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

[Jakub Kicinski <kuba@kernel.org>]

On Wed, 21 May 2025 15:50:19 -0700 Samiullah Khawaja wrote:
> > 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.  
> Wait, the thread will unset STATE_THREADED if kthread_should_stop is
> true. right? Otherwise how would thread know that it has to unset the
> bit and stop?
> 
> As I understand, we should be doing something like following:
> 
> while (true) {
>    state = READ_ONCE()
>    can_stop = false;
> 
>    if (kthread_should_stop) {
>        if (SCHED_THREADED || !SCHED) {
>            state &= !THREADED
>        } else {
>            msleep(1);
>            continue;
>        }
> 
>         if (try_cmpxchg) {
>             can_stop = true;
>             if (!SCHED_THREADED)
>                 break;
>         }
>    }
> 
>    if (SCHED_THREADED)
>        poll()
> 
>    if (can_stop))
>        break;
> }

So moving the stopping logic into the polling thread? I don't think this
helps anything. 

Once we unset the THREADED bit we should wait for the thread to clear
SCHED_THREADED (before trying to stop it). SCHED_THREADED going away
is our signal that it's safe to reap the thread.

If you're afraid that the thread will not terminate (because packets
continue to flow in) - we probably want something like:

diff --git a/net/core/dev.c b/net/core/dev.c
index 2b514d95c528..33d4b726395b 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -7548,6 +7548,13 @@ static void napi_threaded_poll_loop(struct napi_struct *napi)
                if (!repoll)
                        break;
 
+               /* Thread is going away, give up the ownership */
+               if (!likely(READ_ONCE(napi->state) & NAPIF_STATE_THREADED)) {
+                       clear_bit(NAPI_STATE_SCHED_THREADED, &napi->state);
+                       __napi_schedule(napi);
+                       break;
+               }
+
                rcu_softirq_qs_periodic(last_qs);
                cond_resched();
        }