[PATCH v2 2/2] KVM/arm: enable enhanced armv7 fp/simd lazy switch

Christoffer Dall christoffer.dall at linaro.org
Thu Oct 22 14:20:33 PDT 2015


On Tue, Oct 20, 2015 at 06:10:59PM -0700, Mario Smarduch wrote:
> 
> 
> On 10/20/2015 12:24 AM, Christoffer Dall wrote:
> > On Mon, Oct 19, 2015 at 04:25:04PM -0700, Mario Smarduch wrote:
> >>
> >>
> >> On 10/19/2015 3:14 AM, Christoffer Dall wrote:
> >>> On Sat, Sep 26, 2015 at 04:43:29PM -0700, Mario Smarduch wrote:
> >>>> This patch enhances current lazy vfp/simd hardware switch. In addition to
> >>>> current lazy switch, it tracks vfp/simd hardware state with a vcpu 
> >>>> lazy flag. 
> >>>>
> >>>> vcpu lazy flag is set on guest access and trap to vfp/simd hardware switch 
> >>>> handler. On vm-enter if lazy flag is set skip trap enable and saving 
> >>>> host fpexc. On vm-exit if flag is set skip hardware context switch
> >>>> and return to host with guest context.
> >>>>
> >>>> On vcpu_put check if vcpu lazy flag is set, and execute a hardware context 
> >>>> switch to restore host.
> >>>>
> >>>> Signed-off-by: Mario Smarduch <m.smarduch at samsung.com>
> >>>> ---
> >>>>  arch/arm/include/asm/kvm_asm.h |  1 +
> >>>>  arch/arm/kvm/arm.c             | 17 ++++++++++++
> >>>>  arch/arm/kvm/interrupts.S      | 60 +++++++++++++++++++++++++++++++-----------
> >>>>  arch/arm/kvm/interrupts_head.S | 12 ++++++---
> >>>>  4 files changed, 71 insertions(+), 19 deletions(-)
> >>>>
> >>>> diff --git a/arch/arm/include/asm/kvm_asm.h b/arch/arm/include/asm/kvm_asm.h
> >>>> index 194c91b..4b45d79 100644
> >>>> --- a/arch/arm/include/asm/kvm_asm.h
> >>>> +++ b/arch/arm/include/asm/kvm_asm.h
> >>>> @@ -97,6 +97,7 @@ extern char __kvm_hyp_code_end[];
> >>>>  extern void __kvm_flush_vm_context(void);
> >>>>  extern void __kvm_tlb_flush_vmid_ipa(struct kvm *kvm, phys_addr_t ipa);
> >>>>  extern void __kvm_tlb_flush_vmid(struct kvm *kvm);
> >>>> +extern void __kvm_restore_host_vfp_state(struct kvm_vcpu *vcpu);
> >>>>  
> >>>>  extern int __kvm_vcpu_run(struct kvm_vcpu *vcpu);
> >>>>  #endif
> >>>> diff --git a/arch/arm/kvm/arm.c b/arch/arm/kvm/arm.c
> >>>> index ce404a5..79f49c7 100644
> >>>> --- a/arch/arm/kvm/arm.c
> >>>> +++ b/arch/arm/kvm/arm.c
> >>>> @@ -105,6 +105,20 @@ void kvm_arch_check_processor_compat(void *rtn)
> >>>>  	*(int *)rtn = 0;
> >>>>  }
> >>>>  
> >>>> +/**
> >>>> + * kvm_switch_fp_regs() - switch guest/host VFP/SIMD registers
> >>>> + * @vcpu:	pointer to vcpu structure.
> >>>> + *
> >>>
> >>> nit: stray blank line
> >> ok
> >>>
> >>>> + */
> >>>> +static void kvm_switch_fp_regs(struct kvm_vcpu *vcpu)
> >>>> +{
> >>>> +#ifdef CONFIG_ARM
> >>>> +	if (vcpu->arch.vfp_lazy == 1) {
> >>>> +		kvm_call_hyp(__kvm_restore_host_vfp_state, vcpu);
> >>>
> >>> why do you have to do this in HYP mode ?
> >>  Calling it directly works fine. I moved the function outside hyp start/end
> >> range in interrupts.S. Not thinking outside the box, just thought let them all
> >> be hyp calls.
> >>
> >>>
> >>>> +		vcpu->arch.vfp_lazy = 0;
> >>>> +	}
> >>>> +#endif
> >>>
> >>> we've tried to put stuff like this in header files to avoid the ifdefs
> >>> so far.  Could that be done here as well?
> >>
> >> That was a to do, but didn't get around to it.
> >>>
> >>>> +}
> >>>>  
> >>>>  /**
> >>>>   * kvm_arch_init_vm - initializes a VM data structure
> >>>> @@ -295,6 +309,9 @@ void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
> >>>>  
> >>>>  void kvm_arch_vcpu_put(struct kvm_vcpu *vcpu)
> >>>>  {
> >>>> +	/* Check if Guest accessed VFP registers */
> >>>
> >>> misleading comment: this function does more than checking
> >> Yep sure does, will change.
> >>>
> >>>> +	kvm_switch_fp_regs(vcpu);
> >>>> +
> >>>>  	/*
> >>>>  	 * The arch-generic KVM code expects the cpu field of a vcpu to be -1
> >>>>  	 * if the vcpu is no longer assigned to a cpu.  This is used for the
> >>>> diff --git a/arch/arm/kvm/interrupts.S b/arch/arm/kvm/interrupts.S
> >>>> index 900ef6d..6d98232 100644
> >>>> --- a/arch/arm/kvm/interrupts.S
> >>>> +++ b/arch/arm/kvm/interrupts.S
> >>>> @@ -96,6 +96,29 @@ ENTRY(__kvm_flush_vm_context)
> >>>>  	bx	lr
> >>>>  ENDPROC(__kvm_flush_vm_context)
> >>>>  
> >>>> +/**
> >>>> + * void __kvm_restore_host_vfp_state(struct vcpu *vcpu) - Executes a lazy
> >>>> + *     fp/simd switch, saves the guest, restores host.
> >>>> + *
> >>>
> >>> nit: stray blank line
> >> ok.
> >>>
> >>>> + */
> >>>> +ENTRY(__kvm_restore_host_vfp_state)
> >>>> +#ifdef CONFIG_VFPv3
> >>>> +	push	{r3-r7}
> >>>> +
> >>>> +	add	r7, r0, #VCPU_VFP_GUEST
> >>>> +	store_vfp_state r7
> >>>> +
> >>>> +	add	r7, r0, #VCPU_VFP_HOST
> >>>> +	ldr	r7, [r7]
> >>>> +	restore_vfp_state r7
> >>>> +
> >>>> +	ldr	r3, [vcpu, #VCPU_VFP_FPEXC]
> >>>
> >>> either use r0 or vcpu throughout this function please
> >> Yeah that's bad - in the same function to
> >>>
> >>>> +	VFPFMXR FPEXC, r3
> >>>> +
> >>>> +	pop	{r3-r7}
> >>>> +#endif
> >>>> +	bx	lr
> >>>> +ENDPROC(__kvm_restore_host_vfp_state)
> >>>>  
> >>>>  /********************************************************************
> >>>>   *  Hypervisor world-switch code
> >>>> @@ -119,11 +142,15 @@ ENTRY(__kvm_vcpu_run)
> >>>>  	@ If the host kernel has not been configured with VFPv3 support,
> >>>>  	@ then it is safer if we deny guests from using it as well.
> >>>>  #ifdef CONFIG_VFPv3
> >>>> +	@ r7 must be preserved until next vfp lazy check
> >>>
> >>> I don't understand this comment
> >>>
> >>>> +	vfp_inlazy_mode r7, skip_save_host_fpexc
> >>>> +
> >>>>  	@ Set FPEXC_EN so the guest doesn't trap floating point instructions
> >>>>  	VFPFMRX r2, FPEXC		@ VMRS
> >>>> -	push	{r2}
> >>>> +	str     r2, [vcpu, #VCPU_VFP_FPEXC]
> >>>>  	orr	r2, r2, #FPEXC_EN
> >>>>  	VFPFMXR FPEXC, r2		@ VMSR
> >>>> +skip_save_host_fpexc:
> >>>>  #endif
> >>>>  
> >>>>  	@ Configure Hyp-role
> >>>> @@ -131,7 +158,14 @@ ENTRY(__kvm_vcpu_run)
> >>>>  
> >>>>  	@ Trap coprocessor CRx accesses
> >>>>  	set_hstr vmentry
> >>>> -	set_hcptr vmentry, (HCPTR_TTA | HCPTR_TCP(10) | HCPTR_TCP(11))
> >>>> +	set_hcptr vmentry, (HCPTR_TTA)
> >>>> +
> >>>> +	@ check if vfp_lazy flag set
> >>>> +	cmp     r7, #1
> >>>
> >>> if you meant that you need to preserve r7 down to here, could you
> >>> instead just move the VFP logic above down here and do the whole VFP
> >>> logic in one coherent block?
> >>
> >> I reworked the code both fpexc save and trap enable are handled at once.
> >>>
> >>>> +	beq     skip_guest_vfp_trap
> >>>> +	set_hcptr vmentry, (HCPTR_TCP(10) | HCPTR_TCP(11))
> >>>> +skip_guest_vfp_trap:
> >>>> +
> >>>>  	set_hdcr vmentry
> >>>>  
> >>>>  	@ Write configured ID register into MIDR alias
> >>>> @@ -170,22 +204,14 @@ __kvm_vcpu_return:
> >>>>  	@ Don't trap coprocessor accesses for host kernel
> >>>>  	set_hstr vmexit
> >>>>  	set_hdcr vmexit
> >>>> -	set_hcptr vmexit, (HCPTR_TTA | HCPTR_TCP(10) | HCPTR_TCP(11)), after_vfp_restore
> >>>> +	set_hcptr vmexit, (HCPTR_TTA | HCPTR_TCP(10) | HCPTR_TCP(11))
> >>>>  
> >>>>  #ifdef CONFIG_VFPv3
> >>>> -	@ Switch VFP/NEON hardware state to the host's
> >>>> -	add	r7, vcpu, #VCPU_VFP_GUEST
> >>>> -	store_vfp_state r7
> >>>> -	add	r7, vcpu, #VCPU_VFP_HOST
> >>>> -	ldr	r7, [r7]
> >>>> -	restore_vfp_state r7
> >>>> -
> >>>> -after_vfp_restore:
> >>>> -	@ Restore FPEXC_EN which we clobbered on entry
> >>>> -	pop	{r2}
> >>>> +	vfp_inlazy_mode r2, skip_restore_host_fpexc
> >>>> +	@ If vfp_lazy is not set, restore FPEXC_EN which we clobbered on entry
> >>>> +	ldr     r2, [vcpu, #VCPU_VFP_FPEXC]
> >>>
> >>> so we do this restore if, since we scheduled this VCPU thread, the guest
> >>> has not touched any VFP regs, is that correct?
> >> That's right.
> >>>
> >>> Did you measure how often we schedule the guest and run it until we
> >>> schedule another process without the guest touching any VFP regs?  I'm
> >>> just wondering if this complexity is worth it, or if we should just
> >>> switch the VFP regs on vcpu_load/vcpu_put instead?
> >>
> >> The loads I've been running mix of fp operations and lmbench mmu - shows huge
> >> decrease of fp save/restore like from ~30-50%, down to ~2%. What I did is
> >> measured all exits and fp/save restore for both scenarios. So yes it does make a
> >> difference. Of course will depend on the load, but should be never be worse then
> >> now.
> > 
> > True, and with the renaming the complexity shouldn't be that bad.
> > 
> >>>
> >>> Also, what do other architectures do here?
> >>
> >> x86 does a similar thing in it's kvm_arch_vcpu_put().
> >>
> > 
> > ok.
> > 
> >>>
> >>>>  	VFPFMXR FPEXC, r2
> >>>> -#else
> >>>> -after_vfp_restore:
> >>>> +skip_restore_host_fpexc:
> >>>>  #endif
> >>>>  
> >>>>  	@ Reset Hyp-role
> >>>> @@ -485,6 +511,10 @@ switch_to_guest_vfp:
> >>>>  	@ NEON/VFP used.  Turn on VFP access.
> >>>>  	set_hcptr vmtrap, (HCPTR_TCP(10) | HCPTR_TCP(11))
> >>>>  
> >>>> +	@ set lazy mode flag, switch hardware context on vcpu_put
> >>>> +	mov     r1, #1
> >>>> +	str     r1, [vcpu, #VCPU_VFP_LAZY]
> >>>> +
> >>>>  	@ Switch VFP/NEON hardware state to the guest's
> >>>>  	add	r7, r0, #VCPU_VFP_HOST
> >>>>  	ldr	r7, [r7]
> >>>> diff --git a/arch/arm/kvm/interrupts_head.S b/arch/arm/kvm/interrupts_head.S
> >>>> index 702740d..4561171 100644
> >>>> --- a/arch/arm/kvm/interrupts_head.S
> >>>> +++ b/arch/arm/kvm/interrupts_head.S
> >>>> @@ -594,7 +594,7 @@ ARM_BE8(rev	r6, r6  )
> >>>>   * If a label is specified with vmexit, it is branched to if VFP wasn't
> >>>>   * enabled.
> >>>>   */
> >>>> -.macro set_hcptr operation, mask, label = none
> >>>> +.macro set_hcptr operation, mask
> >>>>  	mrc	p15, 4, r2, c1, c1, 2
> >>>>  	ldr	r3, =\mask
> >>>>  	.if \operation == vmentry
> >>>> @@ -609,13 +609,17 @@ ARM_BE8(rev	r6, r6  )
> >>>>  	beq	1f
> >>>>  	.endif
> >>>>  	isb
> >>>> -	.if \label != none
> >>>> -	b	\label
> >>>> -	.endif
> >>>>  1:
> >>>>  	.endif
> >>>>  .endm
> >>>>  
> >>>> +/* Checks if VFP/SIMD lazy flag is set, if it is branch to label. */
> >>>
> >>> I don't easily understand the semantics of the lazy flag.  When set,
> >>> does it mean we've switched the hardware to the guest state?
> >>>
> > 
> > The conclusion here is probably that the lazy flag should instead be
> > called the dirty flag or something where a value of true has some more
> > intuitive meaning.
> > 
> > Thanks,
> > -Christoffer
> 
> So to summarize arm patches will be reworked to include your latest comments.
> arm64 will directly call the host el1 function in vcpu_put. And a retest of both.
> 
> Any cutoff dates in mind?
> 
We're getting close to v4.4, but I'll try to have a review of your arm64
patches soon and if they're similarly simple and I have time to test
them thoroughly, I may consider adding them given the immediate
performance benefit.

-Christoffer



More information about the linux-arm-kernel mailing list