Re: [PATCH v1] net/dev.c : Remove redundant state settings after waking up

[Jakub Kicinski]

On Wed, 11 Jan 2023 08:32:42 +0100 Sebastian Andrzej Siewior wrote:
> It made sense in the beginning but now the suggested patch is a clean
> up. First the `woken' parameter was added in commit
>    cb038357937ee ("net: fix race between napi kthread mode and busy poll")
> 
> and then the `napi_disable_pending' check was removed in commit
>    27f0ad71699de ("net: fix hangup on napi_disable for threaded napi")
> 
> which renders the code to:
> |         while (!kthread_should_stop()) {
> |                 if (test_bit(NAPI_STATE_SCHED_THREADED, &napi->state) || woken) {
> |                         WARN_ON(!list_empty(&napi->poll_list));
> |                         __set_current_state(TASK_RUNNING);
> |                         return 0;
> |                 }
> | 
> |                 schedule();
> |                 /* woken being true indicates this thread owns this napi. */
> |                 woken = true;
> |                 set_current_state(TASK_INTERRUPTIBLE);
> |         }
> |         __set_current_state(TASK_RUNNING);
> 
> so when you get out of schedule() woken is set and even if
> NAPI_STATE_SCHED_THREADED is not set, the while() loop is left due to
> `woken = true'. So changing state to TASK_INTERRUPTIBLE makes no sense
> since it will be set back to TASK_RUNNING cycles later.

Ah, fair point, forgot about the woken optimization.