Open Stack

On 03/23/2018 07:05 AM, Ramon Orru wrote:
> Hi Brian,
> 
> yes, you're right, that's the point, libvirt is sending gARPs too 
> early.... am I misconfiguring something? Seems to me a really strange 
> behaviour...

Hi Ramon,

Sorry, don't know much about the libvirt side of things, maybe someone 
more familiar with the Nova side of the migration would know?

-Brian

> Il 21/03/18 22:35, Brian Haley ha scritto:
>> On 03/20/2018 12:40 PM, Ramon Orru wrote:
>>> Hello everybody, I'm running a fresh queens cluster. I'm using 
>>> bridges to support networking. I'm facing an issue when an instance 
>>> is live migrated.
>>> Suppose we have an instance running with an interface on vlan XXX, 
>>> and we want to migrate it to compute host YYY. We'll call that 
>>> instance ZZZ.
>>> If no other instance is already running on YYY using vlan XXX, no 
>>> bridge called 'br-vlan.XXX at br-vlan' exists yet on YYY.
>>> Now, if I migrate ZZZ on YYY host, a new bridge 'br-vlan.XXX at br-vlan' 
>>> will be created.
>>> During the migration process, ZZZ become unreachable while interfaces 
>>> are going up on YYY (from 10 seconds to about 2 minutes).
>>> After some troubleshooting, we spotted the problem: bridge 
>>> 'br-vlan.XXX at br-vlan' is being created after gratuitous ARP packets 
>>> are sent from migrating machines to advise other devices about new 
>>> position.
>>> The result is: no other device can reach the fresh migrated machine 
>>> until ARP table becomes stale.
>>> This does not happen when an instance with an interface on same vlan 
>>> is already runnning on destination host ('br-vlan.XXX at br-vlan' is 
>>> already up and running, and ARPs can be sent flawlessly).
>>>
>>> Any idea of how to get rid of this? I think it's very unlikely that 
>>> I'm the first to face this problem, but i didn't manage to find 
>>> additional info on this strange behaviour.
>>> Thanks in advance.
>>
>> Just to clarify - are you talking about connectivity to the floating 
>> IP or just on the VLAN itself, i.e. the instance is directly connected 
>> to the VLAN.  I'm thinking it's the latter, which would mean it's 
>> libvirt(?) sending the gARPs before the bridge is up?
>>
>> -Brian
>