[PATCH] sched/fair: Make cfs_rq::decay_counter non-atomic

[PATCH] sched/fair: Make cfs_rq::decay_counter non-atomic

[Kirill Tkhai]

We update decay_counter in update_cfs_rq_blocked_load()
only. This function is always called with rq lock locked,
so we can kill atomic actions.

Signed-off-by: Kirill Tkhai <ktkhai@parallels.com>
---
 kernel/sched/fair.c  |   11 ++++++-----
 kernel/sched/sched.h |    2 +-
 2 files changed, 7 insertions(+), 6 deletions(-)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 5f3b5a7..af990c4 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -2570,7 +2570,7 @@ static __always_inline int __update_entity_runnable_avg(u64 now,
 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
 {
 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
-	u64 decays = atomic64_read(&cfs_rq->decay_counter);
+	u64 decays = ACCESS_ONCE(cfs_rq->decay_counter);
 
 	decays -= se->avg.decay_count;
 	if (!decays)
@@ -2767,7 +2767,8 @@ static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
 	if (decays) {
 		cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
 						      decays);
-		atomic64_add(decays, &cfs_rq->decay_counter);
+		lockdep_assert_held(&rq_of(cfs_rq)->lock);
+		cfs_rq->decay_counter += decays;
 		cfs_rq->last_decay = now;
 	}
 
@@ -2837,7 +2838,7 @@ static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
 	cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
 	if (sleep) {
 		cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
-		se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
+		se->avg.decay_count = ACCESS_ONCE(cfs_rq->decay_counter);
 	} /* migrations, e.g. sleep=0 leave decay_count == 0 */
 }
 
@@ -7876,7 +7877,7 @@ void init_cfs_rq(struct cfs_rq *cfs_rq)
 	cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
 #endif
 #ifdef CONFIG_SMP
-	atomic64_set(&cfs_rq->decay_counter, 1);
+	cfs_rq->decay_counter = 1;
 	atomic_long_set(&cfs_rq->removed_load, 0);
 #endif
 }
@@ -7928,7 +7929,7 @@ static void task_move_group_fair(struct task_struct *p, int queued)
 		 * contribution, but we must synchronize for ongoing future
 		 * decay.
 		 */
-		se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
+		se->avg.decay_count = ACCESS_ONCE(cfs_rq->decay_counter);
 		cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
 #endif
 	}
diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
index 9a2a45c..672f120 100644
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h
@@ -364,7 +364,7 @@ struct cfs_rq {
 	 * the FAIR_GROUP_SCHED case).
 	 */
 	unsigned long runnable_load_avg, blocked_load_avg;
-	atomic64_t decay_counter;
+	u64 decay_counter;
 	u64 last_decay;
 	atomic_long_t removed_load;
 

Re: [PATCH] sched/fair: Make cfs_rq::decay_counter non-atomic

[Kirill Tkhai]

This should go on top of yesterday's "sched/fair:
Fix sched_entity::avg::decay_count initialization"

В Вт, 16/12/2014 в 17:25 +0300, Kirill Tkhai пишет:
> We update decay_counter in update_cfs_rq_blocked_load()
> only. This function is always called with rq lock locked,
> so we can kill atomic actions.
> 
> Signed-off-by: Kirill Tkhai <ktkhai@parallels.com>
> ---
>  kernel/sched/fair.c  |   11 ++++++-----
>  kernel/sched/sched.h |    2 +-
>  2 files changed, 7 insertions(+), 6 deletions(-)
> 
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 5f3b5a7..af990c4 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -2570,7 +2570,7 @@ static __always_inline int __update_entity_runnable_avg(u64 now,
>  static inline u64 __synchronize_entity_decay(struct sched_entity *se)
>  {
>  	struct cfs_rq *cfs_rq = cfs_rq_of(se);
> -	u64 decays = atomic64_read(&cfs_rq->decay_counter);
> +	u64 decays = ACCESS_ONCE(cfs_rq->decay_counter);
>  
>  	decays -= se->avg.decay_count;
>  	if (!decays)
> @@ -2767,7 +2767,8 @@ static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
>  	if (decays) {
>  		cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
>  						      decays);
> -		atomic64_add(decays, &cfs_rq->decay_counter);
> +		lockdep_assert_held(&rq_of(cfs_rq)->lock);
> +		cfs_rq->decay_counter += decays;
>  		cfs_rq->last_decay = now;
>  	}
>  
> @@ -2837,7 +2838,7 @@ static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
>  	cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
>  	if (sleep) {
>  		cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
> -		se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
> +		se->avg.decay_count = ACCESS_ONCE(cfs_rq->decay_counter);
>  	} /* migrations, e.g. sleep=0 leave decay_count == 0 */
>  }
>  
> @@ -7876,7 +7877,7 @@ void init_cfs_rq(struct cfs_rq *cfs_rq)
>  	cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
>  #endif
>  #ifdef CONFIG_SMP
> -	atomic64_set(&cfs_rq->decay_counter, 1);
> +	cfs_rq->decay_counter = 1;
>  	atomic_long_set(&cfs_rq->removed_load, 0);
>  #endif
>  }
> @@ -7928,7 +7929,7 @@ static void task_move_group_fair(struct task_struct *p, int queued)
>  		 * contribution, but we must synchronize for ongoing future
>  		 * decay.
>  		 */
> -		se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
> +		se->avg.decay_count = ACCESS_ONCE(cfs_rq->decay_counter);
>  		cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
>  #endif
>  	}
> diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
> index 9a2a45c..672f120 100644
> --- a/kernel/sched/sched.h
> +++ b/kernel/sched/sched.h
> @@ -364,7 +364,7 @@ struct cfs_rq {
>  	 * the FAIR_GROUP_SCHED case).
>  	 */
>  	unsigned long runnable_load_avg, blocked_load_avg;
> -	atomic64_t decay_counter;
> +	u64 decay_counter;
>  	u64 last_decay;
>  	atomic_long_t removed_load;
>  
> 
> 

Re: [PATCH] sched/fair: Make cfs_rq::decay_counter non-atomic

[bsegall]

Kirill Tkhai <ktkhai@parallels.com> writes:

> We update decay_counter in update_cfs_rq_blocked_load()
> only. This function is always called with rq lock locked,
> so we can kill atomic actions.
>
> Signed-off-by: Kirill Tkhai <ktkhai@parallels.com>
> ---
>  kernel/sched/fair.c  |   11 ++++++-----
>  kernel/sched/sched.h |    2 +-
>  2 files changed, 7 insertions(+), 6 deletions(-)
>
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 5f3b5a7..af990c4 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -2570,7 +2570,7 @@ static __always_inline int __update_entity_runnable_avg(u64 now,
>  static inline u64 __synchronize_entity_decay(struct sched_entity *se)
>  {
>  	struct cfs_rq *cfs_rq = cfs_rq_of(se);
> -	u64 decays = atomic64_read(&cfs_rq->decay_counter);
> +	u64 decays = ACCESS_ONCE(cfs_rq->decay_counter);

This is called without rq lock held from migrate_task_rq_fair. (We could
technically change the atomic_add to atomic64_set(atomic64_read() + x),
but I don't know that that is a win) Now, we could do a
min_vruntime-style two-copy thing if this atomic usage is a hot spot on
32-bit, we just didn't bother initially.

Re: [PATCH] sched/fair: Make cfs_rq::decay_counter non-atomic

[Kirill Tkhai]

В Вт, 16/12/2014 в 10:00 -0800, bsegall@google.com пишет:
> Kirill Tkhai <ktkhai@parallels.com> writes:
> 
> > We update decay_counter in update_cfs_rq_blocked_load()
> > only. This function is always called with rq lock locked,
> > so we can kill atomic actions.
> >
> > Signed-off-by: Kirill Tkhai <ktkhai@parallels.com>
> > ---
> >  kernel/sched/fair.c  |   11 ++++++-----
> >  kernel/sched/sched.h |    2 +-
> >  2 files changed, 7 insertions(+), 6 deletions(-)
> >
> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > index 5f3b5a7..af990c4 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -2570,7 +2570,7 @@ static __always_inline int __update_entity_runnable_avg(u64 now,
> >  static inline u64 __synchronize_entity_decay(struct sched_entity *se)
> >  {
> >  	struct cfs_rq *cfs_rq = cfs_rq_of(se);
> > -	u64 decays = atomic64_read(&cfs_rq->decay_counter);
> > +	u64 decays = ACCESS_ONCE(cfs_rq->decay_counter);
> 
> This is called without rq lock held from migrate_task_rq_fair. (We could
> technically change the atomic_add to atomic64_set(atomic64_read() + x),
> but I don't know that that is a win) Now, we could do a
> min_vruntime-style two-copy thing if this atomic usage is a hot spot on
> 32-bit, we just didn't bother initially.

Oh, I forgot that 64-bit read is not everywhere atomic... Thanks, Ben.

Peter, please, ignore this patch.