Skip to contentSkip to conversation

rcu: Use rcu_gp_kthread_wake() to wake up grace period kthreads

PATCH Applied Pranith Kumar 2 participants 2 messages Aug 12, 2014 – Aug 13, 2014 RCU

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 rcu: Use rcu_gp_kthread_wake() to wake up grace period kthreads

Updated changelog regarding the removed comment about the implied barrier provided by wake_up() logic.

--
Pranith

The rcu_gp_kthread_wake() function checks for three conditions before waking up grace period kthreads:

*  Is the thread we are trying to wake up the current thread?
*  Are the gp_flags zero? (all threads wait on non-zero gp_flags condition)
*  Is there no thread created for this flavour, hence nothing to wake up?

If any one of these condition is true, we do not call wake_up(). It was found that there are quite a few avoidable wake ups both during idle time and under stress induced by rcutorture.

Idle:

Total:66000, unnecessary:66000, case1:61827, case2:66000, case3:0 Total:68000, unnecessary:68000, case1:63696, case2:68000, case3:0

rcutorture:

Total:254000, unnecessary:254000, case1:199913, case2:254000, case3:0 Total:256000, unnecessary:256000, case1:201784, case2:256000, case3:0

Here case{1-3} are the cases listed above. We can avoid these wake ups by using rcu_gp_kthread_wake() to conditionally wake up the grace period kthreads.

There is a comment about an implied barrier supplied by the wake_up() logic. This barrier is necessary for the awakened thread to see the updated ->gp_flags. This flag is always being updated with the root node lock held. Also, the awakened thread tries to acquire the root node lock before reading ->gp_flags because of which there is proper ordering.

Hence this commit tries to avoid calling wake_up() whenever we can by using rcu_gp_kthread_wake() function.

CC Mathieu Desnoyers <[email protected]>
---
 kernel/rcu/tree.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
kernel/rcu/tree.c +2 -2
index b63517c..6b68d2d 100644
@@ -1938,7 +1938,7 @@ static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
 {
 	WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
 	raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
-	wake_up(&rsp->gp_wq);  /* Memory barrier implied by wake_up() path. */
+	rcu_gp_kthread_wake(rsp);
 }
 
 /*
@@ -2516,7 +2516,7 @@ static void force_quiescent_state(struct rcu_state *rsp)
 	ACCESS_ONCE(rsp->gp_flags) =
 		ACCESS_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS;
 	raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
-	wake_up(&rsp->gp_wq);  /* Memory barrier implied by wake_up() path. */
+	rcu_gp_kthread_wake(rsp);
 }
 
 /*
Signature · 1 line
1.9.1
#1 lore raw
Quoted 5 lines · Pranith Kumar

Updated changelog regarding the removed comment about the implied barrier provided by wake_up() logic.

--
Pranith

Queued for 3.18, thank you!

Thanx, Paul

Quoted 64 lines

The rcu_gp_kthread_wake() function checks for three conditions before waking up grace period kthreads:

*  Is the thread we are trying to wake up the current thread?
*  Are the gp_flags zero? (all threads wait on non-zero gp_flags condition)
*  Is there no thread created for this flavour, hence nothing to wake up?

If any one of these condition is true, we do not call wake_up(). It was found that there are quite a few avoidable wake ups both during idle time and under stress induced by rcutorture.

Idle:

Total:66000, unnecessary:66000, case1:61827, case2:66000, case3:0 Total:68000, unnecessary:68000, case1:63696, case2:68000, case3:0

rcutorture:

Total:254000, unnecessary:254000, case1:199913, case2:254000, case3:0 Total:256000, unnecessary:256000, case1:201784, case2:256000, case3:0

Here case{1-3} are the cases listed above. We can avoid these wake ups by using rcu_gp_kthread_wake() to conditionally wake up the grace period kthreads.

There is a comment about an implied barrier supplied by the wake_up() logic. This barrier is necessary for the awakened thread to see the updated ->gp_flags. This flag is always being updated with the root node lock held. Also, the awakened thread tries to acquire the root node lock before reading ->gp_flags because of which there is proper ordering.

Hence this commit tries to avoid calling wake_up() whenever we can by using rcu_gp_kthread_wake() function.

CC Mathieu Desnoyers <[email protected]>
---
 kernel/rcu/tree.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
kernel/rcu/tree.c +2 -2
index b63517c..6b68d2d 100644
@@ -1938,7 +1938,7 @@ static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
 {
 	WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
 	raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
-	wake_up(&rsp->gp_wq);  /* Memory barrier implied by wake_up() path. */
+	rcu_gp_kthread_wake(rsp);
 }
 
 /*
@@ -2516,7 +2516,7 @@ static void force_quiescent_state(struct rcu_state *rsp)
 	ACCESS_ONCE(rsp->gp_flags) =
 		ACCESS_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS;
 	raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
-	wake_up(&rsp->gp_wq);  /* Memory barrier implied by wake_up() path. */
+	rcu_gp_kthread_wake(rsp);
 }
 
 /*
Signature · 1 line
1.9.1