Re: [PATCH] sched/loadavg: Avoid loadavg spikes caused by delayed NO_HZ accounting

[Peter Zijlstra]

On Wed, Feb 08, 2017 at 01:29:24PM +0000, Matt Fleming wrote:
> The calculation for the next sample window when exiting NOH_HZ idle
> does not handle the fact that we may not have reached the next sample
> window yet, i.e. that we came out of idle between sample windows.
> 
> If we wake from NO_HZ idle after the pending this_rq->calc_load_update
> window time when we want idle but before the next sample window, we
> will add an unnecessary LOAD_FREQ delay to the load average
> accounting, delaying any update for potentially ~9seconds.
> 
> This can result in huge spikes in the load average values due to
> per-cpu uninterruptible task counts being out of sync when accumulated
> across all CPUs.
> 
> It's safe to update the per-cpu active count if we wake between sample
> windows because any load that we left in 'calc_load_idle' will have
> been zero'd when the idle load was folded in calc_global_load().

Right, so differently put; the problem is that we check against the
'stale' rq->calc_load_update, while the current and effective period
boundary is 'calc_load_update'.

So, when rq->calc_load_update < jiffies < calc_load_update, we end up
setting the next-update to calc_load_update+LOAD_FREQ, where it should
have been calc_load_update.

> diff --git a/kernel/sched/loadavg.c b/kernel/sched/loadavg.c
> index a2d6eb71f06b..a7a6f3646970 100644
> --- a/kernel/sched/loadavg.c
> +++ b/kernel/sched/loadavg.c

> @@ -210,10 +211,16 @@ void calc_load_exit_idle(void)
>  	 * We woke inside or after the sample window, this means we're already
>  	 * accounted through the nohz accounting, so skip the entire deal and
>  	 * sync up for the next window.
> +	 *
> +	 * The next window is 'calc_load_update' if we haven't reached it yet,
> +	 * and 'calc_load_update + 10' if we're inside the current window.
>  	 */
> +	next_window = calc_load_update;
> +
> +	if (time_in_range_open(jiffies, next_window, next_window + 10)
> +		next_window += LOAD_FREQ;
> +
> +	this_rq->calc_load_update = next_window;
>  }

So I don't much like the time_in_range_open() thing. The simpler patch
which you tested to also work was:

diff --git a/kernel/sched/loadavg.c b/kernel/sched/loadavg.c
index 7296b7308eca..cfb47bd0ee50 100644
--- a/kernel/sched/loadavg.c
+++ b/kernel/sched/loadavg.c
@@ -201,6 +201,8 @@ void calc_load_exit_idle(void)
 {
 	struct rq *this_rq = this_rq();
 
+	this_rq->calc_load_update = calc_load_update;
+
 	/*
 	 * If we're still before the sample window, we're done.
 	 */
@@ -212,7 +214,6 @@ void calc_load_exit_idle(void)
 	 * accounted through the nohz accounting, so skip the entire deal and
 	 * sync up for the next window.
 	 */
-	this_rq->calc_load_update = calc_load_update;
 	if (time_before(jiffies, this_rq->calc_load_update + 10))
 		this_rq->calc_load_update += LOAD_FREQ;
 }

But the problem there is that we unconditionally issue that store. Now
I've no idea how much of a problem that is, and it certainly is the
simplest form (+- comments that need updating), so maybe that makes
sense.

Alternatively, something like:

diff --git a/kernel/sched/loadavg.c b/kernel/sched/loadavg.c
index 7296b7308eca..3dd4ce6fe151 100644
--- a/kernel/sched/loadavg.c
+++ b/kernel/sched/loadavg.c
@@ -207,12 +207,15 @@ void calc_load_exit_idle(void)
 	if (time_before(jiffies, this_rq->calc_load_update))
 		return;
 
+	this_rq->calc_load_update = calc_load_update;
+	if (time_before(jiffies, this_rq->calc_load_update))
+		return;
+
 	/*
 	 * We woke inside or after the sample window, this means we're already
 	 * accounted through the nohz accounting, so skip the entire deal and
 	 * sync up for the next window.
 	 */
-	this_rq->calc_load_update = calc_load_update;
 	if (time_before(jiffies, this_rq->calc_load_update + 10))
 		this_rq->calc_load_update += LOAD_FREQ;
 }

might be another solution.

Irrespective the above though; should we not make this:

+	this_rq->calc_load_update = READ_ONCE(calc_load_update);

because if for some reason we do a double load of calc_load_update and
see two different values, weird stuff could happen.

And because, on general principle, a READ_ONCE() should be paired with a
WRITE_ONCE(), that should be done too I suppose.