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         "os/user"
  23         "path"
  24         "path/filepath"
  25         "regexp"
  26         "runtime"
  27         "runtime/pprof"
  28         "sort"
  29         "strings"
  30         "sync"
  31         "syscall"
  32         "time"
  33
  34         "git.arvados.org/arvados.git/lib/cmd"
  35         "git.arvados.org/arvados.git/lib/config"
  36         "git.arvados.org/arvados.git/lib/crunchstat"
  37         "git.arvados.org/arvados.git/sdk/go/arvados"
  38         "git.arvados.org/arvados.git/sdk/go/arvadosclient"
  39         "git.arvados.org/arvados.git/sdk/go/ctxlog"
  40         "git.arvados.org/arvados.git/sdk/go/keepclient"
  41         "git.arvados.org/arvados.git/sdk/go/manifest"
  42         "golang.org/x/sys/unix"
  43 )
  44
  45 type command struct{}
  46
  47 var Command = command{}
  48
  49 // ConfigData contains environment variables and (when needed) cluster
  50 // configuration, passed from dispatchcloud to crunch-run on stdin.
  51 type ConfigData struct {
  52         Env         map[string]string
  53         KeepBuffers int
  54         Cluster     *arvados.Cluster
  55 }
  56
  57 // IArvadosClient is the minimal Arvados API methods used by crunch-run.
  58 type IArvadosClient interface {
  59         Create(resourceType string, parameters arvadosclient.Dict, output interface{}) error
  60         Get(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  61         Update(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  62         Call(method, resourceType, uuid, action string, parameters arvadosclient.Dict, output interface{}) error
  63         CallRaw(method string, resourceType string, uuid string, action string, parameters arvadosclient.Dict) (reader io.ReadCloser, err error)
  64         Discovery(key string) (interface{}, error)
  65 }
  66
  67 // ErrCancelled is the error returned when the container is cancelled.
  68 var ErrCancelled = errors.New("Cancelled")
  69
  70 // IKeepClient is the minimal Keep API methods used by crunch-run.
  71 type IKeepClient interface {
  72         BlockWrite(context.Context, arvados.BlockWriteOptions) (arvados.BlockWriteResponse, error)
  73         ReadAt(locator string, p []byte, off int) (int, error)
  74         ManifestFileReader(m manifest.Manifest, filename string) (arvados.File, error)
  75         LocalLocator(locator string) (string, error)
  76         ClearBlockCache()
  77         SetStorageClasses(sc []string)
  78 }
  79
  80 // NewLogWriter is a factory function to create a new log writer.
  81 type NewLogWriter func(name string) (io.WriteCloser, error)
  82
  83 type RunArvMount func(cmdline []string, tok string) (*exec.Cmd, error)
  84
  85 type MkTempDir func(string, string) (string, error)
  86
  87 type PsProcess interface {
  88         CmdlineSlice() ([]string, error)
  89 }
  90
  91 // ContainerRunner is the main stateful struct used for a single execution of a
  92 // container.
  93 type ContainerRunner struct {
  94         executor       containerExecutor
  95         executorStdin  io.Closer
  96         executorStdout io.Closer
  97         executorStderr io.Closer
  98
  99         // Dispatcher client is initialized with the Dispatcher token.
 100         // This is a privileged token used to manage container status
 101         // and logs.
 102         //
 103         // We have both dispatcherClient and DispatcherArvClient
 104         // because there are two different incompatible Arvados Go
 105         // SDKs and we have to use both (hopefully this gets fixed in
 106         // #14467)
 107         dispatcherClient     *arvados.Client
 108         DispatcherArvClient  IArvadosClient
 109         DispatcherKeepClient IKeepClient
 110
 111         // Container client is initialized with the Container token
 112         // This token controls the permissions of the container, and
 113         // must be used for operations such as reading collections.
 114         //
 115         // Same comment as above applies to
 116         // containerClient/ContainerArvClient.
 117         containerClient     *arvados.Client
 118         ContainerArvClient  IArvadosClient
 119         ContainerKeepClient IKeepClient
 120
 121         Container     arvados.Container
 122         token         string
 123         ExitCode      *int
 124         NewLogWriter  NewLogWriter
 125         CrunchLog     *ThrottledLogger
 126         logUUID       string
 127         logMtx        sync.Mutex
 128         LogCollection arvados.CollectionFileSystem
 129         LogsPDH       *string
 130         RunArvMount   RunArvMount
 131         MkTempDir     MkTempDir
 132         ArvMount      *exec.Cmd
 133         ArvMountPoint string
 134         HostOutputDir string
 135         Volumes       map[string]struct{}
 136         OutputPDH     *string
 137         SigChan       chan os.Signal
 138         ArvMountExit  chan error
 139         SecretMounts  map[string]arvados.Mount
 140         MkArvClient   func(token string) (IArvadosClient, IKeepClient, *arvados.Client, error)
 141         finalState    string
 142         parentTemp    string
 143
 144         keepstoreLogger  io.WriteCloser
 145         keepstoreLogbuf  *bufThenWrite
 146         statLogger       io.WriteCloser
 147         statReporter     *crunchstat.Reporter
 148         hoststatLogger   io.WriteCloser
 149         hoststatReporter *crunchstat.Reporter
 150         statInterval     time.Duration
 151         cgroupRoot       string
 152         // What we expect the container's cgroup parent to be.
 153         expectCgroupParent string
 154         // What we tell docker to use as the container's cgroup
 155         // parent. Note: Ideally we would use the same field for both
 156         // expectCgroupParent and setCgroupParent, and just make it
 157         // default to "docker". However, when using docker < 1.10 with
 158         // systemd, specifying a non-empty cgroup parent (even the
 159         // default value "docker") hits a docker bug
 160         // (https://github.com/docker/docker/issues/17126). Using two
 161         // separate fields makes it possible to use the "expect cgroup
 162         // parent to be X" feature even on sites where the "specify
 163         // cgroup parent" feature breaks.
 164         setCgroupParent string
 165
 166         cStateLock sync.Mutex
 167         cCancelled bool // StopContainer() invoked
 168
 169         enableMemoryLimit bool
 170         enableNetwork     string // one of "default" or "always"
 171         networkMode       string // "none", "host", or "" -- passed through to executor
 172         brokenNodeHook    string // script to run if node appears to be broken
 173         arvMountLog       *ThrottledLogger
 174
 175         containerWatchdogInterval time.Duration
 176
 177         gateway Gateway
 178 }
 179
 180 // setupSignals sets up signal handling to gracefully terminate the
 181 // underlying container and update state when receiving a TERM, INT or
 182 // QUIT signal.
 183 func (runner *ContainerRunner) setupSignals() {
 184         runner.SigChan = make(chan os.Signal, 1)
 185         signal.Notify(runner.SigChan, syscall.SIGTERM)
 186         signal.Notify(runner.SigChan, syscall.SIGINT)
 187         signal.Notify(runner.SigChan, syscall.SIGQUIT)
 188
 189         go func(sig chan os.Signal) {
 190                 for s := range sig {
 191                         runner.stop(s)
 192                 }
 193         }(runner.SigChan)
 194 }
 195
 196 // stop the underlying container.
 197 func (runner *ContainerRunner) stop(sig os.Signal) {
 198         runner.cStateLock.Lock()
 199         defer runner.cStateLock.Unlock()
 200         if sig != nil {
 201                 runner.CrunchLog.Printf("caught signal: %v", sig)
 202         }
 203         runner.cCancelled = true
 204         runner.CrunchLog.Printf("stopping container")
 205         err := runner.executor.Stop()
 206         if err != nil {
 207                 runner.CrunchLog.Printf("error stopping container: %s", err)
 208         }
 209 }
 210
 211 var errorBlacklist = []string{
 212         "(?ms).*[Cc]annot connect to the Docker daemon.*",
 213         "(?ms).*oci runtime error.*starting container process.*container init.*mounting.*to rootfs.*no such file or directory.*",
 214         "(?ms).*grpc: the connection is unavailable.*",
 215 }
 216
 217 func (runner *ContainerRunner) runBrokenNodeHook() {
 218         if runner.brokenNodeHook == "" {
 219                 path := filepath.Join(lockdir, brokenfile)
 220                 runner.CrunchLog.Printf("Writing %s to mark node as broken", path)
 221                 f, err := os.OpenFile(path, os.O_CREATE|os.O_RDWR, 0700)
 222                 if err != nil {
 223                         runner.CrunchLog.Printf("Error writing %s: %s", path, err)
 224                         return
 225                 }
 226                 f.Close()
 227         } else {
 228                 runner.CrunchLog.Printf("Running broken node hook %q", runner.brokenNodeHook)
 229                 // run killme script
 230                 c := exec.Command(runner.brokenNodeHook)
 231                 c.Stdout = runner.CrunchLog
 232                 c.Stderr = runner.CrunchLog
 233                 err := c.Run()
 234                 if err != nil {
 235                         runner.CrunchLog.Printf("Error running broken node hook: %v", err)
 236                 }
 237         }
 238 }
 239
 240 func (runner *ContainerRunner) checkBrokenNode(goterr error) bool {
 241         for _, d := range errorBlacklist {
 242                 if m, e := regexp.MatchString(d, goterr.Error()); m && e == nil {
 243                         runner.CrunchLog.Printf("Error suggests node is unable to run containers: %v", goterr)
 244                         runner.runBrokenNodeHook()
 245                         return true
 246                 }
 247         }
 248         return false
 249 }
 250
 251 // LoadImage determines the docker image id from the container record and
 252 // checks if it is available in the local Docker image store.  If not, it loads
 253 // the image from Keep.
 254 func (runner *ContainerRunner) LoadImage() (string, error) {
 255         runner.CrunchLog.Printf("Fetching Docker image from collection '%s'", runner.Container.ContainerImage)
 256
 257         d, err := os.Open(runner.ArvMountPoint + "/by_id/" + runner.Container.ContainerImage)
 258         if err != nil {
 259                 return "", err
 260         }
 261         defer d.Close()
 262         allfiles, err := d.Readdirnames(-1)
 263         if err != nil {
 264                 return "", err
 265         }
 266         var tarfiles []string
 267         for _, fnm := range allfiles {
 268                 if strings.HasSuffix(fnm, ".tar") {
 269                         tarfiles = append(tarfiles, fnm)
 270                 }
 271         }
 272         if len(tarfiles) == 0 {
 273                 return "", fmt.Errorf("image collection does not include a .tar image file")
 274         }
 275         if len(tarfiles) > 1 {
 276                 return "", fmt.Errorf("cannot choose from multiple tar files in image collection: %v", tarfiles)
 277         }
 278         imageID := tarfiles[0][:len(tarfiles[0])-4]
 279         imageTarballPath := runner.ArvMountPoint + "/by_id/" + runner.Container.ContainerImage + "/" + imageID + ".tar"
 280         runner.CrunchLog.Printf("Using Docker image id %q", imageID)
 281
 282         runner.CrunchLog.Print("Loading Docker image from keep")
 283         err = runner.executor.LoadImage(imageID, imageTarballPath, runner.Container, runner.ArvMountPoint,
 284                 runner.containerClient)
 285         if err != nil {
 286                 return "", err
 287         }
 288
 289         return imageID, nil
 290 }
 291
 292 func (runner *ContainerRunner) ArvMountCmd(cmdline []string, token string) (c *exec.Cmd, err error) {
 293         c = exec.Command(cmdline[0], cmdline[1:]...)
 294
 295         // Copy our environment, but override ARVADOS_API_TOKEN with
 296         // the container auth token.
 297         c.Env = nil
 298         for _, s := range os.Environ() {
 299                 if !strings.HasPrefix(s, "ARVADOS_API_TOKEN=") {
 300                         c.Env = append(c.Env, s)
 301                 }
 302         }
 303         c.Env = append(c.Env, "ARVADOS_API_TOKEN="+token)
 304
 305         w, err := runner.NewLogWriter("arv-mount")
 306         if err != nil {
 307                 return nil, err
 308         }
 309         runner.arvMountLog = NewThrottledLogger(w)
 310         scanner := logScanner{
 311                 Patterns: []string{
 312                         "Keep write error",
 313                         "Block not found error",
 314                         "Unhandled exception during FUSE operation",
 315                 },
 316                 ReportFunc: runner.reportArvMountWarning,
 317         }
 318         c.Stdout = runner.arvMountLog
 319         c.Stderr = io.MultiWriter(runner.arvMountLog, os.Stderr, &scanner)
 320
 321         runner.CrunchLog.Printf("Running %v", c.Args)
 322
 323         err = c.Start()
 324         if err != nil {
 325                 return nil, err
 326         }
 327
 328         statReadme := make(chan bool)
 329         runner.ArvMountExit = make(chan error)
 330
 331         keepStatting := true
 332         go func() {
 333                 for keepStatting {
 334                         time.Sleep(100 * time.Millisecond)
 335                         _, err = os.Stat(fmt.Sprintf("%s/by_id/README", runner.ArvMountPoint))
 336                         if err == nil {
 337                                 keepStatting = false
 338                                 statReadme <- true
 339                         }
 340                 }
 341                 close(statReadme)
 342         }()
 343
 344         go func() {
 345                 mnterr := c.Wait()
 346                 if mnterr != nil {
 347                         runner.CrunchLog.Printf("Arv-mount exit error: %v", mnterr)
 348                 }
 349                 runner.ArvMountExit <- mnterr
 350                 close(runner.ArvMountExit)
 351         }()
 352
 353         select {
 354         case <-statReadme:
 355                 break
 356         case err := <-runner.ArvMountExit:
 357                 runner.ArvMount = nil
 358                 keepStatting = false
 359                 return nil, err
 360         }
 361
 362         return c, nil
 363 }
 364
 365 func (runner *ContainerRunner) SetupArvMountPoint(prefix string) (err error) {
 366         if runner.ArvMountPoint == "" {
 367                 runner.ArvMountPoint, err = runner.MkTempDir(runner.parentTemp, prefix)
 368         }
 369         return
 370 }
 371
 372 func copyfile(src string, dst string) (err error) {
 373         srcfile, err := os.Open(src)
 374         if err != nil {
 375                 return
 376         }
 377
 378         os.MkdirAll(path.Dir(dst), 0777)
 379
 380         dstfile, err := os.Create(dst)
 381         if err != nil {
 382                 return
 383         }
 384         _, err = io.Copy(dstfile, srcfile)
 385         if err != nil {
 386                 return
 387         }
 388
 389         err = srcfile.Close()
 390         err2 := dstfile.Close()
 391
 392         if err != nil {
 393                 return
 394         }
 395
 396         if err2 != nil {
 397                 return err2
 398         }
 399
 400         return nil
 401 }
 402
 403 func (runner *ContainerRunner) SetupMounts() (map[string]bindmount, error) {
 404         bindmounts := map[string]bindmount{}
 405         err := runner.SetupArvMountPoint("keep")
 406         if err != nil {
 407                 return nil, fmt.Errorf("While creating keep mount temp dir: %v", err)
 408         }
 409
 410         token, err := runner.ContainerToken()
 411         if err != nil {
 412                 return nil, fmt.Errorf("could not get container token: %s", err)
 413         }
 414         runner.CrunchLog.Printf("container token %q", token)
 415
 416         pdhOnly := true
 417         tmpcount := 0
 418         arvMountCmd := []string{
 419                 "arv-mount",
 420                 "--foreground",
 421                 "--read-write",
 422                 "--storage-classes", strings.Join(runner.Container.OutputStorageClasses, ","),
 423                 fmt.Sprintf("--crunchstat-interval=%v", runner.statInterval.Seconds())}
 424
 425         if _, isdocker := runner.executor.(*dockerExecutor); isdocker {
 426                 arvMountCmd = append(arvMountCmd, "--allow-other")
 427         }
 428
 429         if runner.Container.RuntimeConstraints.KeepCacheRAM > 0 {
 430                 arvMountCmd = append(arvMountCmd, "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheRAM))
 431         }
 432
 433         collectionPaths := []string{}
 434         needCertMount := true
 435         type copyFile struct {
 436                 src  string
 437                 bind string
 438         }
 439         var copyFiles []copyFile
 440
 441         var binds []string
 442         for bind := range runner.Container.Mounts {
 443                 binds = append(binds, bind)
 444         }
 445         for bind := range runner.SecretMounts {
 446                 if _, ok := runner.Container.Mounts[bind]; ok {
 447                         return nil, fmt.Errorf("secret mount %q conflicts with regular mount", bind)
 448                 }
 449                 if runner.SecretMounts[bind].Kind != "json" &&
 450                         runner.SecretMounts[bind].Kind != "text" {
 451                         return nil, fmt.Errorf("secret mount %q type is %q but only 'json' and 'text' are permitted",
 452                                 bind, runner.SecretMounts[bind].Kind)
 453                 }
 454                 binds = append(binds, bind)
 455         }
 456         sort.Strings(binds)
 457
 458         for _, bind := range binds {
 459                 mnt, notSecret := runner.Container.Mounts[bind]
 460                 if !notSecret {
 461                         mnt = runner.SecretMounts[bind]
 462                 }
 463                 if bind == "stdout" || bind == "stderr" {
 464                         // Is it a "file" mount kind?
 465                         if mnt.Kind != "file" {
 466                                 return nil, fmt.Errorf("unsupported mount kind '%s' for %s: only 'file' is supported", mnt.Kind, bind)
 467                         }
 468
 469                         // Does path start with OutputPath?
 470                         prefix := runner.Container.OutputPath
 471                         if !strings.HasSuffix(prefix, "/") {
 472                                 prefix += "/"
 473                         }
 474                         if !strings.HasPrefix(mnt.Path, prefix) {
 475                                 return nil, fmt.Errorf("%s path does not start with OutputPath: %s, %s", strings.Title(bind), mnt.Path, prefix)
 476                         }
 477                 }
 478
 479                 if bind == "stdin" {
 480                         // Is it a "collection" mount kind?
 481                         if mnt.Kind != "collection" && mnt.Kind != "json" {
 482                                 return nil, fmt.Errorf("unsupported mount kind '%s' for stdin: only 'collection' and 'json' are supported", mnt.Kind)
 483                         }
 484                 }
 485
 486                 if bind == "/etc/arvados/ca-certificates.crt" {
 487                         needCertMount = false
 488                 }
 489
 490                 if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && bind != runner.Container.OutputPath+"/" {
 491                         if mnt.Kind != "collection" && mnt.Kind != "text" && mnt.Kind != "json" {
 492                                 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)
 493                         }
 494                 }
 495
 496                 switch {
 497                 case mnt.Kind == "collection" && bind != "stdin":
 498                         var src string
 499                         if mnt.UUID != "" && mnt.PortableDataHash != "" {
 500                                 return nil, fmt.Errorf("cannot specify both 'uuid' and 'portable_data_hash' for a collection mount")
 501                         }
 502                         if mnt.UUID != "" {
 503                                 if mnt.Writable {
 504                                         return nil, fmt.Errorf("writing to existing collections currently not permitted")
 505                                 }
 506                                 pdhOnly = false
 507                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.UUID)
 508                         } else if mnt.PortableDataHash != "" {
 509                                 if mnt.Writable && !strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 510                                         return nil, fmt.Errorf("can never write to a collection specified by portable data hash")
 511                                 }
 512                                 idx := strings.Index(mnt.PortableDataHash, "/")
 513                                 if idx > 0 {
 514                                         mnt.Path = path.Clean(mnt.PortableDataHash[idx:])
 515                                         mnt.PortableDataHash = mnt.PortableDataHash[0:idx]
 516                                         runner.Container.Mounts[bind] = mnt
 517                                 }
 518                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.PortableDataHash)
 519                                 if mnt.Path != "" && mnt.Path != "." {
 520                                         if strings.HasPrefix(mnt.Path, "./") {
 521                                                 mnt.Path = mnt.Path[2:]
 522                                         } else if strings.HasPrefix(mnt.Path, "/") {
 523                                                 mnt.Path = mnt.Path[1:]
 524                                         }
 525                                         src += "/" + mnt.Path
 526                                 }
 527                         } else {
 528                                 src = fmt.Sprintf("%s/tmp%d", runner.ArvMountPoint, tmpcount)
 529                                 arvMountCmd = append(arvMountCmd, "--mount-tmp", fmt.Sprintf("tmp%d", tmpcount))
 530                                 tmpcount++
 531                         }
 532                         if mnt.Writable {
 533                                 if bind == runner.Container.OutputPath {
 534                                         runner.HostOutputDir = src
 535                                         bindmounts[bind] = bindmount{HostPath: src}
 536                                 } else if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 537                                         copyFiles = append(copyFiles, copyFile{src, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
 538                                 } else {
 539                                         bindmounts[bind] = bindmount{HostPath: src}
 540                                 }
 541                         } else {
 542                                 bindmounts[bind] = bindmount{HostPath: src, ReadOnly: true}
 543                         }
 544                         collectionPaths = append(collectionPaths, src)
 545
 546                 case mnt.Kind == "tmp":
 547                         var tmpdir string
 548                         tmpdir, err = runner.MkTempDir(runner.parentTemp, "tmp")
 549                         if err != nil {
 550                                 return nil, fmt.Errorf("while creating mount temp dir: %v", err)
 551                         }
 552                         st, staterr := os.Stat(tmpdir)
 553                         if staterr != nil {
 554                                 return nil, fmt.Errorf("while Stat on temp dir: %v", staterr)
 555                         }
 556                         err = os.Chmod(tmpdir, st.Mode()|os.ModeSetgid|0777)
 557                         if staterr != nil {
 558                                 return nil, fmt.Errorf("while Chmod temp dir: %v", err)
 559                         }
 560                         bindmounts[bind] = bindmount{HostPath: tmpdir}
 561                         if bind == runner.Container.OutputPath {
 562                                 runner.HostOutputDir = tmpdir
 563                         }
 564
 565                 case mnt.Kind == "json" || mnt.Kind == "text":
 566                         var filedata []byte
 567                         if mnt.Kind == "json" {
 568                                 filedata, err = json.Marshal(mnt.Content)
 569                                 if err != nil {
 570                                         return nil, fmt.Errorf("encoding json data: %v", err)
 571                                 }
 572                         } else {
 573                                 text, ok := mnt.Content.(string)
 574                                 if !ok {
 575                                         return nil, fmt.Errorf("content for mount %q must be a string", bind)
 576                                 }
 577                                 filedata = []byte(text)
 578                         }
 579
 580                         tmpdir, err := runner.MkTempDir(runner.parentTemp, mnt.Kind)
 581                         if err != nil {
 582                                 return nil, fmt.Errorf("creating temp dir: %v", err)
 583                         }
 584                         tmpfn := filepath.Join(tmpdir, "mountdata."+mnt.Kind)
 585                         err = ioutil.WriteFile(tmpfn, filedata, 0444)
 586                         if err != nil {
 587                                 return nil, fmt.Errorf("writing temp file: %v", err)
 588                         }
 589                         if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && (notSecret || runner.Container.Mounts[runner.Container.OutputPath].Kind != "collection") {
 590                                 // In most cases, if the container
 591                                 // specifies a literal file inside the
 592                                 // output path, we copy it into the
 593                                 // output directory (either a mounted
 594                                 // collection or a staging area on the
 595                                 // host fs). If it's a secret, it will
 596                                 // be skipped when copying output from
 597                                 // staging to Keep later.
 598                                 copyFiles = append(copyFiles, copyFile{tmpfn, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
 599                         } else {
 600                                 // If a secret is outside OutputPath,
 601                                 // we bind mount the secret file
 602                                 // directly just like other mounts. We
 603                                 // also use this strategy when a
 604                                 // secret is inside OutputPath but
 605                                 // OutputPath is a live collection, to
 606                                 // avoid writing the secret to
 607                                 // Keep. Attempting to remove a
 608                                 // bind-mounted secret file from
 609                                 // inside the container will return a
 610                                 // "Device or resource busy" error
 611                                 // that might not be handled well by
 612                                 // the container, which is why we
 613                                 // don't use this strategy when
 614                                 // OutputPath is a staging directory.
 615                                 bindmounts[bind] = bindmount{HostPath: tmpfn, ReadOnly: true}
 616                         }
 617
 618                 case mnt.Kind == "git_tree":
 619                         tmpdir, err := runner.MkTempDir(runner.parentTemp, "git_tree")
 620                         if err != nil {
 621                                 return nil, fmt.Errorf("creating temp dir: %v", err)
 622                         }
 623                         err = gitMount(mnt).extractTree(runner.ContainerArvClient, tmpdir, token)
 624                         if err != nil {
 625                                 return nil, err
 626                         }
 627                         bindmounts[bind] = bindmount{HostPath: tmpdir, ReadOnly: true}
 628                 }
 629         }
 630
 631         if runner.HostOutputDir == "" {
 632                 return nil, fmt.Errorf("output path does not correspond to a writable mount point")
 633         }
 634
 635         if needCertMount && runner.Container.RuntimeConstraints.API {
 636                 for _, certfile := range arvadosclient.CertFiles {
 637                         _, err := os.Stat(certfile)
 638                         if err == nil {
 639                                 bindmounts["/etc/arvados/ca-certificates.crt"] = bindmount{HostPath: certfile, ReadOnly: true}
 640                                 break
 641                         }
 642                 }
 643         }
 644
 645         if pdhOnly {
 646                 // If we are only mounting collections by pdh, make
 647                 // sure we don't subscribe to websocket events to
 648                 // avoid putting undesired load on the API server
 649                 arvMountCmd = append(arvMountCmd, "--mount-by-pdh", "by_id", "--disable-event-listening")
 650         } else {
 651                 arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_id")
 652         }
 653         // the by_uuid mount point is used by singularity when writing
 654         // out docker images converted to SIF
 655         arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_uuid")
 656         arvMountCmd = append(arvMountCmd, runner.ArvMountPoint)
 657
 658         runner.ArvMount, err = runner.RunArvMount(arvMountCmd, token)
 659         if err != nil {
 660                 return nil, fmt.Errorf("while trying to start arv-mount: %v", err)
 661         }
 662
 663         for _, p := range collectionPaths {
 664                 _, err = os.Stat(p)
 665                 if err != nil {
 666                         return nil, fmt.Errorf("while checking that input files exist: %v", err)
 667                 }
 668         }
 669
 670         for _, cp := range copyFiles {
 671                 st, err := os.Stat(cp.src)
 672                 if err != nil {
 673                         return nil, fmt.Errorf("while staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 674                 }
 675                 if st.IsDir() {
 676                         err = filepath.Walk(cp.src, func(walkpath string, walkinfo os.FileInfo, walkerr error) error {
 677                                 if walkerr != nil {
 678                                         return walkerr
 679                                 }
 680                                 target := path.Join(cp.bind, walkpath[len(cp.src):])
 681                                 if walkinfo.Mode().IsRegular() {
 682                                         copyerr := copyfile(walkpath, target)
 683                                         if copyerr != nil {
 684                                                 return copyerr
 685                                         }
 686                                         return os.Chmod(target, walkinfo.Mode()|0777)
 687                                 } else if walkinfo.Mode().IsDir() {
 688                                         mkerr := os.MkdirAll(target, 0777)
 689                                         if mkerr != nil {
 690                                                 return mkerr
 691                                         }
 692                                         return os.Chmod(target, walkinfo.Mode()|os.ModeSetgid|0777)
 693                                 } else {
 694                                         return fmt.Errorf("source %q is not a regular file or directory", cp.src)
 695                                 }
 696                         })
 697                 } else if st.Mode().IsRegular() {
 698                         err = copyfile(cp.src, cp.bind)
 699                         if err == nil {
 700                                 err = os.Chmod(cp.bind, st.Mode()|0777)
 701                         }
 702                 }
 703                 if err != nil {
 704                         return nil, fmt.Errorf("while staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 705                 }
 706         }
 707
 708         return bindmounts, nil
 709 }
 710
 711 func (runner *ContainerRunner) stopHoststat() error {
 712         if runner.hoststatReporter == nil {
 713                 return nil
 714         }
 715         runner.hoststatReporter.Stop()
 716         err := runner.hoststatLogger.Close()
 717         if err != nil {
 718                 return fmt.Errorf("error closing hoststat logs: %v", err)
 719         }
 720         return nil
 721 }
 722
 723 func (runner *ContainerRunner) startHoststat() error {
 724         w, err := runner.NewLogWriter("hoststat")
 725         if err != nil {
 726                 return err
 727         }
 728         runner.hoststatLogger = NewThrottledLogger(w)
 729         runner.hoststatReporter = &crunchstat.Reporter{
 730                 Logger:     log.New(runner.hoststatLogger, "", 0),
 731                 CgroupRoot: runner.cgroupRoot,
 732                 PollPeriod: runner.statInterval,
 733         }
 734         runner.hoststatReporter.Start()
 735         return nil
 736 }
 737
 738 func (runner *ContainerRunner) startCrunchstat() error {
 739         w, err := runner.NewLogWriter("crunchstat")
 740         if err != nil {
 741                 return err
 742         }
 743         runner.statLogger = NewThrottledLogger(w)
 744         runner.statReporter = &crunchstat.Reporter{
 745                 CID:          runner.executor.CgroupID(),
 746                 Logger:       log.New(runner.statLogger, "", 0),
 747                 CgroupParent: runner.expectCgroupParent,
 748                 CgroupRoot:   runner.cgroupRoot,
 749                 PollPeriod:   runner.statInterval,
 750                 TempDir:      runner.parentTemp,
 751         }
 752         runner.statReporter.Start()
 753         return nil
 754 }
 755
 756 type infoCommand struct {
 757         label string
 758         cmd   []string
 759 }
 760
 761 // LogHostInfo logs info about the current host, for debugging and
 762 // accounting purposes. Although it's logged as "node-info", this is
 763 // about the environment where crunch-run is actually running, which
 764 // might differ from what's described in the node record (see
 765 // LogNodeRecord).
 766 func (runner *ContainerRunner) LogHostInfo() (err error) {
 767         w, err := runner.NewLogWriter("node-info")
 768         if err != nil {
 769                 return
 770         }
 771
 772         commands := []infoCommand{
 773                 {
 774                         label: "Host Information",
 775                         cmd:   []string{"uname", "-a"},
 776                 },
 777                 {
 778                         label: "CPU Information",
 779                         cmd:   []string{"cat", "/proc/cpuinfo"},
 780                 },
 781                 {
 782                         label: "Memory Information",
 783                         cmd:   []string{"cat", "/proc/meminfo"},
 784                 },
 785                 {
 786                         label: "Disk Space",
 787                         cmd:   []string{"df", "-m", "/", os.TempDir()},
 788                 },
 789                 {
 790                         label: "Disk INodes",
 791                         cmd:   []string{"df", "-i", "/", os.TempDir()},
 792                 },
 793         }
 794
 795         // Run commands with informational output to be logged.
 796         for _, command := range commands {
 797                 fmt.Fprintln(w, command.label)
 798                 cmd := exec.Command(command.cmd[0], command.cmd[1:]...)
 799                 cmd.Stdout = w
 800                 cmd.Stderr = w
 801                 if err := cmd.Run(); err != nil {
 802                         err = fmt.Errorf("While running command %q: %v", command.cmd, err)
 803                         fmt.Fprintln(w, err)
 804                         return err
 805                 }
 806                 fmt.Fprintln(w, "")
 807         }
 808
 809         err = w.Close()
 810         if err != nil {
 811                 return fmt.Errorf("While closing node-info logs: %v", err)
 812         }
 813         return nil
 814 }
 815
 816 // LogContainerRecord gets and saves the raw JSON container record from the API server
 817 func (runner *ContainerRunner) LogContainerRecord() error {
 818         logged, err := runner.logAPIResponse("container", "containers", map[string]interface{}{"filters": [][]string{{"uuid", "=", runner.Container.UUID}}}, nil)
 819         if !logged && err == nil {
 820                 err = fmt.Errorf("error: no container record found for %s", runner.Container.UUID)
 821         }
 822         return err
 823 }
 824
 825 // LogNodeRecord logs the current host's InstanceType config entry (or
 826 // the arvados#node record, if running via crunch-dispatch-slurm).
 827 func (runner *ContainerRunner) LogNodeRecord() error {
 828         if it := os.Getenv("InstanceType"); it != "" {
 829                 // Dispatched via arvados-dispatch-cloud. Save
 830                 // InstanceType config fragment received from
 831                 // dispatcher on stdin.
 832                 w, err := runner.LogCollection.OpenFile("node.json", os.O_CREATE|os.O_WRONLY, 0666)
 833                 if err != nil {
 834                         return err
 835                 }
 836                 defer w.Close()
 837                 _, err = io.WriteString(w, it)
 838                 if err != nil {
 839                         return err
 840                 }
 841                 return w.Close()
 842         }
 843         // Dispatched via crunch-dispatch-slurm. Look up
 844         // apiserver's node record corresponding to
 845         // $SLURMD_NODENAME.
 846         hostname := os.Getenv("SLURMD_NODENAME")
 847         if hostname == "" {
 848                 hostname, _ = os.Hostname()
 849         }
 850         _, err := runner.logAPIResponse("node", "nodes", map[string]interface{}{"filters": [][]string{{"hostname", "=", hostname}}}, func(resp interface{}) {
 851                 // The "info" field has admin-only info when
 852                 // obtained with a privileged token, and
 853                 // should not be logged.
 854                 node, ok := resp.(map[string]interface{})
 855                 if ok {
 856                         delete(node, "info")
 857                 }
 858         })
 859         return err
 860 }
 861
 862 func (runner *ContainerRunner) logAPIResponse(label, path string, params map[string]interface{}, munge func(interface{})) (logged bool, err error) {
 863         writer, err := runner.LogCollection.OpenFile(label+".json", os.O_CREATE|os.O_WRONLY, 0666)
 864         if err != nil {
 865                 return false, err
 866         }
 867         w := &ArvLogWriter{
 868                 ArvClient:     runner.DispatcherArvClient,
 869                 UUID:          runner.Container.UUID,
 870                 loggingStream: label,
 871                 writeCloser:   writer,
 872         }
 873
 874         reader, err := runner.DispatcherArvClient.CallRaw("GET", path, "", "", arvadosclient.Dict(params))
 875         if err != nil {
 876                 return false, fmt.Errorf("error getting %s record: %v", label, err)
 877         }
 878         defer reader.Close()
 879
 880         dec := json.NewDecoder(reader)
 881         dec.UseNumber()
 882         var resp map[string]interface{}
 883         if err = dec.Decode(&resp); err != nil {
 884                 return false, fmt.Errorf("error decoding %s list response: %v", label, err)
 885         }
 886         items, ok := resp["items"].([]interface{})
 887         if !ok {
 888                 return false, fmt.Errorf("error decoding %s list response: no \"items\" key in API list response", label)
 889         } else if len(items) < 1 {
 890                 return false, nil
 891         }
 892         if munge != nil {
 893                 munge(items[0])
 894         }
 895         // Re-encode it using indentation to improve readability
 896         enc := json.NewEncoder(w)
 897         enc.SetIndent("", "    ")
 898         if err = enc.Encode(items[0]); err != nil {
 899                 return false, fmt.Errorf("error logging %s record: %v", label, err)
 900         }
 901         err = w.Close()
 902         if err != nil {
 903                 return false, fmt.Errorf("error closing %s.json in log collection: %v", label, err)
 904         }
 905         return true, nil
 906 }
 907
 908 func (runner *ContainerRunner) getStdoutFile(mntPath string) (*os.File, error) {
 909         stdoutPath := mntPath[len(runner.Container.OutputPath):]
 910         index := strings.LastIndex(stdoutPath, "/")
 911         if index > 0 {
 912                 subdirs := stdoutPath[:index]
 913                 if subdirs != "" {
 914                         st, err := os.Stat(runner.HostOutputDir)
 915                         if err != nil {
 916                                 return nil, fmt.Errorf("While Stat on temp dir: %v", err)
 917                         }
 918                         stdoutPath := filepath.Join(runner.HostOutputDir, subdirs)
 919                         err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777)
 920                         if err != nil {
 921                                 return nil, fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err)
 922                         }
 923                 }
 924         }
 925         stdoutFile, err := os.Create(filepath.Join(runner.HostOutputDir, stdoutPath))
 926         if err != nil {
 927                 return nil, fmt.Errorf("While creating file %q: %v", stdoutPath, err)
 928         }
 929
 930         return stdoutFile, nil
 931 }
 932
 933 // CreateContainer creates the docker container.
 934 func (runner *ContainerRunner) CreateContainer(imageID string, bindmounts map[string]bindmount) error {
 935         var stdin io.ReadCloser = ioutil.NopCloser(bytes.NewReader(nil))
 936         if mnt, ok := runner.Container.Mounts["stdin"]; ok {
 937                 switch mnt.Kind {
 938                 case "collection":
 939                         var collID string
 940                         if mnt.UUID != "" {
 941                                 collID = mnt.UUID
 942                         } else {
 943                                 collID = mnt.PortableDataHash
 944                         }
 945                         path := runner.ArvMountPoint + "/by_id/" + collID + "/" + mnt.Path
 946                         f, err := os.Open(path)
 947                         if err != nil {
 948                                 return err
 949                         }
 950                         stdin = f
 951                 case "json":
 952                         j, err := json.Marshal(mnt.Content)
 953                         if err != nil {
 954                                 return fmt.Errorf("error encoding stdin json data: %v", err)
 955                         }
 956                         stdin = ioutil.NopCloser(bytes.NewReader(j))
 957                 default:
 958                         return fmt.Errorf("stdin mount has unsupported kind %q", mnt.Kind)
 959                 }
 960         }
 961
 962         var stdout, stderr io.WriteCloser
 963         if mnt, ok := runner.Container.Mounts["stdout"]; ok {
 964                 f, err := runner.getStdoutFile(mnt.Path)
 965                 if err != nil {
 966                         return err
 967                 }
 968                 stdout = f
 969         } else if w, err := runner.NewLogWriter("stdout"); err != nil {
 970                 return err
 971         } else {
 972                 stdout = NewThrottledLogger(w)
 973         }
 974
 975         if mnt, ok := runner.Container.Mounts["stderr"]; ok {
 976                 f, err := runner.getStdoutFile(mnt.Path)
 977                 if err != nil {
 978                         return err
 979                 }
 980                 stderr = f
 981         } else if w, err := runner.NewLogWriter("stderr"); err != nil {
 982                 return err
 983         } else {
 984                 stderr = NewThrottledLogger(w)
 985         }
 986
 987         env := runner.Container.Environment
 988         enableNetwork := runner.enableNetwork == "always"
 989         if runner.Container.RuntimeConstraints.API {
 990                 enableNetwork = true
 991                 tok, err := runner.ContainerToken()
 992                 if err != nil {
 993                         return err
 994                 }
 995                 env = map[string]string{}
 996                 for k, v := range runner.Container.Environment {
 997                         env[k] = v
 998                 }
 999                 env["ARVADOS_API_TOKEN"] = tok
1000                 env["ARVADOS_API_HOST"] = os.Getenv("ARVADOS_API_HOST")
1001                 env["ARVADOS_API_HOST_INSECURE"] = os.Getenv("ARVADOS_API_HOST_INSECURE")
1002         }
1003         workdir := runner.Container.Cwd
1004         if workdir == "." {
1005                 // both "" and "." mean default
1006                 workdir = ""
1007         }
1008         ram := runner.Container.RuntimeConstraints.RAM
1009         if !runner.enableMemoryLimit {
1010                 ram = 0
1011         }
1012         runner.executorStdin = stdin
1013         runner.executorStdout = stdout
1014         runner.executorStderr = stderr
1015
1016         if runner.Container.RuntimeConstraints.CUDA.DeviceCount > 0 {
1017                 nvidiaModprobe(runner.CrunchLog)
1018         }
1019
1020         return runner.executor.Create(containerSpec{
1021                 Image:           imageID,
1022                 VCPUs:           runner.Container.RuntimeConstraints.VCPUs,
1023                 RAM:             ram,
1024                 WorkingDir:      workdir,
1025                 Env:             env,
1026                 BindMounts:      bindmounts,
1027                 Command:         runner.Container.Command,
1028                 EnableNetwork:   enableNetwork,
1029                 CUDADeviceCount: runner.Container.RuntimeConstraints.CUDA.DeviceCount,
1030                 NetworkMode:     runner.networkMode,
1031                 CgroupParent:    runner.setCgroupParent,
1032                 Stdin:           stdin,
1033                 Stdout:          stdout,
1034                 Stderr:          stderr,
1035         })
1036 }
1037
1038 // StartContainer starts the docker container created by CreateContainer.
1039 func (runner *ContainerRunner) StartContainer() error {
1040         runner.CrunchLog.Printf("Starting container")
1041         runner.cStateLock.Lock()
1042         defer runner.cStateLock.Unlock()
1043         if runner.cCancelled {
1044                 return ErrCancelled
1045         }
1046         err := runner.executor.Start()
1047         if err != nil {
1048                 var advice string
1049                 if m, e := regexp.MatchString("(?ms).*(exec|System error).*(no such file or directory|file not found).*", err.Error()); m && e == nil {
1050                         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])
1051                 }
1052                 return fmt.Errorf("could not start container: %v%s", err, advice)
1053         }
1054         return nil
1055 }
1056
1057 // WaitFinish waits for the container to terminate, capture the exit code, and
1058 // close the stdout/stderr logging.
1059 func (runner *ContainerRunner) WaitFinish() error {
1060         runner.CrunchLog.Print("Waiting for container to finish")
1061         var timeout <-chan time.Time
1062         if s := runner.Container.SchedulingParameters.MaxRunTime; s > 0 {
1063                 timeout = time.After(time.Duration(s) * time.Second)
1064         }
1065         ctx, cancel := context.WithCancel(context.Background())
1066         defer cancel()
1067         go func() {
1068                 select {
1069                 case <-timeout:
1070                         runner.CrunchLog.Printf("maximum run time exceeded. Stopping container.")
1071                         runner.stop(nil)
1072                 case <-runner.ArvMountExit:
1073                         runner.CrunchLog.Printf("arv-mount exited while container is still running. Stopping container.")
1074                         runner.stop(nil)
1075                 case <-ctx.Done():
1076                 }
1077         }()
1078         exitcode, err := runner.executor.Wait(ctx)
1079         if err != nil {
1080                 runner.checkBrokenNode(err)
1081                 return err
1082         }
1083         runner.ExitCode = &exitcode
1084
1085         extra := ""
1086         if exitcode&0x80 != 0 {
1087                 // Convert raw exit status (0x80 + signal number) to a
1088                 // string to log after the code, like " (signal 101)"
1089                 // or " (signal 9, killed)"
1090                 sig := syscall.WaitStatus(exitcode).Signal()
1091                 if name := unix.SignalName(sig); name != "" {
1092                         extra = fmt.Sprintf(" (signal %d, %s)", sig, name)
1093                 } else {
1094                         extra = fmt.Sprintf(" (signal %d)", sig)
1095                 }
1096         }
1097         runner.CrunchLog.Printf("Container exited with status code %d%s", exitcode, extra)
1098
1099         var returnErr error
1100         if err = runner.executorStdin.Close(); err != nil {
1101                 err = fmt.Errorf("error closing container stdin: %s", err)
1102                 runner.CrunchLog.Printf("%s", err)
1103                 returnErr = err
1104         }
1105         if err = runner.executorStdout.Close(); err != nil {
1106                 err = fmt.Errorf("error closing container stdout: %s", err)
1107                 runner.CrunchLog.Printf("%s", err)
1108                 if returnErr == nil {
1109                         returnErr = err
1110                 }
1111         }
1112         if err = runner.executorStderr.Close(); err != nil {
1113                 err = fmt.Errorf("error closing container stderr: %s", err)
1114                 runner.CrunchLog.Printf("%s", err)
1115                 if returnErr == nil {
1116                         returnErr = err
1117                 }
1118         }
1119
1120         if runner.statReporter != nil {
1121                 runner.statReporter.Stop()
1122                 err = runner.statLogger.Close()
1123                 if err != nil {
1124                         runner.CrunchLog.Printf("error closing crunchstat logs: %v", err)
1125                 }
1126         }
1127         return returnErr
1128 }
1129
1130 func (runner *ContainerRunner) updateLogs() {
1131         ticker := time.NewTicker(crunchLogUpdatePeriod / 360)
1132         defer ticker.Stop()
1133
1134         sigusr1 := make(chan os.Signal, 1)
1135         signal.Notify(sigusr1, syscall.SIGUSR1)
1136         defer signal.Stop(sigusr1)
1137
1138         saveAtTime := time.Now().Add(crunchLogUpdatePeriod)
1139         saveAtSize := crunchLogUpdateSize
1140         var savedSize int64
1141         for {
1142                 select {
1143                 case <-ticker.C:
1144                 case <-sigusr1:
1145                         saveAtTime = time.Now()
1146                 }
1147                 runner.logMtx.Lock()
1148                 done := runner.LogsPDH != nil
1149                 runner.logMtx.Unlock()
1150                 if done {
1151                         return
1152                 }
1153                 size := runner.LogCollection.Size()
1154                 if size == savedSize || (time.Now().Before(saveAtTime) && size < saveAtSize) {
1155                         continue
1156                 }
1157                 saveAtTime = time.Now().Add(crunchLogUpdatePeriod)
1158                 saveAtSize = runner.LogCollection.Size() + crunchLogUpdateSize
1159                 saved, err := runner.saveLogCollection(false)
1160                 if err != nil {
1161                         runner.CrunchLog.Printf("error updating log collection: %s", err)
1162                         continue
1163                 }
1164
1165                 var updated arvados.Container
1166                 err = runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1167                         "container": arvadosclient.Dict{"log": saved.PortableDataHash},
1168                 }, &updated)
1169                 if err != nil {
1170                         runner.CrunchLog.Printf("error updating container log to %s: %s", saved.PortableDataHash, err)
1171                         continue
1172                 }
1173
1174                 savedSize = size
1175         }
1176 }
1177
1178 func (runner *ContainerRunner) reportArvMountWarning(pattern, text string) {
1179         var updated arvados.Container
1180         err := runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1181                 "container": arvadosclient.Dict{
1182                         "runtime_status": arvadosclient.Dict{
1183                                 "warning":       "arv-mount: " + pattern,
1184                                 "warningDetail": text,
1185                         },
1186                 },
1187         }, &updated)
1188         if err != nil {
1189                 runner.CrunchLog.Printf("error updating container runtime_status: %s", err)
1190         }
1191 }
1192
1193 // CaptureOutput saves data from the container's output directory if
1194 // needed, and updates the container output accordingly.
1195 func (runner *ContainerRunner) CaptureOutput(bindmounts map[string]bindmount) error {
1196         if runner.Container.RuntimeConstraints.API {
1197                 // Output may have been set directly by the container, so
1198                 // refresh the container record to check.
1199                 err := runner.DispatcherArvClient.Get("containers", runner.Container.UUID,
1200                         nil, &runner.Container)
1201                 if err != nil {
1202                         return err
1203                 }
1204                 if runner.Container.Output != "" {
1205                         // Container output is already set.
1206                         runner.OutputPDH = &runner.Container.Output
1207                         return nil
1208                 }
1209         }
1210
1211         txt, err := (&copier{
1212                 client:        runner.containerClient,
1213                 arvClient:     runner.ContainerArvClient,
1214                 keepClient:    runner.ContainerKeepClient,
1215                 hostOutputDir: runner.HostOutputDir,
1216                 ctrOutputDir:  runner.Container.OutputPath,
1217                 bindmounts:    bindmounts,
1218                 mounts:        runner.Container.Mounts,
1219                 secretMounts:  runner.SecretMounts,
1220                 logger:        runner.CrunchLog,
1221         }).Copy()
1222         if err != nil {
1223                 return err
1224         }
1225         if n := len(regexp.MustCompile(` [0-9a-f]+\+\S*\+R`).FindAllStringIndex(txt, -1)); n > 0 {
1226                 runner.CrunchLog.Printf("Copying %d data blocks from remote input collections...", n)
1227                 fs, err := (&arvados.Collection{ManifestText: txt}).FileSystem(runner.containerClient, runner.ContainerKeepClient)
1228                 if err != nil {
1229                         return err
1230                 }
1231                 txt, err = fs.MarshalManifest(".")
1232                 if err != nil {
1233                         return err
1234                 }
1235         }
1236         var resp arvados.Collection
1237         err = runner.ContainerArvClient.Create("collections", arvadosclient.Dict{
1238                 "ensure_unique_name": true,
1239                 "collection": arvadosclient.Dict{
1240                         "is_trashed":    true,
1241                         "name":          "output for " + runner.Container.UUID,
1242                         "manifest_text": txt,
1243                 },
1244         }, &resp)
1245         if err != nil {
1246                 return fmt.Errorf("error creating output collection: %v", err)
1247         }
1248         runner.OutputPDH = &resp.PortableDataHash
1249         return nil
1250 }
1251
1252 func (runner *ContainerRunner) CleanupDirs() {
1253         if runner.ArvMount != nil {
1254                 var delay int64 = 8
1255                 umount := exec.Command("arv-mount", fmt.Sprintf("--unmount-timeout=%d", delay), "--unmount", runner.ArvMountPoint)
1256                 umount.Stdout = runner.CrunchLog
1257                 umount.Stderr = runner.CrunchLog
1258                 runner.CrunchLog.Printf("Running %v", umount.Args)
1259                 umnterr := umount.Start()
1260
1261                 if umnterr != nil {
1262                         runner.CrunchLog.Printf("Error unmounting: %v", umnterr)
1263                         runner.ArvMount.Process.Kill()
1264                 } else {
1265                         // If arv-mount --unmount gets stuck for any reason, we
1266                         // don't want to wait for it forever.  Do Wait() in a goroutine
1267                         // so it doesn't block crunch-run.
1268                         umountExit := make(chan error)
1269                         go func() {
1270                                 mnterr := umount.Wait()
1271                                 if mnterr != nil {
1272                                         runner.CrunchLog.Printf("Error unmounting: %v", mnterr)
1273                                 }
1274                                 umountExit <- mnterr
1275                         }()
1276
1277                         for again := true; again; {
1278                                 again = false
1279                                 select {
1280                                 case <-umountExit:
1281                                         umount = nil
1282                                         again = true
1283                                 case <-runner.ArvMountExit:
1284                                         break
1285                                 case <-time.After(time.Duration((delay + 1) * int64(time.Second))):
1286                                         runner.CrunchLog.Printf("Timed out waiting for unmount")
1287                                         if umount != nil {
1288                                                 umount.Process.Kill()
1289                                         }
1290                                         runner.ArvMount.Process.Kill()
1291                                 }
1292                         }
1293                 }
1294                 runner.ArvMount = nil
1295         }
1296
1297         if runner.ArvMountPoint != "" {
1298                 if rmerr := os.Remove(runner.ArvMountPoint); rmerr != nil {
1299                         runner.CrunchLog.Printf("While cleaning up arv-mount directory %s: %v", runner.ArvMountPoint, rmerr)
1300                 }
1301                 runner.ArvMountPoint = ""
1302         }
1303
1304         if rmerr := os.RemoveAll(runner.parentTemp); rmerr != nil {
1305                 runner.CrunchLog.Printf("While cleaning up temporary directory %s: %v", runner.parentTemp, rmerr)
1306         }
1307 }
1308
1309 // CommitLogs posts the collection containing the final container logs.
1310 func (runner *ContainerRunner) CommitLogs() error {
1311         func() {
1312                 // Hold cStateLock to prevent races on CrunchLog (e.g., stop()).
1313                 runner.cStateLock.Lock()
1314                 defer runner.cStateLock.Unlock()
1315
1316                 runner.CrunchLog.Print(runner.finalState)
1317
1318                 if runner.arvMountLog != nil {
1319                         runner.arvMountLog.Close()
1320                 }
1321                 runner.CrunchLog.Close()
1322
1323                 // Closing CrunchLog above allows them to be committed to Keep at this
1324                 // point, but re-open crunch log with ArvClient in case there are any
1325                 // other further errors (such as failing to write the log to Keep!)
1326                 // while shutting down
1327                 runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{
1328                         ArvClient:     runner.DispatcherArvClient,
1329                         UUID:          runner.Container.UUID,
1330                         loggingStream: "crunch-run",
1331                         writeCloser:   nil,
1332                 })
1333                 runner.CrunchLog.Immediate = log.New(os.Stderr, runner.Container.UUID+" ", 0)
1334         }()
1335
1336         if runner.keepstoreLogger != nil {
1337                 // Flush any buffered logs from our local keepstore
1338                 // process.  Discard anything logged after this point
1339                 // -- it won't end up in the log collection, so
1340                 // there's no point writing it to the collectionfs.
1341                 runner.keepstoreLogbuf.SetWriter(io.Discard)
1342                 runner.keepstoreLogger.Close()
1343                 runner.keepstoreLogger = nil
1344         }
1345
1346         if runner.LogsPDH != nil {
1347                 // If we have already assigned something to LogsPDH,
1348                 // we must be closing the re-opened log, which won't
1349                 // end up getting attached to the container record and
1350                 // therefore doesn't need to be saved as a collection
1351                 // -- it exists only to send logs to other channels.
1352                 return nil
1353         }
1354
1355         saved, err := runner.saveLogCollection(true)
1356         if err != nil {
1357                 return fmt.Errorf("error saving log collection: %s", err)
1358         }
1359         runner.logMtx.Lock()
1360         defer runner.logMtx.Unlock()
1361         runner.LogsPDH = &saved.PortableDataHash
1362         return nil
1363 }
1364
1365 func (runner *ContainerRunner) saveLogCollection(final bool) (response arvados.Collection, err error) {
1366         runner.logMtx.Lock()
1367         defer runner.logMtx.Unlock()
1368         if runner.LogsPDH != nil {
1369                 // Already finalized.
1370                 return
1371         }
1372         updates := arvadosclient.Dict{
1373                 "name": "logs for " + runner.Container.UUID,
1374         }
1375         mt, err1 := runner.LogCollection.MarshalManifest(".")
1376         if err1 == nil {
1377                 // Only send updated manifest text if there was no
1378                 // error.
1379                 updates["manifest_text"] = mt
1380         }
1381
1382         // Even if flushing the manifest had an error, we still want
1383         // to update the log record, if possible, to push the trash_at
1384         // and delete_at times into the future.  Details on bug
1385         // #17293.
1386         if final {
1387                 updates["is_trashed"] = true
1388         } else {
1389                 exp := time.Now().Add(crunchLogUpdatePeriod * 24)
1390                 updates["trash_at"] = exp
1391                 updates["delete_at"] = exp
1392         }
1393         reqBody := arvadosclient.Dict{"collection": updates}
1394         var err2 error
1395         if runner.logUUID == "" {
1396                 reqBody["ensure_unique_name"] = true
1397                 err2 = runner.DispatcherArvClient.Create("collections", reqBody, &response)
1398         } else {
1399                 err2 = runner.DispatcherArvClient.Update("collections", runner.logUUID, reqBody, &response)
1400         }
1401         if err2 == nil {
1402                 runner.logUUID = response.UUID
1403         }
1404
1405         if err1 != nil || err2 != nil {
1406                 err = fmt.Errorf("error recording logs: %q, %q", err1, err2)
1407         }
1408         return
1409 }
1410
1411 // UpdateContainerRunning updates the container state to "Running"
1412 func (runner *ContainerRunner) UpdateContainerRunning() error {
1413         runner.cStateLock.Lock()
1414         defer runner.cStateLock.Unlock()
1415         if runner.cCancelled {
1416                 return ErrCancelled
1417         }
1418         return runner.DispatcherArvClient.Update("containers", runner.Container.UUID,
1419                 arvadosclient.Dict{"container": arvadosclient.Dict{"state": "Running", "gateway_address": runner.gateway.Address}}, nil)
1420 }
1421
1422 // ContainerToken returns the api_token the container (and any
1423 // arv-mount processes) are allowed to use.
1424 func (runner *ContainerRunner) ContainerToken() (string, error) {
1425         if runner.token != "" {
1426                 return runner.token, nil
1427         }
1428
1429         var auth arvados.APIClientAuthorization
1430         err := runner.DispatcherArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth)
1431         if err != nil {
1432                 return "", err
1433         }
1434         runner.token = fmt.Sprintf("v2/%s/%s/%s", auth.UUID, auth.APIToken, runner.Container.UUID)
1435         return runner.token, nil
1436 }
1437
1438 // UpdateContainerFinal updates the container record state on API
1439 // server to "Complete" or "Cancelled"
1440 func (runner *ContainerRunner) UpdateContainerFinal() error {
1441         update := arvadosclient.Dict{}
1442         update["state"] = runner.finalState
1443         if runner.LogsPDH != nil {
1444                 update["log"] = *runner.LogsPDH
1445         }
1446         if runner.finalState == "Complete" {
1447                 if runner.ExitCode != nil {
1448                         update["exit_code"] = *runner.ExitCode
1449                 }
1450                 if runner.OutputPDH != nil {
1451                         update["output"] = *runner.OutputPDH
1452                 }
1453         }
1454         return runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{"container": update}, nil)
1455 }
1456
1457 // IsCancelled returns the value of Cancelled, with goroutine safety.
1458 func (runner *ContainerRunner) IsCancelled() bool {
1459         runner.cStateLock.Lock()
1460         defer runner.cStateLock.Unlock()
1461         return runner.cCancelled
1462 }
1463
1464 // NewArvLogWriter creates an ArvLogWriter
1465 func (runner *ContainerRunner) NewArvLogWriter(name string) (io.WriteCloser, error) {
1466         writer, err := runner.LogCollection.OpenFile(name+".txt", os.O_CREATE|os.O_WRONLY, 0666)
1467         if err != nil {
1468                 return nil, err
1469         }
1470         return &ArvLogWriter{
1471                 ArvClient:     runner.DispatcherArvClient,
1472                 UUID:          runner.Container.UUID,
1473                 loggingStream: name,
1474                 writeCloser:   writer,
1475         }, nil
1476 }
1477
1478 // Run the full container lifecycle.
1479 func (runner *ContainerRunner) Run() (err error) {
1480         runner.CrunchLog.Printf("crunch-run %s started", cmd.Version.String())
1481         runner.CrunchLog.Printf("%s", currentUserAndGroups())
1482         v, _ := exec.Command("arv-mount", "--version").CombinedOutput()
1483         runner.CrunchLog.Printf("Using FUSE mount: %s", v)
1484         runner.CrunchLog.Printf("Using container runtime: %s", runner.executor.Runtime())
1485         runner.CrunchLog.Printf("Executing container: %s", runner.Container.UUID)
1486
1487         hostname, hosterr := os.Hostname()
1488         if hosterr != nil {
1489                 runner.CrunchLog.Printf("Error getting hostname '%v'", hosterr)
1490         } else {
1491                 runner.CrunchLog.Printf("Executing on host '%s'", hostname)
1492         }
1493
1494         runner.finalState = "Queued"
1495
1496         defer func() {
1497                 runner.CleanupDirs()
1498
1499                 runner.CrunchLog.Printf("crunch-run finished")
1500                 runner.CrunchLog.Close()
1501         }()
1502
1503         err = runner.fetchContainerRecord()
1504         if err != nil {
1505                 return
1506         }
1507         if runner.Container.State != "Locked" {
1508                 return fmt.Errorf("dispatch error detected: container %q has state %q", runner.Container.UUID, runner.Container.State)
1509         }
1510
1511         var bindmounts map[string]bindmount
1512         defer func() {
1513                 // checkErr prints e (unless it's nil) and sets err to
1514                 // e (unless err is already non-nil). Thus, if err
1515                 // hasn't already been assigned when Run() returns,
1516                 // this cleanup func will cause Run() to return the
1517                 // first non-nil error that is passed to checkErr().
1518                 checkErr := func(errorIn string, e error) {
1519                         if e == nil {
1520                                 return
1521                         }
1522                         runner.CrunchLog.Printf("error in %s: %v", errorIn, e)
1523                         if err == nil {
1524                                 err = e
1525                         }
1526                         if runner.finalState == "Complete" {
1527                                 // There was an error in the finalization.
1528                                 runner.finalState = "Cancelled"
1529                         }
1530                 }
1531
1532                 // Log the error encountered in Run(), if any
1533                 checkErr("Run", err)
1534
1535                 if runner.finalState == "Queued" {
1536                         runner.UpdateContainerFinal()
1537                         return
1538                 }
1539
1540                 if runner.IsCancelled() {
1541                         runner.finalState = "Cancelled"
1542                         // but don't return yet -- we still want to
1543                         // capture partial output and write logs
1544                 }
1545
1546                 if bindmounts != nil {
1547                         checkErr("CaptureOutput", runner.CaptureOutput(bindmounts))
1548                 }
1549                 checkErr("stopHoststat", runner.stopHoststat())
1550                 checkErr("CommitLogs", runner.CommitLogs())
1551                 runner.CleanupDirs()
1552                 checkErr("UpdateContainerFinal", runner.UpdateContainerFinal())
1553         }()
1554
1555         runner.setupSignals()
1556         err = runner.startHoststat()
1557         if err != nil {
1558                 return
1559         }
1560
1561         // set up FUSE mount and binds
1562         bindmounts, err = runner.SetupMounts()
1563         if err != nil {
1564                 runner.finalState = "Cancelled"
1565                 err = fmt.Errorf("While setting up mounts: %v", err)
1566                 return
1567         }
1568
1569         // check for and/or load image
1570         imageID, err := runner.LoadImage()
1571         if err != nil {
1572                 if !runner.checkBrokenNode(err) {
1573                         // Failed to load image but not due to a "broken node"
1574                         // condition, probably user error.
1575                         runner.finalState = "Cancelled"
1576                 }
1577                 err = fmt.Errorf("While loading container image: %v", err)
1578                 return
1579         }
1580
1581         err = runner.CreateContainer(imageID, bindmounts)
1582         if err != nil {
1583                 return
1584         }
1585         err = runner.LogHostInfo()
1586         if err != nil {
1587                 return
1588         }
1589         err = runner.LogNodeRecord()
1590         if err != nil {
1591                 return
1592         }
1593         err = runner.LogContainerRecord()
1594         if err != nil {
1595                 return
1596         }
1597
1598         if runner.IsCancelled() {
1599                 return
1600         }
1601
1602         err = runner.UpdateContainerRunning()
1603         if err != nil {
1604                 return
1605         }
1606         runner.finalState = "Cancelled"
1607
1608         err = runner.startCrunchstat()
1609         if err != nil {
1610                 return
1611         }
1612
1613         err = runner.StartContainer()
1614         if err != nil {
1615                 runner.checkBrokenNode(err)
1616                 return
1617         }
1618
1619         err = runner.WaitFinish()
1620         if err == nil && !runner.IsCancelled() {
1621                 runner.finalState = "Complete"
1622         }
1623         return
1624 }
1625
1626 // Fetch the current container record (uuid = runner.Container.UUID)
1627 // into runner.Container.
1628 func (runner *ContainerRunner) fetchContainerRecord() error {
1629         reader, err := runner.DispatcherArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
1630         if err != nil {
1631                 return fmt.Errorf("error fetching container record: %v", err)
1632         }
1633         defer reader.Close()
1634
1635         dec := json.NewDecoder(reader)
1636         dec.UseNumber()
1637         err = dec.Decode(&runner.Container)
1638         if err != nil {
1639                 return fmt.Errorf("error decoding container record: %v", err)
1640         }
1641
1642         var sm struct {
1643                 SecretMounts map[string]arvados.Mount `json:"secret_mounts"`
1644         }
1645
1646         containerToken, err := runner.ContainerToken()
1647         if err != nil {
1648                 return fmt.Errorf("error getting container token: %v", err)
1649         }
1650
1651         runner.ContainerArvClient, runner.ContainerKeepClient,
1652                 runner.containerClient, err = runner.MkArvClient(containerToken)
1653         if err != nil {
1654                 return fmt.Errorf("error creating container API client: %v", err)
1655         }
1656
1657         runner.ContainerKeepClient.SetStorageClasses(runner.Container.OutputStorageClasses)
1658         runner.DispatcherKeepClient.SetStorageClasses(runner.Container.OutputStorageClasses)
1659
1660         err = runner.ContainerArvClient.Call("GET", "containers", runner.Container.UUID, "secret_mounts", nil, &sm)
1661         if err != nil {
1662                 if apierr, ok := err.(arvadosclient.APIServerError); !ok || apierr.HttpStatusCode != 404 {
1663                         return fmt.Errorf("error fetching secret_mounts: %v", err)
1664                 }
1665                 // ok && apierr.HttpStatusCode == 404, which means
1666                 // secret_mounts isn't supported by this API server.
1667         }
1668         runner.SecretMounts = sm.SecretMounts
1669
1670         return nil
1671 }
1672
1673 // NewContainerRunner creates a new container runner.
1674 func NewContainerRunner(dispatcherClient *arvados.Client,
1675         dispatcherArvClient IArvadosClient,
1676         dispatcherKeepClient IKeepClient,
1677         containerUUID string) (*ContainerRunner, error) {
1678
1679         cr := &ContainerRunner{
1680                 dispatcherClient:     dispatcherClient,
1681                 DispatcherArvClient:  dispatcherArvClient,
1682                 DispatcherKeepClient: dispatcherKeepClient,
1683         }
1684         cr.NewLogWriter = cr.NewArvLogWriter
1685         cr.RunArvMount = cr.ArvMountCmd
1686         cr.MkTempDir = ioutil.TempDir
1687         cr.MkArvClient = func(token string) (IArvadosClient, IKeepClient, *arvados.Client, error) {
1688                 cl, err := arvadosclient.MakeArvadosClient()
1689                 if err != nil {
1690                         return nil, nil, nil, err
1691                 }
1692                 cl.ApiToken = token
1693                 kc, err := keepclient.MakeKeepClient(cl)
1694                 if err != nil {
1695                         return nil, nil, nil, err
1696                 }
1697                 c2 := arvados.NewClientFromEnv()
1698                 c2.AuthToken = token
1699                 return cl, kc, c2, nil
1700         }
1701         var err error
1702         cr.LogCollection, err = (&arvados.Collection{}).FileSystem(cr.dispatcherClient, cr.DispatcherKeepClient)
1703         if err != nil {
1704                 return nil, err
1705         }
1706         cr.Container.UUID = containerUUID
1707         w, err := cr.NewLogWriter("crunch-run")
1708         if err != nil {
1709                 return nil, err
1710         }
1711         cr.CrunchLog = NewThrottledLogger(w)
1712         cr.CrunchLog.Immediate = log.New(os.Stderr, containerUUID+" ", 0)
1713
1714         loadLogThrottleParams(dispatcherArvClient)
1715         go cr.updateLogs()
1716
1717         return cr, nil
1718 }
1719
1720 func (command) RunCommand(prog string, args []string, stdin io.Reader, stdout, stderr io.Writer) int {
1721         log := log.New(stderr, "", 0)
1722         flags := flag.NewFlagSet(prog, flag.ContinueOnError)
1723         statInterval := flags.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting")
1724         cgroupRoot := flags.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree")
1725         cgroupParent := flags.String("cgroup-parent", "docker", "name of container's parent cgroup (ignored if -cgroup-parent-subsystem is used)")
1726         cgroupParentSubsystem := flags.String("cgroup-parent-subsystem", "", "use current cgroup for given subsystem as parent cgroup for container")
1727         caCertsPath := flags.String("ca-certs", "", "Path to TLS root certificates")
1728         detach := flags.Bool("detach", false, "Detach from parent process and run in the background")
1729         stdinConfig := flags.Bool("stdin-config", false, "Load config and environment variables from JSON message on stdin")
1730         configFile := flags.String("config", arvados.DefaultConfigFile, "filename of cluster config file to try loading if -stdin-config=false (default is $ARVADOS_CONFIG)")
1731         sleep := flags.Duration("sleep", 0, "Delay before starting (testing use only)")
1732         kill := flags.Int("kill", -1, "Send signal to an existing crunch-run process for given UUID")
1733         list := flags.Bool("list", false, "List UUIDs of existing crunch-run processes")
1734         enableMemoryLimit := flags.Bool("enable-memory-limit", true, "tell container runtime to limit container's memory usage")
1735         enableNetwork := flags.String("container-enable-networking", "default", "enable networking \"always\" (for all containers) or \"default\" (for containers that request it)")
1736         networkMode := flags.String("container-network-mode", "default", `Docker network mode for container (use any argument valid for docker --net)`)
1737         memprofile := flags.String("memprofile", "", "write memory profile to `file` after running container")
1738         runtimeEngine := flags.String("runtime-engine", "docker", "container runtime: docker or singularity")
1739         brokenNodeHook := flags.String("broken-node-hook", "", "script to run if node is detected to be broken (for example, Docker daemon is not running)")
1740         flags.Duration("check-containerd", 0, "Ignored. Exists for compatibility with older versions.")
1741
1742         ignoreDetachFlag := false
1743         if len(args) > 0 && args[0] == "-no-detach" {
1744                 // This process was invoked by a parent process, which
1745                 // has passed along its own arguments, including
1746                 // -detach, after the leading -no-detach flag.  Strip
1747                 // the leading -no-detach flag (it's not recognized by
1748                 // flags.Parse()) and ignore the -detach flag that
1749                 // comes later.
1750                 args = args[1:]
1751                 ignoreDetachFlag = true
1752         }
1753
1754         if ok, code := cmd.ParseFlags(flags, prog, args, "container-uuid", stderr); !ok {
1755                 return code
1756         } else if !*list && flags.NArg() != 1 {
1757                 fmt.Fprintf(stderr, "missing required argument: container-uuid (try -help)\n")
1758                 return 2
1759         }
1760
1761         containerUUID := flags.Arg(0)
1762
1763         switch {
1764         case *detach && !ignoreDetachFlag:
1765                 return Detach(containerUUID, prog, args, os.Stdin, os.Stdout, os.Stderr)
1766         case *kill >= 0:
1767                 return KillProcess(containerUUID, syscall.Signal(*kill), os.Stdout, os.Stderr)
1768         case *list:
1769                 return ListProcesses(os.Stdout, os.Stderr)
1770         }
1771
1772         if len(containerUUID) != 27 {
1773                 log.Printf("usage: %s [options] UUID", prog)
1774                 return 1
1775         }
1776
1777         var keepstoreLogbuf bufThenWrite
1778         var conf ConfigData
1779         if *stdinConfig {
1780                 err := json.NewDecoder(stdin).Decode(&conf)
1781                 if err != nil {
1782                         log.Printf("decode stdin: %s", err)
1783                         return 1
1784                 }
1785                 for k, v := range conf.Env {
1786                         err = os.Setenv(k, v)
1787                         if err != nil {
1788                                 log.Printf("setenv(%q): %s", k, err)
1789                                 return 1
1790                         }
1791                 }
1792                 if conf.Cluster != nil {
1793                         // ClusterID is missing from the JSON
1794                         // representation, but we need it to generate
1795                         // a valid config file for keepstore, so we
1796                         // fill it using the container UUID prefix.
1797                         conf.Cluster.ClusterID = containerUUID[:5]
1798                 }
1799         } else {
1800                 conf = hpcConfData(containerUUID, *configFile, io.MultiWriter(&keepstoreLogbuf, stderr))
1801         }
1802
1803         log.Printf("crunch-run %s started", cmd.Version.String())
1804         time.Sleep(*sleep)
1805
1806         if *caCertsPath != "" {
1807                 arvadosclient.CertFiles = []string{*caCertsPath}
1808         }
1809
1810         keepstore, err := startLocalKeepstore(conf, io.MultiWriter(&keepstoreLogbuf, stderr))
1811         if err != nil {
1812                 log.Print(err)
1813                 return 1
1814         }
1815         if keepstore != nil {
1816                 defer keepstore.Process.Kill()
1817         }
1818
1819         api, err := arvadosclient.MakeArvadosClient()
1820         if err != nil {
1821                 log.Printf("%s: %v", containerUUID, err)
1822                 return 1
1823         }
1824         api.Retries = 8
1825
1826         kc, err := keepclient.MakeKeepClient(api)
1827         if err != nil {
1828                 log.Printf("%s: %v", containerUUID, err)
1829                 return 1
1830         }
1831         kc.BlockCache = &keepclient.BlockCache{MaxBlocks: 2}
1832         kc.Retries = 4
1833
1834         cr, err := NewContainerRunner(arvados.NewClientFromEnv(), api, kc, containerUUID)
1835         if err != nil {
1836                 log.Print(err)
1837                 return 1
1838         }
1839
1840         if keepstore == nil {
1841                 // Log explanation (if any) for why we're not running
1842                 // a local keepstore.
1843                 var buf bytes.Buffer
1844                 keepstoreLogbuf.SetWriter(&buf)
1845                 if buf.Len() > 0 {
1846                         cr.CrunchLog.Printf("%s", strings.TrimSpace(buf.String()))
1847                 }
1848         } else if logWhat := conf.Cluster.Containers.LocalKeepLogsToContainerLog; logWhat == "none" {
1849                 cr.CrunchLog.Printf("using local keepstore process (pid %d) at %s", keepstore.Process.Pid, os.Getenv("ARVADOS_KEEP_SERVICES"))
1850                 keepstoreLogbuf.SetWriter(io.Discard)
1851         } else {
1852                 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"))
1853                 logwriter, err := cr.NewLogWriter("keepstore")
1854                 if err != nil {
1855                         log.Print(err)
1856                         return 1
1857                 }
1858                 cr.keepstoreLogger = NewThrottledLogger(logwriter)
1859
1860                 var writer io.WriteCloser = cr.keepstoreLogger
1861                 if logWhat == "errors" {
1862                         writer = &filterKeepstoreErrorsOnly{WriteCloser: writer}
1863                 } else if logWhat != "all" {
1864                         // should have been caught earlier by
1865                         // dispatcher's config loader
1866                         log.Printf("invalid value for Containers.LocalKeepLogsToContainerLog: %q", logWhat)
1867                         return 1
1868                 }
1869                 err = keepstoreLogbuf.SetWriter(writer)
1870                 if err != nil {
1871                         log.Print(err)
1872                         return 1
1873                 }
1874                 cr.keepstoreLogbuf = &keepstoreLogbuf
1875         }
1876
1877         switch *runtimeEngine {
1878         case "docker":
1879                 cr.executor, err = newDockerExecutor(containerUUID, cr.CrunchLog.Printf, cr.containerWatchdogInterval)
1880         case "singularity":
1881                 cr.executor, err = newSingularityExecutor(cr.CrunchLog.Printf)
1882         default:
1883                 cr.CrunchLog.Printf("%s: unsupported RuntimeEngine %q", containerUUID, *runtimeEngine)
1884                 cr.CrunchLog.Close()
1885                 return 1
1886         }
1887         if err != nil {
1888                 cr.CrunchLog.Printf("%s: %v", containerUUID, err)
1889                 cr.checkBrokenNode(err)
1890                 cr.CrunchLog.Close()
1891                 return 1
1892         }
1893         defer cr.executor.Close()
1894
1895         cr.brokenNodeHook = *brokenNodeHook
1896
1897         gwAuthSecret := os.Getenv("GatewayAuthSecret")
1898         os.Unsetenv("GatewayAuthSecret")
1899         if gwAuthSecret == "" {
1900                 // not safe to run a gateway service without an auth
1901                 // secret
1902                 cr.CrunchLog.Printf("Not starting a gateway server (GatewayAuthSecret was not provided by dispatcher)")
1903         } else if gwListen := os.Getenv("GatewayAddress"); gwListen == "" {
1904                 // dispatcher did not tell us which external IP
1905                 // address to advertise --> no gateway service
1906                 cr.CrunchLog.Printf("Not starting a gateway server (GatewayAddress was not provided by dispatcher)")
1907         } else if de, ok := cr.executor.(*dockerExecutor); ok {
1908                 cr.gateway = Gateway{
1909                         Address:            gwListen,
1910                         AuthSecret:         gwAuthSecret,
1911                         ContainerUUID:      containerUUID,
1912                         DockerContainerID:  &de.containerID,
1913                         Log:                cr.CrunchLog,
1914                         ContainerIPAddress: dockerContainerIPAddress(&de.containerID),
1915                 }
1916                 err = cr.gateway.Start()
1917                 if err != nil {
1918                         log.Printf("error starting gateway server: %s", err)
1919                         return 1
1920                 }
1921         }
1922
1923         parentTemp, tmperr := cr.MkTempDir("", "crunch-run."+containerUUID+".")
1924         if tmperr != nil {
1925                 log.Printf("%s: %v", containerUUID, tmperr)
1926                 return 1
1927         }
1928
1929         cr.parentTemp = parentTemp
1930         cr.statInterval = *statInterval
1931         cr.cgroupRoot = *cgroupRoot
1932         cr.expectCgroupParent = *cgroupParent
1933         cr.enableMemoryLimit = *enableMemoryLimit
1934         cr.enableNetwork = *enableNetwork
1935         cr.networkMode = *networkMode
1936         if *cgroupParentSubsystem != "" {
1937                 p, err := findCgroup(*cgroupParentSubsystem)
1938                 if err != nil {
1939                         log.Printf("fatal: cgroup parent subsystem: %s", err)
1940                         return 1
1941                 }
1942                 cr.setCgroupParent = p
1943                 cr.expectCgroupParent = p
1944         }
1945
1946         runerr := cr.Run()
1947
1948         if *memprofile != "" {
1949                 f, err := os.Create(*memprofile)
1950                 if err != nil {
1951                         log.Printf("could not create memory profile: %s", err)
1952                 }
1953                 runtime.GC() // get up-to-date statistics
1954                 if err := pprof.WriteHeapProfile(f); err != nil {
1955                         log.Printf("could not write memory profile: %s", err)
1956                 }
1957                 closeerr := f.Close()
1958                 if closeerr != nil {
1959                         log.Printf("closing memprofile file: %s", err)
1960                 }
1961         }
1962
1963         if runerr != nil {
1964                 log.Printf("%s: %v", containerUUID, runerr)
1965                 return 1
1966         }
1967         return 0
1968 }
1969
1970 // Try to load ConfigData in hpc (slurm/lsf) environment. This means
1971 // loading the cluster config from the specified file and (if that
1972 // works) getting the runtime_constraints container field from
1973 // controller to determine # VCPUs so we can calculate KeepBuffers.
1974 func hpcConfData(uuid string, configFile string, stderr io.Writer) ConfigData {
1975         var conf ConfigData
1976         conf.Cluster = loadClusterConfigFile(configFile, stderr)
1977         if conf.Cluster == nil {
1978                 // skip loading the container record -- we won't be
1979                 // able to start local keepstore anyway.
1980                 return conf
1981         }
1982         arv, err := arvadosclient.MakeArvadosClient()
1983         if err != nil {
1984                 fmt.Fprintf(stderr, "error setting up arvadosclient: %s\n", err)
1985                 return conf
1986         }
1987         arv.Retries = 8
1988         var ctr arvados.Container
1989         err = arv.Call("GET", "containers", uuid, "", arvadosclient.Dict{"select": []string{"runtime_constraints"}}, &ctr)
1990         if err != nil {
1991                 fmt.Fprintf(stderr, "error getting container record: %s\n", err)
1992                 return conf
1993         }
1994         if ctr.RuntimeConstraints.VCPUs > 0 {
1995                 conf.KeepBuffers = ctr.RuntimeConstraints.VCPUs * conf.Cluster.Containers.LocalKeepBlobBuffersPerVCPU
1996         }
1997         return conf
1998 }
1999
2000 // Load cluster config file from given path. If an error occurs, log
2001 // the error to stderr and return nil.
2002 func loadClusterConfigFile(path string, stderr io.Writer) *arvados.Cluster {
2003         ldr := config.NewLoader(&bytes.Buffer{}, ctxlog.New(stderr, "plain", "info"))
2004         ldr.Path = path
2005         cfg, err := ldr.Load()
2006         if err != nil {
2007                 fmt.Fprintf(stderr, "could not load config file %s: %s\n", path, err)
2008                 return nil
2009         }
2010         cluster, err := cfg.GetCluster("")
2011         if err != nil {
2012                 fmt.Fprintf(stderr, "could not use config file %s: %s\n", path, err)
2013                 return nil
2014         }
2015         fmt.Fprintf(stderr, "loaded config file %s\n", path)
2016         return cluster
2017 }
2018
2019 func startLocalKeepstore(configData ConfigData, logbuf io.Writer) (*exec.Cmd, error) {
2020         if configData.KeepBuffers < 1 {
2021                 fmt.Fprintf(logbuf, "not starting a local keepstore process because KeepBuffers=%v in config\n", configData.KeepBuffers)
2022                 return nil, nil
2023         }
2024         if configData.Cluster == nil {
2025                 fmt.Fprint(logbuf, "not starting a local keepstore process because cluster config file was not loaded\n")
2026                 return nil, nil
2027         }
2028         for uuid, vol := range configData.Cluster.Volumes {
2029                 if len(vol.AccessViaHosts) > 0 {
2030                         fmt.Fprintf(logbuf, "not starting a local keepstore process because a volume (%s) uses AccessViaHosts\n", uuid)
2031                         return nil, nil
2032                 }
2033                 if !vol.ReadOnly && vol.Replication < configData.Cluster.Collections.DefaultReplication {
2034                         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)
2035                         return nil, nil
2036                 }
2037         }
2038
2039         // Rather than have an alternate way to tell keepstore how
2040         // many buffers to use when starting it this way, we just
2041         // modify the cluster configuration that we feed it on stdin.
2042         configData.Cluster.API.MaxKeepBlobBuffers = configData.KeepBuffers
2043
2044         ln, err := net.Listen("tcp", "localhost:0")
2045         if err != nil {
2046                 return nil, err
2047         }
2048         _, port, err := net.SplitHostPort(ln.Addr().String())
2049         if err != nil {
2050                 ln.Close()
2051                 return nil, err
2052         }
2053         ln.Close()
2054         url := "http://localhost:" + port
2055
2056         fmt.Fprintf(logbuf, "starting keepstore on %s\n", url)
2057
2058         var confJSON bytes.Buffer
2059         err = json.NewEncoder(&confJSON).Encode(arvados.Config{
2060                 Clusters: map[string]arvados.Cluster{
2061                         configData.Cluster.ClusterID: *configData.Cluster,
2062                 },
2063         })
2064         if err != nil {
2065                 return nil, err
2066         }
2067         cmd := exec.Command("/proc/self/exe", "keepstore", "-config=-")
2068         if target, err := os.Readlink(cmd.Path); err == nil && strings.HasSuffix(target, ".test") {
2069                 // If we're a 'go test' process, running
2070                 // /proc/self/exe would start the test suite in a
2071                 // child process, which is not what we want.
2072                 cmd.Path, _ = exec.LookPath("go")
2073                 cmd.Args = append([]string{"go", "run", "../../cmd/arvados-server"}, cmd.Args[1:]...)
2074                 cmd.Env = os.Environ()
2075         }
2076         cmd.Stdin = &confJSON
2077         cmd.Stdout = logbuf
2078         cmd.Stderr = logbuf
2079         cmd.Env = append(cmd.Env,
2080                 "GOGC=10",
2081                 "ARVADOS_SERVICE_INTERNAL_URL="+url)
2082         err = cmd.Start()
2083         if err != nil {
2084                 return nil, fmt.Errorf("error starting keepstore process: %w", err)
2085         }
2086         cmdExited := false
2087         go func() {
2088                 cmd.Wait()
2089                 cmdExited = true
2090         }()
2091         ctx, cancel := context.WithDeadline(context.Background(), time.Now().Add(time.Second*10))
2092         defer cancel()
2093         poll := time.NewTicker(time.Second / 10)
2094         defer poll.Stop()
2095         client := http.Client{}
2096         for range poll.C {
2097                 testReq, err := http.NewRequestWithContext(ctx, "GET", url+"/_health/ping", nil)
2098                 testReq.Header.Set("Authorization", "Bearer "+configData.Cluster.ManagementToken)
2099                 if err != nil {
2100                         return nil, err
2101                 }
2102                 resp, err := client.Do(testReq)
2103                 if err == nil {
2104                         resp.Body.Close()
2105                         if resp.StatusCode == http.StatusOK {
2106                                 break
2107                         }
2108                 }
2109                 if cmdExited {
2110                         return nil, fmt.Errorf("keepstore child process exited")
2111                 }
2112                 if ctx.Err() != nil {
2113                         return nil, fmt.Errorf("timed out waiting for new keepstore process to report healthy")
2114                 }
2115         }
2116         os.Setenv("ARVADOS_KEEP_SERVICES", url)
2117         return cmd, nil
2118 }
2119
2120 // return current uid, gid, groups in a format suitable for logging:
2121 // "crunch-run process has uid=1234(arvados) gid=1234(arvados)
2122 // groups=1234(arvados),114(fuse)"
2123 func currentUserAndGroups() string {
2124         u, err := user.Current()
2125         if err != nil {
2126                 return fmt.Sprintf("error getting current user ID: %s", err)
2127         }
2128         s := fmt.Sprintf("crunch-run process has uid=%s(%s) gid=%s", u.Uid, u.Username, u.Gid)
2129         if g, err := user.LookupGroupId(u.Gid); err == nil {
2130                 s += fmt.Sprintf("(%s)", g.Name)
2131         }
2132         s += " groups="
2133         if gids, err := u.GroupIds(); err == nil {
2134                 for i, gid := range gids {
2135                         if i > 0 {
2136                                 s += ","
2137                         }
2138                         s += gid
2139                         if g, err := user.LookupGroupId(gid); err == nil {
2140                                 s += fmt.Sprintf("(%s)", g.Name)
2141                         }
2142                 }
2143         }
2144         return s
2145 }