24 "git.curoverse.com/arvados.git/lib/crunchstat"
25 "git.curoverse.com/arvados.git/sdk/go/arvados"
26 "git.curoverse.com/arvados.git/sdk/go/arvadosclient"
27 "git.curoverse.com/arvados.git/sdk/go/keepclient"
28 "git.curoverse.com/arvados.git/sdk/go/manifest"
30 dockertypes "github.com/docker/docker/api/types"
31 dockercontainer "github.com/docker/docker/api/types/container"
32 dockernetwork "github.com/docker/docker/api/types/network"
33 dockerclient "github.com/docker/docker/client"
36 // IArvadosClient is the minimal Arvados API methods used by crunch-run.
37 type IArvadosClient interface {
38 Create(resourceType string, parameters arvadosclient.Dict, output interface{}) error
39 Get(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
40 Update(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
41 Call(method, resourceType, uuid, action string, parameters arvadosclient.Dict, output interface{}) error
42 CallRaw(method string, resourceType string, uuid string, action string, parameters arvadosclient.Dict) (reader io.ReadCloser, err error)
43 Discovery(key string) (interface{}, error)
46 // ErrCancelled is the error returned when the container is cancelled.
47 var ErrCancelled = errors.New("Cancelled")
49 // IKeepClient is the minimal Keep API methods used by crunch-run.
50 type IKeepClient interface {
51 PutHB(hash string, buf []byte) (string, int, error)
52 ManifestFileReader(m manifest.Manifest, filename string) (keepclient.Reader, error)
55 // NewLogWriter is a factory function to create a new log writer.
56 type NewLogWriter func(name string) io.WriteCloser
58 type RunArvMount func(args []string, tok string) (*exec.Cmd, error)
60 type MkTempDir func(string, string) (string, error)
62 // ThinDockerClient is the minimal Docker client interface used by crunch-run.
63 type ThinDockerClient interface {
64 ContainerAttach(ctx context.Context, container string, options dockertypes.ContainerAttachOptions) (dockertypes.HijackedResponse, error)
65 ContainerCreate(ctx context.Context, config *dockercontainer.Config, hostConfig *dockercontainer.HostConfig,
66 networkingConfig *dockernetwork.NetworkingConfig, containerName string) (dockercontainer.ContainerCreateCreatedBody, error)
67 ContainerStart(ctx context.Context, container string, options dockertypes.ContainerStartOptions) error
68 ContainerStop(ctx context.Context, container string, timeout *time.Duration) error
69 ContainerWait(ctx context.Context, container string) (int64, error)
70 ImageInspectWithRaw(ctx context.Context, image string) (dockertypes.ImageInspect, []byte, error)
71 ImageLoad(ctx context.Context, input io.Reader, quiet bool) (dockertypes.ImageLoadResponse, error)
72 ImageRemove(ctx context.Context, image string, options dockertypes.ImageRemoveOptions) ([]dockertypes.ImageDeleteResponseItem, error)
75 // ThinDockerClientProxy is a proxy implementation of ThinDockerClient
76 // that executes the docker requests on dockerclient.Client
77 type ThinDockerClientProxy struct {
78 Docker *dockerclient.Client
81 // ContainerAttach invokes dockerclient.Client.ContainerAttach
82 func (proxy ThinDockerClientProxy) ContainerAttach(ctx context.Context, container string, options dockertypes.ContainerAttachOptions) (dockertypes.HijackedResponse, error) {
83 return proxy.Docker.ContainerAttach(ctx, container, options)
86 // ContainerCreate invokes dockerclient.Client.ContainerCreate
87 func (proxy ThinDockerClientProxy) ContainerCreate(ctx context.Context, config *dockercontainer.Config, hostConfig *dockercontainer.HostConfig,
88 networkingConfig *dockernetwork.NetworkingConfig, containerName string) (dockercontainer.ContainerCreateCreatedBody, error) {
89 return proxy.Docker.ContainerCreate(ctx, config, hostConfig, networkingConfig, containerName)
92 // ContainerStart invokes dockerclient.Client.ContainerStart
93 func (proxy ThinDockerClientProxy) ContainerStart(ctx context.Context, container string, options dockertypes.ContainerStartOptions) error {
94 return proxy.Docker.ContainerStart(ctx, container, options)
97 // ContainerStop invokes dockerclient.Client.ContainerStop
98 func (proxy ThinDockerClientProxy) ContainerStop(ctx context.Context, container string, timeout *time.Duration) error {
99 return proxy.Docker.ContainerStop(ctx, container, timeout)
102 // ContainerWait invokes dockerclient.Client.ContainerWait
103 func (proxy ThinDockerClientProxy) ContainerWait(ctx context.Context, container string) (int64, error) {
104 return proxy.Docker.ContainerWait(ctx, container)
107 // ImageInspectWithRaw invokes dockerclient.Client.ImageInspectWithRaw
108 func (proxy ThinDockerClientProxy) ImageInspectWithRaw(ctx context.Context, image string) (dockertypes.ImageInspect, []byte, error) {
109 return proxy.Docker.ImageInspectWithRaw(ctx, image)
112 // ImageLoad invokes dockerclient.Client.ImageLoad
113 func (proxy ThinDockerClientProxy) ImageLoad(ctx context.Context, input io.Reader, quiet bool) (dockertypes.ImageLoadResponse, error) {
114 return proxy.Docker.ImageLoad(ctx, input, quiet)
117 // ImageRemove invokes dockerclient.Client.ImageRemove
118 func (proxy ThinDockerClientProxy) ImageRemove(ctx context.Context, image string, options dockertypes.ImageRemoveOptions) ([]dockertypes.ImageDeleteResponseItem, error) {
119 return proxy.Docker.ImageRemove(ctx, image, options)
122 // ContainerRunner is the main stateful struct used for a single execution of a
124 type ContainerRunner struct {
125 Docker ThinDockerClient
126 ArvClient IArvadosClient
129 ContainerConfig dockercontainer.Config
130 dockercontainer.HostConfig
135 loggingDone chan bool
136 CrunchLog *ThrottledLogger
137 Stdout io.WriteCloser
138 Stderr io.WriteCloser
139 LogCollection *CollectionWriter
146 CleanupTempDir []string
149 SigChan chan os.Signal
150 ArvMountExit chan error
153 statLogger io.WriteCloser
154 statReporter *crunchstat.Reporter
155 statInterval time.Duration
157 // What we expect the container's cgroup parent to be.
158 expectCgroupParent string
159 // What we tell docker to use as the container's cgroup
160 // parent. Note: Ideally we would use the same field for both
161 // expectCgroupParent and setCgroupParent, and just make it
162 // default to "docker". However, when using docker < 1.10 with
163 // systemd, specifying a non-empty cgroup parent (even the
164 // default value "docker") hits a docker bug
165 // (https://github.com/docker/docker/issues/17126). Using two
166 // separate fields makes it possible to use the "expect cgroup
167 // parent to be X" feature even on sites where the "specify
168 // cgroup parent" feature breaks.
169 setCgroupParent string
171 cStateLock sync.Mutex
172 cStarted bool // StartContainer() succeeded
173 cCancelled bool // StopContainer() invoked
175 enableNetwork string // one of "default" or "always"
176 networkMode string // passed through to HostConfig.NetworkMode
179 // SetupSignals sets up signal handling to gracefully terminate the underlying
180 // Docker container and update state when receiving a TERM, INT or QUIT signal.
181 func (runner *ContainerRunner) SetupSignals() {
182 runner.SigChan = make(chan os.Signal, 1)
183 signal.Notify(runner.SigChan, syscall.SIGTERM)
184 signal.Notify(runner.SigChan, syscall.SIGINT)
185 signal.Notify(runner.SigChan, syscall.SIGQUIT)
187 go func(sig chan os.Signal) {
194 // stop the underlying Docker container.
195 func (runner *ContainerRunner) stop() {
196 runner.cStateLock.Lock()
197 defer runner.cStateLock.Unlock()
198 if runner.cCancelled {
201 runner.cCancelled = true
203 timeout := time.Duration(10)
204 err := runner.Docker.ContainerStop(context.TODO(), runner.ContainerID, &(timeout))
206 log.Printf("StopContainer failed: %s", err)
211 // LoadImage determines the docker image id from the container record and
212 // checks if it is available in the local Docker image store. If not, it loads
213 // the image from Keep.
214 func (runner *ContainerRunner) LoadImage() (err error) {
216 runner.CrunchLog.Printf("Fetching Docker image from collection '%s'", runner.Container.ContainerImage)
218 var collection arvados.Collection
219 err = runner.ArvClient.Get("collections", runner.Container.ContainerImage, nil, &collection)
221 return fmt.Errorf("While getting container image collection: %v", err)
223 manifest := manifest.Manifest{Text: collection.ManifestText}
224 var img, imageID string
225 for ms := range manifest.StreamIter() {
226 img = ms.FileStreamSegments[0].Name
227 if !strings.HasSuffix(img, ".tar") {
228 return fmt.Errorf("First file in the container image collection does not end in .tar")
230 imageID = img[:len(img)-4]
233 runner.CrunchLog.Printf("Using Docker image id '%s'", imageID)
235 _, _, err = runner.Docker.ImageInspectWithRaw(context.TODO(), imageID)
237 runner.CrunchLog.Print("Loading Docker image from keep")
239 var readCloser io.ReadCloser
240 readCloser, err = runner.Kc.ManifestFileReader(manifest, img)
242 return fmt.Errorf("While creating ManifestFileReader for container image: %v", err)
245 response, err := runner.Docker.ImageLoad(context.TODO(), readCloser, false)
247 return fmt.Errorf("While loading container image into Docker: %v", err)
249 response.Body.Close()
251 runner.CrunchLog.Print("Docker image is available")
254 runner.ContainerConfig.Image = imageID
259 func (runner *ContainerRunner) ArvMountCmd(arvMountCmd []string, token string) (c *exec.Cmd, err error) {
260 c = exec.Command("arv-mount", arvMountCmd...)
262 // Copy our environment, but override ARVADOS_API_TOKEN with
263 // the container auth token.
265 for _, s := range os.Environ() {
266 if !strings.HasPrefix(s, "ARVADOS_API_TOKEN=") {
267 c.Env = append(c.Env, s)
270 c.Env = append(c.Env, "ARVADOS_API_TOKEN="+token)
272 nt := NewThrottledLogger(runner.NewLogWriter("arv-mount"))
281 statReadme := make(chan bool)
282 runner.ArvMountExit = make(chan error)
287 time.Sleep(100 * time.Millisecond)
288 _, err = os.Stat(fmt.Sprintf("%s/by_id/README", runner.ArvMountPoint))
298 runner.ArvMountExit <- c.Wait()
299 close(runner.ArvMountExit)
305 case err := <-runner.ArvMountExit:
306 runner.ArvMount = nil
314 func (runner *ContainerRunner) SetupArvMountPoint(prefix string) (err error) {
315 if runner.ArvMountPoint == "" {
316 runner.ArvMountPoint, err = runner.MkTempDir("", prefix)
321 func (runner *ContainerRunner) SetupMounts() (err error) {
322 err = runner.SetupArvMountPoint("keep")
324 return fmt.Errorf("While creating keep mount temp dir: %v", err)
327 runner.CleanupTempDir = append(runner.CleanupTempDir, runner.ArvMountPoint)
331 arvMountCmd := []string{"--foreground", "--allow-other", "--read-write"}
333 if runner.Container.RuntimeConstraints.KeepCacheRAM > 0 {
334 arvMountCmd = append(arvMountCmd, "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheRAM))
337 collectionPaths := []string{}
339 needCertMount := true
342 for bind, _ := range runner.Container.Mounts {
343 binds = append(binds, bind)
347 for _, bind := range binds {
348 mnt := runner.Container.Mounts[bind]
349 if bind == "stdout" || bind == "stderr" {
350 // Is it a "file" mount kind?
351 if mnt.Kind != "file" {
352 return fmt.Errorf("Unsupported mount kind '%s' for %s. Only 'file' is supported.", mnt.Kind, bind)
355 // Does path start with OutputPath?
356 prefix := runner.Container.OutputPath
357 if !strings.HasSuffix(prefix, "/") {
360 if !strings.HasPrefix(mnt.Path, prefix) {
361 return fmt.Errorf("%s path does not start with OutputPath: %s, %s", strings.Title(bind), mnt.Path, prefix)
366 // Is it a "collection" mount kind?
367 if mnt.Kind != "collection" && mnt.Kind != "json" {
368 return fmt.Errorf("Unsupported mount kind '%s' for stdin. Only 'collection' or 'json' are supported.", mnt.Kind)
372 if bind == "/etc/arvados/ca-certificates.crt" {
373 needCertMount = false
376 if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && bind != runner.Container.OutputPath+"/" {
377 if mnt.Kind != "collection" {
378 return fmt.Errorf("Only mount points of kind 'collection' are supported underneath the output_path: %v", bind)
383 case mnt.Kind == "collection" && bind != "stdin":
385 if mnt.UUID != "" && mnt.PortableDataHash != "" {
386 return fmt.Errorf("Cannot specify both 'uuid' and 'portable_data_hash' for a collection mount")
390 return fmt.Errorf("Writing to existing collections currently not permitted.")
393 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.UUID)
394 } else if mnt.PortableDataHash != "" {
396 return fmt.Errorf("Can never write to a collection specified by portable data hash")
398 idx := strings.Index(mnt.PortableDataHash, "/")
400 mnt.Path = path.Clean(mnt.PortableDataHash[idx:])
401 mnt.PortableDataHash = mnt.PortableDataHash[0:idx]
402 runner.Container.Mounts[bind] = mnt
404 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.PortableDataHash)
405 if mnt.Path != "" && mnt.Path != "." {
406 if strings.HasPrefix(mnt.Path, "./") {
407 mnt.Path = mnt.Path[2:]
408 } else if strings.HasPrefix(mnt.Path, "/") {
409 mnt.Path = mnt.Path[1:]
411 src += "/" + mnt.Path
414 src = fmt.Sprintf("%s/tmp%d", runner.ArvMountPoint, tmpcount)
415 arvMountCmd = append(arvMountCmd, "--mount-tmp")
416 arvMountCmd = append(arvMountCmd, fmt.Sprintf("tmp%d", tmpcount))
420 if bind == runner.Container.OutputPath {
421 runner.HostOutputDir = src
422 } else if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
423 return fmt.Errorf("Writable mount points are not permitted underneath the output_path: %v", bind)
425 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", src, bind))
427 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", src, bind))
429 collectionPaths = append(collectionPaths, src)
431 case mnt.Kind == "tmp" && bind == runner.Container.OutputPath:
432 runner.HostOutputDir, err = runner.MkTempDir("", "")
434 return fmt.Errorf("While creating mount temp dir: %v", err)
436 st, staterr := os.Stat(runner.HostOutputDir)
438 return fmt.Errorf("While Stat on temp dir: %v", staterr)
440 err = os.Chmod(runner.HostOutputDir, st.Mode()|os.ModeSetgid|0777)
442 return fmt.Errorf("While Chmod temp dir: %v", err)
444 runner.CleanupTempDir = append(runner.CleanupTempDir, runner.HostOutputDir)
445 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", runner.HostOutputDir, bind))
447 case mnt.Kind == "tmp":
448 runner.Binds = append(runner.Binds, bind)
450 case mnt.Kind == "json":
451 jsondata, err := json.Marshal(mnt.Content)
453 return fmt.Errorf("encoding json data: %v", err)
455 // Create a tempdir with a single file
456 // (instead of just a tempfile): this way we
457 // can ensure the file is world-readable
458 // inside the container, without having to
459 // make it world-readable on the docker host.
460 tmpdir, err := runner.MkTempDir("", "")
462 return fmt.Errorf("creating temp dir: %v", err)
464 runner.CleanupTempDir = append(runner.CleanupTempDir, tmpdir)
465 tmpfn := filepath.Join(tmpdir, "mountdata.json")
466 err = ioutil.WriteFile(tmpfn, jsondata, 0644)
468 return fmt.Errorf("writing temp file: %v", err)
470 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", tmpfn, bind))
474 if runner.HostOutputDir == "" {
475 return fmt.Errorf("Output path does not correspond to a writable mount point")
478 if wantAPI := runner.Container.RuntimeConstraints.API; needCertMount && wantAPI != nil && *wantAPI {
479 for _, certfile := range arvadosclient.CertFiles {
480 _, err := os.Stat(certfile)
482 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:/etc/arvados/ca-certificates.crt:ro", certfile))
489 arvMountCmd = append(arvMountCmd, "--mount-by-pdh", "by_id")
491 arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_id")
493 arvMountCmd = append(arvMountCmd, runner.ArvMountPoint)
495 token, err := runner.ContainerToken()
497 return fmt.Errorf("could not get container token: %s", err)
500 runner.ArvMount, err = runner.RunArvMount(arvMountCmd, token)
502 return fmt.Errorf("While trying to start arv-mount: %v", err)
505 for _, p := range collectionPaths {
508 return fmt.Errorf("While checking that input files exist: %v", err)
515 func (runner *ContainerRunner) ProcessDockerAttach(containerReader io.Reader) {
516 // Handle docker log protocol
517 // https://docs.docker.com/engine/reference/api/docker_remote_api_v1.15/#attach-to-a-container
519 header := make([]byte, 8)
521 _, readerr := io.ReadAtLeast(containerReader, header, 8)
524 readsize := int64(header[7]) | (int64(header[6]) << 8) | (int64(header[5]) << 16) | (int64(header[4]) << 24)
527 _, readerr = io.CopyN(runner.Stdout, containerReader, readsize)
530 _, readerr = io.CopyN(runner.Stderr, containerReader, readsize)
535 if readerr != io.EOF {
536 runner.CrunchLog.Printf("While reading docker logs: %v", readerr)
539 closeerr := runner.Stdout.Close()
541 runner.CrunchLog.Printf("While closing stdout logs: %v", closeerr)
544 closeerr = runner.Stderr.Close()
546 runner.CrunchLog.Printf("While closing stderr logs: %v", closeerr)
549 if runner.statReporter != nil {
550 runner.statReporter.Stop()
551 closeerr = runner.statLogger.Close()
553 runner.CrunchLog.Printf("While closing crunchstat logs: %v", closeerr)
557 runner.loggingDone <- true
558 close(runner.loggingDone)
564 func (runner *ContainerRunner) StartCrunchstat() {
565 runner.statLogger = NewThrottledLogger(runner.NewLogWriter("crunchstat"))
566 runner.statReporter = &crunchstat.Reporter{
567 CID: runner.ContainerID,
568 Logger: log.New(runner.statLogger, "", 0),
569 CgroupParent: runner.expectCgroupParent,
570 CgroupRoot: runner.cgroupRoot,
571 PollPeriod: runner.statInterval,
573 runner.statReporter.Start()
576 type infoCommand struct {
581 // Gather node information and store it on the log for debugging
583 func (runner *ContainerRunner) LogNodeInfo() (err error) {
584 w := runner.NewLogWriter("node-info")
585 logger := log.New(w, "node-info", 0)
587 commands := []infoCommand{
589 label: "Host Information",
590 cmd: []string{"uname", "-a"},
593 label: "CPU Information",
594 cmd: []string{"cat", "/proc/cpuinfo"},
597 label: "Memory Information",
598 cmd: []string{"cat", "/proc/meminfo"},
602 cmd: []string{"df", "-m", "/", os.TempDir()},
605 label: "Disk INodes",
606 cmd: []string{"df", "-i", "/", os.TempDir()},
610 // Run commands with informational output to be logged.
612 for _, command := range commands {
613 out, err = exec.Command(command.cmd[0], command.cmd[1:]...).CombinedOutput()
615 return fmt.Errorf("While running command %q: %v",
618 logger.Println(command.label)
619 for _, line := range strings.Split(string(out), "\n") {
620 logger.Println(" ", line)
626 return fmt.Errorf("While closing node-info logs: %v", err)
631 // Get and save the raw JSON container record from the API server
632 func (runner *ContainerRunner) LogContainerRecord() (err error) {
633 w := NewArvLogWriter(runner.ArvClient, runner.Container.UUID, "container",
634 runner.LogCollection.Open("container.json"))
635 // Get Container record JSON from the API Server
636 reader, err := runner.ArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
638 return fmt.Errorf("While retrieving container record from the API server: %v", err)
641 // Read the API server response as []byte
642 json_bytes, err := ioutil.ReadAll(reader)
644 return fmt.Errorf("While reading container record API server response: %v", err)
646 // Decode the JSON []byte
647 var cr map[string]interface{}
648 if err = json.Unmarshal(json_bytes, &cr); err != nil {
649 return fmt.Errorf("While decoding the container record JSON response: %v", err)
651 // Re-encode it using indentation to improve readability
652 enc := json.NewEncoder(w)
653 enc.SetIndent("", " ")
654 if err = enc.Encode(cr); err != nil {
655 return fmt.Errorf("While logging the JSON container record: %v", err)
659 return fmt.Errorf("While closing container.json log: %v", err)
664 // AttachStreams connects the docker container stdin, stdout and stderr logs
665 // to the Arvados logger which logs to Keep and the API server logs table.
666 func (runner *ContainerRunner) AttachStreams() (err error) {
668 runner.CrunchLog.Print("Attaching container streams")
670 // If stdin mount is provided, attach it to the docker container
671 var stdinRdr keepclient.Reader
673 if stdinMnt, ok := runner.Container.Mounts["stdin"]; ok {
674 if stdinMnt.Kind == "collection" {
675 var stdinColl arvados.Collection
676 collId := stdinMnt.UUID
678 collId = stdinMnt.PortableDataHash
680 err = runner.ArvClient.Get("collections", collId, nil, &stdinColl)
682 return fmt.Errorf("While getting stding collection: %v", err)
685 stdinRdr, err = runner.Kc.ManifestFileReader(manifest.Manifest{Text: stdinColl.ManifestText}, stdinMnt.Path)
686 if os.IsNotExist(err) {
687 return fmt.Errorf("stdin collection path not found: %v", stdinMnt.Path)
688 } else if err != nil {
689 return fmt.Errorf("While getting stdin collection path %v: %v", stdinMnt.Path, err)
691 } else if stdinMnt.Kind == "json" {
692 stdinJson, err = json.Marshal(stdinMnt.Content)
694 return fmt.Errorf("While encoding stdin json data: %v", err)
699 stdinUsed := stdinRdr != nil || len(stdinJson) != 0
700 response, err := runner.Docker.ContainerAttach(context.TODO(), runner.ContainerID,
701 dockertypes.ContainerAttachOptions{Stream: true, Stdin: stdinUsed, Stdout: true, Stderr: true})
703 return fmt.Errorf("While attaching container stdout/stderr streams: %v", err)
706 runner.loggingDone = make(chan bool)
708 if stdoutMnt, ok := runner.Container.Mounts["stdout"]; ok {
709 stdoutFile, err := runner.getStdoutFile(stdoutMnt.Path)
713 runner.Stdout = stdoutFile
715 runner.Stdout = NewThrottledLogger(runner.NewLogWriter("stdout"))
718 if stderrMnt, ok := runner.Container.Mounts["stderr"]; ok {
719 stderrFile, err := runner.getStdoutFile(stderrMnt.Path)
723 runner.Stderr = stderrFile
725 runner.Stderr = NewThrottledLogger(runner.NewLogWriter("stderr"))
730 _, err := io.Copy(response.Conn, stdinRdr)
732 runner.CrunchLog.Print("While writing stdin collection to docker container %q", err)
736 response.CloseWrite()
738 } else if len(stdinJson) != 0 {
740 _, err := io.Copy(response.Conn, bytes.NewReader(stdinJson))
742 runner.CrunchLog.Print("While writing stdin json to docker container %q", err)
745 response.CloseWrite()
749 go runner.ProcessDockerAttach(response.Reader)
754 func (runner *ContainerRunner) getStdoutFile(mntPath string) (*os.File, error) {
755 stdoutPath := mntPath[len(runner.Container.OutputPath):]
756 index := strings.LastIndex(stdoutPath, "/")
758 subdirs := stdoutPath[:index]
760 st, err := os.Stat(runner.HostOutputDir)
762 return nil, fmt.Errorf("While Stat on temp dir: %v", err)
764 stdoutPath := path.Join(runner.HostOutputDir, subdirs)
765 err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777)
767 return nil, fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err)
771 stdoutFile, err := os.Create(path.Join(runner.HostOutputDir, stdoutPath))
773 return nil, fmt.Errorf("While creating file %q: %v", stdoutPath, err)
776 return stdoutFile, nil
779 // CreateContainer creates the docker container.
780 func (runner *ContainerRunner) CreateContainer() error {
781 runner.CrunchLog.Print("Creating Docker container")
783 runner.ContainerConfig.Cmd = runner.Container.Command
784 if runner.Container.Cwd != "." {
785 runner.ContainerConfig.WorkingDir = runner.Container.Cwd
788 for k, v := range runner.Container.Environment {
789 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env, k+"="+v)
792 runner.HostConfig = dockercontainer.HostConfig{
794 Cgroup: dockercontainer.CgroupSpec(runner.setCgroupParent),
795 LogConfig: dockercontainer.LogConfig{
800 if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
801 tok, err := runner.ContainerToken()
805 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env,
806 "ARVADOS_API_TOKEN="+tok,
807 "ARVADOS_API_HOST="+os.Getenv("ARVADOS_API_HOST"),
808 "ARVADOS_API_HOST_INSECURE="+os.Getenv("ARVADOS_API_HOST_INSECURE"),
810 runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
812 if runner.enableNetwork == "always" {
813 runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
815 runner.HostConfig.NetworkMode = dockercontainer.NetworkMode("none")
819 _, stdinUsed := runner.Container.Mounts["stdin"]
820 runner.ContainerConfig.OpenStdin = stdinUsed
821 runner.ContainerConfig.StdinOnce = stdinUsed
822 runner.ContainerConfig.AttachStdin = stdinUsed
823 runner.ContainerConfig.AttachStdout = true
824 runner.ContainerConfig.AttachStderr = true
826 createdBody, err := runner.Docker.ContainerCreate(context.TODO(), &runner.ContainerConfig, &runner.HostConfig, nil, runner.Container.UUID)
828 return fmt.Errorf("While creating container: %v", err)
831 runner.ContainerID = createdBody.ID
833 return runner.AttachStreams()
836 // StartContainer starts the docker container created by CreateContainer.
837 func (runner *ContainerRunner) StartContainer() error {
838 runner.CrunchLog.Printf("Starting Docker container id '%s'", runner.ContainerID)
839 runner.cStateLock.Lock()
840 defer runner.cStateLock.Unlock()
841 if runner.cCancelled {
844 err := runner.Docker.ContainerStart(context.TODO(), runner.ContainerID,
845 dockertypes.ContainerStartOptions{})
847 return fmt.Errorf("could not start container: %v", err)
849 runner.cStarted = true
853 // WaitFinish waits for the container to terminate, capture the exit code, and
854 // close the stdout/stderr logging.
855 func (runner *ContainerRunner) WaitFinish() error {
856 runner.CrunchLog.Print("Waiting for container to finish")
858 waitDocker, err := runner.Docker.ContainerWait(context.TODO(), runner.ContainerID)
860 return fmt.Errorf("container wait: %v", err)
863 runner.CrunchLog.Printf("Container exited with code: %v", waitDocker)
864 code := int(waitDocker)
865 runner.ExitCode = &code
867 waitMount := runner.ArvMountExit
869 case err := <-waitMount:
870 runner.CrunchLog.Printf("arv-mount exited before container finished: %v", err)
876 // wait for stdout/stderr to complete
882 // HandleOutput sets the output, unmounts the FUSE mount, and deletes temporary directories
883 func (runner *ContainerRunner) CaptureOutput() error {
884 if runner.finalState != "Complete" {
888 if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
889 // Output may have been set directly by the container, so
890 // refresh the container record to check.
891 err := runner.ArvClient.Get("containers", runner.Container.UUID,
892 nil, &runner.Container)
896 if runner.Container.Output != "" {
897 // Container output is already set.
898 runner.OutputPDH = &runner.Container.Output
903 if runner.HostOutputDir == "" {
907 _, err := os.Stat(runner.HostOutputDir)
909 return fmt.Errorf("While checking host output path: %v", err)
912 var manifestText string
914 collectionMetafile := fmt.Sprintf("%s/.arvados#collection", runner.HostOutputDir)
915 _, err = os.Stat(collectionMetafile)
918 cw := CollectionWriter{0, runner.Kc, nil, nil, sync.Mutex{}}
919 manifestText, err = cw.WriteTree(runner.HostOutputDir, runner.CrunchLog.Logger)
921 return fmt.Errorf("While uploading output files: %v", err)
924 // FUSE mount directory
925 file, openerr := os.Open(collectionMetafile)
927 return fmt.Errorf("While opening FUSE metafile: %v", err)
931 var rec arvados.Collection
932 err = json.NewDecoder(file).Decode(&rec)
934 return fmt.Errorf("While reading FUSE metafile: %v", err)
936 manifestText = rec.ManifestText
939 // Pre-populate output from the configured mount points
941 for bind, _ := range runner.Container.Mounts {
942 binds = append(binds, bind)
946 for _, bind := range binds {
947 mnt := runner.Container.Mounts[bind]
949 bindSuffix := strings.TrimPrefix(bind, runner.Container.OutputPath)
951 if bindSuffix == bind || len(bindSuffix) <= 0 {
952 // either does not start with OutputPath or is OutputPath itself
956 if mnt.ExcludeFromOutput == true {
960 // append to manifest_text
961 m, err := runner.getCollectionManifestForPath(mnt, bindSuffix)
966 manifestText = manifestText + m
970 var response arvados.Collection
971 manifest := manifest.Manifest{Text: manifestText}
972 manifestText = manifest.Extract(".", ".").Text
973 err = runner.ArvClient.Create("collections",
975 "ensure_unique_name": true,
976 "collection": arvadosclient.Dict{
978 "name": "output for " + runner.Container.UUID,
979 "manifest_text": manifestText}},
982 return fmt.Errorf("While creating output collection: %v", err)
984 runner.OutputPDH = &response.PortableDataHash
988 var outputCollections = make(map[string]arvados.Collection)
990 // Fetch the collection for the mnt.PortableDataHash
991 // Return the manifest_text fragment corresponding to the specified mnt.Path
992 // after making any required updates.
994 // If mnt.Path is not specified,
995 // return the entire manifest_text after replacing any "." with bindSuffix
996 // If mnt.Path corresponds to one stream,
997 // return the manifest_text for that stream after replacing that stream name with bindSuffix
998 // Otherwise, check if a filename in any one stream is being sought. Return the manifest_text
999 // for that stream after replacing stream name with bindSuffix minus the last word
1000 // and the file name with last word of the bindSuffix
1001 // Allowed path examples:
1003 // "path":"/subdir1"
1004 // "path":"/subdir1/subdir2"
1005 // "path":"/subdir/filename" etc
1006 func (runner *ContainerRunner) getCollectionManifestForPath(mnt arvados.Mount, bindSuffix string) (string, error) {
1007 collection := outputCollections[mnt.PortableDataHash]
1008 if collection.PortableDataHash == "" {
1009 err := runner.ArvClient.Get("collections", mnt.PortableDataHash, nil, &collection)
1011 return "", fmt.Errorf("While getting collection for %v: %v", mnt.PortableDataHash, err)
1013 outputCollections[mnt.PortableDataHash] = collection
1016 if collection.ManifestText == "" {
1017 runner.CrunchLog.Printf("No manifest text for collection %v", collection.PortableDataHash)
1021 mft := manifest.Manifest{Text: collection.ManifestText}
1022 extracted := mft.Extract(mnt.Path, bindSuffix)
1023 if extracted.Err != nil {
1024 return "", fmt.Errorf("Error parsing manifest for %v: %v", mnt.PortableDataHash, extracted.Err.Error())
1026 return extracted.Text, nil
1029 func (runner *ContainerRunner) CleanupDirs() {
1030 if runner.ArvMount != nil {
1031 umount := exec.Command("fusermount", "-z", "-u", runner.ArvMountPoint)
1032 umnterr := umount.Run()
1034 runner.CrunchLog.Printf("While running fusermount: %v", umnterr)
1037 mnterr := <-runner.ArvMountExit
1039 runner.CrunchLog.Printf("Arv-mount exit error: %v", mnterr)
1043 for _, tmpdir := range runner.CleanupTempDir {
1044 rmerr := os.RemoveAll(tmpdir)
1046 runner.CrunchLog.Printf("While cleaning up temporary directory %s: %v", tmpdir, rmerr)
1051 // CommitLogs posts the collection containing the final container logs.
1052 func (runner *ContainerRunner) CommitLogs() error {
1053 runner.CrunchLog.Print(runner.finalState)
1054 runner.CrunchLog.Close()
1056 // Closing CrunchLog above allows it to be committed to Keep at this
1057 // point, but re-open crunch log with ArvClient in case there are any
1058 // other further (such as failing to write the log to Keep!) while
1060 runner.CrunchLog = NewThrottledLogger(NewArvLogWriter(runner.ArvClient, runner.Container.UUID, "crunch-run", nil))
1062 if runner.LogsPDH != nil {
1063 // If we have already assigned something to LogsPDH,
1064 // we must be closing the re-opened log, which won't
1065 // end up getting attached to the container record and
1066 // therefore doesn't need to be saved as a collection
1067 // -- it exists only to send logs to other channels.
1071 mt, err := runner.LogCollection.ManifestText()
1073 return fmt.Errorf("While creating log manifest: %v", err)
1076 var response arvados.Collection
1077 err = runner.ArvClient.Create("collections",
1079 "ensure_unique_name": true,
1080 "collection": arvadosclient.Dict{
1082 "name": "logs for " + runner.Container.UUID,
1083 "manifest_text": mt}},
1086 return fmt.Errorf("While creating log collection: %v", err)
1088 runner.LogsPDH = &response.PortableDataHash
1092 // UpdateContainerRunning updates the container state to "Running"
1093 func (runner *ContainerRunner) UpdateContainerRunning() error {
1094 runner.cStateLock.Lock()
1095 defer runner.cStateLock.Unlock()
1096 if runner.cCancelled {
1099 return runner.ArvClient.Update("containers", runner.Container.UUID,
1100 arvadosclient.Dict{"container": arvadosclient.Dict{"state": "Running"}}, nil)
1103 // ContainerToken returns the api_token the container (and any
1104 // arv-mount processes) are allowed to use.
1105 func (runner *ContainerRunner) ContainerToken() (string, error) {
1106 if runner.token != "" {
1107 return runner.token, nil
1110 var auth arvados.APIClientAuthorization
1111 err := runner.ArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth)
1115 runner.token = auth.APIToken
1116 return runner.token, nil
1119 // UpdateContainerComplete updates the container record state on API
1120 // server to "Complete" or "Cancelled"
1121 func (runner *ContainerRunner) UpdateContainerFinal() error {
1122 update := arvadosclient.Dict{}
1123 update["state"] = runner.finalState
1124 if runner.LogsPDH != nil {
1125 update["log"] = *runner.LogsPDH
1127 if runner.finalState == "Complete" {
1128 if runner.ExitCode != nil {
1129 update["exit_code"] = *runner.ExitCode
1131 if runner.OutputPDH != nil {
1132 update["output"] = *runner.OutputPDH
1135 return runner.ArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{"container": update}, nil)
1138 // IsCancelled returns the value of Cancelled, with goroutine safety.
1139 func (runner *ContainerRunner) IsCancelled() bool {
1140 runner.cStateLock.Lock()
1141 defer runner.cStateLock.Unlock()
1142 return runner.cCancelled
1145 // NewArvLogWriter creates an ArvLogWriter
1146 func (runner *ContainerRunner) NewArvLogWriter(name string) io.WriteCloser {
1147 return NewArvLogWriter(runner.ArvClient, runner.Container.UUID, name, runner.LogCollection.Open(name+".txt"))
1150 // Run the full container lifecycle.
1151 func (runner *ContainerRunner) Run() (err error) {
1152 runner.CrunchLog.Printf("Executing container '%s'", runner.Container.UUID)
1154 hostname, hosterr := os.Hostname()
1156 runner.CrunchLog.Printf("Error getting hostname '%v'", hosterr)
1158 runner.CrunchLog.Printf("Executing on host '%s'", hostname)
1161 // Clean up temporary directories _after_ finalizing
1162 // everything (if we've made any by then)
1163 defer runner.CleanupDirs()
1165 runner.finalState = "Queued"
1168 // checkErr prints e (unless it's nil) and sets err to
1169 // e (unless err is already non-nil). Thus, if err
1170 // hasn't already been assigned when Run() returns,
1171 // this cleanup func will cause Run() to return the
1172 // first non-nil error that is passed to checkErr().
1173 checkErr := func(e error) {
1177 runner.CrunchLog.Print(e)
1183 // Log the error encountered in Run(), if any
1186 if runner.finalState == "Queued" {
1187 runner.CrunchLog.Close()
1188 runner.UpdateContainerFinal()
1192 if runner.IsCancelled() {
1193 runner.finalState = "Cancelled"
1194 // but don't return yet -- we still want to
1195 // capture partial output and write logs
1198 checkErr(runner.CaptureOutput())
1199 checkErr(runner.CommitLogs())
1200 checkErr(runner.UpdateContainerFinal())
1202 // The real log is already closed, but then we opened
1203 // a new one in case we needed to log anything while
1205 runner.CrunchLog.Close()
1208 err = runner.ArvClient.Get("containers", runner.Container.UUID, nil, &runner.Container)
1210 err = fmt.Errorf("While getting container record: %v", err)
1214 // setup signal handling
1215 runner.SetupSignals()
1217 // check for and/or load image
1218 err = runner.LoadImage()
1220 runner.finalState = "Cancelled"
1221 err = fmt.Errorf("While loading container image: %v", err)
1225 // set up FUSE mount and binds
1226 err = runner.SetupMounts()
1228 runner.finalState = "Cancelled"
1229 err = fmt.Errorf("While setting up mounts: %v", err)
1233 err = runner.CreateContainer()
1238 // Gather and record node information
1239 err = runner.LogNodeInfo()
1243 // Save container.json record on log collection
1244 err = runner.LogContainerRecord()
1249 runner.StartCrunchstat()
1251 if runner.IsCancelled() {
1255 err = runner.UpdateContainerRunning()
1259 runner.finalState = "Cancelled"
1261 err = runner.StartContainer()
1266 err = runner.WaitFinish()
1268 runner.finalState = "Complete"
1273 // NewContainerRunner creates a new container runner.
1274 func NewContainerRunner(api IArvadosClient,
1276 docker ThinDockerClient,
1277 containerUUID string) *ContainerRunner {
1279 cr := &ContainerRunner{ArvClient: api, Kc: kc, Docker: docker}
1280 cr.NewLogWriter = cr.NewArvLogWriter
1281 cr.RunArvMount = cr.ArvMountCmd
1282 cr.MkTempDir = ioutil.TempDir
1283 cr.LogCollection = &CollectionWriter{0, kc, nil, nil, sync.Mutex{}}
1284 cr.Container.UUID = containerUUID
1285 cr.CrunchLog = NewThrottledLogger(cr.NewLogWriter("crunch-run"))
1286 cr.CrunchLog.Immediate = log.New(os.Stderr, containerUUID+" ", 0)
1291 statInterval := flag.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting")
1292 cgroupRoot := flag.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree")
1293 cgroupParent := flag.String("cgroup-parent", "docker", "name of container's parent cgroup (ignored if -cgroup-parent-subsystem is used)")
1294 cgroupParentSubsystem := flag.String("cgroup-parent-subsystem", "", "use current cgroup for given subsystem as parent cgroup for container")
1295 caCertsPath := flag.String("ca-certs", "", "Path to TLS root certificates")
1296 enableNetwork := flag.String("container-enable-networking", "default",
1297 `Specify if networking should be enabled for container. One of 'default', 'always':
1298 default: only enable networking if container requests it.
1299 always: containers always have networking enabled
1301 networkMode := flag.String("container-network-mode", "default",
1302 `Set networking mode for container. Corresponds to Docker network mode (--net).
1306 containerId := flag.Arg(0)
1308 if *caCertsPath != "" {
1309 arvadosclient.CertFiles = []string{*caCertsPath}
1312 api, err := arvadosclient.MakeArvadosClient()
1314 log.Fatalf("%s: %v", containerId, err)
1318 var kc *keepclient.KeepClient
1319 kc, err = keepclient.MakeKeepClient(api)
1321 log.Fatalf("%s: %v", containerId, err)
1325 var docker *dockerclient.Client
1326 // API version 1.21 corresponds to Docker 1.9, which is currently the
1327 // minimum version we want to support.
1328 docker, err = dockerclient.NewClient(dockerclient.DefaultDockerHost, "1.21", nil, nil)
1330 log.Fatalf("%s: %v", containerId, err)
1333 dockerClientProxy := ThinDockerClientProxy{Docker: docker}
1335 cr := NewContainerRunner(api, kc, dockerClientProxy, containerId)
1336 cr.statInterval = *statInterval
1337 cr.cgroupRoot = *cgroupRoot
1338 cr.expectCgroupParent = *cgroupParent
1339 cr.enableNetwork = *enableNetwork
1340 cr.networkMode = *networkMode
1341 if *cgroupParentSubsystem != "" {
1342 p := findCgroup(*cgroupParentSubsystem)
1343 cr.setCgroupParent = p
1344 cr.expectCgroupParent = p
1349 log.Fatalf("%s: %v", containerId, err)