Re: [PATCH 2/2] sched/idle: Use spin loop primitives for polling idle

[Nicholas Piggin]

On Fri, 1 Sep 2017 14:24:54 +0200
Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:

> On Sun, Aug 20, 2017 at 07:25:02PM +1000, Nicholas Piggin wrote:
> > Signed-off-by: Nicholas Piggin <npiggin@xxxxxxxxx>
> > ---
> >  kernel/sched/idle.c | 7 ++++++-
> >  1 file changed, 6 insertions(+), 1 deletion(-)
> > 
> > diff --git a/kernel/sched/idle.c b/kernel/sched/idle.c
> > index 6c23e30c0e5c..b884980da8ef 100644
> > --- a/kernel/sched/idle.c
> > +++ b/kernel/sched/idle.c
> > @@ -11,6 +11,7 @@
> >  #include <linux/stackprotector.h>
> >  #include <linux/suspend.h>
> >  #include <linux/livepatch.h>
> > +#include <linux/processor.h>
> >  
> >  #include <asm/tlb.h>
> >  
> > @@ -64,9 +65,13 @@ static noinline int __cpuidle cpu_idle_poll(void)
> >  	trace_cpu_idle_rcuidle(0, smp_processor_id());
> >  	local_irq_enable();
> >  	stop_critical_timings();
> > +
> > +	spin_begin();
> >  	while (!tif_need_resched() &&
> >  		(cpu_idle_force_poll || tick_check_broadcast_expired()))
> > -		cpu_relax();
> > +		spin_cpu_relax();
> > +	spin_end();  
> 
> Do we want at least one tif_need_resched() check before we drop into low
> prio mode?

Well we've already done one to get here. My thinking is once we decide
to go idle, get there as fast as we can and avoid too much icache and
branches etc as we're going down. Admittedly the existing cpu_relax code
does a check first, but the difference there is that it can make such a
check without adding extra code. It's more difficult for spin_begin/spin_end
to do the same, so I don't want to add extra junk that other CPUs don't need.
I don't think it will hurt powerpc.

Thanks,
Nick