On Thu, Sep 11, 2014 at 01:34:12PM +0100, Peter Zijlstra wrote:
@@ -4514,6 +4519,17 @@ static int select_idle_sibling(struct task_struct *p, int target) return target; } +static int get_cpu_utilization(int cpu) +{
- unsigned long usage = cpu_rq(cpu)->cfs.usage_load_avg;
- unsigned long capacity = capacity_of(cpu);
- if (usage >= SCHED_LOAD_SCALE)
return capacity + 1;- return (usage * capacity) >> SCHED_LOAD_SHIFT;
+}
So if I understood patch 9 correct, your changelog is iffy. usage_load_avg should never get > 1 (of whatever unit), no matter how many tasks are on the rq. You can only maximally run all the time.
Therefore I can only interpret the if (usage >= SCHED_LOAD_SCALE) as numerical error handling, nothing more.
That is not entirely true unless you also classify transient usage spikes due to task migrations as numerical errors as well.
Since each task sched_entity is carrying around 350ms worth of execution history with it between different cpus and cpu utilization is based on the sum of task entity usage_avg_contrib on the runqueue you may get cfs.usage_load_avg > 1 temporarily after task migrations. It will eventually converge to 1.
The same goes for new tasks which are initialized to have a usage_avg_contrib of 1 and may be queued on cpu with tasks already running. In that case cfs.usage_load_avg is temporarily unbounded.
Also I'm not entirely sure I like the usage, utilization names/metrics. I would suggest to reverse them. Call the pure running number 'utilization' and this scaled with capacity 'usage' or so.
I can agree with calling running for utilization, but I'm not convienced about capacity. What does it exactly cover here? I'm confused and jetlagged.
Morten