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         "git.arvados.org/arvados.git/sdk/go/manifest"
  44         "golang.org/x/sys/unix"
  45 )
  46
  47 type command struct{}
  48
  49 var arvadosCertPath = "/etc/arvados/ca-certificates.crt"
  50
  51 var Command = command{}
  52
  53 // ConfigData contains environment variables and (when needed) cluster
  54 // configuration, passed from dispatchcloud to crunch-run on stdin.
  55 type ConfigData struct {
  56         Env          map[string]string
  57         KeepBuffers  int
  58         EC2SpotCheck bool
  59         Cluster      *arvados.Cluster
  60 }
  61
  62 // IArvadosClient is the minimal Arvados API methods used by crunch-run.
  63 type IArvadosClient interface {
  64         Create(resourceType string, parameters arvadosclient.Dict, output interface{}) error
  65         Get(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  66         Update(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  67         Call(method, resourceType, uuid, action string, parameters arvadosclient.Dict, output interface{}) error
  68         CallRaw(method string, resourceType string, uuid string, action string, parameters arvadosclient.Dict) (reader io.ReadCloser, err error)
  69         Discovery(key string) (interface{}, error)
  70 }
  71
  72 // ErrCancelled is the error returned when the container is cancelled.
  73 var ErrCancelled = errors.New("Cancelled")
  74
  75 // IKeepClient is the minimal Keep API methods used by crunch-run.
  76 type IKeepClient interface {
  77         BlockWrite(context.Context, arvados.BlockWriteOptions) (arvados.BlockWriteResponse, error)
  78         ReadAt(locator string, p []byte, off int) (int, error)
  79         ManifestFileReader(m manifest.Manifest, filename string) (arvados.File, error)
  80         LocalLocator(locator string) (string, error)
  81         SetStorageClasses(sc []string)
  82 }
  83
  84 // NewLogWriter is a factory function to create a new log writer.
  85 type NewLogWriter func(name string) (io.WriteCloser, error)
  86
  87 type RunArvMount func(cmdline []string, tok string) (*exec.Cmd, error)
  88
  89 type MkTempDir func(string, string) (string, error)
  90
  91 type PsProcess interface {
  92         CmdlineSlice() ([]string, error)
  93 }
  94
  95 // ContainerRunner is the main stateful struct used for a single execution of a
  96 // container.
  97 type ContainerRunner struct {
  98         executor       containerExecutor
  99         executorStdin  io.Closer
 100         executorStdout io.Closer
 101         executorStderr io.Closer
 102
 103         // Dispatcher client is initialized with the Dispatcher token.
 104         // This is a privileged token used to manage container status
 105         // and logs.
 106         //
 107         // We have both dispatcherClient and DispatcherArvClient
 108         // because there are two different incompatible Arvados Go
 109         // SDKs and we have to use both (hopefully this gets fixed in
 110         // #14467)
 111         dispatcherClient     *arvados.Client
 112         DispatcherArvClient  IArvadosClient
 113         DispatcherKeepClient IKeepClient
 114
 115         // Container client is initialized with the Container token
 116         // This token controls the permissions of the container, and
 117         // must be used for operations such as reading collections.
 118         //
 119         // Same comment as above applies to
 120         // containerClient/ContainerArvClient.
 121         containerClient     *arvados.Client
 122         ContainerArvClient  IArvadosClient
 123         ContainerKeepClient IKeepClient
 124
 125         Container     arvados.Container
 126         token         string
 127         ExitCode      *int
 128         NewLogWriter  NewLogWriter
 129         CrunchLog     *ThrottledLogger
 130         logUUID       string
 131         logMtx        sync.Mutex
 132         LogCollection arvados.CollectionFileSystem
 133         LogsPDH       *string
 134         RunArvMount   RunArvMount
 135         MkTempDir     MkTempDir
 136         ArvMount      *exec.Cmd
 137         ArvMountPoint string
 138         HostOutputDir string
 139         Volumes       map[string]struct{}
 140         OutputPDH     *string
 141         SigChan       chan os.Signal
 142         ArvMountExit  chan error
 143         SecretMounts  map[string]arvados.Mount
 144         MkArvClient   func(token string) (IArvadosClient, IKeepClient, *arvados.Client, error)
 145         finalState    string
 146         parentTemp    string
 147         costStartTime time.Time
 148
 149         keepstore        *exec.Cmd
 150         keepstoreLogger  io.WriteCloser
 151         keepstoreLogbuf  *bufThenWrite
 152         statLogger       io.WriteCloser
 153         statReporter     *crunchstat.Reporter
 154         hoststatLogger   io.WriteCloser
 155         hoststatReporter *crunchstat.Reporter
 156         statInterval     time.Duration
 157         // What we tell docker to use as the container's cgroup
 158         // parent.
 159         setCgroupParent string
 160         // Fake root dir where crunchstat.Reporter should read OS
 161         // files, for testing.
 162         crunchstatFakeFS fs.FS
 163
 164         cStateLock sync.Mutex
 165         cCancelled bool // StopContainer() invoked
 166
 167         enableMemoryLimit bool
 168         enableNetwork     string // one of "default" or "always"
 169         networkMode       string // "none", "host", or "" -- passed through to executor
 170         brokenNodeHook    string // script to run if node appears to be broken
 171         arvMountLog       *ThrottledLogger
 172
 173         containerWatchdogInterval time.Duration
 174
 175         gateway Gateway
 176
 177         prices     []cloud.InstancePrice
 178         pricesLock sync.Mutex
 179 }
 180
 181 // setupSignals sets up signal handling to gracefully terminate the
 182 // underlying container and update state when receiving a TERM, INT or
 183 // QUIT signal.
 184 func (runner *ContainerRunner) setupSignals() {
 185         runner.SigChan = make(chan os.Signal, 1)
 186         signal.Notify(runner.SigChan, syscall.SIGTERM)
 187         signal.Notify(runner.SigChan, syscall.SIGINT)
 188         signal.Notify(runner.SigChan, syscall.SIGQUIT)
 189
 190         go func(sig chan os.Signal) {
 191                 for s := range sig {
 192                         runner.stop(s)
 193                 }
 194         }(runner.SigChan)
 195 }
 196
 197 // stop the underlying container.
 198 func (runner *ContainerRunner) stop(sig os.Signal) {
 199         runner.cStateLock.Lock()
 200         defer runner.cStateLock.Unlock()
 201         if sig != nil {
 202                 runner.CrunchLog.Printf("caught signal: %v", sig)
 203         }
 204         runner.cCancelled = true
 205         runner.CrunchLog.Printf("stopping container")
 206         err := runner.executor.Stop()
 207         if err != nil {
 208                 runner.CrunchLog.Printf("error stopping container: %s", err)
 209         }
 210 }
 211
 212 var errorBlacklist = []string{
 213         "(?ms).*[Cc]annot connect to the Docker daemon.*",
 214         "(?ms).*oci runtime error.*starting container process.*container init.*mounting.*to rootfs.*no such file or directory.*",
 215         "(?ms).*grpc: the connection is unavailable.*",
 216 }
 217
 218 func (runner *ContainerRunner) runBrokenNodeHook() {
 219         if runner.brokenNodeHook == "" {
 220                 path := filepath.Join(lockdir, brokenfile)
 221                 runner.CrunchLog.Printf("Writing %s to mark node as broken", path)
 222                 f, err := os.OpenFile(path, os.O_CREATE|os.O_RDWR, 0700)
 223                 if err != nil {
 224                         runner.CrunchLog.Printf("Error writing %s: %s", path, err)
 225                         return
 226                 }
 227                 f.Close()
 228         } else {
 229                 runner.CrunchLog.Printf("Running broken node hook %q", runner.brokenNodeHook)
 230                 // run killme script
 231                 c := exec.Command(runner.brokenNodeHook)
 232                 c.Stdout = runner.CrunchLog
 233                 c.Stderr = runner.CrunchLog
 234                 err := c.Run()
 235                 if err != nil {
 236                         runner.CrunchLog.Printf("Error running broken node hook: %v", err)
 237                 }
 238         }
 239 }
 240
 241 func (runner *ContainerRunner) checkBrokenNode(goterr error) bool {
 242         for _, d := range errorBlacklist {
 243                 if m, e := regexp.MatchString(d, goterr.Error()); m && e == nil {
 244                         runner.CrunchLog.Printf("Error suggests node is unable to run containers: %v", goterr)
 245                         runner.runBrokenNodeHook()
 246                         return true
 247                 }
 248         }
 249         return false
 250 }
 251
 252 // LoadImage determines the docker image id from the container record and
 253 // checks if it is available in the local Docker image store.  If not, it loads
 254 // the image from Keep.
 255 func (runner *ContainerRunner) LoadImage() (string, error) {
 256         runner.CrunchLog.Printf("Fetching Docker image from collection '%s'", runner.Container.ContainerImage)
 257
 258         d, err := os.Open(runner.ArvMountPoint + "/by_id/" + runner.Container.ContainerImage)
 259         if err != nil {
 260                 return "", err
 261         }
 262         defer d.Close()
 263         allfiles, err := d.Readdirnames(-1)
 264         if err != nil {
 265                 return "", err
 266         }
 267         var tarfiles []string
 268         for _, fnm := range allfiles {
 269                 if strings.HasSuffix(fnm, ".tar") {
 270                         tarfiles = append(tarfiles, fnm)
 271                 }
 272         }
 273         if len(tarfiles) == 0 {
 274                 return "", fmt.Errorf("image collection does not include a .tar image file")
 275         }
 276         if len(tarfiles) > 1 {
 277                 return "", fmt.Errorf("cannot choose from multiple tar files in image collection: %v", tarfiles)
 278         }
 279         imageID := tarfiles[0][:len(tarfiles[0])-4]
 280         imageTarballPath := runner.ArvMountPoint + "/by_id/" + runner.Container.ContainerImage + "/" + imageID + ".tar"
 281         runner.CrunchLog.Printf("Using Docker image id %q", imageID)
 282
 283         runner.CrunchLog.Print("Loading Docker image from keep")
 284         err = runner.executor.LoadImage(imageID, imageTarballPath, runner.Container, runner.ArvMountPoint,
 285                 runner.containerClient)
 286         if err != nil {
 287                 return "", err
 288         }
 289
 290         return imageID, nil
 291 }
 292
 293 func (runner *ContainerRunner) ArvMountCmd(cmdline []string, token string) (c *exec.Cmd, err error) {
 294         c = exec.Command(cmdline[0], cmdline[1:]...)
 295
 296         // Copy our environment, but override ARVADOS_API_TOKEN with
 297         // the container auth token.
 298         c.Env = nil
 299         for _, s := range os.Environ() {
 300                 if !strings.HasPrefix(s, "ARVADOS_API_TOKEN=") {
 301                         c.Env = append(c.Env, s)
 302                 }
 303         }
 304         c.Env = append(c.Env, "ARVADOS_API_TOKEN="+token)
 305
 306         w, err := runner.NewLogWriter("arv-mount")
 307         if err != nil {
 308                 return nil, err
 309         }
 310         runner.arvMountLog = NewThrottledLogger(w)
 311         scanner := logScanner{
 312                 Patterns: []string{
 313                         "Keep write error",
 314                         "Block not found error",
 315                         "Unhandled exception during FUSE operation",
 316                 },
 317                 ReportFunc: func(pattern, text string) {
 318                         runner.updateRuntimeStatus(arvadosclient.Dict{
 319                                 "warning":       "arv-mount: " + pattern,
 320                                 "warningDetail": text,
 321                         })
 322                 },
 323         }
 324         c.Stdout = runner.arvMountLog
 325         c.Stderr = io.MultiWriter(runner.arvMountLog, os.Stderr, &scanner)
 326
 327         runner.CrunchLog.Printf("Running %v", c.Args)
 328
 329         err = c.Start()
 330         if err != nil {
 331                 return nil, err
 332         }
 333
 334         statReadme := make(chan bool)
 335         runner.ArvMountExit = make(chan error)
 336
 337         keepStatting := true
 338         go func() {
 339                 for keepStatting {
 340                         time.Sleep(100 * time.Millisecond)
 341                         _, err = os.Stat(fmt.Sprintf("%s/by_id/README", runner.ArvMountPoint))
 342                         if err == nil {
 343                                 keepStatting = false
 344                                 statReadme <- true
 345                         }
 346                 }
 347                 close(statReadme)
 348         }()
 349
 350         go func() {
 351                 mnterr := c.Wait()
 352                 if mnterr != nil {
 353                         runner.CrunchLog.Printf("Arv-mount exit error: %v", mnterr)
 354                 }
 355                 runner.ArvMountExit <- mnterr
 356                 close(runner.ArvMountExit)
 357         }()
 358
 359         select {
 360         case <-statReadme:
 361                 break
 362         case err := <-runner.ArvMountExit:
 363                 runner.ArvMount = nil
 364                 keepStatting = false
 365                 return nil, err
 366         }
 367
 368         return c, nil
 369 }
 370
 371 func (runner *ContainerRunner) SetupArvMountPoint(prefix string) (err error) {
 372         if runner.ArvMountPoint == "" {
 373                 runner.ArvMountPoint, err = runner.MkTempDir(runner.parentTemp, prefix)
 374         }
 375         return
 376 }
 377
 378 func copyfile(src string, dst string) (err error) {
 379         srcfile, err := os.Open(src)
 380         if err != nil {
 381                 return
 382         }
 383
 384         os.MkdirAll(path.Dir(dst), 0777)
 385
 386         dstfile, err := os.Create(dst)
 387         if err != nil {
 388                 return
 389         }
 390         _, err = io.Copy(dstfile, srcfile)
 391         if err != nil {
 392                 return
 393         }
 394
 395         err = srcfile.Close()
 396         err2 := dstfile.Close()
 397
 398         if err != nil {
 399                 return
 400         }
 401
 402         if err2 != nil {
 403                 return err2
 404         }
 405
 406         return nil
 407 }
 408
 409 func (runner *ContainerRunner) SetupMounts() (map[string]bindmount, error) {
 410         bindmounts := map[string]bindmount{}
 411         err := runner.SetupArvMountPoint("keep")
 412         if err != nil {
 413                 return nil, fmt.Errorf("While creating keep mount temp dir: %v", err)
 414         }
 415
 416         token, err := runner.ContainerToken()
 417         if err != nil {
 418                 return nil, fmt.Errorf("could not get container token: %s", err)
 419         }
 420         runner.CrunchLog.Printf("container token %q", token)
 421
 422         pdhOnly := true
 423         tmpcount := 0
 424         arvMountCmd := []string{
 425                 "arv-mount",
 426                 "--foreground",
 427                 "--read-write",
 428                 "--storage-classes", strings.Join(runner.Container.OutputStorageClasses, ","),
 429                 fmt.Sprintf("--crunchstat-interval=%v", runner.statInterval.Seconds())}
 430
 431         if _, isdocker := runner.executor.(*dockerExecutor); isdocker {
 432                 arvMountCmd = append(arvMountCmd, "--allow-other")
 433         }
 434
 435         if runner.Container.RuntimeConstraints.KeepCacheDisk > 0 {
 436                 keepcachedir, err := runner.MkTempDir(runner.parentTemp, "keepcache")
 437                 if err != nil {
 438                         return nil, fmt.Errorf("while creating keep cache temp dir: %v", err)
 439                 }
 440                 arvMountCmd = append(arvMountCmd, "--disk-cache", "--disk-cache-dir", keepcachedir, "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheDisk))
 441         } else if runner.Container.RuntimeConstraints.KeepCacheRAM > 0 {
 442                 arvMountCmd = append(arvMountCmd, "--ram-cache", "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheRAM))
 443         }
 444
 445         collectionPaths := []string{}
 446         needCertMount := true
 447         type copyFile struct {
 448                 src  string
 449                 bind string
 450         }
 451         var copyFiles []copyFile
 452
 453         var binds []string
 454         for bind := range runner.Container.Mounts {
 455                 binds = append(binds, bind)
 456         }
 457         for bind := range runner.SecretMounts {
 458                 if _, ok := runner.Container.Mounts[bind]; ok {
 459                         return nil, fmt.Errorf("secret mount %q conflicts with regular mount", bind)
 460                 }
 461                 if runner.SecretMounts[bind].Kind != "json" &&
 462                         runner.SecretMounts[bind].Kind != "text" {
 463                         return nil, fmt.Errorf("secret mount %q type is %q but only 'json' and 'text' are permitted",
 464                                 bind, runner.SecretMounts[bind].Kind)
 465                 }
 466                 binds = append(binds, bind)
 467         }
 468         sort.Strings(binds)
 469
 470         for _, bind := range binds {
 471                 mnt, notSecret := runner.Container.Mounts[bind]
 472                 if !notSecret {
 473                         mnt = runner.SecretMounts[bind]
 474                 }
 475                 if bind == "stdout" || bind == "stderr" {
 476                         // Is it a "file" mount kind?
 477                         if mnt.Kind != "file" {
 478                                 return nil, fmt.Errorf("unsupported mount kind '%s' for %s: only 'file' is supported", mnt.Kind, bind)
 479                         }
 480
 481                         // Does path start with OutputPath?
 482                         prefix := runner.Container.OutputPath
 483                         if !strings.HasSuffix(prefix, "/") {
 484                                 prefix += "/"
 485                         }
 486                         if !strings.HasPrefix(mnt.Path, prefix) {
 487                                 return nil, fmt.Errorf("%s path does not start with OutputPath: %s, %s", strings.Title(bind), mnt.Path, prefix)
 488                         }
 489                 }
 490
 491                 if bind == "stdin" {
 492                         // Is it a "collection" mount kind?
 493                         if mnt.Kind != "collection" && mnt.Kind != "json" {
 494                                 return nil, fmt.Errorf("unsupported mount kind '%s' for stdin: only 'collection' and 'json' are supported", mnt.Kind)
 495                         }
 496                 }
 497
 498                 if bind == arvadosCertPath {
 499                         needCertMount = false
 500                 }
 501
 502                 if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && bind != runner.Container.OutputPath+"/" {
 503                         if mnt.Kind != "collection" && mnt.Kind != "text" && mnt.Kind != "json" {
 504                                 return nil, fmt.Errorf("only mount points of kind 'collection', 'text' or 'json' are supported underneath the output_path for %q, was %q", bind, mnt.Kind)
 505                         }
 506                 }
 507
 508                 switch {
 509                 case mnt.Kind == "collection" && bind != "stdin":
 510                         var src string
 511                         if mnt.UUID != "" && mnt.PortableDataHash != "" {
 512                                 return nil, fmt.Errorf("cannot specify both 'uuid' and 'portable_data_hash' for a collection mount")
 513                         }
 514                         if mnt.UUID != "" {
 515                                 if mnt.Writable {
 516                                         return nil, fmt.Errorf("writing to existing collections currently not permitted")
 517                                 }
 518                                 pdhOnly = false
 519                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.UUID)
 520                         } else if mnt.PortableDataHash != "" {
 521                                 if mnt.Writable && !strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 522                                         return nil, fmt.Errorf("can never write to a collection specified by portable data hash")
 523                                 }
 524                                 idx := strings.Index(mnt.PortableDataHash, "/")
 525                                 if idx > 0 {
 526                                         mnt.Path = path.Clean(mnt.PortableDataHash[idx:])
 527                                         mnt.PortableDataHash = mnt.PortableDataHash[0:idx]
 528                                         runner.Container.Mounts[bind] = mnt
 529                                 }
 530                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.PortableDataHash)
 531                                 if mnt.Path != "" && mnt.Path != "." {
 532                                         if strings.HasPrefix(mnt.Path, "./") {
 533                                                 mnt.Path = mnt.Path[2:]
 534                                         } else if strings.HasPrefix(mnt.Path, "/") {
 535                                                 mnt.Path = mnt.Path[1:]
 536                                         }
 537                                         src += "/" + mnt.Path
 538                                 }
 539                         } else {
 540                                 src = fmt.Sprintf("%s/tmp%d", runner.ArvMountPoint, tmpcount)
 541                                 arvMountCmd = append(arvMountCmd, "--mount-tmp", fmt.Sprintf("tmp%d", tmpcount))
 542                                 tmpcount++
 543                         }
 544                         if mnt.Writable {
 545                                 if bind == runner.Container.OutputPath {
 546                                         runner.HostOutputDir = src
 547                                         bindmounts[bind] = bindmount{HostPath: src}
 548                                 } else if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 549                                         copyFiles = append(copyFiles, copyFile{src, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
 550                                 } else {
 551                                         bindmounts[bind] = bindmount{HostPath: src}
 552                                 }
 553                         } else {
 554                                 bindmounts[bind] = bindmount{HostPath: src, ReadOnly: true}
 555                         }
 556                         collectionPaths = append(collectionPaths, src)
 557
 558                 case mnt.Kind == "tmp":
 559                         var tmpdir string
 560                         tmpdir, err = runner.MkTempDir(runner.parentTemp, "tmp")
 561                         if err != nil {
 562                                 return nil, fmt.Errorf("while creating mount temp dir: %v", err)
 563                         }
 564                         st, staterr := os.Stat(tmpdir)
 565                         if staterr != nil {
 566                                 return nil, fmt.Errorf("while Stat on temp dir: %v", staterr)
 567                         }
 568                         err = os.Chmod(tmpdir, st.Mode()|os.ModeSetgid|0777)
 569                         if staterr != nil {
 570                                 return nil, fmt.Errorf("while Chmod temp dir: %v", err)
 571                         }
 572                         bindmounts[bind] = bindmount{HostPath: tmpdir}
 573                         if bind == runner.Container.OutputPath {
 574                                 runner.HostOutputDir = tmpdir
 575                         }
 576
 577                 case mnt.Kind == "json" || mnt.Kind == "text":
 578                         var filedata []byte
 579                         if mnt.Kind == "json" {
 580                                 filedata, err = json.Marshal(mnt.Content)
 581                                 if err != nil {
 582                                         return nil, fmt.Errorf("encoding json data: %v", err)
 583                                 }
 584                         } else {
 585                                 text, ok := mnt.Content.(string)
 586                                 if !ok {
 587                                         return nil, fmt.Errorf("content for mount %q must be a string", bind)
 588                                 }
 589                                 filedata = []byte(text)
 590                         }
 591
 592                         tmpdir, err := runner.MkTempDir(runner.parentTemp, mnt.Kind)
 593                         if err != nil {
 594                                 return nil, fmt.Errorf("creating temp dir: %v", err)
 595                         }
 596                         tmpfn := filepath.Join(tmpdir, "mountdata."+mnt.Kind)
 597                         err = ioutil.WriteFile(tmpfn, filedata, 0444)
 598                         if err != nil {
 599                                 return nil, fmt.Errorf("writing temp file: %v", err)
 600                         }
 601                         if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && (notSecret || runner.Container.Mounts[runner.Container.OutputPath].Kind != "collection") {
 602                                 // In most cases, if the container
 603                                 // specifies a literal file inside the
 604                                 // output path, we copy it into the
 605                                 // output directory (either a mounted
 606                                 // collection or a staging area on the
 607                                 // host fs). If it's a secret, it will
 608                                 // be skipped when copying output from
 609                                 // staging to Keep later.
 610                                 copyFiles = append(copyFiles, copyFile{tmpfn, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
 611                         } else {
 612                                 // If a secret is outside OutputPath,
 613                                 // we bind mount the secret file
 614                                 // directly just like other mounts. We
 615                                 // also use this strategy when a
 616                                 // secret is inside OutputPath but
 617                                 // OutputPath is a live collection, to
 618                                 // avoid writing the secret to
 619                                 // Keep. Attempting to remove a
 620                                 // bind-mounted secret file from
 621                                 // inside the container will return a
 622                                 // "Device or resource busy" error
 623                                 // that might not be handled well by
 624                                 // the container, which is why we
 625                                 // don't use this strategy when
 626                                 // OutputPath is a staging directory.
 627                                 bindmounts[bind] = bindmount{HostPath: tmpfn, ReadOnly: true}
 628                         }
 629
 630                 case mnt.Kind == "git_tree":
 631                         tmpdir, err := runner.MkTempDir(runner.parentTemp, "git_tree")
 632                         if err != nil {
 633                                 return nil, fmt.Errorf("creating temp dir: %v", err)
 634                         }
 635                         err = gitMount(mnt).extractTree(runner.containerClient, tmpdir, token)
 636                         if err != nil {
 637                                 return nil, err
 638                         }
 639                         bindmounts[bind] = bindmount{HostPath: tmpdir, ReadOnly: true}
 640                 }
 641         }
 642
 643         if runner.HostOutputDir == "" {
 644                 return nil, fmt.Errorf("output path does not correspond to a writable mount point")
 645         }
 646
 647         if needCertMount && runner.Container.RuntimeConstraints.API {
 648                 for _, certfile := range []string{
 649                         // Populated by caller, or sdk/go/arvados init(), or test suite:
 650                         os.Getenv("SSL_CERT_FILE"),
 651                         // Copied from Go 1.21 stdlib (src/crypto/x509/root_linux.go):
 652                         "/etc/ssl/certs/ca-certificates.crt",                // Debian/Ubuntu/Gentoo etc.
 653                         "/etc/pki/tls/certs/ca-bundle.crt",                  // Fedora/RHEL 6
 654                         "/etc/ssl/ca-bundle.pem",                            // OpenSUSE
 655                         "/etc/pki/tls/cacert.pem",                           // OpenELEC
 656                         "/etc/pki/ca-trust/extracted/pem/tls-ca-bundle.pem", // CentOS/RHEL 7
 657                         "/etc/ssl/cert.pem",                                 // Alpine Linux
 658                 } {
 659                         if _, err := os.Stat(certfile); err == nil {
 660                                 bindmounts[arvadosCertPath] = bindmount{HostPath: certfile, ReadOnly: true}
 661                                 break
 662                         }
 663                 }
 664         }
 665
 666         if pdhOnly {
 667                 // If we are only mounting collections by pdh, make
 668                 // sure we don't subscribe to websocket events to
 669                 // avoid putting undesired load on the API server
 670                 arvMountCmd = append(arvMountCmd, "--mount-by-pdh", "by_id", "--disable-event-listening")
 671         } else {
 672                 arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_id")
 673         }
 674         // the by_uuid mount point is used by singularity when writing
 675         // out docker images converted to SIF
 676         arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_uuid")
 677         arvMountCmd = append(arvMountCmd, runner.ArvMountPoint)
 678
 679         runner.ArvMount, err = runner.RunArvMount(arvMountCmd, token)
 680         if err != nil {
 681                 return nil, fmt.Errorf("while trying to start arv-mount: %v", err)
 682         }
 683         if runner.hoststatReporter != nil && runner.ArvMount != nil {
 684                 runner.hoststatReporter.ReportPID("arv-mount", runner.ArvMount.Process.Pid)
 685         }
 686
 687         for _, p := range collectionPaths {
 688                 _, err = os.Stat(p)
 689                 if err != nil {
 690                         return nil, fmt.Errorf("while checking that input files exist: %v", err)
 691                 }
 692         }
 693
 694         for _, cp := range copyFiles {
 695                 st, err := os.Stat(cp.src)
 696                 if err != nil {
 697                         return nil, fmt.Errorf("while staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 698                 }
 699                 if st.IsDir() {
 700                         err = filepath.Walk(cp.src, func(walkpath string, walkinfo os.FileInfo, walkerr error) error {
 701                                 if walkerr != nil {
 702                                         return walkerr
 703                                 }
 704                                 target := path.Join(cp.bind, walkpath[len(cp.src):])
 705                                 if walkinfo.Mode().IsRegular() {
 706                                         copyerr := copyfile(walkpath, target)
 707                                         if copyerr != nil {
 708                                                 return copyerr
 709                                         }
 710                                         return os.Chmod(target, walkinfo.Mode()|0777)
 711                                 } else if walkinfo.Mode().IsDir() {
 712                                         mkerr := os.MkdirAll(target, 0777)
 713                                         if mkerr != nil {
 714                                                 return mkerr
 715                                         }
 716                                         return os.Chmod(target, walkinfo.Mode()|os.ModeSetgid|0777)
 717                                 } else {
 718                                         return fmt.Errorf("source %q is not a regular file or directory", cp.src)
 719                                 }
 720                         })
 721                 } else if st.Mode().IsRegular() {
 722                         err = copyfile(cp.src, cp.bind)
 723                         if err == nil {
 724                                 err = os.Chmod(cp.bind, st.Mode()|0777)
 725                         }
 726                 }
 727                 if err != nil {
 728                         return nil, fmt.Errorf("while staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 729                 }
 730         }
 731
 732         return bindmounts, nil
 733 }
 734
 735 func (runner *ContainerRunner) stopHoststat() error {
 736         if runner.hoststatReporter == nil {
 737                 return nil
 738         }
 739         runner.hoststatReporter.Stop()
 740         runner.hoststatReporter.LogProcessMemMax(runner.CrunchLog)
 741         err := runner.hoststatLogger.Close()
 742         if err != nil {
 743                 return fmt.Errorf("error closing hoststat logs: %v", err)
 744         }
 745         return nil
 746 }
 747
 748 func (runner *ContainerRunner) startHoststat() error {
 749         w, err := runner.NewLogWriter("hoststat")
 750         if err != nil {
 751                 return err
 752         }
 753         runner.hoststatLogger = NewThrottledLogger(w)
 754         runner.hoststatReporter = &crunchstat.Reporter{
 755                 Logger: log.New(runner.hoststatLogger, "", 0),
 756                 // Our own cgroup is the "host" cgroup, in the sense
 757                 // that it accounts for resource usage outside the
 758                 // container. It doesn't count _all_ resource usage on
 759                 // the system.
 760                 //
 761                 // TODO?: Use the furthest ancestor of our own cgroup
 762                 // that has stats available. (Currently crunchstat
 763                 // does not have that capability.)
 764                 Pid:        os.Getpid,
 765                 PollPeriod: runner.statInterval,
 766         }
 767         runner.hoststatReporter.Start()
 768         runner.hoststatReporter.ReportPID("crunch-run", os.Getpid())
 769         return nil
 770 }
 771
 772 func (runner *ContainerRunner) startCrunchstat() error {
 773         w, err := runner.NewLogWriter("crunchstat")
 774         if err != nil {
 775                 return err
 776         }
 777         runner.statLogger = NewThrottledLogger(w)
 778         runner.statReporter = &crunchstat.Reporter{
 779                 Pid:    runner.executor.Pid,
 780                 FS:     runner.crunchstatFakeFS,
 781                 Logger: log.New(runner.statLogger, "", 0),
 782                 MemThresholds: map[string][]crunchstat.Threshold{
 783                         "rss": crunchstat.NewThresholdsFromPercentages(runner.Container.RuntimeConstraints.RAM, []int64{90, 95, 99}),
 784                 },
 785                 PollPeriod:      runner.statInterval,
 786                 TempDir:         runner.parentTemp,
 787                 ThresholdLogger: runner.CrunchLog,
 788         }
 789         runner.statReporter.Start()
 790         return nil
 791 }
 792
 793 type infoCommand struct {
 794         label string
 795         cmd   []string
 796 }
 797
 798 // LogHostInfo logs info about the current host, for debugging and
 799 // accounting purposes. Although it's logged as "node-info", this is
 800 // about the environment where crunch-run is actually running, which
 801 // might differ from what's described in the node record (see
 802 // LogNodeRecord).
 803 func (runner *ContainerRunner) LogHostInfo() (err error) {
 804         w, err := runner.NewLogWriter("node-info")
 805         if err != nil {
 806                 return
 807         }
 808
 809         commands := []infoCommand{
 810                 {
 811                         label: "Host Information",
 812                         cmd:   []string{"uname", "-a"},
 813                 },
 814                 {
 815                         label: "CPU Information",
 816                         cmd:   []string{"cat", "/proc/cpuinfo"},
 817                 },
 818                 {
 819                         label: "Memory Information",
 820                         cmd:   []string{"cat", "/proc/meminfo"},
 821                 },
 822                 {
 823                         label: "Disk Space",
 824                         cmd:   []string{"df", "-m", "/", os.TempDir()},
 825                 },
 826                 {
 827                         label: "Disk INodes",
 828                         cmd:   []string{"df", "-i", "/", os.TempDir()},
 829                 },
 830         }
 831
 832         // Run commands with informational output to be logged.
 833         for _, command := range commands {
 834                 fmt.Fprintln(w, command.label)
 835                 cmd := exec.Command(command.cmd[0], command.cmd[1:]...)
 836                 cmd.Stdout = w
 837                 cmd.Stderr = w
 838                 if err := cmd.Run(); err != nil {
 839                         err = fmt.Errorf("While running command %q: %v", command.cmd, err)
 840                         fmt.Fprintln(w, err)
 841                         return err
 842                 }
 843                 fmt.Fprintln(w, "")
 844         }
 845
 846         err = w.Close()
 847         if err != nil {
 848                 return fmt.Errorf("While closing node-info logs: %v", err)
 849         }
 850         return nil
 851 }
 852
 853 // LogContainerRecord gets and saves the raw JSON container record from the API server
 854 func (runner *ContainerRunner) LogContainerRecord() error {
 855         logged, err := runner.logAPIResponse("container", "containers", map[string]interface{}{"filters": [][]string{{"uuid", "=", runner.Container.UUID}}}, nil)
 856         if !logged && err == nil {
 857                 err = fmt.Errorf("error: no container record found for %s", runner.Container.UUID)
 858         }
 859         return err
 860 }
 861
 862 // LogNodeRecord logs the current host's InstanceType config entry (or
 863 // the arvados#node record, if running via crunch-dispatch-slurm).
 864 func (runner *ContainerRunner) LogNodeRecord() error {
 865         if it := os.Getenv("InstanceType"); it != "" {
 866                 // Dispatched via arvados-dispatch-cloud. Save
 867                 // InstanceType config fragment received from
 868                 // dispatcher on stdin.
 869                 w, err := runner.LogCollection.OpenFile("node.json", os.O_CREATE|os.O_WRONLY, 0666)
 870                 if err != nil {
 871                         return err
 872                 }
 873                 defer w.Close()
 874                 _, err = io.WriteString(w, it)
 875                 if err != nil {
 876                         return err
 877                 }
 878                 return w.Close()
 879         }
 880         // Dispatched via crunch-dispatch-slurm. Look up
 881         // apiserver's node record corresponding to
 882         // $SLURMD_NODENAME.
 883         hostname := os.Getenv("SLURMD_NODENAME")
 884         if hostname == "" {
 885                 hostname, _ = os.Hostname()
 886         }
 887         _, err := runner.logAPIResponse("node", "nodes", map[string]interface{}{"filters": [][]string{{"hostname", "=", hostname}}}, func(resp interface{}) {
 888                 // The "info" field has admin-only info when
 889                 // obtained with a privileged token, and
 890                 // should not be logged.
 891                 node, ok := resp.(map[string]interface{})
 892                 if ok {
 893                         delete(node, "info")
 894                 }
 895         })
 896         return err
 897 }
 898
 899 func (runner *ContainerRunner) logAPIResponse(label, path string, params map[string]interface{}, munge func(interface{})) (logged bool, err error) {
 900         writer, err := runner.LogCollection.OpenFile(label+".json", os.O_CREATE|os.O_WRONLY, 0666)
 901         if err != nil {
 902                 return false, err
 903         }
 904         w := &ArvLogWriter{
 905                 ArvClient:     runner.DispatcherArvClient,
 906                 UUID:          runner.Container.UUID,
 907                 loggingStream: label,
 908                 writeCloser:   writer,
 909         }
 910
 911         reader, err := runner.DispatcherArvClient.CallRaw("GET", path, "", "", arvadosclient.Dict(params))
 912         if err != nil {
 913                 return false, fmt.Errorf("error getting %s record: %v", label, err)
 914         }
 915         defer reader.Close()
 916
 917         dec := json.NewDecoder(reader)
 918         dec.UseNumber()
 919         var resp map[string]interface{}
 920         if err = dec.Decode(&resp); err != nil {
 921                 return false, fmt.Errorf("error decoding %s list response: %v", label, err)
 922         }
 923         items, ok := resp["items"].([]interface{})
 924         if !ok {
 925                 return false, fmt.Errorf("error decoding %s list response: no \"items\" key in API list response", label)
 926         } else if len(items) < 1 {
 927                 return false, nil
 928         }
 929         if munge != nil {
 930                 munge(items[0])
 931         }
 932         // Re-encode it using indentation to improve readability
 933         enc := json.NewEncoder(w)
 934         enc.SetIndent("", "    ")
 935         if err = enc.Encode(items[0]); err != nil {
 936                 return false, fmt.Errorf("error logging %s record: %v", label, err)
 937         }
 938         err = w.Close()
 939         if err != nil {
 940                 return false, fmt.Errorf("error closing %s.json in log collection: %v", label, err)
 941         }
 942         return true, nil
 943 }
 944
 945 func (runner *ContainerRunner) getStdoutFile(mntPath string) (*os.File, error) {
 946         stdoutPath := mntPath[len(runner.Container.OutputPath):]
 947         index := strings.LastIndex(stdoutPath, "/")
 948         if index > 0 {
 949                 subdirs := stdoutPath[:index]
 950                 if subdirs != "" {
 951                         st, err := os.Stat(runner.HostOutputDir)
 952                         if err != nil {
 953                                 return nil, fmt.Errorf("While Stat on temp dir: %v", err)
 954                         }
 955                         stdoutPath := filepath.Join(runner.HostOutputDir, subdirs)
 956                         err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777)
 957                         if err != nil {
 958                                 return nil, fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err)
 959                         }
 960                 }
 961         }
 962         stdoutFile, err := os.Create(filepath.Join(runner.HostOutputDir, stdoutPath))
 963         if err != nil {
 964                 return nil, fmt.Errorf("While creating file %q: %v", stdoutPath, err)
 965         }
 966
 967         return stdoutFile, nil
 968 }
 969
 970 // CreateContainer creates the docker container.
 971 func (runner *ContainerRunner) CreateContainer(imageID string, bindmounts map[string]bindmount) error {
 972         var stdin io.ReadCloser = ioutil.NopCloser(bytes.NewReader(nil))
 973         if mnt, ok := runner.Container.Mounts["stdin"]; ok {
 974                 switch mnt.Kind {
 975                 case "collection":
 976                         var collID string
 977                         if mnt.UUID != "" {
 978                                 collID = mnt.UUID
 979                         } else {
 980                                 collID = mnt.PortableDataHash
 981                         }
 982                         path := runner.ArvMountPoint + "/by_id/" + collID + "/" + mnt.Path
 983                         f, err := os.Open(path)
 984                         if err != nil {
 985                                 return err
 986                         }
 987                         stdin = f
 988                 case "json":
 989                         j, err := json.Marshal(mnt.Content)
 990                         if err != nil {
 991                                 return fmt.Errorf("error encoding stdin json data: %v", err)
 992                         }
 993                         stdin = ioutil.NopCloser(bytes.NewReader(j))
 994                 default:
 995                         return fmt.Errorf("stdin mount has unsupported kind %q", mnt.Kind)
 996                 }
 997         }
 998
 999         var stdout, stderr io.WriteCloser
1000         if mnt, ok := runner.Container.Mounts["stdout"]; ok {
1001                 f, err := runner.getStdoutFile(mnt.Path)
1002                 if err != nil {
1003                         return err
1004                 }
1005                 stdout = f
1006         } else if w, err := runner.NewLogWriter("stdout"); err != nil {
1007                 return err
1008         } else {
1009                 stdout = NewThrottledLogger(w)
1010         }
1011
1012         if mnt, ok := runner.Container.Mounts["stderr"]; ok {
1013                 f, err := runner.getStdoutFile(mnt.Path)
1014                 if err != nil {
1015                         return err
1016                 }
1017                 stderr = f
1018         } else if w, err := runner.NewLogWriter("stderr"); err != nil {
1019                 return err
1020         } else {
1021                 stderr = NewThrottledLogger(w)
1022         }
1023
1024         env := runner.Container.Environment
1025         enableNetwork := runner.enableNetwork == "always"
1026         if runner.Container.RuntimeConstraints.API {
1027                 enableNetwork = true
1028                 tok, err := runner.ContainerToken()
1029                 if err != nil {
1030                         return err
1031                 }
1032                 env = map[string]string{}
1033                 for k, v := range runner.Container.Environment {
1034                         env[k] = v
1035                 }
1036                 env["ARVADOS_API_TOKEN"] = tok
1037                 env["ARVADOS_API_HOST"] = os.Getenv("ARVADOS_API_HOST")
1038                 env["ARVADOS_API_HOST_INSECURE"] = os.Getenv("ARVADOS_API_HOST_INSECURE")
1039                 env["ARVADOS_KEEP_SERVICES"] = os.Getenv("ARVADOS_KEEP_SERVICES")
1040         }
1041         workdir := runner.Container.Cwd
1042         if workdir == "." {
1043                 // both "" and "." mean default
1044                 workdir = ""
1045         }
1046         ram := runner.Container.RuntimeConstraints.RAM
1047         if !runner.enableMemoryLimit {
1048                 ram = 0
1049         }
1050         runner.executorStdin = stdin
1051         runner.executorStdout = stdout
1052         runner.executorStderr = stderr
1053
1054         if runner.Container.RuntimeConstraints.CUDA.DeviceCount > 0 {
1055                 nvidiaModprobe(runner.CrunchLog)
1056         }
1057
1058         return runner.executor.Create(containerSpec{
1059                 Image:           imageID,
1060                 VCPUs:           runner.Container.RuntimeConstraints.VCPUs,
1061                 RAM:             ram,
1062                 WorkingDir:      workdir,
1063                 Env:             env,
1064                 BindMounts:      bindmounts,
1065                 Command:         runner.Container.Command,
1066                 EnableNetwork:   enableNetwork,
1067                 CUDADeviceCount: runner.Container.RuntimeConstraints.CUDA.DeviceCount,
1068                 NetworkMode:     runner.networkMode,
1069                 CgroupParent:    runner.setCgroupParent,
1070                 Stdin:           stdin,
1071                 Stdout:          stdout,
1072                 Stderr:          stderr,
1073         })
1074 }
1075
1076 // StartContainer starts the docker container created by CreateContainer.
1077 func (runner *ContainerRunner) StartContainer() error {
1078         runner.CrunchLog.Printf("Starting container")
1079         runner.cStateLock.Lock()
1080         defer runner.cStateLock.Unlock()
1081         if runner.cCancelled {
1082                 return ErrCancelled
1083         }
1084         err := runner.executor.Start()
1085         if err != nil {
1086                 var advice string
1087                 if m, e := regexp.MatchString("(?ms).*(exec|System error).*(no such file or directory|file not found).*", err.Error()); m && e == nil {
1088                         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])
1089                 }
1090                 return fmt.Errorf("could not start container: %v%s", err, advice)
1091         }
1092         return nil
1093 }
1094
1095 // WaitFinish waits for the container to terminate, capture the exit code, and
1096 // close the stdout/stderr logging.
1097 func (runner *ContainerRunner) WaitFinish() error {
1098         runner.CrunchLog.Print("Waiting for container to finish")
1099         var timeout <-chan time.Time
1100         if s := runner.Container.SchedulingParameters.MaxRunTime; s > 0 {
1101                 timeout = time.After(time.Duration(s) * time.Second)
1102         }
1103         ctx, cancel := context.WithCancel(context.Background())
1104         defer cancel()
1105         go func() {
1106                 select {
1107                 case <-timeout:
1108                         runner.CrunchLog.Printf("maximum run time exceeded. Stopping container.")
1109                         runner.stop(nil)
1110                 case <-runner.ArvMountExit:
1111                         runner.CrunchLog.Printf("arv-mount exited while container is still running. Stopping container.")
1112                         runner.stop(nil)
1113                 case <-ctx.Done():
1114                 }
1115         }()
1116         exitcode, err := runner.executor.Wait(ctx)
1117         if err != nil {
1118                 runner.checkBrokenNode(err)
1119                 return err
1120         }
1121         runner.ExitCode = &exitcode
1122
1123         extra := ""
1124         if exitcode&0x80 != 0 {
1125                 // Convert raw exit status (0x80 + signal number) to a
1126                 // string to log after the code, like " (signal 101)"
1127                 // or " (signal 9, killed)"
1128                 sig := syscall.WaitStatus(exitcode).Signal()
1129                 if name := unix.SignalName(sig); name != "" {
1130                         extra = fmt.Sprintf(" (signal %d, %s)", sig, name)
1131                 } else {
1132                         extra = fmt.Sprintf(" (signal %d)", sig)
1133                 }
1134         }
1135         runner.CrunchLog.Printf("Container exited with status code %d%s", exitcode, extra)
1136         err = runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1137                 "select":    []string{"uuid"},
1138                 "container": arvadosclient.Dict{"exit_code": exitcode},
1139         }, nil)
1140         if err != nil {
1141                 runner.CrunchLog.Printf("ignoring error updating exit_code: %s", err)
1142         }
1143
1144         var returnErr error
1145         if err = runner.executorStdin.Close(); err != nil {
1146                 err = fmt.Errorf("error closing container stdin: %s", err)
1147                 runner.CrunchLog.Printf("%s", err)
1148                 returnErr = err
1149         }
1150         if err = runner.executorStdout.Close(); err != nil {
1151                 err = fmt.Errorf("error closing container stdout: %s", err)
1152                 runner.CrunchLog.Printf("%s", err)
1153                 if returnErr == nil {
1154                         returnErr = err
1155                 }
1156         }
1157         if err = runner.executorStderr.Close(); err != nil {
1158                 err = fmt.Errorf("error closing container stderr: %s", err)
1159                 runner.CrunchLog.Printf("%s", err)
1160                 if returnErr == nil {
1161                         returnErr = err
1162                 }
1163         }
1164
1165         if runner.statReporter != nil {
1166                 runner.statReporter.Stop()
1167                 runner.statReporter.LogMaxima(runner.CrunchLog, map[string]int64{
1168                         "rss": runner.Container.RuntimeConstraints.RAM,
1169                 })
1170                 err = runner.statLogger.Close()
1171                 if err != nil {
1172                         runner.CrunchLog.Printf("error closing crunchstat logs: %v", err)
1173                 }
1174         }
1175         return returnErr
1176 }
1177
1178 func (runner *ContainerRunner) updateLogs() {
1179         ticker := time.NewTicker(crunchLogUpdatePeriod / 360)
1180         defer ticker.Stop()
1181
1182         sigusr1 := make(chan os.Signal, 1)
1183         signal.Notify(sigusr1, syscall.SIGUSR1)
1184         defer signal.Stop(sigusr1)
1185
1186         saveAtTime := time.Now().Add(crunchLogUpdatePeriod)
1187         saveAtSize := crunchLogUpdateSize
1188         var savedSize int64
1189         for {
1190                 select {
1191                 case <-ticker.C:
1192                 case <-sigusr1:
1193                         saveAtTime = time.Now()
1194                 }
1195                 runner.logMtx.Lock()
1196                 done := runner.LogsPDH != nil
1197                 runner.logMtx.Unlock()
1198                 if done {
1199                         return
1200                 }
1201                 size := runner.LogCollection.Size()
1202                 if size == savedSize || (time.Now().Before(saveAtTime) && size < saveAtSize) {
1203                         continue
1204                 }
1205                 saveAtTime = time.Now().Add(crunchLogUpdatePeriod)
1206                 saveAtSize = runner.LogCollection.Size() + crunchLogUpdateSize
1207                 saved, err := runner.saveLogCollection(false)
1208                 if err != nil {
1209                         runner.CrunchLog.Printf("error updating log collection: %s", err)
1210                         continue
1211                 }
1212
1213                 err = runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1214                         "select": []string{"uuid"},
1215                         "container": arvadosclient.Dict{
1216                                 "log": saved.PortableDataHash,
1217                         },
1218                 }, nil)
1219                 if err != nil {
1220                         runner.CrunchLog.Printf("error updating container log to %s: %s", saved.PortableDataHash, err)
1221                         continue
1222                 }
1223
1224                 savedSize = size
1225         }
1226 }
1227
1228 var spotInterruptionCheckInterval = 5 * time.Second
1229 var ec2MetadataBaseURL = "http://169.254.169.254"
1230
1231 const ec2TokenTTL = time.Second * 21600
1232
1233 func (runner *ContainerRunner) checkSpotInterruptionNotices() {
1234         type ec2metadata struct {
1235                 Action string    `json:"action"`
1236                 Time   time.Time `json:"time"`
1237         }
1238         runner.CrunchLog.Printf("Checking for spot interruptions every %v using instance metadata at %s", spotInterruptionCheckInterval, ec2MetadataBaseURL)
1239         var metadata ec2metadata
1240         var token string
1241         var tokenExp time.Time
1242         check := func() error {
1243                 ctx, cancel := context.WithDeadline(context.Background(), time.Now().Add(time.Minute))
1244                 defer cancel()
1245                 if token == "" || tokenExp.Sub(time.Now()) < time.Minute {
1246                         req, err := http.NewRequestWithContext(ctx, http.MethodPut, ec2MetadataBaseURL+"/latest/api/token", nil)
1247                         if err != nil {
1248                                 return err
1249                         }
1250                         req.Header.Set("X-aws-ec2-metadata-token-ttl-seconds", fmt.Sprintf("%d", int(ec2TokenTTL/time.Second)))
1251                         resp, err := http.DefaultClient.Do(req)
1252                         if err != nil {
1253                                 return err
1254                         }
1255                         defer resp.Body.Close()
1256                         if resp.StatusCode != http.StatusOK {
1257                                 return fmt.Errorf("%s", resp.Status)
1258                         }
1259                         newtoken, err := ioutil.ReadAll(resp.Body)
1260                         if err != nil {
1261                                 return err
1262                         }
1263                         token = strings.TrimSpace(string(newtoken))
1264                         tokenExp = time.Now().Add(ec2TokenTTL)
1265                 }
1266                 req, err := http.NewRequestWithContext(ctx, http.MethodGet, ec2MetadataBaseURL+"/latest/meta-data/spot/instance-action", nil)
1267                 if err != nil {
1268                         return err
1269                 }
1270                 req.Header.Set("X-aws-ec2-metadata-token", token)
1271                 resp, err := http.DefaultClient.Do(req)
1272                 if err != nil {
1273                         return err
1274                 }
1275                 defer resp.Body.Close()
1276                 metadata = ec2metadata{}
1277                 switch resp.StatusCode {
1278                 case http.StatusOK:
1279                         break
1280                 case http.StatusNotFound:
1281                         // "If Amazon EC2 is not preparing to stop or
1282                         // terminate the instance, or if you
1283                         // terminated the instance yourself,
1284                         // instance-action is not present in the
1285                         // instance metadata and you receive an HTTP
1286                         // 404 error when you try to retrieve it."
1287                         return nil
1288                 case http.StatusUnauthorized:
1289                         token = ""
1290                         return fmt.Errorf("%s", resp.Status)
1291                 default:
1292                         return fmt.Errorf("%s", resp.Status)
1293                 }
1294                 err = json.NewDecoder(resp.Body).Decode(&metadata)
1295                 if err != nil {
1296                         return err
1297                 }
1298                 return nil
1299         }
1300         failures := 0
1301         var lastmetadata ec2metadata
1302         for range time.NewTicker(spotInterruptionCheckInterval).C {
1303                 err := check()
1304                 if err != nil {
1305                         runner.CrunchLog.Printf("Error checking spot interruptions: %s", err)
1306                         failures++
1307                         if failures > 5 {
1308                                 runner.CrunchLog.Printf("Giving up on checking spot interruptions after too many consecutive failures")
1309                                 return
1310                         }
1311                         continue
1312                 }
1313                 failures = 0
1314                 if metadata != lastmetadata {
1315                         lastmetadata = metadata
1316                         text := fmt.Sprintf("Cloud provider scheduled instance %s at %s", metadata.Action, metadata.Time.UTC().Format(time.RFC3339))
1317                         runner.CrunchLog.Printf("%s", text)
1318                         runner.updateRuntimeStatus(arvadosclient.Dict{
1319                                 "warning":          "preemption notice",
1320                                 "warningDetail":    text,
1321                                 "preemptionNotice": text,
1322                         })
1323                         if proc, err := os.FindProcess(os.Getpid()); err == nil {
1324                                 // trigger updateLogs
1325                                 proc.Signal(syscall.SIGUSR1)
1326                         }
1327                 }
1328         }
1329 }
1330
1331 func (runner *ContainerRunner) updateRuntimeStatus(status arvadosclient.Dict) {
1332         err := runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1333                 "select": []string{"uuid"},
1334                 "container": arvadosclient.Dict{
1335                         "runtime_status": status,
1336                 },
1337         }, nil)
1338         if err != nil {
1339                 runner.CrunchLog.Printf("error updating container runtime_status: %s", err)
1340         }
1341 }
1342
1343 // CaptureOutput saves data from the container's output directory if
1344 // needed, and updates the container output accordingly.
1345 func (runner *ContainerRunner) CaptureOutput(bindmounts map[string]bindmount) error {
1346         if runner.Container.RuntimeConstraints.API {
1347                 // Output may have been set directly by the container, so
1348                 // refresh the container record to check.
1349                 err := runner.DispatcherArvClient.Get("containers", runner.Container.UUID,
1350                         arvadosclient.Dict{
1351                                 "select": []string{"output"},
1352                         }, &runner.Container)
1353                 if err != nil {
1354                         return err
1355                 }
1356                 if runner.Container.Output != "" {
1357                         // Container output is already set.
1358                         runner.OutputPDH = &runner.Container.Output
1359                         return nil
1360                 }
1361         }
1362
1363         txt, err := (&copier{
1364                 client:        runner.containerClient,
1365                 keepClient:    runner.ContainerKeepClient,
1366                 hostOutputDir: runner.HostOutputDir,
1367                 ctrOutputDir:  runner.Container.OutputPath,
1368                 globs:         runner.Container.OutputGlob,
1369                 bindmounts:    bindmounts,
1370                 mounts:        runner.Container.Mounts,
1371                 secretMounts:  runner.SecretMounts,
1372                 logger:        runner.CrunchLog,
1373         }).Copy()
1374         if err != nil {
1375                 return err
1376         }
1377         if n := len(regexp.MustCompile(` [0-9a-f]+\+\S*\+R`).FindAllStringIndex(txt, -1)); n > 0 {
1378                 runner.CrunchLog.Printf("Copying %d data blocks from remote input collections...", n)
1379                 fs, err := (&arvados.Collection{ManifestText: txt}).FileSystem(runner.containerClient, runner.ContainerKeepClient)
1380                 if err != nil {
1381                         return err
1382                 }
1383                 txt, err = fs.MarshalManifest(".")
1384                 if err != nil {
1385                         return err
1386                 }
1387         }
1388         var resp arvados.Collection
1389         err = runner.ContainerArvClient.Create("collections", arvadosclient.Dict{
1390                 "ensure_unique_name": true,
1391                 "select":             []string{"portable_data_hash"},
1392                 "collection": arvadosclient.Dict{
1393                         "is_trashed":    true,
1394                         "name":          "output for " + runner.Container.UUID,
1395                         "manifest_text": txt,
1396                 },
1397         }, &resp)
1398         if err != nil {
1399                 return fmt.Errorf("error creating output collection: %v", err)
1400         }
1401         runner.OutputPDH = &resp.PortableDataHash
1402         return nil
1403 }
1404
1405 func (runner *ContainerRunner) CleanupDirs() {
1406         if runner.ArvMount != nil {
1407                 var delay int64 = 8
1408                 umount := exec.Command("arv-mount", fmt.Sprintf("--unmount-timeout=%d", delay), "--unmount", runner.ArvMountPoint)
1409                 umount.Stdout = runner.CrunchLog
1410                 umount.Stderr = runner.CrunchLog
1411                 runner.CrunchLog.Printf("Running %v", umount.Args)
1412                 umnterr := umount.Start()
1413
1414                 if umnterr != nil {
1415                         runner.CrunchLog.Printf("Error unmounting: %v", umnterr)
1416                         runner.ArvMount.Process.Kill()
1417                 } else {
1418                         // If arv-mount --unmount gets stuck for any reason, we
1419                         // don't want to wait for it forever.  Do Wait() in a goroutine
1420                         // so it doesn't block crunch-run.
1421                         umountExit := make(chan error)
1422                         go func() {
1423                                 mnterr := umount.Wait()
1424                                 if mnterr != nil {
1425                                         runner.CrunchLog.Printf("Error unmounting: %v", mnterr)
1426                                 }
1427                                 umountExit <- mnterr
1428                         }()
1429
1430                         for again := true; again; {
1431                                 again = false
1432                                 select {
1433                                 case <-umountExit:
1434                                         umount = nil
1435                                         again = true
1436                                 case <-runner.ArvMountExit:
1437                                         break
1438                                 case <-time.After(time.Duration((delay + 1) * int64(time.Second))):
1439                                         runner.CrunchLog.Printf("Timed out waiting for unmount")
1440                                         if umount != nil {
1441                                                 umount.Process.Kill()
1442                                         }
1443                                         runner.ArvMount.Process.Kill()
1444                                 }
1445                         }
1446                 }
1447                 runner.ArvMount = nil
1448         }
1449
1450         if runner.ArvMountPoint != "" {
1451                 if rmerr := os.Remove(runner.ArvMountPoint); rmerr != nil {
1452                         runner.CrunchLog.Printf("While cleaning up arv-mount directory %s: %v", runner.ArvMountPoint, rmerr)
1453                 }
1454                 runner.ArvMountPoint = ""
1455         }
1456
1457         if rmerr := os.RemoveAll(runner.parentTemp); rmerr != nil {
1458                 runner.CrunchLog.Printf("While cleaning up temporary directory %s: %v", runner.parentTemp, rmerr)
1459         }
1460 }
1461
1462 // CommitLogs posts the collection containing the final container logs.
1463 func (runner *ContainerRunner) CommitLogs() error {
1464         func() {
1465                 // Hold cStateLock to prevent races on CrunchLog (e.g., stop()).
1466                 runner.cStateLock.Lock()
1467                 defer runner.cStateLock.Unlock()
1468
1469                 runner.CrunchLog.Print(runner.finalState)
1470
1471                 if runner.arvMountLog != nil {
1472                         runner.arvMountLog.Close()
1473                 }
1474                 runner.CrunchLog.Close()
1475
1476                 // Closing CrunchLog above allows them to be committed to Keep at this
1477                 // point, but re-open crunch log with ArvClient in case there are any
1478                 // other further errors (such as failing to write the log to Keep!)
1479                 // while shutting down
1480                 runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{
1481                         ArvClient:     runner.DispatcherArvClient,
1482                         UUID:          runner.Container.UUID,
1483                         loggingStream: "crunch-run",
1484                         writeCloser:   nil,
1485                 })
1486                 runner.CrunchLog.Immediate = log.New(os.Stderr, runner.Container.UUID+" ", 0)
1487         }()
1488
1489         if runner.keepstoreLogger != nil {
1490                 // Flush any buffered logs from our local keepstore
1491                 // process.  Discard anything logged after this point
1492                 // -- it won't end up in the log collection, so
1493                 // there's no point writing it to the collectionfs.
1494                 runner.keepstoreLogbuf.SetWriter(io.Discard)
1495                 runner.keepstoreLogger.Close()
1496                 runner.keepstoreLogger = nil
1497         }
1498
1499         if runner.LogsPDH != nil {
1500                 // If we have already assigned something to LogsPDH,
1501                 // we must be closing the re-opened log, which won't
1502                 // end up getting attached to the container record and
1503                 // therefore doesn't need to be saved as a collection
1504                 // -- it exists only to send logs to other channels.
1505                 return nil
1506         }
1507
1508         saved, err := runner.saveLogCollection(true)
1509         if err != nil {
1510                 return fmt.Errorf("error saving log collection: %s", err)
1511         }
1512         runner.logMtx.Lock()
1513         defer runner.logMtx.Unlock()
1514         runner.LogsPDH = &saved.PortableDataHash
1515         return nil
1516 }
1517
1518 // Create/update the log collection. Return value has UUID and
1519 // PortableDataHash fields populated, but others may be blank.
1520 func (runner *ContainerRunner) saveLogCollection(final bool) (response arvados.Collection, err error) {
1521         runner.logMtx.Lock()
1522         defer runner.logMtx.Unlock()
1523         if runner.LogsPDH != nil {
1524                 // Already finalized.
1525                 return
1526         }
1527         updates := arvadosclient.Dict{
1528                 "name": "logs for " + runner.Container.UUID,
1529         }
1530         mt, err1 := runner.LogCollection.MarshalManifest(".")
1531         if err1 == nil {
1532                 // Only send updated manifest text if there was no
1533                 // error.
1534                 updates["manifest_text"] = mt
1535         }
1536
1537         // Even if flushing the manifest had an error, we still want
1538         // to update the log record, if possible, to push the trash_at
1539         // and delete_at times into the future.  Details on bug
1540         // #17293.
1541         if final {
1542                 updates["is_trashed"] = true
1543         } else {
1544                 // We set trash_at so this collection gets
1545                 // automatically cleaned up eventually.  It used to be
1546                 // 12 hours but we had a situation where the API
1547                 // server was down over a weekend but the containers
1548                 // kept running such that the log collection got
1549                 // trashed, so now we make it 2 weeks.  refs #20378
1550                 exp := time.Now().Add(time.Duration(24*14) * time.Hour)
1551                 updates["trash_at"] = exp
1552                 updates["delete_at"] = exp
1553         }
1554         reqBody := arvadosclient.Dict{
1555                 "select":     []string{"uuid", "portable_data_hash"},
1556                 "collection": updates,
1557         }
1558         var err2 error
1559         if runner.logUUID == "" {
1560                 reqBody["ensure_unique_name"] = true
1561                 err2 = runner.DispatcherArvClient.Create("collections", reqBody, &response)
1562         } else {
1563                 err2 = runner.DispatcherArvClient.Update("collections", runner.logUUID, reqBody, &response)
1564         }
1565         if err2 == nil {
1566                 runner.logUUID = response.UUID
1567         }
1568
1569         if err1 != nil || err2 != nil {
1570                 err = fmt.Errorf("error recording logs: %q, %q", err1, err2)
1571         }
1572         return
1573 }
1574
1575 // UpdateContainerRunning updates the container state to "Running"
1576 func (runner *ContainerRunner) UpdateContainerRunning(logId string) error {
1577         runner.cStateLock.Lock()
1578         defer runner.cStateLock.Unlock()
1579         if runner.cCancelled {
1580                 return ErrCancelled
1581         }
1582         updates := arvadosclient.Dict{
1583                 "gateway_address": runner.gateway.Address,
1584                 "state":           "Running",
1585         }
1586         if logId != "" {
1587                 updates["log"] = logId
1588         }
1589         return runner.DispatcherArvClient.Update(
1590                 "containers",
1591                 runner.Container.UUID,
1592                 arvadosclient.Dict{
1593                         "select":    []string{"uuid"},
1594                         "container": updates,
1595                 },
1596                 nil,
1597         )
1598 }
1599
1600 // ContainerToken returns the api_token the container (and any
1601 // arv-mount processes) are allowed to use.
1602 func (runner *ContainerRunner) ContainerToken() (string, error) {
1603         if runner.token != "" {
1604                 return runner.token, nil
1605         }
1606
1607         var auth arvados.APIClientAuthorization
1608         err := runner.DispatcherArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth)
1609         if err != nil {
1610                 return "", err
1611         }
1612         runner.token = fmt.Sprintf("v2/%s/%s/%s", auth.UUID, auth.APIToken, runner.Container.UUID)
1613         return runner.token, nil
1614 }
1615
1616 // UpdateContainerFinal updates the container record state on API
1617 // server to "Complete" or "Cancelled"
1618 func (runner *ContainerRunner) UpdateContainerFinal() error {
1619         update := arvadosclient.Dict{}
1620         update["state"] = runner.finalState
1621         if runner.LogsPDH != nil {
1622                 update["log"] = *runner.LogsPDH
1623         }
1624         if runner.ExitCode != nil {
1625                 update["exit_code"] = *runner.ExitCode
1626         } else {
1627                 update["exit_code"] = nil
1628         }
1629         if runner.finalState == "Complete" && runner.OutputPDH != nil {
1630                 update["output"] = *runner.OutputPDH
1631         }
1632         update["cost"] = runner.calculateCost(time.Now())
1633         return runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1634                 "select":    []string{"uuid"},
1635                 "container": update,
1636         }, nil)
1637 }
1638
1639 // IsCancelled returns the value of Cancelled, with goroutine safety.
1640 func (runner *ContainerRunner) IsCancelled() bool {
1641         runner.cStateLock.Lock()
1642         defer runner.cStateLock.Unlock()
1643         return runner.cCancelled
1644 }
1645
1646 // NewArvLogWriter creates an ArvLogWriter
1647 func (runner *ContainerRunner) NewArvLogWriter(name string) (io.WriteCloser, error) {
1648         writer, err := runner.LogCollection.OpenFile(name+".txt", os.O_CREATE|os.O_WRONLY, 0666)
1649         if err != nil {
1650                 return nil, err
1651         }
1652         return &ArvLogWriter{
1653                 ArvClient:     runner.DispatcherArvClient,
1654                 UUID:          runner.Container.UUID,
1655                 loggingStream: name,
1656                 writeCloser:   writer,
1657         }, nil
1658 }
1659
1660 // Run the full container lifecycle.
1661 func (runner *ContainerRunner) Run() (err error) {
1662         runner.CrunchLog.Printf("crunch-run %s started", cmd.Version.String())
1663         runner.CrunchLog.Printf("%s", currentUserAndGroups())
1664         v, _ := exec.Command("arv-mount", "--version").CombinedOutput()
1665         runner.CrunchLog.Printf("Using FUSE mount: %s", v)
1666         runner.CrunchLog.Printf("Using container runtime: %s", runner.executor.Runtime())
1667         runner.CrunchLog.Printf("Executing container: %s", runner.Container.UUID)
1668         runner.costStartTime = time.Now()
1669
1670         hostname, hosterr := os.Hostname()
1671         if hosterr != nil {
1672                 runner.CrunchLog.Printf("Error getting hostname '%v'", hosterr)
1673         } else {
1674                 runner.CrunchLog.Printf("Executing on host '%s'", hostname)
1675         }
1676
1677         sigusr2 := make(chan os.Signal, 1)
1678         signal.Notify(sigusr2, syscall.SIGUSR2)
1679         defer signal.Stop(sigusr2)
1680         runner.loadPrices()
1681         go runner.handleSIGUSR2(sigusr2)
1682
1683         runner.finalState = "Queued"
1684
1685         defer func() {
1686                 runner.CleanupDirs()
1687
1688                 runner.CrunchLog.Printf("crunch-run finished")
1689                 runner.CrunchLog.Close()
1690         }()
1691
1692         err = runner.fetchContainerRecord()
1693         if err != nil {
1694                 return
1695         }
1696         if runner.Container.State != "Locked" {
1697                 return fmt.Errorf("dispatch error detected: container %q has state %q", runner.Container.UUID, runner.Container.State)
1698         }
1699
1700         var bindmounts map[string]bindmount
1701         defer func() {
1702                 // checkErr prints e (unless it's nil) and sets err to
1703                 // e (unless err is already non-nil). Thus, if err
1704                 // hasn't already been assigned when Run() returns,
1705                 // this cleanup func will cause Run() to return the
1706                 // first non-nil error that is passed to checkErr().
1707                 checkErr := func(errorIn string, e error) {
1708                         if e == nil {
1709                                 return
1710                         }
1711                         runner.CrunchLog.Printf("error in %s: %v", errorIn, e)
1712                         if err == nil {
1713                                 err = e
1714                         }
1715                         if runner.finalState == "Complete" {
1716                                 // There was an error in the finalization.
1717                                 runner.finalState = "Cancelled"
1718                         }
1719                 }
1720
1721                 // Log the error encountered in Run(), if any
1722                 checkErr("Run", err)
1723
1724                 if runner.finalState == "Queued" {
1725                         runner.UpdateContainerFinal()
1726                         return
1727                 }
1728
1729                 if runner.IsCancelled() {
1730                         runner.finalState = "Cancelled"
1731                         // but don't return yet -- we still want to
1732                         // capture partial output and write logs
1733                 }
1734
1735                 if bindmounts != nil {
1736                         checkErr("CaptureOutput", runner.CaptureOutput(bindmounts))
1737                 }
1738                 checkErr("stopHoststat", runner.stopHoststat())
1739                 checkErr("CommitLogs", runner.CommitLogs())
1740                 runner.CleanupDirs()
1741                 checkErr("UpdateContainerFinal", runner.UpdateContainerFinal())
1742         }()
1743
1744         runner.setupSignals()
1745         err = runner.startHoststat()
1746         if err != nil {
1747                 return
1748         }
1749         if runner.keepstore != nil {
1750                 runner.hoststatReporter.ReportPID("keepstore", runner.keepstore.Process.Pid)
1751         }
1752
1753         // set up FUSE mount and binds
1754         bindmounts, err = runner.SetupMounts()
1755         if err != nil {
1756                 runner.finalState = "Cancelled"
1757                 err = fmt.Errorf("While setting up mounts: %v", err)
1758                 return
1759         }
1760
1761         // check for and/or load image
1762         imageID, err := runner.LoadImage()
1763         if err != nil {
1764                 if !runner.checkBrokenNode(err) {
1765                         // Failed to load image but not due to a "broken node"
1766                         // condition, probably user error.
1767                         runner.finalState = "Cancelled"
1768                 }
1769                 err = fmt.Errorf("While loading container image: %v", err)
1770                 return
1771         }
1772
1773         err = runner.CreateContainer(imageID, bindmounts)
1774         if err != nil {
1775                 return
1776         }
1777         err = runner.LogHostInfo()
1778         if err != nil {
1779                 return
1780         }
1781         err = runner.LogNodeRecord()
1782         if err != nil {
1783                 return
1784         }
1785         err = runner.LogContainerRecord()
1786         if err != nil {
1787                 return
1788         }
1789
1790         if runner.IsCancelled() {
1791                 return
1792         }
1793
1794         logCollection, err := runner.saveLogCollection(false)
1795         var logId string
1796         if err == nil {
1797                 logId = logCollection.PortableDataHash
1798         } else {
1799                 runner.CrunchLog.Printf("Error committing initial log collection: %v", err)
1800         }
1801         err = runner.UpdateContainerRunning(logId)
1802         if err != nil {
1803                 return
1804         }
1805         runner.finalState = "Cancelled"
1806
1807         err = runner.startCrunchstat()
1808         if err != nil {
1809                 return
1810         }
1811
1812         err = runner.StartContainer()
1813         if err != nil {
1814                 runner.checkBrokenNode(err)
1815                 return
1816         }
1817
1818         err = runner.WaitFinish()
1819         if err == nil && !runner.IsCancelled() {
1820                 runner.finalState = "Complete"
1821         }
1822         return
1823 }
1824
1825 // Fetch the current container record (uuid = runner.Container.UUID)
1826 // into runner.Container.
1827 func (runner *ContainerRunner) fetchContainerRecord() error {
1828         reader, err := runner.DispatcherArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
1829         if err != nil {
1830                 return fmt.Errorf("error fetching container record: %v", err)
1831         }
1832         defer reader.Close()
1833
1834         dec := json.NewDecoder(reader)
1835         dec.UseNumber()
1836         err = dec.Decode(&runner.Container)
1837         if err != nil {
1838                 return fmt.Errorf("error decoding container record: %v", err)
1839         }
1840
1841         var sm struct {
1842                 SecretMounts map[string]arvados.Mount `json:"secret_mounts"`
1843         }
1844
1845         containerToken, err := runner.ContainerToken()
1846         if err != nil {
1847                 return fmt.Errorf("error getting container token: %v", err)
1848         }
1849
1850         runner.ContainerArvClient, runner.ContainerKeepClient,
1851                 runner.containerClient, err = runner.MkArvClient(containerToken)
1852         if err != nil {
1853                 return fmt.Errorf("error creating container API client: %v", err)
1854         }
1855
1856         runner.ContainerKeepClient.SetStorageClasses(runner.Container.OutputStorageClasses)
1857         runner.DispatcherKeepClient.SetStorageClasses(runner.Container.OutputStorageClasses)
1858
1859         err = runner.ContainerArvClient.Call("GET", "containers", runner.Container.UUID, "secret_mounts", nil, &sm)
1860         if err != nil {
1861                 if apierr, ok := err.(arvadosclient.APIServerError); !ok || apierr.HttpStatusCode != 404 {
1862                         return fmt.Errorf("error fetching secret_mounts: %v", err)
1863                 }
1864                 // ok && apierr.HttpStatusCode == 404, which means
1865                 // secret_mounts isn't supported by this API server.
1866         }
1867         runner.SecretMounts = sm.SecretMounts
1868
1869         return nil
1870 }
1871
1872 // NewContainerRunner creates a new container runner.
1873 func NewContainerRunner(dispatcherClient *arvados.Client,
1874         dispatcherArvClient IArvadosClient,
1875         dispatcherKeepClient IKeepClient,
1876         containerUUID string) (*ContainerRunner, error) {
1877
1878         cr := &ContainerRunner{
1879                 dispatcherClient:     dispatcherClient,
1880                 DispatcherArvClient:  dispatcherArvClient,
1881                 DispatcherKeepClient: dispatcherKeepClient,
1882         }
1883         cr.NewLogWriter = cr.NewArvLogWriter
1884         cr.RunArvMount = cr.ArvMountCmd
1885         cr.MkTempDir = ioutil.TempDir
1886         cr.MkArvClient = func(token string) (IArvadosClient, IKeepClient, *arvados.Client, error) {
1887                 cl, err := arvadosclient.MakeArvadosClient()
1888                 if err != nil {
1889                         return nil, nil, nil, err
1890                 }
1891                 cl.ApiToken = token
1892                 kc, err := keepclient.MakeKeepClient(cl)
1893                 if err != nil {
1894                         return nil, nil, nil, err
1895                 }
1896                 c2 := arvados.NewClientFromEnv()
1897                 c2.AuthToken = token
1898                 return cl, kc, c2, nil
1899         }
1900         var err error
1901         cr.LogCollection, err = (&arvados.Collection{}).FileSystem(cr.dispatcherClient, cr.DispatcherKeepClient)
1902         if err != nil {
1903                 return nil, err
1904         }
1905         cr.Container.UUID = containerUUID
1906         w, err := cr.NewLogWriter("crunch-run")
1907         if err != nil {
1908                 return nil, err
1909         }
1910         cr.CrunchLog = NewThrottledLogger(w)
1911         cr.CrunchLog.Immediate = log.New(os.Stderr, containerUUID+" ", 0)
1912
1913         loadLogThrottleParams(dispatcherArvClient)
1914         go cr.updateLogs()
1915
1916         return cr, nil
1917 }
1918
1919 func (command) RunCommand(prog string, args []string, stdin io.Reader, stdout, stderr io.Writer) int {
1920         log := log.New(stderr, "", 0)
1921         flags := flag.NewFlagSet(prog, flag.ContinueOnError)
1922         statInterval := flags.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting")
1923         flags.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree (obsolete, ignored)")
1924         flags.String("cgroup-parent", "docker", "name of container's parent cgroup (obsolete, ignored)")
1925         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")
1926         caCertsPath := flags.String("ca-certs", "", "Path to TLS root certificates")
1927         detach := flags.Bool("detach", false, "Detach from parent process and run in the background")
1928         stdinConfig := flags.Bool("stdin-config", false, "Load config and environment variables from JSON message on stdin")
1929         configFile := flags.String("config", arvados.DefaultConfigFile, "filename of cluster config file to try loading if -stdin-config=false (default is $ARVADOS_CONFIG)")
1930         sleep := flags.Duration("sleep", 0, "Delay before starting (testing use only)")
1931         kill := flags.Int("kill", -1, "Send signal to an existing crunch-run process for given UUID")
1932         list := flags.Bool("list", false, "List UUIDs of existing crunch-run processes (and notify them to use price data passed on stdin)")
1933         enableMemoryLimit := flags.Bool("enable-memory-limit", true, "tell container runtime to limit container's memory usage")
1934         enableNetwork := flags.String("container-enable-networking", "default", "enable networking \"always\" (for all containers) or \"default\" (for containers that request it)")
1935         networkMode := flags.String("container-network-mode", "default", `Docker network mode for container (use any argument valid for docker --net)`)
1936         memprofile := flags.String("memprofile", "", "write memory profile to `file` after running container")
1937         runtimeEngine := flags.String("runtime-engine", "docker", "container runtime: docker or singularity")
1938         brokenNodeHook := flags.String("broken-node-hook", "", "script to run if node is detected to be broken (for example, Docker daemon is not running)")
1939         flags.Duration("check-containerd", 0, "Ignored. Exists for compatibility with older versions.")
1940         version := flags.Bool("version", false, "Write version information to stdout and exit 0.")
1941
1942         ignoreDetachFlag := false
1943         if len(args) > 0 && args[0] == "-no-detach" {
1944                 // This process was invoked by a parent process, which
1945                 // has passed along its own arguments, including
1946                 // -detach, after the leading -no-detach flag.  Strip
1947                 // the leading -no-detach flag (it's not recognized by
1948                 // flags.Parse()) and ignore the -detach flag that
1949                 // comes later.
1950                 args = args[1:]
1951                 ignoreDetachFlag = true
1952         }
1953
1954         if ok, code := cmd.ParseFlags(flags, prog, args, "container-uuid", stderr); !ok {
1955                 return code
1956         } else if *version {
1957                 fmt.Fprintln(stdout, prog, cmd.Version.String())
1958                 return 0
1959         } else if !*list && flags.NArg() != 1 {
1960                 fmt.Fprintf(stderr, "missing required argument: container-uuid (try -help)\n")
1961                 return 2
1962         }
1963
1964         containerUUID := flags.Arg(0)
1965
1966         switch {
1967         case *detach && !ignoreDetachFlag:
1968                 return Detach(containerUUID, prog, args, stdin, stdout, stderr)
1969         case *kill >= 0:
1970                 return KillProcess(containerUUID, syscall.Signal(*kill), stdout, stderr)
1971         case *list:
1972                 return ListProcesses(stdin, stdout, stderr)
1973         }
1974
1975         if len(containerUUID) != 27 {
1976                 log.Printf("usage: %s [options] UUID", prog)
1977                 return 1
1978         }
1979
1980         var keepstoreLogbuf bufThenWrite
1981         var conf ConfigData
1982         if *stdinConfig {
1983                 err := json.NewDecoder(stdin).Decode(&conf)
1984                 if err != nil {
1985                         log.Printf("decode stdin: %s", err)
1986                         return 1
1987                 }
1988                 for k, v := range conf.Env {
1989                         err = os.Setenv(k, v)
1990                         if err != nil {
1991                                 log.Printf("setenv(%q): %s", k, err)
1992                                 return 1
1993                         }
1994                 }
1995                 if conf.Cluster != nil {
1996                         // ClusterID is missing from the JSON
1997                         // representation, but we need it to generate
1998                         // a valid config file for keepstore, so we
1999                         // fill it using the container UUID prefix.
2000                         conf.Cluster.ClusterID = containerUUID[:5]
2001                 }
2002         } else {
2003                 conf = hpcConfData(containerUUID, *configFile, io.MultiWriter(&keepstoreLogbuf, stderr))
2004         }
2005
2006         log.Printf("crunch-run %s started", cmd.Version.String())
2007         time.Sleep(*sleep)
2008
2009         if *caCertsPath != "" {
2010                 os.Setenv("SSL_CERT_FILE", *caCertsPath)
2011         }
2012
2013         keepstore, err := startLocalKeepstore(conf, io.MultiWriter(&keepstoreLogbuf, stderr))
2014         if err != nil {
2015                 log.Print(err)
2016                 return 1
2017         }
2018         if keepstore != nil {
2019                 defer keepstore.Process.Kill()
2020         }
2021
2022         api, err := arvadosclient.MakeArvadosClient()
2023         if err != nil {
2024                 log.Printf("%s: %v", containerUUID, err)
2025                 return 1
2026         }
2027         // arvadosclient now interprets Retries=10 to mean
2028         // Timeout=10m, retrying with exponential backoff + jitter.
2029         api.Retries = 10
2030
2031         kc, err := keepclient.MakeKeepClient(api)
2032         if err != nil {
2033                 log.Printf("%s: %v", containerUUID, err)
2034                 return 1
2035         }
2036         kc.Retries = 4
2037
2038         cr, err := NewContainerRunner(arvados.NewClientFromEnv(), api, kc, containerUUID)
2039         if err != nil {
2040                 log.Print(err)
2041                 return 1
2042         }
2043
2044         cr.keepstore = keepstore
2045         if keepstore == nil {
2046                 // Log explanation (if any) for why we're not running
2047                 // a local keepstore.
2048                 var buf bytes.Buffer
2049                 keepstoreLogbuf.SetWriter(&buf)
2050                 if buf.Len() > 0 {
2051                         cr.CrunchLog.Printf("%s", strings.TrimSpace(buf.String()))
2052                 }
2053         } else if logWhat := conf.Cluster.Containers.LocalKeepLogsToContainerLog; logWhat == "none" {
2054                 cr.CrunchLog.Printf("using local keepstore process (pid %d) at %s", keepstore.Process.Pid, os.Getenv("ARVADOS_KEEP_SERVICES"))
2055                 keepstoreLogbuf.SetWriter(io.Discard)
2056         } else {
2057                 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"))
2058                 logwriter, err := cr.NewLogWriter("keepstore")
2059                 if err != nil {
2060                         log.Print(err)
2061                         return 1
2062                 }
2063                 cr.keepstoreLogger = NewThrottledLogger(logwriter)
2064
2065                 var writer io.WriteCloser = cr.keepstoreLogger
2066                 if logWhat == "errors" {
2067                         writer = &filterKeepstoreErrorsOnly{WriteCloser: writer}
2068                 } else if logWhat != "all" {
2069                         // should have been caught earlier by
2070                         // dispatcher's config loader
2071                         log.Printf("invalid value for Containers.LocalKeepLogsToContainerLog: %q", logWhat)
2072                         return 1
2073                 }
2074                 err = keepstoreLogbuf.SetWriter(writer)
2075                 if err != nil {
2076                         log.Print(err)
2077                         return 1
2078                 }
2079                 cr.keepstoreLogbuf = &keepstoreLogbuf
2080         }
2081
2082         switch *runtimeEngine {
2083         case "docker":
2084                 cr.executor, err = newDockerExecutor(containerUUID, cr.CrunchLog.Printf, cr.containerWatchdogInterval)
2085         case "singularity":
2086                 cr.executor, err = newSingularityExecutor(cr.CrunchLog.Printf)
2087         default:
2088                 cr.CrunchLog.Printf("%s: unsupported RuntimeEngine %q", containerUUID, *runtimeEngine)
2089                 cr.CrunchLog.Close()
2090                 return 1
2091         }
2092         if err != nil {
2093                 cr.CrunchLog.Printf("%s: %v", containerUUID, err)
2094                 cr.checkBrokenNode(err)
2095                 cr.CrunchLog.Close()
2096                 return 1
2097         }
2098         defer cr.executor.Close()
2099
2100         cr.brokenNodeHook = *brokenNodeHook
2101
2102         gwAuthSecret := os.Getenv("GatewayAuthSecret")
2103         os.Unsetenv("GatewayAuthSecret")
2104         if gwAuthSecret == "" {
2105                 // not safe to run a gateway service without an auth
2106                 // secret
2107                 cr.CrunchLog.Printf("Not starting a gateway server (GatewayAuthSecret was not provided by dispatcher)")
2108         } else {
2109                 gwListen := os.Getenv("GatewayAddress")
2110                 cr.gateway = Gateway{
2111                         Address:       gwListen,
2112                         AuthSecret:    gwAuthSecret,
2113                         ContainerUUID: containerUUID,
2114                         Target:        cr.executor,
2115                         Log:           cr.CrunchLog,
2116                         LogCollection: cr.LogCollection,
2117                 }
2118                 if gwListen == "" {
2119                         // Direct connection won't work, so we use the
2120                         // gateway_address field to indicate the
2121                         // internalURL of the controller process that
2122                         // has the current tunnel connection.
2123                         cr.gateway.ArvadosClient = cr.dispatcherClient
2124                         cr.gateway.UpdateTunnelURL = func(url string) {
2125                                 cr.gateway.Address = "tunnel " + url
2126                                 cr.DispatcherArvClient.Update("containers", containerUUID,
2127                                         arvadosclient.Dict{
2128                                                 "select":    []string{"uuid"},
2129                                                 "container": arvadosclient.Dict{"gateway_address": cr.gateway.Address},
2130                                         }, nil)
2131                         }
2132                 }
2133                 err = cr.gateway.Start()
2134                 if err != nil {
2135                         log.Printf("error starting gateway server: %s", err)
2136                         return 1
2137                 }
2138         }
2139
2140         parentTemp, tmperr := cr.MkTempDir("", "crunch-run."+containerUUID+".")
2141         if tmperr != nil {
2142                 log.Printf("%s: %v", containerUUID, tmperr)
2143                 return 1
2144         }
2145
2146         cr.parentTemp = parentTemp
2147         cr.statInterval = *statInterval
2148         cr.enableMemoryLimit = *enableMemoryLimit
2149         cr.enableNetwork = *enableNetwork
2150         cr.networkMode = *networkMode
2151         if *cgroupParentSubsystem != "" {
2152                 p, err := findCgroup(os.DirFS("/"), *cgroupParentSubsystem)
2153                 if err != nil {
2154                         log.Printf("fatal: cgroup parent subsystem: %s", err)
2155                         return 1
2156                 }
2157                 cr.setCgroupParent = p
2158         }
2159
2160         if conf.EC2SpotCheck {
2161                 go cr.checkSpotInterruptionNotices()
2162         }
2163
2164         runerr := cr.Run()
2165
2166         if *memprofile != "" {
2167                 f, err := os.Create(*memprofile)
2168                 if err != nil {
2169                         log.Printf("could not create memory profile: %s", err)
2170                 }
2171                 runtime.GC() // get up-to-date statistics
2172                 if err := pprof.WriteHeapProfile(f); err != nil {
2173                         log.Printf("could not write memory profile: %s", err)
2174                 }
2175                 closeerr := f.Close()
2176                 if closeerr != nil {
2177                         log.Printf("closing memprofile file: %s", err)
2178                 }
2179         }
2180
2181         if runerr != nil {
2182                 log.Printf("%s: %v", containerUUID, runerr)
2183                 return 1
2184         }
2185         return 0
2186 }
2187
2188 // Try to load ConfigData in hpc (slurm/lsf) environment. This means
2189 // loading the cluster config from the specified file and (if that
2190 // works) getting the runtime_constraints container field from
2191 // controller to determine # VCPUs so we can calculate KeepBuffers.
2192 func hpcConfData(uuid string, configFile string, stderr io.Writer) ConfigData {
2193         var conf ConfigData
2194         conf.Cluster = loadClusterConfigFile(configFile, stderr)
2195         if conf.Cluster == nil {
2196                 // skip loading the container record -- we won't be
2197                 // able to start local keepstore anyway.
2198                 return conf
2199         }
2200         arv, err := arvadosclient.MakeArvadosClient()
2201         if err != nil {
2202                 fmt.Fprintf(stderr, "error setting up arvadosclient: %s\n", err)
2203                 return conf
2204         }
2205         // arvadosclient now interprets Retries=10 to mean
2206         // Timeout=10m, retrying with exponential backoff + jitter.
2207         arv.Retries = 10
2208         var ctr arvados.Container
2209         err = arv.Call("GET", "containers", uuid, "", arvadosclient.Dict{"select": []string{"runtime_constraints"}}, &ctr)
2210         if err != nil {
2211                 fmt.Fprintf(stderr, "error getting container record: %s\n", err)
2212                 return conf
2213         }
2214         if ctr.RuntimeConstraints.VCPUs > 0 {
2215                 conf.KeepBuffers = ctr.RuntimeConstraints.VCPUs * conf.Cluster.Containers.LocalKeepBlobBuffersPerVCPU
2216         }
2217         return conf
2218 }
2219
2220 // Load cluster config file from given path. If an error occurs, log
2221 // the error to stderr and return nil.
2222 func loadClusterConfigFile(path string, stderr io.Writer) *arvados.Cluster {
2223         ldr := config.NewLoader(&bytes.Buffer{}, ctxlog.New(stderr, "plain", "info"))
2224         ldr.Path = path
2225         cfg, err := ldr.Load()
2226         if err != nil {
2227                 fmt.Fprintf(stderr, "could not load config file %s: %s\n", path, err)
2228                 return nil
2229         }
2230         cluster, err := cfg.GetCluster("")
2231         if err != nil {
2232                 fmt.Fprintf(stderr, "could not use config file %s: %s\n", path, err)
2233                 return nil
2234         }
2235         fmt.Fprintf(stderr, "loaded config file %s\n", path)
2236         return cluster
2237 }
2238
2239 func startLocalKeepstore(configData ConfigData, logbuf io.Writer) (*exec.Cmd, error) {
2240         if configData.KeepBuffers < 1 {
2241                 fmt.Fprintf(logbuf, "not starting a local keepstore process because KeepBuffers=%v in config\n", configData.KeepBuffers)
2242                 return nil, nil
2243         }
2244         if configData.Cluster == nil {
2245                 fmt.Fprint(logbuf, "not starting a local keepstore process because cluster config file was not loaded\n")
2246                 return nil, nil
2247         }
2248         for uuid, vol := range configData.Cluster.Volumes {
2249                 if len(vol.AccessViaHosts) > 0 {
2250                         fmt.Fprintf(logbuf, "not starting a local keepstore process because a volume (%s) uses AccessViaHosts\n", uuid)
2251                         return nil, nil
2252                 }
2253                 if !vol.ReadOnly && vol.Replication < configData.Cluster.Collections.DefaultReplication {
2254                         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)
2255                         return nil, nil
2256                 }
2257         }
2258
2259         // Rather than have an alternate way to tell keepstore how
2260         // many buffers to use when starting it this way, we just
2261         // modify the cluster configuration that we feed it on stdin.
2262         configData.Cluster.API.MaxKeepBlobBuffers = configData.KeepBuffers
2263
2264         localaddr := localKeepstoreAddr()
2265         ln, err := net.Listen("tcp", net.JoinHostPort(localaddr, "0"))
2266         if err != nil {
2267                 return nil, err
2268         }
2269         _, port, err := net.SplitHostPort(ln.Addr().String())
2270         if err != nil {
2271                 ln.Close()
2272                 return nil, err
2273         }
2274         ln.Close()
2275         url := "http://" + net.JoinHostPort(localaddr, port)
2276
2277         fmt.Fprintf(logbuf, "starting keepstore on %s\n", url)
2278
2279         var confJSON bytes.Buffer
2280         err = json.NewEncoder(&confJSON).Encode(arvados.Config{
2281                 Clusters: map[string]arvados.Cluster{
2282                         configData.Cluster.ClusterID: *configData.Cluster,
2283                 },
2284         })
2285         if err != nil {
2286                 return nil, err
2287         }
2288         cmd := exec.Command("/proc/self/exe", "keepstore", "-config=-")
2289         if target, err := os.Readlink(cmd.Path); err == nil && strings.HasSuffix(target, ".test") {
2290                 // If we're a 'go test' process, running
2291                 // /proc/self/exe would start the test suite in a
2292                 // child process, which is not what we want.
2293                 cmd.Path, _ = exec.LookPath("go")
2294                 cmd.Args = append([]string{"go", "run", "../../cmd/arvados-server"}, cmd.Args[1:]...)
2295                 cmd.Env = os.Environ()
2296         }
2297         cmd.Stdin = &confJSON
2298         cmd.Stdout = logbuf
2299         cmd.Stderr = logbuf
2300         cmd.Env = append(cmd.Env,
2301                 "GOGC=10",
2302                 "ARVADOS_SERVICE_INTERNAL_URL="+url)
2303         err = cmd.Start()
2304         if err != nil {
2305                 return nil, fmt.Errorf("error starting keepstore process: %w", err)
2306         }
2307         cmdExited := false
2308         go func() {
2309                 cmd.Wait()
2310                 cmdExited = true
2311         }()
2312         ctx, cancel := context.WithDeadline(context.Background(), time.Now().Add(time.Second*10))
2313         defer cancel()
2314         poll := time.NewTicker(time.Second / 10)
2315         defer poll.Stop()
2316         client := http.Client{}
2317         for range poll.C {
2318                 testReq, err := http.NewRequestWithContext(ctx, "GET", url+"/_health/ping", nil)
2319                 testReq.Header.Set("Authorization", "Bearer "+configData.Cluster.ManagementToken)
2320                 if err != nil {
2321                         return nil, err
2322                 }
2323                 resp, err := client.Do(testReq)
2324                 if err == nil {
2325                         resp.Body.Close()
2326                         if resp.StatusCode == http.StatusOK {
2327                                 break
2328                         }
2329                 }
2330                 if cmdExited {
2331                         return nil, fmt.Errorf("keepstore child process exited")
2332                 }
2333                 if ctx.Err() != nil {
2334                         return nil, fmt.Errorf("timed out waiting for new keepstore process to report healthy")
2335                 }
2336         }
2337         os.Setenv("ARVADOS_KEEP_SERVICES", url)
2338         return cmd, nil
2339 }
2340
2341 // return current uid, gid, groups in a format suitable for logging:
2342 // "crunch-run process has uid=1234(arvados) gid=1234(arvados)
2343 // groups=1234(arvados),114(fuse)"
2344 func currentUserAndGroups() string {
2345         u, err := user.Current()
2346         if err != nil {
2347                 return fmt.Sprintf("error getting current user ID: %s", err)
2348         }
2349         s := fmt.Sprintf("crunch-run process has uid=%s(%s) gid=%s", u.Uid, u.Username, u.Gid)
2350         if g, err := user.LookupGroupId(u.Gid); err == nil {
2351                 s += fmt.Sprintf("(%s)", g.Name)
2352         }
2353         s += " groups="
2354         if gids, err := u.GroupIds(); err == nil {
2355                 for i, gid := range gids {
2356                         if i > 0 {
2357                                 s += ","
2358                         }
2359                         s += gid
2360                         if g, err := user.LookupGroupId(gid); err == nil {
2361                                 s += fmt.Sprintf("(%s)", g.Name)
2362                         }
2363                 }
2364         }
2365         return s
2366 }
2367
2368 // Return a suitable local interface address for a local keepstore
2369 // service. Currently this is the numerically lowest non-loopback ipv4
2370 // address assigned to a local interface that is not in any of the
2371 // link-local/vpn/loopback ranges 169.254/16, 100.64/10, or 127/8.
2372 func localKeepstoreAddr() string {
2373         var ips []net.IP
2374         // Ignore error (proceed with zero IPs)
2375         addrs, _ := processIPs(os.Getpid())
2376         for addr := range addrs {
2377                 ip := net.ParseIP(addr)
2378                 if ip == nil {
2379                         // invalid
2380                         continue
2381                 }
2382                 if ip.Mask(net.CIDRMask(8, 32)).Equal(net.IPv4(127, 0, 0, 0)) ||
2383                         ip.Mask(net.CIDRMask(10, 32)).Equal(net.IPv4(100, 64, 0, 0)) ||
2384                         ip.Mask(net.CIDRMask(16, 32)).Equal(net.IPv4(169, 254, 0, 0)) {
2385                         // unsuitable
2386                         continue
2387                 }
2388                 ips = append(ips, ip)
2389         }
2390         if len(ips) == 0 {
2391                 return "0.0.0.0"
2392         }
2393         sort.Slice(ips, func(ii, jj int) bool {
2394                 i, j := ips[ii], ips[jj]
2395                 if len(i) != len(j) {
2396                         return len(i) < len(j)
2397                 }
2398                 for x := range i {
2399                         if i[x] != j[x] {
2400                                 return i[x] < j[x]
2401                         }
2402                 }
2403                 return false
2404         })
2405         return ips[0].String()
2406 }
2407
2408 func (cr *ContainerRunner) loadPrices() {
2409         buf, err := os.ReadFile(filepath.Join(lockdir, pricesfile))
2410         if err != nil {
2411                 if !os.IsNotExist(err) {
2412                         cr.CrunchLog.Printf("loadPrices: read: %s", err)
2413                 }
2414                 return
2415         }
2416         var prices []cloud.InstancePrice
2417         err = json.Unmarshal(buf, &prices)
2418         if err != nil {
2419                 cr.CrunchLog.Printf("loadPrices: decode: %s", err)
2420                 return
2421         }
2422         cr.pricesLock.Lock()
2423         defer cr.pricesLock.Unlock()
2424         var lastKnown time.Time
2425         if len(cr.prices) > 0 {
2426                 lastKnown = cr.prices[0].StartTime
2427         }
2428         cr.prices = cloud.NormalizePriceHistory(append(prices, cr.prices...))
2429         for i := len(cr.prices) - 1; i >= 0; i-- {
2430                 price := cr.prices[i]
2431                 if price.StartTime.After(lastKnown) {
2432                         cr.CrunchLog.Printf("Instance price changed to %#.3g at %s", price.Price, price.StartTime.UTC())
2433                 }
2434         }
2435 }
2436
2437 func (cr *ContainerRunner) calculateCost(now time.Time) float64 {
2438         cr.pricesLock.Lock()
2439         defer cr.pricesLock.Unlock()
2440
2441         // First, make a "prices" slice with the real data as far back
2442         // as it goes, and (if needed) a "since the beginning of time"
2443         // placeholder containing a reasonable guess about what the
2444         // price was between cr.costStartTime and the earliest real
2445         // data point.
2446         prices := cr.prices
2447         if len(prices) == 0 {
2448                 // use price info in InstanceType record initially
2449                 // provided by cloud dispatcher
2450                 var p float64
2451                 var it arvados.InstanceType
2452                 if j := os.Getenv("InstanceType"); j != "" && json.Unmarshal([]byte(j), &it) == nil && it.Price > 0 {
2453                         p = it.Price
2454                 }
2455                 prices = []cloud.InstancePrice{{Price: p}}
2456         } else if prices[len(prices)-1].StartTime.After(cr.costStartTime) {
2457                 // guess earlier pricing was the same as the earliest
2458                 // price we know about
2459                 filler := prices[len(prices)-1]
2460                 filler.StartTime = time.Time{}
2461                 prices = append(prices, filler)
2462         }
2463
2464         // Now that our history of price changes goes back at least as
2465         // far as cr.costStartTime, add up the costs for each
2466         // interval.
2467         cost := 0.0
2468         spanEnd := now
2469         for _, ip := range prices {
2470                 spanStart := ip.StartTime
2471                 if spanStart.After(now) {
2472                         // pricing information from the future -- not
2473                         // expected from AWS, but possible in
2474                         // principle, and exercised by tests.
2475                         continue
2476                 }
2477                 last := false
2478                 if spanStart.Before(cr.costStartTime) {
2479                         spanStart = cr.costStartTime
2480                         last = true
2481                 }
2482                 cost += ip.Price * spanEnd.Sub(spanStart).Seconds() / 3600
2483                 if last {
2484                         break
2485                 }
2486                 spanEnd = spanStart
2487         }
2488
2489         return cost
2490 }
2491
2492 func (runner *ContainerRunner) handleSIGUSR2(sigchan chan os.Signal) {
2493         for range sigchan {
2494                 runner.loadPrices()
2495                 update := arvadosclient.Dict{
2496                         "select": []string{"uuid"},
2497                         "container": arvadosclient.Dict{
2498                                 "cost": runner.calculateCost(time.Now()),
2499                         },
2500                 }
2501                 runner.DispatcherArvClient.Update("containers", runner.Container.UUID, update, nil)
2502         }
2503 }