.. This work is licensed under a Creative Commons Attribution 3.0 Unported License. http://creativecommons.org/licenses/by/3.0/legalcode ========================================== VLAN aware VMs ========================================== Launchpad blueprint: https://blueprints.launchpad.net/neutron/+spec/vlan-aware-vms This blueprint proposes how to incorporate VLAN aware VMs into OpenStack. In this document a VLAN aware VM is a VM that sends and receives VLAN tagged frames over its vNIC. The main point of that is to overcome the limitations of the current one vNIC per network model. A VLAN (or other encapsulation) aware VM can differentiate between traffic of many networks by different encapsulation types and IDs, instead of using many vNICs. This approach scales to higher number of networks and enables dynamic handling of network attachments (without hotplugging vNICs). Problem Description =================== Currently VLANs in VMs are not integrated with Neutron. There are various use cases where it would be useful to allow a VM to be attached to multiple Neutron networks using VLANs as a local encapsulation to differentiate the traffic for each network as it goes in/out of a single VIF. Use cases: * Some applications have requirements to connect to many (say, hundreds) of Neutron networks. It is more practical to use a single or other small number of VIFs and VLANs to differentiate traffic for each network than to have hundreds of VIFs per VM. * Cloud workloads are often very dynamic. It may be more efficient and/or less complex to add/remove VLANs than to hotplug interfaces in a VM. * A VM could be moved from one network to another without detaching the VIF from the VM. * A VM may be running many containers. Each container may have requirements to be connected to different Neutron networks. Assigning a VLAN (or other encapsulation) id for each container is more efficient and scalable than requiring a vNIC per container. * There are legacy applications that expect to use VLANs as a way to connect to multiple networks. Neutron should provide a way to expose that model to the VM decoupled from how the network is actually implemented. The proposal may look related in some sense to the VLAN-transparent approach (bp/nfv-vlan-trunks), but differs in that it doesn't provide a way to send opaque data through the links, but instead be aware of the encapsulating, provided by the VM (or an appliance inside a VM). Proposed Change =============== The proposal keeps the port resource unchanged. :: $ neutron port-create NETWORK The new API extension looks like this: #. One new trunk resource with member_actions for subport operations: * One trunk resource * Subports encoded in an attribute of the trunk, added and removed by member_actions The proposal is to add an API extension to turn already existing ports into trunk and subports. The best way to think about this is that Neutron needs a method to describe a logical topology such that you have a regular VIF port (the parent) and that a parent may also have child ports. These child ports are used to connect the VIF port to further networks. Both the parent and child ports have all of the attributes that a port has today (however some subport attributes must behave differently, eg. binding:host_id) and can be created on any Neutron network. All needed information beyond legacy port attributes is contained in newly introduced resource(s) referring legacy ports. The first part of the API extension is the logical topology portion. First, we include a method to indicate that a port may have child ports. This isn't strictly necessary, but is included for convenience. It allows Neutron to go ahead and reject the port creation if subports aren't supported by the backend. :: # trunk-create may refer to 0, 1 or more subport(s). $ neutron trunk-create --port-id PORT \ [--subport PORT[,SEGMENTATION-TYPE,SEGMENTATION-ID]] \ [--subport ...] All ports referred must exist. Note that if a given plugin does not implement support for bp/vlan-aware-vms, this request will be rejected. If the backend does not implement the extension any trunk operation will fail (with 404). If a plugin does implement the extension, trunk operations may return a conflict error if the port status is not compatible with the request. In addition to the logical topology piece, we need to define the type of encapsulation (segmentation) used between the hypervisor and VM for targeting packets to a logical subport, as well as for determining the source logical port for a packet arriving from the VM. The encapsulation here is local, used between the VM and hypervisor only. It has nothing to do with how the Neutron network the port is attached to is implemented. SEGMENTATION-TYPE is an enum of strings, currently having one valid value 'vlan'. SEGMENTATION-ID is an unsigned integer. SEGMENTATION-TYPE and SEGMENTATION-ID are optional in both trunk-create and trunk-subport-add in the generic case, allowing neutron server to choose the type and id if needed. On the other hand backends are allowed to reject requests with unspecified segmentation details and make the user supply them. Further segmentation types may be introduced by later specs. :: # trunk-add-subport adds 1 or more subport(s) $ neutron trunk-subport-add TRUNK \ PORT[,SEGMENTATION-TYPE,SEGMENTATION-ID] \ [PORT,...] All ports referred must exist. The CLI has the following new commands: * trunk-create * trunk-delete * trunk-list * trunk-show [should not print subports] * trunk-subport-add * trunk-subport-delete * trunk-subport-list Importantly, after creating trunks and maybe subports, VMs are still booted with a reference to plain old ports: :: $ nova boot --nic port-id=... Nova need not know about trunks and subports. Example logical model:: +-----------+ | | | +-------+ | +------+ | | port1 +-----+ net1 | | +-------+ | +------+ +-----+ | | | | | +-------+ | +------+ | vm0 +------+ port0 +-----+ net0 | | | | +-------+ | +------+ +-----+ | | | +-------+ | +------+ | | port2 +-----+ net2 | | +-------+ | +------+ | | +-----+-----+ ^ | +----------------+--------------+ | | | Ports combined into one VIF | | by turning port0 into a trunk | | and the other ports into | | subports of the trunk. | | | +-------------------------------+ * The traffic of port0 is untagged in vm0. * The traffic of port1 is encapsulated, for example as vlan id 100. * The traffic of port2 is encapsulated, for example as vlan id 200. In the above example, a VM connects to 3 Neutron networks. port0 is a regular port that handles untagged traffic. port0 has two subports, each using VLAN encapsulation with a different VLAN ID. Packets targeted at a subport will be tagged with the appropriate VLAN ID before being sent to the VIF. Packets received on the VIF tagged with a VLAN ID associated with a subport will be treated as if the logical source was that subport and the packet will be sent to the Neutron network that subport is attached to. Subports have independent, arbitrary MAC addresses just like any other Neutron port. If the application requires having subports use the same MAC address as their parent, the port create API supports specifying the MAC address, so it can be specified to match the parent's MAC address. A parent and a subport having the same MAC address are not allowed to be on the same net. Nested subports are not supported. Constraints ----------- * A top level port that may have child ports added must be marked by creating a trunk referring to the port in its `port_id` attribute. * Every subport of a given parent should have unique (segmentation_type, segmentation_id) among its siblings. Otherwise, it would not be possible to properly differentiate the source and destination logical port. * The parent port handles "untagged" traffic. The parent will receive all the packets that do not match a subport, which is no different than how a regular port already behaves today. This also means that every VM has a port for untagged traffic. It doesn't necessarily have to be used though and could be attached to a dummy Neutron network if desired. You could also have a security group set on the parent port to drop all incoming and outgoing traffic. A future enhancement could include the ability to create a Neutron port not yet attached to a network, though it's unclear how valuable that actually is. * A port with a parent logically inherits its binding:host_id from its parent. * An attempt to update binding:host_id of a subport (by booting a Nova VM with a subport UUID, for example) must result in an error. * A normal user can only connect subports to its own parent ports. Admin can connect subports to parent ports of different owners. * Until OVS supports QinQ, an OVS based Neutron backend cannot support having a subport attached to a "vlan_transparent" network. That would require the abilty to transport VLAN tagged traffic over the VLAN tagged subport interface (nested VLANs), which is not yet possible. The following deletion constraints exist: * Deletion of a trunk automatically deletes all of its subports. * Deletion of a (child) port referred by a subport is forbidden. The subport must be deleted first. * Deletion of a (parent) port referred by a trunk is forbidden. The trunk must be deleted first. Nova changes ------------ The Neutron OVS agent uses OVS in such a way that it will likely need to create a new OVS bridge per top level parent port. That requires an enhancement to the existing OVS and vhost-user VIF types in Nova to allow Neutron to tell Nova the name of the bridge to use for the port. Currently, Nova only supports a single bridge name provided by the Nova configuration file. Alternatives ------------ For alternatives please see the history of the spec, including: * v1: https://review.openstack.org/#q,I015762abe197b91916374143410b1f182908909e,n,z patch sets 1 through 10 * v2: https://review.openstack.org/#q,I015762abe197b91916374143410b1f182908909e,n,z patch sets 11 through 21 * v3: https://review.openstack.org/#q,I8563e54cc984b60f2be2eb7cc2e3c6fc60e6de20,n,z patch sets 1 through 5 * v4: https://review.openstack.org/#q,I8563e54cc984b60f2be2eb7cc2e3c6fc60e6de20,n,z patch sets 6 and later (the current) Data Model Impact ----------------- Neutron port: no changes. Neutron DB table: trunks * id: integer primary key * uuid * name * tenant_id * port_id: foreign key to ports.id Neutron DB table: subports * id * port_id: foreign key to ports.id * trunk_id: foreign key to trunks.id * segmentation_type * segmentation_id Neutron DB constraints: * (segmentation_type, segmentation_id, trunk_id) must be unique in the subports table * Ports referred by trunks.port_id must not be referred by subports.port_id and vice versa. REST API Impact --------------- trunk resource: * id * name * tenant_id * port_id :: POST /v2.0/trunks ... PUT /v2.0/trunks/TRUNK-ID/add_subports PUT /v2.0/trunks/TRUNK-ID/delete_subports {'subports': [{'port_id': PORT_ID, 'segmentation_type': SEGMENTATION_TYPE, 'segmentation_id': SEGMENTATION_ID}, ... ]}} GET /v2.0/trunks/TRUNK-ID/subports Security Impact --------------- Only admin can create and attach subports for different tenants. Notifications Impact -------------------- None Other End User Impact --------------------- Tenants should be aware that OpenStack does nothing to enable VMs to handle the tagged traffic, but just provides tagged packets. It is totally up to the user to set VMs up properly. Performance Impact ------------------ The performance of existing functionality should be unaffected. The data path for normal ports is unchanged. In some cases, this change may improve performance. Without this change, connecting a VM to many Neutron networks required a VIF per network. With this change, you could connect to 1000 Neutron networks with very little overhead vs. having to attach 1000 virtual interfaces to your VM before. The tradeoff is that some additional burden is shifted into the VM to deal with that segmentation on a single interface. IPv6 Impact ----------- None. Both parent ports and subports have all of the same attributes as Neutron ports do today, including IPv6 addresses if desired. Other Deployer Impact --------------------- None Developer Impact ---------------- * New neutron resources are to be used * Requires modifications to Neutron plugins to support this model * Requires development of new tests. Community Impact ---------------- Implementation ============== Assignee(s) ----------- * Kevin Benton * Peter V. Saveliev * Russell Bryant * Bence Romsics Work Items ---------- * API extension and DB schema updates * Unit tests for API+DB changes * Tempest tests for creating port topology * Tempest scenario test(s) for doing functional validation * Neutron Plugin support. * networking-ovn (OVN supports this model already) * ml2+ovs * Expose subport information in nova metadata service (likely will require a new spec) Dependencies ============ The change needed for the existing OVS VIF type has `its own blueprint `_. Testing ======= Tempest and functional tests will be created. Full-stack and in-tree API tests -------------------------------- * Create parent ports * Create subports * Bind parent ports * Delete subports * Delete parent ports Functional Tests ---------------- Tests to be implemented: * Boot VM with one parent port and no children. Verify connectivity. * Boot VM with one parent port and multiple subports. Verify connectivity to each logical port. * Boot VM with multiple parent ports and subports and verify connectivity. * Add subport to running VM with a parent port. * Remove subport from running VM with parent port. * Delete VM with parent port including subports. API Tests --------- Tests to be implemented: * Check that subport only can be connected to a parent port. * Check that an invalid segmentation_type is rejected * Check that an invalid segmentation_id is rejected Documentation Impact ==================== The use of parent ports and subports should be documented as a way to create a logical multi-port topology using a single VIF on a VM. Possible scenarios for use cases should be provided with CLI examples. User Documentation ------------------ Update networking API reference. Update admin guide. Developer Documentation ----------------------- The parent port resource description, subports behaviour, API reference. References ========== * `OVS VIF type: pass bridge name from neutron to nova `_ * Subport details may be exposed to VMs via the Nova metadata service in a later spec like this: `Virtual guest device role tagging `_. * `nfv-vlan-trunks `_