-
Bug
-
Resolution: Done
-
Normal
-
None
-
None
The GW being configured over RAs on the workloads is featuring the link-local address of the LRP; when the VM live-migrates to another node, it'll learn a new GW route (since the link-local IP addr of the LRP in the destination node is different.
We need to inform the workload of two things:
- learn the GW information (link-local addr of the LRP for the network on the DST node)
- forget the old GW information (link-local addr of the LRP for the network on the SRC node)
This can be achieved by sending unsolicited RAs from a RAW socket (by ovnkube-node). Making the workload "forget" the old GW can be achieved by issuing a RA with a lifetime of 0.
NOTE:
- this requires an e2e test asserting connectivity to a different network to be added to the kubevirt suite (upstream)
- this solution assumes the link-local addresses are generated from the MAC addresses of the ports - privacy extensions are not supported
- incorporates
-
CORENET-5446 [ovn-kubernetes, migration improvement] remove the old default route when the VM migrates to another node
-
- Closed
-
- is related to
-
CNV-60575 IPv6 does not work properly on dual-stack clusters with layer primary UDN
-
- Closed
-
- split from
-
CORENET-5476 [ovn-kubernetes, IPAM] advertise LSP addresses via DHCP[v6] and prefix via RAs (for IPv6)
-
- Closed
-
- links to