Surprisingly, gdb of all things uses the login shell from /etc/passwd to
launch the program under test, so I was seeing the
"This account is currently not available" message there.
Gtk uses G_DBUS_CALL_FLAGS_NO_AUTO_START in all the (early) portal calls
so we basically get a version of https://gitlab.gnome.org/GNOME/gtk/-/issues/7379
if we try to lazily start the agent..
Add a NixOS option to configure the default command for the VM and
provide a system.build.munix output that wraps munix with the correct
toplevel and default command. This reduces boilerplate in downstream
flakes since they no longer need to manually wrap munix.
The template now uses these new features, significantly simplifying
the apps definition.
This allows using normal nix-managed services that register sockets for
lazy activation. Also brings in some red errors about starting services
that we don't have, but that will be fixed
Instead of interpreting all that shell and running actual tmpfiles, use
a tiny stage before systemd that mounts a tmpfs at /run (preventing
systemd from doing the same), populates it with NixOS symlinks and
preserved resolv.conf, and mounts the immutable /etc overlay before
passing control over to systemd.
Avoid spending ~100ms of userspace boot time on unnecessary regeneration
of entirely static /etc/{passwd,group,shadow}. This will be proposed to nixpkgs.
Relying on udev to assign permissions to the virtgpu was very "proper"
but really excessive. We have a simple static configuration, built into
the kernel, so there are no dynamic shenanigans, we can always reliably
just chmod/chown the permissions. Let's go back to faster booting.
