services/crunch-run/crunchrun.go

   1 // Copyright (C) The Arvados Authors. All rights reserved.
   2 //
   3 // SPDX-License-Identifier: AGPL-3.0
   4
   5 package main
   6
   7 import (
   8         "bytes"
   9         "encoding/json"
  10         "errors"
  11         "flag"
  12         "fmt"
  13         "io"
  14         "io/ioutil"
  15         "log"
  16         "os"
  17         "os/exec"
  18         "os/signal"
  19         "path"
  20         "path/filepath"
  21         "regexp"
  22         "runtime"
  23         "runtime/pprof"
  24         "sort"
  25         "strings"
  26         "sync"
  27         "syscall"
  28         "time"
  29
  30         "git.curoverse.com/arvados.git/lib/crunchstat"
  31         "git.curoverse.com/arvados.git/sdk/go/arvados"
  32         "git.curoverse.com/arvados.git/sdk/go/arvadosclient"
  33         "git.curoverse.com/arvados.git/sdk/go/keepclient"
  34         "git.curoverse.com/arvados.git/sdk/go/manifest"
  35         "golang.org/x/net/context"
  36
  37         dockertypes "github.com/docker/docker/api/types"
  38         dockercontainer "github.com/docker/docker/api/types/container"
  39         dockernetwork "github.com/docker/docker/api/types/network"
  40         dockerclient "github.com/docker/docker/client"
  41 )
  42
  43 var version = "dev"
  44
  45 // IArvadosClient is the minimal Arvados API methods used by crunch-run.
  46 type IArvadosClient interface {
  47         Create(resourceType string, parameters arvadosclient.Dict, output interface{}) error
  48         Get(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  49         Update(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  50         Call(method, resourceType, uuid, action string, parameters arvadosclient.Dict, output interface{}) error
  51         CallRaw(method string, resourceType string, uuid string, action string, parameters arvadosclient.Dict) (reader io.ReadCloser, err error)
  52         Discovery(key string) (interface{}, error)
  53 }
  54
  55 // ErrCancelled is the error returned when the container is cancelled.
  56 var ErrCancelled = errors.New("Cancelled")
  57
  58 // IKeepClient is the minimal Keep API methods used by crunch-run.
  59 type IKeepClient interface {
  60         PutHB(hash string, buf []byte) (string, int, error)
  61         ManifestFileReader(m manifest.Manifest, filename string) (arvados.File, error)
  62         ClearBlockCache()
  63 }
  64
  65 // NewLogWriter is a factory function to create a new log writer.
  66 type NewLogWriter func(name string) io.WriteCloser
  67
  68 type RunArvMount func(args []string, tok string) (*exec.Cmd, error)
  69
  70 type MkTempDir func(string, string) (string, error)
  71
  72 // ThinDockerClient is the minimal Docker client interface used by crunch-run.
  73 type ThinDockerClient interface {
  74         ContainerAttach(ctx context.Context, container string, options dockertypes.ContainerAttachOptions) (dockertypes.HijackedResponse, error)
  75         ContainerCreate(ctx context.Context, config *dockercontainer.Config, hostConfig *dockercontainer.HostConfig,
  76                 networkingConfig *dockernetwork.NetworkingConfig, containerName string) (dockercontainer.ContainerCreateCreatedBody, error)
  77         ContainerStart(ctx context.Context, container string, options dockertypes.ContainerStartOptions) error
  78         ContainerRemove(ctx context.Context, container string, options dockertypes.ContainerRemoveOptions) error
  79         ContainerWait(ctx context.Context, container string, condition dockercontainer.WaitCondition) (<-chan dockercontainer.ContainerWaitOKBody, <-chan error)
  80         ImageInspectWithRaw(ctx context.Context, image string) (dockertypes.ImageInspect, []byte, error)
  81         ImageLoad(ctx context.Context, input io.Reader, quiet bool) (dockertypes.ImageLoadResponse, error)
  82         ImageRemove(ctx context.Context, image string, options dockertypes.ImageRemoveOptions) ([]dockertypes.ImageDeleteResponseItem, error)
  83 }
  84
  85 // ContainerRunner is the main stateful struct used for a single execution of a
  86 // container.
  87 type ContainerRunner struct {
  88         Docker    ThinDockerClient
  89         ArvClient IArvadosClient
  90         Kc        IKeepClient
  91         arvados.Container
  92         ContainerConfig dockercontainer.Config
  93         dockercontainer.HostConfig
  94         token       string
  95         ContainerID string
  96         ExitCode    *int
  97         NewLogWriter
  98         loggingDone   chan bool
  99         CrunchLog     *ThrottledLogger
 100         Stdout        io.WriteCloser
 101         Stderr        io.WriteCloser
 102         LogCollection *CollectionWriter
 103         LogsPDH       *string
 104         RunArvMount
 105         MkTempDir
 106         ArvMount      *exec.Cmd
 107         ArvMountPoint string
 108         HostOutputDir string
 109         Binds         []string
 110         Volumes       map[string]struct{}
 111         OutputPDH     *string
 112         SigChan       chan os.Signal
 113         ArvMountExit  chan error
 114         SecretMounts  map[string]arvados.Mount
 115         MkArvClient   func(token string) (IArvadosClient, error)
 116         finalState    string
 117         parentTemp    string
 118
 119         statLogger       io.WriteCloser
 120         statReporter     *crunchstat.Reporter
 121         hoststatLogger   io.WriteCloser
 122         hoststatReporter *crunchstat.Reporter
 123         statInterval     time.Duration
 124         cgroupRoot       string
 125         // What we expect the container's cgroup parent to be.
 126         expectCgroupParent string
 127         // What we tell docker to use as the container's cgroup
 128         // parent. Note: Ideally we would use the same field for both
 129         // expectCgroupParent and setCgroupParent, and just make it
 130         // default to "docker". However, when using docker < 1.10 with
 131         // systemd, specifying a non-empty cgroup parent (even the
 132         // default value "docker") hits a docker bug
 133         // (https://github.com/docker/docker/issues/17126). Using two
 134         // separate fields makes it possible to use the "expect cgroup
 135         // parent to be X" feature even on sites where the "specify
 136         // cgroup parent" feature breaks.
 137         setCgroupParent string
 138
 139         cStateLock sync.Mutex
 140         cCancelled bool // StopContainer() invoked
 141
 142         enableNetwork string // one of "default" or "always"
 143         networkMode   string // passed through to HostConfig.NetworkMode
 144         arvMountLog   *ThrottledLogger
 145 }
 146
 147 // setupSignals sets up signal handling to gracefully terminate the underlying
 148 // Docker container and update state when receiving a TERM, INT or QUIT signal.
 149 func (runner *ContainerRunner) setupSignals() {
 150         runner.SigChan = make(chan os.Signal, 1)
 151         signal.Notify(runner.SigChan, syscall.SIGTERM)
 152         signal.Notify(runner.SigChan, syscall.SIGINT)
 153         signal.Notify(runner.SigChan, syscall.SIGQUIT)
 154
 155         go func(sig chan os.Signal) {
 156                 for s := range sig {
 157                         runner.stop(s)
 158                 }
 159         }(runner.SigChan)
 160 }
 161
 162 // stop the underlying Docker container.
 163 func (runner *ContainerRunner) stop(sig os.Signal) {
 164         runner.cStateLock.Lock()
 165         defer runner.cStateLock.Unlock()
 166         if sig != nil {
 167                 runner.CrunchLog.Printf("caught signal: %v", sig)
 168         }
 169         if runner.ContainerID == "" {
 170                 return
 171         }
 172         runner.cCancelled = true
 173         runner.CrunchLog.Printf("removing container")
 174         err := runner.Docker.ContainerRemove(context.TODO(), runner.ContainerID, dockertypes.ContainerRemoveOptions{Force: true})
 175         if err != nil {
 176                 runner.CrunchLog.Printf("error removing container: %s", err)
 177         }
 178 }
 179
 180 var errorBlacklist = []string{
 181         "(?ms).*[Cc]annot connect to the Docker daemon.*",
 182         "(?ms).*oci runtime error.*starting container process.*container init.*mounting.*to rootfs.*no such file or directory.*",
 183 }
 184 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)")
 185
 186 func (runner *ContainerRunner) checkBrokenNode(goterr error) bool {
 187         for _, d := range errorBlacklist {
 188                 if m, e := regexp.MatchString(d, goterr.Error()); m && e == nil {
 189                         runner.CrunchLog.Printf("Error suggests node is unable to run containers: %v", goterr)
 190                         if *brokenNodeHook == "" {
 191                                 runner.CrunchLog.Printf("No broken node hook provided, cannot mark node as broken.")
 192                         } else {
 193                                 runner.CrunchLog.Printf("Running broken node hook %q", *brokenNodeHook)
 194                                 // run killme script
 195                                 c := exec.Command(*brokenNodeHook)
 196                                 c.Stdout = runner.CrunchLog
 197                                 c.Stderr = runner.CrunchLog
 198                                 err := c.Run()
 199                                 if err != nil {
 200                                         runner.CrunchLog.Printf("Error running broken node hook: %v", err)
 201                                 }
 202                         }
 203                         return true
 204                 }
 205         }
 206         return false
 207 }
 208
 209 // LoadImage determines the docker image id from the container record and
 210 // checks if it is available in the local Docker image store.  If not, it loads
 211 // the image from Keep.
 212 func (runner *ContainerRunner) LoadImage() (err error) {
 213
 214         runner.CrunchLog.Printf("Fetching Docker image from collection '%s'", runner.Container.ContainerImage)
 215
 216         var collection arvados.Collection
 217         err = runner.ArvClient.Get("collections", runner.Container.ContainerImage, nil, &collection)
 218         if err != nil {
 219                 return fmt.Errorf("While getting container image collection: %v", err)
 220         }
 221         manifest := manifest.Manifest{Text: collection.ManifestText}
 222         var img, imageID string
 223         for ms := range manifest.StreamIter() {
 224                 img = ms.FileStreamSegments[0].Name
 225                 if !strings.HasSuffix(img, ".tar") {
 226                         return fmt.Errorf("First file in the container image collection does not end in .tar")
 227                 }
 228                 imageID = img[:len(img)-4]
 229         }
 230
 231         runner.CrunchLog.Printf("Using Docker image id '%s'", imageID)
 232
 233         _, _, err = runner.Docker.ImageInspectWithRaw(context.TODO(), imageID)
 234         if err != nil {
 235                 runner.CrunchLog.Print("Loading Docker image from keep")
 236
 237                 var readCloser io.ReadCloser
 238                 readCloser, err = runner.Kc.ManifestFileReader(manifest, img)
 239                 if err != nil {
 240                         return fmt.Errorf("While creating ManifestFileReader for container image: %v", err)
 241                 }
 242
 243                 response, err := runner.Docker.ImageLoad(context.TODO(), readCloser, true)
 244                 if err != nil {
 245                         return fmt.Errorf("While loading container image into Docker: %v", err)
 246                 }
 247
 248                 defer response.Body.Close()
 249                 rbody, err := ioutil.ReadAll(response.Body)
 250                 if err != nil {
 251                         return fmt.Errorf("Reading response to image load: %v", err)
 252                 }
 253                 runner.CrunchLog.Printf("Docker response: %s", rbody)
 254         } else {
 255                 runner.CrunchLog.Print("Docker image is available")
 256         }
 257
 258         runner.ContainerConfig.Image = imageID
 259
 260         runner.Kc.ClearBlockCache()
 261
 262         return nil
 263 }
 264
 265 func (runner *ContainerRunner) ArvMountCmd(arvMountCmd []string, token string) (c *exec.Cmd, err error) {
 266         c = exec.Command("arv-mount", arvMountCmd...)
 267
 268         // Copy our environment, but override ARVADOS_API_TOKEN with
 269         // the container auth token.
 270         c.Env = nil
 271         for _, s := range os.Environ() {
 272                 if !strings.HasPrefix(s, "ARVADOS_API_TOKEN=") {
 273                         c.Env = append(c.Env, s)
 274                 }
 275         }
 276         c.Env = append(c.Env, "ARVADOS_API_TOKEN="+token)
 277
 278         runner.arvMountLog = NewThrottledLogger(runner.NewLogWriter("arv-mount"))
 279         c.Stdout = runner.arvMountLog
 280         c.Stderr = runner.arvMountLog
 281
 282         runner.CrunchLog.Printf("Running %v", c.Args)
 283
 284         err = c.Start()
 285         if err != nil {
 286                 return nil, err
 287         }
 288
 289         statReadme := make(chan bool)
 290         runner.ArvMountExit = make(chan error)
 291
 292         keepStatting := true
 293         go func() {
 294                 for keepStatting {
 295                         time.Sleep(100 * time.Millisecond)
 296                         _, err = os.Stat(fmt.Sprintf("%s/by_id/README", runner.ArvMountPoint))
 297                         if err == nil {
 298                                 keepStatting = false
 299                                 statReadme <- true
 300                         }
 301                 }
 302                 close(statReadme)
 303         }()
 304
 305         go func() {
 306                 mnterr := c.Wait()
 307                 if mnterr != nil {
 308                         runner.CrunchLog.Printf("Arv-mount exit error: %v", mnterr)
 309                 }
 310                 runner.ArvMountExit <- mnterr
 311                 close(runner.ArvMountExit)
 312         }()
 313
 314         select {
 315         case <-statReadme:
 316                 break
 317         case err := <-runner.ArvMountExit:
 318                 runner.ArvMount = nil
 319                 keepStatting = false
 320                 return nil, err
 321         }
 322
 323         return c, nil
 324 }
 325
 326 func (runner *ContainerRunner) SetupArvMountPoint(prefix string) (err error) {
 327         if runner.ArvMountPoint == "" {
 328                 runner.ArvMountPoint, err = runner.MkTempDir(runner.parentTemp, prefix)
 329         }
 330         return
 331 }
 332
 333 func copyfile(src string, dst string) (err error) {
 334         srcfile, err := os.Open(src)
 335         if err != nil {
 336                 return
 337         }
 338
 339         os.MkdirAll(path.Dir(dst), 0777)
 340
 341         dstfile, err := os.Create(dst)
 342         if err != nil {
 343                 return
 344         }
 345         _, err = io.Copy(dstfile, srcfile)
 346         if err != nil {
 347                 return
 348         }
 349
 350         err = srcfile.Close()
 351         err2 := dstfile.Close()
 352
 353         if err != nil {
 354                 return
 355         }
 356
 357         if err2 != nil {
 358                 return err2
 359         }
 360
 361         return nil
 362 }
 363
 364 func (runner *ContainerRunner) SetupMounts() (err error) {
 365         err = runner.SetupArvMountPoint("keep")
 366         if err != nil {
 367                 return fmt.Errorf("While creating keep mount temp dir: %v", err)
 368         }
 369
 370         token, err := runner.ContainerToken()
 371         if err != nil {
 372                 return fmt.Errorf("could not get container token: %s", err)
 373         }
 374
 375         pdhOnly := true
 376         tmpcount := 0
 377         arvMountCmd := []string{
 378                 "--foreground",
 379                 "--allow-other",
 380                 "--read-write",
 381                 fmt.Sprintf("--crunchstat-interval=%v", runner.statInterval.Seconds())}
 382
 383         if runner.Container.RuntimeConstraints.KeepCacheRAM > 0 {
 384                 arvMountCmd = append(arvMountCmd, "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheRAM))
 385         }
 386
 387         collectionPaths := []string{}
 388         runner.Binds = nil
 389         runner.Volumes = make(map[string]struct{})
 390         needCertMount := true
 391         type copyFile struct {
 392                 src  string
 393                 bind string
 394         }
 395         var copyFiles []copyFile
 396
 397         var binds []string
 398         for bind := range runner.Container.Mounts {
 399                 binds = append(binds, bind)
 400         }
 401         for bind := range runner.SecretMounts {
 402                 if _, ok := runner.Container.Mounts[bind]; ok {
 403                         return fmt.Errorf("Secret mount %q conflicts with regular mount", bind)
 404                 }
 405                 if runner.SecretMounts[bind].Kind != "json" &&
 406                         runner.SecretMounts[bind].Kind != "text" {
 407                         return fmt.Errorf("Secret mount %q type is %q but only 'json' and 'text' are permitted.",
 408                                 bind, runner.SecretMounts[bind].Kind)
 409                 }
 410                 binds = append(binds, bind)
 411         }
 412         sort.Strings(binds)
 413
 414         for _, bind := range binds {
 415                 mnt, ok := runner.Container.Mounts[bind]
 416                 if !ok {
 417                         mnt = runner.SecretMounts[bind]
 418                 }
 419                 if bind == "stdout" || bind == "stderr" {
 420                         // Is it a "file" mount kind?
 421                         if mnt.Kind != "file" {
 422                                 return fmt.Errorf("Unsupported mount kind '%s' for %s. Only 'file' is supported.", mnt.Kind, bind)
 423                         }
 424
 425                         // Does path start with OutputPath?
 426                         prefix := runner.Container.OutputPath
 427                         if !strings.HasSuffix(prefix, "/") {
 428                                 prefix += "/"
 429                         }
 430                         if !strings.HasPrefix(mnt.Path, prefix) {
 431                                 return fmt.Errorf("%s path does not start with OutputPath: %s, %s", strings.Title(bind), mnt.Path, prefix)
 432                         }
 433                 }
 434
 435                 if bind == "stdin" {
 436                         // Is it a "collection" mount kind?
 437                         if mnt.Kind != "collection" && mnt.Kind != "json" {
 438                                 return fmt.Errorf("Unsupported mount kind '%s' for stdin. Only 'collection' or 'json' are supported.", mnt.Kind)
 439                         }
 440                 }
 441
 442                 if bind == "/etc/arvados/ca-certificates.crt" {
 443                         needCertMount = false
 444                 }
 445
 446                 if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && bind != runner.Container.OutputPath+"/" {
 447                         if mnt.Kind != "collection" && mnt.Kind != "text" && mnt.Kind != "json" {
 448                                 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)
 449                         }
 450                 }
 451
 452                 switch {
 453                 case mnt.Kind == "collection" && bind != "stdin":
 454                         var src string
 455                         if mnt.UUID != "" && mnt.PortableDataHash != "" {
 456                                 return fmt.Errorf("Cannot specify both 'uuid' and 'portable_data_hash' for a collection mount")
 457                         }
 458                         if mnt.UUID != "" {
 459                                 if mnt.Writable {
 460                                         return fmt.Errorf("Writing to existing collections currently not permitted.")
 461                                 }
 462                                 pdhOnly = false
 463                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.UUID)
 464                         } else if mnt.PortableDataHash != "" {
 465                                 if mnt.Writable && !strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 466                                         return fmt.Errorf("Can never write to a collection specified by portable data hash")
 467                                 }
 468                                 idx := strings.Index(mnt.PortableDataHash, "/")
 469                                 if idx > 0 {
 470                                         mnt.Path = path.Clean(mnt.PortableDataHash[idx:])
 471                                         mnt.PortableDataHash = mnt.PortableDataHash[0:idx]
 472                                         runner.Container.Mounts[bind] = mnt
 473                                 }
 474                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.PortableDataHash)
 475                                 if mnt.Path != "" && mnt.Path != "." {
 476                                         if strings.HasPrefix(mnt.Path, "./") {
 477                                                 mnt.Path = mnt.Path[2:]
 478                                         } else if strings.HasPrefix(mnt.Path, "/") {
 479                                                 mnt.Path = mnt.Path[1:]
 480                                         }
 481                                         src += "/" + mnt.Path
 482                                 }
 483                         } else {
 484                                 src = fmt.Sprintf("%s/tmp%d", runner.ArvMountPoint, tmpcount)
 485                                 arvMountCmd = append(arvMountCmd, "--mount-tmp")
 486                                 arvMountCmd = append(arvMountCmd, fmt.Sprintf("tmp%d", tmpcount))
 487                                 tmpcount += 1
 488                         }
 489                         if mnt.Writable {
 490                                 if bind == runner.Container.OutputPath {
 491                                         runner.HostOutputDir = src
 492                                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", src, bind))
 493                                 } else if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 494                                         copyFiles = append(copyFiles, copyFile{src, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
 495                                 } else {
 496                                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", src, bind))
 497                                 }
 498                         } else {
 499                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", src, bind))
 500                         }
 501                         collectionPaths = append(collectionPaths, src)
 502
 503                 case mnt.Kind == "tmp":
 504                         var tmpdir string
 505                         tmpdir, err = runner.MkTempDir(runner.parentTemp, "tmp")
 506                         if err != nil {
 507                                 return fmt.Errorf("While creating mount temp dir: %v", err)
 508                         }
 509                         st, staterr := os.Stat(tmpdir)
 510                         if staterr != nil {
 511                                 return fmt.Errorf("While Stat on temp dir: %v", staterr)
 512                         }
 513                         err = os.Chmod(tmpdir, st.Mode()|os.ModeSetgid|0777)
 514                         if staterr != nil {
 515                                 return fmt.Errorf("While Chmod temp dir: %v", err)
 516                         }
 517                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", tmpdir, bind))
 518                         if bind == runner.Container.OutputPath {
 519                                 runner.HostOutputDir = tmpdir
 520                         }
 521
 522                 case mnt.Kind == "json" || mnt.Kind == "text":
 523                         var filedata []byte
 524                         if mnt.Kind == "json" {
 525                                 filedata, err = json.Marshal(mnt.Content)
 526                                 if err != nil {
 527                                         return fmt.Errorf("encoding json data: %v", err)
 528                                 }
 529                         } else {
 530                                 text, ok := mnt.Content.(string)
 531                                 if !ok {
 532                                         return fmt.Errorf("content for mount %q must be a string", bind)
 533                                 }
 534                                 filedata = []byte(text)
 535                         }
 536
 537                         tmpdir, err := runner.MkTempDir(runner.parentTemp, mnt.Kind)
 538                         if err != nil {
 539                                 return fmt.Errorf("creating temp dir: %v", err)
 540                         }
 541                         tmpfn := filepath.Join(tmpdir, "mountdata."+mnt.Kind)
 542                         err = ioutil.WriteFile(tmpfn, filedata, 0444)
 543                         if err != nil {
 544                                 return fmt.Errorf("writing temp file: %v", err)
 545                         }
 546                         if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 547                                 copyFiles = append(copyFiles, copyFile{tmpfn, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
 548                         } else {
 549                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", tmpfn, bind))
 550                         }
 551
 552                 case mnt.Kind == "git_tree":
 553                         tmpdir, err := runner.MkTempDir(runner.parentTemp, "git_tree")
 554                         if err != nil {
 555                                 return fmt.Errorf("creating temp dir: %v", err)
 556                         }
 557                         err = gitMount(mnt).extractTree(runner.ArvClient, tmpdir, token)
 558                         if err != nil {
 559                                 return err
 560                         }
 561                         runner.Binds = append(runner.Binds, tmpdir+":"+bind+":ro")
 562                 }
 563         }
 564
 565         if runner.HostOutputDir == "" {
 566                 return fmt.Errorf("Output path does not correspond to a writable mount point")
 567         }
 568
 569         if wantAPI := runner.Container.RuntimeConstraints.API; needCertMount && wantAPI != nil && *wantAPI {
 570                 for _, certfile := range arvadosclient.CertFiles {
 571                         _, err := os.Stat(certfile)
 572                         if err == nil {
 573                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:/etc/arvados/ca-certificates.crt:ro", certfile))
 574                                 break
 575                         }
 576                 }
 577         }
 578
 579         if pdhOnly {
 580                 arvMountCmd = append(arvMountCmd, "--mount-by-pdh", "by_id")
 581         } else {
 582                 arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_id")
 583         }
 584         arvMountCmd = append(arvMountCmd, runner.ArvMountPoint)
 585
 586         runner.ArvMount, err = runner.RunArvMount(arvMountCmd, token)
 587         if err != nil {
 588                 return fmt.Errorf("While trying to start arv-mount: %v", err)
 589         }
 590
 591         for _, p := range collectionPaths {
 592                 _, err = os.Stat(p)
 593                 if err != nil {
 594                         return fmt.Errorf("While checking that input files exist: %v", err)
 595                 }
 596         }
 597
 598         for _, cp := range copyFiles {
 599                 st, err := os.Stat(cp.src)
 600                 if err != nil {
 601                         return fmt.Errorf("While staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 602                 }
 603                 if st.IsDir() {
 604                         err = filepath.Walk(cp.src, func(walkpath string, walkinfo os.FileInfo, walkerr error) error {
 605                                 if walkerr != nil {
 606                                         return walkerr
 607                                 }
 608                                 target := path.Join(cp.bind, walkpath[len(cp.src):])
 609                                 if walkinfo.Mode().IsRegular() {
 610                                         copyerr := copyfile(walkpath, target)
 611                                         if copyerr != nil {
 612                                                 return copyerr
 613                                         }
 614                                         return os.Chmod(target, walkinfo.Mode()|0777)
 615                                 } else if walkinfo.Mode().IsDir() {
 616                                         mkerr := os.MkdirAll(target, 0777)
 617                                         if mkerr != nil {
 618                                                 return mkerr
 619                                         }
 620                                         return os.Chmod(target, walkinfo.Mode()|os.ModeSetgid|0777)
 621                                 } else {
 622                                         return fmt.Errorf("Source %q is not a regular file or directory", cp.src)
 623                                 }
 624                         })
 625                 } else if st.Mode().IsRegular() {
 626                         err = copyfile(cp.src, cp.bind)
 627                         if err == nil {
 628                                 err = os.Chmod(cp.bind, st.Mode()|0777)
 629                         }
 630                 }
 631                 if err != nil {
 632                         return fmt.Errorf("While staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 633                 }
 634         }
 635
 636         return nil
 637 }
 638
 639 func (runner *ContainerRunner) ProcessDockerAttach(containerReader io.Reader) {
 640         // Handle docker log protocol
 641         // https://docs.docker.com/engine/reference/api/docker_remote_api_v1.15/#attach-to-a-container
 642         defer close(runner.loggingDone)
 643
 644         header := make([]byte, 8)
 645         var err error
 646         for err == nil {
 647                 _, err = io.ReadAtLeast(containerReader, header, 8)
 648                 if err != nil {
 649                         if err == io.EOF {
 650                                 err = nil
 651                         }
 652                         break
 653                 }
 654                 readsize := int64(header[7]) | (int64(header[6]) << 8) | (int64(header[5]) << 16) | (int64(header[4]) << 24)
 655                 if header[0] == 1 {
 656                         // stdout
 657                         _, err = io.CopyN(runner.Stdout, containerReader, readsize)
 658                 } else {
 659                         // stderr
 660                         _, err = io.CopyN(runner.Stderr, containerReader, readsize)
 661                 }
 662         }
 663
 664         if err != nil {
 665                 runner.CrunchLog.Printf("error reading docker logs: %v", err)
 666         }
 667
 668         err = runner.Stdout.Close()
 669         if err != nil {
 670                 runner.CrunchLog.Printf("error closing stdout logs: %v", err)
 671         }
 672
 673         err = runner.Stderr.Close()
 674         if err != nil {
 675                 runner.CrunchLog.Printf("error closing stderr logs: %v", err)
 676         }
 677
 678         if runner.statReporter != nil {
 679                 runner.statReporter.Stop()
 680                 err = runner.statLogger.Close()
 681                 if err != nil {
 682                         runner.CrunchLog.Printf("error closing crunchstat logs: %v", err)
 683                 }
 684         }
 685 }
 686
 687 func (runner *ContainerRunner) stopHoststat() error {
 688         if runner.hoststatReporter == nil {
 689                 return nil
 690         }
 691         runner.hoststatReporter.Stop()
 692         err := runner.hoststatLogger.Close()
 693         if err != nil {
 694                 return fmt.Errorf("error closing hoststat logs: %v", err)
 695         }
 696         return nil
 697 }
 698
 699 func (runner *ContainerRunner) startHoststat() {
 700         runner.hoststatLogger = NewThrottledLogger(runner.NewLogWriter("hoststat"))
 701         runner.hoststatReporter = &crunchstat.Reporter{
 702                 Logger:     log.New(runner.hoststatLogger, "", 0),
 703                 CgroupRoot: runner.cgroupRoot,
 704                 PollPeriod: runner.statInterval,
 705         }
 706         runner.hoststatReporter.Start()
 707 }
 708
 709 func (runner *ContainerRunner) startCrunchstat() {
 710         runner.statLogger = NewThrottledLogger(runner.NewLogWriter("crunchstat"))
 711         runner.statReporter = &crunchstat.Reporter{
 712                 CID:          runner.ContainerID,
 713                 Logger:       log.New(runner.statLogger, "", 0),
 714                 CgroupParent: runner.expectCgroupParent,
 715                 CgroupRoot:   runner.cgroupRoot,
 716                 PollPeriod:   runner.statInterval,
 717         }
 718         runner.statReporter.Start()
 719 }
 720
 721 type infoCommand struct {
 722         label string
 723         cmd   []string
 724 }
 725
 726 // LogHostInfo logs info about the current host, for debugging and
 727 // accounting purposes. Although it's logged as "node-info", this is
 728 // about the environment where crunch-run is actually running, which
 729 // might differ from what's described in the node record (see
 730 // LogNodeRecord).
 731 func (runner *ContainerRunner) LogHostInfo() (err error) {
 732         w := runner.NewLogWriter("node-info")
 733
 734         commands := []infoCommand{
 735                 {
 736                         label: "Host Information",
 737                         cmd:   []string{"uname", "-a"},
 738                 },
 739                 {
 740                         label: "CPU Information",
 741                         cmd:   []string{"cat", "/proc/cpuinfo"},
 742                 },
 743                 {
 744                         label: "Memory Information",
 745                         cmd:   []string{"cat", "/proc/meminfo"},
 746                 },
 747                 {
 748                         label: "Disk Space",
 749                         cmd:   []string{"df", "-m", "/", os.TempDir()},
 750                 },
 751                 {
 752                         label: "Disk INodes",
 753                         cmd:   []string{"df", "-i", "/", os.TempDir()},
 754                 },
 755         }
 756
 757         // Run commands with informational output to be logged.
 758         for _, command := range commands {
 759                 fmt.Fprintln(w, command.label)
 760                 cmd := exec.Command(command.cmd[0], command.cmd[1:]...)
 761                 cmd.Stdout = w
 762                 cmd.Stderr = w
 763                 if err := cmd.Run(); err != nil {
 764                         err = fmt.Errorf("While running command %q: %v", command.cmd, err)
 765                         fmt.Fprintln(w, err)
 766                         return err
 767                 }
 768                 fmt.Fprintln(w, "")
 769         }
 770
 771         err = w.Close()
 772         if err != nil {
 773                 return fmt.Errorf("While closing node-info logs: %v", err)
 774         }
 775         return nil
 776 }
 777
 778 // LogContainerRecord gets and saves the raw JSON container record from the API server
 779 func (runner *ContainerRunner) LogContainerRecord() error {
 780         logged, err := runner.logAPIResponse("container", "containers", map[string]interface{}{"filters": [][]string{{"uuid", "=", runner.Container.UUID}}}, nil)
 781         if !logged && err == nil {
 782                 err = fmt.Errorf("error: no container record found for %s", runner.Container.UUID)
 783         }
 784         return err
 785 }
 786
 787 // LogNodeRecord logs arvados#node record corresponding to the current host.
 788 func (runner *ContainerRunner) LogNodeRecord() error {
 789         hostname := os.Getenv("SLURMD_NODENAME")
 790         if hostname == "" {
 791                 hostname, _ = os.Hostname()
 792         }
 793         _, err := runner.logAPIResponse("node", "nodes", map[string]interface{}{"filters": [][]string{{"hostname", "=", hostname}}}, func(resp interface{}) {
 794                 // The "info" field has admin-only info when obtained
 795                 // with a privileged token, and should not be logged.
 796                 node, ok := resp.(map[string]interface{})
 797                 if ok {
 798                         delete(node, "info")
 799                 }
 800         })
 801         return err
 802 }
 803
 804 func (runner *ContainerRunner) logAPIResponse(label, path string, params map[string]interface{}, munge func(interface{})) (logged bool, err error) {
 805         w := &ArvLogWriter{
 806                 ArvClient:     runner.ArvClient,
 807                 UUID:          runner.Container.UUID,
 808                 loggingStream: label,
 809                 writeCloser:   runner.LogCollection.Open(label + ".json"),
 810         }
 811
 812         reader, err := runner.ArvClient.CallRaw("GET", path, "", "", arvadosclient.Dict(params))
 813         if err != nil {
 814                 return false, fmt.Errorf("error getting %s record: %v", label, err)
 815         }
 816         defer reader.Close()
 817
 818         dec := json.NewDecoder(reader)
 819         dec.UseNumber()
 820         var resp map[string]interface{}
 821         if err = dec.Decode(&resp); err != nil {
 822                 return false, fmt.Errorf("error decoding %s list response: %v", label, err)
 823         }
 824         items, ok := resp["items"].([]interface{})
 825         if !ok {
 826                 return false, fmt.Errorf("error decoding %s list response: no \"items\" key in API list response", label)
 827         } else if len(items) < 1 {
 828                 return false, nil
 829         }
 830         if munge != nil {
 831                 munge(items[0])
 832         }
 833         // Re-encode it using indentation to improve readability
 834         enc := json.NewEncoder(w)
 835         enc.SetIndent("", "    ")
 836         if err = enc.Encode(items[0]); err != nil {
 837                 return false, fmt.Errorf("error logging %s record: %v", label, err)
 838         }
 839         err = w.Close()
 840         if err != nil {
 841                 return false, fmt.Errorf("error closing %s.json in log collection: %v", label, err)
 842         }
 843         return true, nil
 844 }
 845
 846 // AttachStreams connects the docker container stdin, stdout and stderr logs
 847 // to the Arvados logger which logs to Keep and the API server logs table.
 848 func (runner *ContainerRunner) AttachStreams() (err error) {
 849
 850         runner.CrunchLog.Print("Attaching container streams")
 851
 852         // If stdin mount is provided, attach it to the docker container
 853         var stdinRdr arvados.File
 854         var stdinJson []byte
 855         if stdinMnt, ok := runner.Container.Mounts["stdin"]; ok {
 856                 if stdinMnt.Kind == "collection" {
 857                         var stdinColl arvados.Collection
 858                         collId := stdinMnt.UUID
 859                         if collId == "" {
 860                                 collId = stdinMnt.PortableDataHash
 861                         }
 862                         err = runner.ArvClient.Get("collections", collId, nil, &stdinColl)
 863                         if err != nil {
 864                                 return fmt.Errorf("While getting stding collection: %v", err)
 865                         }
 866
 867                         stdinRdr, err = runner.Kc.ManifestFileReader(manifest.Manifest{Text: stdinColl.ManifestText}, stdinMnt.Path)
 868                         if os.IsNotExist(err) {
 869                                 return fmt.Errorf("stdin collection path not found: %v", stdinMnt.Path)
 870                         } else if err != nil {
 871                                 return fmt.Errorf("While getting stdin collection path %v: %v", stdinMnt.Path, err)
 872                         }
 873                 } else if stdinMnt.Kind == "json" {
 874                         stdinJson, err = json.Marshal(stdinMnt.Content)
 875                         if err != nil {
 876                                 return fmt.Errorf("While encoding stdin json data: %v", err)
 877                         }
 878                 }
 879         }
 880
 881         stdinUsed := stdinRdr != nil || len(stdinJson) != 0
 882         response, err := runner.Docker.ContainerAttach(context.TODO(), runner.ContainerID,
 883                 dockertypes.ContainerAttachOptions{Stream: true, Stdin: stdinUsed, Stdout: true, Stderr: true})
 884         if err != nil {
 885                 return fmt.Errorf("While attaching container stdout/stderr streams: %v", err)
 886         }
 887
 888         runner.loggingDone = make(chan bool)
 889
 890         if stdoutMnt, ok := runner.Container.Mounts["stdout"]; ok {
 891                 stdoutFile, err := runner.getStdoutFile(stdoutMnt.Path)
 892                 if err != nil {
 893                         return err
 894                 }
 895                 runner.Stdout = stdoutFile
 896         } else {
 897                 runner.Stdout = NewThrottledLogger(runner.NewLogWriter("stdout"))
 898         }
 899
 900         if stderrMnt, ok := runner.Container.Mounts["stderr"]; ok {
 901                 stderrFile, err := runner.getStdoutFile(stderrMnt.Path)
 902                 if err != nil {
 903                         return err
 904                 }
 905                 runner.Stderr = stderrFile
 906         } else {
 907                 runner.Stderr = NewThrottledLogger(runner.NewLogWriter("stderr"))
 908         }
 909
 910         if stdinRdr != nil {
 911                 go func() {
 912                         _, err := io.Copy(response.Conn, stdinRdr)
 913                         if err != nil {
 914                                 runner.CrunchLog.Print("While writing stdin collection to docker container %q", err)
 915                                 runner.stop(nil)
 916                         }
 917                         stdinRdr.Close()
 918                         response.CloseWrite()
 919                 }()
 920         } else if len(stdinJson) != 0 {
 921                 go func() {
 922                         _, err := io.Copy(response.Conn, bytes.NewReader(stdinJson))
 923                         if err != nil {
 924                                 runner.CrunchLog.Print("While writing stdin json to docker container %q", err)
 925                                 runner.stop(nil)
 926                         }
 927                         response.CloseWrite()
 928                 }()
 929         }
 930
 931         go runner.ProcessDockerAttach(response.Reader)
 932
 933         return nil
 934 }
 935
 936 func (runner *ContainerRunner) getStdoutFile(mntPath string) (*os.File, error) {
 937         stdoutPath := mntPath[len(runner.Container.OutputPath):]
 938         index := strings.LastIndex(stdoutPath, "/")
 939         if index > 0 {
 940                 subdirs := stdoutPath[:index]
 941                 if subdirs != "" {
 942                         st, err := os.Stat(runner.HostOutputDir)
 943                         if err != nil {
 944                                 return nil, fmt.Errorf("While Stat on temp dir: %v", err)
 945                         }
 946                         stdoutPath := filepath.Join(runner.HostOutputDir, subdirs)
 947                         err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777)
 948                         if err != nil {
 949                                 return nil, fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err)
 950                         }
 951                 }
 952         }
 953         stdoutFile, err := os.Create(filepath.Join(runner.HostOutputDir, stdoutPath))
 954         if err != nil {
 955                 return nil, fmt.Errorf("While creating file %q: %v", stdoutPath, err)
 956         }
 957
 958         return stdoutFile, nil
 959 }
 960
 961 // CreateContainer creates the docker container.
 962 func (runner *ContainerRunner) CreateContainer() error {
 963         runner.CrunchLog.Print("Creating Docker container")
 964
 965         runner.ContainerConfig.Cmd = runner.Container.Command
 966         if runner.Container.Cwd != "." {
 967                 runner.ContainerConfig.WorkingDir = runner.Container.Cwd
 968         }
 969
 970         for k, v := range runner.Container.Environment {
 971                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env, k+"="+v)
 972         }
 973
 974         runner.ContainerConfig.Volumes = runner.Volumes
 975
 976         maxRAM := int64(runner.Container.RuntimeConstraints.RAM)
 977         runner.HostConfig = dockercontainer.HostConfig{
 978                 Binds: runner.Binds,
 979                 LogConfig: dockercontainer.LogConfig{
 980                         Type: "none",
 981                 },
 982                 Resources: dockercontainer.Resources{
 983                         CgroupParent: runner.setCgroupParent,
 984                         NanoCPUs:     int64(runner.Container.RuntimeConstraints.VCPUs) * 1000000000,
 985                         Memory:       maxRAM, // RAM
 986                         MemorySwap:   maxRAM, // RAM+swap
 987                         KernelMemory: maxRAM, // kernel portion
 988                 },
 989         }
 990
 991         if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
 992                 tok, err := runner.ContainerToken()
 993                 if err != nil {
 994                         return err
 995                 }
 996                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env,
 997                         "ARVADOS_API_TOKEN="+tok,
 998                         "ARVADOS_API_HOST="+os.Getenv("ARVADOS_API_HOST"),
 999                         "ARVADOS_API_HOST_INSECURE="+os.Getenv("ARVADOS_API_HOST_INSECURE"),
1000                 )
1001                 runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
1002         } else {
1003                 if runner.enableNetwork == "always" {
1004                         runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
1005                 } else {
1006                         runner.HostConfig.NetworkMode = dockercontainer.NetworkMode("none")
1007                 }
1008         }
1009
1010         _, stdinUsed := runner.Container.Mounts["stdin"]
1011         runner.ContainerConfig.OpenStdin = stdinUsed
1012         runner.ContainerConfig.StdinOnce = stdinUsed
1013         runner.ContainerConfig.AttachStdin = stdinUsed
1014         runner.ContainerConfig.AttachStdout = true
1015         runner.ContainerConfig.AttachStderr = true
1016
1017         createdBody, err := runner.Docker.ContainerCreate(context.TODO(), &runner.ContainerConfig, &runner.HostConfig, nil, runner.Container.UUID)
1018         if err != nil {
1019                 return fmt.Errorf("While creating container: %v", err)
1020         }
1021
1022         runner.ContainerID = createdBody.ID
1023
1024         return runner.AttachStreams()
1025 }
1026
1027 // StartContainer starts the docker container created by CreateContainer.
1028 func (runner *ContainerRunner) StartContainer() error {
1029         runner.CrunchLog.Printf("Starting Docker container id '%s'", runner.ContainerID)
1030         runner.cStateLock.Lock()
1031         defer runner.cStateLock.Unlock()
1032         if runner.cCancelled {
1033                 return ErrCancelled
1034         }
1035         err := runner.Docker.ContainerStart(context.TODO(), runner.ContainerID,
1036                 dockertypes.ContainerStartOptions{})
1037         if err != nil {
1038                 var advice string
1039                 if m, e := regexp.MatchString("(?ms).*(exec|System error).*(no such file or directory|file not found).*", err.Error()); m && e == nil {
1040                         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])
1041                 }
1042                 return fmt.Errorf("could not start container: %v%s", err, advice)
1043         }
1044         return nil
1045 }
1046
1047 // WaitFinish waits for the container to terminate, capture the exit code, and
1048 // close the stdout/stderr logging.
1049 func (runner *ContainerRunner) WaitFinish() error {
1050         runner.CrunchLog.Print("Waiting for container to finish")
1051
1052         waitOk, waitErr := runner.Docker.ContainerWait(context.TODO(), runner.ContainerID, dockercontainer.WaitConditionNotRunning)
1053         arvMountExit := runner.ArvMountExit
1054         for {
1055                 select {
1056                 case waitBody := <-waitOk:
1057                         runner.CrunchLog.Printf("Container exited with code: %v", waitBody.StatusCode)
1058                         code := int(waitBody.StatusCode)
1059                         runner.ExitCode = &code
1060
1061                         // wait for stdout/stderr to complete
1062                         <-runner.loggingDone
1063                         return nil
1064
1065                 case err := <-waitErr:
1066                         return fmt.Errorf("container wait: %v", err)
1067
1068                 case <-arvMountExit:
1069                         runner.CrunchLog.Printf("arv-mount exited while container is still running.  Stopping container.")
1070                         runner.stop(nil)
1071                         // arvMountExit will always be ready now that
1072                         // it's closed, but that doesn't interest us.
1073                         arvMountExit = nil
1074                 }
1075         }
1076 }
1077
1078 var ErrNotInOutputDir = fmt.Errorf("Must point to path within the output directory")
1079
1080 func (runner *ContainerRunner) derefOutputSymlink(path string, startinfo os.FileInfo) (tgt string, readlinktgt string, info os.FileInfo, err error) {
1081         // Follow symlinks if necessary
1082         info = startinfo
1083         tgt = path
1084         readlinktgt = ""
1085         nextlink := path
1086         for followed := 0; info.Mode()&os.ModeSymlink != 0; followed++ {
1087                 if followed >= limitFollowSymlinks {
1088                         // Got stuck in a loop or just a pathological number of links, give up.
1089                         err = fmt.Errorf("Followed more than %v symlinks from path %q", limitFollowSymlinks, path)
1090                         return
1091                 }
1092
1093                 readlinktgt, err = os.Readlink(nextlink)
1094                 if err != nil {
1095                         return
1096                 }
1097
1098                 tgt = readlinktgt
1099                 if !strings.HasPrefix(tgt, "/") {
1100                         // Relative symlink, resolve it to host path
1101                         tgt = filepath.Join(filepath.Dir(path), tgt)
1102                 }
1103                 if strings.HasPrefix(tgt, runner.Container.OutputPath+"/") && !strings.HasPrefix(tgt, runner.HostOutputDir+"/") {
1104                         // Absolute symlink to container output path, adjust it to host output path.
1105                         tgt = filepath.Join(runner.HostOutputDir, tgt[len(runner.Container.OutputPath):])
1106                 }
1107                 if !strings.HasPrefix(tgt, runner.HostOutputDir+"/") {
1108                         // After dereferencing, symlink target must either be
1109                         // within output directory, or must point to a
1110                         // collection mount.
1111                         err = ErrNotInOutputDir
1112                         return
1113                 }
1114
1115                 info, err = os.Lstat(tgt)
1116                 if err != nil {
1117                         // tgt
1118                         err = fmt.Errorf("Symlink in output %q points to invalid location %q: %v",
1119                                 path[len(runner.HostOutputDir):], readlinktgt, err)
1120                         return
1121                 }
1122
1123                 nextlink = tgt
1124         }
1125
1126         return
1127 }
1128
1129 var limitFollowSymlinks = 10
1130
1131 // UploadFile uploads files within the output directory, with special handling
1132 // for symlinks. If the symlink leads to a keep mount, copy the manifest text
1133 // from the keep mount into the output manifestText.  Ensure that whether
1134 // symlinks are relative or absolute, every symlink target (even targets that
1135 // are symlinks themselves) must point to a path in either the output directory
1136 // or a collection mount.
1137 //
1138 // Assumes initial value of "path" is absolute, and located within runner.HostOutputDir.
1139 func (runner *ContainerRunner) UploadOutputFile(
1140         path string,
1141         info os.FileInfo,
1142         infoerr error,
1143         binds []string,
1144         walkUpload *WalkUpload,
1145         relocateFrom string,
1146         relocateTo string,
1147         followed int) (manifestText string, err error) {
1148
1149         if infoerr != nil {
1150                 return "", infoerr
1151         }
1152
1153         if info.Mode().IsDir() {
1154                 // if empty, need to create a .keep file
1155                 dir, direrr := os.Open(path)
1156                 if direrr != nil {
1157                         return "", direrr
1158                 }
1159                 defer dir.Close()
1160                 names, eof := dir.Readdirnames(1)
1161                 if len(names) == 0 && eof == io.EOF && path != runner.HostOutputDir {
1162                         containerPath := runner.OutputPath + path[len(runner.HostOutputDir):]
1163                         for _, bind := range binds {
1164                                 mnt := runner.Container.Mounts[bind]
1165                                 // Check if there is a bind for this
1166                                 // directory, in which case assume we don't need .keep
1167                                 if (containerPath == bind || strings.HasPrefix(containerPath, bind+"/")) && mnt.PortableDataHash != "d41d8cd98f00b204e9800998ecf8427e+0" {
1168                                         return
1169                                 }
1170                         }
1171                         outputSuffix := path[len(runner.HostOutputDir)+1:]
1172                         return fmt.Sprintf("./%v d41d8cd98f00b204e9800998ecf8427e+0 0:0:.keep\n", outputSuffix), nil
1173                 }
1174                 return
1175         }
1176
1177         if followed >= limitFollowSymlinks {
1178                 // Got stuck in a loop or just a pathological number of
1179                 // directory links, give up.
1180                 err = fmt.Errorf("Followed more than %v symlinks from path %q", limitFollowSymlinks, path)
1181                 return
1182         }
1183
1184         // "path" is the actual path we are visiting
1185         // "tgt" is the target of "path" (a non-symlink) after following symlinks
1186         // "relocated" is the path in the output manifest where the file should be placed,
1187         // but has HostOutputDir as a prefix.
1188
1189         // The destination path in the output manifest may need to be
1190         // logically relocated to some other path in order to appear
1191         // in the correct location as a result of following a symlink.
1192         // Remove the relocateFrom prefix and replace it with
1193         // relocateTo.
1194         relocated := relocateTo + path[len(relocateFrom):]
1195
1196         tgt, readlinktgt, info, derefErr := runner.derefOutputSymlink(path, info)
1197         if derefErr != nil && derefErr != ErrNotInOutputDir {
1198                 return "", derefErr
1199         }
1200
1201         // go through mounts and try reverse map to collection reference
1202         for _, bind := range binds {
1203                 mnt := runner.Container.Mounts[bind]
1204                 if (tgt == bind || strings.HasPrefix(tgt, bind+"/")) && !mnt.Writable {
1205                         // get path relative to bind
1206                         targetSuffix := tgt[len(bind):]
1207
1208                         // Copy mount and adjust the path to add path relative to the bind
1209                         adjustedMount := mnt
1210                         adjustedMount.Path = filepath.Join(adjustedMount.Path, targetSuffix)
1211
1212                         // Terminates in this keep mount, so add the
1213                         // manifest text at appropriate location.
1214                         outputSuffix := relocated[len(runner.HostOutputDir):]
1215                         manifestText, err = runner.getCollectionManifestForPath(adjustedMount, outputSuffix)
1216                         return
1217                 }
1218         }
1219
1220         // If target is not a collection mount, it must be located within the
1221         // output directory, otherwise it is an error.
1222         if derefErr == ErrNotInOutputDir {
1223                 err = fmt.Errorf("Symlink in output %q points to invalid location %q, must point to path within the output directory.",
1224                         path[len(runner.HostOutputDir):], readlinktgt)
1225                 return
1226         }
1227
1228         if info.Mode().IsRegular() {
1229                 return "", walkUpload.UploadFile(relocated, tgt)
1230         }
1231
1232         if info.Mode().IsDir() {
1233                 // Symlink leads to directory.  Walk() doesn't follow
1234                 // directory symlinks, so we walk the target directory
1235                 // instead.  Within the walk, file paths are relocated
1236                 // so they appear under the original symlink path.
1237                 err = filepath.Walk(tgt, func(walkpath string, walkinfo os.FileInfo, walkerr error) error {
1238                         var m string
1239                         m, walkerr = runner.UploadOutputFile(walkpath, walkinfo, walkerr,
1240                                 binds, walkUpload, tgt, relocated, followed+1)
1241                         if walkerr == nil {
1242                                 manifestText = manifestText + m
1243                         }
1244                         return walkerr
1245                 })
1246                 return
1247         }
1248
1249         return
1250 }
1251
1252 // HandleOutput sets the output, unmounts the FUSE mount, and deletes temporary directories
1253 func (runner *ContainerRunner) CaptureOutput() error {
1254         if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
1255                 // Output may have been set directly by the container, so
1256                 // refresh the container record to check.
1257                 err := runner.ArvClient.Get("containers", runner.Container.UUID,
1258                         nil, &runner.Container)
1259                 if err != nil {
1260                         return err
1261                 }
1262                 if runner.Container.Output != "" {
1263                         // Container output is already set.
1264                         runner.OutputPDH = &runner.Container.Output
1265                         return nil
1266                 }
1267         }
1268
1269         if runner.HostOutputDir == "" {
1270                 return nil
1271         }
1272
1273         _, err := os.Stat(runner.HostOutputDir)
1274         if err != nil {
1275                 return fmt.Errorf("While checking host output path: %v", err)
1276         }
1277
1278         // Pre-populate output from the configured mount points
1279         var binds []string
1280         for bind, mnt := range runner.Container.Mounts {
1281                 if mnt.Kind == "collection" {
1282                         binds = append(binds, bind)
1283                 }
1284         }
1285         sort.Strings(binds)
1286
1287         // Delete secret mounts so they don't get saved to the output collection.
1288         for bind := range runner.SecretMounts {
1289                 if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
1290                         err = os.Remove(runner.HostOutputDir + bind[len(runner.Container.OutputPath):])
1291                         if err != nil {
1292                                 return fmt.Errorf("Unable to remove secret mount: %v", err)
1293                         }
1294                 }
1295         }
1296
1297         var manifestText string
1298
1299         collectionMetafile := fmt.Sprintf("%s/.arvados#collection", runner.HostOutputDir)
1300         _, err = os.Stat(collectionMetafile)
1301         if err != nil {
1302                 // Regular directory
1303
1304                 cw := CollectionWriter{0, runner.Kc, nil, nil, sync.Mutex{}}
1305                 walkUpload := cw.BeginUpload(runner.HostOutputDir, runner.CrunchLog.Logger)
1306
1307                 var m string
1308                 err = filepath.Walk(runner.HostOutputDir, func(path string, info os.FileInfo, err error) error {
1309                         m, err = runner.UploadOutputFile(path, info, err, binds, walkUpload, "", "", 0)
1310                         if err == nil {
1311                                 manifestText = manifestText + m
1312                         }
1313                         return err
1314                 })
1315
1316                 cw.EndUpload(walkUpload)
1317
1318                 if err != nil {
1319                         return fmt.Errorf("While uploading output files: %v", err)
1320                 }
1321
1322                 m, err = cw.ManifestText()
1323                 manifestText = manifestText + m
1324                 if err != nil {
1325                         return fmt.Errorf("While uploading output files: %v", err)
1326                 }
1327         } else {
1328                 // FUSE mount directory
1329                 file, openerr := os.Open(collectionMetafile)
1330                 if openerr != nil {
1331                         return fmt.Errorf("While opening FUSE metafile: %v", err)
1332                 }
1333                 defer file.Close()
1334
1335                 var rec arvados.Collection
1336                 err = json.NewDecoder(file).Decode(&rec)
1337                 if err != nil {
1338                         return fmt.Errorf("While reading FUSE metafile: %v", err)
1339                 }
1340                 manifestText = rec.ManifestText
1341         }
1342
1343         for _, bind := range binds {
1344                 mnt := runner.Container.Mounts[bind]
1345
1346                 bindSuffix := strings.TrimPrefix(bind, runner.Container.OutputPath)
1347
1348                 if bindSuffix == bind || len(bindSuffix) <= 0 {
1349                         // either does not start with OutputPath or is OutputPath itself
1350                         continue
1351                 }
1352
1353                 if mnt.ExcludeFromOutput == true || mnt.Writable {
1354                         continue
1355                 }
1356
1357                 // append to manifest_text
1358                 m, err := runner.getCollectionManifestForPath(mnt, bindSuffix)
1359                 if err != nil {
1360                         return err
1361                 }
1362
1363                 manifestText = manifestText + m
1364         }
1365
1366         // Save output
1367         var response arvados.Collection
1368         manifest := manifest.Manifest{Text: manifestText}
1369         manifestText = manifest.Extract(".", ".").Text
1370         err = runner.ArvClient.Create("collections",
1371                 arvadosclient.Dict{
1372                         "ensure_unique_name": true,
1373                         "collection": arvadosclient.Dict{
1374                                 "is_trashed":    true,
1375                                 "name":          "output for " + runner.Container.UUID,
1376                                 "manifest_text": manifestText}},
1377                 &response)
1378         if err != nil {
1379                 return fmt.Errorf("While creating output collection: %v", err)
1380         }
1381         runner.OutputPDH = &response.PortableDataHash
1382         return nil
1383 }
1384
1385 var outputCollections = make(map[string]arvados.Collection)
1386
1387 // Fetch the collection for the mnt.PortableDataHash
1388 // Return the manifest_text fragment corresponding to the specified mnt.Path
1389 //  after making any required updates.
1390 //  Ex:
1391 //    If mnt.Path is not specified,
1392 //      return the entire manifest_text after replacing any "." with bindSuffix
1393 //    If mnt.Path corresponds to one stream,
1394 //      return the manifest_text for that stream after replacing that stream name with bindSuffix
1395 //    Otherwise, check if a filename in any one stream is being sought. Return the manifest_text
1396 //      for that stream after replacing stream name with bindSuffix minus the last word
1397 //      and the file name with last word of the bindSuffix
1398 //  Allowed path examples:
1399 //    "path":"/"
1400 //    "path":"/subdir1"
1401 //    "path":"/subdir1/subdir2"
1402 //    "path":"/subdir/filename" etc
1403 func (runner *ContainerRunner) getCollectionManifestForPath(mnt arvados.Mount, bindSuffix string) (string, error) {
1404         collection := outputCollections[mnt.PortableDataHash]
1405         if collection.PortableDataHash == "" {
1406                 err := runner.ArvClient.Get("collections", mnt.PortableDataHash, nil, &collection)
1407                 if err != nil {
1408                         return "", fmt.Errorf("While getting collection for %v: %v", mnt.PortableDataHash, err)
1409                 }
1410                 outputCollections[mnt.PortableDataHash] = collection
1411         }
1412
1413         if collection.ManifestText == "" {
1414                 runner.CrunchLog.Printf("No manifest text for collection %v", collection.PortableDataHash)
1415                 return "", nil
1416         }
1417
1418         mft := manifest.Manifest{Text: collection.ManifestText}
1419         extracted := mft.Extract(mnt.Path, bindSuffix)
1420         if extracted.Err != nil {
1421                 return "", fmt.Errorf("Error parsing manifest for %v: %v", mnt.PortableDataHash, extracted.Err.Error())
1422         }
1423         return extracted.Text, nil
1424 }
1425
1426 func (runner *ContainerRunner) CleanupDirs() {
1427         if runner.ArvMount != nil {
1428                 var delay int64 = 8
1429                 umount := exec.Command("arv-mount", fmt.Sprintf("--unmount-timeout=%d", delay), "--unmount", runner.ArvMountPoint)
1430                 umount.Stdout = runner.CrunchLog
1431                 umount.Stderr = runner.CrunchLog
1432                 runner.CrunchLog.Printf("Running %v", umount.Args)
1433                 umnterr := umount.Start()
1434
1435                 if umnterr != nil {
1436                         runner.CrunchLog.Printf("Error unmounting: %v", umnterr)
1437                 } else {
1438                         // If arv-mount --unmount gets stuck for any reason, we
1439                         // don't want to wait for it forever.  Do Wait() in a goroutine
1440                         // so it doesn't block crunch-run.
1441                         umountExit := make(chan error)
1442                         go func() {
1443                                 mnterr := umount.Wait()
1444                                 if mnterr != nil {
1445                                         runner.CrunchLog.Printf("Error unmounting: %v", mnterr)
1446                                 }
1447                                 umountExit <- mnterr
1448                         }()
1449
1450                         for again := true; again; {
1451                                 again = false
1452                                 select {
1453                                 case <-umountExit:
1454                                         umount = nil
1455                                         again = true
1456                                 case <-runner.ArvMountExit:
1457                                         break
1458                                 case <-time.After(time.Duration((delay + 1) * int64(time.Second))):
1459                                         runner.CrunchLog.Printf("Timed out waiting for unmount")
1460                                         if umount != nil {
1461                                                 umount.Process.Kill()
1462                                         }
1463                                         runner.ArvMount.Process.Kill()
1464                                 }
1465                         }
1466                 }
1467         }
1468
1469         if runner.ArvMountPoint != "" {
1470                 if rmerr := os.Remove(runner.ArvMountPoint); rmerr != nil {
1471                         runner.CrunchLog.Printf("While cleaning up arv-mount directory %s: %v", runner.ArvMountPoint, rmerr)
1472                 }
1473         }
1474
1475         if rmerr := os.RemoveAll(runner.parentTemp); rmerr != nil {
1476                 runner.CrunchLog.Printf("While cleaning up temporary directory %s: %v", runner.parentTemp, rmerr)
1477         }
1478 }
1479
1480 // CommitLogs posts the collection containing the final container logs.
1481 func (runner *ContainerRunner) CommitLogs() error {
1482         func() {
1483                 // Hold cStateLock to prevent races on CrunchLog (e.g., stop()).
1484                 runner.cStateLock.Lock()
1485                 defer runner.cStateLock.Unlock()
1486
1487                 runner.CrunchLog.Print(runner.finalState)
1488
1489                 if runner.arvMountLog != nil {
1490                         runner.arvMountLog.Close()
1491                 }
1492                 runner.CrunchLog.Close()
1493
1494                 // Closing CrunchLog above allows them to be committed to Keep at this
1495                 // point, but re-open crunch log with ArvClient in case there are any
1496                 // other further errors (such as failing to write the log to Keep!)
1497                 // while shutting down
1498                 runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{ArvClient: runner.ArvClient,
1499                         UUID: runner.Container.UUID, loggingStream: "crunch-run", writeCloser: nil})
1500                 runner.CrunchLog.Immediate = log.New(os.Stderr, runner.Container.UUID+" ", 0)
1501         }()
1502
1503         if runner.LogsPDH != nil {
1504                 // If we have already assigned something to LogsPDH,
1505                 // we must be closing the re-opened log, which won't
1506                 // end up getting attached to the container record and
1507                 // therefore doesn't need to be saved as a collection
1508                 // -- it exists only to send logs to other channels.
1509                 return nil
1510         }
1511
1512         mt, err := runner.LogCollection.ManifestText()
1513         if err != nil {
1514                 return fmt.Errorf("While creating log manifest: %v", err)
1515         }
1516
1517         var response arvados.Collection
1518         err = runner.ArvClient.Create("collections",
1519                 arvadosclient.Dict{
1520                         "ensure_unique_name": true,
1521                         "collection": arvadosclient.Dict{
1522                                 "is_trashed":    true,
1523                                 "name":          "logs for " + runner.Container.UUID,
1524                                 "manifest_text": mt}},
1525                 &response)
1526         if err != nil {
1527                 return fmt.Errorf("While creating log collection: %v", err)
1528         }
1529         runner.LogsPDH = &response.PortableDataHash
1530         return nil
1531 }
1532
1533 // UpdateContainerRunning updates the container state to "Running"
1534 func (runner *ContainerRunner) UpdateContainerRunning() error {
1535         runner.cStateLock.Lock()
1536         defer runner.cStateLock.Unlock()
1537         if runner.cCancelled {
1538                 return ErrCancelled
1539         }
1540         return runner.ArvClient.Update("containers", runner.Container.UUID,
1541                 arvadosclient.Dict{"container": arvadosclient.Dict{"state": "Running"}}, nil)
1542 }
1543
1544 // ContainerToken returns the api_token the container (and any
1545 // arv-mount processes) are allowed to use.
1546 func (runner *ContainerRunner) ContainerToken() (string, error) {
1547         if runner.token != "" {
1548                 return runner.token, nil
1549         }
1550
1551         var auth arvados.APIClientAuthorization
1552         err := runner.ArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth)
1553         if err != nil {
1554                 return "", err
1555         }
1556         runner.token = auth.APIToken
1557         return runner.token, nil
1558 }
1559
1560 // UpdateContainerComplete updates the container record state on API
1561 // server to "Complete" or "Cancelled"
1562 func (runner *ContainerRunner) UpdateContainerFinal() error {
1563         update := arvadosclient.Dict{}
1564         update["state"] = runner.finalState
1565         if runner.LogsPDH != nil {
1566                 update["log"] = *runner.LogsPDH
1567         }
1568         if runner.finalState == "Complete" {
1569                 if runner.ExitCode != nil {
1570                         update["exit_code"] = *runner.ExitCode
1571                 }
1572                 if runner.OutputPDH != nil {
1573                         update["output"] = *runner.OutputPDH
1574                 }
1575         }
1576         return runner.ArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{"container": update}, nil)
1577 }
1578
1579 // IsCancelled returns the value of Cancelled, with goroutine safety.
1580 func (runner *ContainerRunner) IsCancelled() bool {
1581         runner.cStateLock.Lock()
1582         defer runner.cStateLock.Unlock()
1583         return runner.cCancelled
1584 }
1585
1586 // NewArvLogWriter creates an ArvLogWriter
1587 func (runner *ContainerRunner) NewArvLogWriter(name string) io.WriteCloser {
1588         return &ArvLogWriter{
1589                 ArvClient:     runner.ArvClient,
1590                 UUID:          runner.Container.UUID,
1591                 loggingStream: name,
1592                 writeCloser:   runner.LogCollection.Open(name + ".txt")}
1593 }
1594
1595 // Run the full container lifecycle.
1596 func (runner *ContainerRunner) Run() (err error) {
1597         runner.CrunchLog.Printf("crunch-run %s started", version)
1598         runner.CrunchLog.Printf("Executing container '%s'", runner.Container.UUID)
1599
1600         hostname, hosterr := os.Hostname()
1601         if hosterr != nil {
1602                 runner.CrunchLog.Printf("Error getting hostname '%v'", hosterr)
1603         } else {
1604                 runner.CrunchLog.Printf("Executing on host '%s'", hostname)
1605         }
1606
1607         runner.finalState = "Queued"
1608
1609         defer func() {
1610                 runner.CleanupDirs()
1611
1612                 runner.CrunchLog.Printf("crunch-run finished")
1613                 runner.CrunchLog.Close()
1614         }()
1615
1616         defer func() {
1617                 // checkErr prints e (unless it's nil) and sets err to
1618                 // e (unless err is already non-nil). Thus, if err
1619                 // hasn't already been assigned when Run() returns,
1620                 // this cleanup func will cause Run() to return the
1621                 // first non-nil error that is passed to checkErr().
1622                 checkErr := func(e error) {
1623                         if e == nil {
1624                                 return
1625                         }
1626                         runner.CrunchLog.Print(e)
1627                         if err == nil {
1628                                 err = e
1629                         }
1630                         if runner.finalState == "Complete" {
1631                                 // There was an error in the finalization.
1632                                 runner.finalState = "Cancelled"
1633                         }
1634                 }
1635
1636                 // Log the error encountered in Run(), if any
1637                 checkErr(err)
1638
1639                 if runner.finalState == "Queued" {
1640                         runner.UpdateContainerFinal()
1641                         return
1642                 }
1643
1644                 if runner.IsCancelled() {
1645                         runner.finalState = "Cancelled"
1646                         // but don't return yet -- we still want to
1647                         // capture partial output and write logs
1648                 }
1649
1650                 checkErr(runner.CaptureOutput())
1651                 checkErr(runner.stopHoststat())
1652                 checkErr(runner.CommitLogs())
1653                 checkErr(runner.UpdateContainerFinal())
1654         }()
1655
1656         err = runner.fetchContainerRecord()
1657         if err != nil {
1658                 return
1659         }
1660         runner.setupSignals()
1661         runner.startHoststat()
1662
1663         // check for and/or load image
1664         err = runner.LoadImage()
1665         if err != nil {
1666                 if !runner.checkBrokenNode(err) {
1667                         // Failed to load image but not due to a "broken node"
1668                         // condition, probably user error.
1669                         runner.finalState = "Cancelled"
1670                 }
1671                 err = fmt.Errorf("While loading container image: %v", err)
1672                 return
1673         }
1674
1675         // set up FUSE mount and binds
1676         err = runner.SetupMounts()
1677         if err != nil {
1678                 runner.finalState = "Cancelled"
1679                 err = fmt.Errorf("While setting up mounts: %v", err)
1680                 return
1681         }
1682
1683         err = runner.CreateContainer()
1684         if err != nil {
1685                 return
1686         }
1687         err = runner.LogHostInfo()
1688         if err != nil {
1689                 return
1690         }
1691         err = runner.LogNodeRecord()
1692         if err != nil {
1693                 return
1694         }
1695         err = runner.LogContainerRecord()
1696         if err != nil {
1697                 return
1698         }
1699
1700         if runner.IsCancelled() {
1701                 return
1702         }
1703
1704         err = runner.UpdateContainerRunning()
1705         if err != nil {
1706                 return
1707         }
1708         runner.finalState = "Cancelled"
1709
1710         runner.startCrunchstat()
1711
1712         err = runner.StartContainer()
1713         if err != nil {
1714                 runner.checkBrokenNode(err)
1715                 return
1716         }
1717
1718         err = runner.WaitFinish()
1719         if err == nil && !runner.IsCancelled() {
1720                 runner.finalState = "Complete"
1721         }
1722         return
1723 }
1724
1725 // Fetch the current container record (uuid = runner.Container.UUID)
1726 // into runner.Container.
1727 func (runner *ContainerRunner) fetchContainerRecord() error {
1728         reader, err := runner.ArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
1729         if err != nil {
1730                 return fmt.Errorf("error fetching container record: %v", err)
1731         }
1732         defer reader.Close()
1733
1734         dec := json.NewDecoder(reader)
1735         dec.UseNumber()
1736         err = dec.Decode(&runner.Container)
1737         if err != nil {
1738                 return fmt.Errorf("error decoding container record: %v", err)
1739         }
1740
1741         var sm struct {
1742                 SecretMounts map[string]arvados.Mount `json:"secret_mounts"`
1743         }
1744
1745         containerToken, err := runner.ContainerToken()
1746         if err != nil {
1747                 return fmt.Errorf("error getting container token: %v", err)
1748         }
1749
1750         containerClient, err := runner.MkArvClient(containerToken)
1751         if err != nil {
1752                 return fmt.Errorf("error creating container API client: %v", err)
1753         }
1754
1755         err = containerClient.Call("GET", "containers", runner.Container.UUID, "secret_mounts", nil, &sm)
1756         if err != nil {
1757                 if apierr, ok := err.(arvadosclient.APIServerError); !ok || apierr.HttpStatusCode != 404 {
1758                         return fmt.Errorf("error fetching secret_mounts: %v", err)
1759                 }
1760                 // ok && apierr.HttpStatusCode == 404, which means
1761                 // secret_mounts isn't supported by this API server.
1762         }
1763         runner.SecretMounts = sm.SecretMounts
1764
1765         return nil
1766 }
1767
1768 // NewContainerRunner creates a new container runner.
1769 func NewContainerRunner(api IArvadosClient,
1770         kc IKeepClient,
1771         docker ThinDockerClient,
1772         containerUUID string) *ContainerRunner {
1773
1774         cr := &ContainerRunner{ArvClient: api, Kc: kc, Docker: docker}
1775         cr.NewLogWriter = cr.NewArvLogWriter
1776         cr.RunArvMount = cr.ArvMountCmd
1777         cr.MkTempDir = ioutil.TempDir
1778         cr.MkArvClient = func(token string) (IArvadosClient, error) {
1779                 cl, err := arvadosclient.MakeArvadosClient()
1780                 if err != nil {
1781                         return nil, err
1782                 }
1783                 cl.ApiToken = token
1784                 return cl, nil
1785         }
1786         cr.LogCollection = &CollectionWriter{0, kc, nil, nil, sync.Mutex{}}
1787         cr.Container.UUID = containerUUID
1788         cr.CrunchLog = NewThrottledLogger(cr.NewLogWriter("crunch-run"))
1789         cr.CrunchLog.Immediate = log.New(os.Stderr, containerUUID+" ", 0)
1790
1791         loadLogThrottleParams(api)
1792
1793         return cr
1794 }
1795
1796 func main() {
1797         statInterval := flag.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting")
1798         cgroupRoot := flag.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree")
1799         cgroupParent := flag.String("cgroup-parent", "docker", "name of container's parent cgroup (ignored if -cgroup-parent-subsystem is used)")
1800         cgroupParentSubsystem := flag.String("cgroup-parent-subsystem", "", "use current cgroup for given subsystem as parent cgroup for container")
1801         caCertsPath := flag.String("ca-certs", "", "Path to TLS root certificates")
1802         enableNetwork := flag.String("container-enable-networking", "default",
1803                 `Specify if networking should be enabled for container.  One of 'default', 'always':
1804         default: only enable networking if container requests it.
1805         always:  containers always have networking enabled
1806         `)
1807         networkMode := flag.String("container-network-mode", "default",
1808                 `Set networking mode for container.  Corresponds to Docker network mode (--net).
1809         `)
1810         memprofile := flag.String("memprofile", "", "write memory profile to `file` after running container")
1811         getVersion := flag.Bool("version", false, "Print version information and exit.")
1812         flag.Parse()
1813
1814         // Print version information if requested
1815         if *getVersion {
1816                 fmt.Printf("crunch-run %s\n", version)
1817                 return
1818         }
1819
1820         log.Printf("crunch-run %s started", version)
1821
1822         containerId := flag.Arg(0)
1823
1824         if *caCertsPath != "" {
1825                 arvadosclient.CertFiles = []string{*caCertsPath}
1826         }
1827
1828         api, err := arvadosclient.MakeArvadosClient()
1829         if err != nil {
1830                 log.Fatalf("%s: %v", containerId, err)
1831         }
1832         api.Retries = 8
1833
1834         kc, kcerr := keepclient.MakeKeepClient(api)
1835         if kcerr != nil {
1836                 log.Fatalf("%s: %v", containerId, kcerr)
1837         }
1838         kc.BlockCache = &keepclient.BlockCache{MaxBlocks: 2}
1839         kc.Retries = 4
1840
1841         // API version 1.21 corresponds to Docker 1.9, which is currently the
1842         // minimum version we want to support.
1843         docker, dockererr := dockerclient.NewClient(dockerclient.DefaultDockerHost, "1.21", nil, nil)
1844
1845         cr := NewContainerRunner(api, kc, docker, containerId)
1846         if dockererr != nil {
1847                 cr.CrunchLog.Printf("%s: %v", containerId, dockererr)
1848                 cr.checkBrokenNode(dockererr)
1849                 cr.CrunchLog.Close()
1850                 os.Exit(1)
1851         }
1852
1853         parentTemp, tmperr := cr.MkTempDir("", "crunch-run."+containerId+".")
1854         if tmperr != nil {
1855                 log.Fatalf("%s: %v", containerId, tmperr)
1856         }
1857
1858         cr.parentTemp = parentTemp
1859         cr.statInterval = *statInterval
1860         cr.cgroupRoot = *cgroupRoot
1861         cr.expectCgroupParent = *cgroupParent
1862         cr.enableNetwork = *enableNetwork
1863         cr.networkMode = *networkMode
1864         if *cgroupParentSubsystem != "" {
1865                 p := findCgroup(*cgroupParentSubsystem)
1866                 cr.setCgroupParent = p
1867                 cr.expectCgroupParent = p
1868         }
1869
1870         runerr := cr.Run()
1871
1872         if *memprofile != "" {
1873                 f, err := os.Create(*memprofile)
1874                 if err != nil {
1875                         log.Printf("could not create memory profile: ", err)
1876                 }
1877                 runtime.GC() // get up-to-date statistics
1878                 if err := pprof.WriteHeapProfile(f); err != nil {
1879                         log.Printf("could not write memory profile: ", err)
1880                 }
1881                 closeerr := f.Close()
1882                 if closeerr != nil {
1883                         log.Printf("closing memprofile file: ", err)
1884                 }
1885         }
1886
1887         if runerr != nil {
1888                 log.Fatalf("%s: %v", containerId, runerr)
1889         }
1890 }