Re: [PATCH RFC V6 1/5] kvm hypervisor : Add a hypercall to KVM hypervisor to support pv-ticketlocks

[Gleb Natapov <gleb@redhat.com>]

On Mon, Apr 30, 2012 at 11:22:34AM +0300, Avi Kivity wrote:
> On 04/29/2012 04:52 PM, Gleb Natapov wrote:
> > On Sun, Apr 29, 2012 at 04:26:21PM +0300, Avi Kivity wrote:
> > > On 04/29/2012 04:20 PM, Gleb Natapov wrote:
> > > > > > This is too similar to kvm_irq_delivery_to_apic(). Why not reuse it. We
> > > > > > can use one of reserved delivery modes as PV delivery mode. We will
> > > > > > disallow guest to trigger it through apic interface, so this will not be
> > > > > > part of ABI and can be changed at will.
> > > > > >
> > > > > 
> > > > > I'm not thrilled about this.  Those delivery modes will eventually
> > > > > become unreserved.  We can have a kvm_lookup_apic_id() that is shared
> > > > > among implementations.
> > > > > 
> > > > This is only internal implementation. If they become unreserved we will
> > > > use something else.
> > > >
> > > 
> > > Yeah, I'm thinking of that time.  Why do something temporary and fragile?
> > > 
> > Why is it fragile? Just by unreserving the value Intel will not break
> > KVM. Only when KVM will implement apic feature that unreserves the value
> > we will have to change internal implementation and use another value,
> > but this will be done by the same patch that does unreserving. The
> > unreserving may even never happen. 
> 
> Some remains of that may leak somewhere.
I do not see where. APIC code should #GP if a guest attempts to set
reserved values through APIC interface, or at least ignore them.

>                                           Why not add an extra
> parameter?
Yes, we can add extra parameter to "struct kvm_lapic_irq" and propagate it to
__apic_accept_irq(). We can do that now, or when Intel unreserve all
reserved values. I prefer the later since I do not believe it will
happen in observable feature.

>              Or do something like
> 
> kvm_for_each_apic_dest(vcpu, apic_destination) {
>       ...
> }
> 
> That can be reused in both the apic code and pv kick.
> 
That's exactly what kvm_irq_delivery_to_apic() is.

> > Meanwhile kvm_irq_delivery_to_apic()
> > will likely be optimized to use hash for unicast delivery and unhalt
> > hypercall will benefit from it immediately.
> 
> Overloading delivery mode is not the only way to achieve sharing.
> 
It is simplest and most straightforward with no demonstratable drawbacks :)
Adding parameter to "struct kvm_lapic_irq" is next best thing.

--
			Gleb.