git.delta.rocks / jrsonnet / refs/commits / 3972fee37ee3

1use std::{env::current_dir, os::unix::fs::symlink, path::PathBuf, time::Duration};23use anyhow::{anyhow, Result};4use clap::{Parser, ValueEnum};5use fleet_base::{6	host::{Config, ConfigHost},7	opts::FleetOpts,8};9use itertools::Itertools as _;10use nix_eval::{nix_go, NixBuildBatch};11use tokio::{task::LocalSet, time::sleep};12use tracing::{error, field, info, info_span, warn, Instrument};1314#[derive(Parser)]15pub struct Deploy {16	/// Disable automatic rollback17	#[clap(long)]18	disable_rollback: bool,19	/// Action to execute after system is built20	action: DeployAction,21}2223#[derive(ValueEnum, Clone, Copy)]24enum DeployAction {25	/// Upload derivation, but do not execute the update.26	Upload,27	/// Upload and execute the activation script, old version will be used after reboot.28	Test,29	/// Upload and set as current system profile, but do not execute activation script.30	Boot,31	/// Upload, set current profile, and execute activation script.32	Switch,33}3435impl DeployAction {36	pub(crate) fn name(&self) -> Option<&'static str> {37		match self {38			Self::Upload => None,39			Self::Test => Some("test"),40			Self::Boot => Some("boot"),41			Self::Switch => Some("switch"),42		}43	}44	pub(crate) fn should_switch_profile(&self) -> bool {45		matches!(self, Self::Switch | Self::Boot)46	}47	pub(crate) fn should_activate(&self) -> bool {48		matches!(self, Self::Switch | Self::Test | Self::Boot)49	}50	pub(crate) fn should_create_rollback_marker(&self) -> bool {51		// Upload does nothing on the target machine, other than uploading the closure.52		// In boot case we want to have rollback marker prepared, so that the system may rollback itself on the next boot.53		!matches!(self, Self::Upload)54	}55	pub(crate) fn should_schedule_rollback_run(&self) -> bool {56		matches!(self, Self::Switch | Self::Test)57	}58}5960#[derive(Parser, Clone)]61pub struct BuildSystems {62	/// Attribute to build. Systems are deployed from "toplevel" attr, well-known used attributes63	/// are "sdImage"/"isoImage", and your configuration may include any other build attributes.64	#[clap(long, default_value = "toplevel")]65	build_attr: String,66}6768struct Generation {69	id: u32,70	current: bool,71	datetime: String,72}7374fn parse_generation_line(g: &str) -> Option<Generation> {75	let mut parts = g.split_whitespace();76	let id = parts.next()?;77	let id: u32 = id.parse().ok()?;78	let date = parts.next()?;79	let time = parts.next()?;80	let current = if let Some(current) = parts.next() {81		if current == "(current)" {82			Some(true)83		} else {84			None85		}86	} else {87		Some(false)88	};89	let current = current?;90	if parts.next().is_some() {91		warn!("unexpected text after generation: {g}");92	}93	Some(Generation {94		id,95		current,96		datetime: format!("{date} {time}"),97	})98}99100async fn get_current_generation(host: &ConfigHost) -> Result<Generation> {101	let mut cmd = host.cmd("nix-env").await?;102	cmd.comparg("--profile", "/nix/var/nix/profiles/system")103		.arg("--list-generations");104	// Sudo is required due to --list-generations acquiring lock on the profile.105	let data = cmd.sudo().run_string().await?;106	let generations = data107		.split('\n')108		.map(|e| e.trim())109		.filter(|&l| !l.is_empty())110		.filter_map(|g| {111			let gen = parse_generation_line(g);112			if gen.is_none() {113				warn!("bad generation: {g}");114			}115			gen116		})117		.collect::<Vec<_>>();118	let current = generations119		.into_iter()120		.filter(|g| g.current)121		.at_most_one()122		.map_err(|_e| anyhow!("bad list-generations output"))?123		.ok_or_else(|| anyhow!("failed to find generation"))?;124	Ok(current)125}126127async fn deploy_task(128	action: DeployAction,129	host: &ConfigHost,130	built: PathBuf,131	specialisation: Option<String>,132	disable_rollback: bool,133) -> Result<()> {134	let mut failed = false;135	// TODO: Lockfile, to prevent concurrent system switch?136	// TODO: If rollback target exists - bail, it should be removed. Lockfile will not work in case if rollback137	// is scheduler on next boot (default behavior). On current boot - rollback activator will fail due to138	// unit name conflict in systemd-run139	// This code is tied to rollback.nix140	if !disable_rollback && action.should_create_rollback_marker() {141		let _span = info_span!("preparing").entered();142		info!("preparing for rollback");143		let generation = get_current_generation(host).await?;144		info!(145			"rollback target would be {} {}",146			generation.id, generation.datetime147		);148		{149			let mut cmd = host.cmd("sh").await?;150			cmd.arg("-c").arg(format!("mark=$(mktemp -p /etc -t fleet_rollback_marker.XXXXX) && echo -n {} > $mark && mv --no-clobber $mark /etc/fleet_rollback_marker", generation.id));151			if let Err(e) = cmd.sudo().run().await {152				error!("failed to set rollback marker: {e}");153				failed = true;154			}155		}156		// Activation script also starts rollback-watchdog.timer, however, it is possible that it won't be started.157		// Kicking it on manually will work best.158		//159		// There wouldn't be conflict, because here we trigger start of the primary service, and systemd will160		// only allow one instance of it.161162		// TODO: We should also watch how this process is going.163		// After running this command, we have less than 3 minutes to deploy everything,164		// if we fail to perform generation switch in time, then we will still call the activation script, and this may break something.165		// Anyway, reboot will still help in this case.166		if action.should_schedule_rollback_run() {167			let mut cmd = host.cmd("systemd-run").await?;168			cmd.comparg("--on-active", "3min")169				.comparg("--unit", "rollback-watchdog-run")170				.arg("systemctl")171				.arg("start")172				.arg("rollback-watchdog.service");173			if let Err(e) = cmd.sudo().run().await {174				error!("failed to schedule rollback run: {e}");175				failed = true;176			}177		}178	}179180	if action.should_switch_profile() && !failed {181		info!("switching system profile generation");182		// It would also be possible to update profile atomically during copy:183		// https://github.com/NixOS/nix/pull/11657184		let mut cmd = host.cmd("nix").await?;185		cmd.arg("build");186		cmd.comparg("--profile", "/nix/var/nix/profiles/system");187		cmd.arg(&built);188		if let Err(e) = cmd.sudo().run_nix().await {189			error!("failed to switch system profile generation: {e}");190			failed = true;191		}192	}193194	// FIXME: Connection might be disconnected after activation run195196	if action.should_activate() && !failed {197		let _span = info_span!("activating").entered();198		info!("executing activation script");199		let specialised = if let Some(specialisation) = specialisation {200			let mut specialised = built.join("specialisation");201			specialised.push(specialisation);202			specialised203		} else {204			built.clone()205		};206		let switch_script = specialised.join("bin/switch-to-configuration");207		let mut cmd = host.cmd(switch_script).in_current_span().await?;208		cmd.arg(action.name().expect("upload.should_activate == false"));209		if let Err(e) = cmd.sudo().run().in_current_span().await {210			error!("failed to activate: {e}");211			failed = true;212		}213	}214	if action.should_create_rollback_marker() {215		if !disable_rollback {216			if failed {217				if action.should_schedule_rollback_run() {218					info!("executing rollback");219					if let Err(e) = host220						.systemctl_start("rollback-watchdog.service")221						.instrument(info_span!("rollback"))222						.await223					{224						error!("failed to trigger rollback: {e}")225					}226				}227			} else {228				info!("trying to mark upgrade as successful");229				if let Err(e) = host230					.rm_file("/etc/fleet_rollback_marker", true)231					.in_current_span()232					.await233				{234					error!("failed to remove rollback marker. This is bad, as the system will be rolled back by watchdog: {e}")235				}236			}237			info!("disarming watchdog, just in case");238			if let Err(_e) = host.systemctl_stop("rollback-watchdog.timer").await {239				// It is ok, if there was no reboot - then timer might not be running.240			}241			if action.should_schedule_rollback_run() {242				if let Err(e) = host.systemctl_stop("rollback-watchdog-run.timer").await {243					error!("failed to disarm rollback run: {e}");244				}245			}246		} else if let Err(_e) = host247			.rm_file("/etc/fleet_rollback_marker", true)248			.in_current_span()249			.await250		{251			// Marker might not exist, yet better try to remove it.252		}253	}254	Ok(())255}256257async fn build_task(258	config: Config,259	hostname: String,260	build_attr: &str,261	batch: Option<NixBuildBatch>,262) -> Result<PathBuf> {263	info!("building");264	let host = config.host(&hostname).await?;265	// let action = Action::from(self.subcommand.clone());266	let nixos = host.nixos_config().await?;267	let drv = nix_go!(nixos.system.build[{ build_attr }]);268	let outputs = drv.build_maybe_batch(batch).await?;269	let out_output = outputs270		.get("out")271		.ok_or_else(|| anyhow!("system build should produce \"out\" output"))?;272273	{274		info!("adding gc root");275		let mut cmd = config.local_host().cmd("nix").await?;276		cmd.arg("build")277			.comparg(278				"--profile",279				format!(280					"/nix/var/nix/profiles/{}-{hostname}",281					config.data().gc_root_prefix282				),283			)284			.arg(out_output);285		cmd.sudo().run_nix().await?;286	}287288	Ok(out_output.clone())289}290291impl BuildSystems {292	pub async fn run(self, config: &Config, opts: &FleetOpts) -> Result<()> {293		let hosts = opts.filter_skipped(config.list_hosts().await?).await?;294		let set = LocalSet::new();295		let build_attr = self.build_attr.clone();296		let batch = (hosts.len() > 1).then(|| {297			config298				.nix_session299				.new_build_batch("build-hosts".to_string())300		});301		for host in hosts {302			let config = config.clone();303			let span = info_span!("build", host = field::display(&host.name));304			let hostname = host.name;305			let build_attr = build_attr.clone();306			let batch = batch.clone();307			set.spawn_local(308				(async move {309					let built = match build_task(config, hostname.clone(), &build_attr, batch).await310					{311						Ok(path) => path,312						Err(e) => {313							error!("failed to deploy host: {}", e);314							return;315						}316					};317					// TODO: Handle error318					let mut out = current_dir().expect("cwd exists");319					out.push(format!("built-{}", hostname));320321					info!("linking iso image to {:?}", out);322					if let Err(e) = symlink(built, out) {323						error!("failed to symlink: {e}")324					}325				})326				.instrument(span),327			);328		}329		drop(batch);330		set.await;331		Ok(())332	}333}334335impl Deploy {336	pub async fn run(self, config: &Config, opts: &FleetOpts) -> Result<()> {337		let hosts = opts.filter_skipped(config.list_hosts().await?).await?;338		let set = LocalSet::new();339		let batch = (hosts.len() > 1).then(|| {340			config341				.nix_session342				.new_build_batch("deploy-hosts".to_string())343		});344		for host in hosts.into_iter() {345			let config = config.clone();346			let span = info_span!("deploy", host = field::display(&host.name));347			let hostname = host.name.clone();348			let local_host = config.local_host();349			let opts = opts.clone();350			let batch = batch.clone();351352			set.spawn_local(353				(async move {354					let built =355						match build_task(config.clone(), hostname.clone(), "toplevel", batch).await356						{357							Ok(path) => path,358							Err(e) => {359								error!("failed to deploy host: {}", e);360								return;361							}362						};363					if !opts.is_local(&hostname) {364						info!("uploading system closure");365						{366							// TODO: Move to remote_derivation method.367							// Alternatively, nix store make-content-addressed can be used,368							// at least for the first deployment, to provide trusted store key.369							//370							// It is much slower, yet doesn't require root on the deployer machine.371							let Ok(mut sign) = local_host.cmd("nix").await else {372								error!("failed to setup local");373								return;374							};375							// Private key for host machine is registered in nix-sign.nix376							sign.arg("store")377								.arg("sign")378								.comparg("--key-file", "/etc/nix/private-key")379								.arg("-r")380								.arg(&built);381							if let Err(e) = sign.sudo().run_nix().await {382								warn!("failed to sign store paths: {e}");383							};384						}385						let mut tries = 0;386						loop {387							match host.remote_derivation(&built).await {388								Ok(remote) => {389									assert!(remote == built, "CA derivations aren't implemented");390									break;391								}392								Err(e) if tries < 3 => {393									tries += 1;394									warn!("copy failure ({}/3): {}", tries, e);395									sleep(Duration::from_millis(5000)).await;396								}397								Err(e) => {398									error!("upload failed: {e}");399									return;400								}401							}402						}403					}404					if let Err(e) = deploy_task(405						self.action,406						&host,407						built,408						if let Ok(v) = opts.action_attr(&host, "specialisation").await {409							v410						} else {411							error!("unreachable? failed to get specialization");412							return;413						},414						self.disable_rollback,415					)416					.await417					{418						error!("activation failed: {e}");419					}420				})421				.instrument(span),422			);423		}424		drop(batch);425		set.await;426		Ok(())427	}428}