hulk inclusion
category: bugfix
bugzilla: 13257
CVE: NA

-------------------------------------------------

Interduce "use-sched-idle-time" option to enable/disable
sched idle time accounting instead of tick based time accounting.

nohz mode: use get_idle_time;
nohz=off mode: use idle time accounting in sched_info_switch function.

And you can also use "use-sched-idle-time=force" force to use
idle time accountint in sched_info_switch function.

guarantee:
cpu_clock(cpu) - idle == utime + stime

We use this time in /proc/stat.

We use get_idle_time instead of collecting idle time when
context switch. And get_idle_time is available when nohz is
enabled. So do not collect idle time when nohz is enabled.

Issue 1:
1) boot euleros
2) bind a kernel thread to a cpu (eg. cpu1), run a long time (more than 10mins)
3) stop this kernel thread
4) run a spin loop process in userspace
5) then, the utime is 0, but the stime is ~100%

for example:
stime = 10000; utime = 1; rtime = 100001;
Run 10s in userspace now, then

rtime = clock - idle , (10001 + 10 = 10011)
stime = 10 * 10000/(1 + 10 + 10000) ≈ 10
utime = 10 * (1 + 10)/(1 + 10 + 10000) ≈ 0

That's the problem.

We just adjust the delta time to reduce the accumulated:
error of stime/utime:
delta = rtime - (utime + stime)
stime += delta * delta_of_stat_stime / (delta_of_stat_stime + delta_of_stat_utime)
utime = rtime - stime

Issue 2:
When cpu is offline we don't need to adjust the user and system time.
Only cpu is online we could asume:

utime% + stime% + idle% = 100%.

Issue 3:
When nohz=off, we must add 'now - last_entry_idle_time' to sum_idle_time.
Tested-by: NXie Zhipeng <xiezhipeng1@huawei.com>
Signed-off-by: NXie XiuQi <xiexiuqi@huawei.com>
Signed-off-by: NYang Yingliang <yangyingliang@huawei.com>