OK, it appears that we are even.  I forgot to attach the promised
analysis of the callers to __kfifo_put() and __kfifo_get(), and
the open-iscsi@googlegroups.com email address listed as maintainer
in drivers/scsi/libiscsi.c bounces complaining that, as a non-member,
I am not allowed to send it email.  ;-)

Anyway, this time the analysis really is attached, sorry for my confusion!

Could someone please let the guys on open-iscsi@googlegroups.com know
that they should take a look at this?

						Thanx, Paul

On Wed, Aug 09, 2006 at 05:18:23PM -0700, Paul E. McKenney wrote:
> Hello!
> 
> Both __kfifo_put() and __kfifo_get() have header comments stating
> that if there is but one concurrent reader and one concurrent writer,
> locking is not necessary.  This is almost the case, but a couple of
> memory barriers are needed.  Another option would be to change the
> header comments to remove the bit about locking not being needed, and
> to change the those callers who currently don't use locking to add
> the required locking.  The attachment analyzes this approach, but the
> patch below seems simpler.
> 
> Signed-off-by: Paul E. McKenney <paulmck@us.ibm.com>
> ---
> 
>  kfifo.c |    4 ++++
>  1 files changed, 4 insertions(+)
> 
> diff -urpNa -X dontdiff linux-2.6.18-rc2/kernel/kfifo.c linux-2.6.18-rc2-kfifo/kernel/kfifo.c
> --- linux-2.6.18-rc2/kernel/kfifo.c	2006-07-15 14:53:08.000000000 -0700
> +++ linux-2.6.18-rc2-kfifo/kernel/kfifo.c	2006-08-09 14:01:53.000000000 -0700
> @@ -129,6 +129,8 @@ unsigned int __kfifo_put(struct kfifo *f
>  	/* then put the rest (if any) at the beginning of the buffer */
>  	memcpy(fifo->buffer, buffer + l, len - l);
>  
> +	smp_wmb();
> +
>  	fifo->in += len;
>  
>  	return len;
> @@ -161,6 +163,8 @@ unsigned int __kfifo_get(struct kfifo *f
>  	/* then get the rest (if any) from the beginning of the buffer */
>  	memcpy(buffer + l, fifo->buffer, len - l);
>  
> +	smp_mb();
> +
>  	fifo->out += len;
>  
>  	return len;