From: Stephen Hemminger Subject: Re: [PATCH net] vxlan: revert per-vxlan port Date: Thu, 23 May 2013 10:08:54 -0700 Message-ID: <20130523100854.6c9749d6@nehalam.linuxnetplumber.net> References: <20130520103017.054ae605@nehalam.linuxnetplumber.net> <20130520113000.0057ce90@nehalam.linuxnetplumber.net> <20130522.150830.1305371698781503084.davem@davemloft.net> <20130522173954.7f75f62d@nehalam.linuxnetplumber.net> Mime-Version: 1.0 Content-Type: text/plain; charset=US-ASCII Content-Transfer-Encoding: 7bit Cc: David Miller , netdev@vger.kernel.org, netdev-owner@vger.kernel.org To: David Stevens Return-path: Received: from mail-pa0-f43.google.com ([209.85.220.43]:32881 "EHLO mail-pa0-f43.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1757953Ab3EWRI7 (ORCPT ); Thu, 23 May 2013 13:08:59 -0400 Received: by mail-pa0-f43.google.com with SMTP id hz11so486440pad.16 for ; Thu, 23 May 2013 10:08:59 -0700 (PDT) In-Reply-To: Sender: netdev-owner@vger.kernel.org List-ID: On Wed, 22 May 2013 22:14:39 -0400 David Stevens wrote: > Stephen Hemminger wrote on 05/22/2013 > 08:39:54 PM: > > > The per-vxlan device stuff is the issue. I don't think it works > > as is in 3.10 because there is no real way to use it to receive. > > And don't want to let it out broken. By reverting that part, we avoid > > raising false expectations. > > > > Of course you can use it to receive; the remote hosts can be > listening on a different port and this is the only mechanism > to send to those if there are not fdb entries on the local machine. > > So, for example, host A: > binds to 1234 > has no fdb entries, but sets the default remote port to 4789 > > host B > binds to 4789 > has no fdb entries and sets the default remote port to 1234 > > ...or you have an agent listening on a multicast address with port 5429 > and the agent's job is to dynamically populate the fdb tables of any > hosts that don't have matching entries. All the VXLAN instances bind > to, say 4789, and the agent binds to port 5429. The agent receives a > packet, and adds an entry to the fdb table on the sending host where > the destination is the actual VXLAN receiver with port 4789. But the > default port on all of them for this kind of set-up is correctly the > different 5429, not 4789. > > These configurations have nothing to do with multiple listen ports, > and reverting the patch prohibits an admin from doing it at all. > In the first set up, you can link two segments on different ports > without knowing any destination MAC addresses. In the second, you > can separate traffic not matching any destination MAC from traffic > with a known destination, and use that to fill forwarding tables. > > I know these configurations don't match your sense of how VXLAN > should be used, but these are not "broken", do receives perfectly fine, > and have no dependency at all on multiple listen ports. I don't know > what "false expectations" you think are there-- I think anyone > changing the default destination port to a different value from > the binding port simply expects that transmits using the default > will go to one port and received packets will go to another. Deciding > in advance that that is never appropriate is short-sighted; that > decision belongs with the admin setting it up. Reverting the patch > removes that choice. > > > +-DLS > Let's go back to basic use cases. User wants to setup a vxlan between two hosts with custom port On both hosts, he/she does: # ip li add vxlan0 type vxlan id 42 group 239.1.1.1 dev eth0 dstport 4789 With net-next it works, with 3.10 it doesn't. Sure it works if they have your agent, or manually configure N-M static FDB entries, but that is corner case. No matter if Dave decides to the revert the kernel part or not, I am pulling the option from 3.10 version of iproute to avoid user confusion.