Re: [PATCH] KVM: x86: Fix the intel_pt PMI handling wrongly considered from guest

[Sean Christopherson]

Please don't top-post.

On Fri, May 20, 2022, Xu, Yanfei wrote:
> From: Sean Christopherson <seanjc@xxxxxxxxxx> 
> On Mon, May 16, 2022, Yanfei Xu wrote:
> > diff --git a/arch/x86/include/asm/kvm_host.h 
> > b/arch/x86/include/asm/kvm_host.h index 4ff36610af6a..308cf19f123d 
> > 100644
> > --- a/arch/x86/include/asm/kvm_host.h
> > +++ b/arch/x86/include/asm/kvm_host.h
> > @@ -1582,8 +1582,14 @@ static inline int kvm_arch_flush_remote_tlb(struct kvm *kvm)
> >  		return -ENOTSUPP;
> >  }
> >  
> > +enum kvm_intr_type {
> > +	/* Values are arbitrary, but must be non-zero. */
> > +	KVM_HANDLING_IRQ = 1,
> > +	KVM_HANDLING_NMI,
> > +};
> > +
> >  #define kvm_arch_pmi_in_guest(vcpu) \
> > -	((vcpu) && (vcpu)->arch.handling_intr_from_guest)
> > +	((vcpu) && (vcpu)->arch.handling_intr_from_guest == 
> > +KVM_HANDLING_NMI)
> 
> My understanding is that this isn't correct as a general change, as perf
> events can use regular IRQs in some cases.  See commit dd60d217062f4 ("KVM:
> x86: Fix perf timer mode IP reporting").
> 
> I assume there's got to be a way to know which mode perf is using, e.g. we
> should be able to make this look something like:
> 
> 	((vcpu) && (vcpu)->arch.handling_intr_from_guest == kvm_pmi_vector)

> Hi Sean,
> You are right, the change of kvm_arch_pmi_in_guest is incorrect, because it should cover two cases of PMI. 
> For the PMI of intel pt, it certainly is the NMI PMI. So how about fixing it like below?

Yep, that works.  I did enough spelunking to figure out how we can fix the generic
issue, but it's per-event and requires a decent amount of plumbing in perf.

perf_guest_handle_intel_pt_intr() doesn't bother with perf_guest_state() since it's
such a specialized event, so fixing this in vmx_handle_intel_pt_intr() would likely
be the long-term solution even if/when the generic case is fixed.