lib/crunchrun/crunchrun.go

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