Re: [ltt-dev] [RFC git tree] Userspace RCU (urcu) for Linux (repost)

[Mathieu Desnoyers <compudj@krystal.dyndns.org>]

* Paul E. McKenney (paulmck@linux.vnet.ibm.com) wrote:
[...]
> diff --git a/urcu.c b/urcu.c
> index f2aae34..a696439 100644
> --- a/urcu.c
> +++ b/urcu.c
> @@ -99,7 +99,8 @@ static void force_mb_single_thread(pthread_t tid)
>  	 * BUSY-LOOP.
>  	 */
>  	while (sig_done < 1)
> -		smp_rmb();	/* ensure we re-read sig-done */
> +		barrier();	/* ensure compiler re-reads sig-done */
> +				/* cache coherence guarantees CPU re-read. */

OK, this is where I think our points of view differ. Please refer to
http://lkml.org/lkml/2007/6/18/299.

Basically, cpu_relax() used in the Linux kernel has an
architecture-specific implementation which *could* include a smp_rmb()
if the architecture doesn't notice writes done by other CPUs. I think
Blackfin is the only architecture currently supported by the Linux
kernel which defines cpu_relax() as a smp_mb(), because it does not have
cache coherency.

Therefore, I propose that we create a memory barrier macro which is
defined as a 
  barrier()   when the cpu has cache coherency
  cache flush when the cpu does not have cache coherency and is
              compiled with smp support.

We could call that

  smp_wmc() (for memory-coherency or memory commit)
  smp_rmc()
  smp_mc()

It would be a good way to identify the location where data exchange
between memory and the local cache are is required in the algorithm.
What do you think ?

Mathieu

-- 
Mathieu Desnoyers
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68