Description of problem:
In our case, there are two cpu/cpuacct/cpuset cgroups A and A/B, cfs_quota of A&B both are 1600000(cfs_period = 100000). All tasks are run in cgroup B.

After running a period of time. We found that idle metric(in cpuacct.proc_stat) of A becomes a really large number, then overflow, finally becomes a large number again.


Version-Release number of selected component (if applicable):


How reproducible:
yield.c

#include <sched.h>
#include <pthread.h>

static void *yield_worker(void *args)
{
	for ( ; ; )
		sched_yield();
}

int main(int argc, const char *argv[])
{
	pthread_t tid[256];
	int i;

	for (i = 0; i < 256; i++)
		pthread_create(&tid[i], NULL, yield_worker, NULL);

	pthread_join(tid[0], NULL);

	return 0;
}

Steps to Reproduce:
1. prepare cpu cgroups 
mkdir /sys/fs/cgroup/cpu/A
echo 0-31 > /sys/fs/cgroup/cpu/A/cpuset.cpus
echo 0 > /sys/fs/cgroup/cpu/A/cpuset.mems
echo 50000 > /sys/fs/cgroup/cpu/A/cpu.cfs_quota_us

mkdir /sys/fs/cgroup/cpu/A/B
echo 0-31 > /sys/fs/cgroup/cpu/A/B/cpuset.cpus
echo 0 > /sys/fs/cgroup/cpu/A/B/cpuset.mems
echo 50000 > /sys/fs/cgroup/cpu/A/B/cpu.cfs_quota_us

2. wait a long time(maybe few of days), we found that more oldder cgroups were easy to reproduce problem.
3. run test case
cgexec -g cpu:/A/B ./yield
4. show results
cat /sys/fs/cgroup/cpu/A/cpuacct.proc_stat ; sleep 5; cat /sys/fs/cgroup/cpu/A/cpuacct.proc_stat

cat /sys/fs/cgroup/cpu/A/B/cpuacct.proc_stat ; sleep 5; cat /sys/fs/cgroup/cpu/A/B/cpuacct.proc_stat

Actual results:
A:
idle 2625715746
steal 49438
idle 3406279083
steal 50520

B:
idle 137536530
steal 847084

idle 137536530
steal 862591

Expected results:
idle in A should not inc two much in this case.

Additional info: