lib/crunchrun/crunchrun.go

   1 // Copyright (C) The Arvados Authors. All rights reserved.
   2 //
   3 // SPDX-License-Identifier: AGPL-3.0
   4
   5 package crunchrun
   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.arvados.org/arvados.git/lib/cmd"
  31         "git.arvados.org/arvados.git/lib/crunchstat"
  32         "git.arvados.org/arvados.git/sdk/go/arvados"
  33         "git.arvados.org/arvados.git/sdk/go/arvadosclient"
  34         "git.arvados.org/arvados.git/sdk/go/keepclient"
  35         "git.arvados.org/arvados.git/sdk/go/manifest"
  36         "golang.org/x/net/context"
  37
  38         dockertypes "github.com/docker/docker/api/types"
  39         dockercontainer "github.com/docker/docker/api/types/container"
  40         dockernetwork "github.com/docker/docker/api/types/network"
  41         dockerclient "github.com/docker/docker/client"
  42 )
  43
  44 type command struct{}
  45
  46 var Command = command{}
  47
  48 // IArvadosClient is the minimal Arvados API methods used by crunch-run.
  49 type IArvadosClient interface {
  50         Create(resourceType string, parameters arvadosclient.Dict, output interface{}) error
  51         Get(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  52         Update(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  53         Call(method, resourceType, uuid, action string, parameters arvadosclient.Dict, output interface{}) error
  54         CallRaw(method string, resourceType string, uuid string, action string, parameters arvadosclient.Dict) (reader io.ReadCloser, err error)
  55         Discovery(key string) (interface{}, error)
  56 }
  57
  58 // ErrCancelled is the error returned when the container is cancelled.
  59 var ErrCancelled = errors.New("Cancelled")
  60
  61 // IKeepClient is the minimal Keep API methods used by crunch-run.
  62 type IKeepClient interface {
  63         PutB(buf []byte) (string, int, error)
  64         ReadAt(locator string, p []byte, off int) (int, error)
  65         ManifestFileReader(m manifest.Manifest, filename string) (arvados.File, error)
  66         LocalLocator(locator string) (string, error)
  67         ClearBlockCache()
  68 }
  69
  70 // NewLogWriter is a factory function to create a new log writer.
  71 type NewLogWriter func(name string) (io.WriteCloser, error)
  72
  73 type RunArvMount func(args []string, tok string) (*exec.Cmd, error)
  74
  75 type MkTempDir func(string, string) (string, error)
  76
  77 // ThinDockerClient is the minimal Docker client interface used by crunch-run.
  78 type ThinDockerClient interface {
  79         ContainerAttach(ctx context.Context, container string, options dockertypes.ContainerAttachOptions) (dockertypes.HijackedResponse, error)
  80         ContainerCreate(ctx context.Context, config *dockercontainer.Config, hostConfig *dockercontainer.HostConfig,
  81                 networkingConfig *dockernetwork.NetworkingConfig, containerName string) (dockercontainer.ContainerCreateCreatedBody, error)
  82         ContainerStart(ctx context.Context, container string, options dockertypes.ContainerStartOptions) error
  83         ContainerRemove(ctx context.Context, container string, options dockertypes.ContainerRemoveOptions) error
  84         ContainerWait(ctx context.Context, container string, condition dockercontainer.WaitCondition) (<-chan dockercontainer.ContainerWaitOKBody, <-chan error)
  85         ContainerInspect(ctx context.Context, id string) (dockertypes.ContainerJSON, error)
  86         ImageInspectWithRaw(ctx context.Context, image string) (dockertypes.ImageInspect, []byte, error)
  87         ImageLoad(ctx context.Context, input io.Reader, quiet bool) (dockertypes.ImageLoadResponse, error)
  88         ImageRemove(ctx context.Context, image string, options dockertypes.ImageRemoveOptions) ([]dockertypes.ImageDeleteResponseItem, error)
  89 }
  90
  91 type PsProcess interface {
  92         CmdlineSlice() ([]string, error)
  93 }
  94
  95 // ContainerRunner is the main stateful struct used for a single execution of a
  96 // container.
  97 type ContainerRunner struct {
  98         ContainerExecRunner ThinContainerExecRunner
  99
 100         //Docker          ThinDockerClient
 101         //ContainerConfig dockercontainer.Config     //FIXME: translate this to the ThinContainerRunner interface
 102         HostConfig dockercontainer.HostConfig //FIXME: translate this to the ThinContainerRunner interface
 103         //--------------
 104
 105         // Dispatcher client is initialized with the Dispatcher token.
 106         // This is a privileged token used to manage container status
 107         // and logs.
 108         //
 109         // We have both dispatcherClient and DispatcherArvClient
 110         // because there are two different incompatible Arvados Go
 111         // SDKs and we have to use both (hopefully this gets fixed in
 112         // #14467)
 113         dispatcherClient     *arvados.Client
 114         DispatcherArvClient  IArvadosClient
 115         DispatcherKeepClient IKeepClient
 116
 117         // Container client is initialized with the Container token
 118         // This token controls the permissions of the container, and
 119         // must be used for operations such as reading collections.
 120         //
 121         // Same comment as above applies to
 122         // containerClient/ContainerArvClient.
 123         containerClient     *arvados.Client
 124         ContainerArvClient  IArvadosClient
 125         ContainerKeepClient IKeepClient
 126
 127         Container     arvados.Container
 128         token         string
 129         ContainerID   string
 130         ExitCode      *int
 131         NewLogWriter  NewLogWriter
 132         loggingDone   chan bool
 133         CrunchLog     *ThrottledLogger
 134         Stdout        io.WriteCloser
 135         Stderr        io.WriteCloser
 136         logUUID       string
 137         logMtx        sync.Mutex
 138         LogCollection arvados.CollectionFileSystem
 139         LogsPDH       *string
 140         RunArvMount   RunArvMount
 141         MkTempDir     MkTempDir
 142         ArvMount      *exec.Cmd
 143         ArvMountPoint string
 144         HostOutputDir string
 145         Binds         []string
 146         Volumes       map[string]struct{}
 147         OutputPDH     *string
 148         SigChan       chan os.Signal
 149         ArvMountExit  chan error
 150         SecretMounts  map[string]arvados.Mount
 151         MkArvClient   func(token string) (IArvadosClient, IKeepClient, *arvados.Client, error)
 152         finalState    string
 153         parentTemp    string
 154
 155         statLogger       io.WriteCloser
 156         statReporter     *crunchstat.Reporter
 157         hoststatLogger   io.WriteCloser
 158         hoststatReporter *crunchstat.Reporter
 159         statInterval     time.Duration
 160         cgroupRoot       string
 161         // What we expect the container's cgroup parent to be.
 162         expectCgroupParent string
 163         // What we tell docker to use as the container's cgroup
 164         // parent. Note: Ideally we would use the same field for both
 165         // expectCgroupParent and setCgroupParent, and just make it
 166         // default to "docker". However, when using docker < 1.10 with
 167         // systemd, specifying a non-empty cgroup parent (even the
 168         // default value "docker") hits a docker bug
 169         // (https://github.com/docker/docker/issues/17126). Using two
 170         // separate fields makes it possible to use the "expect cgroup
 171         // parent to be X" feature even on sites where the "specify
 172         // cgroup parent" feature breaks.
 173         setCgroupParent string
 174
 175         cStateLock sync.Mutex
 176         cCancelled bool // StopContainer() invoked
 177         cRemoved   bool // docker confirmed the container no longer exists
 178
 179         enableNetwork string // one of "default" or "always"
 180         networkMode   string // passed through to HostConfig.NetworkMode
 181         arvMountLog   *ThrottledLogger
 182
 183         containerWatchdogInterval time.Duration
 184
 185         gateway Gateway
 186 }
 187
 188 // setupSignals sets up signal handling to gracefully terminate the underlying
 189 // Docker container and update state when receiving a TERM, INT or QUIT signal.
 190 func (runner *ContainerRunner) setupSignals() {
 191         runner.SigChan = make(chan os.Signal, 1)
 192         signal.Notify(runner.SigChan, syscall.SIGTERM)
 193         signal.Notify(runner.SigChan, syscall.SIGINT)
 194         signal.Notify(runner.SigChan, syscall.SIGQUIT)
 195
 196         go func(sig chan os.Signal) {
 197                 for s := range sig {
 198                         runner.stop(s)
 199                 }
 200         }(runner.SigChan)
 201 }
 202
 203 // stop the underlying Docker container.
 204 func (runner *ContainerRunner) stop(sig os.Signal) {
 205         runner.cStateLock.Lock()
 206         defer runner.cStateLock.Unlock()
 207         if sig != nil {
 208                 runner.CrunchLog.Printf("caught signal: %v", sig)
 209         }
 210         if runner.ContainerID == "" {
 211                 return
 212         }
 213         runner.cCancelled = true
 214         runner.CrunchLog.Printf("removing container")
 215         err := runner.ContainerExecRunner.ContainerRemove(context.TODO(), runner.ContainerID, ContainerRemoveOptions{Force: true})
 216         if err != nil {
 217                 runner.CrunchLog.Printf("error removing container: %s", err)
 218         }
 219         if err == nil || strings.Contains(err.Error(), "No such container: "+runner.ContainerID) {
 220                 runner.cRemoved = true
 221         }
 222 }
 223
 224 var errorBlacklist = []string{
 225         "(?ms).*[Cc]annot connect to the Docker daemon.*",
 226         "(?ms).*oci runtime error.*starting container process.*container init.*mounting.*to rootfs.*no such file or directory.*",
 227         "(?ms).*grpc: the connection is unavailable.*",
 228 }
 229 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)")
 230
 231 func (runner *ContainerRunner) runBrokenNodeHook() {
 232         if *brokenNodeHook == "" {
 233                 path := filepath.Join(lockdir, brokenfile)
 234                 runner.CrunchLog.Printf("Writing %s to mark node as broken", path)
 235                 f, err := os.OpenFile(path, os.O_CREATE|os.O_RDWR, 0700)
 236                 if err != nil {
 237                         runner.CrunchLog.Printf("Error writing %s: %s", path, err)
 238                         return
 239                 }
 240                 f.Close()
 241         } else {
 242                 runner.CrunchLog.Printf("Running broken node hook %q", *brokenNodeHook)
 243                 // run killme script
 244                 c := exec.Command(*brokenNodeHook)
 245                 c.Stdout = runner.CrunchLog
 246                 c.Stderr = runner.CrunchLog
 247                 err := c.Run()
 248                 if err != nil {
 249                         runner.CrunchLog.Printf("Error running broken node hook: %v", err)
 250                 }
 251         }
 252 }
 253
 254 func (runner *ContainerRunner) checkBrokenNode(goterr error) bool {
 255         for _, d := range errorBlacklist {
 256                 if m, e := regexp.MatchString(d, goterr.Error()); m && e == nil {
 257                         runner.CrunchLog.Printf("Error suggests node is unable to run containers: %v", goterr)
 258                         runner.runBrokenNodeHook()
 259                         return true
 260                 }
 261         }
 262         return false
 263 }
 264
 265 // LoadImage determines the docker image id from the container record and
 266 // checks if it is available in the local Docker image store.  If not, it loads
 267 // the image from Keep.
 268 func (runner *ContainerRunner) LoadImage() (err error) {
 269
 270         runner.CrunchLog.Printf("Fetching Docker image from collection '%s'", runner.Container.ContainerImage)
 271
 272         var collection arvados.Collection
 273         err = runner.ContainerArvClient.Get("collections", runner.Container.ContainerImage, nil, &collection)
 274         if err != nil {
 275                 return fmt.Errorf("While getting container image collection: %v", err)
 276         }
 277         manifest := manifest.Manifest{Text: collection.ManifestText}
 278         var img, imageID string
 279         for ms := range manifest.StreamIter() {
 280                 img = ms.FileStreamSegments[0].Name
 281                 if !strings.HasSuffix(img, ".tar") {
 282                         return fmt.Errorf("First file in the container image collection does not end in .tar")
 283                 }
 284                 imageID = img[:len(img)-4]
 285         }
 286
 287         runner.CrunchLog.Printf("Using Docker image id '%s'", imageID)
 288
 289         _, _, err = runner.ContainerExecRunner.ImageInspectWithRaw(context.TODO(), imageID)
 290         if err != nil {
 291                 runner.CrunchLog.Print("Loading Docker image from keep")
 292
 293                 var readCloser io.ReadCloser
 294                 readCloser, err = runner.ContainerKeepClient.ManifestFileReader(manifest, img)
 295                 if err != nil {
 296                         return fmt.Errorf("While creating ManifestFileReader for container image: %v", err)
 297                 }
 298
 299                 response, err := runner.ContainerExecRunner.ImageLoad(context.TODO(), readCloser, true)
 300                 if err != nil {
 301                         return fmt.Errorf("While loading container image into Docker: %v", err)
 302                 }
 303
 304                 defer response.Body.Close()
 305                 rbody, err := ioutil.ReadAll(response.Body)
 306                 if err != nil {
 307                         return fmt.Errorf("Reading response to image load: %v", err)
 308                 }
 309                 runner.CrunchLog.Printf("Docker response: %s", rbody)
 310         } else {
 311                 runner.CrunchLog.Print("Docker image is available")
 312         }
 313
 314         runner.ContainerExecRunner.SetImage(imageID)
 315
 316         runner.ContainerKeepClient.ClearBlockCache()
 317
 318         return nil
 319 }
 320
 321 func (runner *ContainerRunner) ArvMountCmd(arvMountCmd []string, token string) (c *exec.Cmd, err error) {
 322         c = exec.Command("arv-mount", arvMountCmd...)
 323
 324         // Copy our environment, but override ARVADOS_API_TOKEN with
 325         // the container auth token.
 326         c.Env = nil
 327         for _, s := range os.Environ() {
 328                 if !strings.HasPrefix(s, "ARVADOS_API_TOKEN=") {
 329                         c.Env = append(c.Env, s)
 330                 }
 331         }
 332         c.Env = append(c.Env, "ARVADOS_API_TOKEN="+token)
 333
 334         w, err := runner.NewLogWriter("arv-mount")
 335         if err != nil {
 336                 return nil, err
 337         }
 338         runner.arvMountLog = NewThrottledLogger(w)
 339         c.Stdout = runner.arvMountLog
 340         c.Stderr = runner.arvMountLog
 341
 342         runner.CrunchLog.Printf("Running %v", c.Args)
 343
 344         err = c.Start()
 345         if err != nil {
 346                 return nil, err
 347         }
 348
 349         statReadme := make(chan bool)
 350         runner.ArvMountExit = make(chan error)
 351
 352         keepStatting := true
 353         go func() {
 354                 for keepStatting {
 355                         time.Sleep(100 * time.Millisecond)
 356                         _, err = os.Stat(fmt.Sprintf("%s/by_id/README", runner.ArvMountPoint))
 357                         if err == nil {
 358                                 keepStatting = false
 359                                 statReadme <- true
 360                         }
 361                 }
 362                 close(statReadme)
 363         }()
 364
 365         go func() {
 366                 mnterr := c.Wait()
 367                 if mnterr != nil {
 368                         runner.CrunchLog.Printf("Arv-mount exit error: %v", mnterr)
 369                 }
 370                 runner.ArvMountExit <- mnterr
 371                 close(runner.ArvMountExit)
 372         }()
 373
 374         select {
 375         case <-statReadme:
 376                 break
 377         case err := <-runner.ArvMountExit:
 378                 runner.ArvMount = nil
 379                 keepStatting = false
 380                 return nil, err
 381         }
 382
 383         return c, nil
 384 }
 385
 386 func (runner *ContainerRunner) SetupArvMountPoint(prefix string) (err error) {
 387         if runner.ArvMountPoint == "" {
 388                 runner.ArvMountPoint, err = runner.MkTempDir(runner.parentTemp, prefix)
 389         }
 390         return
 391 }
 392
 393 func copyfile(src string, dst string) (err error) {
 394         srcfile, err := os.Open(src)
 395         if err != nil {
 396                 return
 397         }
 398
 399         os.MkdirAll(path.Dir(dst), 0777)
 400
 401         dstfile, err := os.Create(dst)
 402         if err != nil {
 403                 return
 404         }
 405         _, err = io.Copy(dstfile, srcfile)
 406         if err != nil {
 407                 return
 408         }
 409
 410         err = srcfile.Close()
 411         err2 := dstfile.Close()
 412
 413         if err != nil {
 414                 return
 415         }
 416
 417         if err2 != nil {
 418                 return err2
 419         }
 420
 421         return nil
 422 }
 423
 424 func (runner *ContainerRunner) SetupMounts() (err error) {
 425         err = runner.SetupArvMountPoint("keep")
 426         if err != nil {
 427                 return fmt.Errorf("While creating keep mount temp dir: %v", err)
 428         }
 429
 430         token, err := runner.ContainerToken()
 431         if err != nil {
 432                 return fmt.Errorf("could not get container token: %s", err)
 433         }
 434
 435         pdhOnly := true
 436         tmpcount := 0
 437         arvMountCmd := []string{
 438                 "--foreground",
 439                 "--allow-other",
 440                 "--read-write",
 441                 fmt.Sprintf("--crunchstat-interval=%v", runner.statInterval.Seconds())}
 442
 443         if runner.Container.RuntimeConstraints.KeepCacheRAM > 0 {
 444                 arvMountCmd = append(arvMountCmd, "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheRAM))
 445         }
 446
 447         collectionPaths := []string{}
 448         runner.Binds = nil
 449         runner.Volumes = make(map[string]struct{})
 450         needCertMount := true
 451         type copyFile struct {
 452                 src  string
 453                 bind string
 454         }
 455         var copyFiles []copyFile
 456
 457         var binds []string
 458         for bind := range runner.Container.Mounts {
 459                 binds = append(binds, bind)
 460         }
 461         for bind := range runner.SecretMounts {
 462                 if _, ok := runner.Container.Mounts[bind]; ok {
 463                         return fmt.Errorf("secret mount %q conflicts with regular mount", bind)
 464                 }
 465                 if runner.SecretMounts[bind].Kind != "json" &&
 466                         runner.SecretMounts[bind].Kind != "text" {
 467                         return fmt.Errorf("secret mount %q type is %q but only 'json' and 'text' are permitted",
 468                                 bind, runner.SecretMounts[bind].Kind)
 469                 }
 470                 binds = append(binds, bind)
 471         }
 472         sort.Strings(binds)
 473
 474         for _, bind := range binds {
 475                 mnt, ok := runner.Container.Mounts[bind]
 476                 if !ok {
 477                         mnt = runner.SecretMounts[bind]
 478                 }
 479                 if bind == "stdout" || bind == "stderr" {
 480                         // Is it a "file" mount kind?
 481                         if mnt.Kind != "file" {
 482                                 return fmt.Errorf("unsupported mount kind '%s' for %s: only 'file' is supported", mnt.Kind, bind)
 483                         }
 484
 485                         // Does path start with OutputPath?
 486                         prefix := runner.Container.OutputPath
 487                         if !strings.HasSuffix(prefix, "/") {
 488                                 prefix += "/"
 489                         }
 490                         if !strings.HasPrefix(mnt.Path, prefix) {
 491                                 return fmt.Errorf("%s path does not start with OutputPath: %s, %s", strings.Title(bind), mnt.Path, prefix)
 492                         }
 493                 }
 494
 495                 if bind == "stdin" {
 496                         // Is it a "collection" mount kind?
 497                         if mnt.Kind != "collection" && mnt.Kind != "json" {
 498                                 return fmt.Errorf("unsupported mount kind '%s' for stdin: only 'collection' and 'json' are supported", mnt.Kind)
 499                         }
 500                 }
 501
 502                 if bind == "/etc/arvados/ca-certificates.crt" {
 503                         needCertMount = false
 504                 }
 505
 506                 if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && bind != runner.Container.OutputPath+"/" {
 507                         if mnt.Kind != "collection" && mnt.Kind != "text" && mnt.Kind != "json" {
 508                                 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)
 509                         }
 510                 }
 511
 512                 switch {
 513                 case mnt.Kind == "collection" && bind != "stdin":
 514                         var src string
 515                         if mnt.UUID != "" && mnt.PortableDataHash != "" {
 516                                 return fmt.Errorf("cannot specify both 'uuid' and 'portable_data_hash' for a collection mount")
 517                         }
 518                         if mnt.UUID != "" {
 519                                 if mnt.Writable {
 520                                         return fmt.Errorf("writing to existing collections currently not permitted")
 521                                 }
 522                                 pdhOnly = false
 523                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.UUID)
 524                         } else if mnt.PortableDataHash != "" {
 525                                 if mnt.Writable && !strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 526                                         return fmt.Errorf("can never write to a collection specified by portable data hash")
 527                                 }
 528                                 idx := strings.Index(mnt.PortableDataHash, "/")
 529                                 if idx > 0 {
 530                                         mnt.Path = path.Clean(mnt.PortableDataHash[idx:])
 531                                         mnt.PortableDataHash = mnt.PortableDataHash[0:idx]
 532                                         runner.Container.Mounts[bind] = mnt
 533                                 }
 534                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.PortableDataHash)
 535                                 if mnt.Path != "" && mnt.Path != "." {
 536                                         if strings.HasPrefix(mnt.Path, "./") {
 537                                                 mnt.Path = mnt.Path[2:]
 538                                         } else if strings.HasPrefix(mnt.Path, "/") {
 539                                                 mnt.Path = mnt.Path[1:]
 540                                         }
 541                                         src += "/" + mnt.Path
 542                                 }
 543                         } else {
 544                                 src = fmt.Sprintf("%s/tmp%d", runner.ArvMountPoint, tmpcount)
 545                                 arvMountCmd = append(arvMountCmd, "--mount-tmp")
 546                                 arvMountCmd = append(arvMountCmd, fmt.Sprintf("tmp%d", tmpcount))
 547                                 tmpcount++
 548                         }
 549                         if mnt.Writable {
 550                                 if bind == runner.Container.OutputPath {
 551                                         runner.HostOutputDir = src
 552                                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", src, bind))
 553                                 } else if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 554                                         copyFiles = append(copyFiles, copyFile{src, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
 555                                 } else {
 556                                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", src, bind))
 557                                 }
 558                         } else {
 559                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", src, bind))
 560                         }
 561                         collectionPaths = append(collectionPaths, src)
 562
 563                 case mnt.Kind == "tmp":
 564                         var tmpdir string
 565                         tmpdir, err = runner.MkTempDir(runner.parentTemp, "tmp")
 566                         if err != nil {
 567                                 return fmt.Errorf("while creating mount temp dir: %v", err)
 568                         }
 569                         st, staterr := os.Stat(tmpdir)
 570                         if staterr != nil {
 571                                 return fmt.Errorf("while Stat on temp dir: %v", staterr)
 572                         }
 573                         err = os.Chmod(tmpdir, st.Mode()|os.ModeSetgid|0777)
 574                         if staterr != nil {
 575                                 return fmt.Errorf("while Chmod temp dir: %v", err)
 576                         }
 577                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", tmpdir, bind))
 578                         if bind == runner.Container.OutputPath {
 579                                 runner.HostOutputDir = tmpdir
 580                         }
 581
 582                 case mnt.Kind == "json" || mnt.Kind == "text":
 583                         var filedata []byte
 584                         if mnt.Kind == "json" {
 585                                 filedata, err = json.Marshal(mnt.Content)
 586                                 if err != nil {
 587                                         return fmt.Errorf("encoding json data: %v", err)
 588                                 }
 589                         } else {
 590                                 text, ok := mnt.Content.(string)
 591                                 if !ok {
 592                                         return fmt.Errorf("content for mount %q must be a string", bind)
 593                                 }
 594                                 filedata = []byte(text)
 595                         }
 596
 597                         tmpdir, err := runner.MkTempDir(runner.parentTemp, mnt.Kind)
 598                         if err != nil {
 599                                 return fmt.Errorf("creating temp dir: %v", err)
 600                         }
 601                         tmpfn := filepath.Join(tmpdir, "mountdata."+mnt.Kind)
 602                         err = ioutil.WriteFile(tmpfn, filedata, 0444)
 603                         if err != nil {
 604                                 return fmt.Errorf("writing temp file: %v", err)
 605                         }
 606                         if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 607                                 copyFiles = append(copyFiles, copyFile{tmpfn, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
 608                         } else {
 609                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", tmpfn, bind))
 610                         }
 611
 612                 case mnt.Kind == "git_tree":
 613                         tmpdir, err := runner.MkTempDir(runner.parentTemp, "git_tree")
 614                         if err != nil {
 615                                 return fmt.Errorf("creating temp dir: %v", err)
 616                         }
 617                         err = gitMount(mnt).extractTree(runner.ContainerArvClient, tmpdir, token)
 618                         if err != nil {
 619                                 return err
 620                         }
 621                         runner.Binds = append(runner.Binds, tmpdir+":"+bind+":ro")
 622                 }
 623         }
 624
 625         if runner.HostOutputDir == "" {
 626                 return fmt.Errorf("output path does not correspond to a writable mount point")
 627         }
 628
 629         if needCertMount && runner.Container.RuntimeConstraints.API {
 630                 for _, certfile := range arvadosclient.CertFiles {
 631                         _, err := os.Stat(certfile)
 632                         if err == nil {
 633                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:/etc/arvados/ca-certificates.crt:ro", certfile))
 634                                 break
 635                         }
 636                 }
 637         }
 638
 639         if pdhOnly {
 640                 arvMountCmd = append(arvMountCmd, "--mount-by-pdh", "by_id")
 641         } else {
 642                 arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_id")
 643         }
 644         arvMountCmd = append(arvMountCmd, runner.ArvMountPoint)
 645
 646         runner.ArvMount, err = runner.RunArvMount(arvMountCmd, token)
 647         if err != nil {
 648                 return fmt.Errorf("while trying to start arv-mount: %v", err)
 649         }
 650
 651         for _, p := range collectionPaths {
 652                 _, err = os.Stat(p)
 653                 if err != nil {
 654                         return fmt.Errorf("while checking that input files exist: %v", err)
 655                 }
 656         }
 657
 658         for _, cp := range copyFiles {
 659                 st, err := os.Stat(cp.src)
 660                 if err != nil {
 661                         return fmt.Errorf("while staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 662                 }
 663                 if st.IsDir() {
 664                         err = filepath.Walk(cp.src, func(walkpath string, walkinfo os.FileInfo, walkerr error) error {
 665                                 if walkerr != nil {
 666                                         return walkerr
 667                                 }
 668                                 target := path.Join(cp.bind, walkpath[len(cp.src):])
 669                                 if walkinfo.Mode().IsRegular() {
 670                                         copyerr := copyfile(walkpath, target)
 671                                         if copyerr != nil {
 672                                                 return copyerr
 673                                         }
 674                                         return os.Chmod(target, walkinfo.Mode()|0777)
 675                                 } else if walkinfo.Mode().IsDir() {
 676                                         mkerr := os.MkdirAll(target, 0777)
 677                                         if mkerr != nil {
 678                                                 return mkerr
 679                                         }
 680                                         return os.Chmod(target, walkinfo.Mode()|os.ModeSetgid|0777)
 681                                 } else {
 682                                         return fmt.Errorf("source %q is not a regular file or directory", cp.src)
 683                                 }
 684                         })
 685                 } else if st.Mode().IsRegular() {
 686                         err = copyfile(cp.src, cp.bind)
 687                         if err == nil {
 688                                 err = os.Chmod(cp.bind, st.Mode()|0777)
 689                         }
 690                 }
 691                 if err != nil {
 692                         return fmt.Errorf("while staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 693                 }
 694         }
 695
 696         return nil
 697 }
 698
 699 func (runner *ContainerRunner) ProcessDockerAttach(containerReader io.Reader) {
 700         // Handle docker log protocol
 701         // https://docs.docker.com/engine/reference/api/docker_remote_api_v1.15/#attach-to-a-container
 702         defer close(runner.loggingDone)
 703
 704         header := make([]byte, 8)
 705         var err error
 706         for err == nil {
 707                 _, err = io.ReadAtLeast(containerReader, header, 8)
 708                 if err != nil {
 709                         if err == io.EOF {
 710                                 err = nil
 711                         }
 712                         break
 713                 }
 714                 readsize := int64(header[7]) | (int64(header[6]) << 8) | (int64(header[5]) << 16) | (int64(header[4]) << 24)
 715                 if header[0] == 1 {
 716                         // stdout
 717                         _, err = io.CopyN(runner.Stdout, containerReader, readsize)
 718                 } else {
 719                         // stderr
 720                         _, err = io.CopyN(runner.Stderr, containerReader, readsize)
 721                 }
 722         }
 723
 724         if err != nil {
 725                 runner.CrunchLog.Printf("error reading docker logs: %v", err)
 726         }
 727
 728         err = runner.Stdout.Close()
 729         if err != nil {
 730                 runner.CrunchLog.Printf("error closing stdout logs: %v", err)
 731         }
 732
 733         err = runner.Stderr.Close()
 734         if err != nil {
 735                 runner.CrunchLog.Printf("error closing stderr logs: %v", err)
 736         }
 737
 738         if runner.statReporter != nil {
 739                 runner.statReporter.Stop()
 740                 err = runner.statLogger.Close()
 741                 if err != nil {
 742                         runner.CrunchLog.Printf("error closing crunchstat logs: %v", err)
 743                 }
 744         }
 745 }
 746
 747 func (runner *ContainerRunner) stopHoststat() error {
 748         if runner.hoststatReporter == nil {
 749                 return nil
 750         }
 751         runner.hoststatReporter.Stop()
 752         err := runner.hoststatLogger.Close()
 753         if err != nil {
 754                 return fmt.Errorf("error closing hoststat logs: %v", err)
 755         }
 756         return nil
 757 }
 758
 759 func (runner *ContainerRunner) startHoststat() error {
 760         w, err := runner.NewLogWriter("hoststat")
 761         if err != nil {
 762                 return err
 763         }
 764         runner.hoststatLogger = NewThrottledLogger(w)
 765         runner.hoststatReporter = &crunchstat.Reporter{
 766                 Logger:     log.New(runner.hoststatLogger, "", 0),
 767                 CgroupRoot: runner.cgroupRoot,
 768                 PollPeriod: runner.statInterval,
 769         }
 770         runner.hoststatReporter.Start()
 771         return nil
 772 }
 773
 774 func (runner *ContainerRunner) startCrunchstat() error {
 775         w, err := runner.NewLogWriter("crunchstat")
 776         if err != nil {
 777                 return err
 778         }
 779         runner.statLogger = NewThrottledLogger(w)
 780         runner.statReporter = &crunchstat.Reporter{
 781                 CID:          runner.ContainerID,
 782                 Logger:       log.New(runner.statLogger, "", 0),
 783                 CgroupParent: runner.expectCgroupParent,
 784                 CgroupRoot:   runner.cgroupRoot,
 785                 PollPeriod:   runner.statInterval,
 786                 TempDir:      runner.parentTemp,
 787         }
 788         runner.statReporter.Start()
 789         return nil
 790 }
 791
 792 type infoCommand struct {
 793         label string
 794         cmd   []string
 795 }
 796
 797 // LogHostInfo logs info about the current host, for debugging and
 798 // accounting purposes. Although it's logged as "node-info", this is
 799 // about the environment where crunch-run is actually running, which
 800 // might differ from what's described in the node record (see
 801 // LogNodeRecord).
 802 func (runner *ContainerRunner) LogHostInfo() (err error) {
 803         w, err := runner.NewLogWriter("node-info")
 804         if err != nil {
 805                 return
 806         }
 807
 808         commands := []infoCommand{
 809                 {
 810                         label: "Host Information",
 811                         cmd:   []string{"uname", "-a"},
 812                 },
 813                 {
 814                         label: "CPU Information",
 815                         cmd:   []string{"cat", "/proc/cpuinfo"},
 816                 },
 817                 {
 818                         label: "Memory Information",
 819                         cmd:   []string{"cat", "/proc/meminfo"},
 820                 },
 821                 {
 822                         label: "Disk Space",
 823                         cmd:   []string{"df", "-m", "/", os.TempDir()},
 824                 },
 825                 {
 826                         label: "Disk INodes",
 827                         cmd:   []string{"df", "-i", "/", os.TempDir()},
 828                 },
 829         }
 830
 831         // Run commands with informational output to be logged.
 832         for _, command := range commands {
 833                 fmt.Fprintln(w, command.label)
 834                 cmd := exec.Command(command.cmd[0], command.cmd[1:]...)
 835                 cmd.Stdout = w
 836                 cmd.Stderr = w
 837                 if err := cmd.Run(); err != nil {
 838                         err = fmt.Errorf("While running command %q: %v", command.cmd, err)
 839                         fmt.Fprintln(w, err)
 840                         return err
 841                 }
 842                 fmt.Fprintln(w, "")
 843         }
 844
 845         err = w.Close()
 846         if err != nil {
 847                 return fmt.Errorf("While closing node-info logs: %v", err)
 848         }
 849         return nil
 850 }
 851
 852 // LogContainerRecord gets and saves the raw JSON container record from the API server
 853 func (runner *ContainerRunner) LogContainerRecord() error {
 854         logged, err := runner.logAPIResponse("container", "containers", map[string]interface{}{"filters": [][]string{{"uuid", "=", runner.Container.UUID}}}, nil)
 855         if !logged && err == nil {
 856                 err = fmt.Errorf("error: no container record found for %s", runner.Container.UUID)
 857         }
 858         return err
 859 }
 860
 861 // LogNodeRecord logs the current host's InstanceType config entry (or
 862 // the arvados#node record, if running via crunch-dispatch-slurm).
 863 func (runner *ContainerRunner) LogNodeRecord() error {
 864         if it := os.Getenv("InstanceType"); it != "" {
 865                 // Dispatched via arvados-dispatch-cloud. Save
 866                 // InstanceType config fragment received from
 867                 // dispatcher on stdin.
 868                 w, err := runner.LogCollection.OpenFile("node.json", os.O_CREATE|os.O_WRONLY, 0666)
 869                 if err != nil {
 870                         return err
 871                 }
 872                 defer w.Close()
 873                 _, err = io.WriteString(w, it)
 874                 if err != nil {
 875                         return err
 876                 }
 877                 return w.Close()
 878         }
 879         // Dispatched via crunch-dispatch-slurm. Look up
 880         // apiserver's node record corresponding to
 881         // $SLURMD_NODENAME.
 882         hostname := os.Getenv("SLURMD_NODENAME")
 883         if hostname == "" {
 884                 hostname, _ = os.Hostname()
 885         }
 886         _, err := runner.logAPIResponse("node", "nodes", map[string]interface{}{"filters": [][]string{{"hostname", "=", hostname}}}, func(resp interface{}) {
 887                 // The "info" field has admin-only info when
 888                 // obtained with a privileged token, and
 889                 // should not be logged.
 890                 node, ok := resp.(map[string]interface{})
 891                 if ok {
 892                         delete(node, "info")
 893                 }
 894         })
 895         return err
 896 }
 897
 898 func (runner *ContainerRunner) logAPIResponse(label, path string, params map[string]interface{}, munge func(interface{})) (logged bool, err error) {
 899         writer, err := runner.LogCollection.OpenFile(label+".json", os.O_CREATE|os.O_WRONLY, 0666)
 900         if err != nil {
 901                 return false, err
 902         }
 903         w := &ArvLogWriter{
 904                 ArvClient:     runner.DispatcherArvClient,
 905                 UUID:          runner.Container.UUID,
 906                 loggingStream: label,
 907                 writeCloser:   writer,
 908         }
 909
 910         reader, err := runner.DispatcherArvClient.CallRaw("GET", path, "", "", arvadosclient.Dict(params))
 911         if err != nil {
 912                 return false, fmt.Errorf("error getting %s record: %v", label, err)
 913         }
 914         defer reader.Close()
 915
 916         dec := json.NewDecoder(reader)
 917         dec.UseNumber()
 918         var resp map[string]interface{}
 919         if err = dec.Decode(&resp); err != nil {
 920                 return false, fmt.Errorf("error decoding %s list response: %v", label, err)
 921         }
 922         items, ok := resp["items"].([]interface{})
 923         if !ok {
 924                 return false, fmt.Errorf("error decoding %s list response: no \"items\" key in API list response", label)
 925         } else if len(items) < 1 {
 926                 return false, nil
 927         }
 928         if munge != nil {
 929                 munge(items[0])
 930         }
 931         // Re-encode it using indentation to improve readability
 932         enc := json.NewEncoder(w)
 933         enc.SetIndent("", "    ")
 934         if err = enc.Encode(items[0]); err != nil {
 935                 return false, fmt.Errorf("error logging %s record: %v", label, err)
 936         }
 937         err = w.Close()
 938         if err != nil {
 939                 return false, fmt.Errorf("error closing %s.json in log collection: %v", label, err)
 940         }
 941         return true, nil
 942 }
 943
 944 // AttachStreams connects the docker container stdin, stdout and stderr logs
 945 // to the Arvados logger which logs to Keep and the API server logs table.
 946 func (runner *ContainerRunner) AttachStreams() (err error) {
 947
 948         runner.CrunchLog.Print("Attaching container streams")
 949
 950         // If stdin mount is provided, attach it to the docker container
 951         var stdinRdr arvados.File
 952         var stdinJSON []byte
 953         if stdinMnt, ok := runner.Container.Mounts["stdin"]; ok {
 954                 if stdinMnt.Kind == "collection" {
 955                         var stdinColl arvados.Collection
 956                         collID := stdinMnt.UUID
 957                         if collID == "" {
 958                                 collID = stdinMnt.PortableDataHash
 959                         }
 960                         err = runner.ContainerArvClient.Get("collections", collID, nil, &stdinColl)
 961                         if err != nil {
 962                                 return fmt.Errorf("While getting stdin collection: %v", err)
 963                         }
 964
 965                         stdinRdr, err = runner.ContainerKeepClient.ManifestFileReader(
 966                                 manifest.Manifest{Text: stdinColl.ManifestText},
 967                                 stdinMnt.Path)
 968                         if os.IsNotExist(err) {
 969                                 return fmt.Errorf("stdin collection path not found: %v", stdinMnt.Path)
 970                         } else if err != nil {
 971                                 return fmt.Errorf("While getting stdin collection path %v: %v", stdinMnt.Path, err)
 972                         }
 973                 } else if stdinMnt.Kind == "json" {
 974                         stdinJSON, err = json.Marshal(stdinMnt.Content)
 975                         if err != nil {
 976                                 return fmt.Errorf("While encoding stdin json data: %v", err)
 977                         }
 978                 }
 979         }
 980
 981         stdinUsed := stdinRdr != nil || len(stdinJSON) != 0
 982         response, err := runner.ContainerExecRunner.ContainerAttach(context.TODO(), runner.ContainerID,
 983                 ContainerAttachOptions{Stream: true, Stdin: stdinUsed, Stdout: true, Stderr: true})
 984         if err != nil {
 985                 return fmt.Errorf("While attaching container stdout/stderr streams: %v", err)
 986         }
 987
 988         runner.loggingDone = make(chan bool)
 989
 990         if stdoutMnt, ok := runner.Container.Mounts["stdout"]; ok {
 991                 stdoutFile, err := runner.getStdoutFile(stdoutMnt.Path)
 992                 if err != nil {
 993                         return err
 994                 }
 995                 runner.Stdout = stdoutFile
 996         } else if w, err := runner.NewLogWriter("stdout"); err != nil {
 997                 return err
 998         } else {
 999                 runner.Stdout = NewThrottledLogger(w)
1000         }
1001
1002         if stderrMnt, ok := runner.Container.Mounts["stderr"]; ok {
1003                 stderrFile, err := runner.getStdoutFile(stderrMnt.Path)
1004                 if err != nil {
1005                         return err
1006                 }
1007                 runner.Stderr = stderrFile
1008         } else if w, err := runner.NewLogWriter("stderr"); err != nil {
1009                 return err
1010         } else {
1011                 runner.Stderr = NewThrottledLogger(w)
1012         }
1013
1014         if stdinRdr != nil {
1015                 go func() {
1016                         _, err := io.Copy(response.Conn, stdinRdr)
1017                         if err != nil {
1018                                 runner.CrunchLog.Printf("While writing stdin collection to docker container: %v", err)
1019                                 runner.stop(nil)
1020                         }
1021                         stdinRdr.Close()
1022                         response.CloseWrite()
1023                 }()
1024         } else if len(stdinJSON) != 0 {
1025                 go func() {
1026                         _, err := io.Copy(response.Conn, bytes.NewReader(stdinJSON))
1027                         if err != nil {
1028                                 runner.CrunchLog.Printf("While writing stdin json to docker container: %v", err)
1029                                 runner.stop(nil)
1030                         }
1031                         response.CloseWrite()
1032                 }()
1033         }
1034
1035         go runner.ProcessDockerAttach(response.Reader)
1036
1037         return nil
1038 }
1039
1040 func (runner *ContainerRunner) getStdoutFile(mntPath string) (*os.File, error) {
1041         stdoutPath := mntPath[len(runner.Container.OutputPath):]
1042         index := strings.LastIndex(stdoutPath, "/")
1043         if index > 0 {
1044                 subdirs := stdoutPath[:index]
1045                 if subdirs != "" {
1046                         st, err := os.Stat(runner.HostOutputDir)
1047                         if err != nil {
1048                                 return nil, fmt.Errorf("While Stat on temp dir: %v", err)
1049                         }
1050                         stdoutPath := filepath.Join(runner.HostOutputDir, subdirs)
1051                         err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777)
1052                         if err != nil {
1053                                 return nil, fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err)
1054                         }
1055                 }
1056         }
1057         stdoutFile, err := os.Create(filepath.Join(runner.HostOutputDir, stdoutPath))
1058         if err != nil {
1059                 return nil, fmt.Errorf("While creating file %q: %v", stdoutPath, err)
1060         }
1061
1062         return stdoutFile, nil
1063 }
1064
1065 // CreateContainer creates the docker container.
1066 func (runner *ContainerRunner) CreateContainer() error {
1067         runner.CrunchLog.Print("Creating Docker container")
1068
1069         containerConfig, err := runner.ContainerExecRunner.GetContainerConfig()
1070         hostConfig, err := runner.ContainerExecRunner.GetHostConfig()
1071
1072         containerConfig.Cmd = runner.Container.Command
1073         if runner.Container.Cwd != "." {
1074                 containerConfig.WorkingDir = runner.Container.Cwd
1075         }
1076
1077         for k, v := range runner.Container.Environment {
1078                 containerConfig.Env = append(containerConfig.Env, k+"="+v)
1079         }
1080
1081         containerConfig.Volumes = runner.Volumes
1082
1083         maxRAM := int64(runner.Container.RuntimeConstraints.RAM)
1084         minDockerRAM := int64(16)
1085         if maxRAM < minDockerRAM*1024*1024 {
1086                 // Docker daemon won't let you set a limit less than ~10 MiB
1087                 maxRAM = minDockerRAM * 1024 * 1024
1088         }
1089         runner.HostConfig = dockercontainer.HostConfig{
1090                 Binds: runner.Binds,
1091                 LogConfig: dockercontainer.LogConfig{
1092                         Type: "none",
1093                 },
1094                 Resources: dockercontainer.Resources{
1095                         CgroupParent: runner.setCgroupParent,
1096                         NanoCPUs:     int64(runner.Container.RuntimeConstraints.VCPUs) * 1000000000,
1097                         Memory:       maxRAM, // RAM
1098                         MemorySwap:   maxRAM, // RAM+swap
1099                         KernelMemory: maxRAM, // kernel portion
1100                 },
1101         }
1102
1103         if runner.Container.RuntimeConstraints.API {
1104                 tok, err := runner.ContainerToken()
1105                 if err != nil {
1106                         return err
1107                 }
1108                 containerConfig.Env = append(containerConfig.Env,
1109                         "ARVADOS_API_TOKEN="+tok,
1110                         "ARVADOS_API_HOST="+os.Getenv("ARVADOS_API_HOST"),
1111                         "ARVADOS_API_HOST_INSECURE="+os.Getenv("ARVADOS_API_HOST_INSECURE"),
1112                 )
1113                 runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
1114         } else {
1115                 if runner.enableNetwork == "always" {
1116                         runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
1117                 } else {
1118                         runner.HostConfig.NetworkMode = dockercontainer.NetworkMode("none")
1119                 }
1120         }
1121
1122         _, stdinUsed := runner.Container.Mounts["stdin"]
1123         containerConfig.OpenStdin = stdinUsed
1124         containerConfig.StdinOnce = stdinUsed
1125         containerConfig.AttachStdin = stdinUsed
1126         containerConfig.AttachStdout = true
1127         containerConfig.AttachStderr = true
1128
1129         createdBody, err := runner.ContainerExecRunner.ContainerCreate(context.TODO(), containerConfig, hostConfig, nil, runner.Container.UUID)
1130         if err != nil {
1131                 return fmt.Errorf("While creating container: %v", err)
1132         }
1133
1134         runner.ContainerID = createdBody.ID
1135
1136         return runner.AttachStreams()
1137 }
1138
1139 // StartContainer starts the docker container created by CreateContainer.
1140 func (runner *ContainerRunner) StartContainer() error {
1141         runner.CrunchLog.Printf("Starting Docker container id '%s'", runner.ContainerID)
1142         runner.cStateLock.Lock()
1143         defer runner.cStateLock.Unlock()
1144         if runner.cCancelled {
1145                 return ErrCancelled
1146         }
1147         err := runner.ContainerExecRunner.ContainerStart(context.TODO(), runner.ContainerID,
1148                 ContainerStartOptions{})
1149         if err != nil {
1150                 var advice string
1151                 if m, e := regexp.MatchString("(?ms).*(exec|System error).*(no such file or directory|file not found).*", err.Error()); m && e == nil {
1152                         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])
1153                 }
1154                 return fmt.Errorf("could not start container: %v%s", err, advice)
1155         }
1156         return nil
1157 }
1158
1159 // WaitFinish waits for the container to terminate, capture the exit code, and
1160 // close the stdout/stderr logging.
1161 func (runner *ContainerRunner) WaitFinish() error {
1162         var runTimeExceeded <-chan time.Time
1163         runner.CrunchLog.Print("Waiting for container to finish")
1164
1165         waitOk, waitErr := runner.ContainerExecRunner.ContainerWait(context.TODO(), runner.ContainerID, WaitConditionNotRunning)
1166         arvMountExit := runner.ArvMountExit
1167         if timeout := runner.Container.SchedulingParameters.MaxRunTime; timeout > 0 {
1168                 runTimeExceeded = time.After(time.Duration(timeout) * time.Second)
1169         }
1170
1171         containerGone := make(chan struct{})
1172         go func() {
1173                 defer close(containerGone)
1174                 if runner.containerWatchdogInterval < 1 {
1175                         runner.containerWatchdogInterval = time.Minute
1176                 }
1177                 for range time.NewTicker(runner.containerWatchdogInterval).C {
1178                         ctx, cancel := context.WithDeadline(context.Background(), time.Now().Add(runner.containerWatchdogInterval))
1179                         ctr, err := runner.ContainerExecRunner.ContainerInspect(ctx, runner.ContainerID)
1180                         cancel()
1181                         runner.cStateLock.Lock()
1182                         done := runner.cRemoved || runner.ExitCode != nil
1183                         runner.cStateLock.Unlock()
1184                         if done {
1185                                 return
1186                         } else if err != nil {
1187                                 runner.CrunchLog.Printf("Error inspecting container: %s", err)
1188                                 runner.checkBrokenNode(err)
1189                                 return
1190                         } else if ctr.State == nil || !(ctr.State.Running || ctr.State.Status == "created") {
1191                                 runner.CrunchLog.Printf("Container is not running: State=%v", ctr.State)
1192                                 return
1193                         }
1194                 }
1195         }()
1196
1197         for {
1198                 select {
1199                 case waitBody := <-waitOk:
1200                         runner.CrunchLog.Printf("Container exited with code: %v", waitBody.StatusCode)
1201                         code := int(waitBody.StatusCode)
1202                         runner.ExitCode = &code
1203
1204                         // wait for stdout/stderr to complete
1205                         <-runner.loggingDone
1206                         return nil
1207
1208                 case err := <-waitErr:
1209                         return fmt.Errorf("container wait: %v", err)
1210
1211                 case <-arvMountExit:
1212                         runner.CrunchLog.Printf("arv-mount exited while container is still running.  Stopping container.")
1213                         runner.stop(nil)
1214                         // arvMountExit will always be ready now that
1215                         // it's closed, but that doesn't interest us.
1216                         arvMountExit = nil
1217
1218                 case <-runTimeExceeded:
1219                         runner.CrunchLog.Printf("maximum run time exceeded. Stopping container.")
1220                         runner.stop(nil)
1221                         runTimeExceeded = nil
1222
1223                 case <-containerGone:
1224                         return errors.New("docker client never returned status")
1225                 }
1226         }
1227 }
1228
1229 func (runner *ContainerRunner) updateLogs() {
1230         ticker := time.NewTicker(crunchLogUpdatePeriod / 360)
1231         defer ticker.Stop()
1232
1233         sigusr1 := make(chan os.Signal, 1)
1234         signal.Notify(sigusr1, syscall.SIGUSR1)
1235         defer signal.Stop(sigusr1)
1236
1237         saveAtTime := time.Now().Add(crunchLogUpdatePeriod)
1238         saveAtSize := crunchLogUpdateSize
1239         var savedSize int64
1240         for {
1241                 select {
1242                 case <-ticker.C:
1243                 case <-sigusr1:
1244                         saveAtTime = time.Now()
1245                 }
1246                 runner.logMtx.Lock()
1247                 done := runner.LogsPDH != nil
1248                 runner.logMtx.Unlock()
1249                 if done {
1250                         return
1251                 }
1252                 size := runner.LogCollection.Size()
1253                 if size == savedSize || (time.Now().Before(saveAtTime) && size < saveAtSize) {
1254                         continue
1255                 }
1256                 saveAtTime = time.Now().Add(crunchLogUpdatePeriod)
1257                 saveAtSize = runner.LogCollection.Size() + crunchLogUpdateSize
1258                 saved, err := runner.saveLogCollection(false)
1259                 if err != nil {
1260                         runner.CrunchLog.Printf("error updating log collection: %s", err)
1261                         continue
1262                 }
1263
1264                 var updated arvados.Container
1265                 err = runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1266                         "container": arvadosclient.Dict{"log": saved.PortableDataHash},
1267                 }, &updated)
1268                 if err != nil {
1269                         runner.CrunchLog.Printf("error updating container log to %s: %s", saved.PortableDataHash, err)
1270                         continue
1271                 }
1272
1273                 savedSize = size
1274         }
1275 }
1276
1277 // CaptureOutput saves data from the container's output directory if
1278 // needed, and updates the container output accordingly.
1279 func (runner *ContainerRunner) CaptureOutput() error {
1280         if runner.Container.RuntimeConstraints.API {
1281                 // Output may have been set directly by the container, so
1282                 // refresh the container record to check.
1283                 err := runner.DispatcherArvClient.Get("containers", runner.Container.UUID,
1284                         nil, &runner.Container)
1285                 if err != nil {
1286                         return err
1287                 }
1288                 if runner.Container.Output != "" {
1289                         // Container output is already set.
1290                         runner.OutputPDH = &runner.Container.Output
1291                         return nil
1292                 }
1293         }
1294
1295         txt, err := (&copier{
1296                 client:        runner.containerClient,
1297                 arvClient:     runner.ContainerArvClient,
1298                 keepClient:    runner.ContainerKeepClient,
1299                 hostOutputDir: runner.HostOutputDir,
1300                 ctrOutputDir:  runner.Container.OutputPath,
1301                 binds:         runner.Binds,
1302                 mounts:        runner.Container.Mounts,
1303                 secretMounts:  runner.SecretMounts,
1304                 logger:        runner.CrunchLog,
1305         }).Copy()
1306         if err != nil {
1307                 return err
1308         }
1309         if n := len(regexp.MustCompile(` [0-9a-f]+\+\S*\+R`).FindAllStringIndex(txt, -1)); n > 0 {
1310                 runner.CrunchLog.Printf("Copying %d data blocks from remote input collections...", n)
1311                 fs, err := (&arvados.Collection{ManifestText: txt}).FileSystem(runner.containerClient, runner.ContainerKeepClient)
1312                 if err != nil {
1313                         return err
1314                 }
1315                 txt, err = fs.MarshalManifest(".")
1316                 if err != nil {
1317                         return err
1318                 }
1319         }
1320         var resp arvados.Collection
1321         err = runner.ContainerArvClient.Create("collections", arvadosclient.Dict{
1322                 "ensure_unique_name": true,
1323                 "collection": arvadosclient.Dict{
1324                         "is_trashed":    true,
1325                         "name":          "output for " + runner.Container.UUID,
1326                         "manifest_text": txt,
1327                 },
1328         }, &resp)
1329         if err != nil {
1330                 return fmt.Errorf("error creating output collection: %v", err)
1331         }
1332         runner.OutputPDH = &resp.PortableDataHash
1333         return nil
1334 }
1335
1336 func (runner *ContainerRunner) CleanupDirs() {
1337         if runner.ArvMount != nil {
1338                 var delay int64 = 8
1339                 umount := exec.Command("arv-mount", fmt.Sprintf("--unmount-timeout=%d", delay), "--unmount", runner.ArvMountPoint)
1340                 umount.Stdout = runner.CrunchLog
1341                 umount.Stderr = runner.CrunchLog
1342                 runner.CrunchLog.Printf("Running %v", umount.Args)
1343                 umnterr := umount.Start()
1344
1345                 if umnterr != nil {
1346                         runner.CrunchLog.Printf("Error unmounting: %v", umnterr)
1347                 } else {
1348                         // If arv-mount --unmount gets stuck for any reason, we
1349                         // don't want to wait for it forever.  Do Wait() in a goroutine
1350                         // so it doesn't block crunch-run.
1351                         umountExit := make(chan error)
1352                         go func() {
1353                                 mnterr := umount.Wait()
1354                                 if mnterr != nil {
1355                                         runner.CrunchLog.Printf("Error unmounting: %v", mnterr)
1356                                 }
1357                                 umountExit <- mnterr
1358                         }()
1359
1360                         for again := true; again; {
1361                                 again = false
1362                                 select {
1363                                 case <-umountExit:
1364                                         umount = nil
1365                                         again = true
1366                                 case <-runner.ArvMountExit:
1367                                         break
1368                                 case <-time.After(time.Duration((delay + 1) * int64(time.Second))):
1369                                         runner.CrunchLog.Printf("Timed out waiting for unmount")
1370                                         if umount != nil {
1371                                                 umount.Process.Kill()
1372                                         }
1373                                         runner.ArvMount.Process.Kill()
1374                                 }
1375                         }
1376                 }
1377         }
1378
1379         if runner.ArvMountPoint != "" {
1380                 if rmerr := os.Remove(runner.ArvMountPoint); rmerr != nil {
1381                         runner.CrunchLog.Printf("While cleaning up arv-mount directory %s: %v", runner.ArvMountPoint, rmerr)
1382                 }
1383         }
1384
1385         if rmerr := os.RemoveAll(runner.parentTemp); rmerr != nil {
1386                 runner.CrunchLog.Printf("While cleaning up temporary directory %s: %v", runner.parentTemp, rmerr)
1387         }
1388 }
1389
1390 // CommitLogs posts the collection containing the final container logs.
1391 func (runner *ContainerRunner) CommitLogs() error {
1392         func() {
1393                 // Hold cStateLock to prevent races on CrunchLog (e.g., stop()).
1394                 runner.cStateLock.Lock()
1395                 defer runner.cStateLock.Unlock()
1396
1397                 runner.CrunchLog.Print(runner.finalState)
1398
1399                 if runner.arvMountLog != nil {
1400                         runner.arvMountLog.Close()
1401                 }
1402                 runner.CrunchLog.Close()
1403
1404                 // Closing CrunchLog above allows them to be committed to Keep at this
1405                 // point, but re-open crunch log with ArvClient in case there are any
1406                 // other further errors (such as failing to write the log to Keep!)
1407                 // while shutting down
1408                 runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{
1409                         ArvClient:     runner.DispatcherArvClient,
1410                         UUID:          runner.Container.UUID,
1411                         loggingStream: "crunch-run",
1412                         writeCloser:   nil,
1413                 })
1414                 runner.CrunchLog.Immediate = log.New(os.Stderr, runner.Container.UUID+" ", 0)
1415         }()
1416
1417         if runner.LogsPDH != nil {
1418                 // If we have already assigned something to LogsPDH,
1419                 // we must be closing the re-opened log, which won't
1420                 // end up getting attached to the container record and
1421                 // therefore doesn't need to be saved as a collection
1422                 // -- it exists only to send logs to other channels.
1423                 return nil
1424         }
1425         saved, err := runner.saveLogCollection(true)
1426         if err != nil {
1427                 return fmt.Errorf("error saving log collection: %s", err)
1428         }
1429         runner.logMtx.Lock()
1430         defer runner.logMtx.Unlock()
1431         runner.LogsPDH = &saved.PortableDataHash
1432         return nil
1433 }
1434
1435 func (runner *ContainerRunner) saveLogCollection(final bool) (response arvados.Collection, err error) {
1436         runner.logMtx.Lock()
1437         defer runner.logMtx.Unlock()
1438         if runner.LogsPDH != nil {
1439                 // Already finalized.
1440                 return
1441         }
1442         updates := arvadosclient.Dict{
1443                 "name": "logs for " + runner.Container.UUID,
1444         }
1445         mt, err1 := runner.LogCollection.MarshalManifest(".")
1446         if err1 == nil {
1447                 // Only send updated manifest text if there was no
1448                 // error.
1449                 updates["manifest_text"] = mt
1450         }
1451
1452         // Even if flushing the manifest had an error, we still want
1453         // to update the log record, if possible, to push the trash_at
1454         // and delete_at times into the future.  Details on bug
1455         // #17293.
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         var err2 error
1465         if runner.logUUID == "" {
1466                 reqBody["ensure_unique_name"] = true
1467                 err2 = runner.DispatcherArvClient.Create("collections", reqBody, &response)
1468         } else {
1469                 err2 = runner.DispatcherArvClient.Update("collections", runner.logUUID, reqBody, &response)
1470         }
1471         if err2 == nil {
1472                 runner.logUUID = response.UUID
1473         }
1474
1475         if err1 != nil || err2 != nil {
1476                 err = fmt.Errorf("error recording logs: %q, %q", err1, err2)
1477         }
1478         return
1479 }
1480
1481 // UpdateContainerRunning updates the container state to "Running"
1482 func (runner *ContainerRunner) UpdateContainerRunning() error {
1483         runner.cStateLock.Lock()
1484         defer runner.cStateLock.Unlock()
1485         if runner.cCancelled {
1486                 return ErrCancelled
1487         }
1488         return runner.DispatcherArvClient.Update("containers", runner.Container.UUID,
1489                 arvadosclient.Dict{"container": arvadosclient.Dict{"state": "Running", "gateway_address": runner.gateway.Address}}, nil)
1490 }
1491
1492 // ContainerToken returns the api_token the container (and any
1493 // arv-mount processes) are allowed to use.
1494 func (runner *ContainerRunner) ContainerToken() (string, error) {
1495         if runner.token != "" {
1496                 return runner.token, nil
1497         }
1498
1499         var auth arvados.APIClientAuthorization
1500         err := runner.DispatcherArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth)
1501         if err != nil {
1502                 return "", err
1503         }
1504         runner.token = fmt.Sprintf("v2/%s/%s/%s", auth.UUID, auth.APIToken, runner.Container.UUID)
1505         return runner.token, nil
1506 }
1507
1508 // UpdateContainerFinal updates the container record state on API
1509 // server to "Complete" or "Cancelled"
1510 func (runner *ContainerRunner) UpdateContainerFinal() error {
1511         update := arvadosclient.Dict{}
1512         update["state"] = runner.finalState
1513         if runner.LogsPDH != nil {
1514                 update["log"] = *runner.LogsPDH
1515         }
1516         if runner.finalState == "Complete" {
1517                 if runner.ExitCode != nil {
1518                         update["exit_code"] = *runner.ExitCode
1519                 }
1520                 if runner.OutputPDH != nil {
1521                         update["output"] = *runner.OutputPDH
1522                 }
1523         }
1524         return runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{"container": update}, nil)
1525 }
1526
1527 // IsCancelled returns the value of Cancelled, with goroutine safety.
1528 func (runner *ContainerRunner) IsCancelled() bool {
1529         runner.cStateLock.Lock()
1530         defer runner.cStateLock.Unlock()
1531         return runner.cCancelled
1532 }
1533
1534 // NewArvLogWriter creates an ArvLogWriter
1535 func (runner *ContainerRunner) NewArvLogWriter(name string) (io.WriteCloser, error) {
1536         writer, err := runner.LogCollection.OpenFile(name+".txt", os.O_CREATE|os.O_WRONLY, 0666)
1537         if err != nil {
1538                 return nil, err
1539         }
1540         return &ArvLogWriter{
1541                 ArvClient:     runner.DispatcherArvClient,
1542                 UUID:          runner.Container.UUID,
1543                 loggingStream: name,
1544                 writeCloser:   writer,
1545         }, nil
1546 }
1547
1548 // Run the full container lifecycle.
1549 func (runner *ContainerRunner) Run() (err error) {
1550         runner.CrunchLog.Printf("crunch-run %s started", cmd.Version.String())
1551         runner.CrunchLog.Printf("Executing container '%s'", runner.Container.UUID)
1552
1553         hostname, hosterr := os.Hostname()
1554         if hosterr != nil {
1555                 runner.CrunchLog.Printf("Error getting hostname '%v'", hosterr)
1556         } else {
1557                 runner.CrunchLog.Printf("Executing on host '%s'", hostname)
1558         }
1559
1560         runner.finalState = "Queued"
1561
1562         defer func() {
1563                 runner.CleanupDirs()
1564
1565                 runner.CrunchLog.Printf("crunch-run finished")
1566                 runner.CrunchLog.Close()
1567         }()
1568
1569         err = runner.fetchContainerRecord()
1570         if err != nil {
1571                 return
1572         }
1573         if runner.Container.State != "Locked" {
1574                 return fmt.Errorf("dispatch error detected: container %q has state %q", runner.Container.UUID, runner.Container.State)
1575         }
1576
1577         defer func() {
1578                 // checkErr prints e (unless it's nil) and sets err to
1579                 // e (unless err is already non-nil). Thus, if err
1580                 // hasn't already been assigned when Run() returns,
1581                 // this cleanup func will cause Run() to return the
1582                 // first non-nil error that is passed to checkErr().
1583                 checkErr := func(errorIn string, e error) {
1584                         if e == nil {
1585                                 return
1586                         }
1587                         runner.CrunchLog.Printf("error in %s: %v", errorIn, e)
1588                         if err == nil {
1589                                 err = e
1590                         }
1591                         if runner.finalState == "Complete" {
1592                                 // There was an error in the finalization.
1593                                 runner.finalState = "Cancelled"
1594                         }
1595                 }
1596
1597                 // Log the error encountered in Run(), if any
1598                 checkErr("Run", err)
1599
1600                 if runner.finalState == "Queued" {
1601                         runner.UpdateContainerFinal()
1602                         return
1603                 }
1604
1605                 if runner.IsCancelled() {
1606                         runner.finalState = "Cancelled"
1607                         // but don't return yet -- we still want to
1608                         // capture partial output and write logs
1609                 }
1610
1611                 checkErr("CaptureOutput", runner.CaptureOutput())
1612                 checkErr("stopHoststat", runner.stopHoststat())
1613                 checkErr("CommitLogs", runner.CommitLogs())
1614                 checkErr("UpdateContainerFinal", runner.UpdateContainerFinal())
1615         }()
1616
1617         runner.setupSignals()
1618         err = runner.startHoststat()
1619         if err != nil {
1620                 return
1621         }
1622
1623         // check for and/or load image
1624         err = runner.LoadImage()
1625         if err != nil {
1626                 if !runner.checkBrokenNode(err) {
1627                         // Failed to load image but not due to a "broken node"
1628                         // condition, probably user error.
1629                         runner.finalState = "Cancelled"
1630                 }
1631                 err = fmt.Errorf("While loading container image: %v", err)
1632                 return
1633         }
1634
1635         // set up FUSE mount and binds
1636         err = runner.SetupMounts()
1637         if err != nil {
1638                 runner.finalState = "Cancelled"
1639                 err = fmt.Errorf("While setting up mounts: %v", err)
1640                 return
1641         }
1642
1643         err = runner.CreateContainer()
1644         if err != nil {
1645                 return
1646         }
1647         err = runner.LogHostInfo()
1648         if err != nil {
1649                 return
1650         }
1651         err = runner.LogNodeRecord()
1652         if err != nil {
1653                 return
1654         }
1655         err = runner.LogContainerRecord()
1656         if err != nil {
1657                 return
1658         }
1659
1660         if runner.IsCancelled() {
1661                 return
1662         }
1663
1664         err = runner.UpdateContainerRunning()
1665         if err != nil {
1666                 return
1667         }
1668         runner.finalState = "Cancelled"
1669
1670         err = runner.startCrunchstat()
1671         if err != nil {
1672                 return
1673         }
1674
1675         err = runner.StartContainer()
1676         if err != nil {
1677                 runner.checkBrokenNode(err)
1678                 return
1679         }
1680
1681         err = runner.WaitFinish()
1682         if err == nil && !runner.IsCancelled() {
1683                 runner.finalState = "Complete"
1684         }
1685         return
1686 }
1687
1688 // Fetch the current container record (uuid = runner.Container.UUID)
1689 // into runner.Container.
1690 func (runner *ContainerRunner) fetchContainerRecord() error {
1691         reader, err := runner.DispatcherArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
1692         if err != nil {
1693                 return fmt.Errorf("error fetching container record: %v", err)
1694         }
1695         defer reader.Close()
1696
1697         dec := json.NewDecoder(reader)
1698         dec.UseNumber()
1699         err = dec.Decode(&runner.Container)
1700         if err != nil {
1701                 return fmt.Errorf("error decoding container record: %v", err)
1702         }
1703
1704         var sm struct {
1705                 SecretMounts map[string]arvados.Mount `json:"secret_mounts"`
1706         }
1707
1708         containerToken, err := runner.ContainerToken()
1709         if err != nil {
1710                 return fmt.Errorf("error getting container token: %v", err)
1711         }
1712
1713         runner.ContainerArvClient, runner.ContainerKeepClient,
1714                 runner.containerClient, err = runner.MkArvClient(containerToken)
1715         if err != nil {
1716                 return fmt.Errorf("error creating container API client: %v", err)
1717         }
1718
1719         err = runner.ContainerArvClient.Call("GET", "containers", runner.Container.UUID, "secret_mounts", nil, &sm)
1720         if err != nil {
1721                 if apierr, ok := err.(arvadosclient.APIServerError); !ok || apierr.HttpStatusCode != 404 {
1722                         return fmt.Errorf("error fetching secret_mounts: %v", err)
1723                 }
1724                 // ok && apierr.HttpStatusCode == 404, which means
1725                 // secret_mounts isn't supported by this API server.
1726         }
1727         runner.SecretMounts = sm.SecretMounts
1728
1729         return nil
1730 }
1731
1732 // NewContainerRunner creates a new container runner.
1733 func NewContainerRunner(dispatcherClient *arvados.Client,
1734         dispatcherArvClient IArvadosClient,
1735         dispatcherKeepClient IKeepClient,
1736         containerRunner ThinContainerExecRunner,
1737         containerUUID string) (*ContainerRunner, error) {
1738
1739         cr := &ContainerRunner{
1740                 dispatcherClient:     dispatcherClient,
1741                 DispatcherArvClient:  dispatcherArvClient,
1742                 DispatcherKeepClient: dispatcherKeepClient,
1743                 ContainerExecRunner:  containerRunner,
1744         }
1745         cr.NewLogWriter = cr.NewArvLogWriter
1746         cr.RunArvMount = cr.ArvMountCmd
1747         cr.MkTempDir = ioutil.TempDir
1748         cr.MkArvClient = func(token string) (IArvadosClient, IKeepClient, *arvados.Client, error) {
1749                 cl, err := arvadosclient.MakeArvadosClient()
1750                 if err != nil {
1751                         return nil, nil, nil, err
1752                 }
1753                 cl.ApiToken = token
1754                 kc, err := keepclient.MakeKeepClient(cl)
1755                 if err != nil {
1756                         return nil, nil, nil, err
1757                 }
1758                 c2 := arvados.NewClientFromEnv()
1759                 c2.AuthToken = token
1760                 return cl, kc, c2, nil
1761         }
1762         var err error
1763         cr.LogCollection, err = (&arvados.Collection{}).FileSystem(cr.dispatcherClient, cr.DispatcherKeepClient)
1764         if err != nil {
1765                 return nil, err
1766         }
1767         cr.Container.UUID = containerUUID
1768         w, err := cr.NewLogWriter("crunch-run")
1769         if err != nil {
1770                 return nil, err
1771         }
1772         cr.CrunchLog = NewThrottledLogger(w)
1773         cr.CrunchLog.Immediate = log.New(os.Stderr, containerUUID+" ", 0)
1774
1775         loadLogThrottleParams(dispatcherArvClient)
1776         go cr.updateLogs()
1777
1778         return cr, nil
1779 }
1780
1781 func (command) RunCommand(prog string, args []string, stdin io.Reader, stdout, stderr io.Writer) int {
1782         flags := flag.NewFlagSet(prog, flag.ContinueOnError)
1783         statInterval := flags.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting")
1784         cgroupRoot := flags.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree")
1785         cgroupParent := flags.String("cgroup-parent", "docker", "name of container's parent cgroup (ignored if -cgroup-parent-subsystem is used)")
1786         cgroupParentSubsystem := flags.String("cgroup-parent-subsystem", "", "use current cgroup for given subsystem as parent cgroup for container")
1787         caCertsPath := flags.String("ca-certs", "", "Path to TLS root certificates")
1788         detach := flags.Bool("detach", false, "Detach from parent process and run in the background")
1789         stdinEnv := flags.Bool("stdin-env", false, "Load environment variables from JSON message on stdin")
1790         sleep := flags.Duration("sleep", 0, "Delay before starting (testing use only)")
1791         kill := flags.Int("kill", -1, "Send signal to an existing crunch-run process for given UUID")
1792         list := flags.Bool("list", false, "List UUIDs of existing crunch-run processes")
1793         enableNetwork := flags.String("container-enable-networking", "default",
1794                 `Specify if networking should be enabled for container.  One of 'default', 'always':
1795         default: only enable networking if container requests it.
1796         always:  containers always have networking enabled
1797         `)
1798         networkMode := flags.String("container-network-mode", "default",
1799                 `Set networking mode for container.  Corresponds to Docker network mode (--net).
1800         `)
1801         memprofile := flags.String("memprofile", "", "write memory profile to `file` after running container")
1802         flags.Duration("check-containerd", 0, "Ignored. Exists for compatibility with older versions.")
1803         containerRunner := flags.String("container-runner", "docker",
1804                 `Specify the container runner. available options: docker, singularity.
1805         `)
1806         ignoreDetachFlag := false
1807         if len(args) > 0 && args[0] == "-no-detach" {
1808                 // This process was invoked by a parent process, which
1809                 // has passed along its own arguments, including
1810                 // -detach, after the leading -no-detach flag.  Strip
1811                 // the leading -no-detach flag (it's not recognized by
1812                 // flags.Parse()) and ignore the -detach flag that
1813                 // comes later.
1814                 args = args[1:]
1815                 ignoreDetachFlag = true
1816         }
1817
1818         if err := flags.Parse(args); err == flag.ErrHelp {
1819                 return 0
1820         } else if err != nil {
1821                 log.Print(err)
1822                 return 1
1823         }
1824
1825         if *stdinEnv && !ignoreDetachFlag {
1826                 // Load env vars on stdin if asked (but not in a
1827                 // detached child process, in which case stdin is
1828                 // /dev/null).
1829                 err := loadEnv(os.Stdin)
1830                 if err != nil {
1831                         log.Print(err)
1832                         return 1
1833                 }
1834         }
1835
1836         containerID := flags.Arg(0)
1837
1838         switch {
1839         case *detach && !ignoreDetachFlag:
1840                 return Detach(containerID, prog, args, os.Stdout, os.Stderr)
1841         case *kill >= 0:
1842                 return KillProcess(containerID, syscall.Signal(*kill), os.Stdout, os.Stderr)
1843         case *list:
1844                 return ListProcesses(os.Stdout, os.Stderr)
1845         }
1846
1847         if containerID == "" {
1848                 log.Printf("usage: %s [options] UUID", prog)
1849                 return 1
1850         }
1851
1852         log.Printf("crunch-run %s started", cmd.Version.String())
1853         time.Sleep(*sleep)
1854
1855         if *caCertsPath != "" {
1856                 arvadosclient.CertFiles = []string{*caCertsPath}
1857         }
1858
1859         api, err := arvadosclient.MakeArvadosClient()
1860         if err != nil {
1861                 log.Printf("%s: %v", containerID, err)
1862                 return 1
1863         }
1864         api.Retries = 8
1865
1866         kc, kcerr := keepclient.MakeKeepClient(api)
1867         if kcerr != nil {
1868                 log.Printf("%s: %v", containerID, kcerr)
1869                 return 1
1870         }
1871         kc.BlockCache = &keepclient.BlockCache{MaxBlocks: 2}
1872         kc.Retries = 4
1873
1874         var cr *ContainerRunner
1875         if *containerRunner == "docker" {
1876                 // API version 1.21 corresponds to Docker 1.9, which is currently the
1877                 // minimum version we want to support.
1878                 docker, dockererr := dockerclient.NewClient(dockerclient.DefaultDockerHost, "1.21", nil, nil)
1879
1880                 cr, err = NewContainerRunner(arvados.NewClientFromEnv(), api, kc, adapter(docker), containerID)
1881                 if err != nil {
1882                         log.Print(err)
1883                         return 1
1884                 }
1885                 if dockererr != nil {
1886                         cr.CrunchLog.Printf("%s: %v", containerID, dockererr)
1887                         cr.checkBrokenNode(dockererr)
1888                         cr.CrunchLog.Close()
1889                         return 1
1890                 }
1891         } else {
1892                 // Singularity
1893
1894                 singularity, singularityerr := NewSingularityClient()
1895
1896                 cr, err = NewContainerRunner(arvados.NewClientFromEnv(), api, kc, singularity, containerID)
1897                 if err != nil {
1898                         log.Print(err)
1899                         return 1
1900                 }
1901
1902                 if singularityerr != nil {
1903                         cr.CrunchLog.Printf("%s: %v", containerID, singularityerr)
1904                         //cr.checkBrokenNode(singularityrerr) //
1905                         cr.CrunchLog.Close()
1906                         return 1
1907                 }
1908         }
1909         cr.gateway = Gateway{
1910                 Address:           os.Getenv("GatewayAddress"),
1911                 AuthSecret:        os.Getenv("GatewayAuthSecret"),
1912                 ContainerUUID:     containerID,
1913                 DockerContainerID: &cr.ContainerID,
1914                 Log:               cr.CrunchLog,
1915         }
1916         os.Unsetenv("GatewayAuthSecret")
1917         if cr.gateway.Address != "" {
1918                 err = cr.gateway.Start()
1919                 if err != nil {
1920                         log.Printf("error starting gateway server: %s", err)
1921                         return 1
1922                 }
1923         }
1924
1925         parentTemp, tmperr := cr.MkTempDir("", "crunch-run."+containerID+".")
1926         if tmperr != nil {
1927                 log.Printf("%s: %v", containerID, tmperr)
1928                 return 1
1929         }
1930
1931         cr.parentTemp = parentTemp
1932         cr.statInterval = *statInterval
1933         cr.cgroupRoot = *cgroupRoot
1934         cr.expectCgroupParent = *cgroupParent
1935         cr.enableNetwork = *enableNetwork
1936         cr.networkMode = *networkMode
1937         if *cgroupParentSubsystem != "" {
1938                 p := findCgroup(*cgroupParentSubsystem)
1939                 cr.setCgroupParent = p
1940                 cr.expectCgroupParent = p
1941         }
1942
1943         runerr := cr.Run()
1944
1945         if *memprofile != "" {
1946                 f, err := os.Create(*memprofile)
1947                 if err != nil {
1948                         log.Printf("could not create memory profile: %s", err)
1949                 }
1950                 runtime.GC() // get up-to-date statistics
1951                 if err := pprof.WriteHeapProfile(f); err != nil {
1952                         log.Printf("could not write memory profile: %s", err)
1953                 }
1954                 closeerr := f.Close()
1955                 if closeerr != nil {
1956                         log.Printf("closing memprofile file: %s", err)
1957                 }
1958         }
1959
1960         if runerr != nil {
1961                 log.Printf("%s: %v", containerID, runerr)
1962                 return 1
1963         }
1964         return 0
1965 }
1966
1967 func loadEnv(rdr io.Reader) error {
1968         buf, err := ioutil.ReadAll(rdr)
1969         if err != nil {
1970                 return fmt.Errorf("read stdin: %s", err)
1971         }
1972         var env map[string]string
1973         err = json.Unmarshal(buf, &env)
1974         if err != nil {
1975                 return fmt.Errorf("decode stdin: %s", err)
1976         }
1977         for k, v := range env {
1978                 err = os.Setenv(k, v)
1979                 if err != nil {
1980                         return fmt.Errorf("setenv(%q): %s", k, err)
1981                 }
1982         }
1983         return nil
1984 }