Re: pv_ops dom0 clock only runs under load

[Jeremy Fitzhardinge <jeremy@goop.org>]

Milan Holzäpfel wrote:
> On Fri, 6 Mar 2009 21:46:49 +0100
> Milan Holzäpfel <listen@mjh.name> wrote:
>
>   
>> On Thu, 05 Mar 2009 21:41:44 -0800
>> Jeremy Fitzhardinge <jeremy@goop.org> wrote:
>>
>>     
>>> Milan Holzäpfel wrote:
>>>
>>>       
>>>> The clock only seems to run if the system is under load (e.g. run
>>>> "while true; do echo -n; done" in a shell).  
>>>> With the CPU loaded and the clock running everything seems to work
>>>> quite well (hard disk and network access, that is.)
>>>>
>>>> Changing the "xen" clocksource to "tsc" via sysfs didn't make a
>>>> difference.  Enabling or disabling the "tickless" option doesn't make a
>>>> difference either. 
>>>>   
>>>>         
>>> That's pretty mysterious; it does suggest your cpu's tsc is not ticking.
>>>       
>> Excerpt from the dmesg (2.6.29 pvops with / without Xen)
>>
>> [    4.870697] Marking TSC unstable due to TSC halts in idle
>>
>> Saw it scrolling by.
>>
>> [...]
>>
>> By the way, I booted the 2.6.29 pvops enabled kernel without Xen and
>> the tsc timesource only increments when the CPU was loaded, too.  On
>> 2.6.18 (without Xen) it worked normally (also with the BIOS options
>> enabled).
>>
>> [...]
>>
>> I will try whether disabling kernel stuff (there is something called
>> "cpuidle" compiled in) makes the tsc clocksource work again. 
>>     
>
> Booting with idle=poll appended to the kernel command line makes the
> TSC work again. 
>
> The "xen" clocksource of the 2.6.29 pvops kernel (under Xen 3.3.1)
> works again.  2.6.29 pvops on "bare metal" chooses TSC as default time
> source, which works too. 
>
> Doesn't seem like a real solution, does it? The CPU doesn't seem to run
> at full load all the time though (temperature stays quite low).
>   

No, its not a great solution.  But a tsc which only runs while the cpu 
is non-idle isn't a huge problem; it just needs Xen to update the timing 
parameters each time it idles the CPU.  But its a bit of a mystery that 
2.6.18.8 works OK...

> Another question: The xm dmesg I posted reports "(XEN) Platform timer
> is 14.318MHz HPET".  Why is the TSC used for time keeping then? The
> HPET works perfectly on every kernel not running under Xen. 
>   

Xen uses the hpet to maintain a notion of system-wide time.  However its 
fairly expensive to read the hpet, so each CPU uses its own local time 
which is tracked using the tsc and synchronized with the hpet.

    J