On Tue, 2011-02-01 at 19:09 +0200, Avi Kivity wrote:
> On 02/01/2011 05:48 PM, Glauber Costa wrote:
> > > > @@ -2106,6 +2120,25 @@ void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu,
> > > int cpu)
> > > > kvm_migrate_timers(vcpu);
> > > > vcpu->cpu = cpu;
> > > > }
> > > > +
> > > > + if (vcpu->arch.this_time_out) {
> > > > + u64 to = (get_kernel_ns() - vcpu->arch.this_time_out);
> > > > + /*
> > > > + * using nanoseconds introduces noise, which
> > > accumulates easily
> > > > + * leading to big steal time values. We want, however,
> > > to keep the
> > > > + * interface nanosecond-based for future-proofness.
> > > > + */
> > > > + to /= NSEC_PER_USEC;
> > > > + to *= NSEC_PER_USEC;
> > >
> > > Seems there is a real problem and that this is just papering it over.
> > > I'd like to understand the root cause.
> > Okay, my self-explanation seemed reasonable to me, but since both you
> > and Peter dislike it, I think it is important enough to get a more
> > thorough investigation before a second round.
>
> Yes please.
>
> > But in this case,
> > I keep that using nanoseconds may then not be the best approach here. We
> > also have to keep in mind that the host and guest clocks may be running
> > at different resolutions.
>
> We need to choose a resolution for the clock (or negotiate one), an
> nanoseconds seems as good as any from a range and precision
> considerations, and is convenient for the host and Linux guests. So why
> not pick it?
>
> > > > + vcpu->arch.sversion += 2;
> > >
> > > Doesn't survive live migration. You need to use the version from the
> > > guest area.
> > Why not? Who said versions need to always increase? If current version
> > is 102324, and we live migrate and it becomes 0, what is the problem?
>
> Guest reads version (result: 2)
> Guest starts reading data
> Live migration; vcpu->arch.sversion is zeroed
> Steal time update; vcpu->arch.sversion += 2; write to guest
> Guest continues reading data
> Guest reads version (result: 2)
>
> So the guest is unaware that an update has occurred while it was reading
> the data.
Ok, fair.
By the way, kvmclock have the same problem, then
>
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
