On Sat, Sep 22, 2012 at 10:25:59PM +0000, Paul Walmsley wrote:

On Sat, 22 Sep 2012, Paul E. McKenney wrote:

quoted

And here is a patch.  I am still having trouble reproducing the problem,
but figured that I should avoid serializing things.

Thanks, testing this now on v3.6-rc6.

Very cool, thank you!

                                       One question though about the patch 
description:

quoted

All this begs the question of exactly how a callback-free grace period
gets started in the first place.  This can happen due to the fact that
CPUs do not necessarily agree on which grace period is in progress.
If a CPU still believes that the grace period that just completed is
still ongoing, it will believe that it has callbacks that need to wait
for another grace period, never mind the fact that the grace period
that they were waiting for just completed.  This CPU can therefore
erroneously decide to start a new grace period.

Doesn't this imply that this bug would only affect multi-CPU systems?  

Surprisingly not, at least when running TREE_RCU or TREE_PREEMPT_RCU.
In order to keep lock contention down to a dull roar on larger systems,
TREE_RCU keeps three sets of books: (1) the global state in the rcu_state
structure, (2) the combining-tree per-node state in the rcu_node
structure, and the per-CPU state in the rcu_data structure.  A CPU is
not officially aware of the end of a grace period until it is reflected
in its rcu_data structure.  This has the perhaps-surprising consequence
that the CPU that detected the end of the old grace period might start
a new one before becoming officially aware that the old one ended.

Why not have the CPU inform itself immediately upon noticing that the
old grace period ended?  Deadlock.  The rcu_node locks must be acquired
from leaf towards root, and the CPU is holding the root rcu_node lock
when it notices that the grace period has ended.

I have made this a bit less problematic in the bigrt branch, working
towards a goal of getting RCU into a state where automatic formal
validation might one day be possible.  And yes, I am starting to get some
formal-validation people interested in this lofty goal, see for example:
http://sites.google.com/site/popl13grace/paper.pdf.

The recent tests here have been on Pandaboard, which is dual-CPU, but my 
recollection is that I also observed the warnings on a single-core 
Beagleboard.  Will re-test.

Anxiously awaiting the results.  This has been a strange one, even by
RCU's standards.

Plus I need to add a few Reported-by lines.  Next version...

							Thanx, Paul