Re: [PATCH] Read-Copy Update 2.5.34

[Dipankar Sarma <dipankar@in.ibm.com>]

On Wed, Sep 11, 2002 at 01:24:21PM +0200, Ingo Molnar wrote:
> 
> i dont really understand why it has to change the scheduler. You want a
> facility to force a reschedule on any given CPU, correct?
> 

Hi Ingo,

Yes, like force_cpu_reschedule().

Apart from that RCU adds some minor things in the scheduler code -

1. Per-CPU context switch counter increment in the fast path
2. For preemptive kernels, a conditional branch in the voluntary
   context switch code path checking whether the current task may 
   have had an involuntary context switch earlier [rcu_preempt_put()].
3. Adding a field to the task structure - cpu_preempt_cntr.
4. RCU checking hook into the scheduler_tick(), but this is not
   in the fast path.

#2 and #3 are necessary to support call_rcu_preempt() which
allows preemption-safe reads of data protected by RCU. A preempted
task may still contain references to RCU protected data and
the RCU grace period needs to be prolonged until all such tasks
before the update do voluntary context switches.

I did some reflex benchmarking to make sure that I didn't introduce 
any false sharing by mistake in scheduler fast path and the results 
look comparable -

(4CPU P3 Xeon with 1MB L2 + 1GB RAM)

			vanilla-2.5.34	rcu_poll-2.5.34
			--------------  ---------------
80 , 40 , 		1.593		1.569
112 , 40 , 		1.544		1.554
144 , 40 , 		1.595		1.552
176 , 40 , 		1.568		1.605
198 , 40 , 		1.562		1.577
230 , 40 , 		1.563		1.583
244 , 40 , 		1.671		1.638

Not sure how reliable these numbers are.

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