mirror of
https://github.com/rancher/k3k.git
synced 2026-08-19 04:16:16 +00:00
Enrico Candino
d422898d55
Both QEMU workers booted with `-net user` and ended up registering the same InternalIP (10.0.2.15) because each VM gets its own isolated NAT slirp. Flannel propagated this to `public-ip` on both nodes, so VXLAN could not tunnel between workers and any cross-node pod traffic broke (89 failed / 335 passed of 424 conformance specs). Replace user-mode networking with a Linux bridge (k3kbr0, 192.168.100.0/24) and one TAP device per VM, so the two workers share an L2 segment with unique routable IPs. NAT outbound from the bridge keeps internet access working for image pulls. Also set unique hostnames via cloud-init (worker-1/worker-2) and drop the `--node-name` flag from INSTALL_K3S_EXEC, since k3s now picks the correct node name from the OS hostname on its own. Bump hydrophone back to `--parallel 4` to match the single-VM job (parallelism was reduced earlier when the failure was thought to be resource-related).