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
 254         if runner.Container.RuntimeConstraints.KeepCacheRAM > 0 {
 255                 arvMountCmd = append(arvMountCmd, "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheRAM))
 256         }
 257
 258         collectionPaths := []string{}
 259         runner.Binds = nil
 260
 261         for bind, mnt := range runner.Container.Mounts {
 262                 if bind == "stdout" {
 263                         // Is it a "file" mount kind?
 264                         if mnt.Kind != "file" {
 265                                 return fmt.Errorf("Unsupported mount kind '%s' for stdout. Only 'file' is supported.", mnt.Kind)
 266                         }
 267
 268                         // Does path start with OutputPath?
 269                         prefix := runner.Container.OutputPath
 270                         if !strings.HasSuffix(prefix, "/") {
 271                                 prefix += "/"
 272                         }
 273                         if !strings.HasPrefix(mnt.Path, prefix) {
 274                                 return fmt.Errorf("Stdout path does not start with OutputPath: %s, %s", mnt.Path, prefix)
 275                         }
 276                 }
 277
 278                 switch {
 279                 case mnt.Kind == "collection":
 280                         var src string
 281                         if mnt.UUID != "" && mnt.PortableDataHash != "" {
 282                                 return fmt.Errorf("Cannot specify both 'uuid' and 'portable_data_hash' for a collection mount")
 283                         }
 284                         if mnt.UUID != "" {
 285                                 if mnt.Writable {
 286                                         return fmt.Errorf("Writing to existing collections currently not permitted.")
 287                                 }
 288                                 pdhOnly = false
 289                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.UUID)
 290                         } else if mnt.PortableDataHash != "" {
 291                                 if mnt.Writable {
 292                                         return fmt.Errorf("Can never write to a collection specified by portable data hash")
 293                                 }
 294                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.PortableDataHash)
 295                         } else {
 296                                 src = fmt.Sprintf("%s/tmp%d", runner.ArvMountPoint, tmpcount)
 297                                 arvMountCmd = append(arvMountCmd, "--mount-tmp")
 298                                 arvMountCmd = append(arvMountCmd, fmt.Sprintf("tmp%d", tmpcount))
 299                                 tmpcount += 1
 300                         }
 301                         if mnt.Writable {
 302                                 if bind == runner.Container.OutputPath {
 303                                         runner.HostOutputDir = src
 304                                 }
 305                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", src, bind))
 306                         } else {
 307                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", src, bind))
 308                         }
 309                         collectionPaths = append(collectionPaths, src)
 310
 311                 case mnt.Kind == "tmp" && bind == runner.Container.OutputPath:
 312                         runner.HostOutputDir, err = runner.MkTempDir("", "")
 313                         if err != nil {
 314                                 return fmt.Errorf("While creating mount temp dir: %v", err)
 315                         }
 316                         st, staterr := os.Stat(runner.HostOutputDir)
 317                         if staterr != nil {
 318                                 return fmt.Errorf("While Stat on temp dir: %v", staterr)
 319                         }
 320                         err = os.Chmod(runner.HostOutputDir, st.Mode()|os.ModeSetgid|0777)
 321                         if staterr != nil {
 322                                 return fmt.Errorf("While Chmod temp dir: %v", err)
 323                         }
 324                         runner.CleanupTempDir = append(runner.CleanupTempDir, runner.HostOutputDir)
 325                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", runner.HostOutputDir, bind))
 326
 327                 case mnt.Kind == "tmp":
 328                         runner.Binds = append(runner.Binds, bind)
 329
 330                 case mnt.Kind == "json":
 331                         jsondata, err := json.Marshal(mnt.Content)
 332                         if err != nil {
 333                                 return fmt.Errorf("encoding json data: %v", err)
 334                         }
 335                         // Create a tempdir with a single file
 336                         // (instead of just a tempfile): this way we
 337                         // can ensure the file is world-readable
 338                         // inside the container, without having to
 339                         // make it world-readable on the docker host.
 340                         tmpdir, err := runner.MkTempDir("", "")
 341                         if err != nil {
 342                                 return fmt.Errorf("creating temp dir: %v", err)
 343                         }
 344                         runner.CleanupTempDir = append(runner.CleanupTempDir, tmpdir)
 345                         tmpfn := filepath.Join(tmpdir, "mountdata.json")
 346                         err = ioutil.WriteFile(tmpfn, jsondata, 0644)
 347                         if err != nil {
 348                                 return fmt.Errorf("writing temp file: %v", err)
 349                         }
 350                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", tmpfn, bind))
 351                 }
 352         }
 353
 354         if runner.HostOutputDir == "" {
 355                 return fmt.Errorf("Output path does not correspond to a writable mount point")
 356         }
 357
 358         if pdhOnly {
 359                 arvMountCmd = append(arvMountCmd, "--mount-by-pdh", "by_id")
 360         } else {
 361                 arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_id")
 362         }
 363         arvMountCmd = append(arvMountCmd, runner.ArvMountPoint)
 364
 365         token, err := runner.ContainerToken()
 366         if err != nil {
 367                 return fmt.Errorf("could not get container token: %s", err)
 368         }
 369
 370         runner.ArvMount, err = runner.RunArvMount(arvMountCmd, token)
 371         if err != nil {
 372                 return fmt.Errorf("While trying to start arv-mount: %v", err)
 373         }
 374
 375         for _, p := range collectionPaths {
 376                 _, err = os.Stat(p)
 377                 if err != nil {
 378                         return fmt.Errorf("While checking that input files exist: %v", err)
 379                 }
 380         }
 381
 382         return nil
 383 }
 384
 385 func (runner *ContainerRunner) ProcessDockerAttach(containerReader io.Reader) {
 386         // Handle docker log protocol
 387         // https://docs.docker.com/engine/reference/api/docker_remote_api_v1.15/#attach-to-a-container
 388
 389         header := make([]byte, 8)
 390         for {
 391                 _, readerr := io.ReadAtLeast(containerReader, header, 8)
 392
 393                 if readerr == nil {
 394                         readsize := int64(header[7]) | (int64(header[6]) << 8) | (int64(header[5]) << 16) | (int64(header[4]) << 24)
 395                         if header[0] == 1 {
 396                                 // stdout
 397                                 _, readerr = io.CopyN(runner.Stdout, containerReader, readsize)
 398                         } else {
 399                                 // stderr
 400                                 _, readerr = io.CopyN(runner.Stderr, containerReader, readsize)
 401                         }
 402                 }
 403
 404                 if readerr != nil {
 405                         if readerr != io.EOF {
 406                                 runner.CrunchLog.Printf("While reading docker logs: %v", readerr)
 407                         }
 408
 409                         closeerr := runner.Stdout.Close()
 410                         if closeerr != nil {
 411                                 runner.CrunchLog.Printf("While closing stdout logs: %v", closeerr)
 412                         }
 413
 414                         closeerr = runner.Stderr.Close()
 415                         if closeerr != nil {
 416                                 runner.CrunchLog.Printf("While closing stderr logs: %v", closeerr)
 417                         }
 418
 419                         if runner.statReporter != nil {
 420                                 runner.statReporter.Stop()
 421                                 closeerr = runner.statLogger.Close()
 422                                 if closeerr != nil {
 423                                         runner.CrunchLog.Printf("While closing crunchstat logs: %v", closeerr)
 424                                 }
 425                         }
 426
 427                         runner.loggingDone <- true
 428                         close(runner.loggingDone)
 429                         return
 430                 }
 431         }
 432 }
 433
 434 func (runner *ContainerRunner) StartCrunchstat() {
 435         runner.statLogger = NewThrottledLogger(runner.NewLogWriter("crunchstat"))
 436         runner.statReporter = &crunchstat.Reporter{
 437                 CID:          runner.ContainerID,
 438                 Logger:       log.New(runner.statLogger, "", 0),
 439                 CgroupParent: runner.expectCgroupParent,
 440                 CgroupRoot:   runner.cgroupRoot,
 441                 PollPeriod:   runner.statInterval,
 442         }
 443         runner.statReporter.Start()
 444 }
 445
 446 // AttachLogs connects the docker container stdout and stderr logs to the
 447 // Arvados logger which logs to Keep and the API server logs table.
 448 func (runner *ContainerRunner) AttachStreams() (err error) {
 449
 450         runner.CrunchLog.Print("Attaching container streams")
 451
 452         var containerReader io.Reader
 453         containerReader, err = runner.Docker.AttachContainer(runner.ContainerID,
 454                 &dockerclient.AttachOptions{Stream: true, Stdout: true, Stderr: true})
 455         if err != nil {
 456                 return fmt.Errorf("While attaching container stdout/stderr streams: %v", err)
 457         }
 458
 459         runner.loggingDone = make(chan bool)
 460
 461         if stdoutMnt, ok := runner.Container.Mounts["stdout"]; ok {
 462                 stdoutPath := stdoutMnt.Path[len(runner.Container.OutputPath):]
 463                 index := strings.LastIndex(stdoutPath, "/")
 464                 if index > 0 {
 465                         subdirs := stdoutPath[:index]
 466                         if subdirs != "" {
 467                                 st, err := os.Stat(runner.HostOutputDir)
 468                                 if err != nil {
 469                                         return fmt.Errorf("While Stat on temp dir: %v", err)
 470                                 }
 471                                 stdoutPath := path.Join(runner.HostOutputDir, subdirs)
 472                                 err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777)
 473                                 if err != nil {
 474                                         return fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err)
 475                                 }
 476                         }
 477                 }
 478                 stdoutFile, err := os.Create(path.Join(runner.HostOutputDir, stdoutPath))
 479                 if err != nil {
 480                         return fmt.Errorf("While creating stdout file: %v", err)
 481                 }
 482                 runner.Stdout = stdoutFile
 483         } else {
 484                 runner.Stdout = NewThrottledLogger(runner.NewLogWriter("stdout"))
 485         }
 486         runner.Stderr = NewThrottledLogger(runner.NewLogWriter("stderr"))
 487
 488         go runner.ProcessDockerAttach(containerReader)
 489
 490         return nil
 491 }
 492
 493 // CreateContainer creates the docker container.
 494 func (runner *ContainerRunner) CreateContainer() error {
 495         runner.CrunchLog.Print("Creating Docker container")
 496
 497         runner.ContainerConfig.Cmd = runner.Container.Command
 498         if runner.Container.Cwd != "." {
 499                 runner.ContainerConfig.WorkingDir = runner.Container.Cwd
 500         }
 501
 502         for k, v := range runner.Container.Environment {
 503                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env, k+"="+v)
 504         }
 505         if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
 506                 tok, err := runner.ContainerToken()
 507                 if err != nil {
 508                         return err
 509                 }
 510                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env,
 511                         "ARVADOS_API_TOKEN="+tok,
 512                         "ARVADOS_API_HOST="+os.Getenv("ARVADOS_API_HOST"),
 513                         "ARVADOS_API_HOST_INSECURE="+os.Getenv("ARVADOS_API_HOST_INSECURE"),
 514                 )
 515                 runner.ContainerConfig.NetworkDisabled = false
 516         } else {
 517                 runner.ContainerConfig.NetworkDisabled = true
 518         }
 519
 520         var err error
 521         runner.ContainerID, err = runner.Docker.CreateContainer(&runner.ContainerConfig, "", nil)
 522         if err != nil {
 523                 return fmt.Errorf("While creating container: %v", err)
 524         }
 525
 526         runner.HostConfig = dockerclient.HostConfig{
 527                 Binds:        runner.Binds,
 528                 CgroupParent: runner.setCgroupParent,
 529                 LogConfig: dockerclient.LogConfig{
 530                         Type: "none",
 531                 },
 532         }
 533
 534         return runner.AttachStreams()
 535 }
 536
 537 // StartContainer starts the docker container created by CreateContainer.
 538 func (runner *ContainerRunner) StartContainer() error {
 539         runner.CrunchLog.Printf("Starting Docker container id '%s'", runner.ContainerID)
 540         err := runner.Docker.StartContainer(runner.ContainerID, &runner.HostConfig)
 541         if err != nil {
 542                 return fmt.Errorf("could not start container: %v", err)
 543         }
 544         return nil
 545 }
 546
 547 // WaitFinish waits for the container to terminate, capture the exit code, and
 548 // close the stdout/stderr logging.
 549 func (runner *ContainerRunner) WaitFinish() error {
 550         runner.CrunchLog.Print("Waiting for container to finish")
 551
 552         result := runner.Docker.Wait(runner.ContainerID)
 553         wr := <-result
 554         if wr.Error != nil {
 555                 return fmt.Errorf("While waiting for container to finish: %v", wr.Error)
 556         }
 557         runner.ExitCode = &wr.ExitCode
 558
 559         // wait for stdout/stderr to complete
 560         <-runner.loggingDone
 561
 562         return nil
 563 }
 564
 565 // HandleOutput sets the output, unmounts the FUSE mount, and deletes temporary directories
 566 func (runner *ContainerRunner) CaptureOutput() error {
 567         if runner.finalState != "Complete" {
 568                 return nil
 569         }
 570
 571         if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
 572                 // Output may have been set directly by the container, so
 573                 // refresh the container record to check.
 574                 err := runner.ArvClient.Get("containers", runner.Container.UUID,
 575                         nil, &runner.Container)
 576                 if err != nil {
 577                         return err
 578                 }
 579                 if runner.Container.Output != "" {
 580                         // Container output is already set.
 581                         runner.OutputPDH = &runner.Container.Output
 582                         return nil
 583                 }
 584         }
 585
 586         if runner.HostOutputDir == "" {
 587                 return nil
 588         }
 589
 590         _, err := os.Stat(runner.HostOutputDir)
 591         if err != nil {
 592                 return fmt.Errorf("While checking host output path: %v", err)
 593         }
 594
 595         var manifestText string
 596
 597         collectionMetafile := fmt.Sprintf("%s/.arvados#collection", runner.HostOutputDir)
 598         _, err = os.Stat(collectionMetafile)
 599         if err != nil {
 600                 // Regular directory
 601                 cw := CollectionWriter{runner.Kc, nil, sync.Mutex{}}
 602                 manifestText, err = cw.WriteTree(runner.HostOutputDir, runner.CrunchLog.Logger)
 603                 if err != nil {
 604                         return fmt.Errorf("While uploading output files: %v", err)
 605                 }
 606         } else {
 607                 // FUSE mount directory
 608                 file, openerr := os.Open(collectionMetafile)
 609                 if openerr != nil {
 610                         return fmt.Errorf("While opening FUSE metafile: %v", err)
 611                 }
 612                 defer file.Close()
 613
 614                 var rec arvados.Collection
 615                 err = json.NewDecoder(file).Decode(&rec)
 616                 if err != nil {
 617                         return fmt.Errorf("While reading FUSE metafile: %v", err)
 618                 }
 619                 manifestText = rec.ManifestText
 620         }
 621
 622         var response arvados.Collection
 623         err = runner.ArvClient.Create("collections",
 624                 arvadosclient.Dict{
 625                         "collection": arvadosclient.Dict{
 626                                 "expires_at":    time.Now().Add(runner.trashLifetime).Format(time.RFC3339),
 627                                 "name":          "output for " + runner.Container.UUID,
 628                                 "manifest_text": manifestText}},
 629                 &response)
 630         if err != nil {
 631                 return fmt.Errorf("While creating output collection: %v", err)
 632         }
 633         runner.OutputPDH = &response.PortableDataHash
 634         return nil
 635 }
 636
 637 func (runner *ContainerRunner) loadDiscoveryVars() {
 638         tl, err := runner.ArvClient.Discovery("defaultTrashLifetime")
 639         if err != nil {
 640                 log.Fatalf("getting defaultTrashLifetime from discovery document: %s", err)
 641         }
 642         runner.trashLifetime = time.Duration(tl.(float64)) * time.Second
 643 }
 644
 645 func (runner *ContainerRunner) CleanupDirs() {
 646         if runner.ArvMount != nil {
 647                 umount := exec.Command("fusermount", "-z", "-u", runner.ArvMountPoint)
 648                 umnterr := umount.Run()
 649                 if umnterr != nil {
 650                         runner.CrunchLog.Printf("While running fusermount: %v", umnterr)
 651                 }
 652
 653                 mnterr := <-runner.ArvMountExit
 654                 if mnterr != nil {
 655                         runner.CrunchLog.Printf("Arv-mount exit error: %v", mnterr)
 656                 }
 657         }
 658
 659         for _, tmpdir := range runner.CleanupTempDir {
 660                 rmerr := os.RemoveAll(tmpdir)
 661                 if rmerr != nil {
 662                         runner.CrunchLog.Printf("While cleaning up temporary directory %s: %v", tmpdir, rmerr)
 663                 }
 664         }
 665 }
 666
 667 // CommitLogs posts the collection containing the final container logs.
 668 func (runner *ContainerRunner) CommitLogs() error {
 669         runner.CrunchLog.Print(runner.finalState)
 670         runner.CrunchLog.Close()
 671
 672         // Closing CrunchLog above allows it to be committed to Keep at this
 673         // point, but re-open crunch log with ArvClient in case there are any
 674         // other further (such as failing to write the log to Keep!) while
 675         // shutting down
 676         runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{runner.ArvClient, runner.Container.UUID,
 677                 "crunch-run", nil})
 678
 679         if runner.LogsPDH != nil {
 680                 // If we have already assigned something to LogsPDH,
 681                 // we must be closing the re-opened log, which won't
 682                 // end up getting attached to the container record and
 683                 // therefore doesn't need to be saved as a collection
 684                 // -- it exists only to send logs to other channels.
 685                 return nil
 686         }
 687
 688         mt, err := runner.LogCollection.ManifestText()
 689         if err != nil {
 690                 return fmt.Errorf("While creating log manifest: %v", err)
 691         }
 692
 693         var response arvados.Collection
 694         err = runner.ArvClient.Create("collections",
 695                 arvadosclient.Dict{
 696                         "collection": arvadosclient.Dict{
 697                                 "expires_at":    time.Now().Add(runner.trashLifetime).Format(time.RFC3339),
 698                                 "name":          "logs for " + runner.Container.UUID,
 699                                 "manifest_text": mt}},
 700                 &response)
 701         if err != nil {
 702                 return fmt.Errorf("While creating log collection: %v", err)
 703         }
 704         runner.LogsPDH = &response.PortableDataHash
 705         return nil
 706 }
 707
 708 // UpdateContainerRunning updates the container state to "Running"
 709 func (runner *ContainerRunner) UpdateContainerRunning() error {
 710         runner.CancelLock.Lock()
 711         defer runner.CancelLock.Unlock()
 712         if runner.Cancelled {
 713                 return ErrCancelled
 714         }
 715         return runner.ArvClient.Update("containers", runner.Container.UUID,
 716                 arvadosclient.Dict{"container": arvadosclient.Dict{"state": "Running"}}, nil)
 717 }
 718
 719 // ContainerToken returns the api_token the container (and any
 720 // arv-mount processes) are allowed to use.
 721 func (runner *ContainerRunner) ContainerToken() (string, error) {
 722         if runner.token != "" {
 723                 return runner.token, nil
 724         }
 725
 726         var auth arvados.APIClientAuthorization
 727         err := runner.ArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth)
 728         if err != nil {
 729                 return "", err
 730         }
 731         runner.token = auth.APIToken
 732         return runner.token, nil
 733 }
 734
 735 // UpdateContainerComplete updates the container record state on API
 736 // server to "Complete" or "Cancelled"
 737 func (runner *ContainerRunner) UpdateContainerFinal() error {
 738         update := arvadosclient.Dict{}
 739         update["state"] = runner.finalState
 740         if runner.finalState == "Complete" {
 741                 if runner.LogsPDH != nil {
 742                         update["log"] = *runner.LogsPDH
 743                 }
 744                 if runner.ExitCode != nil {
 745                         update["exit_code"] = *runner.ExitCode
 746                 }
 747                 if runner.OutputPDH != nil {
 748                         update["output"] = *runner.OutputPDH
 749                 }
 750         }
 751         return runner.ArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{"container": update}, nil)
 752 }
 753
 754 // IsCancelled returns the value of Cancelled, with goroutine safety.
 755 func (runner *ContainerRunner) IsCancelled() bool {
 756         runner.CancelLock.Lock()
 757         defer runner.CancelLock.Unlock()
 758         return runner.Cancelled
 759 }
 760
 761 // NewArvLogWriter creates an ArvLogWriter
 762 func (runner *ContainerRunner) NewArvLogWriter(name string) io.WriteCloser {
 763         return &ArvLogWriter{runner.ArvClient, runner.Container.UUID, name, runner.LogCollection.Open(name + ".txt")}
 764 }
 765
 766 // Run the full container lifecycle.
 767 func (runner *ContainerRunner) Run() (err error) {
 768         runner.CrunchLog.Printf("Executing container '%s'", runner.Container.UUID)
 769
 770         hostname, hosterr := os.Hostname()
 771         if hosterr != nil {
 772                 runner.CrunchLog.Printf("Error getting hostname '%v'", hosterr)
 773         } else {
 774                 runner.CrunchLog.Printf("Executing on host '%s'", hostname)
 775         }
 776
 777         // Clean up temporary directories _after_ finalizing
 778         // everything (if we've made any by then)
 779         defer runner.CleanupDirs()
 780
 781         runner.finalState = "Queued"
 782
 783         defer func() {
 784                 // checkErr prints e (unless it's nil) and sets err to
 785                 // e (unless err is already non-nil). Thus, if err
 786                 // hasn't already been assigned when Run() returns,
 787                 // this cleanup func will cause Run() to return the
 788                 // first non-nil error that is passed to checkErr().
 789                 checkErr := func(e error) {
 790                         if e == nil {
 791                                 return
 792                         }
 793                         runner.CrunchLog.Print(e)
 794                         if err == nil {
 795                                 err = e
 796                         }
 797                 }
 798
 799                 // Log the error encountered in Run(), if any
 800                 checkErr(err)
 801
 802                 if runner.finalState == "Queued" {
 803                         runner.CrunchLog.Close()
 804                         runner.UpdateContainerFinal()
 805                         return
 806                 }
 807
 808                 if runner.IsCancelled() {
 809                         runner.finalState = "Cancelled"
 810                         // but don't return yet -- we still want to
 811                         // capture partial output and write logs
 812                 }
 813
 814                 checkErr(runner.CaptureOutput())
 815                 checkErr(runner.CommitLogs())
 816                 checkErr(runner.UpdateContainerFinal())
 817
 818                 // The real log is already closed, but then we opened
 819                 // a new one in case we needed to log anything while
 820                 // finalizing.
 821                 runner.CrunchLog.Close()
 822         }()
 823
 824         err = runner.ArvClient.Get("containers", runner.Container.UUID, nil, &runner.Container)
 825         if err != nil {
 826                 err = fmt.Errorf("While getting container record: %v", err)
 827                 return
 828         }
 829
 830         // setup signal handling
 831         runner.SetupSignals()
 832
 833         // check for and/or load image
 834         err = runner.LoadImage()
 835         if err != nil {
 836                 runner.finalState = "Cancelled"
 837                 err = fmt.Errorf("While loading container image: %v", err)
 838                 return
 839         }
 840
 841         // set up FUSE mount and binds
 842         err = runner.SetupMounts()
 843         if err != nil {
 844                 runner.finalState = "Cancelled"
 845                 err = fmt.Errorf("While setting up mounts: %v", err)
 846                 return
 847         }
 848
 849         err = runner.CreateContainer()
 850         if err != nil {
 851                 return
 852         }
 853
 854         runner.StartCrunchstat()
 855
 856         if runner.IsCancelled() {
 857                 return
 858         }
 859
 860         err = runner.UpdateContainerRunning()
 861         if err != nil {
 862                 return
 863         }
 864         runner.finalState = "Cancelled"
 865
 866         err = runner.StartContainer()
 867         if err != nil {
 868                 return
 869         }
 870
 871         err = runner.WaitFinish()
 872         if err == nil {
 873                 runner.finalState = "Complete"
 874         }
 875         return
 876 }
 877
 878 // NewContainerRunner creates a new container runner.
 879 func NewContainerRunner(api IArvadosClient,
 880         kc IKeepClient,
 881         docker ThinDockerClient,
 882         containerUUID string) *ContainerRunner {
 883
 884         cr := &ContainerRunner{ArvClient: api, Kc: kc, Docker: docker}
 885         cr.NewLogWriter = cr.NewArvLogWriter
 886         cr.RunArvMount = cr.ArvMountCmd
 887         cr.MkTempDir = ioutil.TempDir
 888         cr.LogCollection = &CollectionWriter{kc, nil, sync.Mutex{}}
 889         cr.Container.UUID = containerUUID
 890         cr.CrunchLog = NewThrottledLogger(cr.NewLogWriter("crunch-run"))
 891         cr.CrunchLog.Immediate = log.New(os.Stderr, containerUUID+" ", 0)
 892         cr.loadDiscoveryVars()
 893         return cr
 894 }
 895
 896 func main() {
 897         statInterval := flag.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting")
 898         cgroupRoot := flag.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree")
 899         cgroupParent := flag.String("cgroup-parent", "docker", "name of container's parent cgroup (ignored if -cgroup-parent-subsystem is used)")
 900         cgroupParentSubsystem := flag.String("cgroup-parent-subsystem", "", "use current cgroup for given subsystem as parent cgroup for container")
 901         flag.Parse()
 902
 903         containerId := flag.Arg(0)
 904
 905         api, err := arvadosclient.MakeArvadosClient()
 906         if err != nil {
 907                 log.Fatalf("%s: %v", containerId, err)
 908         }
 909         api.Retries = 8
 910
 911         var kc *keepclient.KeepClient
 912         kc, err = keepclient.MakeKeepClient(api)
 913         if err != nil {
 914                 log.Fatalf("%s: %v", containerId, err)
 915         }
 916         kc.Retries = 4
 917
 918         var docker *dockerclient.DockerClient
 919         docker, err = dockerclient.NewDockerClient("unix:///var/run/docker.sock", nil)
 920         if err != nil {
 921                 log.Fatalf("%s: %v", containerId, err)
 922         }
 923
 924         cr := NewContainerRunner(api, kc, docker, containerId)
 925         cr.statInterval = *statInterval
 926         cr.cgroupRoot = *cgroupRoot
 927         cr.expectCgroupParent = *cgroupParent
 928         if *cgroupParentSubsystem != "" {
 929                 p := findCgroup(*cgroupParentSubsystem)
 930                 cr.setCgroupParent = p
 931                 cr.expectCgroupParent = p
 932         }
 933
 934         err = cr.Run()
 935         if err != nil {
 936                 log.Fatalf("%s: %v", containerId, err)
 937         }
 938
 939 }