Re: [PATCH] clocksource/drivers/arm_global_timer: Always use {readl|writel}_relaxed

From: Jisheng Zhang
Date: Fri Nov 13 2015 - 07:24:24 EST


Dear Arnd,

On Fri, 13 Nov 2015 11:33:12 +0100
Arnd Bergmann wrote:

> On Friday 13 November 2015 17:59:48 Jisheng Zhang wrote:
> > On Fri, 13 Nov 2015 10:28:01 +0100
> > Arnd Bergmann wrote:
> > > On Friday 13 November 2015 16:40:25 Jisheng Zhang wrote:
> > > > On Fri, 13 Nov 2015 16:34:38 +0800
> > > > > diff --git a/drivers/clocksource/arm_global_timer.c b/drivers/clocksource/arm_global_timer.c
> > > > > index a2cb6fa..84a5a5d 100644
> > > > > --- a/drivers/clocksource/arm_global_timer.c
> > > > > +++ b/drivers/clocksource/arm_global_timer.c
> > > > > @@ -99,27 +99,27 @@ static void gt_compare_set(unsigned long delta, int periodic)
> > > > >
> > > > > counter += delta;
> > > > > ctrl = GT_CONTROL_TIMER_ENABLE;
> > > > > - writel(ctrl, gt_base + GT_CONTROL);
> > > > > - writel(lower_32_bits(counter), gt_base + GT_COMP0);
> > > > > - writel(upper_32_bits(counter), gt_base + GT_COMP1);
> > > > > + writel_relaxed(ctrl, gt_base + GT_CONTROL);
> > > > > + writel_relaxed(lower_32_bits(counter), gt_base + GT_COMP0);
> > > > > + writel_relaxed(upper_32_bits(counter), gt_base + GT_COMP1);
> > > > >
> > > > > if (periodic) {
> > > > > - writel(delta, gt_base + GT_AUTO_INC);
> > > > > + writel_relaxed(delta, gt_base + GT_AUTO_INC);
> > > > > ctrl |= GT_CONTROL_AUTO_INC;
> > > > > }
> > > > >
> > > > > ctrl |= GT_CONTROL_COMP_ENABLE | GT_CONTROL_IRQ_ENABLE;
> > > > > - writel(ctrl, gt_base + GT_CONTROL);
> > > > > + writel_relaxed(ctrl, gt_base + GT_CONTROL);
> > > > > }
> > >
> > > This seems fine. Do you have any performance numbers to show how much
> > > we save per call on a platform you care about, and how often it is
> > > called for a typical workload?
> >
> > To be honest, all my platforms don't make use of global timer for clockevent,
> > we use dw_apb_timer and twd or arch_timer instead, but one performance impact
> > I saw in our case can also apply for the case with global timer as clokevent:
> >
> > there are 500-1000 short sleeps, yes not good userspace behavior, so we
> > program clockevent device 500-1000 times/s. If the system is powered by CA9
> > with outer L2 cache, the writel will contend for l2x0_lock for 500-1000 times/s.
> > Then the L2 cache maintenance from other device driver have more chance to
> > spinning at the l2x0_lock, so other device driver performance is impacted.
>
> Just to make sure I get this right: which outer cache implementation do you
> use in this case? Most Cortex-A9 use pl310, which does not require l2x0_lock

PL310

> for outer_cache.sync(). The Aurora outer cache sync has a different method
> and also doesn't use l2x0_lock. Finally, tauros3 doesn't need a cache sync
> at all.
>
> Did you look at an older kernel version? We used to do a loop in the

oops, yes. The kernel version in product still needs the spinlock in sync.
I didn't check the L2 cache code for about 1 year, sorry for that.
If we upgrade to newer kernel version, yes, the bit performance bottleneck --
spinlock contention won't exist anymore. Thanks for pointing out this.

But I think we may still see trivial system performance improvement in 500-1000
times/s of clockevent programming case due to the mb() in writel.

Thanks,
Jisheng

> Aurora cache sync operation until I fixed that, so it should be a bit
> faster now. It will still require doing the actual sync, but at least
> there should not be any lock contention these days.
>
> Arnd

--
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/