FWIW, good to see progress, still waiting for you guys to agree :-)
On Mon, Jul 01, 2019 at 01:15:44PM -0700, bsegall@xxxxxxxxxx wrote:
> - Taking up-to-every rq->lock is bad and expensive and 5ms may be too
> short a delay for this. I haven't tried microbenchmarks on the cost of
> this vs min_cfs_rq_runtime = 0 vs baseline.
Yes, that's tricky, SGI/HPE have definite ideas about that.
> @@ -4781,12 +4790,41 @@ static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
> */
> static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
> {
> - u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
> + u64 runtime = 0;
> unsigned long flags;
> u64 expires;
> + struct cfs_rq *cfs_rq, *temp;
> + LIST_HEAD(temp_head);
> +
> + local_irq_save(flags);
> +
> + raw_spin_lock(&cfs_b->lock);
> + cfs_b->slack_started = false;
> + list_splice_init(&cfs_b->slack_cfs_rq, &temp_head);
> + raw_spin_unlock(&cfs_b->lock);
> +
> +
> + /* Gather all left over runtime from all rqs */
> + list_for_each_entry_safe(cfs_rq, temp, &temp_head, slack_list) {
> + struct rq *rq = rq_of(cfs_rq);
> + struct rq_flags rf;
> +
> + rq_lock(rq, &rf);
> +
> + raw_spin_lock(&cfs_b->lock);
> + list_del_init(&cfs_rq->slack_list);
> + if (!cfs_rq->nr_running && cfs_rq->runtime_remaining > 0 &&
> + cfs_rq->runtime_expires == cfs_b->runtime_expires) {
> + cfs_b->runtime += cfs_rq->runtime_remaining;
> + cfs_rq->runtime_remaining = 0;
> + }
> + raw_spin_unlock(&cfs_b->lock);
> +
> + rq_unlock(rq, &rf);
> + }
But worse still, you take possibly every rq->lock without ever
re-enabling IRQs.
>
> /* confirm we're still not at a refresh boundary */
> - raw_spin_lock_irqsave(&cfs_b->lock, flags);
> + raw_spin_lock(&cfs_b->lock);
> cfs_b->slack_started = false;
> if (cfs_b->distribute_running) {
> raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
