RE: [PATCH] rcu: When rcuog kthreads is in polling mode, wakeup waitqueue is not requried

[Zhang, Qiang1]

On Fri, Jan 28, 2022 at 11:13:46AM +0800, Zqiang wrote:
> When grace period cleanup, the rcuog kthreads that waiting in sq 
> waitqueue will be awakened, however if the 'rcu_nocb_poll' is set, the 
> sq waitqueue always empty, so if 'rcu_nocb_poll' is set, return 
> directly.

>This does decrease grace-period-cleanup overhead in kernels built with CONFIG_RCU_NOCB_CPU=y and booted with the rcu_nocb_poll kernel boot parameter set.  On the other hand, it increases grace-period-cleanup overhead in kernels built with CONFIG_RCU_NOCB_CPU=y but booted without the rcu_nocb_poll kernel boot parameter set.
>
>Last I checked, more kernels were booted without the rcu_nocb_poll kernel boot parameter set.  If this is still the case, this patch would slow things down for most systems.
>
>Or are there now lots of systems booted with rcu_nocb_poll?

Hi Paul
I found that some of our customers configured rcu_nocb_poll startup parameters under Preempt-RT kernel.
at each grace period cleanup, we'll all wakeup sq waitqueue, however when rcuog kthreads is in polling mode, 
the wakeup operation doesn't required, because the sq waitqueue always empty.

Thanks,
Zqiang

>
>							Thanx, Paul

> Signed-off-by: Zqiang <qiang1.zhang@xxxxxxxxx>
> ---
>  kernel/rcu/tree_nocb.h | 2 ++
>  1 file changed, 2 insertions(+)
> 
> diff --git a/kernel/rcu/tree_nocb.h b/kernel/rcu/tree_nocb.h index 
> 636d0546a4e9..9e106c590e56 100644
> --- a/kernel/rcu/tree_nocb.h
> +++ b/kernel/rcu/tree_nocb.h
> @@ -201,6 +201,8 @@ static void rcu_lockdep_assert_cblist_protected(struct rcu_data *rdp)
>   */
>  static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)  {
> +	if (rcu_nocb_poll)
> +		return;
>  	swake_up_all(sq);
>  }
>  
> --
> 2.25.1
>