[V3 patch 06/19] cpuidle: make a single register function for all

Daniel Lezcano daniel.lezcano at linaro.org
Tue Apr 23 10:22:07 EDT 2013


On 04/23/2013 03:56 PM, Santosh Shilimkar wrote:
> On Tuesday 23 April 2013 07:13 PM, Daniel Lezcano wrote:
>> On 04/18/2013 10:48 AM, Santosh Shilimkar wrote:
>>> On Friday 12 April 2013 06:05 PM, Daniel Lezcano wrote:
>>>> The usual scheme to initialize a cpuidle driver on a SMP is:
>>>>
>>>> 	cpuidle_register_driver(drv);
>>>> 	for_each_possible_cpu(cpu) {
>>>> 		device = &per_cpu(cpuidle_dev, cpu);
>>>> 		cpuidle_register_device(device);
>>>> 	}
>>>>
>>> Not exactly related to $subject patch but the driver should
>>> be registered after all devices has been registered to avoid
>>> devices start using the idle state data as soon as it is
>>> registered. In multi CPU system, this race can easily happen.
>>
>> Could you elaborate what problems the system will be facing if a cpu
>> starts using the idle state data as soon as it is registered ?
>>
>> Is there a bug related to this ?
>>
> Ofcouse. In multi-CPU scenario, where CPU C-states needs co-ordination
> can just lead into unknown issues if all the CPUs are not already part
> registered.

Hmm, ok. I don't see a scenario, with the current code, where that could
occurs. The coupled idle state will wait for the other cpus to enter
idle before initiating a shutdown sequence and, so far, the other sync
algorithm (last man standing) are doing the same.

There are some systems with 1024 cpus, and I did not heard problems like
this.

Do you know a system where this problem occurred ? Or is it something
you suspect that can happen ?

That would be interesting to have a system where this race occurs in
order to check the modifications will solve the issue.

>>> Current CPUIDLE core layer is also written with the assumption
>>> that driver will be registered first and then the devices which
>>> is not mandatory as per typical drive/device model.
>>
>> Yes, that's true. The framework assumes cpuidle_register_driver is
>> called before cpuidle_register_device.
>>
>>> May be you can fix that part while you are creating this common
>>> wrapper.
>>
>> Personally, as that will modify the cpuidle core layer and the changes
>> are not obvious (because of the design of the code) I prefer to do that
>> in a separate patchset if it is worth to do it - if there is a bug
>> related to it, then there is no discussion, we have to do it :)
>>
> Sure. It would have been nice if you would have clarified that before
> posting the next version.
> 
> You still need to fix the kernel doc in your v4 though.

Which one ? "s/accross/across" ?

Thanks
  -- Daniel

-- 
 <http://www.linaro.org/> Linaro.org │ Open source software for ARM SoCs

Follow Linaro:  <http://www.facebook.com/pages/Linaro> Facebook |
<http://twitter.com/#!/linaroorg> Twitter |
<http://www.linaro.org/linaro-blog/> Blog




More information about the linux-arm-kernel mailing list