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         PutHB(hash string, buf []byte) (string, int, error)
  61         ManifestFileReader(m manifest.Manifest, filename string) (arvados.File, error)
  62         ClearBlockCache()
  63 }
  64
  65 // NewLogWriter is a factory function to create a new log writer.
  66 type NewLogWriter func(name string) io.WriteCloser
  67
  68 type RunArvMount func(args []string, tok string) (*exec.Cmd, error)
  69
  70 type MkTempDir func(string, string) (string, error)
  71
  72 // ThinDockerClient is the minimal Docker client interface used by crunch-run.
  73 type ThinDockerClient interface {
  74         ContainerAttach(ctx context.Context, container string, options dockertypes.ContainerAttachOptions) (dockertypes.HijackedResponse, error)
  75         ContainerCreate(ctx context.Context, config *dockercontainer.Config, hostConfig *dockercontainer.HostConfig,
  76                 networkingConfig *dockernetwork.NetworkingConfig, containerName string) (dockercontainer.ContainerCreateCreatedBody, error)
  77         ContainerStart(ctx context.Context, container string, options dockertypes.ContainerStartOptions) error
  78         ContainerStop(ctx context.Context, container string, timeout *time.Duration) error
  79         ContainerWait(ctx context.Context, container string, condition dockercontainer.WaitCondition) (<-chan dockercontainer.ContainerWaitOKBody, <-chan error)
  80         ImageInspectWithRaw(ctx context.Context, image string) (dockertypes.ImageInspect, []byte, error)
  81         ImageLoad(ctx context.Context, input io.Reader, quiet bool) (dockertypes.ImageLoadResponse, error)
  82         ImageRemove(ctx context.Context, image string, options dockertypes.ImageRemoveOptions) ([]dockertypes.ImageDeleteResponseItem, error)
  83 }
  84
  85 // ContainerRunner is the main stateful struct used for a single execution of a
  86 // container.
  87 type ContainerRunner struct {
  88         Docker    ThinDockerClient
  89         ArvClient IArvadosClient
  90         Kc        IKeepClient
  91         arvados.Container
  92         ContainerConfig dockercontainer.Config
  93         dockercontainer.HostConfig
  94         token       string
  95         ContainerID string
  96         ExitCode    *int
  97         NewLogWriter
  98         loggingDone   chan bool
  99         CrunchLog     *ThrottledLogger
 100         Stdout        io.WriteCloser
 101         Stderr        io.WriteCloser
 102         LogCollection *CollectionWriter
 103         LogsPDH       *string
 104         RunArvMount
 105         MkTempDir
 106         ArvMount       *exec.Cmd
 107         ArvMountPoint  string
 108         HostOutputDir  string
 109         CleanupTempDir []string
 110         Binds          []string
 111         Volumes        map[string]struct{}
 112         OutputPDH      *string
 113         SigChan        chan os.Signal
 114         ArvMountExit   chan error
 115         finalState     string
 116
 117         statLogger   io.WriteCloser
 118         statReporter *crunchstat.Reporter
 119         statInterval time.Duration
 120         cgroupRoot   string
 121         // What we expect the container's cgroup parent to be.
 122         expectCgroupParent string
 123         // What we tell docker to use as the container's cgroup
 124         // parent. Note: Ideally we would use the same field for both
 125         // expectCgroupParent and setCgroupParent, and just make it
 126         // default to "docker". However, when using docker < 1.10 with
 127         // systemd, specifying a non-empty cgroup parent (even the
 128         // default value "docker") hits a docker bug
 129         // (https://github.com/docker/docker/issues/17126). Using two
 130         // separate fields makes it possible to use the "expect cgroup
 131         // parent to be X" feature even on sites where the "specify
 132         // cgroup parent" feature breaks.
 133         setCgroupParent string
 134
 135         cStateLock sync.Mutex
 136         cStarted   bool // StartContainer() succeeded
 137         cCancelled bool // StopContainer() invoked
 138
 139         enableNetwork string // one of "default" or "always"
 140         networkMode   string // passed through to HostConfig.NetworkMode
 141         arvMountLog   *ThrottledLogger
 142 }
 143
 144 // setupSignals sets up signal handling to gracefully terminate the underlying
 145 // Docker container and update state when receiving a TERM, INT or QUIT signal.
 146 func (runner *ContainerRunner) setupSignals() {
 147         runner.SigChan = make(chan os.Signal, 1)
 148         signal.Notify(runner.SigChan, syscall.SIGTERM)
 149         signal.Notify(runner.SigChan, syscall.SIGINT)
 150         signal.Notify(runner.SigChan, syscall.SIGQUIT)
 151
 152         go func(sig chan os.Signal) {
 153                 for s := range sig {
 154                         runner.CrunchLog.Printf("caught signal: %v", s)
 155                         runner.stop()
 156                 }
 157         }(runner.SigChan)
 158 }
 159
 160 // stop the underlying Docker container.
 161 func (runner *ContainerRunner) stop() {
 162         runner.cStateLock.Lock()
 163         defer runner.cStateLock.Unlock()
 164         if !runner.cStarted {
 165                 return
 166         }
 167         runner.cCancelled = true
 168         runner.CrunchLog.Printf("stopping container")
 169         timeout := 10 * time.Second
 170         err := runner.Docker.ContainerStop(context.TODO(), runner.ContainerID, &timeout)
 171         if err != nil {
 172                 runner.CrunchLog.Printf("error stopping container: %s", err)
 173         }
 174 }
 175
 176 func (runner *ContainerRunner) stopSignals() {
 177         if runner.SigChan != nil {
 178                 signal.Stop(runner.SigChan)
 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         runner.arvMountLog = NewThrottledLogger(runner.NewLogWriter("arv-mount"))
 281         c.Stdout = runner.arvMountLog
 282         c.Stderr = runner.arvMountLog
 283
 284         runner.CrunchLog.Printf("Running %v", c.Args)
 285
 286         err = c.Start()
 287         if err != nil {
 288                 return nil, err
 289         }
 290
 291         statReadme := make(chan bool)
 292         runner.ArvMountExit = make(chan error)
 293
 294         keepStatting := true
 295         go func() {
 296                 for keepStatting {
 297                         time.Sleep(100 * time.Millisecond)
 298                         _, err = os.Stat(fmt.Sprintf("%s/by_id/README", runner.ArvMountPoint))
 299                         if err == nil {
 300                                 keepStatting = false
 301                                 statReadme <- true
 302                         }
 303                 }
 304                 close(statReadme)
 305         }()
 306
 307         go func() {
 308                 mnterr := c.Wait()
 309                 if mnterr != nil {
 310                         runner.CrunchLog.Printf("Arv-mount exit error: %v", mnterr)
 311                 }
 312                 runner.ArvMountExit <- mnterr
 313                 close(runner.ArvMountExit)
 314         }()
 315
 316         select {
 317         case <-statReadme:
 318                 break
 319         case err := <-runner.ArvMountExit:
 320                 runner.ArvMount = nil
 321                 keepStatting = false
 322                 return nil, err
 323         }
 324
 325         return c, nil
 326 }
 327
 328 func (runner *ContainerRunner) SetupArvMountPoint(prefix string) (err error) {
 329         if runner.ArvMountPoint == "" {
 330                 runner.ArvMountPoint, err = runner.MkTempDir("", prefix)
 331         }
 332         return
 333 }
 334
 335 func (runner *ContainerRunner) SetupMounts() (err error) {
 336         err = runner.SetupArvMountPoint("keep")
 337         if err != nil {
 338                 return fmt.Errorf("While creating keep mount temp dir: %v", err)
 339         }
 340
 341         token, err := runner.ContainerToken()
 342         if err != nil {
 343                 return fmt.Errorf("could not get container token: %s", err)
 344         }
 345
 346         pdhOnly := true
 347         tmpcount := 0
 348         arvMountCmd := []string{
 349                 "--foreground",
 350                 "--allow-other",
 351                 "--read-write",
 352                 fmt.Sprintf("--crunchstat-interval=%v", runner.statInterval.Seconds())}
 353
 354         if runner.Container.RuntimeConstraints.KeepCacheRAM > 0 {
 355                 arvMountCmd = append(arvMountCmd, "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheRAM))
 356         }
 357
 358         collectionPaths := []string{}
 359         runner.Binds = nil
 360         runner.Volumes = make(map[string]struct{})
 361         needCertMount := true
 362
 363         var binds []string
 364         for bind := range runner.Container.Mounts {
 365                 binds = append(binds, bind)
 366         }
 367         sort.Strings(binds)
 368
 369         for _, bind := range binds {
 370                 mnt := runner.Container.Mounts[bind]
 371                 if bind == "stdout" || bind == "stderr" {
 372                         // Is it a "file" mount kind?
 373                         if mnt.Kind != "file" {
 374                                 return fmt.Errorf("Unsupported mount kind '%s' for %s. Only 'file' is supported.", mnt.Kind, bind)
 375                         }
 376
 377                         // Does path start with OutputPath?
 378                         prefix := runner.Container.OutputPath
 379                         if !strings.HasSuffix(prefix, "/") {
 380                                 prefix += "/"
 381                         }
 382                         if !strings.HasPrefix(mnt.Path, prefix) {
 383                                 return fmt.Errorf("%s path does not start with OutputPath: %s, %s", strings.Title(bind), mnt.Path, prefix)
 384                         }
 385                 }
 386
 387                 if bind == "stdin" {
 388                         // Is it a "collection" mount kind?
 389                         if mnt.Kind != "collection" && mnt.Kind != "json" {
 390                                 return fmt.Errorf("Unsupported mount kind '%s' for stdin. Only 'collection' or 'json' are supported.", mnt.Kind)
 391                         }
 392                 }
 393
 394                 if bind == "/etc/arvados/ca-certificates.crt" {
 395                         needCertMount = false
 396                 }
 397
 398                 if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && bind != runner.Container.OutputPath+"/" {
 399                         if mnt.Kind != "collection" {
 400                                 return fmt.Errorf("Only mount points of kind 'collection' are supported underneath the output_path: %v", bind)
 401                         }
 402                 }
 403
 404                 switch {
 405                 case mnt.Kind == "collection" && bind != "stdin":
 406                         var src string
 407                         if mnt.UUID != "" && mnt.PortableDataHash != "" {
 408                                 return fmt.Errorf("Cannot specify both 'uuid' and 'portable_data_hash' for a collection mount")
 409                         }
 410                         if mnt.UUID != "" {
 411                                 if mnt.Writable {
 412                                         return fmt.Errorf("Writing to existing collections currently not permitted.")
 413                                 }
 414                                 pdhOnly = false
 415                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.UUID)
 416                         } else if mnt.PortableDataHash != "" {
 417                                 if mnt.Writable {
 418                                         return fmt.Errorf("Can never write to a collection specified by portable data hash")
 419                                 }
 420                                 idx := strings.Index(mnt.PortableDataHash, "/")
 421                                 if idx > 0 {
 422                                         mnt.Path = path.Clean(mnt.PortableDataHash[idx:])
 423                                         mnt.PortableDataHash = mnt.PortableDataHash[0:idx]
 424                                         runner.Container.Mounts[bind] = mnt
 425                                 }
 426                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.PortableDataHash)
 427                                 if mnt.Path != "" && mnt.Path != "." {
 428                                         if strings.HasPrefix(mnt.Path, "./") {
 429                                                 mnt.Path = mnt.Path[2:]
 430                                         } else if strings.HasPrefix(mnt.Path, "/") {
 431                                                 mnt.Path = mnt.Path[1:]
 432                                         }
 433                                         src += "/" + mnt.Path
 434                                 }
 435                         } else {
 436                                 src = fmt.Sprintf("%s/tmp%d", runner.ArvMountPoint, tmpcount)
 437                                 arvMountCmd = append(arvMountCmd, "--mount-tmp")
 438                                 arvMountCmd = append(arvMountCmd, fmt.Sprintf("tmp%d", tmpcount))
 439                                 tmpcount += 1
 440                         }
 441                         if mnt.Writable {
 442                                 if bind == runner.Container.OutputPath {
 443                                         runner.HostOutputDir = src
 444                                 } else if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 445                                         return fmt.Errorf("Writable mount points are not permitted underneath the output_path: %v", bind)
 446                                 }
 447                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", src, bind))
 448                         } else {
 449                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", src, bind))
 450                         }
 451                         collectionPaths = append(collectionPaths, src)
 452
 453                 case mnt.Kind == "tmp":
 454                         var tmpdir string
 455                         tmpdir, err = runner.MkTempDir("", "")
 456                         if err != nil {
 457                                 return fmt.Errorf("While creating mount temp dir: %v", err)
 458                         }
 459                         st, staterr := os.Stat(tmpdir)
 460                         if staterr != nil {
 461                                 return fmt.Errorf("While Stat on temp dir: %v", staterr)
 462                         }
 463                         err = os.Chmod(tmpdir, st.Mode()|os.ModeSetgid|0777)
 464                         if staterr != nil {
 465                                 return fmt.Errorf("While Chmod temp dir: %v", err)
 466                         }
 467                         runner.CleanupTempDir = append(runner.CleanupTempDir, tmpdir)
 468                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", tmpdir, bind))
 469                         if bind == runner.Container.OutputPath {
 470                                 runner.HostOutputDir = tmpdir
 471                         }
 472
 473                 case mnt.Kind == "json":
 474                         jsondata, err := json.Marshal(mnt.Content)
 475                         if err != nil {
 476                                 return fmt.Errorf("encoding json data: %v", err)
 477                         }
 478                         // Create a tempdir with a single file
 479                         // (instead of just a tempfile): this way we
 480                         // can ensure the file is world-readable
 481                         // inside the container, without having to
 482                         // make it world-readable on the docker host.
 483                         tmpdir, err := runner.MkTempDir("", "")
 484                         if err != nil {
 485                                 return fmt.Errorf("creating temp dir: %v", err)
 486                         }
 487                         runner.CleanupTempDir = append(runner.CleanupTempDir, tmpdir)
 488                         tmpfn := filepath.Join(tmpdir, "mountdata.json")
 489                         err = ioutil.WriteFile(tmpfn, jsondata, 0644)
 490                         if err != nil {
 491                                 return fmt.Errorf("writing temp file: %v", err)
 492                         }
 493                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", tmpfn, bind))
 494
 495                 case mnt.Kind == "git_tree":
 496                         tmpdir, err := runner.MkTempDir("", "")
 497                         if err != nil {
 498                                 return fmt.Errorf("creating temp dir: %v", err)
 499                         }
 500                         runner.CleanupTempDir = append(runner.CleanupTempDir, tmpdir)
 501                         err = gitMount(mnt).extractTree(runner.ArvClient, tmpdir, token)
 502                         if err != nil {
 503                                 return err
 504                         }
 505                         runner.Binds = append(runner.Binds, tmpdir+":"+bind+":ro")
 506                 }
 507         }
 508
 509         if runner.HostOutputDir == "" {
 510                 return fmt.Errorf("Output path does not correspond to a writable mount point")
 511         }
 512
 513         if wantAPI := runner.Container.RuntimeConstraints.API; needCertMount && wantAPI != nil && *wantAPI {
 514                 for _, certfile := range arvadosclient.CertFiles {
 515                         _, err := os.Stat(certfile)
 516                         if err == nil {
 517                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:/etc/arvados/ca-certificates.crt:ro", certfile))
 518                                 break
 519                         }
 520                 }
 521         }
 522
 523         if pdhOnly {
 524                 arvMountCmd = append(arvMountCmd, "--mount-by-pdh", "by_id")
 525         } else {
 526                 arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_id")
 527         }
 528         arvMountCmd = append(arvMountCmd, runner.ArvMountPoint)
 529
 530         runner.ArvMount, err = runner.RunArvMount(arvMountCmd, token)
 531         if err != nil {
 532                 return fmt.Errorf("While trying to start arv-mount: %v", err)
 533         }
 534
 535         for _, p := range collectionPaths {
 536                 _, err = os.Stat(p)
 537                 if err != nil {
 538                         return fmt.Errorf("While checking that input files exist: %v", err)
 539                 }
 540         }
 541
 542         return nil
 543 }
 544
 545 func (runner *ContainerRunner) ProcessDockerAttach(containerReader io.Reader) {
 546         // Handle docker log protocol
 547         // https://docs.docker.com/engine/reference/api/docker_remote_api_v1.15/#attach-to-a-container
 548
 549         header := make([]byte, 8)
 550         for {
 551                 _, readerr := io.ReadAtLeast(containerReader, header, 8)
 552
 553                 if readerr == nil {
 554                         readsize := int64(header[7]) | (int64(header[6]) << 8) | (int64(header[5]) << 16) | (int64(header[4]) << 24)
 555                         if header[0] == 1 {
 556                                 // stdout
 557                                 _, readerr = io.CopyN(runner.Stdout, containerReader, readsize)
 558                         } else {
 559                                 // stderr
 560                                 _, readerr = io.CopyN(runner.Stderr, containerReader, readsize)
 561                         }
 562                 }
 563
 564                 if readerr != nil {
 565                         if readerr != io.EOF {
 566                                 runner.CrunchLog.Printf("While reading docker logs: %v", readerr)
 567                         }
 568
 569                         closeerr := runner.Stdout.Close()
 570                         if closeerr != nil {
 571                                 runner.CrunchLog.Printf("While closing stdout logs: %v", closeerr)
 572                         }
 573
 574                         closeerr = runner.Stderr.Close()
 575                         if closeerr != nil {
 576                                 runner.CrunchLog.Printf("While closing stderr logs: %v", closeerr)
 577                         }
 578
 579                         if runner.statReporter != nil {
 580                                 runner.statReporter.Stop()
 581                                 closeerr = runner.statLogger.Close()
 582                                 if closeerr != nil {
 583                                         runner.CrunchLog.Printf("While closing crunchstat logs: %v", closeerr)
 584                                 }
 585                         }
 586
 587                         runner.loggingDone <- true
 588                         close(runner.loggingDone)
 589                         return
 590                 }
 591         }
 592 }
 593
 594 func (runner *ContainerRunner) StartCrunchstat() {
 595         runner.statLogger = NewThrottledLogger(runner.NewLogWriter("crunchstat"))
 596         runner.statReporter = &crunchstat.Reporter{
 597                 CID:          runner.ContainerID,
 598                 Logger:       log.New(runner.statLogger, "", 0),
 599                 CgroupParent: runner.expectCgroupParent,
 600                 CgroupRoot:   runner.cgroupRoot,
 601                 PollPeriod:   runner.statInterval,
 602         }
 603         runner.statReporter.Start()
 604 }
 605
 606 type infoCommand struct {
 607         label string
 608         cmd   []string
 609 }
 610
 611 // LogHostInfo logs info about the current host, for debugging and
 612 // accounting purposes. Although it's logged as "node-info", this is
 613 // about the environment where crunch-run is actually running, which
 614 // might differ from what's described in the node record (see
 615 // LogNodeRecord).
 616 func (runner *ContainerRunner) LogHostInfo() (err error) {
 617         w := runner.NewLogWriter("node-info")
 618
 619         commands := []infoCommand{
 620                 {
 621                         label: "Host Information",
 622                         cmd:   []string{"uname", "-a"},
 623                 },
 624                 {
 625                         label: "CPU Information",
 626                         cmd:   []string{"cat", "/proc/cpuinfo"},
 627                 },
 628                 {
 629                         label: "Memory Information",
 630                         cmd:   []string{"cat", "/proc/meminfo"},
 631                 },
 632                 {
 633                         label: "Disk Space",
 634                         cmd:   []string{"df", "-m", "/", os.TempDir()},
 635                 },
 636                 {
 637                         label: "Disk INodes",
 638                         cmd:   []string{"df", "-i", "/", os.TempDir()},
 639                 },
 640         }
 641
 642         // Run commands with informational output to be logged.
 643         for _, command := range commands {
 644                 fmt.Fprintln(w, command.label)
 645                 cmd := exec.Command(command.cmd[0], command.cmd[1:]...)
 646                 cmd.Stdout = w
 647                 cmd.Stderr = w
 648                 if err := cmd.Run(); err != nil {
 649                         err = fmt.Errorf("While running command %q: %v", command.cmd, err)
 650                         fmt.Fprintln(w, err)
 651                         return err
 652                 }
 653                 fmt.Fprintln(w, "")
 654         }
 655
 656         err = w.Close()
 657         if err != nil {
 658                 return fmt.Errorf("While closing node-info logs: %v", err)
 659         }
 660         return nil
 661 }
 662
 663 // LogContainerRecord gets and saves the raw JSON container record from the API server
 664 func (runner *ContainerRunner) LogContainerRecord() error {
 665         logged, err := runner.logAPIResponse("container", "containers", map[string]interface{}{"filters": [][]string{{"uuid", "=", runner.Container.UUID}}}, nil)
 666         if !logged && err == nil {
 667                 err = fmt.Errorf("error: no container record found for %s", runner.Container.UUID)
 668         }
 669         return err
 670 }
 671
 672 // LogNodeRecord logs arvados#node record corresponding to the current host.
 673 func (runner *ContainerRunner) LogNodeRecord() error {
 674         hostname := os.Getenv("SLURMD_NODENAME")
 675         if hostname == "" {
 676                 hostname, _ = os.Hostname()
 677         }
 678         _, err := runner.logAPIResponse("node", "nodes", map[string]interface{}{"filters": [][]string{{"hostname", "=", hostname}}}, func(resp interface{}) {
 679                 // The "info" field has admin-only info when obtained
 680                 // with a privileged token, and should not be logged.
 681                 node, ok := resp.(map[string]interface{})
 682                 if ok {
 683                         delete(node, "info")
 684                 }
 685         })
 686         return err
 687 }
 688
 689 func (runner *ContainerRunner) logAPIResponse(label, path string, params map[string]interface{}, munge func(interface{})) (logged bool, err error) {
 690         w := &ArvLogWriter{
 691                 ArvClient:     runner.ArvClient,
 692                 UUID:          runner.Container.UUID,
 693                 loggingStream: label,
 694                 writeCloser:   runner.LogCollection.Open(label + ".json"),
 695         }
 696
 697         reader, err := runner.ArvClient.CallRaw("GET", path, "", "", arvadosclient.Dict(params))
 698         if err != nil {
 699                 return false, fmt.Errorf("error getting %s record: %v", label, err)
 700         }
 701         defer reader.Close()
 702
 703         dec := json.NewDecoder(reader)
 704         dec.UseNumber()
 705         var resp map[string]interface{}
 706         if err = dec.Decode(&resp); err != nil {
 707                 return false, fmt.Errorf("error decoding %s list response: %v", label, err)
 708         }
 709         items, ok := resp["items"].([]interface{})
 710         if !ok {
 711                 return false, fmt.Errorf("error decoding %s list response: no \"items\" key in API list response", label)
 712         } else if len(items) < 1 {
 713                 return false, nil
 714         }
 715         if munge != nil {
 716                 munge(items[0])
 717         }
 718         // Re-encode it using indentation to improve readability
 719         enc := json.NewEncoder(w)
 720         enc.SetIndent("", "    ")
 721         if err = enc.Encode(items[0]); err != nil {
 722                 return false, fmt.Errorf("error logging %s record: %v", label, err)
 723         }
 724         err = w.Close()
 725         if err != nil {
 726                 return false, fmt.Errorf("error closing %s.json in log collection: %v", label, err)
 727         }
 728         return true, nil
 729 }
 730
 731 // AttachStreams connects the docker container stdin, stdout and stderr logs
 732 // to the Arvados logger which logs to Keep and the API server logs table.
 733 func (runner *ContainerRunner) AttachStreams() (err error) {
 734
 735         runner.CrunchLog.Print("Attaching container streams")
 736
 737         // If stdin mount is provided, attach it to the docker container
 738         var stdinRdr arvados.File
 739         var stdinJson []byte
 740         if stdinMnt, ok := runner.Container.Mounts["stdin"]; ok {
 741                 if stdinMnt.Kind == "collection" {
 742                         var stdinColl arvados.Collection
 743                         collId := stdinMnt.UUID
 744                         if collId == "" {
 745                                 collId = stdinMnt.PortableDataHash
 746                         }
 747                         err = runner.ArvClient.Get("collections", collId, nil, &stdinColl)
 748                         if err != nil {
 749                                 return fmt.Errorf("While getting stding collection: %v", err)
 750                         }
 751
 752                         stdinRdr, err = runner.Kc.ManifestFileReader(manifest.Manifest{Text: stdinColl.ManifestText}, stdinMnt.Path)
 753                         if os.IsNotExist(err) {
 754                                 return fmt.Errorf("stdin collection path not found: %v", stdinMnt.Path)
 755                         } else if err != nil {
 756                                 return fmt.Errorf("While getting stdin collection path %v: %v", stdinMnt.Path, err)
 757                         }
 758                 } else if stdinMnt.Kind == "json" {
 759                         stdinJson, err = json.Marshal(stdinMnt.Content)
 760                         if err != nil {
 761                                 return fmt.Errorf("While encoding stdin json data: %v", err)
 762                         }
 763                 }
 764         }
 765
 766         stdinUsed := stdinRdr != nil || len(stdinJson) != 0
 767         response, err := runner.Docker.ContainerAttach(context.TODO(), runner.ContainerID,
 768                 dockertypes.ContainerAttachOptions{Stream: true, Stdin: stdinUsed, Stdout: true, Stderr: true})
 769         if err != nil {
 770                 return fmt.Errorf("While attaching container stdout/stderr streams: %v", err)
 771         }
 772
 773         runner.loggingDone = make(chan bool)
 774
 775         if stdoutMnt, ok := runner.Container.Mounts["stdout"]; ok {
 776                 stdoutFile, err := runner.getStdoutFile(stdoutMnt.Path)
 777                 if err != nil {
 778                         return err
 779                 }
 780                 runner.Stdout = stdoutFile
 781         } else {
 782                 runner.Stdout = NewThrottledLogger(runner.NewLogWriter("stdout"))
 783         }
 784
 785         if stderrMnt, ok := runner.Container.Mounts["stderr"]; ok {
 786                 stderrFile, err := runner.getStdoutFile(stderrMnt.Path)
 787                 if err != nil {
 788                         return err
 789                 }
 790                 runner.Stderr = stderrFile
 791         } else {
 792                 runner.Stderr = NewThrottledLogger(runner.NewLogWriter("stderr"))
 793         }
 794
 795         if stdinRdr != nil {
 796                 go func() {
 797                         _, err := io.Copy(response.Conn, stdinRdr)
 798                         if err != nil {
 799                                 runner.CrunchLog.Print("While writing stdin collection to docker container %q", err)
 800                                 runner.stop()
 801                         }
 802                         stdinRdr.Close()
 803                         response.CloseWrite()
 804                 }()
 805         } else if len(stdinJson) != 0 {
 806                 go func() {
 807                         _, err := io.Copy(response.Conn, bytes.NewReader(stdinJson))
 808                         if err != nil {
 809                                 runner.CrunchLog.Print("While writing stdin json to docker container %q", err)
 810                                 runner.stop()
 811                         }
 812                         response.CloseWrite()
 813                 }()
 814         }
 815
 816         go runner.ProcessDockerAttach(response.Reader)
 817
 818         return nil
 819 }
 820
 821 func (runner *ContainerRunner) getStdoutFile(mntPath string) (*os.File, error) {
 822         stdoutPath := mntPath[len(runner.Container.OutputPath):]
 823         index := strings.LastIndex(stdoutPath, "/")
 824         if index > 0 {
 825                 subdirs := stdoutPath[:index]
 826                 if subdirs != "" {
 827                         st, err := os.Stat(runner.HostOutputDir)
 828                         if err != nil {
 829                                 return nil, fmt.Errorf("While Stat on temp dir: %v", err)
 830                         }
 831                         stdoutPath := filepath.Join(runner.HostOutputDir, subdirs)
 832                         err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777)
 833                         if err != nil {
 834                                 return nil, fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err)
 835                         }
 836                 }
 837         }
 838         stdoutFile, err := os.Create(filepath.Join(runner.HostOutputDir, stdoutPath))
 839         if err != nil {
 840                 return nil, fmt.Errorf("While creating file %q: %v", stdoutPath, err)
 841         }
 842
 843         return stdoutFile, nil
 844 }
 845
 846 // CreateContainer creates the docker container.
 847 func (runner *ContainerRunner) CreateContainer() error {
 848         runner.CrunchLog.Print("Creating Docker container")
 849
 850         runner.ContainerConfig.Cmd = runner.Container.Command
 851         if runner.Container.Cwd != "." {
 852                 runner.ContainerConfig.WorkingDir = runner.Container.Cwd
 853         }
 854
 855         for k, v := range runner.Container.Environment {
 856                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env, k+"="+v)
 857         }
 858
 859         runner.ContainerConfig.Volumes = runner.Volumes
 860
 861         runner.HostConfig = dockercontainer.HostConfig{
 862                 Binds: runner.Binds,
 863                 LogConfig: dockercontainer.LogConfig{
 864                         Type: "none",
 865                 },
 866                 Resources: dockercontainer.Resources{
 867                         CgroupParent: runner.setCgroupParent,
 868                 },
 869         }
 870
 871         if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
 872                 tok, err := runner.ContainerToken()
 873                 if err != nil {
 874                         return err
 875                 }
 876                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env,
 877                         "ARVADOS_API_TOKEN="+tok,
 878                         "ARVADOS_API_HOST="+os.Getenv("ARVADOS_API_HOST"),
 879                         "ARVADOS_API_HOST_INSECURE="+os.Getenv("ARVADOS_API_HOST_INSECURE"),
 880                 )
 881                 runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
 882         } else {
 883                 if runner.enableNetwork == "always" {
 884                         runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
 885                 } else {
 886                         runner.HostConfig.NetworkMode = dockercontainer.NetworkMode("none")
 887                 }
 888         }
 889
 890         _, stdinUsed := runner.Container.Mounts["stdin"]
 891         runner.ContainerConfig.OpenStdin = stdinUsed
 892         runner.ContainerConfig.StdinOnce = stdinUsed
 893         runner.ContainerConfig.AttachStdin = stdinUsed
 894         runner.ContainerConfig.AttachStdout = true
 895         runner.ContainerConfig.AttachStderr = true
 896
 897         createdBody, err := runner.Docker.ContainerCreate(context.TODO(), &runner.ContainerConfig, &runner.HostConfig, nil, runner.Container.UUID)
 898         if err != nil {
 899                 return fmt.Errorf("While creating container: %v", err)
 900         }
 901
 902         runner.ContainerID = createdBody.ID
 903
 904         return runner.AttachStreams()
 905 }
 906
 907 // StartContainer starts the docker container created by CreateContainer.
 908 func (runner *ContainerRunner) StartContainer() error {
 909         runner.CrunchLog.Printf("Starting Docker container id '%s'", runner.ContainerID)
 910         runner.cStateLock.Lock()
 911         defer runner.cStateLock.Unlock()
 912         if runner.cCancelled {
 913                 return ErrCancelled
 914         }
 915         err := runner.Docker.ContainerStart(context.TODO(), runner.ContainerID,
 916                 dockertypes.ContainerStartOptions{})
 917         if err != nil {
 918                 var advice string
 919                 if m, e := regexp.MatchString("(?ms).*(exec|System error).*(no such file or directory|file not found).*", err.Error()); m && e == nil {
 920                         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])
 921                 }
 922                 return fmt.Errorf("could not start container: %v%s", err, advice)
 923         }
 924         runner.cStarted = true
 925         return nil
 926 }
 927
 928 // WaitFinish waits for the container to terminate, capture the exit code, and
 929 // close the stdout/stderr logging.
 930 func (runner *ContainerRunner) WaitFinish() (err error) {
 931         runner.CrunchLog.Print("Waiting for container to finish")
 932
 933         waitOk, waitErr := runner.Docker.ContainerWait(context.TODO(), runner.ContainerID, "not-running")
 934
 935         go func() {
 936                 <-runner.ArvMountExit
 937                 if runner.cStarted {
 938                         runner.CrunchLog.Printf("arv-mount exited while container is still running.  Stopping container.")
 939                         runner.stop()
 940                 }
 941         }()
 942
 943         var waitBody dockercontainer.ContainerWaitOKBody
 944         select {
 945         case waitBody = <-waitOk:
 946         case err = <-waitErr:
 947         }
 948
 949         // Container isn't running any more
 950         runner.cStarted = false
 951
 952         if err != nil {
 953                 return fmt.Errorf("container wait: %v", err)
 954         }
 955
 956         runner.CrunchLog.Printf("Container exited with code: %v", waitBody.StatusCode)
 957         code := int(waitBody.StatusCode)
 958         runner.ExitCode = &code
 959
 960         // wait for stdout/stderr to complete
 961         <-runner.loggingDone
 962
 963         return nil
 964 }
 965
 966 var ErrNotInOutputDir = fmt.Errorf("Must point to path within the output directory")
 967
 968 func (runner *ContainerRunner) derefOutputSymlink(path string, startinfo os.FileInfo) (tgt string, readlinktgt string, info os.FileInfo, err error) {
 969         // Follow symlinks if necessary
 970         info = startinfo
 971         tgt = path
 972         readlinktgt = ""
 973         nextlink := path
 974         for followed := 0; info.Mode()&os.ModeSymlink != 0; followed++ {
 975                 if followed >= limitFollowSymlinks {
 976                         // Got stuck in a loop or just a pathological number of links, give up.
 977                         err = fmt.Errorf("Followed more than %v symlinks from path %q", limitFollowSymlinks, path)
 978                         return
 979                 }
 980
 981                 readlinktgt, err = os.Readlink(nextlink)
 982                 if err != nil {
 983                         return
 984                 }
 985
 986                 tgt = readlinktgt
 987                 if !strings.HasPrefix(tgt, "/") {
 988                         // Relative symlink, resolve it to host path
 989                         tgt = filepath.Join(filepath.Dir(path), tgt)
 990                 }
 991                 if strings.HasPrefix(tgt, runner.Container.OutputPath+"/") && !strings.HasPrefix(tgt, runner.HostOutputDir+"/") {
 992                         // Absolute symlink to container output path, adjust it to host output path.
 993                         tgt = filepath.Join(runner.HostOutputDir, tgt[len(runner.Container.OutputPath):])
 994                 }
 995                 if !strings.HasPrefix(tgt, runner.HostOutputDir+"/") {
 996                         // After dereferencing, symlink target must either be
 997                         // within output directory, or must point to a
 998                         // collection mount.
 999                         err = ErrNotInOutputDir
1000                         return
1001                 }
1002
1003                 info, err = os.Lstat(tgt)
1004                 if err != nil {
1005                         // tgt
1006                         err = fmt.Errorf("Symlink in output %q points to invalid location %q: %v",
1007                                 path[len(runner.HostOutputDir):], readlinktgt, err)
1008                         return
1009                 }
1010
1011                 nextlink = tgt
1012         }
1013
1014         return
1015 }
1016
1017 var limitFollowSymlinks = 10
1018
1019 // UploadFile uploads files within the output directory, with special handling
1020 // for symlinks. If the symlink leads to a keep mount, copy the manifest text
1021 // from the keep mount into the output manifestText.  Ensure that whether
1022 // symlinks are relative or absolute, every symlink target (even targets that
1023 // are symlinks themselves) must point to a path in either the output directory
1024 // or a collection mount.
1025 //
1026 // Assumes initial value of "path" is absolute, and located within runner.HostOutputDir.
1027 func (runner *ContainerRunner) UploadOutputFile(
1028         path string,
1029         info os.FileInfo,
1030         infoerr error,
1031         binds []string,
1032         walkUpload *WalkUpload,
1033         relocateFrom string,
1034         relocateTo string,
1035         followed int) (manifestText string, err error) {
1036
1037         if infoerr != nil {
1038                 return "", infoerr
1039         }
1040
1041         if info.Mode().IsDir() {
1042                 // if empty, need to create a .keep file
1043                 dir, direrr := os.Open(path)
1044                 if direrr != nil {
1045                         return "", direrr
1046                 }
1047                 defer dir.Close()
1048                 names, eof := dir.Readdirnames(1)
1049                 if len(names) == 0 && eof == io.EOF && path != runner.HostOutputDir {
1050                         containerPath := runner.OutputPath + path[len(runner.HostOutputDir):]
1051                         for _, bind := range binds {
1052                                 mnt := runner.Container.Mounts[bind]
1053                                 // Check if there is a bind for this
1054                                 // directory, in which case assume we don't need .keep
1055                                 if (containerPath == bind || strings.HasPrefix(containerPath, bind+"/")) && mnt.PortableDataHash != "d41d8cd98f00b204e9800998ecf8427e+0" {
1056                                         return
1057                                 }
1058                         }
1059                         outputSuffix := path[len(runner.HostOutputDir)+1:]
1060                         return fmt.Sprintf("./%v d41d8cd98f00b204e9800998ecf8427e+0 0:0:.keep\n", outputSuffix), nil
1061                 }
1062                 return
1063         }
1064
1065         if followed >= limitFollowSymlinks {
1066                 // Got stuck in a loop or just a pathological number of
1067                 // directory links, give up.
1068                 err = fmt.Errorf("Followed more than %v symlinks from path %q", limitFollowSymlinks, path)
1069                 return
1070         }
1071
1072         // "path" is the actual path we are visiting
1073         // "tgt" is the target of "path" (a non-symlink) after following symlinks
1074         // "relocated" is the path in the output manifest where the file should be placed,
1075         // but has HostOutputDir as a prefix.
1076
1077         // The destination path in the output manifest may need to be
1078         // logically relocated to some other path in order to appear
1079         // in the correct location as a result of following a symlink.
1080         // Remove the relocateFrom prefix and replace it with
1081         // relocateTo.
1082         relocated := relocateTo + path[len(relocateFrom):]
1083
1084         tgt, readlinktgt, info, derefErr := runner.derefOutputSymlink(path, info)
1085         if derefErr != nil && derefErr != ErrNotInOutputDir {
1086                 return "", derefErr
1087         }
1088
1089         // go through mounts and try reverse map to collection reference
1090         for _, bind := range binds {
1091                 mnt := runner.Container.Mounts[bind]
1092                 if tgt == bind || strings.HasPrefix(tgt, bind+"/") {
1093                         // get path relative to bind
1094                         targetSuffix := tgt[len(bind):]
1095
1096                         // Copy mount and adjust the path to add path relative to the bind
1097                         adjustedMount := mnt
1098                         adjustedMount.Path = filepath.Join(adjustedMount.Path, targetSuffix)
1099
1100                         // Terminates in this keep mount, so add the
1101                         // manifest text at appropriate location.
1102                         outputSuffix := relocated[len(runner.HostOutputDir):]
1103                         manifestText, err = runner.getCollectionManifestForPath(adjustedMount, outputSuffix)
1104                         return
1105                 }
1106         }
1107
1108         // If target is not a collection mount, it must be located within the
1109         // output directory, otherwise it is an error.
1110         if derefErr == ErrNotInOutputDir {
1111                 err = fmt.Errorf("Symlink in output %q points to invalid location %q, must point to path within the output directory.",
1112                         path[len(runner.HostOutputDir):], readlinktgt)
1113                 return
1114         }
1115
1116         if info.Mode().IsRegular() {
1117                 return "", walkUpload.UploadFile(relocated, tgt)
1118         }
1119
1120         if info.Mode().IsDir() {
1121                 // Symlink leads to directory.  Walk() doesn't follow
1122                 // directory symlinks, so we walk the target directory
1123                 // instead.  Within the walk, file paths are relocated
1124                 // so they appear under the original symlink path.
1125                 err = filepath.Walk(tgt, func(walkpath string, walkinfo os.FileInfo, walkerr error) error {
1126                         var m string
1127                         m, walkerr = runner.UploadOutputFile(walkpath, walkinfo, walkerr,
1128                                 binds, walkUpload, tgt, relocated, followed+1)
1129                         if walkerr == nil {
1130                                 manifestText = manifestText + m
1131                         }
1132                         return walkerr
1133                 })
1134                 return
1135         }
1136
1137         return
1138 }
1139
1140 // HandleOutput sets the output, unmounts the FUSE mount, and deletes temporary directories
1141 func (runner *ContainerRunner) CaptureOutput() error {
1142         if runner.finalState != "Complete" {
1143                 return nil
1144         }
1145
1146         if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
1147                 // Output may have been set directly by the container, so
1148                 // refresh the container record to check.
1149                 err := runner.ArvClient.Get("containers", runner.Container.UUID,
1150                         nil, &runner.Container)
1151                 if err != nil {
1152                         return err
1153                 }
1154                 if runner.Container.Output != "" {
1155                         // Container output is already set.
1156                         runner.OutputPDH = &runner.Container.Output
1157                         return nil
1158                 }
1159         }
1160
1161         if runner.HostOutputDir == "" {
1162                 return nil
1163         }
1164
1165         _, err := os.Stat(runner.HostOutputDir)
1166         if err != nil {
1167                 return fmt.Errorf("While checking host output path: %v", err)
1168         }
1169
1170         // Pre-populate output from the configured mount points
1171         var binds []string
1172         for bind, mnt := range runner.Container.Mounts {
1173                 if mnt.Kind == "collection" {
1174                         binds = append(binds, bind)
1175                 }
1176         }
1177         sort.Strings(binds)
1178
1179         var manifestText string
1180
1181         collectionMetafile := fmt.Sprintf("%s/.arvados#collection", runner.HostOutputDir)
1182         _, err = os.Stat(collectionMetafile)
1183         if err != nil {
1184                 // Regular directory
1185
1186                 cw := CollectionWriter{0, runner.Kc, nil, nil, sync.Mutex{}}
1187                 walkUpload := cw.BeginUpload(runner.HostOutputDir, runner.CrunchLog.Logger)
1188
1189                 var m string
1190                 err = filepath.Walk(runner.HostOutputDir, func(path string, info os.FileInfo, err error) error {
1191                         m, err = runner.UploadOutputFile(path, info, err, binds, walkUpload, "", "", 0)
1192                         if err == nil {
1193                                 manifestText = manifestText + m
1194                         }
1195                         return err
1196                 })
1197
1198                 cw.EndUpload(walkUpload)
1199
1200                 if err != nil {
1201                         return fmt.Errorf("While uploading output files: %v", err)
1202                 }
1203
1204                 m, err = cw.ManifestText()
1205                 manifestText = manifestText + m
1206                 if err != nil {
1207                         return fmt.Errorf("While uploading output files: %v", err)
1208                 }
1209         } else {
1210                 // FUSE mount directory
1211                 file, openerr := os.Open(collectionMetafile)
1212                 if openerr != nil {
1213                         return fmt.Errorf("While opening FUSE metafile: %v", err)
1214                 }
1215                 defer file.Close()
1216
1217                 var rec arvados.Collection
1218                 err = json.NewDecoder(file).Decode(&rec)
1219                 if err != nil {
1220                         return fmt.Errorf("While reading FUSE metafile: %v", err)
1221                 }
1222                 manifestText = rec.ManifestText
1223         }
1224
1225         for _, bind := range binds {
1226                 mnt := runner.Container.Mounts[bind]
1227
1228                 bindSuffix := strings.TrimPrefix(bind, runner.Container.OutputPath)
1229
1230                 if bindSuffix == bind || len(bindSuffix) <= 0 {
1231                         // either does not start with OutputPath or is OutputPath itself
1232                         continue
1233                 }
1234
1235                 if mnt.ExcludeFromOutput == true {
1236                         continue
1237                 }
1238
1239                 // append to manifest_text
1240                 m, err := runner.getCollectionManifestForPath(mnt, bindSuffix)
1241                 if err != nil {
1242                         return err
1243                 }
1244
1245                 manifestText = manifestText + m
1246         }
1247
1248         // Save output
1249         var response arvados.Collection
1250         manifest := manifest.Manifest{Text: manifestText}
1251         manifestText = manifest.Extract(".", ".").Text
1252         err = runner.ArvClient.Create("collections",
1253                 arvadosclient.Dict{
1254                         "ensure_unique_name": true,
1255                         "collection": arvadosclient.Dict{
1256                                 "is_trashed":    true,
1257                                 "name":          "output for " + runner.Container.UUID,
1258                                 "manifest_text": manifestText}},
1259                 &response)
1260         if err != nil {
1261                 return fmt.Errorf("While creating output collection: %v", err)
1262         }
1263         runner.OutputPDH = &response.PortableDataHash
1264         return nil
1265 }
1266
1267 var outputCollections = make(map[string]arvados.Collection)
1268
1269 // Fetch the collection for the mnt.PortableDataHash
1270 // Return the manifest_text fragment corresponding to the specified mnt.Path
1271 //  after making any required updates.
1272 //  Ex:
1273 //    If mnt.Path is not specified,
1274 //      return the entire manifest_text after replacing any "." with bindSuffix
1275 //    If mnt.Path corresponds to one stream,
1276 //      return the manifest_text for that stream after replacing that stream name with bindSuffix
1277 //    Otherwise, check if a filename in any one stream is being sought. Return the manifest_text
1278 //      for that stream after replacing stream name with bindSuffix minus the last word
1279 //      and the file name with last word of the bindSuffix
1280 //  Allowed path examples:
1281 //    "path":"/"
1282 //    "path":"/subdir1"
1283 //    "path":"/subdir1/subdir2"
1284 //    "path":"/subdir/filename" etc
1285 func (runner *ContainerRunner) getCollectionManifestForPath(mnt arvados.Mount, bindSuffix string) (string, error) {
1286         collection := outputCollections[mnt.PortableDataHash]
1287         if collection.PortableDataHash == "" {
1288                 err := runner.ArvClient.Get("collections", mnt.PortableDataHash, nil, &collection)
1289                 if err != nil {
1290                         return "", fmt.Errorf("While getting collection for %v: %v", mnt.PortableDataHash, err)
1291                 }
1292                 outputCollections[mnt.PortableDataHash] = collection
1293         }
1294
1295         if collection.ManifestText == "" {
1296                 runner.CrunchLog.Printf("No manifest text for collection %v", collection.PortableDataHash)
1297                 return "", nil
1298         }
1299
1300         mft := manifest.Manifest{Text: collection.ManifestText}
1301         extracted := mft.Extract(mnt.Path, bindSuffix)
1302         if extracted.Err != nil {
1303                 return "", fmt.Errorf("Error parsing manifest for %v: %v", mnt.PortableDataHash, extracted.Err.Error())
1304         }
1305         return extracted.Text, nil
1306 }
1307
1308 func (runner *ContainerRunner) CleanupDirs() {
1309         if runner.ArvMount != nil {
1310                 var delay int64 = 8
1311                 umount := exec.Command("arv-mount", fmt.Sprintf("--unmount-timeout=%d", delay), "--unmount", runner.ArvMountPoint)
1312                 umount.Stdout = runner.CrunchLog
1313                 umount.Stderr = runner.CrunchLog
1314                 runner.CrunchLog.Printf("Running %v", umount.Args)
1315                 umnterr := umount.Start()
1316
1317                 if umnterr != nil {
1318                         runner.CrunchLog.Printf("Error unmounting: %v", umnterr)
1319                 } else {
1320                         // If arv-mount --unmount gets stuck for any reason, we
1321                         // don't want to wait for it forever.  Do Wait() in a goroutine
1322                         // so it doesn't block crunch-run.
1323                         umountExit := make(chan error)
1324                         go func() {
1325                                 mnterr := umount.Wait()
1326                                 if mnterr != nil {
1327                                         runner.CrunchLog.Printf("Error unmounting: %v", mnterr)
1328                                 }
1329                                 umountExit <- mnterr
1330                         }()
1331
1332                         for again := true; again; {
1333                                 again = false
1334                                 select {
1335                                 case <-umountExit:
1336                                         umount = nil
1337                                         again = true
1338                                 case <-runner.ArvMountExit:
1339                                         break
1340                                 case <-time.After(time.Duration((delay + 1) * int64(time.Second))):
1341                                         runner.CrunchLog.Printf("Timed out waiting for unmount")
1342                                         if umount != nil {
1343                                                 umount.Process.Kill()
1344                                         }
1345                                         runner.ArvMount.Process.Kill()
1346                                 }
1347                         }
1348                 }
1349         }
1350
1351         if runner.ArvMountPoint != "" {
1352                 if rmerr := os.Remove(runner.ArvMountPoint); rmerr != nil {
1353                         runner.CrunchLog.Printf("While cleaning up arv-mount directory %s: %v", runner.ArvMountPoint, rmerr)
1354                 }
1355         }
1356
1357         for _, tmpdir := range runner.CleanupTempDir {
1358                 if rmerr := os.RemoveAll(tmpdir); rmerr != nil {
1359                         runner.CrunchLog.Printf("While cleaning up temporary directory %s: %v", tmpdir, rmerr)
1360                 }
1361         }
1362 }
1363
1364 // CommitLogs posts the collection containing the final container logs.
1365 func (runner *ContainerRunner) CommitLogs() error {
1366         runner.CrunchLog.Print(runner.finalState)
1367
1368         if runner.arvMountLog != nil {
1369                 runner.arvMountLog.Close()
1370         }
1371         runner.CrunchLog.Close()
1372
1373         // Closing CrunchLog above allows them to be committed to Keep at this
1374         // point, but re-open crunch log with ArvClient in case there are any
1375         // other further errors (such as failing to write the log to Keep!)
1376         // while shutting down
1377         runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{ArvClient: runner.ArvClient,
1378                 UUID: runner.Container.UUID, loggingStream: "crunch-run", writeCloser: nil})
1379         runner.CrunchLog.Immediate = log.New(os.Stderr, runner.Container.UUID+" ", 0)
1380
1381         if runner.LogsPDH != nil {
1382                 // If we have already assigned something to LogsPDH,
1383                 // we must be closing the re-opened log, which won't
1384                 // end up getting attached to the container record and
1385                 // therefore doesn't need to be saved as a collection
1386                 // -- it exists only to send logs to other channels.
1387                 return nil
1388         }
1389
1390         mt, err := runner.LogCollection.ManifestText()
1391         if err != nil {
1392                 return fmt.Errorf("While creating log manifest: %v", err)
1393         }
1394
1395         var response arvados.Collection
1396         err = runner.ArvClient.Create("collections",
1397                 arvadosclient.Dict{
1398                         "ensure_unique_name": true,
1399                         "collection": arvadosclient.Dict{
1400                                 "is_trashed":    true,
1401                                 "name":          "logs for " + runner.Container.UUID,
1402                                 "manifest_text": mt}},
1403                 &response)
1404         if err != nil {
1405                 return fmt.Errorf("While creating log collection: %v", err)
1406         }
1407         runner.LogsPDH = &response.PortableDataHash
1408         return nil
1409 }
1410
1411 // UpdateContainerRunning updates the container state to "Running"
1412 func (runner *ContainerRunner) UpdateContainerRunning() error {
1413         runner.cStateLock.Lock()
1414         defer runner.cStateLock.Unlock()
1415         if runner.cCancelled {
1416                 return ErrCancelled
1417         }
1418         return runner.ArvClient.Update("containers", runner.Container.UUID,
1419                 arvadosclient.Dict{"container": arvadosclient.Dict{"state": "Running"}}, nil)
1420 }
1421
1422 // ContainerToken returns the api_token the container (and any
1423 // arv-mount processes) are allowed to use.
1424 func (runner *ContainerRunner) ContainerToken() (string, error) {
1425         if runner.token != "" {
1426                 return runner.token, nil
1427         }
1428
1429         var auth arvados.APIClientAuthorization
1430         err := runner.ArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth)
1431         if err != nil {
1432                 return "", err
1433         }
1434         runner.token = auth.APIToken
1435         return runner.token, nil
1436 }
1437
1438 // UpdateContainerComplete updates the container record state on API
1439 // server to "Complete" or "Cancelled"
1440 func (runner *ContainerRunner) UpdateContainerFinal() error {
1441         update := arvadosclient.Dict{}
1442         update["state"] = runner.finalState
1443         if runner.LogsPDH != nil {
1444                 update["log"] = *runner.LogsPDH
1445         }
1446         if runner.finalState == "Complete" {
1447                 if runner.ExitCode != nil {
1448                         update["exit_code"] = *runner.ExitCode
1449                 }
1450                 if runner.OutputPDH != nil {
1451                         update["output"] = *runner.OutputPDH
1452                 }
1453         }
1454         return runner.ArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{"container": update}, nil)
1455 }
1456
1457 // IsCancelled returns the value of Cancelled, with goroutine safety.
1458 func (runner *ContainerRunner) IsCancelled() bool {
1459         runner.cStateLock.Lock()
1460         defer runner.cStateLock.Unlock()
1461         return runner.cCancelled
1462 }
1463
1464 // NewArvLogWriter creates an ArvLogWriter
1465 func (runner *ContainerRunner) NewArvLogWriter(name string) io.WriteCloser {
1466         return &ArvLogWriter{
1467                 ArvClient:     runner.ArvClient,
1468                 UUID:          runner.Container.UUID,
1469                 loggingStream: name,
1470                 writeCloser:   runner.LogCollection.Open(name + ".txt")}
1471 }
1472
1473 // Run the full container lifecycle.
1474 func (runner *ContainerRunner) Run() (err error) {
1475         runner.CrunchLog.Printf("crunch-run %s started", version)
1476         runner.CrunchLog.Printf("Executing container '%s'", runner.Container.UUID)
1477
1478         hostname, hosterr := os.Hostname()
1479         if hosterr != nil {
1480                 runner.CrunchLog.Printf("Error getting hostname '%v'", hosterr)
1481         } else {
1482                 runner.CrunchLog.Printf("Executing on host '%s'", hostname)
1483         }
1484
1485         runner.finalState = "Queued"
1486
1487         defer func() {
1488                 runner.stopSignals()
1489                 runner.CleanupDirs()
1490
1491                 runner.CrunchLog.Printf("crunch-run finished")
1492                 runner.CrunchLog.Close()
1493         }()
1494
1495         defer func() {
1496                 // checkErr prints e (unless it's nil) and sets err to
1497                 // e (unless err is already non-nil). Thus, if err
1498                 // hasn't already been assigned when Run() returns,
1499                 // this cleanup func will cause Run() to return the
1500                 // first non-nil error that is passed to checkErr().
1501                 checkErr := func(e error) {
1502                         if e == nil {
1503                                 return
1504                         }
1505                         runner.CrunchLog.Print(e)
1506                         if err == nil {
1507                                 err = e
1508                         }
1509                         if runner.finalState == "Complete" {
1510                                 // There was an error in the finalization.
1511                                 runner.finalState = "Cancelled"
1512                         }
1513                 }
1514
1515                 // Log the error encountered in Run(), if any
1516                 checkErr(err)
1517
1518                 if runner.finalState == "Queued" {
1519                         runner.UpdateContainerFinal()
1520                         return
1521                 }
1522
1523                 if runner.IsCancelled() {
1524                         runner.finalState = "Cancelled"
1525                         // but don't return yet -- we still want to
1526                         // capture partial output and write logs
1527                 }
1528
1529                 checkErr(runner.CaptureOutput())
1530                 checkErr(runner.CommitLogs())
1531                 checkErr(runner.UpdateContainerFinal())
1532         }()
1533
1534         err = runner.fetchContainerRecord()
1535         if err != nil {
1536                 return
1537         }
1538
1539         // setup signal handling
1540         runner.setupSignals()
1541
1542         // check for and/or load image
1543         err = runner.LoadImage()
1544         if err != nil {
1545                 if !runner.checkBrokenNode(err) {
1546                         // Failed to load image but not due to a "broken node"
1547                         // condition, probably user error.
1548                         runner.finalState = "Cancelled"
1549                 }
1550                 err = fmt.Errorf("While loading container image: %v", err)
1551                 return
1552         }
1553
1554         // set up FUSE mount and binds
1555         err = runner.SetupMounts()
1556         if err != nil {
1557                 runner.finalState = "Cancelled"
1558                 err = fmt.Errorf("While setting up mounts: %v", err)
1559                 return
1560         }
1561
1562         err = runner.CreateContainer()
1563         if err != nil {
1564                 return
1565         }
1566         err = runner.LogHostInfo()
1567         if err != nil {
1568                 return
1569         }
1570         err = runner.LogNodeRecord()
1571         if err != nil {
1572                 return
1573         }
1574         err = runner.LogContainerRecord()
1575         if err != nil {
1576                 return
1577         }
1578
1579         if runner.IsCancelled() {
1580                 return
1581         }
1582
1583         err = runner.UpdateContainerRunning()
1584         if err != nil {
1585                 return
1586         }
1587         runner.finalState = "Cancelled"
1588
1589         runner.StartCrunchstat()
1590
1591         err = runner.StartContainer()
1592         if err != nil {
1593                 runner.checkBrokenNode(err)
1594                 return
1595         }
1596
1597         err = runner.WaitFinish()
1598         if err == nil {
1599                 runner.finalState = "Complete"
1600         }
1601         return
1602 }
1603
1604 // Fetch the current container record (uuid = runner.Container.UUID)
1605 // into runner.Container.
1606 func (runner *ContainerRunner) fetchContainerRecord() error {
1607         reader, err := runner.ArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
1608         if err != nil {
1609                 return fmt.Errorf("error fetching container record: %v", err)
1610         }
1611         defer reader.Close()
1612
1613         dec := json.NewDecoder(reader)
1614         dec.UseNumber()
1615         err = dec.Decode(&runner.Container)
1616         if err != nil {
1617                 return fmt.Errorf("error decoding container record: %v", err)
1618         }
1619         return nil
1620 }
1621
1622 // NewContainerRunner creates a new container runner.
1623 func NewContainerRunner(api IArvadosClient,
1624         kc IKeepClient,
1625         docker ThinDockerClient,
1626         containerUUID string) *ContainerRunner {
1627
1628         cr := &ContainerRunner{ArvClient: api, Kc: kc, Docker: docker}
1629         cr.NewLogWriter = cr.NewArvLogWriter
1630         cr.RunArvMount = cr.ArvMountCmd
1631         cr.MkTempDir = ioutil.TempDir
1632         cr.LogCollection = &CollectionWriter{0, kc, nil, nil, sync.Mutex{}}
1633         cr.Container.UUID = containerUUID
1634         cr.CrunchLog = NewThrottledLogger(cr.NewLogWriter("crunch-run"))
1635         cr.CrunchLog.Immediate = log.New(os.Stderr, containerUUID+" ", 0)
1636
1637         loadLogThrottleParams(api)
1638
1639         return cr
1640 }
1641
1642 func main() {
1643         statInterval := flag.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting")
1644         cgroupRoot := flag.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree")
1645         cgroupParent := flag.String("cgroup-parent", "docker", "name of container's parent cgroup (ignored if -cgroup-parent-subsystem is used)")
1646         cgroupParentSubsystem := flag.String("cgroup-parent-subsystem", "", "use current cgroup for given subsystem as parent cgroup for container")
1647         caCertsPath := flag.String("ca-certs", "", "Path to TLS root certificates")
1648         enableNetwork := flag.String("container-enable-networking", "default",
1649                 `Specify if networking should be enabled for container.  One of 'default', 'always':
1650         default: only enable networking if container requests it.
1651         always:  containers always have networking enabled
1652         `)
1653         networkMode := flag.String("container-network-mode", "default",
1654                 `Set networking mode for container.  Corresponds to Docker network mode (--net).
1655         `)
1656         memprofile := flag.String("memprofile", "", "write memory profile to `file` after running container")
1657         getVersion := flag.Bool("version", false, "Print version information and exit.")
1658         flag.Parse()
1659
1660         // Print version information if requested
1661         if *getVersion {
1662                 fmt.Printf("crunch-run %s\n", version)
1663                 return
1664         }
1665
1666         log.Printf("crunch-run %s started", version)
1667
1668         containerId := flag.Arg(0)
1669
1670         if *caCertsPath != "" {
1671                 arvadosclient.CertFiles = []string{*caCertsPath}
1672         }
1673
1674         api, err := arvadosclient.MakeArvadosClient()
1675         if err != nil {
1676                 log.Fatalf("%s: %v", containerId, err)
1677         }
1678         api.Retries = 8
1679
1680         kc, kcerr := keepclient.MakeKeepClient(api)
1681         if kcerr != nil {
1682                 log.Fatalf("%s: %v", containerId, kcerr)
1683         }
1684         kc.BlockCache = &keepclient.BlockCache{MaxBlocks: 2}
1685         kc.Retries = 4
1686
1687         // API version 1.21 corresponds to Docker 1.9, which is currently the
1688         // minimum version we want to support.
1689         docker, dockererr := dockerclient.NewClient(dockerclient.DefaultDockerHost, "1.21", nil, nil)
1690
1691         cr := NewContainerRunner(api, kc, docker, containerId)
1692         if dockererr != nil {
1693                 cr.CrunchLog.Printf("%s: %v", containerId, dockererr)
1694                 cr.checkBrokenNode(dockererr)
1695                 cr.CrunchLog.Close()
1696                 os.Exit(1)
1697         }
1698
1699         cr.statInterval = *statInterval
1700         cr.cgroupRoot = *cgroupRoot
1701         cr.expectCgroupParent = *cgroupParent
1702         cr.enableNetwork = *enableNetwork
1703         cr.networkMode = *networkMode
1704         if *cgroupParentSubsystem != "" {
1705                 p := findCgroup(*cgroupParentSubsystem)
1706                 cr.setCgroupParent = p
1707                 cr.expectCgroupParent = p
1708         }
1709
1710         runerr := cr.Run()
1711
1712         if *memprofile != "" {
1713                 f, err := os.Create(*memprofile)
1714                 if err != nil {
1715                         log.Printf("could not create memory profile: ", err)
1716                 }
1717                 runtime.GC() // get up-to-date statistics
1718                 if err := pprof.WriteHeapProfile(f); err != nil {
1719                         log.Printf("could not write memory profile: ", err)
1720                 }
1721                 closeerr := f.Close()
1722                 if closeerr != nil {
1723                         log.Printf("closing memprofile file: ", err)
1724                 }
1725         }
1726
1727         if runerr != nil {
1728                 log.Fatalf("%s: %v", containerId, runerr)
1729         }
1730 }