Pluggable network providers

https://bugs.launchpad.net/ironic/+bug/1526401

Today, Ironic provisions servers on the same (flat) network that tenants run on. Ironic should have the ability to logically separate provisioning and tenant networks, and switch between them as needed.

Note: where this spec says “network”, it generally means “Neutron network”, which may be implemented in different ways.

Problem description

Ironic currently provisions servers on the same (flat) network that tenants use. It is clear that Ironic should allow operators to separate provisioning and tenant networks; where deploy ramdisks run on the “provisioning” network, and instances run on one or more “tenant” networks.

In order to secure the provisioning network, Ironic should be able to switch between these networks when spawning and destroying instances.

This method should also be extended to other management networks that may or may not be separated from the provisioning network; for example the operator may wish to use different management networks for cleaning or rescue tasks.

Proposed change

Ironic should have a pluggable “network provider” that can handle switching nodes between different networks. This provider should be able to be selected per Ironic node object. A new network_provider field will be added to the node object to define which network provider to use for that node. There should also be a configuration option for the default network provider, defaulting to ‘none’ for now. The default value for node.network_provider will be NULL, meaning use the configuration option. Both the node field and the configuration option may be set to ‘none’ or ‘neutron’, mapping to the no-op provider and the Neutron provider, respectively.

This network provider system will not be part of Ironic’s driver interface mixin system; rather it is standalone and is loaded on demand via stevedore. This is similar to how the DHCP provider code works today.

An initial implementation of a no-op network provider should be written, to put the logic in place in Ironic while maintaining compatibility with the existing model.

A second implementation should be written that uses Neutron to attach hardware to networks as needed, and should work as follows.

The network provider should have a concept of a “provisioning network”, where the control plane can communicate with nodes managed by Ironic in order to deploy, tear down, or otherwise manage nodes. It should be able to connect and disconnect nodes from this network.

The network provider should also know how to connect and disconnect nodes from arbitrary tenant networks. These networks are defined by the Nova user and Nova. Nova should continue to create Neutron ports (a logical attachment to a network), but these ports will be left unbound, as they will not have enough information to be plumbed through. Ironic later should send a port-update call to Neutron, to pass the necessary information to complete the binding. This call should happen after deploying the image, between the power off and power on calls that boot to the instance image. This may have implications for “boot from volume” workloads. As Ironic does not yet support these workloads, these are out of scope for the purposes of this spec.

In the case where the Ironic driver is used, Nova should send a null host_id in the binding profile. This will prevent Neutron from binding the port immediately, so we can defer this and allow Ironic to tell Neutron to bind the port when it is ready to do so. Ironic should send the node UUID as the host_id. Ironic will also delete the Neutron port connecting the node to the provisioning network at this time. The reverse will happen at tear down time.

If an older client (e.g. Nova) is in use and does initially send the host_id, Ironic needs to handle this. There are two cases here:

  • The node is using the Neutron network provider. In this case, Ironic should fetch the ports first, and if the ports are already bound with the correct info, do nothing. If binding failed due to missing switchport information, Ironic should update the port appropriately and allow it to be bound.
  • The node is using the ‘none’ network provider. In this case, the node is expected to be on the provisioning network after deployment (today’s current model). In this case, the ports should be treated as they are today, putting DHCP configs on those ports, etc.

Nova and Ironic should both use the binding:profile dictionary to send data such as physical switchport information.

Nova currently has a broken assumption that each Ironic port (physical NIC) may only attach to one network. This assumption will need to be fixed, as hardware (today) cannot spawn arbitrary NICs like virtual servers can. We may, in the future, also need a way for Nova users to define which NIC is attached to which networks. A first version should leave this assumption in place for the sake of simplicity.

If port groups exist for a node, those should be connected to the networks rather than the individual port. This allows for an aggregated connection such as a LAG to connect to the network.

Note that an Ironic environment may be deployed without Nova. In this case, the user should make the same calls Nova would make.

One caveat here is that nodes will not be able to PXE boot the instance image if they cannot reach the conductor (for tftp). Local boot will need to be used for any node deployed outside of the provisioning network. Any deploys outside of the provisioning network that do not use local boot should error.

Deploy drivers should call to this interface at proper times to switch between networks. For example, a driver completing a deploy should power off the node, switch to the instance networks, and power on the node. This will ensure that the deploy ramdisk never runs on the instance network, and the instance image never runs on the provisioning network.

Alternatives

Alternatively, Ironic could continue to prescribe that operators run Ironic on a single flat network shared between tenants and the control plane. This is clearly not viable for many real-world use cases.

Data model impact

A network_provider field will be added to the Node object.

State Machine Impact

None.

REST API impact

Update the REST API for the node object to allow reading and modifying the new network_provider field. This will likely need a version bump.

Client (CLI) impact

Will need to update the CLI to print the new Node.network_provider field, when available.

RPC API impact

None.

Driver API impact

This adds a new interface, NetworkProvider. This interface is not a part of Ironic’s driver composition system, to be clear. This interface will define the following methods:

def add_provisioning_network(self, task):
"""Add the provisioning network to a node."""

def remove_provisioning_network(self, task):
"""Remove the provisioning network from a node."""

def add_cleaning_network(self, task):
"""Add the cleaning network to a node."""

def remove_cleaning_network(self, task):
"""Remove the cleaning network from a node."""

def configure_tenant_networks(self, task):
"""Configure tenant networks (added by Nova/user) for a node."""

def unconfigure_tenant_networks(self, task):
"""Unconfigure tenant networks (to be removed by Nova/user) for a node."""

Nova driver impact

The Nova driver should not be directly impacted here; however, this does depend on changes to the Neutron network driver in Nova as described above.

Ramdisk impact

N/A

Security impact

This potentially improves security by restricting tenant access to the control plane.

Other end user impact

To use this feature, end users will need to:

  • Set nodes to use the Neutron provider.
  • Use local boot for nodes using the Neutron provider.

Scalability impact

When configured to use the Neutron plugin, this will result in additional API calls to Neutron to manage a node. However, impact on scalability should be negligible.

Performance Impact

None.

Other deployer impact

Two new configuration options will be added:

  • CONF.provisioning_network specifies the ID of the provisioning network.
  • CONF.default_network_provider specifies the default network provider to use for nodes with node.network_provider set to NULL.

A new database column (Node.network_provider) is also added, and so deploying this change will require a database migration to be ran.

Deployers will need to deploy a version of Nova that supports this feature, if using Nova.

Deployers will need to deploy an ML2 mechanism driver that supports connecting baremetal resources to Neutron networks.

Developer impact

Driver authors should support this feature by calling the methods provided.

Implementation

Assignee(s)

jroll <jim@jimrollenhagen.com>

And hopefully many others! :)

Work Items

  • Add the Node.network_provider field and the default_network_provider configuration option..
  • Implement the base interface.
  • Implement the no-op provider.
  • Instrument each deploy driver with calls to this interface.
  • Implement the Neutron plugin provider.
  • Modify Nova to send the extra flag discussed above, when creating ports for a machine using the Ironic virt driver.

Dependencies

None.

Testing

The no-op provider will be tested in the gate by default.

Neutron will provide an ML2 mechanism that simulates connecting real hardware to real switches. When that mechanism is available, we can test the Neutron provider in the gate.

Upgrades and Backwards Compatibility

Default behavior is the current behavior, so this change should be fully backwards compatible.

Documentation Impact

This feature will be fully documented.

References

Discussions on the topic include: