1 // Copyright (C) The Arvados Authors. All rights reserved.
3 // SPDX-License-Identifier: AGPL-3.0
30 "git.curoverse.com/arvados.git/lib/crunchstat"
31 "git.curoverse.com/arvados.git/sdk/go/arvados"
32 "git.curoverse.com/arvados.git/sdk/go/arvadosclient"
33 "git.curoverse.com/arvados.git/sdk/go/keepclient"
34 "git.curoverse.com/arvados.git/sdk/go/manifest"
36 dockertypes "github.com/docker/docker/api/types"
37 dockercontainer "github.com/docker/docker/api/types/container"
38 dockernetwork "github.com/docker/docker/api/types/network"
39 dockerclient "github.com/docker/docker/client"
42 // IArvadosClient is the minimal Arvados API methods used by crunch-run.
43 type IArvadosClient interface {
44 Create(resourceType string, parameters arvadosclient.Dict, output interface{}) error
45 Get(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
46 Update(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
47 Call(method, resourceType, uuid, action string, parameters arvadosclient.Dict, output interface{}) error
48 CallRaw(method string, resourceType string, uuid string, action string, parameters arvadosclient.Dict) (reader io.ReadCloser, err error)
49 Discovery(key string) (interface{}, error)
52 // ErrCancelled is the error returned when the container is cancelled.
53 var ErrCancelled = errors.New("Cancelled")
55 // IKeepClient is the minimal Keep API methods used by crunch-run.
56 type IKeepClient interface {
57 PutHB(hash string, buf []byte) (string, int, error)
58 ManifestFileReader(m manifest.Manifest, filename string) (arvados.File, error)
62 // NewLogWriter is a factory function to create a new log writer.
63 type NewLogWriter func(name string) io.WriteCloser
65 type RunArvMount func(args []string, tok string) (*exec.Cmd, error)
67 type MkTempDir func(string, string) (string, error)
69 // ThinDockerClient is the minimal Docker client interface used by crunch-run.
70 type ThinDockerClient interface {
71 ContainerAttach(ctx context.Context, container string, options dockertypes.ContainerAttachOptions) (dockertypes.HijackedResponse, error)
72 ContainerCreate(ctx context.Context, config *dockercontainer.Config, hostConfig *dockercontainer.HostConfig,
73 networkingConfig *dockernetwork.NetworkingConfig, containerName string) (dockercontainer.ContainerCreateCreatedBody, error)
74 ContainerStart(ctx context.Context, container string, options dockertypes.ContainerStartOptions) error
75 ContainerStop(ctx context.Context, container string, timeout *time.Duration) error
76 ContainerWait(ctx context.Context, container string, condition dockercontainer.WaitCondition) (<-chan dockercontainer.ContainerWaitOKBody, <-chan error)
77 ImageInspectWithRaw(ctx context.Context, image string) (dockertypes.ImageInspect, []byte, error)
78 ImageLoad(ctx context.Context, input io.Reader, quiet bool) (dockertypes.ImageLoadResponse, error)
79 ImageRemove(ctx context.Context, image string, options dockertypes.ImageRemoveOptions) ([]dockertypes.ImageDeleteResponseItem, error)
82 // ThinDockerClientProxy is a proxy implementation of ThinDockerClient
83 // that executes the docker requests on dockerclient.Client
84 type ThinDockerClientProxy struct {
85 Docker *dockerclient.Client
88 // ContainerAttach invokes dockerclient.Client.ContainerAttach
89 func (proxy ThinDockerClientProxy) ContainerAttach(ctx context.Context, container string, options dockertypes.ContainerAttachOptions) (dockertypes.HijackedResponse, error) {
90 return proxy.Docker.ContainerAttach(ctx, container, options)
93 // ContainerCreate invokes dockerclient.Client.ContainerCreate
94 func (proxy ThinDockerClientProxy) ContainerCreate(ctx context.Context, config *dockercontainer.Config, hostConfig *dockercontainer.HostConfig,
95 networkingConfig *dockernetwork.NetworkingConfig, containerName string) (dockercontainer.ContainerCreateCreatedBody, error) {
96 return proxy.Docker.ContainerCreate(ctx, config, hostConfig, networkingConfig, containerName)
99 // ContainerStart invokes dockerclient.Client.ContainerStart
100 func (proxy ThinDockerClientProxy) ContainerStart(ctx context.Context, container string, options dockertypes.ContainerStartOptions) error {
101 return proxy.Docker.ContainerStart(ctx, container, options)
104 // ContainerStop invokes dockerclient.Client.ContainerStop
105 func (proxy ThinDockerClientProxy) ContainerStop(ctx context.Context, container string, timeout *time.Duration) error {
106 return proxy.Docker.ContainerStop(ctx, container, timeout)
109 // ContainerWait invokes dockerclient.Client.ContainerWait
110 func (proxy ThinDockerClientProxy) ContainerWait(ctx context.Context, container string, condition dockercontainer.WaitCondition) (<-chan dockercontainer.ContainerWaitOKBody, <-chan error) {
111 return proxy.Docker.ContainerWait(ctx, container, condition)
114 // ImageInspectWithRaw invokes dockerclient.Client.ImageInspectWithRaw
115 func (proxy ThinDockerClientProxy) ImageInspectWithRaw(ctx context.Context, image string) (dockertypes.ImageInspect, []byte, error) {
116 return proxy.Docker.ImageInspectWithRaw(ctx, image)
119 // ImageLoad invokes dockerclient.Client.ImageLoad
120 func (proxy ThinDockerClientProxy) ImageLoad(ctx context.Context, input io.Reader, quiet bool) (dockertypes.ImageLoadResponse, error) {
121 return proxy.Docker.ImageLoad(ctx, input, quiet)
124 // ImageRemove invokes dockerclient.Client.ImageRemove
125 func (proxy ThinDockerClientProxy) ImageRemove(ctx context.Context, image string, options dockertypes.ImageRemoveOptions) ([]dockertypes.ImageDeleteResponseItem, error) {
126 return proxy.Docker.ImageRemove(ctx, image, options)
129 // ContainerRunner is the main stateful struct used for a single execution of a
131 type ContainerRunner struct {
132 Docker ThinDockerClient
133 ArvClient IArvadosClient
136 ContainerConfig dockercontainer.Config
137 dockercontainer.HostConfig
142 loggingDone chan bool
143 CrunchLog *ThrottledLogger
144 Stdout io.WriteCloser
145 Stderr io.WriteCloser
146 LogCollection *CollectionWriter
153 CleanupTempDir []string
155 Volumes map[string]struct{}
157 SigChan chan os.Signal
158 ArvMountExit chan error
161 statLogger io.WriteCloser
162 statReporter *crunchstat.Reporter
163 statInterval time.Duration
165 // What we expect the container's cgroup parent to be.
166 expectCgroupParent string
167 // What we tell docker to use as the container's cgroup
168 // parent. Note: Ideally we would use the same field for both
169 // expectCgroupParent and setCgroupParent, and just make it
170 // default to "docker". However, when using docker < 1.10 with
171 // systemd, specifying a non-empty cgroup parent (even the
172 // default value "docker") hits a docker bug
173 // (https://github.com/docker/docker/issues/17126). Using two
174 // separate fields makes it possible to use the "expect cgroup
175 // parent to be X" feature even on sites where the "specify
176 // cgroup parent" feature breaks.
177 setCgroupParent string
179 cStateLock sync.Mutex
180 cStarted bool // StartContainer() succeeded
181 cCancelled bool // StopContainer() invoked
183 enableNetwork string // one of "default" or "always"
184 networkMode string // passed through to HostConfig.NetworkMode
187 // setupSignals sets up signal handling to gracefully terminate the underlying
188 // Docker container and update state when receiving a TERM, INT or QUIT signal.
189 func (runner *ContainerRunner) setupSignals() {
190 runner.SigChan = make(chan os.Signal, 1)
191 signal.Notify(runner.SigChan, syscall.SIGTERM)
192 signal.Notify(runner.SigChan, syscall.SIGINT)
193 signal.Notify(runner.SigChan, syscall.SIGQUIT)
195 go func(sig chan os.Signal) {
201 // stop the underlying Docker container.
202 func (runner *ContainerRunner) stop() {
203 runner.cStateLock.Lock()
204 defer runner.cStateLock.Unlock()
205 if runner.cCancelled {
208 runner.cCancelled = true
210 timeout := time.Duration(10)
211 err := runner.Docker.ContainerStop(context.TODO(), runner.ContainerID, &(timeout))
213 log.Printf("StopContainer failed: %s", err)
218 func (runner *ContainerRunner) teardown() {
219 if runner.SigChan != nil {
220 signal.Stop(runner.SigChan)
221 close(runner.SigChan)
225 // LoadImage determines the docker image id from the container record and
226 // checks if it is available in the local Docker image store. If not, it loads
227 // the image from Keep.
228 func (runner *ContainerRunner) LoadImage() (err error) {
230 runner.CrunchLog.Printf("Fetching Docker image from collection '%s'", runner.Container.ContainerImage)
232 var collection arvados.Collection
233 err = runner.ArvClient.Get("collections", runner.Container.ContainerImage, nil, &collection)
235 return fmt.Errorf("While getting container image collection: %v", err)
237 manifest := manifest.Manifest{Text: collection.ManifestText}
238 var img, imageID string
239 for ms := range manifest.StreamIter() {
240 img = ms.FileStreamSegments[0].Name
241 if !strings.HasSuffix(img, ".tar") {
242 return fmt.Errorf("First file in the container image collection does not end in .tar")
244 imageID = img[:len(img)-4]
247 runner.CrunchLog.Printf("Using Docker image id '%s'", imageID)
249 _, _, err = runner.Docker.ImageInspectWithRaw(context.TODO(), imageID)
251 runner.CrunchLog.Print("Loading Docker image from keep")
253 var readCloser io.ReadCloser
254 readCloser, err = runner.Kc.ManifestFileReader(manifest, img)
256 return fmt.Errorf("While creating ManifestFileReader for container image: %v", err)
259 response, err := runner.Docker.ImageLoad(context.TODO(), readCloser, true)
261 return fmt.Errorf("While loading container image into Docker: %v", err)
264 defer response.Body.Close()
265 rbody, err := ioutil.ReadAll(response.Body)
267 return fmt.Errorf("Reading response to image load: %v", err)
269 runner.CrunchLog.Printf("Docker response: %s", rbody)
271 runner.CrunchLog.Print("Docker image is available")
274 runner.ContainerConfig.Image = imageID
276 runner.Kc.ClearBlockCache()
281 func (runner *ContainerRunner) ArvMountCmd(arvMountCmd []string, token string) (c *exec.Cmd, err error) {
282 c = exec.Command("arv-mount", arvMountCmd...)
284 // Copy our environment, but override ARVADOS_API_TOKEN with
285 // the container auth token.
287 for _, s := range os.Environ() {
288 if !strings.HasPrefix(s, "ARVADOS_API_TOKEN=") {
289 c.Env = append(c.Env, s)
292 c.Env = append(c.Env, "ARVADOS_API_TOKEN="+token)
294 nt := NewThrottledLogger(runner.NewLogWriter("arv-mount"))
303 statReadme := make(chan bool)
304 runner.ArvMountExit = make(chan error)
309 time.Sleep(100 * time.Millisecond)
310 _, err = os.Stat(fmt.Sprintf("%s/by_id/README", runner.ArvMountPoint))
320 runner.ArvMountExit <- c.Wait()
321 close(runner.ArvMountExit)
327 case err := <-runner.ArvMountExit:
328 runner.ArvMount = nil
336 func (runner *ContainerRunner) SetupArvMountPoint(prefix string) (err error) {
337 if runner.ArvMountPoint == "" {
338 runner.ArvMountPoint, err = runner.MkTempDir("", prefix)
343 func (runner *ContainerRunner) SetupMounts() (err error) {
344 err = runner.SetupArvMountPoint("keep")
346 return fmt.Errorf("While creating keep mount temp dir: %v", err)
349 runner.CleanupTempDir = append(runner.CleanupTempDir, runner.ArvMountPoint)
353 arvMountCmd := []string{"--foreground", "--allow-other", "--read-write"}
355 if runner.Container.RuntimeConstraints.KeepCacheRAM > 0 {
356 arvMountCmd = append(arvMountCmd, "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheRAM))
359 collectionPaths := []string{}
361 runner.Volumes = make(map[string]struct{})
362 needCertMount := true
365 for bind := range runner.Container.Mounts {
366 binds = append(binds, bind)
370 for _, bind := range binds {
371 mnt := runner.Container.Mounts[bind]
372 if bind == "stdout" || bind == "stderr" {
373 // Is it a "file" mount kind?
374 if mnt.Kind != "file" {
375 return fmt.Errorf("Unsupported mount kind '%s' for %s. Only 'file' is supported.", mnt.Kind, bind)
378 // Does path start with OutputPath?
379 prefix := runner.Container.OutputPath
380 if !strings.HasSuffix(prefix, "/") {
383 if !strings.HasPrefix(mnt.Path, prefix) {
384 return fmt.Errorf("%s path does not start with OutputPath: %s, %s", strings.Title(bind), mnt.Path, prefix)
389 // Is it a "collection" mount kind?
390 if mnt.Kind != "collection" && mnt.Kind != "json" {
391 return fmt.Errorf("Unsupported mount kind '%s' for stdin. Only 'collection' or 'json' are supported.", mnt.Kind)
395 if bind == "/etc/arvados/ca-certificates.crt" {
396 needCertMount = false
399 if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && bind != runner.Container.OutputPath+"/" {
400 if mnt.Kind != "collection" {
401 return fmt.Errorf("Only mount points of kind 'collection' are supported underneath the output_path: %v", bind)
406 case mnt.Kind == "collection" && bind != "stdin":
408 if mnt.UUID != "" && mnt.PortableDataHash != "" {
409 return fmt.Errorf("Cannot specify both 'uuid' and 'portable_data_hash' for a collection mount")
413 return fmt.Errorf("Writing to existing collections currently not permitted.")
416 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.UUID)
417 } else if mnt.PortableDataHash != "" {
419 return fmt.Errorf("Can never write to a collection specified by portable data hash")
421 idx := strings.Index(mnt.PortableDataHash, "/")
423 mnt.Path = path.Clean(mnt.PortableDataHash[idx:])
424 mnt.PortableDataHash = mnt.PortableDataHash[0:idx]
425 runner.Container.Mounts[bind] = mnt
427 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.PortableDataHash)
428 if mnt.Path != "" && mnt.Path != "." {
429 if strings.HasPrefix(mnt.Path, "./") {
430 mnt.Path = mnt.Path[2:]
431 } else if strings.HasPrefix(mnt.Path, "/") {
432 mnt.Path = mnt.Path[1:]
434 src += "/" + mnt.Path
437 src = fmt.Sprintf("%s/tmp%d", runner.ArvMountPoint, tmpcount)
438 arvMountCmd = append(arvMountCmd, "--mount-tmp")
439 arvMountCmd = append(arvMountCmd, fmt.Sprintf("tmp%d", tmpcount))
443 if bind == runner.Container.OutputPath {
444 runner.HostOutputDir = src
445 } else if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
446 return fmt.Errorf("Writable mount points are not permitted underneath the output_path: %v", bind)
448 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", src, bind))
450 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", src, bind))
452 collectionPaths = append(collectionPaths, src)
454 case mnt.Kind == "tmp":
456 tmpdir, err = runner.MkTempDir("", "")
458 return fmt.Errorf("While creating mount temp dir: %v", err)
460 st, staterr := os.Stat(tmpdir)
462 return fmt.Errorf("While Stat on temp dir: %v", staterr)
464 err = os.Chmod(tmpdir, st.Mode()|os.ModeSetgid|0777)
466 return fmt.Errorf("While Chmod temp dir: %v", err)
468 runner.CleanupTempDir = append(runner.CleanupTempDir, tmpdir)
469 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", tmpdir, bind))
470 if bind == runner.Container.OutputPath {
471 runner.HostOutputDir = tmpdir
474 case mnt.Kind == "json":
475 jsondata, err := json.Marshal(mnt.Content)
477 return fmt.Errorf("encoding json data: %v", err)
479 // Create a tempdir with a single file
480 // (instead of just a tempfile): this way we
481 // can ensure the file is world-readable
482 // inside the container, without having to
483 // make it world-readable on the docker host.
484 tmpdir, err := runner.MkTempDir("", "")
486 return fmt.Errorf("creating temp dir: %v", err)
488 runner.CleanupTempDir = append(runner.CleanupTempDir, tmpdir)
489 tmpfn := filepath.Join(tmpdir, "mountdata.json")
490 err = ioutil.WriteFile(tmpfn, jsondata, 0644)
492 return fmt.Errorf("writing temp file: %v", err)
494 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", tmpfn, bind))
498 if runner.HostOutputDir == "" {
499 return fmt.Errorf("Output path does not correspond to a writable mount point")
502 if wantAPI := runner.Container.RuntimeConstraints.API; needCertMount && wantAPI != nil && *wantAPI {
503 for _, certfile := range arvadosclient.CertFiles {
504 _, err := os.Stat(certfile)
506 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:/etc/arvados/ca-certificates.crt:ro", certfile))
513 arvMountCmd = append(arvMountCmd, "--mount-by-pdh", "by_id")
515 arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_id")
517 arvMountCmd = append(arvMountCmd, runner.ArvMountPoint)
519 token, err := runner.ContainerToken()
521 return fmt.Errorf("could not get container token: %s", err)
524 runner.ArvMount, err = runner.RunArvMount(arvMountCmd, token)
526 return fmt.Errorf("While trying to start arv-mount: %v", err)
529 for _, p := range collectionPaths {
532 return fmt.Errorf("While checking that input files exist: %v", err)
539 func (runner *ContainerRunner) ProcessDockerAttach(containerReader io.Reader) {
540 // Handle docker log protocol
541 // https://docs.docker.com/engine/reference/api/docker_remote_api_v1.15/#attach-to-a-container
543 header := make([]byte, 8)
545 _, readerr := io.ReadAtLeast(containerReader, header, 8)
548 readsize := int64(header[7]) | (int64(header[6]) << 8) | (int64(header[5]) << 16) | (int64(header[4]) << 24)
551 _, readerr = io.CopyN(runner.Stdout, containerReader, readsize)
554 _, readerr = io.CopyN(runner.Stderr, containerReader, readsize)
559 if readerr != io.EOF {
560 runner.CrunchLog.Printf("While reading docker logs: %v", readerr)
563 closeerr := runner.Stdout.Close()
565 runner.CrunchLog.Printf("While closing stdout logs: %v", closeerr)
568 closeerr = runner.Stderr.Close()
570 runner.CrunchLog.Printf("While closing stderr logs: %v", closeerr)
573 if runner.statReporter != nil {
574 runner.statReporter.Stop()
575 closeerr = runner.statLogger.Close()
577 runner.CrunchLog.Printf("While closing crunchstat logs: %v", closeerr)
581 runner.loggingDone <- true
582 close(runner.loggingDone)
588 func (runner *ContainerRunner) StartCrunchstat() {
589 runner.statLogger = NewThrottledLogger(runner.NewLogWriter("crunchstat"))
590 runner.statReporter = &crunchstat.Reporter{
591 CID: runner.ContainerID,
592 Logger: log.New(runner.statLogger, "", 0),
593 CgroupParent: runner.expectCgroupParent,
594 CgroupRoot: runner.cgroupRoot,
595 PollPeriod: runner.statInterval,
597 runner.statReporter.Start()
600 type infoCommand struct {
605 // Gather node information and store it on the log for debugging
607 func (runner *ContainerRunner) LogNodeInfo() (err error) {
608 w := runner.NewLogWriter("node-info")
609 logger := log.New(w, "node-info", 0)
611 commands := []infoCommand{
613 label: "Host Information",
614 cmd: []string{"uname", "-a"},
617 label: "CPU Information",
618 cmd: []string{"cat", "/proc/cpuinfo"},
621 label: "Memory Information",
622 cmd: []string{"cat", "/proc/meminfo"},
626 cmd: []string{"df", "-m", "/", os.TempDir()},
629 label: "Disk INodes",
630 cmd: []string{"df", "-i", "/", os.TempDir()},
634 // Run commands with informational output to be logged.
636 for _, command := range commands {
637 out, err = exec.Command(command.cmd[0], command.cmd[1:]...).CombinedOutput()
639 return fmt.Errorf("While running command %q: %v",
642 logger.Println(command.label)
643 for _, line := range strings.Split(string(out), "\n") {
644 logger.Println(" ", line)
650 return fmt.Errorf("While closing node-info logs: %v", err)
655 // Get and save the raw JSON container record from the API server
656 func (runner *ContainerRunner) LogContainerRecord() (err error) {
658 ArvClient: runner.ArvClient,
659 UUID: runner.Container.UUID,
660 loggingStream: "container",
661 writeCloser: runner.LogCollection.Open("container.json"),
664 // Get Container record JSON from the API Server
665 reader, err := runner.ArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
667 return fmt.Errorf("While retrieving container record from the API server: %v", err)
671 dec := json.NewDecoder(reader)
673 var cr map[string]interface{}
674 if err = dec.Decode(&cr); err != nil {
675 return fmt.Errorf("While decoding the container record JSON response: %v", err)
677 // Re-encode it using indentation to improve readability
678 enc := json.NewEncoder(w)
679 enc.SetIndent("", " ")
680 if err = enc.Encode(cr); err != nil {
681 return fmt.Errorf("While logging the JSON container record: %v", err)
685 return fmt.Errorf("While closing container.json log: %v", err)
690 // AttachStreams connects the docker container stdin, stdout and stderr logs
691 // to the Arvados logger which logs to Keep and the API server logs table.
692 func (runner *ContainerRunner) AttachStreams() (err error) {
694 runner.CrunchLog.Print("Attaching container streams")
696 // If stdin mount is provided, attach it to the docker container
697 var stdinRdr arvados.File
699 if stdinMnt, ok := runner.Container.Mounts["stdin"]; ok {
700 if stdinMnt.Kind == "collection" {
701 var stdinColl arvados.Collection
702 collId := stdinMnt.UUID
704 collId = stdinMnt.PortableDataHash
706 err = runner.ArvClient.Get("collections", collId, nil, &stdinColl)
708 return fmt.Errorf("While getting stding collection: %v", err)
711 stdinRdr, err = runner.Kc.ManifestFileReader(manifest.Manifest{Text: stdinColl.ManifestText}, stdinMnt.Path)
712 if os.IsNotExist(err) {
713 return fmt.Errorf("stdin collection path not found: %v", stdinMnt.Path)
714 } else if err != nil {
715 return fmt.Errorf("While getting stdin collection path %v: %v", stdinMnt.Path, err)
717 } else if stdinMnt.Kind == "json" {
718 stdinJson, err = json.Marshal(stdinMnt.Content)
720 return fmt.Errorf("While encoding stdin json data: %v", err)
725 stdinUsed := stdinRdr != nil || len(stdinJson) != 0
726 response, err := runner.Docker.ContainerAttach(context.TODO(), runner.ContainerID,
727 dockertypes.ContainerAttachOptions{Stream: true, Stdin: stdinUsed, Stdout: true, Stderr: true})
729 return fmt.Errorf("While attaching container stdout/stderr streams: %v", err)
732 runner.loggingDone = make(chan bool)
734 if stdoutMnt, ok := runner.Container.Mounts["stdout"]; ok {
735 stdoutFile, err := runner.getStdoutFile(stdoutMnt.Path)
739 runner.Stdout = stdoutFile
741 runner.Stdout = NewThrottledLogger(runner.NewLogWriter("stdout"))
744 if stderrMnt, ok := runner.Container.Mounts["stderr"]; ok {
745 stderrFile, err := runner.getStdoutFile(stderrMnt.Path)
749 runner.Stderr = stderrFile
751 runner.Stderr = NewThrottledLogger(runner.NewLogWriter("stderr"))
756 _, err := io.Copy(response.Conn, stdinRdr)
758 runner.CrunchLog.Print("While writing stdin collection to docker container %q", err)
762 response.CloseWrite()
764 } else if len(stdinJson) != 0 {
766 _, err := io.Copy(response.Conn, bytes.NewReader(stdinJson))
768 runner.CrunchLog.Print("While writing stdin json to docker container %q", err)
771 response.CloseWrite()
775 go runner.ProcessDockerAttach(response.Reader)
780 func (runner *ContainerRunner) getStdoutFile(mntPath string) (*os.File, error) {
781 stdoutPath := mntPath[len(runner.Container.OutputPath):]
782 index := strings.LastIndex(stdoutPath, "/")
784 subdirs := stdoutPath[:index]
786 st, err := os.Stat(runner.HostOutputDir)
788 return nil, fmt.Errorf("While Stat on temp dir: %v", err)
790 stdoutPath := filepath.Join(runner.HostOutputDir, subdirs)
791 err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777)
793 return nil, fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err)
797 stdoutFile, err := os.Create(filepath.Join(runner.HostOutputDir, stdoutPath))
799 return nil, fmt.Errorf("While creating file %q: %v", stdoutPath, err)
802 return stdoutFile, nil
805 // CreateContainer creates the docker container.
806 func (runner *ContainerRunner) CreateContainer() error {
807 runner.CrunchLog.Print("Creating Docker container")
809 runner.ContainerConfig.Cmd = runner.Container.Command
810 if runner.Container.Cwd != "." {
811 runner.ContainerConfig.WorkingDir = runner.Container.Cwd
814 for k, v := range runner.Container.Environment {
815 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env, k+"="+v)
818 runner.ContainerConfig.Volumes = runner.Volumes
820 runner.HostConfig = dockercontainer.HostConfig{
822 LogConfig: dockercontainer.LogConfig{
825 Resources: dockercontainer.Resources{
826 CgroupParent: runner.setCgroupParent,
830 if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
831 tok, err := runner.ContainerToken()
835 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env,
836 "ARVADOS_API_TOKEN="+tok,
837 "ARVADOS_API_HOST="+os.Getenv("ARVADOS_API_HOST"),
838 "ARVADOS_API_HOST_INSECURE="+os.Getenv("ARVADOS_API_HOST_INSECURE"),
840 runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
842 if runner.enableNetwork == "always" {
843 runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
845 runner.HostConfig.NetworkMode = dockercontainer.NetworkMode("none")
849 _, stdinUsed := runner.Container.Mounts["stdin"]
850 runner.ContainerConfig.OpenStdin = stdinUsed
851 runner.ContainerConfig.StdinOnce = stdinUsed
852 runner.ContainerConfig.AttachStdin = stdinUsed
853 runner.ContainerConfig.AttachStdout = true
854 runner.ContainerConfig.AttachStderr = true
856 createdBody, err := runner.Docker.ContainerCreate(context.TODO(), &runner.ContainerConfig, &runner.HostConfig, nil, runner.Container.UUID)
858 return fmt.Errorf("While creating container: %v", err)
861 runner.ContainerID = createdBody.ID
863 return runner.AttachStreams()
866 // StartContainer starts the docker container created by CreateContainer.
867 func (runner *ContainerRunner) StartContainer() error {
868 runner.CrunchLog.Printf("Starting Docker container id '%s'", runner.ContainerID)
869 runner.cStateLock.Lock()
870 defer runner.cStateLock.Unlock()
871 if runner.cCancelled {
874 err := runner.Docker.ContainerStart(context.TODO(), runner.ContainerID,
875 dockertypes.ContainerStartOptions{})
878 if strings.Contains(err.Error(), "no such file or directory") {
879 advice = fmt.Sprintf("\nPossible causes: command %q is missing, the interpreter given in #! is missing, or script has Windows line endings.", runner.Container.Command[0])
881 return fmt.Errorf("could not start container: %v%s", err, advice)
883 runner.cStarted = true
887 // WaitFinish waits for the container to terminate, capture the exit code, and
888 // close the stdout/stderr logging.
889 func (runner *ContainerRunner) WaitFinish() (err error) {
890 runner.CrunchLog.Print("Waiting for container to finish")
892 waitOk, waitErr := runner.Docker.ContainerWait(context.TODO(), runner.ContainerID, "not-running")
894 var waitBody dockercontainer.ContainerWaitOKBody
896 case waitBody = <-waitOk:
897 case err = <-waitErr:
901 return fmt.Errorf("container wait: %v", err)
904 runner.CrunchLog.Printf("Container exited with code: %v", waitBody.StatusCode)
905 code := int(waitBody.StatusCode)
906 runner.ExitCode = &code
908 waitMount := runner.ArvMountExit
910 case err = <-waitMount:
911 runner.CrunchLog.Printf("arv-mount exited before container finished: %v", err)
917 // wait for stdout/stderr to complete
923 // HandleOutput sets the output, unmounts the FUSE mount, and deletes temporary directories
924 func (runner *ContainerRunner) CaptureOutput() error {
925 if runner.finalState != "Complete" {
929 if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
930 // Output may have been set directly by the container, so
931 // refresh the container record to check.
932 err := runner.ArvClient.Get("containers", runner.Container.UUID,
933 nil, &runner.Container)
937 if runner.Container.Output != "" {
938 // Container output is already set.
939 runner.OutputPDH = &runner.Container.Output
944 if runner.HostOutputDir == "" {
948 _, err := os.Stat(runner.HostOutputDir)
950 return fmt.Errorf("While checking host output path: %v", err)
953 // Pre-populate output from the configured mount points
955 for bind, mnt := range runner.Container.Mounts {
956 if mnt.Kind == "collection" {
957 binds = append(binds, bind)
962 var manifestText string
964 collectionMetafile := fmt.Sprintf("%s/.arvados#collection", runner.HostOutputDir)
965 _, err = os.Stat(collectionMetafile)
969 // Find symlinks to arv-mounted files & dirs.
970 err = filepath.Walk(runner.HostOutputDir, func(path string, info os.FileInfo, err error) error {
974 if info.Mode()&os.ModeSymlink == 0 {
977 // read link to get container internal path
978 // only support 1 level of symlinking here.
980 tgt, err = os.Readlink(path)
985 // get path relative to output dir
986 outputSuffix := path[len(runner.HostOutputDir):]
988 if strings.HasPrefix(tgt, "/") {
989 // go through mounts and try reverse map to collection reference
990 for _, bind := range binds {
991 mnt := runner.Container.Mounts[bind]
992 if tgt == bind || strings.HasPrefix(tgt, bind+"/") {
993 // get path relative to bind
994 targetSuffix := tgt[len(bind):]
996 // Copy mount and adjust the path to add path relative to the bind
998 adjustedMount.Path = filepath.Join(adjustedMount.Path, targetSuffix)
1000 // get manifest text
1002 m, err = runner.getCollectionManifestForPath(adjustedMount, outputSuffix)
1006 manifestText = manifestText + m
1007 // delete symlink so WriteTree won't try to to dereference it.
1014 // Not a link to a mount. Must be dereferencible and
1015 // point into the output directory.
1016 tgt, err = filepath.EvalSymlinks(path)
1022 // Symlink target must be within the output directory otherwise it's an error.
1023 if !strings.HasPrefix(tgt, runner.HostOutputDir+"/") {
1025 return fmt.Errorf("Output directory symlink %q points to invalid location %q, must point to mount or output directory.",
1031 return fmt.Errorf("While checking output symlinks: %v", err)
1034 cw := CollectionWriter{0, runner.Kc, nil, nil, sync.Mutex{}}
1036 m, err = cw.WriteTree(runner.HostOutputDir, runner.CrunchLog.Logger)
1037 manifestText = manifestText + m
1039 return fmt.Errorf("While uploading output files: %v", err)
1042 // FUSE mount directory
1043 file, openerr := os.Open(collectionMetafile)
1045 return fmt.Errorf("While opening FUSE metafile: %v", err)
1049 var rec arvados.Collection
1050 err = json.NewDecoder(file).Decode(&rec)
1052 return fmt.Errorf("While reading FUSE metafile: %v", err)
1054 manifestText = rec.ManifestText
1057 for _, bind := range binds {
1058 mnt := runner.Container.Mounts[bind]
1060 bindSuffix := strings.TrimPrefix(bind, runner.Container.OutputPath)
1062 if bindSuffix == bind || len(bindSuffix) <= 0 {
1063 // either does not start with OutputPath or is OutputPath itself
1067 if mnt.ExcludeFromOutput == true {
1071 // append to manifest_text
1072 m, err := runner.getCollectionManifestForPath(mnt, bindSuffix)
1077 manifestText = manifestText + m
1081 var response arvados.Collection
1082 manifest := manifest.Manifest{Text: manifestText}
1083 manifestText = manifest.Extract(".", ".").Text
1084 err = runner.ArvClient.Create("collections",
1086 "ensure_unique_name": true,
1087 "collection": arvadosclient.Dict{
1089 "name": "output for " + runner.Container.UUID,
1090 "manifest_text": manifestText}},
1093 return fmt.Errorf("While creating output collection: %v", err)
1095 runner.OutputPDH = &response.PortableDataHash
1099 var outputCollections = make(map[string]arvados.Collection)
1101 // Fetch the collection for the mnt.PortableDataHash
1102 // Return the manifest_text fragment corresponding to the specified mnt.Path
1103 // after making any required updates.
1105 // If mnt.Path is not specified,
1106 // return the entire manifest_text after replacing any "." with bindSuffix
1107 // If mnt.Path corresponds to one stream,
1108 // return the manifest_text for that stream after replacing that stream name with bindSuffix
1109 // Otherwise, check if a filename in any one stream is being sought. Return the manifest_text
1110 // for that stream after replacing stream name with bindSuffix minus the last word
1111 // and the file name with last word of the bindSuffix
1112 // Allowed path examples:
1114 // "path":"/subdir1"
1115 // "path":"/subdir1/subdir2"
1116 // "path":"/subdir/filename" etc
1117 func (runner *ContainerRunner) getCollectionManifestForPath(mnt arvados.Mount, bindSuffix string) (string, error) {
1118 collection := outputCollections[mnt.PortableDataHash]
1119 if collection.PortableDataHash == "" {
1120 err := runner.ArvClient.Get("collections", mnt.PortableDataHash, nil, &collection)
1122 return "", fmt.Errorf("While getting collection for %v: %v", mnt.PortableDataHash, err)
1124 outputCollections[mnt.PortableDataHash] = collection
1127 if collection.ManifestText == "" {
1128 runner.CrunchLog.Printf("No manifest text for collection %v", collection.PortableDataHash)
1132 mft := manifest.Manifest{Text: collection.ManifestText}
1133 extracted := mft.Extract(mnt.Path, bindSuffix)
1134 if extracted.Err != nil {
1135 return "", fmt.Errorf("Error parsing manifest for %v: %v", mnt.PortableDataHash, extracted.Err.Error())
1137 return extracted.Text, nil
1140 func (runner *ContainerRunner) CleanupDirs() {
1141 if runner.ArvMount != nil {
1142 umount := exec.Command("fusermount", "-z", "-u", runner.ArvMountPoint)
1143 umnterr := umount.Run()
1145 runner.CrunchLog.Printf("While running fusermount: %v", umnterr)
1148 mnterr := <-runner.ArvMountExit
1150 runner.CrunchLog.Printf("Arv-mount exit error: %v", mnterr)
1154 for _, tmpdir := range runner.CleanupTempDir {
1155 rmerr := os.RemoveAll(tmpdir)
1157 runner.CrunchLog.Printf("While cleaning up temporary directory %s: %v", tmpdir, rmerr)
1162 // CommitLogs posts the collection containing the final container logs.
1163 func (runner *ContainerRunner) CommitLogs() error {
1164 runner.CrunchLog.Print(runner.finalState)
1165 runner.CrunchLog.Close()
1167 // Closing CrunchLog above allows it to be committed to Keep at this
1168 // point, but re-open crunch log with ArvClient in case there are any
1169 // other further (such as failing to write the log to Keep!) while
1171 runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{ArvClient: runner.ArvClient,
1172 UUID: runner.Container.UUID, loggingStream: "crunch-run", writeCloser: nil})
1174 if runner.LogsPDH != nil {
1175 // If we have already assigned something to LogsPDH,
1176 // we must be closing the re-opened log, which won't
1177 // end up getting attached to the container record and
1178 // therefore doesn't need to be saved as a collection
1179 // -- it exists only to send logs to other channels.
1183 mt, err := runner.LogCollection.ManifestText()
1185 return fmt.Errorf("While creating log manifest: %v", err)
1188 var response arvados.Collection
1189 err = runner.ArvClient.Create("collections",
1191 "ensure_unique_name": true,
1192 "collection": arvadosclient.Dict{
1194 "name": "logs for " + runner.Container.UUID,
1195 "manifest_text": mt}},
1198 return fmt.Errorf("While creating log collection: %v", err)
1200 runner.LogsPDH = &response.PortableDataHash
1204 // UpdateContainerRunning updates the container state to "Running"
1205 func (runner *ContainerRunner) UpdateContainerRunning() error {
1206 runner.cStateLock.Lock()
1207 defer runner.cStateLock.Unlock()
1208 if runner.cCancelled {
1211 return runner.ArvClient.Update("containers", runner.Container.UUID,
1212 arvadosclient.Dict{"container": arvadosclient.Dict{"state": "Running"}}, nil)
1215 // ContainerToken returns the api_token the container (and any
1216 // arv-mount processes) are allowed to use.
1217 func (runner *ContainerRunner) ContainerToken() (string, error) {
1218 if runner.token != "" {
1219 return runner.token, nil
1222 var auth arvados.APIClientAuthorization
1223 err := runner.ArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth)
1227 runner.token = auth.APIToken
1228 return runner.token, nil
1231 // UpdateContainerComplete updates the container record state on API
1232 // server to "Complete" or "Cancelled"
1233 func (runner *ContainerRunner) UpdateContainerFinal() error {
1234 update := arvadosclient.Dict{}
1235 update["state"] = runner.finalState
1236 if runner.LogsPDH != nil {
1237 update["log"] = *runner.LogsPDH
1239 if runner.finalState == "Complete" {
1240 if runner.ExitCode != nil {
1241 update["exit_code"] = *runner.ExitCode
1243 if runner.OutputPDH != nil {
1244 update["output"] = *runner.OutputPDH
1247 return runner.ArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{"container": update}, nil)
1250 // IsCancelled returns the value of Cancelled, with goroutine safety.
1251 func (runner *ContainerRunner) IsCancelled() bool {
1252 runner.cStateLock.Lock()
1253 defer runner.cStateLock.Unlock()
1254 return runner.cCancelled
1257 // NewArvLogWriter creates an ArvLogWriter
1258 func (runner *ContainerRunner) NewArvLogWriter(name string) io.WriteCloser {
1259 return &ArvLogWriter{ArvClient: runner.ArvClient, UUID: runner.Container.UUID, loggingStream: name,
1260 writeCloser: runner.LogCollection.Open(name + ".txt")}
1263 // Run the full container lifecycle.
1264 func (runner *ContainerRunner) Run() (err error) {
1265 runner.CrunchLog.Printf("Executing container '%s'", runner.Container.UUID)
1267 hostname, hosterr := os.Hostname()
1269 runner.CrunchLog.Printf("Error getting hostname '%v'", hosterr)
1271 runner.CrunchLog.Printf("Executing on host '%s'", hostname)
1274 // Clean up temporary directories _after_ finalizing
1275 // everything (if we've made any by then)
1276 defer runner.CleanupDirs()
1278 runner.finalState = "Queued"
1281 // checkErr prints e (unless it's nil) and sets err to
1282 // e (unless err is already non-nil). Thus, if err
1283 // hasn't already been assigned when Run() returns,
1284 // this cleanup func will cause Run() to return the
1285 // first non-nil error that is passed to checkErr().
1286 checkErr := func(e error) {
1290 runner.CrunchLog.Print(e)
1294 if runner.finalState == "Complete" {
1295 // There was an error in the finalization.
1296 runner.finalState = "Cancelled"
1300 // Log the error encountered in Run(), if any
1303 if runner.finalState == "Queued" {
1304 runner.CrunchLog.Close()
1305 runner.UpdateContainerFinal()
1309 if runner.IsCancelled() {
1310 runner.finalState = "Cancelled"
1311 // but don't return yet -- we still want to
1312 // capture partial output and write logs
1315 checkErr(runner.CaptureOutput())
1316 checkErr(runner.CommitLogs())
1317 checkErr(runner.UpdateContainerFinal())
1319 // The real log is already closed, but then we opened
1320 // a new one in case we needed to log anything while
1322 runner.CrunchLog.Close()
1327 err = runner.fetchContainerRecord()
1332 // setup signal handling
1333 runner.setupSignals()
1335 // check for and/or load image
1336 err = runner.LoadImage()
1338 runner.finalState = "Cancelled"
1339 err = fmt.Errorf("While loading container image: %v", err)
1343 // set up FUSE mount and binds
1344 err = runner.SetupMounts()
1346 runner.finalState = "Cancelled"
1347 err = fmt.Errorf("While setting up mounts: %v", err)
1351 err = runner.CreateContainer()
1356 // Gather and record node information
1357 err = runner.LogNodeInfo()
1361 // Save container.json record on log collection
1362 err = runner.LogContainerRecord()
1367 runner.StartCrunchstat()
1369 if runner.IsCancelled() {
1373 err = runner.UpdateContainerRunning()
1377 runner.finalState = "Cancelled"
1379 err = runner.StartContainer()
1384 err = runner.WaitFinish()
1386 runner.finalState = "Complete"
1391 // Fetch the current container record (uuid = runner.Container.UUID)
1392 // into runner.Container.
1393 func (runner *ContainerRunner) fetchContainerRecord() error {
1394 reader, err := runner.ArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
1396 return fmt.Errorf("error fetching container record: %v", err)
1398 defer reader.Close()
1400 dec := json.NewDecoder(reader)
1402 err = dec.Decode(&runner.Container)
1404 return fmt.Errorf("error decoding container record: %v", err)
1409 // NewContainerRunner creates a new container runner.
1410 func NewContainerRunner(api IArvadosClient,
1412 docker ThinDockerClient,
1413 containerUUID string) *ContainerRunner {
1415 cr := &ContainerRunner{ArvClient: api, Kc: kc, Docker: docker}
1416 cr.NewLogWriter = cr.NewArvLogWriter
1417 cr.RunArvMount = cr.ArvMountCmd
1418 cr.MkTempDir = ioutil.TempDir
1419 cr.LogCollection = &CollectionWriter{0, kc, nil, nil, sync.Mutex{}}
1420 cr.Container.UUID = containerUUID
1421 cr.CrunchLog = NewThrottledLogger(cr.NewLogWriter("crunch-run"))
1422 cr.CrunchLog.Immediate = log.New(os.Stderr, containerUUID+" ", 0)
1424 loadLogThrottleParams(api)
1430 statInterval := flag.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting")
1431 cgroupRoot := flag.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree")
1432 cgroupParent := flag.String("cgroup-parent", "docker", "name of container's parent cgroup (ignored if -cgroup-parent-subsystem is used)")
1433 cgroupParentSubsystem := flag.String("cgroup-parent-subsystem", "", "use current cgroup for given subsystem as parent cgroup for container")
1434 caCertsPath := flag.String("ca-certs", "", "Path to TLS root certificates")
1435 enableNetwork := flag.String("container-enable-networking", "default",
1436 `Specify if networking should be enabled for container. One of 'default', 'always':
1437 default: only enable networking if container requests it.
1438 always: containers always have networking enabled
1440 networkMode := flag.String("container-network-mode", "default",
1441 `Set networking mode for container. Corresponds to Docker network mode (--net).
1443 memprofile := flag.String("memprofile", "", "write memory profile to `file` after running container")
1446 containerId := flag.Arg(0)
1448 if *caCertsPath != "" {
1449 arvadosclient.CertFiles = []string{*caCertsPath}
1452 api, err := arvadosclient.MakeArvadosClient()
1454 log.Fatalf("%s: %v", containerId, err)
1458 var kc *keepclient.KeepClient
1459 kc, err = keepclient.MakeKeepClient(api)
1461 log.Fatalf("%s: %v", containerId, err)
1463 kc.BlockCache = &keepclient.BlockCache{MaxBlocks: 2}
1466 var docker *dockerclient.Client
1467 // API version 1.21 corresponds to Docker 1.9, which is currently the
1468 // minimum version we want to support.
1469 docker, err = dockerclient.NewClient(dockerclient.DefaultDockerHost, "1.21", nil, nil)
1471 log.Fatalf("%s: %v", containerId, err)
1474 dockerClientProxy := ThinDockerClientProxy{Docker: docker}
1476 cr := NewContainerRunner(api, kc, dockerClientProxy, containerId)
1477 cr.statInterval = *statInterval
1478 cr.cgroupRoot = *cgroupRoot
1479 cr.expectCgroupParent = *cgroupParent
1480 cr.enableNetwork = *enableNetwork
1481 cr.networkMode = *networkMode
1482 if *cgroupParentSubsystem != "" {
1483 p := findCgroup(*cgroupParentSubsystem)
1484 cr.setCgroupParent = p
1485 cr.expectCgroupParent = p
1490 if *memprofile != "" {
1491 f, err := os.Create(*memprofile)
1493 log.Printf("could not create memory profile: ", err)
1495 runtime.GC() // get up-to-date statistics
1496 if err := pprof.WriteHeapProfile(f); err != nil {
1497 log.Printf("could not write memory profile: ", err)
1499 closeerr := f.Close()
1500 if closeerr != nil {
1501 log.Printf("closing memprofile file: ", err)
1506 log.Fatalf("%s: %v", containerId, runerr)