Jumbo packets being dropped in dpdk bonds

[Andre Ruiz]

During a migration of an queens/ovs openstack cloud to yoga/ovn we noticed that jumbo frames stopped working. This cloud is using DPDK so I'm not sure if this affects non-dpdk too.

All three mtu-related parameters in neutron-api charm are set to 9000 the same as before, but it seems like those values are not being respected after the migration.

Small packets (up to exactly 1500 bytes) go through normally and overall network is working well but jumbo packets just get dropped silently, in both directions (VM to outside and outside to VM).

In this specific scenario there are no routers and VMs are being plugged directly to provider networks (they are set to internal and have dhcp enabled) -- I'm mentioning this only to exclude routers from the equation and make it easier to debug.

From a conversation with Frode Nordahl, it seems like in the old scenario the old neutron-openvswitch charm had a relation with neutron-api where it was getting the MTU settings and configuring them in OVS ports, but the new charm ovn-chassis does not have that relation and therefore is not using those settings to configure ovs ports, which stay in default mtu size of 1500.

I'm going to test (and report back) the suggested workaround which is manually setting the MTU on the ports directly in OVS:

ovs-vsctl set Interface dpdk-xxx mtu_request=9000

[Andre Ruiz (andre-ruiz)]

Manually changing the MTU values for all dpdk-bondX interfaces did work and now jumbo packets are going through.

[Frode Nordahl (fnordahl)]

Thank you for the bug report, Andre, and thank you for confirming the workaround.

This does indeed look like an oversight in the initial implementation of DPDK support in the charm, and should be fixed.

As the OVN charms are CMS agnostic (i.e. can be used with other systems than OpenStack), mimicking the behavior of the neutron-openvswitch charm by adding the relation may not be the preferred solution. An alternative might be a DPDK specific MTU configuration option or something similar.

> During a migration of an queens/ovs openstack cloud to yoga/ovn we noticed that jumbo frames stopped working. This cloud is using DPDK so I'm not sure if this affects non-dpdk too.

To answer the question above, setting MTU on interfaces in OVS is only required for DPDK configurations. For regular configurations, configuring MTU is the responsibility of the operator through MAAS and Netplan.

[Andre Ruiz (andre-ruiz)]

Just for documentation:

This one liner will fix it in all compute nodes:

juju run -a nova-compute "for i in \$(sudo ovs-vsctl show | grep 'Interface dpdk' | awk '{print \$2}'); do sudo ovs-vsctl set Interface \$i mtu_request=9000; done"