Re: [PATCHv4 3/5] kvm: host side for eoi optimization

[Marcelo Tosatti <mtosatti@redhat.com>]

On Wed, May 16, 2012 at 07:22:47PM +0300, Michael S. Tsirkin wrote:
> On Wed, May 16, 2012 at 12:49:40PM -0300, Marcelo Tosatti wrote:
> > > @@ -1245,9 +1306,20 @@ int kvm_get_apic_interrupt(struct kvm_vcpu *vcpu)
> > >  	int vector = kvm_apic_has_interrupt(vcpu);
> > >  	struct kvm_lapic *apic = vcpu->arch.apic;
> > >  
> > > -	if (vector == -1)
> > > +	/* Detect interrupt nesting and disable EOI optimization */
> > > +	if (pv_eoi_enabled(vcpu) && vector == -2)
> > > +		pv_eoi_clr_pending(vcpu);
> > > +
> > > +	if (vector < 0)
> > 
> > With interrupt window exiting, the guest will exit:
> > 
> > - as soon as it sets RFLAGS.IF=1 and there is any 
> > interrupt pending in IRR.
> > - any new interrupt is set in IRR will kick vcpu
> > out of guest mode and recalculate interrupt-window-exiting.
> > 
> > Doesnt this make this bit unnecessary ?
> 
> Looks like we could cut it out.  But I'm not sure how architectural it is
> that we exit on interrupt window.
> I guess there are reasons to exit on interrupt window but
> isn't it better to make the feature independent of it?

Hum... not sure. Is it helpful for the Hyper-V interface?

> This almost never happens in my testing anyway, so
> however we handle it is unlikely to affect performance.

It decreases the amount of state that must be maintained.

BTW there is a bug covered by interrupt window exiting:

vcpu0                               host
- irr 5 set
- isr 5 set, irr 5 cleared
- eoi_skip bit not set, 
no other bit set in irr.
- enter guest

                                    irr 4 set
                                    kick vcpu0 out of guest mode

- eoi pending bit not set
  (previous interrupt injection 
   still pending)
- skip eoi

If it were not for interrupt window exiting, this would 
inject vector 4 on an unrelated exit who knows how long 
in the future.

Also note optimization depends on the fact that the host 
kicks vcpu out unconditionally (so it is dependent on 
certain kvm implementation details).