Re: [RFC PATCH 14/14] sched: implement usage tracking

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Fri, 2012-02-17 at 03:32 -0800, Paul Turner wrote:

> Since we do not know where we slept relative to the tick, unless we
> explicitly track where this occurred we don't know for a "n.y" jiffy
> sleep whether the "0.y" fraction included an extra decay. 

but isn't that the same problem you have in
__update_entity_runnable_avg() where you use ->runnable_avg_period %
1024 ?

Also, shouldn't __synchronize_entity_decay() update
->runnable_avg_period? Surely when you wake up your window fraction
could be different than when you went to sleep?

>  It's
> important to get this right to reduce instability, since when the
> entity wakes back up we need to charge an equivalent decay against
> it's contribution before we can remove it from the blocked_load_sum.
> 
> For the common case of a short sleep whether we charge an extra period
> (if y included a jiffy edge) or not is significant to stability since
> getting it wrong means we leave a little extra load in blocked_load.

Sure..

> Further complicating this we need to be able to do said
> synchronization in the annoying cases of wake-up migration (where we
> don't hold the lock on previous rq) and 32-bit machines (where
> everything sucks).

Ah, wouldn't task_waking_fair() be a better place though?

> >
> > C) 'Migrate'
> >    - uses contributes_blocked_load to check if we actually did migrate?
> >
> 
> We actually use se->avg.decay_count to check whether we were migrated
> (setting it appropriately at migration); contributes_blocked_load is
> just a convenience variable to track whether we're a part of
> cfs_rq->blocked_load_avg or not.

If you use task_waking_fair() you can also use ENQUEUE_WAKING.

> There's a one or two other paths (than ttwu) that this matters for,
> using set_task_rq() gets them all.

What paths are those? Having a cfs hook in the generic code there isn't
pretty an suggests funny things could happen when PI gets done.