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