Re: [lkp] [sched/fair] 41e0d37f7a: divide error: 0000 [#1] SMP

From: Wanpeng Li
Date: Wed May 04 2016 - 08:04:50 EST


2016-05-04 19:56 GMT+08:00 Rafael J. Wysocki <rafael@xxxxxxxxxx>:
> On Wed, May 4, 2016 at 1:51 PM, Wanpeng Li <kernellwp@xxxxxxxxx> wrote:
>> 2016-05-04 19:44 GMT+08:00 Rafael J. Wysocki <rafael@xxxxxxxxxx>:
>>> On Wed, May 4, 2016 at 2:58 AM, Wanpeng Li <kernellwp@xxxxxxxxx> wrote:
>>>> 2016-05-03 20:15 GMT+08:00 Rafael J. Wysocki <rafael@xxxxxxxxxx>:
>>>>> On Tue, May 3, 2016 at 10:32 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>>>>>> On Tue, May 03, 2016 at 09:10:51AM +0800, kernel test robot wrote:
>>>>>>> FYI, we noticed the following commit:
>>>>>>>
>>>>>>> https://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.git sched/core
>>>>>>> commit 41e0d37f7ac81297c07ba311e4ad39465b8c8295 ("sched/fair: Do not call cpufreq hook unless util changed")
>>>>>>
>>>
>>> [cut]
>>>
>>>>>>
>>>>>> That's intel_pstate.c:get_avg_frequency(), which assumes mperf != 0. It
>>>>>> being 0 seems to suggest intel_pstate_sample() hasn't been called yet or
>>>>>> so.
>>>>>
>>>>> Well, what's the tree based on?
>>>>>
>>>>> The mainline does this:
>>>>>
>>>>> bool sample_taken = intel_pstate_sample(cpu, time);
>>>>>
>>>>> if (sample_taken && !hwp_active)
>>>>> intel_pstate_adjust_busy_pstate(cpu);
>>>>>
>>>>> and (the mainline version of) intel_pstate_sample() returns false when
>>>>> it is called for the first time after setting the update_util hook.
>>>>
>>>> The callsites in scheduler will set time to rq_clock(rq) when trigger
>>>> sample, so when time 0 will be used even if it is set just before
>>>> setting the update_util hook?
>>>
>>> I'm not sure what you mean.
>>>
>>> time=0 is special as it will cause intel_pstate_sample() to return
>>> false on the next invocation.
>>
>> Sample is driven by cpufreq_update_util() which uses rq_clock(rq) as
>> time parameter, so there is no opportunity to pass time 0 to
>> intel_pstate_sample().
>
> Right.
>
> So I should have said that had time=0 been passed to
> intel_pstate_sample(), it would have caused it to return false on the
> next invocation. :-)
>
> The way it works is that sample.time is 0 initially, so
> intel_pstate_sample() returns false first time it is called and the
> second invocation gets all of the deltas as needed.

I see, thanks Rafael. :-)

Regards,
Wanpeng Li