On Tue, Jun 25, 2024 at 10:57:42AM +0200, Daniel Wagner wrote:
> On Tue, Jun 25, 2024 at 09:07:30AM GMT, Thomas Gleixner wrote:
> > On Tue, Jun 25 2024 at 08:37, Hannes Reinecke wrote:
> > > On 6/24/24 11:00, Daniel Wagner wrote:
> > >> On Mon, Jun 24, 2024 at 10:47:05AM GMT, Christoph Hellwig wrote:
> > >>>> Do you think we should introduce a new type or just use the existing
> > >>>> managed_irq for this?
> > >>>
> > >>> No idea really. What was the reason for adding a new one?
> > >>
> > >> I've added the new type so that the current behavior of spreading the
> > >> queues over to the isolated CPUs is still possible. I don't know if this
> > >> a valid use case or not. I just didn't wanted to kill this feature it
> > >> without having discussed it before.
> > >>
> > >> But if we agree this doesn't really makes sense with isolcpus, then I
> > >> think we should use the managed_irq one as nvme-pci is using the managed
> > >> IRQ API.
> > >>
> > > I'm in favour in expanding/modifying the managed irq case.
> > > For managed irqs the driver will be running on the housekeeping CPUs
> > > only, and has no way of even installing irq handlers for the isolcpus.
> >
> > Yes, that's preferred, but please double check with the people who
> > introduced that in the first place.
>
> The relevant code was added by Ming:
>
> 11ea68f553e2 ("genirq, sched/isolation: Isolate from handling managed
> interrupts")
>
> [...] it can happen that a managed interrupt whose affinity
> mask contains both isolated and housekeeping CPUs is routed to an isolated
> CPU. As a consequence IO submitted on a housekeeping CPU causes interrupts
> on the isolated CPU.
>
> Add a new sub-parameter 'managed_irq' for 'isolcpus' and the corresponding
> logic in the interrupt affinity selection code.
>
> The subparameter indicates to the interrupt affinity selection logic that
> it should try to avoid the above scenario.
> [...]
>
> From the commit message I read the original indent is that managed_irq
> should avoid speading queues on isolcated CPUs.
>
> Ming, do you agree to use the managed_irq mask to limit the queue
> spreading on isolated CPUs? It would make the io_queue option obsolete.
Yes, managed_irq is introduced for not spreading on isolated CPUs, and
it is supposed to work well.
The only problem of managed_irq is just that isolated CPUs are
spread, but they are excluded from irq effective masks.
Thanks,
Ming
