Re: [PATCH v4 14/16] timer: Implement the hierarchical pull model

[Frederic Weisbecker]

On Thu, Nov 24, 2022 at 08:47:56AM +0100, Anna-Maria Behnsen wrote:
> On Tue, 15 Nov 2022, Frederic Weisbecker wrote:
> 
> > On Fri, Nov 04, 2022 at 03:57:35PM +0100, Anna-Maria Behnsen wrote:
> > > +static int tmigr_cpu_offline(unsigned int cpu)
> > > +{
> > > +	struct tmigr_cpu *tmc = this_cpu_ptr(&tmigr_cpu);
> > > +
> > > +	raw_spin_lock_irq(&tmc->lock);
> > > +	tmc->online = 0;
> > > +	__tmigr_cpu_deactivate(tmc, KTIME_MAX);
> > 
> > This means that if the CPU is going idle for some time during
> > the hotplug process (ie: at some point between CPUHP_AP_TMIGR_ONLINE
> > and CPUHP_TEARDOWN_CPU), then a global timer may be delayed for that long.
> > 
> > I guess it shouldn't be too bad but worth mentioning...
> > 
> > Although if it happens to be a problem it could be solved with simply allowing
> > tmigr_cpu_deactivate() when !tmc->online.
> 
> The plan was (and I broke it) to let the CPU handle global timers by itself
> as long as timer migration hierarchy is not completely initialized and as
> long as CPU is marked offline in timer migration hierarchy. Otherwise
> global timers might be delayed during this period. The proper way would be
> that tmigr_cpu_deactivate(nextexp) directly returns nextexp if !tmc->online
> and tmigr hierarchy is not in place yet. I will have a deeper look if there
> was a reason why I changed the return to KTIME_MAX...

Well, do you need tmc->online at all?

I guess tmigr_cpu_offline() could be simply removed because when the CPU goes
finally down in do_idle() it does:

	tick_nohz_idle_stop_tick();
	cpuhp_report_idle_dead();

So the tick is expected to be stopped (better check it is) and thus the tmc
should be deactivated.

I guess you just need to call tmigr_cpu_activate() from timers_prepare_cpu().

Or am I missing something along the way?