Re: [PATCH 2/2] arm64: dts: mt8173: add timer node

[Yingjoe Chen <yingjoe.chen@mediatek.com>]

On Thu, 2015-10-01 at 16:32 +0100, Sudeep Holla wrote:
> 
> On 01/10/15 15:33, Yingjoe Chen wrote:
> > On Thu, 2015-09-17 at 17:13 +0100, Sudeep Holla wrote:
> >>
> 
> [...]
> 
> >>
> >> I think your are confusing the system counter with arch timers. System
> >> counter is always-on, but the arch timers(logic implementing timers
> >> comparators) might not be off when the processor is powered down.
> >>
> >> I think you need this timer and are using it for low power idle states
> >> in which case you will use this as a clock event and not clock source.
> >> It will be used as a hardware broadcast event source.
> >>
> >> There's no call to sched_clock_register in mtk_timer.c, so it can't be
> >> the sched clock, so you need to fix the commit log.
> >
> > Hi Sudeep,
> >
> > Sorry for late reply.
> >
> > For sched_clock_register, please see
> > http://lists.infradead.org/pipermail/linux-mediatek/2015-July/001547.html
> > which was accepted in
> > https://git.linaro.org/people/daniel.lezcano/linux.git/shortlog/refs/heads/clockevents/4.4
> >
> 
> The commit message makes no sense to me. The counters should continue to
> work as long as they are in always-on domain. Only timers are lost
> when you enter deeper idle states. So I agree with using MTK timer as
> broadcast timer/eventsource. You still didn't answer what's the need
> to use MTK timer as sched clocksource ?


Hi, Sudeep,

ARM ARM said the counter should be in always-on domain, but
unfortunately that not true for mt8173. The last CPU enter idle can
choose to enter deep idle mode and the counter value would be lost. Our
firmware backup/recover the counter so it looks like it is stopped.
That's why I thought we need to use it as sched clocksource.

On mt8173, we will fix the firmware to add missing counts, so it will
looks like the counter keep counting. But other mediatek platform have
similar issue, and the 2 counter have same resolution, so I still want
to keep using GPT as sched clocksource.

Joe.C