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