[libvirt] [ 0/5] netdev ethernet allow to set ip, route and peer address

[Laine Stump]

On 04/07/2016 01:35 PM, Daniel P. Berrange wrote:
> On Mon, Apr 04, 2016 at 09:00:01PM +0000, Vasiliy Tolstov wrote:
>> Some minor improvements and patch split as suggested by Laine Stump
> FYI, make sure you include the word PATCH in mails, so they get
> picked up by out patch tracking too, otherwise they could get
> left without being noticed.
>
>
> Regards,
> Daniel

Something I've found myself worrying about lately while driving in the 
car or nodding off to sleep - are the "address" and "peer" attributes 
effectively used in the same way for all network connection types and 
both hypervisors? I think the answer may be "no", and if so we need to 
fix that before they go out in a release.

In particular, when an lxc domain's interface has:

    <ip address='192.168.128.1'/>

That is the IP address seen by the guest, not the host. So I would 
assume that if an LXC domain had:


    <ip address='192.168.128.1' peer='192.168.128.2'/>

that 192.168.128.1 would still be the IP address see by the guest, and 
192.168.128.2 would be the IP address on the host side; and it should be 
the same for qemu.

 From what I can see of the code, though, on a qemu domain, the IP 
address is set for the tap device's own IP, meaning that it would show 
up on the *host* side, while the peer address would be what the host 
expects to be at the other end of the tap device (i.e. the guest side), 
so the two attributes are used for the *opposite* end of the PTP link in 
lxc vs. qemu.

I think that, instead, the "address" attribute should *always* be the IP 
address that is seen/used by the guest, and the "peer" attribute should 
be the IP address that is seen/used by the host. (perhaps "peer" could 
be replaced with some other name, like "host" or "hostAddress" to avoid 
confusion? (don't like either of those alternatives, but I don't really 
like peer either)).

Aside from that, I can see that these patches have been pushed in the 
code that I'm running, and I've been trying to add "peer='blah'" to 
interface IP addresses on my test machine, but it's just removed from 
the config. Have you tested what got pushed? Has something gone wrong?

Since there hasn't been a release with these patches included yet, there 
is still time to fix it at least to be consistent (assuming that my 
suspicions are correct; I've been unable to test it myself for the 
reason above).