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         Docker ThinDockerClient
  99
 100         // Dispatcher client is initialized with the Dispatcher token.
 101         // This is a privileged token used to manage container status
 102         // and logs.
 103         //
 104         // We have both dispatcherClient and DispatcherArvClient
 105         // because there are two different incompatible Arvados Go
 106         // SDKs and we have to use both (hopefully this gets fixed in
 107         // #14467)
 108         dispatcherClient     *arvados.Client
 109         DispatcherArvClient  IArvadosClient
 110         DispatcherKeepClient IKeepClient
 111
 112         // Container client is initialized with the Container token
 113         // This token controls the permissions of the container, and
 114         // must be used for operations such as reading collections.
 115         //
 116         // Same comment as above applies to
 117         // containerClient/ContainerArvClient.
 118         containerClient     *arvados.Client
 119         ContainerArvClient  IArvadosClient
 120         ContainerKeepClient IKeepClient
 121
 122         Container       arvados.Container
 123         ContainerConfig dockercontainer.Config
 124         HostConfig      dockercontainer.HostConfig
 125         token           string
 126         ContainerID     string
 127         ExitCode        *int
 128         NewLogWriter    NewLogWriter
 129         loggingDone     chan bool
 130         CrunchLog       *ThrottledLogger
 131         Stdout          io.WriteCloser
 132         Stderr          io.WriteCloser
 133         logUUID         string
 134         logMtx          sync.Mutex
 135         LogCollection   arvados.CollectionFileSystem
 136         LogsPDH         *string
 137         RunArvMount     RunArvMount
 138         MkTempDir       MkTempDir
 139         ArvMount        *exec.Cmd
 140         ArvMountPoint   string
 141         HostOutputDir   string
 142         Binds           []string
 143         Volumes         map[string]struct{}
 144         OutputPDH       *string
 145         SigChan         chan os.Signal
 146         ArvMountExit    chan error
 147         SecretMounts    map[string]arvados.Mount
 148         MkArvClient     func(token string) (IArvadosClient, IKeepClient, *arvados.Client, error)
 149         finalState      string
 150         parentTemp      string
 151
 152         statLogger       io.WriteCloser
 153         statReporter     *crunchstat.Reporter
 154         hoststatLogger   io.WriteCloser
 155         hoststatReporter *crunchstat.Reporter
 156         statInterval     time.Duration
 157         cgroupRoot       string
 158         // What we expect the container's cgroup parent to be.
 159         expectCgroupParent string
 160         // What we tell docker to use as the container's cgroup
 161         // parent. Note: Ideally we would use the same field for both
 162         // expectCgroupParent and setCgroupParent, and just make it
 163         // default to "docker". However, when using docker < 1.10 with
 164         // systemd, specifying a non-empty cgroup parent (even the
 165         // default value "docker") hits a docker bug
 166         // (https://github.com/docker/docker/issues/17126). Using two
 167         // separate fields makes it possible to use the "expect cgroup
 168         // parent to be X" feature even on sites where the "specify
 169         // cgroup parent" feature breaks.
 170         setCgroupParent string
 171
 172         cStateLock sync.Mutex
 173         cCancelled bool // StopContainer() invoked
 174         cRemoved   bool // docker confirmed the container no longer exists
 175
 176         enableNetwork string // one of "default" or "always"
 177         networkMode   string // passed through to HostConfig.NetworkMode
 178         arvMountLog   *ThrottledLogger
 179
 180         containerWatchdogInterval time.Duration
 181
 182         gateway Gateway
 183 }
 184
 185 // setupSignals sets up signal handling to gracefully terminate the underlying
 186 // Docker container and update state when receiving a TERM, INT or QUIT signal.
 187 func (runner *ContainerRunner) setupSignals() {
 188         runner.SigChan = make(chan os.Signal, 1)
 189         signal.Notify(runner.SigChan, syscall.SIGTERM)
 190         signal.Notify(runner.SigChan, syscall.SIGINT)
 191         signal.Notify(runner.SigChan, syscall.SIGQUIT)
 192
 193         go func(sig chan os.Signal) {
 194                 for s := range sig {
 195                         runner.stop(s)
 196                 }
 197         }(runner.SigChan)
 198 }
 199
 200 // stop the underlying Docker container.
 201 func (runner *ContainerRunner) stop(sig os.Signal) {
 202         runner.cStateLock.Lock()
 203         defer runner.cStateLock.Unlock()
 204         if sig != nil {
 205                 runner.CrunchLog.Printf("caught signal: %v", sig)
 206         }
 207         if runner.ContainerID == "" {
 208                 return
 209         }
 210         runner.cCancelled = true
 211         runner.CrunchLog.Printf("removing container")
 212         err := runner.Docker.ContainerRemove(context.TODO(), runner.ContainerID, dockertypes.ContainerRemoveOptions{Force: true})
 213         if err != nil {
 214                 runner.CrunchLog.Printf("error removing container: %s", err)
 215         }
 216         if err == nil || strings.Contains(err.Error(), "No such container: "+runner.ContainerID) {
 217                 runner.cRemoved = true
 218         }
 219 }
 220
 221 var errorBlacklist = []string{
 222         "(?ms).*[Cc]annot connect to the Docker daemon.*",
 223         "(?ms).*oci runtime error.*starting container process.*container init.*mounting.*to rootfs.*no such file or directory.*",
 224         "(?ms).*grpc: the connection is unavailable.*",
 225 }
 226 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)")
 227
 228 func (runner *ContainerRunner) runBrokenNodeHook() {
 229         if *brokenNodeHook == "" {
 230                 path := filepath.Join(lockdir, brokenfile)
 231                 runner.CrunchLog.Printf("Writing %s to mark node as broken", path)
 232                 f, err := os.OpenFile(path, os.O_CREATE|os.O_RDWR, 0700)
 233                 if err != nil {
 234                         runner.CrunchLog.Printf("Error writing %s: %s", path, err)
 235                         return
 236                 }
 237                 f.Close()
 238         } else {
 239                 runner.CrunchLog.Printf("Running broken node hook %q", *brokenNodeHook)
 240                 // run killme script
 241                 c := exec.Command(*brokenNodeHook)
 242                 c.Stdout = runner.CrunchLog
 243                 c.Stderr = runner.CrunchLog
 244                 err := c.Run()
 245                 if err != nil {
 246                         runner.CrunchLog.Printf("Error running broken node hook: %v", err)
 247                 }
 248         }
 249 }
 250
 251 func (runner *ContainerRunner) checkBrokenNode(goterr error) bool {
 252         for _, d := range errorBlacklist {
 253                 if m, e := regexp.MatchString(d, goterr.Error()); m && e == nil {
 254                         runner.CrunchLog.Printf("Error suggests node is unable to run containers: %v", goterr)
 255                         runner.runBrokenNodeHook()
 256                         return true
 257                 }
 258         }
 259         return false
 260 }
 261
 262 // LoadImage determines the docker image id from the container record and
 263 // checks if it is available in the local Docker image store.  If not, it loads
 264 // the image from Keep.
 265 func (runner *ContainerRunner) LoadImage() (err error) {
 266
 267         runner.CrunchLog.Printf("Fetching Docker image from collection '%s'", runner.Container.ContainerImage)
 268
 269         var collection arvados.Collection
 270         err = runner.ContainerArvClient.Get("collections", runner.Container.ContainerImage, nil, &collection)
 271         if err != nil {
 272                 return fmt.Errorf("While getting container image collection: %v", err)
 273         }
 274         manifest := manifest.Manifest{Text: collection.ManifestText}
 275         var img, imageID string
 276         for ms := range manifest.StreamIter() {
 277                 img = ms.FileStreamSegments[0].Name
 278                 if !strings.HasSuffix(img, ".tar") {
 279                         return fmt.Errorf("First file in the container image collection does not end in .tar")
 280                 }
 281                 imageID = img[:len(img)-4]
 282         }
 283
 284         runner.CrunchLog.Printf("Using Docker image id '%s'", imageID)
 285
 286         _, _, err = runner.Docker.ImageInspectWithRaw(context.TODO(), imageID)
 287         if err != nil {
 288                 runner.CrunchLog.Print("Loading Docker image from keep")
 289
 290                 var readCloser io.ReadCloser
 291                 readCloser, err = runner.ContainerKeepClient.ManifestFileReader(manifest, img)
 292                 if err != nil {
 293                         return fmt.Errorf("While creating ManifestFileReader for container image: %v", err)
 294                 }
 295
 296                 response, err := runner.Docker.ImageLoad(context.TODO(), readCloser, true)
 297                 if err != nil {
 298                         return fmt.Errorf("While loading container image into Docker: %v", err)
 299                 }
 300
 301                 defer response.Body.Close()
 302                 rbody, err := ioutil.ReadAll(response.Body)
 303                 if err != nil {
 304                         return fmt.Errorf("Reading response to image load: %v", err)
 305                 }
 306                 runner.CrunchLog.Printf("Docker response: %s", rbody)
 307         } else {
 308                 runner.CrunchLog.Print("Docker image is available")
 309         }
 310
 311         runner.ContainerConfig.Image = imageID
 312
 313         runner.ContainerKeepClient.ClearBlockCache()
 314
 315         return nil
 316 }
 317
 318 func (runner *ContainerRunner) ArvMountCmd(arvMountCmd []string, token string) (c *exec.Cmd, err error) {
 319         c = exec.Command("arv-mount", arvMountCmd...)
 320
 321         // Copy our environment, but override ARVADOS_API_TOKEN with
 322         // the container auth token.
 323         c.Env = nil
 324         for _, s := range os.Environ() {
 325                 if !strings.HasPrefix(s, "ARVADOS_API_TOKEN=") {
 326                         c.Env = append(c.Env, s)
 327                 }
 328         }
 329         c.Env = append(c.Env, "ARVADOS_API_TOKEN="+token)
 330
 331         w, err := runner.NewLogWriter("arv-mount")
 332         if err != nil {
 333                 return nil, err
 334         }
 335         runner.arvMountLog = NewThrottledLogger(w)
 336         c.Stdout = runner.arvMountLog
 337         c.Stderr = runner.arvMountLog
 338
 339         runner.CrunchLog.Printf("Running %v", c.Args)
 340
 341         err = c.Start()
 342         if err != nil {
 343                 return nil, err
 344         }
 345
 346         statReadme := make(chan bool)
 347         runner.ArvMountExit = make(chan error)
 348
 349         keepStatting := true
 350         go func() {
 351                 for keepStatting {
 352                         time.Sleep(100 * time.Millisecond)
 353                         _, err = os.Stat(fmt.Sprintf("%s/by_id/README", runner.ArvMountPoint))
 354                         if err == nil {
 355                                 keepStatting = false
 356                                 statReadme <- true
 357                         }
 358                 }
 359                 close(statReadme)
 360         }()
 361
 362         go func() {
 363                 mnterr := c.Wait()
 364                 if mnterr != nil {
 365                         runner.CrunchLog.Printf("Arv-mount exit error: %v", mnterr)
 366                 }
 367                 runner.ArvMountExit <- mnterr
 368                 close(runner.ArvMountExit)
 369         }()
 370
 371         select {
 372         case <-statReadme:
 373                 break
 374         case err := <-runner.ArvMountExit:
 375                 runner.ArvMount = nil
 376                 keepStatting = false
 377                 return nil, err
 378         }
 379
 380         return c, nil
 381 }
 382
 383 func (runner *ContainerRunner) SetupArvMountPoint(prefix string) (err error) {
 384         if runner.ArvMountPoint == "" {
 385                 runner.ArvMountPoint, err = runner.MkTempDir(runner.parentTemp, prefix)
 386         }
 387         return
 388 }
 389
 390 func copyfile(src string, dst string) (err error) {
 391         srcfile, err := os.Open(src)
 392         if err != nil {
 393                 return
 394         }
 395
 396         os.MkdirAll(path.Dir(dst), 0777)
 397
 398         dstfile, err := os.Create(dst)
 399         if err != nil {
 400                 return
 401         }
 402         _, err = io.Copy(dstfile, srcfile)
 403         if err != nil {
 404                 return
 405         }
 406
 407         err = srcfile.Close()
 408         err2 := dstfile.Close()
 409
 410         if err != nil {
 411                 return
 412         }
 413
 414         if err2 != nil {
 415                 return err2
 416         }
 417
 418         return nil
 419 }
 420
 421 func (runner *ContainerRunner) SetupMounts() (err error) {
 422         err = runner.SetupArvMountPoint("keep")
 423         if err != nil {
 424                 return fmt.Errorf("While creating keep mount temp dir: %v", err)
 425         }
 426
 427         token, err := runner.ContainerToken()
 428         if err != nil {
 429                 return fmt.Errorf("could not get container token: %s", err)
 430         }
 431
 432         pdhOnly := true
 433         tmpcount := 0
 434         arvMountCmd := []string{
 435                 "--foreground",
 436                 "--allow-other",
 437                 "--read-write",
 438                 fmt.Sprintf("--crunchstat-interval=%v", runner.statInterval.Seconds())}
 439
 440         if runner.Container.RuntimeConstraints.KeepCacheRAM > 0 {
 441                 arvMountCmd = append(arvMountCmd, "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheRAM))
 442         }
 443
 444         collectionPaths := []string{}
 445         runner.Binds = nil
 446         runner.Volumes = make(map[string]struct{})
 447         needCertMount := true
 448         type copyFile struct {
 449                 src  string
 450                 bind string
 451         }
 452         var copyFiles []copyFile
 453
 454         var binds []string
 455         for bind := range runner.Container.Mounts {
 456                 binds = append(binds, bind)
 457         }
 458         for bind := range runner.SecretMounts {
 459                 if _, ok := runner.Container.Mounts[bind]; ok {
 460                         return fmt.Errorf("secret mount %q conflicts with regular mount", bind)
 461                 }
 462                 if runner.SecretMounts[bind].Kind != "json" &&
 463                         runner.SecretMounts[bind].Kind != "text" {
 464                         return fmt.Errorf("secret mount %q type is %q but only 'json' and 'text' are permitted",
 465                                 bind, runner.SecretMounts[bind].Kind)
 466                 }
 467                 binds = append(binds, bind)
 468         }
 469         sort.Strings(binds)
 470
 471         for _, bind := range binds {
 472                 mnt, ok := runner.Container.Mounts[bind]
 473                 if !ok {
 474                         mnt = runner.SecretMounts[bind]
 475                 }
 476                 if bind == "stdout" || bind == "stderr" {
 477                         // Is it a "file" mount kind?
 478                         if mnt.Kind != "file" {
 479                                 return fmt.Errorf("unsupported mount kind '%s' for %s: only 'file' is supported", mnt.Kind, bind)
 480                         }
 481
 482                         // Does path start with OutputPath?
 483                         prefix := runner.Container.OutputPath
 484                         if !strings.HasSuffix(prefix, "/") {
 485                                 prefix += "/"
 486                         }
 487                         if !strings.HasPrefix(mnt.Path, prefix) {
 488                                 return fmt.Errorf("%s path does not start with OutputPath: %s, %s", strings.Title(bind), mnt.Path, prefix)
 489                         }
 490                 }
 491
 492                 if bind == "stdin" {
 493                         // Is it a "collection" mount kind?
 494                         if mnt.Kind != "collection" && mnt.Kind != "json" {
 495                                 return fmt.Errorf("unsupported mount kind '%s' for stdin: only 'collection' and 'json' are supported", mnt.Kind)
 496                         }
 497                 }
 498
 499                 if bind == "/etc/arvados/ca-certificates.crt" {
 500                         needCertMount = false
 501                 }
 502
 503                 if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && bind != runner.Container.OutputPath+"/" {
 504                         if mnt.Kind != "collection" && mnt.Kind != "text" && mnt.Kind != "json" {
 505                                 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)
 506                         }
 507                 }
 508
 509                 switch {
 510                 case mnt.Kind == "collection" && bind != "stdin":
 511                         var src string
 512                         if mnt.UUID != "" && mnt.PortableDataHash != "" {
 513                                 return fmt.Errorf("cannot specify both 'uuid' and 'portable_data_hash' for a collection mount")
 514                         }
 515                         if mnt.UUID != "" {
 516                                 if mnt.Writable {
 517                                         return fmt.Errorf("writing to existing collections currently not permitted")
 518                                 }
 519                                 pdhOnly = false
 520                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.UUID)
 521                         } else if mnt.PortableDataHash != "" {
 522                                 if mnt.Writable && !strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 523                                         return fmt.Errorf("can never write to a collection specified by portable data hash")
 524                                 }
 525                                 idx := strings.Index(mnt.PortableDataHash, "/")
 526                                 if idx > 0 {
 527                                         mnt.Path = path.Clean(mnt.PortableDataHash[idx:])
 528                                         mnt.PortableDataHash = mnt.PortableDataHash[0:idx]
 529                                         runner.Container.Mounts[bind] = mnt
 530                                 }
 531                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.PortableDataHash)
 532                                 if mnt.Path != "" && mnt.Path != "." {
 533                                         if strings.HasPrefix(mnt.Path, "./") {
 534                                                 mnt.Path = mnt.Path[2:]
 535                                         } else if strings.HasPrefix(mnt.Path, "/") {
 536                                                 mnt.Path = mnt.Path[1:]
 537                                         }
 538                                         src += "/" + mnt.Path
 539                                 }
 540                         } else {
 541                                 src = fmt.Sprintf("%s/tmp%d", runner.ArvMountPoint, tmpcount)
 542                                 arvMountCmd = append(arvMountCmd, "--mount-tmp")
 543                                 arvMountCmd = append(arvMountCmd, fmt.Sprintf("tmp%d", tmpcount))
 544                                 tmpcount++
 545                         }
 546                         if mnt.Writable {
 547                                 if bind == runner.Container.OutputPath {
 548                                         runner.HostOutputDir = src
 549                                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", src, bind))
 550                                 } else if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 551                                         copyFiles = append(copyFiles, copyFile{src, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
 552                                 } else {
 553                                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", src, bind))
 554                                 }
 555                         } else {
 556                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", src, bind))
 557                         }
 558                         collectionPaths = append(collectionPaths, src)
 559
 560                 case mnt.Kind == "tmp":
 561                         var tmpdir string
 562                         tmpdir, err = runner.MkTempDir(runner.parentTemp, "tmp")
 563                         if err != nil {
 564                                 return fmt.Errorf("while creating mount temp dir: %v", err)
 565                         }
 566                         st, staterr := os.Stat(tmpdir)
 567                         if staterr != nil {
 568                                 return fmt.Errorf("while Stat on temp dir: %v", staterr)
 569                         }
 570                         err = os.Chmod(tmpdir, st.Mode()|os.ModeSetgid|0777)
 571                         if staterr != nil {
 572                                 return fmt.Errorf("while Chmod temp dir: %v", err)
 573                         }
 574                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", tmpdir, bind))
 575                         if bind == runner.Container.OutputPath {
 576                                 runner.HostOutputDir = tmpdir
 577                         }
 578
 579                 case mnt.Kind == "json" || mnt.Kind == "text":
 580                         var filedata []byte
 581                         if mnt.Kind == "json" {
 582                                 filedata, err = json.Marshal(mnt.Content)
 583                                 if err != nil {
 584                                         return fmt.Errorf("encoding json data: %v", err)
 585                                 }
 586                         } else {
 587                                 text, ok := mnt.Content.(string)
 588                                 if !ok {
 589                                         return fmt.Errorf("content for mount %q must be a string", bind)
 590                                 }
 591                                 filedata = []byte(text)
 592                         }
 593
 594                         tmpdir, err := runner.MkTempDir(runner.parentTemp, mnt.Kind)
 595                         if err != nil {
 596                                 return fmt.Errorf("creating temp dir: %v", err)
 597                         }
 598                         tmpfn := filepath.Join(tmpdir, "mountdata."+mnt.Kind)
 599                         err = ioutil.WriteFile(tmpfn, filedata, 0444)
 600                         if err != nil {
 601                                 return fmt.Errorf("writing temp file: %v", err)
 602                         }
 603                         if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 604                                 copyFiles = append(copyFiles, copyFile{tmpfn, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
 605                         } else {
 606                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", tmpfn, bind))
 607                         }
 608
 609                 case mnt.Kind == "git_tree":
 610                         tmpdir, err := runner.MkTempDir(runner.parentTemp, "git_tree")
 611                         if err != nil {
 612                                 return fmt.Errorf("creating temp dir: %v", err)
 613                         }
 614                         err = gitMount(mnt).extractTree(runner.ContainerArvClient, tmpdir, token)
 615                         if err != nil {
 616                                 return err
 617                         }
 618                         runner.Binds = append(runner.Binds, tmpdir+":"+bind+":ro")
 619                 }
 620         }
 621
 622         if runner.HostOutputDir == "" {
 623                 return fmt.Errorf("output path does not correspond to a writable mount point")
 624         }
 625
 626         if needCertMount && runner.Container.RuntimeConstraints.API {
 627                 for _, certfile := range arvadosclient.CertFiles {
 628                         _, err := os.Stat(certfile)
 629                         if err == nil {
 630                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:/etc/arvados/ca-certificates.crt:ro", certfile))
 631                                 break
 632                         }
 633                 }
 634         }
 635
 636         if pdhOnly {
 637                 arvMountCmd = append(arvMountCmd, "--mount-by-pdh", "by_id")
 638         } else {
 639                 arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_id")
 640         }
 641         arvMountCmd = append(arvMountCmd, runner.ArvMountPoint)
 642
 643         runner.ArvMount, err = runner.RunArvMount(arvMountCmd, token)
 644         if err != nil {
 645                 return fmt.Errorf("while trying to start arv-mount: %v", err)
 646         }
 647
 648         for _, p := range collectionPaths {
 649                 _, err = os.Stat(p)
 650                 if err != nil {
 651                         return fmt.Errorf("while checking that input files exist: %v", err)
 652                 }
 653         }
 654
 655         for _, cp := range copyFiles {
 656                 st, err := os.Stat(cp.src)
 657                 if err != nil {
 658                         return fmt.Errorf("while staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 659                 }
 660                 if st.IsDir() {
 661                         err = filepath.Walk(cp.src, func(walkpath string, walkinfo os.FileInfo, walkerr error) error {
 662                                 if walkerr != nil {
 663                                         return walkerr
 664                                 }
 665                                 target := path.Join(cp.bind, walkpath[len(cp.src):])
 666                                 if walkinfo.Mode().IsRegular() {
 667                                         copyerr := copyfile(walkpath, target)
 668                                         if copyerr != nil {
 669                                                 return copyerr
 670                                         }
 671                                         return os.Chmod(target, walkinfo.Mode()|0777)
 672                                 } else if walkinfo.Mode().IsDir() {
 673                                         mkerr := os.MkdirAll(target, 0777)
 674                                         if mkerr != nil {
 675                                                 return mkerr
 676                                         }
 677                                         return os.Chmod(target, walkinfo.Mode()|os.ModeSetgid|0777)
 678                                 } else {
 679                                         return fmt.Errorf("source %q is not a regular file or directory", cp.src)
 680                                 }
 681                         })
 682                 } else if st.Mode().IsRegular() {
 683                         err = copyfile(cp.src, cp.bind)
 684                         if err == nil {
 685                                 err = os.Chmod(cp.bind, st.Mode()|0777)
 686                         }
 687                 }
 688                 if err != nil {
 689                         return fmt.Errorf("while staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 690                 }
 691         }
 692
 693         return nil
 694 }
 695
 696 func (runner *ContainerRunner) ProcessDockerAttach(containerReader io.Reader) {
 697         // Handle docker log protocol
 698         // https://docs.docker.com/engine/reference/api/docker_remote_api_v1.15/#attach-to-a-container
 699         defer close(runner.loggingDone)
 700
 701         header := make([]byte, 8)
 702         var err error
 703         for err == nil {
 704                 _, err = io.ReadAtLeast(containerReader, header, 8)
 705                 if err != nil {
 706                         if err == io.EOF {
 707                                 err = nil
 708                         }
 709                         break
 710                 }
 711                 readsize := int64(header[7]) | (int64(header[6]) << 8) | (int64(header[5]) << 16) | (int64(header[4]) << 24)
 712                 if header[0] == 1 {
 713                         // stdout
 714                         _, err = io.CopyN(runner.Stdout, containerReader, readsize)
 715                 } else {
 716                         // stderr
 717                         _, err = io.CopyN(runner.Stderr, containerReader, readsize)
 718                 }
 719         }
 720
 721         if err != nil {
 722                 runner.CrunchLog.Printf("error reading docker logs: %v", err)
 723         }
 724
 725         err = runner.Stdout.Close()
 726         if err != nil {
 727                 runner.CrunchLog.Printf("error closing stdout logs: %v", err)
 728         }
 729
 730         err = runner.Stderr.Close()
 731         if err != nil {
 732                 runner.CrunchLog.Printf("error closing stderr logs: %v", err)
 733         }
 734
 735         if runner.statReporter != nil {
 736                 runner.statReporter.Stop()
 737                 err = runner.statLogger.Close()
 738                 if err != nil {
 739                         runner.CrunchLog.Printf("error closing crunchstat logs: %v", err)
 740                 }
 741         }
 742 }
 743
 744 func (runner *ContainerRunner) stopHoststat() error {
 745         if runner.hoststatReporter == nil {
 746                 return nil
 747         }
 748         runner.hoststatReporter.Stop()
 749         err := runner.hoststatLogger.Close()
 750         if err != nil {
 751                 return fmt.Errorf("error closing hoststat logs: %v", err)
 752         }
 753         return nil
 754 }
 755
 756 func (runner *ContainerRunner) startHoststat() error {
 757         w, err := runner.NewLogWriter("hoststat")
 758         if err != nil {
 759                 return err
 760         }
 761         runner.hoststatLogger = NewThrottledLogger(w)
 762         runner.hoststatReporter = &crunchstat.Reporter{
 763                 Logger:     log.New(runner.hoststatLogger, "", 0),
 764                 CgroupRoot: runner.cgroupRoot,
 765                 PollPeriod: runner.statInterval,
 766         }
 767         runner.hoststatReporter.Start()
 768         return nil
 769 }
 770
 771 func (runner *ContainerRunner) startCrunchstat() error {
 772         w, err := runner.NewLogWriter("crunchstat")
 773         if err != nil {
 774                 return err
 775         }
 776         runner.statLogger = NewThrottledLogger(w)
 777         runner.statReporter = &crunchstat.Reporter{
 778                 CID:          runner.ContainerID,
 779                 Logger:       log.New(runner.statLogger, "", 0),
 780                 CgroupParent: runner.expectCgroupParent,
 781                 CgroupRoot:   runner.cgroupRoot,
 782                 PollPeriod:   runner.statInterval,
 783                 TempDir:      runner.parentTemp,
 784         }
 785         runner.statReporter.Start()
 786         return nil
 787 }
 788
 789 type infoCommand struct {
 790         label string
 791         cmd   []string
 792 }
 793
 794 // LogHostInfo logs info about the current host, for debugging and
 795 // accounting purposes. Although it's logged as "node-info", this is
 796 // about the environment where crunch-run is actually running, which
 797 // might differ from what's described in the node record (see
 798 // LogNodeRecord).
 799 func (runner *ContainerRunner) LogHostInfo() (err error) {
 800         w, err := runner.NewLogWriter("node-info")
 801         if err != nil {
 802                 return
 803         }
 804
 805         commands := []infoCommand{
 806                 {
 807                         label: "Host Information",
 808                         cmd:   []string{"uname", "-a"},
 809                 },
 810                 {
 811                         label: "CPU Information",
 812                         cmd:   []string{"cat", "/proc/cpuinfo"},
 813                 },
 814                 {
 815                         label: "Memory Information",
 816                         cmd:   []string{"cat", "/proc/meminfo"},
 817                 },
 818                 {
 819                         label: "Disk Space",
 820                         cmd:   []string{"df", "-m", "/", os.TempDir()},
 821                 },
 822                 {
 823                         label: "Disk INodes",
 824                         cmd:   []string{"df", "-i", "/", os.TempDir()},
 825                 },
 826         }
 827
 828         // Run commands with informational output to be logged.
 829         for _, command := range commands {
 830                 fmt.Fprintln(w, command.label)
 831                 cmd := exec.Command(command.cmd[0], command.cmd[1:]...)
 832                 cmd.Stdout = w
 833                 cmd.Stderr = w
 834                 if err := cmd.Run(); err != nil {
 835                         err = fmt.Errorf("While running command %q: %v", command.cmd, err)
 836                         fmt.Fprintln(w, err)
 837                         return err
 838                 }
 839                 fmt.Fprintln(w, "")
 840         }
 841
 842         err = w.Close()
 843         if err != nil {
 844                 return fmt.Errorf("While closing node-info logs: %v", err)
 845         }
 846         return nil
 847 }
 848
 849 // LogContainerRecord gets and saves the raw JSON container record from the API server
 850 func (runner *ContainerRunner) LogContainerRecord() error {
 851         logged, err := runner.logAPIResponse("container", "containers", map[string]interface{}{"filters": [][]string{{"uuid", "=", runner.Container.UUID}}}, nil)
 852         if !logged && err == nil {
 853                 err = fmt.Errorf("error: no container record found for %s", runner.Container.UUID)
 854         }
 855         return err
 856 }
 857
 858 // LogNodeRecord logs the current host's InstanceType config entry (or
 859 // the arvados#node record, if running via crunch-dispatch-slurm).
 860 func (runner *ContainerRunner) LogNodeRecord() error {
 861         if it := os.Getenv("InstanceType"); it != "" {
 862                 // Dispatched via arvados-dispatch-cloud. Save
 863                 // InstanceType config fragment received from
 864                 // dispatcher on stdin.
 865                 w, err := runner.LogCollection.OpenFile("node.json", os.O_CREATE|os.O_WRONLY, 0666)
 866                 if err != nil {
 867                         return err
 868                 }
 869                 defer w.Close()
 870                 _, err = io.WriteString(w, it)
 871                 if err != nil {
 872                         return err
 873                 }
 874                 return w.Close()
 875         }
 876         // Dispatched via crunch-dispatch-slurm. Look up
 877         // apiserver's node record corresponding to
 878         // $SLURMD_NODENAME.
 879         hostname := os.Getenv("SLURMD_NODENAME")
 880         if hostname == "" {
 881                 hostname, _ = os.Hostname()
 882         }
 883         _, err := runner.logAPIResponse("node", "nodes", map[string]interface{}{"filters": [][]string{{"hostname", "=", hostname}}}, func(resp interface{}) {
 884                 // The "info" field has admin-only info when
 885                 // obtained with a privileged token, and
 886                 // should not be logged.
 887                 node, ok := resp.(map[string]interface{})
 888                 if ok {
 889                         delete(node, "info")
 890                 }
 891         })
 892         return err
 893 }
 894
 895 func (runner *ContainerRunner) logAPIResponse(label, path string, params map[string]interface{}, munge func(interface{})) (logged bool, err error) {
 896         writer, err := runner.LogCollection.OpenFile(label+".json", os.O_CREATE|os.O_WRONLY, 0666)
 897         if err != nil {
 898                 return false, err
 899         }
 900         w := &ArvLogWriter{
 901                 ArvClient:     runner.DispatcherArvClient,
 902                 UUID:          runner.Container.UUID,
 903                 loggingStream: label,
 904                 writeCloser:   writer,
 905         }
 906
 907         reader, err := runner.DispatcherArvClient.CallRaw("GET", path, "", "", arvadosclient.Dict(params))
 908         if err != nil {
 909                 return false, fmt.Errorf("error getting %s record: %v", label, err)
 910         }
 911         defer reader.Close()
 912
 913         dec := json.NewDecoder(reader)
 914         dec.UseNumber()
 915         var resp map[string]interface{}
 916         if err = dec.Decode(&resp); err != nil {
 917                 return false, fmt.Errorf("error decoding %s list response: %v", label, err)
 918         }
 919         items, ok := resp["items"].([]interface{})
 920         if !ok {
 921                 return false, fmt.Errorf("error decoding %s list response: no \"items\" key in API list response", label)
 922         } else if len(items) < 1 {
 923                 return false, nil
 924         }
 925         if munge != nil {
 926                 munge(items[0])
 927         }
 928         // Re-encode it using indentation to improve readability
 929         enc := json.NewEncoder(w)
 930         enc.SetIndent("", "    ")
 931         if err = enc.Encode(items[0]); err != nil {
 932                 return false, fmt.Errorf("error logging %s record: %v", label, err)
 933         }
 934         err = w.Close()
 935         if err != nil {
 936                 return false, fmt.Errorf("error closing %s.json in log collection: %v", label, err)
 937         }
 938         return true, nil
 939 }
 940
 941 // AttachStreams connects the docker container stdin, stdout and stderr logs
 942 // to the Arvados logger which logs to Keep and the API server logs table.
 943 func (runner *ContainerRunner) AttachStreams() (err error) {
 944
 945         runner.CrunchLog.Print("Attaching container streams")
 946
 947         // If stdin mount is provided, attach it to the docker container
 948         var stdinRdr arvados.File
 949         var stdinJSON []byte
 950         if stdinMnt, ok := runner.Container.Mounts["stdin"]; ok {
 951                 if stdinMnt.Kind == "collection" {
 952                         var stdinColl arvados.Collection
 953                         collID := stdinMnt.UUID
 954                         if collID == "" {
 955                                 collID = stdinMnt.PortableDataHash
 956                         }
 957                         err = runner.ContainerArvClient.Get("collections", collID, nil, &stdinColl)
 958                         if err != nil {
 959                                 return fmt.Errorf("While getting stdin collection: %v", err)
 960                         }
 961
 962                         stdinRdr, err = runner.ContainerKeepClient.ManifestFileReader(
 963                                 manifest.Manifest{Text: stdinColl.ManifestText},
 964                                 stdinMnt.Path)
 965                         if os.IsNotExist(err) {
 966                                 return fmt.Errorf("stdin collection path not found: %v", stdinMnt.Path)
 967                         } else if err != nil {
 968                                 return fmt.Errorf("While getting stdin collection path %v: %v", stdinMnt.Path, err)
 969                         }
 970                 } else if stdinMnt.Kind == "json" {
 971                         stdinJSON, err = json.Marshal(stdinMnt.Content)
 972                         if err != nil {
 973                                 return fmt.Errorf("While encoding stdin json data: %v", err)
 974                         }
 975                 }
 976         }
 977
 978         stdinUsed := stdinRdr != nil || len(stdinJSON) != 0
 979         response, err := runner.Docker.ContainerAttach(context.TODO(), runner.ContainerID,
 980                 dockertypes.ContainerAttachOptions{Stream: true, Stdin: stdinUsed, Stdout: true, Stderr: true})
 981         if err != nil {
 982                 return fmt.Errorf("While attaching container stdout/stderr streams: %v", err)
 983         }
 984
 985         runner.loggingDone = make(chan bool)
 986
 987         if stdoutMnt, ok := runner.Container.Mounts["stdout"]; ok {
 988                 stdoutFile, err := runner.getStdoutFile(stdoutMnt.Path)
 989                 if err != nil {
 990                         return err
 991                 }
 992                 runner.Stdout = stdoutFile
 993         } else if w, err := runner.NewLogWriter("stdout"); err != nil {
 994                 return err
 995         } else {
 996                 runner.Stdout = NewThrottledLogger(w)
 997         }
 998
 999         if stderrMnt, ok := runner.Container.Mounts["stderr"]; ok {
1000                 stderrFile, err := runner.getStdoutFile(stderrMnt.Path)
1001                 if err != nil {
1002                         return err
1003                 }
1004                 runner.Stderr = stderrFile
1005         } else if w, err := runner.NewLogWriter("stderr"); err != nil {
1006                 return err
1007         } else {
1008                 runner.Stderr = NewThrottledLogger(w)
1009         }
1010
1011         if stdinRdr != nil {
1012                 go func() {
1013                         _, err := io.Copy(response.Conn, stdinRdr)
1014                         if err != nil {
1015                                 runner.CrunchLog.Printf("While writing stdin collection to docker container: %v", err)
1016                                 runner.stop(nil)
1017                         }
1018                         stdinRdr.Close()
1019                         response.CloseWrite()
1020                 }()
1021         } else if len(stdinJSON) != 0 {
1022                 go func() {
1023                         _, err := io.Copy(response.Conn, bytes.NewReader(stdinJSON))
1024                         if err != nil {
1025                                 runner.CrunchLog.Printf("While writing stdin json to docker container: %v", err)
1026                                 runner.stop(nil)
1027                         }
1028                         response.CloseWrite()
1029                 }()
1030         }
1031
1032         go runner.ProcessDockerAttach(response.Reader)
1033
1034         return nil
1035 }
1036
1037 func (runner *ContainerRunner) getStdoutFile(mntPath string) (*os.File, error) {
1038         stdoutPath := mntPath[len(runner.Container.OutputPath):]
1039         index := strings.LastIndex(stdoutPath, "/")
1040         if index > 0 {
1041                 subdirs := stdoutPath[:index]
1042                 if subdirs != "" {
1043                         st, err := os.Stat(runner.HostOutputDir)
1044                         if err != nil {
1045                                 return nil, fmt.Errorf("While Stat on temp dir: %v", err)
1046                         }
1047                         stdoutPath := filepath.Join(runner.HostOutputDir, subdirs)
1048                         err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777)
1049                         if err != nil {
1050                                 return nil, fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err)
1051                         }
1052                 }
1053         }
1054         stdoutFile, err := os.Create(filepath.Join(runner.HostOutputDir, stdoutPath))
1055         if err != nil {
1056                 return nil, fmt.Errorf("While creating file %q: %v", stdoutPath, err)
1057         }
1058
1059         return stdoutFile, nil
1060 }
1061
1062 // CreateContainer creates the docker container.
1063 func (runner *ContainerRunner) CreateContainer() error {
1064         runner.CrunchLog.Print("Creating Docker container")
1065
1066         runner.ContainerConfig.Cmd = runner.Container.Command
1067         if runner.Container.Cwd != "." {
1068                 runner.ContainerConfig.WorkingDir = runner.Container.Cwd
1069         }
1070
1071         for k, v := range runner.Container.Environment {
1072                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env, k+"="+v)
1073         }
1074
1075         runner.ContainerConfig.Volumes = runner.Volumes
1076
1077         maxRAM := int64(runner.Container.RuntimeConstraints.RAM)
1078         minDockerRAM := int64(16)
1079         if maxRAM < minDockerRAM*1024*1024 {
1080                 // Docker daemon won't let you set a limit less than ~10 MiB
1081                 maxRAM = minDockerRAM * 1024 * 1024
1082         }
1083         runner.HostConfig = dockercontainer.HostConfig{
1084                 Binds: runner.Binds,
1085                 LogConfig: dockercontainer.LogConfig{
1086                         Type: "none",
1087                 },
1088                 Resources: dockercontainer.Resources{
1089                         CgroupParent: runner.setCgroupParent,
1090                         NanoCPUs:     int64(runner.Container.RuntimeConstraints.VCPUs) * 1000000000,
1091                         Memory:       maxRAM, // RAM
1092                         MemorySwap:   maxRAM, // RAM+swap
1093                         KernelMemory: maxRAM, // kernel portion
1094                 },
1095         }
1096
1097         if runner.Container.RuntimeConstraints.API {
1098                 tok, err := runner.ContainerToken()
1099                 if err != nil {
1100                         return err
1101                 }
1102                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env,
1103                         "ARVADOS_API_TOKEN="+tok,
1104                         "ARVADOS_API_HOST="+os.Getenv("ARVADOS_API_HOST"),
1105                         "ARVADOS_API_HOST_INSECURE="+os.Getenv("ARVADOS_API_HOST_INSECURE"),
1106                 )
1107                 runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
1108         } else {
1109                 if runner.enableNetwork == "always" {
1110                         runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
1111                 } else {
1112                         runner.HostConfig.NetworkMode = dockercontainer.NetworkMode("none")
1113                 }
1114         }
1115
1116         _, stdinUsed := runner.Container.Mounts["stdin"]
1117         runner.ContainerConfig.OpenStdin = stdinUsed
1118         runner.ContainerConfig.StdinOnce = stdinUsed
1119         runner.ContainerConfig.AttachStdin = stdinUsed
1120         runner.ContainerConfig.AttachStdout = true
1121         runner.ContainerConfig.AttachStderr = true
1122
1123         createdBody, err := runner.Docker.ContainerCreate(context.TODO(), &runner.ContainerConfig, &runner.HostConfig, nil, runner.Container.UUID)
1124         if err != nil {
1125                 return fmt.Errorf("While creating container: %v", err)
1126         }
1127
1128         runner.ContainerID = createdBody.ID
1129
1130         return runner.AttachStreams()
1131 }
1132
1133 // StartContainer starts the docker container created by CreateContainer.
1134 func (runner *ContainerRunner) StartContainer() error {
1135         runner.CrunchLog.Printf("Starting Docker container id '%s'", runner.ContainerID)
1136         runner.cStateLock.Lock()
1137         defer runner.cStateLock.Unlock()
1138         if runner.cCancelled {
1139                 return ErrCancelled
1140         }
1141         err := runner.Docker.ContainerStart(context.TODO(), runner.ContainerID,
1142                 dockertypes.ContainerStartOptions{})
1143         if err != nil {
1144                 var advice string
1145                 if m, e := regexp.MatchString("(?ms).*(exec|System error).*(no such file or directory|file not found).*", err.Error()); m && e == nil {
1146                         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])
1147                 }
1148                 return fmt.Errorf("could not start container: %v%s", err, advice)
1149         }
1150         return nil
1151 }
1152
1153 // WaitFinish waits for the container to terminate, capture the exit code, and
1154 // close the stdout/stderr logging.
1155 func (runner *ContainerRunner) WaitFinish() error {
1156         var runTimeExceeded <-chan time.Time
1157         runner.CrunchLog.Print("Waiting for container to finish")
1158
1159         waitOk, waitErr := runner.Docker.ContainerWait(context.TODO(), runner.ContainerID, dockercontainer.WaitConditionNotRunning)
1160         arvMountExit := runner.ArvMountExit
1161         if timeout := runner.Container.SchedulingParameters.MaxRunTime; timeout > 0 {
1162                 runTimeExceeded = time.After(time.Duration(timeout) * time.Second)
1163         }
1164
1165         containerGone := make(chan struct{})
1166         go func() {
1167                 defer close(containerGone)
1168                 if runner.containerWatchdogInterval < 1 {
1169                         runner.containerWatchdogInterval = time.Minute
1170                 }
1171                 for range time.NewTicker(runner.containerWatchdogInterval).C {
1172                         ctx, cancel := context.WithDeadline(context.Background(), time.Now().Add(runner.containerWatchdogInterval))
1173                         ctr, err := runner.Docker.ContainerInspect(ctx, runner.ContainerID)
1174                         cancel()
1175                         runner.cStateLock.Lock()
1176                         done := runner.cRemoved || runner.ExitCode != nil
1177                         runner.cStateLock.Unlock()
1178                         if done {
1179                                 return
1180                         } else if err != nil {
1181                                 runner.CrunchLog.Printf("Error inspecting container: %s", err)
1182                                 runner.checkBrokenNode(err)
1183                                 return
1184                         } else if ctr.State == nil || !(ctr.State.Running || ctr.State.Status == "created") {
1185                                 runner.CrunchLog.Printf("Container is not running: State=%v", ctr.State)
1186                                 return
1187                         }
1188                 }
1189         }()
1190
1191         for {
1192                 select {
1193                 case waitBody := <-waitOk:
1194                         runner.CrunchLog.Printf("Container exited with code: %v", waitBody.StatusCode)
1195                         code := int(waitBody.StatusCode)
1196                         runner.ExitCode = &code
1197
1198                         // wait for stdout/stderr to complete
1199                         <-runner.loggingDone
1200                         return nil
1201
1202                 case err := <-waitErr:
1203                         return fmt.Errorf("container wait: %v", err)
1204
1205                 case <-arvMountExit:
1206                         runner.CrunchLog.Printf("arv-mount exited while container is still running.  Stopping container.")
1207                         runner.stop(nil)
1208                         // arvMountExit will always be ready now that
1209                         // it's closed, but that doesn't interest us.
1210                         arvMountExit = nil
1211
1212                 case <-runTimeExceeded:
1213                         runner.CrunchLog.Printf("maximum run time exceeded. Stopping container.")
1214                         runner.stop(nil)
1215                         runTimeExceeded = nil
1216
1217                 case <-containerGone:
1218                         return errors.New("docker client never returned status")
1219                 }
1220         }
1221 }
1222
1223 func (runner *ContainerRunner) updateLogs() {
1224         ticker := time.NewTicker(crunchLogUpdatePeriod / 360)
1225         defer ticker.Stop()
1226
1227         sigusr1 := make(chan os.Signal, 1)
1228         signal.Notify(sigusr1, syscall.SIGUSR1)
1229         defer signal.Stop(sigusr1)
1230
1231         saveAtTime := time.Now().Add(crunchLogUpdatePeriod)
1232         saveAtSize := crunchLogUpdateSize
1233         var savedSize int64
1234         for {
1235                 select {
1236                 case <-ticker.C:
1237                 case <-sigusr1:
1238                         saveAtTime = time.Now()
1239                 }
1240                 runner.logMtx.Lock()
1241                 done := runner.LogsPDH != nil
1242                 runner.logMtx.Unlock()
1243                 if done {
1244                         return
1245                 }
1246                 size := runner.LogCollection.Size()
1247                 if size == savedSize || (time.Now().Before(saveAtTime) && size < saveAtSize) {
1248                         continue
1249                 }
1250                 saveAtTime = time.Now().Add(crunchLogUpdatePeriod)
1251                 saveAtSize = runner.LogCollection.Size() + crunchLogUpdateSize
1252                 saved, err := runner.saveLogCollection(false)
1253                 if err != nil {
1254                         runner.CrunchLog.Printf("error updating log collection: %s", err)
1255                         continue
1256                 }
1257
1258                 var updated arvados.Container
1259                 err = runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1260                         "container": arvadosclient.Dict{"log": saved.PortableDataHash},
1261                 }, &updated)
1262                 if err != nil {
1263                         runner.CrunchLog.Printf("error updating container log to %s: %s", saved.PortableDataHash, err)
1264                         continue
1265                 }
1266
1267                 savedSize = size
1268         }
1269 }
1270
1271 // CaptureOutput saves data from the container's output directory if
1272 // needed, and updates the container output accordingly.
1273 func (runner *ContainerRunner) CaptureOutput() error {
1274         if runner.Container.RuntimeConstraints.API {
1275                 // Output may have been set directly by the container, so
1276                 // refresh the container record to check.
1277                 err := runner.DispatcherArvClient.Get("containers", runner.Container.UUID,
1278                         nil, &runner.Container)
1279                 if err != nil {
1280                         return err
1281                 }
1282                 if runner.Container.Output != "" {
1283                         // Container output is already set.
1284                         runner.OutputPDH = &runner.Container.Output
1285                         return nil
1286                 }
1287         }
1288
1289         txt, err := (&copier{
1290                 client:        runner.containerClient,
1291                 arvClient:     runner.ContainerArvClient,
1292                 keepClient:    runner.ContainerKeepClient,
1293                 hostOutputDir: runner.HostOutputDir,
1294                 ctrOutputDir:  runner.Container.OutputPath,
1295                 binds:         runner.Binds,
1296                 mounts:        runner.Container.Mounts,
1297                 secretMounts:  runner.SecretMounts,
1298                 logger:        runner.CrunchLog,
1299         }).Copy()
1300         if err != nil {
1301                 return err
1302         }
1303         if n := len(regexp.MustCompile(` [0-9a-f]+\+\S*\+R`).FindAllStringIndex(txt, -1)); n > 0 {
1304                 runner.CrunchLog.Printf("Copying %d data blocks from remote input collections...", n)
1305                 fs, err := (&arvados.Collection{ManifestText: txt}).FileSystem(runner.containerClient, runner.ContainerKeepClient)
1306                 if err != nil {
1307                         return err
1308                 }
1309                 txt, err = fs.MarshalManifest(".")
1310                 if err != nil {
1311                         return err
1312                 }
1313         }
1314         var resp arvados.Collection
1315         err = runner.ContainerArvClient.Create("collections", arvadosclient.Dict{
1316                 "ensure_unique_name": true,
1317                 "collection": arvadosclient.Dict{
1318                         "is_trashed":    true,
1319                         "name":          "output for " + runner.Container.UUID,
1320                         "manifest_text": txt,
1321                 },
1322         }, &resp)
1323         if err != nil {
1324                 return fmt.Errorf("error creating output collection: %v", err)
1325         }
1326         runner.OutputPDH = &resp.PortableDataHash
1327         return nil
1328 }
1329
1330 func (runner *ContainerRunner) CleanupDirs() {
1331         if runner.ArvMount != nil {
1332                 var delay int64 = 8
1333                 umount := exec.Command("arv-mount", fmt.Sprintf("--unmount-timeout=%d", delay), "--unmount", runner.ArvMountPoint)
1334                 umount.Stdout = runner.CrunchLog
1335                 umount.Stderr = runner.CrunchLog
1336                 runner.CrunchLog.Printf("Running %v", umount.Args)
1337                 umnterr := umount.Start()
1338
1339                 if umnterr != nil {
1340                         runner.CrunchLog.Printf("Error unmounting: %v", umnterr)
1341                 } else {
1342                         // If arv-mount --unmount gets stuck for any reason, we
1343                         // don't want to wait for it forever.  Do Wait() in a goroutine
1344                         // so it doesn't block crunch-run.
1345                         umountExit := make(chan error)
1346                         go func() {
1347                                 mnterr := umount.Wait()
1348                                 if mnterr != nil {
1349                                         runner.CrunchLog.Printf("Error unmounting: %v", mnterr)
1350                                 }
1351                                 umountExit <- mnterr
1352                         }()
1353
1354                         for again := true; again; {
1355                                 again = false
1356                                 select {
1357                                 case <-umountExit:
1358                                         umount = nil
1359                                         again = true
1360                                 case <-runner.ArvMountExit:
1361                                         break
1362                                 case <-time.After(time.Duration((delay + 1) * int64(time.Second))):
1363                                         runner.CrunchLog.Printf("Timed out waiting for unmount")
1364                                         if umount != nil {
1365                                                 umount.Process.Kill()
1366                                         }
1367                                         runner.ArvMount.Process.Kill()
1368                                 }
1369                         }
1370                 }
1371         }
1372
1373         if runner.ArvMountPoint != "" {
1374                 if rmerr := os.Remove(runner.ArvMountPoint); rmerr != nil {
1375                         runner.CrunchLog.Printf("While cleaning up arv-mount directory %s: %v", runner.ArvMountPoint, rmerr)
1376                 }
1377         }
1378
1379         if rmerr := os.RemoveAll(runner.parentTemp); rmerr != nil {
1380                 runner.CrunchLog.Printf("While cleaning up temporary directory %s: %v", runner.parentTemp, rmerr)
1381         }
1382 }
1383
1384 // CommitLogs posts the collection containing the final container logs.
1385 func (runner *ContainerRunner) CommitLogs() error {
1386         func() {
1387                 // Hold cStateLock to prevent races on CrunchLog (e.g., stop()).
1388                 runner.cStateLock.Lock()
1389                 defer runner.cStateLock.Unlock()
1390
1391                 runner.CrunchLog.Print(runner.finalState)
1392
1393                 if runner.arvMountLog != nil {
1394                         runner.arvMountLog.Close()
1395                 }
1396                 runner.CrunchLog.Close()
1397
1398                 // Closing CrunchLog above allows them to be committed to Keep at this
1399                 // point, but re-open crunch log with ArvClient in case there are any
1400                 // other further errors (such as failing to write the log to Keep!)
1401                 // while shutting down
1402                 runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{
1403                         ArvClient:     runner.DispatcherArvClient,
1404                         UUID:          runner.Container.UUID,
1405                         loggingStream: "crunch-run",
1406                         writeCloser:   nil,
1407                 })
1408                 runner.CrunchLog.Immediate = log.New(os.Stderr, runner.Container.UUID+" ", 0)
1409         }()
1410
1411         if runner.LogsPDH != nil {
1412                 // If we have already assigned something to LogsPDH,
1413                 // we must be closing the re-opened log, which won't
1414                 // end up getting attached to the container record and
1415                 // therefore doesn't need to be saved as a collection
1416                 // -- it exists only to send logs to other channels.
1417                 return nil
1418         }
1419         saved, err := runner.saveLogCollection(true)
1420         if err != nil {
1421                 return fmt.Errorf("error saving log collection: %s", err)
1422         }
1423         runner.logMtx.Lock()
1424         defer runner.logMtx.Unlock()
1425         runner.LogsPDH = &saved.PortableDataHash
1426         return nil
1427 }
1428
1429 func (runner *ContainerRunner) saveLogCollection(final bool) (response arvados.Collection, err error) {
1430         runner.logMtx.Lock()
1431         defer runner.logMtx.Unlock()
1432         if runner.LogsPDH != nil {
1433                 // Already finalized.
1434                 return
1435         }
1436         updates := arvadosclient.Dict{
1437                 "name": "logs for " + runner.Container.UUID,
1438         }
1439         mt, err1 := runner.LogCollection.MarshalManifest(".")
1440         if err1 == nil {
1441                 // Only send updated manifest text if there was no
1442                 // error.
1443                 updates["manifest_text"] = mt
1444         }
1445
1446         // Even if flushing the manifest had an error, we still want
1447         // to update the log record, if possible, to push the trash_at
1448         // and delete_at times into the future.  Details on bug
1449         // #17293.
1450         if final {
1451                 updates["is_trashed"] = true
1452         } else {
1453                 exp := time.Now().Add(crunchLogUpdatePeriod * 24)
1454                 updates["trash_at"] = exp
1455                 updates["delete_at"] = exp
1456         }
1457         reqBody := arvadosclient.Dict{"collection": updates}
1458         var err2 error
1459         if runner.logUUID == "" {
1460                 reqBody["ensure_unique_name"] = true
1461                 err2 = runner.DispatcherArvClient.Create("collections", reqBody, &response)
1462         } else {
1463                 err2 = runner.DispatcherArvClient.Update("collections", runner.logUUID, reqBody, &response)
1464         }
1465         if err2 == nil {
1466                 runner.logUUID = response.UUID
1467         }
1468
1469         if err1 != nil || err2 != nil {
1470                 err = fmt.Errorf("error recording logs: %q, %q", err1, err2)
1471         }
1472         return
1473 }
1474
1475 // UpdateContainerRunning updates the container state to "Running"
1476 func (runner *ContainerRunner) UpdateContainerRunning() error {
1477         runner.cStateLock.Lock()
1478         defer runner.cStateLock.Unlock()
1479         if runner.cCancelled {
1480                 return ErrCancelled
1481         }
1482         return runner.DispatcherArvClient.Update("containers", runner.Container.UUID,
1483                 arvadosclient.Dict{"container": arvadosclient.Dict{"state": "Running", "gateway_address": runner.gateway.Address}}, nil)
1484 }
1485
1486 // ContainerToken returns the api_token the container (and any
1487 // arv-mount processes) are allowed to use.
1488 func (runner *ContainerRunner) ContainerToken() (string, error) {
1489         if runner.token != "" {
1490                 return runner.token, nil
1491         }
1492
1493         var auth arvados.APIClientAuthorization
1494         err := runner.DispatcherArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth)
1495         if err != nil {
1496                 return "", err
1497         }
1498         runner.token = fmt.Sprintf("v2/%s/%s/%s", auth.UUID, auth.APIToken, runner.Container.UUID)
1499         return runner.token, nil
1500 }
1501
1502 // UpdateContainerFinal updates the container record state on API
1503 // server to "Complete" or "Cancelled"
1504 func (runner *ContainerRunner) UpdateContainerFinal() error {
1505         update := arvadosclient.Dict{}
1506         update["state"] = runner.finalState
1507         if runner.LogsPDH != nil {
1508                 update["log"] = *runner.LogsPDH
1509         }
1510         if runner.finalState == "Complete" {
1511                 if runner.ExitCode != nil {
1512                         update["exit_code"] = *runner.ExitCode
1513                 }
1514                 if runner.OutputPDH != nil {
1515                         update["output"] = *runner.OutputPDH
1516                 }
1517         }
1518         return runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{"container": update}, nil)
1519 }
1520
1521 // IsCancelled returns the value of Cancelled, with goroutine safety.
1522 func (runner *ContainerRunner) IsCancelled() bool {
1523         runner.cStateLock.Lock()
1524         defer runner.cStateLock.Unlock()
1525         return runner.cCancelled
1526 }
1527
1528 // NewArvLogWriter creates an ArvLogWriter
1529 func (runner *ContainerRunner) NewArvLogWriter(name string) (io.WriteCloser, error) {
1530         writer, err := runner.LogCollection.OpenFile(name+".txt", os.O_CREATE|os.O_WRONLY, 0666)
1531         if err != nil {
1532                 return nil, err
1533         }
1534         return &ArvLogWriter{
1535                 ArvClient:     runner.DispatcherArvClient,
1536                 UUID:          runner.Container.UUID,
1537                 loggingStream: name,
1538                 writeCloser:   writer,
1539         }, nil
1540 }
1541
1542 // Run the full container lifecycle.
1543 func (runner *ContainerRunner) Run() (err error) {
1544         runner.CrunchLog.Printf("crunch-run %s started", cmd.Version.String())
1545         runner.CrunchLog.Printf("Executing container '%s'", runner.Container.UUID)
1546
1547         hostname, hosterr := os.Hostname()
1548         if hosterr != nil {
1549                 runner.CrunchLog.Printf("Error getting hostname '%v'", hosterr)
1550         } else {
1551                 runner.CrunchLog.Printf("Executing on host '%s'", hostname)
1552         }
1553
1554         runner.finalState = "Queued"
1555
1556         defer func() {
1557                 runner.CleanupDirs()
1558
1559                 runner.CrunchLog.Printf("crunch-run finished")
1560                 runner.CrunchLog.Close()
1561         }()
1562
1563         err = runner.fetchContainerRecord()
1564         if err != nil {
1565                 return
1566         }
1567         if runner.Container.State != "Locked" {
1568                 return fmt.Errorf("dispatch error detected: container %q has state %q", runner.Container.UUID, runner.Container.State)
1569         }
1570
1571         defer func() {
1572                 // checkErr prints e (unless it's nil) and sets err to
1573                 // e (unless err is already non-nil). Thus, if err
1574                 // hasn't already been assigned when Run() returns,
1575                 // this cleanup func will cause Run() to return the
1576                 // first non-nil error that is passed to checkErr().
1577                 checkErr := func(errorIn string, e error) {
1578                         if e == nil {
1579                                 return
1580                         }
1581                         runner.CrunchLog.Printf("error in %s: %v", errorIn, e)
1582                         if err == nil {
1583                                 err = e
1584                         }
1585                         if runner.finalState == "Complete" {
1586                                 // There was an error in the finalization.
1587                                 runner.finalState = "Cancelled"
1588                         }
1589                 }
1590
1591                 // Log the error encountered in Run(), if any
1592                 checkErr("Run", err)
1593
1594                 if runner.finalState == "Queued" {
1595                         runner.UpdateContainerFinal()
1596                         return
1597                 }
1598
1599                 if runner.IsCancelled() {
1600                         runner.finalState = "Cancelled"
1601                         // but don't return yet -- we still want to
1602                         // capture partial output and write logs
1603                 }
1604
1605                 checkErr("CaptureOutput", runner.CaptureOutput())
1606                 checkErr("stopHoststat", runner.stopHoststat())
1607                 checkErr("CommitLogs", runner.CommitLogs())
1608                 checkErr("UpdateContainerFinal", runner.UpdateContainerFinal())
1609         }()
1610
1611         runner.setupSignals()
1612         err = runner.startHoststat()
1613         if err != nil {
1614                 return
1615         }
1616
1617         // check for and/or load image
1618         err = runner.LoadImage()
1619         if err != nil {
1620                 if !runner.checkBrokenNode(err) {
1621                         // Failed to load image but not due to a "broken node"
1622                         // condition, probably user error.
1623                         runner.finalState = "Cancelled"
1624                 }
1625                 err = fmt.Errorf("While loading container image: %v", err)
1626                 return
1627         }
1628
1629         // set up FUSE mount and binds
1630         err = runner.SetupMounts()
1631         if err != nil {
1632                 runner.finalState = "Cancelled"
1633                 err = fmt.Errorf("While setting up mounts: %v", err)
1634                 return
1635         }
1636
1637         err = runner.CreateContainer()
1638         if err != nil {
1639                 return
1640         }
1641         err = runner.LogHostInfo()
1642         if err != nil {
1643                 return
1644         }
1645         err = runner.LogNodeRecord()
1646         if err != nil {
1647                 return
1648         }
1649         err = runner.LogContainerRecord()
1650         if err != nil {
1651                 return
1652         }
1653
1654         if runner.IsCancelled() {
1655                 return
1656         }
1657
1658         err = runner.UpdateContainerRunning()
1659         if err != nil {
1660                 return
1661         }
1662         runner.finalState = "Cancelled"
1663
1664         err = runner.startCrunchstat()
1665         if err != nil {
1666                 return
1667         }
1668
1669         err = runner.StartContainer()
1670         if err != nil {
1671                 runner.checkBrokenNode(err)
1672                 return
1673         }
1674
1675         err = runner.WaitFinish()
1676         if err == nil && !runner.IsCancelled() {
1677                 runner.finalState = "Complete"
1678         }
1679         return
1680 }
1681
1682 // Fetch the current container record (uuid = runner.Container.UUID)
1683 // into runner.Container.
1684 func (runner *ContainerRunner) fetchContainerRecord() error {
1685         reader, err := runner.DispatcherArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
1686         if err != nil {
1687                 return fmt.Errorf("error fetching container record: %v", err)
1688         }
1689         defer reader.Close()
1690
1691         dec := json.NewDecoder(reader)
1692         dec.UseNumber()
1693         err = dec.Decode(&runner.Container)
1694         if err != nil {
1695                 return fmt.Errorf("error decoding container record: %v", err)
1696         }
1697
1698         var sm struct {
1699                 SecretMounts map[string]arvados.Mount `json:"secret_mounts"`
1700         }
1701
1702         containerToken, err := runner.ContainerToken()
1703         if err != nil {
1704                 return fmt.Errorf("error getting container token: %v", err)
1705         }
1706
1707         runner.ContainerArvClient, runner.ContainerKeepClient,
1708                 runner.containerClient, err = runner.MkArvClient(containerToken)
1709         if err != nil {
1710                 return fmt.Errorf("error creating container API client: %v", err)
1711         }
1712
1713         err = runner.ContainerArvClient.Call("GET", "containers", runner.Container.UUID, "secret_mounts", nil, &sm)
1714         if err != nil {
1715                 if apierr, ok := err.(arvadosclient.APIServerError); !ok || apierr.HttpStatusCode != 404 {
1716                         return fmt.Errorf("error fetching secret_mounts: %v", err)
1717                 }
1718                 // ok && apierr.HttpStatusCode == 404, which means
1719                 // secret_mounts isn't supported by this API server.
1720         }
1721         runner.SecretMounts = sm.SecretMounts
1722
1723         return nil
1724 }
1725
1726 // NewContainerRunner creates a new container runner.
1727 func NewContainerRunner(dispatcherClient *arvados.Client,
1728         dispatcherArvClient IArvadosClient,
1729         dispatcherKeepClient IKeepClient,
1730         docker ThinDockerClient,
1731         containerUUID string) (*ContainerRunner, error) {
1732
1733         cr := &ContainerRunner{
1734                 dispatcherClient:     dispatcherClient,
1735                 DispatcherArvClient:  dispatcherArvClient,
1736                 DispatcherKeepClient: dispatcherKeepClient,
1737                 Docker:               docker,
1738         }
1739         cr.NewLogWriter = cr.NewArvLogWriter
1740         cr.RunArvMount = cr.ArvMountCmd
1741         cr.MkTempDir = ioutil.TempDir
1742         cr.MkArvClient = func(token string) (IArvadosClient, IKeepClient, *arvados.Client, error) {
1743                 cl, err := arvadosclient.MakeArvadosClient()
1744                 if err != nil {
1745                         return nil, nil, nil, err
1746                 }
1747                 cl.ApiToken = token
1748                 kc, err := keepclient.MakeKeepClient(cl)
1749                 if err != nil {
1750                         return nil, nil, nil, err
1751                 }
1752                 c2 := arvados.NewClientFromEnv()
1753                 c2.AuthToken = token
1754                 return cl, kc, c2, nil
1755         }
1756         var err error
1757         cr.LogCollection, err = (&arvados.Collection{}).FileSystem(cr.dispatcherClient, cr.DispatcherKeepClient)
1758         if err != nil {
1759                 return nil, err
1760         }
1761         cr.Container.UUID = containerUUID
1762         w, err := cr.NewLogWriter("crunch-run")
1763         if err != nil {
1764                 return nil, err
1765         }
1766         cr.CrunchLog = NewThrottledLogger(w)
1767         cr.CrunchLog.Immediate = log.New(os.Stderr, containerUUID+" ", 0)
1768
1769         loadLogThrottleParams(dispatcherArvClient)
1770         go cr.updateLogs()
1771
1772         return cr, nil
1773 }
1774
1775 func (command) RunCommand(prog string, args []string, stdin io.Reader, stdout, stderr io.Writer) int {
1776         flags := flag.NewFlagSet(prog, flag.ContinueOnError)
1777         statInterval := flags.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting")
1778         cgroupRoot := flags.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree")
1779         cgroupParent := flags.String("cgroup-parent", "docker", "name of container's parent cgroup (ignored if -cgroup-parent-subsystem is used)")
1780         cgroupParentSubsystem := flags.String("cgroup-parent-subsystem", "", "use current cgroup for given subsystem as parent cgroup for container")
1781         caCertsPath := flags.String("ca-certs", "", "Path to TLS root certificates")
1782         detach := flags.Bool("detach", false, "Detach from parent process and run in the background")
1783         stdinEnv := flags.Bool("stdin-env", false, "Load environment variables from JSON message on stdin")
1784         sleep := flags.Duration("sleep", 0, "Delay before starting (testing use only)")
1785         kill := flags.Int("kill", -1, "Send signal to an existing crunch-run process for given UUID")
1786         list := flags.Bool("list", false, "List UUIDs of existing crunch-run processes")
1787         enableNetwork := flags.String("container-enable-networking", "default",
1788                 `Specify if networking should be enabled for container.  One of 'default', 'always':
1789         default: only enable networking if container requests it.
1790         always:  containers always have networking enabled
1791         `)
1792         networkMode := flags.String("container-network-mode", "default",
1793                 `Set networking mode for container.  Corresponds to Docker network mode (--net).
1794         `)
1795         memprofile := flags.String("memprofile", "", "write memory profile to `file` after running container")
1796         flags.Duration("check-containerd", 0, "Ignored. Exists for compatibility with older versions.")
1797
1798         ignoreDetachFlag := false
1799         if len(args) > 0 && args[0] == "-no-detach" {
1800                 // This process was invoked by a parent process, which
1801                 // has passed along its own arguments, including
1802                 // -detach, after the leading -no-detach flag.  Strip
1803                 // the leading -no-detach flag (it's not recognized by
1804                 // flags.Parse()) and ignore the -detach flag that
1805                 // comes later.
1806                 args = args[1:]
1807                 ignoreDetachFlag = true
1808         }
1809
1810         if err := flags.Parse(args); err == flag.ErrHelp {
1811                 return 0
1812         } else if err != nil {
1813                 log.Print(err)
1814                 return 1
1815         }
1816
1817         if *stdinEnv && !ignoreDetachFlag {
1818                 // Load env vars on stdin if asked (but not in a
1819                 // detached child process, in which case stdin is
1820                 // /dev/null).
1821                 err := loadEnv(os.Stdin)
1822                 if err != nil {
1823                         log.Print(err)
1824                         return 1
1825                 }
1826         }
1827
1828         containerID := flags.Arg(0)
1829
1830         switch {
1831         case *detach && !ignoreDetachFlag:
1832                 return Detach(containerID, prog, args, os.Stdout, os.Stderr)
1833         case *kill >= 0:
1834                 return KillProcess(containerID, syscall.Signal(*kill), os.Stdout, os.Stderr)
1835         case *list:
1836                 return ListProcesses(os.Stdout, os.Stderr)
1837         }
1838
1839         if containerID == "" {
1840                 log.Printf("usage: %s [options] UUID", prog)
1841                 return 1
1842         }
1843
1844         log.Printf("crunch-run %s started", cmd.Version.String())
1845         time.Sleep(*sleep)
1846
1847         if *caCertsPath != "" {
1848                 arvadosclient.CertFiles = []string{*caCertsPath}
1849         }
1850
1851         api, err := arvadosclient.MakeArvadosClient()
1852         if err != nil {
1853                 log.Printf("%s: %v", containerID, err)
1854                 return 1
1855         }
1856         api.Retries = 8
1857
1858         kc, kcerr := keepclient.MakeKeepClient(api)
1859         if kcerr != nil {
1860                 log.Printf("%s: %v", containerID, kcerr)
1861                 return 1
1862         }
1863         kc.BlockCache = &keepclient.BlockCache{MaxBlocks: 2}
1864         kc.Retries = 4
1865
1866         // API version 1.21 corresponds to Docker 1.9, which is currently the
1867         // minimum version we want to support.
1868         docker, dockererr := dockerclient.NewClient(dockerclient.DefaultDockerHost, "1.21", nil, nil)
1869
1870         cr, err := NewContainerRunner(arvados.NewClientFromEnv(), api, kc, docker, containerID)
1871         if err != nil {
1872                 log.Print(err)
1873                 return 1
1874         }
1875         if dockererr != nil {
1876                 cr.CrunchLog.Printf("%s: %v", containerID, dockererr)
1877                 cr.checkBrokenNode(dockererr)
1878                 cr.CrunchLog.Close()
1879                 return 1
1880         }
1881
1882         cr.gateway = Gateway{
1883                 Address:           os.Getenv("GatewayAddress"),
1884                 AuthSecret:        os.Getenv("GatewayAuthSecret"),
1885                 ContainerUUID:     containerID,
1886                 DockerContainerID: &cr.ContainerID,
1887                 Log:               cr.CrunchLog,
1888         }
1889         os.Unsetenv("GatewayAuthSecret")
1890         if cr.gateway.Address != "" {
1891                 err = cr.gateway.Start()
1892                 if err != nil {
1893                         log.Printf("error starting gateway server: %s", err)
1894                         return 1
1895                 }
1896         }
1897
1898         parentTemp, tmperr := cr.MkTempDir("", "crunch-run."+containerID+".")
1899         if tmperr != nil {
1900                 log.Printf("%s: %v", containerID, tmperr)
1901                 return 1
1902         }
1903
1904         cr.parentTemp = parentTemp
1905         cr.statInterval = *statInterval
1906         cr.cgroupRoot = *cgroupRoot
1907         cr.expectCgroupParent = *cgroupParent
1908         cr.enableNetwork = *enableNetwork
1909         cr.networkMode = *networkMode
1910         if *cgroupParentSubsystem != "" {
1911                 p := findCgroup(*cgroupParentSubsystem)
1912                 cr.setCgroupParent = p
1913                 cr.expectCgroupParent = p
1914         }
1915
1916         runerr := cr.Run()
1917
1918         if *memprofile != "" {
1919                 f, err := os.Create(*memprofile)
1920                 if err != nil {
1921                         log.Printf("could not create memory profile: %s", err)
1922                 }
1923                 runtime.GC() // get up-to-date statistics
1924                 if err := pprof.WriteHeapProfile(f); err != nil {
1925                         log.Printf("could not write memory profile: %s", err)
1926                 }
1927                 closeerr := f.Close()
1928                 if closeerr != nil {
1929                         log.Printf("closing memprofile file: %s", err)
1930                 }
1931         }
1932
1933         if runerr != nil {
1934                 log.Printf("%s: %v", containerID, runerr)
1935                 return 1
1936         }
1937         return 0
1938 }
1939
1940 func loadEnv(rdr io.Reader) error {
1941         buf, err := ioutil.ReadAll(rdr)
1942         if err != nil {
1943                 return fmt.Errorf("read stdin: %s", err)
1944         }
1945         var env map[string]string
1946         err = json.Unmarshal(buf, &env)
1947         if err != nil {
1948                 return fmt.Errorf("decode stdin: %s", err)
1949         }
1950         for k, v := range env {
1951                 err = os.Setenv(k, v)
1952                 if err != nil {
1953                         return fmt.Errorf("setenv(%q): %s", k, err)
1954                 }
1955         }
1956         return nil
1957 }