Re: [PATCH v4 4/4] PCI: Limit pci_alloc_irq_vectors() to housekeeping CPUs

[Jakub Kicinski]

On Wed, 21 Oct 2020 22:25:48 +0200 Thomas Gleixner wrote:
> On Tue, Oct 20 2020 at 20:07, Thomas Gleixner wrote:
> > On Tue, Oct 20 2020 at 12:18, Nitesh Narayan Lal wrote:  
> >> However, IMHO we would still need a logic to prevent the devices from
> >> creating excess vectors.  
> >
> > Managed interrupts are preventing exactly that by pinning the interrupts
> > and queues to one or a set of CPUs, which prevents vector exhaustion on
> > CPU hotplug.
> >
> > Non-managed, yes that is and always was a problem. One of the reasons
> > why managed interrupts exist.  
> 
> But why is this only a problem for isolation? The very same problem
> exists vs. CPU hotplug and therefore hibernation.
> 
> On x86 we have at max. 204 vectors available for device interrupts per
> CPU. So assumed the only device interrupt in use is networking then any
> machine which has more than 204 network interrupts (queues, aux ...)
> active will prevent the machine from hibernation.
> 
> Aside of that it's silly to have multiple queues targeted at a single
> CPU in case of hotplug. And that's not a theoretical problem.  Some
> power management schemes shut down sockets when the utilization of a
> system is low enough, e.g. outside of working hours.
> 
> The whole point of multi-queue is to have locality so that traffic from
> a CPU goes through the CPU local queue. What's the point of having two
> or more queues on a CPU in case of hotplug?
> 
> The right answer to this is to utilize managed interrupts and have
> according logic in your network driver to handle CPU hotplug. When a CPU
> goes down, then the queue which is associated to that CPU is quiesced
> and the interrupt core shuts down the relevant interrupt instead of
> moving it to an online CPU (which causes the whole vector exhaustion
> problem on x86). When the CPU comes online again, then the interrupt is
> reenabled in the core and the driver reactivates the queue.

I think Mellanox folks made some forays into managed irqs, but I don't
remember/can't find the details now.

For networking the locality / queue per core does not always work,
since the incoming traffic is usually spread based on a hash. Many
applications perform better when network processing is done on a small
subset of CPUs, and application doesn't get interrupted every 100us. 
So we do need extra user control here.

We have a bit of a uAPI problem since people had grown to depend on
IRQ == queue == NAPI to configure their systems. "The right way" out
would be a proper API which allows associating queues with CPUs rather
than IRQs, then we can use managed IRQs and solve many other problems.

Such new API has been in the works / discussions for a while now.

(Magnus keep me honest here, if you disagree the queue API solves this.)