Re: specjbb2005 and aim7 regression with 2.6.32-rc kernels

From: Mike Galbraith <efault@gmx.de>
To: ego@in.ibm.com
Cc: Peter Zijlstra <peterz@infradead.org>,
	"Zhang, Yanmin" <yanmin_zhang@linux.intel.com>,
	Ingo Molnar <mingo@elte.hu>, LKML <linux-kernel@vger.kernel.org>
Subject: Re: specjbb2005 and aim7 regression with 2.6.32-rc kernels
Date: Mon, 09 Nov 2009 10:57:17 +0100	[thread overview]
Message-ID: <1257760637.7124.26.camel@marge.simson.net> (raw)
In-Reply-To: <20091109094425.GA26733@in.ibm.com>

On Mon, 2009-11-09 at 15:14 +0530, Gautham R Shenoy wrote:
> On Mon, Nov 09, 2009 at 10:15:04AM +0100, Peter Zijlstra wrote:
> > On Mon, 2009-11-09 at 08:09 +0100, Mike Galbraith wrote:
> > > +       smp_read_barrier_depends();
> > >         cpumask_setall(cpus);
> > > +       cpumask_and(cpus, cpus, cpu_online_mask);
> > 
> > 
> > how about: cpumask_copy(cpus, cpu_online_mask); ?
> > 
> > Also, iirc cpu_online_mask is guaranteed stable when preemption is
> > disabled, otherwise you need to use get/put_online_cpus(), an
> > rmb_depends() won't do.
> 
> preempt_disable() guarantees that any cpus won't go offline, since we
> use stop_machine() to take CPUs offline. I don't think it provides cover
> against new cpus coming online.

That's exactly the problem I'm having with newidle.  Without that
barrier, even with the cpumask_and(), it still balances offline cpus.

	-Mike