Re: [PATCH] clockevents: return error from tick_broadcast_oneshot_control if !GENERIC_CLOCKEVENTS_BROADCAST

[Preeti U Murthy <preeti@linux.vnet.ibm.com>]

On 06/25/2015 09:00 PM, Sudeep Holla wrote:
> 
> 
> On 25/06/15 14:55, Thomas Gleixner wrote:
>> On Thu, 25 Jun 2015, Sudeep Holla wrote:
>>
>>> tick_broadcast_enter returns 0 when CPU can switch to broadcast
>>> timer and non-zero otherwise. However when GENERIC_CLOCKEVENTS_BROADCAST
>>> and TICK_ONESHOT are disabled, tick_broadcast_oneshot_control returns 0
>>> which indicates to the CPUIdle framework that the CPU can enter deeper
>>> idle states even when the CPU local timer will be shutdown. If the
>>> target state needs broadcast but not broadcast timer is available, then
>>> the CPU can not resume back from that idle state.
>>>
>>> This patch returns error when there's no broadcast timer support
>>> available so that CPUIdle framework prevents the CPU from entering any
>>> idle states losing the local timer.
>>
>> That's wrong and breaks stuff which does not require the broadcast
>> nonsense.
>>
> 
> OK, sorry for not considering that case.
> 
>> If TICK_ONESHOT is disabled, then everything is in periodic mode and
>> tick_broadcast_enter() rightfully returns 0. Ditto for 'highres=off'
>> on the command line.
>>
>> But there is a case which is not correctly handled right now. That's
>> what you are trying to solve in the wrong way.
>>
> 
> Correct I was trying to solve exactly the case mentioned below.
> 
>> If
>>     GENERIC_CLOCKEVENTS_BROADCAST=n
>>
>> or
>>
>>     GENERIC_CLOCKEVENTS_BROADCAST=y and no broadcast device is available,
>>
>> AND cpu local tick device has the C3STOP flag set,
>>
>> then we have no way to tell the idle code that going deep is not
>> allowed.
>>
>> So we need to be smarter than blindly changing a return
>> value. Completely untested patch below.
>>
> 
> Agreed, thanks for the quick patch, I have tested it and it works fine.
> You can add
> 
> Tested-by: Sudeep Holla <sudeep.holla@arm.com>

What about the case where GENERIC_CLOCKEVENTS_BROADCAST=y and
TICK_ONESHOT=n (HZ_PERIODIC=y) ? Have you tested this ?

This will hang the kernel at boot if you are using the hrtimer mode of
broadcast. This is because the local timers of all cpus are shutdown
when the cpuidle driver registers itself, on finding out that there are
idle states where local tick devices stop. The broadcast tick device is
then in charge of waking up the cpus at every period. In hrtimer mode of
broadcast, there is no such real device and we hang.

There was a patch sent out recently to fix this on powerpc.
https://lkml.org/lkml/2015/6/24/42

Regards
Preeti U Murthy
> 
> Regards,
> Sudeep
>