Re: [PATCH v2 7/7] ARM: smp: Add runtime PM support for CPU hotplug

[Grygorii Strashko <grygorii.strashko@ti.com>]

On 09/07/2015 04:04 PM, Rafael J. Wysocki wrote:
> On Saturday, September 05, 2015 11:39:20 AM Alan Stern wrote:
>> On Sat, 5 Sep 2015, Grygorii Strashko wrote:
>>
>>> On 09/04/2015 09:45 PM, Alan Stern wrote:
>>>> On Fri, 4 Sep 2015, Grygorii Strashko wrote:
>>>>
>>>>> There is one "small" problem with such approach :(
>>>>>
>>>>> - It's incompatible with -RT kernel, because PM runtime can't be used
>>>>> in atomic context on -RT.
>>>>
>>>> Can you explain this more fully?  Why can't runtime PM be used in
>>>> atomic context in the -rt kernels?
>>>>   
>>>
>>> See:
>>>   http://lwn.net/Articles/146861/
>>>   https://rt.wiki.kernel.org/index.php/Frequently_Asked_Questions#How_does_the_CONFIG_PREEMPT_RT_patch_work.3F
>>>
>>> spinlock_t
>>>      Critical sections are preemptible. The _irq operations (e.g., spin_lock_irqsave())
>>>   do -not- disable hardware interrupts. Priority inheritance is used to prevent priority
>>>   inversion. An underlying rt_mutex is used to implement spinlock_t in PREEMPT_RT.
>>>
>>> As result, have to do things like:
>>> https://lkml.org/lkml/2015/8/18/161
>>> https://lkml.org/lkml/2015/8/18/162
>>>
>>> Sorry for brief reply - Friday/Sat night :)
>>
>> I see.  Although we normally think of interrupt contexts as being
>> atomic, in an -rt kernel this isn't true any more because things like
>> spin_lock_irq don't actually disable interrupts.
>>
>> Therefore it would be correct to say that in -rt kernels, runtime PM
>> can be used in interrupt context (if the device is marked as irq-safe),
>> but not in atomic context.  Right?
> 
> Right.
> 
> Whatever is suitable for interrupt context in the mainline, will be suitable
> for that in -rt kernels too. 

Not exactly true :(, since spinlock is converted to [rt_] mutex.
Usually, this difference can't be seen because on -RT kernel all or
mostly all HW IRQ handlers will be forced to be threaded.
For the cases, where such automatic conversion is not working,
(like chained irq handlers or HW-handler+Threaded handler) the code
has to be carefully patched to work properly as for non-RT as for -RT.

Also, this triggers some -RT incompatibility issues, like with PM runtime or
CLK framework (http://www.spinics.net/lists/linux-rt-users/msg13653.html).

> However, what is suitable for the idle loop
> in the mainline, may not be suitable for that in -rt kernels.
> 
> That's why raw_spin_lock/unlock() need to be used within the idle loop.

Indeed. CPU hotplug/CPUIdle is guarded by local_irq_disable()/local_irq_enable().
(example of CPU hotplug RT-issue http://www.spinics.net/lists/arm-kernel/msg438963.html).

I don't want to be the final authority here as my experience with -RT is short.
But, I want to point out on potential issues based on what I've already discovered
and tried to fix.

Thanks & regards,
-grygorii