BGP¶
Note
This page documents the OVN BGP service plugin built into Neutron. This is distinct from the legacy neutron-dynamic-routing BGP speaker, which is an external project with its own agent.
The OVN BGP service plugin integrates BGP route advertisements directly into the Neutron server and the OVN Neutron Agent. It is designed for spine-and-leaf data-centre fabrics where every rack (leaf) is an independent L3 routing domain and inter-rack traffic is forwarded over BGP rather than stretched L2 VLANs.
Architecture overview¶
A spine-and-leaf fabric consists of:
Leaf switches – top-of-rack (ToR) switches, each providing L2 connectivity inside a single rack.
Spine switches – interconnecting all leaf switches via routed (L3) links.
Because every link between a leaf and the spine is a routed link, there is no end-to-end L2 domain across racks. BGP runs on every leaf (and optionally on every compute node) to exchange reachability information.
┌────────┐ ┌────────┐
│ Spine1 │ │ Spine2 │
└───┬┬───┘ └───┬┬───┘
┌─────┘└─────┬─────────┘└─────┐
│ BGP │ BGP │ BGP
┌────┴───┐ ┌────┴───┐ ┌─────┴──┐
│ Leaf 1 │ │ Leaf 2 │ │ Leaf N │
└──┬──┬──┘ └──┬──┬──┘ └──┬──┬──┘
│ │ │ │ │ │
┌─┘ └─┐ ┌─┘ └─┐ ┌─┘ └─┐
CN1 CN2 CN3 CN4 CN.. CN..
Each compute node (CN) has a point-to-point L2 link to its leaf switch port.
On the Neutron server side, a flat provider network is connected to a BGP
logical router topology that the plugin creates in the OVN Northbound
database. On each chassis, a BGP peer bridge (an OVS bridge carrying
both a physical NIC port and a patch port to br-int) bridges the OVN
overlay out to that per-NIC L2 link facing the leaf switch. This way the
relevant IP prefixes (router gateway IPs, floating IPs) are advertised to the
leaf so that return traffic from the spine can be routed back to the correct
host.
Components¶
The feature has two components:
BGP service plugin (ovn-bgp) – runs inside the Neutron server
process. It watches network creation events and manages the BGP topology in
the OVN Northbound database (logical routers, route policies, VRFs).
BGP agent extension (ovn-bgp) – runs inside the OVN Neutron Agent on
every compute/network node. It programs OpenFlow rules on the provider bridge
so that traffic arriving from the leaf switch is correctly steered to the OVN
integration bridge, and vice-versa.
Important
Neither the BGP service plugin nor the BGP agent extension speaks the BGP protocol itself. They only prepare the OVN logical topology and the OpenFlow data-path. An external routing suite such as FRR must be deployed on every node where BGP route advertisement is required (typically every compute and network node). FRR peers with the leaf switch and advertises the routes that OVN makes available through its VRF-based logical routers.
Why VLAN provider networks are not supported¶
When the BGP service plugin is enabled it rejects creation of VLAN provider networks with an HTTP 400 error:
Warning
VLAN provider networks are not supported when the BGP service plugin is enabled. Only flat provider networks are supported.
The reason is architectural. In a spine-and-leaf fabric:
L2 does not cross rack boundaries. Each leaf switch terminates L2 at the rack level. A VLAN-tagged frame leaving one leaf cannot reach a different leaf through the spine because the spine routes at L3.
VLAN tags are a local concern. The leaf switch strips/adds VLAN tags locally; they have no meaning beyond the rack. A Neutron VLAN provider network implies that a given VLAN ID carries the same broadcast domain across all hosts – an assumption that breaks in a routed fabric.
Flat networks model the local segment correctly. A flat provider network does not assume any global L2 scope. Combined with BGP route advertisement the reachability of IPs behind different racks is handled entirely at L3, which matches the physical topology.
Note
If your deployment requires VLAN segmentation within a single rack (e.g. for separating management traffic from tenant traffic on the same physical NIC), that segmentation is handled at the OVS / OS level on the host and does not need a Neutron VLAN provider network.
The existing documentation for
BGP floating IPs over L2 segmented networks describes a different
approach that uses the legacy neutron-dynamic-routing project and VLAN
segments confined to individual racks. That setup is unrelated to the OVN
BGP plugin.
Enabling the plugin¶
Server side¶
Add ovn-bgp to the list of service plugins in neutron.conf:
[DEFAULT]
service_plugins = ovn-router,ovn-bgp
The plugin registers the following configuration options under the [bgp]
section:
Option |
Default |
Description |
|---|---|---|
|
|
VRF ID attached to the main BGP logical router. |
|
|
VRF ID for per-chassis BGP routers. Used to learn default routes for egress traffic leaving the chassis. |
Agent side¶
On every compute or network node, enable the ovn-bgp agent extension in
the OVN agent configuration file (e.g.
/etc/neutron/plugins/ml2/ovn_agent.ini):
[agent]
extensions = ovn-bgp
The agent extension watches OVS and OVN Southbound events to:
Detect provider bridge creation and configure OpenFlow rules that steer ingress traffic from the physical NIC to the OVN integration bridge (via the patch port), and allow egress traffic out.
Track the logical router port (LRP) MAC address bound to the local chassis so that incoming packets are rewritten with the correct destination MAC before entering OVN’s pipeline.
Advertise the set of BGP bridges to the OVN Southbound
Chassistable so that the server-side reconciler can create the appropriate per-chassis logical routers and route policies.
Creating provider networks¶
With the BGP plugin active, provider networks must use the flat type:
$ openstack network create --external \
--provider-physical-network physnet1 \
--provider-network-type flat \
provider
Attempting to create a VLAN provider network will fail:
$ openstack network create --external \
--provider-physical-network physnet1 \
--provider-network-type vlan \
--provider-segment 100 \
provider-vlan
BadRequestException: 400: ...VLAN provider networks are not supported...
Overlay (Geneve, VXLAN, GRE) self-service networks remain fully supported and are unaffected by the BGP plugin.
Leaking tenant subnet routes¶
By default, only the external gateway IP addresses of Neutron routers are
advertised via BGP. The leak_routes subnet attribute allows an
administrator to additionally advertise a tenant (self-service) subnet’s CIDR
to the physical fabric so that external hosts can reach tenant workloads
directly without floating IPs.
Prerequisites¶
Before enabling leak_routes on a subnet, ensure:
The subnet belongs to a Geneve overlay network.
The subnet is attached to a Neutron router that has an external gateway.
Enabling route leaking¶
The leak_routes attribute is set on the subnet via an update operation. It
is an admin-only attribute by default:
$ openstack subnet set --leak-routes <subnet-id>
To disable:
$ openstack subnet set --no-leak-routes <subnet-id>
When leak_routes is enabled the plugin:
Connects the BGP main logical router (
bgp-lr-main) to the tenant logical switch corresponding to the subnet’s network.Adds a static route on
bgp-lr-mainfor the subnet CIDR pointing to the router’s external gateway IP as the nexthop.
The static route is then redistributed into BGP by FRR (or whichever routing suite runs on the compute nodes), making the tenant subnet reachable from the physical fabric.
When leak_routes is disabled (or the subnet is detached from the router)
the static route is removed. If this was the last leaked subnet on the
network, the connection between bgp-lr-main and the tenant logical switch
is also removed.
Known limitation: all IPs on the network are advertised¶
Important
OVN operates at the logical switch (network) level, not at the
individual subnet level. When leak_routes is enabled on any subnet
of a network, the plugin must connect bgp-lr-main to the network’s
logical switch. As a side effect, IP addresses from all subnets on
that network – not just the explicitly leaked subnet(s) – become visible
to bgp-lr-main and are advertised to the physical fabric as host
routes via BGP.
bgp-lr-main does not create routes for non-leaked subnets, so it
will not forward traffic for those IPs. The advertised host routes are
therefore misleading: external hosts learn the routes but traffic sent to
non-leaked IPs will be dropped at the bgp-lr-main router.
This is an inherent limitation of the OVN data model, which does not support connecting a logical router to a single subnet within a logical switch. To avoid advertising unintended IP addresses, administrators should place subnets that require route leaking on dedicated networks that do not share the logical switch with subnets whose addresses should remain hidden from the underlay.