Re: [PATCH] clockevents: Leave broadcast device shtudown only if the current device is always running.

[Suresh Siddha <suresh.b.siddha@intel.com>]

On Sun, 2012-04-15 at 18:27 +0530, Santosh Shilimkar wrote:
> On Tuesday 10 April 2012 04:11 AM, Suresh Siddha wrote:
> > @@ -575,10 +575,12 @@ void tick_broadcast_switch_to_oneshot(void)
> >  	unsigned long flags;
> >  
> >  	raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
> > +
> > +	tick_broadcast_device.mode = TICKDEV_MODE_ONESHOT;
> > +
> >  	if (cpumask_empty(tick_get_broadcast_mask()))
> >  		goto end;
> >  
> > -	tick_broadcast_device.mode = TICKDEV_MODE_ONESHOT;
> >  	bc = tick_broadcast_device.evtdev;
> >  	if (bc)
> >  		tick_broadcast_setup_oneshot(bc);
> > 
> > 
> Last time when I tried your patch, some how he ethernet IRQ masked
> the issue.
> 
> The problem is still as before. If you make a minimal kernel
> configuration and possibly have only local timer, broadcast timer
> and say console device as CPU wakeup sources, you should hang in
> the boot process itself(Assuming CPUIDLE driver is built-in)
> 
> The main issue is, you are not letting the one time setup
> needed for the broadcast device. Even with above change,
> the 'tick_broadcast_device.evtdev' is still in SHUTDOWN
> mode and the event handler is pointing to clockevents_handle_noop()

On x86, idle driver (like drivers/idle/intel_idle.c:
__setup_broadcast_timer()) calls clockevents_notify() with
CLOCK_EVT_NOTIFY_BROADCAST_ON, during which we do
tick_broadcast_setup_oneshot(). See tick_do_broadcast_on_off() which
does this. That should setup the evtdev handler, mode etc.

And during the next CLOCK_EVT_NOTIFY_BROADCAST_ENTER we program the next
broadcast event.

> So the very first 'CLOCK_EVT_NOTIFY_BROADCAST_ENTER' call won't
> program the broadcast device with next event and CPU won't wakeup
> from low power. if you are lucky and have some other wakeup source,
> system will move further and eventually the handler and
> mode of the broad-cast device get set correctly.

I am confused. Can you elaborate on how on the next spurious wakeup
event, broad-cast device handler, mode is set?

It sounds like the CLOCK_EVT_NOTIFY_BROADCAST_ON/OFF notifications are
missing from your idle driver. But I am confused how it will work on the
next spurious wakeup etc with out the calls to
tick_do_broadcast_on_off()? Perhaps the next spurious wakeup correcting
everything is just theory and you didn't see it in practice?

Anyways, can you add CLOCK_EVT_NOTIFY_BROADCAST_ON/OFF notifications in
your idle driver to see if it addresses the problem. That is the correct
thing to do here.

thanks,
suresh