Re: [PATCH V3 4/4] genirq/affinity: irq vector spread among online CPUs as far as possible

From: Thomas Gleixner
Date: Wed Apr 04 2018 - 04:25:28 EST


On Wed, 4 Apr 2018, Ming Lei wrote:
> On Tue, Apr 03, 2018 at 03:32:21PM +0200, Thomas Gleixner wrote:
> > On Thu, 8 Mar 2018, Ming Lei wrote:
> > > 1) before 84676c1f21 ("genirq/affinity: assign vectors to all possible CPUs")
> > > irq 39, cpu list 0
> > > irq 40, cpu list 1
> > > irq 41, cpu list 2
> > > irq 42, cpu list 3
> > >
> > > 2) after 84676c1f21 ("genirq/affinity: assign vectors to all possible CPUs")
> > > irq 39, cpu list 0-2
> > > irq 40, cpu list 3-4,6
> > > irq 41, cpu list 5
> > > irq 42, cpu list 7
> > >
> > > 3) after applying this patch against V4.15+:
> > > irq 39, cpu list 0,4
> > > irq 40, cpu list 1,6
> > > irq 41, cpu list 2,5
> > > irq 42, cpu list 3,7
> >
> > That's more or less window dressing. If the device is already in use when
> > the offline CPUs get hot plugged, then the interrupts still stay on cpu 0-3
> > because the effective affinity of interrupts on X86 (and other
> > architectures) is always a single CPU.
> >
> > So this only might move interrupts to the hotplugged CPUs when the device
> > is initialized after CPU hotplug and the actual vector allocation moves an
> > interrupt out to the higher numbered CPUs if they have less vectors
> > allocated than the lower numbered ones.
>
> It works for blk-mq devices, such as NVMe.
>
> Now NVMe driver creates num_possible_cpus() hw queues, and each
> hw queue is assigned one msix irq vector.
>
> Storage is Client/Server model, that means the interrupt is only
> delivered to CPU after one IO request is submitted to hw queue and
> it is completed by this hw queue.
>
> When CPUs is hotplugged, and there will be IO submitted from these
> CPUs, then finally IOs complete and irq events are generated from
> hw queues, and notify these submission CPU by IRQ finally.

I'm aware how that hw-queue stuff works. But that only works if the
spreading algorithm makes the interrupts affine to offline/not-present CPUs
when the block device is initialized.

In the example above:

> > > irq 39, cpu list 0,4
> > > irq 40, cpu list 1,6
> > > irq 41, cpu list 2,5
> > > irq 42, cpu list 3,7

and assumed that at driver init time only CPU 0-3 are online then the
hotplug of CPU 4-7 will not result in any interrupt delivered to CPU 4-7.

So the extra assignment to CPU 4-7 in the affinity mask has no effect
whatsoever and even if the spreading result is 'perfect' it just looks
perfect as it is not making any difference versus the original result:

> > > irq 39, cpu list 0
> > > irq 40, cpu list 1
> > > irq 41, cpu list 2
> > > irq 42, cpu list 3

Thanks,

tglx