Re: [PATCH RFC 1/1] genirq: Make threaded handler use irq affinity for managed interrupt

[John Garry]

On 10/12/2019 11:36, Marc Zyngier wrote:
> On 2019-12-10 10:59, John Garry wrote:
> > > > There is no lockup, just a potential performance boost in this change.
> > > >
> > > > My colleague Xiang Chen can provide specifics of the test, as he is
> > > > the one running it.
> > > >
> > > > But one key bit of info - which I did not think most relevant before
> > > > - that is we have 2x SAS controllers running the throughput test on
> > > > the same host.
> > > >
> > > > As such, the completion queue interrupts would be spread identically
> > > > over the CPUs for each controller. I notice that ARM GICv3 ITS
> > > > interrupt controller (which we use) does not use the generic irq
> > > > matrix allocator, which I think would really help with this.
> > > >
> > > > Hi Marc,
> > > >
> > > > Is there any reason for which we couldn't utilise of the generic irq
> > > > matrix allocator for GICv3?
> >
> > Hi Marc,
> >
> > > For a start, the ITS code predates the matrix allocator by about three
> > > years. Also, my understanding of this allocator is that it allows
> > > x86 to cope with a very small number of possible interrupt vectors
> > > per CPU. The ITS doesn't have such issue, as:
> > > 1) the namespace is global, and not per CPU
> > > 2) the namespace is *huge*
> > > Now, what property of the matrix allocator is the ITS code missing?
> > > I'd be more than happy to improve it.
> >
> > I think specifically the property that the matrix allocator will try
> > to find a CPU for irq affinity which "has the lowest number of managed
> > IRQs allocated" - I'm quoting the comment on 
> > matrix_find_best_cpu_managed().
>
> But that decision is due to allocation constraints. You can have at most
> 256 interrupts per CPU, so the allocator tries to balance it.
>
> On the contrary, the ITS does care about how many interrupt target any
> given CPU. The whole 2^24 interrupt namespace can be thrown at a single
> CPU.
>
> > The ITS code will make the lowest online CPU in the affinity mask the
> > target CPU for the interrupt, which may result in some CPUs handling
> > so many interrupts.
>
> If what you want is for the *default* affinity to be spread around,
> that should be achieved pretty easily. Let me have a think about how
> to do that.

Cool, I anticipate that it should help my case.

I can also seek out some NVMe cards to see how it would help a more 
"generic" scenario.

Cheers,
John