Re: [PATCH 00/10] timers/cpuidle: Fixes and cleanups

[Frederic Weisbecker]

On Tue, Aug 15, 2023 at 06:10:43PM +0200, Peter Zijlstra wrote:
> On Fri, Aug 11, 2023 at 07:00:39PM +0200, Frederic Weisbecker wrote:
> > Hi,
> > 
> > The first part of the series is cpuidle callback fixes against timers,
> > some of which haven't been Signed by Peter yet.
> > 
> > Another part is removing the overhead of useless TIF_NR_POLLING clearing.
> 
> So I've again forgotten why we don't simply set TIF_NEED_RESCHED if we
> need the timer re-programmed. That is by far the simplest fix.
> 
> I'm sure we talked about it, but this was a long time ago and I can't
> remember :/

I don't think we did but the rationale is that with forcing setting
TIF_NEED_RESCHED, you force a needless timer restart (which is then going
to be cancelled shortly after) and a round trip to the scheduler with the
rq lock overhead, etc...

Just for the fun I just tried the following change:

diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index c52c2eba7c73..ec43d135cf65 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -1132,8 +1132,10 @@ static void wake_up_idle_cpu(int cpu)
 {
 	struct rq *rq = cpu_rq(cpu);
 
-	if (cpu == smp_processor_id())
+	if (cpu == smp_processor_id()) {
+		set_tsk_need_resched(current);
 		return;
+	}
 
 	if (set_nr_and_not_polling(rq->idle))
 		smp_send_reschedule(cpu);


Then I computed the average of 100 runs of "make clean -C tools/perf; make -C
tools/perf/" before and after this patch.

I observed an average regression of 1.27% less time spent in C-states.

So this has a measurable impact.

> 
> Anyway, the patches look good, except I think there's a whole bunch of
> architectures that still need fixing. In particular since loongson
> 'borrowed' the whole lot from MIPS, they need an identical fix. But I'm
> sure there's more architectures affected.

MIPS at least yes, I only did a quick check and it seems that most archs
use a "wfi" like instruction. I'll check for others.

Thanks.