Re: [PATCH v3 2/2] PM / Runtime: Add pm_runtime_enable_recursive

["Rafael J. Wysocki" <rjw@rjwysocki.net>]

On Friday, July 03, 2015 04:22:02 PM Tomeu Vizoso wrote:
> On 3 July 2015 at 16:16, Alan Stern <stern@rowland.harvard.edu> wrote:
> > On Fri, 3 Jul 2015, Tomeu Vizoso wrote:
> >
> >> On 2 July 2015 at 17:21, Alan Stern <stern@rowland.harvard.edu> wrote:
> >> > On Thu, 2 Jul 2015, Tomeu Vizoso wrote:
> >> >
> >> >> > Just because these sub-devices are virtual, it doesn't mean you can
> >> >> > ignore the way they interact with runtime PM.
> >> >>
> >> >> Fair enough, but then, how are we expected to be able to use the
> >> >> direct_complete facility if the core bails out if a descendant doesn't
> >> >> have runtime PM enabled?
> >> >>
> >> >> > In the case of ep_87 this doesn't matter.  Endpoint devices (like all
> >> >> > devices) are in the SUSPENDED state by default when they are created,
> >> >> > and they never leave that state.
> >> >>
> >> >> I don't see why it doesn't matter for endpoints or the others. They
> >> >> don't have runtime PM enabled, so no ancestor will be able to do
> >> >> direct_complete.
> >> >
> >> > Ah, you're concerned about these lines near the start of
> >> > __device_suspend():
> >> >
> >> >         if (dev->power.direct_complete) {
> >> >                 if (pm_runtime_status_suspended(dev)) {
> >> >                         pm_runtime_disable(dev);
> >> >                         if (pm_runtime_suspended_if_enabled(dev))
> >> >                                 goto Complete;
> >> >
> >> >                         pm_runtime_enable(dev);
> >> >                 }
> >> >                 dev->power.direct_complete = false;
> >> >         }
> >> >
> >> > Perhaps the pm_runtime_suspended_if_enabled() test should be changed to
> >> > pm_runtime_status_suspended().  Then it won't matter whether the
> >> > descendant devices are enabled for runtime PM.
> >>
> >> Yeah, that would remove the need for messing with the runtime PM
> >> enable status of descendant devices, but I wonder why Rafael went that
> >> way initially.
> >
> > I forget the details.  Probably it was just to be safe.  We probably
> > thought that if a device was disabled for runtime PM then its runtime
> > PM status might not be accurate.  But if direct_complete is set then it
> > may be reasonable to assume that the runtime PM status _is_ accurate.
> 
> Cool.

We're walking a grey area here.  What exactly does power.direct_complete mean
for devices whose runtime PM is disabled?

> >> >> > A possible way around the problem is to use pm_suspend_ignore_children
> >> >> > on the uvcvideo interface.  But I'm not sure that would be the right
> >> >> > thing to do.
> >> >>
> >> >> Would that mean that if a device has ignore_children then it could
> >> >> still do direct_complete even if its descendants weren't able to?
> >> >
> >> > I think we could justify that.  The ignore_children flag means we can
> >> > communicate with the children even when the device is in runtime
> >> > suspend, so there's no reason to force the device to leave runtime
> >> > suspend during a system sleep.
> >>
> >> IIUIC, what you are proposing is to use ignore_children in a way
> >> similar to how force_direct_complete was used in this patch?
> >>
> >> http://thread.gmane.org/gmane.linux.power-management.general/60198/focus=60292
> >
> > That message doesn't contain a patch.
> 
> The patch is at the top of the thread:
> 
> http://thread.gmane.org/gmane.linux.power-management.general/60198/focus=60292
> 
> >> That should work as well, but Rafael raised some objections and thus I
> >> went with the present direct_complete_default, which should work if we
> >> can relax the check as discussed above.
> >
> > Rafael and I briefly discussed ignore_children while the original
> > direct_complete patch was being designed.  We didn't come to any
> > definite conclusion and decided to forget about it for the time being.
> > Maybe now would be a good time to reconsider it.
> 
> I would prefer to have ignore_children ignore whether the children of
> a device were able to do direct_complete, rather than having a
> direct_complete_default flag (plus not requiring that all its
> descendants have runtime PM enabled).

Why?

Rafael