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         "golang.org/x/sys/unix"
  40 )
  41
  42 type command struct{}
  43
  44 var Command = command{}
  45
  46 // ConfigData contains environment variables and (when needed) cluster
  47 // configuration, passed from dispatchcloud to crunch-run on stdin.
  48 type ConfigData struct {
  49         Env         map[string]string
  50         KeepBuffers int
  51         Cluster     *arvados.Cluster
  52 }
  53
  54 // IArvadosClient is the minimal Arvados API methods used by crunch-run.
  55 type IArvadosClient interface {
  56         Create(resourceType string, parameters arvadosclient.Dict, output interface{}) error
  57         Get(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  58         Update(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  59         Call(method, resourceType, uuid, action string, parameters arvadosclient.Dict, output interface{}) error
  60         CallRaw(method string, resourceType string, uuid string, action string, parameters arvadosclient.Dict) (reader io.ReadCloser, err error)
  61         Discovery(key string) (interface{}, error)
  62 }
  63
  64 // ErrCancelled is the error returned when the container is cancelled.
  65 var ErrCancelled = errors.New("Cancelled")
  66
  67 // IKeepClient is the minimal Keep API methods used by crunch-run.
  68 type IKeepClient interface {
  69         BlockWrite(context.Context, arvados.BlockWriteOptions) (arvados.BlockWriteResponse, error)
  70         ReadAt(locator string, p []byte, off int) (int, error)
  71         ManifestFileReader(m manifest.Manifest, filename string) (arvados.File, error)
  72         LocalLocator(locator string) (string, error)
  73         ClearBlockCache()
  74         SetStorageClasses(sc []string)
  75 }
  76
  77 // NewLogWriter is a factory function to create a new log writer.
  78 type NewLogWriter func(name string) (io.WriteCloser, error)
  79
  80 type RunArvMount func(cmdline []string, tok string) (*exec.Cmd, error)
  81
  82 type MkTempDir func(string, string) (string, error)
  83
  84 type PsProcess interface {
  85         CmdlineSlice() ([]string, error)
  86 }
  87
  88 // ContainerRunner is the main stateful struct used for a single execution of a
  89 // container.
  90 type ContainerRunner struct {
  91         executor       containerExecutor
  92         executorStdin  io.Closer
  93         executorStdout io.Closer
  94         executorStderr io.Closer
  95
  96         // Dispatcher client is initialized with the Dispatcher token.
  97         // This is a privileged token used to manage container status
  98         // and logs.
  99         //
 100         // We have both dispatcherClient and DispatcherArvClient
 101         // because there are two different incompatible Arvados Go
 102         // SDKs and we have to use both (hopefully this gets fixed in
 103         // #14467)
 104         dispatcherClient     *arvados.Client
 105         DispatcherArvClient  IArvadosClient
 106         DispatcherKeepClient IKeepClient
 107
 108         // Container client is initialized with the Container token
 109         // This token controls the permissions of the container, and
 110         // must be used for operations such as reading collections.
 111         //
 112         // Same comment as above applies to
 113         // containerClient/ContainerArvClient.
 114         containerClient     *arvados.Client
 115         ContainerArvClient  IArvadosClient
 116         ContainerKeepClient IKeepClient
 117
 118         Container     arvados.Container
 119         token         string
 120         ExitCode      *int
 121         NewLogWriter  NewLogWriter
 122         CrunchLog     *ThrottledLogger
 123         logUUID       string
 124         logMtx        sync.Mutex
 125         LogCollection arvados.CollectionFileSystem
 126         LogsPDH       *string
 127         RunArvMount   RunArvMount
 128         MkTempDir     MkTempDir
 129         ArvMount      *exec.Cmd
 130         ArvMountPoint string
 131         HostOutputDir string
 132         Volumes       map[string]struct{}
 133         OutputPDH     *string
 134         SigChan       chan os.Signal
 135         ArvMountExit  chan error
 136         SecretMounts  map[string]arvados.Mount
 137         MkArvClient   func(token string) (IArvadosClient, IKeepClient, *arvados.Client, error)
 138         finalState    string
 139         parentTemp    string
 140
 141         keepstoreLogger  io.WriteCloser
 142         keepstoreLogbuf  *bufThenWrite
 143         statLogger       io.WriteCloser
 144         statReporter     *crunchstat.Reporter
 145         hoststatLogger   io.WriteCloser
 146         hoststatReporter *crunchstat.Reporter
 147         statInterval     time.Duration
 148         cgroupRoot       string
 149         // What we expect the container's cgroup parent to be.
 150         expectCgroupParent string
 151         // What we tell docker to use as the container's cgroup
 152         // parent. Note: Ideally we would use the same field for both
 153         // expectCgroupParent and setCgroupParent, and just make it
 154         // default to "docker". However, when using docker < 1.10 with
 155         // systemd, specifying a non-empty cgroup parent (even the
 156         // default value "docker") hits a docker bug
 157         // (https://github.com/docker/docker/issues/17126). Using two
 158         // separate fields makes it possible to use the "expect cgroup
 159         // parent to be X" feature even on sites where the "specify
 160         // cgroup parent" feature breaks.
 161         setCgroupParent string
 162
 163         cStateLock sync.Mutex
 164         cCancelled bool // StopContainer() invoked
 165
 166         enableMemoryLimit bool
 167         enableNetwork     string // one of "default" or "always"
 168         networkMode       string // "none", "host", or "" -- passed through to executor
 169         arvMountLog       *ThrottledLogger
 170
 171         containerWatchdogInterval time.Duration
 172
 173         gateway Gateway
 174 }
 175
 176 // setupSignals sets up signal handling to gracefully terminate the
 177 // underlying container and update state when receiving a TERM, INT or
 178 // QUIT signal.
 179 func (runner *ContainerRunner) setupSignals() {
 180         runner.SigChan = make(chan os.Signal, 1)
 181         signal.Notify(runner.SigChan, syscall.SIGTERM)
 182         signal.Notify(runner.SigChan, syscall.SIGINT)
 183         signal.Notify(runner.SigChan, syscall.SIGQUIT)
 184
 185         go func(sig chan os.Signal) {
 186                 for s := range sig {
 187                         runner.stop(s)
 188                 }
 189         }(runner.SigChan)
 190 }
 191
 192 // stop the underlying container.
 193 func (runner *ContainerRunner) stop(sig os.Signal) {
 194         runner.cStateLock.Lock()
 195         defer runner.cStateLock.Unlock()
 196         if sig != nil {
 197                 runner.CrunchLog.Printf("caught signal: %v", sig)
 198         }
 199         runner.cCancelled = true
 200         runner.CrunchLog.Printf("stopping container")
 201         err := runner.executor.Stop()
 202         if err != nil {
 203                 runner.CrunchLog.Printf("error stopping container: %s", err)
 204         }
 205 }
 206
 207 var errorBlacklist = []string{
 208         "(?ms).*[Cc]annot connect to the Docker daemon.*",
 209         "(?ms).*oci runtime error.*starting container process.*container init.*mounting.*to rootfs.*no such file or directory.*",
 210         "(?ms).*grpc: the connection is unavailable.*",
 211 }
 212 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)")
 213
 214 func (runner *ContainerRunner) runBrokenNodeHook() {
 215         if *brokenNodeHook == "" {
 216                 path := filepath.Join(lockdir, brokenfile)
 217                 runner.CrunchLog.Printf("Writing %s to mark node as broken", path)
 218                 f, err := os.OpenFile(path, os.O_CREATE|os.O_RDWR, 0700)
 219                 if err != nil {
 220                         runner.CrunchLog.Printf("Error writing %s: %s", path, err)
 221                         return
 222                 }
 223                 f.Close()
 224         } else {
 225                 runner.CrunchLog.Printf("Running broken node hook %q", *brokenNodeHook)
 226                 // run killme script
 227                 c := exec.Command(*brokenNodeHook)
 228                 c.Stdout = runner.CrunchLog
 229                 c.Stderr = runner.CrunchLog
 230                 err := c.Run()
 231                 if err != nil {
 232                         runner.CrunchLog.Printf("Error running broken node hook: %v", err)
 233                 }
 234         }
 235 }
 236
 237 func (runner *ContainerRunner) checkBrokenNode(goterr error) bool {
 238         for _, d := range errorBlacklist {
 239                 if m, e := regexp.MatchString(d, goterr.Error()); m && e == nil {
 240                         runner.CrunchLog.Printf("Error suggests node is unable to run containers: %v", goterr)
 241                         runner.runBrokenNodeHook()
 242                         return true
 243                 }
 244         }
 245         return false
 246 }
 247
 248 // LoadImage determines the docker image id from the container record and
 249 // checks if it is available in the local Docker image store.  If not, it loads
 250 // the image from Keep.
 251 func (runner *ContainerRunner) LoadImage() (string, error) {
 252         runner.CrunchLog.Printf("Fetching Docker image from collection '%s'", runner.Container.ContainerImage)
 253
 254         d, err := os.Open(runner.ArvMountPoint + "/by_id/" + runner.Container.ContainerImage)
 255         if err != nil {
 256                 return "", err
 257         }
 258         defer d.Close()
 259         allfiles, err := d.Readdirnames(-1)
 260         if err != nil {
 261                 return "", err
 262         }
 263         var tarfiles []string
 264         for _, fnm := range allfiles {
 265                 if strings.HasSuffix(fnm, ".tar") {
 266                         tarfiles = append(tarfiles, fnm)
 267                 }
 268         }
 269         if len(tarfiles) == 0 {
 270                 return "", fmt.Errorf("image collection does not include a .tar image file")
 271         }
 272         if len(tarfiles) > 1 {
 273                 return "", fmt.Errorf("cannot choose from multiple tar files in image collection: %v", tarfiles)
 274         }
 275         imageID := tarfiles[0][:len(tarfiles[0])-4]
 276         imageTarballPath := runner.ArvMountPoint + "/by_id/" + runner.Container.ContainerImage + "/" + imageID + ".tar"
 277         runner.CrunchLog.Printf("Using Docker image id %q", imageID)
 278
 279         runner.CrunchLog.Print("Loading Docker image from keep")
 280         err = runner.executor.LoadImage(imageID, imageTarballPath, runner.Container, runner.ArvMountPoint,
 281                 runner.containerClient)
 282         if err != nil {
 283                 return "", err
 284         }
 285
 286         return imageID, nil
 287 }
 288
 289 func (runner *ContainerRunner) ArvMountCmd(cmdline []string, token string) (c *exec.Cmd, err error) {
 290         c = exec.Command(cmdline[0], cmdline[1:]...)
 291
 292         // Copy our environment, but override ARVADOS_API_TOKEN with
 293         // the container auth token.
 294         c.Env = nil
 295         for _, s := range os.Environ() {
 296                 if !strings.HasPrefix(s, "ARVADOS_API_TOKEN=") {
 297                         c.Env = append(c.Env, s)
 298                 }
 299         }
 300         c.Env = append(c.Env, "ARVADOS_API_TOKEN="+token)
 301
 302         w, err := runner.NewLogWriter("arv-mount")
 303         if err != nil {
 304                 return nil, err
 305         }
 306         runner.arvMountLog = NewThrottledLogger(w)
 307         scanner := logScanner{
 308                 Patterns: []string{
 309                         "Keep write error",
 310                         "Block not found error",
 311                         "Unhandled exception during FUSE operation",
 312                 },
 313                 ReportFunc: runner.reportArvMountWarning,
 314         }
 315         c.Stdout = runner.arvMountLog
 316         c.Stderr = io.MultiWriter(runner.arvMountLog, os.Stderr, &scanner)
 317
 318         runner.CrunchLog.Printf("Running %v", c.Args)
 319
 320         err = c.Start()
 321         if err != nil {
 322                 return nil, err
 323         }
 324
 325         statReadme := make(chan bool)
 326         runner.ArvMountExit = make(chan error)
 327
 328         keepStatting := true
 329         go func() {
 330                 for keepStatting {
 331                         time.Sleep(100 * time.Millisecond)
 332                         _, err = os.Stat(fmt.Sprintf("%s/by_id/README", runner.ArvMountPoint))
 333                         if err == nil {
 334                                 keepStatting = false
 335                                 statReadme <- true
 336                         }
 337                 }
 338                 close(statReadme)
 339         }()
 340
 341         go func() {
 342                 mnterr := c.Wait()
 343                 if mnterr != nil {
 344                         runner.CrunchLog.Printf("Arv-mount exit error: %v", mnterr)
 345                 }
 346                 runner.ArvMountExit <- mnterr
 347                 close(runner.ArvMountExit)
 348         }()
 349
 350         select {
 351         case <-statReadme:
 352                 break
 353         case err := <-runner.ArvMountExit:
 354                 runner.ArvMount = nil
 355                 keepStatting = false
 356                 return nil, err
 357         }
 358
 359         return c, nil
 360 }
 361
 362 func (runner *ContainerRunner) SetupArvMountPoint(prefix string) (err error) {
 363         if runner.ArvMountPoint == "" {
 364                 runner.ArvMountPoint, err = runner.MkTempDir(runner.parentTemp, prefix)
 365         }
 366         return
 367 }
 368
 369 func copyfile(src string, dst string) (err error) {
 370         srcfile, err := os.Open(src)
 371         if err != nil {
 372                 return
 373         }
 374
 375         os.MkdirAll(path.Dir(dst), 0777)
 376
 377         dstfile, err := os.Create(dst)
 378         if err != nil {
 379                 return
 380         }
 381         _, err = io.Copy(dstfile, srcfile)
 382         if err != nil {
 383                 return
 384         }
 385
 386         err = srcfile.Close()
 387         err2 := dstfile.Close()
 388
 389         if err != nil {
 390                 return
 391         }
 392
 393         if err2 != nil {
 394                 return err2
 395         }
 396
 397         return nil
 398 }
 399
 400 func (runner *ContainerRunner) SetupMounts() (map[string]bindmount, error) {
 401         bindmounts := map[string]bindmount{}
 402         err := runner.SetupArvMountPoint("keep")
 403         if err != nil {
 404                 return nil, fmt.Errorf("While creating keep mount temp dir: %v", err)
 405         }
 406
 407         token, err := runner.ContainerToken()
 408         if err != nil {
 409                 return nil, fmt.Errorf("could not get container token: %s", err)
 410         }
 411         runner.CrunchLog.Printf("container token %q", token)
 412
 413         pdhOnly := true
 414         tmpcount := 0
 415         arvMountCmd := []string{
 416                 "arv-mount",
 417                 "--foreground",
 418                 "--read-write",
 419                 "--storage-classes", strings.Join(runner.Container.OutputStorageClasses, ","),
 420                 fmt.Sprintf("--crunchstat-interval=%v", runner.statInterval.Seconds())}
 421
 422         if runner.executor.Runtime() == "docker" {
 423                 arvMountCmd = append(arvMountCmd, "--allow-other")
 424         }
 425
 426         if runner.Container.RuntimeConstraints.KeepCacheRAM > 0 {
 427                 arvMountCmd = append(arvMountCmd, "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheRAM))
 428         }
 429
 430         collectionPaths := []string{}
 431         needCertMount := true
 432         type copyFile struct {
 433                 src  string
 434                 bind string
 435         }
 436         var copyFiles []copyFile
 437
 438         var binds []string
 439         for bind := range runner.Container.Mounts {
 440                 binds = append(binds, bind)
 441         }
 442         for bind := range runner.SecretMounts {
 443                 if _, ok := runner.Container.Mounts[bind]; ok {
 444                         return nil, fmt.Errorf("secret mount %q conflicts with regular mount", bind)
 445                 }
 446                 if runner.SecretMounts[bind].Kind != "json" &&
 447                         runner.SecretMounts[bind].Kind != "text" {
 448                         return nil, fmt.Errorf("secret mount %q type is %q but only 'json' and 'text' are permitted",
 449                                 bind, runner.SecretMounts[bind].Kind)
 450                 }
 451                 binds = append(binds, bind)
 452         }
 453         sort.Strings(binds)
 454
 455         for _, bind := range binds {
 456                 mnt, notSecret := runner.Container.Mounts[bind]
 457                 if !notSecret {
 458                         mnt = runner.SecretMounts[bind]
 459                 }
 460                 if bind == "stdout" || bind == "stderr" {
 461                         // Is it a "file" mount kind?
 462                         if mnt.Kind != "file" {
 463                                 return nil, fmt.Errorf("unsupported mount kind '%s' for %s: only 'file' is supported", mnt.Kind, bind)
 464                         }
 465
 466                         // Does path start with OutputPath?
 467                         prefix := runner.Container.OutputPath
 468                         if !strings.HasSuffix(prefix, "/") {
 469                                 prefix += "/"
 470                         }
 471                         if !strings.HasPrefix(mnt.Path, prefix) {
 472                                 return nil, fmt.Errorf("%s path does not start with OutputPath: %s, %s", strings.Title(bind), mnt.Path, prefix)
 473                         }
 474                 }
 475
 476                 if bind == "stdin" {
 477                         // Is it a "collection" mount kind?
 478                         if mnt.Kind != "collection" && mnt.Kind != "json" {
 479                                 return nil, fmt.Errorf("unsupported mount kind '%s' for stdin: only 'collection' and 'json' are supported", mnt.Kind)
 480                         }
 481                 }
 482
 483                 if bind == "/etc/arvados/ca-certificates.crt" {
 484                         needCertMount = false
 485                 }
 486
 487                 if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && bind != runner.Container.OutputPath+"/" {
 488                         if mnt.Kind != "collection" && mnt.Kind != "text" && mnt.Kind != "json" {
 489                                 return nil, fmt.Errorf("only mount points of kind 'collection', 'text' or 'json' are supported underneath the output_path for %q, was %q", bind, mnt.Kind)
 490                         }
 491                 }
 492
 493                 switch {
 494                 case mnt.Kind == "collection" && bind != "stdin":
 495                         var src string
 496                         if mnt.UUID != "" && mnt.PortableDataHash != "" {
 497                                 return nil, fmt.Errorf("cannot specify both 'uuid' and 'portable_data_hash' for a collection mount")
 498                         }
 499                         if mnt.UUID != "" {
 500                                 if mnt.Writable {
 501                                         return nil, fmt.Errorf("writing to existing collections currently not permitted")
 502                                 }
 503                                 pdhOnly = false
 504                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.UUID)
 505                         } else if mnt.PortableDataHash != "" {
 506                                 if mnt.Writable && !strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 507                                         return nil, fmt.Errorf("can never write to a collection specified by portable data hash")
 508                                 }
 509                                 idx := strings.Index(mnt.PortableDataHash, "/")
 510                                 if idx > 0 {
 511                                         mnt.Path = path.Clean(mnt.PortableDataHash[idx:])
 512                                         mnt.PortableDataHash = mnt.PortableDataHash[0:idx]
 513                                         runner.Container.Mounts[bind] = mnt
 514                                 }
 515                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.PortableDataHash)
 516                                 if mnt.Path != "" && mnt.Path != "." {
 517                                         if strings.HasPrefix(mnt.Path, "./") {
 518                                                 mnt.Path = mnt.Path[2:]
 519                                         } else if strings.HasPrefix(mnt.Path, "/") {
 520                                                 mnt.Path = mnt.Path[1:]
 521                                         }
 522                                         src += "/" + mnt.Path
 523                                 }
 524                         } else {
 525                                 src = fmt.Sprintf("%s/tmp%d", runner.ArvMountPoint, tmpcount)
 526                                 arvMountCmd = append(arvMountCmd, "--mount-tmp", fmt.Sprintf("tmp%d", tmpcount))
 527                                 tmpcount++
 528                         }
 529                         if mnt.Writable {
 530                                 if bind == runner.Container.OutputPath {
 531                                         runner.HostOutputDir = src
 532                                         bindmounts[bind] = bindmount{HostPath: src}
 533                                 } else if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 534                                         copyFiles = append(copyFiles, copyFile{src, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
 535                                 } else {
 536                                         bindmounts[bind] = bindmount{HostPath: src}
 537                                 }
 538                         } else {
 539                                 bindmounts[bind] = bindmount{HostPath: src, ReadOnly: true}
 540                         }
 541                         collectionPaths = append(collectionPaths, src)
 542
 543                 case mnt.Kind == "tmp":
 544                         var tmpdir string
 545                         tmpdir, err = runner.MkTempDir(runner.parentTemp, "tmp")
 546                         if err != nil {
 547                                 return nil, fmt.Errorf("while creating mount temp dir: %v", err)
 548                         }
 549                         st, staterr := os.Stat(tmpdir)
 550                         if staterr != nil {
 551                                 return nil, fmt.Errorf("while Stat on temp dir: %v", staterr)
 552                         }
 553                         err = os.Chmod(tmpdir, st.Mode()|os.ModeSetgid|0777)
 554                         if staterr != nil {
 555                                 return nil, fmt.Errorf("while Chmod temp dir: %v", err)
 556                         }
 557                         bindmounts[bind] = bindmount{HostPath: tmpdir}
 558                         if bind == runner.Container.OutputPath {
 559                                 runner.HostOutputDir = tmpdir
 560                         }
 561
 562                 case mnt.Kind == "json" || mnt.Kind == "text":
 563                         var filedata []byte
 564                         if mnt.Kind == "json" {
 565                                 filedata, err = json.Marshal(mnt.Content)
 566                                 if err != nil {
 567                                         return nil, fmt.Errorf("encoding json data: %v", err)
 568                                 }
 569                         } else {
 570                                 text, ok := mnt.Content.(string)
 571                                 if !ok {
 572                                         return nil, fmt.Errorf("content for mount %q must be a string", bind)
 573                                 }
 574                                 filedata = []byte(text)
 575                         }
 576
 577                         tmpdir, err := runner.MkTempDir(runner.parentTemp, mnt.Kind)
 578                         if err != nil {
 579                                 return nil, fmt.Errorf("creating temp dir: %v", err)
 580                         }
 581                         tmpfn := filepath.Join(tmpdir, "mountdata."+mnt.Kind)
 582                         err = ioutil.WriteFile(tmpfn, filedata, 0444)
 583                         if err != nil {
 584                                 return nil, fmt.Errorf("writing temp file: %v", err)
 585                         }
 586                         if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && (notSecret || runner.Container.Mounts[runner.Container.OutputPath].Kind != "collection") {
 587                                 // In most cases, if the container
 588                                 // specifies a literal file inside the
 589                                 // output path, we copy it into the
 590                                 // output directory (either a mounted
 591                                 // collection or a staging area on the
 592                                 // host fs). If it's a secret, it will
 593                                 // be skipped when copying output from
 594                                 // staging to Keep later.
 595                                 copyFiles = append(copyFiles, copyFile{tmpfn, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
 596                         } else {
 597                                 // If a secret is outside OutputPath,
 598                                 // we bind mount the secret file
 599                                 // directly just like other mounts. We
 600                                 // also use this strategy when a
 601                                 // secret is inside OutputPath but
 602                                 // OutputPath is a live collection, to
 603                                 // avoid writing the secret to
 604                                 // Keep. Attempting to remove a
 605                                 // bind-mounted secret file from
 606                                 // inside the container will return a
 607                                 // "Device or resource busy" error
 608                                 // that might not be handled well by
 609                                 // the container, which is why we
 610                                 // don't use this strategy when
 611                                 // OutputPath is a staging directory.
 612                                 bindmounts[bind] = bindmount{HostPath: tmpfn, ReadOnly: true}
 613                         }
 614
 615                 case mnt.Kind == "git_tree":
 616                         tmpdir, err := runner.MkTempDir(runner.parentTemp, "git_tree")
 617                         if err != nil {
 618                                 return nil, fmt.Errorf("creating temp dir: %v", err)
 619                         }
 620                         err = gitMount(mnt).extractTree(runner.ContainerArvClient, tmpdir, token)
 621                         if err != nil {
 622                                 return nil, err
 623                         }
 624                         bindmounts[bind] = bindmount{HostPath: tmpdir, ReadOnly: true}
 625                 }
 626         }
 627
 628         if runner.HostOutputDir == "" {
 629                 return nil, fmt.Errorf("output path does not correspond to a writable mount point")
 630         }
 631
 632         if needCertMount && runner.Container.RuntimeConstraints.API {
 633                 for _, certfile := range arvadosclient.CertFiles {
 634                         _, err := os.Stat(certfile)
 635                         if err == nil {
 636                                 bindmounts["/etc/arvados/ca-certificates.crt"] = bindmount{HostPath: certfile, ReadOnly: true}
 637                                 break
 638                         }
 639                 }
 640         }
 641
 642         if pdhOnly {
 643                 // If we are only mounting collections by pdh, make
 644                 // sure we don't subscribe to websocket events to
 645                 // avoid putting undesired load on the API server
 646                 arvMountCmd = append(arvMountCmd, "--mount-by-pdh", "by_id", "--disable-event-listening")
 647         } else {
 648                 arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_id")
 649         }
 650         // the by_uuid mount point is used by singularity when writing
 651         // out docker images converted to SIF
 652         arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_uuid")
 653         arvMountCmd = append(arvMountCmd, runner.ArvMountPoint)
 654
 655         runner.ArvMount, err = runner.RunArvMount(arvMountCmd, token)
 656         if err != nil {
 657                 return nil, fmt.Errorf("while trying to start arv-mount: %v", err)
 658         }
 659
 660         for _, p := range collectionPaths {
 661                 _, err = os.Stat(p)
 662                 if err != nil {
 663                         return nil, fmt.Errorf("while checking that input files exist: %v", err)
 664                 }
 665         }
 666
 667         for _, cp := range copyFiles {
 668                 st, err := os.Stat(cp.src)
 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                 if st.IsDir() {
 673                         err = filepath.Walk(cp.src, func(walkpath string, walkinfo os.FileInfo, walkerr error) error {
 674                                 if walkerr != nil {
 675                                         return walkerr
 676                                 }
 677                                 target := path.Join(cp.bind, walkpath[len(cp.src):])
 678                                 if walkinfo.Mode().IsRegular() {
 679                                         copyerr := copyfile(walkpath, target)
 680                                         if copyerr != nil {
 681                                                 return copyerr
 682                                         }
 683                                         return os.Chmod(target, walkinfo.Mode()|0777)
 684                                 } else if walkinfo.Mode().IsDir() {
 685                                         mkerr := os.MkdirAll(target, 0777)
 686                                         if mkerr != nil {
 687                                                 return mkerr
 688                                         }
 689                                         return os.Chmod(target, walkinfo.Mode()|os.ModeSetgid|0777)
 690                                 } else {
 691                                         return fmt.Errorf("source %q is not a regular file or directory", cp.src)
 692                                 }
 693                         })
 694                 } else if st.Mode().IsRegular() {
 695                         err = copyfile(cp.src, cp.bind)
 696                         if err == nil {
 697                                 err = os.Chmod(cp.bind, st.Mode()|0777)
 698                         }
 699                 }
 700                 if err != nil {
 701                         return nil, fmt.Errorf("while staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 702                 }
 703         }
 704
 705         return bindmounts, nil
 706 }
 707
 708 func (runner *ContainerRunner) stopHoststat() error {
 709         if runner.hoststatReporter == nil {
 710                 return nil
 711         }
 712         runner.hoststatReporter.Stop()
 713         err := runner.hoststatLogger.Close()
 714         if err != nil {
 715                 return fmt.Errorf("error closing hoststat logs: %v", err)
 716         }
 717         return nil
 718 }
 719
 720 func (runner *ContainerRunner) startHoststat() error {
 721         w, err := runner.NewLogWriter("hoststat")
 722         if err != nil {
 723                 return err
 724         }
 725         runner.hoststatLogger = NewThrottledLogger(w)
 726         runner.hoststatReporter = &crunchstat.Reporter{
 727                 Logger:     log.New(runner.hoststatLogger, "", 0),
 728                 CgroupRoot: runner.cgroupRoot,
 729                 PollPeriod: runner.statInterval,
 730         }
 731         runner.hoststatReporter.Start()
 732         return nil
 733 }
 734
 735 func (runner *ContainerRunner) startCrunchstat() error {
 736         w, err := runner.NewLogWriter("crunchstat")
 737         if err != nil {
 738                 return err
 739         }
 740         runner.statLogger = NewThrottledLogger(w)
 741         runner.statReporter = &crunchstat.Reporter{
 742                 CID:          runner.executor.CgroupID(),
 743                 Logger:       log.New(runner.statLogger, "", 0),
 744                 CgroupParent: runner.expectCgroupParent,
 745                 CgroupRoot:   runner.cgroupRoot,
 746                 PollPeriod:   runner.statInterval,
 747                 TempDir:      runner.parentTemp,
 748         }
 749         runner.statReporter.Start()
 750         return nil
 751 }
 752
 753 type infoCommand struct {
 754         label string
 755         cmd   []string
 756 }
 757
 758 // LogHostInfo logs info about the current host, for debugging and
 759 // accounting purposes. Although it's logged as "node-info", this is
 760 // about the environment where crunch-run is actually running, which
 761 // might differ from what's described in the node record (see
 762 // LogNodeRecord).
 763 func (runner *ContainerRunner) LogHostInfo() (err error) {
 764         w, err := runner.NewLogWriter("node-info")
 765         if err != nil {
 766                 return
 767         }
 768
 769         commands := []infoCommand{
 770                 {
 771                         label: "Host Information",
 772                         cmd:   []string{"uname", "-a"},
 773                 },
 774                 {
 775                         label: "CPU Information",
 776                         cmd:   []string{"cat", "/proc/cpuinfo"},
 777                 },
 778                 {
 779                         label: "Memory Information",
 780                         cmd:   []string{"cat", "/proc/meminfo"},
 781                 },
 782                 {
 783                         label: "Disk Space",
 784                         cmd:   []string{"df", "-m", "/", os.TempDir()},
 785                 },
 786                 {
 787                         label: "Disk INodes",
 788                         cmd:   []string{"df", "-i", "/", os.TempDir()},
 789                 },
 790         }
 791
 792         // Run commands with informational output to be logged.
 793         for _, command := range commands {
 794                 fmt.Fprintln(w, command.label)
 795                 cmd := exec.Command(command.cmd[0], command.cmd[1:]...)
 796                 cmd.Stdout = w
 797                 cmd.Stderr = w
 798                 if err := cmd.Run(); err != nil {
 799                         err = fmt.Errorf("While running command %q: %v", command.cmd, err)
 800                         fmt.Fprintln(w, err)
 801                         return err
 802                 }
 803                 fmt.Fprintln(w, "")
 804         }
 805
 806         err = w.Close()
 807         if err != nil {
 808                 return fmt.Errorf("While closing node-info logs: %v", err)
 809         }
 810         return nil
 811 }
 812
 813 // LogContainerRecord gets and saves the raw JSON container record from the API server
 814 func (runner *ContainerRunner) LogContainerRecord() error {
 815         logged, err := runner.logAPIResponse("container", "containers", map[string]interface{}{"filters": [][]string{{"uuid", "=", runner.Container.UUID}}}, nil)
 816         if !logged && err == nil {
 817                 err = fmt.Errorf("error: no container record found for %s", runner.Container.UUID)
 818         }
 819         return err
 820 }
 821
 822 // LogNodeRecord logs the current host's InstanceType config entry (or
 823 // the arvados#node record, if running via crunch-dispatch-slurm).
 824 func (runner *ContainerRunner) LogNodeRecord() error {
 825         if it := os.Getenv("InstanceType"); it != "" {
 826                 // Dispatched via arvados-dispatch-cloud. Save
 827                 // InstanceType config fragment received from
 828                 // dispatcher on stdin.
 829                 w, err := runner.LogCollection.OpenFile("node.json", os.O_CREATE|os.O_WRONLY, 0666)
 830                 if err != nil {
 831                         return err
 832                 }
 833                 defer w.Close()
 834                 _, err = io.WriteString(w, it)
 835                 if err != nil {
 836                         return err
 837                 }
 838                 return w.Close()
 839         }
 840         // Dispatched via crunch-dispatch-slurm. Look up
 841         // apiserver's node record corresponding to
 842         // $SLURMD_NODENAME.
 843         hostname := os.Getenv("SLURMD_NODENAME")
 844         if hostname == "" {
 845                 hostname, _ = os.Hostname()
 846         }
 847         _, err := runner.logAPIResponse("node", "nodes", map[string]interface{}{"filters": [][]string{{"hostname", "=", hostname}}}, func(resp interface{}) {
 848                 // The "info" field has admin-only info when
 849                 // obtained with a privileged token, and
 850                 // should not be logged.
 851                 node, ok := resp.(map[string]interface{})
 852                 if ok {
 853                         delete(node, "info")
 854                 }
 855         })
 856         return err
 857 }
 858
 859 func (runner *ContainerRunner) logAPIResponse(label, path string, params map[string]interface{}, munge func(interface{})) (logged bool, err error) {
 860         writer, err := runner.LogCollection.OpenFile(label+".json", os.O_CREATE|os.O_WRONLY, 0666)
 861         if err != nil {
 862                 return false, err
 863         }
 864         w := &ArvLogWriter{
 865                 ArvClient:     runner.DispatcherArvClient,
 866                 UUID:          runner.Container.UUID,
 867                 loggingStream: label,
 868                 writeCloser:   writer,
 869         }
 870
 871         reader, err := runner.DispatcherArvClient.CallRaw("GET", path, "", "", arvadosclient.Dict(params))
 872         if err != nil {
 873                 return false, fmt.Errorf("error getting %s record: %v", label, err)
 874         }
 875         defer reader.Close()
 876
 877         dec := json.NewDecoder(reader)
 878         dec.UseNumber()
 879         var resp map[string]interface{}
 880         if err = dec.Decode(&resp); err != nil {
 881                 return false, fmt.Errorf("error decoding %s list response: %v", label, err)
 882         }
 883         items, ok := resp["items"].([]interface{})
 884         if !ok {
 885                 return false, fmt.Errorf("error decoding %s list response: no \"items\" key in API list response", label)
 886         } else if len(items) < 1 {
 887                 return false, nil
 888         }
 889         if munge != nil {
 890                 munge(items[0])
 891         }
 892         // Re-encode it using indentation to improve readability
 893         enc := json.NewEncoder(w)
 894         enc.SetIndent("", "    ")
 895         if err = enc.Encode(items[0]); err != nil {
 896                 return false, fmt.Errorf("error logging %s record: %v", label, err)
 897         }
 898         err = w.Close()
 899         if err != nil {
 900                 return false, fmt.Errorf("error closing %s.json in log collection: %v", label, err)
 901         }
 902         return true, nil
 903 }
 904
 905 func (runner *ContainerRunner) getStdoutFile(mntPath string) (*os.File, error) {
 906         stdoutPath := mntPath[len(runner.Container.OutputPath):]
 907         index := strings.LastIndex(stdoutPath, "/")
 908         if index > 0 {
 909                 subdirs := stdoutPath[:index]
 910                 if subdirs != "" {
 911                         st, err := os.Stat(runner.HostOutputDir)
 912                         if err != nil {
 913                                 return nil, fmt.Errorf("While Stat on temp dir: %v", err)
 914                         }
 915                         stdoutPath := filepath.Join(runner.HostOutputDir, subdirs)
 916                         err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777)
 917                         if err != nil {
 918                                 return nil, fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err)
 919                         }
 920                 }
 921         }
 922         stdoutFile, err := os.Create(filepath.Join(runner.HostOutputDir, stdoutPath))
 923         if err != nil {
 924                 return nil, fmt.Errorf("While creating file %q: %v", stdoutPath, err)
 925         }
 926
 927         return stdoutFile, nil
 928 }
 929
 930 // CreateContainer creates the docker container.
 931 func (runner *ContainerRunner) CreateContainer(imageID string, bindmounts map[string]bindmount) error {
 932         var stdin io.ReadCloser = ioutil.NopCloser(bytes.NewReader(nil))
 933         if mnt, ok := runner.Container.Mounts["stdin"]; ok {
 934                 switch mnt.Kind {
 935                 case "collection":
 936                         var collID string
 937                         if mnt.UUID != "" {
 938                                 collID = mnt.UUID
 939                         } else {
 940                                 collID = mnt.PortableDataHash
 941                         }
 942                         path := runner.ArvMountPoint + "/by_id/" + collID + "/" + mnt.Path
 943                         f, err := os.Open(path)
 944                         if err != nil {
 945                                 return err
 946                         }
 947                         stdin = f
 948                 case "json":
 949                         j, err := json.Marshal(mnt.Content)
 950                         if err != nil {
 951                                 return fmt.Errorf("error encoding stdin json data: %v", err)
 952                         }
 953                         stdin = ioutil.NopCloser(bytes.NewReader(j))
 954                 default:
 955                         return fmt.Errorf("stdin mount has unsupported kind %q", mnt.Kind)
 956                 }
 957         }
 958
 959         var stdout, stderr io.WriteCloser
 960         if mnt, ok := runner.Container.Mounts["stdout"]; ok {
 961                 f, err := runner.getStdoutFile(mnt.Path)
 962                 if err != nil {
 963                         return err
 964                 }
 965                 stdout = f
 966         } else if w, err := runner.NewLogWriter("stdout"); err != nil {
 967                 return err
 968         } else {
 969                 stdout = NewThrottledLogger(w)
 970         }
 971
 972         if mnt, ok := runner.Container.Mounts["stderr"]; ok {
 973                 f, err := runner.getStdoutFile(mnt.Path)
 974                 if err != nil {
 975                         return err
 976                 }
 977                 stderr = f
 978         } else if w, err := runner.NewLogWriter("stderr"); err != nil {
 979                 return err
 980         } else {
 981                 stderr = NewThrottledLogger(w)
 982         }
 983
 984         env := runner.Container.Environment
 985         enableNetwork := runner.enableNetwork == "always"
 986         if runner.Container.RuntimeConstraints.API {
 987                 enableNetwork = true
 988                 tok, err := runner.ContainerToken()
 989                 if err != nil {
 990                         return err
 991                 }
 992                 env = map[string]string{}
 993                 for k, v := range runner.Container.Environment {
 994                         env[k] = v
 995                 }
 996                 env["ARVADOS_API_TOKEN"] = tok
 997                 env["ARVADOS_API_HOST"] = os.Getenv("ARVADOS_API_HOST")
 998                 env["ARVADOS_API_HOST_INSECURE"] = os.Getenv("ARVADOS_API_HOST_INSECURE")
 999         }
1000         workdir := runner.Container.Cwd
1001         if workdir == "." {
1002                 // both "" and "." mean default
1003                 workdir = ""
1004         }
1005         ram := runner.Container.RuntimeConstraints.RAM
1006         if !runner.enableMemoryLimit {
1007                 ram = 0
1008         }
1009         runner.executorStdin = stdin
1010         runner.executorStdout = stdout
1011         runner.executorStderr = stderr
1012
1013         if runner.Container.RuntimeConstraints.CUDA.DeviceCount > 0 {
1014                 nvidiaModprobe(runner.CrunchLog)
1015         }
1016
1017         return runner.executor.Create(containerSpec{
1018                 Image:           imageID,
1019                 VCPUs:           runner.Container.RuntimeConstraints.VCPUs,
1020                 RAM:             ram,
1021                 WorkingDir:      workdir,
1022                 Env:             env,
1023                 BindMounts:      bindmounts,
1024                 Command:         runner.Container.Command,
1025                 EnableNetwork:   enableNetwork,
1026                 CUDADeviceCount: runner.Container.RuntimeConstraints.CUDA.DeviceCount,
1027                 NetworkMode:     runner.networkMode,
1028                 CgroupParent:    runner.setCgroupParent,
1029                 Stdin:           stdin,
1030                 Stdout:          stdout,
1031                 Stderr:          stderr,
1032         })
1033 }
1034
1035 // StartContainer starts the docker container created by CreateContainer.
1036 func (runner *ContainerRunner) StartContainer() error {
1037         runner.CrunchLog.Printf("Starting container")
1038         runner.cStateLock.Lock()
1039         defer runner.cStateLock.Unlock()
1040         if runner.cCancelled {
1041                 return ErrCancelled
1042         }
1043         err := runner.executor.Start()
1044         if err != nil {
1045                 var advice string
1046                 if m, e := regexp.MatchString("(?ms).*(exec|System error).*(no such file or directory|file not found).*", err.Error()); m && e == nil {
1047                         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])
1048                 }
1049                 return fmt.Errorf("could not start container: %v%s", err, advice)
1050         }
1051         return nil
1052 }
1053
1054 // WaitFinish waits for the container to terminate, capture the exit code, and
1055 // close the stdout/stderr logging.
1056 func (runner *ContainerRunner) WaitFinish() error {
1057         runner.CrunchLog.Print("Waiting for container to finish")
1058         var timeout <-chan time.Time
1059         if s := runner.Container.SchedulingParameters.MaxRunTime; s > 0 {
1060                 timeout = time.After(time.Duration(s) * time.Second)
1061         }
1062         ctx, cancel := context.WithCancel(context.Background())
1063         defer cancel()
1064         go func() {
1065                 select {
1066                 case <-timeout:
1067                         runner.CrunchLog.Printf("maximum run time exceeded. Stopping container.")
1068                         runner.stop(nil)
1069                 case <-runner.ArvMountExit:
1070                         runner.CrunchLog.Printf("arv-mount exited while container is still running. Stopping container.")
1071                         runner.stop(nil)
1072                 case <-ctx.Done():
1073                 }
1074         }()
1075         exitcode, err := runner.executor.Wait(ctx)
1076         if err != nil {
1077                 runner.checkBrokenNode(err)
1078                 return err
1079         }
1080         runner.ExitCode = &exitcode
1081
1082         extra := ""
1083         if exitcode&0x80 != 0 {
1084                 // Convert raw exit status (0x80 + signal number) to a
1085                 // string to log after the code, like " (signal 101)"
1086                 // or " (signal 9, killed)"
1087                 sig := syscall.WaitStatus(exitcode).Signal()
1088                 if name := unix.SignalName(sig); name != "" {
1089                         extra = fmt.Sprintf(" (signal %d, %s)", sig, name)
1090                 } else {
1091                         extra = fmt.Sprintf(" (signal %d)", sig)
1092                 }
1093         }
1094         runner.CrunchLog.Printf("Container exited with status code %d%s", exitcode, extra)
1095
1096         var returnErr error
1097         if err = runner.executorStdin.Close(); err != nil {
1098                 err = fmt.Errorf("error closing container stdin: %s", err)
1099                 runner.CrunchLog.Printf("%s", err)
1100                 returnErr = err
1101         }
1102         if err = runner.executorStdout.Close(); err != nil {
1103                 err = fmt.Errorf("error closing container stdout: %s", err)
1104                 runner.CrunchLog.Printf("%s", err)
1105                 if returnErr == nil {
1106                         returnErr = err
1107                 }
1108         }
1109         if err = runner.executorStderr.Close(); err != nil {
1110                 err = fmt.Errorf("error closing container stderr: %s", err)
1111                 runner.CrunchLog.Printf("%s", err)
1112                 if returnErr == nil {
1113                         returnErr = err
1114                 }
1115         }
1116
1117         if runner.statReporter != nil {
1118                 runner.statReporter.Stop()
1119                 err = runner.statLogger.Close()
1120                 if err != nil {
1121                         runner.CrunchLog.Printf("error closing crunchstat logs: %v", err)
1122                 }
1123         }
1124         return returnErr
1125 }
1126
1127 func (runner *ContainerRunner) updateLogs() {
1128         ticker := time.NewTicker(crunchLogUpdatePeriod / 360)
1129         defer ticker.Stop()
1130
1131         sigusr1 := make(chan os.Signal, 1)
1132         signal.Notify(sigusr1, syscall.SIGUSR1)
1133         defer signal.Stop(sigusr1)
1134
1135         saveAtTime := time.Now().Add(crunchLogUpdatePeriod)
1136         saveAtSize := crunchLogUpdateSize
1137         var savedSize int64
1138         for {
1139                 select {
1140                 case <-ticker.C:
1141                 case <-sigusr1:
1142                         saveAtTime = time.Now()
1143                 }
1144                 runner.logMtx.Lock()
1145                 done := runner.LogsPDH != nil
1146                 runner.logMtx.Unlock()
1147                 if done {
1148                         return
1149                 }
1150                 size := runner.LogCollection.Size()
1151                 if size == savedSize || (time.Now().Before(saveAtTime) && size < saveAtSize) {
1152                         continue
1153                 }
1154                 saveAtTime = time.Now().Add(crunchLogUpdatePeriod)
1155                 saveAtSize = runner.LogCollection.Size() + crunchLogUpdateSize
1156                 saved, err := runner.saveLogCollection(false)
1157                 if err != nil {
1158                         runner.CrunchLog.Printf("error updating log collection: %s", err)
1159                         continue
1160                 }
1161
1162                 var updated arvados.Container
1163                 err = runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1164                         "container": arvadosclient.Dict{"log": saved.PortableDataHash},
1165                 }, &updated)
1166                 if err != nil {
1167                         runner.CrunchLog.Printf("error updating container log to %s: %s", saved.PortableDataHash, err)
1168                         continue
1169                 }
1170
1171                 savedSize = size
1172         }
1173 }
1174
1175 func (runner *ContainerRunner) reportArvMountWarning(pattern, text string) {
1176         var updated arvados.Container
1177         err := runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1178                 "container": arvadosclient.Dict{
1179                         "runtime_status": arvadosclient.Dict{
1180                                 "warning":       "arv-mount: " + pattern,
1181                                 "warningDetail": text,
1182                         },
1183                 },
1184         }, &updated)
1185         if err != nil {
1186                 runner.CrunchLog.Printf("error updating container runtime_status: %s", err)
1187         }
1188 }
1189
1190 // CaptureOutput saves data from the container's output directory if
1191 // needed, and updates the container output accordingly.
1192 func (runner *ContainerRunner) CaptureOutput(bindmounts map[string]bindmount) error {
1193         if runner.Container.RuntimeConstraints.API {
1194                 // Output may have been set directly by the container, so
1195                 // refresh the container record to check.
1196                 err := runner.DispatcherArvClient.Get("containers", runner.Container.UUID,
1197                         nil, &runner.Container)
1198                 if err != nil {
1199                         return err
1200                 }
1201                 if runner.Container.Output != "" {
1202                         // Container output is already set.
1203                         runner.OutputPDH = &runner.Container.Output
1204                         return nil
1205                 }
1206         }
1207
1208         txt, err := (&copier{
1209                 client:        runner.containerClient,
1210                 arvClient:     runner.ContainerArvClient,
1211                 keepClient:    runner.ContainerKeepClient,
1212                 hostOutputDir: runner.HostOutputDir,
1213                 ctrOutputDir:  runner.Container.OutputPath,
1214                 bindmounts:    bindmounts,
1215                 mounts:        runner.Container.Mounts,
1216                 secretMounts:  runner.SecretMounts,
1217                 logger:        runner.CrunchLog,
1218         }).Copy()
1219         if err != nil {
1220                 return err
1221         }
1222         if n := len(regexp.MustCompile(` [0-9a-f]+\+\S*\+R`).FindAllStringIndex(txt, -1)); n > 0 {
1223                 runner.CrunchLog.Printf("Copying %d data blocks from remote input collections...", n)
1224                 fs, err := (&arvados.Collection{ManifestText: txt}).FileSystem(runner.containerClient, runner.ContainerKeepClient)
1225                 if err != nil {
1226                         return err
1227                 }
1228                 txt, err = fs.MarshalManifest(".")
1229                 if err != nil {
1230                         return err
1231                 }
1232         }
1233         var resp arvados.Collection
1234         err = runner.ContainerArvClient.Create("collections", arvadosclient.Dict{
1235                 "ensure_unique_name": true,
1236                 "collection": arvadosclient.Dict{
1237                         "is_trashed":    true,
1238                         "name":          "output for " + runner.Container.UUID,
1239                         "manifest_text": txt,
1240                 },
1241         }, &resp)
1242         if err != nil {
1243                 return fmt.Errorf("error creating output collection: %v", err)
1244         }
1245         runner.OutputPDH = &resp.PortableDataHash
1246         return nil
1247 }
1248
1249 func (runner *ContainerRunner) CleanupDirs() {
1250         if runner.ArvMount != nil {
1251                 var delay int64 = 8
1252                 umount := exec.Command("arv-mount", fmt.Sprintf("--unmount-timeout=%d", delay), "--unmount", runner.ArvMountPoint)
1253                 umount.Stdout = runner.CrunchLog
1254                 umount.Stderr = runner.CrunchLog
1255                 runner.CrunchLog.Printf("Running %v", umount.Args)
1256                 umnterr := umount.Start()
1257
1258                 if umnterr != nil {
1259                         runner.CrunchLog.Printf("Error unmounting: %v", umnterr)
1260                         runner.ArvMount.Process.Kill()
1261                 } else {
1262                         // If arv-mount --unmount gets stuck for any reason, we
1263                         // don't want to wait for it forever.  Do Wait() in a goroutine
1264                         // so it doesn't block crunch-run.
1265                         umountExit := make(chan error)
1266                         go func() {
1267                                 mnterr := umount.Wait()
1268                                 if mnterr != nil {
1269                                         runner.CrunchLog.Printf("Error unmounting: %v", mnterr)
1270                                 }
1271                                 umountExit <- mnterr
1272                         }()
1273
1274                         for again := true; again; {
1275                                 again = false
1276                                 select {
1277                                 case <-umountExit:
1278                                         umount = nil
1279                                         again = true
1280                                 case <-runner.ArvMountExit:
1281                                         break
1282                                 case <-time.After(time.Duration((delay + 1) * int64(time.Second))):
1283                                         runner.CrunchLog.Printf("Timed out waiting for unmount")
1284                                         if umount != nil {
1285                                                 umount.Process.Kill()
1286                                         }
1287                                         runner.ArvMount.Process.Kill()
1288                                 }
1289                         }
1290                 }
1291                 runner.ArvMount = nil
1292         }
1293
1294         if runner.ArvMountPoint != "" {
1295                 if rmerr := os.Remove(runner.ArvMountPoint); rmerr != nil {
1296                         runner.CrunchLog.Printf("While cleaning up arv-mount directory %s: %v", runner.ArvMountPoint, rmerr)
1297                 }
1298                 runner.ArvMountPoint = ""
1299         }
1300
1301         if rmerr := os.RemoveAll(runner.parentTemp); rmerr != nil {
1302                 runner.CrunchLog.Printf("While cleaning up temporary directory %s: %v", runner.parentTemp, rmerr)
1303         }
1304 }
1305
1306 // CommitLogs posts the collection containing the final container logs.
1307 func (runner *ContainerRunner) CommitLogs() error {
1308         func() {
1309                 // Hold cStateLock to prevent races on CrunchLog (e.g., stop()).
1310                 runner.cStateLock.Lock()
1311                 defer runner.cStateLock.Unlock()
1312
1313                 runner.CrunchLog.Print(runner.finalState)
1314
1315                 if runner.arvMountLog != nil {
1316                         runner.arvMountLog.Close()
1317                 }
1318                 runner.CrunchLog.Close()
1319
1320                 // Closing CrunchLog above allows them to be committed to Keep at this
1321                 // point, but re-open crunch log with ArvClient in case there are any
1322                 // other further errors (such as failing to write the log to Keep!)
1323                 // while shutting down
1324                 runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{
1325                         ArvClient:     runner.DispatcherArvClient,
1326                         UUID:          runner.Container.UUID,
1327                         loggingStream: "crunch-run",
1328                         writeCloser:   nil,
1329                 })
1330                 runner.CrunchLog.Immediate = log.New(os.Stderr, runner.Container.UUID+" ", 0)
1331         }()
1332
1333         if runner.keepstoreLogger != nil {
1334                 // Flush any buffered logs from our local keepstore
1335                 // process.  Discard anything logged after this point
1336                 // -- it won't end up in the log collection, so
1337                 // there's no point writing it to the collectionfs.
1338                 runner.keepstoreLogbuf.SetWriter(io.Discard)
1339                 runner.keepstoreLogger.Close()
1340                 runner.keepstoreLogger = nil
1341         }
1342
1343         if runner.LogsPDH != nil {
1344                 // If we have already assigned something to LogsPDH,
1345                 // we must be closing the re-opened log, which won't
1346                 // end up getting attached to the container record and
1347                 // therefore doesn't need to be saved as a collection
1348                 // -- it exists only to send logs to other channels.
1349                 return nil
1350         }
1351
1352         saved, err := runner.saveLogCollection(true)
1353         if err != nil {
1354                 return fmt.Errorf("error saving log collection: %s", err)
1355         }
1356         runner.logMtx.Lock()
1357         defer runner.logMtx.Unlock()
1358         runner.LogsPDH = &saved.PortableDataHash
1359         return nil
1360 }
1361
1362 func (runner *ContainerRunner) saveLogCollection(final bool) (response arvados.Collection, err error) {
1363         runner.logMtx.Lock()
1364         defer runner.logMtx.Unlock()
1365         if runner.LogsPDH != nil {
1366                 // Already finalized.
1367                 return
1368         }
1369         updates := arvadosclient.Dict{
1370                 "name": "logs for " + runner.Container.UUID,
1371         }
1372         mt, err1 := runner.LogCollection.MarshalManifest(".")
1373         if err1 == nil {
1374                 // Only send updated manifest text if there was no
1375                 // error.
1376                 updates["manifest_text"] = mt
1377         }
1378
1379         // Even if flushing the manifest had an error, we still want
1380         // to update the log record, if possible, to push the trash_at
1381         // and delete_at times into the future.  Details on bug
1382         // #17293.
1383         if final {
1384                 updates["is_trashed"] = true
1385         } else {
1386                 exp := time.Now().Add(crunchLogUpdatePeriod * 24)
1387                 updates["trash_at"] = exp
1388                 updates["delete_at"] = exp
1389         }
1390         reqBody := arvadosclient.Dict{"collection": updates}
1391         var err2 error
1392         if runner.logUUID == "" {
1393                 reqBody["ensure_unique_name"] = true
1394                 err2 = runner.DispatcherArvClient.Create("collections", reqBody, &response)
1395         } else {
1396                 err2 = runner.DispatcherArvClient.Update("collections", runner.logUUID, reqBody, &response)
1397         }
1398         if err2 == nil {
1399                 runner.logUUID = response.UUID
1400         }
1401
1402         if err1 != nil || err2 != nil {
1403                 err = fmt.Errorf("error recording logs: %q, %q", err1, err2)
1404         }
1405         return
1406 }
1407
1408 // UpdateContainerRunning updates the container state to "Running"
1409 func (runner *ContainerRunner) UpdateContainerRunning() error {
1410         runner.cStateLock.Lock()
1411         defer runner.cStateLock.Unlock()
1412         if runner.cCancelled {
1413                 return ErrCancelled
1414         }
1415         return runner.DispatcherArvClient.Update("containers", runner.Container.UUID,
1416                 arvadosclient.Dict{"container": arvadosclient.Dict{"state": "Running", "gateway_address": runner.gateway.Address}}, nil)
1417 }
1418
1419 // ContainerToken returns the api_token the container (and any
1420 // arv-mount processes) are allowed to use.
1421 func (runner *ContainerRunner) ContainerToken() (string, error) {
1422         if runner.token != "" {
1423                 return runner.token, nil
1424         }
1425
1426         var auth arvados.APIClientAuthorization
1427         err := runner.DispatcherArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth)
1428         if err != nil {
1429                 return "", err
1430         }
1431         runner.token = fmt.Sprintf("v2/%s/%s/%s", auth.UUID, auth.APIToken, runner.Container.UUID)
1432         return runner.token, nil
1433 }
1434
1435 // UpdateContainerFinal updates the container record state on API
1436 // server to "Complete" or "Cancelled"
1437 func (runner *ContainerRunner) UpdateContainerFinal() error {
1438         update := arvadosclient.Dict{}
1439         update["state"] = runner.finalState
1440         if runner.LogsPDH != nil {
1441                 update["log"] = *runner.LogsPDH
1442         }
1443         if runner.finalState == "Complete" {
1444                 if runner.ExitCode != nil {
1445                         update["exit_code"] = *runner.ExitCode
1446                 }
1447                 if runner.OutputPDH != nil {
1448                         update["output"] = *runner.OutputPDH
1449                 }
1450         }
1451         return runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{"container": update}, nil)
1452 }
1453
1454 // IsCancelled returns the value of Cancelled, with goroutine safety.
1455 func (runner *ContainerRunner) IsCancelled() bool {
1456         runner.cStateLock.Lock()
1457         defer runner.cStateLock.Unlock()
1458         return runner.cCancelled
1459 }
1460
1461 // NewArvLogWriter creates an ArvLogWriter
1462 func (runner *ContainerRunner) NewArvLogWriter(name string) (io.WriteCloser, error) {
1463         writer, err := runner.LogCollection.OpenFile(name+".txt", os.O_CREATE|os.O_WRONLY, 0666)
1464         if err != nil {
1465                 return nil, err
1466         }
1467         return &ArvLogWriter{
1468                 ArvClient:     runner.DispatcherArvClient,
1469                 UUID:          runner.Container.UUID,
1470                 loggingStream: name,
1471                 writeCloser:   writer,
1472         }, nil
1473 }
1474
1475 // Run the full container lifecycle.
1476 func (runner *ContainerRunner) Run() (err error) {
1477         runner.CrunchLog.Printf("crunch-run %s started", cmd.Version.String())
1478         runner.CrunchLog.Printf("Executing container '%s' using %s runtime", runner.Container.UUID, runner.executor.Runtime())
1479
1480         hostname, hosterr := os.Hostname()
1481         if hosterr != nil {
1482                 runner.CrunchLog.Printf("Error getting hostname '%v'", hosterr)
1483         } else {
1484                 runner.CrunchLog.Printf("Executing on host '%s'", hostname)
1485         }
1486
1487         runner.finalState = "Queued"
1488
1489         defer func() {
1490                 runner.CleanupDirs()
1491
1492                 runner.CrunchLog.Printf("crunch-run finished")
1493                 runner.CrunchLog.Close()
1494         }()
1495
1496         err = runner.fetchContainerRecord()
1497         if err != nil {
1498                 return
1499         }
1500         if runner.Container.State != "Locked" {
1501                 return fmt.Errorf("dispatch error detected: container %q has state %q", runner.Container.UUID, runner.Container.State)
1502         }
1503
1504         var bindmounts map[string]bindmount
1505         defer func() {
1506                 // checkErr prints e (unless it's nil) and sets err to
1507                 // e (unless err is already non-nil). Thus, if err
1508                 // hasn't already been assigned when Run() returns,
1509                 // this cleanup func will cause Run() to return the
1510                 // first non-nil error that is passed to checkErr().
1511                 checkErr := func(errorIn string, e error) {
1512                         if e == nil {
1513                                 return
1514                         }
1515                         runner.CrunchLog.Printf("error in %s: %v", errorIn, e)
1516                         if err == nil {
1517                                 err = e
1518                         }
1519                         if runner.finalState == "Complete" {
1520                                 // There was an error in the finalization.
1521                                 runner.finalState = "Cancelled"
1522                         }
1523                 }
1524
1525                 // Log the error encountered in Run(), if any
1526                 checkErr("Run", err)
1527
1528                 if runner.finalState == "Queued" {
1529                         runner.UpdateContainerFinal()
1530                         return
1531                 }
1532
1533                 if runner.IsCancelled() {
1534                         runner.finalState = "Cancelled"
1535                         // but don't return yet -- we still want to
1536                         // capture partial output and write logs
1537                 }
1538
1539                 if bindmounts != nil {
1540                         checkErr("CaptureOutput", runner.CaptureOutput(bindmounts))
1541                 }
1542                 checkErr("stopHoststat", runner.stopHoststat())
1543                 checkErr("CommitLogs", runner.CommitLogs())
1544                 runner.CleanupDirs()
1545                 checkErr("UpdateContainerFinal", runner.UpdateContainerFinal())
1546         }()
1547
1548         runner.setupSignals()
1549         err = runner.startHoststat()
1550         if err != nil {
1551                 return
1552         }
1553
1554         // set up FUSE mount and binds
1555         bindmounts, err = runner.SetupMounts()
1556         if err != nil {
1557                 runner.finalState = "Cancelled"
1558                 err = fmt.Errorf("While setting up mounts: %v", err)
1559                 return
1560         }
1561
1562         // check for and/or load image
1563         imageID, err := runner.LoadImage()
1564         if err != nil {
1565                 if !runner.checkBrokenNode(err) {
1566                         // Failed to load image but not due to a "broken node"
1567                         // condition, probably user error.
1568                         runner.finalState = "Cancelled"
1569                 }
1570                 err = fmt.Errorf("While loading container image: %v", err)
1571                 return
1572         }
1573
1574         err = runner.CreateContainer(imageID, bindmounts)
1575         if err != nil {
1576                 return
1577         }
1578         err = runner.LogHostInfo()
1579         if err != nil {
1580                 return
1581         }
1582         err = runner.LogNodeRecord()
1583         if err != nil {
1584                 return
1585         }
1586         err = runner.LogContainerRecord()
1587         if err != nil {
1588                 return
1589         }
1590
1591         if runner.IsCancelled() {
1592                 return
1593         }
1594
1595         err = runner.UpdateContainerRunning()
1596         if err != nil {
1597                 return
1598         }
1599         runner.finalState = "Cancelled"
1600
1601         err = runner.startCrunchstat()
1602         if err != nil {
1603                 return
1604         }
1605
1606         err = runner.StartContainer()
1607         if err != nil {
1608                 runner.checkBrokenNode(err)
1609                 return
1610         }
1611
1612         err = runner.WaitFinish()
1613         if err == nil && !runner.IsCancelled() {
1614                 runner.finalState = "Complete"
1615         }
1616         return
1617 }
1618
1619 // Fetch the current container record (uuid = runner.Container.UUID)
1620 // into runner.Container.
1621 func (runner *ContainerRunner) fetchContainerRecord() error {
1622         reader, err := runner.DispatcherArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
1623         if err != nil {
1624                 return fmt.Errorf("error fetching container record: %v", err)
1625         }
1626         defer reader.Close()
1627
1628         dec := json.NewDecoder(reader)
1629         dec.UseNumber()
1630         err = dec.Decode(&runner.Container)
1631         if err != nil {
1632                 return fmt.Errorf("error decoding container record: %v", err)
1633         }
1634
1635         var sm struct {
1636                 SecretMounts map[string]arvados.Mount `json:"secret_mounts"`
1637         }
1638
1639         containerToken, err := runner.ContainerToken()
1640         if err != nil {
1641                 return fmt.Errorf("error getting container token: %v", err)
1642         }
1643
1644         runner.ContainerArvClient, runner.ContainerKeepClient,
1645                 runner.containerClient, err = runner.MkArvClient(containerToken)
1646         if err != nil {
1647                 return fmt.Errorf("error creating container API client: %v", err)
1648         }
1649
1650         runner.ContainerKeepClient.SetStorageClasses(runner.Container.OutputStorageClasses)
1651         runner.DispatcherKeepClient.SetStorageClasses(runner.Container.OutputStorageClasses)
1652
1653         err = runner.ContainerArvClient.Call("GET", "containers", runner.Container.UUID, "secret_mounts", nil, &sm)
1654         if err != nil {
1655                 if apierr, ok := err.(arvadosclient.APIServerError); !ok || apierr.HttpStatusCode != 404 {
1656                         return fmt.Errorf("error fetching secret_mounts: %v", err)
1657                 }
1658                 // ok && apierr.HttpStatusCode == 404, which means
1659                 // secret_mounts isn't supported by this API server.
1660         }
1661         runner.SecretMounts = sm.SecretMounts
1662
1663         return nil
1664 }
1665
1666 // NewContainerRunner creates a new container runner.
1667 func NewContainerRunner(dispatcherClient *arvados.Client,
1668         dispatcherArvClient IArvadosClient,
1669         dispatcherKeepClient IKeepClient,
1670         containerUUID string) (*ContainerRunner, error) {
1671
1672         cr := &ContainerRunner{
1673                 dispatcherClient:     dispatcherClient,
1674                 DispatcherArvClient:  dispatcherArvClient,
1675                 DispatcherKeepClient: dispatcherKeepClient,
1676         }
1677         cr.NewLogWriter = cr.NewArvLogWriter
1678         cr.RunArvMount = cr.ArvMountCmd
1679         cr.MkTempDir = ioutil.TempDir
1680         cr.MkArvClient = func(token string) (IArvadosClient, IKeepClient, *arvados.Client, error) {
1681                 cl, err := arvadosclient.MakeArvadosClient()
1682                 if err != nil {
1683                         return nil, nil, nil, err
1684                 }
1685                 cl.ApiToken = token
1686                 kc, err := keepclient.MakeKeepClient(cl)
1687                 if err != nil {
1688                         return nil, nil, nil, err
1689                 }
1690                 c2 := arvados.NewClientFromEnv()
1691                 c2.AuthToken = token
1692                 return cl, kc, c2, nil
1693         }
1694         var err error
1695         cr.LogCollection, err = (&arvados.Collection{}).FileSystem(cr.dispatcherClient, cr.DispatcherKeepClient)
1696         if err != nil {
1697                 return nil, err
1698         }
1699         cr.Container.UUID = containerUUID
1700         w, err := cr.NewLogWriter("crunch-run")
1701         if err != nil {
1702                 return nil, err
1703         }
1704         cr.CrunchLog = NewThrottledLogger(w)
1705         cr.CrunchLog.Immediate = log.New(os.Stderr, containerUUID+" ", 0)
1706
1707         loadLogThrottleParams(dispatcherArvClient)
1708         go cr.updateLogs()
1709
1710         return cr, nil
1711 }
1712
1713 func (command) RunCommand(prog string, args []string, stdin io.Reader, stdout, stderr io.Writer) int {
1714         log := log.New(stderr, "", 0)
1715         flags := flag.NewFlagSet(prog, flag.ContinueOnError)
1716         statInterval := flags.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting")
1717         cgroupRoot := flags.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree")
1718         cgroupParent := flags.String("cgroup-parent", "docker", "name of container's parent cgroup (ignored if -cgroup-parent-subsystem is used)")
1719         cgroupParentSubsystem := flags.String("cgroup-parent-subsystem", "", "use current cgroup for given subsystem as parent cgroup for container")
1720         caCertsPath := flags.String("ca-certs", "", "Path to TLS root certificates")
1721         detach := flags.Bool("detach", false, "Detach from parent process and run in the background")
1722         stdinConfig := flags.Bool("stdin-config", false, "Load config and environment variables from JSON message on stdin")
1723         sleep := flags.Duration("sleep", 0, "Delay before starting (testing use only)")
1724         kill := flags.Int("kill", -1, "Send signal to an existing crunch-run process for given UUID")
1725         list := flags.Bool("list", false, "List UUIDs of existing crunch-run processes")
1726         enableMemoryLimit := flags.Bool("enable-memory-limit", true, "tell container runtime to limit container's memory usage")
1727         enableNetwork := flags.String("container-enable-networking", "default", "enable networking \"always\" (for all containers) or \"default\" (for containers that request it)")
1728         networkMode := flags.String("container-network-mode", "default", `Docker network mode for container (use any argument valid for docker --net)`)
1729         memprofile := flags.String("memprofile", "", "write memory profile to `file` after running container")
1730         runtimeEngine := flags.String("runtime-engine", "docker", "container runtime: docker or singularity")
1731         flags.Duration("check-containerd", 0, "Ignored. Exists for compatibility with older versions.")
1732
1733         ignoreDetachFlag := false
1734         if len(args) > 0 && args[0] == "-no-detach" {
1735                 // This process was invoked by a parent process, which
1736                 // has passed along its own arguments, including
1737                 // -detach, after the leading -no-detach flag.  Strip
1738                 // the leading -no-detach flag (it's not recognized by
1739                 // flags.Parse()) and ignore the -detach flag that
1740                 // comes later.
1741                 args = args[1:]
1742                 ignoreDetachFlag = true
1743         }
1744
1745         if ok, code := cmd.ParseFlags(flags, prog, args, "container-uuid", stderr); !ok {
1746                 return code
1747         } else if !*list && flags.NArg() != 1 {
1748                 fmt.Fprintf(stderr, "missing required argument: container-uuid (try -help)\n")
1749                 return 2
1750         }
1751
1752         containerUUID := flags.Arg(0)
1753
1754         switch {
1755         case *detach && !ignoreDetachFlag:
1756                 return Detach(containerUUID, prog, args, os.Stdin, os.Stdout, os.Stderr)
1757         case *kill >= 0:
1758                 return KillProcess(containerUUID, syscall.Signal(*kill), os.Stdout, os.Stderr)
1759         case *list:
1760                 return ListProcesses(os.Stdout, os.Stderr)
1761         }
1762
1763         if len(containerUUID) != 27 {
1764                 log.Printf("usage: %s [options] UUID", prog)
1765                 return 1
1766         }
1767
1768         var conf ConfigData
1769         if *stdinConfig {
1770                 err := json.NewDecoder(stdin).Decode(&conf)
1771                 if err != nil {
1772                         log.Printf("decode stdin: %s", err)
1773                         return 1
1774                 }
1775                 for k, v := range conf.Env {
1776                         err = os.Setenv(k, v)
1777                         if err != nil {
1778                                 log.Printf("setenv(%q): %s", k, err)
1779                                 return 1
1780                         }
1781                 }
1782                 if conf.Cluster != nil {
1783                         // ClusterID is missing from the JSON
1784                         // representation, but we need it to generate
1785                         // a valid config file for keepstore, so we
1786                         // fill it using the container UUID prefix.
1787                         conf.Cluster.ClusterID = containerUUID[:5]
1788                 }
1789         }
1790
1791         log.Printf("crunch-run %s started", cmd.Version.String())
1792         time.Sleep(*sleep)
1793
1794         if *caCertsPath != "" {
1795                 arvadosclient.CertFiles = []string{*caCertsPath}
1796         }
1797
1798         var keepstoreLogbuf bufThenWrite
1799         keepstore, err := startLocalKeepstore(conf, io.MultiWriter(&keepstoreLogbuf, stderr))
1800         if err != nil {
1801                 log.Print(err)
1802                 return 1
1803         }
1804         if keepstore != nil {
1805                 defer keepstore.Process.Kill()
1806         }
1807
1808         api, err := arvadosclient.MakeArvadosClient()
1809         if err != nil {
1810                 log.Printf("%s: %v", containerUUID, err)
1811                 return 1
1812         }
1813         api.Retries = 8
1814
1815         kc, err := keepclient.MakeKeepClient(api)
1816         if err != nil {
1817                 log.Printf("%s: %v", containerUUID, err)
1818                 return 1
1819         }
1820         kc.BlockCache = &keepclient.BlockCache{MaxBlocks: 2}
1821         kc.Retries = 4
1822
1823         cr, err := NewContainerRunner(arvados.NewClientFromEnv(), api, kc, containerUUID)
1824         if err != nil {
1825                 log.Print(err)
1826                 return 1
1827         }
1828
1829         if keepstore == nil {
1830                 // Log explanation (if any) for why we're not running
1831                 // a local keepstore.
1832                 var buf bytes.Buffer
1833                 keepstoreLogbuf.SetWriter(&buf)
1834                 if buf.Len() > 0 {
1835                         cr.CrunchLog.Printf("%s", strings.TrimSpace(buf.String()))
1836                 }
1837         } else if logWhat := conf.Cluster.Containers.LocalKeepLogsToContainerLog; logWhat == "none" {
1838                 cr.CrunchLog.Printf("using local keepstore process (pid %d) at %s", keepstore.Process.Pid, os.Getenv("ARVADOS_KEEP_SERVICES"))
1839                 keepstoreLogbuf.SetWriter(io.Discard)
1840         } else {
1841                 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"))
1842                 logwriter, err := cr.NewLogWriter("keepstore")
1843                 if err != nil {
1844                         log.Print(err)
1845                         return 1
1846                 }
1847                 cr.keepstoreLogger = NewThrottledLogger(logwriter)
1848
1849                 var writer io.WriteCloser = cr.keepstoreLogger
1850                 if logWhat == "errors" {
1851                         writer = &filterKeepstoreErrorsOnly{WriteCloser: writer}
1852                 } else if logWhat != "all" {
1853                         // should have been caught earlier by
1854                         // dispatcher's config loader
1855                         log.Printf("invalid value for Containers.LocalKeepLogsToContainerLog: %q", logWhat)
1856                         return 1
1857                 }
1858                 err = keepstoreLogbuf.SetWriter(writer)
1859                 if err != nil {
1860                         log.Print(err)
1861                         return 1
1862                 }
1863                 cr.keepstoreLogbuf = &keepstoreLogbuf
1864         }
1865
1866         switch *runtimeEngine {
1867         case "docker":
1868                 cr.executor, err = newDockerExecutor(containerUUID, cr.CrunchLog.Printf, cr.containerWatchdogInterval)
1869         case "singularity":
1870                 cr.executor, err = newSingularityExecutor(cr.CrunchLog.Printf)
1871         default:
1872                 cr.CrunchLog.Printf("%s: unsupported RuntimeEngine %q", containerUUID, *runtimeEngine)
1873                 cr.CrunchLog.Close()
1874                 return 1
1875         }
1876         if err != nil {
1877                 cr.CrunchLog.Printf("%s: %v", containerUUID, err)
1878                 cr.checkBrokenNode(err)
1879                 cr.CrunchLog.Close()
1880                 return 1
1881         }
1882         defer cr.executor.Close()
1883
1884         gwAuthSecret := os.Getenv("GatewayAuthSecret")
1885         os.Unsetenv("GatewayAuthSecret")
1886         if gwAuthSecret == "" {
1887                 // not safe to run a gateway service without an auth
1888                 // secret
1889                 cr.CrunchLog.Printf("Not starting a gateway server (GatewayAuthSecret was not provided by dispatcher)")
1890         } else if gwListen := os.Getenv("GatewayAddress"); gwListen == "" {
1891                 // dispatcher did not tell us which external IP
1892                 // address to advertise --> no gateway service
1893                 cr.CrunchLog.Printf("Not starting a gateway server (GatewayAddress was not provided by dispatcher)")
1894         } else if de, ok := cr.executor.(*dockerExecutor); ok {
1895                 cr.gateway = Gateway{
1896                         Address:            gwListen,
1897                         AuthSecret:         gwAuthSecret,
1898                         ContainerUUID:      containerUUID,
1899                         DockerContainerID:  &de.containerID,
1900                         Log:                cr.CrunchLog,
1901                         ContainerIPAddress: dockerContainerIPAddress(&de.containerID),
1902                 }
1903                 err = cr.gateway.Start()
1904                 if err != nil {
1905                         log.Printf("error starting gateway server: %s", err)
1906                         return 1
1907                 }
1908         }
1909
1910         parentTemp, tmperr := cr.MkTempDir("", "crunch-run."+containerUUID+".")
1911         if tmperr != nil {
1912                 log.Printf("%s: %v", containerUUID, tmperr)
1913                 return 1
1914         }
1915
1916         cr.parentTemp = parentTemp
1917         cr.statInterval = *statInterval
1918         cr.cgroupRoot = *cgroupRoot
1919         cr.expectCgroupParent = *cgroupParent
1920         cr.enableMemoryLimit = *enableMemoryLimit
1921         cr.enableNetwork = *enableNetwork
1922         cr.networkMode = *networkMode
1923         if *cgroupParentSubsystem != "" {
1924                 p := findCgroup(*cgroupParentSubsystem)
1925                 cr.setCgroupParent = p
1926                 cr.expectCgroupParent = p
1927         }
1928
1929         runerr := cr.Run()
1930
1931         if *memprofile != "" {
1932                 f, err := os.Create(*memprofile)
1933                 if err != nil {
1934                         log.Printf("could not create memory profile: %s", err)
1935                 }
1936                 runtime.GC() // get up-to-date statistics
1937                 if err := pprof.WriteHeapProfile(f); err != nil {
1938                         log.Printf("could not write memory profile: %s", err)
1939                 }
1940                 closeerr := f.Close()
1941                 if closeerr != nil {
1942                         log.Printf("closing memprofile file: %s", err)
1943                 }
1944         }
1945
1946         if runerr != nil {
1947                 log.Printf("%s: %v", containerUUID, runerr)
1948                 return 1
1949         }
1950         return 0
1951 }
1952
1953 func startLocalKeepstore(configData ConfigData, logbuf io.Writer) (*exec.Cmd, error) {
1954         if configData.Cluster == nil || configData.KeepBuffers < 1 {
1955                 return nil, nil
1956         }
1957         for uuid, vol := range configData.Cluster.Volumes {
1958                 if len(vol.AccessViaHosts) > 0 {
1959                         fmt.Fprintf(logbuf, "not starting a local keepstore process because a volume (%s) uses AccessViaHosts\n", uuid)
1960                         return nil, nil
1961                 }
1962                 if !vol.ReadOnly && vol.Replication < configData.Cluster.Collections.DefaultReplication {
1963                         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)
1964                         return nil, nil
1965                 }
1966         }
1967
1968         // Rather than have an alternate way to tell keepstore how
1969         // many buffers to use when starting it this way, we just
1970         // modify the cluster configuration that we feed it on stdin.
1971         configData.Cluster.API.MaxKeepBlobBuffers = configData.KeepBuffers
1972
1973         ln, err := net.Listen("tcp", "localhost:0")
1974         if err != nil {
1975                 return nil, err
1976         }
1977         _, port, err := net.SplitHostPort(ln.Addr().String())
1978         if err != nil {
1979                 ln.Close()
1980                 return nil, err
1981         }
1982         ln.Close()
1983         url := "http://localhost:" + port
1984
1985         fmt.Fprintf(logbuf, "starting keepstore on %s\n", url)
1986
1987         var confJSON bytes.Buffer
1988         err = json.NewEncoder(&confJSON).Encode(arvados.Config{
1989                 Clusters: map[string]arvados.Cluster{
1990                         configData.Cluster.ClusterID: *configData.Cluster,
1991                 },
1992         })
1993         if err != nil {
1994                 return nil, err
1995         }
1996         cmd := exec.Command("/proc/self/exe", "keepstore", "-config=-")
1997         if target, err := os.Readlink(cmd.Path); err == nil && strings.HasSuffix(target, ".test") {
1998                 // If we're a 'go test' process, running
1999                 // /proc/self/exe would start the test suite in a
2000                 // child process, which is not what we want.
2001                 cmd.Path, _ = exec.LookPath("go")
2002                 cmd.Args = append([]string{"go", "run", "../../cmd/arvados-server"}, cmd.Args[1:]...)
2003                 cmd.Env = os.Environ()
2004         }
2005         cmd.Stdin = &confJSON
2006         cmd.Stdout = logbuf
2007         cmd.Stderr = logbuf
2008         cmd.Env = append(cmd.Env,
2009                 "GOGC=10",
2010                 "ARVADOS_SERVICE_INTERNAL_URL="+url)
2011         err = cmd.Start()
2012         if err != nil {
2013                 return nil, fmt.Errorf("error starting keepstore process: %w", err)
2014         }
2015         cmdExited := false
2016         go func() {
2017                 cmd.Wait()
2018                 cmdExited = true
2019         }()
2020         ctx, cancel := context.WithDeadline(context.Background(), time.Now().Add(time.Second*10))
2021         defer cancel()
2022         poll := time.NewTicker(time.Second / 10)
2023         defer poll.Stop()
2024         client := http.Client{}
2025         for range poll.C {
2026                 testReq, err := http.NewRequestWithContext(ctx, "GET", url+"/_health/ping", nil)
2027                 testReq.Header.Set("Authorization", "Bearer "+configData.Cluster.ManagementToken)
2028                 if err != nil {
2029                         return nil, err
2030                 }
2031                 resp, err := client.Do(testReq)
2032                 if err == nil {
2033                         resp.Body.Close()
2034                         if resp.StatusCode == http.StatusOK {
2035                                 break
2036                         }
2037                 }
2038                 if cmdExited {
2039                         return nil, fmt.Errorf("keepstore child process exited")
2040                 }
2041                 if ctx.Err() != nil {
2042                         return nil, fmt.Errorf("timed out waiting for new keepstore process to report healthy")
2043                 }
2044         }
2045         os.Setenv("ARVADOS_KEEP_SERVICES", url)
2046         return cmd, nil
2047 }