On Tue, Oct 15, 2019 at 10:22:18AM +0100, Dr. David Alan Gilbert wrote:
> * Peter Xu (pet...@redhat.com) wrote:
> > Migration is silently broken now with x2apic config like this:
> >
> > -smp 200,maxcpus=288,sockets=2,cores=72,threads=2 \
> > -device intel-iommu,intremap=on,eim=on
> >
> > After migration, the guest kernel could hang at anything, due to
> > x2apic bit not migrated correctly in IA32_APIC_BASE on some vcpus, so
> > any operations related to x2apic could be broken then (e.g., RDMSR on
> > x2apic MSRs could fail because KVM would think that the vcpu hasn't
> > enabled x2apic at all).
> >
> > The issue is that the x2apic bit was never applied correctly for vcpus
> > whose ID > 255 when migrate completes, and that's because when we
> > migrate APIC we use the APICCommonState.id as instance ID of the
> > migration stream, while that's too short for x2apic.
> >
> > Let's use the newly introduced initial_apic_id for that.
>
> I'd like to understand a few things:
> a) Does this change the instance ID of existing APICs on the
> migration stream?
> a1) Ever for <256 CPUs?
No.
> a2) For >=256 CPUs?
Yes.
>
> [Because changing the ID breaks migration]
But if we don't change it, the stream is broken too. :)
Then the destination VM will receive e.g. two apic_id==0 instances (I
think the apic_id==256 instance will wrongly overwrite the apic_id==0
one), while the vcpu with apic_id==256 will use the initial apic
values.
So IMHO we should still fix this, even if it changes the migration
stream. At least we start to make it right.
>
> b) Is the instance ID constant - I can see it's a property on the
> APIC, but I cna't see who sets it
For each vcpu, I think yes it should be a constant as long as the
topology is the same. This is how I understand it to be set:
(1) In pc_cpus_init(), we init these:
possible_cpus = mc->possible_cpu_arch_ids(ms);
for (i = 0; i < ms->smp.cpus; i++) {
pc_new_cpu(pcms, possible_cpus->cpus[i].arch_id, &error_fatal);
}
(2) In x86_cpu_apic_create(), we apply the apic_id to "id" property:
qdev_prop_set_uint32(cpu->apic_state, "id", cpu->apic_id);
>
> c) In the case where it fails, did we end up registering two
> devices with the same name and instance ID? If so, is it worth
> adding a check that would error if we tried?
Sounds doable.
Thanks,
--
Peter Xu
