Re: [RFC PATCH 3/3] sched: Implement shared wakequeue in CFS

[David Vernet]

On Wed, Jun 21, 2023 at 04:20:20PM +0200, Peter Zijlstra wrote:
> On Tue, Jun 13, 2023 at 12:20:04AM -0500, David Vernet wrote:
> > +struct swqueue {
> > +	struct list_head list;
> > +	spinlock_t lock;
> > +} ____cacheline_aligned;
> 
> I'm thinking you can shard this just fine, it makes that pop() needs to
> iterate all shards, but that shouldn't be a problem, and it would still
> only need to take a single lock.

Is the idea here to have multiple sharded queues per LLC, with a single
lock protecting them? Assuming so, is the idea that it would avoid
bouncing the list heads amongst all of the cores' cachelines? I could
see that being useful in some scenarios, but it also feels a bit
complicated for what it gives you. If you have tasks being pulled
quickly I don't think that will help much because the list heads will
just bounce to the pullers. Also, if the lock is already heavily
contended, it seems possible that the cores inside a single shard could
still bounce the head back and forth amongst them, or the cache line
bouncing will be a small-order overhead compared to the lock itself.

Or did I misunderstand your suggestion entirely?

> I'm thinking 4 or 8 shards should be plenty, even for Intel LLC.
> 
> >  #ifdef CONFIG_SMP
> 
> > +static struct task_struct *swqueue_pull_task(struct swqueue *swqueue)
> > +{
> > +	unsigned long flags;
> > +
> > +	struct task_struct *p;
> > +
> > +	spin_lock_irqsave(&swqueue->lock, flags);
> > +	p = list_first_entry_or_null(&swqueue->list, struct task_struct,
> > +				     swqueue_node);
> > +	if (p)
> > +		list_del_init(&p->swqueue_node);
> > +	spin_unlock_irqrestore(&swqueue->lock, flags);
> > +
> > +	return p;
> > +}
> 
> Would this not normally be called pop() or somesuch?

Yes, I'll improve the name in the next iteration. swqueue_dequeue() and
swqueue_enqueue() seem like the most canonical. Let me know if you have another
preference.

> 
> > +static void swqueue_enqueue(struct rq *rq, struct task_struct *p, int enq_flags)
> > +{
> > +	unsigned long flags;
> > +	struct swqueue *swqueue;
> > +	bool task_migrated = enq_flags & ENQUEUE_MIGRATED;
> > +	bool task_wakeup = enq_flags & ENQUEUE_WAKEUP;
> > +
> > +	/*
> > +	 * Only enqueue the task in the shared wakequeue if:
> > +	 *
> > +	 * - SWQUEUE is enabled
> > +	 * - The task is on the wakeup path
> > +	 * - The task wasn't purposefully migrated to the current rq by
> > +	 *   select_task_rq()
> > +	 * - The task isn't pinned to a specific CPU
> > +	 */
> > +	if (!task_wakeup || task_migrated || p->nr_cpus_allowed == 1)
> > +		return;
> 
> Elsewhere you mentioned heuristics, this smells like them. This and the
> is_cpus_allowed() thing makes you loose plenty of opportunities.

Yeah fair enough, these certainly are heuristics as well.

I thought it best to try and avoid swqueue getting in the way of
select_task_rq_fair() (at least to start out with), but we could always
remove that and run other experiments to see how it does.

> > +	swqueue = rq_swqueue(rq);
> > +	spin_lock_irqsave(&swqueue->lock, flags);
> > +	list_add_tail(&p->swqueue_node, &swqueue->list);
> > +	spin_unlock_irqrestore(&swqueue->lock, flags);
> > +}
> > +
> >  static int swqueue_pick_next_task(struct rq *rq, struct rq_flags *rf)
> >  {
> > -	return 0;
> > +	struct swqueue *swqueue;
> > +	struct task_struct *p = NULL;
> > +	struct rq *src_rq;
> > +	struct rq_flags src_rf;
> > +	int ret;
> > +
> > +	swqueue = rq_swqueue(rq);
> > +	if (!list_empty(&swqueue->list))
> > +		p = swqueue_pull_task(swqueue);
> > +
> > +	if (!p)
> > +		return 0;
> 
> At this point you can do the whole is_cpu_allowed() and avoid the whole
> lock dance if not.

Good idea, will incorporate into the next iteration.

> > +
> > +	rq_unpin_lock(rq, rf);
> > +	raw_spin_rq_unlock(rq);
> > +
> > +	src_rq = task_rq_lock(p, &src_rf);
> > +
> > +	if (task_on_rq_queued(p) && !task_on_cpu(rq, p))
> > +		src_rq = migrate_task_to(src_rq, &src_rf, p, cpu_of(rq));
> 
> And then this becomes move_queued_task().

Yep, will make this change per your suggestion in [0].

[0]: https://lore.kernel.org/all/20230621130439.GF2053369@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx/

> > +	if (src_rq->cpu != rq->cpu)
> > +		ret = 1;
> > +	else
> > +		ret = -1;
> > +
> > +	task_rq_unlock(src_rq, p, &src_rf);
> > +
> > +	raw_spin_rq_lock(rq);
> > +	rq_repin_lock(rq, rf);
> > +
> > +	return ret;
> >  }
> >  
> >  static void swqueue_remove_task(struct task_struct *p)
> > +{
> > +	unsigned long flags;
> > +	struct swqueue *swqueue;
> > +
> > +	if (!list_empty(&p->swqueue_node)) {
> > +		swqueue = rq_swqueue(task_rq(p));
> > +		spin_lock_irqsave(&swqueue->lock, flags);
> > +		list_del_init(&p->swqueue_node);
> > +		spin_unlock_irqrestore(&swqueue->lock, flags);
> > +	}
> > +}
> 
> dequeue()

Ack