nova-manage network delete fails with QuantumManager

[Isaku Yamahata]

With QuantumManager network driver and linux_net LinuxOVSInterfaceDriver interface driver
- create network
- create an instance and terminate
- then nova-manage network delete fails.
And the db state remains inconsistent state. It requires manual fix by issuing SQL.

The reason is
- LinuxOVSInterfaceDriver implements unplug as nop.
- It creates gw-xxx and register the port to Quantum at the time first instance is created
- When deleting network, the gw-xxx reamins. So network deletion request to quantum fails
  due to the interface gw-xxx is attached and busy.

The fix would be
- implement unplug and call unattach and delete the port request to quantum
- stop dnsmasq daemon

[Salvatore Orlando (salvatore-orlando)]

Hi Isaku!

Thanks for reporting this bugs, and for proposing a strategy for fixing it.

I'm not really up-to-date on the Quantum-nova integration, but I was wondering why we would need to stop dnsmasq. This might make sense with the VLAN manager, altough I think there is already code in place for destroying resources on the network node, including the dnsmasq instance, when a network is delete.

However, in FlatDHCP mode, if I'm not wrong the same dnsmasq instance is used for all networks (several IP addresses are configured on the bridge).

[dan wendlandt (danwent)]

Hi Isaku,

Thanks for the bug report! Its great to also report such issues in the Quantum project (or send email to the netstack list) as sometimes quantum developers won't see a bug that is just filed under nova. Thankfully Salvatore noticed this one.

This bug is specific to using DHCP / L3 NAT gateways.

The fact that LinuxOVSInterfaceDriver doesn't do anything on unplug is not actually the problem (though its something we should clean up anyway, as there will be stale linux devices lying around...).

I believe the real issue is that QuantumManager needs to clean up the "gateway" port on a Quantum network before attempting to delete it, otherwise Quantum refuses to to delete the network with a NetworkInUse error. This should be a QuantumManager fix.

[Openstack Gerrit (openstack-gerrit)]

Reviewed: https://review.openstack.org/2259
Committed: http://github.com/openstack/nova/commit/724e5e7a5bc065be239b3ededf0609de386c1d6f
Submitter: Jenkins
Branch: master

 status fixcommitted
 done

commit 724e5e7a5bc065be239b3ededf0609de386c1d6f
Author: Brad Hall <email address hidden>
Date: Mon Dec 12 18:24:26 2011 +0000

    Fix for bug 902175

    Remove the gateway port when deleting the network (if it's the only one left),
    and kill dnsmasq if the network is deleted.

    Change-Id: If6b4798ddb4d21fe6c32ac20e6237494b799ecbd

[Openstack Gerrit (openstack-gerrit)]

Reviewed: https://review.openstack.org/2258
Committed: http://github.com/openstack/quantum/commit/b0652c0dff90d68a6a7cf84b536e7e539e344f74
Submitter: Jenkins
Branch: master

 status fixcommitted
 done

commit b0652c0dff90d68a6a7cf84b536e7e539e344f74
Author: Brad Hall <email address hidden>
Date: Mon Dec 12 18:26:21 2011 +0000

    Fix for bug 902175

    We can't use merge if we're removing a field from the port object so use add
    instead. Also, pass the session to port_get so that we don't run into the
    "this port is already bound to session x" error.

    Change-Id: I54a8484c8f6429ad18fb0c5e088720d21fc16299