[ovs-dev] [PATCH 0/7] Enable NSH based Service Function Chaining support in OVS

Liu, Mengke mengke.liu at intel.com
Wed Oct 14 08:28:23 UTC 2015

Hi, Jesse,
Thanks for your comments on this series. Below are our comments and our modification plan for this patch:

  1.  Firstly, we agree that we should operate VxLAN GPE header and NSH header at different protocol layers, which will avoid lots of unnecessary conflations. Then in our modified high level design, we will decouple VxLAN/VxLAN-GPE tunneling header push/pop and NSH header push/pop. As the NSH header is handled as an independent layer, a new push/pop action “push_nsh/pop_nsh” will be added to support push and pop NSH header in both kernel datapath and userspace datapath. In this way, NSH header will not be part of tunnel header, it’s independent and the NSH header will be added/removed by NSH specific actions, for example:
Ovs-ofctl add-flow br-int “in_port=1, actions=set_nsi:1, set_nsp:12345, push_nsh, output:2” (port 2 is the VxLAN GPE port)
The “push_nsh” action is to add the NSH header and “output:2” is to add the tunnel header via the VxLAN-GPE port.
One question we want to discuss here is about how to implement the VxLAN-GPE port, we have two proposals:
1)      Add a new configuration option for current VxLAN vport type: “options:enable_GPE=true”, then the virtual port created with this option enabled will be VxLAN-GPE but not VxLAN, corresponding actions will be VxLAN-GPE specific.
2)      Add a new vport type “VxLAN-GPE” parallel to “VxLAN”.
Which proposal do you prefer to take?

  1.  As discussed above, the NSH layer will be an independent protocol layer but not part of VxLAN tunnel. Then for the lightweight tunneling support in kernel, changes we need to make are to add supports for VxLAN-GPE tunneling. Options here are to add a new vport type for VxLAN GPE or extend current VxLAN port to support VxLAN GPE.
Except for the VxLAN GPE support in lightweight tunneling, we’ll also add the NSH protocol support in the kernel space datapath code for Linux upstream.
BTW, can you give us some guidance regarding the model to upstream lightweight tunneling to kernel.

  1.  For supporting TLV, VxLAN GPE doesn’t have option TLVs in current definition. So for VxLAN GPE, we don’t need to implement the TLV option fields as Geneve.
VxLAN GPE header format:
    0                   1                   2                   3
    0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
   |R|R|Ver|I|P|R|O|       Reserved                |Next Protocol  |
   |                VXLAN Network Identifier (VNI) |   Reserved    |
          (No variable length option header for VxLAN-GPE)

But for NSH header, if the meta-type (MD-type) is 2 as below, we need the option TLV support:
        0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
       |Ver|O|C|R|R|R|R|R|R|   Length  |  MD-type=0x2  | Next Protocol |
       |          Service Path ID                      | Service Index |
       |                                                               |
       ~              Variable Length Context Headers  (opt.)          ~
       |                                                               |

So for NSH header, we need to add the TLV support. For the fixed fields of NSH header like NSI, NSP, we’d like to add specific meta-flow fields for them, for example:
MFF_NSP            Type: NXM_NX_NSP (105),         Length:4 bytes
MFF_NSI             Type: NXM_NX_NSI (106),          Length:1 bytes
But for the variable length context headers of NSH, we’d like to use fields like “tun_metadata” (tnl->metadata) to support it. We also have two options:
1)      Reuse the “tun_metadata” for NSH variable context header, it’s similar to current Geneve TLV support. But it’s a little wield because the NSH header is already an independent protocol layer but not belong to the tunnel layer.
2)      Define a new “nsh_metadata” fields for NSH variable context header.
Which one do you prefer? Please tell us for you inputs on our modification plan.

Best Regards

-----Original Message-----
From: Jesse Gross [mailto:jesse at nicira.com]
Sent: Thursday, October 1, 2015 5:24 AM
To: Liu, Mengke <mengke.liu at intel.com>
Cc: dev at openvswitch.org; Pritesh Kothari (pritkoth) <pritkoth at cisco.com>; Li, Ricky <ricky.li at intel.com>; paulq at cisco.com
Subject: Re: [ovs-dev] [PATCH 0/7] Enable NSH based Service Function Chaining support in OVS

On Wed, Sep 30, 2015 at 2:47 AM, mengke <mengke.liu at intel.com<mailto:mengke.liu at intel.com>> wrote:

> This patch adds support for NSH function implementation which can be

> used for service function chaining. It covers control plane, kernel

> dataplane and DPDK-netdev dataplane.

I have several high level comments on this series that I think it would be best to discuss before going through the code level details.

First and foremost, the kernel tunneling code that you are modifying is compatibility code. Upstream Linux is now using a different model of tunnels and any new features need to be added to that rather than to the old model as is done here. The new model is called "lightweight tunnels" so I would encourage you to look at what has been going on with that upstream.

That being said, I have significant concerns with the layering throughout this patchset and I suspect if that was resolved then at least some of the above changes would not be necessary. In particular, there is a great deal of conflation between VXLAN-GPE and NSH. As a theoretical matter, I believe that these are supposed to operate independently of each other at different protocol layers. And more practically, this seems to result in a lot of special "modes" to handle various specific cases.

It looks like the best way to handle this is by creating a push/pop mechanism for inserting and removing NSH headers over various protocols. As it is, I really don't like that there are a variety of specific operations such as for converting between VXLAN+NSH to

Ethernet+NSH - this should be decomposed. I believe that this

decomposition would also help address some other issues. For example, it's not OK to hard code the UDP port of VXLAN-GPE in the kernel - this needs to be configurable. However, the port for vanilla VXLAN is already configurable so if you can add support for GPE and then layer NSH on top then you don't need to introduce a new configuration mechanism.

The last thing that I wanted to mention is that I would strongly recommend looking at supporting TLVs now. There has been extensive work in supporting them for Geneve and not only should it now be much easier to do than it would have been a little while back, some of these public interfaces have not been part of a released version of OVS yet. Therefore, if there are changes that need to be made to make them support multiple protocols, now would be the time to do it before they can no longer be modified.

More information about the dev mailing list