	save_fpsimd
> -1:	ret
> +	ret
>  
>  __restore_fpsimd:
> -	skip_fpsimd_state x3, 1f
>  	restore_fpsimd
> -1:	ret
> +	ret
>  
>  switch_to_guest_fpsimd:
>  	push	x4, lr
> @@ -688,6 +712,9 @@ switch_to_guest_fpsimd:
>  
>  	mrs	x0, tpidr_el2
>  
> +	mov     w2, #1
> +	strb    w2, [x0, #VCPU_VFP_DIRTY]

hmm, just noticing this.  Are you not writing a 32-bit value to a
potentially 8-bit field (ignoring padding in the struct), as the dirty
flag is declared a bool.

Are you also doing this on the 32-bit side?

> +
>  	ldr	x2, [x0, #VCPU_HOST_CONTEXT]
>  	kern_hyp_va x2
>  	bl __save_fpsimd
> @@ -763,7 +790,6 @@ __kvm_vcpu_return:
>  	add	x2, x0, #VCPU_CONTEXT
>  
>  	save_guest_regs
> -	bl __save_fpsimd
>  	bl __save_sysregs
>  
>  	skip_debug_state x3, 1f
> @@ -784,7 +810,6 @@ __kvm_vcpu_return:
>  	kern_hyp_va x2
>  
>  	bl __restore_sysregs
> -	bl __restore_fpsimd
>  	/* Clear FPSIMD and Trace trapping */
>  	msr     cptr_el2, xzr
>  
> -- 
> 1.9.1
> 

Thanks,
-Christoffer