Re: [XEN PATCH] x86/ACPI: Ignore entries with invalid APIC IDs when parsing MADT

[George Dunlap <george.dunlap@xxxxxxxxx>]

On Thu, Sep 14, 2023 at 12:18 AM Stefano Stabellini
<sstabellini@xxxxxxxxxx> wrote:
>
> On Wed, 13 Sep 2023, George Dunlap wrote:
> > On Tue, Sep 12, 2023 at 8:57 AM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
> > >
> > > On Mon, Sep 11 2023 at 19:24, Andrew Cooper wrote:
> > > > Furthermore, cursory testing that Thomas did for the Linux topology work
> > > > demonstrates that it is broken anyway for reasons unrelated to ACPI 
> > > > parsing.
> > > >
> > > > Even furthermore, it's an area of the Xen / dom0 boundary which is
> > > > fundamentally broken for non-PV cases, and undocumented for the PV case,
> > > > hence why it's broken in Linux.
> > > >
> > > > Physical CPU Hotplug does not pass the bar for being anything more than
> > > > experimental.  It's absolutely not tech-preview level because the only
> > > > demo it has had in an environment (admittedly virtual) which does
> > > > implement the spec in a usable way demonstrates that it doesn't 
> > > > function.
> > > >
> > > > The fact no-one has noticed until now shows that the feature isn't used,
> > > > which comes back around full circle to the fact that Intel never made it
> > > > work and never shipped it.
> > >
> > > OTOH it _is_ used in virtualization. KVM supports it and it just
> > > works. That's how I found out that XEN explodes in colourful ways :)
> >
> > It should be pointed out that there's currently a start-up selling a
> > product that specifically runs Xen under cloud providers -- Exostellar
> > (was Exotanium) [1].  If cloud providers do use ACPI hotplug to allow
> > on-the-fly adjustments of the number of vcpus, Exostellar will
> > probably want support at some point.  (Perhaps it would be good to
> > rope them into testing / maintaining it.)
>
> Supporting CPU hotplug in a nested virtualization setting is a different
> proposition compared to supporting Physical CPU Hotplug. Typically
> virtual firmware (hypervisor-provided firmware) has less unexpected
> behaviors compared to baremetal firmware.
>
> Could you make the distinction in SUPPORT.md?

People say at the moment it doesn't actually work, even under QEMU; so
"Experimental" is probably the right status.

But if someone got it working, we might add "supported, with caveats"
and specify things in a comment.

 -George