diff mbox series

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

Message ID 395187057e486df9a4328bc6d7d4ee912967fdb3.1594790493.git.yu.c.chen@intel.com (mailing list archive)
State Rejected, archived
Headers show
Series Extend trace point to cover all runtime usage count | expand

Commit Message

Chen Yu July 15, 2020, 6:28 a.m. UTC
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(-)

Comments

Greg Kroah-Hartman July 15, 2020, 7:06 a.m. UTC | #1
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?

thanks,

greg k-h
Michał Mirosław July 15, 2020, 7:27 a.m. UTC | #2
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?

I would argue that this is easier mentally: We trace what state the
device is in from now on (a "current state" for the time being) instead
of tracing what it was before (an information that has just expired).

Best Regards,
Michał Mirosław
Greg Kroah-Hartman July 15, 2020, 8:17 a.m. UTC | #3
On Wed, Jul 15, 2020 at 09:27:28AM +0200, Michal Miroslaw wrote:
> 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?
> 
> I would argue that this is easier mentally: We trace what state the
> device is in from now on (a "current state" for the time being) instead
> of tracing what it was before (an information that has just expired).

Is that really the case here and you look at that atomic value somehow
in the trace and need it?

thanks,

greg k-h
Chen Yu July 15, 2020, 8:18 a.m. UTC | #4
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.

thanks,
Chenyu
> thanks,
> 
> greg k-h
Greg Kroah-Hartman July 15, 2020, 8:33 a.m. UTC | #5
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.

thanks,

greg k-h
Rafael J. Wysocki July 15, 2020, 3:47 p.m. UTC | #6
On Wed, Jul 15, 2020 at 8:26 AM Chen Yu <yu.c.chen@intel.com> 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);

It looks like you could move the trace event before the atomic variable check.

The ordering between the two doesn't matter, because usage_count may
change between the check and the trace event anyway.

But then what is the trace event useful for in the first place?

> +               if (non_zero)
>                         return 0;
> -               }
>         }
>
>         might_sleep_if(!(rpmflags & RPM_ASYNC) && !dev->power.irq_safe);
> @@ -1038,10 +1039,12 @@ 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)) {
> -                       trace_rpm_usage_rcuidle(dev, rpmflags);
> +               bool non_zero = !atomic_dec_and_test(&dev->power.usage_count);
> +
> +               trace_rpm_usage_rcuidle(dev, rpmflags);

And the same comments apply here.

> +               if (non_zero)
>                         return 0;
> -               }
> +
>         }
>
>         might_sleep_if(!(rpmflags & RPM_ASYNC) && !dev->power.irq_safe);
> @@ -1073,8 +1076,10 @@ int __pm_runtime_resume(struct device *dev, int rpmflags)
>         might_sleep_if(!(rpmflags & RPM_ASYNC) && !dev->power.irq_safe &&
>                         dev->power.runtime_status != RPM_ACTIVE);
>
> -       if (rpmflags & RPM_GET_PUT)
> +       if (rpmflags & RPM_GET_PUT) {
>                 atomic_inc(&dev->power.usage_count);

So the reason why things like that don't work is because the atomic
variable can change again between the inc and the trace event.

> +               trace_rpm_usage_rcuidle(dev, rpmflags);
> +       }
>
>         spin_lock_irqsave(&dev->power.lock, flags);
>         retval = rpm_resume(dev, rpmflags);
> @@ -1433,6 +1438,7 @@ void pm_runtime_forbid(struct device *dev)
>
>         dev->power.runtime_auto = false;
>         atomic_inc(&dev->power.usage_count);

Analogously here.

> +       trace_rpm_usage_rcuidle(dev, 0);
>         rpm_resume(dev, 0);
>
>   out:
> @@ -1448,16 +1454,17 @@ EXPORT_SYMBOL_GPL(pm_runtime_forbid);
>   */
>  void pm_runtime_allow(struct device *dev)
>  {
> +       bool is_zero;
> +
>         spin_lock_irq(&dev->power.lock);
>         if (dev->power.runtime_auto)
>                 goto out;
>
>         dev->power.runtime_auto = true;
> -       if (atomic_dec_and_test(&dev->power.usage_count))
> +       is_zero = atomic_dec_and_test(&dev->power.usage_count);
> +       trace_rpm_usage_rcuidle(dev, RPM_AUTO | RPM_ASYNC);
> +       if (is_zero)
>                 rpm_idle(dev, RPM_AUTO | RPM_ASYNC);
> -       else
> -               trace_rpm_usage_rcuidle(dev, RPM_AUTO | RPM_ASYNC);

The change of ordering is pointless for the reasons outlined above.

And so on.

> -
>   out:
>         spin_unlock_irq(&dev->power.lock);
>  }
> @@ -1523,9 +1530,8 @@ static void update_autosuspend(struct device *dev, int old_delay, int old_use)
>                 /* If it used to be allowed then prevent it. */
>                 if (!old_use || old_delay >= 0) {
>                         atomic_inc(&dev->power.usage_count);
> -                       rpm_resume(dev, 0);
> -               } else {
>                         trace_rpm_usage_rcuidle(dev, 0);
> +                       rpm_resume(dev, 0);
>                 }
>         }
>
> @@ -1533,8 +1539,10 @@ static void update_autosuspend(struct device *dev, int old_delay, int old_use)
>         else {
>
>                 /* If it used to be prevented then allow it. */
> -               if (old_use && old_delay < 0)
> +               if (old_use && old_delay < 0) {
>                         atomic_dec(&dev->power.usage_count);
> +                       trace_rpm_usage_rcuidle(dev, 0);
> +               }
>
>                 /* Maybe we can autosuspend now. */
>                 rpm_idle(dev, RPM_AUTO);
> @@ -1741,12 +1749,14 @@ void pm_runtime_drop_link(struct device *dev)
>  void pm_runtime_get_noresume(struct device *dev)
>  {
>         atomic_inc(&dev->power.usage_count);
> +       trace_rpm_usage_rcuidle(dev, 0);
>  }

This actually kind of makes sense, as a matter of tracing the
pm_runtime_get_noresume() usage, but not as a matter of tracing the
atomic variable value.

>  EXPORT_SYMBOL_GPL(pm_runtime_get_noresume);
>
>  void pm_runtime_put_noidle(struct device *dev)
>  {
>         atomic_add_unless(&dev->power.usage_count, -1, 0);
> +       trace_rpm_usage_rcuidle(dev, 0);
>  }
>  EXPORT_SYMBOL_GPL(pm_runtime_put_noidle);
>
> --
Chen Yu July 16, 2020, 1:36 a.m. UTC | #7
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
Chen Yu July 16, 2020, 1:49 a.m. UTC | #8
On Wed, Jul 15, 2020 at 05:47:36PM +0200, Rafael J. Wysocki wrote:
> On Wed, Jul 15, 2020 at 8:26 AM Chen Yu <yu.c.chen@intel.com> 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);
> 
> It looks like you could move the trace event before the atomic variable check.
> 
> The ordering between the two doesn't matter, because usage_count may
> change between the check and the trace event anyway.
> 
> But then what is the trace event useful for in the first place?
>
Thanks for explanation, I've changed my mind, it seems that we should not
trace the counter because we don't know who the operator is due to race condition.
> > +               if (non_zero)
> >                         return 0;
> > -               }
> >         }
> >
> >         might_sleep_if(!(rpmflags & RPM_ASYNC) && !dev->power.irq_safe);
> > @@ -1038,10 +1039,12 @@ 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)) {
> > -                       trace_rpm_usage_rcuidle(dev, rpmflags);
> > +               bool non_zero = !atomic_dec_and_test(&dev->power.usage_count);
> > +
> > +               trace_rpm_usage_rcuidle(dev, rpmflags);
> 
> And the same comments apply here.
> 
> > +               if (non_zero)
> >                         return 0;
> > -               }
> > +
> >         }
> >
> >         might_sleep_if(!(rpmflags & RPM_ASYNC) && !dev->power.irq_safe);
> > @@ -1073,8 +1076,10 @@ int __pm_runtime_resume(struct device *dev, int rpmflags)
> >         might_sleep_if(!(rpmflags & RPM_ASYNC) && !dev->power.irq_safe &&
> >                         dev->power.runtime_status != RPM_ACTIVE);
> >
> > -       if (rpmflags & RPM_GET_PUT)
> > +       if (rpmflags & RPM_GET_PUT) {
> >                 atomic_inc(&dev->power.usage_count);
> 
> So the reason why things like that don't work is because the atomic
> variable can change again between the inc and the trace event.
> 
Got it.
> > +               trace_rpm_usage_rcuidle(dev, rpmflags);
> > +       }
> >
> >         spin_lock_irqsave(&dev->power.lock, flags);
> >         retval = rpm_resume(dev, rpmflags);
> > @@ -1433,6 +1438,7 @@ void pm_runtime_forbid(struct device *dev)
> >
> >         dev->power.runtime_auto = false;
> >         atomic_inc(&dev->power.usage_count);
> 
> Analogously here.
> 
> > +       trace_rpm_usage_rcuidle(dev, 0);
> >         rpm_resume(dev, 0);
> >
> >   out:
> > @@ -1448,16 +1454,17 @@ EXPORT_SYMBOL_GPL(pm_runtime_forbid);
> >   */
> >  void pm_runtime_allow(struct device *dev)
> >  {
> > +       bool is_zero;
> > +
> >         spin_lock_irq(&dev->power.lock);
> >         if (dev->power.runtime_auto)
> >                 goto out;
> >
> >         dev->power.runtime_auto = true;
> > -       if (atomic_dec_and_test(&dev->power.usage_count))
> > +       is_zero = atomic_dec_and_test(&dev->power.usage_count);
> > +       trace_rpm_usage_rcuidle(dev, RPM_AUTO | RPM_ASYNC);
> > +       if (is_zero)
> >                 rpm_idle(dev, RPM_AUTO | RPM_ASYNC);
> > -       else
> > -               trace_rpm_usage_rcuidle(dev, RPM_AUTO | RPM_ASYNC);
> 
> The change of ordering is pointless for the reasons outlined above.
> 
> And so on.
> 
> > -
> >   out:
> >         spin_unlock_irq(&dev->power.lock);
> >  }
> > @@ -1523,9 +1530,8 @@ static void update_autosuspend(struct device *dev, int old_delay, int old_use)
> >                 /* If it used to be allowed then prevent it. */
> >                 if (!old_use || old_delay >= 0) {
> >                         atomic_inc(&dev->power.usage_count);
> > -                       rpm_resume(dev, 0);
> > -               } else {
> >                         trace_rpm_usage_rcuidle(dev, 0);
> > +                       rpm_resume(dev, 0);
> >                 }
> >         }
> >
> > @@ -1533,8 +1539,10 @@ static void update_autosuspend(struct device *dev, int old_delay, int old_use)
> >         else {
> >
> >                 /* If it used to be prevented then allow it. */
> > -               if (old_use && old_delay < 0)
> > +               if (old_use && old_delay < 0) {
> >                         atomic_dec(&dev->power.usage_count);
> > +                       trace_rpm_usage_rcuidle(dev, 0);
> > +               }
> >
> >                 /* Maybe we can autosuspend now. */
> >                 rpm_idle(dev, RPM_AUTO);
> > @@ -1741,12 +1749,14 @@ void pm_runtime_drop_link(struct device *dev)
> >  void pm_runtime_get_noresume(struct device *dev)
> >  {
> >         atomic_inc(&dev->power.usage_count);
> > +       trace_rpm_usage_rcuidle(dev, 0);
> >  }
> 
> This actually kind of makes sense, as a matter of tracing the
> pm_runtime_get_noresume() usage, but not as a matter of tracing the
> atomic variable value.
> 
Okay, I'll re-iterate the code and re-think about it on how to
track the runtime process more robustly.

Thanks,
Chenyu
> >  EXPORT_SYMBOL_GPL(pm_runtime_get_noresume);
> >
> >  void pm_runtime_put_noidle(struct device *dev)
> >  {
> >         atomic_add_unless(&dev->power.usage_count, -1, 0);
> > +       trace_rpm_usage_rcuidle(dev, 0);
> >  }
> >  EXPORT_SYMBOL_GPL(pm_runtime_put_noidle);
> >
> > --
diff mbox series

Patch

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);
+		if (non_zero)
 			return 0;
-		}
 	}
 
 	might_sleep_if(!(rpmflags & RPM_ASYNC) && !dev->power.irq_safe);
@@ -1038,10 +1039,12 @@  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)) {
-			trace_rpm_usage_rcuidle(dev, rpmflags);
+		bool non_zero = !atomic_dec_and_test(&dev->power.usage_count);
+
+		trace_rpm_usage_rcuidle(dev, rpmflags);
+		if (non_zero)
 			return 0;
-		}
+
 	}
 
 	might_sleep_if(!(rpmflags & RPM_ASYNC) && !dev->power.irq_safe);
@@ -1073,8 +1076,10 @@  int __pm_runtime_resume(struct device *dev, int rpmflags)
 	might_sleep_if(!(rpmflags & RPM_ASYNC) && !dev->power.irq_safe &&
 			dev->power.runtime_status != RPM_ACTIVE);
 
-	if (rpmflags & RPM_GET_PUT)
+	if (rpmflags & RPM_GET_PUT) {
 		atomic_inc(&dev->power.usage_count);
+		trace_rpm_usage_rcuidle(dev, rpmflags);
+	}
 
 	spin_lock_irqsave(&dev->power.lock, flags);
 	retval = rpm_resume(dev, rpmflags);
@@ -1433,6 +1438,7 @@  void pm_runtime_forbid(struct device *dev)
 
 	dev->power.runtime_auto = false;
 	atomic_inc(&dev->power.usage_count);
+	trace_rpm_usage_rcuidle(dev, 0);
 	rpm_resume(dev, 0);
 
  out:
@@ -1448,16 +1454,17 @@  EXPORT_SYMBOL_GPL(pm_runtime_forbid);
  */
 void pm_runtime_allow(struct device *dev)
 {
+	bool is_zero;
+
 	spin_lock_irq(&dev->power.lock);
 	if (dev->power.runtime_auto)
 		goto out;
 
 	dev->power.runtime_auto = true;
-	if (atomic_dec_and_test(&dev->power.usage_count))
+	is_zero = atomic_dec_and_test(&dev->power.usage_count);
+	trace_rpm_usage_rcuidle(dev, RPM_AUTO | RPM_ASYNC);
+	if (is_zero)
 		rpm_idle(dev, RPM_AUTO | RPM_ASYNC);
-	else
-		trace_rpm_usage_rcuidle(dev, RPM_AUTO | RPM_ASYNC);
-
  out:
 	spin_unlock_irq(&dev->power.lock);
 }
@@ -1523,9 +1530,8 @@  static void update_autosuspend(struct device *dev, int old_delay, int old_use)
 		/* If it used to be allowed then prevent it. */
 		if (!old_use || old_delay >= 0) {
 			atomic_inc(&dev->power.usage_count);
-			rpm_resume(dev, 0);
-		} else {
 			trace_rpm_usage_rcuidle(dev, 0);
+			rpm_resume(dev, 0);
 		}
 	}
 
@@ -1533,8 +1539,10 @@  static void update_autosuspend(struct device *dev, int old_delay, int old_use)
 	else {
 
 		/* If it used to be prevented then allow it. */
-		if (old_use && old_delay < 0)
+		if (old_use && old_delay < 0) {
 			atomic_dec(&dev->power.usage_count);
+			trace_rpm_usage_rcuidle(dev, 0);
+		}
 
 		/* Maybe we can autosuspend now. */
 		rpm_idle(dev, RPM_AUTO);
@@ -1741,12 +1749,14 @@  void pm_runtime_drop_link(struct device *dev)
 void pm_runtime_get_noresume(struct device *dev)
 {
 	atomic_inc(&dev->power.usage_count);
+	trace_rpm_usage_rcuidle(dev, 0);
 }
 EXPORT_SYMBOL_GPL(pm_runtime_get_noresume);
 
 void pm_runtime_put_noidle(struct device *dev)
 {
 	atomic_add_unless(&dev->power.usage_count, -1, 0);
+	trace_rpm_usage_rcuidle(dev, 0);
 }
 EXPORT_SYMBOL_GPL(pm_runtime_put_noidle);