Re: [PATCH net-next 1/2] net: napi: wake up ksoftirqd if needed after scheduling NAPI

[Sebastian Andrzej Siewior <bigeasy@linutronix.de>]

On 2022-02-08 16:57:59 [+0100], Paolo Abeni wrote:
> just for historic reference:
> 
> https://lkml.org/lkml/2016/6/15/460

that is
  https://lore.kernel.org/all/cover.1465996447.git.pabeni@redhat.com

let me digest that later…

> I think that running the thread performing the NAPI loop with
> SCHED_FIFO would be dangerous WRT DDOS. Even the affinity setting can
> give mixed results depending on the workload - unless you do good
> static CPUs allocation pinning each process manually, not really a
> generic setup.

The DDoS part is where I meant we need figure out the details. Since it
is a threaded-interrupt we could do msleep() as a break which is similar
to what softirq does. Detecting such a case might be difficult since the
budget is per-thread only and does not involve other NAPI-structs on the
same CPU like backlog.

The performance is usually best if the IRQ and threaded handler are
running on the same CPU. The win with the NAPI thread seems to be that
if two NAPIs fire on the same CPU then the scheduler will see two tasks
which will be moved (at some point) to different CPUs. And in a DDoS
like situation they can constantly push new skbs into the stack and
SCHED_OTHER ensures that the user can still do something. Without napi
threads, both NAPIs would be processed on the same CPU (in order) and
eventually moved to ksoftirqd where it will take a break under high
load.

> Cheers,
> 
> Paolo

Sebastian