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/fs"
  16         "io/ioutil"
  17         "log"
  18         "net"
  19         "net/http"
  20         "os"
  21         "os/exec"
  22         "os/signal"
  23         "os/user"
  24         "path"
  25         "path/filepath"
  26         "regexp"
  27         "runtime"
  28         "runtime/pprof"
  29         "sort"
  30         "strings"
  31         "sync"
  32         "syscall"
  33         "time"
  34
  35         "git.arvados.org/arvados.git/lib/cloud"
  36         "git.arvados.org/arvados.git/lib/cmd"
  37         "git.arvados.org/arvados.git/lib/config"
  38         "git.arvados.org/arvados.git/lib/crunchstat"
  39         "git.arvados.org/arvados.git/sdk/go/arvados"
  40         "git.arvados.org/arvados.git/sdk/go/arvadosclient"
  41         "git.arvados.org/arvados.git/sdk/go/ctxlog"
  42         "git.arvados.org/arvados.git/sdk/go/keepclient"
  43         "golang.org/x/sys/unix"
  44 )
  45
  46 type command struct{}
  47
  48 var arvadosCertPath = "/etc/arvados/ca-certificates.crt"
  49
  50 var Command = command{}
  51
  52 // ConfigData contains environment variables and (when needed) cluster
  53 // configuration, passed from dispatchcloud to crunch-run on stdin.
  54 type ConfigData struct {
  55         Env          map[string]string
  56         KeepBuffers  int
  57         EC2SpotCheck bool
  58         Cluster      *arvados.Cluster
  59 }
  60
  61 // IArvadosClient is the minimal Arvados API methods used by crunch-run.
  62 type IArvadosClient interface {
  63         Create(resourceType string, parameters arvadosclient.Dict, output interface{}) error
  64         Get(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  65         Update(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  66         Call(method, resourceType, uuid, action string, parameters arvadosclient.Dict, output interface{}) error
  67         CallRaw(method string, resourceType string, uuid string, action string, parameters arvadosclient.Dict) (reader io.ReadCloser, err error)
  68         Discovery(key string) (interface{}, error)
  69 }
  70
  71 // ErrCancelled is the error returned when the container is cancelled.
  72 var ErrCancelled = errors.New("Cancelled")
  73
  74 // IKeepClient is the minimal Keep API methods used by crunch-run.
  75 type IKeepClient interface {
  76         BlockRead(context.Context, arvados.BlockReadOptions) (int, error)
  77         BlockWrite(context.Context, arvados.BlockWriteOptions) (arvados.BlockWriteResponse, error)
  78         ReadAt(locator string, p []byte, off int) (int, error)
  79         LocalLocator(locator string) (string, error)
  80         SetStorageClasses(sc []string)
  81 }
  82
  83 type RunArvMount func(cmdline []string, tok string) (*exec.Cmd, error)
  84
  85 type MkTempDir func(string, string) (string, error)
  86
  87 type PsProcess interface {
  88         CmdlineSlice() ([]string, error)
  89 }
  90
  91 // ContainerRunner is the main stateful struct used for a single execution of a
  92 // container.
  93 type ContainerRunner struct {
  94         executor       containerExecutor
  95         executorStdin  io.Closer
  96         executorStdout io.Closer
  97         executorStderr io.Closer
  98
  99         // Dispatcher client is initialized with the Dispatcher token.
 100         // This is a privileged token used to manage container status
 101         // and logs.
 102         //
 103         // We have both dispatcherClient and DispatcherArvClient
 104         // because there are two different incompatible Arvados Go
 105         // SDKs and we have to use both (hopefully this gets fixed in
 106         // #14467)
 107         dispatcherClient     *arvados.Client
 108         DispatcherArvClient  IArvadosClient
 109         DispatcherKeepClient IKeepClient
 110
 111         // Container client is initialized with the Container token
 112         // This token controls the permissions of the container, and
 113         // must be used for operations such as reading collections.
 114         //
 115         // Same comment as above applies to
 116         // containerClient/ContainerArvClient.
 117         containerClient     *arvados.Client
 118         ContainerArvClient  IArvadosClient
 119         ContainerKeepClient IKeepClient
 120
 121         Container     arvados.Container
 122         token         string
 123         ExitCode      *int
 124         CrunchLog     *logWriter
 125         logUUID       string
 126         logMtx        sync.Mutex
 127         LogCollection arvados.CollectionFileSystem
 128         LogsPDH       *string
 129         RunArvMount   RunArvMount
 130         MkTempDir     MkTempDir
 131         ArvMount      *exec.Cmd
 132         ArvMountPoint string
 133         HostOutputDir string
 134         Volumes       map[string]struct{}
 135         OutputPDH     *string
 136         SigChan       chan os.Signal
 137         ArvMountExit  chan error
 138         SecretMounts  map[string]arvados.Mount
 139         MkArvClient   func(token string) (IArvadosClient, IKeepClient, *arvados.Client, error)
 140         finalState    string
 141         parentTemp    string
 142         costStartTime time.Time
 143
 144         keepstore        *exec.Cmd
 145         keepstoreLogger  io.WriteCloser
 146         keepstoreLogbuf  *bufThenWrite
 147         statLogger       io.WriteCloser
 148         statReporter     *crunchstat.Reporter
 149         hoststatLogger   io.WriteCloser
 150         hoststatReporter *crunchstat.Reporter
 151         statInterval     time.Duration
 152         // What we tell docker to use as the container's cgroup
 153         // parent.
 154         setCgroupParent string
 155         // Fake root dir where crunchstat.Reporter should read OS
 156         // files, for testing.
 157         crunchstatFakeFS fs.FS
 158
 159         cStateLock sync.Mutex
 160         cCancelled bool // StopContainer() invoked
 161
 162         enableMemoryLimit bool
 163         enableNetwork     string // one of "default" or "always"
 164         networkMode       string // "none", "host", or "" -- passed through to executor
 165         brokenNodeHook    string // script to run if node appears to be broken
 166         arvMountLog       io.WriteCloser
 167
 168         containerWatchdogInterval time.Duration
 169
 170         gateway Gateway
 171
 172         prices     []cloud.InstancePrice
 173         pricesLock sync.Mutex
 174 }
 175
 176 // setupSignals sets up signal handling to gracefully terminate the
 177 // underlying container and update state when receiving a TERM, INT or
 178 // QUIT signal.
 179 func (runner *ContainerRunner) setupSignals() {
 180         runner.SigChan = make(chan os.Signal, 1)
 181         signal.Notify(runner.SigChan, syscall.SIGTERM)
 182         signal.Notify(runner.SigChan, syscall.SIGINT)
 183         signal.Notify(runner.SigChan, syscall.SIGQUIT)
 184
 185         go func(sig chan os.Signal) {
 186                 for s := range sig {
 187                         runner.stop(s)
 188                 }
 189         }(runner.SigChan)
 190 }
 191
 192 // stop the underlying container.
 193 func (runner *ContainerRunner) stop(sig os.Signal) {
 194         runner.cStateLock.Lock()
 195         defer runner.cStateLock.Unlock()
 196         if sig != nil {
 197                 runner.CrunchLog.Printf("caught signal: %v", sig)
 198         }
 199         runner.cCancelled = true
 200         runner.CrunchLog.Printf("stopping container")
 201         err := runner.executor.Stop()
 202         if err != nil {
 203                 runner.CrunchLog.Printf("error stopping container: %s", err)
 204         }
 205 }
 206
 207 var errorBlacklist = []string{
 208         "(?ms).*[Cc]annot connect to the Docker daemon.*",
 209         "(?ms).*oci runtime error.*starting container process.*container init.*mounting.*to rootfs.*no such file or directory.*",
 210         "(?ms).*grpc: the connection is unavailable.*",
 211         "(?ms).*no space left on device.*",
 212 }
 213
 214 func (runner *ContainerRunner) runBrokenNodeHook() {
 215         if runner.brokenNodeHook == "" {
 216                 path := filepath.Join(lockdir, brokenfile)
 217                 runner.CrunchLog.Printf("Writing %s to mark node as broken", path)
 218                 f, err := os.OpenFile(path, os.O_CREATE|os.O_RDWR, 0700)
 219                 if err != nil {
 220                         runner.CrunchLog.Printf("Error writing %s: %s", path, err)
 221                         return
 222                 }
 223                 f.Close()
 224         } else {
 225                 runner.CrunchLog.Printf("Running broken node hook %q", runner.brokenNodeHook)
 226                 // run killme script
 227                 c := exec.Command(runner.brokenNodeHook)
 228                 c.Stdout = runner.CrunchLog
 229                 c.Stderr = runner.CrunchLog
 230                 err := c.Run()
 231                 if err != nil {
 232                         runner.CrunchLog.Printf("Error running broken node hook: %v", err)
 233                 }
 234         }
 235 }
 236
 237 func (runner *ContainerRunner) checkBrokenNode(goterr error) bool {
 238         for _, d := range errorBlacklist {
 239                 if m, e := regexp.MatchString(d, goterr.Error()); m && e == nil {
 240                         runner.CrunchLog.Printf("Error suggests node is unable to run containers: %v", goterr)
 241                         runner.runBrokenNodeHook()
 242                         return true
 243                 }
 244         }
 245         return false
 246 }
 247
 248 // LoadImage determines the docker image id from the container record and
 249 // checks if it is available in the local Docker image store.  If not, it loads
 250 // the image from Keep.
 251 func (runner *ContainerRunner) LoadImage() (string, error) {
 252         runner.CrunchLog.Printf("Fetching Docker image from collection '%s'", runner.Container.ContainerImage)
 253
 254         d, err := os.Open(runner.ArvMountPoint + "/by_id/" + runner.Container.ContainerImage)
 255         if err != nil {
 256                 return "", err
 257         }
 258         defer d.Close()
 259         allfiles, err := d.Readdirnames(-1)
 260         if err != nil {
 261                 return "", err
 262         }
 263         var tarfiles []string
 264         for _, fnm := range allfiles {
 265                 if strings.HasSuffix(fnm, ".tar") {
 266                         tarfiles = append(tarfiles, fnm)
 267                 }
 268         }
 269         if len(tarfiles) == 0 {
 270                 return "", fmt.Errorf("image collection does not include a .tar image file")
 271         }
 272         if len(tarfiles) > 1 {
 273                 return "", fmt.Errorf("cannot choose from multiple tar files in image collection: %v", tarfiles)
 274         }
 275         imageID := tarfiles[0][:len(tarfiles[0])-4]
 276         imageTarballPath := runner.ArvMountPoint + "/by_id/" + runner.Container.ContainerImage + "/" + imageID + ".tar"
 277         runner.CrunchLog.Printf("Using Docker image id %q", imageID)
 278
 279         runner.CrunchLog.Print("Loading Docker image from keep")
 280         err = runner.executor.LoadImage(imageID, imageTarballPath, runner.Container, runner.ArvMountPoint,
 281                 runner.containerClient)
 282         if err != nil {
 283                 return "", err
 284         }
 285
 286         return imageID, nil
 287 }
 288
 289 func (runner *ContainerRunner) ArvMountCmd(cmdline []string, token string) (c *exec.Cmd, err error) {
 290         c = exec.Command(cmdline[0], cmdline[1:]...)
 291
 292         // Copy our environment, but override ARVADOS_API_TOKEN with
 293         // the container auth token.
 294         c.Env = nil
 295         for _, s := range os.Environ() {
 296                 if !strings.HasPrefix(s, "ARVADOS_API_TOKEN=") {
 297                         c.Env = append(c.Env, s)
 298                 }
 299         }
 300         c.Env = append(c.Env, "ARVADOS_API_TOKEN="+token)
 301
 302         runner.arvMountLog, err = runner.openLogFile("arv-mount")
 303         if err != nil {
 304                 return nil, err
 305         }
 306         scanner := logScanner{
 307                 Patterns: []string{
 308                         "Keep write error",
 309                         "Block not found error",
 310                         "Unhandled exception during FUSE operation",
 311                 },
 312                 ReportFunc: func(pattern, text string) {
 313                         runner.updateRuntimeStatus(arvadosclient.Dict{
 314                                 "warning":       "arv-mount: " + pattern,
 315                                 "warningDetail": text,
 316                         })
 317                 },
 318         }
 319         c.Stdout = newTimestamper(io.MultiWriter(runner.arvMountLog, os.Stderr))
 320         c.Stderr = io.MultiWriter(&scanner, newTimestamper(io.MultiWriter(runner.arvMountLog, os.Stderr)))
 321
 322         runner.CrunchLog.Printf("Running %v", c.Args)
 323
 324         err = c.Start()
 325         if err != nil {
 326                 return nil, err
 327         }
 328
 329         statReadme := make(chan bool)
 330         runner.ArvMountExit = make(chan error)
 331
 332         keepStatting := true
 333         go func() {
 334                 for keepStatting {
 335                         time.Sleep(100 * time.Millisecond)
 336                         _, err = os.Stat(fmt.Sprintf("%s/by_id/README", runner.ArvMountPoint))
 337                         if err == nil {
 338                                 keepStatting = false
 339                                 statReadme <- true
 340                         }
 341                 }
 342                 close(statReadme)
 343         }()
 344
 345         go func() {
 346                 mnterr := c.Wait()
 347                 if mnterr != nil {
 348                         runner.CrunchLog.Printf("Arv-mount exit error: %v", mnterr)
 349                 }
 350                 runner.ArvMountExit <- mnterr
 351                 close(runner.ArvMountExit)
 352         }()
 353
 354         select {
 355         case <-statReadme:
 356                 break
 357         case err := <-runner.ArvMountExit:
 358                 runner.ArvMount = nil
 359                 keepStatting = false
 360                 return nil, err
 361         }
 362
 363         return c, nil
 364 }
 365
 366 func (runner *ContainerRunner) SetupArvMountPoint(prefix string) (err error) {
 367         if runner.ArvMountPoint == "" {
 368                 runner.ArvMountPoint, err = runner.MkTempDir(runner.parentTemp, prefix)
 369         }
 370         return
 371 }
 372
 373 func copyfile(src string, dst string) (err error) {
 374         srcfile, err := os.Open(src)
 375         if err != nil {
 376                 return
 377         }
 378
 379         os.MkdirAll(path.Dir(dst), 0777)
 380
 381         dstfile, err := os.Create(dst)
 382         if err != nil {
 383                 return
 384         }
 385         _, err = io.Copy(dstfile, srcfile)
 386         if err != nil {
 387                 return
 388         }
 389
 390         err = srcfile.Close()
 391         err2 := dstfile.Close()
 392
 393         if err != nil {
 394                 return
 395         }
 396
 397         if err2 != nil {
 398                 return err2
 399         }
 400
 401         return nil
 402 }
 403
 404 func (runner *ContainerRunner) SetupMounts() (map[string]bindmount, error) {
 405         bindmounts := map[string]bindmount{}
 406         err := runner.SetupArvMountPoint("keep")
 407         if err != nil {
 408                 return nil, fmt.Errorf("While creating keep mount temp dir: %v", err)
 409         }
 410
 411         token, err := runner.ContainerToken()
 412         if err != nil {
 413                 return nil, fmt.Errorf("could not get container token: %s", err)
 414         }
 415         runner.CrunchLog.Printf("container token %q", token)
 416
 417         pdhOnly := true
 418         tmpcount := 0
 419         arvMountCmd := []string{
 420                 "arv-mount",
 421                 "--foreground",
 422                 "--read-write",
 423                 "--storage-classes", strings.Join(runner.Container.OutputStorageClasses, ","),
 424                 fmt.Sprintf("--crunchstat-interval=%v", runner.statInterval.Seconds())}
 425
 426         if _, isdocker := runner.executor.(*dockerExecutor); isdocker {
 427                 arvMountCmd = append(arvMountCmd, "--allow-other")
 428         }
 429
 430         if runner.Container.RuntimeConstraints.KeepCacheDisk > 0 {
 431                 keepcachedir, err := runner.MkTempDir(runner.parentTemp, "keepcache")
 432                 if err != nil {
 433                         return nil, fmt.Errorf("while creating keep cache temp dir: %v", err)
 434                 }
 435                 arvMountCmd = append(arvMountCmd, "--disk-cache", "--disk-cache-dir", keepcachedir, "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheDisk))
 436         } else if runner.Container.RuntimeConstraints.KeepCacheRAM > 0 {
 437                 arvMountCmd = append(arvMountCmd, "--ram-cache", "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheRAM))
 438         }
 439
 440         collectionPaths := []string{}
 441         needCertMount := true
 442         type copyFile struct {
 443                 src  string
 444                 bind string
 445         }
 446         var copyFiles []copyFile
 447
 448         var binds []string
 449         for bind := range runner.Container.Mounts {
 450                 binds = append(binds, bind)
 451         }
 452         for bind := range runner.SecretMounts {
 453                 if _, ok := runner.Container.Mounts[bind]; ok {
 454                         return nil, fmt.Errorf("secret mount %q conflicts with regular mount", bind)
 455                 }
 456                 if runner.SecretMounts[bind].Kind != "json" &&
 457                         runner.SecretMounts[bind].Kind != "text" {
 458                         return nil, fmt.Errorf("secret mount %q type is %q but only 'json' and 'text' are permitted",
 459                                 bind, runner.SecretMounts[bind].Kind)
 460                 }
 461                 binds = append(binds, bind)
 462         }
 463         sort.Strings(binds)
 464
 465         for _, bind := range binds {
 466                 mnt, notSecret := runner.Container.Mounts[bind]
 467                 if !notSecret {
 468                         mnt = runner.SecretMounts[bind]
 469                 }
 470                 if bind == "stdout" || bind == "stderr" {
 471                         // Is it a "file" mount kind?
 472                         if mnt.Kind != "file" {
 473                                 return nil, fmt.Errorf("unsupported mount kind '%s' for %s: only 'file' is supported", mnt.Kind, bind)
 474                         }
 475
 476                         // Does path start with OutputPath?
 477                         prefix := runner.Container.OutputPath
 478                         if !strings.HasSuffix(prefix, "/") {
 479                                 prefix += "/"
 480                         }
 481                         if !strings.HasPrefix(mnt.Path, prefix) {
 482                                 return nil, fmt.Errorf("%s path does not start with OutputPath: %s, %s", strings.Title(bind), mnt.Path, prefix)
 483                         }
 484                 }
 485
 486                 if bind == "stdin" {
 487                         // Is it a "collection" mount kind?
 488                         if mnt.Kind != "collection" && mnt.Kind != "json" {
 489                                 return nil, fmt.Errorf("unsupported mount kind '%s' for stdin: only 'collection' and 'json' are supported", mnt.Kind)
 490                         }
 491                 }
 492
 493                 if bind == arvadosCertPath {
 494                         needCertMount = false
 495                 }
 496
 497                 if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && bind != runner.Container.OutputPath+"/" {
 498                         if mnt.Kind != "collection" && mnt.Kind != "text" && mnt.Kind != "json" {
 499                                 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)
 500                         }
 501                 }
 502
 503                 switch {
 504                 case mnt.Kind == "collection" && bind != "stdin":
 505                         var src string
 506                         if mnt.UUID != "" && mnt.PortableDataHash != "" {
 507                                 return nil, fmt.Errorf("cannot specify both 'uuid' and 'portable_data_hash' for a collection mount")
 508                         }
 509                         if mnt.UUID != "" {
 510                                 if mnt.Writable {
 511                                         return nil, fmt.Errorf("writing to existing collections currently not permitted")
 512                                 }
 513                                 pdhOnly = false
 514                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.UUID)
 515                         } else if mnt.PortableDataHash != "" {
 516                                 if mnt.Writable && !strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 517                                         return nil, fmt.Errorf("can never write to a collection specified by portable data hash")
 518                                 }
 519                                 idx := strings.Index(mnt.PortableDataHash, "/")
 520                                 if idx > 0 {
 521                                         mnt.Path = path.Clean(mnt.PortableDataHash[idx:])
 522                                         mnt.PortableDataHash = mnt.PortableDataHash[0:idx]
 523                                         runner.Container.Mounts[bind] = mnt
 524                                 }
 525                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.PortableDataHash)
 526                                 if mnt.Path != "" && mnt.Path != "." {
 527                                         if strings.HasPrefix(mnt.Path, "./") {
 528                                                 mnt.Path = mnt.Path[2:]
 529                                         } else if strings.HasPrefix(mnt.Path, "/") {
 530                                                 mnt.Path = mnt.Path[1:]
 531                                         }
 532                                         src += "/" + mnt.Path
 533                                 }
 534                         } else {
 535                                 src = fmt.Sprintf("%s/tmp%d", runner.ArvMountPoint, tmpcount)
 536                                 arvMountCmd = append(arvMountCmd, "--mount-tmp", fmt.Sprintf("tmp%d", tmpcount))
 537                                 tmpcount++
 538                         }
 539                         if mnt.Writable {
 540                                 if bind == runner.Container.OutputPath {
 541                                         runner.HostOutputDir = src
 542                                         bindmounts[bind] = bindmount{HostPath: src}
 543                                 } else if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 544                                         copyFiles = append(copyFiles, copyFile{src, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
 545                                 } else {
 546                                         bindmounts[bind] = bindmount{HostPath: src}
 547                                 }
 548                         } else {
 549                                 bindmounts[bind] = bindmount{HostPath: src, ReadOnly: true}
 550                         }
 551                         collectionPaths = append(collectionPaths, src)
 552
 553                 case mnt.Kind == "tmp":
 554                         var tmpdir string
 555                         tmpdir, err = runner.MkTempDir(runner.parentTemp, "tmp")
 556                         if err != nil {
 557                                 return nil, fmt.Errorf("while creating mount temp dir: %v", err)
 558                         }
 559                         st, staterr := os.Stat(tmpdir)
 560                         if staterr != nil {
 561                                 return nil, fmt.Errorf("while Stat on temp dir: %v", staterr)
 562                         }
 563                         err = os.Chmod(tmpdir, st.Mode()|os.ModeSetgid|0777)
 564                         if staterr != nil {
 565                                 return nil, fmt.Errorf("while Chmod temp dir: %v", err)
 566                         }
 567                         bindmounts[bind] = bindmount{HostPath: tmpdir}
 568                         if bind == runner.Container.OutputPath {
 569                                 runner.HostOutputDir = tmpdir
 570                         }
 571
 572                 case mnt.Kind == "json" || mnt.Kind == "text":
 573                         var filedata []byte
 574                         if mnt.Kind == "json" {
 575                                 filedata, err = json.Marshal(mnt.Content)
 576                                 if err != nil {
 577                                         return nil, fmt.Errorf("encoding json data: %v", err)
 578                                 }
 579                         } else {
 580                                 text, ok := mnt.Content.(string)
 581                                 if !ok {
 582                                         return nil, fmt.Errorf("content for mount %q must be a string", bind)
 583                                 }
 584                                 filedata = []byte(text)
 585                         }
 586
 587                         tmpdir, err := runner.MkTempDir(runner.parentTemp, mnt.Kind)
 588                         if err != nil {
 589                                 return nil, fmt.Errorf("creating temp dir: %v", err)
 590                         }
 591                         tmpfn := filepath.Join(tmpdir, "mountdata."+mnt.Kind)
 592                         err = ioutil.WriteFile(tmpfn, filedata, 0444)
 593                         if err != nil {
 594                                 return nil, fmt.Errorf("writing temp file: %v", err)
 595                         }
 596                         if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && (notSecret || runner.Container.Mounts[runner.Container.OutputPath].Kind != "collection") {
 597                                 // In most cases, if the container
 598                                 // specifies a literal file inside the
 599                                 // output path, we copy it into the
 600                                 // output directory (either a mounted
 601                                 // collection or a staging area on the
 602                                 // host fs). If it's a secret, it will
 603                                 // be skipped when copying output from
 604                                 // staging to Keep later.
 605                                 copyFiles = append(copyFiles, copyFile{tmpfn, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
 606                         } else {
 607                                 // If a secret is outside OutputPath,
 608                                 // we bind mount the secret file
 609                                 // directly just like other mounts. We
 610                                 // also use this strategy when a
 611                                 // secret is inside OutputPath but
 612                                 // OutputPath is a live collection, to
 613                                 // avoid writing the secret to
 614                                 // Keep. Attempting to remove a
 615                                 // bind-mounted secret file from
 616                                 // inside the container will return a
 617                                 // "Device or resource busy" error
 618                                 // that might not be handled well by
 619                                 // the container, which is why we
 620                                 // don't use this strategy when
 621                                 // OutputPath is a staging directory.
 622                                 bindmounts[bind] = bindmount{HostPath: tmpfn, ReadOnly: true}
 623                         }
 624                 }
 625         }
 626
 627         if runner.HostOutputDir == "" {
 628                 return nil, fmt.Errorf("output path does not correspond to a writable mount point")
 629         }
 630
 631         if needCertMount && runner.Container.RuntimeConstraints.API {
 632                 for _, certfile := range []string{
 633                         // Populated by caller, or sdk/go/arvados init(), or test suite:
 634                         os.Getenv("SSL_CERT_FILE"),
 635                         // Copied from Go 1.21 stdlib (src/crypto/x509/root_linux.go):
 636                         "/etc/ssl/certs/ca-certificates.crt",                // Debian/Ubuntu/Gentoo etc.
 637                         "/etc/pki/tls/certs/ca-bundle.crt",                  // Fedora/RHEL 6
 638                         "/etc/ssl/ca-bundle.pem",                            // OpenSUSE
 639                         "/etc/pki/tls/cacert.pem",                           // OpenELEC
 640                         "/etc/pki/ca-trust/extracted/pem/tls-ca-bundle.pem", // CentOS/RHEL 7
 641                         "/etc/ssl/cert.pem",                                 // Alpine Linux
 642                 } {
 643                         if _, err := os.Stat(certfile); err == nil {
 644                                 bindmounts[arvadosCertPath] = bindmount{HostPath: certfile, ReadOnly: true}
 645                                 break
 646                         }
 647                 }
 648         }
 649
 650         if pdhOnly {
 651                 // If we are only mounting collections by pdh, make
 652                 // sure we don't subscribe to websocket events to
 653                 // avoid putting undesired load on the API server
 654                 arvMountCmd = append(arvMountCmd, "--mount-by-pdh", "by_id", "--disable-event-listening")
 655         } else {
 656                 arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_id")
 657         }
 658         // the by_uuid mount point is used by singularity when writing
 659         // out docker images converted to SIF
 660         arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_uuid")
 661         arvMountCmd = append(arvMountCmd, runner.ArvMountPoint)
 662
 663         runner.ArvMount, err = runner.RunArvMount(arvMountCmd, token)
 664         if err != nil {
 665                 return nil, fmt.Errorf("while trying to start arv-mount: %v", err)
 666         }
 667         if runner.hoststatReporter != nil && runner.ArvMount != nil {
 668                 runner.hoststatReporter.ReportPID("arv-mount", runner.ArvMount.Process.Pid)
 669         }
 670
 671         for _, p := range collectionPaths {
 672                 _, err = os.Stat(p)
 673                 if err != nil {
 674                         return nil, fmt.Errorf("while checking that input files exist: %v", err)
 675                 }
 676         }
 677
 678         for _, cp := range copyFiles {
 679                 st, err := os.Stat(cp.src)
 680                 if err != nil {
 681                         return nil, fmt.Errorf("while staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 682                 }
 683                 if st.IsDir() {
 684                         err = filepath.Walk(cp.src, func(walkpath string, walkinfo os.FileInfo, walkerr error) error {
 685                                 if walkerr != nil {
 686                                         return walkerr
 687                                 }
 688                                 target := path.Join(cp.bind, walkpath[len(cp.src):])
 689                                 if walkinfo.Mode().IsRegular() {
 690                                         copyerr := copyfile(walkpath, target)
 691                                         if copyerr != nil {
 692                                                 return copyerr
 693                                         }
 694                                         return os.Chmod(target, walkinfo.Mode()|0777)
 695                                 } else if walkinfo.Mode().IsDir() {
 696                                         mkerr := os.MkdirAll(target, 0777)
 697                                         if mkerr != nil {
 698                                                 return mkerr
 699                                         }
 700                                         return os.Chmod(target, walkinfo.Mode()|os.ModeSetgid|0777)
 701                                 } else {
 702                                         return fmt.Errorf("source %q is not a regular file or directory", cp.src)
 703                                 }
 704                         })
 705                 } else if st.Mode().IsRegular() {
 706                         err = copyfile(cp.src, cp.bind)
 707                         if err == nil {
 708                                 err = os.Chmod(cp.bind, st.Mode()|0777)
 709                         }
 710                 }
 711                 if err != nil {
 712                         return nil, fmt.Errorf("while staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 713                 }
 714         }
 715
 716         return bindmounts, nil
 717 }
 718
 719 func (runner *ContainerRunner) stopHoststat() error {
 720         if runner.hoststatReporter == nil {
 721                 return nil
 722         }
 723         runner.hoststatReporter.Stop()
 724         runner.hoststatReporter.LogProcessMemMax(runner.CrunchLog)
 725         err := runner.hoststatLogger.Close()
 726         if err != nil {
 727                 return fmt.Errorf("error closing hoststat logs: %v", err)
 728         }
 729         return nil
 730 }
 731
 732 func (runner *ContainerRunner) startHoststat() error {
 733         var err error
 734         runner.hoststatLogger, err = runner.openLogFile("hoststat")
 735         if err != nil {
 736                 return err
 737         }
 738         runner.hoststatReporter = &crunchstat.Reporter{
 739                 Logger: newLogWriter(newTimestamper(runner.hoststatLogger)),
 740                 // Our own cgroup is the "host" cgroup, in the sense
 741                 // that it accounts for resource usage outside the
 742                 // container. It doesn't count _all_ resource usage on
 743                 // the system.
 744                 //
 745                 // TODO?: Use the furthest ancestor of our own cgroup
 746                 // that has stats available. (Currently crunchstat
 747                 // does not have that capability.)
 748                 Pid:        os.Getpid,
 749                 PollPeriod: runner.statInterval,
 750         }
 751         runner.hoststatReporter.Start()
 752         runner.hoststatReporter.ReportPID("crunch-run", os.Getpid())
 753         return nil
 754 }
 755
 756 func (runner *ContainerRunner) startCrunchstat() error {
 757         var err error
 758         runner.statLogger, err = runner.openLogFile("crunchstat")
 759         if err != nil {
 760                 return err
 761         }
 762         runner.statReporter = &crunchstat.Reporter{
 763                 Pid:    runner.executor.Pid,
 764                 FS:     runner.crunchstatFakeFS,
 765                 Logger: newLogWriter(newTimestamper(runner.statLogger)),
 766                 MemThresholds: map[string][]crunchstat.Threshold{
 767                         "rss": crunchstat.NewThresholdsFromPercentages(runner.Container.RuntimeConstraints.RAM, []int64{90, 95, 99}),
 768                 },
 769                 PollPeriod:      runner.statInterval,
 770                 TempDir:         runner.parentTemp,
 771                 ThresholdLogger: runner.CrunchLog,
 772         }
 773         runner.statReporter.Start()
 774         return nil
 775 }
 776
 777 type infoCommand struct {
 778         label string
 779         cmd   []string
 780 }
 781
 782 // LogHostInfo logs info about the current host, for debugging and
 783 // accounting purposes. Although it's logged as "node-info", this is
 784 // about the environment where crunch-run is actually running, which
 785 // might differ from what's described in the node record (see
 786 // LogNodeRecord).
 787 func (runner *ContainerRunner) LogHostInfo() (err error) {
 788         w, err := runner.openLogFile("node-info")
 789         if err != nil {
 790                 return
 791         }
 792
 793         commands := []infoCommand{
 794                 {
 795                         label: "Host Information",
 796                         cmd:   []string{"uname", "-a"},
 797                 },
 798                 {
 799                         label: "CPU Information",
 800                         cmd:   []string{"cat", "/proc/cpuinfo"},
 801                 },
 802                 {
 803                         label: "Memory Information",
 804                         cmd:   []string{"cat", "/proc/meminfo"},
 805                 },
 806                 {
 807                         label: "Disk Space",
 808                         cmd:   []string{"df", "-m", "/", os.TempDir()},
 809                 },
 810                 {
 811                         label: "Disk INodes",
 812                         cmd:   []string{"df", "-i", "/", os.TempDir()},
 813                 },
 814         }
 815
 816         // Run commands with informational output to be logged.
 817         for _, command := range commands {
 818                 fmt.Fprintln(w, command.label)
 819                 cmd := exec.Command(command.cmd[0], command.cmd[1:]...)
 820                 cmd.Stdout = w
 821                 cmd.Stderr = w
 822                 if err := cmd.Run(); err != nil {
 823                         err = fmt.Errorf("While running command %q: %v", command.cmd, err)
 824                         fmt.Fprintln(w, err)
 825                         return err
 826                 }
 827                 fmt.Fprintln(w, "")
 828         }
 829
 830         err = w.Close()
 831         if err != nil {
 832                 return fmt.Errorf("While closing node-info logs: %v", err)
 833         }
 834         return nil
 835 }
 836
 837 // LogContainerRecord gets and saves the raw JSON container record from the API server
 838 func (runner *ContainerRunner) LogContainerRecord() error {
 839         logged, err := runner.logAPIResponse("container", "containers", map[string]interface{}{"filters": [][]string{{"uuid", "=", runner.Container.UUID}}})
 840         if !logged && err == nil {
 841                 err = fmt.Errorf("error: no container record found for %s", runner.Container.UUID)
 842         }
 843         return err
 844 }
 845
 846 // LogNodeRecord logs the current host's InstanceType config entry, if
 847 // running via arvados-dispatch-cloud.
 848 func (runner *ContainerRunner) LogNodeRecord() error {
 849         it := os.Getenv("InstanceType")
 850         if it == "" {
 851                 // Not dispatched by arvados-dispatch-cloud.
 852                 return nil
 853         }
 854         // Save InstanceType config fragment received from dispatcher
 855         // on stdin.
 856         w, err := runner.LogCollection.OpenFile("node.json", os.O_CREATE|os.O_WRONLY, 0666)
 857         if err != nil {
 858                 return err
 859         }
 860         defer w.Close()
 861         _, err = io.WriteString(w, it)
 862         if err != nil {
 863                 return err
 864         }
 865         return w.Close()
 866 }
 867
 868 func (runner *ContainerRunner) logAPIResponse(label, path string, params map[string]interface{}) (logged bool, err error) {
 869         writer, err := runner.LogCollection.OpenFile(label+".json", os.O_CREATE|os.O_WRONLY, 0666)
 870         if err != nil {
 871                 return false, err
 872         }
 873         reader, err := runner.DispatcherArvClient.CallRaw("GET", path, "", "", arvadosclient.Dict(params))
 874         if err != nil {
 875                 return false, fmt.Errorf("error getting %s record: %v", label, err)
 876         }
 877         defer reader.Close()
 878
 879         dec := json.NewDecoder(reader)
 880         dec.UseNumber()
 881         var resp map[string]interface{}
 882         if err = dec.Decode(&resp); err != nil {
 883                 return false, fmt.Errorf("error decoding %s list response: %v", label, err)
 884         }
 885         items, ok := resp["items"].([]interface{})
 886         if !ok {
 887                 return false, fmt.Errorf("error decoding %s list response: no \"items\" key in API list response", label)
 888         } else if len(items) < 1 {
 889                 return false, nil
 890         }
 891         // Re-encode it using indentation to improve readability
 892         enc := json.NewEncoder(writer)
 893         enc.SetIndent("", "    ")
 894         if err = enc.Encode(items[0]); err != nil {
 895                 return false, fmt.Errorf("error logging %s record: %v", label, err)
 896         }
 897         err = writer.Close()
 898         if err != nil {
 899                 return false, fmt.Errorf("error closing %s.json in log collection: %v", label, err)
 900         }
 901         return true, nil
 902 }
 903
 904 func (runner *ContainerRunner) getStdoutFile(mntPath string) (*os.File, error) {
 905         stdoutPath := mntPath[len(runner.Container.OutputPath):]
 906         index := strings.LastIndex(stdoutPath, "/")
 907         if index > 0 {
 908                 subdirs := stdoutPath[:index]
 909                 if subdirs != "" {
 910                         st, err := os.Stat(runner.HostOutputDir)
 911                         if err != nil {
 912                                 return nil, fmt.Errorf("While Stat on temp dir: %v", err)
 913                         }
 914                         stdoutPath := filepath.Join(runner.HostOutputDir, subdirs)
 915                         err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777)
 916                         if err != nil {
 917                                 return nil, fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err)
 918                         }
 919                 }
 920         }
 921         stdoutFile, err := os.Create(filepath.Join(runner.HostOutputDir, stdoutPath))
 922         if err != nil {
 923                 return nil, fmt.Errorf("While creating file %q: %v", stdoutPath, err)
 924         }
 925
 926         return stdoutFile, nil
 927 }
 928
 929 // CreateContainer creates the docker container.
 930 func (runner *ContainerRunner) CreateContainer(imageID string, bindmounts map[string]bindmount) error {
 931         var stdin io.Reader
 932         if mnt, ok := runner.Container.Mounts["stdin"]; ok {
 933                 switch mnt.Kind {
 934                 case "collection":
 935                         var collID string
 936                         if mnt.UUID != "" {
 937                                 collID = mnt.UUID
 938                         } else {
 939                                 collID = mnt.PortableDataHash
 940                         }
 941                         path := runner.ArvMountPoint + "/by_id/" + collID + "/" + mnt.Path
 942                         f, err := os.Open(path)
 943                         if err != nil {
 944                                 return err
 945                         }
 946                         stdin = f
 947                         runner.executorStdin = f
 948                 case "json":
 949                         j, err := json.Marshal(mnt.Content)
 950                         if err != nil {
 951                                 return fmt.Errorf("error encoding stdin json data: %v", err)
 952                         }
 953                         stdin = bytes.NewReader(j)
 954                         runner.executorStdin = io.NopCloser(nil)
 955                 default:
 956                         return fmt.Errorf("stdin mount has unsupported kind %q", mnt.Kind)
 957                 }
 958         } else {
 959                 stdin = bytes.NewReader(nil)
 960                 runner.executorStdin = ioutil.NopCloser(nil)
 961         }
 962
 963         var stdout, stderr io.Writer
 964         if mnt, ok := runner.Container.Mounts["stdout"]; ok {
 965                 f, err := runner.getStdoutFile(mnt.Path)
 966                 if err != nil {
 967                         return err
 968                 }
 969                 stdout = f
 970                 runner.executorStdout = f
 971         } else if w, err := runner.openLogFile("stdout"); err != nil {
 972                 return err
 973         } else {
 974                 stdout = newTimestamper(w)
 975                 runner.executorStdout = w
 976         }
 977
 978         if mnt, ok := runner.Container.Mounts["stderr"]; ok {
 979                 f, err := runner.getStdoutFile(mnt.Path)
 980                 if err != nil {
 981                         return err
 982                 }
 983                 stderr = f
 984                 runner.executorStderr = f
 985         } else if w, err := runner.openLogFile("stderr"); err != nil {
 986                 return err
 987         } else {
 988                 stderr = newTimestamper(w)
 989                 runner.executorStderr = w
 990         }
 991
 992         env := runner.Container.Environment
 993         enableNetwork := runner.enableNetwork == "always"
 994         if runner.Container.RuntimeConstraints.API {
 995                 enableNetwork = true
 996                 tok, err := runner.ContainerToken()
 997                 if err != nil {
 998                         return err
 999                 }
1000                 env = map[string]string{}
1001                 for k, v := range runner.Container.Environment {
1002                         env[k] = v
1003                 }
1004                 env["ARVADOS_API_TOKEN"] = tok
1005                 env["ARVADOS_API_HOST"] = os.Getenv("ARVADOS_API_HOST")
1006                 env["ARVADOS_API_HOST_INSECURE"] = os.Getenv("ARVADOS_API_HOST_INSECURE")
1007                 env["ARVADOS_KEEP_SERVICES"] = os.Getenv("ARVADOS_KEEP_SERVICES")
1008         }
1009         workdir := runner.Container.Cwd
1010         if workdir == "." {
1011                 // both "" and "." mean default
1012                 workdir = ""
1013         }
1014         ram := runner.Container.RuntimeConstraints.RAM
1015         if !runner.enableMemoryLimit {
1016                 ram = 0
1017         }
1018
1019         if runner.Container.RuntimeConstraints.GPU.Stack == "cuda" {
1020                 nvidiaModprobe(runner.CrunchLog)
1021         }
1022
1023         return runner.executor.Create(containerSpec{
1024                 Image:          imageID,
1025                 VCPUs:          runner.Container.RuntimeConstraints.VCPUs,
1026                 RAM:            ram,
1027                 WorkingDir:     workdir,
1028                 Env:            env,
1029                 BindMounts:     bindmounts,
1030                 Command:        runner.Container.Command,
1031                 EnableNetwork:  enableNetwork,
1032                 GPUStack:       runner.Container.RuntimeConstraints.GPU.Stack,
1033                 GPUDeviceCount: runner.Container.RuntimeConstraints.GPU.DeviceCount,
1034                 NetworkMode:    runner.networkMode,
1035                 CgroupParent:   runner.setCgroupParent,
1036                 Stdin:          stdin,
1037                 Stdout:         stdout,
1038                 Stderr:         stderr,
1039         })
1040 }
1041
1042 // StartContainer starts the docker container created by CreateContainer.
1043 func (runner *ContainerRunner) StartContainer() error {
1044         runner.CrunchLog.Printf("Starting container")
1045         runner.cStateLock.Lock()
1046         defer runner.cStateLock.Unlock()
1047         if runner.cCancelled {
1048                 return ErrCancelled
1049         }
1050         err := runner.executor.Start()
1051         if err != nil {
1052                 var advice string
1053                 if m, e := regexp.MatchString("(?ms).*(exec|System error).*(no such file or directory|file not found).*", err.Error()); m && e == nil {
1054                         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])
1055                 }
1056                 return fmt.Errorf("could not start container: %v%s", err, advice)
1057         }
1058         return nil
1059 }
1060
1061 // WaitFinish waits for the container to terminate, capture the exit code, and
1062 // close the stdout/stderr logging.
1063 func (runner *ContainerRunner) WaitFinish() error {
1064         runner.CrunchLog.Print("Waiting for container to finish")
1065         var timeout <-chan time.Time
1066         if s := runner.Container.SchedulingParameters.MaxRunTime; s > 0 {
1067                 timeout = time.After(time.Duration(s) * time.Second)
1068         }
1069         ctx, cancel := context.WithCancel(context.Background())
1070         defer cancel()
1071         go func() {
1072                 select {
1073                 case <-timeout:
1074                         runner.CrunchLog.Printf("maximum run time exceeded. Stopping container.")
1075                         runner.stop(nil)
1076                 case <-runner.ArvMountExit:
1077                         runner.CrunchLog.Printf("arv-mount exited while container is still running. Stopping container.")
1078                         runner.stop(nil)
1079                 case <-ctx.Done():
1080                 }
1081         }()
1082         exitcode, err := runner.executor.Wait(ctx)
1083         if err != nil {
1084                 runner.checkBrokenNode(err)
1085                 return err
1086         }
1087         runner.ExitCode = &exitcode
1088
1089         extra := ""
1090         if exitcode&0x80 != 0 {
1091                 // Convert raw exit status (0x80 + signal number) to a
1092                 // string to log after the code, like " (signal 101)"
1093                 // or " (signal 9, killed)"
1094                 sig := syscall.WaitStatus(exitcode).Signal()
1095                 if name := unix.SignalName(sig); name != "" {
1096                         extra = fmt.Sprintf(" (signal %d, %s)", sig, name)
1097                 } else {
1098                         extra = fmt.Sprintf(" (signal %d)", sig)
1099                 }
1100         }
1101         runner.CrunchLog.Printf("Container exited with status code %d%s", exitcode, extra)
1102         err = runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1103                 "select":    []string{"uuid"},
1104                 "container": arvadosclient.Dict{"exit_code": exitcode},
1105         }, nil)
1106         if err != nil {
1107                 runner.CrunchLog.Printf("ignoring error updating exit_code: %s", err)
1108         }
1109
1110         var returnErr error
1111         if err = runner.executorStdin.Close(); err != nil {
1112                 err = fmt.Errorf("error closing container stdin: %s", err)
1113                 runner.CrunchLog.Printf("%s", err)
1114                 returnErr = err
1115         }
1116         if err = runner.executorStdout.Close(); err != nil {
1117                 err = fmt.Errorf("error closing container stdout: %s", err)
1118                 runner.CrunchLog.Printf("%s", err)
1119                 if returnErr == nil {
1120                         returnErr = err
1121                 }
1122         }
1123         if err = runner.executorStderr.Close(); err != nil {
1124                 err = fmt.Errorf("error closing container stderr: %s", err)
1125                 runner.CrunchLog.Printf("%s", err)
1126                 if returnErr == nil {
1127                         returnErr = err
1128                 }
1129         }
1130
1131         if runner.statReporter != nil {
1132                 runner.statReporter.Stop()
1133                 runner.statReporter.LogMaxima(runner.CrunchLog, map[string]int64{
1134                         "rss": runner.Container.RuntimeConstraints.RAM,
1135                 })
1136                 err = runner.statLogger.Close()
1137                 if err != nil {
1138                         runner.CrunchLog.Printf("error closing crunchstat logs: %v", err)
1139                 }
1140         }
1141         return returnErr
1142 }
1143
1144 func (runner *ContainerRunner) updateLogs() {
1145         ticker := time.NewTicker(crunchLogUpdatePeriod / 360)
1146         defer ticker.Stop()
1147
1148         sigusr1 := make(chan os.Signal, 1)
1149         signal.Notify(sigusr1, syscall.SIGUSR1)
1150         defer signal.Stop(sigusr1)
1151
1152         saveAtTime := time.Now().Add(crunchLogUpdatePeriod)
1153         saveAtSize := crunchLogUpdateSize
1154         var savedSize int64
1155         for {
1156                 select {
1157                 case <-ticker.C:
1158                 case <-sigusr1:
1159                         saveAtTime = time.Now()
1160                 }
1161                 runner.logMtx.Lock()
1162                 done := runner.LogsPDH != nil
1163                 runner.logMtx.Unlock()
1164                 if done {
1165                         return
1166                 }
1167                 size := runner.LogCollection.Size()
1168                 if size == savedSize || (time.Now().Before(saveAtTime) && size < saveAtSize) {
1169                         continue
1170                 }
1171                 saveAtTime = time.Now().Add(crunchLogUpdatePeriod)
1172                 saveAtSize = runner.LogCollection.Size() + crunchLogUpdateSize
1173                 saved, err := runner.saveLogCollection(false)
1174                 if err != nil {
1175                         runner.CrunchLog.Printf("error updating log collection: %s", err)
1176                         continue
1177                 }
1178
1179                 err = runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1180                         "select": []string{"uuid"},
1181                         "container": arvadosclient.Dict{
1182                                 "log": saved.PortableDataHash,
1183                         },
1184                 }, nil)
1185                 if err != nil {
1186                         runner.CrunchLog.Printf("error updating container log to %s: %s", saved.PortableDataHash, err)
1187                         continue
1188                 }
1189
1190                 savedSize = size
1191         }
1192 }
1193
1194 var spotInterruptionCheckInterval = 5 * time.Second
1195 var ec2MetadataBaseURL = "http://169.254.169.254"
1196
1197 const ec2TokenTTL = time.Second * 21600
1198
1199 func (runner *ContainerRunner) checkSpotInterruptionNotices() {
1200         type ec2metadata struct {
1201                 Action string    `json:"action"`
1202                 Time   time.Time `json:"time"`
1203         }
1204         runner.CrunchLog.Printf("Checking for spot instance interruptions every %v using instance metadata at %s", spotInterruptionCheckInterval, ec2MetadataBaseURL)
1205         var metadata ec2metadata
1206         var token string
1207         var tokenExp time.Time
1208         check := func() error {
1209                 ctx, cancel := context.WithDeadline(context.Background(), time.Now().Add(time.Minute))
1210                 defer cancel()
1211                 if token == "" || tokenExp.Sub(time.Now()) < time.Minute {
1212                         req, err := http.NewRequestWithContext(ctx, http.MethodPut, ec2MetadataBaseURL+"/latest/api/token", nil)
1213                         if err != nil {
1214                                 return err
1215                         }
1216                         req.Header.Set("X-aws-ec2-metadata-token-ttl-seconds", fmt.Sprintf("%d", int(ec2TokenTTL/time.Second)))
1217                         resp, err := http.DefaultClient.Do(req)
1218                         if err != nil {
1219                                 return err
1220                         }
1221                         defer resp.Body.Close()
1222                         if resp.StatusCode != http.StatusOK {
1223                                 return fmt.Errorf("%s", resp.Status)
1224                         }
1225                         newtoken, err := ioutil.ReadAll(resp.Body)
1226                         if err != nil {
1227                                 return err
1228                         }
1229                         token = strings.TrimSpace(string(newtoken))
1230                         tokenExp = time.Now().Add(ec2TokenTTL)
1231                 }
1232                 req, err := http.NewRequestWithContext(ctx, http.MethodGet, ec2MetadataBaseURL+"/latest/meta-data/spot/instance-action", nil)
1233                 if err != nil {
1234                         return err
1235                 }
1236                 req.Header.Set("X-aws-ec2-metadata-token", token)
1237                 resp, err := http.DefaultClient.Do(req)
1238                 if err != nil {
1239                         return err
1240                 }
1241                 defer resp.Body.Close()
1242                 switch resp.StatusCode {
1243                 case http.StatusOK:
1244                         break
1245                 case http.StatusNotFound:
1246                         // "If Amazon EC2 is not preparing to stop or
1247                         // terminate the instance, or if you
1248                         // terminated the instance yourself,
1249                         // instance-action is not present in the
1250                         // instance metadata and you receive an HTTP
1251                         // 404 error when you try to retrieve it."
1252                         metadata = ec2metadata{}
1253                         return nil
1254                 case http.StatusUnauthorized:
1255                         token = ""
1256                         return fmt.Errorf("%s", resp.Status)
1257                 default:
1258                         return fmt.Errorf("%s", resp.Status)
1259                 }
1260                 nextmetadata := ec2metadata{}
1261                 err = json.NewDecoder(resp.Body).Decode(&nextmetadata)
1262                 if err != nil {
1263                         return err
1264                 }
1265                 metadata = nextmetadata
1266                 return nil
1267         }
1268         failures := 0
1269         var lastmetadata ec2metadata
1270         for range time.NewTicker(spotInterruptionCheckInterval).C {
1271                 err := check()
1272                 if err != nil {
1273                         message := fmt.Sprintf("Spot instance interruption check was inconclusive: %s", err)
1274                         if failures++; failures > 5 {
1275                                 runner.CrunchLog.Printf("%s -- now giving up after too many consecutive errors", message)
1276                                 return
1277                         } else {
1278                                 runner.CrunchLog.Printf("%s -- will retry in %v", message, spotInterruptionCheckInterval)
1279                                 continue
1280                         }
1281                 }
1282                 failures = 0
1283                 if metadata.Action != "" && metadata != lastmetadata {
1284                         lastmetadata = metadata
1285                         text := fmt.Sprintf("Cloud provider scheduled instance %s at %s", metadata.Action, metadata.Time.UTC().Format(time.RFC3339))
1286                         runner.CrunchLog.Printf("%s", text)
1287                         runner.updateRuntimeStatus(arvadosclient.Dict{
1288                                 "warning":          "preemption notice",
1289                                 "warningDetail":    text,
1290                                 "preemptionNotice": text,
1291                         })
1292                         if proc, err := os.FindProcess(os.Getpid()); err == nil {
1293                                 // trigger updateLogs
1294                                 proc.Signal(syscall.SIGUSR1)
1295                         }
1296                 }
1297         }
1298 }
1299
1300 func (runner *ContainerRunner) updateRuntimeStatus(status arvadosclient.Dict) {
1301         err := runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1302                 "select": []string{"uuid"},
1303                 "container": arvadosclient.Dict{
1304                         "runtime_status": status,
1305                 },
1306         }, nil)
1307         if err != nil {
1308                 runner.CrunchLog.Printf("error updating container runtime_status: %s", err)
1309         }
1310 }
1311
1312 // CaptureOutput saves data from the container's output directory if
1313 // needed, and updates the container output accordingly.
1314 func (runner *ContainerRunner) CaptureOutput(bindmounts map[string]bindmount) error {
1315         if runner.Container.RuntimeConstraints.API {
1316                 // Output may have been set directly by the container, so
1317                 // refresh the container record to check.
1318                 err := runner.DispatcherArvClient.Get("containers", runner.Container.UUID,
1319                         arvadosclient.Dict{
1320                                 "select": []string{"output"},
1321                         }, &runner.Container)
1322                 if err != nil {
1323                         return err
1324                 }
1325                 if runner.Container.Output != "" {
1326                         // Container output is already set.
1327                         runner.OutputPDH = &runner.Container.Output
1328                         return nil
1329                 }
1330         }
1331
1332         txt, err := (&copier{
1333                 client:        runner.containerClient,
1334                 keepClient:    runner.ContainerKeepClient,
1335                 hostOutputDir: runner.HostOutputDir,
1336                 ctrOutputDir:  runner.Container.OutputPath,
1337                 globs:         runner.Container.OutputGlob,
1338                 bindmounts:    bindmounts,
1339                 mounts:        runner.Container.Mounts,
1340                 secretMounts:  runner.SecretMounts,
1341                 logger:        runner.CrunchLog,
1342         }).Copy()
1343         if err != nil {
1344                 return err
1345         }
1346         if n := len(regexp.MustCompile(` [0-9a-f]+\+\S*\+R`).FindAllStringIndex(txt, -1)); n > 0 {
1347                 runner.CrunchLog.Printf("Copying %d data blocks from remote input collections...", n)
1348                 fs, err := (&arvados.Collection{ManifestText: txt}).FileSystem(runner.containerClient, runner.ContainerKeepClient)
1349                 if err != nil {
1350                         return err
1351                 }
1352                 txt, err = fs.MarshalManifest(".")
1353                 if err != nil {
1354                         return err
1355                 }
1356         }
1357         var resp arvados.Collection
1358         err = runner.ContainerArvClient.Create("collections", arvadosclient.Dict{
1359                 "ensure_unique_name": true,
1360                 "select":             []string{"portable_data_hash"},
1361                 "collection": arvadosclient.Dict{
1362                         "is_trashed":    true,
1363                         "name":          "output for " + runner.Container.UUID,
1364                         "manifest_text": txt,
1365                 },
1366         }, &resp)
1367         if err != nil {
1368                 return fmt.Errorf("error creating output collection: %v", err)
1369         }
1370         runner.OutputPDH = &resp.PortableDataHash
1371         return nil
1372 }
1373
1374 func (runner *ContainerRunner) CleanupDirs() {
1375         if runner.ArvMount != nil {
1376                 var delay int64 = 8
1377                 umount := exec.Command("arv-mount", fmt.Sprintf("--unmount-timeout=%d", delay), "--unmount", runner.ArvMountPoint)
1378                 umount.Stdout = runner.CrunchLog
1379                 umount.Stderr = runner.CrunchLog
1380                 runner.CrunchLog.Printf("Running %v", umount.Args)
1381                 umnterr := umount.Start()
1382
1383                 if umnterr != nil {
1384                         runner.CrunchLog.Printf("Error unmounting: %v", umnterr)
1385                         runner.ArvMount.Process.Kill()
1386                 } else {
1387                         // If arv-mount --unmount gets stuck for any reason, we
1388                         // don't want to wait for it forever.  Do Wait() in a goroutine
1389                         // so it doesn't block crunch-run.
1390                         umountExit := make(chan error)
1391                         go func() {
1392                                 mnterr := umount.Wait()
1393                                 if mnterr != nil {
1394                                         runner.CrunchLog.Printf("Error unmounting: %v", mnterr)
1395                                 }
1396                                 umountExit <- mnterr
1397                         }()
1398
1399                         for again := true; again; {
1400                                 again = false
1401                                 select {
1402                                 case <-umountExit:
1403                                         umount = nil
1404                                         again = true
1405                                 case <-runner.ArvMountExit:
1406                                         break
1407                                 case <-time.After(time.Duration((delay + 1) * int64(time.Second))):
1408                                         runner.CrunchLog.Printf("Timed out waiting for unmount")
1409                                         if umount != nil {
1410                                                 umount.Process.Kill()
1411                                         }
1412                                         runner.ArvMount.Process.Kill()
1413                                 }
1414                         }
1415                 }
1416                 runner.ArvMount = nil
1417         }
1418
1419         if runner.ArvMountPoint != "" {
1420                 if rmerr := os.Remove(runner.ArvMountPoint); rmerr != nil {
1421                         runner.CrunchLog.Printf("While cleaning up arv-mount directory %s: %v", runner.ArvMountPoint, rmerr)
1422                 }
1423                 runner.ArvMountPoint = ""
1424         }
1425
1426         if rmerr := os.RemoveAll(runner.parentTemp); rmerr != nil {
1427                 runner.CrunchLog.Printf("While cleaning up temporary directory %s: %v", runner.parentTemp, rmerr)
1428         }
1429 }
1430
1431 // CommitLogs posts the collection containing the final container logs.
1432 func (runner *ContainerRunner) CommitLogs() error {
1433         func() {
1434                 // Hold cStateLock to prevent races on CrunchLog (e.g., stop()).
1435                 runner.cStateLock.Lock()
1436                 defer runner.cStateLock.Unlock()
1437
1438                 runner.CrunchLog.Print(runner.finalState)
1439
1440                 if runner.arvMountLog != nil {
1441                         runner.arvMountLog.Close()
1442                 }
1443
1444                 // From now on just log to stderr, in case there are
1445                 // any other further errors (such as failing to write
1446                 // the log to Keep!)  while shutting down
1447                 runner.CrunchLog = newLogWriter(newTimestamper(newStringPrefixer(os.Stderr, runner.Container.UUID+" ")))
1448         }()
1449
1450         if runner.keepstoreLogger != nil {
1451                 // Flush any buffered logs from our local keepstore
1452                 // process.  Discard anything logged after this point
1453                 // -- it won't end up in the log collection, so
1454                 // there's no point writing it to the collectionfs.
1455                 runner.keepstoreLogbuf.SetWriter(io.Discard)
1456                 runner.keepstoreLogger.Close()
1457                 runner.keepstoreLogger = nil
1458         }
1459
1460         if runner.LogsPDH != nil {
1461                 // If we have already assigned something to LogsPDH,
1462                 // we must be closing the re-opened log, which won't
1463                 // end up getting attached to the container record and
1464                 // therefore doesn't need to be saved as a collection
1465                 // -- it exists only to send logs to other channels.
1466                 return nil
1467         }
1468
1469         saved, err := runner.saveLogCollection(true)
1470         if err != nil {
1471                 return fmt.Errorf("error saving log collection: %s", err)
1472         }
1473         runner.logMtx.Lock()
1474         defer runner.logMtx.Unlock()
1475         runner.LogsPDH = &saved.PortableDataHash
1476         return nil
1477 }
1478
1479 // Create/update the log collection. Return value has UUID and
1480 // PortableDataHash fields populated, but others may be blank.
1481 func (runner *ContainerRunner) saveLogCollection(final bool) (response arvados.Collection, err error) {
1482         runner.logMtx.Lock()
1483         defer runner.logMtx.Unlock()
1484         if runner.LogsPDH != nil {
1485                 // Already finalized.
1486                 return
1487         }
1488         updates := arvadosclient.Dict{
1489                 "name": "logs for " + runner.Container.UUID,
1490         }
1491         mt, err1 := runner.LogCollection.MarshalManifest(".")
1492         if err1 == nil {
1493                 // Only send updated manifest text if there was no
1494                 // error.
1495                 updates["manifest_text"] = mt
1496         }
1497
1498         // Even if flushing the manifest had an error, we still want
1499         // to update the log record, if possible, to push the trash_at
1500         // and delete_at times into the future.  Details on bug
1501         // #17293.
1502         if final {
1503                 updates["is_trashed"] = true
1504         } else {
1505                 // We set trash_at so this collection gets
1506                 // automatically cleaned up eventually.  It used to be
1507                 // 12 hours but we had a situation where the API
1508                 // server was down over a weekend but the containers
1509                 // kept running such that the log collection got
1510                 // trashed, so now we make it 2 weeks.  refs #20378
1511                 exp := time.Now().Add(time.Duration(24*14) * time.Hour)
1512                 updates["trash_at"] = exp
1513                 updates["delete_at"] = exp
1514         }
1515         reqBody := arvadosclient.Dict{
1516                 "select":     []string{"uuid", "portable_data_hash"},
1517                 "collection": updates,
1518         }
1519         var err2 error
1520         if runner.logUUID == "" {
1521                 reqBody["ensure_unique_name"] = true
1522                 err2 = runner.DispatcherArvClient.Create("collections", reqBody, &response)
1523         } else {
1524                 err2 = runner.DispatcherArvClient.Update("collections", runner.logUUID, reqBody, &response)
1525         }
1526         if err2 == nil {
1527                 runner.logUUID = response.UUID
1528         }
1529
1530         if err1 != nil || err2 != nil {
1531                 err = fmt.Errorf("error recording logs: %q, %q", err1, err2)
1532         }
1533         return
1534 }
1535
1536 // UpdateContainerRunning updates the container state to "Running"
1537 func (runner *ContainerRunner) UpdateContainerRunning(logId string) error {
1538         runner.cStateLock.Lock()
1539         defer runner.cStateLock.Unlock()
1540         if runner.cCancelled {
1541                 return ErrCancelled
1542         }
1543         updates := arvadosclient.Dict{
1544                 "gateway_address": runner.gateway.Address,
1545                 "state":           "Running",
1546         }
1547         if logId != "" {
1548                 updates["log"] = logId
1549         }
1550         return runner.DispatcherArvClient.Update(
1551                 "containers",
1552                 runner.Container.UUID,
1553                 arvadosclient.Dict{
1554                         "select":    []string{"uuid"},
1555                         "container": updates,
1556                 },
1557                 nil,
1558         )
1559 }
1560
1561 // ContainerToken returns the api_token the container (and any
1562 // arv-mount processes) are allowed to use.
1563 func (runner *ContainerRunner) ContainerToken() (string, error) {
1564         if runner.token != "" {
1565                 return runner.token, nil
1566         }
1567
1568         var auth arvados.APIClientAuthorization
1569         err := runner.DispatcherArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth)
1570         if err != nil {
1571                 return "", err
1572         }
1573         runner.token = fmt.Sprintf("v2/%s/%s/%s", auth.UUID, auth.APIToken, runner.Container.UUID)
1574         return runner.token, nil
1575 }
1576
1577 // UpdateContainerFinal updates the container record state on API
1578 // server to "Complete" or "Cancelled"
1579 func (runner *ContainerRunner) UpdateContainerFinal() error {
1580         update := arvadosclient.Dict{}
1581         update["state"] = runner.finalState
1582         if runner.LogsPDH != nil {
1583                 update["log"] = *runner.LogsPDH
1584         }
1585         if runner.ExitCode != nil {
1586                 update["exit_code"] = *runner.ExitCode
1587         } else {
1588                 update["exit_code"] = nil
1589         }
1590         if runner.finalState == "Complete" && runner.OutputPDH != nil {
1591                 update["output"] = *runner.OutputPDH
1592         }
1593         update["cost"] = runner.calculateCost(time.Now())
1594         return runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1595                 "select":    []string{"uuid"},
1596                 "container": update,
1597         }, nil)
1598 }
1599
1600 // IsCancelled returns the value of Cancelled, with goroutine safety.
1601 func (runner *ContainerRunner) IsCancelled() bool {
1602         runner.cStateLock.Lock()
1603         defer runner.cStateLock.Unlock()
1604         return runner.cCancelled
1605 }
1606
1607 func (runner *ContainerRunner) openLogFile(name string) (io.WriteCloser, error) {
1608         return runner.LogCollection.OpenFile(name+".txt", os.O_CREATE|os.O_WRONLY, 0666)
1609 }
1610
1611 // Run the full container lifecycle.
1612 func (runner *ContainerRunner) Run() (err error) {
1613         runner.CrunchLog.Printf("crunch-run %s started", cmd.Version.String())
1614         runner.CrunchLog.Printf("%s", currentUserAndGroups())
1615         v, _ := exec.Command("arv-mount", "--version").CombinedOutput()
1616         runner.CrunchLog.Printf("Using FUSE mount: %s", v)
1617         runner.CrunchLog.Printf("Using container runtime: %s", runner.executor.Runtime())
1618         runner.CrunchLog.Printf("Executing container: %s", runner.Container.UUID)
1619         runner.costStartTime = time.Now()
1620
1621         hostname, hosterr := os.Hostname()
1622         if hosterr != nil {
1623                 runner.CrunchLog.Printf("Error getting hostname '%v'", hosterr)
1624         } else {
1625                 runner.CrunchLog.Printf("Executing on host '%s'", hostname)
1626         }
1627
1628         sigusr2 := make(chan os.Signal, 1)
1629         signal.Notify(sigusr2, syscall.SIGUSR2)
1630         defer signal.Stop(sigusr2)
1631         runner.loadPrices()
1632         go runner.handleSIGUSR2(sigusr2)
1633
1634         runner.finalState = "Queued"
1635
1636         defer func() {
1637                 runner.CleanupDirs()
1638                 runner.CrunchLog.Printf("crunch-run finished")
1639         }()
1640
1641         err = runner.fetchContainerRecord()
1642         if err != nil {
1643                 return
1644         }
1645         if runner.Container.State != "Locked" {
1646                 return fmt.Errorf("dispatch error detected: container %q has state %q", runner.Container.UUID, runner.Container.State)
1647         }
1648
1649         var bindmounts map[string]bindmount
1650         defer func() {
1651                 // checkErr prints e (unless it's nil) and sets err to
1652                 // e (unless err is already non-nil). Thus, if err
1653                 // hasn't already been assigned when Run() returns,
1654                 // this cleanup func will cause Run() to return the
1655                 // first non-nil error that is passed to checkErr().
1656                 checkErr := func(errorIn string, e error) {
1657                         if e == nil {
1658                                 return
1659                         }
1660                         runner.CrunchLog.Printf("error in %s: %v", errorIn, e)
1661                         if err == nil {
1662                                 err = e
1663                         }
1664                         if runner.finalState == "Complete" {
1665                                 // There was an error in the finalization.
1666                                 runner.finalState = "Cancelled"
1667                         }
1668                 }
1669
1670                 // Log the error encountered in Run(), if any
1671                 checkErr("Run", err)
1672
1673                 if runner.finalState == "Queued" {
1674                         runner.UpdateContainerFinal()
1675                         return
1676                 }
1677
1678                 if runner.IsCancelled() {
1679                         runner.finalState = "Cancelled"
1680                         // but don't return yet -- we still want to
1681                         // capture partial output and write logs
1682                 }
1683
1684                 if bindmounts != nil {
1685                         checkErr("CaptureOutput", runner.CaptureOutput(bindmounts))
1686                 }
1687                 checkErr("stopHoststat", runner.stopHoststat())
1688                 checkErr("CommitLogs", runner.CommitLogs())
1689                 runner.CleanupDirs()
1690                 checkErr("UpdateContainerFinal", runner.UpdateContainerFinal())
1691         }()
1692
1693         runner.setupSignals()
1694         err = runner.startHoststat()
1695         if err != nil {
1696                 return
1697         }
1698         if runner.keepstore != nil {
1699                 runner.hoststatReporter.ReportPID("keepstore", runner.keepstore.Process.Pid)
1700         }
1701
1702         // set up FUSE mount and binds
1703         bindmounts, err = runner.SetupMounts()
1704         if err != nil {
1705                 runner.finalState = "Cancelled"
1706                 err = fmt.Errorf("While setting up mounts: %v", err)
1707                 return
1708         }
1709
1710         // check for and/or load image
1711         imageID, err := runner.LoadImage()
1712         if err != nil {
1713                 if !runner.checkBrokenNode(err) {
1714                         // Failed to load image but not due to a "broken node"
1715                         // condition, probably user error.
1716                         runner.finalState = "Cancelled"
1717                 }
1718                 err = fmt.Errorf("failed to load container image: %v", err)
1719                 return
1720         }
1721
1722         err = runner.CreateContainer(imageID, bindmounts)
1723         if err != nil {
1724                 return
1725         }
1726         err = runner.LogHostInfo()
1727         if err != nil {
1728                 return
1729         }
1730         err = runner.LogNodeRecord()
1731         if err != nil {
1732                 return
1733         }
1734         err = runner.LogContainerRecord()
1735         if err != nil {
1736                 return
1737         }
1738
1739         if runner.IsCancelled() {
1740                 return
1741         }
1742
1743         logCollection, err := runner.saveLogCollection(false)
1744         var logId string
1745         if err == nil {
1746                 logId = logCollection.PortableDataHash
1747         } else {
1748                 runner.CrunchLog.Printf("Error committing initial log collection: %v", err)
1749         }
1750         err = runner.UpdateContainerRunning(logId)
1751         if err != nil {
1752                 return
1753         }
1754         runner.finalState = "Cancelled"
1755
1756         err = runner.startCrunchstat()
1757         if err != nil {
1758                 return
1759         }
1760
1761         err = runner.StartContainer()
1762         if err != nil {
1763                 runner.checkBrokenNode(err)
1764                 return
1765         }
1766
1767         err = runner.WaitFinish()
1768         if err == nil && !runner.IsCancelled() {
1769                 runner.finalState = "Complete"
1770         }
1771         return
1772 }
1773
1774 // Fetch the current container record (uuid = runner.Container.UUID)
1775 // into runner.Container.
1776 func (runner *ContainerRunner) fetchContainerRecord() error {
1777         reader, err := runner.DispatcherArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
1778         if err != nil {
1779                 return fmt.Errorf("error fetching container record: %v", err)
1780         }
1781         defer reader.Close()
1782
1783         dec := json.NewDecoder(reader)
1784         dec.UseNumber()
1785         err = dec.Decode(&runner.Container)
1786         if err != nil {
1787                 return fmt.Errorf("error decoding container record: %v", err)
1788         }
1789
1790         var sm struct {
1791                 SecretMounts map[string]arvados.Mount `json:"secret_mounts"`
1792         }
1793
1794         containerToken, err := runner.ContainerToken()
1795         if err != nil {
1796                 return fmt.Errorf("error getting container token: %v", err)
1797         }
1798
1799         runner.ContainerArvClient, runner.ContainerKeepClient,
1800                 runner.containerClient, err = runner.MkArvClient(containerToken)
1801         if err != nil {
1802                 return fmt.Errorf("error creating container API client: %v", err)
1803         }
1804
1805         runner.ContainerKeepClient.SetStorageClasses(runner.Container.OutputStorageClasses)
1806         runner.DispatcherKeepClient.SetStorageClasses(runner.Container.OutputStorageClasses)
1807
1808         err = runner.ContainerArvClient.Call("GET", "containers", runner.Container.UUID, "secret_mounts", nil, &sm)
1809         if err != nil {
1810                 if apierr, ok := err.(arvadosclient.APIServerError); !ok || apierr.HttpStatusCode != 404 {
1811                         return fmt.Errorf("error fetching secret_mounts: %v", err)
1812                 }
1813                 // ok && apierr.HttpStatusCode == 404, which means
1814                 // secret_mounts isn't supported by this API server.
1815         }
1816         runner.SecretMounts = sm.SecretMounts
1817
1818         return nil
1819 }
1820
1821 // NewContainerRunner creates a new container runner.
1822 func NewContainerRunner(dispatcherClient *arvados.Client,
1823         dispatcherArvClient IArvadosClient,
1824         dispatcherKeepClient IKeepClient,
1825         containerUUID string) (*ContainerRunner, error) {
1826
1827         cr := &ContainerRunner{
1828                 dispatcherClient:     dispatcherClient,
1829                 DispatcherArvClient:  dispatcherArvClient,
1830                 DispatcherKeepClient: dispatcherKeepClient,
1831         }
1832         cr.RunArvMount = cr.ArvMountCmd
1833         cr.MkTempDir = ioutil.TempDir
1834         cr.MkArvClient = func(token string) (IArvadosClient, IKeepClient, *arvados.Client, error) {
1835                 cl, err := arvadosclient.MakeArvadosClient()
1836                 if err != nil {
1837                         return nil, nil, nil, err
1838                 }
1839                 cl.Retries = 10
1840                 cl.ApiToken = token
1841                 kc, err := keepclient.MakeKeepClient(cl)
1842                 if err != nil {
1843                         return nil, nil, nil, err
1844                 }
1845                 kc.Retries = 10
1846                 c2 := arvados.NewClientFromEnv()
1847                 c2.AuthToken = token
1848                 return cl, kc, c2, nil
1849         }
1850         var err error
1851         cr.LogCollection, err = (&arvados.Collection{}).FileSystem(cr.dispatcherClient, cr.DispatcherKeepClient)
1852         if err != nil {
1853                 return nil, err
1854         }
1855         cr.Container.UUID = containerUUID
1856         f, err := cr.openLogFile("crunch-run")
1857         if err != nil {
1858                 return nil, err
1859         }
1860         cr.CrunchLog = newLogWriter(newTimestamper(io.MultiWriter(f, newStringPrefixer(os.Stderr, cr.Container.UUID+" "))))
1861
1862         go cr.updateLogs()
1863
1864         return cr, nil
1865 }
1866
1867 func (command) RunCommand(prog string, args []string, stdin io.Reader, stdout, stderr io.Writer) int {
1868         log := log.New(stderr, "", 0)
1869         flags := flag.NewFlagSet(prog, flag.ContinueOnError)
1870         statInterval := flags.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting")
1871         flags.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree (obsolete, ignored)")
1872         flags.String("cgroup-parent", "docker", "name of container's parent cgroup (obsolete, ignored)")
1873         cgroupParentSubsystem := flags.String("cgroup-parent-subsystem", "", "use current cgroup for given `subsystem` as parent cgroup for container (subsystem argument is only relevant for cgroups v1; in cgroups v2 / unified mode, any non-empty value means use current cgroup); if empty, use the docker daemon's default cgroup parent. See https://doc.arvados.org/install/crunch2-slurm/install-dispatch.html#CrunchRunCommand-cgroups")
1874         caCertsPath := flags.String("ca-certs", "", "Path to TLS root certificates")
1875         detach := flags.Bool("detach", false, "Detach from parent process and run in the background")
1876         stdinConfig := flags.Bool("stdin-config", false, "Load config and environment variables from JSON message on stdin")
1877         configFile := flags.String("config", arvados.DefaultConfigFile, "filename of cluster config file to try loading if -stdin-config=false (default is $ARVADOS_CONFIG)")
1878         sleep := flags.Duration("sleep", 0, "Delay before starting (testing use only)")
1879         kill := flags.Int("kill", -1, "Send signal to an existing crunch-run process for given UUID")
1880         list := flags.Bool("list", false, "List UUIDs of existing crunch-run processes (and notify them to use price data passed on stdin)")
1881         enableMemoryLimit := flags.Bool("enable-memory-limit", true, "tell container runtime to limit container's memory usage")
1882         enableNetwork := flags.String("container-enable-networking", "default", "enable networking \"always\" (for all containers) or \"default\" (for containers that request it)")
1883         networkMode := flags.String("container-network-mode", "default", `Docker network mode for container (use any argument valid for docker --net)`)
1884         memprofile := flags.String("memprofile", "", "write memory profile to `file` after running container")
1885         runtimeEngine := flags.String("runtime-engine", "docker", "container runtime: docker or singularity")
1886         brokenNodeHook := flags.String("broken-node-hook", "", "script to run if node is detected to be broken (for example, Docker daemon is not running)")
1887         flags.Duration("check-containerd", 0, "Ignored. Exists for compatibility with older versions.")
1888         version := flags.Bool("version", false, "Write version information to stdout and exit 0.")
1889
1890         ignoreDetachFlag := false
1891         if len(args) > 0 && args[0] == "-no-detach" {
1892                 // This process was invoked by a parent process, which
1893                 // has passed along its own arguments, including
1894                 // -detach, after the leading -no-detach flag.  Strip
1895                 // the leading -no-detach flag (it's not recognized by
1896                 // flags.Parse()) and ignore the -detach flag that
1897                 // comes later.
1898                 args = args[1:]
1899                 ignoreDetachFlag = true
1900         }
1901
1902         if ok, code := cmd.ParseFlags(flags, prog, args, "container-uuid", stderr); !ok {
1903                 return code
1904         } else if *version {
1905                 fmt.Fprintln(stdout, prog, cmd.Version.String())
1906                 return 0
1907         } else if !*list && flags.NArg() != 1 {
1908                 fmt.Fprintf(stderr, "missing required argument: container-uuid (try -help)\n")
1909                 return 2
1910         }
1911
1912         containerUUID := flags.Arg(0)
1913
1914         switch {
1915         case *detach && !ignoreDetachFlag:
1916                 return Detach(containerUUID, prog, args, stdin, stdout, stderr)
1917         case *kill >= 0:
1918                 return KillProcess(containerUUID, syscall.Signal(*kill), stdout, stderr)
1919         case *list:
1920                 return ListProcesses(stdin, stdout, stderr)
1921         }
1922
1923         if len(containerUUID) != 27 {
1924                 log.Printf("usage: %s [options] UUID", prog)
1925                 return 1
1926         }
1927
1928         var keepstoreLogbuf bufThenWrite
1929         var conf ConfigData
1930         if *stdinConfig {
1931                 err := json.NewDecoder(stdin).Decode(&conf)
1932                 if err != nil {
1933                         log.Printf("decode stdin: %s", err)
1934                         return 1
1935                 }
1936                 for k, v := range conf.Env {
1937                         err = os.Setenv(k, v)
1938                         if err != nil {
1939                                 log.Printf("setenv(%q): %s", k, err)
1940                                 return 1
1941                         }
1942                 }
1943                 if conf.Cluster != nil {
1944                         // ClusterID is missing from the JSON
1945                         // representation, but we need it to generate
1946                         // a valid config file for keepstore, so we
1947                         // fill it using the container UUID prefix.
1948                         conf.Cluster.ClusterID = containerUUID[:5]
1949                 }
1950         } else {
1951                 conf = hpcConfData(containerUUID, *configFile, io.MultiWriter(&keepstoreLogbuf, stderr))
1952         }
1953
1954         log.Printf("crunch-run %s started", cmd.Version.String())
1955         time.Sleep(*sleep)
1956
1957         if *caCertsPath != "" {
1958                 os.Setenv("SSL_CERT_FILE", *caCertsPath)
1959         }
1960
1961         keepstore, err := startLocalKeepstore(conf, io.MultiWriter(&keepstoreLogbuf, stderr))
1962         if err != nil {
1963                 log.Print(err)
1964                 return 1
1965         }
1966         if keepstore != nil {
1967                 defer keepstore.Process.Kill()
1968         }
1969
1970         api, err := arvadosclient.MakeArvadosClient()
1971         if err != nil {
1972                 log.Printf("%s: %v", containerUUID, err)
1973                 return 1
1974         }
1975         // arvadosclient now interprets Retries=10 to mean
1976         // Timeout=10m, retrying with exponential backoff + jitter.
1977         api.Retries = 10
1978
1979         kc, err := keepclient.MakeKeepClient(api)
1980         if err != nil {
1981                 log.Printf("%s: %v", containerUUID, err)
1982                 return 1
1983         }
1984         kc.Retries = 10
1985
1986         cr, err := NewContainerRunner(arvados.NewClientFromEnv(), api, kc, containerUUID)
1987         if err != nil {
1988                 log.Print(err)
1989                 return 1
1990         }
1991
1992         cr.keepstore = keepstore
1993         if keepstore == nil {
1994                 // Log explanation (if any) for why we're not running
1995                 // a local keepstore.
1996                 var buf bytes.Buffer
1997                 keepstoreLogbuf.SetWriter(&buf)
1998                 if buf.Len() > 0 {
1999                         cr.CrunchLog.Printf("%s", strings.TrimSpace(buf.String()))
2000                 }
2001         } else if logWhat := conf.Cluster.Containers.LocalKeepLogsToContainerLog; logWhat == "none" {
2002                 cr.CrunchLog.Printf("using local keepstore process (pid %d) at %s", keepstore.Process.Pid, os.Getenv("ARVADOS_KEEP_SERVICES"))
2003                 keepstoreLogbuf.SetWriter(io.Discard)
2004         } else {
2005                 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"))
2006                 cr.keepstoreLogger, err = cr.openLogFile("keepstore")
2007                 if err != nil {
2008                         log.Print(err)
2009                         return 1
2010                 }
2011
2012                 var writer io.WriteCloser = cr.keepstoreLogger
2013                 if logWhat == "errors" {
2014                         writer = &filterKeepstoreErrorsOnly{WriteCloser: writer}
2015                 } else if logWhat != "all" {
2016                         // should have been caught earlier by
2017                         // dispatcher's config loader
2018                         log.Printf("invalid value for Containers.LocalKeepLogsToContainerLog: %q", logWhat)
2019                         return 1
2020                 }
2021                 err = keepstoreLogbuf.SetWriter(writer)
2022                 if err != nil {
2023                         log.Print(err)
2024                         return 1
2025                 }
2026                 cr.keepstoreLogbuf = &keepstoreLogbuf
2027         }
2028
2029         switch *runtimeEngine {
2030         case "docker":
2031                 cr.executor, err = newDockerExecutor(containerUUID, cr.CrunchLog.Printf, cr.containerWatchdogInterval)
2032         case "singularity":
2033                 cr.executor, err = newSingularityExecutor(cr.CrunchLog.Printf)
2034         default:
2035                 cr.CrunchLog.Printf("%s: unsupported RuntimeEngine %q", containerUUID, *runtimeEngine)
2036                 return 1
2037         }
2038         if err != nil {
2039                 cr.CrunchLog.Printf("%s: %v", containerUUID, err)
2040                 cr.checkBrokenNode(err)
2041                 return 1
2042         }
2043         defer cr.executor.Close()
2044
2045         cr.brokenNodeHook = *brokenNodeHook
2046
2047         gwAuthSecret := os.Getenv("GatewayAuthSecret")
2048         os.Unsetenv("GatewayAuthSecret")
2049         if gwAuthSecret == "" {
2050                 // not safe to run a gateway service without an auth
2051                 // secret
2052                 cr.CrunchLog.Printf("Not starting a gateway server (GatewayAuthSecret was not provided by dispatcher)")
2053         } else {
2054                 gwListen := os.Getenv("GatewayAddress")
2055                 cr.gateway = Gateway{
2056                         Address:       gwListen,
2057                         AuthSecret:    gwAuthSecret,
2058                         ContainerUUID: containerUUID,
2059                         Target:        cr.executor,
2060                         Log:           cr.CrunchLog,
2061                         LogCollection: cr.LogCollection,
2062                 }
2063                 if gwListen == "" {
2064                         // Direct connection won't work, so we use the
2065                         // gateway_address field to indicate the
2066                         // internalURL of the controller process that
2067                         // has the current tunnel connection.
2068                         cr.gateway.ArvadosClient = cr.dispatcherClient
2069                         cr.gateway.UpdateTunnelURL = func(url string) {
2070                                 cr.gateway.Address = "tunnel " + url
2071                                 cr.DispatcherArvClient.Update("containers", containerUUID,
2072                                         arvadosclient.Dict{
2073                                                 "select":    []string{"uuid"},
2074                                                 "container": arvadosclient.Dict{"gateway_address": cr.gateway.Address},
2075                                         }, nil)
2076                         }
2077                 }
2078                 err = cr.gateway.Start()
2079                 if err != nil {
2080                         log.Printf("error starting gateway server: %s", err)
2081                         return 1
2082                 }
2083         }
2084
2085         parentTemp, tmperr := cr.MkTempDir("", "crunch-run."+containerUUID+".")
2086         if tmperr != nil {
2087                 log.Printf("%s: %v", containerUUID, tmperr)
2088                 return 1
2089         }
2090
2091         cr.parentTemp = parentTemp
2092         cr.statInterval = *statInterval
2093         cr.enableMemoryLimit = *enableMemoryLimit
2094         cr.enableNetwork = *enableNetwork
2095         cr.networkMode = *networkMode
2096         if *cgroupParentSubsystem != "" {
2097                 p, err := findCgroup(os.DirFS("/"), *cgroupParentSubsystem)
2098                 if err != nil {
2099                         log.Printf("fatal: cgroup parent subsystem: %s", err)
2100                         return 1
2101                 }
2102                 cr.setCgroupParent = p
2103         }
2104
2105         if conf.EC2SpotCheck {
2106                 go cr.checkSpotInterruptionNotices()
2107         }
2108
2109         runerr := cr.Run()
2110
2111         if *memprofile != "" {
2112                 f, err := os.Create(*memprofile)
2113                 if err != nil {
2114                         log.Printf("could not create memory profile: %s", err)
2115                 }
2116                 runtime.GC() // get up-to-date statistics
2117                 if err := pprof.WriteHeapProfile(f); err != nil {
2118                         log.Printf("could not write memory profile: %s", err)
2119                 }
2120                 closeerr := f.Close()
2121                 if closeerr != nil {
2122                         log.Printf("closing memprofile file: %s", err)
2123                 }
2124         }
2125
2126         if runerr != nil {
2127                 log.Printf("%s: %v", containerUUID, runerr)
2128                 return 1
2129         }
2130         return 0
2131 }
2132
2133 // Try to load ConfigData in hpc (slurm/lsf) environment. This means
2134 // loading the cluster config from the specified file and (if that
2135 // works) getting the runtime_constraints container field from
2136 // controller to determine # VCPUs so we can calculate KeepBuffers.
2137 func hpcConfData(uuid string, configFile string, stderr io.Writer) ConfigData {
2138         var conf ConfigData
2139         conf.Cluster = loadClusterConfigFile(configFile, stderr)
2140         if conf.Cluster == nil {
2141                 // skip loading the container record -- we won't be
2142                 // able to start local keepstore anyway.
2143                 return conf
2144         }
2145         arv, err := arvadosclient.MakeArvadosClient()
2146         if err != nil {
2147                 fmt.Fprintf(stderr, "error setting up arvadosclient: %s\n", err)
2148                 return conf
2149         }
2150         // arvadosclient now interprets Retries=10 to mean
2151         // Timeout=10m, retrying with exponential backoff + jitter.
2152         arv.Retries = 10
2153         var ctr arvados.Container
2154         err = arv.Call("GET", "containers", uuid, "", arvadosclient.Dict{"select": []string{"runtime_constraints"}}, &ctr)
2155         if err != nil {
2156                 fmt.Fprintf(stderr, "error getting container record: %s\n", err)
2157                 return conf
2158         }
2159         if ctr.RuntimeConstraints.VCPUs > 0 {
2160                 conf.KeepBuffers = ctr.RuntimeConstraints.VCPUs * conf.Cluster.Containers.LocalKeepBlobBuffersPerVCPU
2161         }
2162         return conf
2163 }
2164
2165 // Load cluster config file from given path. If an error occurs, log
2166 // the error to stderr and return nil.
2167 func loadClusterConfigFile(path string, stderr io.Writer) *arvados.Cluster {
2168         ldr := config.NewLoader(&bytes.Buffer{}, ctxlog.New(stderr, "plain", "info"))
2169         ldr.Path = path
2170         cfg, err := ldr.Load()
2171         if err != nil {
2172                 fmt.Fprintf(stderr, "could not load config file %s: %s\n", path, err)
2173                 return nil
2174         }
2175         cluster, err := cfg.GetCluster("")
2176         if err != nil {
2177                 fmt.Fprintf(stderr, "could not use config file %s: %s\n", path, err)
2178                 return nil
2179         }
2180         fmt.Fprintf(stderr, "loaded config file %s\n", path)
2181         return cluster
2182 }
2183
2184 func startLocalKeepstore(configData ConfigData, logbuf io.Writer) (*exec.Cmd, error) {
2185         if configData.KeepBuffers < 1 {
2186                 fmt.Fprintf(logbuf, "not starting a local keepstore process because KeepBuffers=%v in config\n", configData.KeepBuffers)
2187                 return nil, nil
2188         }
2189         if configData.Cluster == nil {
2190                 fmt.Fprint(logbuf, "not starting a local keepstore process because cluster config file was not loaded\n")
2191                 return nil, nil
2192         }
2193         for uuid, vol := range configData.Cluster.Volumes {
2194                 if len(vol.AccessViaHosts) > 0 {
2195                         fmt.Fprintf(logbuf, "not starting a local keepstore process because a volume (%s) uses AccessViaHosts\n", uuid)
2196                         return nil, nil
2197                 }
2198                 if !vol.ReadOnly && vol.Replication < configData.Cluster.Collections.DefaultReplication {
2199                         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)
2200                         return nil, nil
2201                 }
2202         }
2203
2204         // Rather than have an alternate way to tell keepstore how
2205         // many buffers to use, etc., when starting it this way, we
2206         // just modify the cluster configuration that we feed it on
2207         // stdin.
2208         ccfg := *configData.Cluster
2209         ccfg.API.MaxKeepBlobBuffers = configData.KeepBuffers
2210         ccfg.Collections.BlobTrash = false
2211         ccfg.Collections.BlobTrashConcurrency = 0
2212         ccfg.Collections.BlobDeleteConcurrency = 0
2213
2214         localaddr := localKeepstoreAddr()
2215         ln, err := net.Listen("tcp", net.JoinHostPort(localaddr, "0"))
2216         if err != nil {
2217                 return nil, err
2218         }
2219         _, port, err := net.SplitHostPort(ln.Addr().String())
2220         if err != nil {
2221                 ln.Close()
2222                 return nil, err
2223         }
2224         ln.Close()
2225         url := "http://" + net.JoinHostPort(localaddr, port)
2226
2227         fmt.Fprintf(logbuf, "starting keepstore on %s\n", url)
2228
2229         var confJSON bytes.Buffer
2230         err = json.NewEncoder(&confJSON).Encode(arvados.Config{
2231                 Clusters: map[string]arvados.Cluster{
2232                         ccfg.ClusterID: ccfg,
2233                 },
2234         })
2235         if err != nil {
2236                 return nil, err
2237         }
2238         cmd := exec.Command("/proc/self/exe", "keepstore", "-config=-")
2239         if target, err := os.Readlink(cmd.Path); err == nil && strings.HasSuffix(target, ".test") {
2240                 // If we're a 'go test' process, running
2241                 // /proc/self/exe would start the test suite in a
2242                 // child process, which is not what we want.
2243                 cmd.Path, _ = exec.LookPath("go")
2244                 cmd.Args = append([]string{"go", "run", "../../cmd/arvados-server"}, cmd.Args[1:]...)
2245                 cmd.Env = os.Environ()
2246         }
2247         cmd.Stdin = &confJSON
2248         cmd.Stdout = logbuf
2249         cmd.Stderr = logbuf
2250         cmd.Env = append(cmd.Env,
2251                 "GOGC=10",
2252                 "ARVADOS_SERVICE_INTERNAL_URL="+url)
2253         err = cmd.Start()
2254         if err != nil {
2255                 return nil, fmt.Errorf("error starting keepstore process: %w", err)
2256         }
2257         cmdExited := false
2258         go func() {
2259                 cmd.Wait()
2260                 cmdExited = true
2261         }()
2262         ctx, cancel := context.WithDeadline(context.Background(), time.Now().Add(time.Second*10))
2263         defer cancel()
2264         poll := time.NewTicker(time.Second / 10)
2265         defer poll.Stop()
2266         client := http.Client{}
2267         for range poll.C {
2268                 testReq, err := http.NewRequestWithContext(ctx, "GET", url+"/_health/ping", nil)
2269                 testReq.Header.Set("Authorization", "Bearer "+configData.Cluster.ManagementToken)
2270                 if err != nil {
2271                         return nil, err
2272                 }
2273                 resp, err := client.Do(testReq)
2274                 if err == nil {
2275                         resp.Body.Close()
2276                         if resp.StatusCode == http.StatusOK {
2277                                 break
2278                         }
2279                 }
2280                 if cmdExited {
2281                         return nil, fmt.Errorf("keepstore child process exited")
2282                 }
2283                 if ctx.Err() != nil {
2284                         return nil, fmt.Errorf("timed out waiting for new keepstore process to report healthy")
2285                 }
2286         }
2287         os.Setenv("ARVADOS_KEEP_SERVICES", url)
2288         return cmd, nil
2289 }
2290
2291 // return current uid, gid, groups in a format suitable for logging:
2292 // "crunch-run process has uid=1234(arvados) gid=1234(arvados)
2293 // groups=1234(arvados),114(fuse)"
2294 func currentUserAndGroups() string {
2295         u, err := user.Current()
2296         if err != nil {
2297                 return fmt.Sprintf("error getting current user ID: %s", err)
2298         }
2299         s := fmt.Sprintf("crunch-run process has uid=%s(%s) gid=%s", u.Uid, u.Username, u.Gid)
2300         if g, err := user.LookupGroupId(u.Gid); err == nil {
2301                 s += fmt.Sprintf("(%s)", g.Name)
2302         }
2303         s += " groups="
2304         if gids, err := u.GroupIds(); err == nil {
2305                 for i, gid := range gids {
2306                         if i > 0 {
2307                                 s += ","
2308                         }
2309                         s += gid
2310                         if g, err := user.LookupGroupId(gid); err == nil {
2311                                 s += fmt.Sprintf("(%s)", g.Name)
2312                         }
2313                 }
2314         }
2315         return s
2316 }
2317
2318 // Return a suitable local interface address for a local keepstore
2319 // service. Currently this is the numerically lowest non-loopback ipv4
2320 // address assigned to a local interface that is not in any of the
2321 // link-local/vpn/loopback ranges 169.254/16, 100.64/10, or 127/8.
2322 func localKeepstoreAddr() string {
2323         var ips []net.IP
2324         // Ignore error (proceed with zero IPs)
2325         addrs, _ := processIPs(os.Getpid())
2326         for addr := range addrs {
2327                 ip := net.ParseIP(addr)
2328                 if ip == nil {
2329                         // invalid
2330                         continue
2331                 }
2332                 if ip.Mask(net.CIDRMask(8, 32)).Equal(net.IPv4(127, 0, 0, 0)) ||
2333                         ip.Mask(net.CIDRMask(10, 32)).Equal(net.IPv4(100, 64, 0, 0)) ||
2334                         ip.Mask(net.CIDRMask(16, 32)).Equal(net.IPv4(169, 254, 0, 0)) {
2335                         // unsuitable
2336                         continue
2337                 }
2338                 ips = append(ips, ip)
2339         }
2340         if len(ips) == 0 {
2341                 return "0.0.0.0"
2342         }
2343         sort.Slice(ips, func(ii, jj int) bool {
2344                 i, j := ips[ii], ips[jj]
2345                 if len(i) != len(j) {
2346                         return len(i) < len(j)
2347                 }
2348                 for x := range i {
2349                         if i[x] != j[x] {
2350                                 return i[x] < j[x]
2351                         }
2352                 }
2353                 return false
2354         })
2355         return ips[0].String()
2356 }
2357
2358 func (cr *ContainerRunner) loadPrices() {
2359         buf, err := os.ReadFile(filepath.Join(lockdir, pricesfile))
2360         if err != nil {
2361                 if !os.IsNotExist(err) {
2362                         cr.CrunchLog.Printf("loadPrices: read: %s", err)
2363                 }
2364                 return
2365         }
2366         var prices []cloud.InstancePrice
2367         err = json.Unmarshal(buf, &prices)
2368         if err != nil {
2369                 cr.CrunchLog.Printf("loadPrices: decode: %s", err)
2370                 return
2371         }
2372         cr.pricesLock.Lock()
2373         defer cr.pricesLock.Unlock()
2374         var lastKnown time.Time
2375         if len(cr.prices) > 0 {
2376                 lastKnown = cr.prices[0].StartTime
2377         }
2378         cr.prices = cloud.NormalizePriceHistory(append(prices, cr.prices...))
2379         for i := len(cr.prices) - 1; i >= 0; i-- {
2380                 price := cr.prices[i]
2381                 if price.StartTime.After(lastKnown) {
2382                         cr.CrunchLog.Printf("Instance price changed to %#.3g at %s", price.Price, price.StartTime.UTC())
2383                 }
2384         }
2385 }
2386
2387 func (cr *ContainerRunner) calculateCost(now time.Time) float64 {
2388         cr.pricesLock.Lock()
2389         defer cr.pricesLock.Unlock()
2390
2391         // First, make a "prices" slice with the real data as far back
2392         // as it goes, and (if needed) a "since the beginning of time"
2393         // placeholder containing a reasonable guess about what the
2394         // price was between cr.costStartTime and the earliest real
2395         // data point.
2396         prices := cr.prices
2397         if len(prices) == 0 {
2398                 // use price info in InstanceType record initially
2399                 // provided by cloud dispatcher
2400                 var p float64
2401                 var it arvados.InstanceType
2402                 if j := os.Getenv("InstanceType"); j != "" && json.Unmarshal([]byte(j), &it) == nil && it.Price > 0 {
2403                         p = it.Price
2404                 }
2405                 prices = []cloud.InstancePrice{{Price: p}}
2406         } else if prices[len(prices)-1].StartTime.After(cr.costStartTime) {
2407                 // guess earlier pricing was the same as the earliest
2408                 // price we know about
2409                 filler := prices[len(prices)-1]
2410                 filler.StartTime = time.Time{}
2411                 prices = append(prices, filler)
2412         }
2413
2414         // Now that our history of price changes goes back at least as
2415         // far as cr.costStartTime, add up the costs for each
2416         // interval.
2417         cost := 0.0
2418         spanEnd := now
2419         for _, ip := range prices {
2420                 spanStart := ip.StartTime
2421                 if spanStart.After(now) {
2422                         // pricing information from the future -- not
2423                         // expected from AWS, but possible in
2424                         // principle, and exercised by tests.
2425                         continue
2426                 }
2427                 last := false
2428                 if spanStart.Before(cr.costStartTime) {
2429                         spanStart = cr.costStartTime
2430                         last = true
2431                 }
2432                 cost += ip.Price * spanEnd.Sub(spanStart).Seconds() / 3600
2433                 if last {
2434                         break
2435                 }
2436                 spanEnd = spanStart
2437         }
2438
2439         return cost
2440 }
2441
2442 func (runner *ContainerRunner) handleSIGUSR2(sigchan chan os.Signal) {
2443         for range sigchan {
2444                 runner.loadPrices()
2445                 update := arvadosclient.Dict{
2446                         "select": []string{"uuid"},
2447                         "container": arvadosclient.Dict{
2448                                 "cost": runner.calculateCost(time.Now()),
2449                         },
2450                 }
2451                 runner.DispatcherArvClient.Update("containers", runner.Container.UUID, update, nil)
2452         }
2453 }