use NIRI_SOCKET variable in systemd user services

[TheAngusMcFire]

Hi everybody,

I try to start 2 services which need to communicate with niri as systemd user services.
One is the xwayland-satellite and one is a rust service which uses the niri-ipc crate to communicate with niri.

Both systemd files look like that.

[Unit]
Description=Script Daemon for xwayland-satellite 
After=graphical-session.target
Requisite=graphical-session.target
PartOf=graphical-session.target

[Service]
Type=simple
ExecStart=/usr/bin/xwayland-satellite :1000
Restart=on-failure

The problem I am having is that the services start to fast, so that the NIRI_SOCKET variable is not yet set.
niri-ipc reports the missing env var.
If I restart the 2 services everything is working fine, so I think systemd is not waiting long enough before it starts those 2 services.

Has anybody done something similar or does know how I can configure the systemd service to wait for the env var to be present?

System: Arch Linux
DisplayManager: GDM

[YaLTeR]

Are you on NixOS? It seems to have some problem with graphical-session.target dependencies, so you could try replacing them with niri.service dependencies.

[TheAngusMcFire]

They are as symlinks in the niri.service.wants folder, thats how they get started

[YaLTeR]

Yeah, that's normal. Not sure, I haven't had any failures with the same setup on Fedora.

[TheAngusMcFire]

how does systemd know that niri is started? Is it guaranteed that the environment was published before the service goes into active state?

[YaLTeR]

It uses Type=notify, which means that it waits for a message from niri, which niri sends after publishing the environment.

[TheAngusMcFire]

I see, thank you for the help, I think I have to debug a bit for this one 🙈

[TheAngusMcFire]

I kept having this problem, but I just had time for some additional debugging, and it seems that the NIRI_SOCKET variable contains the wrong or old NIRI_SOCKET value.

NIRI_SOCKET of the service: /run/user/1000/niri.wayland-1.2668771.sock
NIRI_SOCKET of the current instance: /run/user/1000/niri.wayland-1.2685705.sock

Is this a systemd bug maybe?

[TheAngusMcFire]

It seams that the services are not stopped when niri is stopped, what makes it more strange: if I switch from gdm to another tty, the service is suddenly stopped...

[YaLTeR]

FYI, this is the expected order of events (copying from my message in sodiboo/niri-flake#348):

1. niri-session runs and starts niri.service
2. niri.service ExecStart-s /usr/bin/niri --session, then waits for a notification (Type=notify)
3. niri imports WAYLAND_DISPLAY and whatnot into the systemd environment:

   • niri/src/main.rs

     Line 209 in 7b6fa12

     import_environment();

   • niri/src/main.rs

     Lines 271 to 305 in 7b6fa12

     	fn import_environment() {
     	let variables = [
     	"WAYLAND_DISPLAY",
     	"XDG_CURRENT_DESKTOP",
     	"XDG_SESSION_TYPE",
     	niri_ipc::SOCKET_PATH_ENV,
     	]
     	.join(" ");
     	let mut init_system_import = String::new();
     	if cfg!(feature = "systemd") {
     	write!(
     	init_system_import,
     	"systemctl --user import-environment {variables};"
     	)
     	.unwrap();
     	}
     	if cfg!(feature = "dinit") {
     	write!(init_system_import, "dinitctl setenv {variables};").unwrap();
     	}
     	let rv = Command::new("/bin/sh")
     	.args([
     	"-c",
     	&format!(
     	"{init_system_import}\
     	hash dbus-update-activation-environment 2>/dev/null && \
     	dbus-update-activation-environment {variables}"
     	),
     	])
     	.spawn();
     	// Wait for the import process to complete, otherwise services will start too fast without
     	// environment variables available.
     	match rv {
     	Ok(mut child) => match child.wait() {

4. niri notifies systemd that it is ready:

   niri/src/main.rs

   Line 224 in 7b6fa12

   if let Err(err) = sd_notify::notify(true, &[NotifyState::Ready]) {

5. systemd now proceeds to start graphical-session.target (since niri.service says Before=graphical-session.target). At this point the environment is filled in and everything is ready
6. systemd starts everything that has After=graphical-session.target, Requisite=graphical-session.target, i.e. example swaybg.service from the wiki page

[TheAngusMcFire]

Thanks for the detailed explanation. The problem was not related to niri at all.
The services did not get stopped when niri got stopped for some reason. I added a systemctl --user restart as spawn-at-startup this worked perfectly.
Thanks for the help.