On Monday, September 4, 2017 3:21:15 PM CEST Ulf Hansson wrote:

On 2 September 2017 at 17:38, Rafael J. Wysocki [off-list ref] wrote:

quoted

On Friday, September 1, 2017 10:27:05 AM CEST Ulf Hansson wrote:

quoted

On 29 August 2017 at 17:27, Rafael J. Wysocki [off-list ref] wrote:

quoted

On Tuesday, August 29, 2017 4:56:48 PM CEST Ulf Hansson wrote:

quoted

This change enables the ACPI PM domain to cope with drivers that deploys
the runtime PM centric path for system sleep.

[cut]

quoted

@@ -1052,11 +1066,20 @@ EXPORT_SYMBOL_GPL(acpi_subsys_complete);
  * @dev: Device to handle.
  *
  * Follow PCI and resume devices suspended at run time before running their
- * system suspend callbacks.
+ * system suspend callbacks. However, try to avoid it in case the runtime PM
+ * centric path is used for the device and then trust the driver to do the
+ * right thing.
  */
 int acpi_subsys_suspend(struct device *dev)
 {
-     pm_runtime_resume(dev);
+     struct acpi_device *adev = ACPI_COMPANION(dev);
+
+     if (!adev)
+             return 0;
+
+     if (!dev_pm_is_rpm_sleep(dev) || acpi_dev_needs_resume(dev, adev))
+             pm_runtime_resume(dev);
+
      return pm_generic_suspend(dev);
 }
 EXPORT_SYMBOL_GPL(acpi_subsys_suspend);

Well, I tried to avoid calling acpi_dev_needs_resume() for multiple times
and that's why I added the update_state thing.

Moreover, the is_rpm_sleep flag here has to mean not only that
direct_complete should not be used with the device, but also that its driver
is fine with not resuming it.

Let me try to explain this better. I realize the changelog is
misleading around this particular section! Huh, apologize for that!

First, patch1 makes the PM core treat the is_rpm_sleep flag as the
direct_complete isn't allowed for the device.

For that reason, when the is_rpm_sleep is set, there is no point
calling acpi_dev_needs_resume() from acpi_subsys_prepare(), but
instead that can be deferred to acpi_subsys_suspend() - because it
doesn't matter if acpi_subsys_prepare() returns 0 or 1, in either case
the acpi_subsys_suspend() will be called. That's really what goes on
here.

The end result is the same. If the acpi_dev_needs_resume() thinks that
the device needs to be runtime resumed, pm_runtime_resume() is called
for the device in acpi_subsys_suspend().

So, this has nothing to do with whether the driver "is fine with not
resuming it" thing.

No, sorry.

If is_rpm_sleep was not set, the ACPI PM domain would resume the device in
acpi_subsys_suspend() regardless of the acpi_dev_needs_resume() return value.

Yes, I believe I forgot about one scenario, when the direct_complete
path has been abandoned by the PM core, because a child device was
suspend before and it couldn't run the direct_complete path for it?

Just to be sure, that's the case you also had in mind?

Yes.

Thanks,
Rafael