Re: [PATCH] PM / wakeirq: report wakeup events in dedicated wake-IRQs

From: Rafael J. Wysocki
Date: Fri Nov 11 2016 - 16:33:22 EST


On Fri, Nov 11, 2016 at 5:31 PM, Tony Lindgren <tony@xxxxxxxxxxx> wrote:
> * Rafael J. Wysocki <rafael@xxxxxxxxxx> [161110 16:06]:
>> On Thu, Nov 10, 2016 at 7:49 PM, Brian Norris <briannorris@xxxxxxxxxxxx> wrote:
>> > On Thu, Nov 10, 2016 at 10:13:55AM -0800, Dmitry Torokhov wrote:
>> >> On Thu, Nov 10, 2016 at 10:07 AM, Brian Norris <briannorris@xxxxxxxxxxxx> wrote:
>> >> > It's important that user space can figure out what device woke the
>> >> > system from suspend -- e.g., for debugging, or for implementing
>> >> > conditional wake behavior. Dedicated wakeup IRQs don't currently do
>> >> > that.
>> >> >
>> >> > Let's report the event (pm_wakeup_event()) and also allow drivers to
>> >> > synchronize with these events in their resume path (hence, disable_irq()
>> >> > instead of disable_irq_nosync()).
>> >>
>> >> Hmm, dev_pm_disable_wake_irq() is called from
>> >> rpm_suspend()/rpm_resume() that take dev->power.lock spinlock and
>> >> disable interrupts. Dropping _nosync() feels dangerous.
>> >
>> > Indeed. So how do you suggest we get sane wakeup reports? Every device
>> > or bus that's going to use the dedicated wake APIs has to
>> > synchronize_irq() [1] in their resume() routine? Seems like an odd
>> > implementation detail to have to remember (and therefore most drivers
>> > will get it wrong).
>> >
>> > Brian
>> >
>> > [1] Or maybe at least create a helper API that will extract the
>> > dedicated wake IRQ number and do the synchronize_irq() for us, so
>> > drivers don't have to stash this separately (or poke at
>> > dev->power.wakeirq->irq) for no good reason.
>>
>> Well, in the first place, can anyone please refresh my memory on why
>> it is necessary to call dev_pm_disable_wake_irq() under power.lock?
>
> I guess no other reason except we need to manage the wakeirq
> for rpm_callback(). So we dev_pm_enable_wake_irq() before
> rpm_callback() in rpm_suspend(), then disable on resume.

But we drop the lock in rpm_callback(), so can't it be moved to where
the callback is invoked?

Thanks,
Rafael