Skip to main content

runnerd_core/
docker.rs

1//! Runner containers through the Docker Engine API: the connection, container, volume and image requests, and the
2//! `DockerApi` trait the engine and the caches call. It decides nothing about scheduling, and a JIT config only
3//! ever reaches a container as a root-only file, never the create request or a log line.
4
5use std::collections::{BTreeMap, BTreeSet, HashMap};
6use std::fmt;
7use std::time::{Duration, Instant};
8
9use bollard::exec::{CreateExecOptions, StartExecResults};
10use std::pin::Pin;
11
12use bollard::models::{
13    ChangeType, ContainerConfig, ContainerCreateBody, CreateImageInfo, HostConfig, Mount,
14    NetworkCreateRequest, SystemInfo, VolumeCreateRequest,
15};
16use bollard::query_parameters::{
17    AttachContainerOptionsBuilder, CommitContainerOptionsBuilder, CreateContainerOptionsBuilder,
18    CreateImageOptionsBuilder, DataUsageOptionsBuilder, DownloadFromContainerOptionsBuilder,
19    EventsOptionsBuilder, InspectContainerOptionsBuilder, ListContainersOptionsBuilder,
20    ListImagesOptionsBuilder, LogsOptionsBuilder, RemoveContainerOptionsBuilder,
21    RemoveImageOptionsBuilder, RemoveVolumeOptionsBuilder, ResizeContainerTTYOptionsBuilder,
22    UploadToContainerOptionsBuilder, WaitContainerOptionsBuilder,
23};
24use futures_util::stream::BoxStream;
25use futures_util::{StreamExt, TryStreamExt};
26use tracing::info;
27
28use crate::config::ByteSize;
29use crate::isolation::Isolation;
30
31pub use bollard::errors::Error;
32
33/// One runner container: a profile's image stage with the profile's limits.
34#[derive(Debug, Clone)]
35pub struct RunnerSpec {
36    /// The container's name; for a runner, also the runner's.
37    pub name: String,
38    /// The image, such as `runnerd-runner:base`.
39    pub image: String,
40    /// The profile, put on the `runnerd.profile` label.
41    pub profile: String,
42    /// The systemd slice the container is placed in.
43    pub slice: String,
44    /// CPUs the container may use.
45    pub cpus: u32,
46    /// Memory the container may use, swap included.
47    pub memory: ByteSize,
48    /// `NAME=value` environment variables.
49    pub env: Vec<String>,
50    /// Volumes to mount.
51    pub mounts: Vec<Mount>,
52    /// The network the container joins; `None` is Docker's default bridge.
53    pub network: Option<String>,
54    /// Labels on top of `runnerd.profile` and `runnerd.isolation`.
55    pub labels: BTreeMap<String, String>,
56    /// Replaces the image's entrypoint; `None` runs the runner.
57    pub entrypoint: Option<Vec<String>>,
58    /// The runner's just-in-time config, which the entrypoint reads from `JIT_CONFIG_PATH`.
59    pub jit_config: Option<JitConfig>,
60}
61
62/// Where `start` puts the just-in-time config in the container: a root-only file, never the create request,
63/// which Docker stores and clients may log.
64pub const JIT_CONFIG_PATH: &str = "/run/runnerd/jitconfig";
65
66/// A runner's just-in-time config from GitHub: credentials, never printed.
67#[derive(Clone)]
68pub struct JitConfig(pub String);
69
70impl fmt::Debug for JitConfig {
71    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
72        f.write_str("JitConfig(<redacted>)")
73    }
74}
75
76/// What `inspect` reports about a runner container.
77#[derive(Debug, Clone, PartialEq, Eq)]
78pub struct RunnerState {
79    /// Whether the container runs.
80    pub running: bool,
81    /// Its exit code, once it stopped.
82    pub exit_code: Option<i64>,
83    /// Whether the kernel killed a process in it for memory, as Docker saw.
84    pub oom_killed: bool,
85    /// True under `unsafe-dind`: a job in the container can take over the host.
86    pub privileged: bool,
87    /// The runtime Docker runs it with, such as `sysbox-runc`.
88    pub runtime: Option<String>,
89    /// Its main process on the host, while it runs.
90    pub pid: Option<u32>,
91    /// Its address on its first network.
92    pub address: Option<String>,
93    /// When it started, in milliseconds since the epoch.
94    pub started_at: Option<u64>,
95    /// When it stopped, in milliseconds since the epoch; `None` while it runs.
96    pub finished_at: Option<u64>,
97    /// Its memory cap in bytes; `None` when it has none.
98    pub memory: Option<u64>,
99    /// Its CPU limit in millicores; `None` when it has none.
100    pub cpus: Option<u64>,
101}
102
103/// What a command run in a container prints on stdout, as it prints it.
104pub type Output = BoxStream<'static, Result<Vec<u8>, Error>>;
105
106/// An image `images` found.
107#[derive(Debug, Clone, PartialEq, Eq)]
108pub struct Image {
109    /// Its first tag, such as `runnerd-record:42`; its ID when it has none.
110    pub reference: String,
111    /// Its labels.
112    pub labels: HashMap<String, String>,
113    /// The disk it takes beyond its parent image: what removing it frees.
114    pub bytes: u64,
115}
116
117/// An image `tagged_images` found.
118#[derive(Debug, Clone, PartialEq, Eq)]
119pub struct TaggedImage {
120    /// Its ID, such as `sha256:224e…`.
121    pub id: String,
122    /// Its tags, such as `ghcr.io/ayagmar/runnerd-runner:node-0.0.29`.
123    pub tags: Vec<String>,
124    /// The registry digests it was pulled by, such as `ghcr.io/ayagmar/runnerd-runner@sha256:9e79…`.
125    pub digests: Vec<String>,
126}
127
128/// A file `files` read from a container.
129#[derive(Debug, Clone, PartialEq, Eq)]
130pub struct File {
131    /// Its name in the directory read, such as `Worker_20261031-120000-utc.log`.
132    pub name: String,
133    /// What it holds, as text: invalid UTF-8 is replaced. Cut short past `FILES_LIMIT`.
134    pub contents: String,
135}
136
137/// What `files` read of a directory.
138#[derive(Debug, Clone, Default, PartialEq, Eq)]
139pub struct Files {
140    /// The files read, in the archive's order.
141    pub files: Vec<File>,
142    /// Whether they are all there and whole: the archive ended as a tar does, and none was cut at `FILES_LIMIT`. A
143    /// file missing from an incomplete read may be in the directory.
144    pub complete: bool,
145}
146
147/// The most of the files `files` keeps, together. What it skips, such as a runner's job log pages below `_diag`,
148/// does not count: it is read through and dropped.
149pub(crate) const FILES_LIMIT: usize = 16 << 20;
150
151/// A container's terminal, attached once it was created and before it started, so none of its output is missed.
152pub struct Tty {
153    /// The container's ID.
154    pub container: String,
155    /// What the terminal shows, as it shows it; it ends when the container's process exits.
156    pub output: BoxStream<'static, Result<Vec<u8>, Error>>,
157    /// What is typed into it.
158    pub input: Pin<Box<dyn tokio::io::AsyncWrite + Send>>,
159}
160
161impl fmt::Debug for Tty {
162    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
163        f.debug_struct("Tty")
164            .field("container", &self.container)
165            .finish_non_exhaustive()
166    }
167}
168
169/// A terminal's size.
170#[derive(Debug, Clone, Copy, PartialEq, Eq)]
171pub struct TtySize {
172    /// Columns.
173    pub cols: u16,
174    /// Rows.
175    pub rows: u16,
176}
177
178/// The CPU shares of a runner's container or a record's shell with `cpus` CPUs: 1024 a CPU, within the 2 to 262144
179/// Docker takes, which it writes as the container's `cpu.weight`. They sit in the slice beside each other, so when the
180/// slice's CPUs are all in use they share them in proportion to their `cpus`, not evenly; a container with work still
181/// gets at most its `cpus` (`nano_cpus`). The mirror keeps Docker's default weight, which every job's pulls go
182/// through, and programs outside the slice share the host's CPUs with the slice by its own weight (ADR 0079).
183pub fn cpu_shares(cpus: u32) -> i64 {
184    (i64::from(cpus) * 1024).clamp(2, 262_144)
185}
186
187/// The create-container request for a runner container under an isolation backend.
188pub fn create_body(spec: &RunnerSpec, isolation: &dyn Isolation) -> ContainerCreateBody {
189    // Clamped to 8 EiB, the most Docker takes.
190    let memory = i64::try_from(spec.memory.0).unwrap_or(i64::MAX);
191    let mut host_config = HostConfig {
192        cgroup_parent: Some(spec.slice.clone()),
193        memory: Some(memory),
194        // Swap equal to memory: the container gets no swap on top of its memory.
195        memory_swap: Some(memory),
196        nano_cpus: Some(i64::from(spec.cpus) * 1_000_000_000),
197        cpu_shares: Some(cpu_shares(spec.cpus)),
198        init: Some(true),
199        mounts: Some(spec.mounts.clone()),
200        network_mode: spec.network.clone(),
201        ..Default::default()
202    };
203    isolation.configure(&mut host_config);
204    let mut labels: HashMap<String, String> = spec.labels.clone().into_iter().collect();
205    labels.insert("runnerd.profile".to_owned(), spec.profile.clone());
206    labels.insert("runnerd.isolation".to_owned(), isolation.name().to_owned());
207    ContainerCreateBody {
208        image: Some(spec.image.clone()),
209        env: Some(spec.env.clone()),
210        entrypoint: spec.entrypoint.clone(),
211        labels: Some(labels),
212        host_config: Some(host_config),
213        ..Default::default()
214    }
215}
216
217/// A container `list` found.
218#[derive(Debug, Clone, PartialEq, Eq)]
219pub struct Container {
220    /// The Docker container ID.
221    pub id: String,
222    /// Without Docker's leading `/`.
223    pub name: String,
224    /// Whether it runs.
225    pub running: bool,
226    /// Its labels.
227    pub labels: HashMap<String, String>,
228}
229
230/// A runner start that failed, with the container it may have left: Docker may have made, or even started, it before
231/// the request failed, as when its answer is lost.
232#[derive(Debug)]
233pub struct StartError {
234    /// The container the start left, by ID, or by name when Docker could not say whether one exists; `None` when
235    /// none does. It may run.
236    pub container: Option<String>,
237    /// Whether Docker may have started the container before the start failed: its runner may have taken a job,
238    /// even if the container is gone now.
239    pub may_have_run: bool,
240    /// Why the start failed.
241    pub error: Error,
242}
243
244impl fmt::Display for StartError {
245    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
246        self.error.fmt(f)
247    }
248}
249
250impl std::error::Error for StartError {
251    fn source(&self) -> Option<&(dyn std::error::Error + 'static)> {
252        Some(&self.error)
253    }
254}
255
256/// For callers that go on whether or not a container was left.
257impl From<StartError> for Error {
258    fn from(failed: StartError) -> Error {
259        failed.error
260    }
261}
262
263/// Whether Docker answered a request with a refusal, so it did nothing: a 4xx answer. Any other error may come after
264/// Docker carried the request out.
265fn refused(error: &Error) -> bool {
266    matches!(
267        error,
268        Error::DockerResponseServerError {
269            status_code: 400..=499,
270            ..
271        }
272    )
273}
274
275/// Whether a Docker error says the Engine did not answer at all: its socket is gone, refuses connections or closed
276/// the connection, as while the machine shuts down. Not an answer the Engine gave, nor a timeout, which a busy Engine
277/// also gives.
278pub fn is_unreachable(error: &Error) -> bool {
279    matches!(
280        error,
281        Error::HyperLegacyError { .. }
282            | Error::HyperResponseError { .. }
283            | Error::IOError { .. }
284            | Error::SocketNotFoundError(_)
285    )
286}
287
288/// Whether Docker answered that it cannot or will not do what was asked: a 4xx, or 501 for what it lacks, such as
289/// the log of a container whose log driver keeps none. Asking again gets the same answer; any other error may not.
290pub fn is_declined(error: &Error) -> bool {
291    matches!(
292        error,
293        Error::DockerResponseServerError {
294            status_code: 400..=499 | 501,
295            ..
296        }
297    )
298}
299
300/// Whether a Docker error says the container, volume or image does not exist.
301pub fn is_not_found(error: &Error) -> bool {
302    matches!(
303        error,
304        Error::DockerResponseServerError {
305            status_code: 404,
306            ..
307        }
308    )
309}
310
311/// Whether Docker refused because it is removing the container already, as after a `docker rm -f`: it says so only in
312/// the words of its 409, which a running container's refusal shares. It may still be stopping.
313pub fn is_being_removed(error: &Error) -> bool {
314    matches!(
315        error,
316        Error::DockerResponseServerError { status_code: 409, message }
317            if message.contains("is already in progress") || message.contains("marked for removal")
318    )
319}
320
321/// Whether a container's log is out of reach because it is gone or going. A removal under way can still fail, so this
322/// says nothing of whether the container will go.
323pub fn is_log_gone(error: &Error) -> bool {
324    is_not_found(error) || is_being_removed(error)
325}
326
327/// The Docker Engine calls the scheduler and the caches make. `Docker` is the real one; the engine's tests use a
328/// fake.
329pub trait DockerApi: Send + Sync {
330    /// Creates and starts a runner container; returns its ID. A start that fails removes what it made, and says
331    /// which container is left when it cannot.
332    fn start(
333        &self,
334        spec: &RunnerSpec,
335        isolation: &dyn Isolation,
336    ) -> impl Future<Output = Result<String, StartError>> + Send;
337    /// A container's state and isolation.
338    fn inspect(&self, container: &str) -> impl Future<Output = Result<RunnerState, Error>> + Send;
339    /// Removes a runner container, stopping it first if it runs.
340    fn remove(&self, container: &str) -> impl Future<Output = Result<(), Error>> + Send;
341    /// Creates and starts a container that is not a runner, such as the mirror; returns its ID.
342    fn run(
343        &self,
344        name: &str,
345        body: ContainerCreateBody,
346    ) -> impl Future<Output = Result<String, Error>> + Send;
347    /// Every container, running or not, that has the label.
348    fn list(&self, label: &str) -> impl Future<Output = Result<Vec<Container>, Error>> + Send;
349    /// The IDs of runner containers (labelled `runnerd.job`) as each stops, from Docker's events. A hint: Docker sends
350    /// an event once at most, none while the stream is down, and keeps none to replay. The stream fails or ends with
351    /// the connection, or when Docker restarts.
352    fn exits(&self) -> BoxStream<'static, Result<String, Error>>;
353    /// Waits for a container to exit; returns its exit code.
354    fn wait(&self, container: &str) -> impl Future<Output = Result<i64, Error>> + Send;
355    /// A container's output so far, stdout and stderr interleaved.
356    fn logs(&self, container: &str) -> impl Future<Output = Result<String, Error>> + Send;
357    /// A container's output since `since` (seconds since the epoch), each line prefixed with Docker's RFC 3339
358    /// timestamp and a space.
359    fn logs_since(
360        &self,
361        container: &str,
362        since: u64,
363    ) -> impl Future<Output = Result<String, Error>> + Send;
364    /// Runs a command as root in a running container; its stdout streams until it exits or the container stops.
365    /// Its stderr is dropped.
366    fn exec(
367        &self,
368        container: &str,
369        command: Vec<String>,
370    ) -> impl Future<Output = Result<Output, Error>> + Send;
371    /// Whether Docker has an image.
372    fn image_present(&self, image: &str) -> impl Future<Output = Result<bool, Error>> + Send;
373    /// Pulls an image, logging its progress every few seconds.
374    fn pull(&self, image: &str) -> impl Future<Output = Result<(), Error>> + Send;
375    /// Creates a bridge network unless it exists.
376    fn ensure_network(&self, name: &str) -> impl Future<Output = Result<(), Error>> + Send;
377    /// Creates a volume unless it exists.
378    fn ensure_volume(&self, name: &str) -> impl Future<Output = Result<(), Error>> + Send;
379    /// Removes a volume; a volume that does not exist is not an error.
380    fn remove_volume(&self, name: &str) -> impl Future<Output = Result<(), Error>> + Send;
381    /// The disk size of every volume Docker knows, by name. Slow: Docker measures each volume.
382    fn volume_sizes(&self) -> impl Future<Output = Result<BTreeMap<String, u64>, Error>> + Send;
383    /// The paths a container's own layer added or changed, not those it deleted.
384    fn changes(&self, container: &str) -> impl Future<Output = Result<Vec<String>, Error>> + Send;
385    /// The bytes a container's own layer holds.
386    fn written(&self, container: &str) -> impl Future<Output = Result<u64, Error>> + Send;
387    /// The files directly in a container's `directory` whose names start with `prefix`, not those in the
388    /// directories below it, read from its filesystem whether it runs or stopped, without starting it.
389    fn files(
390        &self,
391        container: &str,
392        directory: &str,
393        prefix: &str,
394    ) -> impl Future<Output = Result<Files, Error>> + Send;
395    /// Commits a stopped container's own layer as the image `reference` (`repository:tag`) with `labels`; returns
396    /// the disk the image takes beyond its parent.
397    fn commit(
398        &self,
399        container: &str,
400        reference: &str,
401        labels: BTreeMap<String, String>,
402    ) -> impl Future<Output = Result<u64, Error>> + Send;
403    /// Every image with the label.
404    fn images(&self, label: &str) -> impl Future<Output = Result<Vec<Image>, Error>> + Send;
405    /// The ID of the image a reference (tag, digest or ID) names; `None` when Docker does not have it.
406    fn image_id(
407        &self,
408        reference: &str,
409    ) -> impl Future<Output = Result<Option<String>, Error>> + Send;
410    /// Every image Docker has a tag for.
411    fn tagged_images(&self) -> impl Future<Output = Result<Vec<TaggedImage>, Error>> + Send;
412    /// The image ID of every container, running or stopped.
413    fn container_images(&self) -> impl Future<Output = Result<BTreeSet<String>, Error>> + Send;
414    /// Removes an image and the layers only it used; one that does not exist is not an error.
415    fn remove_image(&self, reference: &str) -> impl Future<Output = Result<(), Error>> + Send;
416    /// The free disk where Docker keeps its images and containers.
417    fn free_space(&self) -> impl Future<Output = Result<u64, Error>> + Send;
418    /// Creates a container with a terminal of `size`, attaches to it and starts it. The container's stdin closes
419    /// when `Tty::input` is dropped.
420    fn shell(
421        &self,
422        name: &str,
423        body: ContainerCreateBody,
424        size: TtySize,
425    ) -> impl Future<Output = Result<Tty, Error>> + Send;
426}
427
428/// The local Docker Engine.
429#[derive(Debug)]
430pub struct Docker {
431    api: bollard::Docker,
432}
433
434impl Docker {
435    /// Removes the container a failed start `made`, if any; the start's `error` is the one worth reporting. The
436    /// container stays in the answer unless its removal is confirmed.
437    async fn clean_up(&self, made: Option<String>, error: Error, may_have_run: bool) -> StartError {
438        let container = match made {
439            Some(container) => match self.remove(&container).await {
440                Err(failed) if !is_not_found(&failed) => Some(container),
441                _ => None,
442            },
443            None => None,
444        };
445        StartError {
446            container,
447            may_have_run,
448            error,
449        }
450    }
451
452    /// Connects to the Docker Engine at `DOCKER_HOST`, or its local socket. Only `unix://` hosts are supported.
453    pub fn connect() -> Result<Docker, Error> {
454        connect_to(std::env::var("DOCKER_HOST").ok().as_deref())
455    }
456
457    /// The Engine's system information.
458    pub async fn info(&self) -> Result<SystemInfo, Error> {
459        self.api.info().await
460    }
461
462    /// An image's labels; `None` when Docker does not have it.
463    pub async fn image_labels(
464        &self,
465        image: &str,
466    ) -> Result<Option<HashMap<String, String>>, Error> {
467        match self.api.inspect_image(image).await {
468            Ok(inspect) => Ok(Some(
469                inspect
470                    .config
471                    .and_then(|config| config.labels)
472                    .unwrap_or_default(),
473            )),
474            Err(error) if is_not_found(&error) => Ok(None),
475            Err(error) => Err(error),
476        }
477    }
478
479    /// Asks the image's registry for its manifest, without pulling it, and returns the platforms it is built for,
480    /// such as `linux/amd64`. An error means `pull` could not get it.
481    pub async fn registry_platforms(&self, image: &str) -> Result<Vec<String>, Error> {
482        let inspect = self.api.inspect_registry_image(image, None).await?;
483        Ok(inspect
484            .platforms
485            .into_iter()
486            .filter_map(|platform| Some(format!("{}/{}", platform.os?, platform.architecture?)))
487            // Attestation manifests in an index carry the platform unknown/unknown.
488            .filter(|platform| platform != "unknown/unknown")
489            .collect())
490    }
491}
492
493impl DockerApi for Docker {
494    async fn start(
495        &self,
496        spec: &RunnerSpec,
497        isolation: &dyn Isolation,
498    ) -> Result<String, StartError> {
499        let options = CreateContainerOptionsBuilder::new()
500            .name(&spec.name)
501            .build();
502        let id = match self
503            .api
504            .create_container(Some(options), create_body(spec, isolation))
505            .await
506        {
507            Ok(created) => created.id,
508            Err(error) if refused(&error) => {
509                return Err(StartError {
510                    container: None,
511                    may_have_run: false,
512                    error,
513                });
514            }
515            // Docker may have made it and lost the answer: its name, unique to this start, finds it.
516            Err(error) => {
517                let made = match self.api.inspect_container(&spec.name, None).await {
518                    Ok(found) => Some(found.id.unwrap_or_else(|| spec.name.clone())),
519                    Err(missing) if is_not_found(&missing) => None,
520                    Err(_) => Some(spec.name.clone()),
521                };
522                // Made at most, never started.
523                return Err(self.clean_up(made, error, false).await);
524            }
525        };
526        let uploaded = async {
527            if let Some(jit_config) = &spec.jit_config {
528                let options = UploadToContainerOptionsBuilder::new().path("/").build();
529                let tar = jit_config_tar(jit_config)?;
530                // As a stream: bollard traces a full body's bytes, but a stream only as "(stream)".
531                let tar = futures_util::stream::once(std::future::ready(tar.into()));
532                self.api
533                    .upload_to_container(&id, Some(options), bollard::body_stream(tar))
534                    .await?;
535            }
536            Ok::<(), Error>(())
537        };
538        if let Err(error) = uploaded.await {
539            return Err(self.clean_up(Some(id), error, false).await);
540        }
541        // A start whose answer is lost may have started it.
542        if let Err(error) = self.api.start_container(&id, None).await {
543            return Err(self.clean_up(Some(id), error, true).await);
544        }
545        Ok(id)
546    }
547
548    async fn inspect(&self, container: &str) -> Result<RunnerState, Error> {
549        let response = self.api.inspect_container(container, None).await?;
550        let state = response.state.unwrap_or_default();
551        let host_config = response.host_config.unwrap_or_default();
552        let address = response
553            .network_settings
554            .and_then(|settings| settings.networks)
555            .and_then(|networks| {
556                networks
557                    .into_values()
558                    .find_map(|network| network.ip_address.filter(|address| !address.is_empty()))
559            });
560        Ok(RunnerState {
561            running: state.running.unwrap_or(false),
562            exit_code: state.exit_code,
563            oom_killed: state.oom_killed.unwrap_or(false),
564            privileged: host_config.privileged.unwrap_or(false),
565            runtime: host_config.runtime,
566            pid: state
567                .pid
568                .and_then(|pid| u32::try_from(pid).ok())
569                .filter(|pid| *pid > 0),
570            address,
571            started_at: state.started_at.as_deref().and_then(crate::time::parse_ms),
572            // Docker says 0001-01-01 for one that never stopped, which reads as none.
573            finished_at: state.finished_at.as_deref().and_then(crate::time::parse_ms),
574            memory: host_config
575                .memory
576                .and_then(|memory| u64::try_from(memory).ok())
577                .filter(|memory| *memory > 0),
578            cpus: host_config
579                .nano_cpus
580                .and_then(|nano| u64::try_from(nano / 1_000_000).ok())
581                .filter(|cpus| *cpus > 0),
582        })
583    }
584
585    async fn remove(&self, container: &str) -> Result<(), Error> {
586        // With its anonymous volumes, such as a shell's /state.
587        let options = RemoveContainerOptionsBuilder::new()
588            .force(true)
589            .v(true)
590            .build();
591        self.api.remove_container(container, Some(options)).await
592    }
593
594    async fn run(&self, name: &str, body: ContainerCreateBody) -> Result<String, Error> {
595        let options = CreateContainerOptionsBuilder::new().name(name).build();
596        let id = self.api.create_container(Some(options), body).await?.id;
597        if let Err(error) = self.api.start_container(&id, None).await {
598            let _ = self.remove(&id).await;
599            return Err(error);
600        }
601        Ok(id)
602    }
603
604    async fn list(&self, label: &str) -> Result<Vec<Container>, Error> {
605        let options = ListContainersOptionsBuilder::new()
606            .all(true)
607            .filters(&HashMap::from([("label", vec![label])]))
608            .build();
609        let containers = self.api.list_containers(Some(options)).await?;
610        Ok(containers
611            .into_iter()
612            .map(|container| Container {
613                id: container.id.unwrap_or_default(),
614                name: container
615                    .names
616                    .and_then(|names| names.into_iter().next())
617                    .unwrap_or_default()
618                    .trim_start_matches('/')
619                    .to_owned(),
620                running: container.state
621                    == Some(bollard::models::ContainerSummaryStateEnum::RUNNING),
622                labels: container.labels.unwrap_or_default(),
623            })
624            .collect())
625    }
626
627    fn exits(&self) -> BoxStream<'static, Result<String, Error>> {
628        let filters = HashMap::from([
629            ("type", vec!["container"]),
630            ("event", vec!["die"]),
631            ("label", vec!["runnerd.job"]),
632        ]);
633        let options = EventsOptionsBuilder::new().filters(&filters).build();
634        self.api
635            .events(Some(options))
636            .try_filter_map(|event| std::future::ready(Ok(event.actor.and_then(|actor| actor.id))))
637            .boxed()
638    }
639
640    async fn wait(&self, container: &str) -> Result<i64, Error> {
641        let options = WaitContainerOptionsBuilder::new()
642            .condition("not-running")
643            .build();
644        let mut responses = self.api.wait_container(container, Some(options));
645        match responses.next().await {
646            Some(Ok(response)) => Ok(response.status_code),
647            // Bollard reports a non-zero exit as an error.
648            Some(Err(Error::DockerContainerWaitError { code, .. })) => Ok(code),
649            Some(Err(error)) => Err(error),
650            None => Ok(0),
651        }
652    }
653
654    async fn logs(&self, container: &str) -> Result<String, Error> {
655        let options = LogsOptionsBuilder::new().stdout(true).stderr(true).build();
656        let chunks: Vec<_> = self
657            .api
658            .logs(container, Some(options))
659            .try_collect()
660            .await?;
661        Ok(chunks.iter().map(ToString::to_string).collect())
662    }
663
664    async fn logs_since(&self, container: &str, since: u64) -> Result<String, Error> {
665        let options = LogsOptionsBuilder::new()
666            .stdout(true)
667            .stderr(true)
668            .timestamps(true)
669            .since(i32::try_from(since).unwrap_or(i32::MAX))
670            .build();
671        let chunks: Vec<_> = self
672            .api
673            .logs(container, Some(options))
674            .try_collect()
675            .await?;
676        Ok(chunks.iter().map(ToString::to_string).collect())
677    }
678
679    async fn exec(&self, container: &str, command: Vec<String>) -> Result<Output, Error> {
680        let options = CreateExecOptions {
681            cmd: Some(command),
682            user: Some("root".to_owned()),
683            attach_stdout: Some(true),
684            attach_stderr: Some(true),
685            ..Default::default()
686        };
687        let exec = self.api.create_exec(container, options).await?;
688        match self.api.start_exec(&exec.id, None).await? {
689            StartExecResults::Attached { output, .. } => Ok(output
690                .try_filter_map(|output| {
691                    std::future::ready(Ok(match output {
692                        bollard::container::LogOutput::StdOut { message } => Some(message.to_vec()),
693                        _ => None,
694                    }))
695                })
696                .boxed()),
697            StartExecResults::Detached => Ok(futures_util::stream::empty().boxed()),
698        }
699    }
700
701    async fn image_present(&self, image: &str) -> Result<bool, Error> {
702        Ok(self.image_labels(image).await?.is_some())
703    }
704
705    async fn pull(&self, image: &str) -> Result<(), Error> {
706        let options = CreateImageOptionsBuilder::new().from_image(image).build();
707        let mut stream = self.api.create_image(Some(options), None, None);
708        let mut progress = PullProgress::default();
709        let mut logged = Instant::now();
710        info!("pulling {image}");
711        while let Some(update) = stream.try_next().await? {
712            progress.update(&update);
713            if logged.elapsed() >= PULL_LOG_INTERVAL {
714                logged = Instant::now();
715                info!("pulling {image}: {progress}");
716            }
717        }
718        info!("pulled {image}");
719        Ok(())
720    }
721
722    async fn ensure_network(&self, name: &str) -> Result<(), Error> {
723        match self.api.inspect_network(name, None).await {
724            Err(error) if is_not_found(&error) => {
725                let request = NetworkCreateRequest {
726                    name: name.to_owned(),
727                    driver: Some("bridge".to_owned()),
728                    ..Default::default()
729                };
730                self.api.create_network(request).await.map(|_| ())
731            }
732            other => other.map(|_| ()),
733        }
734    }
735
736    async fn ensure_volume(&self, name: &str) -> Result<(), Error> {
737        // Creating a volume that exists is a no-op in Docker.
738        let request = VolumeCreateRequest {
739            name: Some(name.to_owned()),
740            ..Default::default()
741        };
742        self.api.create_volume(request).await.map(|_| ())
743    }
744
745    async fn remove_volume(&self, name: &str) -> Result<(), Error> {
746        let options = RemoveVolumeOptionsBuilder::new().build();
747        match self.api.remove_volume(name, Some(options)).await {
748            Err(error) if is_not_found(&error) => Ok(()),
749            other => other,
750        }
751    }
752
753    async fn volume_sizes(&self) -> Result<BTreeMap<String, u64>, Error> {
754        // Without a type filter: bollard cannot encode one in the query.
755        let options = DataUsageOptionsBuilder::new().verbose(true).build();
756        let usage = self.api.df(Some(options)).await?;
757        // API 1.53 leaves the items untyped; they are volumes as `GET /volumes` describes them.
758        Ok(usage
759            .volume_usage
760            .and_then(|volumes| volumes.items)
761            .unwrap_or_default()
762            .into_iter()
763            .filter_map(|volume| {
764                let name = volume.get("Name")?.as_str()?.to_owned();
765                let size = volume.get("UsageData")?.get("Size")?.as_u64()?;
766                Some((name, size))
767            })
768            .collect())
769    }
770
771    async fn changes(&self, container: &str) -> Result<Vec<String>, Error> {
772        Ok(self
773            .api
774            .container_changes(container)
775            .await?
776            .unwrap_or_default()
777            .into_iter()
778            .filter(|change| change.kind != ChangeType::_2)
779            .map(|change| change.path)
780            .collect())
781    }
782
783    async fn written(&self, container: &str) -> Result<u64, Error> {
784        let options = InspectContainerOptionsBuilder::new().size(true).build();
785        let response = self.api.inspect_container(container, Some(options)).await?;
786        Ok(response
787            .size_rw
788            .and_then(|bytes| u64::try_from(bytes).ok())
789            .unwrap_or(0))
790    }
791
792    async fn files(&self, container: &str, directory: &str, prefix: &str) -> Result<Files, Error> {
793        let options = DownloadFromContainerOptionsBuilder::new()
794            .path(directory)
795            .build();
796        let mut archive = self.api.download_from_container(container, Some(options));
797        // The archive is read as it arrives, by a blocking reader, so what it skips is never held.
798        let (chunks, received) = tokio::sync::mpsc::channel(4);
799        let prefix = prefix.to_owned();
800        let reader = tokio::task::spawn_blocking(move || {
801            files_in(
802                Chunks {
803                    received,
804                    chunk: Vec::new(),
805                    at: 0,
806                },
807                &prefix,
808            )
809        });
810        let mut failed = None;
811        while let Some(chunk) = archive.next().await {
812            match chunk {
813                // Fails once the reader is done, at the archive's end or `FILES_LIMIT`.
814                Ok(chunk) => {
815                    if chunks.send(chunk.to_vec()).await.is_err() {
816                        break;
817                    }
818                }
819                Err(error) => {
820                    failed = Some(error);
821                    break;
822                }
823            }
824        }
825        drop(chunks);
826        let files = match reader.await {
827            Ok(files) => files,
828            Err(error) => match error.try_into_panic() {
829                Ok(panic) => std::panic::resume_unwind(panic),
830                // Only while the runtime shuts down.
831                Err(_) => Files::default(),
832            },
833        };
834        match failed {
835            Some(error) => Err(error),
836            None => Ok(files),
837        }
838    }
839
840    async fn commit(
841        &self,
842        container: &str,
843        reference: &str,
844        labels: BTreeMap<String, String>,
845    ) -> Result<u64, Error> {
846        let (repository, tag) = reference.rsplit_once(':').unwrap_or((reference, "latest"));
847        let options = CommitContainerOptionsBuilder::new()
848            .container(container)
849            .repo(repository)
850            .tag(tag)
851            .pause(false)
852            .build();
853        let config = ContainerConfig {
854            labels: Some(labels.into_iter().collect()),
855            ..Default::default()
856        };
857        // The image the container runs is the record's parent.
858        let parent = match self.api.inspect_container(container, None).await?.image {
859            Some(parent) => self.api.inspect_image(&parent).await?.size.unwrap_or(0),
860            None => 0,
861        };
862        let id = self.api.commit_container(options, config).await?.id;
863        let size = self.api.inspect_image(&id).await?.size.unwrap_or(0);
864        Ok(u64::try_from(size - parent).unwrap_or(0))
865    }
866
867    async fn images(&self, label: &str) -> Result<Vec<Image>, Error> {
868        let options = ListImagesOptionsBuilder::new()
869            .filters(&HashMap::from([("label", vec![label])]))
870            .build();
871        let images = self.api.list_images(Some(options)).await?;
872        let mut parents: HashMap<String, i64> = HashMap::new();
873        let mut found = Vec::new();
874        for image in images {
875            let parent = if image.parent_id.is_empty() {
876                0
877            } else if let Some(size) = parents.get(&image.parent_id) {
878                *size
879            } else {
880                let size = self
881                    .api
882                    .inspect_image(&image.parent_id)
883                    .await
884                    .ok()
885                    .and_then(|parent| parent.size)
886                    .unwrap_or(0);
887                parents.insert(image.parent_id.clone(), size);
888                size
889            };
890            found.push(Image {
891                reference: image.repo_tags.into_iter().next().unwrap_or(image.id),
892                labels: image.labels,
893                bytes: u64::try_from(image.size - parent).unwrap_or(0),
894            });
895        }
896        Ok(found)
897    }
898
899    async fn image_id(&self, reference: &str) -> Result<Option<String>, Error> {
900        match self.api.inspect_image(reference).await {
901            Ok(inspect) => Ok(inspect.id),
902            Err(error) if is_not_found(&error) => Ok(None),
903            Err(error) => Err(error),
904        }
905    }
906
907    async fn tagged_images(&self) -> Result<Vec<TaggedImage>, Error> {
908        let images = self
909            .api
910            .list_images(Some(ListImagesOptionsBuilder::new().build()))
911            .await?;
912        Ok(images
913            .into_iter()
914            .map(|image| TaggedImage {
915                id: image.id,
916                tags: image
917                    .repo_tags
918                    .into_iter()
919                    .filter(|tag| tag != "<none>:<none>")
920                    .collect(),
921                digests: image
922                    .repo_digests
923                    .into_iter()
924                    .filter(|digest| digest != "<none>@<none>")
925                    .collect(),
926            })
927            .filter(|image| !image.tags.is_empty())
928            .collect())
929    }
930
931    async fn container_images(&self) -> Result<BTreeSet<String>, Error> {
932        let options = ListContainersOptionsBuilder::new().all(true).build();
933        let containers = self.api.list_containers(Some(options)).await?;
934        Ok(containers
935            .into_iter()
936            .filter_map(|container| container.image_id)
937            .collect())
938    }
939
940    async fn remove_image(&self, reference: &str) -> Result<(), Error> {
941        let options = RemoveImageOptionsBuilder::new().build();
942        match self.api.remove_image(reference, Some(options), None).await {
943            Err(error) if is_not_found(&error) => Ok(()),
944            other => other.map(|_| ()),
945        }
946    }
947
948    async fn free_space(&self) -> Result<u64, Error> {
949        let root = self.api.info().await?.docker_root_dir.unwrap_or_default();
950        // Only the directories above Docker's root need to be searchable: the daemon's user cannot read the root.
951        let stat = rustix::fs::statvfs(root.as_str()).map_err(|error| Error::IOError {
952            err: std::io::Error::other(format!("cannot measure the disk of {root}: {error}")),
953        })?;
954        Ok(stat.f_bavail.saturating_mul(stat.f_frsize))
955    }
956
957    async fn shell(
958        &self,
959        name: &str,
960        body: ContainerCreateBody,
961        size: TtySize,
962    ) -> Result<Tty, Error> {
963        let options = CreateContainerOptionsBuilder::new().name(name).build();
964        let id = self.api.create_container(Some(options), body).await?.id;
965        let attached = async {
966            let options = AttachContainerOptionsBuilder::new()
967                .stream(true)
968                .stdin(true)
969                .stdout(true)
970                .stderr(true)
971                .build();
972            let attached = self.api.attach_container(&id, Some(options)).await?;
973            self.api.start_container(&id, None).await?;
974            let resize = ResizeContainerTTYOptionsBuilder::new()
975                .w(i32::from(size.cols))
976                .h(i32::from(size.rows))
977                .build();
978            self.api.resize_container_tty(&id, resize).await?;
979            Ok::<_, Error>(attached)
980        };
981        match attached.await {
982            Ok(attached) => Ok(Tty {
983                container: id,
984                output: attached
985                    .output
986                    .map_ok(|output| output.into_bytes().to_vec())
987                    .boxed(),
988                input: attached.input,
989            }),
990            Err(error) => {
991                let _ = self.remove(&id).await;
992                Err(error)
993            }
994        }
995    }
996}
997
998/// How often `pull` logs its progress.
999const PULL_LOG_INTERVAL: Duration = Duration::from_secs(10);
1000
1001/// A pull's layers so far, from the statuses Docker streams per layer.
1002#[derive(Debug, Default)]
1003struct PullProgress {
1004    layers: BTreeMap<String, Layer>,
1005}
1006
1007#[derive(Debug, Default)]
1008struct Layer {
1009    downloaded: u64,
1010    size: u64,
1011    done: bool,
1012}
1013
1014impl PullProgress {
1015    fn update(&mut self, update: &CreateImageInfo) {
1016        let (Some(id), Some(status)) = (&update.id, &update.status) else {
1017            return;
1018        };
1019        let bytes = |value: Option<i64>| value.and_then(|value| u64::try_from(value).ok());
1020        let detail = update.progress_detail.as_ref();
1021        match status.as_str() {
1022            "Pulling fs layer" | "Waiting" => {
1023                self.layers.entry(id.clone()).or_default();
1024            }
1025            "Downloading" => {
1026                let layer = self.layers.entry(id.clone()).or_default();
1027                layer.downloaded =
1028                    bytes(detail.and_then(|detail| detail.current)).unwrap_or(layer.downloaded);
1029                layer.size = bytes(detail.and_then(|detail| detail.total)).unwrap_or(layer.size);
1030            }
1031            "Download complete" => {
1032                if let Some(layer) = self.layers.get_mut(id) {
1033                    layer.downloaded = layer.size;
1034                }
1035            }
1036            "Pull complete" | "Already exists" => {
1037                let layer = self.layers.entry(id.clone()).or_default();
1038                layer.downloaded = layer.size;
1039                layer.done = true;
1040            }
1041            _ => {}
1042        }
1043    }
1044}
1045
1046impl fmt::Display for PullProgress {
1047    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
1048        let done = self.layers.values().filter(|layer| layer.done).count();
1049        let downloaded: u64 = self.layers.values().map(|layer| layer.downloaded).sum();
1050        let size: u64 = self.layers.values().map(|layer| layer.size).sum();
1051        write!(
1052            f,
1053            "{done} of {} layers done, {} of {} downloaded",
1054            self.layers.len(),
1055            ByteSize(downloaded),
1056            ByteSize(size)
1057        )
1058    }
1059}
1060
1061fn connect_to(docker_host: Option<&str>) -> Result<Docker, Error> {
1062    let api = match docker_host.filter(|host| !host.is_empty()) {
1063        None => bollard::Docker::connect_with_socket_defaults()?,
1064        Some(host) if host.starts_with("unix://") => {
1065            bollard::Docker::connect_with_socket(host, 120, bollard::API_DEFAULT_VERSION)?
1066        }
1067        // Bollard would quietly fall back to the local socket: a different Engine than the one asked for.
1068        Some(host) => {
1069            return Err(Error::UnsupportedURISchemeError {
1070                uri: host.to_owned(),
1071            });
1072        }
1073    };
1074    Ok(Docker { api })
1075}
1076
1077/// A tar archive, to extract at `/`, holding the config as a root-only file at `JIT_CONFIG_PATH`.
1078/// The chunks of an archive as they arrive, for a blocking reader: it ends when no more come.
1079struct Chunks {
1080    received: tokio::sync::mpsc::Receiver<Vec<u8>>,
1081    chunk: Vec<u8>,
1082    at: usize,
1083}
1084
1085impl std::io::Read for Chunks {
1086    fn read(&mut self, buf: &mut [u8]) -> std::io::Result<usize> {
1087        while self.at == self.chunk.len() {
1088            match self.received.blocking_recv() {
1089                Some(chunk) => {
1090                    self.chunk = chunk;
1091                    self.at = 0;
1092                }
1093                None => return Ok(0),
1094            }
1095        }
1096        let n = buf.len().min(self.chunk.len() - self.at);
1097        buf[..n].copy_from_slice(&self.chunk[self.at..self.at + n]);
1098        self.at += n;
1099        Ok(n)
1100    }
1101}
1102
1103/// A reader that notes whether its last read found the end: `tar` reads past a zero block, the archive's end, no
1104/// further, and takes running out between entries as the end too.
1105struct Ends<R> {
1106    inner: R,
1107    ran_out: bool,
1108}
1109
1110impl<R: std::io::Read> std::io::Read for Ends<R> {
1111    fn read(&mut self, buf: &mut [u8]) -> std::io::Result<usize> {
1112        let n = self.inner.read(buf)?;
1113        self.ran_out = n == 0 && !buf.is_empty();
1114        Ok(n)
1115    }
1116}
1117
1118/// The regular files at the top of a directory's `tar` whose names start with `prefix`. Docker roots the archive at
1119/// the directory's own name. Every other entry is read through and dropped, so only the files kept count against
1120/// `FILES_LIMIT`. Complete only once the archive's end marker is read: one that breaks off, between entries too, or
1121/// a file cut at the limit is not; what was read is kept.
1122pub(crate) fn files_in(tar: impl std::io::Read, prefix: &str) -> Files {
1123    let mut tar = Ends {
1124        inner: tar,
1125        ran_out: false,
1126    };
1127    let mut read = Files::default();
1128    let ended = files_until_end(&mut tar, prefix, &mut read.files);
1129    read.complete = ended && !tar.ran_out;
1130    read
1131}
1132
1133/// Reads `tar` into `files` as `files_in` says; whether it got to the archive's end, which `files_in` checks was its
1134/// marker.
1135fn files_until_end(tar: impl std::io::Read, prefix: &str, files: &mut Vec<File>) -> bool {
1136    let mut archive = tar::Archive::new(tar);
1137    let Ok(entries) = archive.entries() else {
1138        return false;
1139    };
1140    let mut kept = 0;
1141    for entry in entries {
1142        let Ok(mut entry) = entry else {
1143            return false;
1144        };
1145        let name = match entry.path() {
1146            Ok(path) if path.components().count() == 2 => path
1147                .file_name()
1148                .map(|name| name.to_string_lossy().into_owned()),
1149            _ => None,
1150        };
1151        let Some(name) =
1152            name.filter(|name| name.starts_with(prefix) && entry.header().entry_type().is_file())
1153        else {
1154            continue;
1155        };
1156        let room = FILES_LIMIT - kept;
1157        let mut contents = Vec::new();
1158        // One byte past the room tells a file cut at the limit from one that fits it exactly.
1159        let whole = std::io::Read::read_to_end(
1160            &mut std::io::Read::take(&mut entry, room as u64 + 1),
1161            &mut contents,
1162        )
1163        .is_ok()
1164            && contents.len() <= room;
1165        contents.truncate(room);
1166        kept += contents.len();
1167        files.push(File {
1168            name,
1169            contents: String::from_utf8_lossy(&contents).into_owned(),
1170        });
1171        if !whole {
1172            return false;
1173        }
1174    }
1175    true
1176}
1177
1178fn jit_config_tar(jit_config: &JitConfig) -> std::io::Result<Vec<u8>> {
1179    let mut tar = tar::Builder::new(Vec::new());
1180    let path = JIT_CONFIG_PATH.trim_start_matches('/');
1181    let mut dir = tar::Header::new_gnu();
1182    dir.set_entry_type(tar::EntryType::Directory);
1183    dir.set_mode(0o700);
1184    dir.set_size(0);
1185    tar.append_data(
1186        &mut dir,
1187        path.rsplit_once('/').map_or("", |(dir, _)| dir),
1188        std::io::empty(),
1189    )?;
1190    let mut file = tar::Header::new_gnu();
1191    file.set_mode(0o400);
1192    file.set_size(jit_config.0.len() as u64);
1193    tar.append_data(&mut file, path, jit_config.0.as_bytes())?;
1194    tar.into_inner()
1195}
1196
1197#[cfg(test)]
1198pub(crate) mod tests;