Skip to contentSkip to conversation

sched/fair: Fix the wrong sched_stat_wait time

PATCH In review [email protected] 2 participants 6 messages Sep 30, 2020 – Oct 9, 2020 Scheduler

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 sched/fair: Fix the wrong sched_stat_wait time

From: jun qian <[email protected]>

When the sched_schedstat changes from 0 to 1, some sched se maybe already in the runqueue, the se->statistics.wait_start will be 0. So it will let the (rq_of(cfs_rq)) - se->statistics.wait_start) wrong. We need to avoid this scenario.

---
 kernel/sched/fair.c | 8 ++++++++
 1 file changed, 8 insertions(+)
kernel/sched/fair.c +8 -0
index 658aa7a..dd7c3bb 100644
@@ -908,6 +908,14 @@ static void update_curr_fair(struct rq *rq)
 	if (!schedstat_enabled())
 		return;
 
+	/*
+	 * When the sched_schedstat changes from 0 to 1, some sched se maybe
+	 * already in the runqueue, the se->statistics.wait_start will be 0.
+	 * So it will let the delta wrong. We need to avoid this scenario.
+	 */
+	if (unlikely(!schedstat_val(se->statistics.wait_start)))
+		return;
+
 	delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
 
 	if (entity_is_task(se)) {
Signature · 1 line
1.8.3.1
maintainer
#1 lore raw
Quoted 6 lines · On Wed, Sep 30, 2020 at 10:47:12AM +0800, [email protected] wrote:

From: jun qian <[email protected]>

When the sched_schedstat changes from 0 to 1, some sched se maybe already in the runqueue, the se->statistics.wait_start will be 0. So it will let the (rq_of(cfs_rq)) - se->statistics.wait_start) wrong. We need to avoid this scenario.

Is this really the only problem there? Did you do a full audit of that schedstat nonsense?

kernel/sched/fair.c static void update_curr_fair(struct rq *rq)

Quoted 11 lines
---
 kernel/sched/fair.c | 8 ++++++++
 1 file changed, 8 insertions(+)
kernel/sched/fair.c +0 -0
index 658aa7a..dd7c3bb 100644
@@ -908,6 +908,14 @@ static void update_curr_fair(struct rq *rq)

your git-diff is 'funny', it got the function ^ wrong.

Quoted 17 lines
kernel/sched/fair.c +8 -0
 	if (!schedstat_enabled())
 		return;
 
+	/*
+	 * When the sched_schedstat changes from 0 to 1, some sched se maybe
+	 * already in the runqueue, the se->statistics.wait_start will be 0.
+	 * So it will let the delta wrong. We need to avoid this scenario.
+	 */
+	if (unlikely(!schedstat_val(se->statistics.wait_start)))
+		return;
+
 	delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
 
 	if (entity_is_task(se)) {
Signature · 1 line
1.8.3.1
author
#2 lore raw
Quoted 4 lines · Peter Zijlstra <[email protected]> 于2020年9月30日周三 下午4:20写道:
Quoted 2 lines · On Wed, Sep 30, 2020 at 10:47:12AM +0800, [email protected] wrote:

From: jun qian <[email protected]>

Quoted 8 lines
Quoted 4 lines

When the sched_schedstat changes from 0 to 1, some sched se maybe already in the runqueue, the se->statistics.wait_start will be 0. So it will let the (rq_of(cfs_rq)) - se->statistics.wait_start) wrong. We need to avoid this scenario.

Is this really the only problem there? Did you do a full audit of that schedstat nonsense?

Did you mean that the sched_stat_xxx's xxx_start(sched_stat_sleep sched_stat_iowait sched_stat_blocked sched_stat_runtime) may be also depend the schedstat_enabled? I have searched the codes, and found that these sched_stat_xxx's xxx_start don't depend the schedstat_enabled
except the wait_start.

This patch is going to slove the problem that when the schedstat_enabled is enabled, the sched_stat_wait of
the probed process will become unbelievable big probability in the fist time.

Quoted 6 lines
Quoted 6 lines
---
 kernel/sched/fair.c | 8 ++++++++
 1 file changed, 8 insertions(+)
Quoted 8 lines

kernel/sched/fair.c static void update_curr_fair(struct rq *rq)

Quoted 5 lines
kernel/sched/fair.c +0 -0
index 658aa7a..dd7c3bb 100644
@@ -908,6 +908,14 @@ static void update_curr_fair(struct rq *rq)

your git-diff is 'funny', it got the function ^ wrong.

sorry :)

Quoted 17 lines
Quoted 17 lines
kernel/sched/fair.c +8 -1
      if (!schedstat_enabled())
              return;
 
+     /*
+      * When the sched_schedstat changes from 0 to 1, some sched se maybe
+      * already in the runqueue, the se->statistics.wait_start will be 0.
+      * So it will let the delta wrong. We need to avoid this scenario.
+      */
+     if (unlikely(!schedstat_val(se->statistics.wait_start)))
+             return;
+
      delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
 
      if (entity_is_task(se)) {
--

1.8.3.1

maintainer
#3 lore raw
Quoted 10 lines · jun qian
Quoted 9 lines · Peter Zijlstra <[email protected]> 于2020年9月30日周三 下午4:20写道:
Quoted 6 lines · On Wed, Sep 30, 2020 at 10:47:12AM +0800, [email protected] wrote:

From: jun qian <[email protected]>

When the sched_schedstat changes from 0 to 1, some sched se maybe already in the runqueue, the se->statistics.wait_start will be 0. So it will let the (rq_of(cfs_rq)) - se->statistics.wait_start) wrong. We need to avoid this scenario.

Quoted 7 lines
Quoted 3 lines

Is this really the only problem there? Did you do a full audit of that schedstat nonsense?

Did you mean that the sched_stat_xxx's xxx_start(sched_stat_sleep sched_stat_iowait sched_stat_blocked sched_stat_runtime) may be also depend the schedstat_enabled?

Yeah, this runtime schedstat_enabled thing is fairly recent, it used to be an always on/off kinda thing.

At the time we figured inconsistencies from dynamically enabling/disabling it were okay, it's just stats after all.

But if you now want to 'fix' that, then a full audit might be nice.

Quoted 3 lines

I have searched the codes, and found that these sched_stat_xxx's xxx_start don't depend the schedstat_enabled
except the wait_start.

OK, so you did the audit and only found this one issue? That's good Changelog material :-)

Thanks!

author
#4 lore raw
Quoted 26 lines · Peter Zijlstra <[email protected]> 于2020年9月30日周三 下午5:57写道:
Quoted 10 lines · jun qian
Quoted 9 lines · Peter Zijlstra <[email protected]> 于2020年9月30日周三 下午4:20写道:
Quoted 6 lines · On Wed, Sep 30, 2020 at 10:47:12AM +0800, [email protected] wrote:

From: jun qian <[email protected]>

When the sched_schedstat changes from 0 to 1, some sched se maybe already in the runqueue, the se->statistics.wait_start will be 0. So it will let the (rq_of(cfs_rq)) - se->statistics.wait_start) wrong. We need to avoid this scenario.

Quoted 7 lines
Quoted 3 lines

Is this really the only problem there? Did you do a full audit of that schedstat nonsense?

Did you mean that the sched_stat_xxx's xxx_start(sched_stat_sleep sched_stat_iowait sched_stat_blocked sched_stat_runtime) may be also depend the schedstat_enabled?

Yeah, this runtime schedstat_enabled thing is fairly recent, it used to be an always on/off kinda thing.

At the time we figured inconsistencies from dynamically enabling/disabling it were okay, it's just stats after all.

Quoted 9 lines

But if you now want to 'fix' that, then a full audit might be nice.

Quoted 3 lines

I have searched the codes, and found that these sched_stat_xxx's xxx_start don't depend the schedstat_enabled
except the wait_start.

OK, so you did the audit and only found this one issue? That's good Changelog material :-)

I found another problem, when the sched_schedstat changes from 1 to 0, the sched se
which is already in the runqueue, the statistics.wait_start already has a value. At this moment
sched_schedstat changes from 0 to 1 again, the (rq_of(cfs_rq)) - se->statistics.wait_start)
will not be 0 and the wait time will be bigger than the real one. So we need to modify the patch
to resolve the problem with this scenario.

So I really need a full audit might, :-)

Thanks

Quoted 1 line

Thanks!

author
#5 lore raw
Quoted 38 lines · jun qian <[email protected]> 于2020年9月30日周三 下午9:08写道:
Quoted 26 lines · Peter Zijlstra <[email protected]> 于2020年9月30日周三 下午5:57写道:
Quoted 10 lines · jun qian
Quoted 9 lines · Peter Zijlstra <[email protected]> 于2020年9月30日周三 下午4:20写道:
Quoted 6 lines · On Wed, Sep 30, 2020 at 10:47:12AM +0800, [email protected] wrote:

From: jun qian <[email protected]>

When the sched_schedstat changes from 0 to 1, some sched se maybe already in the runqueue, the se->statistics.wait_start will be 0. So it will let the (rq_of(cfs_rq)) - se->statistics.wait_start) wrong. We need to avoid this scenario.

Quoted 7 lines
Quoted 3 lines

Is this really the only problem there? Did you do a full audit of that schedstat nonsense?

Did you mean that the sched_stat_xxx's xxx_start(sched_stat_sleep sched_stat_iowait sched_stat_blocked sched_stat_runtime) may be also depend the schedstat_enabled?

Yeah, this runtime schedstat_enabled thing is fairly recent, it used to be an always on/off kinda thing.

At the time we figured inconsistencies from dynamically enabling/disabling it were okay, it's just stats after all.

Quoted 9 lines

But if you now want to 'fix' that, then a full audit might be nice.

Quoted 3 lines

I have searched the codes, and found that these sched_stat_xxx's xxx_start don't depend the schedstat_enabled
except the wait_start.

OK, so you did the audit and only found this one issue? That's good Changelog material :-)

Quoted 10 lines

I found another problem, when the sched_schedstat changes from 1 to 0, the sched se
which is already in the runqueue, the statistics.wait_start already has a value. At this moment
sched_schedstat changes from 0 to 1 again, the (rq_of(cfs_rq)) - se->statistics.wait_start)
will not be 0 and the wait time will be bigger than the real one. So we need to modify the patch
to resolve the problem with this scenario.

hi Peter

I have sent another patch to improve the accuracy of sched_stat_wait statistics.
I have no good idea to solve the another problem what i described up, but the new
patch can improve the accuracy.

Thanks

Quoted 5 lines

So I really need a full audit might, :-)

Thanks

Quoted 1 line

Thanks!