Re: [Patch net-next v5 0/5] vxlan: add ipv6 support

[Cong Wang <amwang@redhat.com>]

On Mon, 2013-04-22 at 16:08 -0400, David Miller wrote:
> This is broken.  Every time I see someone export new things from IPV6
> and then try to use those symbols in some other unrelated module, it
> is a huge red flag.
> 
> You can't call into IPV6 protected symbols unless VXLAN and IPV6 are
> configured identically.
> 
> So with your changes, with VXLAN=y and IPV6=m, you'll get link errors.
> I could see this just by looking at your patch, I didn't have to even
> try to build it.

Yeah, the IPv6 multicast API's we export indeed have such problem.

> 
> Please do not fix this by adding Kconfig dependencies, you have to
> find another way.  In bonding and bridging, we've made it such that
> you can configure them in any combination whatsoever with ipv6 and
> everything works properly.  Most of them time this can be accomplished
> by moving things into the explicit "obj-y" objects in
> net/ipv6/Makefile

One quick solution is just linking mcast.o statically, because it is not
easy to separate core functions from mcast.c like addrconf_core.c.

> 
> If you are adding stateful dependencies upon ipv6 (you want to inspect
> the ipv6 routes or something like that), I'm sorry but I really don't
> want any hard dependies on ipv6's internal state, we can't export that
> properly.

I don't think we need that.

Thanks!