Re: Nohz_full on boot CPU is broken (was: Re: [PATCH v2 1/1] wq: Avoid using isolated cpus' timers on queue_delayed_work)

[Frederic Weisbecker]

Le Sun, Apr 07, 2024 at 03:52:48PM +0200, Oleg Nesterov a écrit :
> On 04/07, Oleg Nesterov wrote:
> >
> > On 04/05, Frederic Weisbecker wrote:
> > >
> > > +Cc Nick
> > >
> > > Le Fri, Apr 05, 2024 at 04:04:49PM +0200, Oleg Nesterov a écrit :
> > > > On 04/03, Oleg Nesterov wrote:
> > > > >
> > > > > > > OTOH, Documentation/timers/no_hz.rst says
> > > > > > >
> > > > > > > 	Therefore, the
> > > > > > > 	boot CPU is prohibited from entering adaptive-ticks mode.  Specifying a
> > > > > > > 	"nohz_full=" mask that includes the boot CPU will result in a boot-time
> > > > > > > 	error message, and the boot CPU will be removed from the mask.
> > > > > > >
> > > > > > > and this doesn't match the reality.
> > > > > >
> > > > > > Don't some archs allow the boot CPU to go down too tho? If so, this doesn't
> > > > > > really solve the problem, right?
> > > > >
> > > > > I do not know. But I thought about this too.
> > > > >
> > > > > In the context of this discussion we do not care if the boot CPU goes down.
> > > > > But we need at least one housekeeping CPU after cpu_down(). The comment in
> > > > > cpu_down_maps_locked() says
> > > > >
> > > > > 	Also keep at least one housekeeping cpu onlined
> > > > >
> > > > > but it checks HK_TYPE_DOMAIN, and I do not know (and it is too late for me
> > > > > to try to read the code ;) if housekeeping.cpumasks[HK_TYPE_TIMER] can get
> > > > > empty or not.
> > > >
> > > > This nearly killed me, but I managed to convince myself we shouldn't worry
> > > > about cpu_down().
> > > >
> > > > HK_FLAG_TIMER implies HK_FLAG_TICK.
> > > >
> > > > HK_FLAG_TICK implies tick_nohz_full_setup() which sets
> > > > tick_nohz_full_mask = non_housekeeping_mask.
> > > >
> > > > When tick_setup_device() is called on a housekeeping CPU it does
> > > >
> > > > 	else if (tick_do_timer_boot_cpu != -1 &&
> > > > 					!tick_nohz_full_cpu(cpu)) {
> > > > 		tick_take_do_timer_from_boot();
> > > > 		tick_do_timer_boot_cpu = -1;
> > > >
> > > >
> > > > 	and this sets tick_do_timer_cpu = first-housekeeping-cpu.
> > > >
> > > > cpu_down(tick_do_timer_cpu) will fail, tick_nohz_cpu_down() will nack it.
> > > >
> > > > So cpu_down() can't make housekeeping.cpumasks[HK_FLAG_TIMER] empty and I
> > > > still think that the change below is the right approach.
> > > >
> > > > But probably WARN_ON() in housekeeping_any_cpu() makes sense anyway.
> > > >
> > > > What do you think?
> > >
> > > Good analysis on this nasty housekeeping VS tick code. I promised so many
> > > times to cleanup this mess but things keep piling up.
> > >
> > > It is indeed possible for the boot CPU to be a nohz_full CPU and as
> > > you can see, it's only half-working. This is so ever since:
> > >
> > >     08ae95f4fd3b (nohz_full: Allow the boot CPU to be nohz_full)
> >
> > Thanks... So this is intentional. I was confused by
> >
> > 	Therefore, the
> > 	boot CPU is prohibited from entering adaptive-ticks mode.  Specifying a
> > 	"nohz_full=" mask that includes the boot CPU will result in a boot-time
> > 	error message, and the boot CPU will be removed from the mask.
> >
> > from Documentation/timers/no_hz.rst
> >
> > > I would love
> > > to revert that now but I don't know if anyone uses this and have it working
> > > by chance somewhere... Should we continue to support a broken feature? Can we
> > > break user ABI if it's already half-broken?
> >
> > Well, the changelog says
> >
> >     nohz_full has been trialed at a large supercomputer site and found to
> >     significantly reduce jitter. In order to deploy it in production, they
> >     need CPU0 to be nohz_full
> >
> > so I guess this feature has users.
> >
> > But after the commit aae17ebb53cd3da ("workqueue: Avoid using isolated cpus'
> > timers on queue_delayed_work") the kernel will crash at boot time if the boot
> > CPU is nohz_full.
> >
> > So we need a workaround at least. I am starting to think I will send a trivial
> > patch which changes __queue_delayed_work() to validate the cpu returned by
> > housekeeping_any_cpu(HK_TYPE_TIMER).
> >
> > But perhaps something like below makes more sense as a (stupid) workaround?
> 
> Or simply
> 
> 
> diff --git a/kernel/sched/isolation.c b/kernel/sched/isolation.c
> index 373d42c707bc..e912555c6fc8 100644
> --- a/kernel/sched/isolation.c
> +++ b/kernel/sched/isolation.c
> @@ -46,7 +46,11 @@ int housekeeping_any_cpu(enum hk_type type)
>  			if (cpu < nr_cpu_ids)
>  				return cpu;
>  
> -			return cpumask_any_and(housekeeping.cpumasks[type], cpu_online_mask);
> +			cpu = cpumask_any_and(housekeeping.cpumasks[type], cpu_online_mask);
> +			if (cpu < nr_cpu_ids)
> +				return cpu;
> +
> +			WARN_ON_ONCE(system_state == SYSTEM_RUNNING);
>  		}
>  	}
>  	return smp_processor_id();

Yep, looks good!

Thanks.