Re: [PATCH] xen: add steal_clock support on x86

[Juergen Gross]

On 18/05/16 18:13, Boris Ostrovsky wrote:
> On 05/18/2016 12:00 PM, Juergen Gross wrote:
>> On 18/05/16 17:53, Boris Ostrovsky wrote:
>>> On 05/18/2016 11:45 AM, David Vrabel wrote:
>>>> On 18/05/16 16:42, Juergen Gross wrote:
>>>>> On 18/05/16 17:25, Boris Ostrovsky wrote:
>>>>>> On 05/18/2016 10:53 AM, Juergen Gross wrote:
>>>>>>> On 18/05/16 16:46, Boris Ostrovsky wrote:
>>>>>>>> On 05/18/2016 08:15 AM, Juergen Gross wrote:
>>>>>>>>>  }
>>>>>>>>>  
>>>>>>>>> +void __init xen_time_setup_guest(void)
>>>>>>>>> +{
>>>>>>>>> +	pv_time_ops.steal_clock = xen_steal_clock;
>>>>>>>>> +
>>>>>>>>> +	static_key_slow_inc(&paravirt_steal_enabled);
>>>>>>>>> +	/*
>>>>>>>>> +	 * We can't set paravirt_steal_rq_enabled as this would require the
>>>>>>>>> +	 * capability to read another cpu's runstate info.
>>>>>>>>> +	 */
>>>>>>>>> +}
>>>>>>>> Won't we be accounting for stolen cycles twice now --- once from
>>>>>>>> steal_account_process_tick()->steal_clock() and second time from
>>>>>>>> do_stolen_accounting()?
>>>>>>> Uuh, yes.
>>>>>>>
>>>>>>> I guess I should rip do_stolen_accounting() out, too? 
>>>>>> I don't think PARAVIRT_TIME_ACCOUNTING is always selected for Xen. If
>>>>> This is easy to accomplish. :-)
>>>
>>> I looked at KVM code (PARAVIRT_TIME_ACCOUNTING is not selected there
>>> neither) and in their case that's presumably because stealing accounting
>>> is a CPUID bit, i.e. it might not be supported. In Xen case we always
>>> have this interface.
>> So they added it later and the default is to keep the old behavior.
>>
>>>>>> that's indeed the case then we should ifndef do_stolen_accounting(). Or
>>>>>> maybe check for paravirt_steal_enabled.
>>>>> Is this really a sensible thing to do? There is a mechanism used by KVM
>>>>> to do the stolen accounting. I think we should use it instead of having
>>>>> a second implementation doing the same thing in case the generic one
>>>>> isn't enabled.
>>>> I agree.
>>>>
>>>> Although I don't think selecting PARAVIRT_TIME_ACC' is necessary -- I
>>>> don't think it's essential (or is it?).
>>> Looks like it's useful only if paravirt_steal_rq_enabled, which we don't
>>> support yet.
>> I think the patch is still useful. It is reducing code size and
>> it is removing arch-specific Xen-hack(s). With the patch Xen's
>> solution for arm and x86 is common and the same as for KVM. Adding
>> paravirt_steal_rq_enabled later will be much easier as only one
>> function needs substantial modification.
> 
> I am not arguing against having a patch that will remove
> do_stolen_accounting(). I was responding to David's statement about
> whether we need to select CONFIG_PARAVIRT_TIME_ACCOUNTING, and I am not
> sure this is necessary since steal_account_process_tick() (that will
> take case of things that do_stolen_accounting() currently does) doesn't
> need it.

Aah, okay. That's a good reason to not add the Kconfig stuff.

> (And if it is indeed needed --- can we have Xen's Kconfig select it
> instead of "default y if XEN" ?)

I've verified that CONFIG_PARAVIRT_TIME_ACCOUNTING is _not_ needed.
I've removed it from .config and used my patch with
do_stolen_accounting() removed. In an overcommitted guest (4 vcpus on 2
physical cpus) running a parallel make top showed near 50% stolen time.


Juergen