Re: [PATCH] PM / Runtime: honor device autosuspend in pm_generic_runtime_idle()

[Mika Westerberg <mika.westerberg@linux.intel.com>]

On Tue, May 28, 2013 at 02:53:47PM +0200, Rafael J. Wysocki wrote:
> On Monday, May 13, 2013 01:50:18 PM Rafael J. Wysocki wrote:
> > On Monday, May 13, 2013 02:05:27 PM Mika Westerberg wrote:
> > > If the device is using autosuspend we should honor that and call
> > > pm_runtime_autosuspend() instead of pm_runtime_suspend(). Failing to do so
> > > causes the device to be suspended immediately even though it expects to be
> > > suspended only when the autosuspend delay is expired.
> > > 
> > > Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
> > > ---
> > >  drivers/base/power/generic_ops.c |    5 ++++-
> > >  1 file changed, 4 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/drivers/base/power/generic_ops.c b/drivers/base/power/generic_ops.c
> > > index bfd898b..19786ca 100644
> > > --- a/drivers/base/power/generic_ops.c
> > > +++ b/drivers/base/power/generic_ops.c
> > > @@ -29,7 +29,10 @@ int pm_generic_runtime_idle(struct device *dev)
> > >  			return ret;
> > >  	}
> > >  
> > > -	pm_runtime_suspend(dev);
> > > +	if (dev->power.use_autosuspend)
> > > +		pm_runtime_autosuspend(dev);
> > > +	else
> > > +		pm_runtime_suspend(dev);
> > >  	return 0;
> > 
> > First of all, this is racy (power.use_autosuspend shoud be accessed under
> > power.lock).
> > 
> > Second, this is not the only place we'd need to make this change (the analogous
> > function for PCI is one example, but there may be others).
> > 
> > Finally, I'm not sure how to address this problem in general.  It may be better
> > to simply modify rpm_idle() and remove pm_generic_runtime_idle() etc. entirely.
> > 
> > I'll have a look at that, thanks for pointing out the problem.
> 
> I'm not sure if the core is the right place to address this, because it's
> not entirely clear if all drivers using autosuspend will have the same policy
> with respect to pm_runtime_idle() (i.e. to avoid suspending immediately if
> the suspend delay timer is active).
> 
> In my opinion it'd be better to address that in the driver by adding a
> .runtime_idle() callback executing pm_runtime_autosuspend(dev) and returning
> -EBUSY.

Makes sense. I did a quick test with following patch and it seems to fix
the issue for i2c-designware driver.

diff --git a/drivers/i2c/busses/i2c-designware-platdrv.c b/drivers/i2c/busses/i2c-designware-platdrv.c
index 966a554..af8be8f 100644
--- a/drivers/i2c/busses/i2c-designware-platdrv.c
+++ b/drivers/i2c/busses/i2c-designware-platdrv.c
@@ -267,7 +267,22 @@ static int dw_i2c_resume(struct device *dev)
 }
 #endif
 
-static SIMPLE_DEV_PM_OPS(dw_i2c_dev_pm_ops, dw_i2c_suspend, dw_i2c_resume);
+#ifdef CONFIG_PM_RUNTIME
+static int dw_i2c_runtime_idle(struct device *dev)
+{
+	/*
+	 * Always schedule autosuspend regardless of what runtime PM the
+	 * client devices use.
+	 */
+	pm_runtime_autosuspend(dev);
+	return -EBUSY;
+}
+#endif
+
+static const struct dev_pm_ops dw_i2c_dev_pm_ops = {
+	SET_SYSTEM_SLEEP_PM_OPS(dw_i2c_suspend, dw_i2c_resume)
+	SET_RUNTIME_PM_OPS(NULL, NULL, dw_i2c_runtime_idle)
+};
 
 /* work with hotplug and coldplug */
 MODULE_ALIAS("platform:i2c_designware");