Hi Arnoud,

On 05/27/2010 11:14 AM, Arnaud Ebalard wrote:
quoted text
> Hi,
> 
> Thanks for your reply Brian and sorry for the length of this response. If
> Hideaki and David can comment on the IPv6/XFRM and SO_BINDTODEVICE
> aspects discussed below that would be helpful, IMHO.

Thanks for all the analysis, comments below.

quoted text
>> On 05/26/2010 01:01 PM, Arnaud Ebalard wrote:
>>> Hi,
>>>
>>> I just updated my laptop's kernel to 2.6.34 (previously running .33 and
>>> configured to act as an IPsec/IKE-protected MIPv6 Mobile Node using
>>> racoon and umip): after rebooting on the new kernel, the transport mode
>>> SA protecting MIPv6 signaling traffic are missing.
>>>
>>> I bisected the issue down to f4f914b58019f0e50d521bbbadfaee260d766f95
>>> (net: ipv6 bind to device issue) which was added after 2.6.34-rc5: 
>>>
>>> diff --git a/net/ipv6/route.c b/net/ipv6/route.c
>>> index c2438e8..05ebd78 100644
>>> --- a/net/ipv6/route.c
>>> +++ b/net/ipv6/route.c
>>> @@ -815,7 +815,7 @@ struct dst_entry * ip6_route_output(struct net *net, struct sock *sk,
>>>  {
>>>         int flags = 0;
>>>  
>>> -       if (rt6_need_strict(&fl->fl6_dst))
>>> +       if (fl->oif || rt6_need_strict(&fl->fl6_dst))
>>>                 flags |= RT6_LOOKUP_F_IFACE;
>>
>> Can you see if fl->oif is at least a sane value here?  Maybe there's some
>> partially un-initialized flowi getting passed-in, a quick source code check
>> didn't find anything obvious.
> 
> When it's not 0, fl->oif is a sane value: it is set to the index of the
> interface on which the current *Care-of Address* is configured. All the
> traffic is expected to leave the host via this interface. 

Ok, so it's not un-initialized data causing this.

quoted text
> In previous debug outputs, the content of the fl->oif is ok, i.e. it is
> set to the interface on which the CoA is configured, i.e. the output
> interface. But the commit results in flags |= RT6_LOOKUP_F_IFACE.
> Later, in rt6_score_route(), the call to rt6_check_dev() returns 0
> (dev->ifindex is ip6tnl1 but oif is wlan0). Because of the change to flags 
> flags, we quickly return -1 in rt6_score_route():

Ok, so the call to ip6_route_output() was from the tunnel code, which is
using it's cached flowi, which has oif set to the tunnel.  The XFRM code
swaps the addresses, which should invalidate the oif, but it doesn't.

quoted text
> static int rt6_score_route(struct rt6_info *rt, int oif,
> 			   int strict)
> {
> 	int m, n;
> 
> 	m = rt6_check_dev(rt, oif);
> 	if (!m && (strict & RT6_LOOKUP_F_IFACE))
>                 return -1;
>         ...
> 
> Now, I wonder if the following is correct. Don't hesitate to correct me
> if I am wrong:
> 
> Initially (before f4f914b58019f0), the purpose of the test using
> rt6_need_strict() in ip6_route_output() (introduced by c71099ac) was to
> allow the multiple routing table logic to be applied to all global
> addresses but to preserve the addresses for which it would not make
> sense (link-local, multicast, ). The change introduced by f4f914b58019f0
> basically reduces the ability to route traffic as you want and forces
> the traffic to leave the device by the interface on which it is
> configured (if fl->oif is set).

The problem is we assumed the caller's would only set fl->oif if they
wanted it enforced (multicast, link-local, SO_BINDTODEVICE), but it
didn't take into account the tunnel code.  I guess the easy answer
would be to revert this until we can fix it correctly.

quoted text
> From my (very limited and possibly wrong) understanding, the change
> introduced by f4f914b58019f0 looks like a workaround for the 
> SO_BINDTODEVICE issue. Looking at the code, there is something I don't
> understand: if SO_BINDTODEVICE has been used on a socket, the socket
> should have its sk_bound_dev_if attribute set to the correct ifindex
> value. Hence the following (naive) question: why is that information not
> used to inflect the selection of the route cached for the socket? And
> why would the fix be at the adress level instead of being at the
> interface level (ifindex)?

I guess I always believed setting SO_BINDTODEVICE should always force
traffic out that interface, but from Yoshifuji's email it seems that
maybe wasn't the intention, at least for things that don't meet
the rt_need_strict() criteria like globals.  I don't know the history
behind the setsockopt.

The below might actually be what was actually intended, triggering
on what the user forced, rather than assuming all callers require
strict behavior.

-Brian


diff --git a/net/ipv6/route.c b/net/ipv6/route.c
index 294cbe8..252d761 100644
--- a/net/ipv6/route.c
+++ b/net/ipv6/route.c
@@ -814,7 +814,7 @@ struct dst_entry * ip6_route_output(struct net *net, struct sock *sk,
 {
 	int flags = 0;
 
-	if (fl->oif || rt6_need_strict(&fl->fl6_dst))
+	if ((sk && sk->sk_bound_dev_if) || rt6_need_strict(&fl->fl6_dst))
 		flags |= RT6_LOOKUP_F_IFACE;
 
 	if (!ipv6_addr_any(&fl->fl6_src))
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html