ToDo: the plan is to turn this into something that the Release Notes can point at as the nearest thing we've got to an official guide. What if anything does this page need to say about NetworkManager and systemd-networkd?
Anything that changes the names of your network interfaces may result in the machine suddenly not being reachable over SSH, so if you're editing settings on a remote server, plan your changes carefully and doublecheck your safety nets.
- THE ORIGINAL SIMPLE SCHEME
- THE "PERSISTENT NAMES" SCHEME
- MISCELLANEOUS OLD SCHEMES
- THE NEW APPROACH TO NETWORK INTERFACE NAMES
- ALTERNATIVE SCHEMES
- EXTERNAL REFERENCES
THE ORIGINAL SIMPLE SCHEME
Back in the nineties, eth0, eth1, etc were simply assigned by the kernel.
Why it was abandoned
At least in theory, if module probes completed in a different order, eth0 and eth1 might switch places on successive boots. As boot processes became less linear and interfaces became more hotpluggable this became more of a concern.
How to get it back
If you wipe out all other mechanisms (booting with the net.ifnames mechanism deactivated and no /etc/udev/rules.d/75-persistent-net.rules file, or maybe just a udev that has finally stopped supporting it) then presumably you'll be back to this one.
THE "PERSISTENT NAMES" SCHEME
This scheme, introduced somewhere around Debian 5 "Lenny", used udev to identify interfaces by MAC address and assign a fixed interface number to any interface it recognized (writing the rules to /etc/udev/rules.d/75-persistent-net.rules). This could have annoying side-effects (e.g. if you were replacing a machine's sole NIC, you'd also have to take special care to ensure it took over as network interface number 0) but these were minor and predictable.
Why this one was abandoned
This still had subtle race conditions, required /etc to be on a writable file system, and had problems with virtualization, so it's no longer supported upstream; the plan (still taken for granted in most of the documentation) was for it not to be supported in Debian 10 "Buster", but in the end it has scraped through into another release cycle.
How to cling to it for now
First, make sure you've got a working "legacy" /etc/udev/rules.d/75-persistent-net.rules file. Then deactivate the replacement scheme. The simple way of doing that latter (which you might want to try for one-off testing) is just to boot with the kernel parameter net.ifnames=0, which can be set in an interactive grub session at boot or made persistent by editing /etc/default/grub and running update-grub.
Alternatively, you can override /lib/systemd/network/99-default.link, with a custom version in /etc/systemd/network/, or similarly override /lib/udev/rules.d/80-net-setup-link.rules, or mask the latter by using a /dev/null symlink instead of a custom version, or... there seem to be lots of ways of doing this, so make sure you haven't done it in more than one way or it'll trip you up in a couple of years when you try to undo it. Oh, and beware of initrd skew.
Unfortunately, none of this will carry on working beyond the point when udev stops supporting the legacy 75-persistent-net.rules mechanism, so you should be ready to switch to a different system when that happens.
How to let go and move on
If you've currently got a legacy /etc/udev/rules.d/75-persistent-net.rules file but have decided to switch to the new regime, you can do that just by disabling the .rules file; see the udev README.Debian.gz and the more detailed guide below.
MISCELLANEOUS OLD SCHEMES
Several workarounds for renaming interfaces grew up in the early days of hotpluggable wireless interfaces, but if they still work it'll be because like ifrename they now use udev rules under the hood. It's not clear what remaining advantage this has over the canonical .link approach - is it perhaps useful for non-systemd machines?
Old releases of RedHat (among others) used a "biosdevname" system, but that's never been supported under Debian. If you need to know about it there's bound to be documentation somewhere.
THE NEW APPROACH TO NETWORK INTERFACE NAMES
The new "net.ifnames" approach uses names usually derived from the location of the interface in terms of hardware buses etc: eno1, wlp1s3...
(This system is often advertised as providing "Predictable Names", but the main thing that's predictable about it is that calling it this will cause furious users to pop up disputing the appropriateness of that name. Can we just skip all that here, please?)
How to cope with it on fresh installs
This should be easy enough; before you start configuring firewalls etc., just look at (e.g.) the output of ip a.
Unlike the old days, when the only way to guess which cable was plugged into eth0 and which was eth1 was to keep track of MAC addresses, this system provides extra clues in the interface names.
How to switch to this scheme on upgraded systems
It's advisable to do this as a separate thing in its own right, not as part of a general dist-upgrade. However, if your PC only has one network interface and not much is at stake you can try:
wait until udev breaks your networking, if it's going to, or trigger the change yourself by disabling the persistent-names system.
ask ip what new name it's using, and fix your config files.
You should probably at least check in advance to see what files hard-code interface names, by running something like
sudo rgrep wlan0 /etc
(Obvious possibilities include /etc/network/interfaces and configuration files for firewalls, wifi, DHCP...)
This strategy, more or less compulsory for remote servers, runs along the lines of:
- devise a migration plan
consider a dead-man's-handle etckeeper cronjob or the like
consult the list of corner cases below and search the Internet for others
- once you're sure you're safe, implement your migration plan
- you've got backups, right?
To find out what names udev would be choosing between if you switched over to the new system, first get a list of the network devices the system knows about:
For each device path (other than /sys/class/net/lo), ask udevadm what NET_IDs it knows:
udevadm test-builtin net_id /sys/class/net/enp0s1 2>/dev/null
It's likely to tell you about things like ID_OUI_FROM_DATABASE and an ID_NET_NAMING_SCHEME, but the lines that matter are the ones (given in unhelpfully random order) starting with ID_NET_NAME_. One of these is the name that udev will give priority to - the list of candidates may be so short that all you need to know is that ..._PATH beats ..._MAC, but there are also some rarer possibilities, and in general if something unusual shows up then it will take priority.
From highest priority to lowest, the list is:
ID_NET_NAME_FROM_DATABASE= Very rare (possibly mythical) and not to be confused with ID_OUI_FROM_DATABASE; if present, it gets maximum priority. Nobody seems to know why - it's mentioned in systemd.link(5), but otherwise undocumented. The database is hardcoded into udev and has only one known entry, the spooky-sounding idrac.
ID_NET_NAME_ONBOARD= Appears for some but not all kinds of onboard network card (Ethernet only, or wireless too?) - it's usually a nice simple name like eno0.
ID_NET_NAME_SLOT= Appears for some PCI-hotplug cards. Usually looks like ens0 (again, any wifi cases? Does it ever occur along with _ONBOARD?)
ID_NET_NAME_PATH= Always present; usually something just complicated enough to be easy to forget, like wlp3s5 or enp1s3f0. Note that all numbers are in hex.
ID_NET_NAME_MAC= Also always present, but with a low enough priority that by default it won't be used; e.g. wlx800e1319c734
Complications and corner cases
(Additions welcome, but please try to avoid ballooning this section with tales of "I don't know how this happened but it all went wrong for me"...)
- * ANTIQUE SYSTEMS
on Debian 9 "Stretch" or newer, merely booting without a net.ifnames=0 override (and without a 70-persistent-net.rules file) should be enough to let you to run the new scheme, but on Debian 8 "Jessie" or older you'll need to actively set it to net.ifnames=1.
- * TRIGGERING THE SWITCH
you don't need to delete /etc/udev/rules.d/70-persistent-net.rules to test what network interface names you'll get without it. Just renaming it (e.g. to 70-persistent-net.rules.old) or commenting out particular lines should be enough (as long as you update your initrd before you reboot). See the udev README.Debian.gz file.
- * THE ID_NET_NAME_ HIERARCHY
- if you're ignoring ID_NET_NAME_SOMETHING on the assumption that anything you don't understand probably isn't important, you need to reread the above - the general rule is, if you don't recognise it, it'll mess you up.
- * INITRD SKEW
it's all very well having everything sorted out in /etc, but interface renaming has to happen very early during boot; to make sure your initrd doesn't contain out-of-date versions of important systemd files, regenerate it with sudo update-initramfs -u
- * LEFTOVERS
the old persistent-names system first started being publicly deprecated in NEWS files back in mid-2015, so this upgrade has been hanging ominously over people's heads for a long time. Are you sure you didn't do something about it the last time the subject came up, like setting up a net.ifnames=0 kernel parameter, and/or masking some systemd config file? If so, this may result in confusing symptoms when you try to go over to the new system. Check your administrative logbooks. What do you mean, you don't keep logs?
- * USB DEVICES
since they might get plugged into a different socket each time, these use ID_NET_NAME_MAC - automated via /lib/udev/rules.d/73-usb-net-by-mac.rules.
- * VIRTUAL MACHINES
on virtual machines (according to the udev README) you will need to remove the files /etc/systemd/network/99-default.link and (if using virtio network devices) /etc/systemd/network/50-virtio-kernel-names.link, then rebuild the initrd.
- * UNPREDICTABILITY
- it turns out even after all this there are still reported cases of interfaces changing their name on a reboot. All that needs to happen is that some buggy BIOS (or some new, less buggy version of a driver module, or systemd's naming policy) changes its mind about some detail like whether or not your hardware counts as the kind that should have an ONBOARD name.
The above is the new standard default scheme, but there's also a canonical way of overriding the default: you can use .link files to set up naming policies to suit your needs. Thus for instance if you have two PCs each of which has only a single wireless card, but one calls it wlp0s1 and the other wlp1s0, you can arrange for them both to use the name wifi0 to simplify sharing firewall configurations. For details see systemd.link(5).
It's also possible to reorganize the naming policy by overriding the NamePolicy defined in /lib/systemd/network/99-default.link (so that for instance all network cards are named by MAC address).
The nearest thing to a canonical upstream migration-HOWTO used to be https://www.freedesktop.org/wiki/Software/systemd/PredictableNetworkInterfaceNames/. The big problem with this was that it delegated all its technical details to a link pointing at the sourcecode:https://github.com/systemd/systemd/blob/master/src/udev/udev-builtin-net_id.c ...but most of the useful comments that used to be at the top of that file were then thrown out, so you need to find your way back through the git tree to a previous version such as https://github.com/systemd/systemd/blob/eefe36e64c1a583bb9470884ed92115e0ce4647e/src/udev/udev-builtin-net_id.c. Meanwhile, the page now claims to be obsolete, and points instead to https://www.freedesktop.org/software/systemd/man/systemd.net-naming-scheme.html (which has much less useful information). So the nearest thing to an official HOWTO is probably /usr/share/doc/udev/README.Debian.gz (though it doesn't cover the "how to predict the names" part at all).
A guide that mentions ID_NET_NAME_FROM_DATABASE:https://major.io/2015/08/21/understanding-systemds-predictable-network-device-names/
General guides to overriding systemd configuration: https://askubuntu.com/questions/659267/how-do-i-override-or-configure-systemd-services, https://wiki.archlinux.org/index.php/systemd
keywords persistent, predictable, NIC, wlan, eth