Skip to content

Latest commit

 

History

History
122 lines (86 loc) · 3.81 KB

debugging.md

File metadata and controls

122 lines (86 loc) · 3.81 KB

Initial setup issues.

Each host should have unique system-ids.

In a OVN cluster, each host should have a unique system-id. This value is set in external-ids:system-id of the Open_vSwitch table of each host's Open_vSwitch database. You can fetch its value with:

ovs-vsctl get Open_vSwitch . external-ids:system-id

This is automatically set by OVS startup scripts (that come with packages). But if you installed OVS from source, and don't use a startup script, this will be empty. If you clone your VM, then you may endup with same system-id for all your VMs - which is a problem.

All nodes should register with OVN SB database.

On the master, run:

ovn-sbctl list chassis

You should see as many records as the number of nodes in your cluster. The "name" column in each record represents the system-id of each host and they should all be unique.

Hosts should have unique node names.

When you run 'ovnkube -init-node' or 'ovnkube -init-master' commands, it will ask for node names. This should all be unique. This should also be the same as used by "kubelet" on each node. So the names you see when you run the below command should be the same ones that you supply to ovnkube:

kubectl get nodes

Check the presence of "geneve" or "stt" kernel module.

If you cannot ping between pods in 2 nodes, make sure to check whether you have the OVS kernel modules correctly installed by running:

lsmod | grep openvswitch
lsmod | grep geneve  # Or "stt" if your encap type was chosen to be "stt".

If you do not see the kernel module, modprobe them

modprobe openvswitch
modprobe vport-geneve  # or vport-stt for "stt".

Sanity check cross host ping.

Since on each host, a OVS internal device named "k8s-$NODENAME" is created with the second IP address in the subnet assigned for that node, you should be able to check cross host connectivity by pinging these internal devices across hosts.

Firewall blocking overlay networks

If your host has a firewall blocking incoming connections as a default policy, you should open up ports to allow overlay networks. If you use "geneve" as the encapsulation type, you should open up UDP port 6081. If you use "stt" as the encapsulation type, you should open up TCP port 7471.

You can use the following command to achieve it via iptables.

# To open up Geneve port.
/usr/share/openvswitch/scripts/ovs-ctl --protocol=udp \
        --dport=6081 enable-protocol

# To open up STT port.
/usr/share/openvswitch/scripts/ovs-ctl --protocol=tcp \
        --dport=7471 enable-protocol

Check ovn-northd's log file.

On the master, look at /var/log/openvswitch/ovn-northd.log to see for any errors with the setup of the OVN central node.

Runtime issues

Check the watcher's log file.

On the master, check whether ovnkube is running by:

ps -ef | grep ovnkube

Check the watcher's log file at "/var/log/ovn-kubernetes/ovnkube.log" to see whether it is creating logical ports whenever a pod is created and for any obvious errors.

Check the OVN CNI log file.

When you create a pod and it gets scheduled on a particular host, the OVN CNI plugin on that host, tries to access the pod's information from the K8s api server. Specifically, it tries to get the IP address and mac address for that pod. This information is logged in the OVN CNI log file on each node if you have specified a log file via "/etc/openvswitch/ovn_k8s.conf". You can read how to provide a logfile by reading 'man ovn_k8s.conf.5'.

Check the kubelet's log file.

If there were any issues with downloading upstream CNI plugins, then kubelet will complain about them in its log file.

Check ovn-controller's log file.

If you suspect issues on only one of the host, look at the log file of ovn-controller at /var/log/openvswitch/ovn-controller.log to see any obvious error messages.