跳至主要内容

Stretching a PVE LAN to QEMU guests on OCI

This pattern carries one deliberately selected Ethernet broadcast domain from a PVE bridge to QEMU guests on an OCI Compute instance. The OCI VCN remains an L3 transport. It is never joined to the stretched bridge.

Topology​

PVE vmbr0 -- lan0 [routerd VM] wg0 -- VXLAN/200001 -- wg0 [routerd QEMU guest] lan0
                    |                                      |
                  br-l2                                  br-l2
                                                           |
                                              OCI host br-legacy -- Windows taps

OCI VCN VNIC -- OCI host routed/NAT transit -- routerd guest underlay NIC

The OCI routerd appliance is itself a QEMU guest. On VM.Standard.E5.Flex it must initially be treated as a TCG guest unless /dev/kvm is positively verified. Give it two NICs:

  1. an underlay/management NIC on an internal host transit network; and
  2. an unnumbered LAN NIC attached to the host-only br-legacy bridge.

Attach every legacy Windows tap to br-legacy. Never attach the OCI VCN VNIC to that bridge. If the routerd guest uses a private transit address, forward only its WireGuard UDP port on the OCI host and keep VXLAN/UDP 4789 inside WireGuard.

routerd overlay resource​

Use the same VNI and UDP port at both endpoints, exchanging the two WireGuard addresses. The LAN-facing NIC and VXLAN device must share one bridge.

apiVersion: net.routerd.net/v1alpha1
kind: Bridge
metadata:
  name: legacy-l2
spec:
  ifname: br-l2
  members:
    - lan0
  stp: true
  multicastSnooping: false
---
apiVersion: net.routerd.net/v1alpha1
kind: VXLANTunnel
metadata:
  name: legacy-l2
spec:
  ifname: vx-l2
  vni: 200001
  localAddress: 10.254.200.1
  peers:
    - 10.254.200.2
  underlayInterface: wg-l2
  udpPort: 4789
  mtu: 1370
  bridge: br-l2

Swap .1 and .2 on the OCI endpoint. VXLANTunnel installs an all-zero MAC FDB entry per peer, so broadcast and unknown-unicast frames are replicated over the unicast underlay. It does not apply the default VXLANSegment control-plane filter. ARP, DHCPv4, IPv6 RS/RA/NS/NA, and DHCPv6 therefore cross the bridge.

Safety gates​

  • First run sudo tests/netns/vxlan-l2-control-plane-transparency.sh on a disposable Linux host.
  • Next connect both appliances to isolated PVE/QEMU bridges and capture both sides. Do not use vmbr0 yet.
  • Reserve one guest IP and MAC with the network administrator before the final attachment to vmbr0.
  • The routerd appliances must not run DHCP, DNS, or RA services on br-l2.
  • Full transparency is symmetric: a guest could emit DHCP Offers or RAs. Keep the Windows guests trusted and offline until a directional bridge-firewall policy has separately passed positive and negative tests.
  • Verify PVE firewall/MAC filtering and the physical switch allow the remote guest MAC. Watch for MAC flaps, duplicate IPs, and broadcast storms.

Evidence​

Record the routerd commit, rendered plan, ip -d link, bridge link, bridge fdb, WireGuard handshake, VXLAN packet capture, protocol test output, MTU probes, and rollback timestamps. Rollback starts by disconnecting the PVE appliance LAN NIC from vmbr0; then stop the VXLAN and WireGuard resources.