Re: [PATCH] x86: Add irq spillover warning

From: Neil Horman
Date: Tue Jul 16 2019 - 12:08:25 EST


On Tue, Jul 16, 2019 at 05:57:31PM +0200, Thomas Gleixner wrote:
> Neil,
>
> On Tue, 16 Jul 2019, Neil Horman wrote:
>
> > On Intel hardware, cpus are limited in the number of irqs they can
> > have affined to them (currently 240), based on section 10.5.2 of:
> > https://www.intel.com/content/dam/www/public/us/en/documents/manuals/64-ia-32-architectures-software-developer-vol-3a-part-1-manual.pdf
>
> That reference is really not useful to explain the problem and the number
> of vectors is neither. Please explain the conceptual issue.
>
You seem to have already done that below. Not really sure what more you are
asking for here.

> > If a cpu has more than this number of interrupts affined to it, they
> > will spill over to other cpus, which potentially may be outside of their
> > affinity mask.
>
> Spill over?
>
> The kernel decides to pick a vector on a CPU outside of the affinity when
> it runs out of vectors on the CPUs in the affinity mask.
>
Yes.

> Please explain issues technically correct.
>
I don't know what you mean by this. I explained it above, and you clearly
understood it.

> > Given that this might cause unexpected behavior on
> > performance sensitive systems, warn the user should this condition occur
> > so that corrective action can be taken
>
> > @@ -244,6 +244,14 @@ __visible unsigned int __irq_entry do_IRQ(struct pt_regs *regs)
>
> Why on earth warn in the interrupt delivery hotpath? Just because it's the
> place which really needs extra instructions and extra cache lines on
> performance sensitive systems, right?
>
Because theres already a check of the same variety in do_IRQ, but if the
information is available outside the hotpath, I was unaware, and am happy to
update this patch to refelct that.

> The fact that the kernel ran out of vectors for the CPUs in the affinity
> mask is already known when the vector is allocated in activate_reserved().
>
> So there is an obvious place to put such a warning and it's certainly not
> do_IRQ().
>
Sure

Thanks
Neil

> Thanks,
>
> tglx
>