Re: [PATCH] x86/cpuid: Deal with broken firmware once more

[Boris Ostrovsky]

On 11/10/2016 12:13 PM, Thomas Gleixner wrote:
> On Thu, 10 Nov 2016, Boris Ostrovsky wrote:
>> On 11/10/2016 10:12 AM, Thomas Gleixner wrote:
>>> On Thu, 10 Nov 2016, Boris Ostrovsky wrote:
>>>> By firmware you mean ACPI? It is most likely not available to PV guests.
>>> You either have to provide ACPI or MP tables. And either of those has to
>>> provide the intial APIC ids for the CPUs. They are supposed to match the
>>> IDs which are in the CPUID leafs.
>>>
>>>> How about returning cpu_data(cpu).initial_apicid?
>>> For a not yet brought up CPU. That's what the initial ACPI/MP table
>>> enumeration is for.
>> Unfortunately PV guests have neither. So we may need to emulate
>> something in xen_cpu_present_to_apicid().
> SFI does the same thing and according to the dmesg which was posted, this
> is using SFI. We also have devicetree based boot concept which provides the
> APICids in the CPU enumeration at boot time in a way which the whole x86
> machinery is expecting.
>
> So what kind of APICid is XEN handing in via SFI? None, or just an invalid
> one?

None, SFI is not used at all. Or device tree for that matter.

The (PV) guest currently assumes APIC ID (i.e. APIC[0x20]) to always be
zero (we just fixed a bug where CPUID may return an
inconsistent/non-zero initial APIC ID value, but that's a different
issue I think). There is no topology, really, everything is flat.

So this obviously is rather, ahem, fragile and we've been fixing bugs
there, especially with introduction of package map (a6a198bc60e6 that I
mentioned in the other message is one such example). There is work
scheduled on the hypervisor side to properly report initial APIC IDs in
CPUID and that may help with correctly dealing with this on Linux side.

(Longer term we would like to move to PVH where we will have "true" APIC
emulation, with ACPI, just like in HVM guests. You can do your part by
helping: http://marc.info/?l=linux-kernel&m=147791731414152&w=3  ;-))


-boris