Routed is VMStack's default network mode. Servers sit behind the node on the host-only bridge vmbr0; the node routes each of their addresses and answers for them on its uplink with its own MAC. Nothing on the uplink changes, and it works at providers that allow only the node's MAC on the port.
The blocks themselves are added in Infrastructure › IP pools; see IP pools. For servers that need their own MAC on the provider's network, see Bridged networking.
How routed networking works#
- Every IPv4 address of a server is a
/32route onvmbr0(IPv6: a/128, or the server's routed prefix). - On an on-link block, the node answers ARP on its uplink for its servers' addresses (proxy ARP) and IPv6 neighbour solicitations (NDP proxy). A free address has no route, so the node does not answer for it.
- Each server's card is bound to libvirt's
clean-trafficfilter with its own MAC and addresses (anti-spoofing). The bridge ports are isolated: servers reach each other through the node, not directly. - The server is configured statically by its first-boot seed (cloud-init, or the Windows equivalent). A DHCP server on the node also answers each server with its main address, for systems installed from an ISO.
Gateway inside the block#
Use the Gateway inside the block (OVH vRack, or providers that give an on-link gateway) type when the provider's gateway is an address of the block and answers on the node's link: a colocation range, a /27 on a shared segment, an OVHcloud vRack block.
Servers get each address with the block's prefix and gateway. A single address (/31 or /32) cannot be this type.
Provider IPs routed to the main IP#
Many providers sell extra addresses that are routed to the server's main IP, with no usable gateway in them. Use the type Routed to this node's main IP (OVH/Hetzner additional IPs and blocks) for them.
- Go to Infrastructure › IP pools › Add IP block, choose IPv4, and choose the Node.
- Choose Routed to this node's main IP.
- Enter the Network: a block such as
203.0.113.8/29, or a single address such as203.0.113.50. - Check the Gateway: it is filled in with the node's own default gateway. It must be outside the block.
- Click Add IPv4 block.
Every address of a routed block is usable, network and broadcast included (a /29 gives 8 addresses, not 5). Narrow it with First address and Last address if the provider asks you to keep some free.
OVHcloud additional IPs#
An OVHcloud additional IP (a single address or a block) in routed mode is routed to the dedicated server's main IP. Add it as Routed, with the node's own gateway (the .254 of the node's /24 at OVHcloud). Do not use the gateway OVHcloud lists inside the block: that only applies in a vRack or Private Cloud. OVHcloud's router answers ARP for almost any address, so it tells nothing about which type to choose.
If you want the addresses to leave with their own virtual MAC instead, use a bridged block.
Hetzner additional IPs and subnets#
Hetzner's additional single IPs and additional subnets routed to the main IP are the same shape: add them as Routed, with the node's gateway. A Hetzner single IP ordered with a separate MAC is a bridged block with Entered per address. A Hetzner vSwitch subnet is a VLAN.
What the server gets#
Each address of a routed block is configured inside the server as a /32, with a default route via the provider's gateway on-link. The server sends to the gateway's address; the node answers its ARP on vmbr0 and forwards the traffic with its own MAC. The provider sees only the node's MAC.
- The seed writes
on-link: truefor the gateway (netplan needs it). On AlmaLinux, Rocky Linux and Fedora,cloud-init statusmay reportdegraded donebecause of that key; the network works. - DHCP gives the address with netmask
255.255.255.255and the classless static route option 121 (<gateway>/32on-link, then the default route via the gateway). - Routed blocks take no fixed MAC and are not checked by the address check: nothing else on the uplink can hold their addresses.
IPv6#
Add an IPv6 block for the prefix the provider gives the node; see IP pools. Plans give 0 to 16 IPv6 addresses per server (IPv6 addresses) and optionally IPv6 prefix per server (64 to 124, from a routed block).
- On the node's link: the prefix is on the uplink's segment (the usual /64). Servers get single addresses; the node answers neighbour solicitations for each.
- Routed to the node: the provider routes the prefix to the node. Servers can also get a subnet each. Routed IPv6 prefixes are built but have not yet been tested on a real node.
Servers use fe80::1, the node's address on their bridge, as their IPv6 gateway, whatever the provider's gateway is. Their seed sets accept-ra: false. A server's own router advertisements and redirects are dropped by the agent's IPv6 filter.
What the agent changes on the node#
Nothing about IPv6 is touched until the first server with IPv6 is created or given an address. Then the agent:
- sets
net.ipv6.conf.<uplink>.accept_ra = 2before turning forwarding on, so the node keeps learning its own default route from router advertisements; - turns on
net.ipv6.conf.all.forwarding, the uplink'sforwardingandproxy_ndp; - checks the node still has its IPv6 default route, and puts forwarding back if not;
- writes the settings to
/etc/sysctl.d/95-vmstack.confso they survive a reboot; - makes sure the uplink has a link-local address, and on a systemd-networkd uplink writes a drop-in
50-vmstack.conf(withIPv6ProxyNDP=yesand, when networkd takes the router advertisements,IPv6AcceptRA=yes).
Every minute it puts back what a network reload or a link flap removed: the link-local address, forwarding, proxy_ndp and the servers' NDP proxy entries.
A node whose router advertisements are handled by another network manager (NetworkManager, for example) cannot route IPv6 to servers: the IPv6 block is refused with the reason. Give the node a static IPv6 address and gateway, or let the kernel take the advertisements.
The node page shows IPv6 forwarding in the readiness check.
The gateway check#
When you enter The node's gateway (from the provider), the block's page compares it with the IPv6 default route the node reports. When they differ, the page gives the command to run on the node, for example:
ip -6 route replace default via 2001:db8:100:200::1 dev eth0 onlink
Put the same route in the node's own network configuration so it survives a reboot.
IPv6-only servers#
A server with no IPv4 address and at least one IPv6 address is IPv6-only. To offer one, set a plan's Network to Public: addresses of its own, IPv4 addresses to 0 and IPv6 addresses to 1 or more (in Customers › Hosting plans), or choose the same counts for a custom size in the create form.
- The server's network configuration has no IPv4 at all, and the DHCP server has no IPv4 entry for it.
- It cannot reach IPv4-only sites. The create form says so.
- Images that fetch packages over IPv4 during the first boot may finish it with warnings.
- An IPv6-only server cannot be given IPv4 addresses later, and keeps its last address.
DHCP for routed servers#
Every node runs vmstack-dhcp.service, a dnsmasq (from the dnsmasq-base package, no DNS service) that answers only the node's servers, each by MAC, with its main address. Leases are infinite. Additional addresses are not given by DHCP; the seed configures them.
The node page shows DHCP server (running, for how many servers), and the readiness check has a "DHCP server" line. If it fails, check the service:
systemctl status vmstack-dhcp
journalctl -u vmstack-dhcp -n 50
Servers with IPv6 also get router advertisements (from fe80::1) and one DHCPv6 address. This needs dnsmasq 2.81 or later built with DHCPv6; otherwise the node page says why.
Limits#
- A block belongs to one node. Moving a block between nodes is not supported; delete it and add it to the other node.
- A routed or bridged block's addresses must not be on the node's own uplink network.
- The address limit counts public IPv4 only. There is no IPv6 limit.
- Port isolation means servers on the same node reach each other only through the node, by their public addresses.