From mboxrd@z Thu Jan 1 00:00:00 1970 From: Simon Horman Subject: Re: [PATCH v2.39 7/7] datapath: Add basic MPLS support to kernel Date: Sun, 22 Sep 2013 14:34:25 +0900 Message-ID: <20130922053156.GA4849@verge.net.au> References: <1378711207-29890-1-git-send-email-horms@verge.net.au> <1378711207-29890-8-git-send-email-horms@verge.net.au> <20130919155732.GC24919@verge.net.au> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: Ben Pfaff , Pravin Shelar , "dev@openvswitch.org" , netdev , Ravi K , Isaku Yamahata , Joe Stringer To: Jesse Gross Return-path: Received: from kirsty.vergenet.net ([202.4.237.240]:34923 "EHLO kirsty.vergenet.net" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752890Ab3IVLYH (ORCPT ); Sun, 22 Sep 2013 07:24:07 -0400 Content-Disposition: inline In-Reply-To: Sender: netdev-owner@vger.kernel.org List-ID: On Thu, Sep 19, 2013 at 12:21:33PM -0500, Jesse Gross wrote: > On Thu, Sep 19, 2013 at 10:57 AM, Simon Horman wrote: > > On Mon, Sep 16, 2013 at 03:38:21PM -0500, Jesse Gross wrote: > >> On Mon, Sep 9, 2013 at 12:20 AM, Simon Horman wrote: > >> > @@ -616,6 +736,13 @@ int ovs_execute_actions(struct datapath *dp, struct sk_buff *skb) > >> > goto out_loop; > >> > } > >> > > >> > + /* Needed to initialise inner protocol on kernels older > >> > + * than v3.11 where skb->inner_protocol is not present > >> > + * and compatibility code uses the OVS_CB(skb) to store > >> > + * the inner protocol. > >> > + */ > >> > + ovs_skb_set_inner_protocol(skb, skb->protocol); > >> > >> The comment makes it sound like this code should just be deleted when > >> upstreaming. However, I believe that we still need to initialize this > >> field, right? Is this the best place do it or should it be conditional > >> on adding a first MPLS header? (i.e. what happens if inner_protocol is > >> already set and the packet simply passes through OVS?) > > > > I believe there are several problems here. > > > > The first one, which my comment was written around is that I think that if > > inner_protocol is a field of struct sk_buff then we can rely on it already > > being initialised. However, if we are using compatibility code, where > > inner_protcol is called in the callback field of struct sk_buff then I > > think that OVS needs to initialise it. > > I'm not sure that it's true that inner_protocol is already initialized > - I grepped the tree and the only assignment that I found is in > skbuff.c in __copy_skb_header(). My assumption was that it would be initialised to zero, primarily due to the behaviour of __alloc_skb_head(). Perhaps the core code should be fixed to make my assumption true? > > A second problem is one that you raise which I had not considered > > which is how to handle things if inner_protocol is already set. > > > > I believe this should only occur in the case where inner_protocol > > is a field of struct sk_buff and I think it would be most convenient > > to set it conditionally in ovs_skb_reset_inner_protocol(). > > I think that if it is not set it should be zero but it should be > > safe to check for values less than ETH_P_802_3_MIN. > > It's probably OK to check for values less than ETH_P_802_3_MIN but I'm > not sure that it's the most correct thing to do since skb->protocol > could contain these values (such as ETH_P_802_2). It's unlikely that > they will be GSO packets but it seems better to use the more strict > check against zero. Sure, a strict check against zero is fine my me. > One other consideration in the OVS case - with recirculation we may > hit this code multiple times and the difference in behavior could be > surprising. However, on the other hand, we need to be careful because > skb->cb is not guaranteed to be initialized to zero. Thanks, that is also not something that I had considered. I'm not sure, but I think that we can rely on skb->cb not being clobbered between rounds of recirculation. Or at the very least I think we could save and restore it as necessary. So I think if we could be careful to make sure that inner_protocol is in a sane state the first time we see the skb but not each time it is recirculated then I think things should work out. In my current implementation of recirculation the datapath side is driven ovs_dp_process_received_packet(). So by my reasoning above I think it would make sense to reset the inner_protocol there and in ovs_packet_cmd_execute() rather than in ovs_execute_actions() which each of those functions call.