On Wed, Oct 07, 2015 at 11:39:30AM +0300, Pavel Fedin wrote:
> Hello!
>
> > +/* Called with the distributor lock held by the caller. */
> > +void vits_unqueue_lpi(struct kvm_vcpu *vcpu, int lpi)
> > +{
> > + struct vgic_its *its = &vcpu->kvm->arch.vgic.its;
> > + struct its_itte *itte;
> > +
> > + spin_lock(&its->lock);
> > +
> > + /* Find the right ITTE and put the pending state back in there */
> > + itte = find_itte_by_lpi(vcpu->kvm, lpi);
> > + if (itte)
> > + __set_bit(vcpu->vcpu_id, itte->pending);
> > +
> > + spin_unlock(&its->lock);
> > +}
>
> I am working on implementing live migration for the ITS. And here i have one fundamental problem.
> vits_unqueue_lpi() processes only PENDING state. And looks like this corresponds to the HW
> implementation, which has only bitwise pending table. But, in terms of migration, we can actually
> have LPI in active state, while it's being processed.
I thought LPIs had strict fire-and-forget semantics, not allowing any
active state, and that they are either pending or inactive?
> The question is - how can we handle it? Should we have one more bitwise table for active LPIs, or
> is it enough to remember only a single, currently active LPI? Can LPIs be preempted on a real
> hardware, or not?
>
Perhaps you're asking if LPIs have active state semantics on real
hardware and thus supports threaded interrupt handling for LPIs?
That is not supported on real hardware, which I think addresses your
concerns.
Thanks,
-Christoffer
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
