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         "context"
  10         "encoding/json"
  11         "errors"
  12         "flag"
  13         "fmt"
  14         "io"
  15         "io/ioutil"
  16         "log"
  17         "os"
  18         "os/exec"
  19         "os/signal"
  20         "path"
  21         "path/filepath"
  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
  36         dockertypes "github.com/docker/docker/api/types"
  37         dockercontainer "github.com/docker/docker/api/types/container"
  38         dockernetwork "github.com/docker/docker/api/types/network"
  39         dockerclient "github.com/docker/docker/client"
  40 )
  41
  42 // IArvadosClient is the minimal Arvados API methods used by crunch-run.
  43 type IArvadosClient interface {
  44         Create(resourceType string, parameters arvadosclient.Dict, output interface{}) error
  45         Get(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  46         Update(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  47         Call(method, resourceType, uuid, action string, parameters arvadosclient.Dict, output interface{}) error
  48         CallRaw(method string, resourceType string, uuid string, action string, parameters arvadosclient.Dict) (reader io.ReadCloser, err error)
  49         Discovery(key string) (interface{}, error)
  50 }
  51
  52 // ErrCancelled is the error returned when the container is cancelled.
  53 var ErrCancelled = errors.New("Cancelled")
  54
  55 // IKeepClient is the minimal Keep API methods used by crunch-run.
  56 type IKeepClient interface {
  57         PutHB(hash string, buf []byte) (string, int, error)
  58         ManifestFileReader(m manifest.Manifest, filename string) (arvados.File, error)
  59         ClearBlockCache()
  60 }
  61
  62 // NewLogWriter is a factory function to create a new log writer.
  63 type NewLogWriter func(name string) io.WriteCloser
  64
  65 type RunArvMount func(args []string, tok string) (*exec.Cmd, error)
  66
  67 type MkTempDir func(string, string) (string, error)
  68
  69 // ThinDockerClient is the minimal Docker client interface used by crunch-run.
  70 type ThinDockerClient interface {
  71         ContainerAttach(ctx context.Context, container string, options dockertypes.ContainerAttachOptions) (dockertypes.HijackedResponse, error)
  72         ContainerCreate(ctx context.Context, config *dockercontainer.Config, hostConfig *dockercontainer.HostConfig,
  73                 networkingConfig *dockernetwork.NetworkingConfig, containerName string) (dockercontainer.ContainerCreateCreatedBody, error)
  74         ContainerStart(ctx context.Context, container string, options dockertypes.ContainerStartOptions) error
  75         ContainerStop(ctx context.Context, container string, timeout *time.Duration) error
  76         ContainerWait(ctx context.Context, container string, condition dockercontainer.WaitCondition) (<-chan dockercontainer.ContainerWaitOKBody, <-chan error)
  77         ImageInspectWithRaw(ctx context.Context, image string) (dockertypes.ImageInspect, []byte, error)
  78         ImageLoad(ctx context.Context, input io.Reader, quiet bool) (dockertypes.ImageLoadResponse, error)
  79         ImageRemove(ctx context.Context, image string, options dockertypes.ImageRemoveOptions) ([]dockertypes.ImageDeleteResponseItem, error)
  80 }
  81
  82 // ThinDockerClientProxy is a proxy implementation of ThinDockerClient
  83 // that executes the docker requests on dockerclient.Client
  84 type ThinDockerClientProxy struct {
  85         Docker *dockerclient.Client
  86 }
  87
  88 // ContainerAttach invokes dockerclient.Client.ContainerAttach
  89 func (proxy ThinDockerClientProxy) ContainerAttach(ctx context.Context, container string, options dockertypes.ContainerAttachOptions) (dockertypes.HijackedResponse, error) {
  90         return proxy.Docker.ContainerAttach(ctx, container, options)
  91 }
  92
  93 // ContainerCreate invokes dockerclient.Client.ContainerCreate
  94 func (proxy ThinDockerClientProxy) ContainerCreate(ctx context.Context, config *dockercontainer.Config, hostConfig *dockercontainer.HostConfig,
  95         networkingConfig *dockernetwork.NetworkingConfig, containerName string) (dockercontainer.ContainerCreateCreatedBody, error) {
  96         return proxy.Docker.ContainerCreate(ctx, config, hostConfig, networkingConfig, containerName)
  97 }
  98
  99 // ContainerStart invokes dockerclient.Client.ContainerStart
 100 func (proxy ThinDockerClientProxy) ContainerStart(ctx context.Context, container string, options dockertypes.ContainerStartOptions) error {
 101         return proxy.Docker.ContainerStart(ctx, container, options)
 102 }
 103
 104 // ContainerStop invokes dockerclient.Client.ContainerStop
 105 func (proxy ThinDockerClientProxy) ContainerStop(ctx context.Context, container string, timeout *time.Duration) error {
 106         return proxy.Docker.ContainerStop(ctx, container, timeout)
 107 }
 108
 109 // ContainerWait invokes dockerclient.Client.ContainerWait
 110 func (proxy ThinDockerClientProxy) ContainerWait(ctx context.Context, container string, condition dockercontainer.WaitCondition) (<-chan dockercontainer.ContainerWaitOKBody, <-chan error) {
 111         return proxy.Docker.ContainerWait(ctx, container, condition)
 112 }
 113
 114 // ImageInspectWithRaw invokes dockerclient.Client.ImageInspectWithRaw
 115 func (proxy ThinDockerClientProxy) ImageInspectWithRaw(ctx context.Context, image string) (dockertypes.ImageInspect, []byte, error) {
 116         return proxy.Docker.ImageInspectWithRaw(ctx, image)
 117 }
 118
 119 // ImageLoad invokes dockerclient.Client.ImageLoad
 120 func (proxy ThinDockerClientProxy) ImageLoad(ctx context.Context, input io.Reader, quiet bool) (dockertypes.ImageLoadResponse, error) {
 121         return proxy.Docker.ImageLoad(ctx, input, quiet)
 122 }
 123
 124 // ImageRemove invokes dockerclient.Client.ImageRemove
 125 func (proxy ThinDockerClientProxy) ImageRemove(ctx context.Context, image string, options dockertypes.ImageRemoveOptions) ([]dockertypes.ImageDeleteResponseItem, error) {
 126         return proxy.Docker.ImageRemove(ctx, image, options)
 127 }
 128
 129 // ContainerRunner is the main stateful struct used for a single execution of a
 130 // container.
 131 type ContainerRunner struct {
 132         Docker    ThinDockerClient
 133         ArvClient IArvadosClient
 134         Kc        IKeepClient
 135         arvados.Container
 136         ContainerConfig dockercontainer.Config
 137         dockercontainer.HostConfig
 138         token       string
 139         ContainerID string
 140         ExitCode    *int
 141         NewLogWriter
 142         loggingDone   chan bool
 143         CrunchLog     *ThrottledLogger
 144         Stdout        io.WriteCloser
 145         Stderr        io.WriteCloser
 146         LogCollection *CollectionWriter
 147         LogsPDH       *string
 148         RunArvMount
 149         MkTempDir
 150         ArvMount       *exec.Cmd
 151         ArvMountPoint  string
 152         HostOutputDir  string
 153         CleanupTempDir []string
 154         Binds          []string
 155         Volumes        map[string]struct{}
 156         OutputPDH      *string
 157         SigChan        chan os.Signal
 158         ArvMountExit   chan error
 159         finalState     string
 160
 161         statLogger   io.WriteCloser
 162         statReporter *crunchstat.Reporter
 163         statInterval time.Duration
 164         cgroupRoot   string
 165         // What we expect the container's cgroup parent to be.
 166         expectCgroupParent string
 167         // What we tell docker to use as the container's cgroup
 168         // parent. Note: Ideally we would use the same field for both
 169         // expectCgroupParent and setCgroupParent, and just make it
 170         // default to "docker". However, when using docker < 1.10 with
 171         // systemd, specifying a non-empty cgroup parent (even the
 172         // default value "docker") hits a docker bug
 173         // (https://github.com/docker/docker/issues/17126). Using two
 174         // separate fields makes it possible to use the "expect cgroup
 175         // parent to be X" feature even on sites where the "specify
 176         // cgroup parent" feature breaks.
 177         setCgroupParent string
 178
 179         cStateLock sync.Mutex
 180         cStarted   bool // StartContainer() succeeded
 181         cCancelled bool // StopContainer() invoked
 182
 183         enableNetwork string // one of "default" or "always"
 184         networkMode   string // passed through to HostConfig.NetworkMode
 185         arvMountLog   *ThrottledLogger
 186 }
 187
 188 // setupSignals sets up signal handling to gracefully terminate the underlying
 189 // Docker container and update state when receiving a TERM, INT or QUIT signal.
 190 func (runner *ContainerRunner) setupSignals() {
 191         runner.SigChan = make(chan os.Signal, 1)
 192         signal.Notify(runner.SigChan, syscall.SIGTERM)
 193         signal.Notify(runner.SigChan, syscall.SIGINT)
 194         signal.Notify(runner.SigChan, syscall.SIGQUIT)
 195
 196         go func(sig chan os.Signal) {
 197                 s := <-sig
 198                 if s != nil {
 199                         runner.CrunchLog.Printf("Caught signal %v", s)
 200                 }
 201                 runner.stop()
 202         }(runner.SigChan)
 203 }
 204
 205 // stop the underlying Docker container.
 206 func (runner *ContainerRunner) stop() {
 207         runner.cStateLock.Lock()
 208         defer runner.cStateLock.Unlock()
 209         if runner.cCancelled {
 210                 return
 211         }
 212         runner.cCancelled = true
 213         if runner.cStarted {
 214                 timeout := time.Duration(10)
 215                 err := runner.Docker.ContainerStop(context.TODO(), runner.ContainerID, &(timeout))
 216                 if err != nil {
 217                         runner.CrunchLog.Printf("StopContainer failed: %s", err)
 218                 }
 219                 // Suppress multiple calls to stop()
 220                 runner.cStarted = false
 221         }
 222 }
 223
 224 func (runner *ContainerRunner) stopSignals() {
 225         if runner.SigChan != nil {
 226                 signal.Stop(runner.SigChan)
 227                 close(runner.SigChan)
 228         }
 229 }
 230
 231 // LoadImage determines the docker image id from the container record and
 232 // checks if it is available in the local Docker image store.  If not, it loads
 233 // the image from Keep.
 234 func (runner *ContainerRunner) LoadImage() (err error) {
 235
 236         runner.CrunchLog.Printf("Fetching Docker image from collection '%s'", runner.Container.ContainerImage)
 237
 238         var collection arvados.Collection
 239         err = runner.ArvClient.Get("collections", runner.Container.ContainerImage, nil, &collection)
 240         if err != nil {
 241                 return fmt.Errorf("While getting container image collection: %v", err)
 242         }
 243         manifest := manifest.Manifest{Text: collection.ManifestText}
 244         var img, imageID string
 245         for ms := range manifest.StreamIter() {
 246                 img = ms.FileStreamSegments[0].Name
 247                 if !strings.HasSuffix(img, ".tar") {
 248                         return fmt.Errorf("First file in the container image collection does not end in .tar")
 249                 }
 250                 imageID = img[:len(img)-4]
 251         }
 252
 253         runner.CrunchLog.Printf("Using Docker image id '%s'", imageID)
 254
 255         _, _, err = runner.Docker.ImageInspectWithRaw(context.TODO(), imageID)
 256         if err != nil {
 257                 runner.CrunchLog.Print("Loading Docker image from keep")
 258
 259                 var readCloser io.ReadCloser
 260                 readCloser, err = runner.Kc.ManifestFileReader(manifest, img)
 261                 if err != nil {
 262                         return fmt.Errorf("While creating ManifestFileReader for container image: %v", err)
 263                 }
 264
 265                 response, err := runner.Docker.ImageLoad(context.TODO(), readCloser, true)
 266                 if err != nil {
 267                         return fmt.Errorf("While loading container image into Docker: %v", err)
 268                 }
 269
 270                 defer response.Body.Close()
 271                 rbody, err := ioutil.ReadAll(response.Body)
 272                 if err != nil {
 273                         return fmt.Errorf("Reading response to image load: %v", err)
 274                 }
 275                 runner.CrunchLog.Printf("Docker response: %s", rbody)
 276         } else {
 277                 runner.CrunchLog.Print("Docker image is available")
 278         }
 279
 280         runner.ContainerConfig.Image = imageID
 281
 282         runner.Kc.ClearBlockCache()
 283
 284         return nil
 285 }
 286
 287 func (runner *ContainerRunner) ArvMountCmd(arvMountCmd []string, token string) (c *exec.Cmd, err error) {
 288         c = exec.Command("arv-mount", arvMountCmd...)
 289
 290         // Copy our environment, but override ARVADOS_API_TOKEN with
 291         // the container auth token.
 292         c.Env = nil
 293         for _, s := range os.Environ() {
 294                 if !strings.HasPrefix(s, "ARVADOS_API_TOKEN=") {
 295                         c.Env = append(c.Env, s)
 296                 }
 297         }
 298         c.Env = append(c.Env, "ARVADOS_API_TOKEN="+token)
 299
 300         runner.arvMountLog = NewThrottledLogger(runner.NewLogWriter("arv-mount"))
 301         c.Stdout = runner.arvMountLog
 302         c.Stderr = runner.arvMountLog
 303
 304         runner.CrunchLog.Printf("Running %v", c.Args)
 305
 306         err = c.Start()
 307         if err != nil {
 308                 return nil, err
 309         }
 310
 311         statReadme := make(chan bool)
 312         runner.ArvMountExit = make(chan error)
 313
 314         keepStatting := true
 315         go func() {
 316                 for keepStatting {
 317                         time.Sleep(100 * time.Millisecond)
 318                         _, err = os.Stat(fmt.Sprintf("%s/by_id/README", runner.ArvMountPoint))
 319                         if err == nil {
 320                                 keepStatting = false
 321                                 statReadme <- true
 322                         }
 323                 }
 324                 close(statReadme)
 325         }()
 326
 327         go func() {
 328                 mnterr := c.Wait()
 329                 if mnterr != nil {
 330                         runner.CrunchLog.Printf("Arv-mount exit error: %v", mnterr)
 331                 }
 332                 runner.ArvMountExit <- mnterr
 333                 close(runner.ArvMountExit)
 334         }()
 335
 336         select {
 337         case <-statReadme:
 338                 break
 339         case err := <-runner.ArvMountExit:
 340                 runner.ArvMount = nil
 341                 keepStatting = false
 342                 return nil, err
 343         }
 344
 345         return c, nil
 346 }
 347
 348 func (runner *ContainerRunner) SetupArvMountPoint(prefix string) (err error) {
 349         if runner.ArvMountPoint == "" {
 350                 runner.ArvMountPoint, err = runner.MkTempDir("", prefix)
 351         }
 352         return
 353 }
 354
 355 func (runner *ContainerRunner) SetupMounts() (err error) {
 356         err = runner.SetupArvMountPoint("keep")
 357         if err != nil {
 358                 return fmt.Errorf("While creating keep mount temp dir: %v", err)
 359         }
 360
 361         pdhOnly := true
 362         tmpcount := 0
 363         arvMountCmd := []string{
 364                 "--foreground",
 365                 "--allow-other",
 366                 "--read-write",
 367                 fmt.Sprintf("--crunchstat-interval=%v", runner.statInterval.Seconds())}
 368
 369         if runner.Container.RuntimeConstraints.KeepCacheRAM > 0 {
 370                 arvMountCmd = append(arvMountCmd, "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheRAM))
 371         }
 372
 373         collectionPaths := []string{}
 374         runner.Binds = nil
 375         runner.Volumes = make(map[string]struct{})
 376         needCertMount := true
 377
 378         var binds []string
 379         for bind := range runner.Container.Mounts {
 380                 binds = append(binds, bind)
 381         }
 382         sort.Strings(binds)
 383
 384         for _, bind := range binds {
 385                 mnt := runner.Container.Mounts[bind]
 386                 if bind == "stdout" || bind == "stderr" {
 387                         // Is it a "file" mount kind?
 388                         if mnt.Kind != "file" {
 389                                 return fmt.Errorf("Unsupported mount kind '%s' for %s. Only 'file' is supported.", mnt.Kind, bind)
 390                         }
 391
 392                         // Does path start with OutputPath?
 393                         prefix := runner.Container.OutputPath
 394                         if !strings.HasSuffix(prefix, "/") {
 395                                 prefix += "/"
 396                         }
 397                         if !strings.HasPrefix(mnt.Path, prefix) {
 398                                 return fmt.Errorf("%s path does not start with OutputPath: %s, %s", strings.Title(bind), mnt.Path, prefix)
 399                         }
 400                 }
 401
 402                 if bind == "stdin" {
 403                         // Is it a "collection" mount kind?
 404                         if mnt.Kind != "collection" && mnt.Kind != "json" {
 405                                 return fmt.Errorf("Unsupported mount kind '%s' for stdin. Only 'collection' or 'json' are supported.", mnt.Kind)
 406                         }
 407                 }
 408
 409                 if bind == "/etc/arvados/ca-certificates.crt" {
 410                         needCertMount = false
 411                 }
 412
 413                 if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && bind != runner.Container.OutputPath+"/" {
 414                         if mnt.Kind != "collection" {
 415                                 return fmt.Errorf("Only mount points of kind 'collection' are supported underneath the output_path: %v", bind)
 416                         }
 417                 }
 418
 419                 switch {
 420                 case mnt.Kind == "collection" && bind != "stdin":
 421                         var src string
 422                         if mnt.UUID != "" && mnt.PortableDataHash != "" {
 423                                 return fmt.Errorf("Cannot specify both 'uuid' and 'portable_data_hash' for a collection mount")
 424                         }
 425                         if mnt.UUID != "" {
 426                                 if mnt.Writable {
 427                                         return fmt.Errorf("Writing to existing collections currently not permitted.")
 428                                 }
 429                                 pdhOnly = false
 430                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.UUID)
 431                         } else if mnt.PortableDataHash != "" {
 432                                 if mnt.Writable {
 433                                         return fmt.Errorf("Can never write to a collection specified by portable data hash")
 434                                 }
 435                                 idx := strings.Index(mnt.PortableDataHash, "/")
 436                                 if idx > 0 {
 437                                         mnt.Path = path.Clean(mnt.PortableDataHash[idx:])
 438                                         mnt.PortableDataHash = mnt.PortableDataHash[0:idx]
 439                                         runner.Container.Mounts[bind] = mnt
 440                                 }
 441                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.PortableDataHash)
 442                                 if mnt.Path != "" && mnt.Path != "." {
 443                                         if strings.HasPrefix(mnt.Path, "./") {
 444                                                 mnt.Path = mnt.Path[2:]
 445                                         } else if strings.HasPrefix(mnt.Path, "/") {
 446                                                 mnt.Path = mnt.Path[1:]
 447                                         }
 448                                         src += "/" + mnt.Path
 449                                 }
 450                         } else {
 451                                 src = fmt.Sprintf("%s/tmp%d", runner.ArvMountPoint, tmpcount)
 452                                 arvMountCmd = append(arvMountCmd, "--mount-tmp")
 453                                 arvMountCmd = append(arvMountCmd, fmt.Sprintf("tmp%d", tmpcount))
 454                                 tmpcount += 1
 455                         }
 456                         if mnt.Writable {
 457                                 if bind == runner.Container.OutputPath {
 458                                         runner.HostOutputDir = src
 459                                 } else if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 460                                         return fmt.Errorf("Writable mount points are not permitted underneath the output_path: %v", bind)
 461                                 }
 462                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", src, bind))
 463                         } else {
 464                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", src, bind))
 465                         }
 466                         collectionPaths = append(collectionPaths, src)
 467
 468                 case mnt.Kind == "tmp":
 469                         var tmpdir string
 470                         tmpdir, err = runner.MkTempDir("", "")
 471                         if err != nil {
 472                                 return fmt.Errorf("While creating mount temp dir: %v", err)
 473                         }
 474                         st, staterr := os.Stat(tmpdir)
 475                         if staterr != nil {
 476                                 return fmt.Errorf("While Stat on temp dir: %v", staterr)
 477                         }
 478                         err = os.Chmod(tmpdir, st.Mode()|os.ModeSetgid|0777)
 479                         if staterr != nil {
 480                                 return fmt.Errorf("While Chmod temp dir: %v", err)
 481                         }
 482                         runner.CleanupTempDir = append(runner.CleanupTempDir, tmpdir)
 483                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", tmpdir, bind))
 484                         if bind == runner.Container.OutputPath {
 485                                 runner.HostOutputDir = tmpdir
 486                         }
 487
 488                 case mnt.Kind == "json":
 489                         jsondata, err := json.Marshal(mnt.Content)
 490                         if err != nil {
 491                                 return fmt.Errorf("encoding json data: %v", err)
 492                         }
 493                         // Create a tempdir with a single file
 494                         // (instead of just a tempfile): this way we
 495                         // can ensure the file is world-readable
 496                         // inside the container, without having to
 497                         // make it world-readable on the docker host.
 498                         tmpdir, err := runner.MkTempDir("", "")
 499                         if err != nil {
 500                                 return fmt.Errorf("creating temp dir: %v", err)
 501                         }
 502                         runner.CleanupTempDir = append(runner.CleanupTempDir, tmpdir)
 503                         tmpfn := filepath.Join(tmpdir, "mountdata.json")
 504                         err = ioutil.WriteFile(tmpfn, jsondata, 0644)
 505                         if err != nil {
 506                                 return fmt.Errorf("writing temp file: %v", err)
 507                         }
 508                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", tmpfn, bind))
 509                 }
 510         }
 511
 512         if runner.HostOutputDir == "" {
 513                 return fmt.Errorf("Output path does not correspond to a writable mount point")
 514         }
 515
 516         if wantAPI := runner.Container.RuntimeConstraints.API; needCertMount && wantAPI != nil && *wantAPI {
 517                 for _, certfile := range arvadosclient.CertFiles {
 518                         _, err := os.Stat(certfile)
 519                         if err == nil {
 520                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:/etc/arvados/ca-certificates.crt:ro", certfile))
 521                                 break
 522                         }
 523                 }
 524         }
 525
 526         if pdhOnly {
 527                 arvMountCmd = append(arvMountCmd, "--mount-by-pdh", "by_id")
 528         } else {
 529                 arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_id")
 530         }
 531         arvMountCmd = append(arvMountCmd, runner.ArvMountPoint)
 532
 533         token, err := runner.ContainerToken()
 534         if err != nil {
 535                 return fmt.Errorf("could not get container token: %s", err)
 536         }
 537
 538         runner.ArvMount, err = runner.RunArvMount(arvMountCmd, token)
 539         if err != nil {
 540                 return fmt.Errorf("While trying to start arv-mount: %v", err)
 541         }
 542
 543         for _, p := range collectionPaths {
 544                 _, err = os.Stat(p)
 545                 if err != nil {
 546                         return fmt.Errorf("While checking that input files exist: %v", err)
 547                 }
 548         }
 549
 550         return nil
 551 }
 552
 553 func (runner *ContainerRunner) ProcessDockerAttach(containerReader io.Reader) {
 554         // Handle docker log protocol
 555         // https://docs.docker.com/engine/reference/api/docker_remote_api_v1.15/#attach-to-a-container
 556
 557         header := make([]byte, 8)
 558         for {
 559                 _, readerr := io.ReadAtLeast(containerReader, header, 8)
 560
 561                 if readerr == nil {
 562                         readsize := int64(header[7]) | (int64(header[6]) << 8) | (int64(header[5]) << 16) | (int64(header[4]) << 24)
 563                         if header[0] == 1 {
 564                                 // stdout
 565                                 _, readerr = io.CopyN(runner.Stdout, containerReader, readsize)
 566                         } else {
 567                                 // stderr
 568                                 _, readerr = io.CopyN(runner.Stderr, containerReader, readsize)
 569                         }
 570                 }
 571
 572                 if readerr != nil {
 573                         if readerr != io.EOF {
 574                                 runner.CrunchLog.Printf("While reading docker logs: %v", readerr)
 575                         }
 576
 577                         closeerr := runner.Stdout.Close()
 578                         if closeerr != nil {
 579                                 runner.CrunchLog.Printf("While closing stdout logs: %v", closeerr)
 580                         }
 581
 582                         closeerr = runner.Stderr.Close()
 583                         if closeerr != nil {
 584                                 runner.CrunchLog.Printf("While closing stderr logs: %v", closeerr)
 585                         }
 586
 587                         if runner.statReporter != nil {
 588                                 runner.statReporter.Stop()
 589                                 closeerr = runner.statLogger.Close()
 590                                 if closeerr != nil {
 591                                         runner.CrunchLog.Printf("While closing crunchstat logs: %v", closeerr)
 592                                 }
 593                         }
 594
 595                         runner.loggingDone <- true
 596                         close(runner.loggingDone)
 597                         return
 598                 }
 599         }
 600 }
 601
 602 func (runner *ContainerRunner) StartCrunchstat() {
 603         runner.statLogger = NewThrottledLogger(runner.NewLogWriter("crunchstat"))
 604         runner.statReporter = &crunchstat.Reporter{
 605                 CID:          runner.ContainerID,
 606                 Logger:       log.New(runner.statLogger, "", 0),
 607                 CgroupParent: runner.expectCgroupParent,
 608                 CgroupRoot:   runner.cgroupRoot,
 609                 PollPeriod:   runner.statInterval,
 610         }
 611         runner.statReporter.Start()
 612 }
 613
 614 type infoCommand struct {
 615         label string
 616         cmd   []string
 617 }
 618
 619 // Gather node information and store it on the log for debugging
 620 // purposes.
 621 func (runner *ContainerRunner) LogNodeInfo() (err error) {
 622         w := runner.NewLogWriter("node-info")
 623         logger := log.New(w, "node-info", 0)
 624
 625         commands := []infoCommand{
 626                 {
 627                         label: "Host Information",
 628                         cmd:   []string{"uname", "-a"},
 629                 },
 630                 {
 631                         label: "CPU Information",
 632                         cmd:   []string{"cat", "/proc/cpuinfo"},
 633                 },
 634                 {
 635                         label: "Memory Information",
 636                         cmd:   []string{"cat", "/proc/meminfo"},
 637                 },
 638                 {
 639                         label: "Disk Space",
 640                         cmd:   []string{"df", "-m", "/", os.TempDir()},
 641                 },
 642                 {
 643                         label: "Disk INodes",
 644                         cmd:   []string{"df", "-i", "/", os.TempDir()},
 645                 },
 646         }
 647
 648         // Run commands with informational output to be logged.
 649         var out []byte
 650         for _, command := range commands {
 651                 out, err = exec.Command(command.cmd[0], command.cmd[1:]...).CombinedOutput()
 652                 if err != nil {
 653                         return fmt.Errorf("While running command %q: %v",
 654                                 command.cmd, err)
 655                 }
 656                 logger.Println(command.label)
 657                 for _, line := range strings.Split(string(out), "\n") {
 658                         logger.Println(" ", line)
 659                 }
 660         }
 661
 662         err = w.Close()
 663         if err != nil {
 664                 return fmt.Errorf("While closing node-info logs: %v", err)
 665         }
 666         return nil
 667 }
 668
 669 // Get and save the raw JSON container record from the API server
 670 func (runner *ContainerRunner) LogContainerRecord() (err error) {
 671         w := &ArvLogWriter{
 672                 ArvClient:     runner.ArvClient,
 673                 UUID:          runner.Container.UUID,
 674                 loggingStream: "container",
 675                 writeCloser:   runner.LogCollection.Open("container.json"),
 676         }
 677
 678         // Get Container record JSON from the API Server
 679         reader, err := runner.ArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
 680         if err != nil {
 681                 return fmt.Errorf("While retrieving container record from the API server: %v", err)
 682         }
 683         defer reader.Close()
 684
 685         dec := json.NewDecoder(reader)
 686         dec.UseNumber()
 687         var cr map[string]interface{}
 688         if err = dec.Decode(&cr); err != nil {
 689                 return fmt.Errorf("While decoding the container record JSON response: %v", err)
 690         }
 691         // Re-encode it using indentation to improve readability
 692         enc := json.NewEncoder(w)
 693         enc.SetIndent("", "    ")
 694         if err = enc.Encode(cr); err != nil {
 695                 return fmt.Errorf("While logging the JSON container record: %v", err)
 696         }
 697         err = w.Close()
 698         if err != nil {
 699                 return fmt.Errorf("While closing container.json log: %v", err)
 700         }
 701         return nil
 702 }
 703
 704 // AttachStreams connects the docker container stdin, stdout and stderr logs
 705 // to the Arvados logger which logs to Keep and the API server logs table.
 706 func (runner *ContainerRunner) AttachStreams() (err error) {
 707
 708         runner.CrunchLog.Print("Attaching container streams")
 709
 710         // If stdin mount is provided, attach it to the docker container
 711         var stdinRdr arvados.File
 712         var stdinJson []byte
 713         if stdinMnt, ok := runner.Container.Mounts["stdin"]; ok {
 714                 if stdinMnt.Kind == "collection" {
 715                         var stdinColl arvados.Collection
 716                         collId := stdinMnt.UUID
 717                         if collId == "" {
 718                                 collId = stdinMnt.PortableDataHash
 719                         }
 720                         err = runner.ArvClient.Get("collections", collId, nil, &stdinColl)
 721                         if err != nil {
 722                                 return fmt.Errorf("While getting stding collection: %v", err)
 723                         }
 724
 725                         stdinRdr, err = runner.Kc.ManifestFileReader(manifest.Manifest{Text: stdinColl.ManifestText}, stdinMnt.Path)
 726                         if os.IsNotExist(err) {
 727                                 return fmt.Errorf("stdin collection path not found: %v", stdinMnt.Path)
 728                         } else if err != nil {
 729                                 return fmt.Errorf("While getting stdin collection path %v: %v", stdinMnt.Path, err)
 730                         }
 731                 } else if stdinMnt.Kind == "json" {
 732                         stdinJson, err = json.Marshal(stdinMnt.Content)
 733                         if err != nil {
 734                                 return fmt.Errorf("While encoding stdin json data: %v", err)
 735                         }
 736                 }
 737         }
 738
 739         stdinUsed := stdinRdr != nil || len(stdinJson) != 0
 740         response, err := runner.Docker.ContainerAttach(context.TODO(), runner.ContainerID,
 741                 dockertypes.ContainerAttachOptions{Stream: true, Stdin: stdinUsed, Stdout: true, Stderr: true})
 742         if err != nil {
 743                 return fmt.Errorf("While attaching container stdout/stderr streams: %v", err)
 744         }
 745
 746         runner.loggingDone = make(chan bool)
 747
 748         if stdoutMnt, ok := runner.Container.Mounts["stdout"]; ok {
 749                 stdoutFile, err := runner.getStdoutFile(stdoutMnt.Path)
 750                 if err != nil {
 751                         return err
 752                 }
 753                 runner.Stdout = stdoutFile
 754         } else {
 755                 runner.Stdout = NewThrottledLogger(runner.NewLogWriter("stdout"))
 756         }
 757
 758         if stderrMnt, ok := runner.Container.Mounts["stderr"]; ok {
 759                 stderrFile, err := runner.getStdoutFile(stderrMnt.Path)
 760                 if err != nil {
 761                         return err
 762                 }
 763                 runner.Stderr = stderrFile
 764         } else {
 765                 runner.Stderr = NewThrottledLogger(runner.NewLogWriter("stderr"))
 766         }
 767
 768         if stdinRdr != nil {
 769                 go func() {
 770                         _, err := io.Copy(response.Conn, stdinRdr)
 771                         if err != nil {
 772                                 runner.CrunchLog.Print("While writing stdin collection to docker container %q", err)
 773                                 runner.stop()
 774                         }
 775                         stdinRdr.Close()
 776                         response.CloseWrite()
 777                 }()
 778         } else if len(stdinJson) != 0 {
 779                 go func() {
 780                         _, err := io.Copy(response.Conn, bytes.NewReader(stdinJson))
 781                         if err != nil {
 782                                 runner.CrunchLog.Print("While writing stdin json to docker container %q", err)
 783                                 runner.stop()
 784                         }
 785                         response.CloseWrite()
 786                 }()
 787         }
 788
 789         go runner.ProcessDockerAttach(response.Reader)
 790
 791         return nil
 792 }
 793
 794 func (runner *ContainerRunner) getStdoutFile(mntPath string) (*os.File, error) {
 795         stdoutPath := mntPath[len(runner.Container.OutputPath):]
 796         index := strings.LastIndex(stdoutPath, "/")
 797         if index > 0 {
 798                 subdirs := stdoutPath[:index]
 799                 if subdirs != "" {
 800                         st, err := os.Stat(runner.HostOutputDir)
 801                         if err != nil {
 802                                 return nil, fmt.Errorf("While Stat on temp dir: %v", err)
 803                         }
 804                         stdoutPath := filepath.Join(runner.HostOutputDir, subdirs)
 805                         err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777)
 806                         if err != nil {
 807                                 return nil, fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err)
 808                         }
 809                 }
 810         }
 811         stdoutFile, err := os.Create(filepath.Join(runner.HostOutputDir, stdoutPath))
 812         if err != nil {
 813                 return nil, fmt.Errorf("While creating file %q: %v", stdoutPath, err)
 814         }
 815
 816         return stdoutFile, nil
 817 }
 818
 819 // CreateContainer creates the docker container.
 820 func (runner *ContainerRunner) CreateContainer() error {
 821         runner.CrunchLog.Print("Creating Docker container")
 822
 823         runner.ContainerConfig.Cmd = runner.Container.Command
 824         if runner.Container.Cwd != "." {
 825                 runner.ContainerConfig.WorkingDir = runner.Container.Cwd
 826         }
 827
 828         for k, v := range runner.Container.Environment {
 829                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env, k+"="+v)
 830         }
 831
 832         runner.ContainerConfig.Volumes = runner.Volumes
 833
 834         runner.HostConfig = dockercontainer.HostConfig{
 835                 Binds: runner.Binds,
 836                 LogConfig: dockercontainer.LogConfig{
 837                         Type: "none",
 838                 },
 839                 Resources: dockercontainer.Resources{
 840                         CgroupParent: runner.setCgroupParent,
 841                 },
 842         }
 843
 844         if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
 845                 tok, err := runner.ContainerToken()
 846                 if err != nil {
 847                         return err
 848                 }
 849                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env,
 850                         "ARVADOS_API_TOKEN="+tok,
 851                         "ARVADOS_API_HOST="+os.Getenv("ARVADOS_API_HOST"),
 852                         "ARVADOS_API_HOST_INSECURE="+os.Getenv("ARVADOS_API_HOST_INSECURE"),
 853                 )
 854                 runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
 855         } else {
 856                 if runner.enableNetwork == "always" {
 857                         runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
 858                 } else {
 859                         runner.HostConfig.NetworkMode = dockercontainer.NetworkMode("none")
 860                 }
 861         }
 862
 863         _, stdinUsed := runner.Container.Mounts["stdin"]
 864         runner.ContainerConfig.OpenStdin = stdinUsed
 865         runner.ContainerConfig.StdinOnce = stdinUsed
 866         runner.ContainerConfig.AttachStdin = stdinUsed
 867         runner.ContainerConfig.AttachStdout = true
 868         runner.ContainerConfig.AttachStderr = true
 869
 870         createdBody, err := runner.Docker.ContainerCreate(context.TODO(), &runner.ContainerConfig, &runner.HostConfig, nil, runner.Container.UUID)
 871         if err != nil {
 872                 return fmt.Errorf("While creating container: %v", err)
 873         }
 874
 875         runner.ContainerID = createdBody.ID
 876
 877         return runner.AttachStreams()
 878 }
 879
 880 // StartContainer starts the docker container created by CreateContainer.
 881 func (runner *ContainerRunner) StartContainer() error {
 882         runner.CrunchLog.Printf("Starting Docker container id '%s'", runner.ContainerID)
 883         runner.cStateLock.Lock()
 884         defer runner.cStateLock.Unlock()
 885         if runner.cCancelled {
 886                 return ErrCancelled
 887         }
 888         err := runner.Docker.ContainerStart(context.TODO(), runner.ContainerID,
 889                 dockertypes.ContainerStartOptions{})
 890         if err != nil {
 891                 var advice string
 892                 if strings.Contains(err.Error(), "no such file or directory") {
 893                         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])
 894                 }
 895                 return fmt.Errorf("could not start container: %v%s", err, advice)
 896         }
 897         runner.cStarted = true
 898         return nil
 899 }
 900
 901 // WaitFinish waits for the container to terminate, capture the exit code, and
 902 // close the stdout/stderr logging.
 903 func (runner *ContainerRunner) WaitFinish() (err error) {
 904         runner.CrunchLog.Print("Waiting for container to finish")
 905
 906         waitOk, waitErr := runner.Docker.ContainerWait(context.TODO(), runner.ContainerID, "not-running")
 907
 908         go func() {
 909                 <-runner.ArvMountExit
 910                 if runner.cStarted {
 911                         runner.CrunchLog.Printf("arv-mount exited while container is still running.  Stopping container.")
 912                         runner.stop()
 913                 }
 914         }()
 915
 916         var waitBody dockercontainer.ContainerWaitOKBody
 917         select {
 918         case waitBody = <-waitOk:
 919         case err = <-waitErr:
 920         }
 921
 922         // Container isn't running any more
 923         runner.cStarted = false
 924
 925         if err != nil {
 926                 return fmt.Errorf("container wait: %v", err)
 927         }
 928
 929         runner.CrunchLog.Printf("Container exited with code: %v", waitBody.StatusCode)
 930         code := int(waitBody.StatusCode)
 931         runner.ExitCode = &code
 932
 933         // wait for stdout/stderr to complete
 934         <-runner.loggingDone
 935
 936         return nil
 937 }
 938
 939 var ErrNotInOutputDir = fmt.Errorf("Must point to path within the output directory")
 940
 941 func (runner *ContainerRunner) derefOutputSymlink(path string, startinfo os.FileInfo) (tgt string, readlinktgt string, info os.FileInfo, err error) {
 942         // Follow symlinks if necessary
 943         info = startinfo
 944         tgt = path
 945         readlinktgt = ""
 946         nextlink := path
 947         for followed := 0; info.Mode()&os.ModeSymlink != 0; followed++ {
 948                 if followed >= limitFollowSymlinks {
 949                         // Got stuck in a loop or just a pathological number of links, give up.
 950                         err = fmt.Errorf("Followed more than %v symlinks from path %q", limitFollowSymlinks, path)
 951                         return
 952                 }
 953
 954                 readlinktgt, err = os.Readlink(nextlink)
 955                 if err != nil {
 956                         return
 957                 }
 958
 959                 tgt = readlinktgt
 960                 if !strings.HasPrefix(tgt, "/") {
 961                         // Relative symlink, resolve it to host path
 962                         tgt = filepath.Join(filepath.Dir(path), tgt)
 963                 }
 964                 if strings.HasPrefix(tgt, runner.Container.OutputPath+"/") && !strings.HasPrefix(tgt, runner.HostOutputDir+"/") {
 965                         // Absolute symlink to container output path, adjust it to host output path.
 966                         tgt = filepath.Join(runner.HostOutputDir, tgt[len(runner.Container.OutputPath):])
 967                 }
 968                 if !strings.HasPrefix(tgt, runner.HostOutputDir+"/") {
 969                         // After dereferencing, symlink target must either be
 970                         // within output directory, or must point to a
 971                         // collection mount.
 972                         err = ErrNotInOutputDir
 973                         return
 974                 }
 975
 976                 info, err = os.Lstat(tgt)
 977                 if err != nil {
 978                         // tgt
 979                         err = fmt.Errorf("Symlink in output %q points to invalid location %q: %v",
 980                                 path[len(runner.HostOutputDir):], readlinktgt, err)
 981                         return
 982                 }
 983
 984                 nextlink = tgt
 985         }
 986
 987         return
 988 }
 989
 990 var limitFollowSymlinks = 10
 991
 992 // UploadFile uploads files within the output directory, with special handling
 993 // for symlinks. If the symlink leads to a keep mount, copy the manifest text
 994 // from the keep mount into the output manifestText.  Ensure that whether
 995 // symlinks are relative or absolute, every symlink target (even targets that
 996 // are symlinks themselves) must point to a path in either the output directory
 997 // or a collection mount.
 998 //
 999 // Assumes initial value of "path" is absolute, and located within runner.HostOutputDir.
1000 func (runner *ContainerRunner) UploadOutputFile(
1001         path string,
1002         info os.FileInfo,
1003         infoerr error,
1004         binds []string,
1005         walkUpload *WalkUpload,
1006         relocateFrom string,
1007         relocateTo string,
1008         followed int) (manifestText string, err error) {
1009
1010         if info.Mode().IsDir() {
1011                 return
1012         }
1013
1014         if infoerr != nil {
1015                 return "", infoerr
1016         }
1017
1018         if followed >= limitFollowSymlinks {
1019                 // Got stuck in a loop or just a pathological number of
1020                 // directory links, give up.
1021                 err = fmt.Errorf("Followed more than %v symlinks from path %q", limitFollowSymlinks, path)
1022                 return
1023         }
1024
1025         // When following symlinks, the source path may need to be logically
1026         // relocated to some other path within the output collection.  Remove
1027         // the relocateFrom prefix and replace it with relocateTo.
1028         relocated := relocateTo + path[len(relocateFrom):]
1029
1030         tgt, readlinktgt, info, derefErr := runner.derefOutputSymlink(path, info)
1031         if derefErr != nil && derefErr != ErrNotInOutputDir {
1032                 return "", derefErr
1033         }
1034
1035         // go through mounts and try reverse map to collection reference
1036         for _, bind := range binds {
1037                 mnt := runner.Container.Mounts[bind]
1038                 if tgt == bind || strings.HasPrefix(tgt, bind+"/") {
1039                         // get path relative to bind
1040                         targetSuffix := tgt[len(bind):]
1041
1042                         // Copy mount and adjust the path to add path relative to the bind
1043                         adjustedMount := mnt
1044                         adjustedMount.Path = filepath.Join(adjustedMount.Path, targetSuffix)
1045
1046                         // Terminates in this keep mount, so add the
1047                         // manifest text at appropriate location.
1048                         outputSuffix := path[len(runner.HostOutputDir):]
1049                         manifestText, err = runner.getCollectionManifestForPath(adjustedMount, outputSuffix)
1050                         return
1051                 }
1052         }
1053
1054         // If target is not a collection mount, it must be located within the
1055         // output directory, otherwise it is an error.
1056         if derefErr == ErrNotInOutputDir {
1057                 err = fmt.Errorf("Symlink in output %q points to invalid location %q, must point to path within the output directory.",
1058                         path[len(runner.HostOutputDir):], readlinktgt)
1059                 return
1060         }
1061
1062         if info.Mode().IsRegular() {
1063                 return "", walkUpload.UploadFile(relocated, tgt)
1064         }
1065
1066         if info.Mode().IsDir() {
1067                 // Symlink leads to directory.  Walk() doesn't follow
1068                 // directory symlinks, so we walk the target directory
1069                 // instead.  Within the walk, file paths are relocated
1070                 // so they appear under the original symlink path.
1071                 err = filepath.Walk(tgt, func(walkpath string, walkinfo os.FileInfo, walkerr error) error {
1072                         var m string
1073                         m, walkerr = runner.UploadOutputFile(walkpath, walkinfo, walkerr,
1074                                 binds, walkUpload, tgt, relocated, followed+1)
1075                         if walkerr == nil {
1076                                 manifestText = manifestText + m
1077                         }
1078                         return walkerr
1079                 })
1080                 return
1081         }
1082
1083         return
1084 }
1085
1086 // HandleOutput sets the output, unmounts the FUSE mount, and deletes temporary directories
1087 func (runner *ContainerRunner) CaptureOutput() error {
1088         if runner.finalState != "Complete" {
1089                 return nil
1090         }
1091
1092         if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
1093                 // Output may have been set directly by the container, so
1094                 // refresh the container record to check.
1095                 err := runner.ArvClient.Get("containers", runner.Container.UUID,
1096                         nil, &runner.Container)
1097                 if err != nil {
1098                         return err
1099                 }
1100                 if runner.Container.Output != "" {
1101                         // Container output is already set.
1102                         runner.OutputPDH = &runner.Container.Output
1103                         return nil
1104                 }
1105         }
1106
1107         if runner.HostOutputDir == "" {
1108                 return nil
1109         }
1110
1111         _, err := os.Stat(runner.HostOutputDir)
1112         if err != nil {
1113                 return fmt.Errorf("While checking host output path: %v", err)
1114         }
1115
1116         // Pre-populate output from the configured mount points
1117         var binds []string
1118         for bind, mnt := range runner.Container.Mounts {
1119                 if mnt.Kind == "collection" {
1120                         binds = append(binds, bind)
1121                 }
1122         }
1123         sort.Strings(binds)
1124
1125         var manifestText string
1126
1127         collectionMetafile := fmt.Sprintf("%s/.arvados#collection", runner.HostOutputDir)
1128         _, err = os.Stat(collectionMetafile)
1129         if err != nil {
1130                 // Regular directory
1131
1132                 cw := CollectionWriter{0, runner.Kc, nil, nil, sync.Mutex{}}
1133                 walkUpload := cw.BeginUpload(runner.HostOutputDir, runner.CrunchLog.Logger)
1134
1135                 var m string
1136                 err = filepath.Walk(runner.HostOutputDir, func(path string, info os.FileInfo, err error) error {
1137                         m, err = runner.UploadOutputFile(path, info, err, binds, walkUpload, "", "", 0)
1138                         if err == nil {
1139                                 manifestText = manifestText + m
1140                         }
1141                         return err
1142                 })
1143
1144                 cw.EndUpload(walkUpload)
1145
1146                 if err != nil {
1147                         return fmt.Errorf("While uploading output files: %v", err)
1148                 }
1149
1150                 m, err = cw.ManifestText()
1151                 manifestText = manifestText + m
1152                 if err != nil {
1153                         return fmt.Errorf("While uploading output files: %v", err)
1154                 }
1155         } else {
1156                 // FUSE mount directory
1157                 file, openerr := os.Open(collectionMetafile)
1158                 if openerr != nil {
1159                         return fmt.Errorf("While opening FUSE metafile: %v", err)
1160                 }
1161                 defer file.Close()
1162
1163                 var rec arvados.Collection
1164                 err = json.NewDecoder(file).Decode(&rec)
1165                 if err != nil {
1166                         return fmt.Errorf("While reading FUSE metafile: %v", err)
1167                 }
1168                 manifestText = rec.ManifestText
1169         }
1170
1171         for _, bind := range binds {
1172                 mnt := runner.Container.Mounts[bind]
1173
1174                 bindSuffix := strings.TrimPrefix(bind, runner.Container.OutputPath)
1175
1176                 if bindSuffix == bind || len(bindSuffix) <= 0 {
1177                         // either does not start with OutputPath or is OutputPath itself
1178                         continue
1179                 }
1180
1181                 if mnt.ExcludeFromOutput == true {
1182                         continue
1183                 }
1184
1185                 // append to manifest_text
1186                 m, err := runner.getCollectionManifestForPath(mnt, bindSuffix)
1187                 if err != nil {
1188                         return err
1189                 }
1190
1191                 manifestText = manifestText + m
1192         }
1193
1194         // Save output
1195         var response arvados.Collection
1196         manifest := manifest.Manifest{Text: manifestText}
1197         manifestText = manifest.Extract(".", ".").Text
1198         err = runner.ArvClient.Create("collections",
1199                 arvadosclient.Dict{
1200                         "ensure_unique_name": true,
1201                         "collection": arvadosclient.Dict{
1202                                 "is_trashed":    true,
1203                                 "name":          "output for " + runner.Container.UUID,
1204                                 "manifest_text": manifestText}},
1205                 &response)
1206         if err != nil {
1207                 return fmt.Errorf("While creating output collection: %v", err)
1208         }
1209         runner.OutputPDH = &response.PortableDataHash
1210         return nil
1211 }
1212
1213 var outputCollections = make(map[string]arvados.Collection)
1214
1215 // Fetch the collection for the mnt.PortableDataHash
1216 // Return the manifest_text fragment corresponding to the specified mnt.Path
1217 //  after making any required updates.
1218 //  Ex:
1219 //    If mnt.Path is not specified,
1220 //      return the entire manifest_text after replacing any "." with bindSuffix
1221 //    If mnt.Path corresponds to one stream,
1222 //      return the manifest_text for that stream after replacing that stream name with bindSuffix
1223 //    Otherwise, check if a filename in any one stream is being sought. Return the manifest_text
1224 //      for that stream after replacing stream name with bindSuffix minus the last word
1225 //      and the file name with last word of the bindSuffix
1226 //  Allowed path examples:
1227 //    "path":"/"
1228 //    "path":"/subdir1"
1229 //    "path":"/subdir1/subdir2"
1230 //    "path":"/subdir/filename" etc
1231 func (runner *ContainerRunner) getCollectionManifestForPath(mnt arvados.Mount, bindSuffix string) (string, error) {
1232         collection := outputCollections[mnt.PortableDataHash]
1233         if collection.PortableDataHash == "" {
1234                 err := runner.ArvClient.Get("collections", mnt.PortableDataHash, nil, &collection)
1235                 if err != nil {
1236                         return "", fmt.Errorf("While getting collection for %v: %v", mnt.PortableDataHash, err)
1237                 }
1238                 outputCollections[mnt.PortableDataHash] = collection
1239         }
1240
1241         if collection.ManifestText == "" {
1242                 runner.CrunchLog.Printf("No manifest text for collection %v", collection.PortableDataHash)
1243                 return "", nil
1244         }
1245
1246         mft := manifest.Manifest{Text: collection.ManifestText}
1247         extracted := mft.Extract(mnt.Path, bindSuffix)
1248         if extracted.Err != nil {
1249                 return "", fmt.Errorf("Error parsing manifest for %v: %v", mnt.PortableDataHash, extracted.Err.Error())
1250         }
1251         return extracted.Text, nil
1252 }
1253
1254 func (runner *ContainerRunner) CleanupDirs() {
1255         if runner.ArvMount != nil {
1256                 var delay int64 = 8
1257                 umount := exec.Command("arv-mount", fmt.Sprintf("--unmount-timeout=%d", delay), "--unmount", runner.ArvMountPoint)
1258                 umount.Stdout = runner.CrunchLog
1259                 umount.Stderr = runner.CrunchLog
1260                 runner.CrunchLog.Printf("Running %v", umount.Args)
1261                 umnterr := umount.Start()
1262
1263                 if umnterr != nil {
1264                         runner.CrunchLog.Printf("Error unmounting: %v", umnterr)
1265                 } else {
1266                         // If arv-mount --unmount gets stuck for any reason, we
1267                         // don't want to wait for it forever.  Do Wait() in a goroutine
1268                         // so it doesn't block crunch-run.
1269                         umountExit := make(chan error)
1270                         go func() {
1271                                 mnterr := umount.Wait()
1272                                 if mnterr != nil {
1273                                         runner.CrunchLog.Printf("Error unmounting: %v", mnterr)
1274                                 }
1275                                 umountExit <- mnterr
1276                         }()
1277
1278                         for again := true; again; {
1279                                 again = false
1280                                 select {
1281                                 case <-umountExit:
1282                                         umount = nil
1283                                         again = true
1284                                 case <-runner.ArvMountExit:
1285                                         break
1286                                 case <-time.After(time.Duration((delay + 1) * int64(time.Second))):
1287                                         runner.CrunchLog.Printf("Timed out waiting for unmount")
1288                                         if umount != nil {
1289                                                 umount.Process.Kill()
1290                                         }
1291                                         runner.ArvMount.Process.Kill()
1292                                 }
1293                         }
1294                 }
1295         }
1296
1297         if runner.ArvMountPoint != "" {
1298                 if rmerr := os.Remove(runner.ArvMountPoint); rmerr != nil {
1299                         runner.CrunchLog.Printf("While cleaning up arv-mount directory %s: %v", runner.ArvMountPoint, rmerr)
1300                 }
1301         }
1302
1303         for _, tmpdir := range runner.CleanupTempDir {
1304                 if rmerr := os.RemoveAll(tmpdir); rmerr != nil {
1305                         runner.CrunchLog.Printf("While cleaning up temporary directory %s: %v", tmpdir, rmerr)
1306                 }
1307         }
1308 }
1309
1310 // CommitLogs posts the collection containing the final container logs.
1311 func (runner *ContainerRunner) CommitLogs() error {
1312         runner.CrunchLog.Print(runner.finalState)
1313
1314         if runner.arvMountLog != nil {
1315                 runner.arvMountLog.Close()
1316         }
1317         runner.CrunchLog.Close()
1318
1319         // Closing CrunchLog above allows them to be committed to Keep at this
1320         // point, but re-open crunch log with ArvClient in case there are any
1321         // other further errors (such as failing to write the log to Keep!)
1322         // while shutting down
1323         runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{ArvClient: runner.ArvClient,
1324                 UUID: runner.Container.UUID, loggingStream: "crunch-run", writeCloser: nil})
1325         runner.CrunchLog.Immediate = log.New(os.Stderr, runner.Container.UUID+" ", 0)
1326
1327         if runner.LogsPDH != nil {
1328                 // If we have already assigned something to LogsPDH,
1329                 // we must be closing the re-opened log, which won't
1330                 // end up getting attached to the container record and
1331                 // therefore doesn't need to be saved as a collection
1332                 // -- it exists only to send logs to other channels.
1333                 return nil
1334         }
1335
1336         mt, err := runner.LogCollection.ManifestText()
1337         if err != nil {
1338                 return fmt.Errorf("While creating log manifest: %v", err)
1339         }
1340
1341         var response arvados.Collection
1342         err = runner.ArvClient.Create("collections",
1343                 arvadosclient.Dict{
1344                         "ensure_unique_name": true,
1345                         "collection": arvadosclient.Dict{
1346                                 "is_trashed":    true,
1347                                 "name":          "logs for " + runner.Container.UUID,
1348                                 "manifest_text": mt}},
1349                 &response)
1350         if err != nil {
1351                 return fmt.Errorf("While creating log collection: %v", err)
1352         }
1353         runner.LogsPDH = &response.PortableDataHash
1354         return nil
1355 }
1356
1357 // UpdateContainerRunning updates the container state to "Running"
1358 func (runner *ContainerRunner) UpdateContainerRunning() error {
1359         runner.cStateLock.Lock()
1360         defer runner.cStateLock.Unlock()
1361         if runner.cCancelled {
1362                 return ErrCancelled
1363         }
1364         return runner.ArvClient.Update("containers", runner.Container.UUID,
1365                 arvadosclient.Dict{"container": arvadosclient.Dict{"state": "Running"}}, nil)
1366 }
1367
1368 // ContainerToken returns the api_token the container (and any
1369 // arv-mount processes) are allowed to use.
1370 func (runner *ContainerRunner) ContainerToken() (string, error) {
1371         if runner.token != "" {
1372                 return runner.token, nil
1373         }
1374
1375         var auth arvados.APIClientAuthorization
1376         err := runner.ArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth)
1377         if err != nil {
1378                 return "", err
1379         }
1380         runner.token = auth.APIToken
1381         return runner.token, nil
1382 }
1383
1384 // UpdateContainerComplete updates the container record state on API
1385 // server to "Complete" or "Cancelled"
1386 func (runner *ContainerRunner) UpdateContainerFinal() error {
1387         update := arvadosclient.Dict{}
1388         update["state"] = runner.finalState
1389         if runner.LogsPDH != nil {
1390                 update["log"] = *runner.LogsPDH
1391         }
1392         if runner.finalState == "Complete" {
1393                 if runner.ExitCode != nil {
1394                         update["exit_code"] = *runner.ExitCode
1395                 }
1396                 if runner.OutputPDH != nil {
1397                         update["output"] = *runner.OutputPDH
1398                 }
1399         }
1400         return runner.ArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{"container": update}, nil)
1401 }
1402
1403 // IsCancelled returns the value of Cancelled, with goroutine safety.
1404 func (runner *ContainerRunner) IsCancelled() bool {
1405         runner.cStateLock.Lock()
1406         defer runner.cStateLock.Unlock()
1407         return runner.cCancelled
1408 }
1409
1410 // NewArvLogWriter creates an ArvLogWriter
1411 func (runner *ContainerRunner) NewArvLogWriter(name string) io.WriteCloser {
1412         return &ArvLogWriter{
1413                 ArvClient:     runner.ArvClient,
1414                 UUID:          runner.Container.UUID,
1415                 loggingStream: name,
1416                 writeCloser:   runner.LogCollection.Open(name + ".txt")}
1417 }
1418
1419 // Run the full container lifecycle.
1420 func (runner *ContainerRunner) Run() (err error) {
1421         runner.CrunchLog.Printf("Executing container '%s'", runner.Container.UUID)
1422
1423         hostname, hosterr := os.Hostname()
1424         if hosterr != nil {
1425                 runner.CrunchLog.Printf("Error getting hostname '%v'", hosterr)
1426         } else {
1427                 runner.CrunchLog.Printf("Executing on host '%s'", hostname)
1428         }
1429
1430         runner.finalState = "Queued"
1431
1432         defer func() {
1433                 runner.stopSignals()
1434                 runner.CleanupDirs()
1435
1436                 runner.CrunchLog.Printf("crunch-run finished")
1437                 runner.CrunchLog.Close()
1438         }()
1439
1440         defer func() {
1441                 // checkErr prints e (unless it's nil) and sets err to
1442                 // e (unless err is already non-nil). Thus, if err
1443                 // hasn't already been assigned when Run() returns,
1444                 // this cleanup func will cause Run() to return the
1445                 // first non-nil error that is passed to checkErr().
1446                 checkErr := func(e error) {
1447                         if e == nil {
1448                                 return
1449                         }
1450                         runner.CrunchLog.Print(e)
1451                         if err == nil {
1452                                 err = e
1453                         }
1454                         if runner.finalState == "Complete" {
1455                                 // There was an error in the finalization.
1456                                 runner.finalState = "Cancelled"
1457                         }
1458                 }
1459
1460                 // Log the error encountered in Run(), if any
1461                 checkErr(err)
1462
1463                 if runner.finalState == "Queued" {
1464                         runner.UpdateContainerFinal()
1465                         return
1466                 }
1467
1468                 if runner.IsCancelled() {
1469                         runner.finalState = "Cancelled"
1470                         // but don't return yet -- we still want to
1471                         // capture partial output and write logs
1472                 }
1473
1474                 checkErr(runner.CaptureOutput())
1475                 checkErr(runner.CommitLogs())
1476                 checkErr(runner.UpdateContainerFinal())
1477         }()
1478
1479         err = runner.fetchContainerRecord()
1480         if err != nil {
1481                 return
1482         }
1483
1484         // setup signal handling
1485         runner.setupSignals()
1486
1487         // check for and/or load image
1488         err = runner.LoadImage()
1489         if err != nil {
1490                 runner.finalState = "Cancelled"
1491                 err = fmt.Errorf("While loading container image: %v", err)
1492                 return
1493         }
1494
1495         // set up FUSE mount and binds
1496         err = runner.SetupMounts()
1497         if err != nil {
1498                 runner.finalState = "Cancelled"
1499                 err = fmt.Errorf("While setting up mounts: %v", err)
1500                 return
1501         }
1502
1503         err = runner.CreateContainer()
1504         if err != nil {
1505                 return
1506         }
1507
1508         // Gather and record node information
1509         err = runner.LogNodeInfo()
1510         if err != nil {
1511                 return
1512         }
1513         // Save container.json record on log collection
1514         err = runner.LogContainerRecord()
1515         if err != nil {
1516                 return
1517         }
1518
1519         runner.StartCrunchstat()
1520
1521         if runner.IsCancelled() {
1522                 return
1523         }
1524
1525         err = runner.UpdateContainerRunning()
1526         if err != nil {
1527                 return
1528         }
1529         runner.finalState = "Cancelled"
1530
1531         err = runner.StartContainer()
1532         if err != nil {
1533                 return
1534         }
1535
1536         err = runner.WaitFinish()
1537         if err == nil {
1538                 runner.finalState = "Complete"
1539         }
1540         return
1541 }
1542
1543 // Fetch the current container record (uuid = runner.Container.UUID)
1544 // into runner.Container.
1545 func (runner *ContainerRunner) fetchContainerRecord() error {
1546         reader, err := runner.ArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
1547         if err != nil {
1548                 return fmt.Errorf("error fetching container record: %v", err)
1549         }
1550         defer reader.Close()
1551
1552         dec := json.NewDecoder(reader)
1553         dec.UseNumber()
1554         err = dec.Decode(&runner.Container)
1555         if err != nil {
1556                 return fmt.Errorf("error decoding container record: %v", err)
1557         }
1558         return nil
1559 }
1560
1561 // NewContainerRunner creates a new container runner.
1562 func NewContainerRunner(api IArvadosClient,
1563         kc IKeepClient,
1564         docker ThinDockerClient,
1565         containerUUID string) *ContainerRunner {
1566
1567         cr := &ContainerRunner{ArvClient: api, Kc: kc, Docker: docker}
1568         cr.NewLogWriter = cr.NewArvLogWriter
1569         cr.RunArvMount = cr.ArvMountCmd
1570         cr.MkTempDir = ioutil.TempDir
1571         cr.LogCollection = &CollectionWriter{0, kc, nil, nil, sync.Mutex{}}
1572         cr.Container.UUID = containerUUID
1573         cr.CrunchLog = NewThrottledLogger(cr.NewLogWriter("crunch-run"))
1574         cr.CrunchLog.Immediate = log.New(os.Stderr, containerUUID+" ", 0)
1575
1576         loadLogThrottleParams(api)
1577
1578         return cr
1579 }
1580
1581 func main() {
1582         statInterval := flag.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting")
1583         cgroupRoot := flag.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree")
1584         cgroupParent := flag.String("cgroup-parent", "docker", "name of container's parent cgroup (ignored if -cgroup-parent-subsystem is used)")
1585         cgroupParentSubsystem := flag.String("cgroup-parent-subsystem", "", "use current cgroup for given subsystem as parent cgroup for container")
1586         caCertsPath := flag.String("ca-certs", "", "Path to TLS root certificates")
1587         enableNetwork := flag.String("container-enable-networking", "default",
1588                 `Specify if networking should be enabled for container.  One of 'default', 'always':
1589         default: only enable networking if container requests it.
1590         always:  containers always have networking enabled
1591         `)
1592         networkMode := flag.String("container-network-mode", "default",
1593                 `Set networking mode for container.  Corresponds to Docker network mode (--net).
1594         `)
1595         memprofile := flag.String("memprofile", "", "write memory profile to `file` after running container")
1596         flag.Parse()
1597
1598         containerId := flag.Arg(0)
1599
1600         if *caCertsPath != "" {
1601                 arvadosclient.CertFiles = []string{*caCertsPath}
1602         }
1603
1604         api, err := arvadosclient.MakeArvadosClient()
1605         if err != nil {
1606                 log.Fatalf("%s: %v", containerId, err)
1607         }
1608         api.Retries = 8
1609
1610         var kc *keepclient.KeepClient
1611         kc, err = keepclient.MakeKeepClient(api)
1612         if err != nil {
1613                 log.Fatalf("%s: %v", containerId, err)
1614         }
1615         kc.BlockCache = &keepclient.BlockCache{MaxBlocks: 2}
1616         kc.Retries = 4
1617
1618         var docker *dockerclient.Client
1619         // API version 1.21 corresponds to Docker 1.9, which is currently the
1620         // minimum version we want to support.
1621         docker, err = dockerclient.NewClient(dockerclient.DefaultDockerHost, "1.21", nil, nil)
1622         if err != nil {
1623                 log.Fatalf("%s: %v", containerId, err)
1624         }
1625
1626         dockerClientProxy := ThinDockerClientProxy{Docker: docker}
1627
1628         cr := NewContainerRunner(api, kc, dockerClientProxy, containerId)
1629         cr.statInterval = *statInterval
1630         cr.cgroupRoot = *cgroupRoot
1631         cr.expectCgroupParent = *cgroupParent
1632         cr.enableNetwork = *enableNetwork
1633         cr.networkMode = *networkMode
1634         if *cgroupParentSubsystem != "" {
1635                 p := findCgroup(*cgroupParentSubsystem)
1636                 cr.setCgroupParent = p
1637                 cr.expectCgroupParent = p
1638         }
1639
1640         runerr := cr.Run()
1641
1642         if *memprofile != "" {
1643                 f, err := os.Create(*memprofile)
1644                 if err != nil {
1645                         log.Printf("could not create memory profile: ", err)
1646                 }
1647                 runtime.GC() // get up-to-date statistics
1648                 if err := pprof.WriteHeapProfile(f); err != nil {
1649                         log.Printf("could not write memory profile: ", err)
1650                 }
1651                 closeerr := f.Close()
1652                 if closeerr != nil {
1653                         log.Printf("closing memprofile file: ", err)
1654                 }
1655         }
1656
1657         if runerr != nil {
1658                 log.Fatalf("%s: %v", containerId, runerr)
1659         }
1660 }