Re: [PATCH v5] kvm: make vcpu life cycle separated from kvm instance

[Gleb Natapov <gleb@redhat.com>]

On Mon, Dec 26, 2011 at 01:17:39PM +0200, Avi Kivity wrote:
> On 12/26/2011 01:09 PM, Gleb Natapov wrote:
> > > +
> > > +		idx = srcu_read_lock(&kvm->srcu_vcpus);
> > > +		kvm_for_each_vcpu(vcpu, kvm) {
> > > +			if (!pass && !firststart &&
> > > +			    vcpu != kvm->last_boosted_vcpu &&
> > > +			    kvm->last_boosted_vcpu != NULL) {
> > > +				vcpu = kvm->last_boosted_vcpu;
> > You access last_boosted_vcpu as if it is protected by srcu, but it
> > isn't. kvm_vcpu_release() changes it after synchronize_srcu_expedited()
> > call.
> >
> > I do not like this last_boosted_vcpu pointer much. May be we can rid of
> > it by remembering last apic_id and searching for it each time we enter
> > the function. I do not think this function is to performance sensitive.
> > We enter here when vcpu is spinning anyway.
> 
> We aren't guaranteed to have an apic_id, so it has to be done using rcu,
> or maybe vcpu_id.  I prefer using srcu, we can't run away from vcpu
> pointers.
> 
Yeah, I meant vcpu_id (it is used as initial apic_id for x86, but this
code is not x86 specific).

--
			Gleb.