Re: [RFC 15/16] sched/fair: Account kthread runtime debt for CFS bandwidth

[Daniel Jordan <daniel.m.jordan@oracle.com>]

On Fri, Jan 14, 2022 at 10:31:55AM +0100, Peter Zijlstra wrote:
> On Wed, Jan 05, 2022 at 07:46:55PM -0500, Daniel Jordan wrote:
> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > index 44c452072a1b..3c2d7f245c68 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -4655,10 +4655,19 @@ static inline u64 sched_cfs_bandwidth_slice(void)
> >   */
> >  void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
> >  {
> > -	if (unlikely(cfs_b->quota == RUNTIME_INF))
> > +	u64 quota = cfs_b->quota;
> > +	u64 payment;
> > +
> > +	if (unlikely(quota == RUNTIME_INF))
> >  		return;
> >  
> > -	cfs_b->runtime += cfs_b->quota;
> > +	if (cfs_b->debt) {
> > +		payment = min(quota, cfs_b->debt);
> > +		cfs_b->debt -= payment;
> > +		quota -= payment;
> > +	}
> > +
> > +	cfs_b->runtime += quota;
> >  	cfs_b->runtime = min(cfs_b->runtime, cfs_b->quota + cfs_b->burst);
> >  }
> 
> It might be easier to make cfs_bandwidth::runtime an s64 and make it go
> negative.

Yep, nice, no need for a new field in cfs_bandwidth.

> > @@ -5406,6 +5415,32 @@ static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
> >  	rcu_read_unlock();
> >  }
> >  
> > +static void incur_cfs_debt(struct rq *rq, struct sched_entity *se,
> > +			   struct task_group *tg, u64 debt)
> > +{
> > +	if (!cfs_bandwidth_used())
> > +		return;
> > +
> > +	while (tg != &root_task_group) {
> > +		struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
> > +
> > +		if (cfs_rq->runtime_enabled) {
> > +			struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
> > +			u64 payment;
> > +
> > +			raw_spin_lock(&cfs_b->lock);
> > +
> > +			payment = min(cfs_b->runtime, debt);
> > +			cfs_b->runtime -= payment;
> 
> At this point it might hit 0 (or go negative if/when you do the above)
> and you'll need to throttle the group.

I might not be following you, but there could be cfs_rq's with local
runtime_remaining, so even if it goes 0 or negative, the group might
still have quota left and so shouldn't be throttled right away.

I was thinking the throttling would happen as normal, when a cfs_rq ran
out of runtime_remaining and failed to refill it from
cfs_bandwidth::runtime.

> > +			cfs_b->debt += debt - payment;
> > +
> > +			raw_spin_unlock(&cfs_b->lock);
> > +		}
> > +
> > +		tg = tg->parent;
> > +	}
> > +}
> 
> So part of the problem I have with this is that these external things
> can consume all the bandwidth and basically indefinitely starve the
> group.
> 
> This is doulby so if you're going to account things like softirq network
> processing.

Yes.  As Tejun points out, I'll make sure remote charging doesn't run
away.

> Also, why does the whole charging API have a task argument? It either is
> current or NULL in case of things like softirq, neither really make
> sense as an argument.

@task distinguishes between NULL for softirq and current for everybody
else.

It's possible to detect this difference internally though, if that's
what you're saying, so @task can go away.

> Also, by virtue of this being a start-stop annotation interface, the
> accrued time might be arbitrarily large and arbitrarily delayed. I'm not
> sure that's sensible.

Yes, that is a risk.  With start-stop, users need to be careful to
account often enough and have a "reasonable" upper bound on period
length, however that's defined.  Multithreaded jobs are probably the
worst offender since these threads charge a sizable amount at once
compared to the other use cases.

> For tasks it might be better to mark the task and have the tick DTRT
> instead of later trying to 'migrate' the time.

Ok, I'll try that.  The start-stop approach keeps remote charging from
adding overhead in the tick for non-remote-charging things, far and away
the common case, but I'll see how expensive the tick-based approach is.

Can hide it behind a static branch for systems not using the cpu
contoller.