X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/fdea6d02dae5393455683e980c2adec19b4da75c..8ed521f7fd1e48e1e415125745ed8c6627a62c91:/services/crunch-run/crunchrun.go diff --git a/services/crunch-run/crunchrun.go b/services/crunch-run/crunchrun.go index 1b04bb4424..0b59f7df91 100644 --- a/services/crunch-run/crunchrun.go +++ b/services/crunch-run/crunchrun.go @@ -5,6 +5,8 @@ import ( "errors" "flag" "fmt" + "git.curoverse.com/arvados.git/lib/crunchstat" + "git.curoverse.com/arvados.git/sdk/go/arvados" "git.curoverse.com/arvados.git/sdk/go/arvadosclient" "git.curoverse.com/arvados.git/sdk/go/keepclient" "git.curoverse.com/arvados.git/sdk/go/manifest" @@ -15,6 +17,9 @@ import ( "os" "os/exec" "os/signal" + "path" + "path/filepath" + "sort" "strings" "sync" "syscall" @@ -25,7 +30,9 @@ import ( type IArvadosClient interface { Create(resourceType string, parameters arvadosclient.Dict, output interface{}) error Get(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error - Update(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) (err error) + Update(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error + Call(method, resourceType, uuid, action string, parameters arvadosclient.Dict, output interface{}) error + Discovery(key string) (interface{}, error) } // ErrCancelled is the error returned when the container is cancelled. @@ -34,43 +41,13 @@ var ErrCancelled = errors.New("Cancelled") // IKeepClient is the minimal Keep API methods used by crunch-run. type IKeepClient interface { PutHB(hash string, buf []byte) (string, int, error) - ManifestFileReader(m manifest.Manifest, filename string) (keepclient.ReadCloserWithLen, error) -} - -// Mount describes the mount points to create inside the container. -type Mount struct { - Kind string `json:"kind"` - Writable bool `json:"writable"` - PortableDataHash string `json:"portable_data_hash"` - UUID string `json:"uuid"` - DeviceType string `json:"device_type"` -} - -// Collection record returned by the API server. -type CollectionRecord struct { - ManifestText string `json:"manifest_text"` - PortableDataHash string `json:"portable_data_hash"` -} - -// ContainerRecord is the container record returned by the API server. -type ContainerRecord struct { - UUID string `json:"uuid"` - Command []string `json:"command"` - ContainerImage string `json:"container_image"` - Cwd string `json:"cwd"` - Environment map[string]string `json:"environment"` - Mounts map[string]Mount `json:"mounts"` - OutputPath string `json:"output_path"` - Priority int `json:"priority"` - RuntimeConstraints map[string]interface{} `json:"runtime_constraints"` - State string `json:"state"` - Output string `json:"output"` + ManifestFileReader(m manifest.Manifest, filename string) (keepclient.Reader, error) } // NewLogWriter is a factory function to create a new log writer. type NewLogWriter func(name string) io.WriteCloser -type RunArvMount func([]string) (*exec.Cmd, error) +type RunArvMount func(args []string, tok string) (*exec.Cmd, error) type MkTempDir func(string, string) (string, error) @@ -81,7 +58,7 @@ type ThinDockerClient interface { LoadImage(reader io.Reader) error CreateContainer(config *dockerclient.ContainerConfig, name string, authConfig *dockerclient.AuthConfig) (string, error) StartContainer(id string, config *dockerclient.HostConfig) error - ContainerLogs(id string, options *dockerclient.LogOptions) (io.ReadCloser, error) + AttachContainer(id string, options *dockerclient.AttachOptions) (io.ReadCloser, error) Wait(id string) <-chan dockerclient.WaitResult RemoveImage(name string, force bool) ([]*dockerclient.ImageDelete, error) } @@ -92,14 +69,16 @@ type ContainerRunner struct { Docker ThinDockerClient ArvClient IArvadosClient Kc IKeepClient - ContainerRecord + arvados.Container dockerclient.ContainerConfig + dockerclient.HostConfig + token string ContainerID string ExitCode *int NewLogWriter loggingDone chan bool CrunchLog *ThrottledLogger - Stdout *ThrottledLogger + Stdout io.WriteCloser Stderr *ThrottledLogger LogCollection *CollectionWriter LogsPDH *string @@ -116,18 +95,36 @@ type ContainerRunner struct { SigChan chan os.Signal ArvMountExit chan error finalState string + + statLogger io.WriteCloser + statReporter *crunchstat.Reporter + statInterval time.Duration + cgroupRoot string + // What we expect the container's cgroup parent to be. + expectCgroupParent string + // What we tell docker to use as the container's cgroup + // parent. Note: Ideally we would use the same field for both + // expectCgroupParent and setCgroupParent, and just make it + // default to "docker". However, when using docker < 1.10 with + // systemd, specifying a non-empty cgroup parent (even the + // default value "docker") hits a docker bug + // (https://github.com/docker/docker/issues/17126). Using two + // separate fields makes it possible to use the "expect cgroup + // parent to be X" feature even on sites where the "specify + // cgroup parent" feature breaks. + setCgroupParent string } // SetupSignals sets up signal handling to gracefully terminate the underlying // Docker container and update state when receiving a TERM, INT or QUIT signal. -func (runner *ContainerRunner) SetupSignals() error { +func (runner *ContainerRunner) SetupSignals() { runner.SigChan = make(chan os.Signal, 1) signal.Notify(runner.SigChan, syscall.SIGTERM) signal.Notify(runner.SigChan, syscall.SIGINT) signal.Notify(runner.SigChan, syscall.SIGQUIT) go func(sig <-chan os.Signal) { - for _ = range sig { + for range sig { if !runner.Cancelled { runner.CancelLock.Lock() runner.Cancelled = true @@ -138,8 +135,6 @@ func (runner *ContainerRunner) SetupSignals() error { } } }(runner.SigChan) - - return nil } // LoadImage determines the docker image id from the container record and @@ -147,10 +142,10 @@ func (runner *ContainerRunner) SetupSignals() error { // the image from Keep. func (runner *ContainerRunner) LoadImage() (err error) { - runner.CrunchLog.Printf("Fetching Docker image from collection '%s'", runner.ContainerRecord.ContainerImage) + runner.CrunchLog.Printf("Fetching Docker image from collection '%s'", runner.Container.ContainerImage) - var collection CollectionRecord - err = runner.ArvClient.Get("collections", runner.ContainerRecord.ContainerImage, nil, &collection) + var collection arvados.Collection + err = runner.ArvClient.Get("collections", runner.Container.ContainerImage, nil, &collection) if err != nil { return fmt.Errorf("While getting container image collection: %v", err) } @@ -189,8 +184,19 @@ func (runner *ContainerRunner) LoadImage() (err error) { return nil } -func (runner *ContainerRunner) ArvMountCmd(arvMountCmd []string) (c *exec.Cmd, err error) { +func (runner *ContainerRunner) ArvMountCmd(arvMountCmd []string, token string) (c *exec.Cmd, err error) { c = exec.Command("arv-mount", arvMountCmd...) + + // Copy our environment, but override ARVADOS_API_TOKEN with + // the container auth token. + c.Env = nil + for _, s := range os.Environ() { + if !strings.HasPrefix(s, "ARVADOS_API_TOKEN=") { + c.Env = append(c.Env, s) + } + } + c.Env = append(c.Env, "ARVADOS_API_TOKEN="+token) + nt := NewThrottledLogger(runner.NewLogWriter("arv-mount")) c.Stdout = nt c.Stderr = nt @@ -233,8 +239,15 @@ func (runner *ContainerRunner) ArvMountCmd(arvMountCmd []string) (c *exec.Cmd, e return c, nil } +func (runner *ContainerRunner) SetupArvMountPoint(prefix string) (err error) { + if runner.ArvMountPoint == "" { + runner.ArvMountPoint, err = runner.MkTempDir("", prefix) + } + return +} + func (runner *ContainerRunner) SetupMounts() (err error) { - runner.ArvMountPoint, err = runner.MkTempDir("", "keep") + err = runner.SetupArvMountPoint("keep") if err != nil { return fmt.Errorf("While creating keep mount temp dir: %v", err) } @@ -243,12 +256,52 @@ func (runner *ContainerRunner) SetupMounts() (err error) { pdhOnly := true tmpcount := 0 - arvMountCmd := []string{"--foreground", "--allow-other"} + arvMountCmd := []string{"--foreground", "--allow-other", "--read-write"} + + if runner.Container.RuntimeConstraints.KeepCacheRAM > 0 { + arvMountCmd = append(arvMountCmd, "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheRAM)) + } + collectionPaths := []string{} runner.Binds = nil + needCertMount := true + + var binds []string + for bind, _ := range runner.Container.Mounts { + binds = append(binds, bind) + } + sort.Strings(binds) + + for _, bind := range binds { + mnt := runner.Container.Mounts[bind] + if bind == "stdout" { + // Is it a "file" mount kind? + if mnt.Kind != "file" { + return fmt.Errorf("Unsupported mount kind '%s' for stdout. Only 'file' is supported.", mnt.Kind) + } + + // Does path start with OutputPath? + prefix := runner.Container.OutputPath + if !strings.HasSuffix(prefix, "/") { + prefix += "/" + } + if !strings.HasPrefix(mnt.Path, prefix) { + return fmt.Errorf("Stdout path does not start with OutputPath: %s, %s", mnt.Path, prefix) + } + } - for bind, mnt := range runner.ContainerRecord.Mounts { - if mnt.Kind == "collection" { + if bind == "/etc/arvados/ca-certificates.crt" { + needCertMount = false + } + + if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && bind != runner.Container.OutputPath+"/" { + if mnt.Kind != "collection" { + return fmt.Errorf("Only mount points of kind 'collection' are supported underneath the output_path: %v", bind) + } + } + + switch { + case mnt.Kind == "collection": var src string if mnt.UUID != "" && mnt.PortableDataHash != "" { return fmt.Errorf("Cannot specify both 'uuid' and 'portable_data_hash' for a collection mount") @@ -263,7 +316,21 @@ func (runner *ContainerRunner) SetupMounts() (err error) { if mnt.Writable { return fmt.Errorf("Can never write to a collection specified by portable data hash") } + idx := strings.Index(mnt.PortableDataHash, "/") + if idx > 0 { + mnt.Path = path.Clean(mnt.PortableDataHash[idx:]) + mnt.PortableDataHash = mnt.PortableDataHash[0:idx] + runner.Container.Mounts[bind] = mnt + } src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.PortableDataHash) + if mnt.Path != "" && mnt.Path != "." { + if strings.HasPrefix(mnt.Path, "./") { + mnt.Path = mnt.Path[2:] + } else if strings.HasPrefix(mnt.Path, "/") { + mnt.Path = mnt.Path[1:] + } + src += "/" + mnt.Path + } } else { src = fmt.Sprintf("%s/tmp%d", runner.ArvMountPoint, tmpcount) arvMountCmd = append(arvMountCmd, "--mount-tmp") @@ -271,35 +338,57 @@ func (runner *ContainerRunner) SetupMounts() (err error) { tmpcount += 1 } if mnt.Writable { - if bind == runner.ContainerRecord.OutputPath { + if bind == runner.Container.OutputPath { runner.HostOutputDir = src + } else if strings.HasPrefix(bind, runner.Container.OutputPath+"/") { + return fmt.Errorf("Writable mount points are not permitted underneath the output_path: %v", bind) } runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", src, bind)) } else { runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", src, bind)) } collectionPaths = append(collectionPaths, src) - } else if mnt.Kind == "tmp" { - if bind == runner.ContainerRecord.OutputPath { - runner.HostOutputDir, err = runner.MkTempDir("", "") - if err != nil { - return fmt.Errorf("While creating mount temp dir: %v", err) - } - st, staterr := os.Stat(runner.HostOutputDir) - if staterr != nil { - return fmt.Errorf("While Stat on temp dir: %v", staterr) - } - err = os.Chmod(runner.HostOutputDir, st.Mode()|os.ModeSetgid) - if staterr != nil { - return fmt.Errorf("While Chmod temp dir: %v", err) - } - runner.CleanupTempDir = append(runner.CleanupTempDir, runner.HostOutputDir) - runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", runner.HostOutputDir, bind)) - } else { - runner.Binds = append(runner.Binds, bind) + + case mnt.Kind == "tmp" && bind == runner.Container.OutputPath: + runner.HostOutputDir, err = runner.MkTempDir("", "") + if err != nil { + return fmt.Errorf("While creating mount temp dir: %v", err) } - } else { - return fmt.Errorf("Unknown mount kind '%s'", mnt.Kind) + st, staterr := os.Stat(runner.HostOutputDir) + if staterr != nil { + return fmt.Errorf("While Stat on temp dir: %v", staterr) + } + err = os.Chmod(runner.HostOutputDir, st.Mode()|os.ModeSetgid|0777) + if staterr != nil { + return fmt.Errorf("While Chmod temp dir: %v", err) + } + runner.CleanupTempDir = append(runner.CleanupTempDir, runner.HostOutputDir) + runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", runner.HostOutputDir, bind)) + + case mnt.Kind == "tmp": + runner.Binds = append(runner.Binds, bind) + + case mnt.Kind == "json": + jsondata, err := json.Marshal(mnt.Content) + if err != nil { + return fmt.Errorf("encoding json data: %v", err) + } + // Create a tempdir with a single file + // (instead of just a tempfile): this way we + // can ensure the file is world-readable + // inside the container, without having to + // make it world-readable on the docker host. + tmpdir, err := runner.MkTempDir("", "") + if err != nil { + return fmt.Errorf("creating temp dir: %v", err) + } + runner.CleanupTempDir = append(runner.CleanupTempDir, tmpdir) + tmpfn := filepath.Join(tmpdir, "mountdata.json") + err = ioutil.WriteFile(tmpfn, jsondata, 0644) + if err != nil { + return fmt.Errorf("writing temp file: %v", err) + } + runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", tmpfn, bind)) } } @@ -307,6 +396,16 @@ func (runner *ContainerRunner) SetupMounts() (err error) { return fmt.Errorf("Output path does not correspond to a writable mount point") } + if wantAPI := runner.Container.RuntimeConstraints.API; needCertMount && wantAPI != nil && *wantAPI { + for _, certfile := range arvadosclient.CertFiles { + _, err := os.Stat(certfile) + if err == nil { + runner.Binds = append(runner.Binds, fmt.Sprintf("%s:/etc/arvados/ca-certificates.crt:ro", certfile)) + break + } + } + } + if pdhOnly { arvMountCmd = append(arvMountCmd, "--mount-by-pdh", "by_id") } else { @@ -314,7 +413,12 @@ func (runner *ContainerRunner) SetupMounts() (err error) { } arvMountCmd = append(arvMountCmd, runner.ArvMountPoint) - runner.ArvMount, err = runner.RunArvMount(arvMountCmd) + token, err := runner.ContainerToken() + if err != nil { + return fmt.Errorf("could not get container token: %s", err) + } + + runner.ArvMount, err = runner.RunArvMount(arvMountCmd, token) if err != nil { return fmt.Errorf("While trying to start arv-mount: %v", err) } @@ -329,40 +433,6 @@ func (runner *ContainerRunner) SetupMounts() (err error) { return nil } -// StartContainer creates the container and runs it. -func (runner *ContainerRunner) StartContainer() (err error) { - runner.CrunchLog.Print("Creating Docker container") - - runner.CancelLock.Lock() - defer runner.CancelLock.Unlock() - - if runner.Cancelled { - return ErrCancelled - } - - runner.ContainerConfig.Cmd = runner.ContainerRecord.Command - if runner.ContainerRecord.Cwd != "." { - runner.ContainerConfig.WorkingDir = runner.ContainerRecord.Cwd - } - for k, v := range runner.ContainerRecord.Environment { - runner.ContainerConfig.Env = append(runner.ContainerConfig.Env, k+"="+v) - } - runner.ContainerConfig.NetworkDisabled = true - runner.ContainerID, err = runner.Docker.CreateContainer(&runner.ContainerConfig, "", nil) - if err != nil { - return fmt.Errorf("While creating container: %v", err) - } - hostConfig := &dockerclient.HostConfig{Binds: runner.Binds} - - runner.CrunchLog.Printf("Starting Docker container id '%s'", runner.ContainerID) - err = runner.Docker.StartContainer(runner.ContainerID, hostConfig) - if err != nil { - return fmt.Errorf("While starting container: %v", err) - } - - return nil -} - func (runner *ContainerRunner) ProcessDockerAttach(containerReader io.Reader) { // Handle docker log protocol // https://docs.docker.com/engine/reference/api/docker_remote_api_v1.15/#attach-to-a-container @@ -372,7 +442,7 @@ func (runner *ContainerRunner) ProcessDockerAttach(containerReader io.Reader) { _, readerr := io.ReadAtLeast(containerReader, header, 8) if readerr == nil { - readsize := int64(header[4]) | (int64(header[5]) << 8) | (int64(header[6]) << 16) | (int64(header[7]) << 24) + readsize := int64(header[7]) | (int64(header[6]) << 8) | (int64(header[5]) << 16) | (int64(header[4]) << 24) if header[0] == 1 { // stdout _, readerr = io.CopyN(runner.Stdout, containerReader, readsize) @@ -389,12 +459,20 @@ func (runner *ContainerRunner) ProcessDockerAttach(containerReader io.Reader) { closeerr := runner.Stdout.Close() if closeerr != nil { - runner.CrunchLog.Printf("While closing stdout logs: %v", readerr) + runner.CrunchLog.Printf("While closing stdout logs: %v", closeerr) } closeerr = runner.Stderr.Close() if closeerr != nil { - runner.CrunchLog.Printf("While closing stderr logs: %v", readerr) + runner.CrunchLog.Printf("While closing stderr logs: %v", closeerr) + } + + if runner.statReporter != nil { + runner.statReporter.Stop() + closeerr = runner.statLogger.Close() + if closeerr != nil { + runner.CrunchLog.Printf("While closing crunchstat logs: %v", closeerr) + } } runner.loggingDone <- true @@ -404,21 +482,58 @@ func (runner *ContainerRunner) ProcessDockerAttach(containerReader io.Reader) { } } +func (runner *ContainerRunner) StartCrunchstat() { + runner.statLogger = NewThrottledLogger(runner.NewLogWriter("crunchstat")) + runner.statReporter = &crunchstat.Reporter{ + CID: runner.ContainerID, + Logger: log.New(runner.statLogger, "", 0), + CgroupParent: runner.expectCgroupParent, + CgroupRoot: runner.cgroupRoot, + PollPeriod: runner.statInterval, + } + runner.statReporter.Start() +} + // AttachLogs connects the docker container stdout and stderr logs to the // Arvados logger which logs to Keep and the API server logs table. -func (runner *ContainerRunner) AttachLogs() (err error) { +func (runner *ContainerRunner) AttachStreams() (err error) { - runner.CrunchLog.Print("Attaching container logs") + runner.CrunchLog.Print("Attaching container streams") var containerReader io.Reader - containerReader, err = runner.Docker.ContainerLogs(runner.ContainerID, &dockerclient.LogOptions{Follow: true, Stdout: true, Stderr: true}) + containerReader, err = runner.Docker.AttachContainer(runner.ContainerID, + &dockerclient.AttachOptions{Stream: true, Stdout: true, Stderr: true}) if err != nil { - return fmt.Errorf("While attaching container logs: %v", err) + return fmt.Errorf("While attaching container stdout/stderr streams: %v", err) } runner.loggingDone = make(chan bool) - runner.Stdout = NewThrottledLogger(runner.NewLogWriter("stdout")) + if stdoutMnt, ok := runner.Container.Mounts["stdout"]; ok { + stdoutPath := stdoutMnt.Path[len(runner.Container.OutputPath):] + index := strings.LastIndex(stdoutPath, "/") + if index > 0 { + subdirs := stdoutPath[:index] + if subdirs != "" { + st, err := os.Stat(runner.HostOutputDir) + if err != nil { + return fmt.Errorf("While Stat on temp dir: %v", err) + } + stdoutPath := path.Join(runner.HostOutputDir, subdirs) + err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777) + if err != nil { + return fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err) + } + } + } + stdoutFile, err := os.Create(path.Join(runner.HostOutputDir, stdoutPath)) + if err != nil { + return fmt.Errorf("While creating stdout file: %v", err) + } + runner.Stdout = stdoutFile + } else { + runner.Stdout = NewThrottledLogger(runner.NewLogWriter("stdout")) + } runner.Stderr = NewThrottledLogger(runner.NewLogWriter("stderr")) go runner.ProcessDockerAttach(containerReader) @@ -426,6 +541,60 @@ func (runner *ContainerRunner) AttachLogs() (err error) { return nil } +// CreateContainer creates the docker container. +func (runner *ContainerRunner) CreateContainer() error { + runner.CrunchLog.Print("Creating Docker container") + + runner.ContainerConfig.Cmd = runner.Container.Command + if runner.Container.Cwd != "." { + runner.ContainerConfig.WorkingDir = runner.Container.Cwd + } + + for k, v := range runner.Container.Environment { + runner.ContainerConfig.Env = append(runner.ContainerConfig.Env, k+"="+v) + } + if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI { + tok, err := runner.ContainerToken() + if err != nil { + return err + } + runner.ContainerConfig.Env = append(runner.ContainerConfig.Env, + "ARVADOS_API_TOKEN="+tok, + "ARVADOS_API_HOST="+os.Getenv("ARVADOS_API_HOST"), + "ARVADOS_API_HOST_INSECURE="+os.Getenv("ARVADOS_API_HOST_INSECURE"), + ) + runner.ContainerConfig.NetworkDisabled = false + } else { + runner.ContainerConfig.NetworkDisabled = true + } + + var err error + runner.ContainerID, err = runner.Docker.CreateContainer(&runner.ContainerConfig, "", nil) + if err != nil { + return fmt.Errorf("While creating container: %v", err) + } + + runner.HostConfig = dockerclient.HostConfig{ + Binds: runner.Binds, + CgroupParent: runner.setCgroupParent, + LogConfig: dockerclient.LogConfig{ + Type: "none", + }, + } + + return runner.AttachStreams() +} + +// StartContainer starts the docker container created by CreateContainer. +func (runner *ContainerRunner) StartContainer() error { + runner.CrunchLog.Printf("Starting Docker container id '%s'", runner.ContainerID) + err := runner.Docker.StartContainer(runner.ContainerID, &runner.HostConfig) + if err != nil { + return fmt.Errorf("could not start container: %v", err) + } + return nil +} + // WaitFinish waits for the container to terminate, capture the exit code, and // close the stdout/stderr logging. func (runner *ContainerRunner) WaitFinish() error { @@ -450,6 +619,21 @@ func (runner *ContainerRunner) CaptureOutput() error { return nil } + if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI { + // Output may have been set directly by the container, so + // refresh the container record to check. + err := runner.ArvClient.Get("containers", runner.Container.UUID, + nil, &runner.Container) + if err != nil { + return err + } + if runner.Container.Output != "" { + // Container output is already set. + runner.OutputPDH = &runner.Container.Output + return nil + } + } + if runner.HostOutputDir == "" { return nil } @@ -478,7 +662,7 @@ func (runner *ContainerRunner) CaptureOutput() error { } defer file.Close() - rec := CollectionRecord{} + var rec arvados.Collection err = json.NewDecoder(file).Decode(&rec) if err != nil { return fmt.Errorf("While reading FUSE metafile: %v", err) @@ -486,20 +670,93 @@ func (runner *ContainerRunner) CaptureOutput() error { manifestText = rec.ManifestText } - var response CollectionRecord + // Pre-populate output from the configured mount points + var binds []string + for bind, _ := range runner.Container.Mounts { + binds = append(binds, bind) + } + sort.Strings(binds) + + for _, bind := range binds { + mnt := runner.Container.Mounts[bind] + + bindSuffix := strings.TrimPrefix(bind, runner.Container.OutputPath) + + if bindSuffix == bind || len(bindSuffix) <= 0 { + // either does not start with OutputPath or is OutputPath itself + continue + } + + if mnt.ExcludeFromOutput == true { + continue + } + + // append to manifest_text + m, err := runner.getCollectionManifestForPath(mnt, bindSuffix) + if err != nil { + return err + } + + manifestText = manifestText + m + } + + // Save output + var response arvados.Collection + manifest := manifest.Manifest{Text: manifestText} + manifestText = manifest.Extract(".", ".").Text err = runner.ArvClient.Create("collections", arvadosclient.Dict{ "collection": arvadosclient.Dict{ + "is_trashed": true, + "name": "output for " + runner.Container.UUID, "manifest_text": manifestText}}, &response) if err != nil { return fmt.Errorf("While creating output collection: %v", err) } + runner.OutputPDH = &response.PortableDataHash + return nil +} - runner.OutputPDH = new(string) - *runner.OutputPDH = response.PortableDataHash +var outputCollections = make(map[string]arvados.Collection) + +// Fetch the collection for the mnt.PortableDataHash +// Return the manifest_text fragment corresponding to the specified mnt.Path +// after making any required updates. +// Ex: +// If mnt.Path is not specified, +// return the entire manifest_text after replacing any "." with bindSuffix +// If mnt.Path corresponds to one stream, +// return the manifest_text for that stream after replacing that stream name with bindSuffix +// Otherwise, check if a filename in any one stream is being sought. Return the manifest_text +// for that stream after replacing stream name with bindSuffix minus the last word +// and the file name with last word of the bindSuffix +// Allowed path examples: +// "path":"/" +// "path":"/subdir1" +// "path":"/subdir1/subdir2" +// "path":"/subdir/filename" etc +func (runner *ContainerRunner) getCollectionManifestForPath(mnt arvados.Mount, bindSuffix string) (string, error) { + collection := outputCollections[mnt.PortableDataHash] + if collection.PortableDataHash == "" { + err := runner.ArvClient.Get("collections", mnt.PortableDataHash, nil, &collection) + if err != nil { + return "", fmt.Errorf("While getting collection for %v: %v", mnt.PortableDataHash, err) + } + outputCollections[mnt.PortableDataHash] = collection + } - return nil + if collection.ManifestText == "" { + runner.CrunchLog.Printf("No manifest text for collection %v", collection.PortableDataHash) + return "", nil + } + + mft := manifest.Manifest{Text: collection.ManifestText} + extracted := mft.Extract(mnt.Path, bindSuffix) + if extracted.Err != nil { + return "", fmt.Errorf("Error parsing manifest for %v: %v", mnt.PortableDataHash, extracted.Err.Error()) + } + return extracted.Text, nil } func (runner *ContainerRunner) CleanupDirs() { @@ -533,161 +790,205 @@ func (runner *ContainerRunner) CommitLogs() error { // point, but re-open crunch log with ArvClient in case there are any // other further (such as failing to write the log to Keep!) while // shutting down - runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{runner.ArvClient, runner.ContainerRecord.UUID, + runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{runner.ArvClient, runner.Container.UUID, "crunch-run", nil}) + if runner.LogsPDH != nil { + // If we have already assigned something to LogsPDH, + // we must be closing the re-opened log, which won't + // end up getting attached to the container record and + // therefore doesn't need to be saved as a collection + // -- it exists only to send logs to other channels. + return nil + } + mt, err := runner.LogCollection.ManifestText() if err != nil { return fmt.Errorf("While creating log manifest: %v", err) } - var response CollectionRecord + var response arvados.Collection err = runner.ArvClient.Create("collections", arvadosclient.Dict{ "collection": arvadosclient.Dict{ - "name": "logs for " + runner.ContainerRecord.UUID, + "is_trashed": true, + "name": "logs for " + runner.Container.UUID, "manifest_text": mt}}, &response) if err != nil { return fmt.Errorf("While creating log collection: %v", err) } - - runner.LogsPDH = new(string) - *runner.LogsPDH = response.PortableDataHash - + runner.LogsPDH = &response.PortableDataHash return nil } -// UpdateContainerRecordRunning updates the container state to "Running" -func (runner *ContainerRunner) UpdateContainerRecordRunning() error { - return runner.ArvClient.Update("containers", runner.ContainerRecord.UUID, +// UpdateContainerRunning updates the container state to "Running" +func (runner *ContainerRunner) UpdateContainerRunning() error { + runner.CancelLock.Lock() + defer runner.CancelLock.Unlock() + if runner.Cancelled { + return ErrCancelled + } + return runner.ArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{"container": arvadosclient.Dict{"state": "Running"}}, nil) } -// UpdateContainerRecordComplete updates the container record state on API +// ContainerToken returns the api_token the container (and any +// arv-mount processes) are allowed to use. +func (runner *ContainerRunner) ContainerToken() (string, error) { + if runner.token != "" { + return runner.token, nil + } + + var auth arvados.APIClientAuthorization + err := runner.ArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth) + if err != nil { + return "", err + } + runner.token = auth.APIToken + return runner.token, nil +} + +// UpdateContainerComplete updates the container record state on API // server to "Complete" or "Cancelled" -func (runner *ContainerRunner) UpdateContainerRecordComplete() error { +func (runner *ContainerRunner) UpdateContainerFinal() error { update := arvadosclient.Dict{} + update["state"] = runner.finalState if runner.LogsPDH != nil { update["log"] = *runner.LogsPDH } - if runner.ExitCode != nil { - update["exit_code"] = *runner.ExitCode - } - if runner.OutputPDH != nil { - update["output"] = runner.OutputPDH + if runner.finalState == "Complete" { + if runner.ExitCode != nil { + update["exit_code"] = *runner.ExitCode + } + if runner.OutputPDH != nil { + update["output"] = *runner.OutputPDH + } } + return runner.ArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{"container": update}, nil) +} - update["state"] = runner.finalState - - return runner.ArvClient.Update("containers", runner.ContainerRecord.UUID, arvadosclient.Dict{"container": update}, nil) +// IsCancelled returns the value of Cancelled, with goroutine safety. +func (runner *ContainerRunner) IsCancelled() bool { + runner.CancelLock.Lock() + defer runner.CancelLock.Unlock() + return runner.Cancelled } // NewArvLogWriter creates an ArvLogWriter func (runner *ContainerRunner) NewArvLogWriter(name string) io.WriteCloser { - return &ArvLogWriter{runner.ArvClient, runner.ContainerRecord.UUID, name, runner.LogCollection.Open(name + ".txt")} + return &ArvLogWriter{runner.ArvClient, runner.Container.UUID, name, runner.LogCollection.Open(name + ".txt")} } // Run the full container lifecycle. func (runner *ContainerRunner) Run() (err error) { - runner.CrunchLog.Printf("Executing container '%s'", runner.ContainerRecord.UUID) + runner.CrunchLog.Printf("Executing container '%s'", runner.Container.UUID) - var runerr, waiterr error + hostname, hosterr := os.Hostname() + if hosterr != nil { + runner.CrunchLog.Printf("Error getting hostname '%v'", hosterr) + } else { + runner.CrunchLog.Printf("Executing on host '%s'", hostname) + } - defer func() { - if err != nil { - runner.CrunchLog.Print(err) - } + // Clean up temporary directories _after_ finalizing + // everything (if we've made any by then) + defer runner.CleanupDirs() - if runner.Cancelled { - runner.finalState = "Cancelled" - } else { - runner.finalState = "Complete" - } + runner.finalState = "Queued" - // (7) capture output - outputerr := runner.CaptureOutput() - if outputerr != nil { - runner.CrunchLog.Print(outputerr) + defer func() { + // checkErr prints e (unless it's nil) and sets err to + // e (unless err is already non-nil). Thus, if err + // hasn't already been assigned when Run() returns, + // this cleanup func will cause Run() to return the + // first non-nil error that is passed to checkErr(). + checkErr := func(e error) { + if e == nil { + return + } + runner.CrunchLog.Print(e) + if err == nil { + err = e + } } - // (8) clean up temporary directories - runner.CleanupDirs() + // Log the error encountered in Run(), if any + checkErr(err) - // (9) write logs - logerr := runner.CommitLogs() - if logerr != nil { - runner.CrunchLog.Print(logerr) + if runner.finalState == "Queued" { + runner.CrunchLog.Close() + runner.UpdateContainerFinal() + return } - // (10) update container record with results - updateerr := runner.UpdateContainerRecordComplete() - if updateerr != nil { - runner.CrunchLog.Print(updateerr) + if runner.IsCancelled() { + runner.finalState = "Cancelled" + // but don't return yet -- we still want to + // capture partial output and write logs } - runner.CrunchLog.Close() + checkErr(runner.CaptureOutput()) + checkErr(runner.CommitLogs()) + checkErr(runner.UpdateContainerFinal()) - if err == nil { - if runerr != nil { - err = runerr - } else if waiterr != nil { - err = runerr - } else if logerr != nil { - err = logerr - } else if updateerr != nil { - err = updateerr - } - } + // The real log is already closed, but then we opened + // a new one in case we needed to log anything while + // finalizing. + runner.CrunchLog.Close() }() - err = runner.ArvClient.Get("containers", runner.ContainerRecord.UUID, nil, &runner.ContainerRecord) + err = runner.ArvClient.Get("containers", runner.Container.UUID, nil, &runner.Container) if err != nil { - return fmt.Errorf("While getting container record: %v", err) + err = fmt.Errorf("While getting container record: %v", err) + return } - // (1) setup signal handling - err = runner.SetupSignals() - if err != nil { - return fmt.Errorf("While setting up signal handling: %v", err) - } + // setup signal handling + runner.SetupSignals() - // (2) check for and/or load image + // check for and/or load image err = runner.LoadImage() if err != nil { - return fmt.Errorf("While loading container image: %v", err) + runner.finalState = "Cancelled" + err = fmt.Errorf("While loading container image: %v", err) + return } - // (3) set up FUSE mount and binds + // set up FUSE mount and binds err = runner.SetupMounts() if err != nil { - return fmt.Errorf("While setting up mounts: %v", err) + runner.finalState = "Cancelled" + err = fmt.Errorf("While setting up mounts: %v", err) + return } - // (3) create and start container - err = runner.StartContainer() + err = runner.CreateContainer() if err != nil { - if err == ErrCancelled { - err = nil - } return } - // (4) update container record state - err = runner.UpdateContainerRecordRunning() - if err != nil { - runner.CrunchLog.Print(err) + runner.StartCrunchstat() + + if runner.IsCancelled() { + return } - // (5) attach container logs - runerr = runner.AttachLogs() - if runerr != nil { - runner.CrunchLog.Print(runerr) + err = runner.UpdateContainerRunning() + if err != nil { + return } + runner.finalState = "Cancelled" - // (6) wait for container to finish - waiterr = runner.WaitFinish() + err = runner.StartContainer() + if err != nil { + return + } + err = runner.WaitFinish() + if err == nil { + runner.finalState = "Complete" + } return } @@ -702,38 +1003,58 @@ func NewContainerRunner(api IArvadosClient, cr.RunArvMount = cr.ArvMountCmd cr.MkTempDir = ioutil.TempDir cr.LogCollection = &CollectionWriter{kc, nil, sync.Mutex{}} - cr.ContainerRecord.UUID = containerUUID + cr.Container.UUID = containerUUID cr.CrunchLog = NewThrottledLogger(cr.NewLogWriter("crunch-run")) + cr.CrunchLog.Immediate = log.New(os.Stderr, containerUUID+" ", 0) return cr } func main() { + statInterval := flag.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting") + cgroupRoot := flag.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree") + cgroupParent := flag.String("cgroup-parent", "docker", "name of container's parent cgroup (ignored if -cgroup-parent-subsystem is used)") + cgroupParentSubsystem := flag.String("cgroup-parent-subsystem", "", "use current cgroup for given subsystem as parent cgroup for container") + caCertsPath := flag.String("ca-certs", "", "Path to TLS root certificates") flag.Parse() + containerId := flag.Arg(0) + + if *caCertsPath != "" { + arvadosclient.CertFiles = []string{*caCertsPath} + } + api, err := arvadosclient.MakeArvadosClient() if err != nil { - log.Fatalf("%s: %v", flag.Arg(0), err) + log.Fatalf("%s: %v", containerId, err) } api.Retries = 8 var kc *keepclient.KeepClient - kc, err = keepclient.MakeKeepClient(&api) + kc, err = keepclient.MakeKeepClient(api) if err != nil { - log.Fatalf("%s: %v", flag.Arg(0), err) + log.Fatalf("%s: %v", containerId, err) } kc.Retries = 4 var docker *dockerclient.DockerClient docker, err = dockerclient.NewDockerClient("unix:///var/run/docker.sock", nil) if err != nil { - log.Fatalf("%s: %v", flag.Arg(0), err) + log.Fatalf("%s: %v", containerId, err) } - cr := NewContainerRunner(api, kc, docker, flag.Arg(0)) + cr := NewContainerRunner(api, kc, docker, containerId) + cr.statInterval = *statInterval + cr.cgroupRoot = *cgroupRoot + cr.expectCgroupParent = *cgroupParent + if *cgroupParentSubsystem != "" { + p := findCgroup(*cgroupParentSubsystem) + cr.setCgroupParent = p + cr.expectCgroupParent = p + } err = cr.Run() if err != nil { - log.Fatalf("%s: %v", flag.Arg(0), err) + log.Fatalf("%s: %v", containerId, err) } }