Re: [PATCH] sched/idle: Export cpu_idle_poll_ctrl() symbol

[Daniel Lezcano]

On 26/11/2021 06:56, Maulik Shah wrote:
> Hi Daniel,
> 
> On 11/25/2021 10:56 PM, Daniel Lezcano wrote:
>> On 25/11/2021 15:13, Maulik Shah wrote:
>>> Hi Peter,
>>>
>>> On 11/25/2021 3:21 PM, Peter Zijlstra wrote:
>>>> On Thu, Nov 25, 2021 at 02:44:36PM +0530, Maulik Shah wrote:
>>>>> Export cpu_idle_poll_ctrl() so that module drivers can use same.
>>>> This does not seem like a really safe interface to expose to the
>>>> world.
>>> Thanks for the review.
>>>
>>> Keeping the cpuidle enabled from boot up may delay/increase the boot up
>>> time.
>>> Below is our use case to force cpuidle to stay in cpu_idle_poll().
>>>
>>> We keep cpuidle disabled from boot up using "nohlt" option of kernel
>>> command line which internally sets cpu_idle_force_poll = 1;
>>> and once the device bootup reaches till certain point (for example the
>>> android homescreen is up) userspace may notify a
>>> vendor module driver which can invoke cpu_idle_poll_ctrl(false); to come
>>> out of poll mode.
>>> So vendor module driver needs cpu_idle_poll_ctrl() exported symbol.
>>>
>>> We can not take PM-QoS from driver to prevent deep cpuidle since all the
>>> vendor modules are kept in a separate partition and will be loaded only
>>> after kernel boot up is done
>>> and by this time kernel already starts executing deep cpuidle modes.
>>>> Surely the better solution is to rework things to not rely on this. I'm
>>>> fairly sure it's not hard to write a cpuidle driver that does much the
>>>> same.
>>> The other option i think is to pass cpuidle.off=1 in kernel command line
>>> and then add enable_cpuidle() in drivers/cpuidle/cpuidle.c
>>> something similar as below which can be called by vendor module.
>>>
>>> void enable_cpuidle(void)
>>> {
>>>          off = 0;
>>> }
>>> EXPORT_SYMBOL_GPL(enable_cpuidle);
>>>
>>> This may be a good option since we have already disable_cpuidle() but
>>> not enable_cpuidle().
>>>
>>> void disable_cpuidle(void)
>>> {
>>>          off = 1;
>>> }
>>>
>>> Hi Rafael/Daniel, can you please let me know your suggestion on
>>> this/similar implementation?
>> Did you try to use the QoS latency? Sounds like it is exactly for this
>> purpose.
>>
>> Set it to zero to force cpuidle to choose the shallowest idle state and
>> then INT_MAX to disable the constraint.
>>
>>   cpu_latency_qos_add_request();
>>
>> Hope that helps
>>
>>    -- Daniel
> The PM-QoS is not helping here since all the vendor drivers are kept in
> a separate partition
> and will be loaded only after kernel boot up is done and by the time
> vendor kernel modules are inserted
> takes QoS, kernel/menu governor already starts executing deep cpuidle
> modes.
> 
> kernel start (t0)---------Menu governor loads (t1)----------vendor
> modules loaded (t2)----------Usespace ready(t3)
> 
> Untill (t2), its only core kernel/android kernel which don't have any
> vendor driver which can take QoS.
> If we take QoS, it can be taken only from point (t2) but CPUs still
> enter deep idle state between (t1) to (t2).
> 
> So to prevent this passing "cpuidle.off=1" or "nohlt" in kernel command
> line can keep deep cpuidle states disabled from boot up and
> once vendor modules are ready at (t2) or (t3), it can either invoke
> newly added enable_cpuidle() or cpu_idle_poll_ctrl(false);
> to comeout of polling mode and start executing deep low power modes.

I understand. The approach is valid but I'm wondering if it should fall
under a more global feature with a perf <-> power cursor

If the goal is to be as fast as possible at boot time, the cursor should
be set to 'perf and then changed to 'power' after the system has booted.

So not only cpuidle but also any other subsystems can skip the PM path
or go the highest performance state (assuming the PM code is compiled-in
for the devices).




-- 
<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