services/crunch-run/crunchrun.go

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