Re: [Bug #10815] 2.6.26-rc4: RIP find_pid_ns+0x6b/0xa0

From: Paul E. McKenney
Date: Mon Jun 16 2008 - 00:40:05 EST


On Mon, Jun 16, 2008 at 07:01:54AM +0400, Alexey Dobriyan wrote:
> On Sun, Jun 15, 2008 at 04:38:24PM -0700, Linus Torvalds wrote:
> > On Sun, 15 Jun 2008, Paul E. McKenney wrote:
> > >
> > > If that is the right thing to do. I would feel better about it if I
> > > could reproduce the failure.
> >
> > Is there some configuration thing that might make it happen for Alexey but
> > not for you? The only obvious one is the normal PREEMPT thing, but
> > PREEMPT_RCU obviously depends on that, so there must be something else.
> >
> > CONFIG_SMP? CONFIG_NO_HZ? CONFIG_[RT|FAIR]_GROUP_SCHED? There's a number
> > of config options that change scheduling details..
>
> The only way I can "fix" rcutorture here is SMP=n (not maxcpus=1!).

I am running SMP=y without seeing the problem.

> Among things tried and do not matter are NO_HZ, HIGH_RES_TIMERS, turning off
> debugging options (one by one and in bulk), DEBUG_PER_CPU_MAPS, HZ tweaks,
> SCHED_MC.
>
> All of this doesn't help, but SMP=n does.

Attached is the compressed config file I used for a run on a POWER
machine. I have used similar config files for x86. For this run, I built
LTP, installed it, started rcutorture, ran LTP, then ended rcutorture.

The only other thing I can think of is that Alexey's machines are
somehow managing to nest NMIs or something. I can produce a fix for
that, but have no way of testing it.

Thoughts?

Thanx, Paul

Attachment: dotconfig.bz2
Description: dotconfig.bz2