Re: [PATCH 2/2][RESEND v3] PM-runtime: change the tracepoints to cover all usage_count

[Chen Yu <yu.c.chen@intel.com>]

On Wed, Jul 15, 2020 at 10:33:22AM +0200, Greg Kroah-Hartman wrote:
> On Wed, Jul 15, 2020 at 04:18:38PM +0800, Chen Yu wrote:
> > Hi Greg,
> > thanks very much for taking a look,
> > On Wed, Jul 15, 2020 at 09:06:14AM +0200, Greg Kroah-Hartman wrote:
> > > On Wed, Jul 15, 2020 at 02:28:03PM +0800, Chen Yu wrote:
> > > > Commit d229290689ae ("PM-runtime: add tracepoints for usage_count changes")
> > > > has added some tracepoints to monitor the change of runtime usage, and
> > > > there is something to improve:
> > > > 1. There are some places that adjust the usage count not
> > > >    been traced yet. For example, pm_runtime_get_noresume() and
> > > >    pm_runtime_put_noidle()
> > > > 2. The change of the usage count will not be tracked if decreased
> > > >    from 1 to 0.
> > > > 
> > > > This patch intends to adjust the logic to be consistent with the
> > > > change of usage_counter, that is to say, only after the counter has
> > > > been possibly modified, we record it. Besides, all usage changes will
> > > > be shown using rpm_usage even if included by other trace points.
> > > > And these changes has helped track down the e1000e runtime issue.
> > > > 
> > > > Reviewed-by: Michał Mirosław <mirq-linux@rere.qmqm.pl>
> > > > Signed-off-by: Chen Yu <yu.c.chen@intel.com>
> > > > ---
> > > >  drivers/base/power/runtime.c | 38 +++++++++++++++++++++++-------------
> > > >  1 file changed, 24 insertions(+), 14 deletions(-)
> > > > 
> > > > diff --git a/drivers/base/power/runtime.c b/drivers/base/power/runtime.c
> > > > index 85a248e196ca..5789d2624513 100644
> > > > --- a/drivers/base/power/runtime.c
> > > > +++ b/drivers/base/power/runtime.c
> > > > @@ -1004,10 +1004,11 @@ 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)) {
> > > > -			trace_rpm_usage_rcuidle(dev, rpmflags);
> > > > +		bool non_zero = !atomic_dec_and_test(&dev->power.usage_count);
> > > > +
> > > > +		trace_rpm_usage_rcuidle(dev, rpmflags);
> > > 
> > > Why not just call trace everywhere before you do the atomic operations?
> > > Why does the trace need to be called after the operation everywhere?
> > > 
> > If I understand correctly, besides Michal's comments, if we put the trace
> > before the atomic operation, we might be unable to judge whether the counter
> > is going to increase or decrease from rpmflags: it is RPM_GET_PUT which combine
> > the get() and put() together, then it is a little inconvenient for tracking IMO.
> 
> A trace can never know the exact value of an atomic value as it could
> change right before or after the trace function is called, right?
> 
> So why are you caring about that?  Care about the functionality that is
> happening, not a reference count that you do not control at all.
>
Ah I see, thanks for the explanation, I'll re-think about the scenaio.

Thanks,
Chenyu
> thanks,
> 
> greg k-h