Re: [PATCH tip/core/rcu 09/15] rcu: Increasing rcu_barrier() concurrency

[Josh Triplett <josh@joshtriplett.org>]

On Fri, Jun 15, 2012 at 02:06:04PM -0700, Paul E. McKenney wrote:
> From: "Paul E. McKenney" <paul.mckenney@linaro.org>
> 
> The traditional rcu_barrier() implementation has serialized all requests,
> regardless of RCU flavor, and also does not coalesce concurrent requests.
> In the past, this has been good and sufficient.
> 
> However, systems are getting larger and use of rcu_barrier() has been
> increasing.  This commit therefore introduces a counter-based scheme
> that allows _rcu_barrier() calls for the same flavor of RCU to take
> advantage of each others' work.
> 
> Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
> ---
>  kernel/rcutree.c |   27 ++++++++++++++++++++++++++-
>  kernel/rcutree.h |    2 ++
>  2 files changed, 28 insertions(+), 1 deletions(-)
> 
> diff --git a/kernel/rcutree.c b/kernel/rcutree.c
> index 93358d4..7c299d3 100644
> --- a/kernel/rcutree.c
> +++ b/kernel/rcutree.c
> @@ -2291,13 +2291,32 @@ static void _rcu_barrier(struct rcu_state *rsp)
>  	unsigned long flags;
>  	struct rcu_data *rdp;
>  	struct rcu_data rd;
> +	unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
> +	unsigned long snap_done;
>  
>  	init_rcu_head_on_stack(&rd.barrier_head);
>  
>  	/* Take mutex to serialize concurrent rcu_barrier() requests. */
>  	mutex_lock(&rsp->barrier_mutex);
>  
> -	smp_mb();  /* Prevent any prior operations from leaking in. */
> +	/*
> +	 * Ensure tht all prior references, including to ->n_barrier_done,
> +	 * are ordered before the _rcu_barrier() machinery.
> +	 */
> +	smp_mb();  /* See above block comment. */

If checkpatch complains about the lack of a comment to the right of a
barrier even when the barrier has a comment directly above it, that
seems like a bug in checkpatch that needs fixing, to prevent developers
from having to add noise like "See above block comment.". :)

Also: what type of barriers do mutex_lock and mutex_unlock imply?  I
assume they imply some weaker barrier than smp_mb, but I'd still assume
they imply *some* barrier.

> +	/* Recheck ->n_barrier_done to see if others did our work for us. */
> +	snap_done = ACCESS_ONCE(rsp->n_barrier_done);
> +	if (ULONG_CMP_GE(snap_done, ((snap + 1) & ~0x1) + 2)) {

This calculation seems sufficiently clever that it merits an explanatory
comment.

> +		smp_mb();
> +		mutex_unlock(&rsp->barrier_mutex);
> +		return;
> +	}
> +
> +	/* Increment ->n_barrier_done to avoid duplicate work. */
> +	ACCESS_ONCE(rsp->n_barrier_done)++;

Interesting dissonance here: the use of ACCESS_ONCE with ++ implies
exactly two accesses, rather than exactly one.  What makes it safe to
not use atomic_inc here, but not safe to drop the ACCESS_ONCE?
Potential use of a cached value read earlier in the function?

- Josh Triplett