nix-ota/nix/tests/ota.nix

{ pkgs, self, system }:
# NixOS VM test for nix-ota.
#
# Builds three "system closure" stand-ins at evaluation time (each is a
# directory containing a marker file and a `bin/switch-to-configuration`
# stub), then drives the agent through three publishes:
#   1. publish A -> device switches to A
#   2. publish B -> device switches to B
#   3. publish C (broken: agent's healthCmd will fail) -> device rolls back to B
let
  mkClosure = label: extraScript: pkgs.runCommand "sys-${label}" {} ''
    mkdir -p $out/bin
    echo "${label}" > $out/marker
    cat > $out/bin/switch-to-configuration <<'EOF'
    #!/bin/sh
    set -eu
    echo "applied ${label}" >&2
    ${extraScript}
    exit 0
    EOF
    chmod +x $out/bin/switch-to-configuration
  '';

  closureA = mkClosure "a" "touch /run/nix-ota-applied-a";
  closureB = mkClosure "b" "touch /run/nix-ota-applied-b";
  # Closure C activates fine, but the healthCmd checks for /run/nix-ota-broken
  # which we create before publishing C, forcing rollback.
  closureC = mkClosure "c" "touch /run/nix-ota-applied-c";

  # Pre-generated binary cache keypair (test fixture; not secret).
  # Generated with: nix-store --generate-binary-cache-key cache.local sec pub
  cacheKeys = pkgs.runCommand "test-cache-keys" {} ''
    mkdir -p $out
    export HOME=$TMPDIR
    export NIX_STATE_DIR=$TMPDIR/state
    export NIX_STORE_DIR=$TMPDIR/store
    mkdir -p $NIX_STATE_DIR $NIX_STORE_DIR
    ${pkgs.nix}/bin/nix-store --generate-binary-cache-key cache.local $out/secret $out/public
  '';

  pubBin = "${self.packages.${system}.nix-ota-publisher}/bin/nix-ota";
in
pkgs.testers.runNixOSTest {
  name = "nix-ota";
  nodes = {
    server = { config, pkgs, lib, ... }: {
      imports = [ self.nixosModules.server ];
      nix.settings.experimental-features = [ "nix-command" "flakes" ];
      services.nix-ota-server = {
        enable = true;
        listen = "0.0.0.0:8080";
        openFirewall = true;
        publishTokenFile = pkgs.writeText "tok" "test-token";
      };
      services.nix-serve = {
        enable = true;
        port = 5000;
        secretKeyFile = "${cacheKeys}/secret";
      };
      networking.firewall.allowedTCPPorts = [ 5000 ];
      # The closures need to be in the server's store so nix-serve can serve them.
      system.extraDependencies = [ closureA closureB closureC ];
    };

    device = { config, pkgs, lib, ... }: {
      imports = [ self.nixosModules.agent ];
      services.nix-ota-agent = {
        enable = true;
        server = "http://server:8080";
        channel = "prod";
        deviceId = "vm-device-1";
        publicKeyFile = "/var/lib/nix-ota/public.key";
        cacheUrl = "http://server:5000";
        cachePublicKey = builtins.readFile "${cacheKeys}/public";
        interval = 5;
        healthCmd = "test ! -f /run/nix-ota-broken";
      };
      nix.settings.experimental-features = [ "nix-command" "flakes" ];
      nix.settings.trusted-users = [ "root" ];
    };
  };

  testScript = ''
    closureA = "${closureA}"
    closureB = "${closureB}"
    closureC = "${closureC}"
    pubBin = "${pubBin}"

    start_all()
    server.wait_for_unit("nix-ota-server.service")
    server.wait_for_open_port(8080)
    server.wait_for_unit("nix-serve.service")
    server.wait_for_open_port(5000)
    # Drive the agent ourselves; disable the timer for deterministic stepping.
    device.succeed("systemctl stop nix-ota-agent.timer || true")

    # Sign the closures with the binary cache key so the device's Nix will accept them.
    for c in [closureA, closureB, closureC]:
        server.succeed(f"nix store sign --extra-experimental-features nix-command --key-file ${cacheKeys}/secret --recursive {c}")

    # Operator generates a manifest signing key on the server host.
    server.succeed("mkdir -p /root/keys")
    pub = server.succeed(f"{pubBin} keygen --out /root/keys/sign.key").strip()

    # Push pubkey onto the device's writable state dir.
    device.succeed("mkdir -p /var/lib/nix-ota")
    device.succeed(f"echo '{pub}' > /var/lib/nix-ota/public.key")

    def publish(store_path, rev):
        server.succeed(
            f"{pubBin} publish "
            f"--server http://localhost:8080 --token test-token "
            f"--key /root/keys/sign.key --channel prod "
            f"--store-path {store_path} --substituter http://server:5000 --revision {rev}"
        )

    def poll_agent():
        # oneshot service: start and wait for it to finish (success or failure).
        device.succeed("systemctl start --wait nix-ota-agent.service || true")

    # --- Step 1: publish A
    publish(closureA, 1)
    poll_agent()
    device.succeed(f"readlink -f /nix/var/nix/profiles/system | grep -qF {closureA}")
    device.succeed("test -f /run/nix-ota-applied-a")

    # --- Step 2: publish B
    publish(closureB, 2)
    poll_agent()
    device.succeed(f"readlink -f /nix/var/nix/profiles/system | grep -qF {closureB}")
    device.succeed("test -f /run/nix-ota-applied-b")

    # --- Step 3: publish C with health check rigged to fail
    device.succeed("touch /run/nix-ota-broken")
    publish(closureC, 3)
    poll_agent()
    # Agent should have applied C, failed the health check, and rolled back to B.
    device.succeed(f"readlink -f /nix/var/nix/profiles/system | grep -qF {closureB}")
    # The activation script for C did run before health check.
    device.succeed("test -f /run/nix-ota-applied-c")

    # The dashboard should reflect the rolled_back state.
    server.wait_until_succeeds(
        "curl -fsS http://localhost:8080/ | grep -Eq 'rolled_back|failed'", timeout=30
    )
  '';
}