Skip to contentSkip to conversation

sched/stats: Fix rt/dl task's sched latency statistics error in sched_stat_wait trace_point

PATCH Superseded Junwen Wu 2 participants 3 messages Jan 1, 2024 – Jan 10, 2024 Scheduler

v3 of 5 Superseded Latest is v5 →

lore
Share & tools
lore

Junwen Wu

author
#0 lore raw

PATCH v3 1/1 sched/stats: Fix rt/dl task's sched latency statistics error in sched_stat_wait trace_point

When enable sched_stat_wait trace_point, some rt tasks sched latency so long, like this,
sched_stat_wait: comm=rcu_preempt pid=14 delay=4936139545261 [ns]
Rt task has low latency, it must have a bug. When rt task balance off source cpu, dequeue operation not update the sched_statistics, so follow update_stats_wait_end_fair update method, so do dl tasks.

Fixes 57a5c2dafca8 ("sched/rt: Support schedstats for RT sched class")
Fixes b5eb4a5f6521 ("sched/dl: Support schedstats for deadline sched class")
---
 kernel/sched/deadline.c | 8 +++++++-
 kernel/sched/rt.c       | 7 +++++++
 2 files changed, 14 insertions(+), 1 deletion(-)
kernel/sched/deadline.c +7 -1
index b28114478b82..29223163ee22 100644
@@ -1558,10 +1558,16 @@ update_stats_dequeue_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se,
 			int flags)
 {
 	struct task_struct *p = dl_task_of(dl_se);
+	struct rq *rq = rq_of_dl_rq(dl_rq);
 
 	if (!schedstat_enabled())
 		return;
-
+	/*
+	 * Mark the end of the wait period
+	 * if dequeueing a waiting task.
+	 */
+	if (p && (p != rq->curr))
+		 update_stats_wait_end_dl(dl_rq, dl_se);
 	if ((flags & DEQUEUE_SLEEP)) {
 		unsigned int state;
 
kernel/sched/rt.c +7 -0
index 6aaf0a3d6081..6a2600213991 100644
@@ -1360,12 +1360,19 @@ update_stats_dequeue_rt(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se,
 			int flags)
 {
 	struct task_struct *p = NULL;
+	struct rq *rq = rq_of_rt_se(rt_se);
 
 	if (!schedstat_enabled())
 		return;
 
 	if (rt_entity_is_task(rt_se))
 		p = rt_task_of(rt_se);
+	 /*
+	  * Mark the end of the wait period
+	  * if dequeueing a waiting task.
+	  */
+	if (p && (p != rq->curr))
+		update_stats_wait_end_rt(rt_rq, rt_se);
 
 	if ((flags & DEQUEUE_SLEEP) && p) {
 		unsigned int state;
Signature · 1 line
2.34.1
#1 lore raw

Hi Junwen

s/trace_point/tracepoint/g

Quoted 5 lines · Junwen Wu

When enable sched_stat_wait trace_point, some rt tasks sched latency so long, like this,
sched_stat_wait: comm=rcu_preempt pid=14 delay=4936139545261 [ns]
Rt task has low latency, it must have a bug. When rt task balance off source cpu, dequeue operation not update the sched_statistics, so follow update_stats_wait_end_fair update method, so do dl tasks.

This is a little bit hard to read, how about something like this?

The sched_stat_wait tracepoint is showing unreasonably long latencies for real-time tasks. For example:

sched_stat_wait: comm=rcu_preempt pid=14 delay=4936139545261 [ns]

This error happens when the rt task balances off the source CPU because the dequeue operation is not updating the sched_statistics. So, follow update_stats_wait_end_fair() and update the stats. Do the same for SCHED_DEADLINE.

[ feel free to change ]

Quoted 16 lines
Fixes 57a5c2dafca8 ("sched/rt: Support schedstats for RT sched class")
Fixes b5eb4a5f6521 ("sched/dl: Support schedstats for deadline sched class")
---
 kernel/sched/deadline.c | 8 +++++++-
 kernel/sched/rt.c       | 7 +++++++
 2 files changed, 14 insertions(+), 1 deletion(-)
kernel/sched/deadline.c +0 -0
index b28114478b82..29223163ee22 100644
@@ -1558,10 +1558,16 @@ update_stats_dequeue_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se,
 			int flags)
 {

kernel/sched/deadline.c update_stats_dequeue_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se, · line 1568

Quoted 10 lines
kernel/sched/deadline.c +5 -1
 	struct task_struct *p = dl_task_of(dl_se);
+	struct rq *rq = rq_of_dl_rq(dl_rq);
 
 	if (!schedstat_enabled())
 		return;
-
+	/*
+	 * Mark the end of the wait period
+	 * if dequeueing a waiting task.
+	 */

make this a single line comment.

Quoted 14 lines
kernel/sched/deadline.c +2 -0
+	if (p && (p != rq->curr))
+		 update_stats_wait_end_dl(dl_rq, dl_se);
 	if ((flags & DEQUEUE_SLEEP)) {
 		unsigned int state;
 
kernel/sched/rt.c +1 -0
index 6aaf0a3d6081..6a2600213991 100644
@@ -1360,12 +1360,19 @@ update_stats_dequeue_rt(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se,
 			int flags)
 {
 	struct task_struct *p = NULL;
+	struct rq *rq = rq_of_rt_se(rt_se);

kernel/sched/rt.c update_stats_dequeue_rt(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, · line 1373

Quoted 10 lines
kernel/sched/rt.c +4 -0
 
 	if (!schedstat_enabled())
 		return;
 
 	if (rt_entity_is_task(rt_se))
 		p = rt_task_of(rt_se);
+	 /*
+	  * Mark the end of the wait period
+	  * if dequeueing a waiting task.
+	  */

ditto

-- Daniel

author
#2 lore raw
Quoted 6 lines

This error happens when the rt task balances off the source CPU because the dequeue operation is not updating the sched_statistics. So, follow update_stats_wait_end_fair() and update the stats. Do the same for SCHED_DEADLINE.

[ feel free to change ]

ok

Quoted 5 lines

kernel/sched/deadline.c update_stats_dequeue_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se, · line 1568

Quoted 4 lines
kernel/sched/deadline.c +4 -0
+	/*
+	 * Mark the end of the wait period
+	 * if dequeueing a waiting task.
+	 */

make this a single line comment.

ok
--
Best regards