timers & suspend

[Sören Brinkmann]

On Thu, 2014-07-03 at 07:26PM +0200, Daniel Lezcano wrote:
> On 07/03/2014 06:09 PM, Sören Brinkmann wrote:
> >On Thu, 2014-07-03 at 02:21PM +0200, Daniel Lezcano wrote:
> >>On 06/30/2014 08:39 PM, Sören Brinkmann wrote:
> >>>Hi,
> >>>
> >>>I'm currently working on suspend for Zynq and try to track down some
> >>>spurious wakes. It looks like the spurious wakes are caused by timers,
> >>>hence I was wondering whether there are any special requirements for
> >>>timer drivers when it comes to suspend support or if I just missed
> >>>something.
> >>>
> >>>Zynq sets the 'IRQCHIP_MASK_ON_SUSPEND' flag, which should mask all
> >>>interrupts but the wake source. Reading through kernel/irq/pm.c
> >>>indicates, that timer interrupts get some special treatment though.
> >>>Therefore I implemented some suspend/resume callbacks for the
> >>>cadence_ttc which disable and clear the timer's interrupts when going
> >>>into suspend. That seems to mitigate the issue quite a bit, but I still
> >>>saw spurious wakes - just a lot less often.
> >>>Digging a little deeper revealed, the spurious wakes are caused by the
> >>>ARM's smp_twd timer now. Given that that driver is probably used by a few
> >>>more ARM platforms, I get the feeling that I'm missing something.
> >>
> >>Do you receive any interrupt from the cadence_ttc ? (/proc/interrupts)
> >>
> >>That's funny because I realize that you cadence ttc timer is never
> >>used as there are the architected timers. The cadence ttc would be
> >>only useful if there were an idle state powering down the smp_twd
> >>timers but it is not the case on this board, IIUC.
> >Yes they are used. They TTC is the only broadcast capable timer for
> >Zynq. In my experience, I can not even boot without it (may have
> >dependencies on CPUidle or something).
> 
> Actually the cpuidle driver is wrong. It assumes the state #1 will
> power off the different cores with their architected timers and then
> switch to the broadcast timer. But this one is not needed as we
> don't power down the core with the twd timers, so no need to switch
> to a backup timer device.
> 
> The implementation of the DDR self refresh idle state (incoming
> patchset) removes the cpu_pm notifiers + the flag TIMER_STOP. The
> result is 0 interrupts for ttc cadence timer. I removed in the dts
> the cadence ttc and my board booted without problem (it is a zynq
> 702).
> 
> Except I missed something, the cadence ttc is actually not used at all.
You mean with your patches, or without? I'd have to check that.

> 
> >>>It's probably worth mentioning that the suspend state in Zynq does not
> >>>power off the CPU cores. It just asserts the resets on secondary cores
> >>>and the primary one waits in wfi.
> >>
> >>Why do you need to reset them ?
> >CPU0 is not reset, but all secondary cores are. That is as close to
> >power off as we get and works well with hotplug.
> 
> Ok, I was not aware of this approach. Is there any patchset
> available in the net ?
It's in mainline for a while. arch/arm/mach-zynq/platsmp.c defines the
SMP ops. Zynq's cpu_die/kill result in the CPU reset being asserted. And
when the core is plugged in again, it gets deasserted and goes through
the secondary boot path.
That code is a little cleaned up in our vendor tree, but overall it
should be the same functionality wise.

	Sören