Flannel is another example of a dual CNI plugin design:

  • Connectivity is taken care of by the flannel binary. This binary is a metaplugin – a plugin that wraps other reference CNI plugins. In the simplest case, it generates a bridge plugin configuration and “delegates” the connectivity setup to it.

  • Reachability is taken care of by the Daemonset running flanneld. Here’s an approximate sequence of actions of what happens when the daemon starts:

    1. It queries the Kubernetes Node API to discover its local PodCIDR and ClusterCIDR. This information is saved in the /run/flannel/subnet.env and is used by the flannel metaplugin to generate the host-local IPAM configuration.
    2. It creates a vxlan interfaces called flannel.1 and updates the Kubernetes Node object with its MAC address (along with its own Node IP).
    3. Using Kubernetes API, it discovers the VXLAN MAC information of other Nodes and builds a local unicast head-end replication (HER) table for its vxlan interface.

This plugin assumes that daemons have a way to exchange information (e.g. VXLAN MAC). Previously, this required a separate database (hosted etcd) which was considered a big disadvantage. The new version of the plugin uses Kubernetes API to store that information in annotations of a Node API object.

The fully converged IP and MAC tables will look like this:


Assuming that the lab is already setup, flannel can be enabled with the following 3 commands:

make flannel

Check that the flannel daemonset has reached the READY state:

$ kubectl -n kube-system get daemonset -l app=flannel
kube-flannel-ds   3         3         3       3            3           <none>          90s

Now we need to “kick” all Pods to restart and pick up the new CNI plugin:

make nuke-all-pods

Here’s how the information from the diagram can be validated (using worker2 as an example):

  1. Pod IP and default route
$ NODE=k8s-guide-worker2 make tshoot
bash-5.0# ip route get 1.1 via dev eth0 src uid 0 
  1. Node routing table
$ docker exec -it k8s-guide-worker2 ip route
default via dev eth0 via dev flannel.1 onlink via dev flannel.1 onlink dev cni0 proto kernel scope link src dev eth0 proto kernel scope link src 
  1. Static ARP entries for NextHops
$ docker exec -it k8s-guide-worker2 ip neigh | grep PERM dev flannel.1 lladdr ce:0a:4f:22:a4:2a PERMANENT dev flannel.1 lladdr 5a:11:99:ab:8c:22 PERMANENT

  1. VXLAN forwarding database
$ docker exec -it k8s-guide-worker2 bridge fdb show dev flannel.1 
5a:11:99:ab:8c:22 dst self permanent
ce:0a:4f:22:a4:2a dst self permanent

A day in the life of a Packet

Let’s track what happens when Pod-1 tries to talk to Pod-3.

We’ll assume that the ARP and MAC tables are converged and fully populated.

  1. Pod-1 wants to send a packet to Its network stack looks up the routing table to find the NextHop IP:
$ kubectl exec -it net-tshoot-4sg7g -- ip route get via dev eth0 src uid 0
  1. The packet reaches the cbr0 bridge in the root network namespace, where the lookup is performed again:
$ docker exec -it k8s-guide-worker ip route get via dev flannel.1 src uid 0 
  1. The NextHop and the outgoing interfaces are set, the ARP table lookup returns the static entry provisioned by the flanneld:
$ docker exec -it k8s-guide-worker ip neigh get dev flannel.1 dev flannel.1 lladdr 5a:11:99:ab:8c:22 PERMANENT
  1. Next, the FDB of the VXLAN interface is consulted to find out the destination VTEP IP:
$ docker exec -it k8s-guide-worker bridge fdb | grep 5a:11:99:ab:8c:22           
5a:11:99:ab:8c:22 dev flannel.1 dst self permanent
  1. The packet is VXLAN-encapsulated and sent to the control-node where flannel.1 matches the VNI and the VXLAN MAC:
$ docker exec -it k8s-guide-control-plane ip link show flannel.1
3: flannel.1: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1450 qdisc noqueue state UNKNOWN mode DEFAULT group default 
    link/ether 5a:11:99:ab:8c:22 brd ff:ff:ff:ff:ff:ff
  1. The packet gets decapsulated and its original destination IP looked up in the main routing table:
$ docker exec -it k8s-guide-control-plane ip route get   dev cni0 src uid 0 
  1. The ARP and bridge tables are than consulted to find the outgoing veth interface:
$ docker exec -it k8s-guide-control-plane ip neigh get dev cni0 dev cni0 lladdr 7e:46:23:43:6f:ec REACHABLE
$ docker exec -it k8s-guide-control-plane bridge fdb get 7e:46:23:43:6f:ec br cni0 
7e:46:23:43:6f:ec dev vethaabf9eb2 master cni0 
  1. Finally, the packet arrives in the Pod-3’s network namespace where it gets processed by the local network stack:
$ kubectl exec -it net-tshoot-rkg46 -- ip route get
local dev lo src uid 0

SNAT functionality

Similar to kindnet flanneld sets up the SNAT rules to enable egress connectivity for the Pods, the only difference is that it does this directly inside the POSTROUTING chain:

Chain POSTROUTING (policy ACCEPT 327 packets, 20536 bytes)
 pkts bytes target     prot opt in     out     source               destination         
    0     0 RETURN     all  --  *      *       
    0     0 MASQUERADE  all  --  *      *       !          random-fully
    0     0 RETURN     all  --  *      *      !       
    0     0 MASQUERADE  all  --  *      *      !        random-fully

Caveats and Gotchas

  • The official installation manifest does not install the CNI binary by default. This binary is distributed as a part of reference CNI plugins and needs to be installed separately.
  • flannel can run in a direct routing mode, which acts by installing static routes for hosts on the same subnet.
  • flannel can use generic UDP encapsulation instead of VXLAN