Re: [PATCH 1/7] sched/core: uclamp: add CPU clamp groups accounting

[Peter Zijlstra]

On Fri, Apr 13, 2018 at 12:08:48PM +0100, Patrick Bellasi wrote:
> On 13-Apr 11:46, Peter Zijlstra wrote:
> > On Mon, Apr 09, 2018 at 05:56:09PM +0100, Patrick Bellasi wrote:
> > > +static inline void uclamp_cpu_get(struct task_struct *p, int cpu, int clamp_id)
> > > +{
> > > +	struct uclamp_cpu *uc_cpu = &cpu_rq(cpu)->uclamp[clamp_id];
> > > +	int clamp_value;
> > > +	int group_id;
> > > +
> > > +	/* Get task's specific clamp value */
> > > +	clamp_value = p->uclamp[clamp_id].value;
> > > +	group_id = p->uclamp[clamp_id].group_id;
> > > +
> > > +	/* No task specific clamp values: nothing to do */
> > > +	if (group_id == UCLAMP_NONE)
> > > +		return;
> > > +
> > > +	/* Increment the current group_id */
> > 
> > That I think qualifies being called a bad comment.
> 
> my bad :/
> 
> > > +	uc_cpu->group[group_id].tasks += 1;
> > > +
> > > +	/* Mark task as enqueued for this clamp index */
> > > +	p->uclamp_group_id[clamp_id] = group_id;
> > 
> > Why exactly do we need this? we got group_id from @p in the first place.
> 
> The idea is to back-annotate on the task the group in which it has
> been refcounted. That allows a much simpler and less racy refcount
> decrements at dequeue/migration time.

I'm not following; the only possible reason for having this second copy
of group_id is when your original value (p->uclamp[clamp_id].group_id)
can change between enqueue and dequeue.

Why can this happen?

> That's also why we have a single call-back, uclamp_task_update(),
> for both enqueue/dequeue. Depending on the check performed by
> uclamp_task_affects() we know if we have to get or put the refcounter.

But that check is _completely_ redundant, because you already _know_
from being in the en/de-queue path. So having that single callback is
actively harmfull (and confusing).