services/crunch-run/crunchrun.go

   1 // Copyright (C) The Arvados Authors. All rights reserved.
   2 //
   3 // SPDX-License-Identifier: AGPL-3.0
   4
   5 package main
   6
   7 import (
   8         "bytes"
   9         "encoding/json"
  10         "errors"
  11         "flag"
  12         "fmt"
  13         "io"
  14         "io/ioutil"
  15         "log"
  16         "os"
  17         "os/exec"
  18         "os/signal"
  19         "path"
  20         "path/filepath"
  21         "regexp"
  22         "runtime"
  23         "runtime/pprof"
  24         "sort"
  25         "strings"
  26         "sync"
  27         "syscall"
  28         "time"
  29
  30         "git.curoverse.com/arvados.git/lib/crunchstat"
  31         "git.curoverse.com/arvados.git/sdk/go/arvados"
  32         "git.curoverse.com/arvados.git/sdk/go/arvadosclient"
  33         "git.curoverse.com/arvados.git/sdk/go/keepclient"
  34         "git.curoverse.com/arvados.git/sdk/go/manifest"
  35         "golang.org/x/net/context"
  36
  37         dockertypes "github.com/docker/docker/api/types"
  38         dockercontainer "github.com/docker/docker/api/types/container"
  39         dockernetwork "github.com/docker/docker/api/types/network"
  40         dockerclient "github.com/docker/docker/client"
  41 )
  42
  43 var version = "dev"
  44
  45 // IArvadosClient is the minimal Arvados API methods used by crunch-run.
  46 type IArvadosClient interface {
  47         Create(resourceType string, parameters arvadosclient.Dict, output interface{}) error
  48         Get(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  49         Update(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  50         Call(method, resourceType, uuid, action string, parameters arvadosclient.Dict, output interface{}) error
  51         CallRaw(method string, resourceType string, uuid string, action string, parameters arvadosclient.Dict) (reader io.ReadCloser, err error)
  52         Discovery(key string) (interface{}, error)
  53 }
  54
  55 // ErrCancelled is the error returned when the container is cancelled.
  56 var ErrCancelled = errors.New("Cancelled")
  57
  58 // IKeepClient is the minimal Keep API methods used by crunch-run.
  59 type IKeepClient interface {
  60         PutB(buf []byte) (string, int, error)
  61         ReadAt(locator string, p []byte, off int) (int, error)
  62         ManifestFileReader(m manifest.Manifest, filename string) (arvados.File, error)
  63         ClearBlockCache()
  64 }
  65
  66 // NewLogWriter is a factory function to create a new log writer.
  67 type NewLogWriter func(name string) (io.WriteCloser, error)
  68
  69 type RunArvMount func(args []string, tok string) (*exec.Cmd, error)
  70
  71 type MkTempDir func(string, string) (string, error)
  72
  73 // ThinDockerClient is the minimal Docker client interface used by crunch-run.
  74 type ThinDockerClient interface {
  75         ContainerAttach(ctx context.Context, container string, options dockertypes.ContainerAttachOptions) (dockertypes.HijackedResponse, error)
  76         ContainerCreate(ctx context.Context, config *dockercontainer.Config, hostConfig *dockercontainer.HostConfig,
  77                 networkingConfig *dockernetwork.NetworkingConfig, containerName string) (dockercontainer.ContainerCreateCreatedBody, error)
  78         ContainerStart(ctx context.Context, container string, options dockertypes.ContainerStartOptions) error
  79         ContainerRemove(ctx context.Context, container string, options dockertypes.ContainerRemoveOptions) error
  80         ContainerWait(ctx context.Context, container string, condition dockercontainer.WaitCondition) (<-chan dockercontainer.ContainerWaitOKBody, <-chan error)
  81         ImageInspectWithRaw(ctx context.Context, image string) (dockertypes.ImageInspect, []byte, error)
  82         ImageLoad(ctx context.Context, input io.Reader, quiet bool) (dockertypes.ImageLoadResponse, error)
  83         ImageRemove(ctx context.Context, image string, options dockertypes.ImageRemoveOptions) ([]dockertypes.ImageDeleteResponseItem, error)
  84 }
  85
  86 // ContainerRunner is the main stateful struct used for a single execution of a
  87 // container.
  88 type ContainerRunner struct {
  89         Docker    ThinDockerClient
  90         client    *arvados.Client
  91         ArvClient IArvadosClient
  92         Kc        IKeepClient
  93         arvados.Container
  94         ContainerConfig dockercontainer.Config
  95         dockercontainer.HostConfig
  96         token       string
  97         ContainerID string
  98         ExitCode    *int
  99         NewLogWriter
 100         loggingDone   chan bool
 101         CrunchLog     *ThrottledLogger
 102         Stdout        io.WriteCloser
 103         Stderr        io.WriteCloser
 104         LogCollection arvados.CollectionFileSystem
 105         LogsPDH       *string
 106         RunArvMount
 107         MkTempDir
 108         ArvMount      *exec.Cmd
 109         ArvMountPoint string
 110         HostOutputDir string
 111         Binds         []string
 112         Volumes       map[string]struct{}
 113         OutputPDH     *string
 114         SigChan       chan os.Signal
 115         ArvMountExit  chan error
 116         SecretMounts  map[string]arvados.Mount
 117         MkArvClient   func(token string) (IArvadosClient, error)
 118         finalState    string
 119         parentTemp    string
 120
 121         statLogger       io.WriteCloser
 122         statReporter     *crunchstat.Reporter
 123         hoststatLogger   io.WriteCloser
 124         hoststatReporter *crunchstat.Reporter
 125         statInterval     time.Duration
 126         cgroupRoot       string
 127         // What we expect the container's cgroup parent to be.
 128         expectCgroupParent string
 129         // What we tell docker to use as the container's cgroup
 130         // parent. Note: Ideally we would use the same field for both
 131         // expectCgroupParent and setCgroupParent, and just make it
 132         // default to "docker". However, when using docker < 1.10 with
 133         // systemd, specifying a non-empty cgroup parent (even the
 134         // default value "docker") hits a docker bug
 135         // (https://github.com/docker/docker/issues/17126). Using two
 136         // separate fields makes it possible to use the "expect cgroup
 137         // parent to be X" feature even on sites where the "specify
 138         // cgroup parent" feature breaks.
 139         setCgroupParent string
 140
 141         cStateLock sync.Mutex
 142         cCancelled bool // StopContainer() invoked
 143
 144         enableNetwork string // one of "default" or "always"
 145         networkMode   string // passed through to HostConfig.NetworkMode
 146         arvMountLog   *ThrottledLogger
 147 }
 148
 149 // setupSignals sets up signal handling to gracefully terminate the underlying
 150 // Docker container and update state when receiving a TERM, INT or QUIT signal.
 151 func (runner *ContainerRunner) setupSignals() {
 152         runner.SigChan = make(chan os.Signal, 1)
 153         signal.Notify(runner.SigChan, syscall.SIGTERM)
 154         signal.Notify(runner.SigChan, syscall.SIGINT)
 155         signal.Notify(runner.SigChan, syscall.SIGQUIT)
 156
 157         go func(sig chan os.Signal) {
 158                 for s := range sig {
 159                         runner.stop(s)
 160                 }
 161         }(runner.SigChan)
 162 }
 163
 164 // stop the underlying Docker container.
 165 func (runner *ContainerRunner) stop(sig os.Signal) {
 166         runner.cStateLock.Lock()
 167         defer runner.cStateLock.Unlock()
 168         if sig != nil {
 169                 runner.CrunchLog.Printf("caught signal: %v", sig)
 170         }
 171         if runner.ContainerID == "" {
 172                 return
 173         }
 174         runner.cCancelled = true
 175         runner.CrunchLog.Printf("removing container")
 176         err := runner.Docker.ContainerRemove(context.TODO(), runner.ContainerID, dockertypes.ContainerRemoveOptions{Force: true})
 177         if err != nil {
 178                 runner.CrunchLog.Printf("error removing container: %s", err)
 179         }
 180 }
 181
 182 var errorBlacklist = []string{
 183         "(?ms).*[Cc]annot connect to the Docker daemon.*",
 184         "(?ms).*oci runtime error.*starting container process.*container init.*mounting.*to rootfs.*no such file or directory.*",
 185 }
 186 var brokenNodeHook *string = flag.String("broken-node-hook", "", "Script to run if node is detected to be broken (for example, Docker daemon is not running)")
 187
 188 func (runner *ContainerRunner) checkBrokenNode(goterr error) bool {
 189         for _, d := range errorBlacklist {
 190                 if m, e := regexp.MatchString(d, goterr.Error()); m && e == nil {
 191                         runner.CrunchLog.Printf("Error suggests node is unable to run containers: %v", goterr)
 192                         if *brokenNodeHook == "" {
 193                                 runner.CrunchLog.Printf("No broken node hook provided, cannot mark node as broken.")
 194                         } else {
 195                                 runner.CrunchLog.Printf("Running broken node hook %q", *brokenNodeHook)
 196                                 // run killme script
 197                                 c := exec.Command(*brokenNodeHook)
 198                                 c.Stdout = runner.CrunchLog
 199                                 c.Stderr = runner.CrunchLog
 200                                 err := c.Run()
 201                                 if err != nil {
 202                                         runner.CrunchLog.Printf("Error running broken node hook: %v", err)
 203                                 }
 204                         }
 205                         return true
 206                 }
 207         }
 208         return false
 209 }
 210
 211 // LoadImage determines the docker image id from the container record and
 212 // checks if it is available in the local Docker image store.  If not, it loads
 213 // the image from Keep.
 214 func (runner *ContainerRunner) LoadImage() (err error) {
 215
 216         runner.CrunchLog.Printf("Fetching Docker image from collection '%s'", runner.Container.ContainerImage)
 217
 218         var collection arvados.Collection
 219         err = runner.ArvClient.Get("collections", runner.Container.ContainerImage, nil, &collection)
 220         if err != nil {
 221                 return fmt.Errorf("While getting container image collection: %v", err)
 222         }
 223         manifest := manifest.Manifest{Text: collection.ManifestText}
 224         var img, imageID string
 225         for ms := range manifest.StreamIter() {
 226                 img = ms.FileStreamSegments[0].Name
 227                 if !strings.HasSuffix(img, ".tar") {
 228                         return fmt.Errorf("First file in the container image collection does not end in .tar")
 229                 }
 230                 imageID = img[:len(img)-4]
 231         }
 232
 233         runner.CrunchLog.Printf("Using Docker image id '%s'", imageID)
 234
 235         _, _, err = runner.Docker.ImageInspectWithRaw(context.TODO(), imageID)
 236         if err != nil {
 237                 runner.CrunchLog.Print("Loading Docker image from keep")
 238
 239                 var readCloser io.ReadCloser
 240                 readCloser, err = runner.Kc.ManifestFileReader(manifest, img)
 241                 if err != nil {
 242                         return fmt.Errorf("While creating ManifestFileReader for container image: %v", err)
 243                 }
 244
 245                 response, err := runner.Docker.ImageLoad(context.TODO(), readCloser, true)
 246                 if err != nil {
 247                         return fmt.Errorf("While loading container image into Docker: %v", err)
 248                 }
 249
 250                 defer response.Body.Close()
 251                 rbody, err := ioutil.ReadAll(response.Body)
 252                 if err != nil {
 253                         return fmt.Errorf("Reading response to image load: %v", err)
 254                 }
 255                 runner.CrunchLog.Printf("Docker response: %s", rbody)
 256         } else {
 257                 runner.CrunchLog.Print("Docker image is available")
 258         }
 259
 260         runner.ContainerConfig.Image = imageID
 261
 262         runner.Kc.ClearBlockCache()
 263
 264         return nil
 265 }
 266
 267 func (runner *ContainerRunner) ArvMountCmd(arvMountCmd []string, token string) (c *exec.Cmd, err error) {
 268         c = exec.Command("arv-mount", arvMountCmd...)
 269
 270         // Copy our environment, but override ARVADOS_API_TOKEN with
 271         // the container auth token.
 272         c.Env = nil
 273         for _, s := range os.Environ() {
 274                 if !strings.HasPrefix(s, "ARVADOS_API_TOKEN=") {
 275                         c.Env = append(c.Env, s)
 276                 }
 277         }
 278         c.Env = append(c.Env, "ARVADOS_API_TOKEN="+token)
 279
 280         w, err := runner.NewLogWriter("arv-mount")
 281         if err != nil {
 282                 return nil, err
 283         }
 284         runner.arvMountLog = NewThrottledLogger(w)
 285         c.Stdout = runner.arvMountLog
 286         c.Stderr = runner.arvMountLog
 287
 288         runner.CrunchLog.Printf("Running %v", c.Args)
 289
 290         err = c.Start()
 291         if err != nil {
 292                 return nil, err
 293         }
 294
 295         statReadme := make(chan bool)
 296         runner.ArvMountExit = make(chan error)
 297
 298         keepStatting := true
 299         go func() {
 300                 for keepStatting {
 301                         time.Sleep(100 * time.Millisecond)
 302                         _, err = os.Stat(fmt.Sprintf("%s/by_id/README", runner.ArvMountPoint))
 303                         if err == nil {
 304                                 keepStatting = false
 305                                 statReadme <- true
 306                         }
 307                 }
 308                 close(statReadme)
 309         }()
 310
 311         go func() {
 312                 mnterr := c.Wait()
 313                 if mnterr != nil {
 314                         runner.CrunchLog.Printf("Arv-mount exit error: %v", mnterr)
 315                 }
 316                 runner.ArvMountExit <- mnterr
 317                 close(runner.ArvMountExit)
 318         }()
 319
 320         select {
 321         case <-statReadme:
 322                 break
 323         case err := <-runner.ArvMountExit:
 324                 runner.ArvMount = nil
 325                 keepStatting = false
 326                 return nil, err
 327         }
 328
 329         return c, nil
 330 }
 331
 332 func (runner *ContainerRunner) SetupArvMountPoint(prefix string) (err error) {
 333         if runner.ArvMountPoint == "" {
 334                 runner.ArvMountPoint, err = runner.MkTempDir(runner.parentTemp, prefix)
 335         }
 336         return
 337 }
 338
 339 func copyfile(src string, dst string) (err error) {
 340         srcfile, err := os.Open(src)
 341         if err != nil {
 342                 return
 343         }
 344
 345         os.MkdirAll(path.Dir(dst), 0777)
 346
 347         dstfile, err := os.Create(dst)
 348         if err != nil {
 349                 return
 350         }
 351         _, err = io.Copy(dstfile, srcfile)
 352         if err != nil {
 353                 return
 354         }
 355
 356         err = srcfile.Close()
 357         err2 := dstfile.Close()
 358
 359         if err != nil {
 360                 return
 361         }
 362
 363         if err2 != nil {
 364                 return err2
 365         }
 366
 367         return nil
 368 }
 369
 370 func (runner *ContainerRunner) SetupMounts() (err error) {
 371         err = runner.SetupArvMountPoint("keep")
 372         if err != nil {
 373                 return fmt.Errorf("While creating keep mount temp dir: %v", err)
 374         }
 375
 376         token, err := runner.ContainerToken()
 377         if err != nil {
 378                 return fmt.Errorf("could not get container token: %s", err)
 379         }
 380
 381         pdhOnly := true
 382         tmpcount := 0
 383         arvMountCmd := []string{
 384                 "--foreground",
 385                 "--allow-other",
 386                 "--read-write",
 387                 fmt.Sprintf("--crunchstat-interval=%v", runner.statInterval.Seconds())}
 388
 389         if runner.Container.RuntimeConstraints.KeepCacheRAM > 0 {
 390                 arvMountCmd = append(arvMountCmd, "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheRAM))
 391         }
 392
 393         collectionPaths := []string{}
 394         runner.Binds = nil
 395         runner.Volumes = make(map[string]struct{})
 396         needCertMount := true
 397         type copyFile struct {
 398                 src  string
 399                 bind string
 400         }
 401         var copyFiles []copyFile
 402
 403         var binds []string
 404         for bind := range runner.Container.Mounts {
 405                 binds = append(binds, bind)
 406         }
 407         for bind := range runner.SecretMounts {
 408                 if _, ok := runner.Container.Mounts[bind]; ok {
 409                         return fmt.Errorf("Secret mount %q conflicts with regular mount", bind)
 410                 }
 411                 if runner.SecretMounts[bind].Kind != "json" &&
 412                         runner.SecretMounts[bind].Kind != "text" {
 413                         return fmt.Errorf("Secret mount %q type is %q but only 'json' and 'text' are permitted.",
 414                                 bind, runner.SecretMounts[bind].Kind)
 415                 }
 416                 binds = append(binds, bind)
 417         }
 418         sort.Strings(binds)
 419
 420         for _, bind := range binds {
 421                 mnt, ok := runner.Container.Mounts[bind]
 422                 if !ok {
 423                         mnt = runner.SecretMounts[bind]
 424                 }
 425                 if bind == "stdout" || bind == "stderr" {
 426                         // Is it a "file" mount kind?
 427                         if mnt.Kind != "file" {
 428                                 return fmt.Errorf("Unsupported mount kind '%s' for %s. Only 'file' is supported.", mnt.Kind, bind)
 429                         }
 430
 431                         // Does path start with OutputPath?
 432                         prefix := runner.Container.OutputPath
 433                         if !strings.HasSuffix(prefix, "/") {
 434                                 prefix += "/"
 435                         }
 436                         if !strings.HasPrefix(mnt.Path, prefix) {
 437                                 return fmt.Errorf("%s path does not start with OutputPath: %s, %s", strings.Title(bind), mnt.Path, prefix)
 438                         }
 439                 }
 440
 441                 if bind == "stdin" {
 442                         // Is it a "collection" mount kind?
 443                         if mnt.Kind != "collection" && mnt.Kind != "json" {
 444                                 return fmt.Errorf("Unsupported mount kind '%s' for stdin. Only 'collection' or 'json' are supported.", mnt.Kind)
 445                         }
 446                 }
 447
 448                 if bind == "/etc/arvados/ca-certificates.crt" {
 449                         needCertMount = false
 450                 }
 451
 452                 if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && bind != runner.Container.OutputPath+"/" {
 453                         if mnt.Kind != "collection" && mnt.Kind != "text" && mnt.Kind != "json" {
 454                                 return fmt.Errorf("Only mount points of kind 'collection', 'text' or 'json' are supported underneath the output_path for %q, was %q", bind, mnt.Kind)
 455                         }
 456                 }
 457
 458                 switch {
 459                 case mnt.Kind == "collection" && bind != "stdin":
 460                         var src string
 461                         if mnt.UUID != "" && mnt.PortableDataHash != "" {
 462                                 return fmt.Errorf("Cannot specify both 'uuid' and 'portable_data_hash' for a collection mount")
 463                         }
 464                         if mnt.UUID != "" {
 465                                 if mnt.Writable {
 466                                         return fmt.Errorf("Writing to existing collections currently not permitted.")
 467                                 }
 468                                 pdhOnly = false
 469                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.UUID)
 470                         } else if mnt.PortableDataHash != "" {
 471                                 if mnt.Writable && !strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 472                                         return fmt.Errorf("Can never write to a collection specified by portable data hash")
 473                                 }
 474                                 idx := strings.Index(mnt.PortableDataHash, "/")
 475                                 if idx > 0 {
 476                                         mnt.Path = path.Clean(mnt.PortableDataHash[idx:])
 477                                         mnt.PortableDataHash = mnt.PortableDataHash[0:idx]
 478                                         runner.Container.Mounts[bind] = mnt
 479                                 }
 480                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.PortableDataHash)
 481                                 if mnt.Path != "" && mnt.Path != "." {
 482                                         if strings.HasPrefix(mnt.Path, "./") {
 483                                                 mnt.Path = mnt.Path[2:]
 484                                         } else if strings.HasPrefix(mnt.Path, "/") {
 485                                                 mnt.Path = mnt.Path[1:]
 486                                         }
 487                                         src += "/" + mnt.Path
 488                                 }
 489                         } else {
 490                                 src = fmt.Sprintf("%s/tmp%d", runner.ArvMountPoint, tmpcount)
 491                                 arvMountCmd = append(arvMountCmd, "--mount-tmp")
 492                                 arvMountCmd = append(arvMountCmd, fmt.Sprintf("tmp%d", tmpcount))
 493                                 tmpcount += 1
 494                         }
 495                         if mnt.Writable {
 496                                 if bind == runner.Container.OutputPath {
 497                                         runner.HostOutputDir = src
 498                                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", src, bind))
 499                                 } else if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 500                                         copyFiles = append(copyFiles, copyFile{src, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
 501                                 } else {
 502                                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", src, bind))
 503                                 }
 504                         } else {
 505                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", src, bind))
 506                         }
 507                         collectionPaths = append(collectionPaths, src)
 508
 509                 case mnt.Kind == "tmp":
 510                         var tmpdir string
 511                         tmpdir, err = runner.MkTempDir(runner.parentTemp, "tmp")
 512                         if err != nil {
 513                                 return fmt.Errorf("While creating mount temp dir: %v", err)
 514                         }
 515                         st, staterr := os.Stat(tmpdir)
 516                         if staterr != nil {
 517                                 return fmt.Errorf("While Stat on temp dir: %v", staterr)
 518                         }
 519                         err = os.Chmod(tmpdir, st.Mode()|os.ModeSetgid|0777)
 520                         if staterr != nil {
 521                                 return fmt.Errorf("While Chmod temp dir: %v", err)
 522                         }
 523                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", tmpdir, bind))
 524                         if bind == runner.Container.OutputPath {
 525                                 runner.HostOutputDir = tmpdir
 526                         }
 527
 528                 case mnt.Kind == "json" || mnt.Kind == "text":
 529                         var filedata []byte
 530                         if mnt.Kind == "json" {
 531                                 filedata, err = json.Marshal(mnt.Content)
 532                                 if err != nil {
 533                                         return fmt.Errorf("encoding json data: %v", err)
 534                                 }
 535                         } else {
 536                                 text, ok := mnt.Content.(string)
 537                                 if !ok {
 538                                         return fmt.Errorf("content for mount %q must be a string", bind)
 539                                 }
 540                                 filedata = []byte(text)
 541                         }
 542
 543                         tmpdir, err := runner.MkTempDir(runner.parentTemp, mnt.Kind)
 544                         if err != nil {
 545                                 return fmt.Errorf("creating temp dir: %v", err)
 546                         }
 547                         tmpfn := filepath.Join(tmpdir, "mountdata."+mnt.Kind)
 548                         err = ioutil.WriteFile(tmpfn, filedata, 0444)
 549                         if err != nil {
 550                                 return fmt.Errorf("writing temp file: %v", err)
 551                         }
 552                         if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 553                                 copyFiles = append(copyFiles, copyFile{tmpfn, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
 554                         } else {
 555                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", tmpfn, bind))
 556                         }
 557
 558                 case mnt.Kind == "git_tree":
 559                         tmpdir, err := runner.MkTempDir(runner.parentTemp, "git_tree")
 560                         if err != nil {
 561                                 return fmt.Errorf("creating temp dir: %v", err)
 562                         }
 563                         err = gitMount(mnt).extractTree(runner.ArvClient, tmpdir, token)
 564                         if err != nil {
 565                                 return err
 566                         }
 567                         runner.Binds = append(runner.Binds, tmpdir+":"+bind+":ro")
 568                 }
 569         }
 570
 571         if runner.HostOutputDir == "" {
 572                 return fmt.Errorf("Output path does not correspond to a writable mount point")
 573         }
 574
 575         if wantAPI := runner.Container.RuntimeConstraints.API; needCertMount && wantAPI != nil && *wantAPI {
 576                 for _, certfile := range arvadosclient.CertFiles {
 577                         _, err := os.Stat(certfile)
 578                         if err == nil {
 579                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:/etc/arvados/ca-certificates.crt:ro", certfile))
 580                                 break
 581                         }
 582                 }
 583         }
 584
 585         if pdhOnly {
 586                 arvMountCmd = append(arvMountCmd, "--mount-by-pdh", "by_id")
 587         } else {
 588                 arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_id")
 589         }
 590         arvMountCmd = append(arvMountCmd, runner.ArvMountPoint)
 591
 592         runner.ArvMount, err = runner.RunArvMount(arvMountCmd, token)
 593         if err != nil {
 594                 return fmt.Errorf("While trying to start arv-mount: %v", err)
 595         }
 596
 597         for _, p := range collectionPaths {
 598                 _, err = os.Stat(p)
 599                 if err != nil {
 600                         return fmt.Errorf("While checking that input files exist: %v", err)
 601                 }
 602         }
 603
 604         for _, cp := range copyFiles {
 605                 st, err := os.Stat(cp.src)
 606                 if err != nil {
 607                         return fmt.Errorf("While staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 608                 }
 609                 if st.IsDir() {
 610                         err = filepath.Walk(cp.src, func(walkpath string, walkinfo os.FileInfo, walkerr error) error {
 611                                 if walkerr != nil {
 612                                         return walkerr
 613                                 }
 614                                 target := path.Join(cp.bind, walkpath[len(cp.src):])
 615                                 if walkinfo.Mode().IsRegular() {
 616                                         copyerr := copyfile(walkpath, target)
 617                                         if copyerr != nil {
 618                                                 return copyerr
 619                                         }
 620                                         return os.Chmod(target, walkinfo.Mode()|0777)
 621                                 } else if walkinfo.Mode().IsDir() {
 622                                         mkerr := os.MkdirAll(target, 0777)
 623                                         if mkerr != nil {
 624                                                 return mkerr
 625                                         }
 626                                         return os.Chmod(target, walkinfo.Mode()|os.ModeSetgid|0777)
 627                                 } else {
 628                                         return fmt.Errorf("Source %q is not a regular file or directory", cp.src)
 629                                 }
 630                         })
 631                 } else if st.Mode().IsRegular() {
 632                         err = copyfile(cp.src, cp.bind)
 633                         if err == nil {
 634                                 err = os.Chmod(cp.bind, st.Mode()|0777)
 635                         }
 636                 }
 637                 if err != nil {
 638                         return fmt.Errorf("While staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 639                 }
 640         }
 641
 642         return nil
 643 }
 644
 645 func (runner *ContainerRunner) ProcessDockerAttach(containerReader io.Reader) {
 646         // Handle docker log protocol
 647         // https://docs.docker.com/engine/reference/api/docker_remote_api_v1.15/#attach-to-a-container
 648         defer close(runner.loggingDone)
 649
 650         header := make([]byte, 8)
 651         var err error
 652         for err == nil {
 653                 _, err = io.ReadAtLeast(containerReader, header, 8)
 654                 if err != nil {
 655                         if err == io.EOF {
 656                                 err = nil
 657                         }
 658                         break
 659                 }
 660                 readsize := int64(header[7]) | (int64(header[6]) << 8) | (int64(header[5]) << 16) | (int64(header[4]) << 24)
 661                 if header[0] == 1 {
 662                         // stdout
 663                         _, err = io.CopyN(runner.Stdout, containerReader, readsize)
 664                 } else {
 665                         // stderr
 666                         _, err = io.CopyN(runner.Stderr, containerReader, readsize)
 667                 }
 668         }
 669
 670         if err != nil {
 671                 runner.CrunchLog.Printf("error reading docker logs: %v", err)
 672         }
 673
 674         err = runner.Stdout.Close()
 675         if err != nil {
 676                 runner.CrunchLog.Printf("error closing stdout logs: %v", err)
 677         }
 678
 679         err = runner.Stderr.Close()
 680         if err != nil {
 681                 runner.CrunchLog.Printf("error closing stderr logs: %v", err)
 682         }
 683
 684         if runner.statReporter != nil {
 685                 runner.statReporter.Stop()
 686                 err = runner.statLogger.Close()
 687                 if err != nil {
 688                         runner.CrunchLog.Printf("error closing crunchstat logs: %v", err)
 689                 }
 690         }
 691 }
 692
 693 func (runner *ContainerRunner) stopHoststat() error {
 694         if runner.hoststatReporter == nil {
 695                 return nil
 696         }
 697         runner.hoststatReporter.Stop()
 698         err := runner.hoststatLogger.Close()
 699         if err != nil {
 700                 return fmt.Errorf("error closing hoststat logs: %v", err)
 701         }
 702         return nil
 703 }
 704
 705 func (runner *ContainerRunner) startHoststat() error {
 706         w, err := runner.NewLogWriter("hoststat")
 707         if err != nil {
 708                 return err
 709         }
 710         runner.hoststatLogger = NewThrottledLogger(w)
 711         runner.hoststatReporter = &crunchstat.Reporter{
 712                 Logger:     log.New(runner.hoststatLogger, "", 0),
 713                 CgroupRoot: runner.cgroupRoot,
 714                 PollPeriod: runner.statInterval,
 715         }
 716         runner.hoststatReporter.Start()
 717         return nil
 718 }
 719
 720 func (runner *ContainerRunner) startCrunchstat() error {
 721         w, err := runner.NewLogWriter("crunchstat")
 722         if err != nil {
 723                 return err
 724         }
 725         runner.statLogger = NewThrottledLogger(w)
 726         runner.statReporter = &crunchstat.Reporter{
 727                 CID:          runner.ContainerID,
 728                 Logger:       log.New(runner.statLogger, "", 0),
 729                 CgroupParent: runner.expectCgroupParent,
 730                 CgroupRoot:   runner.cgroupRoot,
 731                 PollPeriod:   runner.statInterval,
 732         }
 733         runner.statReporter.Start()
 734         return nil
 735 }
 736
 737 type infoCommand struct {
 738         label string
 739         cmd   []string
 740 }
 741
 742 // LogHostInfo logs info about the current host, for debugging and
 743 // accounting purposes. Although it's logged as "node-info", this is
 744 // about the environment where crunch-run is actually running, which
 745 // might differ from what's described in the node record (see
 746 // LogNodeRecord).
 747 func (runner *ContainerRunner) LogHostInfo() (err error) {
 748         w, err := runner.NewLogWriter("node-info")
 749         if err != nil {
 750                 return
 751         }
 752
 753         commands := []infoCommand{
 754                 {
 755                         label: "Host Information",
 756                         cmd:   []string{"uname", "-a"},
 757                 },
 758                 {
 759                         label: "CPU Information",
 760                         cmd:   []string{"cat", "/proc/cpuinfo"},
 761                 },
 762                 {
 763                         label: "Memory Information",
 764                         cmd:   []string{"cat", "/proc/meminfo"},
 765                 },
 766                 {
 767                         label: "Disk Space",
 768                         cmd:   []string{"df", "-m", "/", os.TempDir()},
 769                 },
 770                 {
 771                         label: "Disk INodes",
 772                         cmd:   []string{"df", "-i", "/", os.TempDir()},
 773                 },
 774         }
 775
 776         // Run commands with informational output to be logged.
 777         for _, command := range commands {
 778                 fmt.Fprintln(w, command.label)
 779                 cmd := exec.Command(command.cmd[0], command.cmd[1:]...)
 780                 cmd.Stdout = w
 781                 cmd.Stderr = w
 782                 if err := cmd.Run(); err != nil {
 783                         err = fmt.Errorf("While running command %q: %v", command.cmd, err)
 784                         fmt.Fprintln(w, err)
 785                         return err
 786                 }
 787                 fmt.Fprintln(w, "")
 788         }
 789
 790         err = w.Close()
 791         if err != nil {
 792                 return fmt.Errorf("While closing node-info logs: %v", err)
 793         }
 794         return nil
 795 }
 796
 797 // LogContainerRecord gets and saves the raw JSON container record from the API server
 798 func (runner *ContainerRunner) LogContainerRecord() error {
 799         logged, err := runner.logAPIResponse("container", "containers", map[string]interface{}{"filters": [][]string{{"uuid", "=", runner.Container.UUID}}}, nil)
 800         if !logged && err == nil {
 801                 err = fmt.Errorf("error: no container record found for %s", runner.Container.UUID)
 802         }
 803         return err
 804 }
 805
 806 // LogNodeRecord logs arvados#node record corresponding to the current host.
 807 func (runner *ContainerRunner) LogNodeRecord() error {
 808         hostname := os.Getenv("SLURMD_NODENAME")
 809         if hostname == "" {
 810                 hostname, _ = os.Hostname()
 811         }
 812         _, err := runner.logAPIResponse("node", "nodes", map[string]interface{}{"filters": [][]string{{"hostname", "=", hostname}}}, func(resp interface{}) {
 813                 // The "info" field has admin-only info when obtained
 814                 // with a privileged token, and should not be logged.
 815                 node, ok := resp.(map[string]interface{})
 816                 if ok {
 817                         delete(node, "info")
 818                 }
 819         })
 820         return err
 821 }
 822
 823 func (runner *ContainerRunner) logAPIResponse(label, path string, params map[string]interface{}, munge func(interface{})) (logged bool, err error) {
 824         writer, err := runner.LogCollection.OpenFile(label+".json", os.O_CREATE|os.O_WRONLY, 0666)
 825         if err != nil {
 826                 return false, err
 827         }
 828         w := &ArvLogWriter{
 829                 ArvClient:     runner.ArvClient,
 830                 UUID:          runner.Container.UUID,
 831                 loggingStream: label,
 832                 writeCloser:   writer,
 833         }
 834
 835         reader, err := runner.ArvClient.CallRaw("GET", path, "", "", arvadosclient.Dict(params))
 836         if err != nil {
 837                 return false, fmt.Errorf("error getting %s record: %v", label, err)
 838         }
 839         defer reader.Close()
 840
 841         dec := json.NewDecoder(reader)
 842         dec.UseNumber()
 843         var resp map[string]interface{}
 844         if err = dec.Decode(&resp); err != nil {
 845                 return false, fmt.Errorf("error decoding %s list response: %v", label, err)
 846         }
 847         items, ok := resp["items"].([]interface{})
 848         if !ok {
 849                 return false, fmt.Errorf("error decoding %s list response: no \"items\" key in API list response", label)
 850         } else if len(items) < 1 {
 851                 return false, nil
 852         }
 853         if munge != nil {
 854                 munge(items[0])
 855         }
 856         // Re-encode it using indentation to improve readability
 857         enc := json.NewEncoder(w)
 858         enc.SetIndent("", "    ")
 859         if err = enc.Encode(items[0]); err != nil {
 860                 return false, fmt.Errorf("error logging %s record: %v", label, err)
 861         }
 862         err = w.Close()
 863         if err != nil {
 864                 return false, fmt.Errorf("error closing %s.json in log collection: %v", label, err)
 865         }
 866         return true, nil
 867 }
 868
 869 // AttachStreams connects the docker container stdin, stdout and stderr logs
 870 // to the Arvados logger which logs to Keep and the API server logs table.
 871 func (runner *ContainerRunner) AttachStreams() (err error) {
 872
 873         runner.CrunchLog.Print("Attaching container streams")
 874
 875         // If stdin mount is provided, attach it to the docker container
 876         var stdinRdr arvados.File
 877         var stdinJson []byte
 878         if stdinMnt, ok := runner.Container.Mounts["stdin"]; ok {
 879                 if stdinMnt.Kind == "collection" {
 880                         var stdinColl arvados.Collection
 881                         collId := stdinMnt.UUID
 882                         if collId == "" {
 883                                 collId = stdinMnt.PortableDataHash
 884                         }
 885                         err = runner.ArvClient.Get("collections", collId, nil, &stdinColl)
 886                         if err != nil {
 887                                 return fmt.Errorf("While getting stding collection: %v", err)
 888                         }
 889
 890                         stdinRdr, err = runner.Kc.ManifestFileReader(manifest.Manifest{Text: stdinColl.ManifestText}, stdinMnt.Path)
 891                         if os.IsNotExist(err) {
 892                                 return fmt.Errorf("stdin collection path not found: %v", stdinMnt.Path)
 893                         } else if err != nil {
 894                                 return fmt.Errorf("While getting stdin collection path %v: %v", stdinMnt.Path, err)
 895                         }
 896                 } else if stdinMnt.Kind == "json" {
 897                         stdinJson, err = json.Marshal(stdinMnt.Content)
 898                         if err != nil {
 899                                 return fmt.Errorf("While encoding stdin json data: %v", err)
 900                         }
 901                 }
 902         }
 903
 904         stdinUsed := stdinRdr != nil || len(stdinJson) != 0
 905         response, err := runner.Docker.ContainerAttach(context.TODO(), runner.ContainerID,
 906                 dockertypes.ContainerAttachOptions{Stream: true, Stdin: stdinUsed, Stdout: true, Stderr: true})
 907         if err != nil {
 908                 return fmt.Errorf("While attaching container stdout/stderr streams: %v", err)
 909         }
 910
 911         runner.loggingDone = make(chan bool)
 912
 913         if stdoutMnt, ok := runner.Container.Mounts["stdout"]; ok {
 914                 stdoutFile, err := runner.getStdoutFile(stdoutMnt.Path)
 915                 if err != nil {
 916                         return err
 917                 }
 918                 runner.Stdout = stdoutFile
 919         } else if w, err := runner.NewLogWriter("stdout"); err != nil {
 920                 return err
 921         } else {
 922                 runner.Stdout = NewThrottledLogger(w)
 923         }
 924
 925         if stderrMnt, ok := runner.Container.Mounts["stderr"]; ok {
 926                 stderrFile, err := runner.getStdoutFile(stderrMnt.Path)
 927                 if err != nil {
 928                         return err
 929                 }
 930                 runner.Stderr = stderrFile
 931         } else if w, err := runner.NewLogWriter("stderr"); err != nil {
 932                 return err
 933         } else {
 934                 runner.Stderr = NewThrottledLogger(w)
 935         }
 936
 937         if stdinRdr != nil {
 938                 go func() {
 939                         _, err := io.Copy(response.Conn, stdinRdr)
 940                         if err != nil {
 941                                 runner.CrunchLog.Print("While writing stdin collection to docker container %q", err)
 942                                 runner.stop(nil)
 943                         }
 944                         stdinRdr.Close()
 945                         response.CloseWrite()
 946                 }()
 947         } else if len(stdinJson) != 0 {
 948                 go func() {
 949                         _, err := io.Copy(response.Conn, bytes.NewReader(stdinJson))
 950                         if err != nil {
 951                                 runner.CrunchLog.Print("While writing stdin json to docker container %q", err)
 952                                 runner.stop(nil)
 953                         }
 954                         response.CloseWrite()
 955                 }()
 956         }
 957
 958         go runner.ProcessDockerAttach(response.Reader)
 959
 960         return nil
 961 }
 962
 963 func (runner *ContainerRunner) getStdoutFile(mntPath string) (*os.File, error) {
 964         stdoutPath := mntPath[len(runner.Container.OutputPath):]
 965         index := strings.LastIndex(stdoutPath, "/")
 966         if index > 0 {
 967                 subdirs := stdoutPath[:index]
 968                 if subdirs != "" {
 969                         st, err := os.Stat(runner.HostOutputDir)
 970                         if err != nil {
 971                                 return nil, fmt.Errorf("While Stat on temp dir: %v", err)
 972                         }
 973                         stdoutPath := filepath.Join(runner.HostOutputDir, subdirs)
 974                         err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777)
 975                         if err != nil {
 976                                 return nil, fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err)
 977                         }
 978                 }
 979         }
 980         stdoutFile, err := os.Create(filepath.Join(runner.HostOutputDir, stdoutPath))
 981         if err != nil {
 982                 return nil, fmt.Errorf("While creating file %q: %v", stdoutPath, err)
 983         }
 984
 985         return stdoutFile, nil
 986 }
 987
 988 // CreateContainer creates the docker container.
 989 func (runner *ContainerRunner) CreateContainer() error {
 990         runner.CrunchLog.Print("Creating Docker container")
 991
 992         runner.ContainerConfig.Cmd = runner.Container.Command
 993         if runner.Container.Cwd != "." {
 994                 runner.ContainerConfig.WorkingDir = runner.Container.Cwd
 995         }
 996
 997         for k, v := range runner.Container.Environment {
 998                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env, k+"="+v)
 999         }
1000
1001         runner.ContainerConfig.Volumes = runner.Volumes
1002
1003         maxRAM := int64(runner.Container.RuntimeConstraints.RAM)
1004         runner.HostConfig = dockercontainer.HostConfig{
1005                 Binds: runner.Binds,
1006                 LogConfig: dockercontainer.LogConfig{
1007                         Type: "none",
1008                 },
1009                 Resources: dockercontainer.Resources{
1010                         CgroupParent: runner.setCgroupParent,
1011                         NanoCPUs:     int64(runner.Container.RuntimeConstraints.VCPUs) * 1000000000,
1012                         Memory:       maxRAM, // RAM
1013                         MemorySwap:   maxRAM, // RAM+swap
1014                         KernelMemory: maxRAM, // kernel portion
1015                 },
1016         }
1017
1018         if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
1019                 tok, err := runner.ContainerToken()
1020                 if err != nil {
1021                         return err
1022                 }
1023                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env,
1024                         "ARVADOS_API_TOKEN="+tok,
1025                         "ARVADOS_API_HOST="+os.Getenv("ARVADOS_API_HOST"),
1026                         "ARVADOS_API_HOST_INSECURE="+os.Getenv("ARVADOS_API_HOST_INSECURE"),
1027                 )
1028                 runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
1029         } else {
1030                 if runner.enableNetwork == "always" {
1031                         runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
1032                 } else {
1033                         runner.HostConfig.NetworkMode = dockercontainer.NetworkMode("none")
1034                 }
1035         }
1036
1037         _, stdinUsed := runner.Container.Mounts["stdin"]
1038         runner.ContainerConfig.OpenStdin = stdinUsed
1039         runner.ContainerConfig.StdinOnce = stdinUsed
1040         runner.ContainerConfig.AttachStdin = stdinUsed
1041         runner.ContainerConfig.AttachStdout = true
1042         runner.ContainerConfig.AttachStderr = true
1043
1044         createdBody, err := runner.Docker.ContainerCreate(context.TODO(), &runner.ContainerConfig, &runner.HostConfig, nil, runner.Container.UUID)
1045         if err != nil {
1046                 return fmt.Errorf("While creating container: %v", err)
1047         }
1048
1049         runner.ContainerID = createdBody.ID
1050
1051         return runner.AttachStreams()
1052 }
1053
1054 // StartContainer starts the docker container created by CreateContainer.
1055 func (runner *ContainerRunner) StartContainer() error {
1056         runner.CrunchLog.Printf("Starting Docker container id '%s'", runner.ContainerID)
1057         runner.cStateLock.Lock()
1058         defer runner.cStateLock.Unlock()
1059         if runner.cCancelled {
1060                 return ErrCancelled
1061         }
1062         err := runner.Docker.ContainerStart(context.TODO(), runner.ContainerID,
1063                 dockertypes.ContainerStartOptions{})
1064         if err != nil {
1065                 var advice string
1066                 if m, e := regexp.MatchString("(?ms).*(exec|System error).*(no such file or directory|file not found).*", err.Error()); m && e == nil {
1067                         advice = fmt.Sprintf("\nPossible causes: command %q is missing, the interpreter given in #! is missing, or script has Windows line endings.", runner.Container.Command[0])
1068                 }
1069                 return fmt.Errorf("could not start container: %v%s", err, advice)
1070         }
1071         return nil
1072 }
1073
1074 // WaitFinish waits for the container to terminate, capture the exit code, and
1075 // close the stdout/stderr logging.
1076 func (runner *ContainerRunner) WaitFinish() error {
1077         var runTimeExceeded <-chan time.Time
1078         runner.CrunchLog.Print("Waiting for container to finish")
1079
1080         waitOk, waitErr := runner.Docker.ContainerWait(context.TODO(), runner.ContainerID, dockercontainer.WaitConditionNotRunning)
1081         arvMountExit := runner.ArvMountExit
1082         if timeout := runner.Container.SchedulingParameters.MaxRunTime; timeout > 0 {
1083                 runTimeExceeded = time.After(time.Duration(timeout) * time.Second)
1084         }
1085         for {
1086                 select {
1087                 case waitBody := <-waitOk:
1088                         runner.CrunchLog.Printf("Container exited with code: %v", waitBody.StatusCode)
1089                         code := int(waitBody.StatusCode)
1090                         runner.ExitCode = &code
1091
1092                         // wait for stdout/stderr to complete
1093                         <-runner.loggingDone
1094                         return nil
1095
1096                 case err := <-waitErr:
1097                         return fmt.Errorf("container wait: %v", err)
1098
1099                 case <-arvMountExit:
1100                         runner.CrunchLog.Printf("arv-mount exited while container is still running.  Stopping container.")
1101                         runner.stop(nil)
1102                         // arvMountExit will always be ready now that
1103                         // it's closed, but that doesn't interest us.
1104                         arvMountExit = nil
1105
1106                 case <-runTimeExceeded:
1107                         runner.CrunchLog.Printf("maximum run time exceeded. Stopping container.")
1108                         runner.stop(nil)
1109                         runTimeExceeded = nil
1110                 }
1111         }
1112 }
1113
1114 // CaptureOutput saves data from the container's output directory if
1115 // needed, and updates the container output accordingly.
1116 func (runner *ContainerRunner) CaptureOutput() error {
1117         if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
1118                 // Output may have been set directly by the container, so
1119                 // refresh the container record to check.
1120                 err := runner.ArvClient.Get("containers", runner.Container.UUID,
1121                         nil, &runner.Container)
1122                 if err != nil {
1123                         return err
1124                 }
1125                 if runner.Container.Output != "" {
1126                         // Container output is already set.
1127                         runner.OutputPDH = &runner.Container.Output
1128                         return nil
1129                 }
1130         }
1131
1132         txt, err := (&copier{
1133                 client:        runner.client,
1134                 arvClient:     runner.ArvClient,
1135                 keepClient:    runner.Kc,
1136                 hostOutputDir: runner.HostOutputDir,
1137                 ctrOutputDir:  runner.Container.OutputPath,
1138                 binds:         runner.Binds,
1139                 mounts:        runner.Container.Mounts,
1140                 secretMounts:  runner.SecretMounts,
1141                 logger:        runner.CrunchLog,
1142         }).Copy()
1143         if err != nil {
1144                 return err
1145         }
1146         var resp arvados.Collection
1147         err = runner.ArvClient.Create("collections", arvadosclient.Dict{
1148                 "ensure_unique_name": true,
1149                 "collection": arvadosclient.Dict{
1150                         "is_trashed":    true,
1151                         "name":          "output for " + runner.Container.UUID,
1152                         "manifest_text": txt,
1153                 },
1154         }, &resp)
1155         if err != nil {
1156                 return fmt.Errorf("error creating output collection: %v", err)
1157         }
1158         runner.OutputPDH = &resp.PortableDataHash
1159         return nil
1160 }
1161
1162 func (runner *ContainerRunner) CleanupDirs() {
1163         if runner.ArvMount != nil {
1164                 var delay int64 = 8
1165                 umount := exec.Command("arv-mount", fmt.Sprintf("--unmount-timeout=%d", delay), "--unmount", runner.ArvMountPoint)
1166                 umount.Stdout = runner.CrunchLog
1167                 umount.Stderr = runner.CrunchLog
1168                 runner.CrunchLog.Printf("Running %v", umount.Args)
1169                 umnterr := umount.Start()
1170
1171                 if umnterr != nil {
1172                         runner.CrunchLog.Printf("Error unmounting: %v", umnterr)
1173                 } else {
1174                         // If arv-mount --unmount gets stuck for any reason, we
1175                         // don't want to wait for it forever.  Do Wait() in a goroutine
1176                         // so it doesn't block crunch-run.
1177                         umountExit := make(chan error)
1178                         go func() {
1179                                 mnterr := umount.Wait()
1180                                 if mnterr != nil {
1181                                         runner.CrunchLog.Printf("Error unmounting: %v", mnterr)
1182                                 }
1183                                 umountExit <- mnterr
1184                         }()
1185
1186                         for again := true; again; {
1187                                 again = false
1188                                 select {
1189                                 case <-umountExit:
1190                                         umount = nil
1191                                         again = true
1192                                 case <-runner.ArvMountExit:
1193                                         break
1194                                 case <-time.After(time.Duration((delay + 1) * int64(time.Second))):
1195                                         runner.CrunchLog.Printf("Timed out waiting for unmount")
1196                                         if umount != nil {
1197                                                 umount.Process.Kill()
1198                                         }
1199                                         runner.ArvMount.Process.Kill()
1200                                 }
1201                         }
1202                 }
1203         }
1204
1205         if runner.ArvMountPoint != "" {
1206                 if rmerr := os.Remove(runner.ArvMountPoint); rmerr != nil {
1207                         runner.CrunchLog.Printf("While cleaning up arv-mount directory %s: %v", runner.ArvMountPoint, rmerr)
1208                 }
1209         }
1210
1211         if rmerr := os.RemoveAll(runner.parentTemp); rmerr != nil {
1212                 runner.CrunchLog.Printf("While cleaning up temporary directory %s: %v", runner.parentTemp, rmerr)
1213         }
1214 }
1215
1216 // CommitLogs posts the collection containing the final container logs.
1217 func (runner *ContainerRunner) CommitLogs() error {
1218         func() {
1219                 // Hold cStateLock to prevent races on CrunchLog (e.g., stop()).
1220                 runner.cStateLock.Lock()
1221                 defer runner.cStateLock.Unlock()
1222
1223                 runner.CrunchLog.Print(runner.finalState)
1224
1225                 if runner.arvMountLog != nil {
1226                         runner.arvMountLog.Close()
1227                 }
1228                 runner.CrunchLog.Close()
1229
1230                 // Closing CrunchLog above allows them to be committed to Keep at this
1231                 // point, but re-open crunch log with ArvClient in case there are any
1232                 // other further errors (such as failing to write the log to Keep!)
1233                 // while shutting down
1234                 runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{
1235                         ArvClient:     runner.ArvClient,
1236                         UUID:          runner.Container.UUID,
1237                         loggingStream: "crunch-run",
1238                         writeCloser:   nil,
1239                 })
1240                 runner.CrunchLog.Immediate = log.New(os.Stderr, runner.Container.UUID+" ", 0)
1241         }()
1242
1243         if runner.LogsPDH != nil {
1244                 // If we have already assigned something to LogsPDH,
1245                 // we must be closing the re-opened log, which won't
1246                 // end up getting attached to the container record and
1247                 // therefore doesn't need to be saved as a collection
1248                 // -- it exists only to send logs to other channels.
1249                 return nil
1250         }
1251
1252         mt, err := runner.LogCollection.MarshalManifest(".")
1253         if err != nil {
1254                 return fmt.Errorf("While creating log manifest: %v", err)
1255         }
1256
1257         var response arvados.Collection
1258         err = runner.ArvClient.Create("collections",
1259                 arvadosclient.Dict{
1260                         "ensure_unique_name": true,
1261                         "collection": arvadosclient.Dict{
1262                                 "is_trashed":    true,
1263                                 "name":          "logs for " + runner.Container.UUID,
1264                                 "manifest_text": mt}},
1265                 &response)
1266         if err != nil {
1267                 return fmt.Errorf("While creating log collection: %v", err)
1268         }
1269         runner.LogsPDH = &response.PortableDataHash
1270         return nil
1271 }
1272
1273 // UpdateContainerRunning updates the container state to "Running"
1274 func (runner *ContainerRunner) UpdateContainerRunning() error {
1275         runner.cStateLock.Lock()
1276         defer runner.cStateLock.Unlock()
1277         if runner.cCancelled {
1278                 return ErrCancelled
1279         }
1280         return runner.ArvClient.Update("containers", runner.Container.UUID,
1281                 arvadosclient.Dict{"container": arvadosclient.Dict{"state": "Running"}}, nil)
1282 }
1283
1284 // ContainerToken returns the api_token the container (and any
1285 // arv-mount processes) are allowed to use.
1286 func (runner *ContainerRunner) ContainerToken() (string, error) {
1287         if runner.token != "" {
1288                 return runner.token, nil
1289         }
1290
1291         var auth arvados.APIClientAuthorization
1292         err := runner.ArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth)
1293         if err != nil {
1294                 return "", err
1295         }
1296         runner.token = auth.APIToken
1297         return runner.token, nil
1298 }
1299
1300 // UpdateContainerComplete updates the container record state on API
1301 // server to "Complete" or "Cancelled"
1302 func (runner *ContainerRunner) UpdateContainerFinal() error {
1303         update := arvadosclient.Dict{}
1304         update["state"] = runner.finalState
1305         if runner.LogsPDH != nil {
1306                 update["log"] = *runner.LogsPDH
1307         }
1308         if runner.finalState == "Complete" {
1309                 if runner.ExitCode != nil {
1310                         update["exit_code"] = *runner.ExitCode
1311                 }
1312                 if runner.OutputPDH != nil {
1313                         update["output"] = *runner.OutputPDH
1314                 }
1315         }
1316         return runner.ArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{"container": update}, nil)
1317 }
1318
1319 // IsCancelled returns the value of Cancelled, with goroutine safety.
1320 func (runner *ContainerRunner) IsCancelled() bool {
1321         runner.cStateLock.Lock()
1322         defer runner.cStateLock.Unlock()
1323         return runner.cCancelled
1324 }
1325
1326 // NewArvLogWriter creates an ArvLogWriter
1327 func (runner *ContainerRunner) NewArvLogWriter(name string) (io.WriteCloser, error) {
1328         writer, err := runner.LogCollection.OpenFile(name+".txt", os.O_CREATE|os.O_WRONLY, 0666)
1329         if err != nil {
1330                 return nil, err
1331         }
1332         return &ArvLogWriter{
1333                 ArvClient:     runner.ArvClient,
1334                 UUID:          runner.Container.UUID,
1335                 loggingStream: name,
1336                 writeCloser:   writer,
1337         }, nil
1338 }
1339
1340 // Run the full container lifecycle.
1341 func (runner *ContainerRunner) Run() (err error) {
1342         runner.CrunchLog.Printf("crunch-run %s started", version)
1343         runner.CrunchLog.Printf("Executing container '%s'", runner.Container.UUID)
1344
1345         hostname, hosterr := os.Hostname()
1346         if hosterr != nil {
1347                 runner.CrunchLog.Printf("Error getting hostname '%v'", hosterr)
1348         } else {
1349                 runner.CrunchLog.Printf("Executing on host '%s'", hostname)
1350         }
1351
1352         runner.finalState = "Queued"
1353
1354         defer func() {
1355                 runner.CleanupDirs()
1356
1357                 runner.CrunchLog.Printf("crunch-run finished")
1358                 runner.CrunchLog.Close()
1359         }()
1360
1361         defer func() {
1362                 // checkErr prints e (unless it's nil) and sets err to
1363                 // e (unless err is already non-nil). Thus, if err
1364                 // hasn't already been assigned when Run() returns,
1365                 // this cleanup func will cause Run() to return the
1366                 // first non-nil error that is passed to checkErr().
1367                 checkErr := func(e error) {
1368                         if e == nil {
1369                                 return
1370                         }
1371                         runner.CrunchLog.Print(e)
1372                         if err == nil {
1373                                 err = e
1374                         }
1375                         if runner.finalState == "Complete" {
1376                                 // There was an error in the finalization.
1377                                 runner.finalState = "Cancelled"
1378                         }
1379                 }
1380
1381                 // Log the error encountered in Run(), if any
1382                 checkErr(err)
1383
1384                 if runner.finalState == "Queued" {
1385                         runner.UpdateContainerFinal()
1386                         return
1387                 }
1388
1389                 if runner.IsCancelled() {
1390                         runner.finalState = "Cancelled"
1391                         // but don't return yet -- we still want to
1392                         // capture partial output and write logs
1393                 }
1394
1395                 checkErr(runner.CaptureOutput())
1396                 checkErr(runner.stopHoststat())
1397                 checkErr(runner.CommitLogs())
1398                 checkErr(runner.UpdateContainerFinal())
1399         }()
1400
1401         err = runner.fetchContainerRecord()
1402         if err != nil {
1403                 return
1404         }
1405         runner.setupSignals()
1406         err = runner.startHoststat()
1407         if err != nil {
1408                 return
1409         }
1410
1411         // check for and/or load image
1412         err = runner.LoadImage()
1413         if err != nil {
1414                 if !runner.checkBrokenNode(err) {
1415                         // Failed to load image but not due to a "broken node"
1416                         // condition, probably user error.
1417                         runner.finalState = "Cancelled"
1418                 }
1419                 err = fmt.Errorf("While loading container image: %v", err)
1420                 return
1421         }
1422
1423         // set up FUSE mount and binds
1424         err = runner.SetupMounts()
1425         if err != nil {
1426                 runner.finalState = "Cancelled"
1427                 err = fmt.Errorf("While setting up mounts: %v", err)
1428                 return
1429         }
1430
1431         err = runner.CreateContainer()
1432         if err != nil {
1433                 return
1434         }
1435         err = runner.LogHostInfo()
1436         if err != nil {
1437                 return
1438         }
1439         err = runner.LogNodeRecord()
1440         if err != nil {
1441                 return
1442         }
1443         err = runner.LogContainerRecord()
1444         if err != nil {
1445                 return
1446         }
1447
1448         if runner.IsCancelled() {
1449                 return
1450         }
1451
1452         err = runner.UpdateContainerRunning()
1453         if err != nil {
1454                 return
1455         }
1456         runner.finalState = "Cancelled"
1457
1458         err = runner.startCrunchstat()
1459         if err != nil {
1460                 return
1461         }
1462
1463         err = runner.StartContainer()
1464         if err != nil {
1465                 runner.checkBrokenNode(err)
1466                 return
1467         }
1468
1469         err = runner.WaitFinish()
1470         if err == nil && !runner.IsCancelled() {
1471                 runner.finalState = "Complete"
1472         }
1473         return
1474 }
1475
1476 // Fetch the current container record (uuid = runner.Container.UUID)
1477 // into runner.Container.
1478 func (runner *ContainerRunner) fetchContainerRecord() error {
1479         reader, err := runner.ArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
1480         if err != nil {
1481                 return fmt.Errorf("error fetching container record: %v", err)
1482         }
1483         defer reader.Close()
1484
1485         dec := json.NewDecoder(reader)
1486         dec.UseNumber()
1487         err = dec.Decode(&runner.Container)
1488         if err != nil {
1489                 return fmt.Errorf("error decoding container record: %v", err)
1490         }
1491
1492         var sm struct {
1493                 SecretMounts map[string]arvados.Mount `json:"secret_mounts"`
1494         }
1495
1496         containerToken, err := runner.ContainerToken()
1497         if err != nil {
1498                 return fmt.Errorf("error getting container token: %v", err)
1499         }
1500
1501         containerClient, err := runner.MkArvClient(containerToken)
1502         if err != nil {
1503                 return fmt.Errorf("error creating container API client: %v", err)
1504         }
1505
1506         err = containerClient.Call("GET", "containers", runner.Container.UUID, "secret_mounts", nil, &sm)
1507         if err != nil {
1508                 if apierr, ok := err.(arvadosclient.APIServerError); !ok || apierr.HttpStatusCode != 404 {
1509                         return fmt.Errorf("error fetching secret_mounts: %v", err)
1510                 }
1511                 // ok && apierr.HttpStatusCode == 404, which means
1512                 // secret_mounts isn't supported by this API server.
1513         }
1514         runner.SecretMounts = sm.SecretMounts
1515
1516         return nil
1517 }
1518
1519 // NewContainerRunner creates a new container runner.
1520 func NewContainerRunner(client *arvados.Client, api IArvadosClient, kc IKeepClient, docker ThinDockerClient, containerUUID string) (*ContainerRunner, error) {
1521         cr := &ContainerRunner{
1522                 client:    client,
1523                 ArvClient: api,
1524                 Kc:        kc,
1525                 Docker:    docker,
1526         }
1527         cr.NewLogWriter = cr.NewArvLogWriter
1528         cr.RunArvMount = cr.ArvMountCmd
1529         cr.MkTempDir = ioutil.TempDir
1530         cr.MkArvClient = func(token string) (IArvadosClient, error) {
1531                 cl, err := arvadosclient.MakeArvadosClient()
1532                 if err != nil {
1533                         return nil, err
1534                 }
1535                 cl.ApiToken = token
1536                 return cl, nil
1537         }
1538         var err error
1539         cr.LogCollection, err = (&arvados.Collection{}).FileSystem(cr.client, cr.Kc)
1540         if err != nil {
1541                 return nil, err
1542         }
1543         cr.Container.UUID = containerUUID
1544         w, err := cr.NewLogWriter("crunch-run")
1545         if err != nil {
1546                 return nil, err
1547         }
1548         cr.CrunchLog = NewThrottledLogger(w)
1549         cr.CrunchLog.Immediate = log.New(os.Stderr, containerUUID+" ", 0)
1550
1551         loadLogThrottleParams(api)
1552
1553         return cr, nil
1554 }
1555
1556 func main() {
1557         statInterval := flag.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting")
1558         cgroupRoot := flag.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree")
1559         cgroupParent := flag.String("cgroup-parent", "docker", "name of container's parent cgroup (ignored if -cgroup-parent-subsystem is used)")
1560         cgroupParentSubsystem := flag.String("cgroup-parent-subsystem", "", "use current cgroup for given subsystem as parent cgroup for container")
1561         caCertsPath := flag.String("ca-certs", "", "Path to TLS root certificates")
1562         enableNetwork := flag.String("container-enable-networking", "default",
1563                 `Specify if networking should be enabled for container.  One of 'default', 'always':
1564         default: only enable networking if container requests it.
1565         always:  containers always have networking enabled
1566         `)
1567         networkMode := flag.String("container-network-mode", "default",
1568                 `Set networking mode for container.  Corresponds to Docker network mode (--net).
1569         `)
1570         memprofile := flag.String("memprofile", "", "write memory profile to `file` after running container")
1571         getVersion := flag.Bool("version", false, "Print version information and exit.")
1572         flag.Parse()
1573
1574         // Print version information if requested
1575         if *getVersion {
1576                 fmt.Printf("crunch-run %s\n", version)
1577                 return
1578         }
1579
1580         log.Printf("crunch-run %s started", version)
1581
1582         containerId := flag.Arg(0)
1583
1584         if *caCertsPath != "" {
1585                 arvadosclient.CertFiles = []string{*caCertsPath}
1586         }
1587
1588         api, err := arvadosclient.MakeArvadosClient()
1589         if err != nil {
1590                 log.Fatalf("%s: %v", containerId, err)
1591         }
1592         api.Retries = 8
1593
1594         kc, kcerr := keepclient.MakeKeepClient(api)
1595         if kcerr != nil {
1596                 log.Fatalf("%s: %v", containerId, kcerr)
1597         }
1598         kc.BlockCache = &keepclient.BlockCache{MaxBlocks: 2}
1599         kc.Retries = 4
1600
1601         // API version 1.21 corresponds to Docker 1.9, which is currently the
1602         // minimum version we want to support.
1603         docker, dockererr := dockerclient.NewClient(dockerclient.DefaultDockerHost, "1.21", nil, nil)
1604
1605         cr, err := NewContainerRunner(arvados.NewClientFromEnv(), api, kc, docker, containerId)
1606         if err != nil {
1607                 log.Fatal(err)
1608         }
1609         if dockererr != nil {
1610                 cr.CrunchLog.Printf("%s: %v", containerId, dockererr)
1611                 cr.checkBrokenNode(dockererr)
1612                 cr.CrunchLog.Close()
1613                 os.Exit(1)
1614         }
1615
1616         parentTemp, tmperr := cr.MkTempDir("", "crunch-run."+containerId+".")
1617         if tmperr != nil {
1618                 log.Fatalf("%s: %v", containerId, tmperr)
1619         }
1620
1621         cr.parentTemp = parentTemp
1622         cr.statInterval = *statInterval
1623         cr.cgroupRoot = *cgroupRoot
1624         cr.expectCgroupParent = *cgroupParent
1625         cr.enableNetwork = *enableNetwork
1626         cr.networkMode = *networkMode
1627         if *cgroupParentSubsystem != "" {
1628                 p := findCgroup(*cgroupParentSubsystem)
1629                 cr.setCgroupParent = p
1630                 cr.expectCgroupParent = p
1631         }
1632
1633         runerr := cr.Run()
1634
1635         if *memprofile != "" {
1636                 f, err := os.Create(*memprofile)
1637                 if err != nil {
1638                         log.Printf("could not create memory profile: %s", err)
1639                 }
1640                 runtime.GC() // get up-to-date statistics
1641                 if err := pprof.WriteHeapProfile(f); err != nil {
1642                         log.Printf("could not write memory profile: %s", err)
1643                 }
1644                 closeerr := f.Close()
1645                 if closeerr != nil {
1646                         log.Printf("closing memprofile file: %s", err)
1647                 }
1648         }
1649
1650         if runerr != nil {
1651                 log.Fatalf("%s: %v", containerId, runerr)
1652         }
1653 }