Re: [PATCH 2/2] x86: introduce NMI_AUTO as nmi_watchdog option

[Ingo Molnar <mingo@elte.hu>]

* Mikael Pettersson <mikpe@it.uu.se> wrote:

> Ingo Molnar writes:
>  > > but that would work only for LAPIC. You're suggesting killing IOAPIC 
>  > > mode too?
>  > 
>  > Would it be a big loss, with all modern systems expected to have a 
>  > working lapic based NMI source?
> 
> The IOAPIC NMI uses a HW resource that is free and reliable, the LAPIC 
> NMI uses a HW resource that is neither free nor reliable (for that 
> application), and for which better applications exist. So I prefer the 
> IOAPIC NMI.

The IO-APIC based NMI watchdog is neither free nor reliable.

Firstly, when we use it we cripple dynticks/NOHZ.

Secondly, there's been cases where activating the IO-APIC NMI watchdog 
causes rare lockups. One of my testsystems is such - once every few 
dozen bootups it would hang hard with the IO-APIC NMI watchdog 
activated.

No other OS uses the IO-APIC in this fashion, so it's a cute legacy 
hack, but we are phasing it out. It's not default.

> (I admit part of that preference is because using the IOAPIC NMI 
> helped stabilize a Dell PE2650 here years ago. Without the IOAPIC NMI 
> the machine would lock up hard within days or a few weeks at the 
> most.)

That looks quite myserious, did you get to the bottom of it? Does it 
reproduce with latest kernels too? It might be one of the dynticks hangs 
we fixed, and such hangs got in essence worked around by the nmi 
watchdog deactivating nohz. The nohz=off boot option would do something 
similar.

	Ingo