mx-networkd: replace mxnetctl/mxvlanctl with a systemd-networkd generator

.gitignore

@@ -1,2 +1,3 @@
 setuid/setuid
 prun/*.1
+mx-networkd/target/

Makefile

@@ -66,10 +66,14 @@ INSTALL_DATA = $(INSTALL) -m 644
 
 manpages += prun/prun.1 prun/pman.1 prun/ptype.1
 
-all: $(manpages)
+CARGO ?= cargo
+mx_networkd = mx-networkd/target/release/mx-networkd
+
+all: $(manpages) $(mx_networkd)
 
 clean:
 	@rm -f $(manpages)
+	@rm -rf mx-networkd/target
 
 install: all
 	@prefix="$(prefix)" usr_prefix="$(usr_prefix)" usrlocal_prefix="$(usrlocal_prefix)" \
@@ -89,3 +93,8 @@ install: all
 %.1: %.md
 	@if [ "$$UID" = 0 ]; then echo "Please build (\`make\`) as non-root before running \`make install\` as root" >&2;exit 1;fi
 	pandoc --standalone --to man $< -o $@
+
+$(mx_networkd): mx-networkd/Cargo.toml $(wildcard mx-networkd/src/*.rs)
+	@if [ "$$UID" = 0 ]; then echo "Please build (\`make\`) as non-root before running \`make install\` as root" >&2;exit 1;fi
+	@command -v $(CARGO) >/dev/null || { echo "$(CARGO) not found - it is needed to build mx-networkd" >&2; exit 1; }
+	$(CARGO) build --release --manifest-path mx-networkd/Cargo.toml

etc/systemd/system/mxvlan.service

@@ -3,6 +3,10 @@ DefaultDependencies=no
 After=sysinit.target
 After=network.service
 Before=network.target
+# Machines carrying the hostconfig tag mx-network-generator are configured
+# by systemd-networkd, fed by the mx-networkd generator instead; see
+# mx-networkd/README.md, section Migration.
+ConditionPathExists=!/node/tags/mx-network-generator
 
 [Service]
 Type=oneshot

etc/systemd/system/network.service

@@ -3,6 +3,10 @@ DefaultDependencies=no
 After=sysinit.target
 Before=network.target
 Wants=network.target
+# Machines carrying the hostconfig tag mx-network-generator are configured
+# by systemd-networkd, fed by the mx-networkd generator instead; see
+# mx-networkd/README.md, section Migration.
+ConditionPathExists=!/node/tags/mx-network-generator
 
 [Service]
 EnvironmentFile=/etc/local/mxhost.conf

etc/systemd/system/systemd-networkd.service.d/mx-network-generator.conf

@@ -0,0 +1,8 @@
+# Run systemd-networkd only on machines that have been switched over to the
+# mx-networkd generator (hostconfig tag mx-network-generator). On all other
+# machines network.service/mxvlan.service keep configuring the network; the
+# two mechanisms must never run at the same time. See mx-networkd/README.md,
+# section Migration. /node/tags/ is populated by startup-tags.service, which
+# runs Before=sysinit.target, i.e. before this unit's condition is checked.
+[Unit]
+ConditionPathExists=/node/tags/mx-network-generator

install.sh

@@ -32,6 +32,7 @@ fi
 : ${sysconfdir:=$prefix/etc}
 
 : ${systemdunitdir:=$sysconfdir/systemd/system}
+: ${systemdgeneratordir:=$usr_libdir/systemd/system-generators}
 : ${udev_rulesdir:=$sysconfdir/udev/rules.d}
 : ${crond_dir:=$sysconfdir/cron.d}
 : ${udev_helperdir:=$prefix/lib/udev}
@@ -145,6 +146,13 @@ install_exec mkmotd/mkmotd.pl                            "$DESTDIR$usr_sbindir/m
 install_data mkmotd/motd.service                         "$DESTDIR$systemdunitdir/motd.service"
 install_exec mxgrub/mxgrub                               "$DESTDIR$usr_sbindir/mxgrub"
 install_exec mxnetctl/mxnetctl                           "$DESTDIR$usr_sbindir/mxnetctl"
+# mx-networkd is built by `make`; it is installed but not enabled -- the
+# generator only does something on a machine that runs systemd-networkd.
+if [ -x mx-networkd/target/release/mx-networkd ]; then
+install_exec mx-networkd/target/release/mx-networkd      "$DESTDIR$usr_sbindir/mx-networkd"
+install_exec mx-networkd/target/release/mx-networkd      "$DESTDIR$systemdgeneratordir/mx-networkd-generator"
+fi
+install_exec mx-networkd/migrate-to-mx-networkd.sh       "$DESTDIR$usr_sbindir/migrate-to-mx-networkd"
 install_exec mxrouter/mxrouterctl                        "$DESTDIR$usr_sbindir/mxrouterctl"
 install_exec mxvlan/mxvlanctl                            "$DESTDIR$usr_sbindir/mxvlanctl"
 install_exec netlog/netlog                               "$DESTDIR$usr_sbindir/netlog"

mx-networkd/Cargo.toml

@@ -0,0 +1,18 @@
+[package]
+name = "mx-networkd"
+version = "0.1.0"
+edition = "2021"
+rust-version = "1.70"
+description = "Translate the central MarIuX network configuration into systemd-networkd and udev configuration"
+license = "MIT"
+publish = false
+
+# No dependencies on purpose: this runs as a systemd generator before any
+# file system other than / is guaranteed to be there, and it has to be
+# auditable by everyone who maintains the network configuration.
+[dependencies]
+
+[profile.release]
+opt-level = "s"
+panic = "abort"
+strip = true

mx-networkd/README.md

@@ -0,0 +1,169 @@
+# mx-networkd
+
+Translates the centrally managed MarIuX network configuration into
+systemd-networkd and udev configuration, so that `mxnetctl`, `mxvlanctl`,
+`network.service` and `mxvlan.service` can be retired without changing any of
+the pushed configuration files.
+
+Status: **not yet in use anywhere.** Nothing in this directory is installed or
+enabled by default; the old units keep running until someone switches a
+machine over by hand (see [Migration](#migration)).
+
+## What it does
+
+It runs as a **systemd generator**. Generators run at every boot and at every
+`systemctl daemon-reload`, before any unit starts — including
+`systemd-udevd.service`. That is early enough for the `.link` files to be in
+place before the first network device is coldplugged, and it means
+`/etc/systemd/network/` stays permanently empty: everything is written to
+`/run/systemd/network/`, which is also a search path for udev and networkd.
+
+    /etc/local/mxnet         ─┐                    ┌─ /run/systemd/network/10-mx-<name>.link
+    /etc/local/mxhost.conf   ─┼─ mx-networkd ──────┼─ /run/systemd/network/10-mx-<vlan>.netdev
+    /etc/mxvlans             ─┘                    └─ /run/systemd/network/20-mx-<iface>.network
+
+The three input files keep their current format and stay the single source of
+truth. No file below `/etc` is written (except by `mx-networkd learn`, see
+below).
+
+### Mapping
+
+| today | with mx-networkd |
+| --- | --- |
+| `mxnetctl start` renames by MAC | `.link` file per interface, applied by udev |
+| `network.service` adds `MX_IPADDR/20` and the default route | `.network` file for `MX_NETDEV` |
+| `mxvlanctl start` creates VLANs and their addresses | `.netdev` + `.network` per VLAN |
+| `mxvlanctl` brings the base device up | `.network` for the base device |
+
+The prefix length (20), the gateway (141.14.16.128) and the broadcast address
+were hard coded in `network.service`. They are defaults here and can be
+overridden per host with `MX_PREFIXLEN`, `MX_GATEWAY` and `MX_BROADCAST` in
+`mxhost.conf` without touching the tool.
+
+### One rule worth knowing
+
+systemd-networkd applies **only the first matching `.network` file** to an
+interface. Everything belonging to one interface therefore has to be in one
+file, even when it comes from two different inputs — the primary address from
+`mxhost.conf` and the `VLAN=` entries from `mxvlans` end up in the same
+`20-mx-net00.network`. That is why this is a program and not a pair of
+`sed` scripts.
+
+## Usage
+
+    mx-networkd generate [--root DIR] [--output DIR] [-n]
+    mx-networkd check    [--root DIR]
+    mx-networkd learn    [--root DIR] [-n]
+
+`generate` is what the generator symlink runs; systemd passes three directory
+arguments, which are accepted and ignored. Output is deterministic — no time
+stamps — so re-running it rewrites nothing and `daemon-reload` does not churn
+the network configuration. Files that mx-networkd generated earlier and no
+longer wants are removed; files in `/run/systemd/network/` belonging to
+anything else are never touched.
+
+`check` parses the same inputs and exits non-zero if anything is wrong. Run it
+on the master before pushing `/etc/mxvlans`, and a typo stops there instead of
+on 300 machines. At boot the same problems are logged and the offending line
+is skipped: a partially configured network beats no network at all.
+
+`learn` replaces the one thing `.link` files cannot do — inventing a name for a
+MAC address nobody has seen before. It uses mxnetctl's algorithm (prefer
+`netNN` matching the kernel's `ethN`, otherwise the lowest free number), writes
+`/etc/local/mxnet`, and like mxnetctl it refuses to write when
+`/etc/local/USB.usb` exists. It does not rename anything; that is udev's job
+after the next `generate`. Only needed on first installation.
+
+To preview on any machine without touching it:
+
+    mx-networkd generate --root / --dry-run
+
+## Building
+
+    cargo build --release
+    cargo test
+
+No dependencies, no build script, `std` only. This is deliberate: the binary
+runs as a generator before most of the system exists, and everyone who
+maintains the network configuration should be able to read all of it.
+
+`make` in the repository root builds it and `make install` installs the binary
+as `/usr/sbin/mx-networkd` and as the generator
+`/usr/lib/systemd/system-generators/mx-networkd-generator`.
+
+## Migration
+
+The switch per machine is the **hostconfig tag `mx-network-generator`**
+(hostconfig allows hyphens in tag names: a tag term matches
+`[a-z][a-z0-9$_-]*`). Tags live in the centrally pushed `/etc/hostconfig`
+and are materialised as files below `/node/tags/` by `startup-tags.service`
+(`hostconfig --populate-node`) at every boot. Everything is conditioned on
+that one tag:
+
+| | with the tag | without the tag |
+| --- | --- | --- |
+| `mx-networkd generate` | writes to `/run/systemd/network/` | writes nothing, prunes its old files |
+| `systemd-networkd.service` | runs (drop-in `ConditionPathExists=/node/tags/mx-network-generator`) | skipped |
+| `network.service`, `mxvlan.service` | skipped (`ConditionPathExists=!/node/tags/mx-network-generator`) | run as before |
+
+`/node/tags/` is on the root file system and persists across boots, so the
+tag file is already there when the generators run — before
+`startup-tags.service` has run again. The old and the new mechanism never
+run at the same time. `--dry-run` ignores the tag, so the preview always
+works.
+
+Per machine, reversible at every step — use the helper script:
+
+    migrate-to-mx-networkd
+
+It validates the prerequisites and the configuration (`mx-networkd check`),
+shows the `--dry-run` preview, verifies the tag is set in `/etc/hostconfig`
+(and tells you what to add on the distmaster if it is not), refreshes
+`/node/tags/` and enables `systemd-networkd.service` for the next boot. It
+never starts, stops or reloads anything, so it is safe to run over SSH and
+safe to run repeatedly. Do **not** flip the services by hand on a running
+machine: `network.service` deletes the primary address on stop
+(`ExecStop=ip addr del …`), so stopping it over SSH cuts the connection and
+locks you out. The switchover is the reboot:
+
+1. Add `tag mx-network-generator` for the host to `/etc/hostconfig` on the
+   distmaster and push it.
+2. Run `migrate-to-mx-networkd`, read its output, compare the preview with
+   `ip -br addr` and `ip -d link show type vlan`.
+3. `reboot`. The generator runs before udev, the `.link` files rename the
+   devices at coldplug, and networkd owns the addresses from the start. A
+   reboot is needed anyway: renaming only happens on device *add* — a
+   running interface is not renamed.
+4. After the reboot, verify: `ip -br addr`, `ip -d link show type vlan`,
+   `networkctl status`.
+
+To go back — also at the next boot, not live, for the same reason in
+reverse (`ip addr add` on an address networkd still holds fails, and the
+units fight over the interface):
+
+1. Remove the tag from `/etc/hostconfig` on the distmaster and push it.
+2. `hostconfig --populate-node`
+3. `systemctl disable systemd-networkd.service`
+4. `reboot` — the old units run again; the generator removes its files.
+
+## Deliberate non-goals
+
+* **No `apply` subcommand.** Applying the configuration is
+  `networkctl reload`. A second implementation of "make the kernel match the
+  files" is exactly the divergence this change is supposed to remove.
+* **No new file format.** If `/etc/mxvlans` needs to grow a field one day,
+  that is a separate discussion; this change must be a no-op for the people
+  who edit it.
+* **No IPv6.** The current scripts do not configure any, so neither does this.
+  `.network` files are the obvious place to add it later.
+
+## Known gaps
+
+* `learn` assumes kernel names of the form `ethN`, i.e. `net.ifnames=0`, which
+  is what mxnetctl assumed too. On a machine with predictable names
+  (`enp1s0f0`) it will find nothing and say so.
+* Renaming still only happens at device add. Changing a name in
+  `/etc/local/mxnet` needs a reboot (or `udevadm trigger` plus taking the
+  interface down), same as before.
+* `/etc/mxvlans` is read for the local host only, matched on the short host
+  name, as mxvlanctl does.

mx-networkd/migrate-to-mx-networkd.sh

@@ -0,0 +1,119 @@
+#! /bin/bash
+
+# migrate-to-mx-networkd - switch this machine from network.service/mxnetctl/
+# mxvlanctl to systemd-networkd fed by the mx-networkd generator.
+#
+# The switchover happens at the NEXT BOOT, never live: network.service
+# deletes the primary address on stop, so flipping services on a running
+# machine over SSH locks you out (see mx-networkd/README.md, Migration).
+# This script therefore never starts, stops or reloads anything -- it only
+# validates the configuration, checks the hostconfig tag, refreshes
+# /node/tags/ and enables systemd-networkd for the next boot. It is
+# idempotent: run it as often as you like, over SSH too.
+#
+# The single switch is the hostconfig tag 'mx-network-generator':
+#   - with the tag:    mx-networkd generates, systemd-networkd runs,
+#                      network.service/mxvlan.service are skipped by their
+#                      ConditionPathExists=!/node/tags/mx-network-generator
+#   - without the tag: mx-networkd writes nothing, systemd-networkd is
+#                      skipped by its drop-in condition, the old units run
+#
+# Written by OnkelClaude (AI) on behalf of Boris Bergenroth.
+
+TAG=mx-network-generator
+TAGFILE=/node/tags/$TAG
+DROPIN=/etc/systemd/system/systemd-networkd.service.d/$TAG.conf
+
+die() {
+    echo "$0: error: $*" >&2
+    exit 1
+}
+
+step() {
+    echo
+    echo "== $*"
+}
+
+[ "$(id -u)" = 0 ] || die "must be run as root"
+
+step "checking prerequisites"
+
+[ -x /usr/sbin/mx-networkd ] \
+    || die "/usr/sbin/mx-networkd is not installed - update mxtools first"
+[ -x /usr/lib/systemd/system-generators/mx-networkd-generator ] \
+    || die "generator /usr/lib/systemd/system-generators/mx-networkd-generator is not installed - update mxtools first"
+[ -x /usr/sbin/hostconfig ] \
+    || die "/usr/sbin/hostconfig is not installed"
+[ -e /etc/local/mxhost.conf ] \
+    || die "/etc/local/mxhost.conf does not exist - this machine does not use the MarIuX network configuration"
+[ -e /etc/local/mxnet ] \
+    || die "/etc/local/mxnet does not exist - run 'mx-networkd learn' (or mxnetctl) first"
+[ -e "$DROPIN" ] \
+    || die "$DROPIN is missing - update mxtools first"
+systemctl cat systemd-networkd.service >/dev/null 2>&1 \
+    || die "systemd-networkd.service does not exist on this machine"
+grep -q 'ConditionPathExists=!'"$TAGFILE" /etc/systemd/system/network.service 2>/dev/null \
+    || die "/etc/systemd/system/network.service is not conditioned on the tag yet - update mxtools first"
+echo "ok"
+
+step "validating the network configuration (mx-networkd check)"
+/usr/sbin/mx-networkd check \
+    || die "configuration problems found - fix them first; nothing was changed"
+
+step "preview of the generated configuration (nothing is applied)"
+/usr/sbin/mx-networkd generate --dry-run
+echo
+echo "Compare with the running state: ip -br addr ; ip -d link show type vlan"
+
+step "checking hostconfig tag '$TAG'"
+if /usr/sbin/hostconfig "$TAG"; then
+    echo "tag '$TAG' is set for this host in /etc/hostconfig"
+else
+    cat >&2 <<__EOF__
+
+The tag '$TAG' is NOT set for this host in /etc/hostconfig.
+/node/tags/ is rebuilt from /etc/hostconfig at every boot, so the tag has
+to come from there to survive. Add it on the distmaster:
+
+    $(uname -n | cut -d. -f1)	tag	$TAG
+
+push /etc/hostconfig to this machine, then run this script again.
+
+Nothing was changed.
+__EOF__
+    exit 1
+fi
+
+step "refreshing /node/tags/ (hostconfig --populate-node)"
+/usr/sbin/hostconfig --populate-node \
+    || die "hostconfig --populate-node failed"
+[ -e "$TAGFILE" ] \
+    || die "$TAGFILE did not appear after populate-node"
+echo "ok: $TAGFILE"
+
+step "enabling systemd-networkd.service for the next boot (not starting it)"
+systemctl enable systemd-networkd.service \
+    || die "systemctl enable systemd-networkd.service failed"
+
+cat <<__EOF__
+
+== done - this machine switches to mx-networkd at the next boot
+
+Nothing has changed on the running system: network.service keeps the
+addresses until shutdown. network.service and mxvlan.service stay enabled;
+from the next boot on they are skipped by their tag condition, and
+systemd-networkd takes over. Reboot when ready, then verify with:
+
+    ip -br addr
+    ip -d link show type vlan
+    networkctl status
+
+Rollback (also reboot-based, do NOT flip services on a running machine):
+
+    1. remove the '$TAG' tag for this host from /etc/hostconfig
+       on the distmaster and push the file
+    2. hostconfig --populate-node
+    3. systemctl disable systemd-networkd.service
+    4. reboot
+
+__EOF__