runnerd_core/
cache.rs

1//! The warm caches runner containers share and keep between jobs: each profile's read-only snapshot, each slot's
2//! volume (its layers' changes and its Docker Engine's data), the shared pnpm store and Maven repository, and the
3//! Docker Hub mirror on the runners' network. Owns their names, mounts and lifecycle; it never decides when a
4//! runner starts, and it never touches a volume a running container uses.
5
6use std::collections::{BTreeMap, HashMap};
7
8use bollard::models::{
9    ContainerCreateBody, EndpointSettings, HostConfig, Mount, MountType, NetworkingConfig,
10};
11
12use crate::api;
13use crate::config::{self, ByteSize};
14use crate::docker::{DockerApi, Error, RunnerSpec, is_not_found};
15use crate::isolation::Isolation;
16
17/// The Docker Hub pull-through cache (`registry` 3.1.2, multi-platform index).
18pub const MIRROR_IMAGE: &str =
19    "registry:3.1.2@sha256:01cd312e67026f06f5d89c3e133a4a9adc7cb9eeb9b1194e0b2954dfd770efd2";
20
21/// The host name runner containers reach the mirror at (`image/daemon.json`).
22const MIRROR_ALIAS: &str = "mirror";
23
24/// The names of everything runnerd creates in Docker, under one prefix: `runnerd`, or another for tests.
25#[derive(Debug, Clone)]
26pub struct Names {
27    prefix: String,
28}
29
30impl Names {
31    /// Names under `prefix`.
32    pub fn new(prefix: &str) -> Names {
33        Names {
34            prefix: prefix.to_owned(),
35        }
36    }
37
38    /// The network runners and the mirror share.
39    pub fn network(&self) -> String {
40        self.prefix.clone()
41    }
42
43    /// The profile's read-only snapshot.
44    pub fn base(&self, profile: &str) -> String {
45        format!("{}-{profile}-base", self.prefix)
46    }
47
48    /// Slot `slot` (1-based) of the profile: its layers' changes and its Docker Engine's data.
49    pub fn slot(&self, profile: &str, slot: usize) -> String {
50        format!("{}-{profile}-slot-{slot}", self.prefix)
51    }
52
53    /// The pnpm store all slots share.
54    pub fn pnpm(&self) -> String {
55        format!("{}-pnpm", self.prefix)
56    }
57
58    /// The Maven repository all slots share.
59    pub fn maven(&self) -> String {
60        format!("{}-maven", self.prefix)
61    }
62
63    /// The container that rebuilds the profile's snapshot.
64    pub fn seed(&self, profile: &str) -> String {
65        format!("{}-{profile}-seed", self.prefix)
66    }
67
68    /// The mirror's container and its storage volume.
69    pub fn mirror(&self) -> String {
70        format!("{}-mirror", self.prefix)
71    }
72}
73
74fn volume(source: String, target: &str, read_only: bool) -> Mount {
75    Mount {
76        target: Some(target.to_owned()),
77        source: Some(source),
78        typ: Some(MountType::VOLUME),
79        read_only: Some(read_only),
80        ..Default::default()
81    }
82}
83
84/// What a runner container in slot `slot` (1-based) of the profile mounts (paths from `image/Dockerfile`).
85pub fn runner_mounts(names: &Names, profile: &str, slot: usize) -> Vec<Mount> {
86    vec![
87        volume(names.base(profile), "/base", true),
88        volume(names.slot(profile, slot), "/state", false),
89        volume(names.pnpm(), "/shared/pnpm", false),
90        volume(names.maven(), "/shared/maven", false),
91    ]
92}
93
94/// The mirror container: Docker Hub through a proxy registry, which refuses pushes, so a job can only make it
95/// fetch what Docker Hub serves. Content not pulled again within the TTL expires. No published port.
96pub fn mirror_body(names: &Names, mirror: &config::Mirror, slice: &str) -> ContainerCreateBody {
97    let memory = i64::try_from(mirror.memory.0).unwrap_or(i64::MAX);
98    ContainerCreateBody {
99        image: Some(MIRROR_IMAGE.to_owned()),
100        env: Some(vec![
101            "REGISTRY_PROXY_REMOTEURL=https://registry-1.docker.io".to_owned(),
102            format!("REGISTRY_PROXY_TTL={}s", mirror.ttl.as_secs()),
103            // Expiry deletes content.
104            "REGISTRY_STORAGE_DELETE_ENABLED=true".to_owned(),
105        ]),
106        labels: Some(HashMap::from([(
107            "runnerd.mirror".to_owned(),
108            String::new(),
109        )])),
110        host_config: Some(HostConfig {
111            cgroup_parent: Some(slice.to_owned()),
112            memory: Some(memory),
113            memory_swap: Some(memory),
114            nano_cpus: Some(i64::from(mirror.cpus) * 1_000_000_000),
115            mounts: Some(vec![volume(names.mirror(), "/var/lib/registry", false)]),
116            network_mode: Some(names.network()),
117            ..Default::default()
118        }),
119        networking_config: Some(NetworkingConfig {
120            endpoints_config: Some(HashMap::from([(
121                names.network(),
122                EndpointSettings {
123                    aliases: Some(vec![MIRROR_ALIAS.to_owned()]),
124                    ..Default::default()
125                },
126            )])),
127        }),
128        ..Default::default()
129    }
130}
131
132/// Creates what a profile's runner in `slot` (1-based) mounts and joins, if missing.
133pub async fn prepare(
134    docker: &impl DockerApi,
135    names: &Names,
136    profile: &str,
137    slot: usize,
138) -> Result<(), Error> {
139    docker.ensure_network(&names.network()).await?;
140    for name in [
141        names.base(profile),
142        names.slot(profile, slot),
143        names.pnpm(),
144        names.maven(),
145    ] {
146        docker.ensure_volume(&name).await?;
147    }
148    Ok(())
149}
150
151/// Starts the mirror unless it runs.
152pub async fn mirror_up(
153    docker: &impl DockerApi,
154    names: &Names,
155    mirror: &config::Mirror,
156    slice: &str,
157) -> Result<(), Error> {
158    match docker.inspect(&names.mirror()).await {
159        Ok(state) if state.running => return Ok(()),
160        // Stopped on its own: started again from scratch, its storage is the volume.
161        Ok(_) => docker.remove(&names.mirror()).await?,
162        Err(error) if is_not_found(&error) => {}
163        Err(error) => return Err(error),
164    }
165    docker.ensure_network(&names.network()).await?;
166    docker.ensure_image(MIRROR_IMAGE).await?;
167    docker
168        .run(&names.mirror(), mirror_body(names, mirror, slice))
169        .await?;
170    Ok(())
171}
172
173/// Stops the mirror; its storage stays.
174pub async fn mirror_down(docker: &impl DockerApi, names: &Names) -> Result<(), Error> {
175    match docker.remove(&names.mirror()).await {
176        Err(error) if is_not_found(&error) => Ok(()),
177        other => other,
178    }
179}
180
181/// What `trim_shared` measured and emptied.
182#[derive(Debug, Clone, PartialEq, Eq)]
183pub struct Trimmed {
184    /// Every volume's size in bytes, by name, as measured before emptying any.
185    pub sizes: BTreeMap<String, u64>,
186    /// The shared caches it emptied.
187    pub emptied: Vec<Emptied>,
188}
189
190/// A shared cache `trim_shared` emptied.
191#[derive(Debug, Clone, PartialEq, Eq)]
192pub struct Emptied {
193    /// The volume.
194    pub name: String,
195    /// How large it had grown.
196    pub size: ByteSize,
197}
198
199/// Empties the shared caches that outgrew their limit; call only while no runner runs.
200pub async fn trim_shared(
201    docker: &impl DockerApi,
202    names: &Names,
203    caches: &config::Caches,
204) -> Result<Trimmed, Error> {
205    let sizes = docker.volume_sizes().await?;
206    let mut emptied = Vec::new();
207    for (name, limit) in [
208        (names.pnpm(), caches.pnpm.max_size),
209        (names.maven(), caches.maven.max_size),
210    ] {
211        if let Some(&size) = sizes.get(&name)
212            && size > limit.0
213        {
214            docker.remove_volume(&name).await?;
215            emptied.push(Emptied {
216                name,
217                size: ByteSize(size),
218            });
219        }
220    }
221    Ok(Trimmed { sizes, emptied })
222}
223
224/// Every cache volume the config's profiles use, what it holds and its limit, with size 0.
225pub fn volumes(names: &Names, config: &config::Config) -> Vec<api::CacheVolume> {
226    let volume = |name: String, kind: api::CacheKind, limit: Option<u64>| api::CacheVolume {
227        name,
228        kind,
229        bytes: 0,
230        limit,
231    };
232    let mut volumes = Vec::new();
233    for (profile, settings) in &config.profiles {
234        volumes.push(volume(
235            names.base(profile),
236            api::CacheKind::Snapshot {
237                profile: profile.clone(),
238            },
239            None,
240        ));
241        for slot in 1..=settings.slots as usize {
242            volumes.push(volume(
243                names.slot(profile, slot),
244                api::CacheKind::Slot {
245                    profile: profile.clone(),
246                    slot,
247                },
248                None,
249            ));
250        }
251    }
252    volumes.push(volume(
253        names.pnpm(),
254        api::CacheKind::Pnpm,
255        Some(config.caches.pnpm.max_size.0),
256    ));
257    volumes.push(volume(
258        names.maven(),
259        api::CacheKind::Maven,
260        Some(config.caches.maven.max_size.0),
261    ));
262    if config.caches.mirror.enabled {
263        volumes.push(volume(names.mirror(), api::CacheKind::Mirror, None));
264    }
265    volumes.sort_by(|a, b| a.name.cmp(&b.name));
266    volumes
267}
268
269/// What `reseed` runs on: the profile's image and limits, its slots and the isolation backend.
270#[derive(Debug)]
271pub struct Reseed<'profile> {
272    /// Where the profile's volumes are.
273    pub names: &'profile Names,
274    /// The profile to reseed.
275    pub profile: &'profile str,
276    /// The profile's runner image, which holds the seed script.
277    pub image: &'profile str,
278    /// How many slots the profile has; each slot's layer changes are dropped.
279    pub slots: usize,
280    /// The systemd slice the seed container runs in.
281    pub slice: &'profile str,
282    /// The profile's CPUs.
283    pub cpus: u32,
284    /// The profile's memory.
285    pub memory: ByteSize,
286    /// Slot 1's layers must have changed at least this much (`caches.reseed_at`).
287    pub min_changes: ByteSize,
288    /// How the seed container is isolated, as runners are.
289    pub isolation: &'profile dyn Isolation,
290}
291
292/// Makes slot 1's layers the profile's new snapshot and drops every slot's own layer changes (`image/seed.sh`),
293/// if slot 1 changed at least `min_changes`. Call only while none of the profile's slots runs. Returns the seed
294/// container's output.
295pub async fn reseed(docker: &impl DockerApi, reseed: &Reseed<'_>) -> Result<String, Error> {
296    let names = reseed.names;
297    let mut mounts = vec![volume(names.base(reseed.profile), "/base", false)];
298    mounts.extend((1..=reseed.slots).map(|slot| {
299        volume(
300            names.slot(reseed.profile, slot),
301            &format!("/workers/{slot}"),
302            false,
303        )
304    }));
305    for mount in &mounts {
306        if let Some(source) = &mount.source {
307            docker.ensure_volume(source).await?;
308        }
309    }
310    let name = names.seed(reseed.profile);
311    let spec = RunnerSpec {
312        name: name.clone(),
313        image: reseed.image.to_owned(),
314        profile: reseed.profile.to_owned(),
315        slice: reseed.slice.to_owned(),
316        cpus: reseed.cpus,
317        memory: reseed.memory,
318        env: Vec::new(),
319        mounts,
320        network: Some("none".to_owned()),
321        labels: BTreeMap::from([("runnerd.seed".to_owned(), String::new())]),
322        entrypoint: Some(vec![
323            "/usr/local/sbin/runnerd-seed".to_owned(),
324            reseed.min_changes.0.to_string(),
325        ]),
326        jit_config: None,
327    };
328    // A seed container left by a daemon that died while seeding.
329    match docker.remove(&name).await {
330        Err(error) if !is_not_found(&error) => return Err(error),
331        _ => {}
332    }
333    let id = docker.start(&spec, reseed.isolation).await?;
334    finish_reseed(docker, &id).await
335}
336
337/// Waits for a seed container to exit, removes it, and returns its output. Also takes over a seed container an
338/// earlier daemon left running: stopping it between its `rm` and `mv` would lose that snapshot layer.
339pub async fn finish_reseed(docker: &impl DockerApi, container: &str) -> Result<String, Error> {
340    let result = async {
341        let code = docker.wait(container).await?;
342        let output = docker.logs(container).await?;
343        if code == 0 {
344            Ok(output)
345        } else {
346            Err(Error::DockerContainerWaitError {
347                error: format!("seeding exited {code}: {}", output.trim()),
348                code,
349            })
350        }
351    }
352    .await;
353    let removed = docker.remove(container).await;
354    let output = result?;
355    removed?;
356    Ok(output)
357}
runnerd_core/cache.rs

runnerd_core/
cache.rs
