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