Re: [PATCH ipsec 1/3] ipv6: wire up skb->encapsulation

[Simon Horman <horms@verge.net.au>]

On Sun, Aug 18, 2013 at 01:00:38PM +0200, Hannes Frederic Sowa wrote:
> On Sun, Aug 18, 2013 at 02:24:16PM +1000, Simon Horman wrote:
> > On Sat, Aug 17, 2013 at 08:07:38PM +0200, Hannes Frederic Sowa wrote:
> > > Hi Simon!
> > > 
> > > On Sat, Aug 17, 2013 at 07:51:16PM +0200, Hannes Frederic Sowa wrote:
> > > > When pushing a new header before current one call skb_reset_inner_headers
> > > > to record the position of the inner headers in the various ipv6 tunnel
> > > > protocols.
> > > > 
> > > > We later need this to correctly identify the addresses needed to send
> > > > back an error in the xfrm layer.
> > > > 
> > > > This change is safe, because skb->protocol is always checked before
> > > > dereferencing data from the inner protocol.
> > > > 
> > > > Cc: Steffen Klassert <steffen.klassert@secunet.com>
> > > > Cc: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
> > > > Cc: Nicolas Dichtel <nicolas.dichtel@6wind.com>
> > > > Signed-off-by: Hannes Frederic Sowa <hannes@stressinduktion.org>
> > > > ---
> > > > 
> > > > This patch is based on Steffen Klassert's ipsec tree.
> > > > 
> > > >  net/ipv6/ip6_gre.c    | 5 +++++
> > > >  net/ipv6/ip6_tunnel.c | 6 ++++++
> > > >  net/ipv6/sit.c        | 5 +++++
> > > >  3 files changed, 16 insertions(+)
> > > > 
> > > > diff --git a/net/ipv6/ip6_gre.c b/net/ipv6/ip6_gre.c
> > > > index ecd6073..90747f1 100644
> > > > --- a/net/ipv6/ip6_gre.c
> > > > +++ b/net/ipv6/ip6_gre.c
> > > > @@ -724,6 +724,11 @@ static netdev_tx_t ip6gre_xmit2(struct sk_buff *skb,
> > > >  		ipv6_push_nfrag_opts(skb, &opt.ops, &proto, NULL);
> > > >  	}
> > > >  
> > > > +	if (likely(!skb->encapsulation)) {
> > > > +		skb_reset_inner_headers(skb);
> > > > +		skb->encapsulation = 1;
> > > > +	}
> > > > +
> > > 
> > > While doing these patches, I wondered how skb->inner_protocol will be
> > > used in future (you added it in 0d89d2035fe063461a5ddb609b2c12e7fb006e44
> > > ("MPLS: Add limited GSO support")). Current use by tunnels seems safe to
> > > me, but I wondered how you would extend its use?
> > 
> > Hi,
> > 
> > I must confess that I'm not entirely sure that I understand the question.
> 
> Sorry to be not clear enough. I try to present my envisioned use case for
> inner_protocol:
> 
> I do think there may be some other corner cases when reporting back
> errors to a local socket when multiple tunnels+ipsecs are stacked into
> each other. The first encapsulation sets skb->inner_network_header, so
> that we can later extract the needed information for the error generation
> via inner_ip{v6,}_hdr. It would be handy here to use inner_protocol to
> mark the type of inner_network_header in advance (all this is because,
> even if we still have a reference to the local socket and know its af,
> it could also have emitted an IPv4 frame, so we have to jump to the ipv4
> error handling path).

That sounds reasonable to me and I don't think it would interfere
with the current usage or vice versa.