runnerd_core/scheduler/
engine.rs

1//! The scheduling loop: polls GitHub, pulls the profiles' images, applies `plan`'s decisions with GitHub and
2//! Docker, removes each runner's container when its job ends, keeps the mirror up while any runner runs, reseeds
3//! idle profiles and recovers runners and containers a previous daemon left. It also feeds the job history: queued
4//! jobs, the job each runner took (from `observe::capture`), steps from GitHub, mirror bytes, cache sizes and slot
5//! and budget changes. It owns slot state; it never serves the API (it publishes `api::Status` and takes commands
6//! through `Control`) and never logs the token or a JIT config.
7//!
8//! This module holds `Engine`'s state, its construction and its loop; the submodules each add one concern's
9//! methods: `recover` (startup), `tick` (polling and planning), `runners` (a slot's runner from start to removal),
10//! `images` (pulling the profiles' images), `caches` (reseeds, the mirror and the shared caches), `history` (what
11//! the store records), `control` (the API's commands) and `status` (what it publishes).
12
13mod caches;
14mod control;
15mod history;
16mod images;
17mod recover;
18mod runners;
19mod status;
20mod tick;
21
22use std::collections::{BTreeMap, BTreeSet};
23use std::sync::Arc;
24use std::time::{Duration, SystemTime};
25
26use tokio::sync::{mpsc, watch};
27use tokio::task::JoinHandle;
28use tracing::{info, warn};
29
30pub use self::control::{
31    Command, Commands, Commit, Control, ControlError, RunningJobs, Shutdown, Target,
32};
33use self::history::{Closing, Found};
34use self::images::ImageState;
35use super::SlotRef;
36use crate::api;
37use crate::cache::{self, Names, Trimmed};
38use crate::config::Config;
39use crate::docker::{self, Docker, DockerApi};
40use crate::github::{Github, GithubApi, QueuedJob};
41use crate::isolation::{self, Isolation};
42use crate::observe::metrics::{self, Paths, Sampler, Tracked};
43use crate::observe::mirror::Leases;
44use crate::store::{MirrorPosition, Store};
45
46/// This machine as its runner names carry it: `runnerd-<id>-<host>-<profile>-<slot>-<secs>`. Only `runnerd-<id>-`
47/// identifies the machine, so a changed host name still matches the runners started before the change, and two
48/// machines with the same or overlapping host names never take each other's runners for their own.
49#[derive(Debug, Clone, PartialEq, Eq)]
50pub struct HostIdentity {
51    /// 8 hex digits: a hash of the systemd machine ID, never the ID itself, as machine-id(5) asks.
52    id: String,
53    /// The host name in lower case, letters, digits and dashes, at most 20 characters: for people reading names.
54    host: String,
55}
56
57impl HostIdentity {
58    /// This machine's identity, from `/etc/machine-id` and its host name.
59    pub fn read() -> std::io::Result<HostIdentity> {
60        let hostname = std::fs::read_to_string("/proc/sys/kernel/hostname").unwrap_or_default();
61        let machine_id = std::fs::read_to_string("/etc/machine-id")
62            .or_else(|_| std::fs::read_to_string("/var/lib/dbus/machine-id"))
63            .map_err(|error| {
64                std::io::Error::new(
65                    error.kind(),
66                    format!("cannot read /etc/machine-id: {error}"),
67                )
68            })?;
69        HostIdentity::new(&hostname, &machine_id).ok_or_else(|| {
70            std::io::Error::new(
71                std::io::ErrorKind::InvalidData,
72                "/etc/machine-id does not hold a machine ID",
73            )
74        })
75    }
76
77    /// The identity for a host name and a machine ID; `None` unless the ID is 32 hex digits.
78    pub fn new(hostname: &str, machine_id: &str) -> Option<HostIdentity> {
79        let machine_id = machine_id.trim();
80        if machine_id.len() != 32 || !machine_id.chars().all(|c| c.is_ascii_hexdigit()) {
81            return None;
82        }
83        // FNV-1a: stable across Rust releases, unlike std's hasher, so a new build recognizes the runners of the old.
84        let hash = format!("runnerd:{}", machine_id.to_ascii_lowercase())
85            .bytes()
86            .fold(0xcbf2_9ce4_8422_2325_u64, |hash, byte| {
87                (hash ^ u64::from(byte)).wrapping_mul(0x0000_0100_0000_01b3)
88            });
89        let host: String = hostname
90            .trim()
91            .chars()
92            .map(|c| {
93                if c.is_ascii_alphanumeric() {
94                    c.to_ascii_lowercase()
95                } else {
96                    '-'
97                }
98            })
99            .take(20)
100            .collect();
101        let host = host.trim_matches('-');
102        Some(HostIdentity {
103            id: format!("{:08x}", hash >> 32),
104            host: if host.is_empty() { "host" } else { host }.to_owned(),
105        })
106    }
107
108    /// What every runner name of this machine starts with. The ID's fixed length keeps it from starting another
109    /// machine's names.
110    pub fn prefix(&self) -> String {
111        format!("runnerd-{}-", self.id)
112    }
113
114    /// The name of a runner, and of its container, started in `slot` (1-based) of `profile` at `secs`.
115    pub fn runner_name(&self, profile: &str, slot: usize, secs: u64) -> String {
116        format!("{}{}-{profile}-{slot}-{secs}", self.prefix(), self.host)
117    }
118}
119
120/// What the API needs from a running engine.
121#[derive(Debug, Clone)]
122pub struct Handle {
123    /// The latest status, replaced after every tick.
124    pub status: watch::Receiver<api::Status>,
125    /// Pause, resume and shutdown requests.
126    pub control: Arc<Control>,
127    /// The job history.
128    pub store: Store,
129    /// Every cache volume of the config, with size 0: the API fills in the sizes it measured. Replaced when the
130    /// config is.
131    pub caches: watch::Receiver<Vec<api::CacheVolume>>,
132}
133
134/// Where the engine runs, besides its config and its GitHub and Docker clients.
135#[derive(Debug)]
136pub struct Setup {
137    /// The names of what runnerd creates in Docker.
138    pub names: Names,
139    /// This machine, which names its runners, so a restarted daemon recognizes them.
140    pub identity: HostIdentity,
141    /// The slice's `MemoryMax`.
142    pub memory_max: Option<u64>,
143    /// The job history.
144    pub store: Store,
145    /// The repositories as the last run announced them (`Store::announced_repositories`), so the first publish
146    /// announces only the difference, including the repositories the config no longer has.
147    pub announced: Vec<api::RepositoryStatus>,
148    /// The repositories recorded paused (`Store::paused_repositories`): they stay paused from the first tick.
149    pub paused: Vec<String>,
150    /// Where the cgroup and proc file systems are.
151    pub paths: Paths,
152}
153
154/// A slot's runner and its container.
155#[derive(Debug, Clone)]
156struct Occupant {
157    repository: String,
158    runner_name: String,
159    runner_id: u64,
160    container: String,
161    busy: bool,
162    started_at: SystemTime,
163    /// Its container's memory cap, which a reload does not change.
164    memory: u64,
165    /// The job it took, once its log capture found it.
166    job: Option<u64>,
167}
168
169impl Occupant {
170    fn running_job(&self) -> api::RunningJob {
171        api::RunningJob {
172            repository: self.repository.clone(),
173            runner: self.runner_name.clone(),
174            job: self.job,
175        }
176    }
177}
178
179#[derive(Debug, Default)]
180struct RepositoryState {
181    queued: Vec<QueuedJob>,
182    /// Whether the last poll of its queued jobs succeeded.
183    listed: bool,
184    held_until: Option<SystemTime>,
185    /// Runners in a row that ended without taking a job.
186    failures: u32,
187}
188
189/// The scheduler with its side effects, on the real GitHub and Docker unless a test passes fakes.
190#[derive(Debug)]
191pub struct Engine<G = Github, D = Docker> {
192    config: Config,
193    github: G,
194    docker: Arc<D>,
195    names: Names,
196    isolation: &'static dyn Isolation,
197    identity: HostIdentity,
198    runner_prefix: String,
199    memory_max: Option<u64>,
200    slots: BTreeMap<String, Vec<Option<Occupant>>>,
201    repositories: BTreeMap<String, RepositoryState>,
202    reseeds: BTreeMap<String, JoinHandle<Result<String, docker::Error>>>,
203    /// The pulled images of the config's profiles, by reference.
204    images: BTreeMap<String, ImageState>,
205    /// The memory cap each running reseed's container started with, which a reload does not change.
206    reseed_memory: BTreeMap<String, u64>,
207    /// Profiles the API asked to reseed: they start nothing until their slots are free and the reseed ran.
208    reseed_requested: BTreeSet<String>,
209    /// Repositories a reload removed, polled until none of this machine's runners is registered on them.
210    draining: BTreeSet<String>,
211    /// Profiles a reload removed, kept while their runners or reseed still run.
212    retired: BTreeMap<String, crate::config::Profile>,
213    /// Profiles whose slot 1 ran a job since their last reseed check.
214    reseed_due: BTreeSet<String>,
215    /// Whether a job ended since the shared caches were last checked.
216    trim_due: bool,
217    /// The running check of the shared caches: Docker takes many seconds to measure volumes.
218    trim: Option<JoinHandle<Result<Trimmed, docker::Error>>>,
219    mirror_up: bool,
220    /// The memory cap the running mirror started with.
221    mirror_memory: u64,
222    rate_limit_logged: bool,
223    /// Runners without a container that GitHub still holds busy, already logged.
224    busy_orphans: BTreeSet<u64>,
225    status: watch::Sender<api::Status>,
226    control: Arc<Control>,
227    commands: Commands,
228    caches: watch::Sender<Vec<api::CacheVolume>>,
229    store: Store,
230    paths: Paths,
231    /// The containers the sampler reads.
232    tracked: watch::Sender<Vec<Tracked>>,
233    found: (mpsc::UnboundedSender<Found>, mpsc::UnboundedReceiver<Found>),
234    captures: BTreeMap<String, JoinHandle<()>>,
235    /// Who held which address, for the mirror's log.
236    leases: Leases,
237    /// How far the mirror's log was counted.
238    mirror_until: MirrorPosition,
239    /// Jobs recorded queued and not yet taken or closed, with their repository.
240    queued_jobs: BTreeMap<u64, String>,
241    /// Runners of runs a forced pause cancelled, with when they are removed if they still run.
242    cancelling: BTreeMap<String, tokio::time::Instant>,
243    /// Jobs whose runner is gone, waiting for GitHub to report them completed.
244    closing: BTreeMap<u64, Closing>,
245    last_prune: Option<SystemTime>,
246}
247
248fn unix_secs(time: SystemTime) -> u64 {
249    time.duration_since(SystemTime::UNIX_EPOCH)
250        .unwrap_or_default()
251        .as_secs()
252}
253
254impl<G: GithubApi, D: DockerApi + 'static> Engine<G, D> {
255    /// An engine for a validated config. `identity` names this machine's runners, so a restarted daemon recognizes
256    /// them; `memory_max` is the slice's.
257    pub fn new(config: Config, github: G, docker: D, setup: Setup) -> (Engine<G, D>, Handle) {
258        let Setup {
259            names,
260            identity,
261            memory_max,
262            store,
263            announced,
264            paused,
265            paths,
266        } = setup;
267        let slots = config
268            .profiles
269            .iter()
270            .map(|(name, profile)| (name.clone(), vec![None; profile.slots as usize]))
271            .collect();
272        let repositories = config
273            .repositories
274            .iter()
275            .map(|repository| (repository.name.clone(), RepositoryState::default()))
276            .collect();
277        let (control, commands) = Control::new(
278            config
279                .repositories
280                .iter()
281                .map(|repository| repository.name.clone()),
282            paused,
283            store.clone(),
284        );
285        let control = Arc::new(control);
286        let isolation = isolation::backend(config.host.isolation);
287        let (status, receiver) = watch::channel(api::Status {
288            version: env!("CARGO_PKG_VERSION").to_owned(),
289            shutting_down: false,
290            budget: api::Budget {
291                slice: config.host.slice.clone(),
292                memory_max,
293                memory_in_use: 0,
294            },
295            profiles: Vec::new(),
296            // What clients resuming after the last run's events have, so the first publish announces the difference.
297            repositories: announced,
298        });
299        let (caches, caches_receiver) = watch::channel(cache::volumes(&names, &config));
300        // Unknown: an earlier run may have left the mirror up, with the configured cap at best.
301        let mirror_memory = if config.caches.mirror.enabled {
302            config.caches.mirror.memory.0
303        } else {
304            0
305        };
306        let mut engine = Engine {
307            config,
308            github,
309            docker: Arc::new(docker),
310            names,
311            isolation,
312            runner_prefix: identity.prefix(),
313            identity,
314            memory_max,
315            slots,
316            repositories,
317            reseeds: BTreeMap::new(),
318            images: BTreeMap::new(),
319            reseed_memory: BTreeMap::new(),
320            reseed_requested: BTreeSet::new(),
321            retired: BTreeMap::new(),
322            draining: BTreeSet::new(),
323            reseed_due: BTreeSet::new(),
324            // So the caches are measured once at start.
325            trim_due: true,
326            trim: None,
327            // Unknown: an earlier run may have left it up, so the first idle tick stops it.
328            mirror_up: true,
329            mirror_memory,
330            rate_limit_logged: false,
331            busy_orphans: BTreeSet::new(),
332            status,
333            control: control.clone(),
334            commands,
335            caches,
336            store: store.clone(),
337            paths,
338            tracked: watch::channel(Vec::new()).0,
339            found: mpsc::unbounded_channel(),
340            captures: BTreeMap::new(),
341            leases: Leases::default(),
342            mirror_until: MirrorPosition::default(),
343            queued_jobs: BTreeMap::new(),
344            cancelling: BTreeMap::new(),
345            closing: BTreeMap::new(),
346            last_prune: None,
347        };
348        engine.publish();
349        (
350            engine,
351            Handle {
352                status: receiver,
353                control,
354                store,
355                caches: caches_receiver,
356            },
357        )
358    }
359
360    /// Runs until a shutdown request is carried out.
361    pub async fn run(mut self) {
362        tokio::spawn(metrics::run(
363            Sampler::new(self.paths.clone(), &self.config.host.slice),
364            self.tracked.subscribe(),
365            self.store.clone(),
366        ));
367        // Until Docker answers: scheduling without knowing the containers already running would start a second
368        // runner on a slot's volume and delete the running one's registration.
369        while let Err(error) = self.recover().await {
370            warn!("cannot list the containers an earlier run left; trying again in 5s: {error}");
371            if self.control.mode() != Shutdown::No {
372                return;
373            }
374            tokio::select! {
375                () = tokio::time::sleep(Duration::from_secs(5)) => {}
376                () = self.control.wake.notified() => {}
377            }
378        }
379        loop {
380            let mode = self.control.mode();
381            self.tick(mode).await;
382            self.publish();
383            let idle = self.slots.values().flatten().all(Option::is_none);
384            if mode != Shutdown::No && idle && self.reseeds.is_empty() && self.trim.is_none() {
385                break;
386            }
387            let interval = if mode == Shutdown::No {
388                self.config.host.poll_interval
389            } else {
390                Duration::from_secs(5)
391            };
392            tokio::select! {
393                () = tokio::time::sleep(interval) => {}
394                () = self.control.wake.notified() => {}
395                Some(command) = self.commands.recv() => {
396                    self.command(command).await;
397                    self.publish();
398                }
399            }
400        }
401        if let Err(error) = cache::mirror_down(&*self.docker, &self.names).await {
402            warn!("cannot stop the mirror: {error}");
403        }
404        info!("stopped");
405    }
406
407    fn occupied(&self) -> impl Iterator<Item = (SlotRef, &Occupant)> {
408        self.slots.iter().flat_map(|(profile, slots)| {
409            slots.iter().enumerate().filter_map(|(index, slot)| {
410                slot.as_ref().map(|occupant| {
411                    (
412                        SlotRef {
413                            profile: profile.clone(),
414                            index,
415                        },
416                        occupant,
417                    )
418                })
419            })
420        })
421    }
422
423    /// Drops the free slots past each profile's configured number, and the profiles a reload removed once nothing
424    /// runs in them.
425    fn drop_closed_slots(&mut self) {
426        for (name, slots) in &mut self.slots {
427            let count = self
428                .config
429                .profiles
430                .get(name)
431                .map_or(0, |profile| profile.slots as usize);
432            while slots.len() > count && slots.last().is_some_and(Option::is_none) {
433                slots.pop();
434            }
435        }
436        let profiles = &self.config.profiles;
437        self.slots
438            .retain(|name, slots| !slots.is_empty() || profiles.contains_key(name));
439        let (slots, reseeds) = (&self.slots, &self.reseeds);
440        self.retired
441            .retain(|name, _| slots.contains_key(name) || reseeds.contains_key(name));
442    }
443}
444
445#[cfg(test)]
446mod tests;
runnerd_core/scheduler/engine.rs

runnerd_core/scheduler/
engine.rs
