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         EC2SpotCheck bool
  56         Cluster      *arvados.Cluster
  57 }
  58
  59 // IArvadosClient is the minimal Arvados API methods used by crunch-run.
  60 type IArvadosClient interface {
  61         Create(resourceType string, parameters arvadosclient.Dict, output interface{}) error
  62         Get(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  63         Update(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  64         Call(method, resourceType, uuid, action string, parameters arvadosclient.Dict, output interface{}) error
  65         CallRaw(method string, resourceType string, uuid string, action string, parameters arvadosclient.Dict) (reader io.ReadCloser, err error)
  66         Discovery(key string) (interface{}, error)
  67 }
  68
  69 // ErrCancelled is the error returned when the container is cancelled.
  70 var ErrCancelled = errors.New("Cancelled")
  71
  72 // IKeepClient is the minimal Keep API methods used by crunch-run.
  73 type IKeepClient interface {
  74         BlockWrite(context.Context, arvados.BlockWriteOptions) (arvados.BlockWriteResponse, error)
  75         ReadAt(locator string, p []byte, off int) (int, error)
  76         ManifestFileReader(m manifest.Manifest, filename string) (arvados.File, error)
  77         LocalLocator(locator string) (string, error)
  78         ClearBlockCache()
  79         SetStorageClasses(sc []string)
  80 }
  81
  82 // NewLogWriter is a factory function to create a new log writer.
  83 type NewLogWriter func(name string) (io.WriteCloser, error)
  84
  85 type RunArvMount func(cmdline []string, tok string) (*exec.Cmd, error)
  86
  87 type MkTempDir func(string, string) (string, error)
  88
  89 type PsProcess interface {
  90         CmdlineSlice() ([]string, error)
  91 }
  92
  93 // ContainerRunner is the main stateful struct used for a single execution of a
  94 // container.
  95 type ContainerRunner struct {
  96         executor       containerExecutor
  97         executorStdin  io.Closer
  98         executorStdout io.Closer
  99         executorStderr io.Closer
 100
 101         // Dispatcher client is initialized with the Dispatcher token.
 102         // This is a privileged token used to manage container status
 103         // and logs.
 104         //
 105         // We have both dispatcherClient and DispatcherArvClient
 106         // because there are two different incompatible Arvados Go
 107         // SDKs and we have to use both (hopefully this gets fixed in
 108         // #14467)
 109         dispatcherClient     *arvados.Client
 110         DispatcherArvClient  IArvadosClient
 111         DispatcherKeepClient IKeepClient
 112
 113         // Container client is initialized with the Container token
 114         // This token controls the permissions of the container, and
 115         // must be used for operations such as reading collections.
 116         //
 117         // Same comment as above applies to
 118         // containerClient/ContainerArvClient.
 119         containerClient     *arvados.Client
 120         ContainerArvClient  IArvadosClient
 121         ContainerKeepClient IKeepClient
 122
 123         Container     arvados.Container
 124         token         string
 125         ExitCode      *int
 126         NewLogWriter  NewLogWriter
 127         CrunchLog     *ThrottledLogger
 128         logUUID       string
 129         logMtx        sync.Mutex
 130         LogCollection arvados.CollectionFileSystem
 131         LogsPDH       *string
 132         RunArvMount   RunArvMount
 133         MkTempDir     MkTempDir
 134         ArvMount      *exec.Cmd
 135         ArvMountPoint string
 136         HostOutputDir string
 137         Volumes       map[string]struct{}
 138         OutputPDH     *string
 139         SigChan       chan os.Signal
 140         ArvMountExit  chan error
 141         SecretMounts  map[string]arvados.Mount
 142         MkArvClient   func(token string) (IArvadosClient, IKeepClient, *arvados.Client, error)
 143         finalState    string
 144         parentTemp    string
 145         costStartTime time.Time
 146
 147         keepstore        *exec.Cmd
 148         keepstoreLogger  io.WriteCloser
 149         keepstoreLogbuf  *bufThenWrite
 150         statLogger       io.WriteCloser
 151         statReporter     *crunchstat.Reporter
 152         hoststatLogger   io.WriteCloser
 153         hoststatReporter *crunchstat.Reporter
 154         statInterval     time.Duration
 155         cgroupRoot       string
 156         // What we expect the container's cgroup parent to be.
 157         expectCgroupParent string
 158         // What we tell docker to use as the container's cgroup
 159         // parent. Note: Ideally we would use the same field for both
 160         // expectCgroupParent and setCgroupParent, and just make it
 161         // default to "docker". However, when using docker < 1.10 with
 162         // systemd, specifying a non-empty cgroup parent (even the
 163         // default value "docker") hits a docker bug
 164         // (https://github.com/docker/docker/issues/17126). Using two
 165         // separate fields makes it possible to use the "expect cgroup
 166         // parent to be X" feature even on sites where the "specify
 167         // cgroup parent" feature breaks.
 168         setCgroupParent string
 169
 170         cStateLock sync.Mutex
 171         cCancelled bool // StopContainer() invoked
 172
 173         enableMemoryLimit bool
 174         enableNetwork     string // one of "default" or "always"
 175         networkMode       string // "none", "host", or "" -- passed through to executor
 176         brokenNodeHook    string // script to run if node appears to be broken
 177         arvMountLog       *ThrottledLogger
 178
 179         containerWatchdogInterval time.Duration
 180
 181         gateway Gateway
 182
 183         prices     []cloud.InstancePrice
 184         pricesLock sync.Mutex
 185 }
 186
 187 // setupSignals sets up signal handling to gracefully terminate the
 188 // underlying container and update state when receiving a TERM, INT or
 189 // QUIT signal.
 190 func (runner *ContainerRunner) setupSignals() {
 191         runner.SigChan = make(chan os.Signal, 1)
 192         signal.Notify(runner.SigChan, syscall.SIGTERM)
 193         signal.Notify(runner.SigChan, syscall.SIGINT)
 194         signal.Notify(runner.SigChan, syscall.SIGQUIT)
 195
 196         go func(sig chan os.Signal) {
 197                 for s := range sig {
 198                         runner.stop(s)
 199                 }
 200         }(runner.SigChan)
 201 }
 202
 203 // stop the underlying container.
 204 func (runner *ContainerRunner) stop(sig os.Signal) {
 205         runner.cStateLock.Lock()
 206         defer runner.cStateLock.Unlock()
 207         if sig != nil {
 208                 runner.CrunchLog.Printf("caught signal: %v", sig)
 209         }
 210         runner.cCancelled = true
 211         runner.CrunchLog.Printf("stopping container")
 212         err := runner.executor.Stop()
 213         if err != nil {
 214                 runner.CrunchLog.Printf("error stopping container: %s", err)
 215         }
 216 }
 217
 218 var errorBlacklist = []string{
 219         "(?ms).*[Cc]annot connect to the Docker daemon.*",
 220         "(?ms).*oci runtime error.*starting container process.*container init.*mounting.*to rootfs.*no such file or directory.*",
 221         "(?ms).*grpc: the connection is unavailable.*",
 222 }
 223
 224 func (runner *ContainerRunner) runBrokenNodeHook() {
 225         if runner.brokenNodeHook == "" {
 226                 path := filepath.Join(lockdir, brokenfile)
 227                 runner.CrunchLog.Printf("Writing %s to mark node as broken", path)
 228                 f, err := os.OpenFile(path, os.O_CREATE|os.O_RDWR, 0700)
 229                 if err != nil {
 230                         runner.CrunchLog.Printf("Error writing %s: %s", path, err)
 231                         return
 232                 }
 233                 f.Close()
 234         } else {
 235                 runner.CrunchLog.Printf("Running broken node hook %q", runner.brokenNodeHook)
 236                 // run killme script
 237                 c := exec.Command(runner.brokenNodeHook)
 238                 c.Stdout = runner.CrunchLog
 239                 c.Stderr = runner.CrunchLog
 240                 err := c.Run()
 241                 if err != nil {
 242                         runner.CrunchLog.Printf("Error running broken node hook: %v", err)
 243                 }
 244         }
 245 }
 246
 247 func (runner *ContainerRunner) checkBrokenNode(goterr error) bool {
 248         for _, d := range errorBlacklist {
 249                 if m, e := regexp.MatchString(d, goterr.Error()); m && e == nil {
 250                         runner.CrunchLog.Printf("Error suggests node is unable to run containers: %v", goterr)
 251                         runner.runBrokenNodeHook()
 252                         return true
 253                 }
 254         }
 255         return false
 256 }
 257
 258 // LoadImage determines the docker image id from the container record and
 259 // checks if it is available in the local Docker image store.  If not, it loads
 260 // the image from Keep.
 261 func (runner *ContainerRunner) LoadImage() (string, error) {
 262         runner.CrunchLog.Printf("Fetching Docker image from collection '%s'", runner.Container.ContainerImage)
 263
 264         d, err := os.Open(runner.ArvMountPoint + "/by_id/" + runner.Container.ContainerImage)
 265         if err != nil {
 266                 return "", err
 267         }
 268         defer d.Close()
 269         allfiles, err := d.Readdirnames(-1)
 270         if err != nil {
 271                 return "", err
 272         }
 273         var tarfiles []string
 274         for _, fnm := range allfiles {
 275                 if strings.HasSuffix(fnm, ".tar") {
 276                         tarfiles = append(tarfiles, fnm)
 277                 }
 278         }
 279         if len(tarfiles) == 0 {
 280                 return "", fmt.Errorf("image collection does not include a .tar image file")
 281         }
 282         if len(tarfiles) > 1 {
 283                 return "", fmt.Errorf("cannot choose from multiple tar files in image collection: %v", tarfiles)
 284         }
 285         imageID := tarfiles[0][:len(tarfiles[0])-4]
 286         imageTarballPath := runner.ArvMountPoint + "/by_id/" + runner.Container.ContainerImage + "/" + imageID + ".tar"
 287         runner.CrunchLog.Printf("Using Docker image id %q", imageID)
 288
 289         runner.CrunchLog.Print("Loading Docker image from keep")
 290         err = runner.executor.LoadImage(imageID, imageTarballPath, runner.Container, runner.ArvMountPoint,
 291                 runner.containerClient)
 292         if err != nil {
 293                 return "", err
 294         }
 295
 296         return imageID, nil
 297 }
 298
 299 func (runner *ContainerRunner) ArvMountCmd(cmdline []string, token string) (c *exec.Cmd, err error) {
 300         c = exec.Command(cmdline[0], cmdline[1:]...)
 301
 302         // Copy our environment, but override ARVADOS_API_TOKEN with
 303         // the container auth token.
 304         c.Env = nil
 305         for _, s := range os.Environ() {
 306                 if !strings.HasPrefix(s, "ARVADOS_API_TOKEN=") {
 307                         c.Env = append(c.Env, s)
 308                 }
 309         }
 310         c.Env = append(c.Env, "ARVADOS_API_TOKEN="+token)
 311
 312         w, err := runner.NewLogWriter("arv-mount")
 313         if err != nil {
 314                 return nil, err
 315         }
 316         runner.arvMountLog = NewThrottledLogger(w)
 317         scanner := logScanner{
 318                 Patterns: []string{
 319                         "Keep write error",
 320                         "Block not found error",
 321                         "Unhandled exception during FUSE operation",
 322                 },
 323                 ReportFunc: func(pattern, text string) {
 324                         runner.updateRuntimeStatus(arvadosclient.Dict{
 325                                 "warning":       "arv-mount: " + pattern,
 326                                 "warningDetail": text,
 327                         })
 328                 },
 329         }
 330         c.Stdout = runner.arvMountLog
 331         c.Stderr = io.MultiWriter(runner.arvMountLog, os.Stderr, &scanner)
 332
 333         runner.CrunchLog.Printf("Running %v", c.Args)
 334
 335         err = c.Start()
 336         if err != nil {
 337                 return nil, err
 338         }
 339
 340         statReadme := make(chan bool)
 341         runner.ArvMountExit = make(chan error)
 342
 343         keepStatting := true
 344         go func() {
 345                 for keepStatting {
 346                         time.Sleep(100 * time.Millisecond)
 347                         _, err = os.Stat(fmt.Sprintf("%s/by_id/README", runner.ArvMountPoint))
 348                         if err == nil {
 349                                 keepStatting = false
 350                                 statReadme <- true
 351                         }
 352                 }
 353                 close(statReadme)
 354         }()
 355
 356         go func() {
 357                 mnterr := c.Wait()
 358                 if mnterr != nil {
 359                         runner.CrunchLog.Printf("Arv-mount exit error: %v", mnterr)
 360                 }
 361                 runner.ArvMountExit <- mnterr
 362                 close(runner.ArvMountExit)
 363         }()
 364
 365         select {
 366         case <-statReadme:
 367                 break
 368         case err := <-runner.ArvMountExit:
 369                 runner.ArvMount = nil
 370                 keepStatting = false
 371                 return nil, err
 372         }
 373
 374         return c, nil
 375 }
 376
 377 func (runner *ContainerRunner) SetupArvMountPoint(prefix string) (err error) {
 378         if runner.ArvMountPoint == "" {
 379                 runner.ArvMountPoint, err = runner.MkTempDir(runner.parentTemp, prefix)
 380         }
 381         return
 382 }
 383
 384 func copyfile(src string, dst string) (err error) {
 385         srcfile, err := os.Open(src)
 386         if err != nil {
 387                 return
 388         }
 389
 390         os.MkdirAll(path.Dir(dst), 0777)
 391
 392         dstfile, err := os.Create(dst)
 393         if err != nil {
 394                 return
 395         }
 396         _, err = io.Copy(dstfile, srcfile)
 397         if err != nil {
 398                 return
 399         }
 400
 401         err = srcfile.Close()
 402         err2 := dstfile.Close()
 403
 404         if err != nil {
 405                 return
 406         }
 407
 408         if err2 != nil {
 409                 return err2
 410         }
 411
 412         return nil
 413 }
 414
 415 func (runner *ContainerRunner) SetupMounts() (map[string]bindmount, error) {
 416         bindmounts := map[string]bindmount{}
 417         err := runner.SetupArvMountPoint("keep")
 418         if err != nil {
 419                 return nil, fmt.Errorf("While creating keep mount temp dir: %v", err)
 420         }
 421
 422         token, err := runner.ContainerToken()
 423         if err != nil {
 424                 return nil, fmt.Errorf("could not get container token: %s", err)
 425         }
 426         runner.CrunchLog.Printf("container token %q", token)
 427
 428         pdhOnly := true
 429         tmpcount := 0
 430         arvMountCmd := []string{
 431                 "arv-mount",
 432                 "--foreground",
 433                 "--read-write",
 434                 "--storage-classes", strings.Join(runner.Container.OutputStorageClasses, ","),
 435                 fmt.Sprintf("--crunchstat-interval=%v", runner.statInterval.Seconds())}
 436
 437         if _, isdocker := runner.executor.(*dockerExecutor); isdocker {
 438                 arvMountCmd = append(arvMountCmd, "--allow-other")
 439         }
 440
 441         if runner.Container.RuntimeConstraints.KeepCacheDisk > 0 {
 442                 keepcachedir, err := runner.MkTempDir(runner.parentTemp, "keepcache")
 443                 if err != nil {
 444                         return nil, fmt.Errorf("while creating keep cache temp dir: %v", err)
 445                 }
 446                 arvMountCmd = append(arvMountCmd, "--disk-cache", "--disk-cache-dir", keepcachedir, "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheDisk))
 447         } else if runner.Container.RuntimeConstraints.KeepCacheRAM > 0 {
 448                 arvMountCmd = append(arvMountCmd, "--ram-cache", "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheRAM))
 449         }
 450
 451         collectionPaths := []string{}
 452         needCertMount := true
 453         type copyFile struct {
 454                 src  string
 455                 bind string
 456         }
 457         var copyFiles []copyFile
 458
 459         var binds []string
 460         for bind := range runner.Container.Mounts {
 461                 binds = append(binds, bind)
 462         }
 463         for bind := range runner.SecretMounts {
 464                 if _, ok := runner.Container.Mounts[bind]; ok {
 465                         return nil, fmt.Errorf("secret mount %q conflicts with regular mount", bind)
 466                 }
 467                 if runner.SecretMounts[bind].Kind != "json" &&
 468                         runner.SecretMounts[bind].Kind != "text" {
 469                         return nil, fmt.Errorf("secret mount %q type is %q but only 'json' and 'text' are permitted",
 470                                 bind, runner.SecretMounts[bind].Kind)
 471                 }
 472                 binds = append(binds, bind)
 473         }
 474         sort.Strings(binds)
 475
 476         for _, bind := range binds {
 477                 mnt, notSecret := runner.Container.Mounts[bind]
 478                 if !notSecret {
 479                         mnt = runner.SecretMounts[bind]
 480                 }
 481                 if bind == "stdout" || bind == "stderr" {
 482                         // Is it a "file" mount kind?
 483                         if mnt.Kind != "file" {
 484                                 return nil, fmt.Errorf("unsupported mount kind '%s' for %s: only 'file' is supported", mnt.Kind, bind)
 485                         }
 486
 487                         // Does path start with OutputPath?
 488                         prefix := runner.Container.OutputPath
 489                         if !strings.HasSuffix(prefix, "/") {
 490                                 prefix += "/"
 491                         }
 492                         if !strings.HasPrefix(mnt.Path, prefix) {
 493                                 return nil, fmt.Errorf("%s path does not start with OutputPath: %s, %s", strings.Title(bind), mnt.Path, prefix)
 494                         }
 495                 }
 496
 497                 if bind == "stdin" {
 498                         // Is it a "collection" mount kind?
 499                         if mnt.Kind != "collection" && mnt.Kind != "json" {
 500                                 return nil, fmt.Errorf("unsupported mount kind '%s' for stdin: only 'collection' and 'json' are supported", mnt.Kind)
 501                         }
 502                 }
 503
 504                 if bind == "/etc/arvados/ca-certificates.crt" {
 505                         needCertMount = false
 506                 }
 507
 508                 if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && bind != runner.Container.OutputPath+"/" {
 509                         if mnt.Kind != "collection" && mnt.Kind != "text" && mnt.Kind != "json" {
 510                                 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)
 511                         }
 512                 }
 513
 514                 switch {
 515                 case mnt.Kind == "collection" && bind != "stdin":
 516                         var src string
 517                         if mnt.UUID != "" && mnt.PortableDataHash != "" {
 518                                 return nil, fmt.Errorf("cannot specify both 'uuid' and 'portable_data_hash' for a collection mount")
 519                         }
 520                         if mnt.UUID != "" {
 521                                 if mnt.Writable {
 522                                         return nil, fmt.Errorf("writing to existing collections currently not permitted")
 523                                 }
 524                                 pdhOnly = false
 525                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.UUID)
 526                         } else if mnt.PortableDataHash != "" {
 527                                 if mnt.Writable && !strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 528                                         return nil, fmt.Errorf("can never write to a collection specified by portable data hash")
 529                                 }
 530                                 idx := strings.Index(mnt.PortableDataHash, "/")
 531                                 if idx > 0 {
 532                                         mnt.Path = path.Clean(mnt.PortableDataHash[idx:])
 533                                         mnt.PortableDataHash = mnt.PortableDataHash[0:idx]
 534                                         runner.Container.Mounts[bind] = mnt
 535                                 }
 536                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.PortableDataHash)
 537                                 if mnt.Path != "" && mnt.Path != "." {
 538                                         if strings.HasPrefix(mnt.Path, "./") {
 539                                                 mnt.Path = mnt.Path[2:]
 540                                         } else if strings.HasPrefix(mnt.Path, "/") {
 541                                                 mnt.Path = mnt.Path[1:]
 542                                         }
 543                                         src += "/" + mnt.Path
 544                                 }
 545                         } else {
 546                                 src = fmt.Sprintf("%s/tmp%d", runner.ArvMountPoint, tmpcount)
 547                                 arvMountCmd = append(arvMountCmd, "--mount-tmp", fmt.Sprintf("tmp%d", tmpcount))
 548                                 tmpcount++
 549                         }
 550                         if mnt.Writable {
 551                                 if bind == runner.Container.OutputPath {
 552                                         runner.HostOutputDir = src
 553                                         bindmounts[bind] = bindmount{HostPath: src}
 554                                 } else if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 555                                         copyFiles = append(copyFiles, copyFile{src, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
 556                                 } else {
 557                                         bindmounts[bind] = bindmount{HostPath: src}
 558                                 }
 559                         } else {
 560                                 bindmounts[bind] = bindmount{HostPath: src, ReadOnly: true}
 561                         }
 562                         collectionPaths = append(collectionPaths, src)
 563
 564                 case mnt.Kind == "tmp":
 565                         var tmpdir string
 566                         tmpdir, err = runner.MkTempDir(runner.parentTemp, "tmp")
 567                         if err != nil {
 568                                 return nil, fmt.Errorf("while creating mount temp dir: %v", err)
 569                         }
 570                         st, staterr := os.Stat(tmpdir)
 571                         if staterr != nil {
 572                                 return nil, fmt.Errorf("while Stat on temp dir: %v", staterr)
 573                         }
 574                         err = os.Chmod(tmpdir, st.Mode()|os.ModeSetgid|0777)
 575                         if staterr != nil {
 576                                 return nil, fmt.Errorf("while Chmod temp dir: %v", err)
 577                         }
 578                         bindmounts[bind] = bindmount{HostPath: tmpdir}
 579                         if bind == runner.Container.OutputPath {
 580                                 runner.HostOutputDir = tmpdir
 581                         }
 582
 583                 case mnt.Kind == "json" || mnt.Kind == "text":
 584                         var filedata []byte
 585                         if mnt.Kind == "json" {
 586                                 filedata, err = json.Marshal(mnt.Content)
 587                                 if err != nil {
 588                                         return nil, fmt.Errorf("encoding json data: %v", err)
 589                                 }
 590                         } else {
 591                                 text, ok := mnt.Content.(string)
 592                                 if !ok {
 593                                         return nil, fmt.Errorf("content for mount %q must be a string", bind)
 594                                 }
 595                                 filedata = []byte(text)
 596                         }
 597
 598                         tmpdir, err := runner.MkTempDir(runner.parentTemp, mnt.Kind)
 599                         if err != nil {
 600                                 return nil, fmt.Errorf("creating temp dir: %v", err)
 601                         }
 602                         tmpfn := filepath.Join(tmpdir, "mountdata."+mnt.Kind)
 603                         err = ioutil.WriteFile(tmpfn, filedata, 0444)
 604                         if err != nil {
 605                                 return nil, fmt.Errorf("writing temp file: %v", err)
 606                         }
 607                         if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && (notSecret || runner.Container.Mounts[runner.Container.OutputPath].Kind != "collection") {
 608                                 // In most cases, if the container
 609                                 // specifies a literal file inside the
 610                                 // output path, we copy it into the
 611                                 // output directory (either a mounted
 612                                 // collection or a staging area on the
 613                                 // host fs). If it's a secret, it will
 614                                 // be skipped when copying output from
 615                                 // staging to Keep later.
 616                                 copyFiles = append(copyFiles, copyFile{tmpfn, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
 617                         } else {
 618                                 // If a secret is outside OutputPath,
 619                                 // we bind mount the secret file
 620                                 // directly just like other mounts. We
 621                                 // also use this strategy when a
 622                                 // secret is inside OutputPath but
 623                                 // OutputPath is a live collection, to
 624                                 // avoid writing the secret to
 625                                 // Keep. Attempting to remove a
 626                                 // bind-mounted secret file from
 627                                 // inside the container will return a
 628                                 // "Device or resource busy" error
 629                                 // that might not be handled well by
 630                                 // the container, which is why we
 631                                 // don't use this strategy when
 632                                 // OutputPath is a staging directory.
 633                                 bindmounts[bind] = bindmount{HostPath: tmpfn, ReadOnly: true}
 634                         }
 635
 636                 case mnt.Kind == "git_tree":
 637                         tmpdir, err := runner.MkTempDir(runner.parentTemp, "git_tree")
 638                         if err != nil {
 639                                 return nil, fmt.Errorf("creating temp dir: %v", err)
 640                         }
 641                         err = gitMount(mnt).extractTree(runner.ContainerArvClient, tmpdir, token)
 642                         if err != nil {
 643                                 return nil, err
 644                         }
 645                         bindmounts[bind] = bindmount{HostPath: tmpdir, ReadOnly: true}
 646                 }
 647         }
 648
 649         if runner.HostOutputDir == "" {
 650                 return nil, fmt.Errorf("output path does not correspond to a writable mount point")
 651         }
 652
 653         if needCertMount && runner.Container.RuntimeConstraints.API {
 654                 for _, certfile := range arvadosclient.CertFiles {
 655                         _, err := os.Stat(certfile)
 656                         if err == nil {
 657                                 bindmounts["/etc/arvados/ca-certificates.crt"] = bindmount{HostPath: certfile, ReadOnly: true}
 658                                 break
 659                         }
 660                 }
 661         }
 662
 663         if pdhOnly {
 664                 // If we are only mounting collections by pdh, make
 665                 // sure we don't subscribe to websocket events to
 666                 // avoid putting undesired load on the API server
 667                 arvMountCmd = append(arvMountCmd, "--mount-by-pdh", "by_id", "--disable-event-listening")
 668         } else {
 669                 arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_id")
 670         }
 671         // the by_uuid mount point is used by singularity when writing
 672         // out docker images converted to SIF
 673         arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_uuid")
 674         arvMountCmd = append(arvMountCmd, runner.ArvMountPoint)
 675
 676         runner.ArvMount, err = runner.RunArvMount(arvMountCmd, token)
 677         if err != nil {
 678                 return nil, fmt.Errorf("while trying to start arv-mount: %v", err)
 679         }
 680         if runner.hoststatReporter != nil && runner.ArvMount != nil {
 681                 runner.hoststatReporter.ReportPID("arv-mount", runner.ArvMount.Process.Pid)
 682         }
 683
 684         for _, p := range collectionPaths {
 685                 _, err = os.Stat(p)
 686                 if err != nil {
 687                         return nil, fmt.Errorf("while checking that input files exist: %v", err)
 688                 }
 689         }
 690
 691         for _, cp := range copyFiles {
 692                 st, err := os.Stat(cp.src)
 693                 if err != nil {
 694                         return nil, fmt.Errorf("while staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 695                 }
 696                 if st.IsDir() {
 697                         err = filepath.Walk(cp.src, func(walkpath string, walkinfo os.FileInfo, walkerr error) error {
 698                                 if walkerr != nil {
 699                                         return walkerr
 700                                 }
 701                                 target := path.Join(cp.bind, walkpath[len(cp.src):])
 702                                 if walkinfo.Mode().IsRegular() {
 703                                         copyerr := copyfile(walkpath, target)
 704                                         if copyerr != nil {
 705                                                 return copyerr
 706                                         }
 707                                         return os.Chmod(target, walkinfo.Mode()|0777)
 708                                 } else if walkinfo.Mode().IsDir() {
 709                                         mkerr := os.MkdirAll(target, 0777)
 710                                         if mkerr != nil {
 711                                                 return mkerr
 712                                         }
 713                                         return os.Chmod(target, walkinfo.Mode()|os.ModeSetgid|0777)
 714                                 } else {
 715                                         return fmt.Errorf("source %q is not a regular file or directory", cp.src)
 716                                 }
 717                         })
 718                 } else if st.Mode().IsRegular() {
 719                         err = copyfile(cp.src, cp.bind)
 720                         if err == nil {
 721                                 err = os.Chmod(cp.bind, st.Mode()|0777)
 722                         }
 723                 }
 724                 if err != nil {
 725                         return nil, fmt.Errorf("while staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 726                 }
 727         }
 728
 729         return bindmounts, nil
 730 }
 731
 732 func (runner *ContainerRunner) stopHoststat() error {
 733         if runner.hoststatReporter == nil {
 734                 return nil
 735         }
 736         runner.hoststatReporter.Stop()
 737         err := runner.hoststatLogger.Close()
 738         if err != nil {
 739                 return fmt.Errorf("error closing hoststat logs: %v", err)
 740         }
 741         return nil
 742 }
 743
 744 func (runner *ContainerRunner) startHoststat() error {
 745         w, err := runner.NewLogWriter("hoststat")
 746         if err != nil {
 747                 return err
 748         }
 749         runner.hoststatLogger = NewThrottledLogger(w)
 750         runner.hoststatReporter = &crunchstat.Reporter{
 751                 Logger:     log.New(runner.hoststatLogger, "", 0),
 752                 CgroupRoot: runner.cgroupRoot,
 753                 PollPeriod: runner.statInterval,
 754         }
 755         runner.hoststatReporter.Start()
 756         runner.hoststatReporter.ReportPID("crunch-run", os.Getpid())
 757         return nil
 758 }
 759
 760 func (runner *ContainerRunner) startCrunchstat() error {
 761         w, err := runner.NewLogWriter("crunchstat")
 762         if err != nil {
 763                 return err
 764         }
 765         runner.statLogger = NewThrottledLogger(w)
 766         runner.statReporter = &crunchstat.Reporter{
 767                 CID:          runner.executor.CgroupID(),
 768                 Logger:       log.New(runner.statLogger, "", 0),
 769                 CgroupParent: runner.expectCgroupParent,
 770                 CgroupRoot:   runner.cgroupRoot,
 771                 PollPeriod:   runner.statInterval,
 772                 TempDir:      runner.parentTemp,
 773         }
 774         runner.statReporter.Start()
 775         return nil
 776 }
 777
 778 type infoCommand struct {
 779         label string
 780         cmd   []string
 781 }
 782
 783 // LogHostInfo logs info about the current host, for debugging and
 784 // accounting purposes. Although it's logged as "node-info", this is
 785 // about the environment where crunch-run is actually running, which
 786 // might differ from what's described in the node record (see
 787 // LogNodeRecord).
 788 func (runner *ContainerRunner) LogHostInfo() (err error) {
 789         w, err := runner.NewLogWriter("node-info")
 790         if err != nil {
 791                 return
 792         }
 793
 794         commands := []infoCommand{
 795                 {
 796                         label: "Host Information",
 797                         cmd:   []string{"uname", "-a"},
 798                 },
 799                 {
 800                         label: "CPU Information",
 801                         cmd:   []string{"cat", "/proc/cpuinfo"},
 802                 },
 803                 {
 804                         label: "Memory Information",
 805                         cmd:   []string{"cat", "/proc/meminfo"},
 806                 },
 807                 {
 808                         label: "Disk Space",
 809                         cmd:   []string{"df", "-m", "/", os.TempDir()},
 810                 },
 811                 {
 812                         label: "Disk INodes",
 813                         cmd:   []string{"df", "-i", "/", os.TempDir()},
 814                 },
 815         }
 816
 817         // Run commands with informational output to be logged.
 818         for _, command := range commands {
 819                 fmt.Fprintln(w, command.label)
 820                 cmd := exec.Command(command.cmd[0], command.cmd[1:]...)
 821                 cmd.Stdout = w
 822                 cmd.Stderr = w
 823                 if err := cmd.Run(); err != nil {
 824                         err = fmt.Errorf("While running command %q: %v", command.cmd, err)
 825                         fmt.Fprintln(w, err)
 826                         return err
 827                 }
 828                 fmt.Fprintln(w, "")
 829         }
 830
 831         err = w.Close()
 832         if err != nil {
 833                 return fmt.Errorf("While closing node-info logs: %v", err)
 834         }
 835         return nil
 836 }
 837
 838 // LogContainerRecord gets and saves the raw JSON container record from the API server
 839 func (runner *ContainerRunner) LogContainerRecord() error {
 840         logged, err := runner.logAPIResponse("container", "containers", map[string]interface{}{"filters": [][]string{{"uuid", "=", runner.Container.UUID}}}, nil)
 841         if !logged && err == nil {
 842                 err = fmt.Errorf("error: no container record found for %s", runner.Container.UUID)
 843         }
 844         return err
 845 }
 846
 847 // LogNodeRecord logs the current host's InstanceType config entry (or
 848 // the arvados#node record, if running via crunch-dispatch-slurm).
 849 func (runner *ContainerRunner) LogNodeRecord() error {
 850         if it := os.Getenv("InstanceType"); it != "" {
 851                 // Dispatched via arvados-dispatch-cloud. Save
 852                 // InstanceType config fragment received from
 853                 // dispatcher on stdin.
 854                 w, err := runner.LogCollection.OpenFile("node.json", os.O_CREATE|os.O_WRONLY, 0666)
 855                 if err != nil {
 856                         return err
 857                 }
 858                 defer w.Close()
 859                 _, err = io.WriteString(w, it)
 860                 if err != nil {
 861                         return err
 862                 }
 863                 return w.Close()
 864         }
 865         // Dispatched via crunch-dispatch-slurm. Look up
 866         // apiserver's node record corresponding to
 867         // $SLURMD_NODENAME.
 868         hostname := os.Getenv("SLURMD_NODENAME")
 869         if hostname == "" {
 870                 hostname, _ = os.Hostname()
 871         }
 872         _, err := runner.logAPIResponse("node", "nodes", map[string]interface{}{"filters": [][]string{{"hostname", "=", hostname}}}, func(resp interface{}) {
 873                 // The "info" field has admin-only info when
 874                 // obtained with a privileged token, and
 875                 // should not be logged.
 876                 node, ok := resp.(map[string]interface{})
 877                 if ok {
 878                         delete(node, "info")
 879                 }
 880         })
 881         return err
 882 }
 883
 884 func (runner *ContainerRunner) logAPIResponse(label, path string, params map[string]interface{}, munge func(interface{})) (logged bool, err error) {
 885         writer, err := runner.LogCollection.OpenFile(label+".json", os.O_CREATE|os.O_WRONLY, 0666)
 886         if err != nil {
 887                 return false, err
 888         }
 889         w := &ArvLogWriter{
 890                 ArvClient:     runner.DispatcherArvClient,
 891                 UUID:          runner.Container.UUID,
 892                 loggingStream: label,
 893                 writeCloser:   writer,
 894         }
 895
 896         reader, err := runner.DispatcherArvClient.CallRaw("GET", path, "", "", arvadosclient.Dict(params))
 897         if err != nil {
 898                 return false, fmt.Errorf("error getting %s record: %v", label, err)
 899         }
 900         defer reader.Close()
 901
 902         dec := json.NewDecoder(reader)
 903         dec.UseNumber()
 904         var resp map[string]interface{}
 905         if err = dec.Decode(&resp); err != nil {
 906                 return false, fmt.Errorf("error decoding %s list response: %v", label, err)
 907         }
 908         items, ok := resp["items"].([]interface{})
 909         if !ok {
 910                 return false, fmt.Errorf("error decoding %s list response: no \"items\" key in API list response", label)
 911         } else if len(items) < 1 {
 912                 return false, nil
 913         }
 914         if munge != nil {
 915                 munge(items[0])
 916         }
 917         // Re-encode it using indentation to improve readability
 918         enc := json.NewEncoder(w)
 919         enc.SetIndent("", "    ")
 920         if err = enc.Encode(items[0]); err != nil {
 921                 return false, fmt.Errorf("error logging %s record: %v", label, err)
 922         }
 923         err = w.Close()
 924         if err != nil {
 925                 return false, fmt.Errorf("error closing %s.json in log collection: %v", label, err)
 926         }
 927         return true, nil
 928 }
 929
 930 func (runner *ContainerRunner) getStdoutFile(mntPath string) (*os.File, error) {
 931         stdoutPath := mntPath[len(runner.Container.OutputPath):]
 932         index := strings.LastIndex(stdoutPath, "/")
 933         if index > 0 {
 934                 subdirs := stdoutPath[:index]
 935                 if subdirs != "" {
 936                         st, err := os.Stat(runner.HostOutputDir)
 937                         if err != nil {
 938                                 return nil, fmt.Errorf("While Stat on temp dir: %v", err)
 939                         }
 940                         stdoutPath := filepath.Join(runner.HostOutputDir, subdirs)
 941                         err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777)
 942                         if err != nil {
 943                                 return nil, fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err)
 944                         }
 945                 }
 946         }
 947         stdoutFile, err := os.Create(filepath.Join(runner.HostOutputDir, stdoutPath))
 948         if err != nil {
 949                 return nil, fmt.Errorf("While creating file %q: %v", stdoutPath, err)
 950         }
 951
 952         return stdoutFile, nil
 953 }
 954
 955 // CreateContainer creates the docker container.
 956 func (runner *ContainerRunner) CreateContainer(imageID string, bindmounts map[string]bindmount) error {
 957         var stdin io.ReadCloser = ioutil.NopCloser(bytes.NewReader(nil))
 958         if mnt, ok := runner.Container.Mounts["stdin"]; ok {
 959                 switch mnt.Kind {
 960                 case "collection":
 961                         var collID string
 962                         if mnt.UUID != "" {
 963                                 collID = mnt.UUID
 964                         } else {
 965                                 collID = mnt.PortableDataHash
 966                         }
 967                         path := runner.ArvMountPoint + "/by_id/" + collID + "/" + mnt.Path
 968                         f, err := os.Open(path)
 969                         if err != nil {
 970                                 return err
 971                         }
 972                         stdin = f
 973                 case "json":
 974                         j, err := json.Marshal(mnt.Content)
 975                         if err != nil {
 976                                 return fmt.Errorf("error encoding stdin json data: %v", err)
 977                         }
 978                         stdin = ioutil.NopCloser(bytes.NewReader(j))
 979                 default:
 980                         return fmt.Errorf("stdin mount has unsupported kind %q", mnt.Kind)
 981                 }
 982         }
 983
 984         var stdout, stderr io.WriteCloser
 985         if mnt, ok := runner.Container.Mounts["stdout"]; ok {
 986                 f, err := runner.getStdoutFile(mnt.Path)
 987                 if err != nil {
 988                         return err
 989                 }
 990                 stdout = f
 991         } else if w, err := runner.NewLogWriter("stdout"); err != nil {
 992                 return err
 993         } else {
 994                 stdout = NewThrottledLogger(w)
 995         }
 996
 997         if mnt, ok := runner.Container.Mounts["stderr"]; ok {
 998                 f, err := runner.getStdoutFile(mnt.Path)
 999                 if err != nil {
1000                         return err
1001                 }
1002                 stderr = f
1003         } else if w, err := runner.NewLogWriter("stderr"); err != nil {
1004                 return err
1005         } else {
1006                 stderr = NewThrottledLogger(w)
1007         }
1008
1009         env := runner.Container.Environment
1010         enableNetwork := runner.enableNetwork == "always"
1011         if runner.Container.RuntimeConstraints.API {
1012                 enableNetwork = true
1013                 tok, err := runner.ContainerToken()
1014                 if err != nil {
1015                         return err
1016                 }
1017                 env = map[string]string{}
1018                 for k, v := range runner.Container.Environment {
1019                         env[k] = v
1020                 }
1021                 env["ARVADOS_API_TOKEN"] = tok
1022                 env["ARVADOS_API_HOST"] = os.Getenv("ARVADOS_API_HOST")
1023                 env["ARVADOS_API_HOST_INSECURE"] = os.Getenv("ARVADOS_API_HOST_INSECURE")
1024                 env["ARVADOS_KEEP_SERVICES"] = os.Getenv("ARVADOS_KEEP_SERVICES")
1025         }
1026         workdir := runner.Container.Cwd
1027         if workdir == "." {
1028                 // both "" and "." mean default
1029                 workdir = ""
1030         }
1031         ram := runner.Container.RuntimeConstraints.RAM
1032         if !runner.enableMemoryLimit {
1033                 ram = 0
1034         }
1035         runner.executorStdin = stdin
1036         runner.executorStdout = stdout
1037         runner.executorStderr = stderr
1038
1039         if runner.Container.RuntimeConstraints.CUDA.DeviceCount > 0 {
1040                 nvidiaModprobe(runner.CrunchLog)
1041         }
1042
1043         return runner.executor.Create(containerSpec{
1044                 Image:           imageID,
1045                 VCPUs:           runner.Container.RuntimeConstraints.VCPUs,
1046                 RAM:             ram,
1047                 WorkingDir:      workdir,
1048                 Env:             env,
1049                 BindMounts:      bindmounts,
1050                 Command:         runner.Container.Command,
1051                 EnableNetwork:   enableNetwork,
1052                 CUDADeviceCount: runner.Container.RuntimeConstraints.CUDA.DeviceCount,
1053                 NetworkMode:     runner.networkMode,
1054                 CgroupParent:    runner.setCgroupParent,
1055                 Stdin:           stdin,
1056                 Stdout:          stdout,
1057                 Stderr:          stderr,
1058         })
1059 }
1060
1061 // StartContainer starts the docker container created by CreateContainer.
1062 func (runner *ContainerRunner) StartContainer() error {
1063         runner.CrunchLog.Printf("Starting container")
1064         runner.cStateLock.Lock()
1065         defer runner.cStateLock.Unlock()
1066         if runner.cCancelled {
1067                 return ErrCancelled
1068         }
1069         err := runner.executor.Start()
1070         if err != nil {
1071                 var advice string
1072                 if m, e := regexp.MatchString("(?ms).*(exec|System error).*(no such file or directory|file not found).*", err.Error()); m && e == nil {
1073                         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])
1074                 }
1075                 return fmt.Errorf("could not start container: %v%s", err, advice)
1076         }
1077         return nil
1078 }
1079
1080 // WaitFinish waits for the container to terminate, capture the exit code, and
1081 // close the stdout/stderr logging.
1082 func (runner *ContainerRunner) WaitFinish() error {
1083         runner.CrunchLog.Print("Waiting for container to finish")
1084         var timeout <-chan time.Time
1085         if s := runner.Container.SchedulingParameters.MaxRunTime; s > 0 {
1086                 timeout = time.After(time.Duration(s) * time.Second)
1087         }
1088         ctx, cancel := context.WithCancel(context.Background())
1089         defer cancel()
1090         go func() {
1091                 select {
1092                 case <-timeout:
1093                         runner.CrunchLog.Printf("maximum run time exceeded. Stopping container.")
1094                         runner.stop(nil)
1095                 case <-runner.ArvMountExit:
1096                         runner.CrunchLog.Printf("arv-mount exited while container is still running. Stopping container.")
1097                         runner.stop(nil)
1098                 case <-ctx.Done():
1099                 }
1100         }()
1101         exitcode, err := runner.executor.Wait(ctx)
1102         if err != nil {
1103                 runner.checkBrokenNode(err)
1104                 return err
1105         }
1106         runner.ExitCode = &exitcode
1107
1108         extra := ""
1109         if exitcode&0x80 != 0 {
1110                 // Convert raw exit status (0x80 + signal number) to a
1111                 // string to log after the code, like " (signal 101)"
1112                 // or " (signal 9, killed)"
1113                 sig := syscall.WaitStatus(exitcode).Signal()
1114                 if name := unix.SignalName(sig); name != "" {
1115                         extra = fmt.Sprintf(" (signal %d, %s)", sig, name)
1116                 } else {
1117                         extra = fmt.Sprintf(" (signal %d)", sig)
1118                 }
1119         }
1120         runner.CrunchLog.Printf("Container exited with status code %d%s", exitcode, extra)
1121         err = runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1122                 "container": arvadosclient.Dict{"exit_code": exitcode},
1123         }, nil)
1124         if err != nil {
1125                 runner.CrunchLog.Printf("ignoring error updating exit_code: %s", err)
1126         }
1127
1128         var returnErr error
1129         if err = runner.executorStdin.Close(); err != nil {
1130                 err = fmt.Errorf("error closing container stdin: %s", err)
1131                 runner.CrunchLog.Printf("%s", err)
1132                 returnErr = err
1133         }
1134         if err = runner.executorStdout.Close(); err != nil {
1135                 err = fmt.Errorf("error closing container stdout: %s", err)
1136                 runner.CrunchLog.Printf("%s", err)
1137                 if returnErr == nil {
1138                         returnErr = err
1139                 }
1140         }
1141         if err = runner.executorStderr.Close(); err != nil {
1142                 err = fmt.Errorf("error closing container stderr: %s", err)
1143                 runner.CrunchLog.Printf("%s", err)
1144                 if returnErr == nil {
1145                         returnErr = err
1146                 }
1147         }
1148
1149         if runner.statReporter != nil {
1150                 runner.statReporter.Stop()
1151                 err = runner.statLogger.Close()
1152                 if err != nil {
1153                         runner.CrunchLog.Printf("error closing crunchstat logs: %v", err)
1154                 }
1155         }
1156         return returnErr
1157 }
1158
1159 func (runner *ContainerRunner) updateLogs() {
1160         ticker := time.NewTicker(crunchLogUpdatePeriod / 360)
1161         defer ticker.Stop()
1162
1163         sigusr1 := make(chan os.Signal, 1)
1164         signal.Notify(sigusr1, syscall.SIGUSR1)
1165         defer signal.Stop(sigusr1)
1166
1167         saveAtTime := time.Now().Add(crunchLogUpdatePeriod)
1168         saveAtSize := crunchLogUpdateSize
1169         var savedSize int64
1170         for {
1171                 select {
1172                 case <-ticker.C:
1173                 case <-sigusr1:
1174                         saveAtTime = time.Now()
1175                 }
1176                 runner.logMtx.Lock()
1177                 done := runner.LogsPDH != nil
1178                 runner.logMtx.Unlock()
1179                 if done {
1180                         return
1181                 }
1182                 size := runner.LogCollection.Size()
1183                 if size == savedSize || (time.Now().Before(saveAtTime) && size < saveAtSize) {
1184                         continue
1185                 }
1186                 saveAtTime = time.Now().Add(crunchLogUpdatePeriod)
1187                 saveAtSize = runner.LogCollection.Size() + crunchLogUpdateSize
1188                 saved, err := runner.saveLogCollection(false)
1189                 if err != nil {
1190                         runner.CrunchLog.Printf("error updating log collection: %s", err)
1191                         continue
1192                 }
1193
1194                 err = runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1195                         "container": arvadosclient.Dict{"log": saved.PortableDataHash},
1196                 }, nil)
1197                 if err != nil {
1198                         runner.CrunchLog.Printf("error updating container log to %s: %s", saved.PortableDataHash, err)
1199                         continue
1200                 }
1201
1202                 savedSize = size
1203         }
1204 }
1205
1206 var spotInterruptionCheckInterval = 5 * time.Second
1207 var ec2MetadataBaseURL = "http://169.254.169.254"
1208
1209 const ec2TokenTTL = time.Second * 21600
1210
1211 func (runner *ContainerRunner) checkSpotInterruptionNotices() {
1212         type ec2metadata struct {
1213                 Action string    `json:"action"`
1214                 Time   time.Time `json:"time"`
1215         }
1216         runner.CrunchLog.Printf("Checking for spot interruptions every %v using instance metadata at %s", spotInterruptionCheckInterval, ec2MetadataBaseURL)
1217         var metadata ec2metadata
1218         var token string
1219         var tokenExp time.Time
1220         check := func() error {
1221                 ctx, cancel := context.WithDeadline(context.Background(), time.Now().Add(time.Minute))
1222                 defer cancel()
1223                 if token == "" || tokenExp.Sub(time.Now()) < time.Minute {
1224                         req, err := http.NewRequestWithContext(ctx, http.MethodPut, ec2MetadataBaseURL+"/latest/api/token", nil)
1225                         if err != nil {
1226                                 return err
1227                         }
1228                         req.Header.Set("X-aws-ec2-metadata-token-ttl-seconds", fmt.Sprintf("%d", int(ec2TokenTTL/time.Second)))
1229                         resp, err := http.DefaultClient.Do(req)
1230                         if err != nil {
1231                                 return err
1232                         }
1233                         defer resp.Body.Close()
1234                         if resp.StatusCode != http.StatusOK {
1235                                 return fmt.Errorf("%s", resp.Status)
1236                         }
1237                         newtoken, err := ioutil.ReadAll(resp.Body)
1238                         if err != nil {
1239                                 return err
1240                         }
1241                         token = strings.TrimSpace(string(newtoken))
1242                         tokenExp = time.Now().Add(ec2TokenTTL)
1243                 }
1244                 req, err := http.NewRequestWithContext(ctx, http.MethodGet, ec2MetadataBaseURL+"/latest/meta-data/spot/instance-action", nil)
1245                 if err != nil {
1246                         return err
1247                 }
1248                 req.Header.Set("X-aws-ec2-metadata-token", token)
1249                 resp, err := http.DefaultClient.Do(req)
1250                 if err != nil {
1251                         return err
1252                 }
1253                 defer resp.Body.Close()
1254                 metadata = ec2metadata{}
1255                 switch resp.StatusCode {
1256                 case http.StatusOK:
1257                         break
1258                 case http.StatusNotFound:
1259                         // "If Amazon EC2 is not preparing to stop or
1260                         // terminate the instance, or if you
1261                         // terminated the instance yourself,
1262                         // instance-action is not present in the
1263                         // instance metadata and you receive an HTTP
1264                         // 404 error when you try to retrieve it."
1265                         return nil
1266                 case http.StatusUnauthorized:
1267                         token = ""
1268                         return fmt.Errorf("%s", resp.Status)
1269                 default:
1270                         return fmt.Errorf("%s", resp.Status)
1271                 }
1272                 err = json.NewDecoder(resp.Body).Decode(&metadata)
1273                 if err != nil {
1274                         return err
1275                 }
1276                 return nil
1277         }
1278         failures := 0
1279         var lastmetadata ec2metadata
1280         for range time.NewTicker(spotInterruptionCheckInterval).C {
1281                 err := check()
1282                 if err != nil {
1283                         runner.CrunchLog.Printf("Error checking spot interruptions: %s", err)
1284                         failures++
1285                         if failures > 5 {
1286                                 runner.CrunchLog.Printf("Giving up on checking spot interruptions after too many consecutive failures")
1287                                 return
1288                         }
1289                         continue
1290                 }
1291                 failures = 0
1292                 if metadata != lastmetadata {
1293                         lastmetadata = metadata
1294                         text := fmt.Sprintf("Cloud provider indicates instance action %q scheduled for time %q", metadata.Action, metadata.Time.UTC().Format(time.RFC3339))
1295                         runner.CrunchLog.Printf("%s", text)
1296                         runner.updateRuntimeStatus(arvadosclient.Dict{
1297                                 "warning":          "preemption notice",
1298                                 "warningDetail":    text,
1299                                 "preemptionNotice": text,
1300                         })
1301                         if proc, err := os.FindProcess(os.Getpid()); err == nil {
1302                                 // trigger updateLogs
1303                                 proc.Signal(syscall.SIGUSR1)
1304                         }
1305                 }
1306         }
1307 }
1308
1309 func (runner *ContainerRunner) updateRuntimeStatus(status arvadosclient.Dict) {
1310         err := runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1311                 "container": arvadosclient.Dict{
1312                         "runtime_status": status,
1313                 },
1314         }, nil)
1315         if err != nil {
1316                 runner.CrunchLog.Printf("error updating container runtime_status: %s", err)
1317         }
1318 }
1319
1320 // CaptureOutput saves data from the container's output directory if
1321 // needed, and updates the container output accordingly.
1322 func (runner *ContainerRunner) CaptureOutput(bindmounts map[string]bindmount) error {
1323         if runner.Container.RuntimeConstraints.API {
1324                 // Output may have been set directly by the container, so
1325                 // refresh the container record to check.
1326                 err := runner.DispatcherArvClient.Get("containers", runner.Container.UUID,
1327                         nil, &runner.Container)
1328                 if err != nil {
1329                         return err
1330                 }
1331                 if runner.Container.Output != "" {
1332                         // Container output is already set.
1333                         runner.OutputPDH = &runner.Container.Output
1334                         return nil
1335                 }
1336         }
1337
1338         txt, err := (&copier{
1339                 client:        runner.containerClient,
1340                 arvClient:     runner.ContainerArvClient,
1341                 keepClient:    runner.ContainerKeepClient,
1342                 hostOutputDir: runner.HostOutputDir,
1343                 ctrOutputDir:  runner.Container.OutputPath,
1344                 bindmounts:    bindmounts,
1345                 mounts:        runner.Container.Mounts,
1346                 secretMounts:  runner.SecretMounts,
1347                 logger:        runner.CrunchLog,
1348         }).Copy()
1349         if err != nil {
1350                 return err
1351         }
1352         if n := len(regexp.MustCompile(` [0-9a-f]+\+\S*\+R`).FindAllStringIndex(txt, -1)); n > 0 {
1353                 runner.CrunchLog.Printf("Copying %d data blocks from remote input collections...", n)
1354                 fs, err := (&arvados.Collection{ManifestText: txt}).FileSystem(runner.containerClient, runner.ContainerKeepClient)
1355                 if err != nil {
1356                         return err
1357                 }
1358                 txt, err = fs.MarshalManifest(".")
1359                 if err != nil {
1360                         return err
1361                 }
1362         }
1363         var resp arvados.Collection
1364         err = runner.ContainerArvClient.Create("collections", arvadosclient.Dict{
1365                 "ensure_unique_name": true,
1366                 "collection": arvadosclient.Dict{
1367                         "is_trashed":    true,
1368                         "name":          "output for " + runner.Container.UUID,
1369                         "manifest_text": txt,
1370                 },
1371         }, &resp)
1372         if err != nil {
1373                 return fmt.Errorf("error creating output collection: %v", err)
1374         }
1375         runner.OutputPDH = &resp.PortableDataHash
1376         return nil
1377 }
1378
1379 func (runner *ContainerRunner) CleanupDirs() {
1380         if runner.ArvMount != nil {
1381                 var delay int64 = 8
1382                 umount := exec.Command("arv-mount", fmt.Sprintf("--unmount-timeout=%d", delay), "--unmount", runner.ArvMountPoint)
1383                 umount.Stdout = runner.CrunchLog
1384                 umount.Stderr = runner.CrunchLog
1385                 runner.CrunchLog.Printf("Running %v", umount.Args)
1386                 umnterr := umount.Start()
1387
1388                 if umnterr != nil {
1389                         runner.CrunchLog.Printf("Error unmounting: %v", umnterr)
1390                         runner.ArvMount.Process.Kill()
1391                 } else {
1392                         // If arv-mount --unmount gets stuck for any reason, we
1393                         // don't want to wait for it forever.  Do Wait() in a goroutine
1394                         // so it doesn't block crunch-run.
1395                         umountExit := make(chan error)
1396                         go func() {
1397                                 mnterr := umount.Wait()
1398                                 if mnterr != nil {
1399                                         runner.CrunchLog.Printf("Error unmounting: %v", mnterr)
1400                                 }
1401                                 umountExit <- mnterr
1402                         }()
1403
1404                         for again := true; again; {
1405                                 again = false
1406                                 select {
1407                                 case <-umountExit:
1408                                         umount = nil
1409                                         again = true
1410                                 case <-runner.ArvMountExit:
1411                                         break
1412                                 case <-time.After(time.Duration((delay + 1) * int64(time.Second))):
1413                                         runner.CrunchLog.Printf("Timed out waiting for unmount")
1414                                         if umount != nil {
1415                                                 umount.Process.Kill()
1416                                         }
1417                                         runner.ArvMount.Process.Kill()
1418                                 }
1419                         }
1420                 }
1421                 runner.ArvMount = nil
1422         }
1423
1424         if runner.ArvMountPoint != "" {
1425                 if rmerr := os.Remove(runner.ArvMountPoint); rmerr != nil {
1426                         runner.CrunchLog.Printf("While cleaning up arv-mount directory %s: %v", runner.ArvMountPoint, rmerr)
1427                 }
1428                 runner.ArvMountPoint = ""
1429         }
1430
1431         if rmerr := os.RemoveAll(runner.parentTemp); rmerr != nil {
1432                 runner.CrunchLog.Printf("While cleaning up temporary directory %s: %v", runner.parentTemp, rmerr)
1433         }
1434 }
1435
1436 // CommitLogs posts the collection containing the final container logs.
1437 func (runner *ContainerRunner) CommitLogs() error {
1438         func() {
1439                 // Hold cStateLock to prevent races on CrunchLog (e.g., stop()).
1440                 runner.cStateLock.Lock()
1441                 defer runner.cStateLock.Unlock()
1442
1443                 runner.CrunchLog.Print(runner.finalState)
1444
1445                 if runner.arvMountLog != nil {
1446                         runner.arvMountLog.Close()
1447                 }
1448                 runner.CrunchLog.Close()
1449
1450                 // Closing CrunchLog above allows them to be committed to Keep at this
1451                 // point, but re-open crunch log with ArvClient in case there are any
1452                 // other further errors (such as failing to write the log to Keep!)
1453                 // while shutting down
1454                 runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{
1455                         ArvClient:     runner.DispatcherArvClient,
1456                         UUID:          runner.Container.UUID,
1457                         loggingStream: "crunch-run",
1458                         writeCloser:   nil,
1459                 })
1460                 runner.CrunchLog.Immediate = log.New(os.Stderr, runner.Container.UUID+" ", 0)
1461         }()
1462
1463         if runner.keepstoreLogger != nil {
1464                 // Flush any buffered logs from our local keepstore
1465                 // process.  Discard anything logged after this point
1466                 // -- it won't end up in the log collection, so
1467                 // there's no point writing it to the collectionfs.
1468                 runner.keepstoreLogbuf.SetWriter(io.Discard)
1469                 runner.keepstoreLogger.Close()
1470                 runner.keepstoreLogger = nil
1471         }
1472
1473         if runner.LogsPDH != nil {
1474                 // If we have already assigned something to LogsPDH,
1475                 // we must be closing the re-opened log, which won't
1476                 // end up getting attached to the container record and
1477                 // therefore doesn't need to be saved as a collection
1478                 // -- it exists only to send logs to other channels.
1479                 return nil
1480         }
1481
1482         saved, err := runner.saveLogCollection(true)
1483         if err != nil {
1484                 return fmt.Errorf("error saving log collection: %s", err)
1485         }
1486         runner.logMtx.Lock()
1487         defer runner.logMtx.Unlock()
1488         runner.LogsPDH = &saved.PortableDataHash
1489         return nil
1490 }
1491
1492 func (runner *ContainerRunner) saveLogCollection(final bool) (response arvados.Collection, err error) {
1493         runner.logMtx.Lock()
1494         defer runner.logMtx.Unlock()
1495         if runner.LogsPDH != nil {
1496                 // Already finalized.
1497                 return
1498         }
1499         updates := arvadosclient.Dict{
1500                 "name": "logs for " + runner.Container.UUID,
1501         }
1502         mt, err1 := runner.LogCollection.MarshalManifest(".")
1503         if err1 == nil {
1504                 // Only send updated manifest text if there was no
1505                 // error.
1506                 updates["manifest_text"] = mt
1507         }
1508
1509         // Even if flushing the manifest had an error, we still want
1510         // to update the log record, if possible, to push the trash_at
1511         // and delete_at times into the future.  Details on bug
1512         // #17293.
1513         if final {
1514                 updates["is_trashed"] = true
1515         } else {
1516                 exp := time.Now().Add(crunchLogUpdatePeriod * 24)
1517                 updates["trash_at"] = exp
1518                 updates["delete_at"] = exp
1519         }
1520         reqBody := arvadosclient.Dict{"collection": updates}
1521         var err2 error
1522         if runner.logUUID == "" {
1523                 reqBody["ensure_unique_name"] = true
1524                 err2 = runner.DispatcherArvClient.Create("collections", reqBody, &response)
1525         } else {
1526                 err2 = runner.DispatcherArvClient.Update("collections", runner.logUUID, reqBody, &response)
1527         }
1528         if err2 == nil {
1529                 runner.logUUID = response.UUID
1530         }
1531
1532         if err1 != nil || err2 != nil {
1533                 err = fmt.Errorf("error recording logs: %q, %q", err1, err2)
1534         }
1535         return
1536 }
1537
1538 // UpdateContainerRunning updates the container state to "Running"
1539 func (runner *ContainerRunner) UpdateContainerRunning(logId string) error {
1540         runner.cStateLock.Lock()
1541         defer runner.cStateLock.Unlock()
1542         if runner.cCancelled {
1543                 return ErrCancelled
1544         }
1545         updates := arvadosclient.Dict{
1546                 "gateway_address": runner.gateway.Address,
1547                 "state":           "Running",
1548         }
1549         if logId != "" {
1550                 updates["log"] = logId
1551         }
1552         return runner.DispatcherArvClient.Update(
1553                 "containers",
1554                 runner.Container.UUID,
1555                 arvadosclient.Dict{"container": updates},
1556                 nil,
1557         )
1558 }
1559
1560 // ContainerToken returns the api_token the container (and any
1561 // arv-mount processes) are allowed to use.
1562 func (runner *ContainerRunner) ContainerToken() (string, error) {
1563         if runner.token != "" {
1564                 return runner.token, nil
1565         }
1566
1567         var auth arvados.APIClientAuthorization
1568         err := runner.DispatcherArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth)
1569         if err != nil {
1570                 return "", err
1571         }
1572         runner.token = fmt.Sprintf("v2/%s/%s/%s", auth.UUID, auth.APIToken, runner.Container.UUID)
1573         return runner.token, nil
1574 }
1575
1576 // UpdateContainerFinal updates the container record state on API
1577 // server to "Complete" or "Cancelled"
1578 func (runner *ContainerRunner) UpdateContainerFinal() error {
1579         update := arvadosclient.Dict{}
1580         update["state"] = runner.finalState
1581         if runner.LogsPDH != nil {
1582                 update["log"] = *runner.LogsPDH
1583         }
1584         if runner.ExitCode != nil {
1585                 update["exit_code"] = *runner.ExitCode
1586         } else {
1587                 update["exit_code"] = nil
1588         }
1589         if runner.finalState == "Complete" && runner.OutputPDH != nil {
1590                 update["output"] = *runner.OutputPDH
1591         }
1592         update["cost"] = runner.calculateCost(time.Now())
1593         return runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{"container": update}, nil)
1594 }
1595
1596 // IsCancelled returns the value of Cancelled, with goroutine safety.
1597 func (runner *ContainerRunner) IsCancelled() bool {
1598         runner.cStateLock.Lock()
1599         defer runner.cStateLock.Unlock()
1600         return runner.cCancelled
1601 }
1602
1603 // NewArvLogWriter creates an ArvLogWriter
1604 func (runner *ContainerRunner) NewArvLogWriter(name string) (io.WriteCloser, error) {
1605         writer, err := runner.LogCollection.OpenFile(name+".txt", os.O_CREATE|os.O_WRONLY, 0666)
1606         if err != nil {
1607                 return nil, err
1608         }
1609         return &ArvLogWriter{
1610                 ArvClient:     runner.DispatcherArvClient,
1611                 UUID:          runner.Container.UUID,
1612                 loggingStream: name,
1613                 writeCloser:   writer,
1614         }, nil
1615 }
1616
1617 // Run the full container lifecycle.
1618 func (runner *ContainerRunner) Run() (err error) {
1619         runner.CrunchLog.Printf("crunch-run %s started", cmd.Version.String())
1620         runner.CrunchLog.Printf("%s", currentUserAndGroups())
1621         v, _ := exec.Command("arv-mount", "--version").CombinedOutput()
1622         runner.CrunchLog.Printf("Using FUSE mount: %s", v)
1623         runner.CrunchLog.Printf("Using container runtime: %s", runner.executor.Runtime())
1624         runner.CrunchLog.Printf("Executing container: %s", runner.Container.UUID)
1625         runner.costStartTime = time.Now()
1626
1627         hostname, hosterr := os.Hostname()
1628         if hosterr != nil {
1629                 runner.CrunchLog.Printf("Error getting hostname '%v'", hosterr)
1630         } else {
1631                 runner.CrunchLog.Printf("Executing on host '%s'", hostname)
1632         }
1633
1634         sigusr2 := make(chan os.Signal, 1)
1635         signal.Notify(sigusr2, syscall.SIGUSR2)
1636         defer signal.Stop(sigusr2)
1637         runner.loadPrices()
1638         go func() {
1639                 for range sigusr2 {
1640                         runner.loadPrices()
1641                 }
1642         }()
1643
1644         runner.finalState = "Queued"
1645
1646         defer func() {
1647                 runner.CleanupDirs()
1648
1649                 runner.CrunchLog.Printf("crunch-run finished")
1650                 runner.CrunchLog.Close()
1651         }()
1652
1653         err = runner.fetchContainerRecord()
1654         if err != nil {
1655                 return
1656         }
1657         if runner.Container.State != "Locked" {
1658                 return fmt.Errorf("dispatch error detected: container %q has state %q", runner.Container.UUID, runner.Container.State)
1659         }
1660
1661         var bindmounts map[string]bindmount
1662         defer func() {
1663                 // checkErr prints e (unless it's nil) and sets err to
1664                 // e (unless err is already non-nil). Thus, if err
1665                 // hasn't already been assigned when Run() returns,
1666                 // this cleanup func will cause Run() to return the
1667                 // first non-nil error that is passed to checkErr().
1668                 checkErr := func(errorIn string, e error) {
1669                         if e == nil {
1670                                 return
1671                         }
1672                         runner.CrunchLog.Printf("error in %s: %v", errorIn, e)
1673                         if err == nil {
1674                                 err = e
1675                         }
1676                         if runner.finalState == "Complete" {
1677                                 // There was an error in the finalization.
1678                                 runner.finalState = "Cancelled"
1679                         }
1680                 }
1681
1682                 // Log the error encountered in Run(), if any
1683                 checkErr("Run", err)
1684
1685                 if runner.finalState == "Queued" {
1686                         runner.UpdateContainerFinal()
1687                         return
1688                 }
1689
1690                 if runner.IsCancelled() {
1691                         runner.finalState = "Cancelled"
1692                         // but don't return yet -- we still want to
1693                         // capture partial output and write logs
1694                 }
1695
1696                 if bindmounts != nil {
1697                         checkErr("CaptureOutput", runner.CaptureOutput(bindmounts))
1698                 }
1699                 checkErr("stopHoststat", runner.stopHoststat())
1700                 checkErr("CommitLogs", runner.CommitLogs())
1701                 runner.CleanupDirs()
1702                 checkErr("UpdateContainerFinal", runner.UpdateContainerFinal())
1703         }()
1704
1705         runner.setupSignals()
1706         err = runner.startHoststat()
1707         if err != nil {
1708                 return
1709         }
1710         if runner.keepstore != nil {
1711                 runner.hoststatReporter.ReportPID("keepstore", runner.keepstore.Process.Pid)
1712         }
1713
1714         // set up FUSE mount and binds
1715         bindmounts, err = runner.SetupMounts()
1716         if err != nil {
1717                 runner.finalState = "Cancelled"
1718                 err = fmt.Errorf("While setting up mounts: %v", err)
1719                 return
1720         }
1721
1722         // check for and/or load image
1723         imageID, err := runner.LoadImage()
1724         if err != nil {
1725                 if !runner.checkBrokenNode(err) {
1726                         // Failed to load image but not due to a "broken node"
1727                         // condition, probably user error.
1728                         runner.finalState = "Cancelled"
1729                 }
1730                 err = fmt.Errorf("While loading container image: %v", err)
1731                 return
1732         }
1733
1734         err = runner.CreateContainer(imageID, bindmounts)
1735         if err != nil {
1736                 return
1737         }
1738         err = runner.LogHostInfo()
1739         if err != nil {
1740                 return
1741         }
1742         err = runner.LogNodeRecord()
1743         if err != nil {
1744                 return
1745         }
1746         err = runner.LogContainerRecord()
1747         if err != nil {
1748                 return
1749         }
1750
1751         if runner.IsCancelled() {
1752                 return
1753         }
1754
1755         logCollection, err := runner.saveLogCollection(false)
1756         var logId string
1757         if err == nil {
1758                 logId = logCollection.PortableDataHash
1759         } else {
1760                 runner.CrunchLog.Printf("Error committing initial log collection: %v", err)
1761         }
1762         err = runner.UpdateContainerRunning(logId)
1763         if err != nil {
1764                 return
1765         }
1766         runner.finalState = "Cancelled"
1767
1768         err = runner.startCrunchstat()
1769         if err != nil {
1770                 return
1771         }
1772
1773         err = runner.StartContainer()
1774         if err != nil {
1775                 runner.checkBrokenNode(err)
1776                 return
1777         }
1778
1779         err = runner.WaitFinish()
1780         if err == nil && !runner.IsCancelled() {
1781                 runner.finalState = "Complete"
1782         }
1783         return
1784 }
1785
1786 // Fetch the current container record (uuid = runner.Container.UUID)
1787 // into runner.Container.
1788 func (runner *ContainerRunner) fetchContainerRecord() error {
1789         reader, err := runner.DispatcherArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
1790         if err != nil {
1791                 return fmt.Errorf("error fetching container record: %v", err)
1792         }
1793         defer reader.Close()
1794
1795         dec := json.NewDecoder(reader)
1796         dec.UseNumber()
1797         err = dec.Decode(&runner.Container)
1798         if err != nil {
1799                 return fmt.Errorf("error decoding container record: %v", err)
1800         }
1801
1802         var sm struct {
1803                 SecretMounts map[string]arvados.Mount `json:"secret_mounts"`
1804         }
1805
1806         containerToken, err := runner.ContainerToken()
1807         if err != nil {
1808                 return fmt.Errorf("error getting container token: %v", err)
1809         }
1810
1811         runner.ContainerArvClient, runner.ContainerKeepClient,
1812                 runner.containerClient, err = runner.MkArvClient(containerToken)
1813         if err != nil {
1814                 return fmt.Errorf("error creating container API client: %v", err)
1815         }
1816
1817         runner.ContainerKeepClient.SetStorageClasses(runner.Container.OutputStorageClasses)
1818         runner.DispatcherKeepClient.SetStorageClasses(runner.Container.OutputStorageClasses)
1819
1820         err = runner.ContainerArvClient.Call("GET", "containers", runner.Container.UUID, "secret_mounts", nil, &sm)
1821         if err != nil {
1822                 if apierr, ok := err.(arvadosclient.APIServerError); !ok || apierr.HttpStatusCode != 404 {
1823                         return fmt.Errorf("error fetching secret_mounts: %v", err)
1824                 }
1825                 // ok && apierr.HttpStatusCode == 404, which means
1826                 // secret_mounts isn't supported by this API server.
1827         }
1828         runner.SecretMounts = sm.SecretMounts
1829
1830         return nil
1831 }
1832
1833 // NewContainerRunner creates a new container runner.
1834 func NewContainerRunner(dispatcherClient *arvados.Client,
1835         dispatcherArvClient IArvadosClient,
1836         dispatcherKeepClient IKeepClient,
1837         containerUUID string) (*ContainerRunner, error) {
1838
1839         cr := &ContainerRunner{
1840                 dispatcherClient:     dispatcherClient,
1841                 DispatcherArvClient:  dispatcherArvClient,
1842                 DispatcherKeepClient: dispatcherKeepClient,
1843         }
1844         cr.NewLogWriter = cr.NewArvLogWriter
1845         cr.RunArvMount = cr.ArvMountCmd
1846         cr.MkTempDir = ioutil.TempDir
1847         cr.MkArvClient = func(token string) (IArvadosClient, IKeepClient, *arvados.Client, error) {
1848                 cl, err := arvadosclient.MakeArvadosClient()
1849                 if err != nil {
1850                         return nil, nil, nil, err
1851                 }
1852                 cl.ApiToken = token
1853                 kc, err := keepclient.MakeKeepClient(cl)
1854                 if err != nil {
1855                         return nil, nil, nil, err
1856                 }
1857                 c2 := arvados.NewClientFromEnv()
1858                 c2.AuthToken = token
1859                 return cl, kc, c2, nil
1860         }
1861         var err error
1862         cr.LogCollection, err = (&arvados.Collection{}).FileSystem(cr.dispatcherClient, cr.DispatcherKeepClient)
1863         if err != nil {
1864                 return nil, err
1865         }
1866         cr.Container.UUID = containerUUID
1867         w, err := cr.NewLogWriter("crunch-run")
1868         if err != nil {
1869                 return nil, err
1870         }
1871         cr.CrunchLog = NewThrottledLogger(w)
1872         cr.CrunchLog.Immediate = log.New(os.Stderr, containerUUID+" ", 0)
1873
1874         loadLogThrottleParams(dispatcherArvClient)
1875         go cr.updateLogs()
1876
1877         return cr, nil
1878 }
1879
1880 func (command) RunCommand(prog string, args []string, stdin io.Reader, stdout, stderr io.Writer) int {
1881         log := log.New(stderr, "", 0)
1882         flags := flag.NewFlagSet(prog, flag.ContinueOnError)
1883         statInterval := flags.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting")
1884         cgroupRoot := flags.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree")
1885         cgroupParent := flags.String("cgroup-parent", "docker", "name of container's parent cgroup (ignored if -cgroup-parent-subsystem is used)")
1886         cgroupParentSubsystem := flags.String("cgroup-parent-subsystem", "", "use current cgroup for given subsystem as parent cgroup for container")
1887         caCertsPath := flags.String("ca-certs", "", "Path to TLS root certificates")
1888         detach := flags.Bool("detach", false, "Detach from parent process and run in the background")
1889         stdinConfig := flags.Bool("stdin-config", false, "Load config and environment variables from JSON message on stdin")
1890         configFile := flags.String("config", arvados.DefaultConfigFile, "filename of cluster config file to try loading if -stdin-config=false (default is $ARVADOS_CONFIG)")
1891         sleep := flags.Duration("sleep", 0, "Delay before starting (testing use only)")
1892         kill := flags.Int("kill", -1, "Send signal to an existing crunch-run process for given UUID")
1893         list := flags.Bool("list", false, "List UUIDs of existing crunch-run processes (and notify them to use price data passed on stdin)")
1894         enableMemoryLimit := flags.Bool("enable-memory-limit", true, "tell container runtime to limit container's memory usage")
1895         enableNetwork := flags.String("container-enable-networking", "default", "enable networking \"always\" (for all containers) or \"default\" (for containers that request it)")
1896         networkMode := flags.String("container-network-mode", "default", `Docker network mode for container (use any argument valid for docker --net)`)
1897         memprofile := flags.String("memprofile", "", "write memory profile to `file` after running container")
1898         runtimeEngine := flags.String("runtime-engine", "docker", "container runtime: docker or singularity")
1899         brokenNodeHook := flags.String("broken-node-hook", "", "script to run if node is detected to be broken (for example, Docker daemon is not running)")
1900         flags.Duration("check-containerd", 0, "Ignored. Exists for compatibility with older versions.")
1901         version := flags.Bool("version", false, "Write version information to stdout and exit 0.")
1902
1903         ignoreDetachFlag := false
1904         if len(args) > 0 && args[0] == "-no-detach" {
1905                 // This process was invoked by a parent process, which
1906                 // has passed along its own arguments, including
1907                 // -detach, after the leading -no-detach flag.  Strip
1908                 // the leading -no-detach flag (it's not recognized by
1909                 // flags.Parse()) and ignore the -detach flag that
1910                 // comes later.
1911                 args = args[1:]
1912                 ignoreDetachFlag = true
1913         }
1914
1915         if ok, code := cmd.ParseFlags(flags, prog, args, "container-uuid", stderr); !ok {
1916                 return code
1917         } else if *version {
1918                 fmt.Fprintln(stdout, prog, cmd.Version.String())
1919                 return 0
1920         } else if !*list && flags.NArg() != 1 {
1921                 fmt.Fprintf(stderr, "missing required argument: container-uuid (try -help)\n")
1922                 return 2
1923         }
1924
1925         containerUUID := flags.Arg(0)
1926
1927         switch {
1928         case *detach && !ignoreDetachFlag:
1929                 return Detach(containerUUID, prog, args, stdin, stdout, stderr)
1930         case *kill >= 0:
1931                 return KillProcess(containerUUID, syscall.Signal(*kill), stdout, stderr)
1932         case *list:
1933                 return ListProcesses(stdin, stdout, stderr)
1934         }
1935
1936         if len(containerUUID) != 27 {
1937                 log.Printf("usage: %s [options] UUID", prog)
1938                 return 1
1939         }
1940
1941         var keepstoreLogbuf bufThenWrite
1942         var conf ConfigData
1943         if *stdinConfig {
1944                 err := json.NewDecoder(stdin).Decode(&conf)
1945                 if err != nil {
1946                         log.Printf("decode stdin: %s", err)
1947                         return 1
1948                 }
1949                 for k, v := range conf.Env {
1950                         err = os.Setenv(k, v)
1951                         if err != nil {
1952                                 log.Printf("setenv(%q): %s", k, err)
1953                                 return 1
1954                         }
1955                 }
1956                 if conf.Cluster != nil {
1957                         // ClusterID is missing from the JSON
1958                         // representation, but we need it to generate
1959                         // a valid config file for keepstore, so we
1960                         // fill it using the container UUID prefix.
1961                         conf.Cluster.ClusterID = containerUUID[:5]
1962                 }
1963         } else {
1964                 conf = hpcConfData(containerUUID, *configFile, io.MultiWriter(&keepstoreLogbuf, stderr))
1965         }
1966
1967         log.Printf("crunch-run %s started", cmd.Version.String())
1968         time.Sleep(*sleep)
1969
1970         if *caCertsPath != "" {
1971                 arvadosclient.CertFiles = []string{*caCertsPath}
1972         }
1973
1974         keepstore, err := startLocalKeepstore(conf, io.MultiWriter(&keepstoreLogbuf, stderr))
1975         if err != nil {
1976                 log.Print(err)
1977                 return 1
1978         }
1979         if keepstore != nil {
1980                 defer keepstore.Process.Kill()
1981         }
1982
1983         api, err := arvadosclient.MakeArvadosClient()
1984         if err != nil {
1985                 log.Printf("%s: %v", containerUUID, err)
1986                 return 1
1987         }
1988         api.Retries = 8
1989
1990         kc, err := keepclient.MakeKeepClient(api)
1991         if err != nil {
1992                 log.Printf("%s: %v", containerUUID, err)
1993                 return 1
1994         }
1995         kc.BlockCache = &keepclient.BlockCache{MaxBlocks: 2}
1996         kc.Retries = 4
1997
1998         cr, err := NewContainerRunner(arvados.NewClientFromEnv(), api, kc, containerUUID)
1999         if err != nil {
2000                 log.Print(err)
2001                 return 1
2002         }
2003
2004         cr.keepstore = keepstore
2005         if keepstore == nil {
2006                 // Log explanation (if any) for why we're not running
2007                 // a local keepstore.
2008                 var buf bytes.Buffer
2009                 keepstoreLogbuf.SetWriter(&buf)
2010                 if buf.Len() > 0 {
2011                         cr.CrunchLog.Printf("%s", strings.TrimSpace(buf.String()))
2012                 }
2013         } else if logWhat := conf.Cluster.Containers.LocalKeepLogsToContainerLog; logWhat == "none" {
2014                 cr.CrunchLog.Printf("using local keepstore process (pid %d) at %s", keepstore.Process.Pid, os.Getenv("ARVADOS_KEEP_SERVICES"))
2015                 keepstoreLogbuf.SetWriter(io.Discard)
2016         } else {
2017                 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"))
2018                 logwriter, err := cr.NewLogWriter("keepstore")
2019                 if err != nil {
2020                         log.Print(err)
2021                         return 1
2022                 }
2023                 cr.keepstoreLogger = NewThrottledLogger(logwriter)
2024
2025                 var writer io.WriteCloser = cr.keepstoreLogger
2026                 if logWhat == "errors" {
2027                         writer = &filterKeepstoreErrorsOnly{WriteCloser: writer}
2028                 } else if logWhat != "all" {
2029                         // should have been caught earlier by
2030                         // dispatcher's config loader
2031                         log.Printf("invalid value for Containers.LocalKeepLogsToContainerLog: %q", logWhat)
2032                         return 1
2033                 }
2034                 err = keepstoreLogbuf.SetWriter(writer)
2035                 if err != nil {
2036                         log.Print(err)
2037                         return 1
2038                 }
2039                 cr.keepstoreLogbuf = &keepstoreLogbuf
2040         }
2041
2042         switch *runtimeEngine {
2043         case "docker":
2044                 cr.executor, err = newDockerExecutor(containerUUID, cr.CrunchLog.Printf, cr.containerWatchdogInterval)
2045         case "singularity":
2046                 cr.executor, err = newSingularityExecutor(cr.CrunchLog.Printf)
2047         default:
2048                 cr.CrunchLog.Printf("%s: unsupported RuntimeEngine %q", containerUUID, *runtimeEngine)
2049                 cr.CrunchLog.Close()
2050                 return 1
2051         }
2052         if err != nil {
2053                 cr.CrunchLog.Printf("%s: %v", containerUUID, err)
2054                 cr.checkBrokenNode(err)
2055                 cr.CrunchLog.Close()
2056                 return 1
2057         }
2058         defer cr.executor.Close()
2059
2060         cr.brokenNodeHook = *brokenNodeHook
2061
2062         gwAuthSecret := os.Getenv("GatewayAuthSecret")
2063         os.Unsetenv("GatewayAuthSecret")
2064         if gwAuthSecret == "" {
2065                 // not safe to run a gateway service without an auth
2066                 // secret
2067                 cr.CrunchLog.Printf("Not starting a gateway server (GatewayAuthSecret was not provided by dispatcher)")
2068         } else {
2069                 gwListen := os.Getenv("GatewayAddress")
2070                 cr.gateway = Gateway{
2071                         Address:       gwListen,
2072                         AuthSecret:    gwAuthSecret,
2073                         ContainerUUID: containerUUID,
2074                         Target:        cr.executor,
2075                         Log:           cr.CrunchLog,
2076                 }
2077                 if gwListen == "" {
2078                         // Direct connection won't work, so we use the
2079                         // gateway_address field to indicate the
2080                         // internalURL of the controller process that
2081                         // has the current tunnel connection.
2082                         cr.gateway.ArvadosClient = cr.dispatcherClient
2083                         cr.gateway.UpdateTunnelURL = func(url string) {
2084                                 cr.gateway.Address = "tunnel " + url
2085                                 cr.DispatcherArvClient.Update("containers", containerUUID,
2086                                         arvadosclient.Dict{"container": arvadosclient.Dict{"gateway_address": cr.gateway.Address}}, nil)
2087                         }
2088                 }
2089                 err = cr.gateway.Start()
2090                 if err != nil {
2091                         log.Printf("error starting gateway server: %s", err)
2092                         return 1
2093                 }
2094         }
2095
2096         parentTemp, tmperr := cr.MkTempDir("", "crunch-run."+containerUUID+".")
2097         if tmperr != nil {
2098                 log.Printf("%s: %v", containerUUID, tmperr)
2099                 return 1
2100         }
2101
2102         cr.parentTemp = parentTemp
2103         cr.statInterval = *statInterval
2104         cr.cgroupRoot = *cgroupRoot
2105         cr.expectCgroupParent = *cgroupParent
2106         cr.enableMemoryLimit = *enableMemoryLimit
2107         cr.enableNetwork = *enableNetwork
2108         cr.networkMode = *networkMode
2109         if *cgroupParentSubsystem != "" {
2110                 p, err := findCgroup(*cgroupParentSubsystem)
2111                 if err != nil {
2112                         log.Printf("fatal: cgroup parent subsystem: %s", err)
2113                         return 1
2114                 }
2115                 cr.setCgroupParent = p
2116                 cr.expectCgroupParent = p
2117         }
2118
2119         if conf.EC2SpotCheck {
2120                 go cr.checkSpotInterruptionNotices()
2121         }
2122
2123         runerr := cr.Run()
2124
2125         if *memprofile != "" {
2126                 f, err := os.Create(*memprofile)
2127                 if err != nil {
2128                         log.Printf("could not create memory profile: %s", err)
2129                 }
2130                 runtime.GC() // get up-to-date statistics
2131                 if err := pprof.WriteHeapProfile(f); err != nil {
2132                         log.Printf("could not write memory profile: %s", err)
2133                 }
2134                 closeerr := f.Close()
2135                 if closeerr != nil {
2136                         log.Printf("closing memprofile file: %s", err)
2137                 }
2138         }
2139
2140         if runerr != nil {
2141                 log.Printf("%s: %v", containerUUID, runerr)
2142                 return 1
2143         }
2144         return 0
2145 }
2146
2147 // Try to load ConfigData in hpc (slurm/lsf) environment. This means
2148 // loading the cluster config from the specified file and (if that
2149 // works) getting the runtime_constraints container field from
2150 // controller to determine # VCPUs so we can calculate KeepBuffers.
2151 func hpcConfData(uuid string, configFile string, stderr io.Writer) ConfigData {
2152         var conf ConfigData
2153         conf.Cluster = loadClusterConfigFile(configFile, stderr)
2154         if conf.Cluster == nil {
2155                 // skip loading the container record -- we won't be
2156                 // able to start local keepstore anyway.
2157                 return conf
2158         }
2159         arv, err := arvadosclient.MakeArvadosClient()
2160         if err != nil {
2161                 fmt.Fprintf(stderr, "error setting up arvadosclient: %s\n", err)
2162                 return conf
2163         }
2164         arv.Retries = 8
2165         var ctr arvados.Container
2166         err = arv.Call("GET", "containers", uuid, "", arvadosclient.Dict{"select": []string{"runtime_constraints"}}, &ctr)
2167         if err != nil {
2168                 fmt.Fprintf(stderr, "error getting container record: %s\n", err)
2169                 return conf
2170         }
2171         if ctr.RuntimeConstraints.VCPUs > 0 {
2172                 conf.KeepBuffers = ctr.RuntimeConstraints.VCPUs * conf.Cluster.Containers.LocalKeepBlobBuffersPerVCPU
2173         }
2174         return conf
2175 }
2176
2177 // Load cluster config file from given path. If an error occurs, log
2178 // the error to stderr and return nil.
2179 func loadClusterConfigFile(path string, stderr io.Writer) *arvados.Cluster {
2180         ldr := config.NewLoader(&bytes.Buffer{}, ctxlog.New(stderr, "plain", "info"))
2181         ldr.Path = path
2182         cfg, err := ldr.Load()
2183         if err != nil {
2184                 fmt.Fprintf(stderr, "could not load config file %s: %s\n", path, err)
2185                 return nil
2186         }
2187         cluster, err := cfg.GetCluster("")
2188         if err != nil {
2189                 fmt.Fprintf(stderr, "could not use config file %s: %s\n", path, err)
2190                 return nil
2191         }
2192         fmt.Fprintf(stderr, "loaded config file %s\n", path)
2193         return cluster
2194 }
2195
2196 func startLocalKeepstore(configData ConfigData, logbuf io.Writer) (*exec.Cmd, error) {
2197         if configData.KeepBuffers < 1 {
2198                 fmt.Fprintf(logbuf, "not starting a local keepstore process because KeepBuffers=%v in config\n", configData.KeepBuffers)
2199                 return nil, nil
2200         }
2201         if configData.Cluster == nil {
2202                 fmt.Fprint(logbuf, "not starting a local keepstore process because cluster config file was not loaded\n")
2203                 return nil, nil
2204         }
2205         for uuid, vol := range configData.Cluster.Volumes {
2206                 if len(vol.AccessViaHosts) > 0 {
2207                         fmt.Fprintf(logbuf, "not starting a local keepstore process because a volume (%s) uses AccessViaHosts\n", uuid)
2208                         return nil, nil
2209                 }
2210                 if !vol.ReadOnly && vol.Replication < configData.Cluster.Collections.DefaultReplication {
2211                         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)
2212                         return nil, nil
2213                 }
2214         }
2215
2216         // Rather than have an alternate way to tell keepstore how
2217         // many buffers to use when starting it this way, we just
2218         // modify the cluster configuration that we feed it on stdin.
2219         configData.Cluster.API.MaxKeepBlobBuffers = configData.KeepBuffers
2220
2221         localaddr := localKeepstoreAddr()
2222         ln, err := net.Listen("tcp", net.JoinHostPort(localaddr, "0"))
2223         if err != nil {
2224                 return nil, err
2225         }
2226         _, port, err := net.SplitHostPort(ln.Addr().String())
2227         if err != nil {
2228                 ln.Close()
2229                 return nil, err
2230         }
2231         ln.Close()
2232         url := "http://" + net.JoinHostPort(localaddr, port)
2233
2234         fmt.Fprintf(logbuf, "starting keepstore on %s\n", url)
2235
2236         var confJSON bytes.Buffer
2237         err = json.NewEncoder(&confJSON).Encode(arvados.Config{
2238                 Clusters: map[string]arvados.Cluster{
2239                         configData.Cluster.ClusterID: *configData.Cluster,
2240                 },
2241         })
2242         if err != nil {
2243                 return nil, err
2244         }
2245         cmd := exec.Command("/proc/self/exe", "keepstore", "-config=-")
2246         if target, err := os.Readlink(cmd.Path); err == nil && strings.HasSuffix(target, ".test") {
2247                 // If we're a 'go test' process, running
2248                 // /proc/self/exe would start the test suite in a
2249                 // child process, which is not what we want.
2250                 cmd.Path, _ = exec.LookPath("go")
2251                 cmd.Args = append([]string{"go", "run", "../../cmd/arvados-server"}, cmd.Args[1:]...)
2252                 cmd.Env = os.Environ()
2253         }
2254         cmd.Stdin = &confJSON
2255         cmd.Stdout = logbuf
2256         cmd.Stderr = logbuf
2257         cmd.Env = append(cmd.Env,
2258                 "GOGC=10",
2259                 "ARVADOS_SERVICE_INTERNAL_URL="+url)
2260         err = cmd.Start()
2261         if err != nil {
2262                 return nil, fmt.Errorf("error starting keepstore process: %w", err)
2263         }
2264         cmdExited := false
2265         go func() {
2266                 cmd.Wait()
2267                 cmdExited = true
2268         }()
2269         ctx, cancel := context.WithDeadline(context.Background(), time.Now().Add(time.Second*10))
2270         defer cancel()
2271         poll := time.NewTicker(time.Second / 10)
2272         defer poll.Stop()
2273         client := http.Client{}
2274         for range poll.C {
2275                 testReq, err := http.NewRequestWithContext(ctx, "GET", url+"/_health/ping", nil)
2276                 testReq.Header.Set("Authorization", "Bearer "+configData.Cluster.ManagementToken)
2277                 if err != nil {
2278                         return nil, err
2279                 }
2280                 resp, err := client.Do(testReq)
2281                 if err == nil {
2282                         resp.Body.Close()
2283                         if resp.StatusCode == http.StatusOK {
2284                                 break
2285                         }
2286                 }
2287                 if cmdExited {
2288                         return nil, fmt.Errorf("keepstore child process exited")
2289                 }
2290                 if ctx.Err() != nil {
2291                         return nil, fmt.Errorf("timed out waiting for new keepstore process to report healthy")
2292                 }
2293         }
2294         os.Setenv("ARVADOS_KEEP_SERVICES", url)
2295         return cmd, nil
2296 }
2297
2298 // return current uid, gid, groups in a format suitable for logging:
2299 // "crunch-run process has uid=1234(arvados) gid=1234(arvados)
2300 // groups=1234(arvados),114(fuse)"
2301 func currentUserAndGroups() string {
2302         u, err := user.Current()
2303         if err != nil {
2304                 return fmt.Sprintf("error getting current user ID: %s", err)
2305         }
2306         s := fmt.Sprintf("crunch-run process has uid=%s(%s) gid=%s", u.Uid, u.Username, u.Gid)
2307         if g, err := user.LookupGroupId(u.Gid); err == nil {
2308                 s += fmt.Sprintf("(%s)", g.Name)
2309         }
2310         s += " groups="
2311         if gids, err := u.GroupIds(); err == nil {
2312                 for i, gid := range gids {
2313                         if i > 0 {
2314                                 s += ","
2315                         }
2316                         s += gid
2317                         if g, err := user.LookupGroupId(gid); err == nil {
2318                                 s += fmt.Sprintf("(%s)", g.Name)
2319                         }
2320                 }
2321         }
2322         return s
2323 }
2324
2325 // Return a suitable local interface address for a local keepstore
2326 // service. Currently this is the numerically lowest non-loopback ipv4
2327 // address assigned to a local interface that is not in any of the
2328 // link-local/vpn/loopback ranges 169.254/16, 100.64/10, or 127/8.
2329 func localKeepstoreAddr() string {
2330         var ips []net.IP
2331         // Ignore error (proceed with zero IPs)
2332         addrs, _ := processIPs(os.Getpid())
2333         for addr := range addrs {
2334                 ip := net.ParseIP(addr)
2335                 if ip == nil {
2336                         // invalid
2337                         continue
2338                 }
2339                 if ip.Mask(net.CIDRMask(8, 32)).Equal(net.IPv4(127, 0, 0, 0)) ||
2340                         ip.Mask(net.CIDRMask(10, 32)).Equal(net.IPv4(100, 64, 0, 0)) ||
2341                         ip.Mask(net.CIDRMask(16, 32)).Equal(net.IPv4(169, 254, 0, 0)) {
2342                         // unsuitable
2343                         continue
2344                 }
2345                 ips = append(ips, ip)
2346         }
2347         if len(ips) == 0 {
2348                 return "0.0.0.0"
2349         }
2350         sort.Slice(ips, func(ii, jj int) bool {
2351                 i, j := ips[ii], ips[jj]
2352                 if len(i) != len(j) {
2353                         return len(i) < len(j)
2354                 }
2355                 for x := range i {
2356                         if i[x] != j[x] {
2357                                 return i[x] < j[x]
2358                         }
2359                 }
2360                 return false
2361         })
2362         return ips[0].String()
2363 }
2364
2365 func (cr *ContainerRunner) loadPrices() {
2366         buf, err := os.ReadFile(filepath.Join(lockdir, pricesfile))
2367         if err != nil {
2368                 if !os.IsNotExist(err) {
2369                         cr.CrunchLog.Printf("loadPrices: read: %s", err)
2370                 }
2371                 return
2372         }
2373         var prices []cloud.InstancePrice
2374         err = json.Unmarshal(buf, &prices)
2375         if err != nil {
2376                 cr.CrunchLog.Printf("loadPrices: decode: %s", err)
2377                 return
2378         }
2379         cr.pricesLock.Lock()
2380         defer cr.pricesLock.Unlock()
2381         var lastKnown time.Time
2382         if len(cr.prices) > 0 {
2383                 lastKnown = cr.prices[0].StartTime
2384         }
2385         cr.prices = cloud.NormalizePriceHistory(append(prices, cr.prices...))
2386         for i := len(cr.prices) - 1; i >= 0; i-- {
2387                 price := cr.prices[i]
2388                 if price.StartTime.After(lastKnown) {
2389                         cr.CrunchLog.Printf("Instance price changed to %#.3g at %s", price.Price, price.StartTime.UTC())
2390                 }
2391         }
2392 }
2393
2394 func (cr *ContainerRunner) calculateCost(now time.Time) float64 {
2395         cr.pricesLock.Lock()
2396         defer cr.pricesLock.Unlock()
2397
2398         // First, make a "prices" slice with the real data as far back
2399         // as it goes, and (if needed) a "since the beginning of time"
2400         // placeholder containing a reasonable guess about what the
2401         // price was between cr.costStartTime and the earliest real
2402         // data point.
2403         prices := cr.prices
2404         if len(prices) == 0 {
2405                 // use price info in InstanceType record initially
2406                 // provided by cloud dispatcher
2407                 var p float64
2408                 var it arvados.InstanceType
2409                 if j := os.Getenv("InstanceType"); j != "" && json.Unmarshal([]byte(j), &it) == nil && it.Price > 0 {
2410                         p = it.Price
2411                 }
2412                 prices = []cloud.InstancePrice{{Price: p}}
2413         } else if prices[len(prices)-1].StartTime.After(cr.costStartTime) {
2414                 // guess earlier pricing was the same as the earliest
2415                 // price we know about
2416                 filler := prices[len(prices)-1]
2417                 filler.StartTime = time.Time{}
2418                 prices = append(prices, filler)
2419         }
2420
2421         // Now that our history of price changes goes back at least as
2422         // far as cr.costStartTime, add up the costs for each
2423         // interval.
2424         cost := 0.0
2425         spanEnd := now
2426         for _, ip := range prices {
2427                 spanStart := ip.StartTime
2428                 if spanStart.After(now) {
2429                         // pricing information from the future -- not
2430                         // expected from AWS, but possible in
2431                         // principle, and exercised by tests.
2432                         continue
2433                 }
2434                 last := false
2435                 if spanStart.Before(cr.costStartTime) {
2436                         spanStart = cr.costStartTime
2437                         last = true
2438                 }
2439                 cost += ip.Price * spanEnd.Sub(spanStart).Seconds() / 3600
2440                 if last {
2441                         break
2442                 }
2443                 spanEnd = spanStart
2444         }
2445
2446         return cost
2447 }