From mboxrd@z Thu Jan  1 00:00:00 1970
Return-Path: <linux-kernel-owner+w=401wt.eu-S1755196AbYFJPvh@vger.kernel.org>
Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand
	id S1755196AbYFJPvh (ORCPT <rfc822;w@1wt.eu>);
	Tue, 10 Jun 2008 11:51:37 -0400
Received: (majordomo@vger.kernel.org) by vger.kernel.org id S1754458AbYFJPvY
	(ORCPT <rfc822;linux-kernel-outgoing>);
	Tue, 10 Jun 2008 11:51:24 -0400
Received: from viefep18-int.chello.at ([213.46.255.22]:5389 "EHLO
	viefep14-int.chello.at" rhost-flags-OK-OK-OK-FAIL) by vger.kernel.org
	with ESMTP id S1754450AbYFJPvX (ORCPT
	<rfc822;linux-kernel@vger.kernel.org>);
	Tue, 10 Jun 2008 11:51:23 -0400
Subject: Re: [PATCH -rt 5/5] cpu-hotplug: cpu_down vs preempt-rt
From: Peter Zijlstra <a.p.zijlstra@chello.nl>
To: paulmck@linux.vnet.ibm.com
Cc: linux-kernel@vger.kernel.org, Ingo Molnar <mingo@elte.hu>,
       Thomas Gleixner <tglx@linutronix.de>,
       Steven Rostedt <rostedt@goodmis.org>,
       Clark Williams <williams@redhat.com>,
       Gregory Haskins <ghaskins@novell.com>,
       Gautham R Shenoy <ego@in.ibm.com>,
       Pekka Enberg <penberg@cs.helsinki.fi>,
       Arnaldo Carvalho de Melo <acme@redhat.com>
In-Reply-To: <20080610153301.GD15481@linux.vnet.ibm.com>
References: <20080610111259.766940257@chello.nl>
	 <20080610111832.969119014@chello.nl>
	 <20080610153301.GD15481@linux.vnet.ibm.com>
Content-Type: text/plain
Date: Tue, 10 Jun 2008 17:51:18 +0200
Message-Id: <1213113078.31518.16.camel@twins>
Mime-Version: 1.0
X-Mailer: Evolution 2.22.1.1 
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
List-ID: <linux-kernel.vger.kernel.org>
X-Mailing-List: linux-kernel@vger.kernel.org

On Tue, 2008-06-10 at 08:33 -0700, Paul E. McKenney wrote:
> On Tue, Jun 10, 2008 at 01:13:04PM +0200, Peter Zijlstra wrote:
> > idle_task_exit() calls mmdrop() from the idle thread, but in PREEMPT_RT all the
> > allocator locks are sleeping locks - for obvious reasons scheduling away the
> > idle thread gives some curious problems.
> > 
> > Solve this by pushing the mmdrop() into an RCU callback, however we can't use
> > RCU because the CPU is already down and all the local RCU state has been
> > destroyed.
> > 
> > Therefore create a new call_rcu() variant that enqueues the callback on an
> > online cpu.
> 
> I am a bit nervous about the non-determinism, but on the other hand
> CPU online/offline events can only happen so often due to the locking.
> 
> So...
> 
> Reviewed-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>

Thanks!

Yesterday you suggested using rcu_cpu_online_map and fliplock to avoid
the loop here:

> > +void fastcall call_rcu_preempt_online(struct rcu_head *head,
> > +		void (*func)(struct rcu_head *rcu))
> > +{
> > +	struct rcu_data *rdp;
> > +	unsigned long flags;
> > +	int cpu;
> > +
> > +	head->func = func;
> > +	head->next = NULL;
> > +again:
> > +	cpu = first_cpu(cpu_online_map);
> > +	rdp = RCU_DATA_CPU(cpu);
> > +
> > +	spin_lock_irqsave(&rdp->lock, flags);
> > +	if (unlikely(!cpu_online(cpu))) {
> > +		/*
> > +		 * cpu is removed from the online map before rcu_offline_cpu
> > +		 * is called.
> > +		 */
> > +		spin_unlock_irqrestore(&rdp->lock, flags);
> > +		goto again;
> > +	}
> > +
> > +	*rdp->nexttail = head;
> > +	rdp->nexttail = &head->next;
> > +	spin_unlock_irqrestore(&rdp->lock, flags);
> > +
> > +}

But then the code would look like:

  spin_lock_irqsave(&rcu_ctrlblk.fliplock, flags);
  cpu = first_cpu(rcu_cpu_online_map);
  rdp = RCU_DATA_CPU(cpu);
  spin_lock(&rdp->lock);

creating a nesting between these two locks, where I could not find one.

Do you still prefer I look into changing it into such a form, or are you
sufficiently non-caring that the current code can stand? :-)