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

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

[Zqiang]

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>
---
 kernel/rcu/tree_nocb.h | 14 +++++++++-----
 1 file changed, 9 insertions(+), 5 deletions(-)

diff --git a/kernel/rcu/tree_nocb.h b/kernel/rcu/tree_nocb.h
index fa8e4f82e60c..c25c3e3dd22d 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));
+		} else {
+			/* 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);
+		}
 	} else if (!needwait_gp) {
 		/* Wait for callbacks to appear. */
 		trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("Sleep"));
@@ -980,7 +984,7 @@ static int rdp_offload_toggle(struct rcu_data *rdp,
 	raw_spin_lock_irqsave(&rdp_gp->nocb_gp_lock, flags);
 	// Queue this rdp for add/del to/from the list to iterate on rcuog
 	WRITE_ONCE(rdp_gp->nocb_toggling_rdp, rdp);
-	if (rdp_gp->nocb_gp_sleep) {
+	if (rcu_nocb_poll || rdp_gp->nocb_gp_sleep) {
 		rdp_gp->nocb_gp_sleep = false;
 		wake_gp = true;
 	}
-- 
2.25.1

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

[Paul E. McKenney]

On Mon, Jun 06, 2022 at 02:51:43PM +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>

Looks promising!  A couple of questions below.

> ---
>  kernel/rcu/tree_nocb.h | 14 +++++++++-----
>  1 file changed, 9 insertions(+), 5 deletions(-)
> 
> diff --git a/kernel/rcu/tree_nocb.h b/kernel/rcu/tree_nocb.h
> index fa8e4f82e60c..c25c3e3dd22d 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)) {

Do we need to set rdp_gp->nocb_gp_sleep to true in here somewhere?

> +			rcu_wait(READ_ONCE(my_rdp->nocb_toggling_rdp));
> +		} else {
> +			/* 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);
> +		}
>  	} else if (!needwait_gp) {
>  		/* Wait for callbacks to appear. */
>  		trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("Sleep"));
> @@ -980,7 +984,7 @@ static int rdp_offload_toggle(struct rcu_data *rdp,
>  	raw_spin_lock_irqsave(&rdp_gp->nocb_gp_lock, flags);
>  	// Queue this rdp for add/del to/from the list to iterate on rcuog
>  	WRITE_ONCE(rdp_gp->nocb_toggling_rdp, rdp);
> -	if (rdp_gp->nocb_gp_sleep) {
> +	if (rcu_nocb_poll || rdp_gp->nocb_gp_sleep) {

Is it the case that other than when (de-)offloading, the rcuog kthread
is never awakened when rcu_nocb_poll is set?  If this is not the case,
we would have spurious wakeups.  (I believe that rcu_nocb_poll suppresses
those wakeups, but it would be good to check.)

>  		rdp_gp->nocb_gp_sleep = false;
>  		wake_gp = true;
>  	}
> -- 
> 2.25.1
> 

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

[Zhang, Qiang1]

On Mon, Jun 06, 2022 at 02:51:43PM +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>

>Looks promising!  A couple of questions below.

> ---
>  kernel/rcu/tree_nocb.h | 14 +++++++++-----
>  1 file changed, 9 insertions(+), 5 deletions(-)
> 
> diff --git a/kernel/rcu/tree_nocb.h b/kernel/rcu/tree_nocb.h
> index fa8e4f82e60c..c25c3e3dd22d 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)) {

>Do we need to set rdp_gp->nocb_gp_sleep to true in here somewhere?

When rcu_nocb_poll is set,  the rdp_gp->nocb_gp_sleep flags is not used.

> +			rcu_wait(READ_ONCE(my_rdp->nocb_toggling_rdp));
> +		} else {
> +			/* 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);
> +		}
>  	} else if (!needwait_gp) {
>  		/* Wait for callbacks to appear. */
>  		trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("Sleep"));
> @@ -980,7 +984,7 @@ static int rdp_offload_toggle(struct rcu_data *rdp,
>  	raw_spin_lock_irqsave(&rdp_gp->nocb_gp_lock, flags);
>  	// Queue this rdp for add/del to/from the list to iterate on rcuog
>  	WRITE_ONCE(rdp_gp->nocb_toggling_rdp, rdp);
> -	if (rdp_gp->nocb_gp_sleep) {
> +	if (rcu_nocb_poll || rdp_gp->nocb_gp_sleep) {

>Is it the case that other than when (de-)offloading, the rcuog kthread
>is never awakened when rcu_nocb_poll is set?  If this is not the case,
>we would have spurious wakeups.  (I believe that rcu_nocb_poll suppresses
>those wakeups, but it would be good to check.)

When rcu_nocb_poll is set, only when (de-)offloading,  actively call 
wake_up_process(rdp_gp->nocb_gp_kthread). in other cases 
the __call_rcu_nocb_wake() will wakeup nocb_gp_kthread, but it met rcu_nocb_poll flags
directly return. 
In do_nocb_deferred_wakeup(), although there is no rcu_nocb_poll check, but
rdp->nocb_defer_wakeup is always RCU_NOCB_WAKE_NOT, also directly return.

Due to rcu_nocb_poll is set,  the rdp_gp->nocb_gp_sleep flags  is not used, and 
Not should be set.  move rcu_nocb_poll  from rdp_offload_toggle() to rcu_nocb_rdp_offload/deoffload()

Thanks
Zqiang

>  		rdp_gp->nocb_gp_sleep = false;
>  		wake_gp = true;
>  	}
> -- 
> 2.25.1
>