Re: [PATCH v2] PM / runtime: Rework pm_runtime_force_suspend/resume()

[Rafael J. Wysocki]

On Tuesday, January 9, 2018 7:08:21 AM CET Ulf Hansson wrote:
> On 3 January 2018 at 12:06, Rafael J. Wysocki <rjw@xxxxxxxxxxxxx> wrote:
> > From: Rafael J. Wysocki <rafael.j.wysocki@xxxxxxxxx>
> >
> > One of the limitations of pm_runtime_force_suspend/resume() is that
> > if a parent driver wants to use these functions, all of its child
> > drivers have to do that too because of the parent usage counter
> > manipulations necessary to get the correct state of the parent during
> > system-wide transitions to the working state (system resume).
> 
> I understand your point, however this isn't describing the full story,
> because there are a some more alternatives when
> pm_runtime_force_suspend|resume() works well, when used for parent
> devices. Let me clarify, just to make sure we get this correct.
> 
> 1) If the child device isn't managed by runtime PM at all (it always
> has runtime PM disabled), pm_runtime_force_suspend|resume() can be
> used for the parent device. This works because the runtime PM status
> if the child remains in the default status, RPM_SUSPENDED.
> 
> 2) If, somehow, during system suspend/resume, the driver for the child
> make sure to synchronize the runtime PM status of the child device,
> according to the state of the HW, this should work too.
> 
> I leave it to you how to take this fact into consideration, if and how
> to update the changelog.

OK, let me see.

> > However, that limitation turns out to be artificial, so remove it.
> 
> Agree!
> 
> >
> > Namely, pm_runtime_force_suspend() only needs to update the children
> > counter of its parent (if there's is a parent) when the device can
> > stay in suspend after the subsequent system resume transition, as
> > that counter is correct already otherwise.  Now, if the parent's
> > children counter is not updated, it is not necessary to increment
> > the parent's usage counter in that case any more, as long as the
> > children counters of devices are checked along with their usage
> > counters in order to decide whether or not the devices may be left
> > in suspend after the subsequent system resume transition.
> >
> > Accordingly, modify pm_runtime_force_suspend() to only call
> > pm_runtime_set_suspended() for devices whose usage and children
> > counters are at the "no references" level (the runtime PM status
> > of the device needs to be updated to "suspended" anyway in case
> > this function is called once again for the same device during the
> > transition under way), drop the parent usage counter incrementation
> > from it and update pm_runtime_force_resume() to compensate for these
> > changes.
> >
> > Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@xxxxxxxxx>
> > ---
> >  drivers/base/power/runtime.c |   74 +++++++++++++++++++------------------------
> >  1 file changed, 34 insertions(+), 40 deletions(-)
> >
> > Index: linux-pm/drivers/base/power/runtime.c
> > ===================================================================
> > --- linux-pm.orig/drivers/base/power/runtime.c
> > +++ linux-pm/drivers/base/power/runtime.c
> > @@ -1613,17 +1613,28 @@ void pm_runtime_drop_link(struct device
> >         spin_unlock_irq(&dev->power.lock);
> >  }
> >
> > +static bool pm_runtime_need_not_resume(struct device *dev)
> > +{
> > +       return atomic_read(&dev->power.usage_count) <= 1 &&
> > +               atomic_read(&dev->power.child_count) == 0;
> 
> We should take into account the ignore_children flag here, I think.
> Something like this:
> 
> return atomic_read(&dev->power.usage_count) <= 1 &&
> (atomic_read(&dev->power.child_count) == 0 ||
> dev->power.ignore_children)

The current code doesn't quite take ignore_children into account, however.

Regardless of which way the change is made, there will be one corner case
that is not going to be covered.  First, if ignore_children is taken into
account, some cases in which the current code increments the parent's
usage counter will be treated as "need not resume".  Second, if ignore_children
is ignored, some cases in which the parent's usage counter is not incremented
today will cause the parent to resume after the change.

Frankly, I prefer to ignore ignore_children at least for the time being,
because resuming the parent unnecessarily is not a tragedy (it will likely
suspend shortly anyway), but if it is necessary to resume it and it is not
be resumed, things will visibly break.

So, I'd prefer to leave this patch as is and do a second one adding the
ignore_children check.  Then, if things break due to the second patch, it
can be reverted easily (unless that can be fixed differently).

> > +}
> > +
> >  /**
> >   * pm_runtime_force_suspend - Force a device into suspend state if needed.
> >   * @dev: Device to suspend.
> >   *
> >   * Disable runtime PM so we safely can check the device's runtime PM status and
> > - * if it is active, invoke it's .runtime_suspend callback to bring it into
> > - * suspend state. Keep runtime PM disabled to preserve the state unless we
> > - * encounter errors.
> > + * if it is active, invoke its ->runtime_suspend callback to suspend it and
> > + * change its runtime PM status field to RPM_SUSPENDED.  Also, if the device's
> > + * usage and children counters don't indicate that the device was in use before
> > + * the system-wide transition under way, decrement its parent's children counter
> > + * (if there is a parent).  Keep runtime PM disabled to preserve the state
> > + * unless we encounter errors.
> >   *
> >   * Typically this function may be invoked from a system suspend callback to make
> > - * sure the device is put into low power state.
> > + * sure the device is put into low power state and it should only be used during
> > + * system-wide PM transitions to sleep states.  It assumes that the analogous
> > + * pm_runtime_force_resume() will be used to resume the device.
> >   */
> >  int pm_runtime_force_suspend(struct device *dev)
> >  {
> > @@ -1646,17 +1657,18 @@ int pm_runtime_force_suspend(struct devi
> >                 goto err;
> >
> >         /*
> > -        * Increase the runtime PM usage count for the device's parent, in case
> > -        * when we find the device being used when system suspend was invoked.
> > -        * This informs pm_runtime_force_resume() to resume the parent
> > -        * immediately, which is needed to be able to resume its children,
> > -        * when not deferring the resume to be managed via runtime PM.
> > +        * If the device can stay in suspend after the system-wide transition
> > +        * to the working state that will follow, drop the children counter of
> > +        * its parent, but set its status to RPM_SUSPENDED anyway in case this
> > +        * function will be called again for it in the meantime.
> >          */
> > -       if (dev->parent && atomic_read(&dev->power.usage_count) > 1)
> > -               pm_runtime_get_noresume(dev->parent);
> > +       if (pm_runtime_need_not_resume(dev))
> > +               pm_runtime_set_suspended(dev);
> > +       else
> > +               __update_runtime_status(dev, RPM_SUSPENDED);
> >
> 
> This is clever!
> 
> I recall when working on commit 1d9174fbc55e ("PM / Runtime: Defer
> resuming of the device in pm_runtime_force_resume()"), that I wanted
> to take into account the dev->power.child_count, but couldn't figure
> out exactly how. :-)
> 
> > -       pm_runtime_set_suspended(dev);
> >         return 0;
> > +
> >  err:
> >         pm_runtime_enable(dev);
> >         return ret;
> > @@ -1669,13 +1681,9 @@ EXPORT_SYMBOL_GPL(pm_runtime_force_suspe
> >   *
> >   * Prior invoking this function we expect the user to have brought the device
> >   * into low power state by a call to pm_runtime_force_suspend(). Here we reverse
> > - * those actions and brings the device into full power, if it is expected to be
> > - * used on system resume. To distinguish that, we check whether the runtime PM
> > - * usage count is greater than 1 (the PM core increases the usage count in the
> > - * system PM prepare phase), as that indicates a real user (such as a subsystem,
> > - * driver, userspace, etc.) is using it. If that is the case, the device is
> > - * expected to be used on system resume as well, so then we resume it. In the
> > - * other case, we defer the resume to be managed via runtime PM.
> > + * those actions and bring the device into full power, if it is expected to be
> > + * used on system resume.  In the other case, we defer the resume to be managed
> > + * via runtime PM.
> >   *
> >   * Typically this function may be invoked from a system resume callback.
> >   */
> > @@ -1684,32 +1692,18 @@ int pm_runtime_force_resume(struct devic
> >         int (*callback)(struct device *);
> >         int ret = 0;
> >
> > -       callback = RPM_GET_CALLBACK(dev, runtime_resume);
> > -
> > -       if (!callback) {
> > -               ret = -ENOSYS;
> > -               goto out;
> > -       }
> 
> This make me realize that, I think this is probably for legacy reasons.
> 
> I think we should allow the callback to be NULL, don't you think?

Well, it probably doesn't matter too much, but OK.

> Of course, this deserves its own change.

Right.

> > -
> > -       if (!pm_runtime_status_suspended(dev))
> > +       if (!pm_runtime_status_suspended(dev) || pm_runtime_need_not_resume(dev))
> >                 goto out;
> >
> >         /*
> > -        * Decrease the parent's runtime PM usage count, if we increased it
> > -        * during system suspend in pm_runtime_force_suspend().
> > -       */
> > -       if (atomic_read(&dev->power.usage_count) > 1) {
> > -               if (dev->parent)
> > -                       pm_runtime_put_noidle(dev->parent);
> > -       } else {
> > -               goto out;
> > -       }
> > +        * The value of the parent's children counter is correct already, so
> > +        * just update the status of the device.
> > +        */
> > +       __update_runtime_status(dev, RPM_ACTIVE);
> >
> > -       ret = pm_runtime_set_active(dev);
> > -       if (ret)
> > -               goto out;
> > +       callback = RPM_GET_CALLBACK(dev, runtime_resume);
> >
> > -       ret = callback(dev);
> > +       ret = callback ? callback(dev) : -ENOSYS;
> >         if (ret) {
> >                 pm_runtime_set_suspended(dev);
> >                 goto out;
> >
> 
> Thanks for looking into this!

No problem.

Thanks,
Rafael