[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-Id: <D1KDV2ZCWZF5.29STW6VA0LPX8@gmail.com>
Date: Mon, 27 May 2024 21:16:57 +1000
From: "Nicholas Piggin" <npiggin@...il.com>
To: "Thomas Gleixner" <tglx@...utronix.de>, "Oleg Nesterov"
<oleg@...hat.com>, "Frederic Weisbecker" <frederic@...nel.org>
Cc: "Ingo Molnar" <mingo@...hat.com>, "Peter Zijlstra"
<peterz@...radead.org>, "Phil Auld" <pauld@...hat.com>, "Chris von
Recklinghausen" <crecklin@...hat.com>, <linux-kernel@...r.kernel.org>
Subject: Re: sched/isolation: tick_take_do_timer_from_boot() calls
smp_call_function_single() with irqs disabled
On Mon May 27, 2024 at 8:23 PM AEST, Thomas Gleixner wrote:
> On Mon, May 27 2024 at 19:10, Nicholas Piggin wrote:
> > On Sat May 25, 2024 at 8:06 AM AEST, Thomas Gleixner wrote:
> >> along with the removal of the SMP function call voodoo programming gunk,
> >> a lengthy changelog and a bunch of useful comments.
> >
> > I might not have tested that path on powerpc since it should not
> > switch clockevent driver (or clocksource either I think) at least
> > on 64-bit. Explains the smp_call_function warning if you are
> > testing on x86 :/
>
> Even on PowerPC the per CPU clockevent driver is registered during early
> boot for the boot CPU and for the APs when they are onlined. Before your
> change the boot CPU was unconditionally taking over the do_timer duty
> and never gave up on it in the NOHZ full case.
>
> The logic you added allows that the duty is taken by the first
> housekeeping CPU in the case that the boot CPU is marked NOHZ full.
>
> So yes, that function call _is_ invoked on PowerPC too if the boot CPU
> is NOHZ full. There is absolutely nothing x86 specific.
Okay, right you are. Indeed it does trigger.
> The difference is that PowerPC registers the per CPU clockevent _before_
> setting the CPU online and x86 does it afterwards.
>
> So the warning does not trigger on PowerPC because:
>
> WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled() && ...);
That explains it.
Thanks,
Nick
Powered by blists - more mailing lists