Re: [PATCH v11 8/8] timers: Exclude isolated cpus from timer migration

[Frederic Weisbecker <frederic@kernel.org>]

Le Tue, Sep 02, 2025 at 01:08:25PM +0200, Gabriele Monaco a écrit :
> On Mon, 2025-09-01 at 23:11 +0200, Frederic Weisbecker wrote:
> > Le Mon, Sep 01, 2025 at 03:48:15PM +0200, Gabriele Monaco a écrit :
> > > On Mon, 2025-09-01 at 14:41 +0200, Frederic Weisbecker wrote:
> > > > Why not evaluate tick_nohz_cpu_hotpluggable() from
> > > > tmigr_clear_cpu_available() instead of this force IPI?
> > > 
> > > The idea is that this IPI runs once during late boot only for the
> > > tick CPU, while the call to tick_nohz_cpu_hotpluggable() would be
> > > running at every hotplug event if I move it to
> > > tmigr_clear_cpu_available. In that scenario, it's guaranteed to
> > > return true (besides the very first call).
> > > 
> > > I don't have a strong opinion against running that check every time
> > > although it's needed only at boot time and remove this IPI, but in
> > > my understanding that's one of the thing Thomas was finding
> > > confusing [1].
> > > 
> > > Am I missing anything here?
> > 
> > Right, Thomas didn't like it, but the organization of the code has
> > changed a bit since then with the late initcall. If the best we can
> > do to workaround the situation is to make the CPU unavailable
> > regardless and then undo that right after with an IPI, then it's a
> > good sign that we should just simplify and eventually check
> > tick_nohz_cpu_hotpluggable() from tmigr_is_isolated().
> 
> Makes sense.
> I'd be tempted using a static branch but since the call to
> tick_nohz_cpu_hotpluggable() isn't really heavy, we can just be fine
> including it in the tmigr_is_isolated() check.

Right.

> 
> 
> > > > But if I understand correctly, this will be handled by cpuset,
> > > > right?
> > > 
> > > Currently tick_nohz_cpu_hotpluggable() is called by
> > > tmigr_should_isolate_cpu() and that is called by cpuset code,
> > > changing cpuset would save that call but won't deal with the tick
> > > CPU not enabled at boot time, unless I'm misunderstanding what
> > > Waiman implied.
> > 
> > Good point!
> 
> Here I'm a bit unsure how to proceed though. We want to fail any single
> isolated cpuset that includes the tick CPU under nohz_full. I can do it
> directly in isolcpus_nohz_conflict and that looks easy.
> 
> But is that going to be clear for the user?
> Can the user even know what the tick CPU is? Besides /assuming/ 0.

You're right the user can't know in advance which CPU is the tick.
I don't mind if we prevent or not cpuset from allowing to isolate
the timekeeper but either way, a pr_info() could be helpful to tell
that either:

* The isolated timekeeper will have limited isolation
or
* The timekeeper can't be isolated.

Thanks.

> 
> Thanks,
> Gabriele
> 
> > Thanks.
> > 
> > > 
> > > Thanks,
> > > Gabriele
> > > 
> > > [1] - https://lore.kernel.org/lkml/875xgqqrel.ffs@tglx
> > > 
> 

-- 
Frederic Weisbecker
SUSE Labs