Re: [PATCH 2/2] x86/hvm: fix corrupt ACPI PM-Timer during live migration

[Kouya Shimura <kouya@jp.fujitsu.com>]

On 05/15/2013 11:54 PM, Jan Beulich wrote:
>>>> On 15.05.13 at 16:49, Alex Bligh <alex@alex.org.uk> wrote:
>> --On 15 May 2013 15:31:19 +0100 Jan Beulich <JBeulich@suse.com> wrote:
>>
>>>> Did this ever get committed? I can't immediately find a commit.
>>>
>>> No, it didn't - no-one knowing that code well enough ever acked
>>> patch 1, and without that we can't apply the patch here.
>>
>> I certainly am not someone who knows that code well, so can't help
>> with that. But I (or more accurately Diana) can reliably replicate
>> live migrate on HVM and qemu-upstream DM causing (a) ACPI entries
>> to disappear from xenstore and (b) walltime to fail to advance in
>> the migrated domain until the walltime is manually set (stuck
>> clock).
>>
>> Is this likely to be related?
>
> I'd like to defer to Kouya to tell whether that matches the
> symptoms he saw.

I don't think my fix is related to Alex's problem.
I observed that gettimeofday() sometimes goes backward *a few seconds*
on migration.

In linux OSs, ACPI Timer value is masked by 24bit.
The valid range:
      (1 / 3.579545MHz) * 0xffffff = 4.7sec
So, the effect of the corrupt ACPI timer is at most 4.7 sec.
The clock shouldn't be frozen for a long time.

Apparently (a) sounds like a problem of xl toolstack.

--
Kouya