Re: [PATCH] TOMOYO: Add garbage collector support. (v3)

[Peter Zijlstra <peterz@infradead.org>]

On Wed, 2009-06-17 at 20:19 +0900, Tetsuo Handa wrote:
> Paul E. McKenney wrote ( http://lkml.org/lkml/2009/5/27/2 ) :
> > I would also recommend the three-part LWN series as a starting point:
> > 
> > #       http://lwn.net/Articles/262464/ (What is RCU, Fundamentally?)
> > #       http://lwn.net/Articles/263130/ (What is RCU's Usage?)
> > #       http://lwn.net/Articles/264090/ (What is RCU's API?)
> I've read these articles. They are very good.
> 
> I came up with an idea that we may be able to implement GC while readers are
> permitted to sleep but no read locks are required.
> 
> The idea is to have two counters which hold the number of readers currently
> reading the list, one is active and the other is inactive. Reader increments
> the currently active counter before starts reading and decrements that counter
> after finished reading. GC swaps active counter and inactive counter and waits
> for previously active counter's count to become 0 before releasing elements
> removed from the list.
> Code is shown below.
> 
> atomic_t users_counter[2];
> atomic_t users_counter_idx;
> DEFINE_MUTEX(updator_mutex);
> DEFINE_MUTEX(gc_mutex);

Sounds like an utter scalability nightmare to me though.

Why not 'simply' use SRCU or always provide an preemptible RCU domain
using:

rcu_read_lock_preempt()
rcu_read_unlock_preempt()
call_rcu_preempt()
etc.

along with the already existing

*{,_bh,_sched} variants

That way PREEMPT_RCU would only affect the implementation of the regular
RCU implementation, it being either _sched or _preempt.