[PATCH] PM / Domains: Power on the PM domain right after attach completes

[Rafael J. Wysocki]

On Tuesday, November 18, 2014 01:26:38 AM Rafael J. Wysocki wrote:
> On Monday, November 17, 2014 03:26:04 PM Dmitry Torokhov wrote:
> > On Tue, Nov 18, 2014 at 12:28:26AM +0100, Rafael J. Wysocki wrote:
> > > On Monday, November 17, 2014 02:17:00 PM Dmitry Torokhov wrote:
> > > > On Mon, Nov 17, 2014 at 05:12:35PM -0500, Alan Stern wrote:
> > > > > On Mon, 17 Nov 2014, Dmitry Torokhov wrote:
> > > > > 
> > > > > > > >  Except buses
> > > > > > > > usually call pm_generic_runtime_resume() which ends up fetching driver's
> > > > > > > > callbacks. Maybe pm_generic_runtime_*() need be a bit smarter?
> > > > > > > 
> > > > > > > No, the bus subsystem needs to be smarter.  It shouldn't call 
> > > > > > > pm_generic_runtime_resume() if the driver hasn't been probed yet, or if 
> > > > > > > the driver has already been unbound from the device.
> > > > > > 
> > > > > > But that code wold be exactly the same for all buses, right? So why
> > > > > > can't pm_generic_runtime_resume() be smarter?
> > > > > 
> > > > > It would not be the same for all buses.  Each bus will have its own way 
> > > > > of recognizing whether or not a driver has been probed (i.e., by 
> > > > > checking some field in the bus-specific part of the device structure).
> > > > > 
> > > > > > However, is it allowed to call pm_runtime_get_sync() on devices that
> > > > > > didn't issue pm_runtime_enable()?
> > > > > 
> > > > > Yes.  But the bus has to issue pm_runtime_enable() before probing the 
> > > > > driver, because the driver will expect runtime PM to work properly 
> > > > > while its probe routine runs.  For example, the probe routine might 
> > > > > want to leave the device in a runtime-suspended state.  It can't do 
> > > > > that if the device isn't enabled for runtime PM.
> > > > 
> > > > That means that runtime PM will be enabled for all devices on given bus
> > > > while up till now drivers were deciding if their devices should be
> > > > runtime-pm-managed or not.
> > > 
> > > That's not the case for PCI drivers.
> > > 
> > > > I do not think we are quite ready for this.
> > > 
> > > We have to do that if power domains are in use, however, because if at least
> > > one device in a power domain in enabled for runtime PM, that will affect the
> > > other devices in that domain.
> > > 
> > > We could make a rule to keep a domail always up if at least one device in it
> > > has runtime PM disabled, but that is equivalent to enabling runtime PM for
> > > that device, powering the domain up and bumping up the device's usage counter.
> > 
> > What will driver will see if it tries to check pm_runtime_active()?
> > Would not it get unexpected result if the driver did not call
> > pm_runtime_enable() on it's device?
> 
> Well, that part was supposed to depend on the bus type.  For example, it
> won't be unexpected for PCI drivers, because runtime PM is always enabled
> for PCI devices (although it may be blocked as I described).
> 
> A problem arises if a power domain is used along with a bus type that doesn't
> enable runtime PM for devices by default and drivers are expected to do that.
> That could be addressed by powering up a PM domain (and bumping up its usage
> counter) when adding a device to it and keeping it up until all of the devices
> in it are runtime suspended, but then it also should be turned off eventually
> if there are no drivers for any of those devices.

In other words, I agree with the Ulf's approach in the $subject patch, although
the changelog needs to be updated.

Rafael