Re: proc/stat: idle goes backward

[Frederic Weisbecker <fweisbec@gmail.com>]

On Fri, Aug 16, 2013 at 04:47:04PM +0200, Oleg Nesterov wrote:
> Hello.
> 
> Out customer reports that "idle" field is not monotonic. So far this
> is all I know. I do not know how to reproduce, etc.
> 
> But when I look at this code, this looks really possible even
> ignoring drivers/cpuidle/ which plays with update_ts_time_stats().
> 
> So, get_cpu_idle_time_us(last_update_time => NULL) does:
> 
> 	if (ts->idle_active && !nr_iowait_cpu(cpu)) {
> 		ktime_t delta = ktime_sub(now, ts->idle_entrytime);
> 
> 		idle = ktime_add(ts->idle_sleeptime, delta);
> 	} else {
> 		idle = ts->idle_sleeptime;
> 	}
> 
> 
> Suppose that ts->idle_active == T. By the time we calculate
> 
> 	idle = ktime_add(ts->idle_sleeptime, delta);
> 
> this cpu can be already non-idle and ->idle_sleeptime can be already
> updated by tick_nohz_stop_idle(), we return the wrong value.
> 
> If user-space reads /proc/stat again after that, "idle" can obviously
> go back.
> 
> get_cpu_iowait_time_us() has the same problem.
> 
> Plus nr_iowait_cpu() can change in between even if cpu stays idle,
> io_schedule() can return on another CPU.
> 
> Questions:
> 
> 	- Any other reason why it can be non-monotonic?
> 
> 	- Should we fix this or should we document that userspace
> 	  should handle this itself?
> 
> 	  IOW, is this is bug or not?

I don't know if we want to fix it (I personally think we should because it is not the
first time I hear complains about this) but if we do, here is a possible fix:

https://lkml.org/lkml/2013/8/8/638

Thanks.