At OLS, we presented some data on latency and overhead for
different approaches to RCU. After some discussions with Andrea and Rusty,
it seemed that it is a good idea to use this and come out with
a patch. So here is a new patch. The main features are -

1. Based on rcu_poll which is already in production kernels (SuSE)
   and has low latency of grace period.
2. In order to reduce overheads, I added per-CPU queue support to
   it.
3. Works in preemptive kernel environment by either requiring
   the use of rcu_read_lock/unlock or using call_rcu_preempt().
4. synchronize_kernel() uses call_rcu_preempt() which requires
   no disabling of preemption on read-side making it useful for
   things like CPU hotplug.
5. Added docbook tags in headers.

Future plans are -

1. More overhead and latency measurements and tweaking around if
   necessary. But for infrequent things like CPU hotplug, this
   is not important.
2. Cleanups/bugfixes etc.

Comments, suggestions and even flames are welcome.

Thanks
-- 
Dipankar Sarma  <dipankar@in.ibm.com> http://lse.sourceforge.net
Linux Technology Center, IBM Software Lab, Bangalore, India.