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/cloud"
  35         "git.arvados.org/arvados.git/lib/cmd"
  36         "git.arvados.org/arvados.git/lib/config"
  37         "git.arvados.org/arvados.git/lib/crunchstat"
  38         "git.arvados.org/arvados.git/sdk/go/arvados"
  39         "git.arvados.org/arvados.git/sdk/go/arvadosclient"
  40         "git.arvados.org/arvados.git/sdk/go/ctxlog"
  41         "git.arvados.org/arvados.git/sdk/go/keepclient"
  42         "git.arvados.org/arvados.git/sdk/go/manifest"
  43         "golang.org/x/sys/unix"
  44 )
  45
  46 type command struct{}
  47
  48 var Command = command{}
  49
  50 // ConfigData contains environment variables and (when needed) cluster
  51 // configuration, passed from dispatchcloud to crunch-run on stdin.
  52 type ConfigData struct {
  53         Env         map[string]string
  54         KeepBuffers int
  55         Cluster     *arvados.Cluster
  56 }
  57
  58 // IArvadosClient is the minimal Arvados API methods used by crunch-run.
  59 type IArvadosClient interface {
  60         Create(resourceType string, parameters arvadosclient.Dict, output interface{}) error
  61         Get(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  62         Update(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  63         Call(method, resourceType, uuid, action string, parameters arvadosclient.Dict, output interface{}) error
  64         CallRaw(method string, resourceType string, uuid string, action string, parameters arvadosclient.Dict) (reader io.ReadCloser, err error)
  65         Discovery(key string) (interface{}, error)
  66 }
  67
  68 // ErrCancelled is the error returned when the container is cancelled.
  69 var ErrCancelled = errors.New("Cancelled")
  70
  71 // IKeepClient is the minimal Keep API methods used by crunch-run.
  72 type IKeepClient interface {
  73         BlockWrite(context.Context, arvados.BlockWriteOptions) (arvados.BlockWriteResponse, error)
  74         ReadAt(locator string, p []byte, off int) (int, error)
  75         ManifestFileReader(m manifest.Manifest, filename string) (arvados.File, error)
  76         LocalLocator(locator string) (string, error)
  77         ClearBlockCache()
  78         SetStorageClasses(sc []string)
  79 }
  80
  81 // NewLogWriter is a factory function to create a new log writer.
  82 type NewLogWriter func(name string) (io.WriteCloser, error)
  83
  84 type RunArvMount func(cmdline []string, tok string) (*exec.Cmd, error)
  85
  86 type MkTempDir func(string, string) (string, error)
  87
  88 type PsProcess interface {
  89         CmdlineSlice() ([]string, error)
  90 }
  91
  92 // ContainerRunner is the main stateful struct used for a single execution of a
  93 // container.
  94 type ContainerRunner struct {
  95         executor       containerExecutor
  96         executorStdin  io.Closer
  97         executorStdout io.Closer
  98         executorStderr io.Closer
  99
 100         // Dispatcher client is initialized with the Dispatcher token.
 101         // This is a privileged token used to manage container status
 102         // and logs.
 103         //
 104         // We have both dispatcherClient and DispatcherArvClient
 105         // because there are two different incompatible Arvados Go
 106         // SDKs and we have to use both (hopefully this gets fixed in
 107         // #14467)
 108         dispatcherClient     *arvados.Client
 109         DispatcherArvClient  IArvadosClient
 110         DispatcherKeepClient IKeepClient
 111
 112         // Container client is initialized with the Container token
 113         // This token controls the permissions of the container, and
 114         // must be used for operations such as reading collections.
 115         //
 116         // Same comment as above applies to
 117         // containerClient/ContainerArvClient.
 118         containerClient     *arvados.Client
 119         ContainerArvClient  IArvadosClient
 120         ContainerKeepClient IKeepClient
 121
 122         Container     arvados.Container
 123         token         string
 124         ExitCode      *int
 125         NewLogWriter  NewLogWriter
 126         CrunchLog     *ThrottledLogger
 127         logUUID       string
 128         logMtx        sync.Mutex
 129         LogCollection arvados.CollectionFileSystem
 130         LogsPDH       *string
 131         RunArvMount   RunArvMount
 132         MkTempDir     MkTempDir
 133         ArvMount      *exec.Cmd
 134         ArvMountPoint string
 135         HostOutputDir string
 136         Volumes       map[string]struct{}
 137         OutputPDH     *string
 138         SigChan       chan os.Signal
 139         ArvMountExit  chan error
 140         SecretMounts  map[string]arvados.Mount
 141         MkArvClient   func(token string) (IArvadosClient, IKeepClient, *arvados.Client, error)
 142         finalState    string
 143         parentTemp    string
 144         costStartTime time.Time
 145
 146         keepstore        *exec.Cmd
 147         keepstoreLogger  io.WriteCloser
 148         keepstoreLogbuf  *bufThenWrite
 149         statLogger       io.WriteCloser
 150         statReporter     *crunchstat.Reporter
 151         hoststatLogger   io.WriteCloser
 152         hoststatReporter *crunchstat.Reporter
 153         statInterval     time.Duration
 154         cgroupRoot       string
 155         // What we expect the container's cgroup parent to be.
 156         expectCgroupParent string
 157         // What we tell docker to use as the container's cgroup
 158         // parent. Note: Ideally we would use the same field for both
 159         // expectCgroupParent and setCgroupParent, and just make it
 160         // default to "docker". However, when using docker < 1.10 with
 161         // systemd, specifying a non-empty cgroup parent (even the
 162         // default value "docker") hits a docker bug
 163         // (https://github.com/docker/docker/issues/17126). Using two
 164         // separate fields makes it possible to use the "expect cgroup
 165         // parent to be X" feature even on sites where the "specify
 166         // cgroup parent" feature breaks.
 167         setCgroupParent string
 168
 169         cStateLock sync.Mutex
 170         cCancelled bool // StopContainer() invoked
 171
 172         enableMemoryLimit bool
 173         enableNetwork     string // one of "default" or "always"
 174         networkMode       string // "none", "host", or "" -- passed through to executor
 175         brokenNodeHook    string // script to run if node appears to be broken
 176         arvMountLog       *ThrottledLogger
 177
 178         containerWatchdogInterval time.Duration
 179
 180         gateway Gateway
 181
 182         prices     []cloud.InstancePrice
 183         pricesLock sync.Mutex
 184 }
 185
 186 // setupSignals sets up signal handling to gracefully terminate the
 187 // underlying container and update state when receiving a TERM, INT or
 188 // QUIT signal.
 189 func (runner *ContainerRunner) setupSignals() {
 190         runner.SigChan = make(chan os.Signal, 1)
 191         signal.Notify(runner.SigChan, syscall.SIGTERM)
 192         signal.Notify(runner.SigChan, syscall.SIGINT)
 193         signal.Notify(runner.SigChan, syscall.SIGQUIT)
 194
 195         go func(sig chan os.Signal) {
 196                 for s := range sig {
 197                         runner.stop(s)
 198                 }
 199         }(runner.SigChan)
 200 }
 201
 202 // stop the underlying container.
 203 func (runner *ContainerRunner) stop(sig os.Signal) {
 204         runner.cStateLock.Lock()
 205         defer runner.cStateLock.Unlock()
 206         if sig != nil {
 207                 runner.CrunchLog.Printf("caught signal: %v", sig)
 208         }
 209         runner.cCancelled = true
 210         runner.CrunchLog.Printf("stopping container")
 211         err := runner.executor.Stop()
 212         if err != nil {
 213                 runner.CrunchLog.Printf("error stopping container: %s", err)
 214         }
 215 }
 216
 217 var errorBlacklist = []string{
 218         "(?ms).*[Cc]annot connect to the Docker daemon.*",
 219         "(?ms).*oci runtime error.*starting container process.*container init.*mounting.*to rootfs.*no such file or directory.*",
 220         "(?ms).*grpc: the connection is unavailable.*",
 221 }
 222
 223 func (runner *ContainerRunner) runBrokenNodeHook() {
 224         if runner.brokenNodeHook == "" {
 225                 path := filepath.Join(lockdir, brokenfile)
 226                 runner.CrunchLog.Printf("Writing %s to mark node as broken", path)
 227                 f, err := os.OpenFile(path, os.O_CREATE|os.O_RDWR, 0700)
 228                 if err != nil {
 229                         runner.CrunchLog.Printf("Error writing %s: %s", path, err)
 230                         return
 231                 }
 232                 f.Close()
 233         } else {
 234                 runner.CrunchLog.Printf("Running broken node hook %q", runner.brokenNodeHook)
 235                 // run killme script
 236                 c := exec.Command(runner.brokenNodeHook)
 237                 c.Stdout = runner.CrunchLog
 238                 c.Stderr = runner.CrunchLog
 239                 err := c.Run()
 240                 if err != nil {
 241                         runner.CrunchLog.Printf("Error running broken node hook: %v", err)
 242                 }
 243         }
 244 }
 245
 246 func (runner *ContainerRunner) checkBrokenNode(goterr error) bool {
 247         for _, d := range errorBlacklist {
 248                 if m, e := regexp.MatchString(d, goterr.Error()); m && e == nil {
 249                         runner.CrunchLog.Printf("Error suggests node is unable to run containers: %v", goterr)
 250                         runner.runBrokenNodeHook()
 251                         return true
 252                 }
 253         }
 254         return false
 255 }
 256
 257 // LoadImage determines the docker image id from the container record and
 258 // checks if it is available in the local Docker image store.  If not, it loads
 259 // the image from Keep.
 260 func (runner *ContainerRunner) LoadImage() (string, error) {
 261         runner.CrunchLog.Printf("Fetching Docker image from collection '%s'", runner.Container.ContainerImage)
 262
 263         d, err := os.Open(runner.ArvMountPoint + "/by_id/" + runner.Container.ContainerImage)
 264         if err != nil {
 265                 return "", err
 266         }
 267         defer d.Close()
 268         allfiles, err := d.Readdirnames(-1)
 269         if err != nil {
 270                 return "", err
 271         }
 272         var tarfiles []string
 273         for _, fnm := range allfiles {
 274                 if strings.HasSuffix(fnm, ".tar") {
 275                         tarfiles = append(tarfiles, fnm)
 276                 }
 277         }
 278         if len(tarfiles) == 0 {
 279                 return "", fmt.Errorf("image collection does not include a .tar image file")
 280         }
 281         if len(tarfiles) > 1 {
 282                 return "", fmt.Errorf("cannot choose from multiple tar files in image collection: %v", tarfiles)
 283         }
 284         imageID := tarfiles[0][:len(tarfiles[0])-4]
 285         imageTarballPath := runner.ArvMountPoint + "/by_id/" + runner.Container.ContainerImage + "/" + imageID + ".tar"
 286         runner.CrunchLog.Printf("Using Docker image id %q", imageID)
 287
 288         runner.CrunchLog.Print("Loading Docker image from keep")
 289         err = runner.executor.LoadImage(imageID, imageTarballPath, runner.Container, runner.ArvMountPoint,
 290                 runner.containerClient)
 291         if err != nil {
 292                 return "", err
 293         }
 294
 295         return imageID, nil
 296 }
 297
 298 func (runner *ContainerRunner) ArvMountCmd(cmdline []string, token string) (c *exec.Cmd, err error) {
 299         c = exec.Command(cmdline[0], cmdline[1:]...)
 300
 301         // Copy our environment, but override ARVADOS_API_TOKEN with
 302         // the container auth token.
 303         c.Env = nil
 304         for _, s := range os.Environ() {
 305                 if !strings.HasPrefix(s, "ARVADOS_API_TOKEN=") {
 306                         c.Env = append(c.Env, s)
 307                 }
 308         }
 309         c.Env = append(c.Env, "ARVADOS_API_TOKEN="+token)
 310
 311         w, err := runner.NewLogWriter("arv-mount")
 312         if err != nil {
 313                 return nil, err
 314         }
 315         runner.arvMountLog = NewThrottledLogger(w)
 316         scanner := logScanner{
 317                 Patterns: []string{
 318                         "Keep write error",
 319                         "Block not found error",
 320                         "Unhandled exception during FUSE operation",
 321                 },
 322                 ReportFunc: runner.reportArvMountWarning,
 323         }
 324         c.Stdout = runner.arvMountLog
 325         c.Stderr = io.MultiWriter(runner.arvMountLog, os.Stderr, &scanner)
 326
 327         runner.CrunchLog.Printf("Running %v", c.Args)
 328
 329         err = c.Start()
 330         if err != nil {
 331                 return nil, err
 332         }
 333
 334         statReadme := make(chan bool)
 335         runner.ArvMountExit = make(chan error)
 336
 337         keepStatting := true
 338         go func() {
 339                 for keepStatting {
 340                         time.Sleep(100 * time.Millisecond)
 341                         _, err = os.Stat(fmt.Sprintf("%s/by_id/README", runner.ArvMountPoint))
 342                         if err == nil {
 343                                 keepStatting = false
 344                                 statReadme <- true
 345                         }
 346                 }
 347                 close(statReadme)
 348         }()
 349
 350         go func() {
 351                 mnterr := c.Wait()
 352                 if mnterr != nil {
 353                         runner.CrunchLog.Printf("Arv-mount exit error: %v", mnterr)
 354                 }
 355                 runner.ArvMountExit <- mnterr
 356                 close(runner.ArvMountExit)
 357         }()
 358
 359         select {
 360         case <-statReadme:
 361                 break
 362         case err := <-runner.ArvMountExit:
 363                 runner.ArvMount = nil
 364                 keepStatting = false
 365                 return nil, err
 366         }
 367
 368         return c, nil
 369 }
 370
 371 func (runner *ContainerRunner) SetupArvMountPoint(prefix string) (err error) {
 372         if runner.ArvMountPoint == "" {
 373                 runner.ArvMountPoint, err = runner.MkTempDir(runner.parentTemp, prefix)
 374         }
 375         return
 376 }
 377
 378 func copyfile(src string, dst string) (err error) {
 379         srcfile, err := os.Open(src)
 380         if err != nil {
 381                 return
 382         }
 383
 384         os.MkdirAll(path.Dir(dst), 0777)
 385
 386         dstfile, err := os.Create(dst)
 387         if err != nil {
 388                 return
 389         }
 390         _, err = io.Copy(dstfile, srcfile)
 391         if err != nil {
 392                 return
 393         }
 394
 395         err = srcfile.Close()
 396         err2 := dstfile.Close()
 397
 398         if err != nil {
 399                 return
 400         }
 401
 402         if err2 != nil {
 403                 return err2
 404         }
 405
 406         return nil
 407 }
 408
 409 func (runner *ContainerRunner) SetupMounts() (map[string]bindmount, error) {
 410         bindmounts := map[string]bindmount{}
 411         err := runner.SetupArvMountPoint("keep")
 412         if err != nil {
 413                 return nil, fmt.Errorf("While creating keep mount temp dir: %v", err)
 414         }
 415
 416         token, err := runner.ContainerToken()
 417         if err != nil {
 418                 return nil, fmt.Errorf("could not get container token: %s", err)
 419         }
 420         runner.CrunchLog.Printf("container token %q", token)
 421
 422         pdhOnly := true
 423         tmpcount := 0
 424         arvMountCmd := []string{
 425                 "arv-mount",
 426                 "--foreground",
 427                 "--read-write",
 428                 "--storage-classes", strings.Join(runner.Container.OutputStorageClasses, ","),
 429                 fmt.Sprintf("--crunchstat-interval=%v", runner.statInterval.Seconds())}
 430
 431         if _, isdocker := runner.executor.(*dockerExecutor); isdocker {
 432                 arvMountCmd = append(arvMountCmd, "--allow-other")
 433         }
 434
 435         if runner.Container.RuntimeConstraints.KeepCacheDisk > 0 {
 436                 keepcachedir, err := runner.MkTempDir(runner.parentTemp, "keepcache")
 437                 if err != nil {
 438                         return nil, fmt.Errorf("while creating keep cache temp dir: %v", err)
 439                 }
 440                 arvMountCmd = append(arvMountCmd, "--disk-cache", "--disk-cache-dir", keepcachedir, "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheDisk))
 441         } else if runner.Container.RuntimeConstraints.KeepCacheRAM > 0 {
 442                 arvMountCmd = append(arvMountCmd, "--ram-cache", "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheRAM))
 443         }
 444
 445         collectionPaths := []string{}
 446         needCertMount := true
 447         type copyFile struct {
 448                 src  string
 449                 bind string
 450         }
 451         var copyFiles []copyFile
 452
 453         var binds []string
 454         for bind := range runner.Container.Mounts {
 455                 binds = append(binds, bind)
 456         }
 457         for bind := range runner.SecretMounts {
 458                 if _, ok := runner.Container.Mounts[bind]; ok {
 459                         return nil, fmt.Errorf("secret mount %q conflicts with regular mount", bind)
 460                 }
 461                 if runner.SecretMounts[bind].Kind != "json" &&
 462                         runner.SecretMounts[bind].Kind != "text" {
 463                         return nil, fmt.Errorf("secret mount %q type is %q but only 'json' and 'text' are permitted",
 464                                 bind, runner.SecretMounts[bind].Kind)
 465                 }
 466                 binds = append(binds, bind)
 467         }
 468         sort.Strings(binds)
 469
 470         for _, bind := range binds {
 471                 mnt, notSecret := runner.Container.Mounts[bind]
 472                 if !notSecret {
 473                         mnt = runner.SecretMounts[bind]
 474                 }
 475                 if bind == "stdout" || bind == "stderr" {
 476                         // Is it a "file" mount kind?
 477                         if mnt.Kind != "file" {
 478                                 return nil, fmt.Errorf("unsupported mount kind '%s' for %s: only 'file' is supported", mnt.Kind, bind)
 479                         }
 480
 481                         // Does path start with OutputPath?
 482                         prefix := runner.Container.OutputPath
 483                         if !strings.HasSuffix(prefix, "/") {
 484                                 prefix += "/"
 485                         }
 486                         if !strings.HasPrefix(mnt.Path, prefix) {
 487                                 return nil, fmt.Errorf("%s path does not start with OutputPath: %s, %s", strings.Title(bind), mnt.Path, prefix)
 488                         }
 489                 }
 490
 491                 if bind == "stdin" {
 492                         // Is it a "collection" mount kind?
 493                         if mnt.Kind != "collection" && mnt.Kind != "json" {
 494                                 return nil, fmt.Errorf("unsupported mount kind '%s' for stdin: only 'collection' and 'json' are supported", mnt.Kind)
 495                         }
 496                 }
 497
 498                 if bind == "/etc/arvados/ca-certificates.crt" {
 499                         needCertMount = false
 500                 }
 501
 502                 if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && bind != runner.Container.OutputPath+"/" {
 503                         if mnt.Kind != "collection" && mnt.Kind != "text" && mnt.Kind != "json" {
 504                                 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)
 505                         }
 506                 }
 507
 508                 switch {
 509                 case mnt.Kind == "collection" && bind != "stdin":
 510                         var src string
 511                         if mnt.UUID != "" && mnt.PortableDataHash != "" {
 512                                 return nil, fmt.Errorf("cannot specify both 'uuid' and 'portable_data_hash' for a collection mount")
 513                         }
 514                         if mnt.UUID != "" {
 515                                 if mnt.Writable {
 516                                         return nil, fmt.Errorf("writing to existing collections currently not permitted")
 517                                 }
 518                                 pdhOnly = false
 519                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.UUID)
 520                         } else if mnt.PortableDataHash != "" {
 521                                 if mnt.Writable && !strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 522                                         return nil, fmt.Errorf("can never write to a collection specified by portable data hash")
 523                                 }
 524                                 idx := strings.Index(mnt.PortableDataHash, "/")
 525                                 if idx > 0 {
 526                                         mnt.Path = path.Clean(mnt.PortableDataHash[idx:])
 527                                         mnt.PortableDataHash = mnt.PortableDataHash[0:idx]
 528                                         runner.Container.Mounts[bind] = mnt
 529                                 }
 530                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.PortableDataHash)
 531                                 if mnt.Path != "" && mnt.Path != "." {
 532                                         if strings.HasPrefix(mnt.Path, "./") {
 533                                                 mnt.Path = mnt.Path[2:]
 534                                         } else if strings.HasPrefix(mnt.Path, "/") {
 535                                                 mnt.Path = mnt.Path[1:]
 536                                         }
 537                                         src += "/" + mnt.Path
 538                                 }
 539                         } else {
 540                                 src = fmt.Sprintf("%s/tmp%d", runner.ArvMountPoint, tmpcount)
 541                                 arvMountCmd = append(arvMountCmd, "--mount-tmp", fmt.Sprintf("tmp%d", tmpcount))
 542                                 tmpcount++
 543                         }
 544                         if mnt.Writable {
 545                                 if bind == runner.Container.OutputPath {
 546                                         runner.HostOutputDir = src
 547                                         bindmounts[bind] = bindmount{HostPath: src}
 548                                 } else if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 549                                         copyFiles = append(copyFiles, copyFile{src, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
 550                                 } else {
 551                                         bindmounts[bind] = bindmount{HostPath: src}
 552                                 }
 553                         } else {
 554                                 bindmounts[bind] = bindmount{HostPath: src, ReadOnly: true}
 555                         }
 556                         collectionPaths = append(collectionPaths, src)
 557
 558                 case mnt.Kind == "tmp":
 559                         var tmpdir string
 560                         tmpdir, err = runner.MkTempDir(runner.parentTemp, "tmp")
 561                         if err != nil {
 562                                 return nil, fmt.Errorf("while creating mount temp dir: %v", err)
 563                         }
 564                         st, staterr := os.Stat(tmpdir)
 565                         if staterr != nil {
 566                                 return nil, fmt.Errorf("while Stat on temp dir: %v", staterr)
 567                         }
 568                         err = os.Chmod(tmpdir, st.Mode()|os.ModeSetgid|0777)
 569                         if staterr != nil {
 570                                 return nil, fmt.Errorf("while Chmod temp dir: %v", err)
 571                         }
 572                         bindmounts[bind] = bindmount{HostPath: tmpdir}
 573                         if bind == runner.Container.OutputPath {
 574                                 runner.HostOutputDir = tmpdir
 575                         }
 576
 577                 case mnt.Kind == "json" || mnt.Kind == "text":
 578                         var filedata []byte
 579                         if mnt.Kind == "json" {
 580                                 filedata, err = json.Marshal(mnt.Content)
 581                                 if err != nil {
 582                                         return nil, fmt.Errorf("encoding json data: %v", err)
 583                                 }
 584                         } else {
 585                                 text, ok := mnt.Content.(string)
 586                                 if !ok {
 587                                         return nil, fmt.Errorf("content for mount %q must be a string", bind)
 588                                 }
 589                                 filedata = []byte(text)
 590                         }
 591
 592                         tmpdir, err := runner.MkTempDir(runner.parentTemp, mnt.Kind)
 593                         if err != nil {
 594                                 return nil, fmt.Errorf("creating temp dir: %v", err)
 595                         }
 596                         tmpfn := filepath.Join(tmpdir, "mountdata."+mnt.Kind)
 597                         err = ioutil.WriteFile(tmpfn, filedata, 0444)
 598                         if err != nil {
 599                                 return nil, fmt.Errorf("writing temp file: %v", err)
 600                         }
 601                         if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && (notSecret || runner.Container.Mounts[runner.Container.OutputPath].Kind != "collection") {
 602                                 // In most cases, if the container
 603                                 // specifies a literal file inside the
 604                                 // output path, we copy it into the
 605                                 // output directory (either a mounted
 606                                 // collection or a staging area on the
 607                                 // host fs). If it's a secret, it will
 608                                 // be skipped when copying output from
 609                                 // staging to Keep later.
 610                                 copyFiles = append(copyFiles, copyFile{tmpfn, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
 611                         } else {
 612                                 // If a secret is outside OutputPath,
 613                                 // we bind mount the secret file
 614                                 // directly just like other mounts. We
 615                                 // also use this strategy when a
 616                                 // secret is inside OutputPath but
 617                                 // OutputPath is a live collection, to
 618                                 // avoid writing the secret to
 619                                 // Keep. Attempting to remove a
 620                                 // bind-mounted secret file from
 621                                 // inside the container will return a
 622                                 // "Device or resource busy" error
 623                                 // that might not be handled well by
 624                                 // the container, which is why we
 625                                 // don't use this strategy when
 626                                 // OutputPath is a staging directory.
 627                                 bindmounts[bind] = bindmount{HostPath: tmpfn, ReadOnly: true}
 628                         }
 629
 630                 case mnt.Kind == "git_tree":
 631                         tmpdir, err := runner.MkTempDir(runner.parentTemp, "git_tree")
 632                         if err != nil {
 633                                 return nil, fmt.Errorf("creating temp dir: %v", err)
 634                         }
 635                         err = gitMount(mnt).extractTree(runner.ContainerArvClient, tmpdir, token)
 636                         if err != nil {
 637                                 return nil, err
 638                         }
 639                         bindmounts[bind] = bindmount{HostPath: tmpdir, ReadOnly: true}
 640                 }
 641         }
 642
 643         if runner.HostOutputDir == "" {
 644                 return nil, fmt.Errorf("output path does not correspond to a writable mount point")
 645         }
 646
 647         if needCertMount && runner.Container.RuntimeConstraints.API {
 648                 for _, certfile := range arvadosclient.CertFiles {
 649                         _, err := os.Stat(certfile)
 650                         if err == nil {
 651                                 bindmounts["/etc/arvados/ca-certificates.crt"] = bindmount{HostPath: certfile, ReadOnly: true}
 652                                 break
 653                         }
 654                 }
 655         }
 656
 657         if pdhOnly {
 658                 // If we are only mounting collections by pdh, make
 659                 // sure we don't subscribe to websocket events to
 660                 // avoid putting undesired load on the API server
 661                 arvMountCmd = append(arvMountCmd, "--mount-by-pdh", "by_id", "--disable-event-listening")
 662         } else {
 663                 arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_id")
 664         }
 665         // the by_uuid mount point is used by singularity when writing
 666         // out docker images converted to SIF
 667         arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_uuid")
 668         arvMountCmd = append(arvMountCmd, runner.ArvMountPoint)
 669
 670         runner.ArvMount, err = runner.RunArvMount(arvMountCmd, token)
 671         if err != nil {
 672                 return nil, fmt.Errorf("while trying to start arv-mount: %v", err)
 673         }
 674         if runner.hoststatReporter != nil && runner.ArvMount != nil {
 675                 runner.hoststatReporter.ReportPID("arv-mount", runner.ArvMount.Process.Pid)
 676         }
 677
 678         for _, p := range collectionPaths {
 679                 _, err = os.Stat(p)
 680                 if err != nil {
 681                         return nil, fmt.Errorf("while checking that input files exist: %v", err)
 682                 }
 683         }
 684
 685         for _, cp := range copyFiles {
 686                 st, err := os.Stat(cp.src)
 687                 if err != nil {
 688                         return nil, fmt.Errorf("while staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 689                 }
 690                 if st.IsDir() {
 691                         err = filepath.Walk(cp.src, func(walkpath string, walkinfo os.FileInfo, walkerr error) error {
 692                                 if walkerr != nil {
 693                                         return walkerr
 694                                 }
 695                                 target := path.Join(cp.bind, walkpath[len(cp.src):])
 696                                 if walkinfo.Mode().IsRegular() {
 697                                         copyerr := copyfile(walkpath, target)
 698                                         if copyerr != nil {
 699                                                 return copyerr
 700                                         }
 701                                         return os.Chmod(target, walkinfo.Mode()|0777)
 702                                 } else if walkinfo.Mode().IsDir() {
 703                                         mkerr := os.MkdirAll(target, 0777)
 704                                         if mkerr != nil {
 705                                                 return mkerr
 706                                         }
 707                                         return os.Chmod(target, walkinfo.Mode()|os.ModeSetgid|0777)
 708                                 } else {
 709                                         return fmt.Errorf("source %q is not a regular file or directory", cp.src)
 710                                 }
 711                         })
 712                 } else if st.Mode().IsRegular() {
 713                         err = copyfile(cp.src, cp.bind)
 714                         if err == nil {
 715                                 err = os.Chmod(cp.bind, st.Mode()|0777)
 716                         }
 717                 }
 718                 if err != nil {
 719                         return nil, fmt.Errorf("while staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 720                 }
 721         }
 722
 723         return bindmounts, nil
 724 }
 725
 726 func (runner *ContainerRunner) stopHoststat() error {
 727         if runner.hoststatReporter == nil {
 728                 return nil
 729         }
 730         runner.hoststatReporter.Stop()
 731         err := runner.hoststatLogger.Close()
 732         if err != nil {
 733                 return fmt.Errorf("error closing hoststat logs: %v", err)
 734         }
 735         return nil
 736 }
 737
 738 func (runner *ContainerRunner) startHoststat() error {
 739         w, err := runner.NewLogWriter("hoststat")
 740         if err != nil {
 741                 return err
 742         }
 743         runner.hoststatLogger = NewThrottledLogger(w)
 744         runner.hoststatReporter = &crunchstat.Reporter{
 745                 Logger:     log.New(runner.hoststatLogger, "", 0),
 746                 CgroupRoot: runner.cgroupRoot,
 747                 PollPeriod: runner.statInterval,
 748         }
 749         runner.hoststatReporter.Start()
 750         runner.hoststatReporter.ReportPID("crunch-run", os.Getpid())
 751         return nil
 752 }
 753
 754 func (runner *ContainerRunner) startCrunchstat() error {
 755         w, err := runner.NewLogWriter("crunchstat")
 756         if err != nil {
 757                 return err
 758         }
 759         runner.statLogger = NewThrottledLogger(w)
 760         runner.statReporter = &crunchstat.Reporter{
 761                 CID:          runner.executor.CgroupID(),
 762                 Logger:       log.New(runner.statLogger, "", 0),
 763                 CgroupParent: runner.expectCgroupParent,
 764                 CgroupRoot:   runner.cgroupRoot,
 765                 PollPeriod:   runner.statInterval,
 766                 TempDir:      runner.parentTemp,
 767         }
 768         runner.statReporter.Start()
 769         return nil
 770 }
 771
 772 type infoCommand struct {
 773         label string
 774         cmd   []string
 775 }
 776
 777 // LogHostInfo logs info about the current host, for debugging and
 778 // accounting purposes. Although it's logged as "node-info", this is
 779 // about the environment where crunch-run is actually running, which
 780 // might differ from what's described in the node record (see
 781 // LogNodeRecord).
 782 func (runner *ContainerRunner) LogHostInfo() (err error) {
 783         w, err := runner.NewLogWriter("node-info")
 784         if err != nil {
 785                 return
 786         }
 787
 788         commands := []infoCommand{
 789                 {
 790                         label: "Host Information",
 791                         cmd:   []string{"uname", "-a"},
 792                 },
 793                 {
 794                         label: "CPU Information",
 795                         cmd:   []string{"cat", "/proc/cpuinfo"},
 796                 },
 797                 {
 798                         label: "Memory Information",
 799                         cmd:   []string{"cat", "/proc/meminfo"},
 800                 },
 801                 {
 802                         label: "Disk Space",
 803                         cmd:   []string{"df", "-m", "/", os.TempDir()},
 804                 },
 805                 {
 806                         label: "Disk INodes",
 807                         cmd:   []string{"df", "-i", "/", os.TempDir()},
 808                 },
 809         }
 810
 811         // Run commands with informational output to be logged.
 812         for _, command := range commands {
 813                 fmt.Fprintln(w, command.label)
 814                 cmd := exec.Command(command.cmd[0], command.cmd[1:]...)
 815                 cmd.Stdout = w
 816                 cmd.Stderr = w
 817                 if err := cmd.Run(); err != nil {
 818                         err = fmt.Errorf("While running command %q: %v", command.cmd, err)
 819                         fmt.Fprintln(w, err)
 820                         return err
 821                 }
 822                 fmt.Fprintln(w, "")
 823         }
 824
 825         err = w.Close()
 826         if err != nil {
 827                 return fmt.Errorf("While closing node-info logs: %v", err)
 828         }
 829         return nil
 830 }
 831
 832 // LogContainerRecord gets and saves the raw JSON container record from the API server
 833 func (runner *ContainerRunner) LogContainerRecord() error {
 834         logged, err := runner.logAPIResponse("container", "containers", map[string]interface{}{"filters": [][]string{{"uuid", "=", runner.Container.UUID}}}, nil)
 835         if !logged && err == nil {
 836                 err = fmt.Errorf("error: no container record found for %s", runner.Container.UUID)
 837         }
 838         return err
 839 }
 840
 841 // LogNodeRecord logs the current host's InstanceType config entry (or
 842 // the arvados#node record, if running via crunch-dispatch-slurm).
 843 func (runner *ContainerRunner) LogNodeRecord() error {
 844         if it := os.Getenv("InstanceType"); it != "" {
 845                 // Dispatched via arvados-dispatch-cloud. Save
 846                 // InstanceType config fragment received from
 847                 // dispatcher on stdin.
 848                 w, err := runner.LogCollection.OpenFile("node.json", os.O_CREATE|os.O_WRONLY, 0666)
 849                 if err != nil {
 850                         return err
 851                 }
 852                 defer w.Close()
 853                 _, err = io.WriteString(w, it)
 854                 if err != nil {
 855                         return err
 856                 }
 857                 return w.Close()
 858         }
 859         // Dispatched via crunch-dispatch-slurm. Look up
 860         // apiserver's node record corresponding to
 861         // $SLURMD_NODENAME.
 862         hostname := os.Getenv("SLURMD_NODENAME")
 863         if hostname == "" {
 864                 hostname, _ = os.Hostname()
 865         }
 866         _, err := runner.logAPIResponse("node", "nodes", map[string]interface{}{"filters": [][]string{{"hostname", "=", hostname}}}, func(resp interface{}) {
 867                 // The "info" field has admin-only info when
 868                 // obtained with a privileged token, and
 869                 // should not be logged.
 870                 node, ok := resp.(map[string]interface{})
 871                 if ok {
 872                         delete(node, "info")
 873                 }
 874         })
 875         return err
 876 }
 877
 878 func (runner *ContainerRunner) logAPIResponse(label, path string, params map[string]interface{}, munge func(interface{})) (logged bool, err error) {
 879         writer, err := runner.LogCollection.OpenFile(label+".json", os.O_CREATE|os.O_WRONLY, 0666)
 880         if err != nil {
 881                 return false, err
 882         }
 883         w := &ArvLogWriter{
 884                 ArvClient:     runner.DispatcherArvClient,
 885                 UUID:          runner.Container.UUID,
 886                 loggingStream: label,
 887                 writeCloser:   writer,
 888         }
 889
 890         reader, err := runner.DispatcherArvClient.CallRaw("GET", path, "", "", arvadosclient.Dict(params))
 891         if err != nil {
 892                 return false, fmt.Errorf("error getting %s record: %v", label, err)
 893         }
 894         defer reader.Close()
 895
 896         dec := json.NewDecoder(reader)
 897         dec.UseNumber()
 898         var resp map[string]interface{}
 899         if err = dec.Decode(&resp); err != nil {
 900                 return false, fmt.Errorf("error decoding %s list response: %v", label, err)
 901         }
 902         items, ok := resp["items"].([]interface{})
 903         if !ok {
 904                 return false, fmt.Errorf("error decoding %s list response: no \"items\" key in API list response", label)
 905         } else if len(items) < 1 {
 906                 return false, nil
 907         }
 908         if munge != nil {
 909                 munge(items[0])
 910         }
 911         // Re-encode it using indentation to improve readability
 912         enc := json.NewEncoder(w)
 913         enc.SetIndent("", "    ")
 914         if err = enc.Encode(items[0]); err != nil {
 915                 return false, fmt.Errorf("error logging %s record: %v", label, err)
 916         }
 917         err = w.Close()
 918         if err != nil {
 919                 return false, fmt.Errorf("error closing %s.json in log collection: %v", label, err)
 920         }
 921         return true, nil
 922 }
 923
 924 func (runner *ContainerRunner) getStdoutFile(mntPath string) (*os.File, error) {
 925         stdoutPath := mntPath[len(runner.Container.OutputPath):]
 926         index := strings.LastIndex(stdoutPath, "/")
 927         if index > 0 {
 928                 subdirs := stdoutPath[:index]
 929                 if subdirs != "" {
 930                         st, err := os.Stat(runner.HostOutputDir)
 931                         if err != nil {
 932                                 return nil, fmt.Errorf("While Stat on temp dir: %v", err)
 933                         }
 934                         stdoutPath := filepath.Join(runner.HostOutputDir, subdirs)
 935                         err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777)
 936                         if err != nil {
 937                                 return nil, fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err)
 938                         }
 939                 }
 940         }
 941         stdoutFile, err := os.Create(filepath.Join(runner.HostOutputDir, stdoutPath))
 942         if err != nil {
 943                 return nil, fmt.Errorf("While creating file %q: %v", stdoutPath, err)
 944         }
 945
 946         return stdoutFile, nil
 947 }
 948
 949 // CreateContainer creates the docker container.
 950 func (runner *ContainerRunner) CreateContainer(imageID string, bindmounts map[string]bindmount) error {
 951         var stdin io.ReadCloser = ioutil.NopCloser(bytes.NewReader(nil))
 952         if mnt, ok := runner.Container.Mounts["stdin"]; ok {
 953                 switch mnt.Kind {
 954                 case "collection":
 955                         var collID string
 956                         if mnt.UUID != "" {
 957                                 collID = mnt.UUID
 958                         } else {
 959                                 collID = mnt.PortableDataHash
 960                         }
 961                         path := runner.ArvMountPoint + "/by_id/" + collID + "/" + mnt.Path
 962                         f, err := os.Open(path)
 963                         if err != nil {
 964                                 return err
 965                         }
 966                         stdin = f
 967                 case "json":
 968                         j, err := json.Marshal(mnt.Content)
 969                         if err != nil {
 970                                 return fmt.Errorf("error encoding stdin json data: %v", err)
 971                         }
 972                         stdin = ioutil.NopCloser(bytes.NewReader(j))
 973                 default:
 974                         return fmt.Errorf("stdin mount has unsupported kind %q", mnt.Kind)
 975                 }
 976         }
 977
 978         var stdout, stderr io.WriteCloser
 979         if mnt, ok := runner.Container.Mounts["stdout"]; ok {
 980                 f, err := runner.getStdoutFile(mnt.Path)
 981                 if err != nil {
 982                         return err
 983                 }
 984                 stdout = f
 985         } else if w, err := runner.NewLogWriter("stdout"); err != nil {
 986                 return err
 987         } else {
 988                 stdout = NewThrottledLogger(w)
 989         }
 990
 991         if mnt, ok := runner.Container.Mounts["stderr"]; ok {
 992                 f, err := runner.getStdoutFile(mnt.Path)
 993                 if err != nil {
 994                         return err
 995                 }
 996                 stderr = f
 997         } else if w, err := runner.NewLogWriter("stderr"); err != nil {
 998                 return err
 999         } else {
1000                 stderr = NewThrottledLogger(w)
1001         }
1002
1003         env := runner.Container.Environment
1004         enableNetwork := runner.enableNetwork == "always"
1005         if runner.Container.RuntimeConstraints.API {
1006                 enableNetwork = true
1007                 tok, err := runner.ContainerToken()
1008                 if err != nil {
1009                         return err
1010                 }
1011                 env = map[string]string{}
1012                 for k, v := range runner.Container.Environment {
1013                         env[k] = v
1014                 }
1015                 env["ARVADOS_API_TOKEN"] = tok
1016                 env["ARVADOS_API_HOST"] = os.Getenv("ARVADOS_API_HOST")
1017                 env["ARVADOS_API_HOST_INSECURE"] = os.Getenv("ARVADOS_API_HOST_INSECURE")
1018                 env["ARVADOS_KEEP_SERVICES"] = os.Getenv("ARVADOS_KEEP_SERVICES")
1019         }
1020         workdir := runner.Container.Cwd
1021         if workdir == "." {
1022                 // both "" and "." mean default
1023                 workdir = ""
1024         }
1025         ram := runner.Container.RuntimeConstraints.RAM
1026         if !runner.enableMemoryLimit {
1027                 ram = 0
1028         }
1029         runner.executorStdin = stdin
1030         runner.executorStdout = stdout
1031         runner.executorStderr = stderr
1032
1033         if runner.Container.RuntimeConstraints.CUDA.DeviceCount > 0 {
1034                 nvidiaModprobe(runner.CrunchLog)
1035         }
1036
1037         return runner.executor.Create(containerSpec{
1038                 Image:           imageID,
1039                 VCPUs:           runner.Container.RuntimeConstraints.VCPUs,
1040                 RAM:             ram,
1041                 WorkingDir:      workdir,
1042                 Env:             env,
1043                 BindMounts:      bindmounts,
1044                 Command:         runner.Container.Command,
1045                 EnableNetwork:   enableNetwork,
1046                 CUDADeviceCount: runner.Container.RuntimeConstraints.CUDA.DeviceCount,
1047                 NetworkMode:     runner.networkMode,
1048                 CgroupParent:    runner.setCgroupParent,
1049                 Stdin:           stdin,
1050                 Stdout:          stdout,
1051                 Stderr:          stderr,
1052         })
1053 }
1054
1055 // StartContainer starts the docker container created by CreateContainer.
1056 func (runner *ContainerRunner) StartContainer() error {
1057         runner.CrunchLog.Printf("Starting container")
1058         runner.cStateLock.Lock()
1059         defer runner.cStateLock.Unlock()
1060         if runner.cCancelled {
1061                 return ErrCancelled
1062         }
1063         err := runner.executor.Start()
1064         if err != nil {
1065                 var advice string
1066                 if m, e := regexp.MatchString("(?ms).*(exec|System error).*(no such file or directory|file not found).*", err.Error()); m && e == nil {
1067                         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])
1068                 }
1069                 return fmt.Errorf("could not start container: %v%s", err, advice)
1070         }
1071         return nil
1072 }
1073
1074 // WaitFinish waits for the container to terminate, capture the exit code, and
1075 // close the stdout/stderr logging.
1076 func (runner *ContainerRunner) WaitFinish() error {
1077         runner.CrunchLog.Print("Waiting for container to finish")
1078         var timeout <-chan time.Time
1079         if s := runner.Container.SchedulingParameters.MaxRunTime; s > 0 {
1080                 timeout = time.After(time.Duration(s) * time.Second)
1081         }
1082         ctx, cancel := context.WithCancel(context.Background())
1083         defer cancel()
1084         go func() {
1085                 select {
1086                 case <-timeout:
1087                         runner.CrunchLog.Printf("maximum run time exceeded. Stopping container.")
1088                         runner.stop(nil)
1089                 case <-runner.ArvMountExit:
1090                         runner.CrunchLog.Printf("arv-mount exited while container is still running. Stopping container.")
1091                         runner.stop(nil)
1092                 case <-ctx.Done():
1093                 }
1094         }()
1095         exitcode, err := runner.executor.Wait(ctx)
1096         if err != nil {
1097                 runner.checkBrokenNode(err)
1098                 return err
1099         }
1100         runner.ExitCode = &exitcode
1101
1102         extra := ""
1103         if exitcode&0x80 != 0 {
1104                 // Convert raw exit status (0x80 + signal number) to a
1105                 // string to log after the code, like " (signal 101)"
1106                 // or " (signal 9, killed)"
1107                 sig := syscall.WaitStatus(exitcode).Signal()
1108                 if name := unix.SignalName(sig); name != "" {
1109                         extra = fmt.Sprintf(" (signal %d, %s)", sig, name)
1110                 } else {
1111                         extra = fmt.Sprintf(" (signal %d)", sig)
1112                 }
1113         }
1114         runner.CrunchLog.Printf("Container exited with status code %d%s", exitcode, extra)
1115         err = runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1116                 "container": arvadosclient.Dict{"exit_code": exitcode},
1117         }, nil)
1118         if err != nil {
1119                 runner.CrunchLog.Printf("ignoring error updating exit_code: %s", err)
1120         }
1121
1122         var returnErr error
1123         if err = runner.executorStdin.Close(); err != nil {
1124                 err = fmt.Errorf("error closing container stdin: %s", err)
1125                 runner.CrunchLog.Printf("%s", err)
1126                 returnErr = err
1127         }
1128         if err = runner.executorStdout.Close(); err != nil {
1129                 err = fmt.Errorf("error closing container stdout: %s", err)
1130                 runner.CrunchLog.Printf("%s", err)
1131                 if returnErr == nil {
1132                         returnErr = err
1133                 }
1134         }
1135         if err = runner.executorStderr.Close(); err != nil {
1136                 err = fmt.Errorf("error closing container stderr: %s", err)
1137                 runner.CrunchLog.Printf("%s", err)
1138                 if returnErr == nil {
1139                         returnErr = err
1140                 }
1141         }
1142
1143         if runner.statReporter != nil {
1144                 runner.statReporter.Stop()
1145                 err = runner.statLogger.Close()
1146                 if err != nil {
1147                         runner.CrunchLog.Printf("error closing crunchstat logs: %v", err)
1148                 }
1149         }
1150         return returnErr
1151 }
1152
1153 func (runner *ContainerRunner) updateLogs() {
1154         ticker := time.NewTicker(crunchLogUpdatePeriod / 360)
1155         defer ticker.Stop()
1156
1157         sigusr1 := make(chan os.Signal, 1)
1158         signal.Notify(sigusr1, syscall.SIGUSR1)
1159         defer signal.Stop(sigusr1)
1160
1161         saveAtTime := time.Now().Add(crunchLogUpdatePeriod)
1162         saveAtSize := crunchLogUpdateSize
1163         var savedSize int64
1164         for {
1165                 select {
1166                 case <-ticker.C:
1167                 case <-sigusr1:
1168                         saveAtTime = time.Now()
1169                 }
1170                 runner.logMtx.Lock()
1171                 done := runner.LogsPDH != nil
1172                 runner.logMtx.Unlock()
1173                 if done {
1174                         return
1175                 }
1176                 size := runner.LogCollection.Size()
1177                 if size == savedSize || (time.Now().Before(saveAtTime) && size < saveAtSize) {
1178                         continue
1179                 }
1180                 saveAtTime = time.Now().Add(crunchLogUpdatePeriod)
1181                 saveAtSize = runner.LogCollection.Size() + crunchLogUpdateSize
1182                 saved, err := runner.saveLogCollection(false)
1183                 if err != nil {
1184                         runner.CrunchLog.Printf("error updating log collection: %s", err)
1185                         continue
1186                 }
1187
1188                 err = runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1189                         "container": arvadosclient.Dict{"log": saved.PortableDataHash},
1190                 }, nil)
1191                 if err != nil {
1192                         runner.CrunchLog.Printf("error updating container log to %s: %s", saved.PortableDataHash, err)
1193                         continue
1194                 }
1195
1196                 savedSize = size
1197         }
1198 }
1199
1200 func (runner *ContainerRunner) reportArvMountWarning(pattern, text string) {
1201         var updated arvados.Container
1202         err := runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1203                 "container": arvadosclient.Dict{
1204                         "runtime_status": arvadosclient.Dict{
1205                                 "warning":       "arv-mount: " + pattern,
1206                                 "warningDetail": text,
1207                         },
1208                 },
1209         }, &updated)
1210         if err != nil {
1211                 runner.CrunchLog.Printf("error updating container runtime_status: %s", err)
1212         }
1213 }
1214
1215 // CaptureOutput saves data from the container's output directory if
1216 // needed, and updates the container output accordingly.
1217 func (runner *ContainerRunner) CaptureOutput(bindmounts map[string]bindmount) error {
1218         if runner.Container.RuntimeConstraints.API {
1219                 // Output may have been set directly by the container, so
1220                 // refresh the container record to check.
1221                 err := runner.DispatcherArvClient.Get("containers", runner.Container.UUID,
1222                         nil, &runner.Container)
1223                 if err != nil {
1224                         return err
1225                 }
1226                 if runner.Container.Output != "" {
1227                         // Container output is already set.
1228                         runner.OutputPDH = &runner.Container.Output
1229                         return nil
1230                 }
1231         }
1232
1233         txt, err := (&copier{
1234                 client:        runner.containerClient,
1235                 arvClient:     runner.ContainerArvClient,
1236                 keepClient:    runner.ContainerKeepClient,
1237                 hostOutputDir: runner.HostOutputDir,
1238                 ctrOutputDir:  runner.Container.OutputPath,
1239                 bindmounts:    bindmounts,
1240                 mounts:        runner.Container.Mounts,
1241                 secretMounts:  runner.SecretMounts,
1242                 logger:        runner.CrunchLog,
1243         }).Copy()
1244         if err != nil {
1245                 return err
1246         }
1247         if n := len(regexp.MustCompile(` [0-9a-f]+\+\S*\+R`).FindAllStringIndex(txt, -1)); n > 0 {
1248                 runner.CrunchLog.Printf("Copying %d data blocks from remote input collections...", n)
1249                 fs, err := (&arvados.Collection{ManifestText: txt}).FileSystem(runner.containerClient, runner.ContainerKeepClient)
1250                 if err != nil {
1251                         return err
1252                 }
1253                 txt, err = fs.MarshalManifest(".")
1254                 if err != nil {
1255                         return err
1256                 }
1257         }
1258         var resp arvados.Collection
1259         err = runner.ContainerArvClient.Create("collections", arvadosclient.Dict{
1260                 "ensure_unique_name": true,
1261                 "collection": arvadosclient.Dict{
1262                         "is_trashed":    true,
1263                         "name":          "output for " + runner.Container.UUID,
1264                         "manifest_text": txt,
1265                 },
1266         }, &resp)
1267         if err != nil {
1268                 return fmt.Errorf("error creating output collection: %v", err)
1269         }
1270         runner.OutputPDH = &resp.PortableDataHash
1271         return nil
1272 }
1273
1274 func (runner *ContainerRunner) CleanupDirs() {
1275         if runner.ArvMount != nil {
1276                 var delay int64 = 8
1277                 umount := exec.Command("arv-mount", fmt.Sprintf("--unmount-timeout=%d", delay), "--unmount", runner.ArvMountPoint)
1278                 umount.Stdout = runner.CrunchLog
1279                 umount.Stderr = runner.CrunchLog
1280                 runner.CrunchLog.Printf("Running %v", umount.Args)
1281                 umnterr := umount.Start()
1282
1283                 if umnterr != nil {
1284                         runner.CrunchLog.Printf("Error unmounting: %v", umnterr)
1285                         runner.ArvMount.Process.Kill()
1286                 } else {
1287                         // If arv-mount --unmount gets stuck for any reason, we
1288                         // don't want to wait for it forever.  Do Wait() in a goroutine
1289                         // so it doesn't block crunch-run.
1290                         umountExit := make(chan error)
1291                         go func() {
1292                                 mnterr := umount.Wait()
1293                                 if mnterr != nil {
1294                                         runner.CrunchLog.Printf("Error unmounting: %v", mnterr)
1295                                 }
1296                                 umountExit <- mnterr
1297                         }()
1298
1299                         for again := true; again; {
1300                                 again = false
1301                                 select {
1302                                 case <-umountExit:
1303                                         umount = nil
1304                                         again = true
1305                                 case <-runner.ArvMountExit:
1306                                         break
1307                                 case <-time.After(time.Duration((delay + 1) * int64(time.Second))):
1308                                         runner.CrunchLog.Printf("Timed out waiting for unmount")
1309                                         if umount != nil {
1310                                                 umount.Process.Kill()
1311                                         }
1312                                         runner.ArvMount.Process.Kill()
1313                                 }
1314                         }
1315                 }
1316                 runner.ArvMount = nil
1317         }
1318
1319         if runner.ArvMountPoint != "" {
1320                 if rmerr := os.Remove(runner.ArvMountPoint); rmerr != nil {
1321                         runner.CrunchLog.Printf("While cleaning up arv-mount directory %s: %v", runner.ArvMountPoint, rmerr)
1322                 }
1323                 runner.ArvMountPoint = ""
1324         }
1325
1326         if rmerr := os.RemoveAll(runner.parentTemp); rmerr != nil {
1327                 runner.CrunchLog.Printf("While cleaning up temporary directory %s: %v", runner.parentTemp, rmerr)
1328         }
1329 }
1330
1331 // CommitLogs posts the collection containing the final container logs.
1332 func (runner *ContainerRunner) CommitLogs() error {
1333         func() {
1334                 // Hold cStateLock to prevent races on CrunchLog (e.g., stop()).
1335                 runner.cStateLock.Lock()
1336                 defer runner.cStateLock.Unlock()
1337
1338                 runner.CrunchLog.Print(runner.finalState)
1339
1340                 if runner.arvMountLog != nil {
1341                         runner.arvMountLog.Close()
1342                 }
1343                 runner.CrunchLog.Close()
1344
1345                 // Closing CrunchLog above allows them to be committed to Keep at this
1346                 // point, but re-open crunch log with ArvClient in case there are any
1347                 // other further errors (such as failing to write the log to Keep!)
1348                 // while shutting down
1349                 runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{
1350                         ArvClient:     runner.DispatcherArvClient,
1351                         UUID:          runner.Container.UUID,
1352                         loggingStream: "crunch-run",
1353                         writeCloser:   nil,
1354                 })
1355                 runner.CrunchLog.Immediate = log.New(os.Stderr, runner.Container.UUID+" ", 0)
1356         }()
1357
1358         if runner.keepstoreLogger != nil {
1359                 // Flush any buffered logs from our local keepstore
1360                 // process.  Discard anything logged after this point
1361                 // -- it won't end up in the log collection, so
1362                 // there's no point writing it to the collectionfs.
1363                 runner.keepstoreLogbuf.SetWriter(io.Discard)
1364                 runner.keepstoreLogger.Close()
1365                 runner.keepstoreLogger = nil
1366         }
1367
1368         if runner.LogsPDH != nil {
1369                 // If we have already assigned something to LogsPDH,
1370                 // we must be closing the re-opened log, which won't
1371                 // end up getting attached to the container record and
1372                 // therefore doesn't need to be saved as a collection
1373                 // -- it exists only to send logs to other channels.
1374                 return nil
1375         }
1376
1377         saved, err := runner.saveLogCollection(true)
1378         if err != nil {
1379                 return fmt.Errorf("error saving log collection: %s", err)
1380         }
1381         runner.logMtx.Lock()
1382         defer runner.logMtx.Unlock()
1383         runner.LogsPDH = &saved.PortableDataHash
1384         return nil
1385 }
1386
1387 func (runner *ContainerRunner) saveLogCollection(final bool) (response arvados.Collection, err error) {
1388         runner.logMtx.Lock()
1389         defer runner.logMtx.Unlock()
1390         if runner.LogsPDH != nil {
1391                 // Already finalized.
1392                 return
1393         }
1394         updates := arvadosclient.Dict{
1395                 "name": "logs for " + runner.Container.UUID,
1396         }
1397         mt, err1 := runner.LogCollection.MarshalManifest(".")
1398         if err1 == nil {
1399                 // Only send updated manifest text if there was no
1400                 // error.
1401                 updates["manifest_text"] = mt
1402         }
1403
1404         // Even if flushing the manifest had an error, we still want
1405         // to update the log record, if possible, to push the trash_at
1406         // and delete_at times into the future.  Details on bug
1407         // #17293.
1408         if final {
1409                 updates["is_trashed"] = true
1410         } else {
1411                 exp := time.Now().Add(crunchLogUpdatePeriod * 24)
1412                 updates["trash_at"] = exp
1413                 updates["delete_at"] = exp
1414         }
1415         reqBody := arvadosclient.Dict{"collection": updates}
1416         var err2 error
1417         if runner.logUUID == "" {
1418                 reqBody["ensure_unique_name"] = true
1419                 err2 = runner.DispatcherArvClient.Create("collections", reqBody, &response)
1420         } else {
1421                 err2 = runner.DispatcherArvClient.Update("collections", runner.logUUID, reqBody, &response)
1422         }
1423         if err2 == nil {
1424                 runner.logUUID = response.UUID
1425         }
1426
1427         if err1 != nil || err2 != nil {
1428                 err = fmt.Errorf("error recording logs: %q, %q", err1, err2)
1429         }
1430         return
1431 }
1432
1433 // UpdateContainerRunning updates the container state to "Running"
1434 func (runner *ContainerRunner) UpdateContainerRunning(logId string) error {
1435         runner.cStateLock.Lock()
1436         defer runner.cStateLock.Unlock()
1437         if runner.cCancelled {
1438                 return ErrCancelled
1439         }
1440         updates := arvadosclient.Dict{
1441                 "gateway_address": runner.gateway.Address,
1442                 "state":           "Running",
1443         }
1444         if logId != "" {
1445                 updates["log"] = logId
1446         }
1447         return runner.DispatcherArvClient.Update(
1448                 "containers",
1449                 runner.Container.UUID,
1450                 arvadosclient.Dict{"container": updates},
1451                 nil,
1452         )
1453 }
1454
1455 // ContainerToken returns the api_token the container (and any
1456 // arv-mount processes) are allowed to use.
1457 func (runner *ContainerRunner) ContainerToken() (string, error) {
1458         if runner.token != "" {
1459                 return runner.token, nil
1460         }
1461
1462         var auth arvados.APIClientAuthorization
1463         err := runner.DispatcherArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth)
1464         if err != nil {
1465                 return "", err
1466         }
1467         runner.token = fmt.Sprintf("v2/%s/%s/%s", auth.UUID, auth.APIToken, runner.Container.UUID)
1468         return runner.token, nil
1469 }
1470
1471 // UpdateContainerFinal updates the container record state on API
1472 // server to "Complete" or "Cancelled"
1473 func (runner *ContainerRunner) UpdateContainerFinal() error {
1474         update := arvadosclient.Dict{}
1475         update["state"] = runner.finalState
1476         if runner.LogsPDH != nil {
1477                 update["log"] = *runner.LogsPDH
1478         }
1479         if runner.ExitCode != nil {
1480                 update["exit_code"] = *runner.ExitCode
1481         } else {
1482                 update["exit_code"] = nil
1483         }
1484         if runner.finalState == "Complete" && runner.OutputPDH != nil {
1485                 update["output"] = *runner.OutputPDH
1486         }
1487         update["cost"] = runner.calculateCost(time.Now())
1488         return runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{"container": update}, nil)
1489 }
1490
1491 // IsCancelled returns the value of Cancelled, with goroutine safety.
1492 func (runner *ContainerRunner) IsCancelled() bool {
1493         runner.cStateLock.Lock()
1494         defer runner.cStateLock.Unlock()
1495         return runner.cCancelled
1496 }
1497
1498 // NewArvLogWriter creates an ArvLogWriter
1499 func (runner *ContainerRunner) NewArvLogWriter(name string) (io.WriteCloser, error) {
1500         writer, err := runner.LogCollection.OpenFile(name+".txt", os.O_CREATE|os.O_WRONLY, 0666)
1501         if err != nil {
1502                 return nil, err
1503         }
1504         return &ArvLogWriter{
1505                 ArvClient:     runner.DispatcherArvClient,
1506                 UUID:          runner.Container.UUID,
1507                 loggingStream: name,
1508                 writeCloser:   writer,
1509         }, nil
1510 }
1511
1512 // Run the full container lifecycle.
1513 func (runner *ContainerRunner) Run() (err error) {
1514         runner.CrunchLog.Printf("crunch-run %s started", cmd.Version.String())
1515         runner.CrunchLog.Printf("%s", currentUserAndGroups())
1516         v, _ := exec.Command("arv-mount", "--version").CombinedOutput()
1517         runner.CrunchLog.Printf("Using FUSE mount: %s", v)
1518         runner.CrunchLog.Printf("Using container runtime: %s", runner.executor.Runtime())
1519         runner.CrunchLog.Printf("Executing container: %s", runner.Container.UUID)
1520         runner.costStartTime = time.Now()
1521
1522         hostname, hosterr := os.Hostname()
1523         if hosterr != nil {
1524                 runner.CrunchLog.Printf("Error getting hostname '%v'", hosterr)
1525         } else {
1526                 runner.CrunchLog.Printf("Executing on host '%s'", hostname)
1527         }
1528
1529         sigusr2 := make(chan os.Signal, 1)
1530         signal.Notify(sigusr2, syscall.SIGUSR2)
1531         defer signal.Stop(sigusr2)
1532         runner.loadPrices()
1533         go func() {
1534                 for range sigusr2 {
1535                         runner.loadPrices()
1536                 }
1537         }()
1538
1539         runner.finalState = "Queued"
1540
1541         defer func() {
1542                 runner.CleanupDirs()
1543
1544                 runner.CrunchLog.Printf("crunch-run finished")
1545                 runner.CrunchLog.Close()
1546         }()
1547
1548         err = runner.fetchContainerRecord()
1549         if err != nil {
1550                 return
1551         }
1552         if runner.Container.State != "Locked" {
1553                 return fmt.Errorf("dispatch error detected: container %q has state %q", runner.Container.UUID, runner.Container.State)
1554         }
1555
1556         var bindmounts map[string]bindmount
1557         defer func() {
1558                 // checkErr prints e (unless it's nil) and sets err to
1559                 // e (unless err is already non-nil). Thus, if err
1560                 // hasn't already been assigned when Run() returns,
1561                 // this cleanup func will cause Run() to return the
1562                 // first non-nil error that is passed to checkErr().
1563                 checkErr := func(errorIn string, e error) {
1564                         if e == nil {
1565                                 return
1566                         }
1567                         runner.CrunchLog.Printf("error in %s: %v", errorIn, e)
1568                         if err == nil {
1569                                 err = e
1570                         }
1571                         if runner.finalState == "Complete" {
1572                                 // There was an error in the finalization.
1573                                 runner.finalState = "Cancelled"
1574                         }
1575                 }
1576
1577                 // Log the error encountered in Run(), if any
1578                 checkErr("Run", err)
1579
1580                 if runner.finalState == "Queued" {
1581                         runner.UpdateContainerFinal()
1582                         return
1583                 }
1584
1585                 if runner.IsCancelled() {
1586                         runner.finalState = "Cancelled"
1587                         // but don't return yet -- we still want to
1588                         // capture partial output and write logs
1589                 }
1590
1591                 if bindmounts != nil {
1592                         checkErr("CaptureOutput", runner.CaptureOutput(bindmounts))
1593                 }
1594                 checkErr("stopHoststat", runner.stopHoststat())
1595                 checkErr("CommitLogs", runner.CommitLogs())
1596                 runner.CleanupDirs()
1597                 checkErr("UpdateContainerFinal", runner.UpdateContainerFinal())
1598         }()
1599
1600         runner.setupSignals()
1601         err = runner.startHoststat()
1602         if err != nil {
1603                 return
1604         }
1605         if runner.keepstore != nil {
1606                 runner.hoststatReporter.ReportPID("keepstore", runner.keepstore.Process.Pid)
1607         }
1608
1609         // set up FUSE mount and binds
1610         bindmounts, err = runner.SetupMounts()
1611         if err != nil {
1612                 runner.finalState = "Cancelled"
1613                 err = fmt.Errorf("While setting up mounts: %v", err)
1614                 return
1615         }
1616
1617         // check for and/or load image
1618         imageID, err := runner.LoadImage()
1619         if err != nil {
1620                 if !runner.checkBrokenNode(err) {
1621                         // Failed to load image but not due to a "broken node"
1622                         // condition, probably user error.
1623                         runner.finalState = "Cancelled"
1624                 }
1625                 err = fmt.Errorf("While loading container image: %v", err)
1626                 return
1627         }
1628
1629         err = runner.CreateContainer(imageID, bindmounts)
1630         if err != nil {
1631                 return
1632         }
1633         err = runner.LogHostInfo()
1634         if err != nil {
1635                 return
1636         }
1637         err = runner.LogNodeRecord()
1638         if err != nil {
1639                 return
1640         }
1641         err = runner.LogContainerRecord()
1642         if err != nil {
1643                 return
1644         }
1645
1646         if runner.IsCancelled() {
1647                 return
1648         }
1649
1650         logCollection, err := runner.saveLogCollection(false)
1651         var logId string
1652         if err == nil {
1653                 logId = logCollection.PortableDataHash
1654         } else {
1655                 runner.CrunchLog.Printf("Error committing initial log collection: %v", err)
1656         }
1657         err = runner.UpdateContainerRunning(logId)
1658         if err != nil {
1659                 return
1660         }
1661         runner.finalState = "Cancelled"
1662
1663         err = runner.startCrunchstat()
1664         if err != nil {
1665                 return
1666         }
1667
1668         err = runner.StartContainer()
1669         if err != nil {
1670                 runner.checkBrokenNode(err)
1671                 return
1672         }
1673
1674         err = runner.WaitFinish()
1675         if err == nil && !runner.IsCancelled() {
1676                 runner.finalState = "Complete"
1677         }
1678         return
1679 }
1680
1681 // Fetch the current container record (uuid = runner.Container.UUID)
1682 // into runner.Container.
1683 func (runner *ContainerRunner) fetchContainerRecord() error {
1684         reader, err := runner.DispatcherArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
1685         if err != nil {
1686                 return fmt.Errorf("error fetching container record: %v", err)
1687         }
1688         defer reader.Close()
1689
1690         dec := json.NewDecoder(reader)
1691         dec.UseNumber()
1692         err = dec.Decode(&runner.Container)
1693         if err != nil {
1694                 return fmt.Errorf("error decoding container record: %v", err)
1695         }
1696
1697         var sm struct {
1698                 SecretMounts map[string]arvados.Mount `json:"secret_mounts"`
1699         }
1700
1701         containerToken, err := runner.ContainerToken()
1702         if err != nil {
1703                 return fmt.Errorf("error getting container token: %v", err)
1704         }
1705
1706         runner.ContainerArvClient, runner.ContainerKeepClient,
1707                 runner.containerClient, err = runner.MkArvClient(containerToken)
1708         if err != nil {
1709                 return fmt.Errorf("error creating container API client: %v", err)
1710         }
1711
1712         runner.ContainerKeepClient.SetStorageClasses(runner.Container.OutputStorageClasses)
1713         runner.DispatcherKeepClient.SetStorageClasses(runner.Container.OutputStorageClasses)
1714
1715         err = runner.ContainerArvClient.Call("GET", "containers", runner.Container.UUID, "secret_mounts", nil, &sm)
1716         if err != nil {
1717                 if apierr, ok := err.(arvadosclient.APIServerError); !ok || apierr.HttpStatusCode != 404 {
1718                         return fmt.Errorf("error fetching secret_mounts: %v", err)
1719                 }
1720                 // ok && apierr.HttpStatusCode == 404, which means
1721                 // secret_mounts isn't supported by this API server.
1722         }
1723         runner.SecretMounts = sm.SecretMounts
1724
1725         return nil
1726 }
1727
1728 // NewContainerRunner creates a new container runner.
1729 func NewContainerRunner(dispatcherClient *arvados.Client,
1730         dispatcherArvClient IArvadosClient,
1731         dispatcherKeepClient IKeepClient,
1732         containerUUID string) (*ContainerRunner, error) {
1733
1734         cr := &ContainerRunner{
1735                 dispatcherClient:     dispatcherClient,
1736                 DispatcherArvClient:  dispatcherArvClient,
1737                 DispatcherKeepClient: dispatcherKeepClient,
1738         }
1739         cr.NewLogWriter = cr.NewArvLogWriter
1740         cr.RunArvMount = cr.ArvMountCmd
1741         cr.MkTempDir = ioutil.TempDir
1742         cr.MkArvClient = func(token string) (IArvadosClient, IKeepClient, *arvados.Client, error) {
1743                 cl, err := arvadosclient.MakeArvadosClient()
1744                 if err != nil {
1745                         return nil, nil, nil, err
1746                 }
1747                 cl.ApiToken = token
1748                 kc, err := keepclient.MakeKeepClient(cl)
1749                 if err != nil {
1750                         return nil, nil, nil, err
1751                 }
1752                 c2 := arvados.NewClientFromEnv()
1753                 c2.AuthToken = token
1754                 return cl, kc, c2, nil
1755         }
1756         var err error
1757         cr.LogCollection, err = (&arvados.Collection{}).FileSystem(cr.dispatcherClient, cr.DispatcherKeepClient)
1758         if err != nil {
1759                 return nil, err
1760         }
1761         cr.Container.UUID = containerUUID
1762         w, err := cr.NewLogWriter("crunch-run")
1763         if err != nil {
1764                 return nil, err
1765         }
1766         cr.CrunchLog = NewThrottledLogger(w)
1767         cr.CrunchLog.Immediate = log.New(os.Stderr, containerUUID+" ", 0)
1768
1769         loadLogThrottleParams(dispatcherArvClient)
1770         go cr.updateLogs()
1771
1772         return cr, nil
1773 }
1774
1775 func (command) RunCommand(prog string, args []string, stdin io.Reader, stdout, stderr io.Writer) int {
1776         log := log.New(stderr, "", 0)
1777         flags := flag.NewFlagSet(prog, flag.ContinueOnError)
1778         statInterval := flags.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting")
1779         cgroupRoot := flags.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree")
1780         cgroupParent := flags.String("cgroup-parent", "docker", "name of container's parent cgroup (ignored if -cgroup-parent-subsystem is used)")
1781         cgroupParentSubsystem := flags.String("cgroup-parent-subsystem", "", "use current cgroup for given subsystem as parent cgroup for container")
1782         caCertsPath := flags.String("ca-certs", "", "Path to TLS root certificates")
1783         detach := flags.Bool("detach", false, "Detach from parent process and run in the background")
1784         stdinConfig := flags.Bool("stdin-config", false, "Load config and environment variables from JSON message on stdin")
1785         configFile := flags.String("config", arvados.DefaultConfigFile, "filename of cluster config file to try loading if -stdin-config=false (default is $ARVADOS_CONFIG)")
1786         sleep := flags.Duration("sleep", 0, "Delay before starting (testing use only)")
1787         kill := flags.Int("kill", -1, "Send signal to an existing crunch-run process for given UUID")
1788         list := flags.Bool("list", false, "List UUIDs of existing crunch-run processes (and notify them to use price data passed on stdin)")
1789         enableMemoryLimit := flags.Bool("enable-memory-limit", true, "tell container runtime to limit container's memory usage")
1790         enableNetwork := flags.String("container-enable-networking", "default", "enable networking \"always\" (for all containers) or \"default\" (for containers that request it)")
1791         networkMode := flags.String("container-network-mode", "default", `Docker network mode for container (use any argument valid for docker --net)`)
1792         memprofile := flags.String("memprofile", "", "write memory profile to `file` after running container")
1793         runtimeEngine := flags.String("runtime-engine", "docker", "container runtime: docker or singularity")
1794         brokenNodeHook := flags.String("broken-node-hook", "", "script to run if node is detected to be broken (for example, Docker daemon is not running)")
1795         flags.Duration("check-containerd", 0, "Ignored. Exists for compatibility with older versions.")
1796         version := flags.Bool("version", false, "Write version information to stdout and exit 0.")
1797
1798         ignoreDetachFlag := false
1799         if len(args) > 0 && args[0] == "-no-detach" {
1800                 // This process was invoked by a parent process, which
1801                 // has passed along its own arguments, including
1802                 // -detach, after the leading -no-detach flag.  Strip
1803                 // the leading -no-detach flag (it's not recognized by
1804                 // flags.Parse()) and ignore the -detach flag that
1805                 // comes later.
1806                 args = args[1:]
1807                 ignoreDetachFlag = true
1808         }
1809
1810         if ok, code := cmd.ParseFlags(flags, prog, args, "container-uuid", stderr); !ok {
1811                 return code
1812         } else if *version {
1813                 fmt.Fprintln(stdout, prog, cmd.Version.String())
1814                 return 0
1815         } else if !*list && flags.NArg() != 1 {
1816                 fmt.Fprintf(stderr, "missing required argument: container-uuid (try -help)\n")
1817                 return 2
1818         }
1819
1820         containerUUID := flags.Arg(0)
1821
1822         switch {
1823         case *detach && !ignoreDetachFlag:
1824                 return Detach(containerUUID, prog, args, stdin, stdout, stderr)
1825         case *kill >= 0:
1826                 return KillProcess(containerUUID, syscall.Signal(*kill), stdout, stderr)
1827         case *list:
1828                 return ListProcesses(stdin, stdout, stderr)
1829         }
1830
1831         if len(containerUUID) != 27 {
1832                 log.Printf("usage: %s [options] UUID", prog)
1833                 return 1
1834         }
1835
1836         var keepstoreLogbuf bufThenWrite
1837         var conf ConfigData
1838         if *stdinConfig {
1839                 err := json.NewDecoder(stdin).Decode(&conf)
1840                 if err != nil {
1841                         log.Printf("decode stdin: %s", err)
1842                         return 1
1843                 }
1844                 for k, v := range conf.Env {
1845                         err = os.Setenv(k, v)
1846                         if err != nil {
1847                                 log.Printf("setenv(%q): %s", k, err)
1848                                 return 1
1849                         }
1850                 }
1851                 if conf.Cluster != nil {
1852                         // ClusterID is missing from the JSON
1853                         // representation, but we need it to generate
1854                         // a valid config file for keepstore, so we
1855                         // fill it using the container UUID prefix.
1856                         conf.Cluster.ClusterID = containerUUID[:5]
1857                 }
1858         } else {
1859                 conf = hpcConfData(containerUUID, *configFile, io.MultiWriter(&keepstoreLogbuf, stderr))
1860         }
1861
1862         log.Printf("crunch-run %s started", cmd.Version.String())
1863         time.Sleep(*sleep)
1864
1865         if *caCertsPath != "" {
1866                 arvadosclient.CertFiles = []string{*caCertsPath}
1867         }
1868
1869         keepstore, err := startLocalKeepstore(conf, io.MultiWriter(&keepstoreLogbuf, stderr))
1870         if err != nil {
1871                 log.Print(err)
1872                 return 1
1873         }
1874         if keepstore != nil {
1875                 defer keepstore.Process.Kill()
1876         }
1877
1878         api, err := arvadosclient.MakeArvadosClient()
1879         if err != nil {
1880                 log.Printf("%s: %v", containerUUID, err)
1881                 return 1
1882         }
1883         api.Retries = 8
1884
1885         kc, err := keepclient.MakeKeepClient(api)
1886         if err != nil {
1887                 log.Printf("%s: %v", containerUUID, err)
1888                 return 1
1889         }
1890         kc.BlockCache = &keepclient.BlockCache{MaxBlocks: 2}
1891         kc.Retries = 4
1892
1893         cr, err := NewContainerRunner(arvados.NewClientFromEnv(), api, kc, containerUUID)
1894         if err != nil {
1895                 log.Print(err)
1896                 return 1
1897         }
1898
1899         cr.keepstore = keepstore
1900         if keepstore == nil {
1901                 // Log explanation (if any) for why we're not running
1902                 // a local keepstore.
1903                 var buf bytes.Buffer
1904                 keepstoreLogbuf.SetWriter(&buf)
1905                 if buf.Len() > 0 {
1906                         cr.CrunchLog.Printf("%s", strings.TrimSpace(buf.String()))
1907                 }
1908         } else if logWhat := conf.Cluster.Containers.LocalKeepLogsToContainerLog; logWhat == "none" {
1909                 cr.CrunchLog.Printf("using local keepstore process (pid %d) at %s", keepstore.Process.Pid, os.Getenv("ARVADOS_KEEP_SERVICES"))
1910                 keepstoreLogbuf.SetWriter(io.Discard)
1911         } else {
1912                 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"))
1913                 logwriter, err := cr.NewLogWriter("keepstore")
1914                 if err != nil {
1915                         log.Print(err)
1916                         return 1
1917                 }
1918                 cr.keepstoreLogger = NewThrottledLogger(logwriter)
1919
1920                 var writer io.WriteCloser = cr.keepstoreLogger
1921                 if logWhat == "errors" {
1922                         writer = &filterKeepstoreErrorsOnly{WriteCloser: writer}
1923                 } else if logWhat != "all" {
1924                         // should have been caught earlier by
1925                         // dispatcher's config loader
1926                         log.Printf("invalid value for Containers.LocalKeepLogsToContainerLog: %q", logWhat)
1927                         return 1
1928                 }
1929                 err = keepstoreLogbuf.SetWriter(writer)
1930                 if err != nil {
1931                         log.Print(err)
1932                         return 1
1933                 }
1934                 cr.keepstoreLogbuf = &keepstoreLogbuf
1935         }
1936
1937         switch *runtimeEngine {
1938         case "docker":
1939                 cr.executor, err = newDockerExecutor(containerUUID, cr.CrunchLog.Printf, cr.containerWatchdogInterval)
1940         case "singularity":
1941                 cr.executor, err = newSingularityExecutor(cr.CrunchLog.Printf)
1942         default:
1943                 cr.CrunchLog.Printf("%s: unsupported RuntimeEngine %q", containerUUID, *runtimeEngine)
1944                 cr.CrunchLog.Close()
1945                 return 1
1946         }
1947         if err != nil {
1948                 cr.CrunchLog.Printf("%s: %v", containerUUID, err)
1949                 cr.checkBrokenNode(err)
1950                 cr.CrunchLog.Close()
1951                 return 1
1952         }
1953         defer cr.executor.Close()
1954
1955         cr.brokenNodeHook = *brokenNodeHook
1956
1957         gwAuthSecret := os.Getenv("GatewayAuthSecret")
1958         os.Unsetenv("GatewayAuthSecret")
1959         if gwAuthSecret == "" {
1960                 // not safe to run a gateway service without an auth
1961                 // secret
1962                 cr.CrunchLog.Printf("Not starting a gateway server (GatewayAuthSecret was not provided by dispatcher)")
1963         } else {
1964                 gwListen := os.Getenv("GatewayAddress")
1965                 cr.gateway = Gateway{
1966                         Address:       gwListen,
1967                         AuthSecret:    gwAuthSecret,
1968                         ContainerUUID: containerUUID,
1969                         Target:        cr.executor,
1970                         Log:           cr.CrunchLog,
1971                 }
1972                 if gwListen == "" {
1973                         // Direct connection won't work, so we use the
1974                         // gateway_address field to indicate the
1975                         // internalURL of the controller process that
1976                         // has the current tunnel connection.
1977                         cr.gateway.ArvadosClient = cr.dispatcherClient
1978                         cr.gateway.UpdateTunnelURL = func(url string) {
1979                                 cr.gateway.Address = "tunnel " + url
1980                                 cr.DispatcherArvClient.Update("containers", containerUUID,
1981                                         arvadosclient.Dict{"container": arvadosclient.Dict{"gateway_address": cr.gateway.Address}}, nil)
1982                         }
1983                 }
1984                 err = cr.gateway.Start()
1985                 if err != nil {
1986                         log.Printf("error starting gateway server: %s", err)
1987                         return 1
1988                 }
1989         }
1990
1991         parentTemp, tmperr := cr.MkTempDir("", "crunch-run."+containerUUID+".")
1992         if tmperr != nil {
1993                 log.Printf("%s: %v", containerUUID, tmperr)
1994                 return 1
1995         }
1996
1997         cr.parentTemp = parentTemp
1998         cr.statInterval = *statInterval
1999         cr.cgroupRoot = *cgroupRoot
2000         cr.expectCgroupParent = *cgroupParent
2001         cr.enableMemoryLimit = *enableMemoryLimit
2002         cr.enableNetwork = *enableNetwork
2003         cr.networkMode = *networkMode
2004         if *cgroupParentSubsystem != "" {
2005                 p, err := findCgroup(*cgroupParentSubsystem)
2006                 if err != nil {
2007                         log.Printf("fatal: cgroup parent subsystem: %s", err)
2008                         return 1
2009                 }
2010                 cr.setCgroupParent = p
2011                 cr.expectCgroupParent = p
2012         }
2013
2014         runerr := cr.Run()
2015
2016         if *memprofile != "" {
2017                 f, err := os.Create(*memprofile)
2018                 if err != nil {
2019                         log.Printf("could not create memory profile: %s", err)
2020                 }
2021                 runtime.GC() // get up-to-date statistics
2022                 if err := pprof.WriteHeapProfile(f); err != nil {
2023                         log.Printf("could not write memory profile: %s", err)
2024                 }
2025                 closeerr := f.Close()
2026                 if closeerr != nil {
2027                         log.Printf("closing memprofile file: %s", err)
2028                 }
2029         }
2030
2031         if runerr != nil {
2032                 log.Printf("%s: %v", containerUUID, runerr)
2033                 return 1
2034         }
2035         return 0
2036 }
2037
2038 // Try to load ConfigData in hpc (slurm/lsf) environment. This means
2039 // loading the cluster config from the specified file and (if that
2040 // works) getting the runtime_constraints container field from
2041 // controller to determine # VCPUs so we can calculate KeepBuffers.
2042 func hpcConfData(uuid string, configFile string, stderr io.Writer) ConfigData {
2043         var conf ConfigData
2044         conf.Cluster = loadClusterConfigFile(configFile, stderr)
2045         if conf.Cluster == nil {
2046                 // skip loading the container record -- we won't be
2047                 // able to start local keepstore anyway.
2048                 return conf
2049         }
2050         arv, err := arvadosclient.MakeArvadosClient()
2051         if err != nil {
2052                 fmt.Fprintf(stderr, "error setting up arvadosclient: %s\n", err)
2053                 return conf
2054         }
2055         arv.Retries = 8
2056         var ctr arvados.Container
2057         err = arv.Call("GET", "containers", uuid, "", arvadosclient.Dict{"select": []string{"runtime_constraints"}}, &ctr)
2058         if err != nil {
2059                 fmt.Fprintf(stderr, "error getting container record: %s\n", err)
2060                 return conf
2061         }
2062         if ctr.RuntimeConstraints.VCPUs > 0 {
2063                 conf.KeepBuffers = ctr.RuntimeConstraints.VCPUs * conf.Cluster.Containers.LocalKeepBlobBuffersPerVCPU
2064         }
2065         return conf
2066 }
2067
2068 // Load cluster config file from given path. If an error occurs, log
2069 // the error to stderr and return nil.
2070 func loadClusterConfigFile(path string, stderr io.Writer) *arvados.Cluster {
2071         ldr := config.NewLoader(&bytes.Buffer{}, ctxlog.New(stderr, "plain", "info"))
2072         ldr.Path = path
2073         cfg, err := ldr.Load()
2074         if err != nil {
2075                 fmt.Fprintf(stderr, "could not load config file %s: %s\n", path, err)
2076                 return nil
2077         }
2078         cluster, err := cfg.GetCluster("")
2079         if err != nil {
2080                 fmt.Fprintf(stderr, "could not use config file %s: %s\n", path, err)
2081                 return nil
2082         }
2083         fmt.Fprintf(stderr, "loaded config file %s\n", path)
2084         return cluster
2085 }
2086
2087 func startLocalKeepstore(configData ConfigData, logbuf io.Writer) (*exec.Cmd, error) {
2088         if configData.KeepBuffers < 1 {
2089                 fmt.Fprintf(logbuf, "not starting a local keepstore process because KeepBuffers=%v in config\n", configData.KeepBuffers)
2090                 return nil, nil
2091         }
2092         if configData.Cluster == nil {
2093                 fmt.Fprint(logbuf, "not starting a local keepstore process because cluster config file was not loaded\n")
2094                 return nil, nil
2095         }
2096         for uuid, vol := range configData.Cluster.Volumes {
2097                 if len(vol.AccessViaHosts) > 0 {
2098                         fmt.Fprintf(logbuf, "not starting a local keepstore process because a volume (%s) uses AccessViaHosts\n", uuid)
2099                         return nil, nil
2100                 }
2101                 if !vol.ReadOnly && vol.Replication < configData.Cluster.Collections.DefaultReplication {
2102                         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)
2103                         return nil, nil
2104                 }
2105         }
2106
2107         // Rather than have an alternate way to tell keepstore how
2108         // many buffers to use when starting it this way, we just
2109         // modify the cluster configuration that we feed it on stdin.
2110         configData.Cluster.API.MaxKeepBlobBuffers = configData.KeepBuffers
2111
2112         localaddr := localKeepstoreAddr()
2113         ln, err := net.Listen("tcp", net.JoinHostPort(localaddr, "0"))
2114         if err != nil {
2115                 return nil, err
2116         }
2117         _, port, err := net.SplitHostPort(ln.Addr().String())
2118         if err != nil {
2119                 ln.Close()
2120                 return nil, err
2121         }
2122         ln.Close()
2123         url := "http://" + net.JoinHostPort(localaddr, port)
2124
2125         fmt.Fprintf(logbuf, "starting keepstore on %s\n", url)
2126
2127         var confJSON bytes.Buffer
2128         err = json.NewEncoder(&confJSON).Encode(arvados.Config{
2129                 Clusters: map[string]arvados.Cluster{
2130                         configData.Cluster.ClusterID: *configData.Cluster,
2131                 },
2132         })
2133         if err != nil {
2134                 return nil, err
2135         }
2136         cmd := exec.Command("/proc/self/exe", "keepstore", "-config=-")
2137         if target, err := os.Readlink(cmd.Path); err == nil && strings.HasSuffix(target, ".test") {
2138                 // If we're a 'go test' process, running
2139                 // /proc/self/exe would start the test suite in a
2140                 // child process, which is not what we want.
2141                 cmd.Path, _ = exec.LookPath("go")
2142                 cmd.Args = append([]string{"go", "run", "../../cmd/arvados-server"}, cmd.Args[1:]...)
2143                 cmd.Env = os.Environ()
2144         }
2145         cmd.Stdin = &confJSON
2146         cmd.Stdout = logbuf
2147         cmd.Stderr = logbuf
2148         cmd.Env = append(cmd.Env,
2149                 "GOGC=10",
2150                 "ARVADOS_SERVICE_INTERNAL_URL="+url)
2151         err = cmd.Start()
2152         if err != nil {
2153                 return nil, fmt.Errorf("error starting keepstore process: %w", err)
2154         }
2155         cmdExited := false
2156         go func() {
2157                 cmd.Wait()
2158                 cmdExited = true
2159         }()
2160         ctx, cancel := context.WithDeadline(context.Background(), time.Now().Add(time.Second*10))
2161         defer cancel()
2162         poll := time.NewTicker(time.Second / 10)
2163         defer poll.Stop()
2164         client := http.Client{}
2165         for range poll.C {
2166                 testReq, err := http.NewRequestWithContext(ctx, "GET", url+"/_health/ping", nil)
2167                 testReq.Header.Set("Authorization", "Bearer "+configData.Cluster.ManagementToken)
2168                 if err != nil {
2169                         return nil, err
2170                 }
2171                 resp, err := client.Do(testReq)
2172                 if err == nil {
2173                         resp.Body.Close()
2174                         if resp.StatusCode == http.StatusOK {
2175                                 break
2176                         }
2177                 }
2178                 if cmdExited {
2179                         return nil, fmt.Errorf("keepstore child process exited")
2180                 }
2181                 if ctx.Err() != nil {
2182                         return nil, fmt.Errorf("timed out waiting for new keepstore process to report healthy")
2183                 }
2184         }
2185         os.Setenv("ARVADOS_KEEP_SERVICES", url)
2186         return cmd, nil
2187 }
2188
2189 // return current uid, gid, groups in a format suitable for logging:
2190 // "crunch-run process has uid=1234(arvados) gid=1234(arvados)
2191 // groups=1234(arvados),114(fuse)"
2192 func currentUserAndGroups() string {
2193         u, err := user.Current()
2194         if err != nil {
2195                 return fmt.Sprintf("error getting current user ID: %s", err)
2196         }
2197         s := fmt.Sprintf("crunch-run process has uid=%s(%s) gid=%s", u.Uid, u.Username, u.Gid)
2198         if g, err := user.LookupGroupId(u.Gid); err == nil {
2199                 s += fmt.Sprintf("(%s)", g.Name)
2200         }
2201         s += " groups="
2202         if gids, err := u.GroupIds(); err == nil {
2203                 for i, gid := range gids {
2204                         if i > 0 {
2205                                 s += ","
2206                         }
2207                         s += gid
2208                         if g, err := user.LookupGroupId(gid); err == nil {
2209                                 s += fmt.Sprintf("(%s)", g.Name)
2210                         }
2211                 }
2212         }
2213         return s
2214 }
2215
2216 // Return a suitable local interface address for a local keepstore
2217 // service. Currently this is the numerically lowest non-loopback ipv4
2218 // address assigned to a local interface that is not in any of the
2219 // link-local/vpn/loopback ranges 169.254/16, 100.64/10, or 127/8.
2220 func localKeepstoreAddr() string {
2221         var ips []net.IP
2222         // Ignore error (proceed with zero IPs)
2223         addrs, _ := processIPs(os.Getpid())
2224         for addr := range addrs {
2225                 ip := net.ParseIP(addr)
2226                 if ip == nil {
2227                         // invalid
2228                         continue
2229                 }
2230                 if ip.Mask(net.CIDRMask(8, 32)).Equal(net.IPv4(127, 0, 0, 0)) ||
2231                         ip.Mask(net.CIDRMask(10, 32)).Equal(net.IPv4(100, 64, 0, 0)) ||
2232                         ip.Mask(net.CIDRMask(16, 32)).Equal(net.IPv4(169, 254, 0, 0)) {
2233                         // unsuitable
2234                         continue
2235                 }
2236                 ips = append(ips, ip)
2237         }
2238         if len(ips) == 0 {
2239                 return "0.0.0.0"
2240         }
2241         sort.Slice(ips, func(ii, jj int) bool {
2242                 i, j := ips[ii], ips[jj]
2243                 if len(i) != len(j) {
2244                         return len(i) < len(j)
2245                 }
2246                 for x := range i {
2247                         if i[x] != j[x] {
2248                                 return i[x] < j[x]
2249                         }
2250                 }
2251                 return false
2252         })
2253         return ips[0].String()
2254 }
2255
2256 func (cr *ContainerRunner) loadPrices() {
2257         buf, err := os.ReadFile(filepath.Join(lockdir, pricesfile))
2258         if err != nil {
2259                 if !os.IsNotExist(err) {
2260                         cr.CrunchLog.Printf("loadPrices: read: %s", err)
2261                 }
2262                 return
2263         }
2264         var prices []cloud.InstancePrice
2265         err = json.Unmarshal(buf, &prices)
2266         if err != nil {
2267                 cr.CrunchLog.Printf("loadPrices: decode: %s", err)
2268                 return
2269         }
2270         cr.pricesLock.Lock()
2271         defer cr.pricesLock.Unlock()
2272         var lastKnown time.Time
2273         if len(cr.prices) > 0 {
2274                 lastKnown = cr.prices[0].StartTime
2275         }
2276         cr.prices = cloud.NormalizePriceHistory(append(prices, cr.prices...))
2277         for i := len(cr.prices) - 1; i >= 0; i-- {
2278                 price := cr.prices[i]
2279                 if price.StartTime.After(lastKnown) {
2280                         cr.CrunchLog.Printf("Instance price changed to %#.3g at %s", price.Price, price.StartTime.UTC())
2281                 }
2282         }
2283 }
2284
2285 func (cr *ContainerRunner) calculateCost(now time.Time) float64 {
2286         cr.pricesLock.Lock()
2287         defer cr.pricesLock.Unlock()
2288
2289         // First, make a "prices" slice with the real data as far back
2290         // as it goes, and (if needed) a "since the beginning of time"
2291         // placeholder containing a reasonable guess about what the
2292         // price was between cr.costStartTime and the earliest real
2293         // data point.
2294         prices := cr.prices
2295         if len(prices) == 0 {
2296                 // use price info in InstanceType record initially
2297                 // provided by cloud dispatcher
2298                 var p float64
2299                 var it arvados.InstanceType
2300                 if j := os.Getenv("InstanceType"); j != "" && json.Unmarshal([]byte(j), &it) == nil && it.Price > 0 {
2301                         p = it.Price
2302                 }
2303                 prices = []cloud.InstancePrice{{Price: p}}
2304         } else if prices[len(prices)-1].StartTime.After(cr.costStartTime) {
2305                 // guess earlier pricing was the same as the earliest
2306                 // price we know about
2307                 filler := prices[len(prices)-1]
2308                 filler.StartTime = time.Time{}
2309                 prices = append(prices, filler)
2310         }
2311
2312         // Now that our history of price changes goes back at least as
2313         // far as cr.costStartTime, add up the costs for each
2314         // interval.
2315         cost := 0.0
2316         spanEnd := now
2317         for _, ip := range prices {
2318                 spanStart := ip.StartTime
2319                 if spanStart.After(now) {
2320                         // pricing information from the future -- not
2321                         // expected from AWS, but possible in
2322                         // principle, and exercised by tests.
2323                         continue
2324                 }
2325                 last := false
2326                 if spanStart.Before(cr.costStartTime) {
2327                         spanStart = cr.costStartTime
2328                         last = true
2329                 }
2330                 cost += ip.Price * spanEnd.Sub(spanStart).Seconds() / 3600
2331                 if last {
2332                         break
2333                 }
2334                 spanEnd = spanStart
2335         }
2336
2337         return cost
2338 }