Re: [PATCH v3 2/7] sched: accumulate per-cfs_rq cpu usage

[Peter Zijlstra <peterz@infradead.org>]

On Thu, 2010-10-14 at 02:14 -0700, Paul Turner wrote:
> On Thu, Oct 14, 2010 at 2:01 AM, Peter Zijlstra <peterz@infradead.org> wrote:
> > On Tue, 2010-10-12 at 13:21 +0530, Bharata B Rao wrote:
> >> +static void account_cfs_rq_quota(struct cfs_rq *cfs_rq,
> >> +               unsigned long delta_exec)
> >> +{
> >> +       if (cfs_rq->quota_assigned == RUNTIME_INF)
> >> +               return;
> >> +
> >> +       cfs_rq->quota_used += delta_exec;
> >> +
> >> +       if (cfs_rq->quota_used < cfs_rq->quota_assigned)
> >> +               return;
> >> +
> >> +       cfs_rq->quota_assigned += tg_request_cfs_quota(cfs_rq->tg);
> >> +}
> >
> > That looks iffy, quota_assigned is only ever incremented and can wrap.
> 
> This can't advance at a rate faster than ~vruntime and we can't handle
> wrapping there anyway (fortunately it would take something like 35k
> years?)

You can't go faster than wall-time, vruntime can actually go a lot
faster and can deal with wrapping.

> > Why not subtract delta_exec and replenish when <0? That keeps the
> > numbers small.
> >
> 
> Accounting in the opposite direction allows us to catch-up in
> subsequent periods when a task exceeds its bandwidth across an
> interval where we are not able to immediately throttle it (e.g. costly
> syscall without config_prempt).  Since we'll continue to accrue the
> execution time in this case it will be effectively pre-charged against
> the next slice received.

Humm, how so, that's a simply matter of the quota going negative, right?