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.containerClient, 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         runner.hoststatReporter.LogProcessMemMax(runner.CrunchLog)
 738         err := runner.hoststatLogger.Close()
 739         if err != nil {
 740                 return fmt.Errorf("error closing hoststat logs: %v", err)
 741         }
 742         return nil
 743 }
 744
 745 func (runner *ContainerRunner) startHoststat() error {
 746         w, err := runner.NewLogWriter("hoststat")
 747         if err != nil {
 748                 return err
 749         }
 750         runner.hoststatLogger = NewThrottledLogger(w)
 751         runner.hoststatReporter = &crunchstat.Reporter{
 752                 Logger:     log.New(runner.hoststatLogger, "", 0),
 753                 CgroupRoot: runner.cgroupRoot,
 754                 PollPeriod: runner.statInterval,
 755         }
 756         runner.hoststatReporter.Start()
 757         runner.hoststatReporter.ReportPID("crunch-run", os.Getpid())
 758         return nil
 759 }
 760
 761 func (runner *ContainerRunner) startCrunchstat() error {
 762         w, err := runner.NewLogWriter("crunchstat")
 763         if err != nil {
 764                 return err
 765         }
 766         runner.statLogger = NewThrottledLogger(w)
 767         runner.statReporter = &crunchstat.Reporter{
 768                 CgroupParent: runner.expectCgroupParent,
 769                 CgroupRoot:   runner.cgroupRoot,
 770                 CID:          runner.executor.CgroupID(),
 771                 Logger:       log.New(runner.statLogger, "", 0),
 772                 MemThresholds: map[string][]crunchstat.Threshold{
 773                         "rss": crunchstat.NewThresholdsFromPercentages(runner.Container.RuntimeConstraints.RAM, []int64{90, 95, 99}),
 774                 },
 775                 PollPeriod:      runner.statInterval,
 776                 TempDir:         runner.parentTemp,
 777                 ThresholdLogger: runner.CrunchLog,
 778         }
 779         runner.statReporter.Start()
 780         return nil
 781 }
 782
 783 type infoCommand struct {
 784         label string
 785         cmd   []string
 786 }
 787
 788 // LogHostInfo logs info about the current host, for debugging and
 789 // accounting purposes. Although it's logged as "node-info", this is
 790 // about the environment where crunch-run is actually running, which
 791 // might differ from what's described in the node record (see
 792 // LogNodeRecord).
 793 func (runner *ContainerRunner) LogHostInfo() (err error) {
 794         w, err := runner.NewLogWriter("node-info")
 795         if err != nil {
 796                 return
 797         }
 798
 799         commands := []infoCommand{
 800                 {
 801                         label: "Host Information",
 802                         cmd:   []string{"uname", "-a"},
 803                 },
 804                 {
 805                         label: "CPU Information",
 806                         cmd:   []string{"cat", "/proc/cpuinfo"},
 807                 },
 808                 {
 809                         label: "Memory Information",
 810                         cmd:   []string{"cat", "/proc/meminfo"},
 811                 },
 812                 {
 813                         label: "Disk Space",
 814                         cmd:   []string{"df", "-m", "/", os.TempDir()},
 815                 },
 816                 {
 817                         label: "Disk INodes",
 818                         cmd:   []string{"df", "-i", "/", os.TempDir()},
 819                 },
 820         }
 821
 822         // Run commands with informational output to be logged.
 823         for _, command := range commands {
 824                 fmt.Fprintln(w, command.label)
 825                 cmd := exec.Command(command.cmd[0], command.cmd[1:]...)
 826                 cmd.Stdout = w
 827                 cmd.Stderr = w
 828                 if err := cmd.Run(); err != nil {
 829                         err = fmt.Errorf("While running command %q: %v", command.cmd, err)
 830                         fmt.Fprintln(w, err)
 831                         return err
 832                 }
 833                 fmt.Fprintln(w, "")
 834         }
 835
 836         err = w.Close()
 837         if err != nil {
 838                 return fmt.Errorf("While closing node-info logs: %v", err)
 839         }
 840         return nil
 841 }
 842
 843 // LogContainerRecord gets and saves the raw JSON container record from the API server
 844 func (runner *ContainerRunner) LogContainerRecord() error {
 845         logged, err := runner.logAPIResponse("container", "containers", map[string]interface{}{"filters": [][]string{{"uuid", "=", runner.Container.UUID}}}, nil)
 846         if !logged && err == nil {
 847                 err = fmt.Errorf("error: no container record found for %s", runner.Container.UUID)
 848         }
 849         return err
 850 }
 851
 852 // LogNodeRecord logs the current host's InstanceType config entry (or
 853 // the arvados#node record, if running via crunch-dispatch-slurm).
 854 func (runner *ContainerRunner) LogNodeRecord() error {
 855         if it := os.Getenv("InstanceType"); it != "" {
 856                 // Dispatched via arvados-dispatch-cloud. Save
 857                 // InstanceType config fragment received from
 858                 // dispatcher on stdin.
 859                 w, err := runner.LogCollection.OpenFile("node.json", os.O_CREATE|os.O_WRONLY, 0666)
 860                 if err != nil {
 861                         return err
 862                 }
 863                 defer w.Close()
 864                 _, err = io.WriteString(w, it)
 865                 if err != nil {
 866                         return err
 867                 }
 868                 return w.Close()
 869         }
 870         // Dispatched via crunch-dispatch-slurm. Look up
 871         // apiserver's node record corresponding to
 872         // $SLURMD_NODENAME.
 873         hostname := os.Getenv("SLURMD_NODENAME")
 874         if hostname == "" {
 875                 hostname, _ = os.Hostname()
 876         }
 877         _, err := runner.logAPIResponse("node", "nodes", map[string]interface{}{"filters": [][]string{{"hostname", "=", hostname}}}, func(resp interface{}) {
 878                 // The "info" field has admin-only info when
 879                 // obtained with a privileged token, and
 880                 // should not be logged.
 881                 node, ok := resp.(map[string]interface{})
 882                 if ok {
 883                         delete(node, "info")
 884                 }
 885         })
 886         return err
 887 }
 888
 889 func (runner *ContainerRunner) logAPIResponse(label, path string, params map[string]interface{}, munge func(interface{})) (logged bool, err error) {
 890         writer, err := runner.LogCollection.OpenFile(label+".json", os.O_CREATE|os.O_WRONLY, 0666)
 891         if err != nil {
 892                 return false, err
 893         }
 894         w := &ArvLogWriter{
 895                 ArvClient:     runner.DispatcherArvClient,
 896                 UUID:          runner.Container.UUID,
 897                 loggingStream: label,
 898                 writeCloser:   writer,
 899         }
 900
 901         reader, err := runner.DispatcherArvClient.CallRaw("GET", path, "", "", arvadosclient.Dict(params))
 902         if err != nil {
 903                 return false, fmt.Errorf("error getting %s record: %v", label, err)
 904         }
 905         defer reader.Close()
 906
 907         dec := json.NewDecoder(reader)
 908         dec.UseNumber()
 909         var resp map[string]interface{}
 910         if err = dec.Decode(&resp); err != nil {
 911                 return false, fmt.Errorf("error decoding %s list response: %v", label, err)
 912         }
 913         items, ok := resp["items"].([]interface{})
 914         if !ok {
 915                 return false, fmt.Errorf("error decoding %s list response: no \"items\" key in API list response", label)
 916         } else if len(items) < 1 {
 917                 return false, nil
 918         }
 919         if munge != nil {
 920                 munge(items[0])
 921         }
 922         // Re-encode it using indentation to improve readability
 923         enc := json.NewEncoder(w)
 924         enc.SetIndent("", "    ")
 925         if err = enc.Encode(items[0]); err != nil {
 926                 return false, fmt.Errorf("error logging %s record: %v", label, err)
 927         }
 928         err = w.Close()
 929         if err != nil {
 930                 return false, fmt.Errorf("error closing %s.json in log collection: %v", label, err)
 931         }
 932         return true, nil
 933 }
 934
 935 func (runner *ContainerRunner) getStdoutFile(mntPath string) (*os.File, error) {
 936         stdoutPath := mntPath[len(runner.Container.OutputPath):]
 937         index := strings.LastIndex(stdoutPath, "/")
 938         if index > 0 {
 939                 subdirs := stdoutPath[:index]
 940                 if subdirs != "" {
 941                         st, err := os.Stat(runner.HostOutputDir)
 942                         if err != nil {
 943                                 return nil, fmt.Errorf("While Stat on temp dir: %v", err)
 944                         }
 945                         stdoutPath := filepath.Join(runner.HostOutputDir, subdirs)
 946                         err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777)
 947                         if err != nil {
 948                                 return nil, fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err)
 949                         }
 950                 }
 951         }
 952         stdoutFile, err := os.Create(filepath.Join(runner.HostOutputDir, stdoutPath))
 953         if err != nil {
 954                 return nil, fmt.Errorf("While creating file %q: %v", stdoutPath, err)
 955         }
 956
 957         return stdoutFile, nil
 958 }
 959
 960 // CreateContainer creates the docker container.
 961 func (runner *ContainerRunner) CreateContainer(imageID string, bindmounts map[string]bindmount) error {
 962         var stdin io.ReadCloser = ioutil.NopCloser(bytes.NewReader(nil))
 963         if mnt, ok := runner.Container.Mounts["stdin"]; ok {
 964                 switch mnt.Kind {
 965                 case "collection":
 966                         var collID string
 967                         if mnt.UUID != "" {
 968                                 collID = mnt.UUID
 969                         } else {
 970                                 collID = mnt.PortableDataHash
 971                         }
 972                         path := runner.ArvMountPoint + "/by_id/" + collID + "/" + mnt.Path
 973                         f, err := os.Open(path)
 974                         if err != nil {
 975                                 return err
 976                         }
 977                         stdin = f
 978                 case "json":
 979                         j, err := json.Marshal(mnt.Content)
 980                         if err != nil {
 981                                 return fmt.Errorf("error encoding stdin json data: %v", err)
 982                         }
 983                         stdin = ioutil.NopCloser(bytes.NewReader(j))
 984                 default:
 985                         return fmt.Errorf("stdin mount has unsupported kind %q", mnt.Kind)
 986                 }
 987         }
 988
 989         var stdout, stderr io.WriteCloser
 990         if mnt, ok := runner.Container.Mounts["stdout"]; ok {
 991                 f, err := runner.getStdoutFile(mnt.Path)
 992                 if err != nil {
 993                         return err
 994                 }
 995                 stdout = f
 996         } else if w, err := runner.NewLogWriter("stdout"); err != nil {
 997                 return err
 998         } else {
 999                 stdout = NewThrottledLogger(w)
1000         }
1001
1002         if mnt, ok := runner.Container.Mounts["stderr"]; ok {
1003                 f, err := runner.getStdoutFile(mnt.Path)
1004                 if err != nil {
1005                         return err
1006                 }
1007                 stderr = f
1008         } else if w, err := runner.NewLogWriter("stderr"); err != nil {
1009                 return err
1010         } else {
1011                 stderr = NewThrottledLogger(w)
1012         }
1013
1014         env := runner.Container.Environment
1015         enableNetwork := runner.enableNetwork == "always"
1016         if runner.Container.RuntimeConstraints.API {
1017                 enableNetwork = true
1018                 tok, err := runner.ContainerToken()
1019                 if err != nil {
1020                         return err
1021                 }
1022                 env = map[string]string{}
1023                 for k, v := range runner.Container.Environment {
1024                         env[k] = v
1025                 }
1026                 env["ARVADOS_API_TOKEN"] = tok
1027                 env["ARVADOS_API_HOST"] = os.Getenv("ARVADOS_API_HOST")
1028                 env["ARVADOS_API_HOST_INSECURE"] = os.Getenv("ARVADOS_API_HOST_INSECURE")
1029                 env["ARVADOS_KEEP_SERVICES"] = os.Getenv("ARVADOS_KEEP_SERVICES")
1030         }
1031         workdir := runner.Container.Cwd
1032         if workdir == "." {
1033                 // both "" and "." mean default
1034                 workdir = ""
1035         }
1036         ram := runner.Container.RuntimeConstraints.RAM
1037         if !runner.enableMemoryLimit {
1038                 ram = 0
1039         }
1040         runner.executorStdin = stdin
1041         runner.executorStdout = stdout
1042         runner.executorStderr = stderr
1043
1044         if runner.Container.RuntimeConstraints.CUDA.DeviceCount > 0 {
1045                 nvidiaModprobe(runner.CrunchLog)
1046         }
1047
1048         return runner.executor.Create(containerSpec{
1049                 Image:           imageID,
1050                 VCPUs:           runner.Container.RuntimeConstraints.VCPUs,
1051                 RAM:             ram,
1052                 WorkingDir:      workdir,
1053                 Env:             env,
1054                 BindMounts:      bindmounts,
1055                 Command:         runner.Container.Command,
1056                 EnableNetwork:   enableNetwork,
1057                 CUDADeviceCount: runner.Container.RuntimeConstraints.CUDA.DeviceCount,
1058                 NetworkMode:     runner.networkMode,
1059                 CgroupParent:    runner.setCgroupParent,
1060                 Stdin:           stdin,
1061                 Stdout:          stdout,
1062                 Stderr:          stderr,
1063         })
1064 }
1065
1066 // StartContainer starts the docker container created by CreateContainer.
1067 func (runner *ContainerRunner) StartContainer() error {
1068         runner.CrunchLog.Printf("Starting container")
1069         runner.cStateLock.Lock()
1070         defer runner.cStateLock.Unlock()
1071         if runner.cCancelled {
1072                 return ErrCancelled
1073         }
1074         err := runner.executor.Start()
1075         if err != nil {
1076                 var advice string
1077                 if m, e := regexp.MatchString("(?ms).*(exec|System error).*(no such file or directory|file not found).*", err.Error()); m && e == nil {
1078                         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])
1079                 }
1080                 return fmt.Errorf("could not start container: %v%s", err, advice)
1081         }
1082         return nil
1083 }
1084
1085 // WaitFinish waits for the container to terminate, capture the exit code, and
1086 // close the stdout/stderr logging.
1087 func (runner *ContainerRunner) WaitFinish() error {
1088         runner.CrunchLog.Print("Waiting for container to finish")
1089         var timeout <-chan time.Time
1090         if s := runner.Container.SchedulingParameters.MaxRunTime; s > 0 {
1091                 timeout = time.After(time.Duration(s) * time.Second)
1092         }
1093         ctx, cancel := context.WithCancel(context.Background())
1094         defer cancel()
1095         go func() {
1096                 select {
1097                 case <-timeout:
1098                         runner.CrunchLog.Printf("maximum run time exceeded. Stopping container.")
1099                         runner.stop(nil)
1100                 case <-runner.ArvMountExit:
1101                         runner.CrunchLog.Printf("arv-mount exited while container is still running. Stopping container.")
1102                         runner.stop(nil)
1103                 case <-ctx.Done():
1104                 }
1105         }()
1106         exitcode, err := runner.executor.Wait(ctx)
1107         if err != nil {
1108                 runner.checkBrokenNode(err)
1109                 return err
1110         }
1111         runner.ExitCode = &exitcode
1112
1113         extra := ""
1114         if exitcode&0x80 != 0 {
1115                 // Convert raw exit status (0x80 + signal number) to a
1116                 // string to log after the code, like " (signal 101)"
1117                 // or " (signal 9, killed)"
1118                 sig := syscall.WaitStatus(exitcode).Signal()
1119                 if name := unix.SignalName(sig); name != "" {
1120                         extra = fmt.Sprintf(" (signal %d, %s)", sig, name)
1121                 } else {
1122                         extra = fmt.Sprintf(" (signal %d)", sig)
1123                 }
1124         }
1125         runner.CrunchLog.Printf("Container exited with status code %d%s", exitcode, extra)
1126         err = runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1127                 "container": arvadosclient.Dict{"exit_code": exitcode},
1128         }, nil)
1129         if err != nil {
1130                 runner.CrunchLog.Printf("ignoring error updating exit_code: %s", err)
1131         }
1132
1133         var returnErr error
1134         if err = runner.executorStdin.Close(); err != nil {
1135                 err = fmt.Errorf("error closing container stdin: %s", err)
1136                 runner.CrunchLog.Printf("%s", err)
1137                 returnErr = err
1138         }
1139         if err = runner.executorStdout.Close(); err != nil {
1140                 err = fmt.Errorf("error closing container stdout: %s", err)
1141                 runner.CrunchLog.Printf("%s", err)
1142                 if returnErr == nil {
1143                         returnErr = err
1144                 }
1145         }
1146         if err = runner.executorStderr.Close(); err != nil {
1147                 err = fmt.Errorf("error closing container stderr: %s", err)
1148                 runner.CrunchLog.Printf("%s", err)
1149                 if returnErr == nil {
1150                         returnErr = err
1151                 }
1152         }
1153
1154         if runner.statReporter != nil {
1155                 runner.statReporter.Stop()
1156                 runner.statReporter.LogMaxima(runner.CrunchLog, map[string]int64{
1157                         "rss": runner.Container.RuntimeConstraints.RAM,
1158                 })
1159                 err = runner.statLogger.Close()
1160                 if err != nil {
1161                         runner.CrunchLog.Printf("error closing crunchstat logs: %v", err)
1162                 }
1163         }
1164         return returnErr
1165 }
1166
1167 func (runner *ContainerRunner) updateLogs() {
1168         ticker := time.NewTicker(crunchLogUpdatePeriod / 360)
1169         defer ticker.Stop()
1170
1171         sigusr1 := make(chan os.Signal, 1)
1172         signal.Notify(sigusr1, syscall.SIGUSR1)
1173         defer signal.Stop(sigusr1)
1174
1175         saveAtTime := time.Now().Add(crunchLogUpdatePeriod)
1176         saveAtSize := crunchLogUpdateSize
1177         var savedSize int64
1178         for {
1179                 select {
1180                 case <-ticker.C:
1181                 case <-sigusr1:
1182                         saveAtTime = time.Now()
1183                 }
1184                 runner.logMtx.Lock()
1185                 done := runner.LogsPDH != nil
1186                 runner.logMtx.Unlock()
1187                 if done {
1188                         return
1189                 }
1190                 size := runner.LogCollection.Size()
1191                 if size == savedSize || (time.Now().Before(saveAtTime) && size < saveAtSize) {
1192                         continue
1193                 }
1194                 saveAtTime = time.Now().Add(crunchLogUpdatePeriod)
1195                 saveAtSize = runner.LogCollection.Size() + crunchLogUpdateSize
1196                 saved, err := runner.saveLogCollection(false)
1197                 if err != nil {
1198                         runner.CrunchLog.Printf("error updating log collection: %s", err)
1199                         continue
1200                 }
1201
1202                 err = runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1203                         "container": arvadosclient.Dict{"log": saved.PortableDataHash},
1204                 }, nil)
1205                 if err != nil {
1206                         runner.CrunchLog.Printf("error updating container log to %s: %s", saved.PortableDataHash, err)
1207                         continue
1208                 }
1209
1210                 savedSize = size
1211         }
1212 }
1213
1214 var spotInterruptionCheckInterval = 5 * time.Second
1215 var ec2MetadataBaseURL = "http://169.254.169.254"
1216
1217 const ec2TokenTTL = time.Second * 21600
1218
1219 func (runner *ContainerRunner) checkSpotInterruptionNotices() {
1220         type ec2metadata struct {
1221                 Action string    `json:"action"`
1222                 Time   time.Time `json:"time"`
1223         }
1224         runner.CrunchLog.Printf("Checking for spot interruptions every %v using instance metadata at %s", spotInterruptionCheckInterval, ec2MetadataBaseURL)
1225         var metadata ec2metadata
1226         var token string
1227         var tokenExp time.Time
1228         check := func() error {
1229                 ctx, cancel := context.WithDeadline(context.Background(), time.Now().Add(time.Minute))
1230                 defer cancel()
1231                 if token == "" || tokenExp.Sub(time.Now()) < time.Minute {
1232                         req, err := http.NewRequestWithContext(ctx, http.MethodPut, ec2MetadataBaseURL+"/latest/api/token", nil)
1233                         if err != nil {
1234                                 return err
1235                         }
1236                         req.Header.Set("X-aws-ec2-metadata-token-ttl-seconds", fmt.Sprintf("%d", int(ec2TokenTTL/time.Second)))
1237                         resp, err := http.DefaultClient.Do(req)
1238                         if err != nil {
1239                                 return err
1240                         }
1241                         defer resp.Body.Close()
1242                         if resp.StatusCode != http.StatusOK {
1243                                 return fmt.Errorf("%s", resp.Status)
1244                         }
1245                         newtoken, err := ioutil.ReadAll(resp.Body)
1246                         if err != nil {
1247                                 return err
1248                         }
1249                         token = strings.TrimSpace(string(newtoken))
1250                         tokenExp = time.Now().Add(ec2TokenTTL)
1251                 }
1252                 req, err := http.NewRequestWithContext(ctx, http.MethodGet, ec2MetadataBaseURL+"/latest/meta-data/spot/instance-action", nil)
1253                 if err != nil {
1254                         return err
1255                 }
1256                 req.Header.Set("X-aws-ec2-metadata-token", token)
1257                 resp, err := http.DefaultClient.Do(req)
1258                 if err != nil {
1259                         return err
1260                 }
1261                 defer resp.Body.Close()
1262                 metadata = ec2metadata{}
1263                 switch resp.StatusCode {
1264                 case http.StatusOK:
1265                         break
1266                 case http.StatusNotFound:
1267                         // "If Amazon EC2 is not preparing to stop or
1268                         // terminate the instance, or if you
1269                         // terminated the instance yourself,
1270                         // instance-action is not present in the
1271                         // instance metadata and you receive an HTTP
1272                         // 404 error when you try to retrieve it."
1273                         return nil
1274                 case http.StatusUnauthorized:
1275                         token = ""
1276                         return fmt.Errorf("%s", resp.Status)
1277                 default:
1278                         return fmt.Errorf("%s", resp.Status)
1279                 }
1280                 err = json.NewDecoder(resp.Body).Decode(&metadata)
1281                 if err != nil {
1282                         return err
1283                 }
1284                 return nil
1285         }
1286         failures := 0
1287         var lastmetadata ec2metadata
1288         for range time.NewTicker(spotInterruptionCheckInterval).C {
1289                 err := check()
1290                 if err != nil {
1291                         runner.CrunchLog.Printf("Error checking spot interruptions: %s", err)
1292                         failures++
1293                         if failures > 5 {
1294                                 runner.CrunchLog.Printf("Giving up on checking spot interruptions after too many consecutive failures")
1295                                 return
1296                         }
1297                         continue
1298                 }
1299                 failures = 0
1300                 if metadata != lastmetadata {
1301                         lastmetadata = metadata
1302                         text := fmt.Sprintf("Cloud provider scheduled instance %s at %s", metadata.Action, metadata.Time.UTC().Format(time.RFC3339))
1303                         runner.CrunchLog.Printf("%s", text)
1304                         runner.updateRuntimeStatus(arvadosclient.Dict{
1305                                 "warning":          "preemption notice",
1306                                 "warningDetail":    text,
1307                                 "preemptionNotice": text,
1308                         })
1309                         if proc, err := os.FindProcess(os.Getpid()); err == nil {
1310                                 // trigger updateLogs
1311                                 proc.Signal(syscall.SIGUSR1)
1312                         }
1313                 }
1314         }
1315 }
1316
1317 func (runner *ContainerRunner) updateRuntimeStatus(status arvadosclient.Dict) {
1318         err := runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1319                 "container": arvadosclient.Dict{
1320                         "runtime_status": status,
1321                 },
1322         }, nil)
1323         if err != nil {
1324                 runner.CrunchLog.Printf("error updating container runtime_status: %s", err)
1325         }
1326 }
1327
1328 // CaptureOutput saves data from the container's output directory if
1329 // needed, and updates the container output accordingly.
1330 func (runner *ContainerRunner) CaptureOutput(bindmounts map[string]bindmount) error {
1331         if runner.Container.RuntimeConstraints.API {
1332                 // Output may have been set directly by the container, so
1333                 // refresh the container record to check.
1334                 err := runner.DispatcherArvClient.Get("containers", runner.Container.UUID,
1335                         nil, &runner.Container)
1336                 if err != nil {
1337                         return err
1338                 }
1339                 if runner.Container.Output != "" {
1340                         // Container output is already set.
1341                         runner.OutputPDH = &runner.Container.Output
1342                         return nil
1343                 }
1344         }
1345
1346         txt, err := (&copier{
1347                 client:        runner.containerClient,
1348                 keepClient:    runner.ContainerKeepClient,
1349                 hostOutputDir: runner.HostOutputDir,
1350                 ctrOutputDir:  runner.Container.OutputPath,
1351                 bindmounts:    bindmounts,
1352                 mounts:        runner.Container.Mounts,
1353                 secretMounts:  runner.SecretMounts,
1354                 logger:        runner.CrunchLog,
1355         }).Copy()
1356         if err != nil {
1357                 return err
1358         }
1359         if n := len(regexp.MustCompile(` [0-9a-f]+\+\S*\+R`).FindAllStringIndex(txt, -1)); n > 0 {
1360                 runner.CrunchLog.Printf("Copying %d data blocks from remote input collections...", n)
1361                 fs, err := (&arvados.Collection{ManifestText: txt}).FileSystem(runner.containerClient, runner.ContainerKeepClient)
1362                 if err != nil {
1363                         return err
1364                 }
1365                 txt, err = fs.MarshalManifest(".")
1366                 if err != nil {
1367                         return err
1368                 }
1369         }
1370         var resp arvados.Collection
1371         err = runner.ContainerArvClient.Create("collections", arvadosclient.Dict{
1372                 "ensure_unique_name": true,
1373                 "collection": arvadosclient.Dict{
1374                         "is_trashed":    true,
1375                         "name":          "output for " + runner.Container.UUID,
1376                         "manifest_text": txt,
1377                 },
1378         }, &resp)
1379         if err != nil {
1380                 return fmt.Errorf("error creating output collection: %v", err)
1381         }
1382         runner.OutputPDH = &resp.PortableDataHash
1383         return nil
1384 }
1385
1386 func (runner *ContainerRunner) CleanupDirs() {
1387         if runner.ArvMount != nil {
1388                 var delay int64 = 8
1389                 umount := exec.Command("arv-mount", fmt.Sprintf("--unmount-timeout=%d", delay), "--unmount", runner.ArvMountPoint)
1390                 umount.Stdout = runner.CrunchLog
1391                 umount.Stderr = runner.CrunchLog
1392                 runner.CrunchLog.Printf("Running %v", umount.Args)
1393                 umnterr := umount.Start()
1394
1395                 if umnterr != nil {
1396                         runner.CrunchLog.Printf("Error unmounting: %v", umnterr)
1397                         runner.ArvMount.Process.Kill()
1398                 } else {
1399                         // If arv-mount --unmount gets stuck for any reason, we
1400                         // don't want to wait for it forever.  Do Wait() in a goroutine
1401                         // so it doesn't block crunch-run.
1402                         umountExit := make(chan error)
1403                         go func() {
1404                                 mnterr := umount.Wait()
1405                                 if mnterr != nil {
1406                                         runner.CrunchLog.Printf("Error unmounting: %v", mnterr)
1407                                 }
1408                                 umountExit <- mnterr
1409                         }()
1410
1411                         for again := true; again; {
1412                                 again = false
1413                                 select {
1414                                 case <-umountExit:
1415                                         umount = nil
1416                                         again = true
1417                                 case <-runner.ArvMountExit:
1418                                         break
1419                                 case <-time.After(time.Duration((delay + 1) * int64(time.Second))):
1420                                         runner.CrunchLog.Printf("Timed out waiting for unmount")
1421                                         if umount != nil {
1422                                                 umount.Process.Kill()
1423                                         }
1424                                         runner.ArvMount.Process.Kill()
1425                                 }
1426                         }
1427                 }
1428                 runner.ArvMount = nil
1429         }
1430
1431         if runner.ArvMountPoint != "" {
1432                 if rmerr := os.Remove(runner.ArvMountPoint); rmerr != nil {
1433                         runner.CrunchLog.Printf("While cleaning up arv-mount directory %s: %v", runner.ArvMountPoint, rmerr)
1434                 }
1435                 runner.ArvMountPoint = ""
1436         }
1437
1438         if rmerr := os.RemoveAll(runner.parentTemp); rmerr != nil {
1439                 runner.CrunchLog.Printf("While cleaning up temporary directory %s: %v", runner.parentTemp, rmerr)
1440         }
1441 }
1442
1443 // CommitLogs posts the collection containing the final container logs.
1444 func (runner *ContainerRunner) CommitLogs() error {
1445         func() {
1446                 // Hold cStateLock to prevent races on CrunchLog (e.g., stop()).
1447                 runner.cStateLock.Lock()
1448                 defer runner.cStateLock.Unlock()
1449
1450                 runner.CrunchLog.Print(runner.finalState)
1451
1452                 if runner.arvMountLog != nil {
1453                         runner.arvMountLog.Close()
1454                 }
1455                 runner.CrunchLog.Close()
1456
1457                 // Closing CrunchLog above allows them to be committed to Keep at this
1458                 // point, but re-open crunch log with ArvClient in case there are any
1459                 // other further errors (such as failing to write the log to Keep!)
1460                 // while shutting down
1461                 runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{
1462                         ArvClient:     runner.DispatcherArvClient,
1463                         UUID:          runner.Container.UUID,
1464                         loggingStream: "crunch-run",
1465                         writeCloser:   nil,
1466                 })
1467                 runner.CrunchLog.Immediate = log.New(os.Stderr, runner.Container.UUID+" ", 0)
1468         }()
1469
1470         if runner.keepstoreLogger != nil {
1471                 // Flush any buffered logs from our local keepstore
1472                 // process.  Discard anything logged after this point
1473                 // -- it won't end up in the log collection, so
1474                 // there's no point writing it to the collectionfs.
1475                 runner.keepstoreLogbuf.SetWriter(io.Discard)
1476                 runner.keepstoreLogger.Close()
1477                 runner.keepstoreLogger = nil
1478         }
1479
1480         if runner.LogsPDH != nil {
1481                 // If we have already assigned something to LogsPDH,
1482                 // we must be closing the re-opened log, which won't
1483                 // end up getting attached to the container record and
1484                 // therefore doesn't need to be saved as a collection
1485                 // -- it exists only to send logs to other channels.
1486                 return nil
1487         }
1488
1489         saved, err := runner.saveLogCollection(true)
1490         if err != nil {
1491                 return fmt.Errorf("error saving log collection: %s", err)
1492         }
1493         runner.logMtx.Lock()
1494         defer runner.logMtx.Unlock()
1495         runner.LogsPDH = &saved.PortableDataHash
1496         return nil
1497 }
1498
1499 func (runner *ContainerRunner) saveLogCollection(final bool) (response arvados.Collection, err error) {
1500         runner.logMtx.Lock()
1501         defer runner.logMtx.Unlock()
1502         if runner.LogsPDH != nil {
1503                 // Already finalized.
1504                 return
1505         }
1506         updates := arvadosclient.Dict{
1507                 "name": "logs for " + runner.Container.UUID,
1508         }
1509         mt, err1 := runner.LogCollection.MarshalManifest(".")
1510         if err1 == nil {
1511                 // Only send updated manifest text if there was no
1512                 // error.
1513                 updates["manifest_text"] = mt
1514         }
1515
1516         // Even if flushing the manifest had an error, we still want
1517         // to update the log record, if possible, to push the trash_at
1518         // and delete_at times into the future.  Details on bug
1519         // #17293.
1520         if final {
1521                 updates["is_trashed"] = true
1522         } else {
1523                 // We set trash_at so this collection gets
1524                 // automatically cleaned up eventually.  It used to be
1525                 // 12 hours but we had a situation where the API
1526                 // server was down over a weekend but the containers
1527                 // kept running such that the log collection got
1528                 // trashed, so now we make it 2 weeks.  refs #20378
1529                 exp := time.Now().Add(time.Duration(24*14) * time.Hour)
1530                 updates["trash_at"] = exp
1531                 updates["delete_at"] = exp
1532         }
1533         reqBody := arvadosclient.Dict{"collection": updates}
1534         var err2 error
1535         if runner.logUUID == "" {
1536                 reqBody["ensure_unique_name"] = true
1537                 err2 = runner.DispatcherArvClient.Create("collections", reqBody, &response)
1538         } else {
1539                 err2 = runner.DispatcherArvClient.Update("collections", runner.logUUID, reqBody, &response)
1540         }
1541         if err2 == nil {
1542                 runner.logUUID = response.UUID
1543         }
1544
1545         if err1 != nil || err2 != nil {
1546                 err = fmt.Errorf("error recording logs: %q, %q", err1, err2)
1547         }
1548         return
1549 }
1550
1551 // UpdateContainerRunning updates the container state to "Running"
1552 func (runner *ContainerRunner) UpdateContainerRunning(logId string) error {
1553         runner.cStateLock.Lock()
1554         defer runner.cStateLock.Unlock()
1555         if runner.cCancelled {
1556                 return ErrCancelled
1557         }
1558         updates := arvadosclient.Dict{
1559                 "gateway_address": runner.gateway.Address,
1560                 "state":           "Running",
1561         }
1562         if logId != "" {
1563                 updates["log"] = logId
1564         }
1565         return runner.DispatcherArvClient.Update(
1566                 "containers",
1567                 runner.Container.UUID,
1568                 arvadosclient.Dict{"container": updates},
1569                 nil,
1570         )
1571 }
1572
1573 // ContainerToken returns the api_token the container (and any
1574 // arv-mount processes) are allowed to use.
1575 func (runner *ContainerRunner) ContainerToken() (string, error) {
1576         if runner.token != "" {
1577                 return runner.token, nil
1578         }
1579
1580         var auth arvados.APIClientAuthorization
1581         err := runner.DispatcherArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth)
1582         if err != nil {
1583                 return "", err
1584         }
1585         runner.token = fmt.Sprintf("v2/%s/%s/%s", auth.UUID, auth.APIToken, runner.Container.UUID)
1586         return runner.token, nil
1587 }
1588
1589 // UpdateContainerFinal updates the container record state on API
1590 // server to "Complete" or "Cancelled"
1591 func (runner *ContainerRunner) UpdateContainerFinal() error {
1592         update := arvadosclient.Dict{}
1593         update["state"] = runner.finalState
1594         if runner.LogsPDH != nil {
1595                 update["log"] = *runner.LogsPDH
1596         }
1597         if runner.ExitCode != nil {
1598                 update["exit_code"] = *runner.ExitCode
1599         } else {
1600                 update["exit_code"] = nil
1601         }
1602         if runner.finalState == "Complete" && runner.OutputPDH != nil {
1603                 update["output"] = *runner.OutputPDH
1604         }
1605         update["cost"] = runner.calculateCost(time.Now())
1606         return runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{"container": update}, nil)
1607 }
1608
1609 // IsCancelled returns the value of Cancelled, with goroutine safety.
1610 func (runner *ContainerRunner) IsCancelled() bool {
1611         runner.cStateLock.Lock()
1612         defer runner.cStateLock.Unlock()
1613         return runner.cCancelled
1614 }
1615
1616 // NewArvLogWriter creates an ArvLogWriter
1617 func (runner *ContainerRunner) NewArvLogWriter(name string) (io.WriteCloser, error) {
1618         writer, err := runner.LogCollection.OpenFile(name+".txt", os.O_CREATE|os.O_WRONLY, 0666)
1619         if err != nil {
1620                 return nil, err
1621         }
1622         return &ArvLogWriter{
1623                 ArvClient:     runner.DispatcherArvClient,
1624                 UUID:          runner.Container.UUID,
1625                 loggingStream: name,
1626                 writeCloser:   writer,
1627         }, nil
1628 }
1629
1630 // Run the full container lifecycle.
1631 func (runner *ContainerRunner) Run() (err error) {
1632         runner.CrunchLog.Printf("crunch-run %s started", cmd.Version.String())
1633         runner.CrunchLog.Printf("%s", currentUserAndGroups())
1634         v, _ := exec.Command("arv-mount", "--version").CombinedOutput()
1635         runner.CrunchLog.Printf("Using FUSE mount: %s", v)
1636         runner.CrunchLog.Printf("Using container runtime: %s", runner.executor.Runtime())
1637         runner.CrunchLog.Printf("Executing container: %s", runner.Container.UUID)
1638         runner.costStartTime = time.Now()
1639
1640         hostname, hosterr := os.Hostname()
1641         if hosterr != nil {
1642                 runner.CrunchLog.Printf("Error getting hostname '%v'", hosterr)
1643         } else {
1644                 runner.CrunchLog.Printf("Executing on host '%s'", hostname)
1645         }
1646
1647         sigusr2 := make(chan os.Signal, 1)
1648         signal.Notify(sigusr2, syscall.SIGUSR2)
1649         defer signal.Stop(sigusr2)
1650         runner.loadPrices()
1651         go runner.handleSIGUSR2(sigusr2)
1652
1653         runner.finalState = "Queued"
1654
1655         defer func() {
1656                 runner.CleanupDirs()
1657
1658                 runner.CrunchLog.Printf("crunch-run finished")
1659                 runner.CrunchLog.Close()
1660         }()
1661
1662         err = runner.fetchContainerRecord()
1663         if err != nil {
1664                 return
1665         }
1666         if runner.Container.State != "Locked" {
1667                 return fmt.Errorf("dispatch error detected: container %q has state %q", runner.Container.UUID, runner.Container.State)
1668         }
1669
1670         var bindmounts map[string]bindmount
1671         defer func() {
1672                 // checkErr prints e (unless it's nil) and sets err to
1673                 // e (unless err is already non-nil). Thus, if err
1674                 // hasn't already been assigned when Run() returns,
1675                 // this cleanup func will cause Run() to return the
1676                 // first non-nil error that is passed to checkErr().
1677                 checkErr := func(errorIn string, e error) {
1678                         if e == nil {
1679                                 return
1680                         }
1681                         runner.CrunchLog.Printf("error in %s: %v", errorIn, e)
1682                         if err == nil {
1683                                 err = e
1684                         }
1685                         if runner.finalState == "Complete" {
1686                                 // There was an error in the finalization.
1687                                 runner.finalState = "Cancelled"
1688                         }
1689                 }
1690
1691                 // Log the error encountered in Run(), if any
1692                 checkErr("Run", err)
1693
1694                 if runner.finalState == "Queued" {
1695                         runner.UpdateContainerFinal()
1696                         return
1697                 }
1698
1699                 if runner.IsCancelled() {
1700                         runner.finalState = "Cancelled"
1701                         // but don't return yet -- we still want to
1702                         // capture partial output and write logs
1703                 }
1704
1705                 if bindmounts != nil {
1706                         checkErr("CaptureOutput", runner.CaptureOutput(bindmounts))
1707                 }
1708                 checkErr("stopHoststat", runner.stopHoststat())
1709                 checkErr("CommitLogs", runner.CommitLogs())
1710                 runner.CleanupDirs()
1711                 checkErr("UpdateContainerFinal", runner.UpdateContainerFinal())
1712         }()
1713
1714         runner.setupSignals()
1715         err = runner.startHoststat()
1716         if err != nil {
1717                 return
1718         }
1719         if runner.keepstore != nil {
1720                 runner.hoststatReporter.ReportPID("keepstore", runner.keepstore.Process.Pid)
1721         }
1722
1723         // set up FUSE mount and binds
1724         bindmounts, err = runner.SetupMounts()
1725         if err != nil {
1726                 runner.finalState = "Cancelled"
1727                 err = fmt.Errorf("While setting up mounts: %v", err)
1728                 return
1729         }
1730
1731         // check for and/or load image
1732         imageID, err := runner.LoadImage()
1733         if err != nil {
1734                 if !runner.checkBrokenNode(err) {
1735                         // Failed to load image but not due to a "broken node"
1736                         // condition, probably user error.
1737                         runner.finalState = "Cancelled"
1738                 }
1739                 err = fmt.Errorf("While loading container image: %v", err)
1740                 return
1741         }
1742
1743         err = runner.CreateContainer(imageID, bindmounts)
1744         if err != nil {
1745                 return
1746         }
1747         err = runner.LogHostInfo()
1748         if err != nil {
1749                 return
1750         }
1751         err = runner.LogNodeRecord()
1752         if err != nil {
1753                 return
1754         }
1755         err = runner.LogContainerRecord()
1756         if err != nil {
1757                 return
1758         }
1759
1760         if runner.IsCancelled() {
1761                 return
1762         }
1763
1764         logCollection, err := runner.saveLogCollection(false)
1765         var logId string
1766         if err == nil {
1767                 logId = logCollection.PortableDataHash
1768         } else {
1769                 runner.CrunchLog.Printf("Error committing initial log collection: %v", err)
1770         }
1771         err = runner.UpdateContainerRunning(logId)
1772         if err != nil {
1773                 return
1774         }
1775         runner.finalState = "Cancelled"
1776
1777         err = runner.startCrunchstat()
1778         if err != nil {
1779                 return
1780         }
1781
1782         err = runner.StartContainer()
1783         if err != nil {
1784                 runner.checkBrokenNode(err)
1785                 return
1786         }
1787
1788         err = runner.WaitFinish()
1789         if err == nil && !runner.IsCancelled() {
1790                 runner.finalState = "Complete"
1791         }
1792         return
1793 }
1794
1795 // Fetch the current container record (uuid = runner.Container.UUID)
1796 // into runner.Container.
1797 func (runner *ContainerRunner) fetchContainerRecord() error {
1798         reader, err := runner.DispatcherArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
1799         if err != nil {
1800                 return fmt.Errorf("error fetching container record: %v", err)
1801         }
1802         defer reader.Close()
1803
1804         dec := json.NewDecoder(reader)
1805         dec.UseNumber()
1806         err = dec.Decode(&runner.Container)
1807         if err != nil {
1808                 return fmt.Errorf("error decoding container record: %v", err)
1809         }
1810
1811         var sm struct {
1812                 SecretMounts map[string]arvados.Mount `json:"secret_mounts"`
1813         }
1814
1815         containerToken, err := runner.ContainerToken()
1816         if err != nil {
1817                 return fmt.Errorf("error getting container token: %v", err)
1818         }
1819
1820         runner.ContainerArvClient, runner.ContainerKeepClient,
1821                 runner.containerClient, err = runner.MkArvClient(containerToken)
1822         if err != nil {
1823                 return fmt.Errorf("error creating container API client: %v", err)
1824         }
1825
1826         runner.ContainerKeepClient.SetStorageClasses(runner.Container.OutputStorageClasses)
1827         runner.DispatcherKeepClient.SetStorageClasses(runner.Container.OutputStorageClasses)
1828
1829         err = runner.ContainerArvClient.Call("GET", "containers", runner.Container.UUID, "secret_mounts", nil, &sm)
1830         if err != nil {
1831                 if apierr, ok := err.(arvadosclient.APIServerError); !ok || apierr.HttpStatusCode != 404 {
1832                         return fmt.Errorf("error fetching secret_mounts: %v", err)
1833                 }
1834                 // ok && apierr.HttpStatusCode == 404, which means
1835                 // secret_mounts isn't supported by this API server.
1836         }
1837         runner.SecretMounts = sm.SecretMounts
1838
1839         return nil
1840 }
1841
1842 // NewContainerRunner creates a new container runner.
1843 func NewContainerRunner(dispatcherClient *arvados.Client,
1844         dispatcherArvClient IArvadosClient,
1845         dispatcherKeepClient IKeepClient,
1846         containerUUID string) (*ContainerRunner, error) {
1847
1848         cr := &ContainerRunner{
1849                 dispatcherClient:     dispatcherClient,
1850                 DispatcherArvClient:  dispatcherArvClient,
1851                 DispatcherKeepClient: dispatcherKeepClient,
1852         }
1853         cr.NewLogWriter = cr.NewArvLogWriter
1854         cr.RunArvMount = cr.ArvMountCmd
1855         cr.MkTempDir = ioutil.TempDir
1856         cr.MkArvClient = func(token string) (IArvadosClient, IKeepClient, *arvados.Client, error) {
1857                 cl, err := arvadosclient.MakeArvadosClient()
1858                 if err != nil {
1859                         return nil, nil, nil, err
1860                 }
1861                 cl.ApiToken = token
1862                 kc, err := keepclient.MakeKeepClient(cl)
1863                 if err != nil {
1864                         return nil, nil, nil, err
1865                 }
1866                 c2 := arvados.NewClientFromEnv()
1867                 c2.AuthToken = token
1868                 return cl, kc, c2, nil
1869         }
1870         var err error
1871         cr.LogCollection, err = (&arvados.Collection{}).FileSystem(cr.dispatcherClient, cr.DispatcherKeepClient)
1872         if err != nil {
1873                 return nil, err
1874         }
1875         cr.Container.UUID = containerUUID
1876         w, err := cr.NewLogWriter("crunch-run")
1877         if err != nil {
1878                 return nil, err
1879         }
1880         cr.CrunchLog = NewThrottledLogger(w)
1881         cr.CrunchLog.Immediate = log.New(os.Stderr, containerUUID+" ", 0)
1882
1883         loadLogThrottleParams(dispatcherArvClient)
1884         go cr.updateLogs()
1885
1886         return cr, nil
1887 }
1888
1889 func (command) RunCommand(prog string, args []string, stdin io.Reader, stdout, stderr io.Writer) int {
1890         log := log.New(stderr, "", 0)
1891         flags := flag.NewFlagSet(prog, flag.ContinueOnError)
1892         statInterval := flags.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting")
1893         cgroupRoot := flags.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree")
1894         cgroupParent := flags.String("cgroup-parent", "docker", "name of container's parent cgroup (ignored if -cgroup-parent-subsystem is used)")
1895         cgroupParentSubsystem := flags.String("cgroup-parent-subsystem", "", "use current cgroup for given subsystem as parent cgroup for container")
1896         caCertsPath := flags.String("ca-certs", "", "Path to TLS root certificates")
1897         detach := flags.Bool("detach", false, "Detach from parent process and run in the background")
1898         stdinConfig := flags.Bool("stdin-config", false, "Load config and environment variables from JSON message on stdin")
1899         configFile := flags.String("config", arvados.DefaultConfigFile, "filename of cluster config file to try loading if -stdin-config=false (default is $ARVADOS_CONFIG)")
1900         sleep := flags.Duration("sleep", 0, "Delay before starting (testing use only)")
1901         kill := flags.Int("kill", -1, "Send signal to an existing crunch-run process for given UUID")
1902         list := flags.Bool("list", false, "List UUIDs of existing crunch-run processes (and notify them to use price data passed on stdin)")
1903         enableMemoryLimit := flags.Bool("enable-memory-limit", true, "tell container runtime to limit container's memory usage")
1904         enableNetwork := flags.String("container-enable-networking", "default", "enable networking \"always\" (for all containers) or \"default\" (for containers that request it)")
1905         networkMode := flags.String("container-network-mode", "default", `Docker network mode for container (use any argument valid for docker --net)`)
1906         memprofile := flags.String("memprofile", "", "write memory profile to `file` after running container")
1907         runtimeEngine := flags.String("runtime-engine", "docker", "container runtime: docker or singularity")
1908         brokenNodeHook := flags.String("broken-node-hook", "", "script to run if node is detected to be broken (for example, Docker daemon is not running)")
1909         flags.Duration("check-containerd", 0, "Ignored. Exists for compatibility with older versions.")
1910         version := flags.Bool("version", false, "Write version information to stdout and exit 0.")
1911
1912         ignoreDetachFlag := false
1913         if len(args) > 0 && args[0] == "-no-detach" {
1914                 // This process was invoked by a parent process, which
1915                 // has passed along its own arguments, including
1916                 // -detach, after the leading -no-detach flag.  Strip
1917                 // the leading -no-detach flag (it's not recognized by
1918                 // flags.Parse()) and ignore the -detach flag that
1919                 // comes later.
1920                 args = args[1:]
1921                 ignoreDetachFlag = true
1922         }
1923
1924         if ok, code := cmd.ParseFlags(flags, prog, args, "container-uuid", stderr); !ok {
1925                 return code
1926         } else if *version {
1927                 fmt.Fprintln(stdout, prog, cmd.Version.String())
1928                 return 0
1929         } else if !*list && flags.NArg() != 1 {
1930                 fmt.Fprintf(stderr, "missing required argument: container-uuid (try -help)\n")
1931                 return 2
1932         }
1933
1934         containerUUID := flags.Arg(0)
1935
1936         switch {
1937         case *detach && !ignoreDetachFlag:
1938                 return Detach(containerUUID, prog, args, stdin, stdout, stderr)
1939         case *kill >= 0:
1940                 return KillProcess(containerUUID, syscall.Signal(*kill), stdout, stderr)
1941         case *list:
1942                 return ListProcesses(stdin, stdout, stderr)
1943         }
1944
1945         if len(containerUUID) != 27 {
1946                 log.Printf("usage: %s [options] UUID", prog)
1947                 return 1
1948         }
1949
1950         var keepstoreLogbuf bufThenWrite
1951         var conf ConfigData
1952         if *stdinConfig {
1953                 err := json.NewDecoder(stdin).Decode(&conf)
1954                 if err != nil {
1955                         log.Printf("decode stdin: %s", err)
1956                         return 1
1957                 }
1958                 for k, v := range conf.Env {
1959                         err = os.Setenv(k, v)
1960                         if err != nil {
1961                                 log.Printf("setenv(%q): %s", k, err)
1962                                 return 1
1963                         }
1964                 }
1965                 if conf.Cluster != nil {
1966                         // ClusterID is missing from the JSON
1967                         // representation, but we need it to generate
1968                         // a valid config file for keepstore, so we
1969                         // fill it using the container UUID prefix.
1970                         conf.Cluster.ClusterID = containerUUID[:5]
1971                 }
1972         } else {
1973                 conf = hpcConfData(containerUUID, *configFile, io.MultiWriter(&keepstoreLogbuf, stderr))
1974         }
1975
1976         log.Printf("crunch-run %s started", cmd.Version.String())
1977         time.Sleep(*sleep)
1978
1979         if *caCertsPath != "" {
1980                 arvadosclient.CertFiles = []string{*caCertsPath}
1981         }
1982
1983         keepstore, err := startLocalKeepstore(conf, io.MultiWriter(&keepstoreLogbuf, stderr))
1984         if err != nil {
1985                 log.Print(err)
1986                 return 1
1987         }
1988         if keepstore != nil {
1989                 defer keepstore.Process.Kill()
1990         }
1991
1992         api, err := arvadosclient.MakeArvadosClient()
1993         if err != nil {
1994                 log.Printf("%s: %v", containerUUID, err)
1995                 return 1
1996         }
1997         // arvadosclient now interprets Retries=10 to mean
1998         // Timeout=10m, retrying with exponential backoff + jitter.
1999         api.Retries = 10
2000
2001         kc, err := keepclient.MakeKeepClient(api)
2002         if err != nil {
2003                 log.Printf("%s: %v", containerUUID, err)
2004                 return 1
2005         }
2006         kc.BlockCache = &keepclient.BlockCache{MaxBlocks: 2}
2007         kc.Retries = 4
2008
2009         cr, err := NewContainerRunner(arvados.NewClientFromEnv(), api, kc, containerUUID)
2010         if err != nil {
2011                 log.Print(err)
2012                 return 1
2013         }
2014
2015         cr.keepstore = keepstore
2016         if keepstore == nil {
2017                 // Log explanation (if any) for why we're not running
2018                 // a local keepstore.
2019                 var buf bytes.Buffer
2020                 keepstoreLogbuf.SetWriter(&buf)
2021                 if buf.Len() > 0 {
2022                         cr.CrunchLog.Printf("%s", strings.TrimSpace(buf.String()))
2023                 }
2024         } else if logWhat := conf.Cluster.Containers.LocalKeepLogsToContainerLog; logWhat == "none" {
2025                 cr.CrunchLog.Printf("using local keepstore process (pid %d) at %s", keepstore.Process.Pid, os.Getenv("ARVADOS_KEEP_SERVICES"))
2026                 keepstoreLogbuf.SetWriter(io.Discard)
2027         } else {
2028                 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"))
2029                 logwriter, err := cr.NewLogWriter("keepstore")
2030                 if err != nil {
2031                         log.Print(err)
2032                         return 1
2033                 }
2034                 cr.keepstoreLogger = NewThrottledLogger(logwriter)
2035
2036                 var writer io.WriteCloser = cr.keepstoreLogger
2037                 if logWhat == "errors" {
2038                         writer = &filterKeepstoreErrorsOnly{WriteCloser: writer}
2039                 } else if logWhat != "all" {
2040                         // should have been caught earlier by
2041                         // dispatcher's config loader
2042                         log.Printf("invalid value for Containers.LocalKeepLogsToContainerLog: %q", logWhat)
2043                         return 1
2044                 }
2045                 err = keepstoreLogbuf.SetWriter(writer)
2046                 if err != nil {
2047                         log.Print(err)
2048                         return 1
2049                 }
2050                 cr.keepstoreLogbuf = &keepstoreLogbuf
2051         }
2052
2053         switch *runtimeEngine {
2054         case "docker":
2055                 cr.executor, err = newDockerExecutor(containerUUID, cr.CrunchLog.Printf, cr.containerWatchdogInterval)
2056         case "singularity":
2057                 cr.executor, err = newSingularityExecutor(cr.CrunchLog.Printf)
2058         default:
2059                 cr.CrunchLog.Printf("%s: unsupported RuntimeEngine %q", containerUUID, *runtimeEngine)
2060                 cr.CrunchLog.Close()
2061                 return 1
2062         }
2063         if err != nil {
2064                 cr.CrunchLog.Printf("%s: %v", containerUUID, err)
2065                 cr.checkBrokenNode(err)
2066                 cr.CrunchLog.Close()
2067                 return 1
2068         }
2069         defer cr.executor.Close()
2070
2071         cr.brokenNodeHook = *brokenNodeHook
2072
2073         gwAuthSecret := os.Getenv("GatewayAuthSecret")
2074         os.Unsetenv("GatewayAuthSecret")
2075         if gwAuthSecret == "" {
2076                 // not safe to run a gateway service without an auth
2077                 // secret
2078                 cr.CrunchLog.Printf("Not starting a gateway server (GatewayAuthSecret was not provided by dispatcher)")
2079         } else {
2080                 gwListen := os.Getenv("GatewayAddress")
2081                 cr.gateway = Gateway{
2082                         Address:       gwListen,
2083                         AuthSecret:    gwAuthSecret,
2084                         ContainerUUID: containerUUID,
2085                         Target:        cr.executor,
2086                         Log:           cr.CrunchLog,
2087                         LogCollection: cr.LogCollection,
2088                 }
2089                 if gwListen == "" {
2090                         // Direct connection won't work, so we use the
2091                         // gateway_address field to indicate the
2092                         // internalURL of the controller process that
2093                         // has the current tunnel connection.
2094                         cr.gateway.ArvadosClient = cr.dispatcherClient
2095                         cr.gateway.UpdateTunnelURL = func(url string) {
2096                                 cr.gateway.Address = "tunnel " + url
2097                                 cr.DispatcherArvClient.Update("containers", containerUUID,
2098                                         arvadosclient.Dict{"container": arvadosclient.Dict{"gateway_address": cr.gateway.Address}}, nil)
2099                         }
2100                 }
2101                 err = cr.gateway.Start()
2102                 if err != nil {
2103                         log.Printf("error starting gateway server: %s", err)
2104                         return 1
2105                 }
2106         }
2107
2108         parentTemp, tmperr := cr.MkTempDir("", "crunch-run."+containerUUID+".")
2109         if tmperr != nil {
2110                 log.Printf("%s: %v", containerUUID, tmperr)
2111                 return 1
2112         }
2113
2114         cr.parentTemp = parentTemp
2115         cr.statInterval = *statInterval
2116         cr.cgroupRoot = *cgroupRoot
2117         cr.expectCgroupParent = *cgroupParent
2118         cr.enableMemoryLimit = *enableMemoryLimit
2119         cr.enableNetwork = *enableNetwork
2120         cr.networkMode = *networkMode
2121         if *cgroupParentSubsystem != "" {
2122                 p, err := findCgroup(*cgroupParentSubsystem)
2123                 if err != nil {
2124                         log.Printf("fatal: cgroup parent subsystem: %s", err)
2125                         return 1
2126                 }
2127                 cr.setCgroupParent = p
2128                 cr.expectCgroupParent = p
2129         }
2130
2131         if conf.EC2SpotCheck {
2132                 go cr.checkSpotInterruptionNotices()
2133         }
2134
2135         runerr := cr.Run()
2136
2137         if *memprofile != "" {
2138                 f, err := os.Create(*memprofile)
2139                 if err != nil {
2140                         log.Printf("could not create memory profile: %s", err)
2141                 }
2142                 runtime.GC() // get up-to-date statistics
2143                 if err := pprof.WriteHeapProfile(f); err != nil {
2144                         log.Printf("could not write memory profile: %s", err)
2145                 }
2146                 closeerr := f.Close()
2147                 if closeerr != nil {
2148                         log.Printf("closing memprofile file: %s", err)
2149                 }
2150         }
2151
2152         if runerr != nil {
2153                 log.Printf("%s: %v", containerUUID, runerr)
2154                 return 1
2155         }
2156         return 0
2157 }
2158
2159 // Try to load ConfigData in hpc (slurm/lsf) environment. This means
2160 // loading the cluster config from the specified file and (if that
2161 // works) getting the runtime_constraints container field from
2162 // controller to determine # VCPUs so we can calculate KeepBuffers.
2163 func hpcConfData(uuid string, configFile string, stderr io.Writer) ConfigData {
2164         var conf ConfigData
2165         conf.Cluster = loadClusterConfigFile(configFile, stderr)
2166         if conf.Cluster == nil {
2167                 // skip loading the container record -- we won't be
2168                 // able to start local keepstore anyway.
2169                 return conf
2170         }
2171         arv, err := arvadosclient.MakeArvadosClient()
2172         if err != nil {
2173                 fmt.Fprintf(stderr, "error setting up arvadosclient: %s\n", err)
2174                 return conf
2175         }
2176         // arvadosclient now interprets Retries=10 to mean
2177         // Timeout=10m, retrying with exponential backoff + jitter.
2178         arv.Retries = 10
2179         var ctr arvados.Container
2180         err = arv.Call("GET", "containers", uuid, "", arvadosclient.Dict{"select": []string{"runtime_constraints"}}, &ctr)
2181         if err != nil {
2182                 fmt.Fprintf(stderr, "error getting container record: %s\n", err)
2183                 return conf
2184         }
2185         if ctr.RuntimeConstraints.VCPUs > 0 {
2186                 conf.KeepBuffers = ctr.RuntimeConstraints.VCPUs * conf.Cluster.Containers.LocalKeepBlobBuffersPerVCPU
2187         }
2188         return conf
2189 }
2190
2191 // Load cluster config file from given path. If an error occurs, log
2192 // the error to stderr and return nil.
2193 func loadClusterConfigFile(path string, stderr io.Writer) *arvados.Cluster {
2194         ldr := config.NewLoader(&bytes.Buffer{}, ctxlog.New(stderr, "plain", "info"))
2195         ldr.Path = path
2196         cfg, err := ldr.Load()
2197         if err != nil {
2198                 fmt.Fprintf(stderr, "could not load config file %s: %s\n", path, err)
2199                 return nil
2200         }
2201         cluster, err := cfg.GetCluster("")
2202         if err != nil {
2203                 fmt.Fprintf(stderr, "could not use config file %s: %s\n", path, err)
2204                 return nil
2205         }
2206         fmt.Fprintf(stderr, "loaded config file %s\n", path)
2207         return cluster
2208 }
2209
2210 func startLocalKeepstore(configData ConfigData, logbuf io.Writer) (*exec.Cmd, error) {
2211         if configData.KeepBuffers < 1 {
2212                 fmt.Fprintf(logbuf, "not starting a local keepstore process because KeepBuffers=%v in config\n", configData.KeepBuffers)
2213                 return nil, nil
2214         }
2215         if configData.Cluster == nil {
2216                 fmt.Fprint(logbuf, "not starting a local keepstore process because cluster config file was not loaded\n")
2217                 return nil, nil
2218         }
2219         for uuid, vol := range configData.Cluster.Volumes {
2220                 if len(vol.AccessViaHosts) > 0 {
2221                         fmt.Fprintf(logbuf, "not starting a local keepstore process because a volume (%s) uses AccessViaHosts\n", uuid)
2222                         return nil, nil
2223                 }
2224                 if !vol.ReadOnly && vol.Replication < configData.Cluster.Collections.DefaultReplication {
2225                         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)
2226                         return nil, nil
2227                 }
2228         }
2229
2230         // Rather than have an alternate way to tell keepstore how
2231         // many buffers to use when starting it this way, we just
2232         // modify the cluster configuration that we feed it on stdin.
2233         configData.Cluster.API.MaxKeepBlobBuffers = configData.KeepBuffers
2234
2235         localaddr := localKeepstoreAddr()
2236         ln, err := net.Listen("tcp", net.JoinHostPort(localaddr, "0"))
2237         if err != nil {
2238                 return nil, err
2239         }
2240         _, port, err := net.SplitHostPort(ln.Addr().String())
2241         if err != nil {
2242                 ln.Close()
2243                 return nil, err
2244         }
2245         ln.Close()
2246         url := "http://" + net.JoinHostPort(localaddr, port)
2247
2248         fmt.Fprintf(logbuf, "starting keepstore on %s\n", url)
2249
2250         var confJSON bytes.Buffer
2251         err = json.NewEncoder(&confJSON).Encode(arvados.Config{
2252                 Clusters: map[string]arvados.Cluster{
2253                         configData.Cluster.ClusterID: *configData.Cluster,
2254                 },
2255         })
2256         if err != nil {
2257                 return nil, err
2258         }
2259         cmd := exec.Command("/proc/self/exe", "keepstore", "-config=-")
2260         if target, err := os.Readlink(cmd.Path); err == nil && strings.HasSuffix(target, ".test") {
2261                 // If we're a 'go test' process, running
2262                 // /proc/self/exe would start the test suite in a
2263                 // child process, which is not what we want.
2264                 cmd.Path, _ = exec.LookPath("go")
2265                 cmd.Args = append([]string{"go", "run", "../../cmd/arvados-server"}, cmd.Args[1:]...)
2266                 cmd.Env = os.Environ()
2267         }
2268         cmd.Stdin = &confJSON
2269         cmd.Stdout = logbuf
2270         cmd.Stderr = logbuf
2271         cmd.Env = append(cmd.Env,
2272                 "GOGC=10",
2273                 "ARVADOS_SERVICE_INTERNAL_URL="+url)
2274         err = cmd.Start()
2275         if err != nil {
2276                 return nil, fmt.Errorf("error starting keepstore process: %w", err)
2277         }
2278         cmdExited := false
2279         go func() {
2280                 cmd.Wait()
2281                 cmdExited = true
2282         }()
2283         ctx, cancel := context.WithDeadline(context.Background(), time.Now().Add(time.Second*10))
2284         defer cancel()
2285         poll := time.NewTicker(time.Second / 10)
2286         defer poll.Stop()
2287         client := http.Client{}
2288         for range poll.C {
2289                 testReq, err := http.NewRequestWithContext(ctx, "GET", url+"/_health/ping", nil)
2290                 testReq.Header.Set("Authorization", "Bearer "+configData.Cluster.ManagementToken)
2291                 if err != nil {
2292                         return nil, err
2293                 }
2294                 resp, err := client.Do(testReq)
2295                 if err == nil {
2296                         resp.Body.Close()
2297                         if resp.StatusCode == http.StatusOK {
2298                                 break
2299                         }
2300                 }
2301                 if cmdExited {
2302                         return nil, fmt.Errorf("keepstore child process exited")
2303                 }
2304                 if ctx.Err() != nil {
2305                         return nil, fmt.Errorf("timed out waiting for new keepstore process to report healthy")
2306                 }
2307         }
2308         os.Setenv("ARVADOS_KEEP_SERVICES", url)
2309         return cmd, nil
2310 }
2311
2312 // return current uid, gid, groups in a format suitable for logging:
2313 // "crunch-run process has uid=1234(arvados) gid=1234(arvados)
2314 // groups=1234(arvados),114(fuse)"
2315 func currentUserAndGroups() string {
2316         u, err := user.Current()
2317         if err != nil {
2318                 return fmt.Sprintf("error getting current user ID: %s", err)
2319         }
2320         s := fmt.Sprintf("crunch-run process has uid=%s(%s) gid=%s", u.Uid, u.Username, u.Gid)
2321         if g, err := user.LookupGroupId(u.Gid); err == nil {
2322                 s += fmt.Sprintf("(%s)", g.Name)
2323         }
2324         s += " groups="
2325         if gids, err := u.GroupIds(); err == nil {
2326                 for i, gid := range gids {
2327                         if i > 0 {
2328                                 s += ","
2329                         }
2330                         s += gid
2331                         if g, err := user.LookupGroupId(gid); err == nil {
2332                                 s += fmt.Sprintf("(%s)", g.Name)
2333                         }
2334                 }
2335         }
2336         return s
2337 }
2338
2339 // Return a suitable local interface address for a local keepstore
2340 // service. Currently this is the numerically lowest non-loopback ipv4
2341 // address assigned to a local interface that is not in any of the
2342 // link-local/vpn/loopback ranges 169.254/16, 100.64/10, or 127/8.
2343 func localKeepstoreAddr() string {
2344         var ips []net.IP
2345         // Ignore error (proceed with zero IPs)
2346         addrs, _ := processIPs(os.Getpid())
2347         for addr := range addrs {
2348                 ip := net.ParseIP(addr)
2349                 if ip == nil {
2350                         // invalid
2351                         continue
2352                 }
2353                 if ip.Mask(net.CIDRMask(8, 32)).Equal(net.IPv4(127, 0, 0, 0)) ||
2354                         ip.Mask(net.CIDRMask(10, 32)).Equal(net.IPv4(100, 64, 0, 0)) ||
2355                         ip.Mask(net.CIDRMask(16, 32)).Equal(net.IPv4(169, 254, 0, 0)) {
2356                         // unsuitable
2357                         continue
2358                 }
2359                 ips = append(ips, ip)
2360         }
2361         if len(ips) == 0 {
2362                 return "0.0.0.0"
2363         }
2364         sort.Slice(ips, func(ii, jj int) bool {
2365                 i, j := ips[ii], ips[jj]
2366                 if len(i) != len(j) {
2367                         return len(i) < len(j)
2368                 }
2369                 for x := range i {
2370                         if i[x] != j[x] {
2371                                 return i[x] < j[x]
2372                         }
2373                 }
2374                 return false
2375         })
2376         return ips[0].String()
2377 }
2378
2379 func (cr *ContainerRunner) loadPrices() {
2380         buf, err := os.ReadFile(filepath.Join(lockdir, pricesfile))
2381         if err != nil {
2382                 if !os.IsNotExist(err) {
2383                         cr.CrunchLog.Printf("loadPrices: read: %s", err)
2384                 }
2385                 return
2386         }
2387         var prices []cloud.InstancePrice
2388         err = json.Unmarshal(buf, &prices)
2389         if err != nil {
2390                 cr.CrunchLog.Printf("loadPrices: decode: %s", err)
2391                 return
2392         }
2393         cr.pricesLock.Lock()
2394         defer cr.pricesLock.Unlock()
2395         var lastKnown time.Time
2396         if len(cr.prices) > 0 {
2397                 lastKnown = cr.prices[0].StartTime
2398         }
2399         cr.prices = cloud.NormalizePriceHistory(append(prices, cr.prices...))
2400         for i := len(cr.prices) - 1; i >= 0; i-- {
2401                 price := cr.prices[i]
2402                 if price.StartTime.After(lastKnown) {
2403                         cr.CrunchLog.Printf("Instance price changed to %#.3g at %s", price.Price, price.StartTime.UTC())
2404                 }
2405         }
2406 }
2407
2408 func (cr *ContainerRunner) calculateCost(now time.Time) float64 {
2409         cr.pricesLock.Lock()
2410         defer cr.pricesLock.Unlock()
2411
2412         // First, make a "prices" slice with the real data as far back
2413         // as it goes, and (if needed) a "since the beginning of time"
2414         // placeholder containing a reasonable guess about what the
2415         // price was between cr.costStartTime and the earliest real
2416         // data point.
2417         prices := cr.prices
2418         if len(prices) == 0 {
2419                 // use price info in InstanceType record initially
2420                 // provided by cloud dispatcher
2421                 var p float64
2422                 var it arvados.InstanceType
2423                 if j := os.Getenv("InstanceType"); j != "" && json.Unmarshal([]byte(j), &it) == nil && it.Price > 0 {
2424                         p = it.Price
2425                 }
2426                 prices = []cloud.InstancePrice{{Price: p}}
2427         } else if prices[len(prices)-1].StartTime.After(cr.costStartTime) {
2428                 // guess earlier pricing was the same as the earliest
2429                 // price we know about
2430                 filler := prices[len(prices)-1]
2431                 filler.StartTime = time.Time{}
2432                 prices = append(prices, filler)
2433         }
2434
2435         // Now that our history of price changes goes back at least as
2436         // far as cr.costStartTime, add up the costs for each
2437         // interval.
2438         cost := 0.0
2439         spanEnd := now
2440         for _, ip := range prices {
2441                 spanStart := ip.StartTime
2442                 if spanStart.After(now) {
2443                         // pricing information from the future -- not
2444                         // expected from AWS, but possible in
2445                         // principle, and exercised by tests.
2446                         continue
2447                 }
2448                 last := false
2449                 if spanStart.Before(cr.costStartTime) {
2450                         spanStart = cr.costStartTime
2451                         last = true
2452                 }
2453                 cost += ip.Price * spanEnd.Sub(spanStart).Seconds() / 3600
2454                 if last {
2455                         break
2456                 }
2457                 spanEnd = spanStart
2458         }
2459
2460         return cost
2461 }
2462
2463 func (runner *ContainerRunner) handleSIGUSR2(sigchan chan os.Signal) {
2464         for range sigchan {
2465                 runner.loadPrices()
2466                 update := arvadosclient.Dict{
2467                         "container": arvadosclient.Dict{
2468                                 "cost": runner.calculateCost(time.Now()),
2469                         },
2470                 }
2471                 runner.DispatcherArvClient.Update("containers", runner.Container.UUID, update, nil)
2472         }
2473 }