Re: [PATCH 2/2 v2] net: Remove VRF change to udp_sendmsg

[David Ahern <dsa@cumulusnetworks.com>]

On 9/9/15 6:51 PM, Tom Herbert wrote:
> It is NAT since you are changing the source address and modifying the
> transport protocol checksum below IP and transport layer. There are a
> bunch of side effects that you would need to consider. This is
> creating custom APIs changing the semantics of address selection, and
> also creates inconsistency between how addresses may be selected
> between a connected and unconnected sockets. Consider that
> ip_local_out_sk calls netfilter NF_INET_LOCAL_OUT hook before
> dst->output, so then netfilter would start seeing packets with zero
> source address???

understood.

>
> A lot of design in the stack is predicated on inet_select_addr
> returning the source address to use for sending a packet. This should
> always return a reasonable address as an invariant, if someone wishes
> to rewrite addresses at a lower layer that's fine, but that should be
> defined as a NAT operation. If a device wants to weigh in on address
> selection then we can define an ndo function for that as I mentioned
> before.

I am floating an idea internally that re-implements how VRF impacts the 
stack. It's 4.4 material and essentially adds dev_xxxx() / ndo functions 
for the intrusions. With net-next closed no since throwing them out yet 
and Nikolay always has good comments on my wild ass ideas.

David