Re: nmi_watchdog fix for x86_64 to be more like i386

[Thomas Gleixner <tglx@linutronix.de>]

On Mon, 1 Oct 2007, Andi Kleen wrote:
> 
> > IRQ_NOBALANCING is not preventing cpu unplug. It moves the affinity to the
> > next CPU, but the check in NMI watchdog for CPU == 0 would not longer
> > work.
> 
> That cannot happen right now because cpu_disable() on both i386/x86-64
> reject CPU #0. So just setting IRQ_NOBALANCING is sufficient and both
> do that already. I was wrong earlier in being concerned about this.
> 
> >  int tick_do_broadcast(cpumask_t mask)
> > @@ -137,6 +147,7 @@ int tick_do_broadcast(cpumask_t mask)
> >  		cpu_clear(cpu, mask);
> >  		td = &per_cpu(tick_cpu_device, cpu);
> >  		td->evtdev->event_handler(td->evtdev);
> > +		tick_broadcast_account(cpu);
> 
> That would not handle the case with a single CPU running only
> irq  0 but not broadcasting I think.

Hmm. The only situation where this can happen is when you add
"nolapic_timer" to the command line on a single CPU system. We do not
register the lapic "dummy" clock event device then.

> I believe 
> ftp://ftp.firstfloor.org/pub/ak/x86_64/quilt/patches/fix-watchdog
> is the correct fix

Yup, I completely missed the fact, that we reject CPU#0 unplugging, so
your fix seems indeed to be more correct and simpler.

OTOH, the accounting hook would allow us to remove the IRQ#0 -> CPU#0
restriction. Not sure whether it's worth the trouble.

     tglx