Re: [PATCH tip/core/rcu 0/3] rcu: prevent hangs by simplifying rcu_barrier/CPU-hotplug, fix lockdep complaint

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

On Wed, Oct 07, 2009 at 08:14:48AM +0200, Ingo Molnar wrote:
> 
> * Paul E. McKenney <paulmck@linux.vnet.ibm.com> wrote:
> 
> > This patchset fixes a couple of issues with TREE_PREEMPT_RCU:
> > 
> > 1.	The current implementation of rcu_barrier() waits for any
> > 	prior CPU-hotplug operation to complete.  This code path has
> > 	resulted in some hangs, so first move this code from rcupdate.c
> > 	to rcutree.c and then use a simpler algorithm that avoids the
> > 	waiting, eliminating the possibility of such hangs.
> > 
> > 2.	The lockdep facility complains from time to time due to the
> > 	fact that certain combinations of CPU-hotplug operations can
> > 	cause the TREE_PREEMPT_RCU code to acquire the root rcu_node
> > 	structure's lock while holding a leaf rcu_node structure's
> > 	lock.  This patch places the root rcu_node structure's lock
> > 	into its own lockdep class to prevent this false positive.
> > 
> > Ingo, please rewind tip/core/rcu to commit #135c8aea before applying this
> > series, as these patches really need to go into 2.6.32.  Reworked and
> > retested rcutiny (for 2.6.33, as Linus requested) will follow in a
> > few days.  Or more patches for more bug fixes, as the case may be.  ;-)
> > 
> >  b/kernel/rcupdate.c       |  120 -----------------------------------
> >  b/kernel/rcutree.c        |  120 +++++++++++++++++++++++++++++++++++
> >  b/kernel/rcutree.h        |   11 ++-
> >  b/kernel/rcutree_plugin.h |   34 ++++++++++
> >  b/kernel/rcutree_trace.c  |    5 -
> >  kernel/rcutree.c          |  156 ++++++++++++++++++++++++----------------------
> >  6 files changed, 248 insertions(+), 198 deletions(-)
> 
> Applied, thanks Paul!
> 
> I dropped these commits for now:
> 
>  3ffea79: rcu: Make hot-unplugged CPU relinquish its own RCU callbacks
>  eddd962: rcu: Move rcu_barrier() to rcutree, make lightweight rcu_barrier() for rcutiny
>  a39e7d5: rcu-tiny: The Bloatwatch Edition, v6
> 
> We can apply rcu-tiny for .33 again, once the dust has settled.

Sounds good!  I am placing rcu-tiny and synchronize_srcu_expedited()
on a separate branch in my local git tree for 2.6.33.

							Thanx, Paul