From mboxrd@z Thu Jan 1 00:00:00 1970 From: Vlad Yasevich Subject: Re: [RFC v2 PATCH 0/2] Add RCU locking to SCTP address management Date: Wed, 12 Sep 2007 16:59:48 -0400 Message-ID: <46E85344.1030402@hp.com> References: <11896263983281-git-send-email-vladislav.yasevich@hp.com> Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1 Content-Transfer-Encoding: 7bit Cc: lksctp-developers@lists.sourceforge.net To: netdev@vger.kernel.org Return-path: Received: from atlrel7.hp.com ([156.153.255.213]:46461 "EHLO atlrel7.hp.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1761568AbXILVBT (ORCPT ); Wed, 12 Sep 2007 17:01:19 -0400 In-Reply-To: <11896263983281-git-send-email-vladislav.yasevich@hp.com> Sender: netdev-owner@vger.kernel.org List-Id: netdev.vger.kernel.org Vlad Yasevich wrote: > Ok, this is version 2 of the patch that incorporates comments from > Sridhar Samudrala and Paul McKenney. > > The changes icorporated are: > 1. Add locking around the modification of the global sctp_local_addr_list > when processing the notifiers. After looking around, it is possible for > the IPv4 and IPv6 notifiers to be called at the same time, which means that > we need a spin lock. > > 2. After the Paul's explanation of why writers would would to call > rcu_read_lock, it's apparent that we really don't need that in our usage. > I've removed all that I could find and conser safe. > > 3. I took Paul's suggestiong of passing an explicit rcu callback when > removing entries from the list since these can be done it different > contexts. This made the removal code rather simple. Paul Moore just pointed out that using rcu versions of list traversal macros inside the writer protected sections is just plain silly, so consider that another change. I'll update patch 2/2 with that comment. -vlad > > Things I've left behind: > 1. The valid flag remains. After discussing the virtues with Paul Moore > (who used the same functionality in Netlabel code), I think that the > valid flag slightly reduces the possibility that the reader will use > an entry that's about to be removed. It's a good thing in our case. > It doesn't really harm anything if a reader used a !valid entry, but > I'd like to reduce that chance. > > I would appreciate any further comments