Re: [COUNTERPATCH] mm: avoid overflowing preempt_count() in mmu_take_all_locks()

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

On Thu, Apr 01, 2010 at 06:36:14PM +0200, Peter Zijlstra wrote:
> On Thu, 2010-04-01 at 09:15 -0700, Paul E. McKenney wrote:
> > > I don't understand.  I thought the problem was that the locks were
> > > taken inside an rcu critical section; switching to srcu would fix
> > > that.  But how is call_rcu_preempt() related?  Grepping a bit, what
> > > is call_rcu_preempt()?  my tree doesn't have it.
> > 
> > I believe that Peter is referring to the RCU implementation you get
> > with CONFIG_TREE_PREEMPT_RCU, which currently depends on CONFIG_PREEMPT.
> > The other implementation is CONFIG_TREE_RCU, which is usually called
> > "classic RCU". 
> 
> Right, so I've been nudging Paul a while to make it so that we always
> have preemptible rcu available and that only the default interface
> switches between sched/classic and preempt.
> 
> Currently we already have:
> 
> call_rcu_sched()
> call_rcu_bh()
> call_rcu()	(depends on CONFIG_PREEMPT_RCU)
> 
> I'm saying it would be nice to also have:
> 
> call_rcu_preempt()

And, given the !CONFIG_PREEMPT issue, along with the issue of
sleeping forever in RCU read-side critical sections, my counteroffer
has been to integrate SRCU into the treercu (and of course the
tinyrcu) implementations, thus getting roughly the same performance
as CONFIG_TREE_RCU.

Delivering on this counteroffer has proven to be another kettle of fish,
although I am making some progress.  It will be several months, best case.

							Thanx, Paul