Re: Resend: 3.10.32-rt31: problem with "rcu: Eliminate softirq processing from rcutree"

[Steven Rostedt <rostedt@goodmis.org>]

On Tue, 18 Mar 2014 12:05:31 +0100
Sebastian Andrzej Siewior <bigeasy@linutronix.de> wrote:

> * Paul E. McKenney | 2014-03-08 12:04:00 [-0800]:
> 
> >On Sat, Mar 08, 2014 at 08:08:35PM +0100, Helmut Buchsbaum wrote:
> >> This is the correct point to tweak the system, I was not yet aware of!
> >> Thanks a lot for the hint. I just enabled RCU_BOOST with its default
> >> settings and the system runs smooth and stable again even in heavy
> >> load situations. Maybe the default for RCU_BOOST should be set to Y on
> >> systems using PREEMPT_RT_FULL ?
> >
> >I must defer to Sebastian on this, but seems like a good approach.
> 
> Hmm. There are a few ways to break an RT system. Afaik RCU_BOOST was
> recommended even before this change. 
> 
> Steven, tglx any opinion on that?
> 
> Another thing we could do is to register_shrinker() which is called
> before OOM. But then boosting is simple enough.

I thought RCU_BOOST was already set for default y when PREEMPT_RT is
set. If not, then please add it. Note, the default should be set, but
not selected. Still let the user disable it.

 default y if PREEMPT_RT_FULL

Thanks,

-- Steve