Re: [PATCHv2 2/9] KVM: Expose a version 2 architectural PMU to a guests

[Gleb Natapov <gleb@redhat.com>]

On Mon, Nov 07, 2011 at 05:13:17PM +0200, Avi Kivity wrote:
> On 11/07/2011 04:59 PM, Peter Zijlstra wrote:
> > > > > +                */
> > > > > +               if (!kvm_is_in_guest())
> > > > > +                       irq_work_queue(&pmc->vcpu->arch.pmu.irq_work);
> > > > > +               else
> > > > > +                       kvm_make_request(KVM_REQ_PMI, pmc->vcpu);
> > > > > +       }
> > > > > +} 
> > > >
> > > > I'm not sure I get this, since the counters are vcpu task bound and only
> > > > count while the guest is running as per (144d31e6f19) how can we get an
> > > > NMI outside of guest context with the vcpu not halted?
> > > 
> > > PMI skew.  Do we know how bad it can get?
> >
> > You're talking about the PMI getting asserted before leaving guest mode
> > but not getting ran until after we left guest mode?
> 
> Yes (here, "guest mode" is not the hardware guest mode, but what
> kvm_is_in_guest() returns).
> 
> > But in that case we know the guest is stopped, right? So we can simply
> > set that REQ_PMI bit and have guest entry sort things, no?
> 
> Unless there is no guest entry, due to the guest sleeping.
> 
> Consider the sequence:
> 
> guest:
>    nop
> hardware:
>    counter overflow; PMI sent to APIC
> guest:
>    hlt
> hardware:
>    begin vmexit due to HLT
> host:
>    process vmexit
>    mark as outside guest mode (i.e. kvm_is_in_guest() now returns false)
>    schedule(), since the guest is sleeping (and assume no interrupts
> pending)
> hardware:
>    APIC finally posts NMI
> host:
>    process NMI; set KVM_REQ_PMI
>    iret
> host:
>    happily do other things, ignoring that we need to post a PMI to the guest
> 
> note, this needs a fairly huge PMI skew to happen.
> 
No, it need not. It is enough to get exit reason as hlt instead of nmi
for a vcpu to go to blocking state instead of reentering guest mode.
Note that we do not check request flags in kvm_arch_vcpu_runnable().

> (btw, like userspace, kvm PMIs may as well be ordinary interrupts. 
> Should we consider some logic to only make them NMIs if some counter
> requires it?  NMI costs are set to increase dramatically with the
> check-all-possible-sources patchset).
> 
> -- 
> error compiling committee.c: too many arguments to function

--
			Gleb.