Open Stack

Hi, 
On Live Migration, I want to update the domain.xml with the required
network interface definition for the target node. For code, please see
the prototype [2]

I'm reaching our for feedback about the right way to implement this!


Use Cases/Problem:
=================
#1 Live Migration with Neutron Macvtap agent (see bug [1])
#2 Live Migration cross compute nodes that run different l2 agents for
agent transitioning.

More details on the use cases see further below.

Today, both is not working! 
Reason: To get the interface information for the target node, nova needs
to query Neutron. But Neutron can return this information only for the
case, when the binding:host_id attribute of the corresponding port is
set to the target host. This update happens during live migration
process, but it happens too late - in post_live_migration. I need this
kind of information in pre_live_migration!


Proposal
========
Update the portbinding in pre livemigration instead of in
post_live_migration. Then the virt driver can just query the ports for
the instance to receive the network interface  information for the
target node and update the migration xml accordingly.

I posted a working prototype here [2]

Open Questions: 
* Is there a reason for doing the portbinding after migration? Some
races or so?
* Where to do the cleanup on a failed migration (set the binding:host_id
to the migration source again)?
* Status of the port during migraton. binding.host_id = dest_host
  * ovs-hybrid plug: Status = Active
  * ovs-non-hybrid, lb, macvtap plug: Status = BUILD  (Goes to active as
soon as libvirt created the instance container and with it the network
device on the target. Is this a problem?


Alternatives
============
* Let Neutron implement a new API to request the port details for a
certain migration target
  * without changing it
  * storing it internally as additional in port.migration_port or so
* Allow portbinding to 2 hosts in parallel (rkukura proposed a patchset
some time ago - need to contact him)



More details on use Cases/Problems:
=================================
#1 For correct live migration with the Neutron Macvtap agent, the same
physical_interface_mapping must be deployed on each node. If one node
wants to use another mapping, migration fails or the instances is
migrated into a wrong network. This happens, as the for macvtap, the
interface name to place the macvtap upon is hard coded into the
domain.xml. For proper migration, I must be able to specifiy the
interface that is used on the target side! More details see [1]

#2 Live Migration cross compute nodes that run different l2 agents.
E.g. one node runs the ovs agent, another one runs the lb agent. I want
to be able to live migrate between those 2 nodes. This could be
interesting as transition strategy from one l2 agent to another one
without the need of shutting down an instance. (Assuming the ML2 Neutron
plugin is being used)

Any feedback is welcome! 
Thank you!




[1] https://bugs.launchpad.net/neutron/+bug/1550400
[2] https://review.openstack.org/#/c/297100/



-- 
-----
Andreas (IRC: scheuran)