Re: [PATCH 1/5] kvm/x86: fix inversed check for async_pf MSR

[Roman Kagan <rkagan@virtuozzo.com>]

On Fri, Dec 02, 2016 at 05:24:55PM +0200, Mika Penttilä wrote:
> > > > > > --- a/arch/x86/kvm/x86.c
> > > > > > +++ b/arch/x86/kvm/x86.c
> > > > > > @@ -8402,7 +8402,7 @@ void kvm_arch_async_page_present(struct kvm_vcpu *vcpu,
> > > > > > bool kvm_arch_can_inject_async_page_present(struct kvm_vcpu *vcpu)
> > > > > > {
> > > > > > 	if (!(vcpu->arch.apf.msr_val & KVM_ASYNC_PF_ENABLED))
> > > > > > -		return true;
> > > > > > +		return false;
> > > > > > 
> > > > > > 
> > > > > Why do you make this change?
> > > 
> > > > Because the code does the opposite of what it's meant to do.
> > > 
> > > It could have a better name but returning "true" is right. See below.
> > > 
> > >  > > I think kvm_arch_async_page_present() is not
> > > > > ever called now and neither kvm_del_async_pf_gfn(vcpu, work->arch.gfn);
> > > > I wonder how you came to such a conclusion?  I certainly see them called
> > > > on my test machine (you need to have the guest memory swapped out for
> > > > that, that can be forced e.g. using a memory cgroup).
> > > if !KVM_ASYNC_PF_ENABLED then kvm_check_async_pf_completion(), it's only call site, never calls it.
> 
> > How's that?  I don't see any check for it in
> > kvm_check_async_pf_completion().  Moreover, that's exactly where it does
> > that check.
> 
> void kvm_check_async_pf_completion(struct kvm_vcpu *vcpu)
> {
>         struct kvm_async_pf *work;
> 
>         while (!list_empty_careful(&vcpu->async_pf.done) &&
>               kvm_arch_can_inject_async_page_present(vcpu)) {
>                 spin_lock(&vcpu->async_pf.lock);
>   
> 
> and you made kvm_arch_can_inject_async_page_present(vcpu) return false if !KVM_ASYNC_PF_ENABLED (i.e. not enabled)
> 
> 
> > > Maybe you had KVM_ASYNC_PF_ENABLED?
> 
> > Of course I did.  Not sure I get what you mean...
> 
> I mean if the guest support for async pf is not enabled you maybe broke that?

Ah I finally see the light, thank you.

Apparently disabling async_pf by the guest should result in draining all
accumulated async pagefaults but without injecting #PF in the guest.
The latter is taken care of by another check for the msr value in
kvm_arch_async_page_present().

My patch is wrong indeed; I was misled by the "self-explanatory"
function name which was bogus.

I'll cook up another patch renaming the function and putting a comment
there, too.

Thanks,
Roman.