Subnet Service Types

RFE: https://bugs.launchpad.net/neutron/+bug/1544768

Launchpad blueprint: https://blueprints.launchpad.net/neutron/+spec/service-subnets

The current IP allocation method for ports on the external network utilizes the same strategy done for internal networks - choose an IP from the subnet(s) associated with the network. In this scenario, all subnets are created equal, and there is no way to give precedence to one over the other.

Since these IP addresses are typically Internet-routable, or at least routable within the datacenter they are used in, they should be considered scarce, and be assigned sparingly.

This blueprint aims to change the allocation strategy to better utilize this scarce resource, and only use it for ports that must have Internet-routable IP addresses, using secondary subnets for IP allocations for other ports.

Problem Description

In Neutron, each router with an interface on the external network consumes one (or more) IP addresses from the subnet configured on it. For example, on the Network Node, an IP address is assigned to a project’s (aka tenant’s) router in order to allow instances to access the external network using this single SNAT IP address.

In the case of a router on a distributed (compute) node, which is also connected to the external network, a single Floating IP (FIP) namespace is created in order for each instance with a floating IP address to be able to directly reach the external network. The floating IP namespace also consumes an IP address from the external subnet to use for all tenant routers residing on this node, even though it is never used for north-south routing of traffic - it is mainly used for reachability testing (aka - can I reach that distributed router).

In this second case, there is no reason an IP address from a different, secondary subnet, could not be used for the primary IP address inside the FIP namespace, such that reachability can still be confirmed, but public IP consumption is not impacted.

Also, when a router is just forwarding packets onto a transit network for further processing by another device, like a carrier-grade NAT, no NAT needs to be performed in the router, the packet just follows the standard forwarding data path. So using a subnet that exists only within the datacenter is possible.

Being able to choose the subnet used for IP allocation in all these cases is highly-desired by operators.

Proposed Change

This spec proposes adding a new logical item, called ‘service_type’, to the Subnet class object in Neutron, a list attribute that can specify the usage type for the subnet. A type must be one of the currently valid values for the ‘device_owner’ field in the Port model (port[‘device_owner’]), for example:

  • ‘network:floatingip_agent_gateway’
  • ‘network:router_gateway’
  • ...

Since it is a list, it can contain more than one value in order to specify IP addresses for multiple device_owner types that should be allocated from the same subnet.

This list will be instantiated as a new table containing rows of (subnet_id, service_type) values, such that a join from Subnets.id can be made to determine the service_types associated with the subnet.

When an IP allocation is performed, for example, when an external gateway is added, and the plugin supports the ‘service-type’ extension, the IPAM driver will return an address from the Subnet that has a ‘service-type’ matching the port’s device_owner field.

In the case where no Subnets with the given ‘service-type’ exist for the port’s device_owner, or the IP address pool of all matching subnets is exhausted, the fallback is to use a Subnet without any type for IP allocation. This preserves backwards-compatibility with existing deployments.

When all Subnets are created with a ‘service-type’, there is no fallback strategy since a Subnet without a type does not exist. This can be used to enforce strict IP allocation from only Subnets where the device_owner matches.

When multiple Subnets exist with the given ‘service-type’, or there are both a matching and one without a type, then they will be tried in order, with the matching ‘service-type’ Subnet(s) being tried first. For example, with a Floating IP agent gateway port, the precedence order would be [‘network:floatingip_agent_gateway’, None].

If a Subnet is specified explicitly in the port create or update request, the logic to select candidate Subnets described above will be skipped. The specified Subnet will be the only candidate regardless of its ‘service-type’.

With the data model proposed, all of this can be accomplished by simply ordering all the matching subnets correctly, putting ‘service-type’ subnets first, followed by non-‘service-type’ subnets, in a single query.

Data Model Impact

  1. Service-type subnet extension is added
  2. New table is created linking Subnet IDs with Service Types. It will have the columns (subnet_id, service_type), where the subnet_id has a foreign key constraint on Subnets.id.
  3. The primary key will be (subnet_id, service_type) to provide an index and uniqueness.

REST API Impact

  • ‘service_type’ will be exposed to users via the extension API.
  • Ability to create or update subnets with a service-type will be
    limited to the admin user, since it will only be affecting IP allocation on external networks. Validation will be done to restrict the values to a known set of values supported by the extension.
  • List Subnets of a given service-type.

Sample REST calls:

POST /v2.0/subnets
{
      'subnet': {
          'network_id': '<NETWORK-ID>',
          'ip_version': 4,
          'cidr': '10.0.3.0/24',
          'allocation_pools': [{'start': '10.0.3.20', 'end': '10.0.3.150'}],
          'service_type': ['network:floatingip_agent_gateway']
      }
}

PUT /v2.0/subnets/SUBNET-ID/service_types/
{
    'service_type': ['network:floatingip_agent_gateway']
}

GET /v2.0/subnets?service_type='network:floatingip_agent_gateway'

Command Line Client Impact

A new option will be added to the openstackclient, for example:

$ openstack subnet create {ARGS} –service-type SERVICE-TYPE SUBNET-NAME

Security Impact

None.

Notifications Impact

None.

Other End User Impact

Horizon will need to at least expose the service_type attribute on the subnet.

Performance Impact

There should be almost no performance impact as this should just be an additional table join in the IPAM code.

IPv6 Impact

This change will consider IPv4 and IPv6 equally.

Other Deployer Impact

Deployers will not be required to do anything when upgrading code until they want to make use of the new features provided by this spec.

New or existing deployments can make use of this new feature after upgrading their code by adding a new Neutron subnet to the existing external network, or change the ‘service-type’ of an existing subnet on that network if required.

They might need to change configuration in their top-of-rack switches, or whatever else needs to be done to have a secondary subnet on their existing physical network. That work is outside the scope of this spec.

Developer Impact

This change will have an effect on developers. This blueprint contains API and model changes that plugin providers will need to look at implementing.

Community Impact

Yes. This change has been discussed on the ML, in Neutron meetings (especially L3), at mid-cycles, and at the design summit.

Alternatives

An alternative is to not use Public IP addresses as Floating IP addresses, but instead use IP addresses that are routable within the datacenter infrastructure. Access to the public Internet is then accomplished via a carrier-grade NAT device that sits between the two networks.

Implementation

Assignee(s)

Primary assignee:

Other contributors:

Please add your name here and attend the L3 Subteam Meeting if you’d like to contribute. If your name is here and you don’t have any time to contribute, let me know. This section is really just my sticky note of contributors who can play a role in this effort.

Work Items

  1. Subnet extension
    • For ML2, use the multi-provider extension data model.
    • Add new ‘Subnet_service_types’ table
    • Add logical service_type value to Subnet
    • API tests
  2. Client support
  3. IPAM support
    • When choosing an IP address to allocate to a port, IPAM needs to take into account the device_owner when selecting the subnet.

Dependencies

None.

Testing

Tempest Tests

Unknown

Functional Tests

Unknown

API Tests

  • Subnet service-type resource (CRUD)

Documentation Impact

Yes

User Documentation

Document the usage of the value in the Subnet

Developer Documentation

Yes

References