Re: [PATCH RFC] KVM: Fix race in apic->pending_events processing

[Gleb Natapov <gleb@redhat.com>]

On Thu, May 30, 2013 at 02:58:09PM +0200, Paolo Bonzini wrote:
> Il 30/05/2013 14:34, Gleb Natapov ha scritto:
> >>> > >
> >>> > > Ah, we check kvm_apic_has_events() in runnable. Then yes, we will not
> >>> > > lose the event.
> >> > 
> >> > Ok, then I'd prefer to have the cmpxchg directly in the if, as in
> >> > http://article.gmane.org/gmane.comp.emulators.kvm.devel/110505
> >> > 
> > I still do not. Both of them are tricky, mine does not coalesce events
> > needlessly.
> 
> Agreed that both are tricky, but I don't think my patch is coalescing
> events.  If you have
> 
>     INIT    SIPI     INIT     SIPI
>                   ^                           ^
>                   INIT bit cleared here       SIPI bit checked here
> 
Not sure I understand what you are trying to say here.

> my patch KVM sees apic_events = INIT | SIPI and deduces that the SIPI
> bit was set by the second SIPI, not by the first.  In fact the first
> SIPI was cancelled by the second INIT, and thus should not be processed
> at all.
That is called coalesced.

> 
> Instead, with your patch KVM will service all four events; strictly
> speaking it is wrong to service the first SIPI, which is why I prefer
> having the cmpxchg in the beginning.
> y
Why is it wrong?

I do not see what are you arguing about.

--
			Gleb.