Re: [GIT PULL rcu/next] rcu commits for 2.6.40

[Ingo Molnar <mingo@elte.hu>]

* Paul E. McKenney <paulmck@linux.vnet.ibm.com> wrote:

> > Would it have been possible to split it in two, one for the movement of the 
> > notifiers, the other for the barrier changes?
> > 
> > That way the bisection would have fingered the movement commit. Or so.
> 
> In hindsight, that certainly would have been better.

This is the Linux kernel and we *can* turn back the clock!

> I was afraid of that...
> 
> On the off-chance that moving the memory barriers was at fault, the following 
> patch restores all of them that don't have in situ replacements.  Grasping at 
> straws, admittedly.

Well, the nice thing is that we really do not have to grasp at straws, and even 
while we have no good ideas we can debug this *much* better.

Could you please do a simple test-tree that does has 3 commits:

 first one reverts the offending commit
 second one applies the barrier part of it
 this one applies the need_resched part of it

( You can do even more finegrained steps, if you find harmless-looking bits of 
  it that can be applied separately! )

Note, the important thing is that the tree should be a 'null pull' - i.e. the 
revert plus the patches applied will not change anything in core/rcu.

Obviously it would be nice if each step built fine - no need to boot test each 
step as long as you are reasonably sure it will boot fine.

Then i could take my reproducer and come up with a very precise bisection 
result for you, with just a couple of minutes time spent on testing. One of the 
commits after the revert will trigger the hang/slowdown.

My prediction is that we will be much wiser after that! :-)

Thanks,

	Ingo