question regarding intel_idle inside kvm

question regarding intel_idle inside kvm

[Daniel Lezcano]

Hi all,

I recently did some modification in the cpuidle core and the patches 
were merge to linux-next.
Someone reported a problem with the intel_idle cpuidle driver.

I tried to reproduce the problem with kvm but the kernel fails to 
intialize the driver because of intel_intel_init function fails in the 
processor probe.

After digging a bit, I found it fails at:

drivers/idle/intel_idle.c

static int intel_idle_probe(void)
{

...


         if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF)
                 return -ENODEV;

         ^^^^^

  ...

}

I assumed the virtualized processor does not support this, so I 
specified the "-cpu host" because the host was running the intel_idle 
driver. But the driver still fails in kvm.

I was wondering why that happens ? Does anyone have an idea of this 
problem ?

Thanks in advance
   -- Daniel

Re: question regarding intel_idle inside kvm

[Avi Kivity]

On 03/16/2012 12:19 AM, Daniel Lezcano wrote:
> Hi all,
>
> I recently did some modification in the cpuidle core and the patches
> were merge to linux-next.
> Someone reported a problem with the intel_idle cpuidle driver.
>
> I tried to reproduce the problem with kvm but the kernel fails to
> intialize the driver because of intel_intel_init function fails in the
> processor probe.
>
> After digging a bit, I found it fails at:
>
> drivers/idle/intel_idle.c
>
> static int intel_idle_probe(void)
> {
>
> ...
>
>
>         if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF)
>                 return -ENODEV;
>
>         ^^^^^
>
>  ...
>
> }
>
> I assumed the virtualized processor does not support this, so I
> specified the "-cpu host" because the host was running the intel_idle
> driver. But the driver still fails in kvm.
>
> I was wondering why that happens ? Does anyone have an idea of this
> problem ?

intel_idle() uses mwait, which kvm does not virtualize (it's very
expensive to do so and brings no benefits).

-- 
error compiling committee.c: too many arguments to function

Re: question regarding intel_idle inside kvm

[Daniel Lezcano]

On 03/19/2012 01:31 PM, Avi Kivity wrote:
> On 03/16/2012 12:19 AM, Daniel Lezcano wrote:
>> Hi all,
>>
>> I recently did some modification in the cpuidle core and the patches
>> were merge to linux-next.
>> Someone reported a problem with the intel_idle cpuidle driver.
>>
>> I tried to reproduce the problem with kvm but the kernel fails to
>> intialize the driver because of intel_intel_init function fails in the
>> processor probe.
>>
>> After digging a bit, I found it fails at:
>>
>> drivers/idle/intel_idle.c
>>
>> static int intel_idle_probe(void)
>> {
>>
>> ...
>>
>>
>>          if (boot_cpu_data.cpuid_level<  CPUID_MWAIT_LEAF)
>>                  return -ENODEV;
>>
>>          ^^^^^
>>
>>   ...
>>
>> }
>>
>> I assumed the virtualized processor does not support this, so I
>> specified the "-cpu host" because the host was running the intel_idle
>> driver. But the driver still fails in kvm.
>>
>> I was wondering why that happens ? Does anyone have an idea of this
>> problem ?
> intel_idle() uses mwait, which kvm does not virtualize (it's very
> expensive to do so and brings no benefits).

Ok, thanks for the information. I was afraid of that :/
I will go to for a real host then :)

Thanks !
   -- Daniel