lib/crunchrun/crunchrun.go

   1 // Copyright (C) The Arvados Authors. All rights reserved.
   2 //
   3 // SPDX-License-Identifier: AGPL-3.0
   4
   5 package crunchrun
   6
   7 import (
   8         "bytes"
   9         "context"
  10         "encoding/json"
  11         "errors"
  12         "flag"
  13         "fmt"
  14         "io"
  15         "io/ioutil"
  16         "log"
  17         "net"
  18         "net/http"
  19         "os"
  20         "os/exec"
  21         "os/signal"
  22         "path"
  23         "path/filepath"
  24         "regexp"
  25         "runtime"
  26         "runtime/pprof"
  27         "sort"
  28         "strings"
  29         "sync"
  30         "syscall"
  31         "time"
  32
  33         "git.arvados.org/arvados.git/lib/cmd"
  34         "git.arvados.org/arvados.git/lib/crunchstat"
  35         "git.arvados.org/arvados.git/sdk/go/arvados"
  36         "git.arvados.org/arvados.git/sdk/go/arvadosclient"
  37         "git.arvados.org/arvados.git/sdk/go/keepclient"
  38         "git.arvados.org/arvados.git/sdk/go/manifest"
  39 )
  40
  41 type command struct{}
  42
  43 var Command = command{}
  44
  45 // ConfigData contains environment variables and (when needed) cluster
  46 // configuration, passed from dispatchcloud to crunch-run on stdin.
  47 type ConfigData struct {
  48         Env         map[string]string
  49         KeepBuffers int
  50         Cluster     *arvados.Cluster
  51 }
  52
  53 // IArvadosClient is the minimal Arvados API methods used by crunch-run.
  54 type IArvadosClient interface {
  55         Create(resourceType string, parameters arvadosclient.Dict, output interface{}) error
  56         Get(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  57         Update(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  58         Call(method, resourceType, uuid, action string, parameters arvadosclient.Dict, output interface{}) error
  59         CallRaw(method string, resourceType string, uuid string, action string, parameters arvadosclient.Dict) (reader io.ReadCloser, err error)
  60         Discovery(key string) (interface{}, error)
  61 }
  62
  63 // ErrCancelled is the error returned when the container is cancelled.
  64 var ErrCancelled = errors.New("Cancelled")
  65
  66 // IKeepClient is the minimal Keep API methods used by crunch-run.
  67 type IKeepClient interface {
  68         BlockWrite(context.Context, arvados.BlockWriteOptions) (arvados.BlockWriteResponse, error)
  69         ReadAt(locator string, p []byte, off int) (int, error)
  70         ManifestFileReader(m manifest.Manifest, filename string) (arvados.File, error)
  71         LocalLocator(locator string) (string, error)
  72         ClearBlockCache()
  73         SetStorageClasses(sc []string)
  74 }
  75
  76 // NewLogWriter is a factory function to create a new log writer.
  77 type NewLogWriter func(name string) (io.WriteCloser, error)
  78
  79 type RunArvMount func(cmdline []string, tok string) (*exec.Cmd, error)
  80
  81 type MkTempDir func(string, string) (string, error)
  82
  83 type PsProcess interface {
  84         CmdlineSlice() ([]string, error)
  85 }
  86
  87 // ContainerRunner is the main stateful struct used for a single execution of a
  88 // container.
  89 type ContainerRunner struct {
  90         executor       containerExecutor
  91         executorStdin  io.Closer
  92         executorStdout io.Closer
  93         executorStderr io.Closer
  94
  95         // Dispatcher client is initialized with the Dispatcher token.
  96         // This is a privileged token used to manage container status
  97         // and logs.
  98         //
  99         // We have both dispatcherClient and DispatcherArvClient
 100         // because there are two different incompatible Arvados Go
 101         // SDKs and we have to use both (hopefully this gets fixed in
 102         // #14467)
 103         dispatcherClient     *arvados.Client
 104         DispatcherArvClient  IArvadosClient
 105         DispatcherKeepClient IKeepClient
 106
 107         // Container client is initialized with the Container token
 108         // This token controls the permissions of the container, and
 109         // must be used for operations such as reading collections.
 110         //
 111         // Same comment as above applies to
 112         // containerClient/ContainerArvClient.
 113         containerClient     *arvados.Client
 114         ContainerArvClient  IArvadosClient
 115         ContainerKeepClient IKeepClient
 116
 117         Container     arvados.Container
 118         token         string
 119         ExitCode      *int
 120         NewLogWriter  NewLogWriter
 121         CrunchLog     *ThrottledLogger
 122         logUUID       string
 123         logMtx        sync.Mutex
 124         LogCollection arvados.CollectionFileSystem
 125         LogsPDH       *string
 126         RunArvMount   RunArvMount
 127         MkTempDir     MkTempDir
 128         ArvMount      *exec.Cmd
 129         ArvMountPoint string
 130         HostOutputDir string
 131         Volumes       map[string]struct{}
 132         OutputPDH     *string
 133         SigChan       chan os.Signal
 134         ArvMountExit  chan error
 135         SecretMounts  map[string]arvados.Mount
 136         MkArvClient   func(token string) (IArvadosClient, IKeepClient, *arvados.Client, error)
 137         finalState    string
 138         parentTemp    string
 139
 140         keepstoreLogger  io.WriteCloser
 141         keepstoreLogbuf  *bufThenWrite
 142         statLogger       io.WriteCloser
 143         statReporter     *crunchstat.Reporter
 144         hoststatLogger   io.WriteCloser
 145         hoststatReporter *crunchstat.Reporter
 146         statInterval     time.Duration
 147         cgroupRoot       string
 148         // What we expect the container's cgroup parent to be.
 149         expectCgroupParent string
 150         // What we tell docker to use as the container's cgroup
 151         // parent. Note: Ideally we would use the same field for both
 152         // expectCgroupParent and setCgroupParent, and just make it
 153         // default to "docker". However, when using docker < 1.10 with
 154         // systemd, specifying a non-empty cgroup parent (even the
 155         // default value "docker") hits a docker bug
 156         // (https://github.com/docker/docker/issues/17126). Using two
 157         // separate fields makes it possible to use the "expect cgroup
 158         // parent to be X" feature even on sites where the "specify
 159         // cgroup parent" feature breaks.
 160         setCgroupParent string
 161
 162         cStateLock sync.Mutex
 163         cCancelled bool // StopContainer() invoked
 164
 165         enableMemoryLimit bool
 166         enableNetwork     string // one of "default" or "always"
 167         networkMode       string // "none", "host", or "" -- passed through to executor
 168         arvMountLog       *ThrottledLogger
 169
 170         containerWatchdogInterval time.Duration
 171
 172         gateway Gateway
 173 }
 174
 175 // setupSignals sets up signal handling to gracefully terminate the
 176 // underlying container and update state when receiving a TERM, INT or
 177 // QUIT signal.
 178 func (runner *ContainerRunner) setupSignals() {
 179         runner.SigChan = make(chan os.Signal, 1)
 180         signal.Notify(runner.SigChan, syscall.SIGTERM)
 181         signal.Notify(runner.SigChan, syscall.SIGINT)
 182         signal.Notify(runner.SigChan, syscall.SIGQUIT)
 183
 184         go func(sig chan os.Signal) {
 185                 for s := range sig {
 186                         runner.stop(s)
 187                 }
 188         }(runner.SigChan)
 189 }
 190
 191 // stop the underlying container.
 192 func (runner *ContainerRunner) stop(sig os.Signal) {
 193         runner.cStateLock.Lock()
 194         defer runner.cStateLock.Unlock()
 195         if sig != nil {
 196                 runner.CrunchLog.Printf("caught signal: %v", sig)
 197         }
 198         runner.cCancelled = true
 199         runner.CrunchLog.Printf("stopping container")
 200         err := runner.executor.Stop()
 201         if err != nil {
 202                 runner.CrunchLog.Printf("error stopping container: %s", err)
 203         }
 204 }
 205
 206 var errorBlacklist = []string{
 207         "(?ms).*[Cc]annot connect to the Docker daemon.*",
 208         "(?ms).*oci runtime error.*starting container process.*container init.*mounting.*to rootfs.*no such file or directory.*",
 209         "(?ms).*grpc: the connection is unavailable.*",
 210 }
 211 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)")
 212
 213 func (runner *ContainerRunner) runBrokenNodeHook() {
 214         if *brokenNodeHook == "" {
 215                 path := filepath.Join(lockdir, brokenfile)
 216                 runner.CrunchLog.Printf("Writing %s to mark node as broken", path)
 217                 f, err := os.OpenFile(path, os.O_CREATE|os.O_RDWR, 0700)
 218                 if err != nil {
 219                         runner.CrunchLog.Printf("Error writing %s: %s", path, err)
 220                         return
 221                 }
 222                 f.Close()
 223         } else {
 224                 runner.CrunchLog.Printf("Running broken node hook %q", *brokenNodeHook)
 225                 // run killme script
 226                 c := exec.Command(*brokenNodeHook)
 227                 c.Stdout = runner.CrunchLog
 228                 c.Stderr = runner.CrunchLog
 229                 err := c.Run()
 230                 if err != nil {
 231                         runner.CrunchLog.Printf("Error running broken node hook: %v", err)
 232                 }
 233         }
 234 }
 235
 236 func (runner *ContainerRunner) checkBrokenNode(goterr error) bool {
 237         for _, d := range errorBlacklist {
 238                 if m, e := regexp.MatchString(d, goterr.Error()); m && e == nil {
 239                         runner.CrunchLog.Printf("Error suggests node is unable to run containers: %v", goterr)
 240                         runner.runBrokenNodeHook()
 241                         return true
 242                 }
 243         }
 244         return false
 245 }
 246
 247 // LoadImage determines the docker image id from the container record and
 248 // checks if it is available in the local Docker image store.  If not, it loads
 249 // the image from Keep.
 250 func (runner *ContainerRunner) LoadImage() (string, error) {
 251         runner.CrunchLog.Printf("Fetching Docker image from collection '%s'", runner.Container.ContainerImage)
 252
 253         d, err := os.Open(runner.ArvMountPoint + "/by_id/" + runner.Container.ContainerImage)
 254         if err != nil {
 255                 return "", err
 256         }
 257         defer d.Close()
 258         allfiles, err := d.Readdirnames(-1)
 259         if err != nil {
 260                 return "", err
 261         }
 262         var tarfiles []string
 263         for _, fnm := range allfiles {
 264                 if strings.HasSuffix(fnm, ".tar") {
 265                         tarfiles = append(tarfiles, fnm)
 266                 }
 267         }
 268         if len(tarfiles) == 0 {
 269                 return "", fmt.Errorf("image collection does not include a .tar image file")
 270         }
 271         if len(tarfiles) > 1 {
 272                 return "", fmt.Errorf("cannot choose from multiple tar files in image collection: %v", tarfiles)
 273         }
 274         imageID := tarfiles[0][:len(tarfiles[0])-4]
 275         imageTarballPath := runner.ArvMountPoint + "/by_id/" + runner.Container.ContainerImage + "/" + imageID + ".tar"
 276         runner.CrunchLog.Printf("Using Docker image id %q", imageID)
 277
 278         runner.CrunchLog.Print("Loading Docker image from keep")
 279         err = runner.executor.LoadImage(imageID, imageTarballPath, runner.Container, runner.ArvMountPoint,
 280                 runner.containerClient)
 281         if err != nil {
 282                 return "", err
 283         }
 284
 285         return imageID, nil
 286 }
 287
 288 func (runner *ContainerRunner) ArvMountCmd(cmdline []string, token string) (c *exec.Cmd, err error) {
 289         c = exec.Command(cmdline[0], cmdline[1:]...)
 290
 291         // Copy our environment, but override ARVADOS_API_TOKEN with
 292         // the container auth token.
 293         c.Env = nil
 294         for _, s := range os.Environ() {
 295                 if !strings.HasPrefix(s, "ARVADOS_API_TOKEN=") {
 296                         c.Env = append(c.Env, s)
 297                 }
 298         }
 299         c.Env = append(c.Env, "ARVADOS_API_TOKEN="+token)
 300
 301         w, err := runner.NewLogWriter("arv-mount")
 302         if err != nil {
 303                 return nil, err
 304         }
 305         runner.arvMountLog = NewThrottledLogger(w)
 306         scanner := logScanner{
 307                 Patterns: []string{
 308                         "Keep write error",
 309                         "Block not found error",
 310                         "Unhandled exception during FUSE operation",
 311                 },
 312                 ReportFunc: runner.reportArvMountWarning,
 313         }
 314         c.Stdout = runner.arvMountLog
 315         c.Stderr = io.MultiWriter(runner.arvMountLog, os.Stderr, &scanner)
 316
 317         runner.CrunchLog.Printf("Running %v", c.Args)
 318
 319         err = c.Start()
 320         if err != nil {
 321                 return nil, err
 322         }
 323
 324         statReadme := make(chan bool)
 325         runner.ArvMountExit = make(chan error)
 326
 327         keepStatting := true
 328         go func() {
 329                 for keepStatting {
 330                         time.Sleep(100 * time.Millisecond)
 331                         _, err = os.Stat(fmt.Sprintf("%s/by_id/README", runner.ArvMountPoint))
 332                         if err == nil {
 333                                 keepStatting = false
 334                                 statReadme <- true
 335                         }
 336                 }
 337                 close(statReadme)
 338         }()
 339
 340         go func() {
 341                 mnterr := c.Wait()
 342                 if mnterr != nil {
 343                         runner.CrunchLog.Printf("Arv-mount exit error: %v", mnterr)
 344                 }
 345                 runner.ArvMountExit <- mnterr
 346                 close(runner.ArvMountExit)
 347         }()
 348
 349         select {
 350         case <-statReadme:
 351                 break
 352         case err := <-runner.ArvMountExit:
 353                 runner.ArvMount = nil
 354                 keepStatting = false
 355                 return nil, err
 356         }
 357
 358         return c, nil
 359 }
 360
 361 func (runner *ContainerRunner) SetupArvMountPoint(prefix string) (err error) {
 362         if runner.ArvMountPoint == "" {
 363                 runner.ArvMountPoint, err = runner.MkTempDir(runner.parentTemp, prefix)
 364         }
 365         return
 366 }
 367
 368 func copyfile(src string, dst string) (err error) {
 369         srcfile, err := os.Open(src)
 370         if err != nil {
 371                 return
 372         }
 373
 374         os.MkdirAll(path.Dir(dst), 0777)
 375
 376         dstfile, err := os.Create(dst)
 377         if err != nil {
 378                 return
 379         }
 380         _, err = io.Copy(dstfile, srcfile)
 381         if err != nil {
 382                 return
 383         }
 384
 385         err = srcfile.Close()
 386         err2 := dstfile.Close()
 387
 388         if err != nil {
 389                 return
 390         }
 391
 392         if err2 != nil {
 393                 return err2
 394         }
 395
 396         return nil
 397 }
 398
 399 func (runner *ContainerRunner) SetupMounts() (map[string]bindmount, error) {
 400         bindmounts := map[string]bindmount{}
 401         err := runner.SetupArvMountPoint("keep")
 402         if err != nil {
 403                 return nil, fmt.Errorf("While creating keep mount temp dir: %v", err)
 404         }
 405
 406         token, err := runner.ContainerToken()
 407         if err != nil {
 408                 return nil, fmt.Errorf("could not get container token: %s", err)
 409         }
 410         runner.CrunchLog.Printf("container token %q", token)
 411
 412         pdhOnly := true
 413         tmpcount := 0
 414         arvMountCmd := []string{
 415                 "arv-mount",
 416                 "--foreground",
 417                 "--read-write",
 418                 "--storage-classes", strings.Join(runner.Container.OutputStorageClasses, ","),
 419                 fmt.Sprintf("--crunchstat-interval=%v", runner.statInterval.Seconds())}
 420
 421         if runner.executor.Runtime() == "docker" {
 422                 arvMountCmd = append(arvMountCmd, "--allow-other")
 423         }
 424
 425         if runner.Container.RuntimeConstraints.KeepCacheRAM > 0 {
 426                 arvMountCmd = append(arvMountCmd, "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheRAM))
 427         }
 428
 429         collectionPaths := []string{}
 430         needCertMount := true
 431         type copyFile struct {
 432                 src  string
 433                 bind string
 434         }
 435         var copyFiles []copyFile
 436
 437         var binds []string
 438         for bind := range runner.Container.Mounts {
 439                 binds = append(binds, bind)
 440         }
 441         for bind := range runner.SecretMounts {
 442                 if _, ok := runner.Container.Mounts[bind]; ok {
 443                         return nil, fmt.Errorf("secret mount %q conflicts with regular mount", bind)
 444                 }
 445                 if runner.SecretMounts[bind].Kind != "json" &&
 446                         runner.SecretMounts[bind].Kind != "text" {
 447                         return nil, fmt.Errorf("secret mount %q type is %q but only 'json' and 'text' are permitted",
 448                                 bind, runner.SecretMounts[bind].Kind)
 449                 }
 450                 binds = append(binds, bind)
 451         }
 452         sort.Strings(binds)
 453
 454         for _, bind := range binds {
 455                 mnt, ok := runner.Container.Mounts[bind]
 456                 if !ok {
 457                         mnt = runner.SecretMounts[bind]
 458                 }
 459                 if bind == "stdout" || bind == "stderr" {
 460                         // Is it a "file" mount kind?
 461                         if mnt.Kind != "file" {
 462                                 return nil, fmt.Errorf("unsupported mount kind '%s' for %s: only 'file' is supported", mnt.Kind, bind)
 463                         }
 464
 465                         // Does path start with OutputPath?
 466                         prefix := runner.Container.OutputPath
 467                         if !strings.HasSuffix(prefix, "/") {
 468                                 prefix += "/"
 469                         }
 470                         if !strings.HasPrefix(mnt.Path, prefix) {
 471                                 return nil, fmt.Errorf("%s path does not start with OutputPath: %s, %s", strings.Title(bind), mnt.Path, prefix)
 472                         }
 473                 }
 474
 475                 if bind == "stdin" {
 476                         // Is it a "collection" mount kind?
 477                         if mnt.Kind != "collection" && mnt.Kind != "json" {
 478                                 return nil, fmt.Errorf("unsupported mount kind '%s' for stdin: only 'collection' and 'json' are supported", mnt.Kind)
 479                         }
 480                 }
 481
 482                 if bind == "/etc/arvados/ca-certificates.crt" {
 483                         needCertMount = false
 484                 }
 485
 486                 if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && bind != runner.Container.OutputPath+"/" {
 487                         if mnt.Kind != "collection" && mnt.Kind != "text" && mnt.Kind != "json" {
 488                                 return nil, fmt.Errorf("only mount points of kind 'collection', 'text' or 'json' are supported underneath the output_path for %q, was %q", bind, mnt.Kind)
 489                         }
 490                 }
 491
 492                 switch {
 493                 case mnt.Kind == "collection" && bind != "stdin":
 494                         var src string
 495                         if mnt.UUID != "" && mnt.PortableDataHash != "" {
 496                                 return nil, fmt.Errorf("cannot specify both 'uuid' and 'portable_data_hash' for a collection mount")
 497                         }
 498                         if mnt.UUID != "" {
 499                                 if mnt.Writable {
 500                                         return nil, fmt.Errorf("writing to existing collections currently not permitted")
 501                                 }
 502                                 pdhOnly = false
 503                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.UUID)
 504                         } else if mnt.PortableDataHash != "" {
 505                                 if mnt.Writable && !strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 506                                         return nil, fmt.Errorf("can never write to a collection specified by portable data hash")
 507                                 }
 508                                 idx := strings.Index(mnt.PortableDataHash, "/")
 509                                 if idx > 0 {
 510                                         mnt.Path = path.Clean(mnt.PortableDataHash[idx:])
 511                                         mnt.PortableDataHash = mnt.PortableDataHash[0:idx]
 512                                         runner.Container.Mounts[bind] = mnt
 513                                 }
 514                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.PortableDataHash)
 515                                 if mnt.Path != "" && mnt.Path != "." {
 516                                         if strings.HasPrefix(mnt.Path, "./") {
 517                                                 mnt.Path = mnt.Path[2:]
 518                                         } else if strings.HasPrefix(mnt.Path, "/") {
 519                                                 mnt.Path = mnt.Path[1:]
 520                                         }
 521                                         src += "/" + mnt.Path
 522                                 }
 523                         } else {
 524                                 src = fmt.Sprintf("%s/tmp%d", runner.ArvMountPoint, tmpcount)
 525                                 arvMountCmd = append(arvMountCmd, "--mount-tmp")
 526                                 arvMountCmd = append(arvMountCmd, fmt.Sprintf("tmp%d", tmpcount))
 527                                 tmpcount++
 528                         }
 529                         if mnt.Writable {
 530                                 if bind == runner.Container.OutputPath {
 531                                         runner.HostOutputDir = src
 532                                         bindmounts[bind] = bindmount{HostPath: src}
 533                                 } else if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 534                                         copyFiles = append(copyFiles, copyFile{src, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
 535                                 } else {
 536                                         bindmounts[bind] = bindmount{HostPath: src}
 537                                 }
 538                         } else {
 539                                 bindmounts[bind] = bindmount{HostPath: src, ReadOnly: true}
 540                         }
 541                         collectionPaths = append(collectionPaths, src)
 542
 543                 case mnt.Kind == "tmp":
 544                         var tmpdir string
 545                         tmpdir, err = runner.MkTempDir(runner.parentTemp, "tmp")
 546                         if err != nil {
 547                                 return nil, fmt.Errorf("while creating mount temp dir: %v", err)
 548                         }
 549                         st, staterr := os.Stat(tmpdir)
 550                         if staterr != nil {
 551                                 return nil, fmt.Errorf("while Stat on temp dir: %v", staterr)
 552                         }
 553                         err = os.Chmod(tmpdir, st.Mode()|os.ModeSetgid|0777)
 554                         if staterr != nil {
 555                                 return nil, fmt.Errorf("while Chmod temp dir: %v", err)
 556                         }
 557                         bindmounts[bind] = bindmount{HostPath: tmpdir}
 558                         if bind == runner.Container.OutputPath {
 559                                 runner.HostOutputDir = tmpdir
 560                         }
 561
 562                 case mnt.Kind == "json" || mnt.Kind == "text":
 563                         var filedata []byte
 564                         if mnt.Kind == "json" {
 565                                 filedata, err = json.Marshal(mnt.Content)
 566                                 if err != nil {
 567                                         return nil, fmt.Errorf("encoding json data: %v", err)
 568                                 }
 569                         } else {
 570                                 text, ok := mnt.Content.(string)
 571                                 if !ok {
 572                                         return nil, fmt.Errorf("content for mount %q must be a string", bind)
 573                                 }
 574                                 filedata = []byte(text)
 575                         }
 576
 577                         tmpdir, err := runner.MkTempDir(runner.parentTemp, mnt.Kind)
 578                         if err != nil {
 579                                 return nil, fmt.Errorf("creating temp dir: %v", err)
 580                         }
 581                         tmpfn := filepath.Join(tmpdir, "mountdata."+mnt.Kind)
 582                         err = ioutil.WriteFile(tmpfn, filedata, 0444)
 583                         if err != nil {
 584                                 return nil, fmt.Errorf("writing temp file: %v", err)
 585                         }
 586                         if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 587                                 copyFiles = append(copyFiles, copyFile{tmpfn, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
 588                         } else {
 589                                 bindmounts[bind] = bindmount{HostPath: tmpfn, ReadOnly: true}
 590                         }
 591
 592                 case mnt.Kind == "git_tree":
 593                         tmpdir, err := runner.MkTempDir(runner.parentTemp, "git_tree")
 594                         if err != nil {
 595                                 return nil, fmt.Errorf("creating temp dir: %v", err)
 596                         }
 597                         err = gitMount(mnt).extractTree(runner.ContainerArvClient, tmpdir, token)
 598                         if err != nil {
 599                                 return nil, err
 600                         }
 601                         bindmounts[bind] = bindmount{HostPath: tmpdir, ReadOnly: true}
 602                 }
 603         }
 604
 605         if runner.HostOutputDir == "" {
 606                 return nil, fmt.Errorf("output path does not correspond to a writable mount point")
 607         }
 608
 609         if needCertMount && runner.Container.RuntimeConstraints.API {
 610                 for _, certfile := range arvadosclient.CertFiles {
 611                         _, err := os.Stat(certfile)
 612                         if err == nil {
 613                                 bindmounts["/etc/arvados/ca-certificates.crt"] = bindmount{HostPath: certfile, ReadOnly: true}
 614                                 break
 615                         }
 616                 }
 617         }
 618
 619         if pdhOnly {
 620                 // If we are only mounting collections by pdh, make
 621                 // sure we don't subscribe to websocket events to
 622                 // avoid putting undesired load on the API server
 623                 arvMountCmd = append(arvMountCmd, "--mount-by-pdh", "by_id", "--disable-event-listening")
 624         } else {
 625                 arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_id")
 626         }
 627         // the by_uuid mount point is used by singularity when writing
 628         // out docker images converted to SIF
 629         arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_uuid")
 630         arvMountCmd = append(arvMountCmd, runner.ArvMountPoint)
 631
 632         runner.ArvMount, err = runner.RunArvMount(arvMountCmd, token)
 633         if err != nil {
 634                 return nil, fmt.Errorf("while trying to start arv-mount: %v", err)
 635         }
 636
 637         for _, p := range collectionPaths {
 638                 _, err = os.Stat(p)
 639                 if err != nil {
 640                         return nil, fmt.Errorf("while checking that input files exist: %v", err)
 641                 }
 642         }
 643
 644         for _, cp := range copyFiles {
 645                 st, err := os.Stat(cp.src)
 646                 if err != nil {
 647                         return nil, fmt.Errorf("while staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 648                 }
 649                 if st.IsDir() {
 650                         err = filepath.Walk(cp.src, func(walkpath string, walkinfo os.FileInfo, walkerr error) error {
 651                                 if walkerr != nil {
 652                                         return walkerr
 653                                 }
 654                                 target := path.Join(cp.bind, walkpath[len(cp.src):])
 655                                 if walkinfo.Mode().IsRegular() {
 656                                         copyerr := copyfile(walkpath, target)
 657                                         if copyerr != nil {
 658                                                 return copyerr
 659                                         }
 660                                         return os.Chmod(target, walkinfo.Mode()|0777)
 661                                 } else if walkinfo.Mode().IsDir() {
 662                                         mkerr := os.MkdirAll(target, 0777)
 663                                         if mkerr != nil {
 664                                                 return mkerr
 665                                         }
 666                                         return os.Chmod(target, walkinfo.Mode()|os.ModeSetgid|0777)
 667                                 } else {
 668                                         return fmt.Errorf("source %q is not a regular file or directory", cp.src)
 669                                 }
 670                         })
 671                 } else if st.Mode().IsRegular() {
 672                         err = copyfile(cp.src, cp.bind)
 673                         if err == nil {
 674                                 err = os.Chmod(cp.bind, st.Mode()|0777)
 675                         }
 676                 }
 677                 if err != nil {
 678                         return nil, fmt.Errorf("while staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 679                 }
 680         }
 681
 682         return bindmounts, nil
 683 }
 684
 685 func (runner *ContainerRunner) stopHoststat() error {
 686         if runner.hoststatReporter == nil {
 687                 return nil
 688         }
 689         runner.hoststatReporter.Stop()
 690         err := runner.hoststatLogger.Close()
 691         if err != nil {
 692                 return fmt.Errorf("error closing hoststat logs: %v", err)
 693         }
 694         return nil
 695 }
 696
 697 func (runner *ContainerRunner) startHoststat() error {
 698         w, err := runner.NewLogWriter("hoststat")
 699         if err != nil {
 700                 return err
 701         }
 702         runner.hoststatLogger = NewThrottledLogger(w)
 703         runner.hoststatReporter = &crunchstat.Reporter{
 704                 Logger:     log.New(runner.hoststatLogger, "", 0),
 705                 CgroupRoot: runner.cgroupRoot,
 706                 PollPeriod: runner.statInterval,
 707         }
 708         runner.hoststatReporter.Start()
 709         return nil
 710 }
 711
 712 func (runner *ContainerRunner) startCrunchstat() error {
 713         w, err := runner.NewLogWriter("crunchstat")
 714         if err != nil {
 715                 return err
 716         }
 717         runner.statLogger = NewThrottledLogger(w)
 718         runner.statReporter = &crunchstat.Reporter{
 719                 CID:          runner.executor.CgroupID(),
 720                 Logger:       log.New(runner.statLogger, "", 0),
 721                 CgroupParent: runner.expectCgroupParent,
 722                 CgroupRoot:   runner.cgroupRoot,
 723                 PollPeriod:   runner.statInterval,
 724                 TempDir:      runner.parentTemp,
 725         }
 726         runner.statReporter.Start()
 727         return nil
 728 }
 729
 730 type infoCommand struct {
 731         label string
 732         cmd   []string
 733 }
 734
 735 // LogHostInfo logs info about the current host, for debugging and
 736 // accounting purposes. Although it's logged as "node-info", this is
 737 // about the environment where crunch-run is actually running, which
 738 // might differ from what's described in the node record (see
 739 // LogNodeRecord).
 740 func (runner *ContainerRunner) LogHostInfo() (err error) {
 741         w, err := runner.NewLogWriter("node-info")
 742         if err != nil {
 743                 return
 744         }
 745
 746         commands := []infoCommand{
 747                 {
 748                         label: "Host Information",
 749                         cmd:   []string{"uname", "-a"},
 750                 },
 751                 {
 752                         label: "CPU Information",
 753                         cmd:   []string{"cat", "/proc/cpuinfo"},
 754                 },
 755                 {
 756                         label: "Memory Information",
 757                         cmd:   []string{"cat", "/proc/meminfo"},
 758                 },
 759                 {
 760                         label: "Disk Space",
 761                         cmd:   []string{"df", "-m", "/", os.TempDir()},
 762                 },
 763                 {
 764                         label: "Disk INodes",
 765                         cmd:   []string{"df", "-i", "/", os.TempDir()},
 766                 },
 767         }
 768
 769         // Run commands with informational output to be logged.
 770         for _, command := range commands {
 771                 fmt.Fprintln(w, command.label)
 772                 cmd := exec.Command(command.cmd[0], command.cmd[1:]...)
 773                 cmd.Stdout = w
 774                 cmd.Stderr = w
 775                 if err := cmd.Run(); err != nil {
 776                         err = fmt.Errorf("While running command %q: %v", command.cmd, err)
 777                         fmt.Fprintln(w, err)
 778                         return err
 779                 }
 780                 fmt.Fprintln(w, "")
 781         }
 782
 783         err = w.Close()
 784         if err != nil {
 785                 return fmt.Errorf("While closing node-info logs: %v", err)
 786         }
 787         return nil
 788 }
 789
 790 // LogContainerRecord gets and saves the raw JSON container record from the API server
 791 func (runner *ContainerRunner) LogContainerRecord() error {
 792         logged, err := runner.logAPIResponse("container", "containers", map[string]interface{}{"filters": [][]string{{"uuid", "=", runner.Container.UUID}}}, nil)
 793         if !logged && err == nil {
 794                 err = fmt.Errorf("error: no container record found for %s", runner.Container.UUID)
 795         }
 796         return err
 797 }
 798
 799 // LogNodeRecord logs the current host's InstanceType config entry (or
 800 // the arvados#node record, if running via crunch-dispatch-slurm).
 801 func (runner *ContainerRunner) LogNodeRecord() error {
 802         if it := os.Getenv("InstanceType"); it != "" {
 803                 // Dispatched via arvados-dispatch-cloud. Save
 804                 // InstanceType config fragment received from
 805                 // dispatcher on stdin.
 806                 w, err := runner.LogCollection.OpenFile("node.json", os.O_CREATE|os.O_WRONLY, 0666)
 807                 if err != nil {
 808                         return err
 809                 }
 810                 defer w.Close()
 811                 _, err = io.WriteString(w, it)
 812                 if err != nil {
 813                         return err
 814                 }
 815                 return w.Close()
 816         }
 817         // Dispatched via crunch-dispatch-slurm. Look up
 818         // apiserver's node record corresponding to
 819         // $SLURMD_NODENAME.
 820         hostname := os.Getenv("SLURMD_NODENAME")
 821         if hostname == "" {
 822                 hostname, _ = os.Hostname()
 823         }
 824         _, err := runner.logAPIResponse("node", "nodes", map[string]interface{}{"filters": [][]string{{"hostname", "=", hostname}}}, func(resp interface{}) {
 825                 // The "info" field has admin-only info when
 826                 // obtained with a privileged token, and
 827                 // should not be logged.
 828                 node, ok := resp.(map[string]interface{})
 829                 if ok {
 830                         delete(node, "info")
 831                 }
 832         })
 833         return err
 834 }
 835
 836 func (runner *ContainerRunner) logAPIResponse(label, path string, params map[string]interface{}, munge func(interface{})) (logged bool, err error) {
 837         writer, err := runner.LogCollection.OpenFile(label+".json", os.O_CREATE|os.O_WRONLY, 0666)
 838         if err != nil {
 839                 return false, err
 840         }
 841         w := &ArvLogWriter{
 842                 ArvClient:     runner.DispatcherArvClient,
 843                 UUID:          runner.Container.UUID,
 844                 loggingStream: label,
 845                 writeCloser:   writer,
 846         }
 847
 848         reader, err := runner.DispatcherArvClient.CallRaw("GET", path, "", "", arvadosclient.Dict(params))
 849         if err != nil {
 850                 return false, fmt.Errorf("error getting %s record: %v", label, err)
 851         }
 852         defer reader.Close()
 853
 854         dec := json.NewDecoder(reader)
 855         dec.UseNumber()
 856         var resp map[string]interface{}
 857         if err = dec.Decode(&resp); err != nil {
 858                 return false, fmt.Errorf("error decoding %s list response: %v", label, err)
 859         }
 860         items, ok := resp["items"].([]interface{})
 861         if !ok {
 862                 return false, fmt.Errorf("error decoding %s list response: no \"items\" key in API list response", label)
 863         } else if len(items) < 1 {
 864                 return false, nil
 865         }
 866         if munge != nil {
 867                 munge(items[0])
 868         }
 869         // Re-encode it using indentation to improve readability
 870         enc := json.NewEncoder(w)
 871         enc.SetIndent("", "    ")
 872         if err = enc.Encode(items[0]); err != nil {
 873                 return false, fmt.Errorf("error logging %s record: %v", label, err)
 874         }
 875         err = w.Close()
 876         if err != nil {
 877                 return false, fmt.Errorf("error closing %s.json in log collection: %v", label, err)
 878         }
 879         return true, nil
 880 }
 881
 882 func (runner *ContainerRunner) getStdoutFile(mntPath string) (*os.File, error) {
 883         stdoutPath := mntPath[len(runner.Container.OutputPath):]
 884         index := strings.LastIndex(stdoutPath, "/")
 885         if index > 0 {
 886                 subdirs := stdoutPath[:index]
 887                 if subdirs != "" {
 888                         st, err := os.Stat(runner.HostOutputDir)
 889                         if err != nil {
 890                                 return nil, fmt.Errorf("While Stat on temp dir: %v", err)
 891                         }
 892                         stdoutPath := filepath.Join(runner.HostOutputDir, subdirs)
 893                         err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777)
 894                         if err != nil {
 895                                 return nil, fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err)
 896                         }
 897                 }
 898         }
 899         stdoutFile, err := os.Create(filepath.Join(runner.HostOutputDir, stdoutPath))
 900         if err != nil {
 901                 return nil, fmt.Errorf("While creating file %q: %v", stdoutPath, err)
 902         }
 903
 904         return stdoutFile, nil
 905 }
 906
 907 // CreateContainer creates the docker container.
 908 func (runner *ContainerRunner) CreateContainer(imageID string, bindmounts map[string]bindmount) error {
 909         var stdin io.ReadCloser = ioutil.NopCloser(bytes.NewReader(nil))
 910         if mnt, ok := runner.Container.Mounts["stdin"]; ok {
 911                 switch mnt.Kind {
 912                 case "collection":
 913                         var collID string
 914                         if mnt.UUID != "" {
 915                                 collID = mnt.UUID
 916                         } else {
 917                                 collID = mnt.PortableDataHash
 918                         }
 919                         path := runner.ArvMountPoint + "/by_id/" + collID + "/" + mnt.Path
 920                         f, err := os.Open(path)
 921                         if err != nil {
 922                                 return err
 923                         }
 924                         stdin = f
 925                 case "json":
 926                         j, err := json.Marshal(mnt.Content)
 927                         if err != nil {
 928                                 return fmt.Errorf("error encoding stdin json data: %v", err)
 929                         }
 930                         stdin = ioutil.NopCloser(bytes.NewReader(j))
 931                 default:
 932                         return fmt.Errorf("stdin mount has unsupported kind %q", mnt.Kind)
 933                 }
 934         }
 935
 936         var stdout, stderr io.WriteCloser
 937         if mnt, ok := runner.Container.Mounts["stdout"]; ok {
 938                 f, err := runner.getStdoutFile(mnt.Path)
 939                 if err != nil {
 940                         return err
 941                 }
 942                 stdout = f
 943         } else if w, err := runner.NewLogWriter("stdout"); err != nil {
 944                 return err
 945         } else {
 946                 stdout = NewThrottledLogger(w)
 947         }
 948
 949         if mnt, ok := runner.Container.Mounts["stderr"]; ok {
 950                 f, err := runner.getStdoutFile(mnt.Path)
 951                 if err != nil {
 952                         return err
 953                 }
 954                 stderr = f
 955         } else if w, err := runner.NewLogWriter("stderr"); err != nil {
 956                 return err
 957         } else {
 958                 stderr = NewThrottledLogger(w)
 959         }
 960
 961         env := runner.Container.Environment
 962         enableNetwork := runner.enableNetwork == "always"
 963         if runner.Container.RuntimeConstraints.API {
 964                 enableNetwork = true
 965                 tok, err := runner.ContainerToken()
 966                 if err != nil {
 967                         return err
 968                 }
 969                 env = map[string]string{}
 970                 for k, v := range runner.Container.Environment {
 971                         env[k] = v
 972                 }
 973                 env["ARVADOS_API_TOKEN"] = tok
 974                 env["ARVADOS_API_HOST"] = os.Getenv("ARVADOS_API_HOST")
 975                 env["ARVADOS_API_HOST_INSECURE"] = os.Getenv("ARVADOS_API_HOST_INSECURE")
 976         }
 977         workdir := runner.Container.Cwd
 978         if workdir == "." {
 979                 // both "" and "." mean default
 980                 workdir = ""
 981         }
 982         ram := runner.Container.RuntimeConstraints.RAM
 983         if !runner.enableMemoryLimit {
 984                 ram = 0
 985         }
 986         runner.executorStdin = stdin
 987         runner.executorStdout = stdout
 988         runner.executorStderr = stderr
 989         return runner.executor.Create(containerSpec{
 990                 Image:         imageID,
 991                 VCPUs:         runner.Container.RuntimeConstraints.VCPUs,
 992                 RAM:           ram,
 993                 WorkingDir:    workdir,
 994                 Env:           env,
 995                 BindMounts:    bindmounts,
 996                 Command:       runner.Container.Command,
 997                 EnableNetwork: enableNetwork,
 998                 NetworkMode:   runner.networkMode,
 999                 CgroupParent:  runner.setCgroupParent,
1000                 Stdin:         stdin,
1001                 Stdout:        stdout,
1002                 Stderr:        stderr,
1003         })
1004 }
1005
1006 // StartContainer starts the docker container created by CreateContainer.
1007 func (runner *ContainerRunner) StartContainer() error {
1008         runner.CrunchLog.Printf("Starting container")
1009         runner.cStateLock.Lock()
1010         defer runner.cStateLock.Unlock()
1011         if runner.cCancelled {
1012                 return ErrCancelled
1013         }
1014         err := runner.executor.Start()
1015         if err != nil {
1016                 var advice string
1017                 if m, e := regexp.MatchString("(?ms).*(exec|System error).*(no such file or directory|file not found).*", err.Error()); m && e == nil {
1018                         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])
1019                 }
1020                 return fmt.Errorf("could not start container: %v%s", err, advice)
1021         }
1022         return nil
1023 }
1024
1025 // WaitFinish waits for the container to terminate, capture the exit code, and
1026 // close the stdout/stderr logging.
1027 func (runner *ContainerRunner) WaitFinish() error {
1028         runner.CrunchLog.Print("Waiting for container to finish")
1029         var timeout <-chan time.Time
1030         if s := runner.Container.SchedulingParameters.MaxRunTime; s > 0 {
1031                 timeout = time.After(time.Duration(s) * time.Second)
1032         }
1033         ctx, cancel := context.WithCancel(context.Background())
1034         defer cancel()
1035         go func() {
1036                 select {
1037                 case <-timeout:
1038                         runner.CrunchLog.Printf("maximum run time exceeded. Stopping container.")
1039                         runner.stop(nil)
1040                 case <-runner.ArvMountExit:
1041                         runner.CrunchLog.Printf("arv-mount exited while container is still running. Stopping container.")
1042                         runner.stop(nil)
1043                 case <-ctx.Done():
1044                 }
1045         }()
1046         exitcode, err := runner.executor.Wait(ctx)
1047         if err != nil {
1048                 runner.checkBrokenNode(err)
1049                 return err
1050         }
1051         runner.ExitCode = &exitcode
1052
1053         var returnErr error
1054         if err = runner.executorStdin.Close(); err != nil {
1055                 err = fmt.Errorf("error closing container stdin: %s", err)
1056                 runner.CrunchLog.Printf("%s", err)
1057                 returnErr = err
1058         }
1059         if err = runner.executorStdout.Close(); err != nil {
1060                 err = fmt.Errorf("error closing container stdout: %s", err)
1061                 runner.CrunchLog.Printf("%s", err)
1062                 if returnErr == nil {
1063                         returnErr = err
1064                 }
1065         }
1066         if err = runner.executorStderr.Close(); err != nil {
1067                 err = fmt.Errorf("error closing container stderr: %s", err)
1068                 runner.CrunchLog.Printf("%s", err)
1069                 if returnErr == nil {
1070                         returnErr = err
1071                 }
1072         }
1073
1074         if runner.statReporter != nil {
1075                 runner.statReporter.Stop()
1076                 err = runner.statLogger.Close()
1077                 if err != nil {
1078                         runner.CrunchLog.Printf("error closing crunchstat logs: %v", err)
1079                 }
1080         }
1081         return returnErr
1082 }
1083
1084 func (runner *ContainerRunner) updateLogs() {
1085         ticker := time.NewTicker(crunchLogUpdatePeriod / 360)
1086         defer ticker.Stop()
1087
1088         sigusr1 := make(chan os.Signal, 1)
1089         signal.Notify(sigusr1, syscall.SIGUSR1)
1090         defer signal.Stop(sigusr1)
1091
1092         saveAtTime := time.Now().Add(crunchLogUpdatePeriod)
1093         saveAtSize := crunchLogUpdateSize
1094         var savedSize int64
1095         for {
1096                 select {
1097                 case <-ticker.C:
1098                 case <-sigusr1:
1099                         saveAtTime = time.Now()
1100                 }
1101                 runner.logMtx.Lock()
1102                 done := runner.LogsPDH != nil
1103                 runner.logMtx.Unlock()
1104                 if done {
1105                         return
1106                 }
1107                 size := runner.LogCollection.Size()
1108                 if size == savedSize || (time.Now().Before(saveAtTime) && size < saveAtSize) {
1109                         continue
1110                 }
1111                 saveAtTime = time.Now().Add(crunchLogUpdatePeriod)
1112                 saveAtSize = runner.LogCollection.Size() + crunchLogUpdateSize
1113                 saved, err := runner.saveLogCollection(false)
1114                 if err != nil {
1115                         runner.CrunchLog.Printf("error updating log collection: %s", err)
1116                         continue
1117                 }
1118
1119                 var updated arvados.Container
1120                 err = runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1121                         "container": arvadosclient.Dict{"log": saved.PortableDataHash},
1122                 }, &updated)
1123                 if err != nil {
1124                         runner.CrunchLog.Printf("error updating container log to %s: %s", saved.PortableDataHash, err)
1125                         continue
1126                 }
1127
1128                 savedSize = size
1129         }
1130 }
1131
1132 func (runner *ContainerRunner) reportArvMountWarning(pattern, text string) {
1133         var updated arvados.Container
1134         err := runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1135                 "container": arvadosclient.Dict{
1136                         "runtime_status": arvadosclient.Dict{
1137                                 "warning":       "arv-mount: " + pattern,
1138                                 "warningDetail": text,
1139                         },
1140                 },
1141         }, &updated)
1142         if err != nil {
1143                 runner.CrunchLog.Printf("error updating container runtime_status: %s", err)
1144         }
1145 }
1146
1147 // CaptureOutput saves data from the container's output directory if
1148 // needed, and updates the container output accordingly.
1149 func (runner *ContainerRunner) CaptureOutput(bindmounts map[string]bindmount) error {
1150         if runner.Container.RuntimeConstraints.API {
1151                 // Output may have been set directly by the container, so
1152                 // refresh the container record to check.
1153                 err := runner.DispatcherArvClient.Get("containers", runner.Container.UUID,
1154                         nil, &runner.Container)
1155                 if err != nil {
1156                         return err
1157                 }
1158                 if runner.Container.Output != "" {
1159                         // Container output is already set.
1160                         runner.OutputPDH = &runner.Container.Output
1161                         return nil
1162                 }
1163         }
1164
1165         txt, err := (&copier{
1166                 client:        runner.containerClient,
1167                 arvClient:     runner.ContainerArvClient,
1168                 keepClient:    runner.ContainerKeepClient,
1169                 hostOutputDir: runner.HostOutputDir,
1170                 ctrOutputDir:  runner.Container.OutputPath,
1171                 bindmounts:    bindmounts,
1172                 mounts:        runner.Container.Mounts,
1173                 secretMounts:  runner.SecretMounts,
1174                 logger:        runner.CrunchLog,
1175         }).Copy()
1176         if err != nil {
1177                 return err
1178         }
1179         if n := len(regexp.MustCompile(` [0-9a-f]+\+\S*\+R`).FindAllStringIndex(txt, -1)); n > 0 {
1180                 runner.CrunchLog.Printf("Copying %d data blocks from remote input collections...", n)
1181                 fs, err := (&arvados.Collection{ManifestText: txt}).FileSystem(runner.containerClient, runner.ContainerKeepClient)
1182                 if err != nil {
1183                         return err
1184                 }
1185                 txt, err = fs.MarshalManifest(".")
1186                 if err != nil {
1187                         return err
1188                 }
1189         }
1190         var resp arvados.Collection
1191         err = runner.ContainerArvClient.Create("collections", arvadosclient.Dict{
1192                 "ensure_unique_name": true,
1193                 "collection": arvadosclient.Dict{
1194                         "is_trashed":    true,
1195                         "name":          "output for " + runner.Container.UUID,
1196                         "manifest_text": txt,
1197                 },
1198         }, &resp)
1199         if err != nil {
1200                 return fmt.Errorf("error creating output collection: %v", err)
1201         }
1202         runner.OutputPDH = &resp.PortableDataHash
1203         return nil
1204 }
1205
1206 func (runner *ContainerRunner) CleanupDirs() {
1207         if runner.ArvMount != nil {
1208                 var delay int64 = 8
1209                 umount := exec.Command("arv-mount", fmt.Sprintf("--unmount-timeout=%d", delay), "--unmount", runner.ArvMountPoint)
1210                 umount.Stdout = runner.CrunchLog
1211                 umount.Stderr = runner.CrunchLog
1212                 runner.CrunchLog.Printf("Running %v", umount.Args)
1213                 umnterr := umount.Start()
1214
1215                 if umnterr != nil {
1216                         runner.CrunchLog.Printf("Error unmounting: %v", umnterr)
1217                         runner.ArvMount.Process.Kill()
1218                 } else {
1219                         // If arv-mount --unmount gets stuck for any reason, we
1220                         // don't want to wait for it forever.  Do Wait() in a goroutine
1221                         // so it doesn't block crunch-run.
1222                         umountExit := make(chan error)
1223                         go func() {
1224                                 mnterr := umount.Wait()
1225                                 if mnterr != nil {
1226                                         runner.CrunchLog.Printf("Error unmounting: %v", mnterr)
1227                                 }
1228                                 umountExit <- mnterr
1229                         }()
1230
1231                         for again := true; again; {
1232                                 again = false
1233                                 select {
1234                                 case <-umountExit:
1235                                         umount = nil
1236                                         again = true
1237                                 case <-runner.ArvMountExit:
1238                                         break
1239                                 case <-time.After(time.Duration((delay + 1) * int64(time.Second))):
1240                                         runner.CrunchLog.Printf("Timed out waiting for unmount")
1241                                         if umount != nil {
1242                                                 umount.Process.Kill()
1243                                         }
1244                                         runner.ArvMount.Process.Kill()
1245                                 }
1246                         }
1247                 }
1248                 runner.ArvMount = nil
1249         }
1250
1251         if runner.ArvMountPoint != "" {
1252                 if rmerr := os.Remove(runner.ArvMountPoint); rmerr != nil {
1253                         runner.CrunchLog.Printf("While cleaning up arv-mount directory %s: %v", runner.ArvMountPoint, rmerr)
1254                 }
1255                 runner.ArvMountPoint = ""
1256         }
1257
1258         if rmerr := os.RemoveAll(runner.parentTemp); rmerr != nil {
1259                 runner.CrunchLog.Printf("While cleaning up temporary directory %s: %v", runner.parentTemp, rmerr)
1260         }
1261 }
1262
1263 // CommitLogs posts the collection containing the final container logs.
1264 func (runner *ContainerRunner) CommitLogs() error {
1265         func() {
1266                 // Hold cStateLock to prevent races on CrunchLog (e.g., stop()).
1267                 runner.cStateLock.Lock()
1268                 defer runner.cStateLock.Unlock()
1269
1270                 runner.CrunchLog.Print(runner.finalState)
1271
1272                 if runner.arvMountLog != nil {
1273                         runner.arvMountLog.Close()
1274                 }
1275                 runner.CrunchLog.Close()
1276
1277                 // Closing CrunchLog above allows them to be committed to Keep at this
1278                 // point, but re-open crunch log with ArvClient in case there are any
1279                 // other further errors (such as failing to write the log to Keep!)
1280                 // while shutting down
1281                 runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{
1282                         ArvClient:     runner.DispatcherArvClient,
1283                         UUID:          runner.Container.UUID,
1284                         loggingStream: "crunch-run",
1285                         writeCloser:   nil,
1286                 })
1287                 runner.CrunchLog.Immediate = log.New(os.Stderr, runner.Container.UUID+" ", 0)
1288         }()
1289
1290         if runner.keepstoreLogger != nil {
1291                 // Flush any buffered logs from our local keepstore
1292                 // process.  Discard anything logged after this point
1293                 // -- it won't end up in the log collection, so
1294                 // there's no point writing it to the collectionfs.
1295                 runner.keepstoreLogbuf.SetWriter(io.Discard)
1296                 runner.keepstoreLogger.Close()
1297                 runner.keepstoreLogger = nil
1298         }
1299
1300         if runner.LogsPDH != nil {
1301                 // If we have already assigned something to LogsPDH,
1302                 // we must be closing the re-opened log, which won't
1303                 // end up getting attached to the container record and
1304                 // therefore doesn't need to be saved as a collection
1305                 // -- it exists only to send logs to other channels.
1306                 return nil
1307         }
1308
1309         saved, err := runner.saveLogCollection(true)
1310         if err != nil {
1311                 return fmt.Errorf("error saving log collection: %s", err)
1312         }
1313         runner.logMtx.Lock()
1314         defer runner.logMtx.Unlock()
1315         runner.LogsPDH = &saved.PortableDataHash
1316         return nil
1317 }
1318
1319 func (runner *ContainerRunner) saveLogCollection(final bool) (response arvados.Collection, err error) {
1320         runner.logMtx.Lock()
1321         defer runner.logMtx.Unlock()
1322         if runner.LogsPDH != nil {
1323                 // Already finalized.
1324                 return
1325         }
1326         updates := arvadosclient.Dict{
1327                 "name": "logs for " + runner.Container.UUID,
1328         }
1329         mt, err1 := runner.LogCollection.MarshalManifest(".")
1330         if err1 == nil {
1331                 // Only send updated manifest text if there was no
1332                 // error.
1333                 updates["manifest_text"] = mt
1334         }
1335
1336         // Even if flushing the manifest had an error, we still want
1337         // to update the log record, if possible, to push the trash_at
1338         // and delete_at times into the future.  Details on bug
1339         // #17293.
1340         if final {
1341                 updates["is_trashed"] = true
1342         } else {
1343                 exp := time.Now().Add(crunchLogUpdatePeriod * 24)
1344                 updates["trash_at"] = exp
1345                 updates["delete_at"] = exp
1346         }
1347         reqBody := arvadosclient.Dict{"collection": updates}
1348         var err2 error
1349         if runner.logUUID == "" {
1350                 reqBody["ensure_unique_name"] = true
1351                 err2 = runner.DispatcherArvClient.Create("collections", reqBody, &response)
1352         } else {
1353                 err2 = runner.DispatcherArvClient.Update("collections", runner.logUUID, reqBody, &response)
1354         }
1355         if err2 == nil {
1356                 runner.logUUID = response.UUID
1357         }
1358
1359         if err1 != nil || err2 != nil {
1360                 err = fmt.Errorf("error recording logs: %q, %q", err1, err2)
1361         }
1362         return
1363 }
1364
1365 // UpdateContainerRunning updates the container state to "Running"
1366 func (runner *ContainerRunner) UpdateContainerRunning() error {
1367         runner.cStateLock.Lock()
1368         defer runner.cStateLock.Unlock()
1369         if runner.cCancelled {
1370                 return ErrCancelled
1371         }
1372         return runner.DispatcherArvClient.Update("containers", runner.Container.UUID,
1373                 arvadosclient.Dict{"container": arvadosclient.Dict{"state": "Running", "gateway_address": runner.gateway.Address}}, nil)
1374 }
1375
1376 // ContainerToken returns the api_token the container (and any
1377 // arv-mount processes) are allowed to use.
1378 func (runner *ContainerRunner) ContainerToken() (string, error) {
1379         if runner.token != "" {
1380                 return runner.token, nil
1381         }
1382
1383         var auth arvados.APIClientAuthorization
1384         err := runner.DispatcherArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth)
1385         if err != nil {
1386                 return "", err
1387         }
1388         runner.token = fmt.Sprintf("v2/%s/%s/%s", auth.UUID, auth.APIToken, runner.Container.UUID)
1389         return runner.token, nil
1390 }
1391
1392 // UpdateContainerFinal updates the container record state on API
1393 // server to "Complete" or "Cancelled"
1394 func (runner *ContainerRunner) UpdateContainerFinal() error {
1395         update := arvadosclient.Dict{}
1396         update["state"] = runner.finalState
1397         if runner.LogsPDH != nil {
1398                 update["log"] = *runner.LogsPDH
1399         }
1400         if runner.finalState == "Complete" {
1401                 if runner.ExitCode != nil {
1402                         update["exit_code"] = *runner.ExitCode
1403                 }
1404                 if runner.OutputPDH != nil {
1405                         update["output"] = *runner.OutputPDH
1406                 }
1407         }
1408         return runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{"container": update}, nil)
1409 }
1410
1411 // IsCancelled returns the value of Cancelled, with goroutine safety.
1412 func (runner *ContainerRunner) IsCancelled() bool {
1413         runner.cStateLock.Lock()
1414         defer runner.cStateLock.Unlock()
1415         return runner.cCancelled
1416 }
1417
1418 // NewArvLogWriter creates an ArvLogWriter
1419 func (runner *ContainerRunner) NewArvLogWriter(name string) (io.WriteCloser, error) {
1420         writer, err := runner.LogCollection.OpenFile(name+".txt", os.O_CREATE|os.O_WRONLY, 0666)
1421         if err != nil {
1422                 return nil, err
1423         }
1424         return &ArvLogWriter{
1425                 ArvClient:     runner.DispatcherArvClient,
1426                 UUID:          runner.Container.UUID,
1427                 loggingStream: name,
1428                 writeCloser:   writer,
1429         }, nil
1430 }
1431
1432 // Run the full container lifecycle.
1433 func (runner *ContainerRunner) Run() (err error) {
1434         runner.CrunchLog.Printf("crunch-run %s started", cmd.Version.String())
1435         runner.CrunchLog.Printf("Executing container '%s' using %s runtime", runner.Container.UUID, runner.executor.Runtime())
1436
1437         hostname, hosterr := os.Hostname()
1438         if hosterr != nil {
1439                 runner.CrunchLog.Printf("Error getting hostname '%v'", hosterr)
1440         } else {
1441                 runner.CrunchLog.Printf("Executing on host '%s'", hostname)
1442         }
1443
1444         runner.finalState = "Queued"
1445
1446         defer func() {
1447                 runner.CleanupDirs()
1448
1449                 runner.CrunchLog.Printf("crunch-run finished")
1450                 runner.CrunchLog.Close()
1451         }()
1452
1453         err = runner.fetchContainerRecord()
1454         if err != nil {
1455                 return
1456         }
1457         if runner.Container.State != "Locked" {
1458                 return fmt.Errorf("dispatch error detected: container %q has state %q", runner.Container.UUID, runner.Container.State)
1459         }
1460
1461         var bindmounts map[string]bindmount
1462         defer func() {
1463                 // checkErr prints e (unless it's nil) and sets err to
1464                 // e (unless err is already non-nil). Thus, if err
1465                 // hasn't already been assigned when Run() returns,
1466                 // this cleanup func will cause Run() to return the
1467                 // first non-nil error that is passed to checkErr().
1468                 checkErr := func(errorIn string, e error) {
1469                         if e == nil {
1470                                 return
1471                         }
1472                         runner.CrunchLog.Printf("error in %s: %v", errorIn, e)
1473                         if err == nil {
1474                                 err = e
1475                         }
1476                         if runner.finalState == "Complete" {
1477                                 // There was an error in the finalization.
1478                                 runner.finalState = "Cancelled"
1479                         }
1480                 }
1481
1482                 // Log the error encountered in Run(), if any
1483                 checkErr("Run", err)
1484
1485                 if runner.finalState == "Queued" {
1486                         runner.UpdateContainerFinal()
1487                         return
1488                 }
1489
1490                 if runner.IsCancelled() {
1491                         runner.finalState = "Cancelled"
1492                         // but don't return yet -- we still want to
1493                         // capture partial output and write logs
1494                 }
1495
1496                 if bindmounts != nil {
1497                         checkErr("CaptureOutput", runner.CaptureOutput(bindmounts))
1498                 }
1499                 checkErr("stopHoststat", runner.stopHoststat())
1500                 checkErr("CommitLogs", runner.CommitLogs())
1501                 runner.CleanupDirs()
1502                 checkErr("UpdateContainerFinal", runner.UpdateContainerFinal())
1503         }()
1504
1505         runner.setupSignals()
1506         err = runner.startHoststat()
1507         if err != nil {
1508                 return
1509         }
1510
1511         // set up FUSE mount and binds
1512         bindmounts, err = runner.SetupMounts()
1513         if err != nil {
1514                 runner.finalState = "Cancelled"
1515                 err = fmt.Errorf("While setting up mounts: %v", err)
1516                 return
1517         }
1518
1519         // check for and/or load image
1520         imageID, err := runner.LoadImage()
1521         if err != nil {
1522                 if !runner.checkBrokenNode(err) {
1523                         // Failed to load image but not due to a "broken node"
1524                         // condition, probably user error.
1525                         runner.finalState = "Cancelled"
1526                 }
1527                 err = fmt.Errorf("While loading container image: %v", err)
1528                 return
1529         }
1530
1531         err = runner.CreateContainer(imageID, bindmounts)
1532         if err != nil {
1533                 return
1534         }
1535         err = runner.LogHostInfo()
1536         if err != nil {
1537                 return
1538         }
1539         err = runner.LogNodeRecord()
1540         if err != nil {
1541                 return
1542         }
1543         err = runner.LogContainerRecord()
1544         if err != nil {
1545                 return
1546         }
1547
1548         if runner.IsCancelled() {
1549                 return
1550         }
1551
1552         err = runner.UpdateContainerRunning()
1553         if err != nil {
1554                 return
1555         }
1556         runner.finalState = "Cancelled"
1557
1558         err = runner.startCrunchstat()
1559         if err != nil {
1560                 return
1561         }
1562
1563         err = runner.StartContainer()
1564         if err != nil {
1565                 runner.checkBrokenNode(err)
1566                 return
1567         }
1568
1569         err = runner.WaitFinish()
1570         if err == nil && !runner.IsCancelled() {
1571                 runner.finalState = "Complete"
1572         }
1573         return
1574 }
1575
1576 // Fetch the current container record (uuid = runner.Container.UUID)
1577 // into runner.Container.
1578 func (runner *ContainerRunner) fetchContainerRecord() error {
1579         reader, err := runner.DispatcherArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
1580         if err != nil {
1581                 return fmt.Errorf("error fetching container record: %v", err)
1582         }
1583         defer reader.Close()
1584
1585         dec := json.NewDecoder(reader)
1586         dec.UseNumber()
1587         err = dec.Decode(&runner.Container)
1588         if err != nil {
1589                 return fmt.Errorf("error decoding container record: %v", err)
1590         }
1591
1592         var sm struct {
1593                 SecretMounts map[string]arvados.Mount `json:"secret_mounts"`
1594         }
1595
1596         containerToken, err := runner.ContainerToken()
1597         if err != nil {
1598                 return fmt.Errorf("error getting container token: %v", err)
1599         }
1600
1601         runner.ContainerArvClient, runner.ContainerKeepClient,
1602                 runner.containerClient, err = runner.MkArvClient(containerToken)
1603         if err != nil {
1604                 return fmt.Errorf("error creating container API client: %v", err)
1605         }
1606
1607         runner.ContainerKeepClient.SetStorageClasses(runner.Container.OutputStorageClasses)
1608         runner.DispatcherKeepClient.SetStorageClasses(runner.Container.OutputStorageClasses)
1609
1610         err = runner.ContainerArvClient.Call("GET", "containers", runner.Container.UUID, "secret_mounts", nil, &sm)
1611         if err != nil {
1612                 if apierr, ok := err.(arvadosclient.APIServerError); !ok || apierr.HttpStatusCode != 404 {
1613                         return fmt.Errorf("error fetching secret_mounts: %v", err)
1614                 }
1615                 // ok && apierr.HttpStatusCode == 404, which means
1616                 // secret_mounts isn't supported by this API server.
1617         }
1618         runner.SecretMounts = sm.SecretMounts
1619
1620         return nil
1621 }
1622
1623 // NewContainerRunner creates a new container runner.
1624 func NewContainerRunner(dispatcherClient *arvados.Client,
1625         dispatcherArvClient IArvadosClient,
1626         dispatcherKeepClient IKeepClient,
1627         containerUUID string) (*ContainerRunner, error) {
1628
1629         cr := &ContainerRunner{
1630                 dispatcherClient:     dispatcherClient,
1631                 DispatcherArvClient:  dispatcherArvClient,
1632                 DispatcherKeepClient: dispatcherKeepClient,
1633         }
1634         cr.NewLogWriter = cr.NewArvLogWriter
1635         cr.RunArvMount = cr.ArvMountCmd
1636         cr.MkTempDir = ioutil.TempDir
1637         cr.MkArvClient = func(token string) (IArvadosClient, IKeepClient, *arvados.Client, error) {
1638                 cl, err := arvadosclient.MakeArvadosClient()
1639                 if err != nil {
1640                         return nil, nil, nil, err
1641                 }
1642                 cl.ApiToken = token
1643                 kc, err := keepclient.MakeKeepClient(cl)
1644                 if err != nil {
1645                         return nil, nil, nil, err
1646                 }
1647                 c2 := arvados.NewClientFromEnv()
1648                 c2.AuthToken = token
1649                 return cl, kc, c2, nil
1650         }
1651         var err error
1652         cr.LogCollection, err = (&arvados.Collection{}).FileSystem(cr.dispatcherClient, cr.DispatcherKeepClient)
1653         if err != nil {
1654                 return nil, err
1655         }
1656         cr.Container.UUID = containerUUID
1657         w, err := cr.NewLogWriter("crunch-run")
1658         if err != nil {
1659                 return nil, err
1660         }
1661         cr.CrunchLog = NewThrottledLogger(w)
1662         cr.CrunchLog.Immediate = log.New(os.Stderr, containerUUID+" ", 0)
1663
1664         loadLogThrottleParams(dispatcherArvClient)
1665         go cr.updateLogs()
1666
1667         return cr, nil
1668 }
1669
1670 func (command) RunCommand(prog string, args []string, stdin io.Reader, stdout, stderr io.Writer) int {
1671         log := log.New(stderr, "", 0)
1672         flags := flag.NewFlagSet(prog, flag.ContinueOnError)
1673         statInterval := flags.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting")
1674         cgroupRoot := flags.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree")
1675         cgroupParent := flags.String("cgroup-parent", "docker", "name of container's parent cgroup (ignored if -cgroup-parent-subsystem is used)")
1676         cgroupParentSubsystem := flags.String("cgroup-parent-subsystem", "", "use current cgroup for given subsystem as parent cgroup for container")
1677         caCertsPath := flags.String("ca-certs", "", "Path to TLS root certificates")
1678         detach := flags.Bool("detach", false, "Detach from parent process and run in the background")
1679         stdinConfig := flags.Bool("stdin-config", false, "Load config and environment variables from JSON message on stdin")
1680         sleep := flags.Duration("sleep", 0, "Delay before starting (testing use only)")
1681         kill := flags.Int("kill", -1, "Send signal to an existing crunch-run process for given UUID")
1682         list := flags.Bool("list", false, "List UUIDs of existing crunch-run processes")
1683         enableMemoryLimit := flags.Bool("enable-memory-limit", true, "tell container runtime to limit container's memory usage")
1684         enableNetwork := flags.String("container-enable-networking", "default", "enable networking \"always\" (for all containers) or \"default\" (for containers that request it)")
1685         networkMode := flags.String("container-network-mode", "default", `Docker network mode for container (use any argument valid for docker --net)`)
1686         memprofile := flags.String("memprofile", "", "write memory profile to `file` after running container")
1687         runtimeEngine := flags.String("runtime-engine", "docker", "container runtime: docker or singularity")
1688         flags.Duration("check-containerd", 0, "Ignored. Exists for compatibility with older versions.")
1689
1690         ignoreDetachFlag := false
1691         if len(args) > 0 && args[0] == "-no-detach" {
1692                 // This process was invoked by a parent process, which
1693                 // has passed along its own arguments, including
1694                 // -detach, after the leading -no-detach flag.  Strip
1695                 // the leading -no-detach flag (it's not recognized by
1696                 // flags.Parse()) and ignore the -detach flag that
1697                 // comes later.
1698                 args = args[1:]
1699                 ignoreDetachFlag = true
1700         }
1701
1702         if err := flags.Parse(args); err == flag.ErrHelp {
1703                 return 0
1704         } else if err != nil {
1705                 log.Print(err)
1706                 return 1
1707         }
1708
1709         containerUUID := flags.Arg(0)
1710
1711         switch {
1712         case *detach && !ignoreDetachFlag:
1713                 return Detach(containerUUID, prog, args, os.Stdin, os.Stdout, os.Stderr)
1714         case *kill >= 0:
1715                 return KillProcess(containerUUID, syscall.Signal(*kill), os.Stdout, os.Stderr)
1716         case *list:
1717                 return ListProcesses(os.Stdout, os.Stderr)
1718         }
1719
1720         if len(containerUUID) != 27 {
1721                 log.Printf("usage: %s [options] UUID", prog)
1722                 return 1
1723         }
1724
1725         var conf ConfigData
1726         if *stdinConfig {
1727                 err := json.NewDecoder(stdin).Decode(&conf)
1728                 if err != nil {
1729                         log.Printf("decode stdin: %s", err)
1730                         return 1
1731                 }
1732                 for k, v := range conf.Env {
1733                         err = os.Setenv(k, v)
1734                         if err != nil {
1735                                 log.Printf("setenv(%q): %s", k, err)
1736                                 return 1
1737                         }
1738                 }
1739                 if conf.Cluster != nil {
1740                         // ClusterID is missing from the JSON
1741                         // representation, but we need it to generate
1742                         // a valid config file for keepstore, so we
1743                         // fill it using the container UUID prefix.
1744                         conf.Cluster.ClusterID = containerUUID[:5]
1745                 }
1746         }
1747
1748         log.Printf("crunch-run %s started", cmd.Version.String())
1749         time.Sleep(*sleep)
1750
1751         if *caCertsPath != "" {
1752                 arvadosclient.CertFiles = []string{*caCertsPath}
1753         }
1754
1755         var keepstoreLogbuf bufThenWrite
1756         keepstore, err := startLocalKeepstore(conf, io.MultiWriter(&keepstoreLogbuf, stderr))
1757         if err != nil {
1758                 log.Print(err)
1759                 return 1
1760         }
1761         if keepstore != nil {
1762                 defer keepstore.Process.Kill()
1763         }
1764
1765         api, err := arvadosclient.MakeArvadosClient()
1766         if err != nil {
1767                 log.Printf("%s: %v", containerUUID, err)
1768                 return 1
1769         }
1770         api.Retries = 8
1771
1772         kc, err := keepclient.MakeKeepClient(api)
1773         if err != nil {
1774                 log.Printf("%s: %v", containerUUID, err)
1775                 return 1
1776         }
1777         kc.BlockCache = &keepclient.BlockCache{MaxBlocks: 2}
1778         kc.Retries = 4
1779
1780         cr, err := NewContainerRunner(arvados.NewClientFromEnv(), api, kc, containerUUID)
1781         if err != nil {
1782                 log.Print(err)
1783                 return 1
1784         }
1785
1786         if keepstore == nil {
1787                 // Log explanation (if any) for why we're not running
1788                 // a local keepstore.
1789                 var buf bytes.Buffer
1790                 keepstoreLogbuf.SetWriter(&buf)
1791                 if buf.Len() > 0 {
1792                         cr.CrunchLog.Printf("%s", strings.TrimSpace(buf.String()))
1793                 }
1794         } else if logWhat := conf.Cluster.Containers.LocalKeepLogsToContainerLog; logWhat == "none" {
1795                 cr.CrunchLog.Printf("using local keepstore process (pid %d) at %s", keepstore.Process.Pid, os.Getenv("ARVADOS_KEEP_SERVICES"))
1796                 keepstoreLogbuf.SetWriter(io.Discard)
1797         } else {
1798                 cr.CrunchLog.Printf("using local keepstore process (pid %d) at %s, writing logs to keepstore.txt in log collection", keepstore.Process.Pid, os.Getenv("ARVADOS_KEEP_SERVICES"))
1799                 logwriter, err := cr.NewLogWriter("keepstore")
1800                 if err != nil {
1801                         log.Print(err)
1802                         return 1
1803                 }
1804                 cr.keepstoreLogger = NewThrottledLogger(logwriter)
1805
1806                 var writer io.WriteCloser = cr.keepstoreLogger
1807                 if logWhat == "errors" {
1808                         writer = &filterKeepstoreErrorsOnly{WriteCloser: writer}
1809                 } else if logWhat != "all" {
1810                         // should have been caught earlier by
1811                         // dispatcher's config loader
1812                         log.Printf("invalid value for Containers.LocalKeepLogsToContainerLog: %q", logWhat)
1813                         return 1
1814                 }
1815                 err = keepstoreLogbuf.SetWriter(writer)
1816                 if err != nil {
1817                         log.Print(err)
1818                         return 1
1819                 }
1820                 cr.keepstoreLogbuf = &keepstoreLogbuf
1821         }
1822
1823         switch *runtimeEngine {
1824         case "docker":
1825                 cr.executor, err = newDockerExecutor(containerUUID, cr.CrunchLog.Printf, cr.containerWatchdogInterval)
1826         case "singularity":
1827                 cr.executor, err = newSingularityExecutor(cr.CrunchLog.Printf)
1828         default:
1829                 cr.CrunchLog.Printf("%s: unsupported RuntimeEngine %q", containerUUID, *runtimeEngine)
1830                 cr.CrunchLog.Close()
1831                 return 1
1832         }
1833         if err != nil {
1834                 cr.CrunchLog.Printf("%s: %v", containerUUID, err)
1835                 cr.checkBrokenNode(err)
1836                 cr.CrunchLog.Close()
1837                 return 1
1838         }
1839         defer cr.executor.Close()
1840
1841         gwAuthSecret := os.Getenv("GatewayAuthSecret")
1842         os.Unsetenv("GatewayAuthSecret")
1843         if gwAuthSecret == "" {
1844                 // not safe to run a gateway service without an auth
1845                 // secret
1846                 cr.CrunchLog.Printf("Not starting a gateway server (GatewayAuthSecret was not provided by dispatcher)")
1847         } else if gwListen := os.Getenv("GatewayAddress"); gwListen == "" {
1848                 // dispatcher did not tell us which external IP
1849                 // address to advertise --> no gateway service
1850                 cr.CrunchLog.Printf("Not starting a gateway server (GatewayAddress was not provided by dispatcher)")
1851         } else if de, ok := cr.executor.(*dockerExecutor); ok {
1852                 cr.gateway = Gateway{
1853                         Address:            gwListen,
1854                         AuthSecret:         gwAuthSecret,
1855                         ContainerUUID:      containerUUID,
1856                         DockerContainerID:  &de.containerID,
1857                         Log:                cr.CrunchLog,
1858                         ContainerIPAddress: dockerContainerIPAddress(&de.containerID),
1859                 }
1860                 err = cr.gateway.Start()
1861                 if err != nil {
1862                         log.Printf("error starting gateway server: %s", err)
1863                         return 1
1864                 }
1865         }
1866
1867         parentTemp, tmperr := cr.MkTempDir("", "crunch-run."+containerUUID+".")
1868         if tmperr != nil {
1869                 log.Printf("%s: %v", containerUUID, tmperr)
1870                 return 1
1871         }
1872
1873         cr.parentTemp = parentTemp
1874         cr.statInterval = *statInterval
1875         cr.cgroupRoot = *cgroupRoot
1876         cr.expectCgroupParent = *cgroupParent
1877         cr.enableMemoryLimit = *enableMemoryLimit
1878         cr.enableNetwork = *enableNetwork
1879         cr.networkMode = *networkMode
1880         if *cgroupParentSubsystem != "" {
1881                 p := findCgroup(*cgroupParentSubsystem)
1882                 cr.setCgroupParent = p
1883                 cr.expectCgroupParent = p
1884         }
1885
1886         runerr := cr.Run()
1887
1888         if *memprofile != "" {
1889                 f, err := os.Create(*memprofile)
1890                 if err != nil {
1891                         log.Printf("could not create memory profile: %s", err)
1892                 }
1893                 runtime.GC() // get up-to-date statistics
1894                 if err := pprof.WriteHeapProfile(f); err != nil {
1895                         log.Printf("could not write memory profile: %s", err)
1896                 }
1897                 closeerr := f.Close()
1898                 if closeerr != nil {
1899                         log.Printf("closing memprofile file: %s", err)
1900                 }
1901         }
1902
1903         if runerr != nil {
1904                 log.Printf("%s: %v", containerUUID, runerr)
1905                 return 1
1906         }
1907         return 0
1908 }
1909
1910 func startLocalKeepstore(configData ConfigData, logbuf io.Writer) (*exec.Cmd, error) {
1911         if configData.Cluster == nil || configData.KeepBuffers < 1 {
1912                 return nil, nil
1913         }
1914         for uuid, vol := range configData.Cluster.Volumes {
1915                 if len(vol.AccessViaHosts) > 0 {
1916                         fmt.Fprintf(logbuf, "not starting a local keepstore process because a volume (%s) uses AccessViaHosts\n", uuid)
1917                         return nil, nil
1918                 }
1919                 if !vol.ReadOnly && vol.Replication < configData.Cluster.Collections.DefaultReplication {
1920                         fmt.Fprintf(logbuf, "not starting a local keepstore process because a writable volume (%s) has replication less than Collections.DefaultReplication (%d < %d)\n", uuid, vol.Replication, configData.Cluster.Collections.DefaultReplication)
1921                         return nil, nil
1922                 }
1923         }
1924
1925         // Rather than have an alternate way to tell keepstore how
1926         // many buffers to use when starting it this way, we just
1927         // modify the cluster configuration that we feed it on stdin.
1928         configData.Cluster.API.MaxKeepBlobBuffers = configData.KeepBuffers
1929
1930         ln, err := net.Listen("tcp", "localhost:0")
1931         if err != nil {
1932                 return nil, err
1933         }
1934         _, port, err := net.SplitHostPort(ln.Addr().String())
1935         if err != nil {
1936                 ln.Close()
1937                 return nil, err
1938         }
1939         ln.Close()
1940         url := "http://localhost:" + port
1941
1942         fmt.Fprintf(logbuf, "starting keepstore on %s\n", url)
1943
1944         var confJSON bytes.Buffer
1945         err = json.NewEncoder(&confJSON).Encode(arvados.Config{
1946                 Clusters: map[string]arvados.Cluster{
1947                         configData.Cluster.ClusterID: *configData.Cluster,
1948                 },
1949         })
1950         if err != nil {
1951                 return nil, err
1952         }
1953         cmd := exec.Command("/proc/self/exe", "keepstore", "-config=-")
1954         if target, err := os.Readlink(cmd.Path); err == nil && strings.HasSuffix(target, ".test") {
1955                 // If we're a 'go test' process, running
1956                 // /proc/self/exe would start the test suite in a
1957                 // child process, which is not what we want.
1958                 cmd.Path, _ = exec.LookPath("go")
1959                 cmd.Args = append([]string{"go", "run", "../../cmd/arvados-server"}, cmd.Args[1:]...)
1960                 cmd.Env = os.Environ()
1961         }
1962         cmd.Stdin = &confJSON
1963         cmd.Stdout = logbuf
1964         cmd.Stderr = logbuf
1965         cmd.Env = append(cmd.Env,
1966                 "GOGC=10",
1967                 "ARVADOS_SERVICE_INTERNAL_URL="+url)
1968         err = cmd.Start()
1969         if err != nil {
1970                 return nil, fmt.Errorf("error starting keepstore process: %w", err)
1971         }
1972         cmdExited := false
1973         go func() {
1974                 cmd.Wait()
1975                 cmdExited = true
1976         }()
1977         ctx, cancel := context.WithDeadline(context.Background(), time.Now().Add(time.Second*10))
1978         defer cancel()
1979         poll := time.NewTicker(time.Second / 10)
1980         defer poll.Stop()
1981         client := http.Client{}
1982         for range poll.C {
1983                 testReq, err := http.NewRequestWithContext(ctx, "GET", url+"/_health/ping", nil)
1984                 testReq.Header.Set("Authorization", "Bearer "+configData.Cluster.ManagementToken)
1985                 if err != nil {
1986                         return nil, err
1987                 }
1988                 resp, err := client.Do(testReq)
1989                 if err == nil {
1990                         resp.Body.Close()
1991                         if resp.StatusCode == http.StatusOK {
1992                                 break
1993                         }
1994                 }
1995                 if cmdExited {
1996                         return nil, fmt.Errorf("keepstore child process exited")
1997                 }
1998                 if ctx.Err() != nil {
1999                         return nil, fmt.Errorf("timed out waiting for new keepstore process to report healthy")
2000                 }
2001         }
2002         os.Setenv("ARVADOS_KEEP_SERVICES", url)
2003         return cmd, nil
2004 }