services/crunch-run/crunchrun.go

   1 package main
   2
   3 import (
   4         "encoding/json"
   5         "errors"
   6         "flag"
   7         "fmt"
   8         "git.curoverse.com/arvados.git/lib/crunchstat"
   9         "git.curoverse.com/arvados.git/sdk/go/arvados"
  10         "git.curoverse.com/arvados.git/sdk/go/arvadosclient"
  11         "git.curoverse.com/arvados.git/sdk/go/keepclient"
  12         "git.curoverse.com/arvados.git/sdk/go/manifest"
  13         "github.com/curoverse/dockerclient"
  14         "io"
  15         "io/ioutil"
  16         "log"
  17         "os"
  18         "os/exec"
  19         "os/signal"
  20         "path"
  21         "path/filepath"
  22         "strings"
  23         "sync"
  24         "syscall"
  25         "time"
  26 )
  27
  28 // IArvadosClient is the minimal Arvados API methods used by crunch-run.
  29 type IArvadosClient interface {
  30         Create(resourceType string, parameters arvadosclient.Dict, output interface{}) error
  31         Get(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  32         Update(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  33         Call(method, resourceType, uuid, action string, parameters arvadosclient.Dict, output interface{}) error
  34         Discovery(key string) (interface{}, error)
  35 }
  36
  37 // ErrCancelled is the error returned when the container is cancelled.
  38 var ErrCancelled = errors.New("Cancelled")
  39
  40 // IKeepClient is the minimal Keep API methods used by crunch-run.
  41 type IKeepClient interface {
  42         PutHB(hash string, buf []byte) (string, int, error)
  43         ManifestFileReader(m manifest.Manifest, filename string) (keepclient.ReadCloserWithLen, error)
  44 }
  45
  46 // NewLogWriter is a factory function to create a new log writer.
  47 type NewLogWriter func(name string) io.WriteCloser
  48
  49 type RunArvMount func(args []string, tok string) (*exec.Cmd, error)
  50
  51 type MkTempDir func(string, string) (string, error)
  52
  53 // ThinDockerClient is the minimal Docker client interface used by crunch-run.
  54 type ThinDockerClient interface {
  55         StopContainer(id string, timeout int) error
  56         InspectImage(id string) (*dockerclient.ImageInfo, error)
  57         LoadImage(reader io.Reader) error
  58         CreateContainer(config *dockerclient.ContainerConfig, name string, authConfig *dockerclient.AuthConfig) (string, error)
  59         StartContainer(id string, config *dockerclient.HostConfig) error
  60         AttachContainer(id string, options *dockerclient.AttachOptions) (io.ReadCloser, error)
  61         Wait(id string) <-chan dockerclient.WaitResult
  62         RemoveImage(name string, force bool) ([]*dockerclient.ImageDelete, error)
  63 }
  64
  65 // ContainerRunner is the main stateful struct used for a single execution of a
  66 // container.
  67 type ContainerRunner struct {
  68         Docker    ThinDockerClient
  69         ArvClient IArvadosClient
  70         Kc        IKeepClient
  71         arvados.Container
  72         dockerclient.ContainerConfig
  73         dockerclient.HostConfig
  74         token       string
  75         ContainerID string
  76         ExitCode    *int
  77         NewLogWriter
  78         loggingDone   chan bool
  79         CrunchLog     *ThrottledLogger
  80         Stdout        io.WriteCloser
  81         Stderr        *ThrottledLogger
  82         LogCollection *CollectionWriter
  83         LogsPDH       *string
  84         RunArvMount
  85         MkTempDir
  86         ArvMount       *exec.Cmd
  87         ArvMountPoint  string
  88         HostOutputDir  string
  89         CleanupTempDir []string
  90         Binds          []string
  91         OutputPDH      *string
  92         CancelLock     sync.Mutex
  93         Cancelled      bool
  94         SigChan        chan os.Signal
  95         ArvMountExit   chan error
  96         finalState     string
  97         trashLifetime  time.Duration
  98
  99         statLogger   io.WriteCloser
 100         statReporter *crunchstat.Reporter
 101         statInterval time.Duration
 102         cgroupRoot   string
 103         // What we expect the container's cgroup parent to be.
 104         expectCgroupParent string
 105         // What we tell docker to use as the container's cgroup
 106         // parent. Note: Ideally we would use the same field for both
 107         // expectCgroupParent and setCgroupParent, and just make it
 108         // default to "docker". However, when using docker < 1.10 with
 109         // systemd, specifying a non-empty cgroup parent (even the
 110         // default value "docker") hits a docker bug
 111         // (https://github.com/docker/docker/issues/17126). Using two
 112         // separate fields makes it possible to use the "expect cgroup
 113         // parent to be X" feature even on sites where the "specify
 114         // cgroup parent" feature breaks.
 115         setCgroupParent string
 116 }
 117
 118 // SetupSignals sets up signal handling to gracefully terminate the underlying
 119 // Docker container and update state when receiving a TERM, INT or QUIT signal.
 120 func (runner *ContainerRunner) SetupSignals() {
 121         runner.SigChan = make(chan os.Signal, 1)
 122         signal.Notify(runner.SigChan, syscall.SIGTERM)
 123         signal.Notify(runner.SigChan, syscall.SIGINT)
 124         signal.Notify(runner.SigChan, syscall.SIGQUIT)
 125
 126         go func(sig <-chan os.Signal) {
 127                 for range sig {
 128                         if !runner.Cancelled {
 129                                 runner.CancelLock.Lock()
 130                                 runner.Cancelled = true
 131                                 if runner.ContainerID != "" {
 132                                         runner.Docker.StopContainer(runner.ContainerID, 10)
 133                                 }
 134                                 runner.CancelLock.Unlock()
 135                         }
 136                 }
 137         }(runner.SigChan)
 138 }
 139
 140 // LoadImage determines the docker image id from the container record and
 141 // checks if it is available in the local Docker image store.  If not, it loads
 142 // the image from Keep.
 143 func (runner *ContainerRunner) LoadImage() (err error) {
 144
 145         runner.CrunchLog.Printf("Fetching Docker image from collection '%s'", runner.Container.ContainerImage)
 146
 147         var collection arvados.Collection
 148         err = runner.ArvClient.Get("collections", runner.Container.ContainerImage, nil, &collection)
 149         if err != nil {
 150                 return fmt.Errorf("While getting container image collection: %v", err)
 151         }
 152         manifest := manifest.Manifest{Text: collection.ManifestText}
 153         var img, imageID string
 154         for ms := range manifest.StreamIter() {
 155                 img = ms.FileStreamSegments[0].Name
 156                 if !strings.HasSuffix(img, ".tar") {
 157                         return fmt.Errorf("First file in the container image collection does not end in .tar")
 158                 }
 159                 imageID = img[:len(img)-4]
 160         }
 161
 162         runner.CrunchLog.Printf("Using Docker image id '%s'", imageID)
 163
 164         _, err = runner.Docker.InspectImage(imageID)
 165         if err != nil {
 166                 runner.CrunchLog.Print("Loading Docker image from keep")
 167
 168                 var readCloser io.ReadCloser
 169                 readCloser, err = runner.Kc.ManifestFileReader(manifest, img)
 170                 if err != nil {
 171                         return fmt.Errorf("While creating ManifestFileReader for container image: %v", err)
 172                 }
 173
 174                 err = runner.Docker.LoadImage(readCloser)
 175                 if err != nil {
 176                         return fmt.Errorf("While loading container image into Docker: %v", err)
 177                 }
 178         } else {
 179                 runner.CrunchLog.Print("Docker image is available")
 180         }
 181
 182         runner.ContainerConfig.Image = imageID
 183
 184         return nil
 185 }
 186
 187 func (runner *ContainerRunner) ArvMountCmd(arvMountCmd []string, token string) (c *exec.Cmd, err error) {
 188         c = exec.Command("arv-mount", arvMountCmd...)
 189
 190         // Copy our environment, but override ARVADOS_API_TOKEN with
 191         // the container auth token.
 192         c.Env = nil
 193         for _, s := range os.Environ() {
 194                 if !strings.HasPrefix(s, "ARVADOS_API_TOKEN=") {
 195                         c.Env = append(c.Env, s)
 196                 }
 197         }
 198         c.Env = append(c.Env, "ARVADOS_API_TOKEN="+token)
 199
 200         nt := NewThrottledLogger(runner.NewLogWriter("arv-mount"))
 201         c.Stdout = nt
 202         c.Stderr = nt
 203
 204         err = c.Start()
 205         if err != nil {
 206                 return nil, err
 207         }
 208
 209         statReadme := make(chan bool)
 210         runner.ArvMountExit = make(chan error)
 211
 212         keepStatting := true
 213         go func() {
 214                 for keepStatting {
 215                         time.Sleep(100 * time.Millisecond)
 216                         _, err = os.Stat(fmt.Sprintf("%s/by_id/README", runner.ArvMountPoint))
 217                         if err == nil {
 218                                 keepStatting = false
 219                                 statReadme <- true
 220                         }
 221                 }
 222                 close(statReadme)
 223         }()
 224
 225         go func() {
 226                 runner.ArvMountExit <- c.Wait()
 227                 close(runner.ArvMountExit)
 228         }()
 229
 230         select {
 231         case <-statReadme:
 232                 break
 233         case err := <-runner.ArvMountExit:
 234                 runner.ArvMount = nil
 235                 keepStatting = false
 236                 return nil, err
 237         }
 238
 239         return c, nil
 240 }
 241
 242 func (runner *ContainerRunner) SetupMounts() (err error) {
 243         runner.ArvMountPoint, err = runner.MkTempDir("", "keep")
 244         if err != nil {
 245                 return fmt.Errorf("While creating keep mount temp dir: %v", err)
 246         }
 247
 248         runner.CleanupTempDir = append(runner.CleanupTempDir, runner.ArvMountPoint)
 249
 250         pdhOnly := true
 251         tmpcount := 0
 252         arvMountCmd := []string{"--foreground", "--allow-other", "--read-write"}
 253         collectionPaths := []string{}
 254         runner.Binds = nil
 255
 256         for bind, mnt := range runner.Container.Mounts {
 257                 if bind == "stdout" {
 258                         // Is it a "file" mount kind?
 259                         if mnt.Kind != "file" {
 260                                 return fmt.Errorf("Unsupported mount kind '%s' for stdout. Only 'file' is supported.", mnt.Kind)
 261                         }
 262
 263                         // Does path start with OutputPath?
 264                         prefix := runner.Container.OutputPath
 265                         if !strings.HasSuffix(prefix, "/") {
 266                                 prefix += "/"
 267                         }
 268                         if !strings.HasPrefix(mnt.Path, prefix) {
 269                                 return fmt.Errorf("Stdout path does not start with OutputPath: %s, %s", mnt.Path, prefix)
 270                         }
 271                 }
 272
 273                 switch {
 274                 case mnt.Kind == "collection":
 275                         var src string
 276                         if mnt.UUID != "" && mnt.PortableDataHash != "" {
 277                                 return fmt.Errorf("Cannot specify both 'uuid' and 'portable_data_hash' for a collection mount")
 278                         }
 279                         if mnt.UUID != "" {
 280                                 if mnt.Writable {
 281                                         return fmt.Errorf("Writing to existing collections currently not permitted.")
 282                                 }
 283                                 pdhOnly = false
 284                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.UUID)
 285                         } else if mnt.PortableDataHash != "" {
 286                                 if mnt.Writable {
 287                                         return fmt.Errorf("Can never write to a collection specified by portable data hash")
 288                                 }
 289                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.PortableDataHash)
 290                         } else {
 291                                 src = fmt.Sprintf("%s/tmp%d", runner.ArvMountPoint, tmpcount)
 292                                 arvMountCmd = append(arvMountCmd, "--mount-tmp")
 293                                 arvMountCmd = append(arvMountCmd, fmt.Sprintf("tmp%d", tmpcount))
 294                                 tmpcount += 1
 295                         }
 296                         if mnt.Writable {
 297                                 if bind == runner.Container.OutputPath {
 298                                         runner.HostOutputDir = src
 299                                 }
 300                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", src, bind))
 301                         } else {
 302                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", src, bind))
 303                         }
 304                         collectionPaths = append(collectionPaths, src)
 305
 306                 case mnt.Kind == "tmp" && bind == runner.Container.OutputPath:
 307                         runner.HostOutputDir, err = runner.MkTempDir("", "")
 308                         if err != nil {
 309                                 return fmt.Errorf("While creating mount temp dir: %v", err)
 310                         }
 311                         st, staterr := os.Stat(runner.HostOutputDir)
 312                         if staterr != nil {
 313                                 return fmt.Errorf("While Stat on temp dir: %v", staterr)
 314                         }
 315                         err = os.Chmod(runner.HostOutputDir, st.Mode()|os.ModeSetgid|0777)
 316                         if staterr != nil {
 317                                 return fmt.Errorf("While Chmod temp dir: %v", err)
 318                         }
 319                         runner.CleanupTempDir = append(runner.CleanupTempDir, runner.HostOutputDir)
 320                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", runner.HostOutputDir, bind))
 321
 322                 case mnt.Kind == "tmp":
 323                         runner.Binds = append(runner.Binds, bind)
 324
 325                 case mnt.Kind == "json":
 326                         jsondata, err := json.Marshal(mnt.Content)
 327                         if err != nil {
 328                                 return fmt.Errorf("encoding json data: %v", err)
 329                         }
 330                         // Create a tempdir with a single file
 331                         // (instead of just a tempfile): this way we
 332                         // can ensure the file is world-readable
 333                         // inside the container, without having to
 334                         // make it world-readable on the docker host.
 335                         tmpdir, err := runner.MkTempDir("", "")
 336                         if err != nil {
 337                                 return fmt.Errorf("creating temp dir: %v", err)
 338                         }
 339                         runner.CleanupTempDir = append(runner.CleanupTempDir, tmpdir)
 340                         tmpfn := filepath.Join(tmpdir, "mountdata.json")
 341                         err = ioutil.WriteFile(tmpfn, jsondata, 0644)
 342                         if err != nil {
 343                                 return fmt.Errorf("writing temp file: %v", err)
 344                         }
 345                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", tmpfn, bind))
 346                 }
 347         }
 348
 349         if runner.HostOutputDir == "" {
 350                 return fmt.Errorf("Output path does not correspond to a writable mount point")
 351         }
 352
 353         if pdhOnly {
 354                 arvMountCmd = append(arvMountCmd, "--mount-by-pdh", "by_id")
 355         } else {
 356                 arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_id")
 357         }
 358         arvMountCmd = append(arvMountCmd, runner.ArvMountPoint)
 359
 360         if runner.Container.RuntimeConstraints.KeepCacheRAM > 0 {
 361                 arvMountCmd = append(arvMountCmd, "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheRAM))
 362         }
 363
 364         token, err := runner.ContainerToken()
 365         if err != nil {
 366                 return fmt.Errorf("could not get container token: %s", err)
 367         }
 368
 369         runner.ArvMount, err = runner.RunArvMount(arvMountCmd, token)
 370         if err != nil {
 371                 return fmt.Errorf("While trying to start arv-mount: %v", err)
 372         }
 373
 374         for _, p := range collectionPaths {
 375                 _, err = os.Stat(p)
 376                 if err != nil {
 377                         return fmt.Errorf("While checking that input files exist: %v", err)
 378                 }
 379         }
 380
 381         return nil
 382 }
 383
 384 func (runner *ContainerRunner) ProcessDockerAttach(containerReader io.Reader) {
 385         // Handle docker log protocol
 386         // https://docs.docker.com/engine/reference/api/docker_remote_api_v1.15/#attach-to-a-container
 387
 388         header := make([]byte, 8)
 389         for {
 390                 _, readerr := io.ReadAtLeast(containerReader, header, 8)
 391
 392                 if readerr == nil {
 393                         readsize := int64(header[7]) | (int64(header[6]) << 8) | (int64(header[5]) << 16) | (int64(header[4]) << 24)
 394                         if header[0] == 1 {
 395                                 // stdout
 396                                 _, readerr = io.CopyN(runner.Stdout, containerReader, readsize)
 397                         } else {
 398                                 // stderr
 399                                 _, readerr = io.CopyN(runner.Stderr, containerReader, readsize)
 400                         }
 401                 }
 402
 403                 if readerr != nil {
 404                         if readerr != io.EOF {
 405                                 runner.CrunchLog.Printf("While reading docker logs: %v", readerr)
 406                         }
 407
 408                         closeerr := runner.Stdout.Close()
 409                         if closeerr != nil {
 410                                 runner.CrunchLog.Printf("While closing stdout logs: %v", closeerr)
 411                         }
 412
 413                         closeerr = runner.Stderr.Close()
 414                         if closeerr != nil {
 415                                 runner.CrunchLog.Printf("While closing stderr logs: %v", closeerr)
 416                         }
 417
 418                         if runner.statReporter != nil {
 419                                 runner.statReporter.Stop()
 420                                 closeerr = runner.statLogger.Close()
 421                                 if closeerr != nil {
 422                                         runner.CrunchLog.Printf("While closing crunchstat logs: %v", closeerr)
 423                                 }
 424                         }
 425
 426                         runner.loggingDone <- true
 427                         close(runner.loggingDone)
 428                         return
 429                 }
 430         }
 431 }
 432
 433 func (runner *ContainerRunner) StartCrunchstat() {
 434         runner.statLogger = NewThrottledLogger(runner.NewLogWriter("crunchstat"))
 435         runner.statReporter = &crunchstat.Reporter{
 436                 CID:          runner.ContainerID,
 437                 Logger:       log.New(runner.statLogger, "", 0),
 438                 CgroupParent: runner.expectCgroupParent,
 439                 CgroupRoot:   runner.cgroupRoot,
 440                 PollPeriod:   runner.statInterval,
 441         }
 442         runner.statReporter.Start()
 443 }
 444
 445 // AttachLogs connects the docker container stdout and stderr logs to the
 446 // Arvados logger which logs to Keep and the API server logs table.
 447 func (runner *ContainerRunner) AttachStreams() (err error) {
 448
 449         runner.CrunchLog.Print("Attaching container streams")
 450
 451         var containerReader io.Reader
 452         containerReader, err = runner.Docker.AttachContainer(runner.ContainerID,
 453                 &dockerclient.AttachOptions{Stream: true, Stdout: true, Stderr: true})
 454         if err != nil {
 455                 return fmt.Errorf("While attaching container stdout/stderr streams: %v", err)
 456         }
 457
 458         runner.loggingDone = make(chan bool)
 459
 460         if stdoutMnt, ok := runner.Container.Mounts["stdout"]; ok {
 461                 stdoutPath := stdoutMnt.Path[len(runner.Container.OutputPath):]
 462                 index := strings.LastIndex(stdoutPath, "/")
 463                 if index > 0 {
 464                         subdirs := stdoutPath[:index]
 465                         if subdirs != "" {
 466                                 st, err := os.Stat(runner.HostOutputDir)
 467                                 if err != nil {
 468                                         return fmt.Errorf("While Stat on temp dir: %v", err)
 469                                 }
 470                                 stdoutPath := path.Join(runner.HostOutputDir, subdirs)
 471                                 err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777)
 472                                 if err != nil {
 473                                         return fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err)
 474                                 }
 475                         }
 476                 }
 477                 stdoutFile, err := os.Create(path.Join(runner.HostOutputDir, stdoutPath))
 478                 if err != nil {
 479                         return fmt.Errorf("While creating stdout file: %v", err)
 480                 }
 481                 runner.Stdout = stdoutFile
 482         } else {
 483                 runner.Stdout = NewThrottledLogger(runner.NewLogWriter("stdout"))
 484         }
 485         runner.Stderr = NewThrottledLogger(runner.NewLogWriter("stderr"))
 486
 487         go runner.ProcessDockerAttach(containerReader)
 488
 489         return nil
 490 }
 491
 492 // CreateContainer creates the docker container.
 493 func (runner *ContainerRunner) CreateContainer() error {
 494         runner.CrunchLog.Print("Creating Docker container")
 495
 496         runner.ContainerConfig.Cmd = runner.Container.Command
 497         if runner.Container.Cwd != "." {
 498                 runner.ContainerConfig.WorkingDir = runner.Container.Cwd
 499         }
 500
 501         for k, v := range runner.Container.Environment {
 502                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env, k+"="+v)
 503         }
 504         if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
 505                 tok, err := runner.ContainerToken()
 506                 if err != nil {
 507                         return err
 508                 }
 509                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env,
 510                         "ARVADOS_API_TOKEN="+tok,
 511                         "ARVADOS_API_HOST="+os.Getenv("ARVADOS_API_HOST"),
 512                         "ARVADOS_API_HOST_INSECURE="+os.Getenv("ARVADOS_API_HOST_INSECURE"),
 513                 )
 514                 runner.ContainerConfig.NetworkDisabled = false
 515         } else {
 516                 runner.ContainerConfig.NetworkDisabled = true
 517         }
 518
 519         var err error
 520         runner.ContainerID, err = runner.Docker.CreateContainer(&runner.ContainerConfig, "", nil)
 521         if err != nil {
 522                 return fmt.Errorf("While creating container: %v", err)
 523         }
 524
 525         runner.HostConfig = dockerclient.HostConfig{
 526                 Binds:        runner.Binds,
 527                 CgroupParent: runner.setCgroupParent,
 528                 LogConfig: dockerclient.LogConfig{
 529                         Type: "none",
 530                 },
 531         }
 532
 533         return runner.AttachStreams()
 534 }
 535
 536 // StartContainer starts the docker container created by CreateContainer.
 537 func (runner *ContainerRunner) StartContainer() error {
 538         runner.CrunchLog.Printf("Starting Docker container id '%s'", runner.ContainerID)
 539         err := runner.Docker.StartContainer(runner.ContainerID, &runner.HostConfig)
 540         if err != nil {
 541                 return fmt.Errorf("could not start container: %v", err)
 542         }
 543         return nil
 544 }
 545
 546 // WaitFinish waits for the container to terminate, capture the exit code, and
 547 // close the stdout/stderr logging.
 548 func (runner *ContainerRunner) WaitFinish() error {
 549         runner.CrunchLog.Print("Waiting for container to finish")
 550
 551         result := runner.Docker.Wait(runner.ContainerID)
 552         wr := <-result
 553         if wr.Error != nil {
 554                 return fmt.Errorf("While waiting for container to finish: %v", wr.Error)
 555         }
 556         runner.ExitCode = &wr.ExitCode
 557
 558         // wait for stdout/stderr to complete
 559         <-runner.loggingDone
 560
 561         return nil
 562 }
 563
 564 // HandleOutput sets the output, unmounts the FUSE mount, and deletes temporary directories
 565 func (runner *ContainerRunner) CaptureOutput() error {
 566         if runner.finalState != "Complete" {
 567                 return nil
 568         }
 569
 570         if runner.HostOutputDir == "" {
 571                 return nil
 572         }
 573
 574         _, err := os.Stat(runner.HostOutputDir)
 575         if err != nil {
 576                 return fmt.Errorf("While checking host output path: %v", err)
 577         }
 578
 579         var manifestText string
 580
 581         collectionMetafile := fmt.Sprintf("%s/.arvados#collection", runner.HostOutputDir)
 582         _, err = os.Stat(collectionMetafile)
 583         if err != nil {
 584                 // Regular directory
 585                 cw := CollectionWriter{runner.Kc, nil, sync.Mutex{}}
 586                 manifestText, err = cw.WriteTree(runner.HostOutputDir, runner.CrunchLog.Logger)
 587                 if err != nil {
 588                         return fmt.Errorf("While uploading output files: %v", err)
 589                 }
 590         } else {
 591                 // FUSE mount directory
 592                 file, openerr := os.Open(collectionMetafile)
 593                 if openerr != nil {
 594                         return fmt.Errorf("While opening FUSE metafile: %v", err)
 595                 }
 596                 defer file.Close()
 597
 598                 var rec arvados.Collection
 599                 err = json.NewDecoder(file).Decode(&rec)
 600                 if err != nil {
 601                         return fmt.Errorf("While reading FUSE metafile: %v", err)
 602                 }
 603                 manifestText = rec.ManifestText
 604         }
 605
 606         var response arvados.Collection
 607         err = runner.ArvClient.Create("collections",
 608                 arvadosclient.Dict{
 609                         "collection": arvadosclient.Dict{
 610                                 "expires_at":    time.Now().Add(runner.trashLifetime).Format(time.RFC3339),
 611                                 "name":          "output for " + runner.Container.UUID,
 612                                 "manifest_text": manifestText}},
 613                 &response)
 614         if err != nil {
 615                 return fmt.Errorf("While creating output collection: %v", err)
 616         }
 617         runner.OutputPDH = &response.PortableDataHash
 618         return nil
 619 }
 620
 621 func (runner *ContainerRunner) loadDiscoveryVars() {
 622         tl, err := runner.ArvClient.Discovery("defaultTrashLifetime")
 623         if err != nil {
 624                 log.Fatalf("getting defaultTrashLifetime from discovery document: %s", err)
 625         }
 626         runner.trashLifetime = time.Duration(tl.(float64)) * time.Second
 627 }
 628
 629 func (runner *ContainerRunner) CleanupDirs() {
 630         if runner.ArvMount != nil {
 631                 umount := exec.Command("fusermount", "-z", "-u", runner.ArvMountPoint)
 632                 umnterr := umount.Run()
 633                 if umnterr != nil {
 634                         runner.CrunchLog.Printf("While running fusermount: %v", umnterr)
 635                 }
 636
 637                 mnterr := <-runner.ArvMountExit
 638                 if mnterr != nil {
 639                         runner.CrunchLog.Printf("Arv-mount exit error: %v", mnterr)
 640                 }
 641         }
 642
 643         for _, tmpdir := range runner.CleanupTempDir {
 644                 rmerr := os.RemoveAll(tmpdir)
 645                 if rmerr != nil {
 646                         runner.CrunchLog.Printf("While cleaning up temporary directory %s: %v", tmpdir, rmerr)
 647                 }
 648         }
 649 }
 650
 651 // CommitLogs posts the collection containing the final container logs.
 652 func (runner *ContainerRunner) CommitLogs() error {
 653         runner.CrunchLog.Print(runner.finalState)
 654         runner.CrunchLog.Close()
 655
 656         // Closing CrunchLog above allows it to be committed to Keep at this
 657         // point, but re-open crunch log with ArvClient in case there are any
 658         // other further (such as failing to write the log to Keep!) while
 659         // shutting down
 660         runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{runner.ArvClient, runner.Container.UUID,
 661                 "crunch-run", nil})
 662
 663         if runner.LogsPDH != nil {
 664                 // If we have already assigned something to LogsPDH,
 665                 // we must be closing the re-opened log, which won't
 666                 // end up getting attached to the container record and
 667                 // therefore doesn't need to be saved as a collection
 668                 // -- it exists only to send logs to other channels.
 669                 return nil
 670         }
 671
 672         mt, err := runner.LogCollection.ManifestText()
 673         if err != nil {
 674                 return fmt.Errorf("While creating log manifest: %v", err)
 675         }
 676
 677         var response arvados.Collection
 678         err = runner.ArvClient.Create("collections",
 679                 arvadosclient.Dict{
 680                         "collection": arvadosclient.Dict{
 681                                 "expires_at":    time.Now().Add(runner.trashLifetime).Format(time.RFC3339),
 682                                 "name":          "logs for " + runner.Container.UUID,
 683                                 "manifest_text": mt}},
 684                 &response)
 685         if err != nil {
 686                 return fmt.Errorf("While creating log collection: %v", err)
 687         }
 688         runner.LogsPDH = &response.PortableDataHash
 689         return nil
 690 }
 691
 692 // UpdateContainerRunning updates the container state to "Running"
 693 func (runner *ContainerRunner) UpdateContainerRunning() error {
 694         runner.CancelLock.Lock()
 695         defer runner.CancelLock.Unlock()
 696         if runner.Cancelled {
 697                 return ErrCancelled
 698         }
 699         return runner.ArvClient.Update("containers", runner.Container.UUID,
 700                 arvadosclient.Dict{"container": arvadosclient.Dict{"state": "Running"}}, nil)
 701 }
 702
 703 // ContainerToken returns the api_token the container (and any
 704 // arv-mount processes) are allowed to use.
 705 func (runner *ContainerRunner) ContainerToken() (string, error) {
 706         if runner.token != "" {
 707                 return runner.token, nil
 708         }
 709
 710         var auth arvados.APIClientAuthorization
 711         err := runner.ArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth)
 712         if err != nil {
 713                 return "", err
 714         }
 715         runner.token = auth.APIToken
 716         return runner.token, nil
 717 }
 718
 719 // UpdateContainerComplete updates the container record state on API
 720 // server to "Complete" or "Cancelled"
 721 func (runner *ContainerRunner) UpdateContainerFinal() error {
 722         update := arvadosclient.Dict{}
 723         update["state"] = runner.finalState
 724         if runner.finalState == "Complete" {
 725                 if runner.LogsPDH != nil {
 726                         update["log"] = *runner.LogsPDH
 727                 }
 728                 if runner.ExitCode != nil {
 729                         update["exit_code"] = *runner.ExitCode
 730                 }
 731                 if runner.OutputPDH != nil {
 732                         update["output"] = *runner.OutputPDH
 733                 }
 734         }
 735         return runner.ArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{"container": update}, nil)
 736 }
 737
 738 // IsCancelled returns the value of Cancelled, with goroutine safety.
 739 func (runner *ContainerRunner) IsCancelled() bool {
 740         runner.CancelLock.Lock()
 741         defer runner.CancelLock.Unlock()
 742         return runner.Cancelled
 743 }
 744
 745 // NewArvLogWriter creates an ArvLogWriter
 746 func (runner *ContainerRunner) NewArvLogWriter(name string) io.WriteCloser {
 747         return &ArvLogWriter{runner.ArvClient, runner.Container.UUID, name, runner.LogCollection.Open(name + ".txt")}
 748 }
 749
 750 // Run the full container lifecycle.
 751 func (runner *ContainerRunner) Run() (err error) {
 752         runner.CrunchLog.Printf("Executing container '%s'", runner.Container.UUID)
 753
 754         hostname, hosterr := os.Hostname()
 755         if hosterr != nil {
 756                 runner.CrunchLog.Printf("Error getting hostname '%v'", hosterr)
 757         } else {
 758                 runner.CrunchLog.Printf("Executing on host '%s'", hostname)
 759         }
 760
 761         // Clean up temporary directories _after_ finalizing
 762         // everything (if we've made any by then)
 763         defer runner.CleanupDirs()
 764
 765         runner.finalState = "Queued"
 766
 767         defer func() {
 768                 // checkErr prints e (unless it's nil) and sets err to
 769                 // e (unless err is already non-nil). Thus, if err
 770                 // hasn't already been assigned when Run() returns,
 771                 // this cleanup func will cause Run() to return the
 772                 // first non-nil error that is passed to checkErr().
 773                 checkErr := func(e error) {
 774                         if e == nil {
 775                                 return
 776                         }
 777                         runner.CrunchLog.Print(e)
 778                         if err == nil {
 779                                 err = e
 780                         }
 781                 }
 782
 783                 // Log the error encountered in Run(), if any
 784                 checkErr(err)
 785
 786                 if runner.finalState == "Queued" {
 787                         runner.UpdateContainerFinal()
 788                         return
 789                 }
 790
 791                 if runner.IsCancelled() {
 792                         runner.finalState = "Cancelled"
 793                         // but don't return yet -- we still want to
 794                         // capture partial output and write logs
 795                 }
 796
 797                 checkErr(runner.CaptureOutput())
 798                 checkErr(runner.CommitLogs())
 799                 checkErr(runner.UpdateContainerFinal())
 800
 801                 // The real log is already closed, but then we opened
 802                 // a new one in case we needed to log anything while
 803                 // finalizing.
 804                 runner.CrunchLog.Close()
 805         }()
 806
 807         err = runner.ArvClient.Get("containers", runner.Container.UUID, nil, &runner.Container)
 808         if err != nil {
 809                 err = fmt.Errorf("While getting container record: %v", err)
 810                 return
 811         }
 812
 813         // setup signal handling
 814         runner.SetupSignals()
 815
 816         // check for and/or load image
 817         err = runner.LoadImage()
 818         if err != nil {
 819                 err = fmt.Errorf("While loading container image: %v", err)
 820                 return
 821         }
 822
 823         // set up FUSE mount and binds
 824         err = runner.SetupMounts()
 825         if err != nil {
 826                 err = fmt.Errorf("While setting up mounts: %v", err)
 827                 return
 828         }
 829
 830         err = runner.CreateContainer()
 831         if err != nil {
 832                 return
 833         }
 834
 835         runner.StartCrunchstat()
 836
 837         if runner.IsCancelled() {
 838                 return
 839         }
 840
 841         err = runner.UpdateContainerRunning()
 842         if err != nil {
 843                 return
 844         }
 845         runner.finalState = "Cancelled"
 846
 847         err = runner.StartContainer()
 848         if err != nil {
 849                 return
 850         }
 851
 852         err = runner.WaitFinish()
 853         if err == nil {
 854                 runner.finalState = "Complete"
 855         }
 856         return
 857 }
 858
 859 // NewContainerRunner creates a new container runner.
 860 func NewContainerRunner(api IArvadosClient,
 861         kc IKeepClient,
 862         docker ThinDockerClient,
 863         containerUUID string) *ContainerRunner {
 864
 865         cr := &ContainerRunner{ArvClient: api, Kc: kc, Docker: docker}
 866         cr.NewLogWriter = cr.NewArvLogWriter
 867         cr.RunArvMount = cr.ArvMountCmd
 868         cr.MkTempDir = ioutil.TempDir
 869         cr.LogCollection = &CollectionWriter{kc, nil, sync.Mutex{}}
 870         cr.Container.UUID = containerUUID
 871         cr.CrunchLog = NewThrottledLogger(cr.NewLogWriter("crunch-run"))
 872         cr.CrunchLog.Immediate = log.New(os.Stderr, containerUUID+" ", 0)
 873         cr.loadDiscoveryVars()
 874         return cr
 875 }
 876
 877 func main() {
 878         statInterval := flag.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting")
 879         cgroupRoot := flag.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree")
 880         cgroupParent := flag.String("cgroup-parent", "docker", "name of container's parent cgroup (ignored if -cgroup-parent-subsystem is used)")
 881         cgroupParentSubsystem := flag.String("cgroup-parent-subsystem", "", "use current cgroup for given subsystem as parent cgroup for container")
 882         flag.Parse()
 883
 884         containerId := flag.Arg(0)
 885
 886         api, err := arvadosclient.MakeArvadosClient()
 887         if err != nil {
 888                 log.Fatalf("%s: %v", containerId, err)
 889         }
 890         api.Retries = 8
 891
 892         var kc *keepclient.KeepClient
 893         kc, err = keepclient.MakeKeepClient(api)
 894         if err != nil {
 895                 log.Fatalf("%s: %v", containerId, err)
 896         }
 897         kc.Retries = 4
 898
 899         var docker *dockerclient.DockerClient
 900         docker, err = dockerclient.NewDockerClient("unix:///var/run/docker.sock", nil)
 901         if err != nil {
 902                 log.Fatalf("%s: %v", containerId, err)
 903         }
 904
 905         cr := NewContainerRunner(api, kc, docker, containerId)
 906         cr.statInterval = *statInterval
 907         cr.cgroupRoot = *cgroupRoot
 908         cr.expectCgroupParent = *cgroupParent
 909         if *cgroupParentSubsystem != "" {
 910                 p := findCgroup(*cgroupParentSubsystem)
 911                 cr.setCgroupParent = p
 912                 cr.expectCgroupParent = p
 913         }
 914
 915         err = cr.Run()
 916         if err != nil {
 917                 log.Fatalf("%s: %v", containerId, err)
 918         }
 919
 920 }