Re: A clocksource question

[Joanna Rutkowska <joanna@invisiblethingslab.com>]

[-- Attachment #1.1: Type: text/plain, Size: 2308 bytes --]

On 03/11/2010 01:44 AM, Dan Magenheimer wrote:
>> From: Ian Pratt [mailto:Ian.Pratt@eu.citrix.com]
>> Sent: Wednesday, March 10, 2010 5:22 PM
>> To: Dan Magenheimer; Joanna Rutkowska; Jeremy Fitzhardinge
>> Cc: Ian Pratt; xen-devel@lists.xensource.com
>> Subject: RE: [Xen-devel] A clocksource question
>>
>>> The TSC delta is troubling... if you hadn't said you had turned
>>> off all power management, I would have guessed a problem with
>>> C-state management.  Maybe Xen is discovering some power management
>>> capability not visible in BIOS settings?
>>
>> Xen uses the architectural C state mechanism, bypassing ACPI.
>> Use "max_cstate=0"
> 
> Yes!  The Core 2 Duo fooled me.  There are some versions
> of Core 2 Duo ("Conroe") that don't support C3-state and
> some ("Merom") that do support C3-state.  And the code
> to "recover" from C3, which I think was added before 3.4
> was released, has been observed to be very poor in its
> attempt to reset TSC after C3 to a reasonable value. I
> didn't think it was THAT poor though!  See:
> http://lists.xensource.com/archives/html/xen-devel/2009-10/msg01414.html 
> 
> (AFAIK, this is not fixed in 4.0, though the new default
> rdtsc emulation may mask the problem by ensuring TSC always
> moves forward, even across different processors.)
> 
> So that may explain the TSC delta.  Since the pv clocksource
> algorithm is dependent in part on the hardware TSC being
> synced reasonably well by Xen, that may also explain
> other clock strangeness.
> 
> P.S. Joanna -- max_cstate=0 must be specified on the
> Xen boot line in grub.conf, not in the vm.cfg or grub.conf
> of the guest.

Dan, I tried passing this argument to Xen, of course and it didn't help.
I still think it's a Dom0-related problem rather than a hypervisor,
because when I was using xen/stable-2.6.32-based kernel for a moment,
I'm pretty sure it worked fine and I didn't have any problem with the
clocksource. On the same hypervisor, when I used xen/stable-2.6.31
kernel, the problem was present. (I had to switch back to 2.6.31
because, back then, the 2.6.32 didn't have pciback, which was important
for me; and recently I decided to downgrade to Xen 3.4.2, for its
presumed stability, and so 2.6.31 is the only option now).

joanna.


[-- Attachment #1.2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 226 bytes --]

[-- Attachment #2: Type: text/plain, Size: 138 bytes --]

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel