Re: [PATCH 3/3] trace points: power: remove 'cpu_id' from trace_cpu_idle

[Ming Lei <tom.leiming@gmail.com>]

Hi,

2011/8/20 Thomas Renninger <trenn@suse.de>:
> On Friday, August 19, 2011 05:04:04 PM tom.leiming@gmail.com wrote:
>> From: Ming Lei <tom.leiming@gmail.com>
>>
>> This patch removes the 'cpu_id' parameter of the cpu_idle
>> trace point, based on the ideas below:
>>
>> - the cpu_id which is passed to trace point is always the current
>>   cpu
> Are you sure this will always be true?

It is sure at least now,  the only place to pass 'dev->cpu' is inside
cpuidle_idle_call,
but the cpuidle_device is got from __this_cpu_read(cpuidle_devices),
so it should
be true for such case.

>
>> - the current cpu info has been included into the trace result
>>   already
>> - smp_processor_id() can't be used safely in preemptible context.
>
> The cpuid has been added to idle events on purpose for example to be
> able to pass C-state dependencies.
> 2 cores may only enter a deeper sleep state if the 2nd core enters it
> as well.
> There may be architectures where you could trigger a sleep state on
> a foreign cpu.
>
> This may not be used now, but if the kernel does want to use it, you do
> not want to adjust a dozen userspace apps.
>
> Not sure how to quickly solve the:
> "smp_processor_id() can't be used safely in preemptible context."
> problem, though.
> A convention could be declared that if -1 is passed, it's the same cpu
> entering the sleep state as the running one. This will probably
> break userspace apps as well...
>
> Best would be to clean up x86 and let idle routines always be entered
> from cpuidle subsystem which knows the cpu id already and eliminate
> all idle functions in arch/x86/kernel/process.c.
>
>   Thomas
>
> PS: I do not care about patch 1 and 2 as these events are ARM specific
> afaik. But I vote against this one.
>


thanks,
-- 
Ming Lei