Add MacVTap as new virtual interface type for libvirt virtualization

Add MacVTap as new virtual interface type for libvirt virtualization

https://blueprints.launchpad.net/nova/+spec/libvirt-macvtap-vif

Add macvtap as new virtual interface (vif) type to Nova’s libvirt driver. This is required to attach libvirt managed KVM guests via macvtap to the hosts network. It will be exploited by a new macvtap ml2 plugin and agent. The current spec is hosted on github [1] but the final goal is to have this under the big tent.

Macvtap guest attachments come with the following values
  • Significantly higher throughput than the reference implementation ovs

  • Significantly lower latency than the reference implementation ovs

  • Significantly lower CPU consumption per throuput in the hypervisor than the reference implementation ovs

  • Built into each kernel - no additional packages required

  • Very less configuration required

However, the disadvantage of an macvtap attachment is, that Security Groups and Rules can technically not be supported for now. More details see section ‘Security impact’.

Problem description

The new macvtap neutron driver and agent [1] requires nova vif driver integration. The corresponding vif_type should represent a general macvtap device. It is responsible for * creating the xml definition accordingly * the correct plug/unplug operations

Use Cases

  • Attaching libvirt/KVM guests via macvtap to the hosts network

No additional impact on actors from the Nova side (configuration is done in Neutron)

Project Priority

None

Proposed change

Adding support for a general macvtap vif_type that is represented by the following domain.xml:

<interface type='direct'>
  <source dev='<macvtap_src>' mode=’<mode>’/>
  <model type='virtio' />
</interface>

The following attributes are variables that must be propagated by neutron to nova using the vif dictionary:

  • macvtap_src: This is the source device for the macvtap (the device the macvtap sits on).

  • mode: The mode, in which the macvtap device should be instantiated. e.g. ‘bridge’ or ‘vepa’

Bandwith configurations are supported like with other vif types using a macvtap based connection (e.g. ‘hw_veb’).

This proposed change will consider the ongoing os-vif-library discussion ([5] or alternative [3]) and

  • implement things in this new way, if it gets approved.

  • implement things in the old way, if it will be moved out.

The corresponding neutron code will use the vnic_type ‘normal’.

Alternatives

Reusing an exsiting vif_type

Reusing an already existing type is not possible:

  • It depends on [5] to allow different plug/unplug operations

  • It requires a refactoring of existing vif_types xml generation method

  • An existing vif_type would have to be renamed which breaks other ml2 plugins

Creating a new generic vif_type for direct connections

This will be part of the os-vif-library effort [5].

Supported vnic types

The plan is to support only the vnic_type ‘normal’.

However two other vnic_types do exist: ‘direct’ and ‘macvtap’. These other vnic_types both trigger PCI Requests in nova, so they cannot be reused. The new macvtap support should be hardware independet. But I see the confusion with the existing ‘macvtap’ type by just looking at the name.An idea would be to rename the type ‘macvtap’ to somehting like ‘sriov-macvtap’. But this requires another blueprint and probbly an nova api change.

Data model impact

None

REST API impact

None

Security impact

Macvtap does not provide a hook to apply iptables or ebtables on the guest’s traffic, and therefore does not support Neutron Security Groups and anti-spoofing-rules. Technically, the linux macvtap driver could implement such a hook in the future. But even without such a hook, macvtap prevents MAC spoofing already today.

As a consequence, configuration of the NoopFirewallDriver is desired, like it is with other integrated direct attachment types.

Notifications impact

None

Other end user impact

None

Performance Impact

None from a Openstack code point of view.

But the network performance of the guest should increase compared to the ovs reference implementation.

Other deployer impact

None for Nova

Developer impact

None

Implementation

Assignee(s)

Primary assignee:

scheuran (andreas.scheuring@de.ibm.com)

Other contributors:

None

Work Items

Dependencies

New vif-plug approach:

Testing

  • Unittest

  • No tempest tests. They will be added with the corresponding neutron code and run by a neutron Thirdparty CI.

Documentation Impact

None for Nova

References

History

Revisions

Release Name

Description

Liberty

Introduced

Creative Commons Attribution 3.0 License

Except where otherwise noted, this document is licensed under Creative Commons Attribution 3.0 License. See all OpenStack Legal Documents.

nova-specs