Re: [RFC v5 9/9] sched/deadline: also reclaim bandwidth not used by dl tasks

From: Peter Zijlstra <peterz@infradead.org>
To: luca abeni <luca.abeni@santannapisa.it>
Cc: linux-kernel@vger.kernel.org, Ingo Molnar <mingo@redhat.com>,
	Juri Lelli <juri.lelli@arm.com>,
	Claudio Scordino <claudio@evidence.eu.com>,
	Steven Rostedt <rostedt@goodmis.org>,
	Tommaso Cucinotta <tommaso.cucinotta@sssup.it>,
	Daniel Bristot de Oliveira <bristot@redhat.com>,
	Joel Fernandes <joelaf@google.com>,
	Mathieu Poirier <mathieu.poirier@linaro.org>
Subject: Re: [RFC v5 9/9] sched/deadline: also reclaim bandwidth not used by dl tasks
Date: Mon, 27 Mar 2017 16:03:41 +0200	[thread overview]
Message-ID: <20170327140341.yvjjr6hbow2jug3t@hirez.programming.kicks-ass.net> (raw)
In-Reply-To: <1490327582-4376-10-git-send-email-luca.abeni@santannapisa.it>

On Fri, Mar 24, 2017 at 04:53:02AM +0100, luca abeni wrote:

> +static inline
> +void __dl_update(struct dl_bw *dl_b, s64 bw)
> +{
> +	struct root_domain *rd = container_of(dl_b, struct root_domain, dl_bw);
> +	int i;
> +
> +	RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
> +			 "sched RCU must be held");
> +	for_each_cpu_and(i, rd->span, cpu_active_mask) {
> +		struct rq *rq = cpu_rq(i);
> +
> +		rq->dl.extra_bw += bw;
> +	}

So this is unfortunate (and we already have one such instance).

It effectively does an for_each_online_cpu() with IRQs disabled, and on
SGI class hardware that takes _forever_.

This is also what I got stuck on trying to rewrite AC to use Tommaso's
recoverable thing. In the end I had to do a 2 stage try/commit variant.
Which ended up being a pain and I didn't finish.

I'm not saying this patch is bad, but this is something we need to thing
about.