On Mon, 28 Jul 2014 16:39:23 -0700
Cong Wang [off-list ref] wrote:

On Mon, Jul 28, 2014 at 4:07 PM, Dmitry Popov [off-list ref] wrote:

quoted

--- a/include/net/ip_tunnels.h
+++ b/include/net/ip_tunnels.h

@@ -40,6 +40,7 @@ struct ip_tunnel_prl_entry {

 struct ip_tunnel_dst {
        struct dst_entry __rcu          *dst;
+       __be32                           saddr;
 };

It looks odd to save a source address in some struct named dst,
maybe it's the time to rename it.

Yes, maybe a bit odd, but ip_tunnel_flow/ip_tunnel_route doesn't look much 
better for me, so I decided to keep it with old name. David is ok with this
name too.

quoted

        if (dst) {
-               if (dst->obsolete && dst->ops->check(dst, cookie) == NULL) {
+               if (!dst->obsolete || dst->ops->check(dst, cookie)) {
+                       *saddr = idst->saddr;
+               } else {

I am wondering if this always works, that is, if saddr could be still valid
as long as dst is valid?

Right, this is not a very smooth place, but a) it should work with current ipv4
route implementation (it obsoletes all dst_entries on any route/addr change).
Yes, we shouldn't rely on implementation, but b) I don't know how (and why) this
implementation should be changed to break it. I think the assumption that
dst_entry is obsoleted when saddr is changed is pretty natural.
Please correct me if I'm wrong.