Re: [PATCH v2] rcu/nocb: Avoid polling when myrdp->nocb_head_rdp list is empty

[Frederic Weisbecker <frederic@kernel.org>]

On Tue, Jun 07, 2022 at 03:50:57PM +0800, Zqiang wrote:
> Currently, If the 'rcu_nocb_poll' bootargs is enable, all rcuog kthreads
> enter polling mode. however, due to only insert CPU's rdp which belong to
> rcu_nocb_mask to 'nocb_head_rdp' list or all CPU's rdp served by rcuog
> kthread have been de-offloaded, these cause the 'nocb_head_rdp' list
> served by rcuog kthread is empty, when the 'nocb_head_rdp' is empty,
> the rcuog kthread in polling mode not actually do anything. fix it by
> exiting polling mode when the 'nocb_head_rdp'list is empty, otherwise
> entering polling mode.
> 
> Signed-off-by: Zqiang <qiang1.zhang@intel.com>
> ---
>  v1->v2:
>  Move rcu_nocb_poll flags check from rdp_offload_toggle() to
>  rcu_nocb_rdp_offload/deoffload(), avoid unnecessary setting of
>  rdp_gp->nocb_gp_sleep flags, because when rcu_nocb_poll is set
>  the rdp_gp->nocb_gp_sleep is not used.
> 
>  kernel/rcu/tree_nocb.h | 16 ++++++++++------
>  1 file changed, 10 insertions(+), 6 deletions(-)
> 
> diff --git a/kernel/rcu/tree_nocb.h b/kernel/rcu/tree_nocb.h
> index fa8e4f82e60c..2a52c9abc681 100644
> --- a/kernel/rcu/tree_nocb.h
> +++ b/kernel/rcu/tree_nocb.h
> @@ -698,10 +698,14 @@ static void nocb_gp_wait(struct rcu_data *my_rdp)
>  				   TPS("WakeBypassIsDeferred"));
>  	}
>  	if (rcu_nocb_poll) {
> -		/* Polling, so trace if first poll in the series. */
> -		if (gotcbs)
> -			trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("Poll"));
> -		schedule_timeout_idle(1);
> +		if (list_empty(&my_rdp->nocb_head_rdp)) {
> +			rcu_wait(READ_ONCE(my_rdp->nocb_toggling_rdp));

I suspect you have based your patch on upstream tree which doesn't seem
to have this one yet:

   "rcu/nocb: Add/del rdp to iterate from rcuog itself"

With this patch you can't wait on changes to my_rdp->nocb_toggling_rdp because
nocb_gp_wait() now performs the list_add/list_del itself.

Please rebase your patch on top of latest rcu:dev from Paul's tree. Then
all you need to do is to change the wait side, something like this (untested):

diff --git a/kernel/rcu/tree_nocb.h b/kernel/rcu/tree_nocb.h
index fa8e4f82e60c..f36d6be4f372 100644
--- a/kernel/rcu/tree_nocb.h
+++ b/kernel/rcu/tree_nocb.h
@@ -584,6 +584,15 @@ static int nocb_gp_toggle_rdp(struct rcu_data *rdp,
 	return ret;
 }
 
+static void nocb_gp_sleep(struct rdp *my_rdp, int cpu)
+{
+	trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("Sleep"));
+	swait_event_interruptible_exclusive(my_rdp->nocb_gp_wq,
+					    !READ_ONCE(my_rdp->nocb_gp_sleep));
+	trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("EndSleep"));
+}
+
+
 /*
  * No-CBs GP kthreads come here to wait for additional callbacks to show up
  * or for grace periods to end.
@@ -701,13 +710,19 @@ static void nocb_gp_wait(struct rcu_data *my_rdp)
 		/* Polling, so trace if first poll in the series. */
 		if (gotcbs)
 			trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("Poll"));
-		schedule_timeout_idle(1);
+		if (list_empty(&my_rdp->nocb_head_rdp)) {
+			raw_spin_lock_irqsave(&my_rdp->nocb_gp_lock, flags);
+			if (!my_rdp->nocb_toggling_rdp)
+				WRITE_ONCE(my_rdp->nocb_gp_sleep, true);
+			raw_spin_unlock_irqrestore(&my_rdp->nocb_gp_lock, flags);
+			/* Wait for any offloading rdp */
+			rdp_gp_sleep(my_rdp, cpu);
+		} else {
+			schedule_timeout_idle(1);
+		}
 	} else if (!needwait_gp) {
 		/* Wait for callbacks to appear. */
-		trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("Sleep"));
-		swait_event_interruptible_exclusive(my_rdp->nocb_gp_wq,
-				!READ_ONCE(my_rdp->nocb_gp_sleep));
-		trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("EndSleep"));
+		rdp_gp_sleep(my_rdp, cpu);
 	} else {
 		rnp = my_rdp->mynode;
 		trace_rcu_this_gp(rnp, my_rdp, wait_gp_seq, TPS("StartWait"));