Group Policy ID BGP Extended Community

Terminology The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP14 when, and only when, they appear in all capitals, as shown here. EVPN: Ethernet Virtual Private Networks, as per . GBP: Group Based Policy VXLAN: Virtual Extensible LAN NVO: Network Virtualization Overlay NVE: Network Virtualization Edge DCI: Data Center Interconnect

Introduction In the virtualized overlay network where EVPN with VXLAN encapsulation is used as the overlay solution, without external management software or controller, the propagation of a Group Policy ID is done through the data plane. The source Group Policy ID is encoded in the VXLAN header before the user traffic is sent to the VXLAN tunnel. The encoding format of a Group Policy ID in the VXLAN header is specified in . When the source Group Policy ID is propagated through the data plane to the remote VXLAN tunnel endpoint, the policy enforcement is carried out at the egress node based on both the source and destination Group Policy tags. The policy rule for the source and destination Group Policy tags may result in the traffic being dropped at the remote VXLAN tunnel endpoint which is the egress node. To send the traffic all the way from an ingress node and then drop it at an egress node is an inefficient use of the network bandwidth. To optimize the network bandwidth usage, it may be desirable to have policy enforcement done at the head-end of a VXLAN tunnel that is the ingress node for the user traffic. To accomplish this, there is a need to communicate the destination Group Policy ID from the egress node to the ingress node. This document defines a Group Policy ID BGP Extended Community that can be used in the control plane to achieve the propagation of Group Policy ID from an egress node to an ingress node.

NVO Use Case In an EVPN VXLAN overlay network, a policy group tag may be assigned based on the MAC, IP, port, VLAN, etc, or a combination of the above. Similar to the MAC/IP addresses in the EVPN network, once the Policy Group ID is known for a local host/server/VM attached to an EVPN network, its Group Policy ID can be advertised to other Network Virtualization Edge devices in the control plane through the Group Policy ID extended community. The scheme used for classification and allocation of Policy Group IDs used for GBP in an EVPN overlay network with VXLAN encapsulation is outside the scope of this document. Policy group tag propagation in the EVPN/BGP control plane can be applied to the EVPN type-2 MAC/IP route, EVPN type-3 Ethernet Inclusive Multicast route or EVPN type-5 IP host and prefix route . If Policy Group ID is allocated for a MAC address, IP host or prefix address through the GBP classification scheme, EVPN can encode its Group Policy ID through the Group Policy ID extended community and advertise it alongside its corresponding EVPN route. For the flows that the ingress VXLAN tunnel endpoint has learned its destination group policy tag through EVPN/BGP control plane signaling, the policy enforcement can be thus carried out right at the ingress node. Otherwise, policy enforcement can be carried out at the egress node. If policy enforcement is carried out at the head-end VXLAN tunnel, the ingress node MUST set the GBP applied bit, the A-bit as it is specified in , to 1 in the VXLAN header before forwarding the traffic to the VXLAN tunnel. Otherwise, the ingress node sets the A-bit to 0 in the VXLAN header.

Interconnecting multiple EVPN VXLAN domains EVPN VXLAN based deployments may comprise of multiple EVPN networks, domains or sites. In such cases, a VXLAN overlay may extend from an ingress node to an egress node across different domains; or it may be divided into multiple stitched overlay segments that are interconnected via DCI through gateway devices. In this document, we simply refer to each EVPN network or site as a EVPN domain or domain for short unless it is explicitly specified otherwise. From a control plane point of view, border GWs in each domain may learn routes of other domains either via direct peering sessions or via a set of external route reflectors. In such deployments, the allocation and management of Group Policy IDs may be done independently in different domains, and consequently the allocated values scoped to each domain. Therefore, when a group policy tag is signaled with routes to a different domain, the tag needs to be translated to a value local to the receiving domain before it can be used in a group based policy at an ingress node in that domain. A domain may receive routes from multiple sender domains. In order to facilitate simpler and flexible application of translation policies regardless of the deployed overlay design or control plane peering model, the advertised Policy ID may also carry with it a Scope, which identifies the allocation domain. Any suitable BGP node in the route distribution path can then consistently translate a received Policy ID based on the scope. Scope assignment is done by the administrator or orchestration system managing the multi-domain deployment. The exact mechanism is out of the purview of this document.

EVPN Interworking with IPVPN In the EVPN interworking use case as it is specified in the , two or more EVPN networks/domains are interconnected by a layer 3 IP-VPN network with VPN-IPv4/VPN-IPv6 BGP address families. To support ingress policy enforcement, the Policy Group ID extended community needs to be propagated by the GW PEs sitting at the border of an EVPN domain and IP-VPN domain from one domain to another. For the Uniform-Propagation-Mode defined in the , when propagating an EVPN IP prefix route across the domain boundary to IP-VPN network, the Gateway PE SHOULD propagate communities, extended communities and large communities except for all the EVPN extended communities. The Policy Group ID extended community defined in this document is a new transitive Opaque Extend Community. It is not subject to stripping at the GW PE when the Uniform-Propagation-Mode is used, and SHOULD be propagated.

The Group Policy ID Extended Community The Group Policy ID BGP Extended Community is a new transitive Opaque Extended Community with a Type value of 0x03. This extended community may be advertised along with an EVPN type-2 MAC/IP route, EVPN type-3 Ethernet Inclusive Multicast route, and EVPN type-5 IP prefix route. This new Opaque Extended Community enables the EVPN route it is attached to propagate the Group Policy ID used for Group Based Policy in the control plane. When the "Uniformed-Propagation-Mode" is used under the EVPN and IPVPN interworking use case, the Group Policy ID extended community is carried over by the GW PE when a route for a given IP or IPv6 prefix is propagated from one domain to another with a different address family. Policy ID Scope: The Policy ID Scope field is 16-bit long, and is an optional field. Group Policy ID (GPI): The GPI field is 16-bit long and it encodes the value of a Group Policy ID. The reserved fields MUST be set to zero by the sender and ignored by the receiver. If the Policy ID Scope is not set, any EVPN VXLAN NVE node that receives a route with a Group Policy ID may use the received value as is. If the Scope is set, a node that has the same locally configured Scope in the received route may use the received Policy ID value. A node that has a different local Scope than in the received route may need to translate the received Policy ID to a locally assigned value.

IANA Considerations For the Group Policy ID extended community defined in this document, IANA has allocated the following codepoint in the Sub-type registry of Type 0x03 Transitive Opaque Extended Community.

Security Considerations All the security considerations for BGP extended communities can be applied there. Attackers may alter the value carried in a BGP extended community. In this case, the Group Policy ID carried in the Group Policy ID field can be altered by attackers, which could lead to the wrong policy rule being enforced on the user traffic.

Acknowledgements The authors would like to thank Jeffrey Zhang, Jeff Haas for their careful review and valuable feedbacks. We also would like to thank Prasad Miriyala and Selvakumar Sivaraj for their contributions.