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-0.0.28`.
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/// The Docker Engine calls the scheduler and the caches make. `Docker` is the real one; the engine's tests use a
312/// fake.
313pub trait DockerApi: Send + Sync {
314    /// Creates and starts a runner container; returns its ID. A start that fails removes what it made, and says
315    /// which container is left when it cannot.
316    fn start(
317        &self,
318        spec: &RunnerSpec,
319        isolation: &dyn Isolation,
320    ) -> impl Future<Output = Result<String, StartError>> + Send;
321    /// A container's state and isolation.
322    fn inspect(&self, container: &str) -> impl Future<Output = Result<RunnerState, Error>> + Send;
323    /// Removes a runner container, stopping it first if it runs.
324    fn remove(&self, container: &str) -> impl Future<Output = Result<(), Error>> + Send;
325    /// Creates and starts a container that is not a runner, such as the mirror; returns its ID.
326    fn run(
327        &self,
328        name: &str,
329        body: ContainerCreateBody,
330    ) -> impl Future<Output = Result<String, Error>> + Send;
331    /// Every container, running or not, that has the label.
332    fn list(&self, label: &str) -> impl Future<Output = Result<Vec<Container>, Error>> + Send;
333    /// The IDs of runner containers (labelled `runnerd.job`) as each stops, from Docker's events. A hint: Docker sends
334    /// an event once at most, none while the stream is down, and keeps none to replay. The stream fails or ends with
335    /// the connection, or when Docker restarts.
336    fn exits(&self) -> BoxStream<'static, Result<String, Error>>;
337    /// Waits for a container to exit; returns its exit code.
338    fn wait(&self, container: &str) -> impl Future<Output = Result<i64, Error>> + Send;
339    /// A container's output so far, stdout and stderr interleaved.
340    fn logs(&self, container: &str) -> impl Future<Output = Result<String, Error>> + Send;
341    /// A container's output since `since` (seconds since the epoch), each line prefixed with Docker's RFC 3339
342    /// timestamp and a space.
343    fn logs_since(
344        &self,
345        container: &str,
346        since: u64,
347    ) -> impl Future<Output = Result<String, Error>> + Send;
348    /// Runs a command as root in a running container; its stdout streams until it exits or the container stops.
349    /// Its stderr is dropped.
350    fn exec(
351        &self,
352        container: &str,
353        command: Vec<String>,
354    ) -> impl Future<Output = Result<Output, Error>> + Send;
355    /// Whether Docker has an image.
356    fn image_present(&self, image: &str) -> impl Future<Output = Result<bool, Error>> + Send;
357    /// Pulls an image, logging its progress every few seconds.
358    fn pull(&self, image: &str) -> impl Future<Output = Result<(), Error>> + Send;
359    /// Creates a bridge network unless it exists.
360    fn ensure_network(&self, name: &str) -> impl Future<Output = Result<(), Error>> + Send;
361    /// Creates a volume unless it exists.
362    fn ensure_volume(&self, name: &str) -> impl Future<Output = Result<(), Error>> + Send;
363    /// Removes a volume; a volume that does not exist is not an error.
364    fn remove_volume(&self, name: &str) -> impl Future<Output = Result<(), Error>> + Send;
365    /// The disk size of every volume Docker knows, by name. Slow: Docker measures each volume.
366    fn volume_sizes(&self) -> impl Future<Output = Result<BTreeMap<String, u64>, Error>> + Send;
367    /// The paths a container's own layer added or changed, not those it deleted.
368    fn changes(&self, container: &str) -> impl Future<Output = Result<Vec<String>, Error>> + Send;
369    /// The bytes a container's own layer holds.
370    fn written(&self, container: &str) -> impl Future<Output = Result<u64, Error>> + Send;
371    /// The files directly in a container's `directory` whose names start with `prefix`, not those in the
372    /// directories below it, read from its filesystem whether it runs or stopped, without starting it.
373    fn files(
374        &self,
375        container: &str,
376        directory: &str,
377        prefix: &str,
378    ) -> impl Future<Output = Result<Files, Error>> + Send;
379    /// Commits a stopped container's own layer as the image `reference` (`repository:tag`) with `labels`; returns
380    /// the disk the image takes beyond its parent.
381    fn commit(
382        &self,
383        container: &str,
384        reference: &str,
385        labels: BTreeMap<String, String>,
386    ) -> impl Future<Output = Result<u64, Error>> + Send;
387    /// Every image with the label.
388    fn images(&self, label: &str) -> impl Future<Output = Result<Vec<Image>, Error>> + Send;
389    /// The ID of the image a reference (tag, digest or ID) names; `None` when Docker does not have it.
390    fn image_id(
391        &self,
392        reference: &str,
393    ) -> impl Future<Output = Result<Option<String>, Error>> + Send;
394    /// Every image Docker has a tag for.
395    fn tagged_images(&self) -> impl Future<Output = Result<Vec<TaggedImage>, Error>> + Send;
396    /// The image ID of every container, running or stopped.
397    fn container_images(&self) -> impl Future<Output = Result<BTreeSet<String>, Error>> + Send;
398    /// Removes an image and the layers only it used; one that does not exist is not an error.
399    fn remove_image(&self, reference: &str) -> impl Future<Output = Result<(), Error>> + Send;
400    /// The free disk where Docker keeps its images and containers.
401    fn free_space(&self) -> impl Future<Output = Result<u64, Error>> + Send;
402    /// Creates a container with a terminal of `size`, attaches to it and starts it. The container's stdin closes
403    /// when `Tty::input` is dropped.
404    fn shell(
405        &self,
406        name: &str,
407        body: ContainerCreateBody,
408        size: TtySize,
409    ) -> impl Future<Output = Result<Tty, Error>> + Send;
410}
411
412/// The local Docker Engine.
413#[derive(Debug)]
414pub struct Docker {
415    api: bollard::Docker,
416}
417
418impl Docker {
419    /// Removes the container a failed start `made`, if any; the start's `error` is the one worth reporting. The
420    /// container stays in the answer unless its removal is confirmed.
421    async fn clean_up(&self, made: Option<String>, error: Error, may_have_run: bool) -> StartError {
422        let container = match made {
423            Some(container) => match self.remove(&container).await {
424                Err(failed) if !is_not_found(&failed) => Some(container),
425                _ => None,
426            },
427            None => None,
428        };
429        StartError {
430            container,
431            may_have_run,
432            error,
433        }
434    }
435
436    /// Connects to the Docker Engine at `DOCKER_HOST`, or its local socket. Only `unix://` hosts are supported.
437    pub fn connect() -> Result<Docker, Error> {
438        connect_to(std::env::var("DOCKER_HOST").ok().as_deref())
439    }
440
441    /// The Engine's system information.
442    pub async fn info(&self) -> Result<SystemInfo, Error> {
443        self.api.info().await
444    }
445
446    /// An image's labels; `None` when Docker does not have it.
447    pub async fn image_labels(
448        &self,
449        image: &str,
450    ) -> Result<Option<HashMap<String, String>>, Error> {
451        match self.api.inspect_image(image).await {
452            Ok(inspect) => Ok(Some(
453                inspect
454                    .config
455                    .and_then(|config| config.labels)
456                    .unwrap_or_default(),
457            )),
458            Err(error) if is_not_found(&error) => Ok(None),
459            Err(error) => Err(error),
460        }
461    }
462
463    /// Asks the image's registry for its manifest, without pulling it, and returns the platforms it is built for,
464    /// such as `linux/amd64`. An error means `pull` could not get it.
465    pub async fn registry_platforms(&self, image: &str) -> Result<Vec<String>, Error> {
466        let inspect = self.api.inspect_registry_image(image, None).await?;
467        Ok(inspect
468            .platforms
469            .into_iter()
470            .filter_map(|platform| Some(format!("{}/{}", platform.os?, platform.architecture?)))
471            // Attestation manifests in an index carry the platform unknown/unknown.
472            .filter(|platform| platform != "unknown/unknown")
473            .collect())
474    }
475}
476
477impl DockerApi for Docker {
478    async fn start(
479        &self,
480        spec: &RunnerSpec,
481        isolation: &dyn Isolation,
482    ) -> Result<String, StartError> {
483        let options = CreateContainerOptionsBuilder::new()
484            .name(&spec.name)
485            .build();
486        let id = match self
487            .api
488            .create_container(Some(options), create_body(spec, isolation))
489            .await
490        {
491            Ok(created) => created.id,
492            Err(error) if refused(&error) => {
493                return Err(StartError {
494                    container: None,
495                    may_have_run: false,
496                    error,
497                });
498            }
499            // Docker may have made it and lost the answer: its name, unique to this start, finds it.
500            Err(error) => {
501                let made = match self.api.inspect_container(&spec.name, None).await {
502                    Ok(found) => Some(found.id.unwrap_or_else(|| spec.name.clone())),
503                    Err(missing) if is_not_found(&missing) => None,
504                    Err(_) => Some(spec.name.clone()),
505                };
506                // Made at most, never started.
507                return Err(self.clean_up(made, error, false).await);
508            }
509        };
510        let uploaded = async {
511            if let Some(jit_config) = &spec.jit_config {
512                let options = UploadToContainerOptionsBuilder::new().path("/").build();
513                let tar = jit_config_tar(jit_config)?;
514                // As a stream: bollard traces a full body's bytes, but a stream only as "(stream)".
515                let tar = futures_util::stream::once(std::future::ready(tar.into()));
516                self.api
517                    .upload_to_container(&id, Some(options), bollard::body_stream(tar))
518                    .await?;
519            }
520            Ok::<(), Error>(())
521        };
522        if let Err(error) = uploaded.await {
523            return Err(self.clean_up(Some(id), error, false).await);
524        }
525        // A start whose answer is lost may have started it.
526        if let Err(error) = self.api.start_container(&id, None).await {
527            return Err(self.clean_up(Some(id), error, true).await);
528        }
529        Ok(id)
530    }
531
532    async fn inspect(&self, container: &str) -> Result<RunnerState, Error> {
533        let response = self.api.inspect_container(container, None).await?;
534        let state = response.state.unwrap_or_default();
535        let host_config = response.host_config.unwrap_or_default();
536        let address = response
537            .network_settings
538            .and_then(|settings| settings.networks)
539            .and_then(|networks| {
540                networks
541                    .into_values()
542                    .find_map(|network| network.ip_address.filter(|address| !address.is_empty()))
543            });
544        Ok(RunnerState {
545            running: state.running.unwrap_or(false),
546            exit_code: state.exit_code,
547            oom_killed: state.oom_killed.unwrap_or(false),
548            privileged: host_config.privileged.unwrap_or(false),
549            runtime: host_config.runtime,
550            pid: state
551                .pid
552                .and_then(|pid| u32::try_from(pid).ok())
553                .filter(|pid| *pid > 0),
554            address,
555            started_at: state.started_at.as_deref().and_then(crate::time::parse_ms),
556            // Docker says 0001-01-01 for one that never stopped, which reads as none.
557            finished_at: state.finished_at.as_deref().and_then(crate::time::parse_ms),
558            memory: host_config
559                .memory
560                .and_then(|memory| u64::try_from(memory).ok())
561                .filter(|memory| *memory > 0),
562            cpus: host_config
563                .nano_cpus
564                .and_then(|nano| u64::try_from(nano / 1_000_000).ok())
565                .filter(|cpus| *cpus > 0),
566        })
567    }
568
569    async fn remove(&self, container: &str) -> Result<(), Error> {
570        // With its anonymous volumes, such as a shell's /state.
571        let options = RemoveContainerOptionsBuilder::new()
572            .force(true)
573            .v(true)
574            .build();
575        self.api.remove_container(container, Some(options)).await
576    }
577
578    async fn run(&self, name: &str, body: ContainerCreateBody) -> Result<String, Error> {
579        let options = CreateContainerOptionsBuilder::new().name(name).build();
580        let id = self.api.create_container(Some(options), body).await?.id;
581        if let Err(error) = self.api.start_container(&id, None).await {
582            let _ = self.remove(&id).await;
583            return Err(error);
584        }
585        Ok(id)
586    }
587
588    async fn list(&self, label: &str) -> Result<Vec<Container>, Error> {
589        let options = ListContainersOptionsBuilder::new()
590            .all(true)
591            .filters(&HashMap::from([("label", vec![label])]))
592            .build();
593        let containers = self.api.list_containers(Some(options)).await?;
594        Ok(containers
595            .into_iter()
596            .map(|container| Container {
597                id: container.id.unwrap_or_default(),
598                name: container
599                    .names
600                    .and_then(|names| names.into_iter().next())
601                    .unwrap_or_default()
602                    .trim_start_matches('/')
603                    .to_owned(),
604                running: container.state
605                    == Some(bollard::models::ContainerSummaryStateEnum::RUNNING),
606                labels: container.labels.unwrap_or_default(),
607            })
608            .collect())
609    }
610
611    fn exits(&self) -> BoxStream<'static, Result<String, Error>> {
612        let filters = HashMap::from([
613            ("type", vec!["container"]),
614            ("event", vec!["die"]),
615            ("label", vec!["runnerd.job"]),
616        ]);
617        let options = EventsOptionsBuilder::new().filters(&filters).build();
618        self.api
619            .events(Some(options))
620            .try_filter_map(|event| std::future::ready(Ok(event.actor.and_then(|actor| actor.id))))
621            .boxed()
622    }
623
624    async fn wait(&self, container: &str) -> Result<i64, Error> {
625        let options = WaitContainerOptionsBuilder::new()
626            .condition("not-running")
627            .build();
628        let mut responses = self.api.wait_container(container, Some(options));
629        match responses.next().await {
630            Some(Ok(response)) => Ok(response.status_code),
631            // Bollard reports a non-zero exit as an error.
632            Some(Err(Error::DockerContainerWaitError { code, .. })) => Ok(code),
633            Some(Err(error)) => Err(error),
634            None => Ok(0),
635        }
636    }
637
638    async fn logs(&self, container: &str) -> Result<String, Error> {
639        let options = LogsOptionsBuilder::new().stdout(true).stderr(true).build();
640        let chunks: Vec<_> = self
641            .api
642            .logs(container, Some(options))
643            .try_collect()
644            .await?;
645        Ok(chunks.iter().map(ToString::to_string).collect())
646    }
647
648    async fn logs_since(&self, container: &str, since: u64) -> Result<String, Error> {
649        let options = LogsOptionsBuilder::new()
650            .stdout(true)
651            .stderr(true)
652            .timestamps(true)
653            .since(i32::try_from(since).unwrap_or(i32::MAX))
654            .build();
655        let chunks: Vec<_> = self
656            .api
657            .logs(container, Some(options))
658            .try_collect()
659            .await?;
660        Ok(chunks.iter().map(ToString::to_string).collect())
661    }
662
663    async fn exec(&self, container: &str, command: Vec<String>) -> Result<Output, Error> {
664        let options = CreateExecOptions {
665            cmd: Some(command),
666            user: Some("root".to_owned()),
667            attach_stdout: Some(true),
668            attach_stderr: Some(true),
669            ..Default::default()
670        };
671        let exec = self.api.create_exec(container, options).await?;
672        match self.api.start_exec(&exec.id, None).await? {
673            StartExecResults::Attached { output, .. } => Ok(output
674                .try_filter_map(|output| {
675                    std::future::ready(Ok(match output {
676                        bollard::container::LogOutput::StdOut { message } => Some(message.to_vec()),
677                        _ => None,
678                    }))
679                })
680                .boxed()),
681            StartExecResults::Detached => Ok(futures_util::stream::empty().boxed()),
682        }
683    }
684
685    async fn image_present(&self, image: &str) -> Result<bool, Error> {
686        Ok(self.image_labels(image).await?.is_some())
687    }
688
689    async fn pull(&self, image: &str) -> Result<(), Error> {
690        let options = CreateImageOptionsBuilder::new().from_image(image).build();
691        let mut stream = self.api.create_image(Some(options), None, None);
692        let mut progress = PullProgress::default();
693        let mut logged = Instant::now();
694        info!("pulling {image}");
695        while let Some(update) = stream.try_next().await? {
696            progress.update(&update);
697            if logged.elapsed() >= PULL_LOG_INTERVAL {
698                logged = Instant::now();
699                info!("pulling {image}: {progress}");
700            }
701        }
702        info!("pulled {image}");
703        Ok(())
704    }
705
706    async fn ensure_network(&self, name: &str) -> Result<(), Error> {
707        match self.api.inspect_network(name, None).await {
708            Err(error) if is_not_found(&error) => {
709                let request = NetworkCreateRequest {
710                    name: name.to_owned(),
711                    driver: Some("bridge".to_owned()),
712                    ..Default::default()
713                };
714                self.api.create_network(request).await.map(|_| ())
715            }
716            other => other.map(|_| ()),
717        }
718    }
719
720    async fn ensure_volume(&self, name: &str) -> Result<(), Error> {
721        // Creating a volume that exists is a no-op in Docker.
722        let request = VolumeCreateRequest {
723            name: Some(name.to_owned()),
724            ..Default::default()
725        };
726        self.api.create_volume(request).await.map(|_| ())
727    }
728
729    async fn remove_volume(&self, name: &str) -> Result<(), Error> {
730        let options = RemoveVolumeOptionsBuilder::new().build();
731        match self.api.remove_volume(name, Some(options)).await {
732            Err(error) if is_not_found(&error) => Ok(()),
733            other => other,
734        }
735    }
736
737    async fn volume_sizes(&self) -> Result<BTreeMap<String, u64>, Error> {
738        // Without a type filter: bollard cannot encode one in the query.
739        let options = DataUsageOptionsBuilder::new().verbose(true).build();
740        let usage = self.api.df(Some(options)).await?;
741        // API 1.53 leaves the items untyped; they are volumes as `GET /volumes` describes them.
742        Ok(usage
743            .volume_usage
744            .and_then(|volumes| volumes.items)
745            .unwrap_or_default()
746            .into_iter()
747            .filter_map(|volume| {
748                let name = volume.get("Name")?.as_str()?.to_owned();
749                let size = volume.get("UsageData")?.get("Size")?.as_u64()?;
750                Some((name, size))
751            })
752            .collect())
753    }
754
755    async fn changes(&self, container: &str) -> Result<Vec<String>, Error> {
756        Ok(self
757            .api
758            .container_changes(container)
759            .await?
760            .unwrap_or_default()
761            .into_iter()
762            .filter(|change| change.kind != ChangeType::_2)
763            .map(|change| change.path)
764            .collect())
765    }
766
767    async fn written(&self, container: &str) -> Result<u64, Error> {
768        let options = InspectContainerOptionsBuilder::new().size(true).build();
769        let response = self.api.inspect_container(container, Some(options)).await?;
770        Ok(response
771            .size_rw
772            .and_then(|bytes| u64::try_from(bytes).ok())
773            .unwrap_or(0))
774    }
775
776    async fn files(&self, container: &str, directory: &str, prefix: &str) -> Result<Files, Error> {
777        let options = DownloadFromContainerOptionsBuilder::new()
778            .path(directory)
779            .build();
780        let mut archive = self.api.download_from_container(container, Some(options));
781        // The archive is read as it arrives, by a blocking reader, so what it skips is never held.
782        let (chunks, received) = tokio::sync::mpsc::channel(4);
783        let prefix = prefix.to_owned();
784        let reader = tokio::task::spawn_blocking(move || {
785            files_in(
786                Chunks {
787                    received,
788                    chunk: Vec::new(),
789                    at: 0,
790                },
791                &prefix,
792            )
793        });
794        let mut failed = None;
795        while let Some(chunk) = archive.next().await {
796            match chunk {
797                // Fails once the reader is done, at the archive's end or `FILES_LIMIT`.
798                Ok(chunk) => {
799                    if chunks.send(chunk.to_vec()).await.is_err() {
800                        break;
801                    }
802                }
803                Err(error) => {
804                    failed = Some(error);
805                    break;
806                }
807            }
808        }
809        drop(chunks);
810        let files = match reader.await {
811            Ok(files) => files,
812            Err(error) => match error.try_into_panic() {
813                Ok(panic) => std::panic::resume_unwind(panic),
814                // Only while the runtime shuts down.
815                Err(_) => Files::default(),
816            },
817        };
818        match failed {
819            Some(error) => Err(error),
820            None => Ok(files),
821        }
822    }
823
824    async fn commit(
825        &self,
826        container: &str,
827        reference: &str,
828        labels: BTreeMap<String, String>,
829    ) -> Result<u64, Error> {
830        let (repository, tag) = reference.rsplit_once(':').unwrap_or((reference, "latest"));
831        let options = CommitContainerOptionsBuilder::new()
832            .container(container)
833            .repo(repository)
834            .tag(tag)
835            .pause(false)
836            .build();
837        let config = ContainerConfig {
838            labels: Some(labels.into_iter().collect()),
839            ..Default::default()
840        };
841        // The image the container runs is the record's parent.
842        let parent = match self.api.inspect_container(container, None).await?.image {
843            Some(parent) => self.api.inspect_image(&parent).await?.size.unwrap_or(0),
844            None => 0,
845        };
846        let id = self.api.commit_container(options, config).await?.id;
847        let size = self.api.inspect_image(&id).await?.size.unwrap_or(0);
848        Ok(u64::try_from(size - parent).unwrap_or(0))
849    }
850
851    async fn images(&self, label: &str) -> Result<Vec<Image>, Error> {
852        let options = ListImagesOptionsBuilder::new()
853            .filters(&HashMap::from([("label", vec![label])]))
854            .build();
855        let images = self.api.list_images(Some(options)).await?;
856        let mut parents: HashMap<String, i64> = HashMap::new();
857        let mut found = Vec::new();
858        for image in images {
859            let parent = if image.parent_id.is_empty() {
860                0
861            } else if let Some(size) = parents.get(&image.parent_id) {
862                *size
863            } else {
864                let size = self
865                    .api
866                    .inspect_image(&image.parent_id)
867                    .await
868                    .ok()
869                    .and_then(|parent| parent.size)
870                    .unwrap_or(0);
871                parents.insert(image.parent_id.clone(), size);
872                size
873            };
874            found.push(Image {
875                reference: image.repo_tags.into_iter().next().unwrap_or(image.id),
876                labels: image.labels,
877                bytes: u64::try_from(image.size - parent).unwrap_or(0),
878            });
879        }
880        Ok(found)
881    }
882
883    async fn image_id(&self, reference: &str) -> Result<Option<String>, Error> {
884        match self.api.inspect_image(reference).await {
885            Ok(inspect) => Ok(inspect.id),
886            Err(error) if is_not_found(&error) => Ok(None),
887            Err(error) => Err(error),
888        }
889    }
890
891    async fn tagged_images(&self) -> Result<Vec<TaggedImage>, Error> {
892        let images = self
893            .api
894            .list_images(Some(ListImagesOptionsBuilder::new().build()))
895            .await?;
896        Ok(images
897            .into_iter()
898            .map(|image| TaggedImage {
899                id: image.id,
900                tags: image
901                    .repo_tags
902                    .into_iter()
903                    .filter(|tag| tag != "<none>:<none>")
904                    .collect(),
905                digests: image
906                    .repo_digests
907                    .into_iter()
908                    .filter(|digest| digest != "<none>@<none>")
909                    .collect(),
910            })
911            .filter(|image| !image.tags.is_empty())
912            .collect())
913    }
914
915    async fn container_images(&self) -> Result<BTreeSet<String>, Error> {
916        let options = ListContainersOptionsBuilder::new().all(true).build();
917        let containers = self.api.list_containers(Some(options)).await?;
918        Ok(containers
919            .into_iter()
920            .filter_map(|container| container.image_id)
921            .collect())
922    }
923
924    async fn remove_image(&self, reference: &str) -> Result<(), Error> {
925        let options = RemoveImageOptionsBuilder::new().build();
926        match self.api.remove_image(reference, Some(options), None).await {
927            Err(error) if is_not_found(&error) => Ok(()),
928            other => other.map(|_| ()),
929        }
930    }
931
932    async fn free_space(&self) -> Result<u64, Error> {
933        let root = self.api.info().await?.docker_root_dir.unwrap_or_default();
934        // Only the directories above Docker's root need to be searchable: the daemon's user cannot read the root.
935        let stat = rustix::fs::statvfs(root.as_str()).map_err(|error| Error::IOError {
936            err: std::io::Error::other(format!("cannot measure the disk of {root}: {error}")),
937        })?;
938        Ok(stat.f_bavail.saturating_mul(stat.f_frsize))
939    }
940
941    async fn shell(
942        &self,
943        name: &str,
944        body: ContainerCreateBody,
945        size: TtySize,
946    ) -> Result<Tty, Error> {
947        let options = CreateContainerOptionsBuilder::new().name(name).build();
948        let id = self.api.create_container(Some(options), body).await?.id;
949        let attached = async {
950            let options = AttachContainerOptionsBuilder::new()
951                .stream(true)
952                .stdin(true)
953                .stdout(true)
954                .stderr(true)
955                .build();
956            let attached = self.api.attach_container(&id, Some(options)).await?;
957            self.api.start_container(&id, None).await?;
958            let resize = ResizeContainerTTYOptionsBuilder::new()
959                .w(i32::from(size.cols))
960                .h(i32::from(size.rows))
961                .build();
962            self.api.resize_container_tty(&id, resize).await?;
963            Ok::<_, Error>(attached)
964        };
965        match attached.await {
966            Ok(attached) => Ok(Tty {
967                container: id,
968                output: attached
969                    .output
970                    .map_ok(|output| output.into_bytes().to_vec())
971                    .boxed(),
972                input: attached.input,
973            }),
974            Err(error) => {
975                let _ = self.remove(&id).await;
976                Err(error)
977            }
978        }
979    }
980}
981
982/// How often `pull` logs its progress.
983const PULL_LOG_INTERVAL: Duration = Duration::from_secs(10);
984
985/// A pull's layers so far, from the statuses Docker streams per layer.
986#[derive(Debug, Default)]
987struct PullProgress {
988    layers: BTreeMap<String, Layer>,
989}
990
991#[derive(Debug, Default)]
992struct Layer {
993    downloaded: u64,
994    size: u64,
995    done: bool,
996}
997
998impl PullProgress {
999    fn update(&mut self, update: &CreateImageInfo) {
1000        let (Some(id), Some(status)) = (&update.id, &update.status) else {
1001            return;
1002        };
1003        let bytes = |value: Option<i64>| value.and_then(|value| u64::try_from(value).ok());
1004        let detail = update.progress_detail.as_ref();
1005        match status.as_str() {
1006            "Pulling fs layer" | "Waiting" => {
1007                self.layers.entry(id.clone()).or_default();
1008            }
1009            "Downloading" => {
1010                let layer = self.layers.entry(id.clone()).or_default();
1011                layer.downloaded =
1012                    bytes(detail.and_then(|detail| detail.current)).unwrap_or(layer.downloaded);
1013                layer.size = bytes(detail.and_then(|detail| detail.total)).unwrap_or(layer.size);
1014            }
1015            "Download complete" => {
1016                if let Some(layer) = self.layers.get_mut(id) {
1017                    layer.downloaded = layer.size;
1018                }
1019            }
1020            "Pull complete" | "Already exists" => {
1021                let layer = self.layers.entry(id.clone()).or_default();
1022                layer.downloaded = layer.size;
1023                layer.done = true;
1024            }
1025            _ => {}
1026        }
1027    }
1028}
1029
1030impl fmt::Display for PullProgress {
1031    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
1032        let done = self.layers.values().filter(|layer| layer.done).count();
1033        let downloaded: u64 = self.layers.values().map(|layer| layer.downloaded).sum();
1034        let size: u64 = self.layers.values().map(|layer| layer.size).sum();
1035        write!(
1036            f,
1037            "{done} of {} layers done, {} of {} downloaded",
1038            self.layers.len(),
1039            ByteSize(downloaded),
1040            ByteSize(size)
1041        )
1042    }
1043}
1044
1045fn connect_to(docker_host: Option<&str>) -> Result<Docker, Error> {
1046    let api = match docker_host.filter(|host| !host.is_empty()) {
1047        None => bollard::Docker::connect_with_socket_defaults()?,
1048        Some(host) if host.starts_with("unix://") => {
1049            bollard::Docker::connect_with_socket(host, 120, bollard::API_DEFAULT_VERSION)?
1050        }
1051        // Bollard would quietly fall back to the local socket: a different Engine than the one asked for.
1052        Some(host) => {
1053            return Err(Error::UnsupportedURISchemeError {
1054                uri: host.to_owned(),
1055            });
1056        }
1057    };
1058    Ok(Docker { api })
1059}
1060
1061/// A tar archive, to extract at `/`, holding the config as a root-only file at `JIT_CONFIG_PATH`.
1062/// The chunks of an archive as they arrive, for a blocking reader: it ends when no more come.
1063struct Chunks {
1064    received: tokio::sync::mpsc::Receiver<Vec<u8>>,
1065    chunk: Vec<u8>,
1066    at: usize,
1067}
1068
1069impl std::io::Read for Chunks {
1070    fn read(&mut self, buf: &mut [u8]) -> std::io::Result<usize> {
1071        while self.at == self.chunk.len() {
1072            match self.received.blocking_recv() {
1073                Some(chunk) => {
1074                    self.chunk = chunk;
1075                    self.at = 0;
1076                }
1077                None => return Ok(0),
1078            }
1079        }
1080        let n = buf.len().min(self.chunk.len() - self.at);
1081        buf[..n].copy_from_slice(&self.chunk[self.at..self.at + n]);
1082        self.at += n;
1083        Ok(n)
1084    }
1085}
1086
1087/// A reader that notes whether its last read found the end: `tar` reads past a zero block, the archive's end, no
1088/// further, and takes running out between entries as the end too.
1089struct Ends<R> {
1090    inner: R,
1091    ran_out: bool,
1092}
1093
1094impl<R: std::io::Read> std::io::Read for Ends<R> {
1095    fn read(&mut self, buf: &mut [u8]) -> std::io::Result<usize> {
1096        let n = self.inner.read(buf)?;
1097        self.ran_out = n == 0 && !buf.is_empty();
1098        Ok(n)
1099    }
1100}
1101
1102/// The regular files at the top of a directory's `tar` whose names start with `prefix`. Docker roots the archive at
1103/// the directory's own name. Every other entry is read through and dropped, so only the files kept count against
1104/// `FILES_LIMIT`. Complete only once the archive's end marker is read: one that breaks off, between entries too, or
1105/// a file cut at the limit is not; what was read is kept.
1106pub(crate) fn files_in(tar: impl std::io::Read, prefix: &str) -> Files {
1107    let mut tar = Ends {
1108        inner: tar,
1109        ran_out: false,
1110    };
1111    let mut read = Files::default();
1112    let ended = files_until_end(&mut tar, prefix, &mut read.files);
1113    read.complete = ended && !tar.ran_out;
1114    read
1115}
1116
1117/// Reads `tar` into `files` as `files_in` says; whether it got to the archive's end, which `files_in` checks was its
1118/// marker.
1119fn files_until_end(tar: impl std::io::Read, prefix: &str, files: &mut Vec<File>) -> bool {
1120    let mut archive = tar::Archive::new(tar);
1121    let Ok(entries) = archive.entries() else {
1122        return false;
1123    };
1124    let mut kept = 0;
1125    for entry in entries {
1126        let Ok(mut entry) = entry else {
1127            return false;
1128        };
1129        let name = match entry.path() {
1130            Ok(path) if path.components().count() == 2 => path
1131                .file_name()
1132                .map(|name| name.to_string_lossy().into_owned()),
1133            _ => None,
1134        };
1135        let Some(name) =
1136            name.filter(|name| name.starts_with(prefix) && entry.header().entry_type().is_file())
1137        else {
1138            continue;
1139        };
1140        let room = FILES_LIMIT - kept;
1141        let mut contents = Vec::new();
1142        // One byte past the room tells a file cut at the limit from one that fits it exactly.
1143        let whole = std::io::Read::read_to_end(
1144            &mut std::io::Read::take(&mut entry, room as u64 + 1),
1145            &mut contents,
1146        )
1147        .is_ok()
1148            && contents.len() <= room;
1149        contents.truncate(room);
1150        kept += contents.len();
1151        files.push(File {
1152            name,
1153            contents: String::from_utf8_lossy(&contents).into_owned(),
1154        });
1155        if !whole {
1156            return false;
1157        }
1158    }
1159    true
1160}
1161
1162fn jit_config_tar(jit_config: &JitConfig) -> std::io::Result<Vec<u8>> {
1163    let mut tar = tar::Builder::new(Vec::new());
1164    let path = JIT_CONFIG_PATH.trim_start_matches('/');
1165    let mut dir = tar::Header::new_gnu();
1166    dir.set_entry_type(tar::EntryType::Directory);
1167    dir.set_mode(0o700);
1168    dir.set_size(0);
1169    tar.append_data(
1170        &mut dir,
1171        path.rsplit_once('/').map_or("", |(dir, _)| dir),
1172        std::io::empty(),
1173    )?;
1174    let mut file = tar::Header::new_gnu();
1175    file.set_mode(0o400);
1176    file.set_size(jit_config.0.len() as u64);
1177    tar.append_data(&mut file, path, jit_config.0.as_bytes())?;
1178    tar.into_inner()
1179}
1180
1181#[cfg(test)]
1182pub(crate) mod tests;