Re: [RFC][PATCH 2/3] PM / sleep: Mechanism to avoid resuming runtime-suspended devices unnecessarily

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

On Tuesday, May 13, 2014 10:49:32 AM Alan Stern wrote:
> On Tue, 13 May 2014, Rafael J. Wysocki wrote:
> 
> > From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> > 
> > Currently, some subsystems (e.g. PCI and the ACPI PM domain) have to
> > resume all runtime-suspended devices during system suspend, mostly
> > because those devices may need to be reprogrammed due to different
> > wakeup settings for system sleep and for runtime PM.
> > 
> > For some devices, though, it's OK to remain in runtime suspend 
> > throughout a complete system suspend/resume cycle (if the device was in
> > runtime suspend at the start of the cycle).  We would like to do this
> > whenever possible, to avoid the overhead of extra power-up and power-down
> > events.
> > 
> > However, problems may arise because the device's descendants may require
> > it to be at full power at various points during the cycle.  Therefore the
> > most straightforward way to do this safely is if the device and all its
> > descendants can remain runtime suspended until the complete stage of
> > system resume.
> > 
> > To this end, introduce a new device PM flag, power.direct_complete
> > and modify the PM core to use that flag as follows.
> > 
> > If the ->prepare() callback of a device returns a positive number,
> > the PM core will regard that as an indication that it may leave the
> > device runtime-suspended.  It will then check if the system power
> > transition in progress is a suspend (and not hibernation in particular)
> > and if the device is, indeed, runtime-suspended.  In that case, the PM
> > core will set the device's power.direct_complete flag.  Otherwise it
> > will clear power.direct_complete for the device and it also will later
> > clear it for the device's parent (if there's one).
> > 
> > Next, the PM core will not invoke the ->suspend() ->suspend_late(),
> > ->suspend_irq(), ->resume_irq(), ->resume_early(), or ->resume()
> > callbacks for all devices having power.direct_complete set.  It
> > will invoke their ->complete() callbacks, however, and those
> > callbacks are then responsible for resuming the devices as
> > appropriate, if necessary.
> 
> Perhaps you should mention here (and maybe even as a comment in the 
> code) that ->complete() callbacks may want to call pm_request_resume() 
> if dev->power.direct_resume is set, but they shouldn't call 
> pm_runtime_resume().

OK

> > Changelog partly based on an Alan Stern's description of the idea
> > (http://marc.info/?l=linux-pm&m=139940466625569&w=2).
> > 
> > Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> 
> ...
> 
> > @@ -1518,17 +1527,19 @@ static int device_prepare(struct device
> >  		callback = dev->driver->pm->prepare;
> >  	}
> >  
> > -	if (callback) {
> > -		error = callback(dev);
> > -		suspend_report_result(callback, error);
> > -	}
> > +	if (callback)
> > +		ret = callback(dev);
> >  
> >  	device_unlock(dev);
> >  
> > -	if (error)
> > +	if (ret < 0) {
> > +		suspend_report_result(callback, ret);
> >  		pm_runtime_put(dev);
> > -
> > -	return error;
> > +		return ret;
> > +	}
> > +	dev->power.direct_complete = ret > 0 && state.event == PM_EVENT_SUSPEND
> > +					&& pm_runtime_suspended(dev);
> 
> Shouldn't the flag be set under the spinlock?

I guess you're worried about runtime PM flags being modified in parallel to
this?  But we've just done the barrier a while ago, so is that still a concern
here?

This won't run in parallel with device_prepare() for any other devices, because
the "complete" phase is sequential.

Rafael