Re: [Xen-devel] [PATCH v2 16/25] x86/pv: Use per-domain policy information in pv_cpuid()

[Boris Ostrovsky <boris.ostrovsky@xxxxxxxxxx>]

On 01/12/2017 03:51 PM, Boris Ostrovsky wrote:
> On 01/12/2017 03:48 PM, Andrew Cooper wrote:
>> On 12/01/17 20:46, Boris Ostrovsky wrote:
>>> On 01/12/2017 02:27 PM, Andrew Cooper wrote:
>>>> On 12/01/17 18:00, Boris Ostrovsky wrote:
>>>>>> Ahh! found it.  This is a side effect of starting to generate the dom0
>>>>>> policy in Xen.
>>>>>>
>>>>>> Can you try this patch?
>>>>> Intel/AMD HVM/PV 64/32bit all look good. So
>>>>>
>>>>> Tested-by: Boris Ostrovsky <boris.ostrovsky@xxxxxxxxxx>
>>>> Does this mean that newer versions of Linux more picky about what they
>>>> tolerate in cpuid?
>>> We started to fail after change in Xen so I am not sure it's something
>>> new in Linux.
>> Right, but Linux 4.4 was entirely happy with this bug, both with and
>> without having CPUID faulting imposed on it.
> Oh, I see. My tests (typically) build and run the latest Linux tree (and
> Xen staging) every morning.
>
> I am trying to see what part of Linux caused the crash.


So the problem starts in Linux ht_detect(), where we check
X86_FEATURE_CMP_LEGACY. On Intel this is supposed to be clear and we
should end up setting phys_proc_id below. This value is then used in
topology_update_package_map(). If the value is incorrect (which it will
be if we bail early in ht_detect()) we may get a BUG_ON() at the caller.
Unfortunately we were too early to see the splat from the BUG_ON so it
wasn't clear right away why we were dying.

On AMD phys_proc_id is set elsewhere.

And the reason you haven't seen problems with earlier versions of Linux
is because the last two or so kernel releases saw major changes in
topology discovery (and, more importantly, topology validation). There
have been a bunch of Xen regressions due to that (the most recent is the
one Konrad reported a few days ago with 32 cores). This all is very
fragile for Xen guests due to bogus CPUID/APICID values.

(+Mohit who has been looking into another problem related to topology)

-boris


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
https://lists.xen.org/xen-devel