[Team2013] [libvirt] Integrating LiSA into libvirt

[Radu Rendec]

On Wed, 2013-09-18 at 09:25 -0400, Laine Stump wrote:
> On 09/16/2013 06:32 AM, Radu Rendec wrote:
> > We've recently changed LiSA to be switching engine agnostic. Now it can
> > work with the "bridge" and "8021q" modules, which are part of the stock
> > linux kernel.
> 
> 1) What are the advantages of using LiSA vs. using the standard Linux
> host bridge without LiSA? vs. openvswitch?

The main advantage is the LiSA CLI, which provides a unified environment
for managing the virtual switch, very similar to a real (hardware)
switch. It may also be very useful in "no-such-usual" scenarios, for
instance when you want to configure a (8021q) tagged port for a VM NIC
or configure link aggregation for 2 VM NICs - I can explain this more
thoroughly, if needed.

Since LiSA is now a switching engine abstraction, it may provide support
in libvirt for other switching technologies, without having to integrate
each of them into libvirt directly.

Obviously there's no advantage in the switching process itself when the
bridge module is used.

> 2) Since LiSA can use the standard Linux host bridge, is it possible to
> use the LiSA userspace utilities on a bridge that was created in the
> traditional manner (i.e. one of the bridge's that libvirt creates
> today)? This may actually provide all the functionality that you're
> looking for.

Yes, theoretically it can. But (when used with bridge+8021q) LiSA
creates its own bridges and automatically sets them up as needed.

> (I can imagine though that, for example, if LiSA is able to support
> per-port vlan configuration, it might be nice to be able to translate
> that information from the interface (or network or portgroup) config
> into the proper LiSA command/ioctl/whatever to setup the vlan tag for a
> port.)

Perfect example! If you want to setup a port in tagged mode, and provide
access for vlans 2 and 3, for instance, all you have to do with LiSA is:
	switchport mode trunk
	switchport trunk allowed vlans 2,3

Assuming that the corresponding VM port is seen as tap0 in the host
machine, LiSA would automatically create tap0.2 and tap0.3 (8021q
devices) and add them to the appropriate bridges.

> 3) Since LiSA is apparently using standard tap devices for all its
> connections, if libvirt did add any sort of support for it, I think it
> should be in a form parallel to the openvswitch/802.1Qb[gh] support,
> which continues to use <interface type='bridge'> (or <interface
> type='network'>) with switching technology-specific details in the
> <virtualport type='<openvswitch|8021Qbg|8021Qbh)'> sub-element of the
> <interface> or <network>.

Honestly, I don't think I know enough about libvirt internals to have an
opinion about this.

However, if we get to the point where we implement any support for LiSA
in libvirt, we'll discuss any technical details on the list before
actually implementing anything.

> > I completely understand your point, but maybe you'll reconsider now that
> > you know an out-of-tree patch is not actually required.
> 
> Purely from the point of view of "let's see if we've abstracted the
> networking well enough to cleanly support a new technology" this would
> be interesting. But I am in agreement with Dan that we need to be sure
> that it is something that is "standards track", has good momentum, has a
> stable ABI, and offers a definite gain in functionality and/or user-base
> prior to committing to supporting it in libvirt - once something goes
> into the libvirt API, we guarantee that it will continue to be supported
> effectively "forever", so we can't make the decision too lightly.

This is like a chicken and egg problem. LiSA (or any project, after all)
cannot have a large user base unless it's known to the public and people
can easily experiment with it. I cannot go out and say "hey, here's this
new thing you can play with, but you need to apply this out-of-tree
patch, recompile libvirt, mess up your whole environment and spend 3
days wondering why it doesn't work out of the box". Nobody would even
try that.

On the other hand, I understand your point about not rushing into
integrating new technology. I can guarantee a stable ABI and long-term
support, but I don't expect you (or anybody else) to take this for
granted.

Perhaps a good approach is to adopt LiSA as "experimental technology"
and distribute it accordingly - for instance in "playground" distros
like Fedora. I'm really not sure what can be done (or not), so some good
advice in this direction would be highly appreciated. 

> > As for our out-of-tree module, we actually tried to push it upstream a
> > couple of years ago, but the kernel networking team said that it didn't
> > make sense to add a new module with similar functionality to already
> > existing modules and that we should have fixed those instead (if
> > something had been wrong with them).
> >
> > That was even before OpenVSwitch was released as open source. I'm
> > actually disappointed about this. LiSA was out in 2005 - that's 4 years
> > before OpenVSwitch, but nobody cared.
> 
> Yes, unfortunately marketing is very often just as important as (or even
> more important than) the quality of the code.

I agree. After all, size does matter :) Meaning that OpenVSwitch was
(is?) developped/supported by Citrix, and the LiSA team is just a small
team of programmers.