Re: [PATCH v2] KVM: MMU: Don't use RCU for lockless shadow walking

[Avi Kivity <avi@redhat.com>]

On 04/28/2012 12:49 AM, Marcelo Tosatti wrote:
> > >> +	smp_mb();
> > >> +	vcpu->mode = OUTSIDE_GUEST_MODE;
> > > 
> > > Don't you want 
> > > 
> > > vcpu->mode = OUTSIDE_GUEST_MODE;
> > > smp_mb();
> > > 
> > 
> > 
> > It is unsafe i think, it is a problem if spte read / spte update is ordered
> > to the behind of vcpu->mode = OUTSIDE_GUEST_MODE, like below:
> > 
> > VCPU 0                                  VCPU 1
> >                                    commit_zapped_page:
> > /*
> >  * setting vcpu->mode is reordered
> >  * to the head of read spte.
> >  */
> > vcpu->mode = OUTSIDE_GUEST_MODE;
> > 
> >                                    see VCPU 0 is out-of-guest-mode, IPI is
> >                                    not sent, and the sp is free immediately.
> > 
> > read spte;
> >  OOPS!!!

Ouch!

> > 
> > (It is invalid since spte is freed.)
> > 
> > smp_mb
>
> Right. In that case a compiler barrier is sufficient (stores are not 
> reordered with earlier loads on x86).

I think it's customary not to depend on arch specific memory barrier
behaviour (but I don't have anything to base this on, just a feeling).

-- 
error compiling committee.c: too many arguments to function