Re: [PATCH] rcu/tree: Improve comments in rcu_report_qs_rdp()

[Frederic Weisbecker]

On Sat, Feb 04, 2023 at 02:20:50AM +0000, Joel Fernandes (Google) wrote:
> Recent discussion triggered due to a patch linked below, from Qiang,
> shed light on the need to accelerate from QS reporting paths.
> 
> Update the comments to capture this piece of knowledge.
> 
> Link: https://lore.kernel.org/all/20230118073014.2020743-1-qiang1.zhang@xxxxxxxxx/
> Cc: Qiang Zhang <Qiang1.zhang@xxxxxxxxx>
> Cc: Frederic Weisbecker <frederic@xxxxxxxxxx>
> Signed-off-by: Joel Fernandes (Google) <joel@xxxxxxxxxxxxxxxxx>
> 
> ---
>  kernel/rcu/tree.c | 13 ++++++++++++-
>  1 file changed, 12 insertions(+), 1 deletion(-)
> 
> diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> index 93eb03f8ed99..713eb6ca6902 100644
> --- a/kernel/rcu/tree.c
> +++ b/kernel/rcu/tree.c
> @@ -1983,7 +1983,12 @@ rcu_report_qs_rdp(struct rcu_data *rdp)
>  	} else {
>  		/*
>  		 * This GP can't end until cpu checks in, so all of our
> -		 * callbacks can be processed during the next GP.
> +		 * callbacks can be processed during the next GP. Do
> +		 * the acceleration from here otherwise there may be extra
> +		 * grace period delays, as any accelerations from rcu_core()
> +		 * or note_gp_changes() may happen only after the GP after the
> +		 * current one has already started. Further, rcu_core()
> +		 * only accelerates if RCU is idle (no GP in progress).

Actually note_gp_changes() should take care of that. My gut feeling is that the
acceleration in rcu_report_qs_rdp() only stands for:

* callbacks that may be enqueued from an IRQ firing during the small window
  between the RNP unlock in note_gp_changes() and the RNP lock in
  rcu_report_qs_rdp()

* __note_gp_changes() got called even before from the GP kthread, and callbacks
  got enqueued between that and rcu_core().
  
Thanks.