From: Keir Fraser <keir@xensource.com>
Subject: Re: Re: [patch 21/26] Xen-paravirt_ops: Use the hvc
	console infrastructure for Xen console
Date: Fri, 16 Mar 2007 10:31:49 +0000
Message-ID: <C2202295.BA49%keir@xensource.com>
References: <20070316095851.GB301@elte.hu>
Mime-Version: 1.0
Content-Type: text/plain;
	charset="US-ASCII"
Content-Transfer-Encoding: 7bit
In-Reply-To: <20070316095851.GB301@elte.hu>
Sender: xen-devel-bounces@lists.xensource.com
Errors-To: xen-devel-bounces@lists.xensource.com
To: Ingo Molnar <mingo@elte.hu>, Keir Fraser <Keir.Fraser@xensource.com>
Cc: xen-devel@lists.xensource.com, Andi Kleen <ak@muc.de>, linux-kernel@vger.kernel.org, Chris Wright <chrisw@sous-sol.org>, virtualization@lists.osdl.org, Andrew Morton <akpm@linux-foundation.org>

On 16/3/07 09:58, "Ingo Molnar" <mingo@elte.hu> wrote:

>> These mb()'s are pretty standard for lock-free producer/consumer
>> rings. Write descriptor /then/ write the updated producer. Read the
>> producer /then/ read any descriptors revealed by this new producer
>> value.
> 
> then use rmb()/wmb(). Rarely does a ring protocol truly need mb().

It's needed for writing data /after/ reading the consumer index that shows
you have space to write. Looking through xenbus_comms.c I think all the
barriers are correct except there is a spurious extra mb() in xb_read(),
where there is a later rmb() which is sufficient by itself. All the others
have a purpose.

 -- Keir