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 admission;
14mod caches;
15mod control;
16mod history;
17mod images;
18mod records;
19mod recover;
20mod runners;
21mod status;
22mod tick;
23
24use std::collections::{BTreeMap, BTreeSet};
25use std::sync::Arc;
26use std::time::{Duration, SystemTime};
27
28use tokio::sync::{mpsc, watch};
29use tokio::task::JoinHandle;
30use tracing::{info, warn};
31
32pub use self::control::{
33    Command, Commands, Commit, Control, ControlError, RunningJobs, Shutdown, Target,
34};
35use self::history::{Closing, Found};
36use self::images::ImageState;
37pub use self::records::RecordSelection;
38use super::SlotRef;
39use crate::admission::Key;
40use crate::api;
41use crate::cache::{self, Names, Trimmed};
42use crate::config::Config;
43use crate::docker::{self, Docker, DockerApi};
44use crate::github::{Github, GithubApi, QueuedJob};
45use crate::isolation::{self, Isolation};
46use crate::observe::metrics::{self, Paths, Sampler, Tracked};
47use crate::observe::mirror::Leases;
48use crate::store::{Announced, MirrorPosition, Store};
49
50/// This machine as its runner names carry it: `runnerd-<id>-<host>-<profile>-<slot>-<secs>`. Only `runnerd-<id>-`
51/// identifies the machine, so a changed host name still matches the runners started before the change, and two
52/// machines with the same or overlapping host names never take each other's runners for their own.
53#[derive(Debug, Clone, PartialEq, Eq)]
54pub struct HostIdentity {
55    /// 8 hex digits: a hash of the systemd machine ID, never the ID itself, as machine-id(5) asks.
56    id: String,
57    /// The host name in lower case, letters, digits and dashes, at most 20 characters: for people reading names.
58    host: String,
59}
60
61impl HostIdentity {
62    /// This machine's identity, from `/etc/machine-id` and its host name.
63    pub fn read() -> std::io::Result<HostIdentity> {
64        let hostname = std::fs::read_to_string("/proc/sys/kernel/hostname").unwrap_or_default();
65        let machine_id = std::fs::read_to_string("/etc/machine-id")
66            .or_else(|_| std::fs::read_to_string("/var/lib/dbus/machine-id"))
67            .map_err(|error| {
68                std::io::Error::new(
69                    error.kind(),
70                    format!("cannot read /etc/machine-id: {error}"),
71                )
72            })?;
73        HostIdentity::new(&hostname, &machine_id).ok_or_else(|| {
74            std::io::Error::new(
75                std::io::ErrorKind::InvalidData,
76                "/etc/machine-id does not hold a machine ID",
77            )
78        })
79    }
80
81    /// The identity for a host name and a machine ID; `None` unless the ID is 32 hex digits.
82    pub fn new(hostname: &str, machine_id: &str) -> Option<HostIdentity> {
83        let machine_id = machine_id.trim();
84        if machine_id.len() != 32 || !machine_id.chars().all(|c| c.is_ascii_hexdigit()) {
85            return None;
86        }
87        // FNV-1a: stable across Rust releases, unlike std's hasher, so a new build recognizes the runners of the old.
88        let hash = format!("runnerd:{}", machine_id.to_ascii_lowercase())
89            .bytes()
90            .fold(0xcbf2_9ce4_8422_2325_u64, |hash, byte| {
91                (hash ^ u64::from(byte)).wrapping_mul(0x0000_0100_0000_01b3)
92            });
93        let host: String = hostname
94            .trim()
95            .chars()
96            .map(|c| {
97                if c.is_ascii_alphanumeric() {
98                    c.to_ascii_lowercase()
99                } else {
100                    '-'
101                }
102            })
103            .take(20)
104            .collect();
105        let host = host.trim_matches('-');
106        Some(HostIdentity {
107            id: format!("{:08x}", hash >> 32),
108            host: if host.is_empty() { "host" } else { host }.to_owned(),
109        })
110    }
111
112    /// What every runner name of this machine starts with. The ID's fixed length keeps it from starting another
113    /// machine's names.
114    pub fn prefix(&self) -> String {
115        format!("runnerd-{}-", self.id)
116    }
117
118    /// The name of a runner, and of its container, started in `slot` (1-based) of `profile` at `secs`.
119    pub fn runner_name(&self, profile: &str, slot: usize, secs: u64) -> String {
120        format!("{}{}-{profile}-{slot}-{secs}", self.prefix(), self.host)
121    }
122}
123
124/// What the API needs from a running engine.
125#[derive(Debug, Clone)]
126pub struct Handle {
127    /// The latest status, replaced after every tick.
128    pub status: watch::Receiver<api::Status>,
129    /// Pause, resume and shutdown requests.
130    pub control: Arc<Control>,
131    /// The job history.
132    pub store: Store,
133    /// Every cache volume of the config, with size 0: the API fills in the sizes it measured. Replaced when the
134    /// config is.
135    pub caches: watch::Receiver<Vec<api::CacheVolume>>,
136    /// The records of failed jobs, replaced whenever one is made or removed.
137    pub records: watch::Receiver<api::Records>,
138}
139
140/// Where the engine runs, besides its config and its GitHub and Docker clients.
141#[derive(Debug)]
142pub struct Setup {
143    /// The names of what runnerd creates in Docker.
144    pub names: Names,
145    /// This machine, which names its runners, so a restarted daemon recognizes them.
146    pub identity: HostIdentity,
147    /// The slice's `MemoryMax`.
148    pub memory_max: Option<u64>,
149    /// The job history.
150    pub store: Store,
151    /// The profiles and repositories as the last run announced them (`Store::announced`), so the first publish
152    /// announces only the difference, including what the config no longer has.
153    pub announced: Announced,
154    /// The repositories recorded paused (`Store::paused_repositories`): they stay paused from the first tick.
155    pub paused: Vec<String>,
156    /// Where the cgroup and proc file systems are.
157    pub paths: Paths,
158}
159
160/// A slot's runner and its container.
161#[derive(Debug, Clone)]
162struct Occupant {
163    repository: String,
164    runner_name: String,
165    runner_id: u64,
166    container: String,
167    busy: bool,
168    started_at: SystemTime,
169    /// Its container's memory cap, which a reload does not change.
170    memory: u64,
171    /// How it was admitted, with what it reserved of the budget; `None` for one a restart adopted, which reserves
172    /// its cap.
173    admission: Option<api::Admission>,
174    /// The job it took, once its log capture found it.
175    job: Option<u64>,
176    /// That job as admission tells jobs apart, once the history had its record.
177    key: Option<Key>,
178}
179
180impl Occupant {
181    /// What it reserves of the budget under measured admission: its cap if a restart adopted it.
182    fn reserved(&self) -> u64 {
183        self.admission
184            .map_or(self.memory, |admission| admission.reserved)
185    }
186
187    fn running_job(&self) -> api::RunningJob {
188        api::RunningJob {
189            repository: self.repository.clone(),
190            runner: self.runner_name.clone(),
191            job: self.job,
192        }
193    }
194}
195
196#[derive(Debug, Default)]
197struct RepositoryState {
198    queued: Vec<QueuedJob>,
199    /// Whether the last poll of its queued jobs succeeded.
200    listed: bool,
201    held_until: Option<SystemTime>,
202    /// Runners in a row that ended without taking a job.
203    failures: u32,
204}
205
206/// The scheduler with its side effects, on the real GitHub and Docker unless a test passes fakes.
207#[derive(Debug)]
208pub struct Engine<G = Github, D = Docker> {
209    config: Config,
210    /// Shared with the listings a tick reads while it answers commands.
211    github: Arc<G>,
212    docker: Arc<D>,
213    names: Names,
214    isolation: &'static dyn Isolation,
215    identity: HostIdentity,
216    runner_prefix: String,
217    memory_max: Option<u64>,
218    slots: BTreeMap<String, Vec<Option<Occupant>>>,
219    repositories: BTreeMap<String, RepositoryState>,
220    reseeds: BTreeMap<String, JoinHandle<Result<String, docker::Error>>>,
221    /// The pulled images of the config's profiles, by reference.
222    images: BTreeMap<String, ImageState>,
223    /// The memory cap each running reseed's container started with, which a reload does not change.
224    reseed_memory: BTreeMap<String, u64>,
225    /// Profiles the API asked to reseed: they start nothing until their slots are free and the reseed ran.
226    reseed_requested: BTreeSet<String>,
227    /// Repositories a reload removed, polled until none of this machine's runners is registered on them.
228    draining: BTreeSet<String>,
229    /// Profiles a reload removed, kept while their runners or reseed still run.
230    retired: BTreeMap<String, crate::config::Profile>,
231    /// Profiles whose slot 1 ran a job since their last reseed check.
232    reseed_due: BTreeSet<String>,
233    /// Whether a job ended since the shared caches were last checked.
234    trim_due: bool,
235    /// The running check of the shared caches: Docker takes many seconds to measure volumes.
236    trim: Option<JoinHandle<Result<Trimmed, docker::Error>>>,
237    mirror_up: bool,
238    /// The memory cap the running mirror started with.
239    mirror_memory: u64,
240    rate_limit_logged: bool,
241    /// Runners without a container that GitHub still holds busy, already logged.
242    busy_orphans: BTreeSet<u64>,
243    status: watch::Sender<api::Status>,
244    /// The profiles as the last run announced them, until the first publish compares the profiles with them. Their
245    /// slots are older than the `slot.changed` events clients have, so that publish announces every slot instead.
246    announced_profiles: Option<Vec<api::ProfileStatus>>,
247    control: Arc<Control>,
248    commands: Commands,
249    caches: watch::Sender<Vec<api::CacheVolume>>,
250    store: Store,
251    paths: Paths,
252    /// The containers the sampler reads.
253    tracked: watch::Sender<Vec<Tracked>>,
254    found: (mpsc::UnboundedSender<Found>, mpsc::UnboundedReceiver<Found>),
255    captures: BTreeMap<String, JoinHandle<()>>,
256    /// Who held which address, for the mirror's log.
257    leases: Leases,
258    /// How far the mirror's log was counted.
259    mirror_until: MirrorPosition,
260    /// Jobs recorded queued and not yet taken or closed, with their repository.
261    queued_jobs: BTreeMap<u64, String>,
262    /// Runners of runs a forced pause cancelled, with when they are removed if they still run.
263    cancelling: BTreeMap<String, tokio::time::Instant>,
264    /// Jobs whose runner is gone, waiting for GitHub to report them completed.
265    closing: BTreeMap<u64, Closing>,
266    last_prune: Option<SystemTime>,
267    /// The records of failed jobs and the shells open in them.
268    records: records::RecordState,
269    /// Under measured admission, the predictor and the high-water mark.
270    measured: admission::Measured,
271}
272
273fn unix_secs(time: SystemTime) -> u64 {
274    time.duration_since(SystemTime::UNIX_EPOCH)
275        .unwrap_or_default()
276        .as_secs()
277}
278
279impl<G: GithubApi + 'static, D: DockerApi + 'static> Engine<G, D> {
280    /// An engine for a validated config. `identity` names this machine's runners, so a restarted daemon recognizes
281    /// them; `memory_max` is the slice's.
282    pub fn new(config: Config, github: G, docker: D, setup: Setup) -> (Engine<G, D>, Handle) {
283        let Setup {
284            names,
285            identity,
286            memory_max,
287            store,
288            announced,
289            paused,
290            paths,
291        } = setup;
292        let slots = config
293            .profiles
294            .iter()
295            .map(|(name, profile)| (name.clone(), vec![None; profile.slots as usize]))
296            .collect();
297        let repositories = config
298            .repositories
299            .iter()
300            .map(|repository| (repository.name.clone(), RepositoryState::default()))
301            .collect();
302        let (control, commands) = Control::new(
303            config
304                .repositories
305                .iter()
306                .map(|repository| repository.name.clone()),
307            paused,
308            store.clone(),
309        );
310        let control = Arc::new(control);
311        let (status, receiver) = watch::channel(api::Status {
312            version: env!("CARGO_PKG_VERSION").to_owned(),
313            shutting_down: false,
314            budget: api::Budget {
315                slice: config.host.slice.clone(),
316                memory_max,
317                ..api::Budget::default()
318            },
319            profiles: Vec::new(),
320            // What clients resuming after the last run's events have, so the first publish announces the difference.
321            repositories: announced.repositories,
322            queued: Vec::new(),
323        });
324        let (caches, caches_receiver) = watch::channel(cache::volumes(&names, &config));
325        let (records, records_receiver) = records::RecordState::new();
326        // Unknown: an earlier run may have left the mirror up, with the configured cap at best.
327        let mirror = &config.caches.mirror;
328        let mirror_memory = if mirror.enabled { mirror.memory.0 } else { 0 };
329        let mut engine = Engine {
330            isolation: isolation::backend(config.host.isolation),
331            config,
332            github: Arc::new(github),
333            docker: Arc::new(docker),
334            names,
335            runner_prefix: identity.prefix(),
336            identity,
337            memory_max,
338            slots,
339            repositories,
340            reseeds: BTreeMap::new(),
341            images: BTreeMap::new(),
342            reseed_memory: BTreeMap::new(),
343            reseed_requested: BTreeSet::new(),
344            retired: BTreeMap::new(),
345            draining: BTreeSet::new(),
346            reseed_due: BTreeSet::new(),
347            // So the caches are measured once at start.
348            trim_due: true,
349            trim: None,
350            // Unknown: an earlier run may have left it up, so the first idle tick stops it.
351            mirror_up: true,
352            mirror_memory,
353            rate_limit_logged: false,
354            busy_orphans: BTreeSet::new(),
355            status,
356            announced_profiles: Some(announced.profiles),
357            control: control.clone(),
358            commands,
359            caches,
360            store: store.clone(),
361            paths,
362            tracked: watch::channel(Vec::new()).0,
363            found: mpsc::unbounded_channel(),
364            captures: BTreeMap::new(),
365            leases: Leases::default(),
366            mirror_until: MirrorPosition::default(),
367            queued_jobs: BTreeMap::new(),
368            cancelling: BTreeMap::new(),
369            closing: BTreeMap::new(),
370            last_prune: None,
371            measured: admission::Measured::default(),
372            records,
373        };
374        engine.publish();
375        (
376            engine,
377            Handle {
378                status: receiver,
379                control,
380                store,
381                caches: caches_receiver,
382                records: records_receiver,
383            },
384        )
385    }
386
387    /// Runs until a shutdown request is carried out.
388    pub async fn run(mut self) {
389        tokio::spawn(metrics::run(
390            Sampler::new(self.paths.clone(), &self.config.host.slice),
391            self.tracked.subscribe(),
392            self.store.clone(),
393        ));
394        // Until Docker answers: scheduling without knowing the containers already running would start a second
395        // runner on a slot's volume and delete the running one's registration.
396        while let Err(error) = self.recover().await {
397            warn!("cannot list the containers an earlier run left; trying again in 5s: {error}");
398            if self.control.mode() != Shutdown::No {
399                return;
400            }
401            tokio::select! {
402                () = tokio::time::sleep(Duration::from_secs(5)) => {}
403                () = self.control.wake.notified() => {}
404            }
405        }
406        loop {
407            let mode = self.control.mode();
408            self.tick(mode).await;
409            self.publish();
410            let idle = self.slots.values().flatten().all(Option::is_none);
411            if mode != Shutdown::No
412                && idle
413                && self.reseeds.is_empty()
414                && self.trim.is_none()
415                && self.records.recordings.is_empty()
416            {
417                break;
418            }
419            let interval = if mode == Shutdown::No {
420                self.config.host.poll_interval
421            } else {
422                Duration::from_secs(5)
423            };
424            tokio::select! {
425                () = tokio::time::sleep(interval) => {}
426                () = self.control.wake.notified() => {}
427                Some(command) = self.commands.recv() => {
428                    self.command(command).await;
429                    self.publish();
430                }
431                Some(()) = self.records.finished.recv() => {
432                    self.collect_recordings().await;
433                    self.publish();
434                }
435            }
436        }
437        self.close_shells().await;
438        if let Err(error) = cache::mirror_down(&*self.docker, &self.names).await {
439            warn!("cannot stop the mirror: {error}");
440        }
441        info!("stopped");
442    }
443
444    /// Runs `work` as a task of its own, answering commands and publishing after each until it is done, so a command
445    /// waits for a slow GitHub no longer than for the command before it. The task goes on while a command runs: a
446    /// request's timeout runs on the clock, not on polls.
447    async fn serving<T: Send + 'static>(
448        &mut self,
449        work: impl Future<Output = T> + Send + 'static,
450    ) -> T {
451        let mut task = tokio::spawn(work);
452        loop {
453            tokio::select! {
454                biased;
455                done = &mut task => match done.map_err(tokio::task::JoinError::try_into_panic) {
456                    Ok(output) => return output,
457                    // As if it had panicked here.
458                    Err(Ok(panic)) => std::panic::resume_unwind(panic),
459                    // Only while the runtime shuts down, which drops this task too.
460                    Err(Err(_)) => std::future::pending::<()>().await,
461                },
462                Some(command) = self.commands.recv() => {
463                    self.command(command).await;
464                    self.publish();
465                }
466            }
467        }
468    }
469
470    fn occupied(&self) -> impl Iterator<Item = (SlotRef, &Occupant)> {
471        self.slots.iter().flat_map(|(profile, slots)| {
472            slots.iter().enumerate().filter_map(|(index, slot)| {
473                slot.as_ref().map(|occupant| {
474                    (
475                        SlotRef {
476                            profile: profile.clone(),
477                            index,
478                        },
479                        occupant,
480                    )
481                })
482            })
483        })
484    }
485
486    /// Drops the free slots past each profile's configured number, and the profiles a reload removed once nothing
487    /// runs in them.
488    fn drop_closed_slots(&mut self) {
489        for (name, slots) in &mut self.slots {
490            let count = self
491                .config
492                .profiles
493                .get(name)
494                .map_or(0, |profile| profile.slots as usize);
495            while slots.len() > count && slots.last().is_some_and(Option::is_none) {
496                slots.pop();
497            }
498        }
499        let profiles = &self.config.profiles;
500        self.slots
501            .retain(|name, slots| !slots.is_empty() || profiles.contains_key(name));
502        let (slots, reseeds) = (&self.slots, &self.reseeds);
503        self.retired
504            .retain(|name, _| slots.contains_key(name) || reseeds.contains_key(name));
505    }
506}
507
508#[cfg(test)]
509mod tests;
runnerd_core/scheduler/engine.rs

runnerd_core/scheduler/
engine.rs
