Cumulus Linux provides virtual routing and forwarding (VRF) to allow for the presence of multiple independent routing tables working simultaneously on the same router or switch. This permits multiple network paths without the need for multiple switches. Think of this feature as VLAN for layer 3, but unlike VLANs, there is no field in the IP header carrying it. Other implementations call this feature VRF-Lite.
The primary use cases for VRF in a data center are similar to VLANs at layer 2: using common physical infrastructure to carry multiple isolated traffic streams for multi-tenant environments, where these streams are allowed to cross over only at configured boundary points, typically firewalls or IDS. You can also use it to burst traffic from private clouds to enterprise networks where the burst point is at layer 3. Or you can use it in an OpenStack deployment.
VRF is fully supported in the Linux kernel, so it has the following characteristics:
- The VRF is presented as a layer 3 master network device with its own associated routing table.
- The layer 3 interfaces (VLAN interfaces, bonds, switch virtual interfaces/SVIs) associated with the VRF are enslaved to that VRF; IP rules direct FIB (forwarding information base) lookups to the routing table for the VRF device.
- The VRF device can have its own IP address, known as a VRF-local loopback.
- Applications can use existing interfaces to operate in a VRF context — by binding sockets to the VRF device or passing the
cmsg. By default, applications on the switch run against the default VRF. Services started by
systemdrun in the default VRF unless the VRF instance is used. If management VRF is enabled, logins to the switch default to the management VRF. This is a convenience for users to not have to specify management VRF for each command.
- Listen sockets used by services are VRF-global by default unless the application is configured to use a more limited scope — for example, read about services in the management VRF. Connected sockets (like TCP) are then bound to the VRF domain in which the connection originates. The kernel provides a sysctl that allows a single instance to accept connections over all VRFs. For TCP, connected sockets are bound to the VRF the first packet was received. This sysctl is enabled for Cumulus Linux.
- Connected and local routes are placed in appropriate VRF tables.
- Neighbor entries continue to be per-interface, and you can view all entries associated with the VRF device.
- A VRF does not map to its own network namespace; however, you can nest VRFs in a network namespace.
- You can use existing Linux tools to interact with it, such as
You configure VRF by associating each subset of interfaces to a VRF routing table, and configuring an instance of the routing protocol — BGP — for each routing table.
Each routing table is called a VRF table, and has its own table ID. You configure VRF using NCLU, then place the layer 3 interface in the VRF. You can have a maximum of 64 VRFs on a switch.
When you configure a VRF, you follow a similar process to other network interfaces. Keep in mind the following for a VRF table:
- It can have an IP address, a loopback interface for the VRF.
- Associated rules are added automatically.
- You can also add a default route to avoid skipping across tables when the kernel forwards the packet.
Names for VRF tables can be up to 15 characters. However, you cannot use the name mgmt, as this name can only be used for management VRF.
To configure a VRF, run:
These commands result in the following VRF configuration in the
Specifying a Table ID
Instead of having Cumulus Linux assign a table ID for the VRF table, you can specify your own table ID in the configuration. The table ID to name mapping is saved in
/etc/iproute2/rt_tables.d/ for name-based references. So instead of using the
auto option above, specify the table ID like this:
If you do specify a table ID, it must be in the range of 1001 to 1255 which is reserved in Cumulus Linux for VRF table IDs.
Configure Route Leaking
Route leaking enables route prefixes from one VRF instance to be leaked to another VRF instance on the same switch. The example command below adds a new route specification (
126.96.36.199/24) to a VRF named
blue, with a device named
swp2. For this example,
swp2 is assumed to be part of a second VRF, named
Bringing a VRF Up after Downing It with ifdown
If you take down a VRF using
, to bring it back up you need to do one of two things:
ifup --with-depends <vrf>
Using the vrf Command
vrf command returns information about VRF tables that is otherwise not available in other Linux commands, such as
iproute. You can also use it to execute non-VRF-specific commands and perform other tasks related to VRF tables.
To get a list of VRF tables, run:
To return a list of processes and PIDs associated with a specific VRF table, run
vrf task list <vrf-name>. For example:
To determine which VRF table is associated with a particular PID, run
vrf task identify <pid>. For example:
Running IPv4 and IPv6 Commands in a VRF Context
You can execute non-VRF-specific Linux commands and perform other tasks against a given VRF table. This typically applies to single-use commands started from a login shell, as they affect only AF_INET and AF_INET6 sockets opened by the command that gets executed; it has no impact on netlink sockets, associated with the
To execute such a command against a VRF table, run
vrf task exec <vrf-name> <command>. For example, to SSH from the switch to a device accessible through VRF red:
You should manage long-running services with
systemd using the service@vrf notation; for example,
systemctl start ntp@mgmt.
systemd-based services are stopped when a VRF is deleted and started when the VRF is created. For example, restarting networking or running an
Services in VRFs
For services that need to run against a specific VRF, Cumulus Linux uses
systemd instances, where the instance is the VRF. In general, you start a service within a VRF like this:
For example, you can run the NTP service in the blue VRF using:
In most cases, the instance running in the default VRF needs to be stopped before a VRF instance can start. This is because the instance running in the default VRF owns the port across all VRFs — that is, it is VRF global.
systemd-based services are stopped when the VRF is deleted and started when the VRF is created. For example, when you restart networking or run an
ifup sequence — as mentioned above. The management VRF chapter details how to do this.
In Cumulus Linux, the following services work with VRF instances:
There are cases where
systemd instances do not work; you must use a service-specific configuration option instead. For example, you can configure
rsyslogd to send messages to remote systems over a VRF:
FRRouting Operation in a VRF
In Cumulus Linux 3.0 and later, BGP and static routing (IPv4 and IPv6) are supported within a VRF context. Various FRRouting routing constructs, such as routing tables, nexthops, router-id, and related processing are also VRF-aware. However, OSPFv2 and OSPFv3 are not VRF-aware and can only operate in the default VRF.
FRRouting learns of VRFs provisioned on the system as well as interface attachment to a VRF through notifications from the kernel.
You can assign switch ports to each VRF table with an interface-level configuration, and BGP instances can be assigned to the table with a BGP router-level command. Note that OSPFv2 and OSPFv3 are not VRF-aware.
Because BGP is VRF-aware, it supports per-VRF neighbors, both iBGP and eBGP as well as numbered and unnumbered interfaces. Non-interface-based VRF neighbors are bound to the VRF, which is how you can have overlapping address spaces in different VRFs. Each VRF can have its own parameters, such as address families and redistribution. Incoming connections rely on the Linux kernel for VRF-global sockets. BGP neighbors can be tracked using BFD, both for single and multiple hops. You can configure multiple BGP instances, associating each with a VRF.
As mentioned above, VRFs are provisioned through
/etc/network/interfaces. VRFs can be pre-provisioned in FRRouting too, but they become active only when configured through
- A VRF can be pre-provisioned in FRRouting by running the command
- A BGP instance corresponding to a VRF can be pre-provisioned by configuring
router bgp asn vrf vrf-name. Under this context, all existing BGP parameters can be configured - neighbors, peer-groups, address-family configuration, redistribution etc.
- Static routes (IPv4 and IPv6) can be provisioned in a VRF by specifying the VRF along with the static route configuration. For example,
ip route prefix nexthop vrf vrf-name. The VRF has to exist for this configuration to be accepted - either already defined through
/etc/network/interfacesor pre-provisioned in FRRouting.
Here's an example VRF configuration in BGP:
These commands produce the following configuration in the
Example Commands to Show VRF Data
There are a number of ways to interact with VRFs, including NCLU,
vtysh (the FRRouting CLI) and
Showing VRF Data Using NCLU Commands
To show the routes in the VRF:
To show the BGP summary for the VRF:
To show BGP (IPv4) routes in the VRF:
However, to show BGP IPv6 routes in the VRF, you need to use
vtysh, the FRRouting CLI:
Showing VRF Data Using FRRouting Commands
Show all VRFs learned by FRRouting from the kernel. The table ID shows the corresponding routing table in the kernel either automatically assigned or manually defined:
Show VRFs configured in BGP, including the default. A non-zero ID is a VRF that has also been actually provisioned — that is, defined in
Display interfaces known to FRRouting and attached to this VRF:
Showing VRF Data Using ip Commands
To list all VRFs provisioned, showing the VRF ID (vrf1012, vrf1013 and vrf1014 below) as well as the table ID:
To list the interfaces attached to a specific VRF:
To show IPv4 routes in a VRF:
To show IPv6 routes in a VRF:
To see a list of links associated with a particular VRF table,
run ip link list <vrf-name>. For example:
To see a list of routes associated with a particular VRF table, run
ip route list <vrf-name>. For example:
You can also show routes in a VRF using
ip [-6] route show vrf <name>. This command omits local and broadcast routes, which can clutter the output.
Using BGP Unnumbered Interfaces with VRF
BGP unnumbered interface configurations are supported with VRF. In BGP unnumbered, there are no addresses on any interface. However, debugging tools like
traceroute need at least a single IP address per node as the node's source IP address. Typically, this address was assigned to the loopback device. With VRF, you need a loopback device for each VRF table since VRF is based on interfaces, not IP addresses. While Linux does not support multiple loopback devices, it does support the concept of a dummy interface, which is used to achieve the same goal.
An IP address can be associated with the VRF device, which will then act as the dummy (loopback-like) interface for that VRF.
Configure the BGP unnumbered configuration. The BGP unnumbered configuration is the same for a non-VRF, applied under the VRF context (
router bgp asn vrf <vrf-name>).
These commands create the following configuration in the
Here is the FRRouting BGP configuration:
These commands create the following configuration in the
Using DHCP with VRF
Since you can use VRF to bind IPv4 and IPv6 sockets to non-default VRF tables, you have the ability to start DHCP servers and relays in any non-default VRF table using the
dhcrelay services, respectively. These services must be managed by
systemd in order to run in a VRF context; in addition, the services must be listed in
/etc/vrf/systemd.conf. By default, this file already lists these two services, as well as others like
snmpd. You can add more services as needed, such as
dhcrelay6 for IPv6.
If you edit
sudo systemctl daemon-reload to generate the
systemd instance files for the newly added service(s). Then you can start the service in the VRF using
systemctl start <service>@<vrf-name>.service, where
<service> is the name of the service — such as
dhcrelay — and
<vrf-name> is the name of the VRF.
For example, to start the
dhcrelay service after you configured a VRF named blue, run:
To enable the service at boot time you should also run
systemctl enable <service>@<vrf-name>. To continue with the previous example:
In addition, you need to create a separate default file in
/etc/default for every instance of a DHCP server and/or relay in a non-default VRF; this is where you set the server and relay options. To run multiple instances of any of these services, you need a separate file for each instance. The files must be named as follows:
See the example configuration below for more details.
Caveats for DHCP with VRF
- Cumulus Linux does not support DHCP server and relay across VRFs, so the server and host cannot be in different VRF tables. In addition, the server and relay cannot be in different VRF tables.
- Typically a service running in the default VRF owns a port across all VRFs. If the VRF local instance is preferred, the global one may need to be disabled and stopped first.
- VRF is a layer 3 routing feature. It only makes sense to run programs that use AF_INET and AF_INET6 sockets in a VRF. VRF context does not affect any other aspects of the operation of a program.
- This method only works with
In the following example, there is one IPv4 network with a VRF named red and one IPv6 network with a VRF named blue.
The IPv4 DHCP server/relay network looks like this:
The IPv6 DHCP server/relay network looks like this:
Configure each DHCP server and relay as follows:
Sample DHCP Server Configuration
Sample DHCP6 Server Configuration
Sample DHCP Relay Configuration
Sample DHCP6 Relay Configuration
Using ping or traceroute
If you wish to use
traceroute on a VRF, use the
-I <vrf> flag for ping and
-i <vrf> for
Caveats and Errata
- The Penguin Computing Arctica 4804IP switch does not support VRFs.
- Table selection based on the incoming interface only; currently, packet attributes or output-interface-based selection are not available.
- BGP (IPv4/IPv6) is the only routing protocol supported currently. There is no support for OSPF (IPv4/IPv6) at this time.
- Setting the router ID outside of BGP via the
router-idoption causes all BGP instances to get the same router ID. If you want each BGP instance to have its own router ID, specify the
router-idunder the BGP instance using
bgp router-id. If both are specified, the one under the BGP instance overrides the one provided outside BGP.
- You cannot configure EVPN address families within a VRF.