[PATCH v2 5/9] PM / ACPI: Provide option to disable direct_complete for ACPI devices

[Ulf Hansson]

On 24 August 2017 at 16:57, Rafael J. Wysocki <rjw at rjwysocki.net> wrote:
> On Thursday, August 24, 2017 10:19:43 AM CEST Ulf Hansson wrote:
>> On 24 August 2017 at 01:39, Rafael J. Wysocki <rafael at kernel.org> wrote:
>> > On Wed, Aug 23, 2017 at 4:42 PM, Ulf Hansson <ulf.hansson at linaro.org> wrote:
>> >> In some cases a driver for an ACPI device needs to be able to prevent the
>> >> ACPI PM domain from using the direct_complete path during system sleep.
>> >>
>> >> One typical case is when the driver for the device needs its device to stay
>> >> runtime enabled, during the __device_suspend phase. This isn't the case
>> >> when the direct_complete path is being executed by the PM core, as it then
>> >> disables runtime PM for the device in __device_suspend(). Any following
>> >> attempts to runtime resume the device after that point, just fails.
>> >
>> > OK, that is a problem.
>> >
>> >> A workaround to this problem is to let the driver runtime resume its device
>> >> from its ->prepare() callback, as that would prevent the direct_complete
>> >> path from being executed. However, that may often be a waste, especially if
>> >> it turned out that no one really needed the device.
>> >>
>> >> For this reason, invent acpi_dev_disable|enable_direct_complete(), to allow
>> >> drivers to inform the ACPI PM domain to change its default behaviour during
>> >> system sleep, and thus control whether it may use the direct_complete path
>> >> or not.
>> >
>> > But I'm not sure this is the right place to address it as it very well
>> > may affect a PCI device too.
>> >
>> > Also, this is about the device and not about its ACPI companion
>> > object, so putting the flag in there is somewhat unclean, so to speak.
>> >
>> > It looks like we need a flag in dev_pm_info saying something along the
>> > lines of "my system suspend callback can deal with runtime PM" that
>> > will cause the direct_complete update to occur in
>> > __device_suspend_late() instead of __device_suspend().
>>
>> I realize that in the end this turns out to be a comparison between
>> the runtime PM centric path and the direct_complete path while
>> implementing system sleep. In patch 9, there is some more explanation
>> around this, however if you like I can elaborate even more about
>> this!?
>>
>> Regarding making changes to the PM core and adding more flags to the
>> dev_pm_info etc, I am not sure we really want that. Isn't it already
>> complex enough?
>
> Maybe it is.
>
>> My point is, that I am trying to improve the behavior of the ACPI PM
>> domain by enabling the runtime PM centric path for it, and even if
>> there is something similar that could be done for PCI, I don't think
>> we should need involvement by the PM core.
>
> Well, this generally simply doesn't work.
>
> The whole "runtime PM centric approach" idea is generally fine by me,
> but the fact today is that there are drivers not ready for it.  Which
> is why there is the direct_complete thing (it may be regarded as a
> sort-of workaround for the unreadiness of drivers if you will).

This is how I see it:

The runtime PM centric path is being widely deployed, however it takes
time to convert drivers.

The direct_complete path offers a great intermediate step for the ACPI
PM domain as it affects all its devices - while we wait for further
optimizations being deployed using the runtime PM centric path.

>
> Now, buy adding the no_direct_complete flag just to the ACPI PM domain
> you basically overlook the fact that this potentially affects the parents
> of the devices in question by preventing direct_complete from being set
> for them.  And those parents may not be in the ACPI PM domain in principle,
> so the problem needs to be addressed in the core.

Okay, let's move the flag to the dev_pm* structures, to not limit this
to the ACPI PM domain.

However in the current approach taken in this series, as it's coded as
opt-in to use for drivers, I am questioning how big of a problem
parent devices not being able to use the direct_complete path could
be!?
Couldn't it be good enough to just adopt the behavior of the ACPI PM
domain and more or less leave the core out of it - at least for now!?

I am also thinking, that for those parent devices that potentially may
suffer from not be able to use the direct_complete path, those can be
fixed by deploying the runtime PM centric path in their
subsystems/drivers. That would even mean that the parent devices get
the additional benefits that the runtime PM centric path offers. So,
in the end we would end up having a better optimized solution than we
had before.

What do you think?

Kind regards
Uffe