Hi Thomas, sorry for the delay, but I couldn't find time any earlier. On Thu, Nov 28, 2013 at 08:07:10PM +0100, Thomas Gleixner wrote: > On Thu, 28 Nov 2013, Sören Brinkmann wrote: > > On Thu, Nov 28, 2013 at 03:18:50PM +0100, Thomas Gleixner wrote: > > > Now the problem with this device is that it is not a per cpu > > > device. It's a global device, so this update can conflict with a > > > parallel access on the other CPU. Now the disable_irq() only prevents > > > that the other CPU can handle a device interrupt from that timer. But > > > it does not prevent any parallel access from e.g. the idle code path > > > which will try to reprogram it. > > > > Does that mean interrupts need to be disabled globally? Also, does the > > Globally disabling interrupts is not going to work, except you want to > use stomp_machine(). But that would be overkill. > > > cpuidle path depend on interrupts or can it interfere no matter what? > > It can interfere no matter what. The broadcast is modified for the cpu > which loses its per cpu timer due to the idle state via > > clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_ENTER, ...); > > > > Soren, is that timer used as the broadcast device ? > > > > Yes, this is the only broadcast capable timer on Zynq, AFAIK. Other than > > the TTC we only have the arm_global_timer and smp_twd timers, which both > > are per_cpu devices and thus not broadcast capable. > > There is a solution to this. We can identify the broadcast device in > the core and serialize all callers including interrupts on a different > cpu against the update. So no need for the disable/enable_irq() dance. IIUC, and please correct me if I'm wrong, with the patch I'd simply call 'clockevents_update_freq() without having to disable IRQs. But I'm not sure whether periodic mode is covered. I found, that I had to reprogram the timer interval in my clock notifier callback when the timer frequency changes. I think 'clockevents_update_freq()' only handles oneshot mode. For that reason I call 'ttc_set_interval()' in the clock notifier in case the timer is in periodic mode. For that call we'd still have possible races. I guess the best solution would be to move that functionality into 'clockevents_update_freq()'? Sören -- To unsubscribe from this list: send the line "unsubscribe devicetree" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
