Re: [PATCH net-next v2] net: stop napi kthreads when THREADED napi is disabled

[Jakub Kicinski <kuba@kernel.org>]

On Wed, 21 May 2025 12:51:33 -0700 Samiullah Khawaja wrote:
> > This might suffer from the problem you highlighted earlier,
> > CPU 0 (IRQ)             CPU 1 (NAPI thr)          CPU 2 (config)
> >
> >   ____napi_schedule()
> >     if (test_bit(NAPI_STATE_THREADED))
> >     if (thread) {
> >
> >  kthread_stop()
> >                               if (state & SCHED_THREADED || !(state & SCHED)) {
> >                                    state &= ~THREADED;
> >                               if (try_cmp_xchg())
> >                                    break
> >
> >        set_bit(NAPI_STATE_SCHED_THREADED)
> >        wake_up_process(thread);

This got a bit line wrapped for me so can't judge :(

> > This would happen without the try_cmp_xchg logic that I added in my
> > patch in the __napi_schedule (in the fast path). __napi_schedule would
> > have to make sure that the kthread is not stopping while it is trying
> > to do SCHED. This is similar to the logic we have in
> > napi_schedule_prep that handles the STATE_DISABLE, STATE_SCHED and
> > STATE_MISSED scenarios. Also if it falls back to normal softirq, it
> > needs to make sure that the kthread is not polling at the same time.  
> Discard this as the SCHED would be set in napi_schedule_prepare before
> __napi_schedule is called in IRQ, so try_cmp_xchg would return false.
> I think if the thread stops if the napi is idle(SCHED is not) set then
> it should do. This should make sure any pending SCHED_THREADED are
> also done. The existing logic in napi_schedulle_prep should handle all
> the cases.

I think we're on the same page. We're clearing the THREADED bit 
(not SCHED_THREADED). Only napi_schedule() path looks at that bit, 
after setting SCHED. So if we cmpxchg on a state where SCHED was 
clear - we can't race with anything that cares about THREADED bit.

Just to be clear - the stopping of the thread has to be after the
proposed loop, so kthread_should_stop() does not come into play.

And FWIW my understanding is that we don't need any barriers on the
fast path (SCHED vs checking THREADED) because memory ordering is
a thing which exists only between distinct memory words.