preemptible RCU

Hello!

This patchset is an update of that posted by Dipankar last January
(http://lkml.org/lkml/2007/1/15/133).  This is work in progress, not yet
ready for inclusion.  It passes rcutorture on i386, x86_64, and ppc64
boxes as well as kernbench, so should be safe for experimentation.  As
with Dipankar's previous post, this variant of preemptible rcu_read_lock()
and rcu_read_unlock may be invoked from NMI/SMI handlers, and do not
contain any heavyweight atomic operations or memory barriers (although
they do still momentarily disable IRQs).  This patchset features
a fully parallel grace-period computation, which will become increasingly
important with upcoming multicore/multi-threaded CPUs.  In addition,
this patchset provides a preemptible-RCU variant of synchronize_sched()
that avoids the previous deadlock with CPU hotplug -- this variant may
eventually prove unnecessary, but is offered in the spirit of separating
concerns.

Next steps:  (1) Integrate with CPU hotplug.  (2) Re-merge RCU priority
boosting.  (3) Fix some naming issues.  Longer term work includes
optimized dyntick operation and eliminating the interrupt disabling
in rcu_read_lock() and rcu_read_unlock().

						Thanx, Paul