[libvirt] [PATCH] lxc: support <interface type='ethernet'>

[Laine Stump]

On 05/25/2016 11:13 AM, Vasiliy Tolstov wrote:
> 2016-05-25 18:00 GMT+03:00 Laine Stump <laine at laine.org>:
>> So $VM_GW is on the same subnet as $VM_IP/24 ? Is that also the same subnet
>> as $HOST_IP? Or is that on a completely different network?
>>
>> (This is all very useful, because it's pointing out that the config of the
>> two ends definitely aren't mirror images, so we really do need independent
>> settings for both).
>>
> VM_GW on the same subnet as PEER , but on host side i apply /32
> address, but inside vm /24.
>
>>> peer only used on host machine to determine on which interface kernel
>>> needs to send traffic
>>
>> Well, you *could* do the same thing with a /30 subnet for each tap (so that
>> your routing daemon would get a bunch of /30 routes pointing to your host
>> for all of them), but that would use up a lot more address space.
>>
>>>> (I'm not sure what the guest does with 169.254.169.254)
>>>>
>>>> What if you instead set the host to:
>>>>
>>>>           $HOST_IP peer $VM_IP/32
>>>>
>>>> and set the guest to:
>>>>
>>>>           $VM_IP peer $HOST_IP/32
>>>>           route add default $HOST_IP
>>>>
>>>> ??
>>>>
>>> I don't use this scheme... If i set $VM_IP peer $HOST_IP/32 on host
>>> machine i think kernel recieves packet and drop it =)
>>
>> No - you would set that on the *guest*, not the host. I setup an LXC
>> container in this manner and everything seemed to work properly.
>>
> May be, but in case of qemu i use simple dhcp server and on vm simple
> dhcp client. It can't assign peer addresses

Ah, right. I haven't looked, but doubt there is a dhcp option to specify 
a peer address.


> may be this is works. I
> don't try.
> In my setup guest vm does not know host side of network, so it does
> not need know about peer... Also i can live migrate guest without
> changing....
> (Changes only host side of network)
>
>
>>>> Anyway, it's important to know if you set the IP config on host and guest
>>>> to
>>>> exact mirrors of each other. It seems like the answer is "no", though, so
>>>> i'm going to make a patch that allows what I was talking about last week:
>>>>
>>>>
>>>>          <interface type='ethernet'>
>>>>             <source>
>>>>               <ip address='HOST_IP' family='ipv4' peer='VM_IP'
>>>> prefix='32'/>
>>>>               <ip address='VM_GW' family='ipv4' prefix='32'/>
>>>>             </source>
>>>>             <ip address='VM_IP' family='ipv4' peer='HOST_IP' prefix='24'/>
>>>>             <route family='ipv4' address='0.0.0.0' gateway='HOST_IP'/>
>>>>             ...
>>>>         </interface>
>>>>
>>>> On qemu only the address info inside <source> would be used, since we
>>>> don't
>>>> have control over the guest's network config. On LXC, we can set both.
>>>>
>>>> Does that sound usable?
>>> Yes, but don't cleanup please  <ip address='VM_IP' family='ipv4'
>>> peer='HOST_IP' prefix='24'/> in qemu case, because on somedays we can
>>> launch dnsmasq or orhet software that can provide dhcp and use this
>>> element..
>>
>> An interesting point - even if libvirt itself doesn't launch the dnsmasq, it
>> could be done by a more intelligent script (one that received the full
>> interface XML on stdin, as I discussed in my last mail), so we should
>> probably just ignore it rather than complaining that it's unsupported...
>>
>> Thanks for the info!
> Yes, i handmade my own dhcp server on golang what listens libvirt
> hooks and serves needed addresses for needed tap devices.

Yeah, I had considered that if we were to support such a thing in 
libvirt, it would be best to do it by having a single dnsmasq instance 
for all tap-only guest interfaces.

> So i have only one daemon that have stateless config (it get it via
> libvirt xml). So i don't need database or something else.

How does it gather than config? With a qemu hook script?