Re: [PATCH] clockevents: return error from tick_broadcast_oneshot_control if !GENERIC_CLOCKEVENTS_BROADCAST

[Thomas Gleixner]

On Fri, 26 Jun 2015, Preeti U Murthy wrote:
> On 06/26/2015 01:17 PM, Thomas Gleixner wrote:
> > On Fri, 26 Jun 2015, Preeti U Murthy wrote:
> >> What about the case where GENERIC_CLOCKEVENTS_BROADCAST=y and
> >> TICK_ONESHOT=n (HZ_PERIODIC=y) ? Have you tested this ?
> >>
> >> This will hang the kernel at boot if you are using the hrtimer mode of
> >> broadcast. This is because the local timers of all cpus are shutdown
> >> when the cpuidle driver registers itself, on finding out that there are
> >> idle states where local tick devices stop. The broadcast tick device is
> >> then in charge of waking up the cpus at every period. In hrtimer mode of
> >> broadcast, there is no such real device and we hang.
> > 
> > Hmm, no. tick-broadcast-hrtimer.o depends on TICK_ONESHOT=y. So this
> > is covered by the check for the broadcast device, which is NULL.
> > 
> > But there is another variant:
> > 
> > GENERIC_CLOCKEVENTS_BROADCAST=y TICK_ONESHOT=y and 'highres=off
> > nohz=off' on the kernel command line. 
> 
> Can this happen at all? It is during tick_init_highres() or
> tick_nohz_switch_to_nohz() that we switch to oneshot mode, not otherwise
> AFAICT.

And how does that matter? If 'highres=off nohz=off' is on the kernel
command line none of the switchovers happens. So system stays in
periodic mode and the broadcast hrtimer thing is registered, right?
 
> I was actually talking of the following scenario. In periodic mode,
> where GENERIC_CLOCKEVENTS_BROADCAST=y, the arch can execute
> tick_setup_hrtimer_broadcast(), which will return nothing as you point
> out above. So there is no broadcast clockevent device.
> 
> When the cpuidle driver registers with the cpuidle core however,
> cpuidle_setup_broadcast_timer() on every cpu is executed if it finds
> that there is an idle state where ticks stop.
> 
> cpuidle_setup_broadcast_timer()
>   tick_broadcast_enable()
>     tick_broadcast_control(BROADCAST_ON)
>        bc = tick_broadcast_device.evtdev which is NULL in this case
> 
>          TICK_BROADCAST_ON:
>          checks for periodic mode of the broadcast device - succeeds
>          although we haven't registered a broadcast device because
>          value of TICKDEV_PERIODIC is 0, the default value of td.mode.
> 
>          clockevents_shutdown(dev)
> 
> At this point all cpus stop.

Right. That's a different one, if tick_broadcast_device.evtdev == NULL.

We should not stop any cpu local timer in that case. Combined with the
patch I sent, we prevent the idle stuff from going into a state where
the cpu local timers stop.

Thanks,

	tglx
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/