Subj : Re: full-blown experiential smr-based reference counted pointer impl...
To   : comp.programming.threads
From : Joe Seigh
Date : Fri Apr 01 2005 05:39 pm

On Fri, 1 Apr 2005 12:27:52 -0800, Chris Thomasson <_no_damn_spam_cristom@_no_damn_comcast.net_spam> wrote:

>> Yes.  Though with classical RCU the quiesent states aren't so near the
>> critical
>> regions and the membars are around the quiesent state itself, so they're
>> not
>> near the critical regions either.  But the RCU implementation I did uses
>> membars like your example shows.
>
> Yeah. So does mine, I basically have it down to something like this:
>
>
> // Acquire RCU
> 1. store 1 into per-thread rcu_quiescent
> 2. ( acquire barrier )
> 3. increment per-thread rcu_generation
>
>
>   //  gc region
>
>
> // Release RCU
> 4. ( release barrier )
> 5. store 0 into per-thread rcu_quiescent
>
>
> As long as step 3 is made visible before step 5 is made visible, everything
> rocks. Plus, its loop-less and requires fewer memory barriers. Sure seems
> there would be a lot less overhead when protecting the increments of a
> single-word reference count...
>

With classical RCU, the quiescent states are voluntary context switches, so
the RCU read lock and unlock macros are no-ops.  That's about as less
overhead as you can get.

"classical RCU".  It's been around for about 20 years now.  :)

-- 
Joe Seigh

.