Re: [PATCH 04/12] ia64: Replace cpu_down with freeze_secondary_cpus

[Qais Yousef]

On 11/19/19 23:59, Thomas Gleixner wrote:
> On Tue, 19 Nov 2019, Qais Yousef wrote:
> > On 11/19/19 23:21, Thomas Gleixner wrote:
> > > On Wed, 30 Oct 2019, Qais Yousef wrote:
> > > >  void machine_shutdown(void)
> > > >  {
> > > >  #ifdef CONFIG_HOTPLUG_CPU
> > > > -	int cpu;
> > > > -
> > > > -	for_each_online_cpu(cpu) {
> > > > -		if (cpu != smp_processor_id())
> > > > -			cpu_down(cpu);
> > > > -	}
> > > > +	/* TODO: Can we use disable_nonboot_cpus()? */
> > > > +	freeze_secondary_cpus(smp_processor_id());
> > > 
> > > freeze_secondary_cpus() is only available for CONFIG_PM_SLEEP_SMP=y and
> > > disable_nonboot_cpus() is a NOOP for CONFIG_PM_SLEEP_SMP=n :)
> > 
> > I thought I replied to this :-(
> > 
> > My plan was to simply make freeze_secondary_cpus() available and protected by
> > CONFIG_SMP only instead.
> > 
> > Good plan?
> 
> No. freeze_secondary_cpus() is really for hibernation. Look at the exit
> conditions there.

Hmm do you mean the pm_wakeup_pending() abort?

In arm64 we machine_shutdown() calls disable_nonboot_cpus(), which in turn
a wrapper around freeze_secondary_cpus() with 0 passed as an argument.

IIUC this means arm64 could fail to offline all CPUs on machine_shutdown(),
correct?

> 
> So you really want a seperate function which depends on CONFIG_HOTPLUG_CPU
> and provides an inline stub for the CONFIG_HOTPLUG_CPU=n case.
> 
> But I have a hard time to see how that stuff works at all on
> ia64:
> 
>   cpu_down() might sleep, i.e. it must be called from preemptible
>   context. smp_processor_id() is invalid from preemtible context.
> 
> As this is obviously broken and ia64 is in keep compile mode only, it
> should just go away.

If arm64 is doing the wrong thing, then we need a new function anyway.

Thanks

--
Qais Yousef