Re: [PATCH v6 6/6] arm/arm64: KVM: Enable armv8 fp/simd enhanced context switch

[Mario Smarduch <m.smarduch@samsung.com>]

On 1/10/2016 8:32 AM, Christoffer Dall wrote:
> On Sat, Dec 26, 2015 at 01:56:56PM -0800, Mario Smarduch wrote:
>> Enable armv8 enhanced fp/simd context switch. Guest and host registers are only
>> context switched on first access and vcpu put.
>>
>> Signed-off-by: Mario Smarduch <m.smarduch@samsung.com>
>> ---
>>  arch/arm/kvm/arm.c              | 13 +++++++++++--
>>  arch/arm64/kernel/asm-offsets.c |  1 +
>>  arch/arm64/kvm/hyp/entry.S      |  1 +
>>  arch/arm64/kvm/hyp/switch.c     | 26 ++------------------------
>>  4 files changed, 15 insertions(+), 26 deletions(-)
>>
>> diff --git a/arch/arm/kvm/arm.c b/arch/arm/kvm/arm.c
>> index b16ed98..633a208 100644
>> --- a/arch/arm/kvm/arm.c
>> +++ b/arch/arm/kvm/arm.c
>> @@ -316,10 +316,19 @@ void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
>>  void kvm_arch_vcpu_put(struct kvm_vcpu *vcpu)
>>  {
>>  	/* If the fp/simd registers are dirty save guest, restore host. */
>> -	if (vcpu_vfp_isdirty(vcpu))
>> +	if (vcpu_vfp_isdirty(vcpu)) {
>> +
>>  		vcpu_restore_host_vfp_state(vcpu);
>>  
>> -	/* Restore host FPEXC trashed in vcpu_load */
>> +		/*
>> +		 * For 32bit guest on arm64 save the guest fpexc register
>> +		 * in EL2 mode.
>> +		 */
>> +		if (vcpu_guest_is_32bit(vcpu))
>> +			vcpu_save_fpexc(vcpu);
>> +	}
>> +
>> +	/* For arm32 restore host FPEXC trashed in vcpu_load. */
>>  	vcpu_restore_host_fpexc(vcpu);
>>  
>>  	/*
>> diff --git a/arch/arm64/kernel/asm-offsets.c b/arch/arm64/kernel/asm-offsets.c
>> index 94090a6..d69145c 100644
>> --- a/arch/arm64/kernel/asm-offsets.c
>> +++ b/arch/arm64/kernel/asm-offsets.c
>> @@ -112,6 +112,7 @@ int main(void)
>>    DEFINE(VCPU_ESR_EL2,		offsetof(struct kvm_vcpu, arch.fault.esr_el2));
>>    DEFINE(VCPU_FAR_EL2,		offsetof(struct kvm_vcpu, arch.fault.far_el2));
>>    DEFINE(VCPU_HPFAR_EL2,	offsetof(struct kvm_vcpu, arch.fault.hpfar_el2));
>> +  DEFINE(VCPU_CPTR_EL2,		offsetof(struct kvm_vcpu, arch.cptr_el2));
>>    DEFINE(VCPU_HOST_CONTEXT,	offsetof(struct kvm_vcpu, arch.host_cpu_context));
>>  #endif
>>  #ifdef CONFIG_CPU_PM
>> diff --git a/arch/arm64/kvm/hyp/entry.S b/arch/arm64/kvm/hyp/entry.S
>> index fd0fbe9..ce7e903 100644
>> --- a/arch/arm64/kvm/hyp/entry.S
>> +++ b/arch/arm64/kvm/hyp/entry.S
>> @@ -136,6 +136,7 @@ ENTRY(__fpsimd_guest_restore)
>>  	isb
>>  
>>  	mrs	x3, tpidr_el2
>> +	str	w2, [x3, #VCPU_CPTR_EL2]
> 
> I'm confused here, why do we need to do this now and not in the previous
> patch?

There should be no harm doing it in previous patch, but this patch
activates the lazy switch and I thought this would be a better
place for it.

> 
> Maybe it helps if we merge these last two patches into one.
> 
>>  
>>  	ldr	x0, [x3, #VCPU_HOST_CONTEXT]
>>  	kern_hyp_va x0
>> diff --git a/arch/arm64/kvm/hyp/switch.c b/arch/arm64/kvm/hyp/switch.c
>> index ca8f5a5..962d179 100644
>> --- a/arch/arm64/kvm/hyp/switch.c
>> +++ b/arch/arm64/kvm/hyp/switch.c
>> @@ -19,24 +19,10 @@
>>  
>>  static void __hyp_text __activate_traps(struct kvm_vcpu *vcpu)
>>  {
>> -	u64 val;
>> -
>> -	/*
>> -	 * We are about to set CPTR_EL2.TFP to trap all floating point
>> -	 * register accesses to EL2, however, the ARM ARM clearly states that
>> -	 * traps are only taken to EL2 if the operation would not otherwise
>> -	 * trap to EL1.  Therefore, always make sure that for 32-bit guests,
>> -	 * we set FPEXC.EN to prevent traps to EL1, when setting the TFP bit.
>> -	 */
>> -	val = vcpu->arch.hcr_el2;
>> -	if (!(val & HCR_RW)) {
>> -		write_sysreg(1 << 30, fpexc32_el2);
>> -		isb();
>> -	}
>> -	write_sysreg(val, hcr_el2);
>> +	write_sysreg(vcpu->arch.hcr_el2, hcr_el2);
>>  	/* Trap on AArch32 cp15 c15 accesses (EL1 or EL0) */
>>  	write_sysreg(1 << 15, hstr_el2);
>> -	write_sysreg(CPTR_EL2_TTA | CPTR_EL2_TFP, cptr_el2);
>> +	write_sysreg(vcpu->arch.cptr_el2, cptr_el2);
>>  	write_sysreg(vcpu->arch.mdcr_el2, mdcr_el2);
>>  }
>>  
>> @@ -89,7 +75,6 @@ static int __hyp_text __guest_run(struct kvm_vcpu *vcpu)
>>  {
>>  	struct kvm_cpu_context *host_ctxt;
>>  	struct kvm_cpu_context *guest_ctxt;
>> -	bool fp_enabled;
>>  	u64 exit_code;
>>  
>>  	vcpu = kern_hyp_va(vcpu);
>> @@ -119,8 +104,6 @@ static int __hyp_text __guest_run(struct kvm_vcpu *vcpu)
>>  	exit_code = __guest_enter(vcpu, host_ctxt);
>>  	/* And we're baaack! */
>>  
>> -	fp_enabled = __fpsimd_enabled();
>> -
>>  	__sysreg_save_state(guest_ctxt);
>>  	__sysreg32_save_state(vcpu);
>>  	__timer_save_state(vcpu);
>> @@ -131,11 +114,6 @@ static int __hyp_text __guest_run(struct kvm_vcpu *vcpu)
>>  
>>  	__sysreg_restore_state(host_ctxt);
>>  
>> -	if (fp_enabled) {
>> -		__fpsimd_save_state(&guest_ctxt->gp_regs.fp_regs);
>> -		__fpsimd_restore_state(&host_ctxt->gp_regs.fp_regs);
>> -	}
>> -
> 
> why do we remove this logic here but preserve something in
> __sysreg32_save_state() ?

Missed it, fpexec code should be removed, that's taken care of
in vcpu_put which stores it to same memory. Thanks for spotting it.

> 
> 
> 
>>  	__debug_save_state(vcpu, kern_hyp_va(vcpu->arch.debug_ptr), guest_ctxt);
>>  	__debug_cond_restore_host_state(vcpu);
>>  
>> -- 
>> 1.9.1
>>
> Thanks,
> -Christoffer
>