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