[PATCHv2 1/4] clockevents: Add generic timer broadcast receiver

[Thomas Gleixner]

On Mon, 14 Jan 2013, Mark Rutland wrote:
> On Mon, Jan 14, 2013 at 11:50:55AM +0000, Thomas Gleixner wrote:
> > On Mon, 14 Jan 2013, Mark Rutland wrote:
> > 
> > > On Mon, Jan 14, 2013 at 11:06:31AM +0000, Thomas Gleixner wrote:
> > > > On Wed, 9 Jan 2013, Mark Rutland wrote:
> > > > > +#ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
> > > > > +extern int tick_receive_broadcast(void);
> > > > > +#else
> > > > > +static inline int tick_receive_broadcast(void)
> > > > > +{
> > > > > +	return 0;
> > > > > +}
> > > > 
> > > > What's the inline function for? If an arch does not have broadcasting
> > > > support it should not have a receive broadcast function call either.
> > > 
> > > That was how this was originally structured [1], but Santosh suggested this
> > > would break the build for !GENERIC_CLOCKEVENTS_BROADCAST [1]. It means that the
> > > arch-specific receive path (i.e. IPI handler) doesn't have to be #ifdef'd,
> > > which makes it less ugly.
> > 
> > Hmm. If you want to keep the IPI around unconditionally the inline
> > makes some sense, though the question is whether keeping an unused IPI
> > around makes sense in the first place. I'd rather see a warning that
> > an unexpected IPI happened than a silent inline function being called.
> 
> How about I add a warning (e.g. "Impossible timer broadcast received.") and
> return -EOPNOTSUPP when !GENERIC_CLOCKEVENTS_BROADCAST?

You still need to do something with the return value in the arch IPI
code, right?

> > > > Is anything going to use the return value?
> > > 
> > > I'd added this after looking at the x86 lapic timers, where interrupts might
> > > remain pending over a kexec, and lapic interrupts come up before timers are
> > > registered. The return value is useful for shutting down the timer in that case
> > > (see x86's local_apic_timer_interrupt).
> > 
> > Right, though then you need to check for evt->event_handler as well.
> 
> I thought this previously also [1], but I couldn't find any path such that a
> tick_cpu_device would have an evtdev without an event_handler. We always set the
> handler before setting evtdev, and alway wipe evtdev before wiping the handler.
> 
> Have I missed something?

That's an x86 specific issue. Though we could try and make that
functionality completely generic.

Thanks,

	tglx