Re: Question on handling managed IRQs when hotplugging CPUs

From: Thomas Gleixner
Date: Tue Jan 29 2019 - 11:27:46 EST


On Tue, 29 Jan 2019, John Garry wrote:
> On 29/01/2019 12:01, Thomas Gleixner wrote:
> > If the last CPU which is associated to a queue (and the corresponding
> > interrupt) goes offline, then the subsytem/driver code has to make sure
> > that:
> >
> > 1) No more requests can be queued on that queue
> >
> > 2) All outstanding of that queue have been completed or redirected
> > (don't know if that's possible at all) to some other queue.
>
> This may not be possible. For the HW I deal with, we have symmetrical delivery
> and completion queues, and a command delivered on DQx will always complete on
> CQx. Each completion queue has a dedicated IRQ.

So you can stop queueing on DQx and wait for all outstanding ones to come
in on CQx, right?

> > That has to be done in that order obviously. Whether any of the
> > subsystems/drivers actually implements this, I can't tell.
>
> Going back to c5cb83bb337c25, it seems to me that the change was made with the
> idea that we can maintain the affinity for the IRQ as we're shutting it down
> as no interrupts should occur.
>
> However I don't see why we can't instead keep the IRQ up and set the affinity
> to all online CPUs in offline path, and restore the original affinity in
> online path. The reason we set the queue affinity to specific CPUs is for
> performance, but I would not say that this matters for handling residual IRQs.

Oh yes it does. The problem is especially on x86, that if you have a large
number of queues and you take a large number of CPUs offline, then you run
into vector space exhaustion on the remaining online CPUs.

In the worst case a single CPU on x86 has only 186 vectors available for
device interrupts. So just take a quad socket machine with 144 CPUs and two
multiqueue devices with a queue per cpu. ---> FAIL

It probably fails already with one device because there are lots of other
devices which have regular interrupt which cannot be shut down.

Thanks,

tglx