Re: [RFC PATCH 1/4] genirq: Provide basic NMI management for interrupt lines

[Julien Thierry]

On 17/07/18 20:46, Thomas Gleixner wrote:
> On Tue, 17 Jul 2018, Julien Thierry wrote:
> > On 17/07/18 16:51, Peter Zijlstra wrote:
> > > On Tue, Jul 17, 2018 at 04:48:17PM +0100, Julien Thierry wrote:
> > > > Supporting desc->affinity_hint for NMIs shouldn't be an issue, right?
> > >
> > > Shouldn't NMIs be strictly per cpu interrupt sources?
> >
> > That's a good question, and I don't have a proper answer right now...
> >
> > However, per cpu or not, the affinity still needs to be set for that
> > interrupt. It seemed to me that the desc->affinity_hint was just a way to keep
> > track of the interrupt affinity when it gets set through
> > irq_set_affinity_hint.
> >
> > Is that not the case?
>
> No the affinity hint is a way to tell user space irqbalanced what the
> driver thinks is the best placement. Affinity settings are done by the
> kernel internal interfaces and those do not care about the hint at all.

Oh, I see. Thanks for explaining

I already do set the IRQF_NOBALANCING for NMIs, so fiddling with that 
should be already limited.

I'll consider setting the IRQF_PERCPU (or mandating it for the NMI 
requests).

Thanks,

--
Julien Thierry