New VIF type to allow routing VM data instead of bridging it¶
We propose to add a new VIF type, VIF_TYPE_TAP, whose meaning is that the host side of a VNIC to a VM is - at least initially - a simple TAP interface that is not plugged into a bridge or vSwitch. A Neutron agent can continue the process of deciding how to handle data from that interface, and how to deliver data to it - but this is beyond the scope of the initial TAP interface setup that Nova needs to provide.
For Project Calico (http://www.projectcalico.org/), we’d like the host side of the pipe into a VM to be a simple, unbridged TAP interface, and currently there is no VIF_TYPE that I can use to get this.
VIF_TYPE_MIDONET, VIF_TYPE_IOVISOR and VIF_TYPE_IVS (without firewall or hybrid plug) all create an unbridged TAP interface, but then do additional things to it (within the Nova code) to connect that TAP interface into the host’s networking system. Other VIF_TYPEs involve bridges or vSwitches, or ‘direct’ attachments to physical host interfaces.
One application is that VIF_TYPE_TAP makes it possible for data to/from VMs, and also between VMs on the same host, to be routed by their immediate compute host instead of being bridged. This is of interest in deployments where VMs only require services at layer 3 (IP) and above, and it is still possible to implement, through iptables and route distribution filters, all of the detailed connectivity and security policies that are implied by any given set of OpenStack’s networking, security group and router configurations. For more on how that can work please see Project Calico (http://www.projectcalico.org/) and https://blueprints.launchpad.net/neutron/+spec/calico-mechanism-driver.
The applicability of VIF_TYPE_TAP should however be wider than just that one project. It enables a class of experimental future networking implementations to be explored in Neutron (with plugin, mechanism driver and agent code) without needing to change or patch any Nova code.
Not applicable. (As advised by John Garbutt.)
(However we would argue that VIF_TYPE_TAP supports the “Nova-network/Neutron migration” priority listed at https://github.com/openstack/nova-specs/blob/master/priorities/kilo-priorities.rst, in that it allows some forms of network connectivity to be explored and developed without needing further changes to the Nova code.)
Add VIF_TYPE_TAP in nova/network/model.py.
Add get_config_tap, plug_tap and unplug_tap methods in nova/virt/libvirt/vif.py, with implementations that simply configure, create and delete a TAP device.
The libvirt config for VIF_TYPE_TAP would be an <interface type=”ethernet”> element with a null script, just as for the existing VIF_TYPE_MIDONET and VIF_TYPE_IOVISOR cases, prepared by calling self.get_base_config and designer.set_vif_host_backend_ethernet_config.
When a VM is to be launched using VIF_TYPE_TAP:
Nova creates the TAP interface, in plug_tap(), by calling linux_net.create_tap_dev
libvirt launches the VM, with the config described just above, to match the created TAP interface.
None. All existing VIF_TYPEs whose plugging is implemented in nova/virt/libvirt/vif.py are unsuitable, as described in Problem Description above. Pre-Juno it was possible to configure use of an out-of-tree VIF driver (with the virt_driver setting), but this was deprecated and has now been removed.
Data model impact¶
REST API impact¶
A compute host can guard against IP address spoofing (by a VM) on a VIF_TYPE_TAP interface by installing iptables rules that require each packet from a VM to have the expected source IP address.
If a VIF_TYPE_TAP interface is not plugged into a bridge, MAC address spoofing by a VM has no impact. If a VIF_TYPE_TAP interface _is_ plugged into a bridge, that bridge can implement similar protections against MAC spoofing as for existing bridged VIF_TYPEs.
Other end user impact¶
Unless the VIF_TYPE_TAP vif type is explicitly requested (e.g. by a Neutron/ML2 mechanism driver class), there is no possible performance impact on a standard OpenStack system.
Other deployer impact¶
The Nova extension proposed here will have no effect on existing or newly deployed OpenStack systems, unless the VIF_TYPE_TAP vif type is explicitly requested somewhere (e.g. by a Neutron/ML2 mechanism driver class).
- Primary assignee:
- Other contributors:
The changes required for this spec have already been implemented by us, based on the Icehouse release code, and fairly extensively tested.
An up to date base is of course appropriate for this spec, so the changes rebased onto the current proposed/juno branch can be seen at the following URL:
Remaining work items are as follows:
Implement unit and 3rd party CI tests as described below.
Verify that proposed changes pass all existing tests (including code style), as well as new tests.
Submit changes formally for review.
Participate in resulting discussions, mark up and re-review processes.
Repeat until done!
There are two related Neutron specs that I have proposed as part of the Project Calico approach, and that build on the capability that VIF_TYPE_TAP provides.
https://blueprints.launchpad.net/neutron/+spec/dhcp-for-routed-ifs enhances the Neutron DHCP agent code to handle DHCP for routed TAP interfaces.
https://blueprints.launchpad.net/neutron/+spec/calico-mechanism-driver provides a Neutron/ML2 mechanism driver that implements routed networking by using VIF_TYPE_TAP.
As already stated above, however, I expect the long term applicability of VIF_TYPE_TAP to be wider than just for Project Calico.
Within the OpenStack ecosystem, this change will be tested at unit-test level, by adding a test case to nova/tests/virt/libvirt/test_vif.py, that creates and verifies a virtual interface with type VIF_TYPE_TAP.
(It will also be extensively tested at the system level by continuing related development and testing at Project Calico (http://www.projectcalico.org/), which uses VIF_TYPE_TAP, and such work will generally be conducted and reported in public.
We understand, though, that this is not formally verifiable testing within the OpenStack ecosystem; so it is mentioned here for information only.)
No documentation changes for Nova are anticipated. VIF_TYPE_TAP will be automatically enabled by a related Neutron/ML2 driver, where appropriate.