Re: [PATCH 0/4] PM / core: Direct handling of DPM_FLAG_SMART_SUSPEND and DPM_FLAG_LEAVE_SUSPENDED

[Ulf Hansson]

On 2 January 2018 at 12:32, Rafael J. Wysocki <rjw@xxxxxxxxxxxxx> wrote:
> On Sunday, December 10, 2017 12:55:23 AM CET Rafael J. Wysocki wrote:
>> Hi All,
>>
>> This series is a follow-up for
>>
>> https://marc.info/?l=linux-doc&m=151101644105835&w=2
>>
>> Patches[1-3/6] from the above have been reviewed and agreed on, so
>> they are in linux-next now and here's a next version of the rest.
>>
>> Patches [1-2/4] are preparatory.  The first one is just really small
>> code duplication avoidance on top of this recent fix:
>>
>> https://patchwork.kernel.org/patch/10097563/
>>
>> and the second one simply moves some code to separate functions.
>>
>> Patch [3/4] causes the PM core to carry out some optimizations for
>> drivers of devices with DPM_FLAG_SMART_SUSPEND set whose "late"
>> and "noirq" suspend (or equivalent) driver callbacks are invoked
>> directly by the core.
>>
>> The underlying observation is that if the device is suspended (via
>> runtime PM) during the "late suspend" phase of a system transition,
>> invoking the "late" and "noirq" callbacks from the driver for it is not
>> going to make it more suspended, so to speak, so it doesn't make sense to
>> invoke them at all.
>>
>> [That optimization is only done for devices with DPM_FLAG_SMART_SUSPEND
>> set, because drivers setting that flag are expected to be prepared for
>> skipping their "late" and "noirq" callbacks if the device is already
>> suspended.]
>>
>> Patch [4/4] makes the core do an analogous thing for devices with
>> DPM_FLAG_LEAVE_SUSPENDED set whose "noirq" and "early" resume (or
>> equivalent) driver callbacks are directly invoked by the core.
>>
>> In that case the observation is that if such devices can be left in
>> suspend after the system transition to the working state, running
>> resume callbacks from their drivers is simply not necessary.
>>
>> Pathes [3-4/4] have been reoredered and reworked a bit since the last
>> iteration, so they are regarded as new.
>>
>> The series is on top of the linux-next branch of the linux-pm.git tree
>> that should be merged into linux-next on Monday.
>>
>> [I have developed debug bus type and driver modules to test that code,
>> but they are not ready to be made available at this point.]
>
> While I acknowledge that Ulf doesn't appear to be convinced by my
> arguments, I also see no technical reason why this cannot go in.

Correct, I am not convinced this is the right path as a general
optimization, at least in it's current form. The main argument is
about skipping invoking callbacks, as I have stated.

Moreover, I think we are lacking important input from some more
experienced PM core code contributors, like Alan, Kevin etc. If any of
those guys would give an ack, that would also make me more comfortable
with this.

On the other hand, I realize that we can't wait forever for that to happen.

>
> As I said during the discussion, I have tested it and it works for me
> as expected.  I also need it to make progress on the drivers front.
>
> Moreover, it should not matter for any drivers that don't set the flags
> in question, so the optimizations introduced here are super-easy to avoid
> by leaving those flags unset.

What prevents you from folding in some changes to a few drivers as
apart of the $subject series?

I have asked for that, as to get a better picture of how this is going
to work in the end.

[...]

Kind regards
Uffe