Ulf Hansson [off-list ref] writes:

On 3 March 2017 at 21:41, Lina Iyer [off-list ref] wrote:

quoted

Notify runtime PM when the CPU is going to be powered off in the idle
state. This allows for runtime PM suspend/resume of the CPU as well as
its PM domain.

Cc: Kevin Hilman <khilman@kernel.org>
Cc: Rafael J. Wysocki <redacted>
Signed-off-by: Lina Iyer <redacted>

[...]

quoted

@@ -99,6 +102,7 @@ int cpu_pm_enter(void)
 {
        int nr_calls;
        int ret = 0;
+       struct device *dev = get_cpu_device(smp_processor_id());

        read_lock(&cpu_pm_notifier_lock);
        ret = cpu_pm_notify(CPU_PM_ENTER, -1, &nr_calls);

@@ -110,6 +114,10 @@ int cpu_pm_enter(void)
                cpu_pm_notify(CPU_PM_ENTER_FAILED, nr_calls - 1, NULL);
        read_unlock(&cpu_pm_notifier_lock);

+       /* Notify Runtime PM that we are suspending the CPU */
+       if (!ret && dev)
+               RCU_NONIDLE(pm_runtime_put_sync_suspend(dev));
+

I am trying to understand how the runtime PM usage count becomes
properly balanced.

I believe you could you end up first calling a
pm_runtime_put_sync_suspend(), without earlier having called
pm_runtime_get*(). I am not sure though, but perhaps you can
elaborate.

Anyway, in patch2/9, where you enable runtime PM there is only a call
to pm_runtime_set_active(), which doesn't increase the usage count. To
me, it seems like that change also needs a pm_runtime_get_noresume().

IIUC, the CPU hotplug callback below (cpu_pm_cpu_starting) will do the
first _get_sync(), and I'm assuming that will happen before any of the
CPU PM notifiers get called, so I think the usecount will always be at
least 1 by the time any CPU PM callbacks happen.

[...]

quoted

+#ifdef CONFIG_HOTPLUG_CPU
+static int cpu_pm_cpu_dying(unsigned int cpu)
+{
+       struct device *dev = get_cpu_device(cpu);
+
+       if (dev)
+               pm_runtime_put_sync_suspend(dev);
+
+       return 0;
+}
+
+static int cpu_pm_cpu_starting(unsigned int cpu)
+{
+       struct device *dev = get_cpu_device(cpu);
+
+       if (dev)
+               pm_runtime_get_sync(dev);

I assume that according to my comment above, you somehow need to
compensate for either of the cases when CONFIG_HOTPLUG_CPU is set or
unset. Right?

Right, if for some reason CONFIG_HOTPLUG_CPU=n, we'll have a problem
where there is never an initial _get() so the usecount will be zero when
CPU PM notifiers get called the first time.

Kevin