sched/fair: Apply more PELT fixes

A kernel bug report was opened against Ubuntu [0]. After a kernel bisect, it was found that reverting the following commit resolved this bug:

Quoted 2 lines · Joseph Salisbury

A kernel bug report was opened against Ubuntu [0]. After a kernel bisect, it was found that reverting the following commit resolved this bug:

Hmm. Interesting, and it sounds like we should revert that unless somebody figures out *why* following the rules wrt cfq updates causes problems. But I also wonder what the Ubuntu kernel config is. Does Ubuntu enable CONFIG_SCHED_AUTOGROUP=y, for example? Because regardless of any other scheduler issues, autogrouping *should* mean that when you run some CPU hogger in one session, that should still balance all CPU time with other sessions..

I'm not seeing anything odd on my xps13, which should have a similar CPU to the X1 Carbon.

Linus

Yes, CONFIG_SCHED_AUTOGROUP is enabled in the Ubuntu kernel. However, that config was also enable in the Ubuntu 4.4 kerrnels without seeing
this issue. I can try disabling the config in the 4.8 based kernel and
see if that changes things.

Hello Peter,

A kernel bug report was opened against Ubuntu [0]. After a kernel bisect, it was found that reverting the following commit resolved this bug:

commit 3d30544f02120b884bba2a9466c87dba980e3be5

Author: Peter Zijlstra <[email protected]>
Date:   Tue Jun 21 14:27:50 2016 +0200

sched/fair: Apply more PELT fixes

The regression was introduced as of v4.8-rc1. The bug can be reproduced on an X1 Carbon with the following:
stress -c $your_total_cpu_cores

I was hoping to get your feedback, since you are the patch author. Do you think gathering any additional data will help diagnose this issue, or would it be best to submit a revert request?

Quoted 5 lines · Joseph Salisbury

Hello Peter,

A kernel bug report was opened against Ubuntu [0]. After a kernel bisect, it was found that reverting the following commit resolved this bug:

Quoted 5 lines

commit 3d30544f02120b884bba2a9466c87dba980e3be5

Author: Peter Zijlstra <[email protected]>
Date:   Tue Jun 21 14:27:50 2016 +0200

sched/fair: Apply more PELT fixes

That commit doesn't revert cleanly, did you take out more?

I'll try and have a look asap, but I'm traveling next week so it might a tad slower than normal.

If you could provide a /proc/sched_debug dump while the thing is running that'd might be useful.

Quoted 3 lines · Peter Zijlstra

Quoted 2 lines · Joseph Salisbury

Hello Peter,

Quoted 8 lines

Quoted 8 lines

A kernel bug report was opened against Ubuntu [0]. After a kernel bisect, it was found that reverting the following commit resolved this bug:

commit 3d30544f02120b884bba2a9466c87dba980e3be5

Author: Peter Zijlstra <[email protected]>
Date:   Tue Jun 21 14:27:50 2016 +0200

sched/fair: Apply more PELT fixes

Quoted 2 lines

That commit doesn't revert cleanly, did you take out more?

Note that it reverts cleanly from v4.8 - while it does to revert from current upstream that did more changes in that area.

I suspect Josheph tested a v4.8-ish kernel.

Quoted 5 lines

I'll try and have a look asap, but I'm traveling next week so it might a tad slower than normal.

If you could provide a /proc/sched_debug dump while the thing is running that'd might be useful.

Also, running the latest scheduler bits would be useful:

git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.git

... just on the off chance that another change in this area fixed the bug, plus to make it easier to send test patches.

Upstream merge commit af79ad2b1f33 should also be pretty safe to try - it's v4.8 with the v4.9 scheduler bits applied.

Thanks,

Ingo

Quoted 15 lines · Ingo Molnar

Quoted 12 lines · Peter Zijlstra

Quoted 11 lines · Joseph Salisbury

Hello Peter,

A kernel bug report was opened against Ubuntu [0]. After a kernel
bisect, it was found that reverting the following commit resolved this bug:

commit 3d30544f02120b884bba2a9466c87dba980e3be5

Author: Peter Zijlstra <[email protected]>
Date:   Tue Jun 21 14:27:50 2016 +0200

Quoted 1 line

Quoted 1 line

Quoted 1 line

sched/fair: Apply more PELT fixes

This patch only speeds up the update of task group load in order to reflect the new load balance but It should not change the final value and as a result the final behavior. I will try to reproduce it in my target later today

Quoted 16 lines · Vincent Guittot

Quoted 15 lines · Ingo Molnar

Quoted 12 lines · Peter Zijlstra

Quoted 11 lines · Joseph Salisbury

Hello Peter,

A kernel bug report was opened against Ubuntu [0]. After a kernel
bisect, it was found that reverting the following commit resolved this bug:

commit 3d30544f02120b884bba2a9466c87dba980e3be5

Author: Peter Zijlstra <[email protected]>
Date:   Tue Jun 21 14:27:50 2016 +0200

Quoted 6 lines

Quoted 1 line

Quoted 1 line

Quoted 1 line

sched/fair: Apply more PELT fixes

This patch only speeds up the update of task group load in order to reflect the new load balance but It should not change the final value and as a result the final behavior. I will try to reproduce it in my target later today

FWIW, I tried and failed w/wo autogroup on 4.8 and master.

-Mike

Quoted 17 lines · Mike Galbraith

Quoted 16 lines · Vincent Guittot

Quoted 15 lines · Ingo Molnar

Quoted 13 lines · Peter Zijlstra

Quoted 12 lines · Joseph Salisbury

Hello Peter,

A kernel bug report was opened against Ubuntu [0]. After a kernel
bisect, it was found that reverting the following commit resolved this bug:

commit 3d30544f02120b884bba2a9466c87dba980e3be5

Author: Peter Zijlstra <[email protected]>
Date:   Tue Jun 21 14:27:50 2016 +0200

sched/fair: Apply more PELT fixes

Quoted 5 lines

Quoted 4 lines

This patch only speeds up the update of task group load in order to reflect the new load balance but It should not change the final value and as a result the final behavior. I will try to reproduce it in my target later today

FWIW, I tried and failed w/wo autogroup on 4.8 and master.

Me too

Is it possible to get some dump of /proc/sched_debug while the problem occurs ?

Vincent

Quoted 1 line

-Mike

Quoted 21 lines · Mike Galbraith

Quoted 20 lines · Vincent Guittot

Quoted 10 lines · Ingo Molnar

Quoted 8 lines · Peter Zijlstra

Quoted 7 lines · Joseph Salisbury

Hello Peter,

A kernel bug report was opened against Ubuntu [0]. After a kernel
bisect, it was found that reverting the following commit resolved this bug:

Quoted 5 lines

Quoted 5 lines

Quoted 5 lines

commit 3d30544f02120b884bba2a9466c87dba980e3be5

Author: Peter Zijlstra <[email protected]>
Date:   Tue Jun 21 14:27:50 2016 +0200

sched/fair: Apply more PELT fixes

This patch only speeds up the update of task group load in order to reflect the new load balance but It should not change the final value and as a result the final behavior. I will try to reproduce it in my target later today

Quoted 1 line

FWIW, I tried and failed w/wo autogroup on 4.8 and master.

Me too

Is it possible to get some dump of /proc/sched_debug while the problem occurs ?

Vincent

Quoted 1 line

-Mike

Quoted 25 lines · Vincent Guittot

Quoted 16 lines · Mike Galbraith

Quoted 15 lines · Vincent Guittot

Quoted 14 lines · Ingo Molnar

Quoted 12 lines · Peter Zijlstra

Quoted 11 lines · Joseph Salisbury

Hello Peter,

A kernel bug report was opened against Ubuntu [0]. After a kernel
bisect, it was found that reverting the following commit resolved this bug:

commit 3d30544f02120b884bba2a9466c87dba980e3be5

Author: Peter Zijlstra <[email protected]>
Date:   Tue Jun 21 14:27:50 2016 +0200

Quoted 6 lines

Quoted 5 lines

Quoted 1 line

Quoted 1 line

Quoted 1 line

sched/fair: Apply more PELT fixes

This patch only speeds up the update of task group load in order to reflect the new load balance but It should not change the final value and as a result the final behavior. I will try to reproduce it in my target later today

FWIW, I tried and failed w/wo autogroup on 4.8 and master.

Me too

Quoted 5 lines

Is it possible to get some dump of /proc/sched_debug while the problem occurs ?

Vincent

Quoted 1 line

-Mike

The output from /proc/shed_debug can be seen here:
http://paste.ubuntu.com/23312351/

Ingo, the latest scheduler bits also still exhibit the bug:
git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.git

Hi,

Quoted 26 lines · Joseph Salisbury

Quoted 25 lines · Vincent Guittot

Quoted 16 lines · Mike Galbraith

Quoted 15 lines · Vincent Guittot

Quoted 14 lines · Ingo Molnar

Quoted 12 lines · Peter Zijlstra

Quoted 11 lines · Joseph Salisbury

Hello Peter,

A kernel bug report was opened against Ubuntu [0]. After a kernel
bisect, it was found that reverting the following commit resolved this bug:

commit 3d30544f02120b884bba2a9466c87dba980e3be5

Author: Peter Zijlstra <[email protected]>
Date:   Tue Jun 21 14:27:50 2016 +0200

Quoted 6 lines

Quoted 5 lines

Quoted 1 line

Quoted 1 line

Quoted 1 line

sched/fair: Apply more PELT fixes

This patch only speeds up the update of task group load in order to reflect the new load balance but It should not change the final value and as a result the final behavior. I will try to reproduce it in my target later today

FWIW, I tried and failed w/wo autogroup on 4.8 and master.

Me too

Quoted 7 lines

Quoted 5 lines

Is it possible to get some dump of /proc/sched_debug while the problem occurs ?

Vincent

Quoted 1 line

-Mike

The output from /proc/shed_debug can be seen here:
http://paste.ubuntu.com/23312351/

I have looked at the dump and there is something very odd for system.slice task group where the display manager is running. system.slice->tg_load_avg is around 381697 but tg_load_avg is normally equal to Sum of system.slice[cpu]->tg_load_avg_contrib whereas Sum of system.slice[cpu]->tg_load_avg_contrib = 1013 in our case. We can have some differences because the dump of /proc/shed_debug is not atomic and some changes can happen but nothing like this difference.

The main effect of this quite high value is that the weight/prio of the sched_entity that represents system.slice in root cfs_rq is very low (lower than task with the smallest nice prio) so the system.slice task group will not get the CPU quite often compared to the user.slice task group: less than 1% for the system.slice where lightDM and xorg are running compared 99% for the user.slice where the stress tasks are running. This is confirmed by the se->avg.util_avg value of the task groups which reflect how much time each task group is effectively running on a CPU:
system.slice[CPU3].se->avg.util_avg = 8 whereas user.slice[CPU3].se->avg.util_avg = 991

This difference of weight/priority explains why the system becomes unresponsive. For now, I can't explain is why system.slice->tg_load_avg = 381697 whereas is should be around 1013 and how the patch can generate this situation.

Is it possible to have a dump of /proc/sched_debug before starting stress command ? to check if the problem is there from the beginning but not seen because not overloaded. Or if it the problem comes when user starts to load the system

Thanks,

Quoted 4 lines

Ingo, the latest scheduler bits also still exhibit the bug:
git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.git

Quoted 63 lines · Vincent Guittot

Hi,

Quoted 26 lines · Joseph Salisbury

Quoted 25 lines · Vincent Guittot

Quoted 16 lines · Mike Galbraith

Quoted 15 lines · Vincent Guittot

Quoted 14 lines · Ingo Molnar

Quoted 12 lines · Peter Zijlstra

Quoted 11 lines · Joseph Salisbury

Hello Peter,

A kernel bug report was opened against Ubuntu [0]. After a kernel
bisect, it was found that reverting the following commit resolved this bug:

commit 3d30544f02120b884bba2a9466c87dba980e3be5

Author: Peter Zijlstra <[email protected]>
Date:   Tue Jun 21 14:27:50 2016 +0200

Quoted 6 lines

Quoted 5 lines

Quoted 1 line

Quoted 1 line

Quoted 1 line

sched/fair: Apply more PELT fixes

This patch only speeds up the update of task group load in order to reflect the new load balance but It should not change the final value and as a result the final behavior. I will try to reproduce it in my target later today

FWIW, I tried and failed w/wo autogroup on 4.8 and master.

Me too

Quoted 7 lines

Quoted 5 lines

Is it possible to get some dump of /proc/sched_debug while the problem occurs ?

Vincent

Quoted 1 line

-Mike

The output from /proc/shed_debug can be seen here:
http://paste.ubuntu.com/23312351/

I have looked at the dump and there is something very odd for system.slice task group where the display manager is running. system.slice->tg_load_avg is around 381697 but tg_load_avg is normally equal to Sum of system.slice[cpu]->tg_load_avg_contrib whereas Sum of system.slice[cpu]->tg_load_avg_contrib = 1013 in our case. We can have some differences because the dump of /proc/shed_debug is not atomic and some changes can happen but nothing like this difference.

The main effect of this quite high value is that the weight/prio of the sched_entity that represents system.slice in root cfs_rq is very low (lower than task with the smallest nice prio) so the system.slice task group will not get the CPU quite often compared to the user.slice task group: less than 1% for the system.slice where lightDM and xorg are running compared 99% for the user.slice where the stress tasks are running. This is confirmed by the se->avg.util_avg value of the task groups which reflect how much time each task group is effectively running on a CPU:
system.slice[CPU3].se->avg.util_avg = 8 whereas user.slice[CPU3].se->avg.util_avg = 991

This difference of weight/priority explains why the system becomes unresponsive. For now, I can't explain is why system.slice->tg_load_avg = 381697 whereas is should be around 1013 and how the patch can generate this situation.

Quoted 4 lines

Is it possible to have a dump of /proc/sched_debug before starting stress command ? to check if the problem is there from the beginning but not seen because not overloaded. Or if it the problem comes when user starts to load the system

Here is the dump before stress is started:
https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1627108/+attachment/4760437/+files/dump_nonbuggy

Here it is after:
https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1627108/+attachment/4760436/+files/dump_buggy

Quoted 8 lines

Thanks,

Quoted 4 lines

Ingo, the latest scheduler bits also still exhibit the bug:
git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.git

Quoted 64 lines · Joseph Salisbury

Quoted 63 lines · Vincent Guittot

Hi,

Quoted 26 lines · Joseph Salisbury

Quoted 25 lines · Vincent Guittot

Quoted 16 lines · Mike Galbraith

Quoted 15 lines · Vincent Guittot

Quoted 14 lines · Ingo Molnar

Quoted 12 lines · Peter Zijlstra

Quoted 11 lines · Joseph Salisbury

Hello Peter,

A kernel bug report was opened against Ubuntu [0]. After a kernel
bisect, it was found that reverting the following commit resolved this bug:

commit 3d30544f02120b884bba2a9466c87dba980e3be5

Author: Peter Zijlstra <[email protected]>
Date:   Tue Jun 21 14:27:50 2016 +0200

Quoted 6 lines

Quoted 5 lines

Quoted 1 line

Quoted 1 line

Quoted 1 line

sched/fair: Apply more PELT fixes

This patch only speeds up the update of task group load in order to reflect the new load balance but It should not change the final value and as a result the final behavior. I will try to reproduce it in my target later today

FWIW, I tried and failed w/wo autogroup on 4.8 and master.

Me too

Quoted 7 lines

Quoted 5 lines

Is it possible to get some dump of /proc/sched_debug while the problem occurs ?

Vincent

Quoted 1 line

-Mike

The output from /proc/shed_debug can be seen here:
http://paste.ubuntu.com/23312351/

I have looked at the dump and there is something very odd for system.slice task group where the display manager is running. system.slice->tg_load_avg is around 381697 but tg_load_avg is normally equal to Sum of system.slice[cpu]->tg_load_avg_contrib whereas Sum of system.slice[cpu]->tg_load_avg_contrib = 1013 in our case. We can have some differences because the dump of /proc/shed_debug is not atomic and some changes can happen but nothing like this difference.

The main effect of this quite high value is that the weight/prio of the sched_entity that represents system.slice in root cfs_rq is very low (lower than task with the smallest nice prio) so the system.slice task group will not get the CPU quite often compared to the user.slice task group: less than 1% for the system.slice where lightDM and xorg are running compared 99% for the user.slice where the stress tasks are running. This is confirmed by the se->avg.util_avg value of the task groups which reflect how much time each task group is effectively running on a CPU:
system.slice[CPU3].se->avg.util_avg = 8 whereas user.slice[CPU3].se->avg.util_avg = 991

This difference of weight/priority explains why the system becomes unresponsive. For now, I can't explain is why system.slice->tg_load_avg = 381697 whereas is should be around 1013 and how the patch can generate this situation.

Quoted 6 lines

Quoted 4 lines

Is it possible to have a dump of /proc/sched_debug before starting stress command ? to check if the problem is there from the beginning but not seen because not overloaded. Or if it the problem comes when user starts to load the system

Here is the dump before stress is started:
https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1627108/+attachment/4760437/+files/dump_nonbuggy

This one is ok.
The dump indicates Sched Debug Version: v0.11, 4.8.0-11-generic #12~lp1627108Commit3d30544Reverted
so this is without the culprit commit

Quoted 4 lines

Here it is after:
https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1627108/+attachment/4760436/+files/dump_buggy

This one has the exact same odds values for system.slice->tg_load_avg than the 1st dump that you sent yesterday
The dump indicates Sched Debug Version: v0.11, 4.8.0-22-generic #24-Ubuntu So this dump has been done with a different kernel than for the dump above. As I can't find any stress task in the dump, i tend to believe that the dump has been done before starting the stress tasks and not after starting them. Can you confirm ?

If i'm right, it mean that the problem was already there before starting stress tasks.

Quoted 91 lines · Vincent Guittot

Quoted 64 lines · Joseph Salisbury

Quoted 63 lines · Vincent Guittot

Hi,

Quoted 26 lines · Joseph Salisbury

Quoted 25 lines · Vincent Guittot

Quoted 16 lines · Mike Galbraith

Quoted 15 lines · Vincent Guittot

Quoted 14 lines · Ingo Molnar

Quoted 12 lines · Peter Zijlstra

Quoted 11 lines · Joseph Salisbury

Hello Peter,

A kernel bug report was opened against Ubuntu [0]. After a kernel
bisect, it was found that reverting the following commit resolved this bug:

commit 3d30544f02120b884bba2a9466c87dba980e3be5

Author: Peter Zijlstra <[email protected]>
Date:   Tue Jun 21 14:27:50 2016 +0200

Quoted 6 lines

Quoted 5 lines

Quoted 1 line

Quoted 1 line

Quoted 1 line

sched/fair: Apply more PELT fixes

This patch only speeds up the update of task group load in order to reflect the new load balance but It should not change the final value and as a result the final behavior. I will try to reproduce it in my target later today

FWIW, I tried and failed w/wo autogroup on 4.8 and master.

Me too

Quoted 7 lines

Quoted 5 lines

Is it possible to get some dump of /proc/sched_debug while the problem occurs ?

Vincent

Quoted 1 line

-Mike

The output from /proc/shed_debug can be seen here:
http://paste.ubuntu.com/23312351/

I have looked at the dump and there is something very odd for system.slice task group where the display manager is running. system.slice->tg_load_avg is around 381697 but tg_load_avg is normally equal to Sum of system.slice[cpu]->tg_load_avg_contrib whereas Sum of system.slice[cpu]->tg_load_avg_contrib = 1013 in our case. We can have some differences because the dump of /proc/shed_debug is not atomic and some changes can happen but nothing like this difference.

The main effect of this quite high value is that the weight/prio of the sched_entity that represents system.slice in root cfs_rq is very low (lower than task with the smallest nice prio) so the system.slice task group will not get the CPU quite often compared to the user.slice task group: less than 1% for the system.slice where lightDM and xorg are running compared 99% for the user.slice where the stress tasks are running. This is confirmed by the se->avg.util_avg value of the task groups which reflect how much time each task group is effectively running on a CPU:
system.slice[CPU3].se->avg.util_avg = 8 whereas user.slice[CPU3].se->avg.util_avg = 991

This difference of weight/priority explains why the system becomes unresponsive. For now, I can't explain is why system.slice->tg_load_avg = 381697 whereas is should be around 1013 and how the patch can generate this situation.

Quoted 6 lines

Quoted 4 lines

Is it possible to have a dump of /proc/sched_debug before starting stress command ? to check if the problem is there from the beginning but not seen because not overloaded. Or if it the problem comes when user starts to load the system

Here is the dump before stress is started:
https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1627108/+attachment/4760437/+files/dump_nonbuggy

This one is ok.
The dump indicates Sched Debug Version: v0.11, 4.8.0-11-generic #12~lp1627108Commit3d30544Reverted
so this is without the culprit commit

Quoted 4 lines

Here it is after:
https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1627108/+attachment/4760436/+files/dump_buggy

This one has the exact same odds values for system.slice->tg_load_avg than the 1st dump that you sent yesterday
The dump indicates Sched Debug Version: v0.11, 4.8.0-22-generic #24-Ubuntu So this dump has been done with a different kernel than for the dump above. As I can't find any stress task in the dump, i tend to believe that the dump has been done before starting the stress tasks and not after starting them. Can you confirm ?

Quoted 2 lines

If i'm right, it mean that the problem was already there before starting stress tasks.

Could it be a problem I'm also seeing on my ARM64 Juno (6 logical cpus) w/o systemd and w/o autogroup (tip/sched/core 447976ef4fd0):

When I create a tg_root/tg_x/tg_y_1 and a tg_root/tg_x/tg_y_2 group, the tg_x->load_avg becomes > 6*1024 before any tasks ran in it.

tg_x : 0xffff800975800d80
tg_y_1 : 0xffff800975800c00
tg_y_2 : 0xffff80097543d200

mkdir-2177 [002] 117.235241: bprint: sched_online_group: tg=0xffff800975800d80 tg->parent=0xffff000008fd0300 mkdir-2177 [002] 117.235244: bprint: online_fair_sched_group: tg=0xffff800975800d80 cpu=0 mkdir-2177 [002] 117.235247: bprint: online_fair_sched_group: tg=0xffff800975800d80 cpu=1 mkdir-2177 [002] 117.235249: bprint: online_fair_sched_group: tg=0xffff800975800d80 cpu=2 mkdir-2177 [002] 117.235251: bprint: online_fair_sched_group: tg=0xffff800975800d80 cpu=3 mkdir-2177 [002] 117.235253: bprint: online_fair_sched_group: tg=0xffff800975800d80 cpu=4 mkdir-2177 [002] 117.235255: bprint: online_fair_sched_group: tg=0xffff800975800d80 cpu=5

mkdir-2181 [002] 117.353155: bprint: sched_online_group: tg=0xffff800975800c00 tg->parent=0xffff800975800d80
mkdir-2181 [002] 117.353158: bprint: online_fair_sched_group: tg=0xffff800975800c00 cpu=0
mkdir-2181 [002] 117.353162: bprint: post_init_entity_util_avg: cpu=0 tg=0xffff800975800d80 tg_css_id=2 cfs_rq->tg->load_avg=0 cfs_rq->avg.load_avg=1024 cfs_rq->tg_load_avg_contrib=0 delta=1024 mkdir-2181 [002] 117.353164: bprint: online_fair_sched_group: tg=0xffff800975800c00 cpu=1
mkdir-2181 [002] 117.353167: bprint: post_init_entity_util_avg: cpu=1 tg=0xffff800975800d80 tg_css_id=2 cfs_rq->tg->load_avg=1024 cfs_rq->avg.load_avg=1024 cfs_rq->tg_load_avg_contrib=0 delta=1024 mkdir-2181 [002] 117.353168: bprint: online_fair_sched_group: tg=0xffff800975800c00 cpu=2
mkdir-2181 [002] 117.353171: bprint: post_init_entity_util_avg: cpu=2 tg=0xffff800975800d80 tg_css_id=2 cfs_rq->tg->load_avg=2048 cfs_rq->avg.load_avg=1024 cfs_rq->tg_load_avg_contrib=0 delta=1024 mkdir-2181 [002] 117.353173: bprint: online_fair_sched_group: tg=0xffff800975800c00 cpu=3
mkdir-2181 [002] 117.353175: bprint: post_init_entity_util_avg: cpu=3 tg=0xffff800975800d80 tg_css_id=2 cfs_rq->tg->load_avg=3072 cfs_rq->avg.load_avg=1024 cfs_rq->tg_load_avg_contrib=0 delta=1024 mkdir-2181 [002] 117.353177: bprint: online_fair_sched_group: tg=0xffff800975800c00 cpu=4
mkdir-2181 [002] 117.353179: bprint: post_init_entity_util_avg: cpu=4 tg=0xffff800975800d80 tg_css_id=2 cfs_rq->tg->load_avg=4096 cfs_rq->avg.load_avg=1024 cfs_rq->tg_load_avg_contrib=0 delta=1024 mkdir-2181 [002] 117.353180: bprint: online_fair_sched_group: tg=0xffff800975800c00 cpu=5
mkdir-2181 [002] 117.353183: bprint: post_init_entity_util_avg: cpu=5 tg=0xffff800975800d80 tg_css_id=2 cfs_rq->tg->load_avg=5120 cfs_rq->avg.load_avg=1024 cfs_rq->tg_load_avg_contrib=0 delta=1024

mkdir-2185 [001] 117.502980: bprint: sched_online_group: tg=0xffff80097543d200 tg->parent=0xffff800975800d80
mkdir-2185 [001] 117.502982: bprint: online_fair_sched_group: tg=0xffff80097543d200 cpu=0
mkdir-2185 [001] 117.502987: bprint: post_init_entity_util_avg: cpu=0 tg=0xffff800975800d80 tg_css_id=2 cfs_rq->tg->load_avg=6144 cfs_rq->avg.load_avg=1068 cfs_rq->tg_load_avg_contrib=1024 delta=44 mkdir-2185 [001] 117.502988: bprint: online_fair_sched_group: tg=0xffff80097543d200 cpu=1
mkdir-2185 [001] 117.502992: bprint: post_init_entity_util_avg: cpu=1 tg=0xffff800975800d80 tg_css_id=2 cfs_rq->tg->load_avg=6188 cfs_rq->avg.load_avg=1058 cfs_rq->tg_load_avg_contrib=1024 delta=34 mkdir-2185 [001] 117.502993: bprint: online_fair_sched_group: tg=0xffff80097543d200 cpu=2
mkdir-2185 [001] 117.502996: bprint: post_init_entity_util_avg: cpu=2 tg=0xffff800975800d80 tg_css_id=2 cfs_rq->tg->load_avg=6222 cfs_rq->avg.load_avg=1092 cfs_rq->tg_load_avg_contrib=1024 delta=68 mkdir-2185 [001] 117.502998: bprint: online_fair_sched_group: tg=0xffff80097543d200 cpu=3
mkdir-2185 [001] 117.503001: bprint: post_init_entity_util_avg: cpu=3 tg=0xffff800975800d80 tg_css_id=2 cfs_rq->tg->load_avg=6290 cfs_rq->avg.load_avg=1069 cfs_rq->tg_load_avg_contrib=1024 delta=45 mkdir-2185 [001] 117.503002: bprint: online_fair_sched_group: tg=0xffff80097543d200 cpu=4
mkdir-2185 [001] 117.503005: bprint: post_init_entity_util_avg: cpu=4 tg=0xffff800975800d80 tg_css_id=2 cfs_rq->tg->load_avg=6335 cfs_rq->avg.load_avg=1064 cfs_rq->tg_load_avg_contrib=1024 delta=40 mkdir-2185 [001] 117.503006: bprint: online_fair_sched_group: tg=0xffff80097543d200 cpu=5
mkdir-2185 [001] 117.503009: bprint: post_init_entity_util_avg: cpu=5 tg=0xffff800975800d80 tg_css_id=2 cfs_rq->tg->load_avg=*6375* cfs_rq->avg.load_avg=1061 cfs_rq->tg_load_avg_contrib=1024 delta=37

We add delta=1024 to tg_x->load_avg during tg_root/tg_x/tg_y_1 initialization but only delta=~40 for tg_root/tg_x/tg_y_2.

system.slice->tg_load_avg = 381697 is still pretty much higher and AFAICS, it's an i5-5300U CPU so only 4 logical cpus.

How many system.slice/foo's actually exists in this system? I only see user.slice

related cfs_rq[x]:/autogroups-xxx on my Ubuntu-16.04 Desktop system.

[...]

Quoted 92 lines · Dietmar Eggemann

Quoted 91 lines · Vincent Guittot

Quoted 64 lines · Joseph Salisbury

Quoted 63 lines · Vincent Guittot

Hi,

Quoted 26 lines · Joseph Salisbury

Quoted 25 lines · Vincent Guittot

Quoted 16 lines · Mike Galbraith

Quoted 15 lines · Vincent Guittot

Quoted 14 lines · Ingo Molnar

Quoted 12 lines · Peter Zijlstra

Quoted 11 lines · Joseph Salisbury

Hello Peter,

A kernel bug report was opened against Ubuntu [0]. After a kernel
bisect, it was found that reverting the following commit resolved this bug:

commit 3d30544f02120b884bba2a9466c87dba980e3be5

Author: Peter Zijlstra <[email protected]>
Date:   Tue Jun 21 14:27:50 2016 +0200

Quoted 6 lines

Quoted 5 lines

Quoted 1 line

Quoted 1 line

Quoted 1 line

sched/fair: Apply more PELT fixes

This patch only speeds up the update of task group load in order to reflect the new load balance but It should not change the final value and as a result the final behavior. I will try to reproduce it in my target later today

FWIW, I tried and failed w/wo autogroup on 4.8 and master.

Me too

Quoted 7 lines

Quoted 5 lines

Is it possible to get some dump of /proc/sched_debug while the problem occurs ?

Vincent

Quoted 1 line

-Mike

The output from /proc/shed_debug can be seen here:
http://paste.ubuntu.com/23312351/

I have looked at the dump and there is something very odd for system.slice task group where the display manager is running. system.slice->tg_load_avg is around 381697 but tg_load_avg is normally equal to Sum of system.slice[cpu]->tg_load_avg_contrib whereas Sum of system.slice[cpu]->tg_load_avg_contrib = 1013 in our case. We can have some differences because the dump of /proc/shed_debug is not atomic and some changes can happen but nothing like this difference.

The main effect of this quite high value is that the weight/prio of the sched_entity that represents system.slice in root cfs_rq is very low (lower than task with the smallest nice prio) so the system.slice task group will not get the CPU quite often compared to the user.slice task group: less than 1% for the system.slice where lightDM and xorg are running compared 99% for the user.slice where the stress tasks are running. This is confirmed by the se->avg.util_avg value of the task groups which reflect how much time each task group is effectively running on a CPU:
system.slice[CPU3].se->avg.util_avg = 8 whereas user.slice[CPU3].se->avg.util_avg = 991

This difference of weight/priority explains why the system becomes unresponsive. For now, I can't explain is why system.slice->tg_load_avg = 381697 whereas is should be around 1013 and how the patch can generate this situation.

Quoted 6 lines

Quoted 4 lines

Is it possible to have a dump of /proc/sched_debug before starting stress command ? to check if the problem is there from the beginning but not seen because not overloaded. Or if it the problem comes when user starts to load the system

Here is the dump before stress is started:
https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1627108/+attachment/4760437/+files/dump_nonbuggy

This one is ok.
The dump indicates Sched Debug Version: v0.11, 4.8.0-11-generic #12~lp1627108Commit3d30544Reverted
so this is without the culprit commit

Quoted 4 lines

Here it is after:
https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1627108/+attachment/4760436/+files/dump_buggy

This one has the exact same odds values for system.slice->tg_load_avg than the 1st dump that you sent yesterday
The dump indicates Sched Debug Version: v0.11, 4.8.0-22-generic #24-Ubuntu So this dump has been done with a different kernel than for the dump above. As I can't find any stress task in the dump, i tend to believe that the dump has been done before starting the stress tasks and not after starting them. Can you confirm ?

Quoted 8 lines

Quoted 2 lines

If i'm right, it mean that the problem was already there before starting stress tasks.

Could it be a problem I'm also seeing on my ARM64 Juno (6 logical cpus) w/o systemd and w/o autogroup (tip/sched/core 447976ef4fd0):

When I create a tg_root/tg_x/tg_y_1 and a tg_root/tg_x/tg_y_2 group, the tg_x->load_avg becomes > 6*1024 before any tasks ran in it.

This is normal as se->avg.load_avg is initialized to scale_load_down(se->load.weight) and this se->avg.load_avg will be added to tg_x[cpu]->cfs_rq->avg.load_avg when attached to the cfs_rq

Quoted 13 lines

tg_x : 0xffff800975800d80
tg_y_1 : 0xffff800975800c00
tg_y_2 : 0xffff80097543d200

mkdir-2177 [002] 117.235241: bprint: sched_online_group: tg=0xffff800975800d80 tg->parent=0xffff000008fd0300 mkdir-2177 [002] 117.235244: bprint: online_fair_sched_group: tg=0xffff800975800d80 cpu=0 mkdir-2177 [002] 117.235247: bprint: online_fair_sched_group: tg=0xffff800975800d80 cpu=1 mkdir-2177 [002] 117.235249: bprint: online_fair_sched_group: tg=0xffff800975800d80 cpu=2 mkdir-2177 [002] 117.235251: bprint: online_fair_sched_group: tg=0xffff800975800d80 cpu=3 mkdir-2177 [002] 117.235253: bprint: online_fair_sched_group: tg=0xffff800975800d80 cpu=4 mkdir-2177 [002] 117.235255: bprint: online_fair_sched_group: tg=0xffff800975800d80 cpu=5

Quoted 14 lines

mkdir-2181 [002] 117.353155: bprint: sched_online_group: tg=0xffff800975800c00 tg->parent=0xffff800975800d80
mkdir-2181 [002] 117.353158: bprint: online_fair_sched_group: tg=0xffff800975800c00 cpu=0
mkdir-2181 [002] 117.353162: bprint: post_init_entity_util_avg: cpu=0 tg=0xffff800975800d80 tg_css_id=2 cfs_rq->tg->load_avg=0 cfs_rq->avg.load_avg=1024 cfs_rq->tg_load_avg_contrib=0 delta=1024 mkdir-2181 [002] 117.353164: bprint: online_fair_sched_group: tg=0xffff800975800c00 cpu=1
mkdir-2181 [002] 117.353167: bprint: post_init_entity_util_avg: cpu=1 tg=0xffff800975800d80 tg_css_id=2 cfs_rq->tg->load_avg=1024 cfs_rq->avg.load_avg=1024 cfs_rq->tg_load_avg_contrib=0 delta=1024 mkdir-2181 [002] 117.353168: bprint: online_fair_sched_group: tg=0xffff800975800c00 cpu=2
mkdir-2181 [002] 117.353171: bprint: post_init_entity_util_avg: cpu=2 tg=0xffff800975800d80 tg_css_id=2 cfs_rq->tg->load_avg=2048 cfs_rq->avg.load_avg=1024 cfs_rq->tg_load_avg_contrib=0 delta=1024 mkdir-2181 [002] 117.353173: bprint: online_fair_sched_group: tg=0xffff800975800c00 cpu=3
mkdir-2181 [002] 117.353175: bprint: post_init_entity_util_avg: cpu=3 tg=0xffff800975800d80 tg_css_id=2 cfs_rq->tg->load_avg=3072 cfs_rq->avg.load_avg=1024 cfs_rq->tg_load_avg_contrib=0 delta=1024 mkdir-2181 [002] 117.353177: bprint: online_fair_sched_group: tg=0xffff800975800c00 cpu=4
mkdir-2181 [002] 117.353179: bprint: post_init_entity_util_avg: cpu=4 tg=0xffff800975800d80 tg_css_id=2 cfs_rq->tg->load_avg=4096 cfs_rq->avg.load_avg=1024 cfs_rq->tg_load_avg_contrib=0 delta=1024 mkdir-2181 [002] 117.353180: bprint: online_fair_sched_group: tg=0xffff800975800c00 cpu=5
mkdir-2181 [002] 117.353183: bprint: post_init_entity_util_avg: cpu=5 tg=0xffff800975800d80 tg_css_id=2 cfs_rq->tg->load_avg=5120 cfs_rq->avg.load_avg=1024 cfs_rq->tg_load_avg_contrib=0 delta=1024

This looks coherent as tg_x->load_avg = 5120 == sum of cfs_rq[cpu]->load_avg and cfs_rq->tg_load_avg_contrib will be set to cfs_rq[cpu]->load_avg at the end of the function call
In the buggy case, tg->load_avg = 381697 but sum of cfs_rq[cpu]->tg_load_avg_contrib = 1013 = cfs_rq[cpu]->avg.load_avg so there is no way to remove the 381697

Quoted 14 lines

mkdir-2185 [001] 117.502980: bprint: sched_online_group: tg=0xffff80097543d200 tg->parent=0xffff800975800d80
mkdir-2185 [001] 117.502982: bprint: online_fair_sched_group: tg=0xffff80097543d200 cpu=0
mkdir-2185 [001] 117.502987: bprint: post_init_entity_util_avg: cpu=0 tg=0xffff800975800d80 tg_css_id=2 cfs_rq->tg->load_avg=6144 cfs_rq->avg.load_avg=1068 cfs_rq->tg_load_avg_contrib=1024 delta=44 mkdir-2185 [001] 117.502988: bprint: online_fair_sched_group: tg=0xffff80097543d200 cpu=1
mkdir-2185 [001] 117.502992: bprint: post_init_entity_util_avg: cpu=1 tg=0xffff800975800d80 tg_css_id=2 cfs_rq->tg->load_avg=6188 cfs_rq->avg.load_avg=1058 cfs_rq->tg_load_avg_contrib=1024 delta=34 mkdir-2185 [001] 117.502993: bprint: online_fair_sched_group: tg=0xffff80097543d200 cpu=2
mkdir-2185 [001] 117.502996: bprint: post_init_entity_util_avg: cpu=2 tg=0xffff800975800d80 tg_css_id=2 cfs_rq->tg->load_avg=6222 cfs_rq->avg.load_avg=1092 cfs_rq->tg_load_avg_contrib=1024 delta=68 mkdir-2185 [001] 117.502998: bprint: online_fair_sched_group: tg=0xffff80097543d200 cpu=3
mkdir-2185 [001] 117.503001: bprint: post_init_entity_util_avg: cpu=3 tg=0xffff800975800d80 tg_css_id=2 cfs_rq->tg->load_avg=6290 cfs_rq->avg.load_avg=1069 cfs_rq->tg_load_avg_contrib=1024 delta=45 mkdir-2185 [001] 117.503002: bprint: online_fair_sched_group: tg=0xffff80097543d200 cpu=4
mkdir-2185 [001] 117.503005: bprint: post_init_entity_util_avg: cpu=4 tg=0xffff800975800d80 tg_css_id=2 cfs_rq->tg->load_avg=6335 cfs_rq->avg.load_avg=1064 cfs_rq->tg_load_avg_contrib=1024 delta=40 mkdir-2185 [001] 117.503006: bprint: online_fair_sched_group: tg=0xffff80097543d200 cpu=5
mkdir-2185 [001] 117.503009: bprint: post_init_entity_util_avg: cpu=5 tg=0xffff800975800d80 tg_css_id=2 cfs_rq->tg->load_avg=*6375* cfs_rq->avg.load_avg=1061 cfs_rq->tg_load_avg_contrib=1024 delta=37

Quoted 2 lines

We add delta=1024 to tg_x->load_avg during tg_root/tg_x/tg_y_1 initialization but only delta=~40 for tg_root/tg_x/tg_y_2.

IMO, this comes from the decay of tg_y_2->se[cpu]->avg.load_avg before being attached to tg_x

But I don't think that it's the same issue as the tg_x->load_avg is still coherent with the sum of cfs_rq->avg.load_avg and the sume of cfs_rq->tg_load_avg_contrib

Quoted 1 line

Quoted 4 lines

system.slice->tg_load_avg = 381697 is still pretty much higher and AFAICS, it's an i5-5300U CPU so only 4 logical cpus.

How many system.slice/foo's actually exists in this system? I only see user.slice

There is several dozen of childs in system.slice but all

Quoted 4 lines

related cfs_rq[x]:/autogroups-xxx on my Ubuntu-16.04 Desktop system.

[...]

Quoted 16 lines · Vincent Guittot

Quoted 15 lines · Dietmar Eggemann

Quoted 14 lines · Vincent Guittot

Quoted 12 lines · Joseph Salisbury

Quoted 11 lines · Vincent Guittot

Hi,

Quoted 7 lines · Joseph Salisbury

Quoted 6 lines · Vincent Guittot

Quoted 5 lines · Mike Galbraith

Quoted 4 lines · Vincent Guittot

Quoted 3 lines · Ingo Molnar

Quoted 1 line · Peter Zijlstra

On Fri, Oct 07, 2016 at 03:38:23PM -0400, Joseph Salisbury wrote:

Quoted 6 lines

Quoted 2 lines

When I create a tg_root/tg_x/tg_y_1 and a tg_root/tg_x/tg_y_2 group, the tg_x->load_avg becomes > 6*1024 before any tasks ran in it.

This is normal as se->avg.load_avg is initialized to scale_load_down(se->load.weight) and this se->avg.load_avg will be added to tg_x[cpu]->cfs_rq->avg.load_avg when attached to the cfs_rq

Quoted 17 lines · Vincent Guittot

Quoted 16 lines · Vincent Guittot

Quoted 15 lines · Dietmar Eggemann

Quoted 14 lines · Vincent Guittot

Quoted 12 lines · Joseph Salisbury

Quoted 11 lines · Vincent Guittot

Hi,

Quoted 7 lines · Joseph Salisbury

Quoted 6 lines · Vincent Guittot

Quoted 5 lines · Mike Galbraith

Quoted 4 lines · Vincent Guittot

Quoted 3 lines · Ingo Molnar

Quoted 1 line · Peter Zijlstra

On Fri, Oct 07, 2016 at 03:38:23PM -0400, Joseph Salisbury wrote:

[...]

Quoted 6 lines

Quoted 6 lines

Quoted 2 lines

When I create a tg_root/tg_x/tg_y_1 and a tg_root/tg_x/tg_y_2 group, the tg_x->load_avg becomes > 6*1024 before any tasks ran in it.

This is normal as se->avg.load_avg is initialized to scale_load_down(se->load.weight) and this se->avg.load_avg will be added to tg_x[cpu]->cfs_rq->avg.load_avg when attached to the cfs_rq

Yeah, you right, even when I've created 50 second level groups, tg_x->load_avg is ~6800.

Could it have something to do with the fact that .se->load.weight = 2 for all these task groups? on a 64bit system?

In case we call __update_load_avg(..., se->on_rq * scale_load_down(se->load.weight), ...) we pass a weight argument of 0 for these se's.

Does not happen with:

(quoted diff) +2 -2

-       if (shares < MIN_SHARES)
-               shares = MIN_SHARES;
+       if (shares < scale_load(MIN_SHARES))
+               shares = scale_load(MIN_SHARES);

in calc_cfs_shares().

[...]

Quoted 28 lines · Dietmar Eggemann

Quoted 17 lines · Vincent Guittot

Quoted 16 lines · Vincent Guittot

Quoted 15 lines · Dietmar Eggemann

Quoted 14 lines · Vincent Guittot

Quoted 12 lines · Joseph Salisbury

Quoted 11 lines · Vincent Guittot

Hi,

Quoted 7 lines · Joseph Salisbury

Quoted 6 lines · Vincent Guittot

Quoted 5 lines · Mike Galbraith

Quoted 4 lines · Vincent Guittot

Quoted 3 lines · Ingo Molnar

* Peter Zijlstra <[email protected]> wrote:

On Fri, Oct 07, 2016 at 03:38:23PM -0400, Joseph Salisbury wrote:

[...]

Quoted 5 lines

Quoted 5 lines

Quoted 2 lines

When I create a tg_root/tg_x/tg_y_1 and a tg_root/tg_x/tg_y_2 group, the tg_x->load_avg becomes > 6*1024 before any tasks ran in it.

This is normal as se->avg.load_avg is initialized to scale_load_down(se->load.weight) and this se->avg.load_avg will be added to tg_x[cpu]->cfs_rq->avg.load_avg when attached to the cfs_rq

Yeah, you right, even when I've created 50 second level groups, tg_x->load_avg is ~6800.

Quoted 2 lines

Could it have something to do with the fact that .se->load.weight = 2 for all these task groups? on a 64bit system?

I don't think so, the problem really comes from tg->load_avg = 381697 but sum of cfs_rq[cpu]->tg_load_avg_contrib = 1013 which is << tg->load_avg and cfs_rq[cpu]->tg_load_avg_contrib == cfs_rq[cpu]->avg.load_avg so we can't expect any negative delta to remove this large value

Quoted 16 lines

In case we call __update_load_avg(..., se->on_rq * scale_load_down(se->load.weight), ...) we pass a weight argument of 0 for these se's.

Does not happen with:

(quoted diff) +2 -2

-       if (shares < MIN_SHARES)
-               shares = MIN_SHARES;
+       if (shares < scale_load(MIN_SHARES))
+               shares = scale_load(MIN_SHARES);

in calc_cfs_shares().

[...]

Quoted 47 lines · Vincent Guittot

Quoted 22 lines · Dietmar Eggemann

Quoted 17 lines · Vincent Guittot

Quoted 16 lines · Vincent Guittot

Quoted 15 lines · Dietmar Eggemann

Quoted 14 lines · Vincent Guittot

Quoted 12 lines · Joseph Salisbury

Quoted 11 lines · Vincent Guittot

Hi,

Quoted 7 lines · Joseph Salisbury

Quoted 6 lines · Vincent Guittot

Quoted 5 lines · Mike Galbraith

Quoted 4 lines · Vincent Guittot

On 8 October 2016 at 10:39, Ingo Molnar <[email protected]> wrote:
* Peter Zijlstra <[email protected]> wrote:

On Fri, Oct 07, 2016 at 03:38:23PM -0400, Joseph Salisbury wrote:

[...]

Quoted 2 lines

Quoted 2 lines

Quoted 2 lines

When I create a tg_root/tg_x/tg_y_1 and a tg_root/tg_x/tg_y_2 group, the tg_x->load_avg becomes > 6*1024 before any tasks ran in it.

Quoted 8 lines

Quoted 3 lines

Quoted 3 lines

This is normal as se->avg.load_avg is initialized to scale_load_down(se->load.weight) and this se->avg.load_avg will be added to tg_x[cpu]->cfs_rq->avg.load_avg when attached to the cfs_rq

Yeah, you right, even when I've created 50 second level groups, tg_x->load_avg is ~6800.

Could it have something to do with the fact that .se->load.weight = 2 for all these task groups? on a 64bit system?

I don't think so, the problem really comes from tg->load_avg = 381697 but sum of cfs_rq[cpu]->tg_load_avg_contrib = 1013 which is << tg->load_avg and cfs_rq[cpu]->tg_load_avg_contrib == cfs_rq[cpu]->avg.load_avg so we can't expect any negative delta to remove this large value

Quoted 11 lines

In case we call __update_load_avg(..., se->on_rq * scale_load_down(se->load.weight), ...) we pass a weight argument of 0 for these se's.

Does not happen with:

(quoted diff) +2 -2

-       if (shares < MIN_SHARES)
-               shares = MIN_SHARES;
+       if (shares < scale_load(MIN_SHARES))
+               shares = scale_load(MIN_SHARES);

Quoted 5 lines

Quoted 5 lines

in calc_cfs_shares().

[...]

Adding Omer to CC list, as he is able to reproduce this bug.

Quoted 3 lines · Joseph Salisbury

Quoted 2 lines · Vincent Guittot

Quoted 1 line · Dietmar Eggemann

On 14/10/16 09:24, Vincent Guittot wrote:

Could you try the patch below on top of the faulty kernel ?

---
 kernel/sched/fair.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

kernel/sched/fair.c +2 -1

index 8b03fb5..8926685 100644
@@ -2902,7 +2902,8 @@ __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
  */
 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
 {
-	long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
+	unsigned long load_avg = READ_ONCE(cfs_rq->avg.load_avg);
+	long delta = load_avg - cfs_rq->tg_load_avg_contrib;
 
 	/*
 	 * No need to update load_avg for root_task_group as it is not used.

Quoted 16 lines

kernel/sched/fair.c +2 -1

index 8b03fb5..8926685 100644
@@ -2902,7 +2902,8 @@ __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
  */
 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
 {
-	long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
+	unsigned long load_avg = READ_ONCE(cfs_rq->avg.load_avg);
+	long delta = load_avg - cfs_rq->tg_load_avg_contrib;
 
 	/*
 	 * No need to update load_avg for root_task_group as it is not used.

@@ -2912,7 +2913,7 @@ static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
 

kernel/sched/fair.c static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) · line 2918

Quoted 6 lines

kernel/sched/fair.c +1 -1

 	if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
 		atomic_long_add(delta, &cfs_rq->tg->load_avg);
-		cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
+		cfs_rq->tg_load_avg_contrib = load_avg;
 	}
 }

Tested it on an Ubuntu 16.10 Server (on top of the default 4.8.0-22-generic kernel) on a Lenovo T430 and it didn't help.

What seems to cure it is to get rid of this snippet (part of the commit mentioned earlier in this thread: 3d30544f0212):

kernel/sched/fair.c +1 -4

index 039de34f1521..16c692049fbf 100644
@@ -726,7 +726,6 @@ void post_init_entity_util_avg(struct sched_entity *se)
        struct sched_avg *sa = &se->avg;
        long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2;
        u64 now = cfs_rq_clock_task(cfs_rq);
-       int tg_update;
 
        if (cap > 0) {
                if (cfs_rq->avg.util_avg != 0) {

@@ -759,10 +758,8 @@ void post_init_entity_util_avg(struct sched_entity *se)
                }
        }
 
-       tg_update = update_cfs_rq_load_avg(now, cfs_rq, false);
+       update_cfs_rq_load_avg(now, cfs_rq, false);
        attach_entity_load_avg(cfs_rq, se);
-       if (tg_update)
-               update_tg_load_avg(cfs_rq, false);
 }
 
 #else /* !CONFIG_SMP */

BTW, I guess we can reach .tg_load_avg up to ~300000-400000 on such a system initially because systemd will create all ~100 services (and therefore the corresponding 2. level tg's) at once. In my previous example, there was 500ms between the creation of 2 tg's so there was a lot of decaying going on in between.

Quoted 16 lines · Dietmar Eggemann

Quoted 16 lines

kernel/sched/fair.c +2 -1

index 8b03fb5..8926685 100644
@@ -2902,7 +2902,8 @@ __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
  */
 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
 {
-   long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
+   unsigned long load_avg = READ_ONCE(cfs_rq->avg.load_avg);
+   long delta = load_avg - cfs_rq->tg_load_avg_contrib;
 
    /*
     * No need to update load_avg for root_task_group as it is not used.

@@ -2912,7 +2913,7 @@ static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
 

Quoted 9 lines

kernel/sched/fair.c static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) · line 2918

Quoted 6 lines

kernel/sched/fair.c +1 -1

    if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
            atomic_long_add(delta, &cfs_rq->tg->load_avg);
-           cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
+           cfs_rq->tg_load_avg_contrib = load_avg;
    }
 }

Tested it on an Ubuntu 16.10 Server (on top of the default 4.8.0-22-generic kernel) on a Lenovo T430 and it didn't help.

Right, I don't think that race exists, we update cfs_rq->avg.load_avg with rq->lock held and have it held here, so it cannot change under us.

This might do with a few lockdep_assert_held() instances to clarify this though.

Quoted 28 lines

What seems to cure it is to get rid of this snippet (part of the commit mentioned earlier in this thread: 3d30544f0212):

kernel/sched/fair.c +1 -4

index 039de34f1521..16c692049fbf 100644
@@ -726,7 +726,6 @@ void post_init_entity_util_avg(struct sched_entity *se)
        struct sched_avg *sa = &se->avg;
        long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2;
        u64 now = cfs_rq_clock_task(cfs_rq);
-       int tg_update;
 
        if (cap > 0) {
                if (cfs_rq->avg.util_avg != 0) {

@@ -759,10 +758,8 @@ void post_init_entity_util_avg(struct sched_entity *se)
                }
        }
 
-       tg_update = update_cfs_rq_load_avg(now, cfs_rq, false);
+       update_cfs_rq_load_avg(now, cfs_rq, false);
        attach_entity_load_avg(cfs_rq, se);
-       if (tg_update)
-               update_tg_load_avg(cfs_rq, false);
 }
 
 #else /* !CONFIG_SMP */

Quoted 4 lines

BTW, I guess we can reach .tg_load_avg up to ~300000-400000 on such a system initially because systemd will create all ~100 services (and therefore the corresponding 2. level tg's) at once. In my previous example, there was 500ms between the creation of 2 tg's so there was a lot of decaying going on in between.

Cute... on current kernels that translates to simply removing the call to update_tg_load_avg(), lets see if we can figure out what goes sideways first though, because it _should_ decay back out. And if that

can fail here, I'm not seeing why that wouldn't fail elsewhere either.

I'll see if I can reproduce this with a script creating heaps of cgroups in a hurry, I have a total lack of system-disease on all my machines.

/me goes prod..

Quoted 1 line · Peter Zijlstra

On Mon, Oct 17, 2016 at 12:49:55PM +0100, Dietmar Eggemann wrote:

Quoted 5 lines

Quoted 4 lines

BTW, I guess we can reach .tg_load_avg up to ~300000-400000 on such a system initially because systemd will create all ~100 services (and therefore the corresponding 2. level tg's) at once. In my previous example, there was 500ms between the creation of 2 tg's so there was a lot of decaying going on in between.

Quoted 3 lines

Cute... on current kernels that translates to simply removing the call to update_tg_load_avg(), lets see if we can figure out what goes sideways first though, because it _should_ decay back out. And if that

yes, Reaching ~300000-400000 is not an issue in itself, the problem is that load_avg has decayed but it has not been reflected in tg->load_avg in the buggy case

Quoted 4 lines

can fail here, I'm not seeing why that wouldn't fail elsewhere either.

I'll see if I can reproduce this with a script creating heaps of cgroups in a hurry, I have a total lack of system-disease on all my machines.

Quoted 2 lines · Vincent Guittot

Quoted 1 line · Peter Zijlstra

On Mon, Oct 17, 2016 at 12:49:55PM +0100, Dietmar Eggemann wrote:

[...]

Quoted 9 lines

Quoted 8 lines

Quoted 4 lines

BTW, I guess we can reach .tg_load_avg up to ~300000-400000 on such a system initially because systemd will create all ~100 services (and therefore the corresponding 2. level tg's) at once. In my previous example, there was 500ms between the creation of 2 tg's so there was a lot of decaying going on in between.

Cute... on current kernels that translates to simply removing the call to update_tg_load_avg(), lets see if we can figure out what goes sideways first though, because it _should_ decay back out. And if that

Quoted 8 lines

yes, Reaching ~300000-400000 is not an issue in itself, the problem is that load_avg has decayed but it has not been reflected in tg->load_avg in the buggy case

Quoted 4 lines

can fail here, I'm not seeing why that wouldn't fail elsewhere either.

I'll see if I can reproduce this with a script creating heaps of cgroups in a hurry, I have a total lack of system-disease on all my machines.

Something looks weird related to the use of for_each_possible_cpu(i) in online_fair_sched_group() on my i5-3320M CPU (4 logical cpus).

In case I print out cpu id and the cpu masks inside the for_each_possible_cpu(i) I get:

[ 5.462368]  cpu=0  cpu_possible_mask=0-7 cpu_online_mask=0-3 cpu_present_mask=0-3 cpu_active_mask=0-3
[ 5.462370]  cpu=1  cpu_possible_mask=0-7 cpu_online_mask=0-3 cpu_present_mask=0-3 cpu_active_mask=0-3
[ 5.462370]  cpu=2  cpu_possible_mask=0-7 cpu_online_mask=0-3 cpu_present_mask=0-3 cpu_active_mask=0-3
[ 5.462371]  cpu=3  cpu_possible_mask=0-7 cpu_online_mask=0-3 cpu_present_mask=0-3 cpu_active_mask=0-3
[ 5.462372] *cpu=4* cpu_possible_mask=0-7 cpu_online_mask=0-3 cpu_present_mask=0-3 cpu_active_mask=0-3
[ 5.462373] *cpu=5* cpu_possible_mask=0-7 cpu_online_mask=0-3 cpu_present_mask=0-3 cpu_active_mask=0-3
[ 5.462374] *cpu=6* cpu_possible_mask=0-7 cpu_online_mask=0-3 cpu_present_mask=0-3 cpu_active_mask=0-3
[ 5.462375] *cpu=7* cpu_possible_mask=0-7 cpu_online_mask=0-3 cpu_present_mask=0-3 cpu_active_mask=0-3

T430:/sys/fs/cgroup/cpu,cpuacct/system.slice# ls -l | grep '^d' | wc -l
80

/proc/sched_debug:

cfs_rq[0]:/system.slice
...
.tg_load_avg : 323584
...

80 * 1024 * 4 (not existent cpu4-cpu7) = 327680 (with a little bit of decay, this could be this extra load on the systen.slice tg)

Using for_each_online_cpu(i) instead of for_each_possible_cpu(i) in online_fair_sched_group() works on this machine, i.e. the .tg_load_avg of system.slice tg is 0 after startup.

Something looks weird related to the use of for_each_possible_cpu(i) in online_fair_sched_group() on my i5-3320M CPU (4 logical cpus).

In case I print out cpu id and the cpu masks inside the for_each_possible_cpu(i) I get:

[ 5.462368]  cpu=0  cpu_possible_mask=0-7 cpu_online_mask=0-3 cpu_present_mask=0-3 cpu_active_mask=0-3

[ 5.462370]  cpu=1  cpu_possible_mask=0-7 cpu_online_mask=0-3 cpu_present_mask=0-3 cpu_active_mask=0-3
[ 5.462370]  cpu=2  cpu_possible_mask=0-7 cpu_online_mask=0-3 cpu_present_mask=0-3 cpu_active_mask=0-3
[ 5.462371]  cpu=3  cpu_possible_mask=0-7 cpu_online_mask=0-3 cpu_present_mask=0-3 cpu_active_mask=0-3
[ 5.462372] *cpu=4* cpu_possible_mask=0-7 cpu_online_mask=0-3 cpu_present_mask=0-3 cpu_active_mask=0-3
[ 5.462373] *cpu=5* cpu_possible_mask=0-7 cpu_online_mask=0-3 cpu_present_mask=0-3 cpu_active_mask=0-3
[ 5.462374] *cpu=6* cpu_possible_mask=0-7 cpu_online_mask=0-3 cpu_present_mask=0-3 cpu_active_mask=0-3
[ 5.462375] *cpu=7* cpu_possible_mask=0-7 cpu_online_mask=0-3 cpu_present_mask=0-3 cpu_active_mask=0-3

T430:/sys/fs/cgroup/cpu,cpuacct/system.slice# ls -l | grep '^d' | wc -l
80

/proc/sched_debug:

cfs_rq[0]:/system.slice
...
.tg_load_avg : 323584
...

80 * 1024 * 4 (not existent cpu4-cpu7) = 327680 (with a little bit of decay, this could be this extra load on the systen.slice tg)

Using for_each_online_cpu(i) instead of for_each_possible_cpu(i) in online_fair_sched_group() works on this machine, i.e. the .tg_load_avg of system.slice tg is 0 after startup.

Quoted 8 lines · Dietmar Eggemann

Something looks weird related to the use of for_each_possible_cpu(i) in online_fair_sched_group() on my i5-3320M CPU (4 logical cpus).

In case I print out cpu id and the cpu masks inside the for_each_possible_cpu(i) I get:

[ 5.462368]  cpu=0  cpu_possible_mask=0-7 cpu_online_mask=0-3 cpu_present_mask=0-3 cpu_active_mask=0-3

OK, you have a buggy BIOS :-) It enumerates too many CPU slots. There is no reason to have 4 empty CPU slots on a machine that cannot do physical hotplug.

This also explains why it doesn't show on many machines, most machines will not have this and possible_mask == present_mask == online_mask for most all cases.

x86 folk, can we detect the lack of physical hotplug capability and FW_WARN on this and lowering possible_mask to present_mask?

Quoted 18 lines

[ 5.462370]  cpu=1  cpu_possible_mask=0-7 cpu_online_mask=0-3 cpu_present_mask=0-3 cpu_active_mask=0-3
[ 5.462370]  cpu=2  cpu_possible_mask=0-7 cpu_online_mask=0-3 cpu_present_mask=0-3 cpu_active_mask=0-3
[ 5.462371]  cpu=3  cpu_possible_mask=0-7 cpu_online_mask=0-3 cpu_present_mask=0-3 cpu_active_mask=0-3
[ 5.462372] *cpu=4* cpu_possible_mask=0-7 cpu_online_mask=0-3 cpu_present_mask=0-3 cpu_active_mask=0-3
[ 5.462373] *cpu=5* cpu_possible_mask=0-7 cpu_online_mask=0-3 cpu_present_mask=0-3 cpu_active_mask=0-3
[ 5.462374] *cpu=6* cpu_possible_mask=0-7 cpu_online_mask=0-3 cpu_present_mask=0-3 cpu_active_mask=0-3
[ 5.462375] *cpu=7* cpu_possible_mask=0-7 cpu_online_mask=0-3 cpu_present_mask=0-3 cpu_active_mask=0-3

T430:/sys/fs/cgroup/cpu,cpuacct/system.slice# ls -l | grep '^d' | wc -l
80

/proc/sched_debug:

cfs_rq[0]:/system.slice
...
.tg_load_avg : 323584
...

Quoted 6 lines

80 * 1024 * 4 (not existent cpu4-cpu7) = 327680 (with a little bit of decay, this could be this extra load on the systen.slice tg)

Using for_each_online_cpu(i) instead of for_each_possible_cpu(i) in online_fair_sched_group() works on this machine, i.e. the .tg_load_avg of system.slice tg is 0 after startup.

Right, so the reason for using present_mask is that it avoids having to deal with hotplug, also all the per-cpu memory is allocated and present for !online CPUs anyway, so might as well set it up properly anyway.

(You might want to start booting your laptop with "possible_cpus=4" to save some memory FWIW.)

But yes, we have a bug here too... /me ponders

So aside from funny BIOSes, this should also show up when creating cgroups when you have offlined a few CPUs, which is far more common I'd think.

On IRC you mentioned that adding list_add_leaf_cfs_rq() to online_fair_sched_group() cures this, this would actually match with unregister_fair_sched_group() doing list_del_leaf_cfs_rq() and avoid a few instructions on the enqueue path, so that's all good.

I'm just not immediately seeing how that cures things. The only relevant user of the leaf_cfs_rq list seems to be update_blocked_averages() which is called from the balance code (idle_balance() and rebalance_domains()). But neither should call that for offline (or !present) CPUs.

Humm..

Quoted 3 lines · Peter Zijlstra

So aside from funny BIOSes, this should also show up when creating cgroups when you have offlined a few CPUs, which is far more common I'd think.

The problem is also that the load of the tg->se[cpu] that represents the tg->cfs_rq[cpu] is initialized to 1024 in:
alloc_fair_sched_group

     for_each_possible_cpu(i) {
         init_entity_runnable_average(se);
            sa->load_avg = scale_load_down(se->load.weight);

Initializing sa->load_avg to 1024 for a newly created task makes sense as we don't know yet what will be its real load but i'm not sure that we have to do the same for se that represents a task group. This load should be initialized to 0 and it will increase when task will be moved/attached into task group

Quoted 12 lines · Vincent Guittot

Quoted 3 lines · Peter Zijlstra

So aside from funny BIOSes, this should also show up when creating cgroups when you have offlined a few CPUs, which is far more common I'd think.

The problem is also that the load of the tg->se[cpu] that represents the tg->cfs_rq[cpu] is initialized to 1024 in:
alloc_fair_sched_group

     for_each_possible_cpu(i) {
         init_entity_runnable_average(se);
            sa->load_avg = scale_load_down(se->load.weight);

Quoted 5 lines

Initializing sa->load_avg to 1024 for a newly created task makes sense as we don't know yet what will be its real load but i'm not sure that we have to do the same for se that represents a task group. This load should be initialized to 0 and it will increase when task will be moved/attached into task group

Yes, I think that makes sense, not sure how horrible that is with the

current state of things, but after your propagate patch, that reinstates the interactivity hack that should work for sure.

Quoted 13 lines · Peter Zijlstra

Quoted 12 lines · Vincent Guittot

Quoted 3 lines · Peter Zijlstra

So aside from funny BIOSes, this should also show up when creating cgroups when you have offlined a few CPUs, which is far more common I'd think.

The problem is also that the load of the tg->se[cpu] that represents the tg->cfs_rq[cpu] is initialized to 1024 in:
alloc_fair_sched_group

     for_each_possible_cpu(i) {
         init_entity_runnable_average(se);
            sa->load_avg = scale_load_down(se->load.weight);

Quoted 7 lines

Quoted 5 lines

Initializing sa->load_avg to 1024 for a newly created task makes sense as we don't know yet what will be its real load but i'm not sure that we have to do the same for se that represents a task group. This load should be initialized to 0 and it will increase when task will be moved/attached into task group

Yes, I think that makes sense, not sure how horrible that is with the

That should not be that bad because this initial value is only useful for the few dozens of ms that follow the creation of the task group

Quoted 1 line

Quoted 2 lines

current state of things, but after your propagate patch, that reinstates the interactivity hack that should work for sure.

The patch below fixes the issue on my platform:

Dietmar, Omer can you confirm that this fix the problem of your platform too ?

---
 kernel/sched/fair.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

kernel/sched/fair.c +8 -1

index 8b03fb5..89776ac 100644
@@ -690,7 +690,14 @@ void init_entity_runnable_average(struct sched_entity *se)
 	 * will definitely be update (after enqueue).
 	 */
 	sa->period_contrib = 1023;
-	sa->load_avg = scale_load_down(se->load.weight);
+	/*
+	 * Tasks are intialized with full load to be seen as heavy task until
+	 * they get a chance to stabilize to their real load level.
+	 * group entity are intialized with null load to reflect the fact that
+	 * nothing has been attached yet to the task group.
+	 */
+	if (entity_is_task(se))
+		sa->load_avg = scale_load_down(se->load.weight);
 	sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
 	/*
 	 * At this point, util_avg won't be used in select_task_rq_fair anyway

Quoted 34 lines · Vincent Guittot

Quoted 18 lines · Peter Zijlstra

Quoted 16 lines · Vincent Guittot

Quoted 4 lines · Peter Zijlstra

So aside from funny BIOSes, this should also show up when creating cgroups when you have offlined a few CPUs, which is far more common I'd think.

The problem is also that the load of the tg->se[cpu] that represents the tg->cfs_rq[cpu] is initialized to 1024 in:
alloc_fair_sched_group

     for_each_possible_cpu(i) {
         init_entity_runnable_average(se);
            sa->load_avg = scale_load_down(se->load.weight);

Quoted 8 lines

Quoted 5 lines

Initializing sa->load_avg to 1024 for a newly created task makes sense as we don't know yet what will be its real load but i'm not sure that we have to do the same for se that represents a task group. This load should be initialized to 0 and it will increase when task will be moved/attached into task group

Yes, I think that makes sense, not sure how horrible that is with the

That should not be that bad because this initial value is only useful for the few dozens of ms that follow the creation of the task group

Quoted 1 line

Quoted 8 lines

Quoted 2 lines

current state of things, but after your propagate patch, that reinstates the interactivity hack that should work for sure.

The patch below fixes the issue on my platform:

Dietmar, Omer can you confirm that this fix the problem of your platform too ?

I just noticed this thread after posting
https://lkml.org/lkml/2016/10/18/719...
Noticed this bug while a ago and had the patch above at least a week but unfortunately didn't have time to post...
I think Omer had same problem I was trying to fix and I believe patch I post should address it.

Vincent, your version fixes my test case as well.

This is sched_stat from the same test case I had in my changelog. Note dd-2030 which is in root cgroup had same runtime as dd-2033 which is in child cgroup.

dd (2030, #threads: 1)
-------------------------------------------------------------------

se.exec_start                                :        275700.024137
se.vruntime                                  :         10589.114654
se.sum_exec_runtime                          :          1576.837993
se.nr_migrations                             :                    0
nr_switches                                  :                  159
nr_voluntary_switches                        :                    0
nr_involuntary_switches                      :                  159
se.load.weight                               :              1048576
se.avg.load_sum                              :             48840575
se.avg.util_sum                              :             19741820
se.avg.load_avg                              :                 1022
se.avg.util_avg                              :                  413
se.avg.last_update_time                      :         275700024137
policy                                       :                    0
prio                                         :                  120
clock-delta                                  :                   34
dd (2033, #threads: 1)
-------------------------------------------------------------------
se.exec_start                                :        275710.037178
se.vruntime                                  :          2383.802868
se.sum_exec_runtime                          :          1576.547591
se.nr_migrations                             :                    0
nr_switches                                  :                  162
nr_voluntary_switches                        :                    0
nr_involuntary_switches                      :                  162
se.load.weight                               :              1048576
se.avg.load_sum                              :             48316646
se.avg.util_sum                              :             21235249
se.avg.load_avg                              :                 1011
se.avg.util_avg                              :                  444
se.avg.last_update_time                      :         275710037178
policy                                       :                    0
prio                                         :                  120
clock-delta                                  :                   36

Thanks,
Joonwoo

Quoted 33 lines

---
 kernel/sched/fair.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

kernel/sched/fair.c +2 -1

index 8b03fb5..89776ac 100644
@@ -690,7 +690,14 @@ void init_entity_runnable_average(struct sched_entity
 *se)
         * will definitely be update (after enqueue).
         */
        sa->period_contrib = 1023;
-       sa->load_avg = scale_load_down(se->load.weight);
+       /*
+        * Tasks are intialized with full load to be seen as heavy task

until

kernel/sched/fair.c +2 -0

+        * they get a chance to stabilize to their real load level.
+        * group entity are intialized with null load to reflect the fact

that

kernel/sched/fair.c +4 -0

+        * nothing has been attached yet to the task group.
+        */
+       if (entity_is_task(se))
+               sa->load_avg = scale_load_down(se->load.weight);
        sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
        /*
         * At this point, util_avg won't be used in select_task_rq_fair

anyway

Quoted 19 lines · Peter Zijlstra

Quoted 17 lines · Vincent Guittot

Quoted 3 lines · Peter Zijlstra

So aside from funny BIOSes, this should also show up when creating cgroups when you have offlined a few CPUs, which is far more common I'd think.

The problem is also that the load of the tg->se[cpu] that represents the tg->cfs_rq[cpu] is initialized to 1024 in:
alloc_fair_sched_group

     for_each_possible_cpu(i) {
         init_entity_runnable_average(se);
            sa->load_avg = scale_load_down(se->load.weight);

Initializing sa->load_avg to 1024 for a newly created task makes sense as we don't know yet what will be its real load but i'm not sure that we have to do the same for se that represents a task group. This load should be initialized to 0 and it will increase when task will be moved/attached into task group

Quoted 1 line

Yes, I think that makes sense, not sure how horrible that is with the

That should not be that bad because this initial value is only useful for the few dozens of ms that follow the creation of the task group

Quoted 2 lines

current state of things, but after your propagate patch, that reinstates the interactivity hack that should work for sure.

 kernel/sched/fair.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

kernel/sched/fair.c +1 -1

index 8b03fb5..89776ac 100644
@@ -690,7 +690,14 @@ void init_entity_runnable_average(struct sched_entity *se)
 	 * will definitely be update (after enqueue).
 	 */
 	sa->period_contrib = 1023;
-	sa->load_avg = scale_load_down(se->load.weight);
+	/*

Quoted 20 lines · Vincent Guittot

Quoted 19 lines · Peter Zijlstra

Quoted 17 lines · Vincent Guittot

Quoted 3 lines · Peter Zijlstra

So aside from funny BIOSes, this should also show up when creating cgroups when you have offlined a few CPUs, which is far more common I'd think.

The problem is also that the load of the tg->se[cpu] that represents the tg->cfs_rq[cpu] is initialized to 1024 in:
alloc_fair_sched_group

     for_each_possible_cpu(i) {
         init_entity_runnable_average(se);
            sa->load_avg = scale_load_down(se->load.weight);

Initializing sa->load_avg to 1024 for a newly created task makes sense as we don't know yet what will be its real load but i'm not sure that we have to do the same for se that represents a task group. This load should be initialized to 0 and it will increase when task will be moved/attached into task group

Quoted 4 lines

Quoted 1 line

Yes, I think that makes sense, not sure how horrible that is with the

That should not be that bad because this initial value is only useful for the few dozens of ms that follow the creation of the task group

IMHO, it doesn't make much sense to initialize empty containers, which group sched_entities really are, to 1024. It is meant to represent what is in it, and a creation it is empty, so in my opinion initializing it to zero make sense.

Quoted 2 lines

Quoted 2 lines

current state of things, but after your propagate patch, that reinstates the interactivity hack that should work for sure.

It actually works on mainline/tip as well.

As I see it, the fundamental problem is keeping group entities up to date. Because the load_weight and hence se->avg.load_avg each per-cpu group sched_entity depends on the group cfs_rq->tg_load_avg_contrib for all cpus (tg->load_avg), including those that might be empty and therefore not enqueued, we must ensure that they are updated some other way. Most naturally as part of update_blocked_averages().

To guarantee that, it basically boils down to making sure:
Any cfs_rq with a non-zero tg_load_avg_contrib must be on the leaf_cfs_rq_list.

We can do that in different ways: 1) Add all cfs_rqs to the leaf_cfs_rq_list at task group creation, or 2) initialize group sched_entity contributions to zero and make sure that they are added to leaf_cfs_rq_list as soon as a sched_entity (task or group) is enqueued on it.

Vincent patch below gives us the second option.

Quoted 13 lines

 kernel/sched/fair.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

kernel/sched/fair.c +1 -1

index 8b03fb5..89776ac 100644
@@ -690,7 +690,14 @@ void init_entity_runnable_average(struct sched_entity *se)
       * will definitely be update (after enqueue).
       */
      sa->period_contrib = 1023;
-     sa->load_avg = scale_load_down(se->load.weight);
+     /*

kernel/sched/fair.c void init_entity_runnable_average(struct sched_entity *se) · line 703

Quoted 10 lines

kernel/sched/fair.c +7 -0

+      * Tasks are intialized with full load to be seen as heavy task until
+      * they get a chance to stabilize to their real load level.
+      * group entity are intialized with null load to reflect the fact that
+      * nothing has been attached yet to the task group.
+      */
+     if (entity_is_task(se))
+             sa->load_avg = scale_load_down(se->load.weight);
      sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
      /*
       * At this point, util_avg won't be used in select_task_rq_fair anyway

I would suggest adding a comment somewhere stating that we need to keep group cfs_rqs up to date:

-----

kernel/sched/fair.c +0 -0

index abb3763dff69..2b820d489be0 100644

kernel/sched/fair.c +0 -1

                if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true))
                        update_tg_load_avg(cfs_rq, 0);
        }
-----

I did a couple of simple tests on tip/sched/core to test whether Vincent's fix works even without reflecting group load/util in the group hierarchy:

Juno (2xA57+4xA53)

tip:

        grouped hog(1) alone: 2841
        non-grouped hogs(6) alone: 40830
        grouped hog(1): 218
        non-grouped hogs(6): 40580

tip+vg:

        grouped hog alone: 2849
        non-grouped hogs(6) alone: 40831
        grouped hog: 2363
        non-grouped hogs: 38418

See script below for details, but we basically see that the grouped task is not getting its 'fair' share on tip, while it does with Vincent's patch.

To summarize, I think Vincent's patch makes sense and works :-) More testing is needed of cause to see if there are other problems.

-----

# Create 100 task groups:
for i in seq 1 100;
do
cgcreate -g cpu:/root/test$i
done

NCPUS=$(grep -c ^processor /proc/cpuinfo)

# Run single cpu hog inside task group on first cpu _alone_:
cgexec -g cpu:/root/test100 taskset 0x01 sysbench --test=cpu \
--num-threads=1 --max-time=5 --max-requests=1000000 run | \
awk '{if ($4=="events:") {print "grouped hog(1) alone: " $5}}'

# Run cpu hogs outside task group _alone_:
sysbench --test=cpu --num-threads=$NCPUS --max-time=10 \
--max-requests=1000000 run | awk '{if ($4=="events:") \
{print "non-grouped hogs('$NCPUS') alone: " $5}}'

# Run cpu hogs outside task group:
sysbench --test=cpu --num-threads=$NCPUS --max-time=10 \
--max-requests=1000000 run | awk '{if ($4=="events:") \
{print "non-grouped hogs('$NCPUS'): " $5}}' &

# Run single cpu hog inside task group on first cpu:
cgexec -g cpu:/root/test100 taskset 0x01 sysbench \
--test=cpu --num-threads=1 --max-time=5 \
--max-requests=1000000 run | awk '{if ($4=="events:") \
{print "grouped hog(1): " $5}}'

wait

# Delete task groups:
for i in seq 1 100;
do
cgdelete -g cpu:/root/test$i
done

Quoted 91 lines · Morten Rasmussen

Quoted 20 lines · Vincent Guittot

Quoted 19 lines · Peter Zijlstra

Quoted 17 lines · Vincent Guittot

Quoted 3 lines · Peter Zijlstra

So aside from funny BIOSes, this should also show up when creating cgroups when you have offlined a few CPUs, which is far more common I'd think.

The problem is also that the load of the tg->se[cpu] that represents the tg->cfs_rq[cpu] is initialized to 1024 in:
alloc_fair_sched_group

     for_each_possible_cpu(i) {
         init_entity_runnable_average(se);
            sa->load_avg = scale_load_down(se->load.weight);

Initializing sa->load_avg to 1024 for a newly created task makes sense as we don't know yet what will be its real load but i'm not sure that we have to do the same for se that represents a task group. This load should be initialized to 0 and it will increase when task will be moved/attached into task group

Quoted 4 lines

Quoted 1 line

Yes, I think that makes sense, not sure how horrible that is with the

That should not be that bad because this initial value is only useful for the few dozens of ms that follow the creation of the task group

IMHO, it doesn't make much sense to initialize empty containers, which group sched_entities really are, to 1024. It is meant to represent what is in it, and a creation it is empty, so in my opinion initializing it to zero make sense.

Quoted 2 lines

Quoted 2 lines

current state of things, but after your propagate patch, that reinstates the interactivity hack that should work for sure.

It actually works on mainline/tip as well.

As I see it, the fundamental problem is keeping group entities up to date. Because the load_weight and hence se->avg.load_avg each per-cpu group sched_entity depends on the group cfs_rq->tg_load_avg_contrib for all cpus (tg->load_avg), including those that might be empty and therefore not enqueued, we must ensure that they are updated some other way. Most naturally as part of update_blocked_averages().

To guarantee that, it basically boils down to making sure:
Any cfs_rq with a non-zero tg_load_avg_contrib must be on the leaf_cfs_rq_list.

We can do that in different ways: 1) Add all cfs_rqs to the leaf_cfs_rq_list at task group creation, or 2) initialize group sched_entity contributions to zero and make sure that they are added to leaf_cfs_rq_list as soon as a sched_entity (task or group) is enqueued on it.

Vincent patch below gives us the second option.

Quoted 13 lines

 kernel/sched/fair.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

kernel/sched/fair.c +1 -1

index 8b03fb5..89776ac 100644
@@ -690,7 +690,14 @@ void init_entity_runnable_average(struct sched_entity *se)
       * will definitely be update (after enqueue).
       */
      sa->period_contrib = 1023;
-     sa->load_avg = scale_load_down(se->load.weight);
+     /*

kernel/sched/fair.c void init_entity_runnable_average(struct sched_entity *se) · line 703

Quoted 10 lines

kernel/sched/fair.c +7 -0

+      * Tasks are intialized with full load to be seen as heavy task until
+      * they get a chance to stabilize to their real load level.
+      * group entity are intialized with null load to reflect the fact that
+      * nothing has been attached yet to the task group.
+      */
+     if (entity_is_task(se))
+             sa->load_avg = scale_load_down(se->load.weight);
      sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
      /*
       * At this point, util_avg won't be used in select_task_rq_fair anyway

I would suggest adding a comment somewhere stating that we need to keep group cfs_rqs up to date:

-----

kernel/sched/fair.c +0 -0

index abb3763dff69..2b820d489be0 100644
@@ -6641,6 +6641,11 @@ static void update_blocked_averages(int cpu)
                if (throttled_hierarchy(cfs_rq))
                        continue;
 

kernel/sched/fair.c static void update_blocked_averages(int cpu) · line 6648

Quoted 5 lines

kernel/sched/fair.c +5 -0

+               /*
+                * Note that _any_ leaf cfs_rq with a non-zero tg_load_avg_contrib
+                * _must_ be on the leaf_cfs_rq_list to ensure that group shares
+                * are updated correctly.
+                */

As discussed on IRC, the point is that even if the leaf cfs_rq is added to the leaf_cfs_rq_list, it doesn't ensure that it will be updated correctly for unplugged CPUs

---
 kernel/sched/fair.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

kernel/sched/fair.c +1 -1

index 8b03fb5..89776ac 100644
@@ -690,7 +690,14 @@ void init_entity_runnable_average(struct sched_entity *se)
 	 * will definitely be update (after enqueue).
 	 */
 	sa->period_contrib = 1023;
-	sa->load_avg = scale_load_down(se->load.weight);
+	/*

Quoted 20 lines · Vincent Guittot

---
 kernel/sched/fair.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

kernel/sched/fair.c +7 -1

index 8b03fb5..89776ac 100644
@@ -690,7 +690,14 @@ void init_entity_runnable_average(struct sched_entity *se)
       * will definitely be update (after enqueue).
       */
      sa->period_contrib = 1023;
-     sa->load_avg = scale_load_down(se->load.weight);
+     /*
+      * Tasks are intialized with full load to be seen as heavy task until
+      * they get a chance to stabilize to their real load level.
+      * group entity are intialized with null load to reflect the fact that
+      * nothing has been attached yet to the task group.
+      */
+     if (entity_is_task(se))

kernel/sched/fair.c void init_entity_runnable_average(struct sched_entity *se) · line 703

Quoted 5 lines

kernel/sched/fair.c +1 -0

+             sa->load_avg = scale_load_down(se->load.weight);
      sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
      /*
       * At this point, util_avg won't be used in select_task_rq_fair anyway

Vince, could you post a proper version of this patch with changelogs and tags so that we can get that merged into Linus' tree and stable for 4.8?

Quoted 12 lines · Peter Zijlstra

Quoted 11 lines · Vincent Guittot

Quoted 3 lines · Peter Zijlstra

So aside from funny BIOSes, this should also show up when creating cgroups when you have offlined a few CPUs, which is far more common I'd think.

The problem is also that the load of the tg->se[cpu] that represents the tg->cfs_rq[cpu] is initialized to 1024 in:
alloc_fair_sched_group

     for_each_possible_cpu(i) {
         init_entity_runnable_average(se);
            sa->load_avg = scale_load_down(se->load.weight);

Quoted 6 lines

Quoted 5 lines

Initializing sa->load_avg to 1024 for a newly created task makes sense as we don't know yet what will be its real load but i'm not sure that we have to do the same for se that represents a task group. This load should be initialized to 0 and it will increase when task will be moved/attached into task group

Yes, I think that makes sense, not sure how horrible that is with the

That should not be that bad because this initial value is only useful for the few dozens of ms that follow the creation of the task group

Quoted 2 lines

current state of things, but after your propagate patch, that reinstates the interactivity hack that should work for sure.

The patch below fixes the issue on my platform:

Dietmar, Omer can you confirm that this fix the problem of your platform too ?

---
 kernel/sched/fair.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)Vinc

kernel/sched/fair.c +1 -1

index 8b03fb5..89776ac 100644
@@ -690,7 +690,14 @@ void init_entity_runnable_average(struct sched_entity *se)
 	 * will definitely be update (after enqueue).
 	 */
 	sa->period_contrib = 1023;
-	sa->load_avg = scale_load_down(se->load.weight);
+	/*

Quoted 50 lines · Vincent Guittot

Quoted 12 lines · Peter Zijlstra

Quoted 11 lines · Vincent Guittot

Quoted 3 lines · Peter Zijlstra

So aside from funny BIOSes, this should also show up when creating cgroups when you have offlined a few CPUs, which is far more common I'd think.

The problem is also that the load of the tg->se[cpu] that represents the tg->cfs_rq[cpu] is initialized to 1024 in:
alloc_fair_sched_group

     for_each_possible_cpu(i) {
         init_entity_runnable_average(se);
            sa->load_avg = scale_load_down(se->load.weight);

Quoted 6 lines

Quoted 5 lines

Initializing sa->load_avg to 1024 for a newly created task makes sense as we don't know yet what will be its real load but i'm not sure that we have to do the same for se that represents a task group. This load should be initialized to 0 and it will increase when task will be moved/attached into task group

Yes, I think that makes sense, not sure how horrible that is with the

That should not be that bad because this initial value is only useful for the few dozens of ms that follow the creation of the task group

Quoted 2 lines

current state of things, but after your propagate patch, that reinstates the interactivity hack that should work for sure.

The patch below fixes the issue on my platform:

Dietmar, Omer can you confirm that this fix the problem of your platform too ?

---
 kernel/sched/fair.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)Vinc

kernel/sched/fair.c +8 -1

index 8b03fb5..89776ac 100644
@@ -690,7 +690,14 @@ void init_entity_runnable_average(struct sched_entity *se)
       * will definitely be update (after enqueue).
       */
      sa->period_contrib = 1023;
-     sa->load_avg = scale_load_down(se->load.weight);
+     /*
+      * Tasks are intialized with full load to be seen as heavy task until
+      * they get a chance to stabilize to their real load level.
+      * group entity are intialized with null load to reflect the fact that
+      * nothing has been attached yet to the task group.
+      */
+     if (entity_is_task(se))
+             sa->load_avg = scale_load_down(se->load.weight);
      sa->load_sum = sa->load_avg * LOAD_AVG_MAX;

kernel/sched/fair.c void init_entity_runnable_average(struct sched_entity *se) · line 703

Quoted 6 lines

kernel/sched/fair.c +0 -0

      /*
       * At this point, util_avg won't be used in select_task_rq_fair anyway

Omer also reports that this patch fixes the bug for him as well. Thanks for the great work, Vincent!

Quoted 2 lines · Peter Zijlstra

Quoted 1 line · Vincent Guittot

On 18 October 2016 at 11:07, Peter Zijlstra <[email protected]> wrote:

The patch below fixes the issue on my platform:

Dietmar, Omer can you confirm that this fix the problem of your platform too ?

---
 kernel/sched/fair.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

kernel/sched/fair.c +8 -1

index 8b03fb5..89776ac 100644
@@ -690,7 +690,14 @@ void init_entity_runnable_average(struct sched_entity *se)
 	 * will definitely be update (after enqueue).
 	 */
 	sa->period_contrib = 1023;
-	sa->load_avg = scale_load_down(se->load.weight);
+	/*
+	 * Tasks are intialized with full load to be seen as heavy task until
+	 * they get a chance to stabilize to their real load level.
+	 * group entity are intialized with null load to reflect the fact that
+	 * nothing has been attached yet to the task group.
+	 */
+	if (entity_is_task(se))
+		sa->load_avg = scale_load_down(se->load.weight);
 	sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
 	/*
 	 * At this point, util_avg won't be used in select_task_rq_fair anyway

Quoted 3 lines · Vincent Guittot

Quoted 2 lines · Peter Zijlstra

Quoted 1 line · Vincent Guittot

On 18 October 2016 at 11:07, Peter Zijlstra <[email protected]> wrote:

[...]

Quoted 3 lines

The patch below fixes the issue on my platform:

Quoted 1 line

Dietmar, Omer can you confirm that this fix the problem of your platform too ?

It fixes this broken BIOS issue on my T430 ( cpu_possible_mask > cpu_online_mask). I ran the original test with the cpu hogs (stress -c 4). Launch time of applications becomes normal again.

Tested-by Dietmar Eggemann <[email protected]>

But this test only makes sure that we don't see any ghost contribution (from non-existing cpus) any more.

We should study the tg->se[i]->avg.load_avg for the hierarchy of tg's (with the highest tg having a task enqueued) a little bit more, with and without your v5 'sched: reflect sched_entity move into task_group's load'.

Quoted 25 lines

---
 kernel/sched/fair.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

kernel/sched/fair.c +8 -1

index 8b03fb5..89776ac 100644
@@ -690,7 +690,14 @@ void init_entity_runnable_average(struct sched_entity *se)
       * will definitely be update (after enqueue).
       */
      sa->period_contrib = 1023;
-     sa->load_avg = scale_load_down(se->load.weight);
+     /*
+      * Tasks are intialized with full load to be seen as heavy task until
+      * they get a chance to stabilize to their real load level.
+      * group entity are intialized with null load to reflect the fact that
+      * nothing has been attached yet to the task group.
+      */
+     if (entity_is_task(se))
+             sa->load_avg = scale_load_down(se->load.weight);
      sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
      /*
       * At this point, util_avg won't be used in select_task_rq_fair anyway

Quoted 4 lines · Dietmar Eggemann

Quoted 3 lines · Vincent Guittot

Quoted 2 lines · Peter Zijlstra

Quoted 1 line · Vincent Guittot

On 18 October 2016 at 11:07, Peter Zijlstra <[email protected]> wrote:

Quoted 6 lines

But this test only makes sure that we don't see any ghost contribution (from non-existing cpus) any more.

We should study the tg->se[i]->avg.load_avg for the hierarchy of tg's (with the highest tg having a task enqueued) a little bit more, with and without your v5 'sched: reflect sched_entity move into task_group's load'.

Can you elaborate ?

Quoted 5 lines · Vincent Guittot

Quoted 4 lines · Dietmar Eggemann

Quoted 3 lines · Vincent Guittot

Quoted 2 lines · Peter Zijlstra

Quoted 1 line · Vincent Guittot

On 18 October 2016 at 11:07, Peter Zijlstra <[email protected]> wrote:

[...]

Quoted 7 lines

Quoted 6 lines

But this test only makes sure that we don't see any ghost contribution (from non-existing cpus) any more.

We should study the tg->se[i]->avg.load_avg for the hierarchy of tg's (with the highest tg having a task enqueued) a little bit more, with and without your v5 'sched: reflect sched_entity move into task_group's load'.

Quoted 1 line

Can you elaborate ?

I try :-)

I thought I will see some different behaviour because of the fact that the tg se's are initialized differently [1024 versus 0].

But I can't spot any difference. The test case is running a sysbench thread affine to cpu1 in tg_root/tg_1/tg_11/tg_111 on tip/sched/core on an ARM64 Juno (6 logical cpus).
The moment the sysbench task is put into tg_111 tg_111->se[1]->avg.load_avg gets updated to 0 any way because of the huge time difference between creating this tg and attaching a task to it. So the tg->se[2]->avg.load_avg signals for tg_111, tg_11 and tg_1 look exactly the same w/o and w/ your patch.

But your patch helps in this (very synthetic) test case as well. W/o your patch I see remaining tg->load_avg for tg_1 and tg_11 after the test case has finished because the tg's were exclusively used on cpu1.

# cat /proc/sched_debug

cfs_rq[1]:/tg_1
.tg_load_avg_contrib : 0
.tg_load_avg : 5120 (5 (unused cpus) * 1024 * 1)
cfs_rq[1]:/tg_1/tg_11/tg_111
.tg_load_avg_contrib : 0
.tg_load_avg : 0
cfs_rq[1]:/tg_1/tg_11
.tg_load_avg_contrib : 0
.tg_load_avg : 5120

With your patch applied all the .tg_load_avg are 0.

Quoted 5 lines · Vincent Guittot

Quoted 4 lines · Dietmar Eggemann

Quoted 3 lines · Vincent Guittot

Quoted 2 lines · Peter Zijlstra

Quoted 1 line · Vincent Guittot

On 18 October 2016 at 11:07, Peter Zijlstra <[email protected]> wrote:

[...]

Quoted 7 lines

Quoted 6 lines

But this test only makes sure that we don't see any ghost contribution (from non-existing cpus) any more.

We should study the tg->se[i]->avg.load_avg for the hierarchy of tg's (with the highest tg having a task enqueued) a little bit more, with and without your v5 'sched: reflect sched_entity move into task_group's load'.

Quoted 1 line

Can you elaborate ?

I try :-)

I thought I will see some different behaviour because of the fact that the tg se's are initialized differently [1024 versus 0].

But I can't spot any difference. The test case is running a sysbench thread affine to cpu1 in tg_root/tg_1/tg_11/tg_111 on tip/sched/core on an ARM64 Juno (6 logical cpus).
The moment the sysbench task is put into tg_111 tg_111->se[1]->avg.load_avg gets updated to 0 any way because of the huge time difference between creating this tg and attaching a task to it. So the tg->se[2]->avg.load_avg signals for tg_111, tg_11 and tg_1 look exactly the same w/o and w/ your patch.

But your patch helps in this (very synthetic) test case as well. W/o your patch I see remaining tg->load_avg for tg_1 and tg_11 after the test case has finished because the tg's were exclusively used on cpu1.

# cat /proc/sched_debug

cfs_rq[1]:/tg_1
.tg_load_avg_contrib : 0
.tg_load_avg : 5120 (5 (unused cpus) * 1024 * 1)
cfs_rq[1]:/tg_1/tg_11/tg_111
.tg_load_avg_contrib : 0
.tg_load_avg : 0
cfs_rq[1]:/tg_1/tg_11
.tg_load_avg_contrib : 0
.tg_load_avg : 5120

With your patch applied all the .tg_load_avg are 0.

On Mon, Oct 17, 2016 at 11:52:39PM +0100, Dietmar Eggemann wrote:

Quoted 3 lines

Using for_each_online_cpu(i) instead of for_each_possible_cpu(i) in online_fair_sched_group() works on this machine, i.e. the .tg_load_avg of system.slice tg is 0 after startup.

Right, so the reason for using present_mask is that it avoids having to deal with hotplug, also all the per-cpu memory is allocated and present for !online CPUs anyway, so might as well set it up properly anyway.

(You might want to start booting your laptop with "possible_cpus=4" to save some memory FWIW.)

But yes, we have a bug here too... /me ponders

So aside from funny BIOSes, this should also show up when creating cgroups when you have offlined a few CPUs, which is far more common I'd think.

On IRC you mentioned that adding list_add_leaf_cfs_rq() to online_fair_sched_group() cures this, this would actually match with unregister_fair_sched_group() doing list_del_leaf_cfs_rq() and avoid a few instructions on the enqueue path, so that's all good.

I'm just not immediately seeing how that cures things. The only relevant user of the leaf_cfs_rq list seems to be update_blocked_averages() which is called from the balance code (idle_balance() and rebalance_domains()). But neither should call that for offline (or !present) CPUs.

Quoted 6 lines · Peter Zijlstra

On Mon, Oct 17, 2016 at 11:52:39PM +0100, Dietmar Eggemann wrote:

On IRC you mentioned that adding list_add_leaf_cfs_rq() to online_fair_sched_group() cures this, this would actually match with unregister_fair_sched_group() doing list_del_leaf_cfs_rq() and avoid a few instructions on the enqueue path, so that's all good.

Yes, I was able to recreate a similar problem (not related to the cpu masks) on ARM64 (6 logical cpus). I created 100 2. level tg's but only put one task (no cpu affinity, so it could run on multiple cpus) in one of these tg's (mainly to see the related cfs_rq's in /proc/sched_debug).

I get a remaining .tg_load_avg : 49898 for cfs_rq[x]:/tg_1

Quoted 5 lines

I'm just not immediately seeing how that cures things. The only relevant user of the leaf_cfs_rq list seems to be update_blocked_averages() which is called from the balance code (idle_balance() and rebalance_domains()). But neither should call that for offline (or !present) CPUs.

Assuming this is load from the 99 2. level tg's which never had a task running, putting list_add_leaf_cfs_rq() into online_fair_sched_group() for all cpus makes sure that all the 'blocked load' get's decayed.

Doing what Vincent just suggested, not initializing tg se's w/ 1024 but w/ 0 instead prevents this from being necessary.