Re: [PATCH v2 4/8] KVM: VMX: Optimize vmx_set_rflags() for unrestricted guest

From: Sean Christopherson <sean.j.christopherson@intel.com>
To: Vitaly Kuznetsov <vkuznets@redhat.com>
Cc: "Paolo Bonzini" <pbonzini@redhat.com>,
	"Radim Krčmář" <rkrcmar@redhat.com>,
	"Wanpeng Li" <wanpengli@tencent.com>,
	"Jim Mattson" <jmattson@google.com>,
	"Joerg Roedel" <joro@8bytes.org>,
	kvm@vger.kernel.org, linux-kernel@vger.kernel.org,
	"Reto Buerki" <reet@codelabs.ch>,
	"Liran Alon" <liran.alon@oracle.com>
Subject: Re: [PATCH v2 4/8] KVM: VMX: Optimize vmx_set_rflags() for unrestricted guest
Date: Mon, 30 Sep 2019 08:19:45 -0700	[thread overview]
Message-ID: <20190930151945.GB14693@linux.intel.com> (raw)
In-Reply-To: <87muem40wi.fsf@vitty.brq.redhat.com>

On Mon, Sep 30, 2019 at 10:57:17AM +0200, Vitaly Kuznetsov wrote:
> Sean Christopherson <sean.j.christopherson@intel.com> writes:
> 
> > Rework vmx_set_rflags() to avoid the extra code need to handle emulation
> > of real mode and invalid state when unrestricted guest is disabled.  The
> > primary reason for doing so is to avoid the call to vmx_get_rflags(),
> > which will incur a VMREAD when RFLAGS is not already available.  When
> > running nested VMs, the majority of calls to vmx_set_rflags() will occur
> > without an associated vmx_get_rflags(), i.e. when stuffing GUEST_RFLAGS
> > during transitions between vmcs01 and vmcs02.
> >
> > Note, vmx_get_rflags() guarantees RFLAGS is marked available.
> >
> > Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
> > ---
> >  arch/x86/kvm/vmx/vmx.c | 28 ++++++++++++++++++----------
> >  1 file changed, 18 insertions(+), 10 deletions(-)
> >
> > diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
> > index 83fe8b02b732..814d3e6d0264 100644
> > --- a/arch/x86/kvm/vmx/vmx.c
> > +++ b/arch/x86/kvm/vmx/vmx.c
> > @@ -1426,18 +1426,26 @@ unsigned long vmx_get_rflags(struct kvm_vcpu *vcpu)
> >  void vmx_set_rflags(struct kvm_vcpu *vcpu, unsigned long rflags)
> >  {
> >  	struct vcpu_vmx *vmx = to_vmx(vcpu);
> > -	unsigned long old_rflags = vmx_get_rflags(vcpu);
> > +	unsigned long old_rflags;
> >  
> > -	__set_bit(VCPU_EXREG_RFLAGS, (ulong *)&vcpu->arch.regs_avail);
> > -	vmx->rflags = rflags;
> > -	if (vmx->rmode.vm86_active) {
> > -		vmx->rmode.save_rflags = rflags;
> > -		rflags |= X86_EFLAGS_IOPL | X86_EFLAGS_VM;
> > +	if (enable_unrestricted_guest) {
> > +		__set_bit(VCPU_EXREG_RFLAGS, (ulong *)&vcpu->arch.regs_avail);
> > +
> > +		vmx->rflags = rflags;
> > +		vmcs_writel(GUEST_RFLAGS, rflags);
> > +	} else {
> > +		old_rflags = vmx_get_rflags(vcpu);
> > +
> > +		vmx->rflags = rflags;
> > +		if (vmx->rmode.vm86_active) {
> > +			vmx->rmode.save_rflags = rflags;
> > +			rflags |= X86_EFLAGS_IOPL | X86_EFLAGS_VM;
> > +		}
> > +		vmcs_writel(GUEST_RFLAGS, rflags);
> > +
> > +		if ((old_rflags ^ vmx->rflags) & X86_EFLAGS_VM)
> > +			vmx->emulation_required = emulation_required(vcpu);
> >  	}
> > -	vmcs_writel(GUEST_RFLAGS, rflags);
> 
> We're doing vmcs_writel() in both branches so it could've stayed here, right?

Yes, but the resulting code is a bit ugly.  emulation_required() consumes
vmcs.GUEST_RFLAGS, i.e. the if statement that reads old_rflags would also
need to be outside of the else{} case.  

This isn't too bad:

	if (!enable_unrestricted_guest && 
	    ((old_rflags ^ vmx->rflags) & X86_EFLAGS_VM))
		vmx->emulation_required = emulation_required(vcpu);

but gcc isn't smart enough to understand old_rflags won't be used if
enable_unrestricted_guest, so old_rflags either needs to be tagged with
uninitialized_var() or explicitly initialized in the if(){} case.

Duplicating a small amount of code felt like the lesser of two evils.

> > -
> > -	if ((old_rflags ^ vmx->rflags) & X86_EFLAGS_VM)
> > -		vmx->emulation_required = emulation_required(vcpu);
> >  }
> >  
> >  u32 vmx_get_interrupt_shadow(struct kvm_vcpu *vcpu)
> 
> Reviewed-by: Vitaly Kuznetsov <vkuznets@redhat.com>
> 
> -- 
> Vitaly