On 01/18/2016 07:42 PM, Alexandre Belloni wrote:

On 18/01/2016 at 18:25:22 +0100, Sebastian Andrzej Siewior wrote :

quoted

* Alexandre Belloni | 2016-01-17 03:23:14 [+0100]:

quoted

index 80d74c4adcbe..43b50634d640 100644

--- a/drivers/clocksource/timer-atmel-pit.c
+++ b/drivers/clocksource/timer-atmel-pit.c

@@ -96,11 +96,44 @@ static int pit_clkevt_shutdown(struct clock_event_device *dev)

	/* disable irq, leaving the clocksource active */
	pit_write(data->base, AT91_PIT_MR, (data->cycle - 1) | AT91_PIT_PITEN);
-	free_irq(atmel_pit_irq, data);
+	if (!clockevent_state_detached(dev))
+		free_irq(data->irq, data);

I did it in the meantime without clockevent_state_detached(). From what
it looks, it first sets the state and then invokes
pit_clkevt_shutdown(). Any particular reason for this?

Yeah, I forgot to mention that. Freeing the irq unconditionally
results in:

…

My understanding is that clockevents_exchange_device() changes the state
from detached to shutdown and so at that point the IRQ has never been
requested.

I see. So we get shutdown called twice while set_periodic was only
called once. In that case I would suggest to have internal bookkeeping
instead of relying on current core's behavior when it is time free the
irq.

Sebastian