Re: [PATCH] PM-runtime: add tracepoints for usage_count changes

[Rafael J. Wysocki]

On Monday, January 6, 2020 11:00:04 AM CET MichaÅ MirosÅaw wrote:
> On Sat, Jan 04, 2020 at 07:21:23PM +0100, Greg Kroah-Hartman wrote:
> > On Sat, Jan 04, 2020 at 05:27:57PM +0100, MichaÅ MirosÅaw wrote:
> > > Add tracepoints to remaining places where device's power.usage_count
> > > is changed. This helps debugging where and why autosuspend is prevented.
> > > 
> > > Signed-off-by: MichaÅ MirosÅaw <mirq-linux@xxxxxxxxxxxx>
> > > ---
> > >  drivers/base/power/runtime.c | 13 +++++++++++--
> > >  include/trace/events/rpm.h   |  6 ++++++
> > >  2 files changed, 17 insertions(+), 2 deletions(-)
> > > 
> > > diff --git a/drivers/base/power/runtime.c b/drivers/base/power/runtime.c
> > > index 48616f358854..16134a69bf6f 100644
> > > --- a/drivers/base/power/runtime.c
> > > +++ b/drivers/base/power/runtime.c
> > > @@ -1006,8 +1006,10 @@ int __pm_runtime_idle(struct device *dev, int rpmflags)
> > >  	int retval;
> > >  
> > >  	if (rpmflags & RPM_GET_PUT) {
> > > -		if (!atomic_dec_and_test(&dev->power.usage_count))
> > > +		if (!atomic_dec_and_test(&dev->power.usage_count)) {
> > > +			trace_rpm_usage_rcuidle(dev, rpmflags);
> > 
> > Who and what is really going to use these tracepoints?
> > 
> > And putting them in these if statements seems odd, are you sure that's
> > the correct place?  What do these show to userspace?
> > 
> > >  			return 0;
> > > +		}
> > >  	}
> > >  
> > >  	might_sleep_if(!(rpmflags & RPM_ASYNC) && !dev->power.irq_safe);
> > > @@ -1038,8 +1040,10 @@ int __pm_runtime_suspend(struct device *dev, int rpmflags)
> > >  	int retval;
> > >  
> > >  	if (rpmflags & RPM_GET_PUT) {
> > > -		if (!atomic_dec_and_test(&dev->power.usage_count))
> > > +		if (!atomic_dec_and_test(&dev->power.usage_count)) {
> > > +			trace_rpm_usage_rcuidle(dev, rpmflags);
> > >  			return 0;
> > > +		}
> > >  	}
> > >  
> > >  	might_sleep_if(!(rpmflags & RPM_ASYNC) && !dev->power.irq_safe);
> > > @@ -1101,6 +1105,7 @@ int pm_runtime_get_if_in_use(struct device *dev)
> > >  	retval = dev->power.disable_depth > 0 ? -EINVAL :
> > >  		dev->power.runtime_status == RPM_ACTIVE
> > >  			&& atomic_inc_not_zero(&dev->power.usage_count);
> > > +	trace_rpm_usage_rcuidle(dev, 0);
> > 
> > Why this one?
> > 
> > 
> > >  	spin_unlock_irqrestore(&dev->power.lock, flags);
> > >  	return retval;
> > >  }
> > > @@ -1434,6 +1439,8 @@ void pm_runtime_allow(struct device *dev)
> > >  	dev->power.runtime_auto = true;
> > >  	if (atomic_dec_and_test(&dev->power.usage_count))
> > >  		rpm_idle(dev, RPM_AUTO | RPM_ASYNC);
> > > +	else
> > > +		trace_rpm_usage_rcuidle(dev, RPM_AUTO | RPM_ASYNC);
> > 
> > Are you sure this is correct?
> > 
> > These feel odd...
> 
> This covers all places where power.usage_count might have changed.
> If atomic_dec_and_test() decrements usage_count but not to zero,
> the new value will be traced in rpm_idle(). But if the value drops to
> zero, then we need to trace it explicitly to be able to have all changes
> accounted for in the trace.
> 
> I actually used this patch to track down why USB storage device was
> not autosuspending correctly.

Fair enough, and the patch makes sense to me, so I'm queuing it up as 5.6
material.

Thanks!