Re: 2.6.39.4: Oops in rcu_read_unlock_special()/_raw_spin_lock()

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Thu, Aug 25, 2011 at 09:20:51AM -0400, Will Simoneau wrote:
> On 14:27 Wed 24 Aug     , Paul E. McKenney wrote:
> > On Wed, Aug 24, 2011 at 05:19:07PM -0400, Will Simoneau wrote:
> > > The below Oops/BUGs were captured on a serial console during a large
> > > rsync job. I do not know of a way to reproduce the Oops, I've only seen
> > > it once. Some recent changes have been made suspiciously close to the
> > > exploding code, which makes me think that maybe 2.6.39-stable is lacking
> > > some fixes? The following commits from Linus' git seem vaguely related,
> > > although I have no idea how relevant they are to 2.6.39.4:
> > > 
> > >    ec433f0c (softirq,rcu: Inform RCU of irq_exit() activity)
> > >    10f39bb1 (rcu: protect __rcu_read_unlock() against scheduler-using
> > >              irq handlers)
> > 
> > If this failure mechanism really is the culprit, you should be able
> > to make failure happen much more frequently by inserting a delay in
> > __rcu_read_unlock() just prior to the call to rcu_read_unlock_special().
> > I would suggest starting with a few tens to hundreds of microseconds
> > worth of delay.
> > 
> > If this does make the failure reproducible, then it would make sense
> > to try applying the two patches you identified.
> 
> Hmm. I tried adding progressively larger delays in the spot you
> indicated. I went from 100uS to an entire 1S (!) and got no crash or
> deadlock. The target runs at 40MHz so the delays do need to be
> relatively long compared to modern machines.
> 
> My hardware breakpoint as well as printk tests confirm that
> rcu_read_unlock_special() really does get called multiple times per
> second, and the 1S delay makes it painfully obvious as well. But, no
> dice.

Well, you can always apply the two patches above anyway, but it is hard
to prove what the underlying problem really is in your case.

							Thanx, Paul