Persistent network configuration

ip, and the commands it replaced

The iproute2 tools are the current ones. ip addr show lists addresses, ip addr add 192.168.1.10/24 dev eth0 assigns one, ip link set eth0 up brings an interface up, and ip route shows or sets routing — ip route add default via 192.168.1.1 sets a default gateway. Their predecessors, ifconfig, route and netstat, are deprecated but still appear in questions and still installed on plenty of machines; the equivalences are worth knowing in both directions. ss has replaced netstat for listing sockets: ss -tulpn shows listening TCP and UDP sockets with the process behind each.

Everything set with ip is lost at reboot. Persisting it depends on the family: Debian traditionally uses /etc/network/interfaces, Red Hat uses ifcfg files under /etc/sysconfig/network-scripts/, and both are increasingly replaced by NetworkManager (nmcli on the command line) or systemd-networkd. DHCP is the alternative to configuring anything, with a client requesting an address, mask, gateway and DNS servers from a server on the network.

The Debian file is worth being able to read and write, because it is short enough to be quoted whole in a question. It is a series of two-line stanzas: auto eth0 says bring this interface up at boot, and iface eth0 inet dhcp says how — inet for IPv4, then dhcp or static. A static stanza indents address, netmask and gateway lines underneath it. allow-hotplug is the alternative to auto and means "when the kernel says this interface appeared" rather than "at boot", which is what you want for a cable that may not be plugged in. Miss the auto line and the stanza is correct and never applied — ifup eth0 by hand would still work, and the machine would come up with no address.

Editing that file changes nothing on its own, and the way to apply it is the part candidates get wrong. ifup and ifdown are wrappers that read the stanzas and record which interfaces they have configured, historically under /run/network, and ifup refuses to touch an interface it believes is already up. So a changed address is applied by cycling the interface through the wrappers — ifdown eth0 && ifup eth0 — and NOT by ip link set eth0 down and up again, which toggles the link without ever re-reading the file. ifquery eth0 prints what ifup would apply, which is the safe way to check a stanza before committing to it.

The hostname is separate from any of this. hostname prints the running one and setting it that way lasts until reboot; the persistent name is in /etc/hostname. On a systemd machine hostnamectl set-hostname myserver does both at once — it writes the file and changes the running system — which is why it is the answer to a question that asks for a change that survives a reboot without saying which file.

# in /etc/network/interfacesauto eth0iface eth0 inet dhcp# or, for a fixed addressauto eth0iface eth0 inet static    address 192.168.1.10/24    gateway 192.168.1.1$ hostnamectl set-hostname myserver$ cat /etc/hostnamemyserver

systemd-networkd, the third place persistence can live

On a minimal server with neither ifupdown nor NetworkManager, the configuration is systemd-networkd's, and it reads three KINDS of file from /etc/systemd/network/, told apart by their extension: .link for udev-level device properties such as the name or MAC address, .netdev to CREATE a virtual device such as a bridge or a VLAN, and .network to configure addressing on an interface that already exists. Files are processed in lexicographic order of filename, which is why a numeric prefix is conventional, and the first .network file whose [Match] section matches an interface is the one applied.

So a static address is /etc/systemd/network/10-wired.network holding a [Match] section with Name=enp1s0 and a [Network] section with Address=10.0.0.10/24 and Gateway=10.0.0.1. For DHCP the [Network] section takes DHCP=yes in place of those two lines. Getting the extension wrong is the trap the distractors are built on: the same content in a .netdev or a .link file configures nothing, because neither of those is the file that assigns addresses.

networkctl is the client for the daemon, in the way nmcli is the client for NetworkManager. networkctl list gives the per-link summary and, crucially, says whether networkd has CONFIGURED each link or is leaving it unmanaged — a link showing unmanaged has no matching .network file. ip link show cannot answer that, because it reports the kernel's view of the interface and not who is looking after it. networkctl status enp1s0 expands one link with its addresses, routes and the .network file that matched, and networkctl reload picks up edited files.

Turning a name into an address

/etc/hosts is a static table of addresses and names, and it is consulted before DNS on a normally configured system — which makes it both the quick fix for a machine with no DNS entry and the reason a stale line there can override a correct DNS record. /etc/resolv.conf names the DNS servers to ask, with nameserver lines and a search list for unqualified names; on modern systems it is often generated by NetworkManager or systemd-resolved rather than edited.

/etc/nsswitch.conf is what decides the ORDER — the hosts line typically reads files dns, meaning /etc/hosts first and DNS second, and it is the file to point at when a question asks how the resolution order is configured. host and dig query DNS directly and therefore ignore /etc/hosts entirely, while getent hosts goes through the full name service and does not. That difference is the fastest way to prove which layer is answering.