services/crunch-run/crunchrun.go

   1 // Copyright (C) The Arvados Authors. All rights reserved.
   2 //
   3 // SPDX-License-Identifier: AGPL-3.0
   4
   5 package main
   6
   7 import (
   8         "bytes"
   9         "encoding/json"
  10         "errors"
  11         "flag"
  12         "fmt"
  13         "io"
  14         "io/ioutil"
  15         "log"
  16         "os"
  17         "os/exec"
  18         "os/signal"
  19         "path"
  20         "path/filepath"
  21         "regexp"
  22         "runtime"
  23         "runtime/pprof"
  24         "sort"
  25         "strings"
  26         "sync"
  27         "syscall"
  28         "time"
  29
  30         "git.curoverse.com/arvados.git/lib/crunchstat"
  31         "git.curoverse.com/arvados.git/sdk/go/arvados"
  32         "git.curoverse.com/arvados.git/sdk/go/arvadosclient"
  33         "git.curoverse.com/arvados.git/sdk/go/keepclient"
  34         "git.curoverse.com/arvados.git/sdk/go/manifest"
  35         "golang.org/x/net/context"
  36
  37         dockertypes "github.com/docker/docker/api/types"
  38         dockercontainer "github.com/docker/docker/api/types/container"
  39         dockernetwork "github.com/docker/docker/api/types/network"
  40         dockerclient "github.com/docker/docker/client"
  41 )
  42
  43 var version = "dev"
  44
  45 // IArvadosClient is the minimal Arvados API methods used by crunch-run.
  46 type IArvadosClient interface {
  47         Create(resourceType string, parameters arvadosclient.Dict, output interface{}) error
  48         Get(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  49         Update(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  50         Call(method, resourceType, uuid, action string, parameters arvadosclient.Dict, output interface{}) error
  51         CallRaw(method string, resourceType string, uuid string, action string, parameters arvadosclient.Dict) (reader io.ReadCloser, err error)
  52         Discovery(key string) (interface{}, error)
  53 }
  54
  55 // ErrCancelled is the error returned when the container is cancelled.
  56 var ErrCancelled = errors.New("Cancelled")
  57
  58 // IKeepClient is the minimal Keep API methods used by crunch-run.
  59 type IKeepClient interface {
  60         PutB(buf []byte) (string, int, error)
  61         ReadAt(locator string, p []byte, off int) (int, error)
  62         ManifestFileReader(m manifest.Manifest, filename string) (arvados.File, error)
  63         LocalLocator(locator string) (string, error)
  64         ClearBlockCache()
  65 }
  66
  67 // NewLogWriter is a factory function to create a new log writer.
  68 type NewLogWriter func(name string) (io.WriteCloser, error)
  69
  70 type RunArvMount func(args []string, tok string) (*exec.Cmd, error)
  71
  72 type MkTempDir func(string, string) (string, error)
  73
  74 // ThinDockerClient is the minimal Docker client interface used by crunch-run.
  75 type ThinDockerClient interface {
  76         ContainerAttach(ctx context.Context, container string, options dockertypes.ContainerAttachOptions) (dockertypes.HijackedResponse, error)
  77         ContainerCreate(ctx context.Context, config *dockercontainer.Config, hostConfig *dockercontainer.HostConfig,
  78                 networkingConfig *dockernetwork.NetworkingConfig, containerName string) (dockercontainer.ContainerCreateCreatedBody, error)
  79         ContainerStart(ctx context.Context, container string, options dockertypes.ContainerStartOptions) error
  80         ContainerRemove(ctx context.Context, container string, options dockertypes.ContainerRemoveOptions) error
  81         ContainerWait(ctx context.Context, container string, condition dockercontainer.WaitCondition) (<-chan dockercontainer.ContainerWaitOKBody, <-chan error)
  82         ContainerInspect(ctx context.Context, id string) (dockertypes.ContainerJSON, error)
  83         ImageInspectWithRaw(ctx context.Context, image string) (dockertypes.ImageInspect, []byte, error)
  84         ImageLoad(ctx context.Context, input io.Reader, quiet bool) (dockertypes.ImageLoadResponse, error)
  85         ImageRemove(ctx context.Context, image string, options dockertypes.ImageRemoveOptions) ([]dockertypes.ImageDeleteResponseItem, error)
  86 }
  87
  88 type PsProcess interface {
  89         CmdlineSlice() ([]string, error)
  90 }
  91
  92 // ContainerRunner is the main stateful struct used for a single execution of a
  93 // container.
  94 type ContainerRunner struct {
  95         Docker ThinDockerClient
  96
  97         // Dispatcher client is initialized with the Dispatcher token.
  98         // This is a privileged token used to manage container status
  99         // and logs.
 100         //
 101         // We have both dispatcherClient and DispatcherArvClient
 102         // because there are two different incompatible Arvados Go
 103         // SDKs and we have to use both (hopefully this gets fixed in
 104         // #14467)
 105         dispatcherClient     *arvados.Client
 106         DispatcherArvClient  IArvadosClient
 107         DispatcherKeepClient IKeepClient
 108
 109         // Container client is initialized with the Container token
 110         // This token controls the permissions of the container, and
 111         // must be used for operations such as reading collections.
 112         //
 113         // Same comment as above applies to
 114         // containerClient/ContainerArvClient.
 115         containerClient     *arvados.Client
 116         ContainerArvClient  IArvadosClient
 117         ContainerKeepClient IKeepClient
 118
 119         // environment provided by arvados-dispatch-cloud
 120         dispatchEnv map[string]interface{}
 121
 122         Container       arvados.Container
 123         ContainerConfig dockercontainer.Config
 124         HostConfig      dockercontainer.HostConfig
 125         token           string
 126         ContainerID     string
 127         ExitCode        *int
 128         NewLogWriter    NewLogWriter
 129         loggingDone     chan bool
 130         CrunchLog       *ThrottledLogger
 131         Stdout          io.WriteCloser
 132         Stderr          io.WriteCloser
 133         logUUID         string
 134         logMtx          sync.Mutex
 135         LogCollection   arvados.CollectionFileSystem
 136         LogsPDH         *string
 137         RunArvMount     RunArvMount
 138         MkTempDir       MkTempDir
 139         ArvMount        *exec.Cmd
 140         ArvMountPoint   string
 141         HostOutputDir   string
 142         Binds           []string
 143         Volumes         map[string]struct{}
 144         OutputPDH       *string
 145         SigChan         chan os.Signal
 146         ArvMountExit    chan error
 147         SecretMounts    map[string]arvados.Mount
 148         MkArvClient     func(token string) (IArvadosClient, IKeepClient, *arvados.Client, error)
 149         finalState      string
 150         parentTemp      string
 151
 152         statLogger       io.WriteCloser
 153         statReporter     *crunchstat.Reporter
 154         hoststatLogger   io.WriteCloser
 155         hoststatReporter *crunchstat.Reporter
 156         statInterval     time.Duration
 157         cgroupRoot       string
 158         // What we expect the container's cgroup parent to be.
 159         expectCgroupParent string
 160         // What we tell docker to use as the container's cgroup
 161         // parent. Note: Ideally we would use the same field for both
 162         // expectCgroupParent and setCgroupParent, and just make it
 163         // default to "docker". However, when using docker < 1.10 with
 164         // systemd, specifying a non-empty cgroup parent (even the
 165         // default value "docker") hits a docker bug
 166         // (https://github.com/docker/docker/issues/17126). Using two
 167         // separate fields makes it possible to use the "expect cgroup
 168         // parent to be X" feature even on sites where the "specify
 169         // cgroup parent" feature breaks.
 170         setCgroupParent string
 171
 172         cStateLock sync.Mutex
 173         cCancelled bool // StopContainer() invoked
 174         cRemoved   bool // docker confirmed the container no longer exists
 175
 176         enableNetwork string // one of "default" or "always"
 177         networkMode   string // passed through to HostConfig.NetworkMode
 178         arvMountLog   *ThrottledLogger
 179
 180         containerWatchdogInterval time.Duration
 181 }
 182
 183 // setupSignals sets up signal handling to gracefully terminate the underlying
 184 // Docker container and update state when receiving a TERM, INT or QUIT signal.
 185 func (runner *ContainerRunner) setupSignals() {
 186         runner.SigChan = make(chan os.Signal, 1)
 187         signal.Notify(runner.SigChan, syscall.SIGTERM)
 188         signal.Notify(runner.SigChan, syscall.SIGINT)
 189         signal.Notify(runner.SigChan, syscall.SIGQUIT)
 190
 191         go func(sig chan os.Signal) {
 192                 for s := range sig {
 193                         runner.stop(s)
 194                 }
 195         }(runner.SigChan)
 196 }
 197
 198 // stop the underlying Docker container.
 199 func (runner *ContainerRunner) stop(sig os.Signal) {
 200         runner.cStateLock.Lock()
 201         defer runner.cStateLock.Unlock()
 202         if sig != nil {
 203                 runner.CrunchLog.Printf("caught signal: %v", sig)
 204         }
 205         if runner.ContainerID == "" {
 206                 return
 207         }
 208         runner.cCancelled = true
 209         runner.CrunchLog.Printf("removing container")
 210         err := runner.Docker.ContainerRemove(context.TODO(), runner.ContainerID, dockertypes.ContainerRemoveOptions{Force: true})
 211         if err != nil {
 212                 runner.CrunchLog.Printf("error removing container: %s", err)
 213         }
 214         if err == nil || strings.Contains(err.Error(), "No such container: "+runner.ContainerID) {
 215                 runner.cRemoved = true
 216         }
 217 }
 218
 219 var errorBlacklist = []string{
 220         "(?ms).*[Cc]annot connect to the Docker daemon.*",
 221         "(?ms).*oci runtime error.*starting container process.*container init.*mounting.*to rootfs.*no such file or directory.*",
 222         "(?ms).*grpc: the connection is unavailable.*",
 223 }
 224 var brokenNodeHook *string = flag.String("broken-node-hook", "", "Script to run if node is detected to be broken (for example, Docker daemon is not running)")
 225
 226 func (runner *ContainerRunner) runBrokenNodeHook() {
 227         if *brokenNodeHook == "" {
 228                 path := filepath.Join(lockdir, brokenfile)
 229                 runner.CrunchLog.Printf("Writing %s to mark node as broken", path)
 230                 f, err := os.OpenFile(path, os.O_CREATE|os.O_RDWR, 0700)
 231                 if err != nil {
 232                         runner.CrunchLog.Printf("Error writing %s: %s", path, err)
 233                         return
 234                 }
 235                 f.Close()
 236         } else {
 237                 runner.CrunchLog.Printf("Running broken node hook %q", *brokenNodeHook)
 238                 // run killme script
 239                 c := exec.Command(*brokenNodeHook)
 240                 c.Stdout = runner.CrunchLog
 241                 c.Stderr = runner.CrunchLog
 242                 err := c.Run()
 243                 if err != nil {
 244                         runner.CrunchLog.Printf("Error running broken node hook: %v", err)
 245                 }
 246         }
 247 }
 248
 249 func (runner *ContainerRunner) checkBrokenNode(goterr error) bool {
 250         for _, d := range errorBlacklist {
 251                 if m, e := regexp.MatchString(d, goterr.Error()); m && e == nil {
 252                         runner.CrunchLog.Printf("Error suggests node is unable to run containers: %v", goterr)
 253                         runner.runBrokenNodeHook()
 254                         return true
 255                 }
 256         }
 257         return false
 258 }
 259
 260 // LoadImage determines the docker image id from the container record and
 261 // checks if it is available in the local Docker image store.  If not, it loads
 262 // the image from Keep.
 263 func (runner *ContainerRunner) LoadImage() (err error) {
 264
 265         runner.CrunchLog.Printf("Fetching Docker image from collection '%s'", runner.Container.ContainerImage)
 266
 267         var collection arvados.Collection
 268         err = runner.ContainerArvClient.Get("collections", runner.Container.ContainerImage, nil, &collection)
 269         if err != nil {
 270                 return fmt.Errorf("While getting container image collection: %v", err)
 271         }
 272         manifest := manifest.Manifest{Text: collection.ManifestText}
 273         var img, imageID string
 274         for ms := range manifest.StreamIter() {
 275                 img = ms.FileStreamSegments[0].Name
 276                 if !strings.HasSuffix(img, ".tar") {
 277                         return fmt.Errorf("First file in the container image collection does not end in .tar")
 278                 }
 279                 imageID = img[:len(img)-4]
 280         }
 281
 282         runner.CrunchLog.Printf("Using Docker image id '%s'", imageID)
 283
 284         _, _, err = runner.Docker.ImageInspectWithRaw(context.TODO(), imageID)
 285         if err != nil {
 286                 runner.CrunchLog.Print("Loading Docker image from keep")
 287
 288                 var readCloser io.ReadCloser
 289                 readCloser, err = runner.ContainerKeepClient.ManifestFileReader(manifest, img)
 290                 if err != nil {
 291                         return fmt.Errorf("While creating ManifestFileReader for container image: %v", err)
 292                 }
 293
 294                 response, err := runner.Docker.ImageLoad(context.TODO(), readCloser, true)
 295                 if err != nil {
 296                         return fmt.Errorf("While loading container image into Docker: %v", err)
 297                 }
 298
 299                 defer response.Body.Close()
 300                 rbody, err := ioutil.ReadAll(response.Body)
 301                 if err != nil {
 302                         return fmt.Errorf("Reading response to image load: %v", err)
 303                 }
 304                 runner.CrunchLog.Printf("Docker response: %s", rbody)
 305         } else {
 306                 runner.CrunchLog.Print("Docker image is available")
 307         }
 308
 309         runner.ContainerConfig.Image = imageID
 310
 311         runner.ContainerKeepClient.ClearBlockCache()
 312
 313         return nil
 314 }
 315
 316 func (runner *ContainerRunner) ArvMountCmd(arvMountCmd []string, token string) (c *exec.Cmd, err error) {
 317         c = exec.Command("arv-mount", arvMountCmd...)
 318
 319         // Copy our environment, but override ARVADOS_API_TOKEN with
 320         // the container auth token.
 321         c.Env = nil
 322         for _, s := range os.Environ() {
 323                 if !strings.HasPrefix(s, "ARVADOS_API_TOKEN=") {
 324                         c.Env = append(c.Env, s)
 325                 }
 326         }
 327         c.Env = append(c.Env, "ARVADOS_API_TOKEN="+token)
 328
 329         w, err := runner.NewLogWriter("arv-mount")
 330         if err != nil {
 331                 return nil, err
 332         }
 333         runner.arvMountLog = NewThrottledLogger(w)
 334         c.Stdout = runner.arvMountLog
 335         c.Stderr = runner.arvMountLog
 336
 337         runner.CrunchLog.Printf("Running %v", c.Args)
 338
 339         err = c.Start()
 340         if err != nil {
 341                 return nil, err
 342         }
 343
 344         statReadme := make(chan bool)
 345         runner.ArvMountExit = make(chan error)
 346
 347         keepStatting := true
 348         go func() {
 349                 for keepStatting {
 350                         time.Sleep(100 * time.Millisecond)
 351                         _, err = os.Stat(fmt.Sprintf("%s/by_id/README", runner.ArvMountPoint))
 352                         if err == nil {
 353                                 keepStatting = false
 354                                 statReadme <- true
 355                         }
 356                 }
 357                 close(statReadme)
 358         }()
 359
 360         go func() {
 361                 mnterr := c.Wait()
 362                 if mnterr != nil {
 363                         runner.CrunchLog.Printf("Arv-mount exit error: %v", mnterr)
 364                 }
 365                 runner.ArvMountExit <- mnterr
 366                 close(runner.ArvMountExit)
 367         }()
 368
 369         select {
 370         case <-statReadme:
 371                 break
 372         case err := <-runner.ArvMountExit:
 373                 runner.ArvMount = nil
 374                 keepStatting = false
 375                 return nil, err
 376         }
 377
 378         return c, nil
 379 }
 380
 381 func (runner *ContainerRunner) SetupArvMountPoint(prefix string) (err error) {
 382         if runner.ArvMountPoint == "" {
 383                 runner.ArvMountPoint, err = runner.MkTempDir(runner.parentTemp, prefix)
 384         }
 385         return
 386 }
 387
 388 func copyfile(src string, dst string) (err error) {
 389         srcfile, err := os.Open(src)
 390         if err != nil {
 391                 return
 392         }
 393
 394         os.MkdirAll(path.Dir(dst), 0777)
 395
 396         dstfile, err := os.Create(dst)
 397         if err != nil {
 398                 return
 399         }
 400         _, err = io.Copy(dstfile, srcfile)
 401         if err != nil {
 402                 return
 403         }
 404
 405         err = srcfile.Close()
 406         err2 := dstfile.Close()
 407
 408         if err != nil {
 409                 return
 410         }
 411
 412         if err2 != nil {
 413                 return err2
 414         }
 415
 416         return nil
 417 }
 418
 419 func (runner *ContainerRunner) SetupMounts() (err error) {
 420         err = runner.SetupArvMountPoint("keep")
 421         if err != nil {
 422                 return fmt.Errorf("While creating keep mount temp dir: %v", err)
 423         }
 424
 425         token, err := runner.ContainerToken()
 426         if err != nil {
 427                 return fmt.Errorf("could not get container token: %s", err)
 428         }
 429
 430         pdhOnly := true
 431         tmpcount := 0
 432         arvMountCmd := []string{
 433                 "--foreground",
 434                 "--allow-other",
 435                 "--read-write",
 436                 fmt.Sprintf("--crunchstat-interval=%v", runner.statInterval.Seconds())}
 437
 438         if runner.Container.RuntimeConstraints.KeepCacheRAM > 0 {
 439                 arvMountCmd = append(arvMountCmd, "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheRAM))
 440         }
 441
 442         collectionPaths := []string{}
 443         runner.Binds = nil
 444         runner.Volumes = make(map[string]struct{})
 445         needCertMount := true
 446         type copyFile struct {
 447                 src  string
 448                 bind string
 449         }
 450         var copyFiles []copyFile
 451
 452         var binds []string
 453         for bind := range runner.Container.Mounts {
 454                 binds = append(binds, bind)
 455         }
 456         for bind := range runner.SecretMounts {
 457                 if _, ok := runner.Container.Mounts[bind]; ok {
 458                         return fmt.Errorf("Secret mount %q conflicts with regular mount", bind)
 459                 }
 460                 if runner.SecretMounts[bind].Kind != "json" &&
 461                         runner.SecretMounts[bind].Kind != "text" {
 462                         return fmt.Errorf("Secret mount %q type is %q but only 'json' and 'text' are permitted.",
 463                                 bind, runner.SecretMounts[bind].Kind)
 464                 }
 465                 binds = append(binds, bind)
 466         }
 467         sort.Strings(binds)
 468
 469         for _, bind := range binds {
 470                 mnt, ok := runner.Container.Mounts[bind]
 471                 if !ok {
 472                         mnt = runner.SecretMounts[bind]
 473                 }
 474                 if bind == "stdout" || bind == "stderr" {
 475                         // Is it a "file" mount kind?
 476                         if mnt.Kind != "file" {
 477                                 return fmt.Errorf("Unsupported mount kind '%s' for %s. Only 'file' is supported.", mnt.Kind, bind)
 478                         }
 479
 480                         // Does path start with OutputPath?
 481                         prefix := runner.Container.OutputPath
 482                         if !strings.HasSuffix(prefix, "/") {
 483                                 prefix += "/"
 484                         }
 485                         if !strings.HasPrefix(mnt.Path, prefix) {
 486                                 return fmt.Errorf("%s path does not start with OutputPath: %s, %s", strings.Title(bind), mnt.Path, prefix)
 487                         }
 488                 }
 489
 490                 if bind == "stdin" {
 491                         // Is it a "collection" mount kind?
 492                         if mnt.Kind != "collection" && mnt.Kind != "json" {
 493                                 return fmt.Errorf("Unsupported mount kind '%s' for stdin. Only 'collection' or 'json' are supported.", mnt.Kind)
 494                         }
 495                 }
 496
 497                 if bind == "/etc/arvados/ca-certificates.crt" {
 498                         needCertMount = false
 499                 }
 500
 501                 if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && bind != runner.Container.OutputPath+"/" {
 502                         if mnt.Kind != "collection" && mnt.Kind != "text" && mnt.Kind != "json" {
 503                                 return fmt.Errorf("Only mount points of kind 'collection', 'text' or 'json' are supported underneath the output_path for %q, was %q", bind, mnt.Kind)
 504                         }
 505                 }
 506
 507                 switch {
 508                 case mnt.Kind == "collection" && bind != "stdin":
 509                         var src string
 510                         if mnt.UUID != "" && mnt.PortableDataHash != "" {
 511                                 return fmt.Errorf("Cannot specify both 'uuid' and 'portable_data_hash' for a collection mount")
 512                         }
 513                         if mnt.UUID != "" {
 514                                 if mnt.Writable {
 515                                         return fmt.Errorf("Writing to existing collections currently not permitted.")
 516                                 }
 517                                 pdhOnly = false
 518                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.UUID)
 519                         } else if mnt.PortableDataHash != "" {
 520                                 if mnt.Writable && !strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 521                                         return fmt.Errorf("Can never write to a collection specified by portable data hash")
 522                                 }
 523                                 idx := strings.Index(mnt.PortableDataHash, "/")
 524                                 if idx > 0 {
 525                                         mnt.Path = path.Clean(mnt.PortableDataHash[idx:])
 526                                         mnt.PortableDataHash = mnt.PortableDataHash[0:idx]
 527                                         runner.Container.Mounts[bind] = mnt
 528                                 }
 529                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.PortableDataHash)
 530                                 if mnt.Path != "" && mnt.Path != "." {
 531                                         if strings.HasPrefix(mnt.Path, "./") {
 532                                                 mnt.Path = mnt.Path[2:]
 533                                         } else if strings.HasPrefix(mnt.Path, "/") {
 534                                                 mnt.Path = mnt.Path[1:]
 535                                         }
 536                                         src += "/" + mnt.Path
 537                                 }
 538                         } else {
 539                                 src = fmt.Sprintf("%s/tmp%d", runner.ArvMountPoint, tmpcount)
 540                                 arvMountCmd = append(arvMountCmd, "--mount-tmp")
 541                                 arvMountCmd = append(arvMountCmd, fmt.Sprintf("tmp%d", tmpcount))
 542                                 tmpcount += 1
 543                         }
 544                         if mnt.Writable {
 545                                 if bind == runner.Container.OutputPath {
 546                                         runner.HostOutputDir = src
 547                                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", src, bind))
 548                                 } else if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 549                                         copyFiles = append(copyFiles, copyFile{src, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
 550                                 } else {
 551                                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", src, bind))
 552                                 }
 553                         } else {
 554                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", src, bind))
 555                         }
 556                         collectionPaths = append(collectionPaths, src)
 557
 558                 case mnt.Kind == "tmp":
 559                         var tmpdir string
 560                         tmpdir, err = runner.MkTempDir(runner.parentTemp, "tmp")
 561                         if err != nil {
 562                                 return fmt.Errorf("While creating mount temp dir: %v", err)
 563                         }
 564                         st, staterr := os.Stat(tmpdir)
 565                         if staterr != nil {
 566                                 return fmt.Errorf("While Stat on temp dir: %v", staterr)
 567                         }
 568                         err = os.Chmod(tmpdir, st.Mode()|os.ModeSetgid|0777)
 569                         if staterr != nil {
 570                                 return fmt.Errorf("While Chmod temp dir: %v", err)
 571                         }
 572                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", tmpdir, bind))
 573                         if bind == runner.Container.OutputPath {
 574                                 runner.HostOutputDir = tmpdir
 575                         }
 576
 577                 case mnt.Kind == "json" || mnt.Kind == "text":
 578                         var filedata []byte
 579                         if mnt.Kind == "json" {
 580                                 filedata, err = json.Marshal(mnt.Content)
 581                                 if err != nil {
 582                                         return fmt.Errorf("encoding json data: %v", err)
 583                                 }
 584                         } else {
 585                                 text, ok := mnt.Content.(string)
 586                                 if !ok {
 587                                         return fmt.Errorf("content for mount %q must be a string", bind)
 588                                 }
 589                                 filedata = []byte(text)
 590                         }
 591
 592                         tmpdir, err := runner.MkTempDir(runner.parentTemp, mnt.Kind)
 593                         if err != nil {
 594                                 return fmt.Errorf("creating temp dir: %v", err)
 595                         }
 596                         tmpfn := filepath.Join(tmpdir, "mountdata."+mnt.Kind)
 597                         err = ioutil.WriteFile(tmpfn, filedata, 0444)
 598                         if err != nil {
 599                                 return fmt.Errorf("writing temp file: %v", err)
 600                         }
 601                         if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 602                                 copyFiles = append(copyFiles, copyFile{tmpfn, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
 603                         } else {
 604                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", tmpfn, bind))
 605                         }
 606
 607                 case mnt.Kind == "git_tree":
 608                         tmpdir, err := runner.MkTempDir(runner.parentTemp, "git_tree")
 609                         if err != nil {
 610                                 return fmt.Errorf("creating temp dir: %v", err)
 611                         }
 612                         err = gitMount(mnt).extractTree(runner.ContainerArvClient, tmpdir, token)
 613                         if err != nil {
 614                                 return err
 615                         }
 616                         runner.Binds = append(runner.Binds, tmpdir+":"+bind+":ro")
 617                 }
 618         }
 619
 620         if runner.HostOutputDir == "" {
 621                 return fmt.Errorf("Output path does not correspond to a writable mount point")
 622         }
 623
 624         if wantAPI := runner.Container.RuntimeConstraints.API; needCertMount && wantAPI != nil && *wantAPI {
 625                 for _, certfile := range arvadosclient.CertFiles {
 626                         _, err := os.Stat(certfile)
 627                         if err == nil {
 628                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:/etc/arvados/ca-certificates.crt:ro", certfile))
 629                                 break
 630                         }
 631                 }
 632         }
 633
 634         if pdhOnly {
 635                 arvMountCmd = append(arvMountCmd, "--mount-by-pdh", "by_id")
 636         } else {
 637                 arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_id")
 638         }
 639         arvMountCmd = append(arvMountCmd, runner.ArvMountPoint)
 640
 641         runner.ArvMount, err = runner.RunArvMount(arvMountCmd, token)
 642         if err != nil {
 643                 return fmt.Errorf("While trying to start arv-mount: %v", err)
 644         }
 645
 646         for _, p := range collectionPaths {
 647                 _, err = os.Stat(p)
 648                 if err != nil {
 649                         return fmt.Errorf("While checking that input files exist: %v", err)
 650                 }
 651         }
 652
 653         for _, cp := range copyFiles {
 654                 st, err := os.Stat(cp.src)
 655                 if err != nil {
 656                         return fmt.Errorf("While staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 657                 }
 658                 if st.IsDir() {
 659                         err = filepath.Walk(cp.src, func(walkpath string, walkinfo os.FileInfo, walkerr error) error {
 660                                 if walkerr != nil {
 661                                         return walkerr
 662                                 }
 663                                 target := path.Join(cp.bind, walkpath[len(cp.src):])
 664                                 if walkinfo.Mode().IsRegular() {
 665                                         copyerr := copyfile(walkpath, target)
 666                                         if copyerr != nil {
 667                                                 return copyerr
 668                                         }
 669                                         return os.Chmod(target, walkinfo.Mode()|0777)
 670                                 } else if walkinfo.Mode().IsDir() {
 671                                         mkerr := os.MkdirAll(target, 0777)
 672                                         if mkerr != nil {
 673                                                 return mkerr
 674                                         }
 675                                         return os.Chmod(target, walkinfo.Mode()|os.ModeSetgid|0777)
 676                                 } else {
 677                                         return fmt.Errorf("Source %q is not a regular file or directory", cp.src)
 678                                 }
 679                         })
 680                 } else if st.Mode().IsRegular() {
 681                         err = copyfile(cp.src, cp.bind)
 682                         if err == nil {
 683                                 err = os.Chmod(cp.bind, st.Mode()|0777)
 684                         }
 685                 }
 686                 if err != nil {
 687                         return fmt.Errorf("While staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 688                 }
 689         }
 690
 691         return nil
 692 }
 693
 694 func (runner *ContainerRunner) ProcessDockerAttach(containerReader io.Reader) {
 695         // Handle docker log protocol
 696         // https://docs.docker.com/engine/reference/api/docker_remote_api_v1.15/#attach-to-a-container
 697         defer close(runner.loggingDone)
 698
 699         header := make([]byte, 8)
 700         var err error
 701         for err == nil {
 702                 _, err = io.ReadAtLeast(containerReader, header, 8)
 703                 if err != nil {
 704                         if err == io.EOF {
 705                                 err = nil
 706                         }
 707                         break
 708                 }
 709                 readsize := int64(header[7]) | (int64(header[6]) << 8) | (int64(header[5]) << 16) | (int64(header[4]) << 24)
 710                 if header[0] == 1 {
 711                         // stdout
 712                         _, err = io.CopyN(runner.Stdout, containerReader, readsize)
 713                 } else {
 714                         // stderr
 715                         _, err = io.CopyN(runner.Stderr, containerReader, readsize)
 716                 }
 717         }
 718
 719         if err != nil {
 720                 runner.CrunchLog.Printf("error reading docker logs: %v", err)
 721         }
 722
 723         err = runner.Stdout.Close()
 724         if err != nil {
 725                 runner.CrunchLog.Printf("error closing stdout logs: %v", err)
 726         }
 727
 728         err = runner.Stderr.Close()
 729         if err != nil {
 730                 runner.CrunchLog.Printf("error closing stderr logs: %v", err)
 731         }
 732
 733         if runner.statReporter != nil {
 734                 runner.statReporter.Stop()
 735                 err = runner.statLogger.Close()
 736                 if err != nil {
 737                         runner.CrunchLog.Printf("error closing crunchstat logs: %v", err)
 738                 }
 739         }
 740 }
 741
 742 func (runner *ContainerRunner) stopHoststat() error {
 743         if runner.hoststatReporter == nil {
 744                 return nil
 745         }
 746         runner.hoststatReporter.Stop()
 747         err := runner.hoststatLogger.Close()
 748         if err != nil {
 749                 return fmt.Errorf("error closing hoststat logs: %v", err)
 750         }
 751         return nil
 752 }
 753
 754 func (runner *ContainerRunner) startHoststat() error {
 755         w, err := runner.NewLogWriter("hoststat")
 756         if err != nil {
 757                 return err
 758         }
 759         runner.hoststatLogger = NewThrottledLogger(w)
 760         runner.hoststatReporter = &crunchstat.Reporter{
 761                 Logger:     log.New(runner.hoststatLogger, "", 0),
 762                 CgroupRoot: runner.cgroupRoot,
 763                 PollPeriod: runner.statInterval,
 764         }
 765         runner.hoststatReporter.Start()
 766         return nil
 767 }
 768
 769 func (runner *ContainerRunner) startCrunchstat() error {
 770         w, err := runner.NewLogWriter("crunchstat")
 771         if err != nil {
 772                 return err
 773         }
 774         runner.statLogger = NewThrottledLogger(w)
 775         runner.statReporter = &crunchstat.Reporter{
 776                 CID:          runner.ContainerID,
 777                 Logger:       log.New(runner.statLogger, "", 0),
 778                 CgroupParent: runner.expectCgroupParent,
 779                 CgroupRoot:   runner.cgroupRoot,
 780                 PollPeriod:   runner.statInterval,
 781                 TempDir:      runner.parentTemp,
 782         }
 783         runner.statReporter.Start()
 784         return nil
 785 }
 786
 787 type infoCommand struct {
 788         label string
 789         cmd   []string
 790 }
 791
 792 // LogHostInfo logs info about the current host, for debugging and
 793 // accounting purposes. Although it's logged as "node-info", this is
 794 // about the environment where crunch-run is actually running, which
 795 // might differ from what's described in the node record (see
 796 // LogNodeRecord).
 797 func (runner *ContainerRunner) LogHostInfo() (err error) {
 798         w, err := runner.NewLogWriter("node-info")
 799         if err != nil {
 800                 return
 801         }
 802
 803         commands := []infoCommand{
 804                 {
 805                         label: "Host Information",
 806                         cmd:   []string{"uname", "-a"},
 807                 },
 808                 {
 809                         label: "CPU Information",
 810                         cmd:   []string{"cat", "/proc/cpuinfo"},
 811                 },
 812                 {
 813                         label: "Memory Information",
 814                         cmd:   []string{"cat", "/proc/meminfo"},
 815                 },
 816                 {
 817                         label: "Disk Space",
 818                         cmd:   []string{"df", "-m", "/", os.TempDir()},
 819                 },
 820                 {
 821                         label: "Disk INodes",
 822                         cmd:   []string{"df", "-i", "/", os.TempDir()},
 823                 },
 824         }
 825
 826         // Run commands with informational output to be logged.
 827         for _, command := range commands {
 828                 fmt.Fprintln(w, command.label)
 829                 cmd := exec.Command(command.cmd[0], command.cmd[1:]...)
 830                 cmd.Stdout = w
 831                 cmd.Stderr = w
 832                 if err := cmd.Run(); err != nil {
 833                         err = fmt.Errorf("While running command %q: %v", command.cmd, err)
 834                         fmt.Fprintln(w, err)
 835                         return err
 836                 }
 837                 fmt.Fprintln(w, "")
 838         }
 839
 840         err = w.Close()
 841         if err != nil {
 842                 return fmt.Errorf("While closing node-info logs: %v", err)
 843         }
 844         return nil
 845 }
 846
 847 // LogContainerRecord gets and saves the raw JSON container record from the API server
 848 func (runner *ContainerRunner) LogContainerRecord() error {
 849         logged, err := runner.logAPIResponse("container", "containers", map[string]interface{}{"filters": [][]string{{"uuid", "=", runner.Container.UUID}}}, nil)
 850         if !logged && err == nil {
 851                 err = fmt.Errorf("error: no container record found for %s", runner.Container.UUID)
 852         }
 853         return err
 854 }
 855
 856 // LogNodeRecord logs the current host's InstanceType config entry (or
 857 // the arvados#node record, if running via crunch-dispatch-slurm).
 858 func (runner *ContainerRunner) LogNodeRecord() error {
 859         if it, ok := runner.dispatchEnv["InstanceType"]; ok {
 860                 // Dispatched via arvados-dispatch-cloud. Save
 861                 // InstanceType config fragment received from
 862                 // dispatcher on stdin.
 863                 w, err := runner.LogCollection.OpenFile("node.json", os.O_CREATE|os.O_WRONLY, 0666)
 864                 if err != nil {
 865                         return err
 866                 }
 867                 defer w.Close()
 868                 if it, ok := it.(string); ok {
 869                         // dispatcher supplied JSON data (in order to
 870                         // stay compatible with old crunch-run
 871                         // versions)
 872                         _, err = io.WriteString(w, it)
 873                         if err != nil {
 874                                 return err
 875                         }
 876                 } else {
 877                         // dispatcher supplied struct
 878                         enc := json.NewEncoder(w)
 879                         enc.SetIndent("", "    ")
 880                         err = enc.Encode(it)
 881                         if err != nil {
 882                                 return err
 883                         }
 884                 }
 885                 return w.Close()
 886         } else {
 887                 // Dispatched via crunch-dispatch-slurm. Look up
 888                 // apiserver's node record corresponding to
 889                 // $SLURMD_NODENAME.
 890                 hostname := os.Getenv("SLURMD_NODENAME")
 891                 if hostname == "" {
 892                         hostname, _ = os.Hostname()
 893                 }
 894                 _, err := runner.logAPIResponse("node", "nodes", map[string]interface{}{"filters": [][]string{{"hostname", "=", hostname}}}, func(resp interface{}) {
 895                         // The "info" field has admin-only info when
 896                         // obtained with a privileged token, and
 897                         // should not be logged.
 898                         node, ok := resp.(map[string]interface{})
 899                         if ok {
 900                                 delete(node, "info")
 901                         }
 902                 })
 903                 return err
 904         }
 905 }
 906
 907 func (runner *ContainerRunner) logAPIResponse(label, path string, params map[string]interface{}, munge func(interface{})) (logged bool, err error) {
 908         writer, err := runner.LogCollection.OpenFile(label+".json", os.O_CREATE|os.O_WRONLY, 0666)
 909         if err != nil {
 910                 return false, err
 911         }
 912         w := &ArvLogWriter{
 913                 ArvClient:     runner.DispatcherArvClient,
 914                 UUID:          runner.Container.UUID,
 915                 loggingStream: label,
 916                 writeCloser:   writer,
 917         }
 918
 919         reader, err := runner.DispatcherArvClient.CallRaw("GET", path, "", "", arvadosclient.Dict(params))
 920         if err != nil {
 921                 return false, fmt.Errorf("error getting %s record: %v", label, err)
 922         }
 923         defer reader.Close()
 924
 925         dec := json.NewDecoder(reader)
 926         dec.UseNumber()
 927         var resp map[string]interface{}
 928         if err = dec.Decode(&resp); err != nil {
 929                 return false, fmt.Errorf("error decoding %s list response: %v", label, err)
 930         }
 931         items, ok := resp["items"].([]interface{})
 932         if !ok {
 933                 return false, fmt.Errorf("error decoding %s list response: no \"items\" key in API list response", label)
 934         } else if len(items) < 1 {
 935                 return false, nil
 936         }
 937         if munge != nil {
 938                 munge(items[0])
 939         }
 940         // Re-encode it using indentation to improve readability
 941         enc := json.NewEncoder(w)
 942         enc.SetIndent("", "    ")
 943         if err = enc.Encode(items[0]); err != nil {
 944                 return false, fmt.Errorf("error logging %s record: %v", label, err)
 945         }
 946         err = w.Close()
 947         if err != nil {
 948                 return false, fmt.Errorf("error closing %s.json in log collection: %v", label, err)
 949         }
 950         return true, nil
 951 }
 952
 953 // AttachStreams connects the docker container stdin, stdout and stderr logs
 954 // to the Arvados logger which logs to Keep and the API server logs table.
 955 func (runner *ContainerRunner) AttachStreams() (err error) {
 956
 957         runner.CrunchLog.Print("Attaching container streams")
 958
 959         // If stdin mount is provided, attach it to the docker container
 960         var stdinRdr arvados.File
 961         var stdinJson []byte
 962         if stdinMnt, ok := runner.Container.Mounts["stdin"]; ok {
 963                 if stdinMnt.Kind == "collection" {
 964                         var stdinColl arvados.Collection
 965                         collId := stdinMnt.UUID
 966                         if collId == "" {
 967                                 collId = stdinMnt.PortableDataHash
 968                         }
 969                         err = runner.ContainerArvClient.Get("collections", collId, nil, &stdinColl)
 970                         if err != nil {
 971                                 return fmt.Errorf("While getting stdin collection: %v", err)
 972                         }
 973
 974                         stdinRdr, err = runner.ContainerKeepClient.ManifestFileReader(
 975                                 manifest.Manifest{Text: stdinColl.ManifestText},
 976                                 stdinMnt.Path)
 977                         if os.IsNotExist(err) {
 978                                 return fmt.Errorf("stdin collection path not found: %v", stdinMnt.Path)
 979                         } else if err != nil {
 980                                 return fmt.Errorf("While getting stdin collection path %v: %v", stdinMnt.Path, err)
 981                         }
 982                 } else if stdinMnt.Kind == "json" {
 983                         stdinJson, err = json.Marshal(stdinMnt.Content)
 984                         if err != nil {
 985                                 return fmt.Errorf("While encoding stdin json data: %v", err)
 986                         }
 987                 }
 988         }
 989
 990         stdinUsed := stdinRdr != nil || len(stdinJson) != 0
 991         response, err := runner.Docker.ContainerAttach(context.TODO(), runner.ContainerID,
 992                 dockertypes.ContainerAttachOptions{Stream: true, Stdin: stdinUsed, Stdout: true, Stderr: true})
 993         if err != nil {
 994                 return fmt.Errorf("While attaching container stdout/stderr streams: %v", err)
 995         }
 996
 997         runner.loggingDone = make(chan bool)
 998
 999         if stdoutMnt, ok := runner.Container.Mounts["stdout"]; ok {
1000                 stdoutFile, err := runner.getStdoutFile(stdoutMnt.Path)
1001                 if err != nil {
1002                         return err
1003                 }
1004                 runner.Stdout = stdoutFile
1005         } else if w, err := runner.NewLogWriter("stdout"); err != nil {
1006                 return err
1007         } else {
1008                 runner.Stdout = NewThrottledLogger(w)
1009         }
1010
1011         if stderrMnt, ok := runner.Container.Mounts["stderr"]; ok {
1012                 stderrFile, err := runner.getStdoutFile(stderrMnt.Path)
1013                 if err != nil {
1014                         return err
1015                 }
1016                 runner.Stderr = stderrFile
1017         } else if w, err := runner.NewLogWriter("stderr"); err != nil {
1018                 return err
1019         } else {
1020                 runner.Stderr = NewThrottledLogger(w)
1021         }
1022
1023         if stdinRdr != nil {
1024                 go func() {
1025                         _, err := io.Copy(response.Conn, stdinRdr)
1026                         if err != nil {
1027                                 runner.CrunchLog.Printf("While writing stdin collection to docker container: %v", err)
1028                                 runner.stop(nil)
1029                         }
1030                         stdinRdr.Close()
1031                         response.CloseWrite()
1032                 }()
1033         } else if len(stdinJson) != 0 {
1034                 go func() {
1035                         _, err := io.Copy(response.Conn, bytes.NewReader(stdinJson))
1036                         if err != nil {
1037                                 runner.CrunchLog.Printf("While writing stdin json to docker container: %v", err)
1038                                 runner.stop(nil)
1039                         }
1040                         response.CloseWrite()
1041                 }()
1042         }
1043
1044         go runner.ProcessDockerAttach(response.Reader)
1045
1046         return nil
1047 }
1048
1049 func (runner *ContainerRunner) getStdoutFile(mntPath string) (*os.File, error) {
1050         stdoutPath := mntPath[len(runner.Container.OutputPath):]
1051         index := strings.LastIndex(stdoutPath, "/")
1052         if index > 0 {
1053                 subdirs := stdoutPath[:index]
1054                 if subdirs != "" {
1055                         st, err := os.Stat(runner.HostOutputDir)
1056                         if err != nil {
1057                                 return nil, fmt.Errorf("While Stat on temp dir: %v", err)
1058                         }
1059                         stdoutPath := filepath.Join(runner.HostOutputDir, subdirs)
1060                         err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777)
1061                         if err != nil {
1062                                 return nil, fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err)
1063                         }
1064                 }
1065         }
1066         stdoutFile, err := os.Create(filepath.Join(runner.HostOutputDir, stdoutPath))
1067         if err != nil {
1068                 return nil, fmt.Errorf("While creating file %q: %v", stdoutPath, err)
1069         }
1070
1071         return stdoutFile, nil
1072 }
1073
1074 // CreateContainer creates the docker container.
1075 func (runner *ContainerRunner) CreateContainer() error {
1076         runner.CrunchLog.Print("Creating Docker container")
1077
1078         runner.ContainerConfig.Cmd = runner.Container.Command
1079         if runner.Container.Cwd != "." {
1080                 runner.ContainerConfig.WorkingDir = runner.Container.Cwd
1081         }
1082
1083         for k, v := range runner.Container.Environment {
1084                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env, k+"="+v)
1085         }
1086
1087         runner.ContainerConfig.Volumes = runner.Volumes
1088
1089         maxRAM := int64(runner.Container.RuntimeConstraints.RAM)
1090         if maxRAM < 4*1024*1024 {
1091                 // Docker daemon won't let you set a limit less than 4 MiB
1092                 maxRAM = 4 * 1024 * 1024
1093         }
1094         runner.HostConfig = dockercontainer.HostConfig{
1095                 Binds: runner.Binds,
1096                 LogConfig: dockercontainer.LogConfig{
1097                         Type: "none",
1098                 },
1099                 Resources: dockercontainer.Resources{
1100                         CgroupParent: runner.setCgroupParent,
1101                         NanoCPUs:     int64(runner.Container.RuntimeConstraints.VCPUs) * 1000000000,
1102                         Memory:       maxRAM, // RAM
1103                         MemorySwap:   maxRAM, // RAM+swap
1104                         KernelMemory: maxRAM, // kernel portion
1105                 },
1106         }
1107
1108         if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
1109                 tok, err := runner.ContainerToken()
1110                 if err != nil {
1111                         return err
1112                 }
1113                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env,
1114                         "ARVADOS_API_TOKEN="+tok,
1115                         "ARVADOS_API_HOST="+os.Getenv("ARVADOS_API_HOST"),
1116                         "ARVADOS_API_HOST_INSECURE="+os.Getenv("ARVADOS_API_HOST_INSECURE"),
1117                 )
1118                 runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
1119         } else {
1120                 if runner.enableNetwork == "always" {
1121                         runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
1122                 } else {
1123                         runner.HostConfig.NetworkMode = dockercontainer.NetworkMode("none")
1124                 }
1125         }
1126
1127         _, stdinUsed := runner.Container.Mounts["stdin"]
1128         runner.ContainerConfig.OpenStdin = stdinUsed
1129         runner.ContainerConfig.StdinOnce = stdinUsed
1130         runner.ContainerConfig.AttachStdin = stdinUsed
1131         runner.ContainerConfig.AttachStdout = true
1132         runner.ContainerConfig.AttachStderr = true
1133
1134         createdBody, err := runner.Docker.ContainerCreate(context.TODO(), &runner.ContainerConfig, &runner.HostConfig, nil, runner.Container.UUID)
1135         if err != nil {
1136                 return fmt.Errorf("While creating container: %v", err)
1137         }
1138
1139         runner.ContainerID = createdBody.ID
1140
1141         return runner.AttachStreams()
1142 }
1143
1144 // StartContainer starts the docker container created by CreateContainer.
1145 func (runner *ContainerRunner) StartContainer() error {
1146         runner.CrunchLog.Printf("Starting Docker container id '%s'", runner.ContainerID)
1147         runner.cStateLock.Lock()
1148         defer runner.cStateLock.Unlock()
1149         if runner.cCancelled {
1150                 return ErrCancelled
1151         }
1152         err := runner.Docker.ContainerStart(context.TODO(), runner.ContainerID,
1153                 dockertypes.ContainerStartOptions{})
1154         if err != nil {
1155                 var advice string
1156                 if m, e := regexp.MatchString("(?ms).*(exec|System error).*(no such file or directory|file not found).*", err.Error()); m && e == nil {
1157                         advice = fmt.Sprintf("\nPossible causes: command %q is missing, the interpreter given in #! is missing, or script has Windows line endings.", runner.Container.Command[0])
1158                 }
1159                 return fmt.Errorf("could not start container: %v%s", err, advice)
1160         }
1161         return nil
1162 }
1163
1164 // WaitFinish waits for the container to terminate, capture the exit code, and
1165 // close the stdout/stderr logging.
1166 func (runner *ContainerRunner) WaitFinish() error {
1167         var runTimeExceeded <-chan time.Time
1168         runner.CrunchLog.Print("Waiting for container to finish")
1169
1170         waitOk, waitErr := runner.Docker.ContainerWait(context.TODO(), runner.ContainerID, dockercontainer.WaitConditionNotRunning)
1171         arvMountExit := runner.ArvMountExit
1172         if timeout := runner.Container.SchedulingParameters.MaxRunTime; timeout > 0 {
1173                 runTimeExceeded = time.After(time.Duration(timeout) * time.Second)
1174         }
1175
1176         containerGone := make(chan struct{})
1177         go func() {
1178                 defer close(containerGone)
1179                 if runner.containerWatchdogInterval < 1 {
1180                         runner.containerWatchdogInterval = time.Minute
1181                 }
1182                 for range time.NewTicker(runner.containerWatchdogInterval).C {
1183                         ctx, cancel := context.WithDeadline(context.Background(), time.Now().Add(runner.containerWatchdogInterval))
1184                         ctr, err := runner.Docker.ContainerInspect(ctx, runner.ContainerID)
1185                         cancel()
1186                         runner.cStateLock.Lock()
1187                         done := runner.cRemoved || runner.ExitCode != nil
1188                         runner.cStateLock.Unlock()
1189                         if done {
1190                                 return
1191                         } else if err != nil {
1192                                 runner.CrunchLog.Printf("Error inspecting container: %s", err)
1193                                 runner.checkBrokenNode(err)
1194                                 return
1195                         } else if ctr.State == nil || !(ctr.State.Running || ctr.State.Status == "created") {
1196                                 runner.CrunchLog.Printf("Container is not running: State=%v", ctr.State)
1197                                 return
1198                         }
1199                 }
1200         }()
1201
1202         for {
1203                 select {
1204                 case waitBody := <-waitOk:
1205                         runner.CrunchLog.Printf("Container exited with code: %v", waitBody.StatusCode)
1206                         code := int(waitBody.StatusCode)
1207                         runner.ExitCode = &code
1208
1209                         // wait for stdout/stderr to complete
1210                         <-runner.loggingDone
1211                         return nil
1212
1213                 case err := <-waitErr:
1214                         return fmt.Errorf("container wait: %v", err)
1215
1216                 case <-arvMountExit:
1217                         runner.CrunchLog.Printf("arv-mount exited while container is still running.  Stopping container.")
1218                         runner.stop(nil)
1219                         // arvMountExit will always be ready now that
1220                         // it's closed, but that doesn't interest us.
1221                         arvMountExit = nil
1222
1223                 case <-runTimeExceeded:
1224                         runner.CrunchLog.Printf("maximum run time exceeded. Stopping container.")
1225                         runner.stop(nil)
1226                         runTimeExceeded = nil
1227
1228                 case <-containerGone:
1229                         return errors.New("docker client never returned status")
1230                 }
1231         }
1232 }
1233
1234 func (runner *ContainerRunner) updateLogs() {
1235         ticker := time.NewTicker(crunchLogUpdatePeriod / 360)
1236         defer ticker.Stop()
1237
1238         sigusr1 := make(chan os.Signal, 1)
1239         signal.Notify(sigusr1, syscall.SIGUSR1)
1240         defer signal.Stop(sigusr1)
1241
1242         saveAtTime := time.Now().Add(crunchLogUpdatePeriod)
1243         saveAtSize := crunchLogUpdateSize
1244         var savedSize int64
1245         for {
1246                 select {
1247                 case <-ticker.C:
1248                 case <-sigusr1:
1249                         saveAtTime = time.Now()
1250                 }
1251                 runner.logMtx.Lock()
1252                 done := runner.LogsPDH != nil
1253                 runner.logMtx.Unlock()
1254                 if done {
1255                         return
1256                 }
1257                 size := runner.LogCollection.Size()
1258                 if size == savedSize || (time.Now().Before(saveAtTime) && size < saveAtSize) {
1259                         continue
1260                 }
1261                 saveAtTime = time.Now().Add(crunchLogUpdatePeriod)
1262                 saveAtSize = runner.LogCollection.Size() + crunchLogUpdateSize
1263                 saved, err := runner.saveLogCollection(false)
1264                 if err != nil {
1265                         runner.CrunchLog.Printf("error updating log collection: %s", err)
1266                         continue
1267                 }
1268
1269                 var updated arvados.Container
1270                 err = runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1271                         "container": arvadosclient.Dict{"log": saved.PortableDataHash},
1272                 }, &updated)
1273                 if err != nil {
1274                         runner.CrunchLog.Printf("error updating container log to %s: %s", saved.PortableDataHash, err)
1275                         continue
1276                 }
1277
1278                 savedSize = size
1279         }
1280 }
1281
1282 // CaptureOutput saves data from the container's output directory if
1283 // needed, and updates the container output accordingly.
1284 func (runner *ContainerRunner) CaptureOutput() error {
1285         if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
1286                 // Output may have been set directly by the container, so
1287                 // refresh the container record to check.
1288                 err := runner.DispatcherArvClient.Get("containers", runner.Container.UUID,
1289                         nil, &runner.Container)
1290                 if err != nil {
1291                         return err
1292                 }
1293                 if runner.Container.Output != "" {
1294                         // Container output is already set.
1295                         runner.OutputPDH = &runner.Container.Output
1296                         return nil
1297                 }
1298         }
1299
1300         txt, err := (&copier{
1301                 client:        runner.containerClient,
1302                 arvClient:     runner.ContainerArvClient,
1303                 keepClient:    runner.ContainerKeepClient,
1304                 hostOutputDir: runner.HostOutputDir,
1305                 ctrOutputDir:  runner.Container.OutputPath,
1306                 binds:         runner.Binds,
1307                 mounts:        runner.Container.Mounts,
1308                 secretMounts:  runner.SecretMounts,
1309                 logger:        runner.CrunchLog,
1310         }).Copy()
1311         if err != nil {
1312                 return err
1313         }
1314         if n := len(regexp.MustCompile(` [0-9a-f]+\+\S*\+R`).FindAllStringIndex(txt, -1)); n > 0 {
1315                 runner.CrunchLog.Printf("Copying %d data blocks from remote input collections...", n)
1316                 fs, err := (&arvados.Collection{ManifestText: txt}).FileSystem(runner.containerClient, runner.ContainerKeepClient)
1317                 if err != nil {
1318                         return err
1319                 }
1320                 txt, err = fs.MarshalManifest(".")
1321                 if err != nil {
1322                         return err
1323                 }
1324         }
1325         var resp arvados.Collection
1326         err = runner.ContainerArvClient.Create("collections", arvadosclient.Dict{
1327                 "ensure_unique_name": true,
1328                 "collection": arvadosclient.Dict{
1329                         "is_trashed":    true,
1330                         "name":          "output for " + runner.Container.UUID,
1331                         "manifest_text": txt,
1332                 },
1333         }, &resp)
1334         if err != nil {
1335                 return fmt.Errorf("error creating output collection: %v", err)
1336         }
1337         runner.OutputPDH = &resp.PortableDataHash
1338         return nil
1339 }
1340
1341 func (runner *ContainerRunner) CleanupDirs() {
1342         if runner.ArvMount != nil {
1343                 var delay int64 = 8
1344                 umount := exec.Command("arv-mount", fmt.Sprintf("--unmount-timeout=%d", delay), "--unmount", runner.ArvMountPoint)
1345                 umount.Stdout = runner.CrunchLog
1346                 umount.Stderr = runner.CrunchLog
1347                 runner.CrunchLog.Printf("Running %v", umount.Args)
1348                 umnterr := umount.Start()
1349
1350                 if umnterr != nil {
1351                         runner.CrunchLog.Printf("Error unmounting: %v", umnterr)
1352                 } else {
1353                         // If arv-mount --unmount gets stuck for any reason, we
1354                         // don't want to wait for it forever.  Do Wait() in a goroutine
1355                         // so it doesn't block crunch-run.
1356                         umountExit := make(chan error)
1357                         go func() {
1358                                 mnterr := umount.Wait()
1359                                 if mnterr != nil {
1360                                         runner.CrunchLog.Printf("Error unmounting: %v", mnterr)
1361                                 }
1362                                 umountExit <- mnterr
1363                         }()
1364
1365                         for again := true; again; {
1366                                 again = false
1367                                 select {
1368                                 case <-umountExit:
1369                                         umount = nil
1370                                         again = true
1371                                 case <-runner.ArvMountExit:
1372                                         break
1373                                 case <-time.After(time.Duration((delay + 1) * int64(time.Second))):
1374                                         runner.CrunchLog.Printf("Timed out waiting for unmount")
1375                                         if umount != nil {
1376                                                 umount.Process.Kill()
1377                                         }
1378                                         runner.ArvMount.Process.Kill()
1379                                 }
1380                         }
1381                 }
1382         }
1383
1384         if runner.ArvMountPoint != "" {
1385                 if rmerr := os.Remove(runner.ArvMountPoint); rmerr != nil {
1386                         runner.CrunchLog.Printf("While cleaning up arv-mount directory %s: %v", runner.ArvMountPoint, rmerr)
1387                 }
1388         }
1389
1390         if rmerr := os.RemoveAll(runner.parentTemp); rmerr != nil {
1391                 runner.CrunchLog.Printf("While cleaning up temporary directory %s: %v", runner.parentTemp, rmerr)
1392         }
1393 }
1394
1395 // CommitLogs posts the collection containing the final container logs.
1396 func (runner *ContainerRunner) CommitLogs() error {
1397         func() {
1398                 // Hold cStateLock to prevent races on CrunchLog (e.g., stop()).
1399                 runner.cStateLock.Lock()
1400                 defer runner.cStateLock.Unlock()
1401
1402                 runner.CrunchLog.Print(runner.finalState)
1403
1404                 if runner.arvMountLog != nil {
1405                         runner.arvMountLog.Close()
1406                 }
1407                 runner.CrunchLog.Close()
1408
1409                 // Closing CrunchLog above allows them to be committed to Keep at this
1410                 // point, but re-open crunch log with ArvClient in case there are any
1411                 // other further errors (such as failing to write the log to Keep!)
1412                 // while shutting down
1413                 runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{
1414                         ArvClient:     runner.DispatcherArvClient,
1415                         UUID:          runner.Container.UUID,
1416                         loggingStream: "crunch-run",
1417                         writeCloser:   nil,
1418                 })
1419                 runner.CrunchLog.Immediate = log.New(os.Stderr, runner.Container.UUID+" ", 0)
1420         }()
1421
1422         if runner.LogsPDH != nil {
1423                 // If we have already assigned something to LogsPDH,
1424                 // we must be closing the re-opened log, which won't
1425                 // end up getting attached to the container record and
1426                 // therefore doesn't need to be saved as a collection
1427                 // -- it exists only to send logs to other channels.
1428                 return nil
1429         }
1430         saved, err := runner.saveLogCollection(true)
1431         if err != nil {
1432                 return fmt.Errorf("error saving log collection: %s", err)
1433         }
1434         runner.logMtx.Lock()
1435         defer runner.logMtx.Unlock()
1436         runner.LogsPDH = &saved.PortableDataHash
1437         return nil
1438 }
1439
1440 func (runner *ContainerRunner) saveLogCollection(final bool) (response arvados.Collection, err error) {
1441         runner.logMtx.Lock()
1442         defer runner.logMtx.Unlock()
1443         if runner.LogsPDH != nil {
1444                 // Already finalized.
1445                 return
1446         }
1447         mt, err := runner.LogCollection.MarshalManifest(".")
1448         if err != nil {
1449                 err = fmt.Errorf("error creating log manifest: %v", err)
1450                 return
1451         }
1452         updates := arvadosclient.Dict{
1453                 "name":          "logs for " + runner.Container.UUID,
1454                 "manifest_text": mt,
1455         }
1456         if final {
1457                 updates["is_trashed"] = true
1458         } else {
1459                 exp := time.Now().Add(crunchLogUpdatePeriod * 24)
1460                 updates["trash_at"] = exp
1461                 updates["delete_at"] = exp
1462         }
1463         reqBody := arvadosclient.Dict{"collection": updates}
1464         if runner.logUUID == "" {
1465                 reqBody["ensure_unique_name"] = true
1466                 err = runner.DispatcherArvClient.Create("collections", reqBody, &response)
1467         } else {
1468                 err = runner.DispatcherArvClient.Update("collections", runner.logUUID, reqBody, &response)
1469         }
1470         if err != nil {
1471                 return
1472         }
1473         runner.logUUID = response.UUID
1474         return
1475 }
1476
1477 // UpdateContainerRunning updates the container state to "Running"
1478 func (runner *ContainerRunner) UpdateContainerRunning() error {
1479         runner.cStateLock.Lock()
1480         defer runner.cStateLock.Unlock()
1481         if runner.cCancelled {
1482                 return ErrCancelled
1483         }
1484         return runner.DispatcherArvClient.Update("containers", runner.Container.UUID,
1485                 arvadosclient.Dict{"container": arvadosclient.Dict{"state": "Running"}}, nil)
1486 }
1487
1488 // ContainerToken returns the api_token the container (and any
1489 // arv-mount processes) are allowed to use.
1490 func (runner *ContainerRunner) ContainerToken() (string, error) {
1491         if runner.token != "" {
1492                 return runner.token, nil
1493         }
1494
1495         var auth arvados.APIClientAuthorization
1496         err := runner.DispatcherArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth)
1497         if err != nil {
1498                 return "", err
1499         }
1500         runner.token = fmt.Sprintf("v2/%s/%s/%s", auth.UUID, auth.APIToken, runner.Container.UUID)
1501         return runner.token, nil
1502 }
1503
1504 // UpdateContainerComplete updates the container record state on API
1505 // server to "Complete" or "Cancelled"
1506 func (runner *ContainerRunner) UpdateContainerFinal() error {
1507         update := arvadosclient.Dict{}
1508         update["state"] = runner.finalState
1509         if runner.LogsPDH != nil {
1510                 update["log"] = *runner.LogsPDH
1511         }
1512         if runner.finalState == "Complete" {
1513                 if runner.ExitCode != nil {
1514                         update["exit_code"] = *runner.ExitCode
1515                 }
1516                 if runner.OutputPDH != nil {
1517                         update["output"] = *runner.OutputPDH
1518                 }
1519         }
1520         return runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{"container": update}, nil)
1521 }
1522
1523 // IsCancelled returns the value of Cancelled, with goroutine safety.
1524 func (runner *ContainerRunner) IsCancelled() bool {
1525         runner.cStateLock.Lock()
1526         defer runner.cStateLock.Unlock()
1527         return runner.cCancelled
1528 }
1529
1530 // NewArvLogWriter creates an ArvLogWriter
1531 func (runner *ContainerRunner) NewArvLogWriter(name string) (io.WriteCloser, error) {
1532         writer, err := runner.LogCollection.OpenFile(name+".txt", os.O_CREATE|os.O_WRONLY, 0666)
1533         if err != nil {
1534                 return nil, err
1535         }
1536         return &ArvLogWriter{
1537                 ArvClient:     runner.DispatcherArvClient,
1538                 UUID:          runner.Container.UUID,
1539                 loggingStream: name,
1540                 writeCloser:   writer,
1541         }, nil
1542 }
1543
1544 // Run the full container lifecycle.
1545 func (runner *ContainerRunner) Run() (err error) {
1546         runner.CrunchLog.Printf("crunch-run %s started", version)
1547         runner.CrunchLog.Printf("Executing container '%s'", runner.Container.UUID)
1548
1549         hostname, hosterr := os.Hostname()
1550         if hosterr != nil {
1551                 runner.CrunchLog.Printf("Error getting hostname '%v'", hosterr)
1552         } else {
1553                 runner.CrunchLog.Printf("Executing on host '%s'", hostname)
1554         }
1555
1556         runner.finalState = "Queued"
1557
1558         defer func() {
1559                 runner.CleanupDirs()
1560
1561                 runner.CrunchLog.Printf("crunch-run finished")
1562                 runner.CrunchLog.Close()
1563         }()
1564
1565         err = runner.fetchContainerRecord()
1566         if err != nil {
1567                 return
1568         }
1569         if runner.Container.State != "Locked" {
1570                 return fmt.Errorf("dispatch error detected: container %q has state %q", runner.Container.UUID, runner.Container.State)
1571         }
1572
1573         defer func() {
1574                 // checkErr prints e (unless it's nil) and sets err to
1575                 // e (unless err is already non-nil). Thus, if err
1576                 // hasn't already been assigned when Run() returns,
1577                 // this cleanup func will cause Run() to return the
1578                 // first non-nil error that is passed to checkErr().
1579                 checkErr := func(errorIn string, e error) {
1580                         if e == nil {
1581                                 return
1582                         }
1583                         runner.CrunchLog.Printf("error in %s: %v", errorIn, e)
1584                         if err == nil {
1585                                 err = e
1586                         }
1587                         if runner.finalState == "Complete" {
1588                                 // There was an error in the finalization.
1589                                 runner.finalState = "Cancelled"
1590                         }
1591                 }
1592
1593                 // Log the error encountered in Run(), if any
1594                 checkErr("Run", err)
1595
1596                 if runner.finalState == "Queued" {
1597                         runner.UpdateContainerFinal()
1598                         return
1599                 }
1600
1601                 if runner.IsCancelled() {
1602                         runner.finalState = "Cancelled"
1603                         // but don't return yet -- we still want to
1604                         // capture partial output and write logs
1605                 }
1606
1607                 checkErr("CaptureOutput", runner.CaptureOutput())
1608                 checkErr("stopHoststat", runner.stopHoststat())
1609                 checkErr("CommitLogs", runner.CommitLogs())
1610                 checkErr("UpdateContainerFinal", runner.UpdateContainerFinal())
1611         }()
1612
1613         runner.setupSignals()
1614         err = runner.startHoststat()
1615         if err != nil {
1616                 return
1617         }
1618
1619         // check for and/or load image
1620         err = runner.LoadImage()
1621         if err != nil {
1622                 if !runner.checkBrokenNode(err) {
1623                         // Failed to load image but not due to a "broken node"
1624                         // condition, probably user error.
1625                         runner.finalState = "Cancelled"
1626                 }
1627                 err = fmt.Errorf("While loading container image: %v", err)
1628                 return
1629         }
1630
1631         // set up FUSE mount and binds
1632         err = runner.SetupMounts()
1633         if err != nil {
1634                 runner.finalState = "Cancelled"
1635                 err = fmt.Errorf("While setting up mounts: %v", err)
1636                 return
1637         }
1638
1639         err = runner.CreateContainer()
1640         if err != nil {
1641                 return
1642         }
1643         err = runner.LogHostInfo()
1644         if err != nil {
1645                 return
1646         }
1647         err = runner.LogNodeRecord()
1648         if err != nil {
1649                 return
1650         }
1651         err = runner.LogContainerRecord()
1652         if err != nil {
1653                 return
1654         }
1655
1656         if runner.IsCancelled() {
1657                 return
1658         }
1659
1660         err = runner.UpdateContainerRunning()
1661         if err != nil {
1662                 return
1663         }
1664         runner.finalState = "Cancelled"
1665
1666         err = runner.startCrunchstat()
1667         if err != nil {
1668                 return
1669         }
1670
1671         err = runner.StartContainer()
1672         if err != nil {
1673                 runner.checkBrokenNode(err)
1674                 return
1675         }
1676
1677         err = runner.WaitFinish()
1678         if err == nil && !runner.IsCancelled() {
1679                 runner.finalState = "Complete"
1680         }
1681         return
1682 }
1683
1684 // Fetch the current container record (uuid = runner.Container.UUID)
1685 // into runner.Container.
1686 func (runner *ContainerRunner) fetchContainerRecord() error {
1687         reader, err := runner.DispatcherArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
1688         if err != nil {
1689                 return fmt.Errorf("error fetching container record: %v", err)
1690         }
1691         defer reader.Close()
1692
1693         dec := json.NewDecoder(reader)
1694         dec.UseNumber()
1695         err = dec.Decode(&runner.Container)
1696         if err != nil {
1697                 return fmt.Errorf("error decoding container record: %v", err)
1698         }
1699
1700         var sm struct {
1701                 SecretMounts map[string]arvados.Mount `json:"secret_mounts"`
1702         }
1703
1704         containerToken, err := runner.ContainerToken()
1705         if err != nil {
1706                 return fmt.Errorf("error getting container token: %v", err)
1707         }
1708
1709         runner.ContainerArvClient, runner.ContainerKeepClient,
1710                 runner.containerClient, err = runner.MkArvClient(containerToken)
1711         if err != nil {
1712                 return fmt.Errorf("error creating container API client: %v", err)
1713         }
1714
1715         err = runner.ContainerArvClient.Call("GET", "containers", runner.Container.UUID, "secret_mounts", nil, &sm)
1716         if err != nil {
1717                 if apierr, ok := err.(arvadosclient.APIServerError); !ok || apierr.HttpStatusCode != 404 {
1718                         return fmt.Errorf("error fetching secret_mounts: %v", err)
1719                 }
1720                 // ok && apierr.HttpStatusCode == 404, which means
1721                 // secret_mounts isn't supported by this API server.
1722         }
1723         runner.SecretMounts = sm.SecretMounts
1724
1725         return nil
1726 }
1727
1728 // NewContainerRunner creates a new container runner.
1729 func NewContainerRunner(dispatcherClient *arvados.Client,
1730         dispatcherArvClient IArvadosClient,
1731         dispatcherKeepClient IKeepClient,
1732         docker ThinDockerClient,
1733         containerUUID string) (*ContainerRunner, error) {
1734
1735         cr := &ContainerRunner{
1736                 dispatcherClient:     dispatcherClient,
1737                 DispatcherArvClient:  dispatcherArvClient,
1738                 DispatcherKeepClient: dispatcherKeepClient,
1739                 Docker:               docker,
1740         }
1741         cr.NewLogWriter = cr.NewArvLogWriter
1742         cr.RunArvMount = cr.ArvMountCmd
1743         cr.MkTempDir = ioutil.TempDir
1744         cr.MkArvClient = func(token string) (IArvadosClient, IKeepClient, *arvados.Client, error) {
1745                 cl, err := arvadosclient.MakeArvadosClient()
1746                 if err != nil {
1747                         return nil, nil, nil, err
1748                 }
1749                 cl.ApiToken = token
1750                 kc, err := keepclient.MakeKeepClient(cl)
1751                 if err != nil {
1752                         return nil, nil, nil, err
1753                 }
1754                 c2 := arvados.NewClientFromEnv()
1755                 c2.AuthToken = token
1756                 return cl, kc, c2, nil
1757         }
1758         var err error
1759         cr.LogCollection, err = (&arvados.Collection{}).FileSystem(cr.dispatcherClient, cr.DispatcherKeepClient)
1760         if err != nil {
1761                 return nil, err
1762         }
1763         cr.Container.UUID = containerUUID
1764         w, err := cr.NewLogWriter("crunch-run")
1765         if err != nil {
1766                 return nil, err
1767         }
1768         cr.CrunchLog = NewThrottledLogger(w)
1769         cr.CrunchLog.Immediate = log.New(os.Stderr, containerUUID+" ", 0)
1770
1771         loadLogThrottleParams(dispatcherArvClient)
1772         go cr.updateLogs()
1773
1774         return cr, nil
1775 }
1776
1777 func main() {
1778         statInterval := flag.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting")
1779         cgroupRoot := flag.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree")
1780         cgroupParent := flag.String("cgroup-parent", "docker", "name of container's parent cgroup (ignored if -cgroup-parent-subsystem is used)")
1781         cgroupParentSubsystem := flag.String("cgroup-parent-subsystem", "", "use current cgroup for given subsystem as parent cgroup for container")
1782         caCertsPath := flag.String("ca-certs", "", "Path to TLS root certificates")
1783         detach := flag.Bool("detach", false, "Detach from parent process and run in the background")
1784         stdinEnv := flag.Bool("stdin-env", false, "Load environment variables from JSON message on stdin")
1785         sleep := flag.Duration("sleep", 0, "Delay before starting (testing use only)")
1786         kill := flag.Int("kill", -1, "Send signal to an existing crunch-run process for given UUID")
1787         list := flag.Bool("list", false, "List UUIDs of existing crunch-run processes")
1788         enableNetwork := flag.String("container-enable-networking", "default",
1789                 `Specify if networking should be enabled for container.  One of 'default', 'always':
1790         default: only enable networking if container requests it.
1791         always:  containers always have networking enabled
1792         `)
1793         networkMode := flag.String("container-network-mode", "default",
1794                 `Set networking mode for container.  Corresponds to Docker network mode (--net).
1795         `)
1796         memprofile := flag.String("memprofile", "", "write memory profile to `file` after running container")
1797         getVersion := flag.Bool("version", false, "Print version information and exit.")
1798         flag.Duration("check-containerd", 0, "Ignored. Exists for compatibility with older versions.")
1799
1800         ignoreDetachFlag := false
1801         if len(os.Args) > 1 && os.Args[1] == "-no-detach" {
1802                 // This process was invoked by a parent process, which
1803                 // has passed along its own arguments, including
1804                 // -detach, after the leading -no-detach flag.  Strip
1805                 // the leading -no-detach flag (it's not recognized by
1806                 // flag.Parse()) and ignore the -detach flag that
1807                 // comes later.
1808                 os.Args = append([]string{os.Args[0]}, os.Args[2:]...)
1809                 ignoreDetachFlag = true
1810         }
1811
1812         flag.Parse()
1813
1814         var env map[string]interface{}
1815         if *stdinEnv && !ignoreDetachFlag {
1816                 // Load env vars on stdin if asked (but not in a
1817                 // detached child process, in which case stdin is
1818                 // /dev/null).
1819                 env = loadEnv(os.Stdin)
1820         }
1821
1822         switch {
1823         case *detach && !ignoreDetachFlag:
1824                 os.Exit(Detach(flag.Arg(0), os.Args, os.Stdout, os.Stderr))
1825         case *kill >= 0:
1826                 os.Exit(KillProcess(flag.Arg(0), syscall.Signal(*kill), os.Stdout, os.Stderr))
1827         case *list:
1828                 os.Exit(ListProcesses(os.Stdout, os.Stderr))
1829         }
1830
1831         // Print version information if requested
1832         if *getVersion {
1833                 fmt.Printf("crunch-run %s\n", version)
1834                 return
1835         }
1836
1837         log.Printf("crunch-run %s started", version)
1838         time.Sleep(*sleep)
1839
1840         containerId := flag.Arg(0)
1841
1842         if *caCertsPath != "" {
1843                 arvadosclient.CertFiles = []string{*caCertsPath}
1844         }
1845
1846         api, err := arvadosclient.MakeArvadosClient()
1847         if err != nil {
1848                 log.Fatalf("%s: %v", containerId, err)
1849         }
1850         api.Retries = 8
1851
1852         kc, kcerr := keepclient.MakeKeepClient(api)
1853         if kcerr != nil {
1854                 log.Fatalf("%s: %v", containerId, kcerr)
1855         }
1856         kc.BlockCache = &keepclient.BlockCache{MaxBlocks: 2}
1857         kc.Retries = 4
1858
1859         // API version 1.21 corresponds to Docker 1.9, which is currently the
1860         // minimum version we want to support.
1861         docker, dockererr := dockerclient.NewClient(dockerclient.DefaultDockerHost, "1.21", nil, nil)
1862
1863         cr, err := NewContainerRunner(arvados.NewClientFromEnv(), api, kc, docker, containerId)
1864         if err != nil {
1865                 log.Fatal(err)
1866         }
1867         if dockererr != nil {
1868                 cr.CrunchLog.Printf("%s: %v", containerId, dockererr)
1869                 cr.checkBrokenNode(dockererr)
1870                 cr.CrunchLog.Close()
1871                 os.Exit(1)
1872         }
1873
1874         cr.dispatchEnv = env
1875
1876         parentTemp, tmperr := cr.MkTempDir("", "crunch-run."+containerId+".")
1877         if tmperr != nil {
1878                 log.Fatalf("%s: %v", containerId, tmperr)
1879         }
1880
1881         cr.parentTemp = parentTemp
1882         cr.statInterval = *statInterval
1883         cr.cgroupRoot = *cgroupRoot
1884         cr.expectCgroupParent = *cgroupParent
1885         cr.enableNetwork = *enableNetwork
1886         cr.networkMode = *networkMode
1887         if *cgroupParentSubsystem != "" {
1888                 p := findCgroup(*cgroupParentSubsystem)
1889                 cr.setCgroupParent = p
1890                 cr.expectCgroupParent = p
1891         }
1892
1893         runerr := cr.Run()
1894
1895         if *memprofile != "" {
1896                 f, err := os.Create(*memprofile)
1897                 if err != nil {
1898                         log.Printf("could not create memory profile: %s", err)
1899                 }
1900                 runtime.GC() // get up-to-date statistics
1901                 if err := pprof.WriteHeapProfile(f); err != nil {
1902                         log.Printf("could not write memory profile: %s", err)
1903                 }
1904                 closeerr := f.Close()
1905                 if closeerr != nil {
1906                         log.Printf("closing memprofile file: %s", err)
1907                 }
1908         }
1909
1910         if runerr != nil {
1911                 log.Fatalf("%s: %v", containerId, runerr)
1912         }
1913 }
1914
1915 func loadEnv(rdr io.Reader) map[string]interface{} {
1916         buf, err := ioutil.ReadAll(rdr)
1917         if err != nil {
1918                 log.Fatalf("read stdin: %s", err)
1919         }
1920         var env map[string]interface{}
1921         err = json.Unmarshal(buf, &env)
1922         if err != nil {
1923                 log.Fatalf("decode stdin: %s", err)
1924         }
1925         for k, v := range env {
1926                 if v, ok := v.(string); ok {
1927                         err = os.Setenv(k, v)
1928                         if err != nil {
1929                                 log.Fatalf("setenv(%q): %s", k, err)
1930                         }
1931                 }
1932         }
1933         return env
1934 }