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 var version = "dev"
  43
  44 // IArvadosClient is the minimal Arvados API methods used by crunch-run.
  45 type IArvadosClient interface {
  46         Create(resourceType string, parameters arvadosclient.Dict, output interface{}) error
  47         Get(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  48         Update(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  49         Call(method, resourceType, uuid, action string, parameters arvadosclient.Dict, output interface{}) error
  50         CallRaw(method string, resourceType string, uuid string, action string, parameters arvadosclient.Dict) (reader io.ReadCloser, err error)
  51         Discovery(key string) (interface{}, error)
  52 }
  53
  54 // ErrCancelled is the error returned when the container is cancelled.
  55 var ErrCancelled = errors.New("Cancelled")
  56
  57 // IKeepClient is the minimal Keep API methods used by crunch-run.
  58 type IKeepClient interface {
  59         PutHB(hash string, buf []byte) (string, int, error)
  60         ManifestFileReader(m manifest.Manifest, filename string) (arvados.File, error)
  61         ClearBlockCache()
  62 }
  63
  64 // NewLogWriter is a factory function to create a new log writer.
  65 type NewLogWriter func(name string) io.WriteCloser
  66
  67 type RunArvMount func(args []string, tok string) (*exec.Cmd, error)
  68
  69 type MkTempDir func(string, string) (string, error)
  70
  71 // ThinDockerClient is the minimal Docker client interface used by crunch-run.
  72 type ThinDockerClient interface {
  73         ContainerAttach(ctx context.Context, container string, options dockertypes.ContainerAttachOptions) (dockertypes.HijackedResponse, error)
  74         ContainerCreate(ctx context.Context, config *dockercontainer.Config, hostConfig *dockercontainer.HostConfig,
  75                 networkingConfig *dockernetwork.NetworkingConfig, containerName string) (dockercontainer.ContainerCreateCreatedBody, error)
  76         ContainerStart(ctx context.Context, container string, options dockertypes.ContainerStartOptions) error
  77         ContainerStop(ctx context.Context, container string, timeout *time.Duration) error
  78         ContainerWait(ctx context.Context, container string, condition dockercontainer.WaitCondition) (<-chan dockercontainer.ContainerWaitOKBody, <-chan error)
  79         ImageInspectWithRaw(ctx context.Context, image string) (dockertypes.ImageInspect, []byte, error)
  80         ImageLoad(ctx context.Context, input io.Reader, quiet bool) (dockertypes.ImageLoadResponse, error)
  81         ImageRemove(ctx context.Context, image string, options dockertypes.ImageRemoveOptions) ([]dockertypes.ImageDeleteResponseItem, error)
  82 }
  83
  84 // ThinDockerClientProxy is a proxy implementation of ThinDockerClient
  85 // that executes the docker requests on dockerclient.Client
  86 type ThinDockerClientProxy struct {
  87         Docker *dockerclient.Client
  88 }
  89
  90 // ContainerAttach invokes dockerclient.Client.ContainerAttach
  91 func (proxy ThinDockerClientProxy) ContainerAttach(ctx context.Context, container string, options dockertypes.ContainerAttachOptions) (dockertypes.HijackedResponse, error) {
  92         return proxy.Docker.ContainerAttach(ctx, container, options)
  93 }
  94
  95 // ContainerCreate invokes dockerclient.Client.ContainerCreate
  96 func (proxy ThinDockerClientProxy) ContainerCreate(ctx context.Context, config *dockercontainer.Config, hostConfig *dockercontainer.HostConfig,
  97         networkingConfig *dockernetwork.NetworkingConfig, containerName string) (dockercontainer.ContainerCreateCreatedBody, error) {
  98         return proxy.Docker.ContainerCreate(ctx, config, hostConfig, networkingConfig, containerName)
  99 }
 100
 101 // ContainerStart invokes dockerclient.Client.ContainerStart
 102 func (proxy ThinDockerClientProxy) ContainerStart(ctx context.Context, container string, options dockertypes.ContainerStartOptions) error {
 103         return proxy.Docker.ContainerStart(ctx, container, options)
 104 }
 105
 106 // ContainerStop invokes dockerclient.Client.ContainerStop
 107 func (proxy ThinDockerClientProxy) ContainerStop(ctx context.Context, container string, timeout *time.Duration) error {
 108         return proxy.Docker.ContainerStop(ctx, container, timeout)
 109 }
 110
 111 // ContainerWait invokes dockerclient.Client.ContainerWait
 112 func (proxy ThinDockerClientProxy) ContainerWait(ctx context.Context, container string, condition dockercontainer.WaitCondition) (<-chan dockercontainer.ContainerWaitOKBody, <-chan error) {
 113         return proxy.Docker.ContainerWait(ctx, container, condition)
 114 }
 115
 116 // ImageInspectWithRaw invokes dockerclient.Client.ImageInspectWithRaw
 117 func (proxy ThinDockerClientProxy) ImageInspectWithRaw(ctx context.Context, image string) (dockertypes.ImageInspect, []byte, error) {
 118         return proxy.Docker.ImageInspectWithRaw(ctx, image)
 119 }
 120
 121 // ImageLoad invokes dockerclient.Client.ImageLoad
 122 func (proxy ThinDockerClientProxy) ImageLoad(ctx context.Context, input io.Reader, quiet bool) (dockertypes.ImageLoadResponse, error) {
 123         return proxy.Docker.ImageLoad(ctx, input, quiet)
 124 }
 125
 126 // ImageRemove invokes dockerclient.Client.ImageRemove
 127 func (proxy ThinDockerClientProxy) ImageRemove(ctx context.Context, image string, options dockertypes.ImageRemoveOptions) ([]dockertypes.ImageDeleteResponseItem, error) {
 128         return proxy.Docker.ImageRemove(ctx, image, options)
 129 }
 130
 131 // ContainerRunner is the main stateful struct used for a single execution of a
 132 // container.
 133 type ContainerRunner struct {
 134         Docker    ThinDockerClient
 135         ArvClient IArvadosClient
 136         Kc        IKeepClient
 137         arvados.Container
 138         ContainerConfig dockercontainer.Config
 139         dockercontainer.HostConfig
 140         token       string
 141         ContainerID string
 142         ExitCode    *int
 143         NewLogWriter
 144         loggingDone   chan bool
 145         CrunchLog     *ThrottledLogger
 146         Stdout        io.WriteCloser
 147         Stderr        io.WriteCloser
 148         LogCollection *CollectionWriter
 149         LogsPDH       *string
 150         RunArvMount
 151         MkTempDir
 152         ArvMount       *exec.Cmd
 153         ArvMountPoint  string
 154         HostOutputDir  string
 155         CleanupTempDir []string
 156         Binds          []string
 157         Volumes        map[string]struct{}
 158         OutputPDH      *string
 159         SigChan        chan os.Signal
 160         ArvMountExit   chan error
 161         finalState     string
 162
 163         statLogger   io.WriteCloser
 164         statReporter *crunchstat.Reporter
 165         statInterval time.Duration
 166         cgroupRoot   string
 167         // What we expect the container's cgroup parent to be.
 168         expectCgroupParent string
 169         // What we tell docker to use as the container's cgroup
 170         // parent. Note: Ideally we would use the same field for both
 171         // expectCgroupParent and setCgroupParent, and just make it
 172         // default to "docker". However, when using docker < 1.10 with
 173         // systemd, specifying a non-empty cgroup parent (even the
 174         // default value "docker") hits a docker bug
 175         // (https://github.com/docker/docker/issues/17126). Using two
 176         // separate fields makes it possible to use the "expect cgroup
 177         // parent to be X" feature even on sites where the "specify
 178         // cgroup parent" feature breaks.
 179         setCgroupParent string
 180
 181         cStateLock sync.Mutex
 182         cStarted   bool // StartContainer() succeeded
 183         cCancelled bool // StopContainer() invoked
 184
 185         enableNetwork string // one of "default" or "always"
 186         networkMode   string // passed through to HostConfig.NetworkMode
 187         arvMountLog   *ThrottledLogger
 188 }
 189
 190 // setupSignals sets up signal handling to gracefully terminate the underlying
 191 // Docker container and update state when receiving a TERM, INT or QUIT signal.
 192 func (runner *ContainerRunner) setupSignals() {
 193         runner.SigChan = make(chan os.Signal, 1)
 194         signal.Notify(runner.SigChan, syscall.SIGTERM)
 195         signal.Notify(runner.SigChan, syscall.SIGINT)
 196         signal.Notify(runner.SigChan, syscall.SIGQUIT)
 197
 198         go func(sig chan os.Signal) {
 199                 s := <-sig
 200                 if s != nil {
 201                         runner.CrunchLog.Printf("Caught signal %v", s)
 202                 }
 203                 runner.stop()
 204         }(runner.SigChan)
 205 }
 206
 207 // stop the underlying Docker container.
 208 func (runner *ContainerRunner) stop() {
 209         runner.cStateLock.Lock()
 210         defer runner.cStateLock.Unlock()
 211         if runner.cCancelled {
 212                 return
 213         }
 214         runner.cCancelled = true
 215         if runner.cStarted {
 216                 timeout := time.Duration(10)
 217                 err := runner.Docker.ContainerStop(context.TODO(), runner.ContainerID, &(timeout))
 218                 if err != nil {
 219                         runner.CrunchLog.Printf("StopContainer failed: %s", err)
 220                 }
 221                 // Suppress multiple calls to stop()
 222                 runner.cStarted = false
 223         }
 224 }
 225
 226 func (runner *ContainerRunner) stopSignals() {
 227         if runner.SigChan != nil {
 228                 signal.Stop(runner.SigChan)
 229                 close(runner.SigChan)
 230         }
 231 }
 232
 233 var errorBlacklist = []string{"Cannot connect to the Docker daemon"}
 234 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)")
 235
 236 func (runner *ContainerRunner) checkBrokenNode(goterr error) bool {
 237         for _, d := range errorBlacklist {
 238                 if strings.Index(goterr.Error(), d) != -1 {
 239                         runner.CrunchLog.Printf("Error suggests node is unable to run containers: %v", goterr)
 240                         if *brokenNodeHook == "" {
 241                                 runner.CrunchLog.Printf("No broken node hook provided, cannot mark node as broken.")
 242                         } else {
 243                                 runner.CrunchLog.Printf("Running broken node hook %q", *brokenNodeHook)
 244                                 // run killme script
 245                                 c := exec.Command(*brokenNodeHook)
 246                                 c.Stdout = runner.CrunchLog
 247                                 c.Stderr = runner.CrunchLog
 248                                 err := c.Run()
 249                                 if err != nil {
 250                                         runner.CrunchLog.Printf("Error running broken node hook: %v", err)
 251                                 }
 252                         }
 253                         return true
 254                 }
 255         }
 256         return false
 257 }
 258
 259 // LoadImage determines the docker image id from the container record and
 260 // checks if it is available in the local Docker image store.  If not, it loads
 261 // the image from Keep.
 262 func (runner *ContainerRunner) LoadImage() (err error) {
 263
 264         runner.CrunchLog.Printf("Fetching Docker image from collection '%s'", runner.Container.ContainerImage)
 265
 266         var collection arvados.Collection
 267         err = runner.ArvClient.Get("collections", runner.Container.ContainerImage, nil, &collection)
 268         if err != nil {
 269                 return fmt.Errorf("While getting container image collection: %v", err)
 270         }
 271         manifest := manifest.Manifest{Text: collection.ManifestText}
 272         var img, imageID string
 273         for ms := range manifest.StreamIter() {
 274                 img = ms.FileStreamSegments[0].Name
 275                 if !strings.HasSuffix(img, ".tar") {
 276                         return fmt.Errorf("First file in the container image collection does not end in .tar")
 277                 }
 278                 imageID = img[:len(img)-4]
 279         }
 280
 281         runner.CrunchLog.Printf("Using Docker image id '%s'", imageID)
 282
 283         _, _, err = runner.Docker.ImageInspectWithRaw(context.TODO(), imageID)
 284         if err != nil {
 285                 runner.CrunchLog.Print("Loading Docker image from keep")
 286
 287                 var readCloser io.ReadCloser
 288                 readCloser, err = runner.Kc.ManifestFileReader(manifest, img)
 289                 if err != nil {
 290                         return fmt.Errorf("While creating ManifestFileReader for container image: %v", err)
 291                 }
 292
 293                 response, err := runner.Docker.ImageLoad(context.TODO(), readCloser, true)
 294                 if err != nil {
 295                         return fmt.Errorf("While loading container image into Docker: %v", err)
 296                 }
 297
 298                 defer response.Body.Close()
 299                 rbody, err := ioutil.ReadAll(response.Body)
 300                 if err != nil {
 301                         return fmt.Errorf("Reading response to image load: %v", err)
 302                 }
 303                 runner.CrunchLog.Printf("Docker response: %s", rbody)
 304         } else {
 305                 runner.CrunchLog.Print("Docker image is available")
 306         }
 307
 308         runner.ContainerConfig.Image = imageID
 309
 310         runner.Kc.ClearBlockCache()
 311
 312         return nil
 313 }
 314
 315 func (runner *ContainerRunner) ArvMountCmd(arvMountCmd []string, token string) (c *exec.Cmd, err error) {
 316         c = exec.Command("arv-mount", arvMountCmd...)
 317
 318         // Copy our environment, but override ARVADOS_API_TOKEN with
 319         // the container auth token.
 320         c.Env = nil
 321         for _, s := range os.Environ() {
 322                 if !strings.HasPrefix(s, "ARVADOS_API_TOKEN=") {
 323                         c.Env = append(c.Env, s)
 324                 }
 325         }
 326         c.Env = append(c.Env, "ARVADOS_API_TOKEN="+token)
 327
 328         runner.arvMountLog = NewThrottledLogger(runner.NewLogWriter("arv-mount"))
 329         c.Stdout = runner.arvMountLog
 330         c.Stderr = runner.arvMountLog
 331
 332         runner.CrunchLog.Printf("Running %v", c.Args)
 333
 334         err = c.Start()
 335         if err != nil {
 336                 return nil, err
 337         }
 338
 339         statReadme := make(chan bool)
 340         runner.ArvMountExit = make(chan error)
 341
 342         keepStatting := true
 343         go func() {
 344                 for keepStatting {
 345                         time.Sleep(100 * time.Millisecond)
 346                         _, err = os.Stat(fmt.Sprintf("%s/by_id/README", runner.ArvMountPoint))
 347                         if err == nil {
 348                                 keepStatting = false
 349                                 statReadme <- true
 350                         }
 351                 }
 352                 close(statReadme)
 353         }()
 354
 355         go func() {
 356                 mnterr := c.Wait()
 357                 if mnterr != nil {
 358                         runner.CrunchLog.Printf("Arv-mount exit error: %v", mnterr)
 359                 }
 360                 runner.ArvMountExit <- mnterr
 361                 close(runner.ArvMountExit)
 362         }()
 363
 364         select {
 365         case <-statReadme:
 366                 break
 367         case err := <-runner.ArvMountExit:
 368                 runner.ArvMount = nil
 369                 keepStatting = false
 370                 return nil, err
 371         }
 372
 373         return c, nil
 374 }
 375
 376 func (runner *ContainerRunner) SetupArvMountPoint(prefix string) (err error) {
 377         if runner.ArvMountPoint == "" {
 378                 runner.ArvMountPoint, err = runner.MkTempDir("", prefix)
 379         }
 380         return
 381 }
 382
 383 func (runner *ContainerRunner) SetupMounts() (err error) {
 384         err = runner.SetupArvMountPoint("keep")
 385         if err != nil {
 386                 return fmt.Errorf("While creating keep mount temp dir: %v", err)
 387         }
 388
 389         pdhOnly := true
 390         tmpcount := 0
 391         arvMountCmd := []string{
 392                 "--foreground",
 393                 "--allow-other",
 394                 "--read-write",
 395                 fmt.Sprintf("--crunchstat-interval=%v", runner.statInterval.Seconds())}
 396
 397         if runner.Container.RuntimeConstraints.KeepCacheRAM > 0 {
 398                 arvMountCmd = append(arvMountCmd, "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheRAM))
 399         }
 400
 401         collectionPaths := []string{}
 402         runner.Binds = nil
 403         runner.Volumes = make(map[string]struct{})
 404         needCertMount := true
 405
 406         var binds []string
 407         for bind := range runner.Container.Mounts {
 408                 binds = append(binds, bind)
 409         }
 410         sort.Strings(binds)
 411
 412         for _, bind := range binds {
 413                 mnt := runner.Container.Mounts[bind]
 414                 if bind == "stdout" || bind == "stderr" {
 415                         // Is it a "file" mount kind?
 416                         if mnt.Kind != "file" {
 417                                 return fmt.Errorf("Unsupported mount kind '%s' for %s. Only 'file' is supported.", mnt.Kind, bind)
 418                         }
 419
 420                         // Does path start with OutputPath?
 421                         prefix := runner.Container.OutputPath
 422                         if !strings.HasSuffix(prefix, "/") {
 423                                 prefix += "/"
 424                         }
 425                         if !strings.HasPrefix(mnt.Path, prefix) {
 426                                 return fmt.Errorf("%s path does not start with OutputPath: %s, %s", strings.Title(bind), mnt.Path, prefix)
 427                         }
 428                 }
 429
 430                 if bind == "stdin" {
 431                         // Is it a "collection" mount kind?
 432                         if mnt.Kind != "collection" && mnt.Kind != "json" {
 433                                 return fmt.Errorf("Unsupported mount kind '%s' for stdin. Only 'collection' or 'json' are supported.", mnt.Kind)
 434                         }
 435                 }
 436
 437                 if bind == "/etc/arvados/ca-certificates.crt" {
 438                         needCertMount = false
 439                 }
 440
 441                 if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && bind != runner.Container.OutputPath+"/" {
 442                         if mnt.Kind != "collection" {
 443                                 return fmt.Errorf("Only mount points of kind 'collection' are supported underneath the output_path: %v", bind)
 444                         }
 445                 }
 446
 447                 switch {
 448                 case mnt.Kind == "collection" && bind != "stdin":
 449                         var src string
 450                         if mnt.UUID != "" && mnt.PortableDataHash != "" {
 451                                 return fmt.Errorf("Cannot specify both 'uuid' and 'portable_data_hash' for a collection mount")
 452                         }
 453                         if mnt.UUID != "" {
 454                                 if mnt.Writable {
 455                                         return fmt.Errorf("Writing to existing collections currently not permitted.")
 456                                 }
 457                                 pdhOnly = false
 458                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.UUID)
 459                         } else if mnt.PortableDataHash != "" {
 460                                 if mnt.Writable {
 461                                         return fmt.Errorf("Can never write to a collection specified by portable data hash")
 462                                 }
 463                                 idx := strings.Index(mnt.PortableDataHash, "/")
 464                                 if idx > 0 {
 465                                         mnt.Path = path.Clean(mnt.PortableDataHash[idx:])
 466                                         mnt.PortableDataHash = mnt.PortableDataHash[0:idx]
 467                                         runner.Container.Mounts[bind] = mnt
 468                                 }
 469                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.PortableDataHash)
 470                                 if mnt.Path != "" && mnt.Path != "." {
 471                                         if strings.HasPrefix(mnt.Path, "./") {
 472                                                 mnt.Path = mnt.Path[2:]
 473                                         } else if strings.HasPrefix(mnt.Path, "/") {
 474                                                 mnt.Path = mnt.Path[1:]
 475                                         }
 476                                         src += "/" + mnt.Path
 477                                 }
 478                         } else {
 479                                 src = fmt.Sprintf("%s/tmp%d", runner.ArvMountPoint, tmpcount)
 480                                 arvMountCmd = append(arvMountCmd, "--mount-tmp")
 481                                 arvMountCmd = append(arvMountCmd, fmt.Sprintf("tmp%d", tmpcount))
 482                                 tmpcount += 1
 483                         }
 484                         if mnt.Writable {
 485                                 if bind == runner.Container.OutputPath {
 486                                         runner.HostOutputDir = src
 487                                 } else if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 488                                         return fmt.Errorf("Writable mount points are not permitted underneath the output_path: %v", bind)
 489                                 }
 490                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", src, bind))
 491                         } else {
 492                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", src, bind))
 493                         }
 494                         collectionPaths = append(collectionPaths, src)
 495
 496                 case mnt.Kind == "tmp":
 497                         var tmpdir string
 498                         tmpdir, err = runner.MkTempDir("", "")
 499                         if err != nil {
 500                                 return fmt.Errorf("While creating mount temp dir: %v", err)
 501                         }
 502                         st, staterr := os.Stat(tmpdir)
 503                         if staterr != nil {
 504                                 return fmt.Errorf("While Stat on temp dir: %v", staterr)
 505                         }
 506                         err = os.Chmod(tmpdir, st.Mode()|os.ModeSetgid|0777)
 507                         if staterr != nil {
 508                                 return fmt.Errorf("While Chmod temp dir: %v", err)
 509                         }
 510                         runner.CleanupTempDir = append(runner.CleanupTempDir, tmpdir)
 511                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", tmpdir, bind))
 512                         if bind == runner.Container.OutputPath {
 513                                 runner.HostOutputDir = tmpdir
 514                         }
 515
 516                 case mnt.Kind == "json":
 517                         jsondata, err := json.Marshal(mnt.Content)
 518                         if err != nil {
 519                                 return fmt.Errorf("encoding json data: %v", err)
 520                         }
 521                         // Create a tempdir with a single file
 522                         // (instead of just a tempfile): this way we
 523                         // can ensure the file is world-readable
 524                         // inside the container, without having to
 525                         // make it world-readable on the docker host.
 526                         tmpdir, err := runner.MkTempDir("", "")
 527                         if err != nil {
 528                                 return fmt.Errorf("creating temp dir: %v", err)
 529                         }
 530                         runner.CleanupTempDir = append(runner.CleanupTempDir, tmpdir)
 531                         tmpfn := filepath.Join(tmpdir, "mountdata.json")
 532                         err = ioutil.WriteFile(tmpfn, jsondata, 0644)
 533                         if err != nil {
 534                                 return fmt.Errorf("writing temp file: %v", err)
 535                         }
 536                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", tmpfn, bind))
 537                 }
 538         }
 539
 540         if runner.HostOutputDir == "" {
 541                 return fmt.Errorf("Output path does not correspond to a writable mount point")
 542         }
 543
 544         if wantAPI := runner.Container.RuntimeConstraints.API; needCertMount && wantAPI != nil && *wantAPI {
 545                 for _, certfile := range arvadosclient.CertFiles {
 546                         _, err := os.Stat(certfile)
 547                         if err == nil {
 548                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:/etc/arvados/ca-certificates.crt:ro", certfile))
 549                                 break
 550                         }
 551                 }
 552         }
 553
 554         if pdhOnly {
 555                 arvMountCmd = append(arvMountCmd, "--mount-by-pdh", "by_id")
 556         } else {
 557                 arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_id")
 558         }
 559         arvMountCmd = append(arvMountCmd, runner.ArvMountPoint)
 560
 561         token, err := runner.ContainerToken()
 562         if err != nil {
 563                 return fmt.Errorf("could not get container token: %s", err)
 564         }
 565
 566         runner.ArvMount, err = runner.RunArvMount(arvMountCmd, token)
 567         if err != nil {
 568                 return fmt.Errorf("While trying to start arv-mount: %v", err)
 569         }
 570
 571         for _, p := range collectionPaths {
 572                 _, err = os.Stat(p)
 573                 if err != nil {
 574                         return fmt.Errorf("While checking that input files exist: %v", err)
 575                 }
 576         }
 577
 578         return nil
 579 }
 580
 581 func (runner *ContainerRunner) ProcessDockerAttach(containerReader io.Reader) {
 582         // Handle docker log protocol
 583         // https://docs.docker.com/engine/reference/api/docker_remote_api_v1.15/#attach-to-a-container
 584
 585         header := make([]byte, 8)
 586         for {
 587                 _, readerr := io.ReadAtLeast(containerReader, header, 8)
 588
 589                 if readerr == nil {
 590                         readsize := int64(header[7]) | (int64(header[6]) << 8) | (int64(header[5]) << 16) | (int64(header[4]) << 24)
 591                         if header[0] == 1 {
 592                                 // stdout
 593                                 _, readerr = io.CopyN(runner.Stdout, containerReader, readsize)
 594                         } else {
 595                                 // stderr
 596                                 _, readerr = io.CopyN(runner.Stderr, containerReader, readsize)
 597                         }
 598                 }
 599
 600                 if readerr != nil {
 601                         if readerr != io.EOF {
 602                                 runner.CrunchLog.Printf("While reading docker logs: %v", readerr)
 603                         }
 604
 605                         closeerr := runner.Stdout.Close()
 606                         if closeerr != nil {
 607                                 runner.CrunchLog.Printf("While closing stdout logs: %v", closeerr)
 608                         }
 609
 610                         closeerr = runner.Stderr.Close()
 611                         if closeerr != nil {
 612                                 runner.CrunchLog.Printf("While closing stderr logs: %v", closeerr)
 613                         }
 614
 615                         if runner.statReporter != nil {
 616                                 runner.statReporter.Stop()
 617                                 closeerr = runner.statLogger.Close()
 618                                 if closeerr != nil {
 619                                         runner.CrunchLog.Printf("While closing crunchstat logs: %v", closeerr)
 620                                 }
 621                         }
 622
 623                         runner.loggingDone <- true
 624                         close(runner.loggingDone)
 625                         return
 626                 }
 627         }
 628 }
 629
 630 func (runner *ContainerRunner) StartCrunchstat() {
 631         runner.statLogger = NewThrottledLogger(runner.NewLogWriter("crunchstat"))
 632         runner.statReporter = &crunchstat.Reporter{
 633                 CID:          runner.ContainerID,
 634                 Logger:       log.New(runner.statLogger, "", 0),
 635                 CgroupParent: runner.expectCgroupParent,
 636                 CgroupRoot:   runner.cgroupRoot,
 637                 PollPeriod:   runner.statInterval,
 638         }
 639         runner.statReporter.Start()
 640 }
 641
 642 type infoCommand struct {
 643         label string
 644         cmd   []string
 645 }
 646
 647 // LogNodeInfo gathers node information and store it on the log for debugging
 648 // purposes.
 649 func (runner *ContainerRunner) LogNodeInfo() (err error) {
 650         w := runner.NewLogWriter("node-info")
 651         logger := log.New(w, "node-info", 0)
 652
 653         commands := []infoCommand{
 654                 {
 655                         label: "Host Information",
 656                         cmd:   []string{"uname", "-a"},
 657                 },
 658                 {
 659                         label: "CPU Information",
 660                         cmd:   []string{"cat", "/proc/cpuinfo"},
 661                 },
 662                 {
 663                         label: "Memory Information",
 664                         cmd:   []string{"cat", "/proc/meminfo"},
 665                 },
 666                 {
 667                         label: "Disk Space",
 668                         cmd:   []string{"df", "-m", "/", os.TempDir()},
 669                 },
 670                 {
 671                         label: "Disk INodes",
 672                         cmd:   []string{"df", "-i", "/", os.TempDir()},
 673                 },
 674         }
 675
 676         // Run commands with informational output to be logged.
 677         var out []byte
 678         for _, command := range commands {
 679                 out, err = exec.Command(command.cmd[0], command.cmd[1:]...).CombinedOutput()
 680                 if err != nil {
 681                         return fmt.Errorf("While running command %q: %v",
 682                                 command.cmd, err)
 683                 }
 684                 logger.Println(command.label)
 685                 for _, line := range strings.Split(string(out), "\n") {
 686                         logger.Println(" ", line)
 687                 }
 688         }
 689
 690         err = w.Close()
 691         if err != nil {
 692                 return fmt.Errorf("While closing node-info logs: %v", err)
 693         }
 694         return nil
 695 }
 696
 697 // LogContainerRecord gets and saves the raw JSON container record from the API server
 698 func (runner *ContainerRunner) LogContainerRecord() (err error) {
 699         w := &ArvLogWriter{
 700                 ArvClient:     runner.ArvClient,
 701                 UUID:          runner.Container.UUID,
 702                 loggingStream: "container",
 703                 writeCloser:   runner.LogCollection.Open("container.json"),
 704         }
 705
 706         // Get Container record JSON from the API Server
 707         reader, err := runner.ArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
 708         if err != nil {
 709                 return fmt.Errorf("While retrieving container record from the API server: %v", err)
 710         }
 711         defer reader.Close()
 712
 713         dec := json.NewDecoder(reader)
 714         dec.UseNumber()
 715         var cr map[string]interface{}
 716         if err = dec.Decode(&cr); err != nil {
 717                 return fmt.Errorf("While decoding the container record JSON response: %v", err)
 718         }
 719         // Re-encode it using indentation to improve readability
 720         enc := json.NewEncoder(w)
 721         enc.SetIndent("", "    ")
 722         if err = enc.Encode(cr); err != nil {
 723                 return fmt.Errorf("While logging the JSON container record: %v", err)
 724         }
 725         err = w.Close()
 726         if err != nil {
 727                 return fmt.Errorf("While closing container.json log: %v", err)
 728         }
 729         return nil
 730 }
 731
 732 // AttachStreams connects the docker container stdin, stdout and stderr logs
 733 // to the Arvados logger which logs to Keep and the API server logs table.
 734 func (runner *ContainerRunner) AttachStreams() (err error) {
 735
 736         runner.CrunchLog.Print("Attaching container streams")
 737
 738         // If stdin mount is provided, attach it to the docker container
 739         var stdinRdr arvados.File
 740         var stdinJson []byte
 741         if stdinMnt, ok := runner.Container.Mounts["stdin"]; ok {
 742                 if stdinMnt.Kind == "collection" {
 743                         var stdinColl arvados.Collection
 744                         collId := stdinMnt.UUID
 745                         if collId == "" {
 746                                 collId = stdinMnt.PortableDataHash
 747                         }
 748                         err = runner.ArvClient.Get("collections", collId, nil, &stdinColl)
 749                         if err != nil {
 750                                 return fmt.Errorf("While getting stding collection: %v", err)
 751                         }
 752
 753                         stdinRdr, err = runner.Kc.ManifestFileReader(manifest.Manifest{Text: stdinColl.ManifestText}, stdinMnt.Path)
 754                         if os.IsNotExist(err) {
 755                                 return fmt.Errorf("stdin collection path not found: %v", stdinMnt.Path)
 756                         } else if err != nil {
 757                                 return fmt.Errorf("While getting stdin collection path %v: %v", stdinMnt.Path, err)
 758                         }
 759                 } else if stdinMnt.Kind == "json" {
 760                         stdinJson, err = json.Marshal(stdinMnt.Content)
 761                         if err != nil {
 762                                 return fmt.Errorf("While encoding stdin json data: %v", err)
 763                         }
 764                 }
 765         }
 766
 767         stdinUsed := stdinRdr != nil || len(stdinJson) != 0
 768         response, err := runner.Docker.ContainerAttach(context.TODO(), runner.ContainerID,
 769                 dockertypes.ContainerAttachOptions{Stream: true, Stdin: stdinUsed, Stdout: true, Stderr: true})
 770         if err != nil {
 771                 return fmt.Errorf("While attaching container stdout/stderr streams: %v", err)
 772         }
 773
 774         runner.loggingDone = make(chan bool)
 775
 776         if stdoutMnt, ok := runner.Container.Mounts["stdout"]; ok {
 777                 stdoutFile, err := runner.getStdoutFile(stdoutMnt.Path)
 778                 if err != nil {
 779                         return err
 780                 }
 781                 runner.Stdout = stdoutFile
 782         } else {
 783                 runner.Stdout = NewThrottledLogger(runner.NewLogWriter("stdout"))
 784         }
 785
 786         if stderrMnt, ok := runner.Container.Mounts["stderr"]; ok {
 787                 stderrFile, err := runner.getStdoutFile(stderrMnt.Path)
 788                 if err != nil {
 789                         return err
 790                 }
 791                 runner.Stderr = stderrFile
 792         } else {
 793                 runner.Stderr = NewThrottledLogger(runner.NewLogWriter("stderr"))
 794         }
 795
 796         if stdinRdr != nil {
 797                 go func() {
 798                         _, err := io.Copy(response.Conn, stdinRdr)
 799                         if err != nil {
 800                                 runner.CrunchLog.Print("While writing stdin collection to docker container %q", err)
 801                                 runner.stop()
 802                         }
 803                         stdinRdr.Close()
 804                         response.CloseWrite()
 805                 }()
 806         } else if len(stdinJson) != 0 {
 807                 go func() {
 808                         _, err := io.Copy(response.Conn, bytes.NewReader(stdinJson))
 809                         if err != nil {
 810                                 runner.CrunchLog.Print("While writing stdin json to docker container %q", err)
 811                                 runner.stop()
 812                         }
 813                         response.CloseWrite()
 814                 }()
 815         }
 816
 817         go runner.ProcessDockerAttach(response.Reader)
 818
 819         return nil
 820 }
 821
 822 func (runner *ContainerRunner) getStdoutFile(mntPath string) (*os.File, error) {
 823         stdoutPath := mntPath[len(runner.Container.OutputPath):]
 824         index := strings.LastIndex(stdoutPath, "/")
 825         if index > 0 {
 826                 subdirs := stdoutPath[:index]
 827                 if subdirs != "" {
 828                         st, err := os.Stat(runner.HostOutputDir)
 829                         if err != nil {
 830                                 return nil, fmt.Errorf("While Stat on temp dir: %v", err)
 831                         }
 832                         stdoutPath := filepath.Join(runner.HostOutputDir, subdirs)
 833                         err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777)
 834                         if err != nil {
 835                                 return nil, fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err)
 836                         }
 837                 }
 838         }
 839         stdoutFile, err := os.Create(filepath.Join(runner.HostOutputDir, stdoutPath))
 840         if err != nil {
 841                 return nil, fmt.Errorf("While creating file %q: %v", stdoutPath, err)
 842         }
 843
 844         return stdoutFile, nil
 845 }
 846
 847 // CreateContainer creates the docker container.
 848 func (runner *ContainerRunner) CreateContainer() error {
 849         runner.CrunchLog.Print("Creating Docker container")
 850
 851         runner.ContainerConfig.Cmd = runner.Container.Command
 852         if runner.Container.Cwd != "." {
 853                 runner.ContainerConfig.WorkingDir = runner.Container.Cwd
 854         }
 855
 856         for k, v := range runner.Container.Environment {
 857                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env, k+"="+v)
 858         }
 859
 860         runner.ContainerConfig.Volumes = runner.Volumes
 861
 862         runner.HostConfig = dockercontainer.HostConfig{
 863                 Binds: runner.Binds,
 864                 LogConfig: dockercontainer.LogConfig{
 865                         Type: "none",
 866                 },
 867                 Resources: dockercontainer.Resources{
 868                         CgroupParent: runner.setCgroupParent,
 869                 },
 870         }
 871
 872         if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
 873                 tok, err := runner.ContainerToken()
 874                 if err != nil {
 875                         return err
 876                 }
 877                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env,
 878                         "ARVADOS_API_TOKEN="+tok,
 879                         "ARVADOS_API_HOST="+os.Getenv("ARVADOS_API_HOST"),
 880                         "ARVADOS_API_HOST_INSECURE="+os.Getenv("ARVADOS_API_HOST_INSECURE"),
 881                 )
 882                 runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
 883         } else {
 884                 if runner.enableNetwork == "always" {
 885                         runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
 886                 } else {
 887                         runner.HostConfig.NetworkMode = dockercontainer.NetworkMode("none")
 888                 }
 889         }
 890
 891         _, stdinUsed := runner.Container.Mounts["stdin"]
 892         runner.ContainerConfig.OpenStdin = stdinUsed
 893         runner.ContainerConfig.StdinOnce = stdinUsed
 894         runner.ContainerConfig.AttachStdin = stdinUsed
 895         runner.ContainerConfig.AttachStdout = true
 896         runner.ContainerConfig.AttachStderr = true
 897
 898         createdBody, err := runner.Docker.ContainerCreate(context.TODO(), &runner.ContainerConfig, &runner.HostConfig, nil, runner.Container.UUID)
 899         if err != nil {
 900                 return fmt.Errorf("While creating container: %v", err)
 901         }
 902
 903         runner.ContainerID = createdBody.ID
 904
 905         return runner.AttachStreams()
 906 }
 907
 908 // StartContainer starts the docker container created by CreateContainer.
 909 func (runner *ContainerRunner) StartContainer() error {
 910         runner.CrunchLog.Printf("Starting Docker container id '%s'", runner.ContainerID)
 911         runner.cStateLock.Lock()
 912         defer runner.cStateLock.Unlock()
 913         if runner.cCancelled {
 914                 return ErrCancelled
 915         }
 916         err := runner.Docker.ContainerStart(context.TODO(), runner.ContainerID,
 917                 dockertypes.ContainerStartOptions{})
 918         if err != nil {
 919                 var advice string
 920                 if strings.Contains(err.Error(), "no such file or directory") {
 921                         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])
 922                 }
 923                 return fmt.Errorf("could not start container: %v%s", err, advice)
 924         }
 925         runner.cStarted = true
 926         return nil
 927 }
 928
 929 // WaitFinish waits for the container to terminate, capture the exit code, and
 930 // close the stdout/stderr logging.
 931 func (runner *ContainerRunner) WaitFinish() (err error) {
 932         runner.CrunchLog.Print("Waiting for container to finish")
 933
 934         waitOk, waitErr := runner.Docker.ContainerWait(context.TODO(), runner.ContainerID, "not-running")
 935
 936         go func() {
 937                 <-runner.ArvMountExit
 938                 if runner.cStarted {
 939                         runner.CrunchLog.Printf("arv-mount exited while container is still running.  Stopping container.")
 940                         runner.stop()
 941                 }
 942         }()
 943
 944         var waitBody dockercontainer.ContainerWaitOKBody
 945         select {
 946         case waitBody = <-waitOk:
 947         case err = <-waitErr:
 948         }
 949
 950         // Container isn't running any more
 951         runner.cStarted = false
 952
 953         if err != nil {
 954                 return fmt.Errorf("container wait: %v", err)
 955         }
 956
 957         runner.CrunchLog.Printf("Container exited with code: %v", waitBody.StatusCode)
 958         code := int(waitBody.StatusCode)
 959         runner.ExitCode = &code
 960
 961         // wait for stdout/stderr to complete
 962         <-runner.loggingDone
 963
 964         return nil
 965 }
 966
 967 var ErrNotInOutputDir = fmt.Errorf("Must point to path within the output directory")
 968
 969 func (runner *ContainerRunner) derefOutputSymlink(path string, startinfo os.FileInfo) (tgt string, readlinktgt string, info os.FileInfo, err error) {
 970         // Follow symlinks if necessary
 971         info = startinfo
 972         tgt = path
 973         readlinktgt = ""
 974         nextlink := path
 975         for followed := 0; info.Mode()&os.ModeSymlink != 0; followed++ {
 976                 if followed >= limitFollowSymlinks {
 977                         // Got stuck in a loop or just a pathological number of links, give up.
 978                         err = fmt.Errorf("Followed more than %v symlinks from path %q", limitFollowSymlinks, path)
 979                         return
 980                 }
 981
 982                 readlinktgt, err = os.Readlink(nextlink)
 983                 if err != nil {
 984                         return
 985                 }
 986
 987                 tgt = readlinktgt
 988                 if !strings.HasPrefix(tgt, "/") {
 989                         // Relative symlink, resolve it to host path
 990                         tgt = filepath.Join(filepath.Dir(path), tgt)
 991                 }
 992                 if strings.HasPrefix(tgt, runner.Container.OutputPath+"/") && !strings.HasPrefix(tgt, runner.HostOutputDir+"/") {
 993                         // Absolute symlink to container output path, adjust it to host output path.
 994                         tgt = filepath.Join(runner.HostOutputDir, tgt[len(runner.Container.OutputPath):])
 995                 }
 996                 if !strings.HasPrefix(tgt, runner.HostOutputDir+"/") {
 997                         // After dereferencing, symlink target must either be
 998                         // within output directory, or must point to a
 999                         // collection mount.
1000                         err = ErrNotInOutputDir
1001                         return
1002                 }
1003
1004                 info, err = os.Lstat(tgt)
1005                 if err != nil {
1006                         // tgt
1007                         err = fmt.Errorf("Symlink in output %q points to invalid location %q: %v",
1008                                 path[len(runner.HostOutputDir):], readlinktgt, err)
1009                         return
1010                 }
1011
1012                 nextlink = tgt
1013         }
1014
1015         return
1016 }
1017
1018 var limitFollowSymlinks = 10
1019
1020 // UploadFile uploads files within the output directory, with special handling
1021 // for symlinks. If the symlink leads to a keep mount, copy the manifest text
1022 // from the keep mount into the output manifestText.  Ensure that whether
1023 // symlinks are relative or absolute, every symlink target (even targets that
1024 // are symlinks themselves) must point to a path in either the output directory
1025 // or a collection mount.
1026 //
1027 // Assumes initial value of "path" is absolute, and located within runner.HostOutputDir.
1028 func (runner *ContainerRunner) UploadOutputFile(
1029         path string,
1030         info os.FileInfo,
1031         infoerr error,
1032         binds []string,
1033         walkUpload *WalkUpload,
1034         relocateFrom string,
1035         relocateTo string,
1036         followed int) (manifestText string, err error) {
1037
1038         if info.Mode().IsDir() {
1039                 return
1040         }
1041
1042         if infoerr != nil {
1043                 return "", infoerr
1044         }
1045
1046         if followed >= limitFollowSymlinks {
1047                 // Got stuck in a loop or just a pathological number of
1048                 // directory links, give up.
1049                 err = fmt.Errorf("Followed more than %v symlinks from path %q", limitFollowSymlinks, path)
1050                 return
1051         }
1052
1053         // When following symlinks, the source path may need to be logically
1054         // relocated to some other path within the output collection.  Remove
1055         // the relocateFrom prefix and replace it with relocateTo.
1056         relocated := relocateTo + path[len(relocateFrom):]
1057
1058         tgt, readlinktgt, info, derefErr := runner.derefOutputSymlink(path, info)
1059         if derefErr != nil && derefErr != ErrNotInOutputDir {
1060                 return "", derefErr
1061         }
1062
1063         // go through mounts and try reverse map to collection reference
1064         for _, bind := range binds {
1065                 mnt := runner.Container.Mounts[bind]
1066                 if tgt == bind || strings.HasPrefix(tgt, bind+"/") {
1067                         // get path relative to bind
1068                         targetSuffix := tgt[len(bind):]
1069
1070                         // Copy mount and adjust the path to add path relative to the bind
1071                         adjustedMount := mnt
1072                         adjustedMount.Path = filepath.Join(adjustedMount.Path, targetSuffix)
1073
1074                         // Terminates in this keep mount, so add the
1075                         // manifest text at appropriate location.
1076                         outputSuffix := path[len(runner.HostOutputDir):]
1077                         manifestText, err = runner.getCollectionManifestForPath(adjustedMount, outputSuffix)
1078                         return
1079                 }
1080         }
1081
1082         // If target is not a collection mount, it must be located within the
1083         // output directory, otherwise it is an error.
1084         if derefErr == ErrNotInOutputDir {
1085                 err = fmt.Errorf("Symlink in output %q points to invalid location %q, must point to path within the output directory.",
1086                         path[len(runner.HostOutputDir):], readlinktgt)
1087                 return
1088         }
1089
1090         if info.Mode().IsRegular() {
1091                 return "", walkUpload.UploadFile(relocated, tgt)
1092         }
1093
1094         if info.Mode().IsDir() {
1095                 // Symlink leads to directory.  Walk() doesn't follow
1096                 // directory symlinks, so we walk the target directory
1097                 // instead.  Within the walk, file paths are relocated
1098                 // so they appear under the original symlink path.
1099                 err = filepath.Walk(tgt, func(walkpath string, walkinfo os.FileInfo, walkerr error) error {
1100                         var m string
1101                         m, walkerr = runner.UploadOutputFile(walkpath, walkinfo, walkerr,
1102                                 binds, walkUpload, tgt, relocated, followed+1)
1103                         if walkerr == nil {
1104                                 manifestText = manifestText + m
1105                         }
1106                         return walkerr
1107                 })
1108                 return
1109         }
1110
1111         return
1112 }
1113
1114 // HandleOutput sets the output, unmounts the FUSE mount, and deletes temporary directories
1115 func (runner *ContainerRunner) CaptureOutput() error {
1116         if runner.finalState != "Complete" {
1117                 return nil
1118         }
1119
1120         if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
1121                 // Output may have been set directly by the container, so
1122                 // refresh the container record to check.
1123                 err := runner.ArvClient.Get("containers", runner.Container.UUID,
1124                         nil, &runner.Container)
1125                 if err != nil {
1126                         return err
1127                 }
1128                 if runner.Container.Output != "" {
1129                         // Container output is already set.
1130                         runner.OutputPDH = &runner.Container.Output
1131                         return nil
1132                 }
1133         }
1134
1135         if runner.HostOutputDir == "" {
1136                 return nil
1137         }
1138
1139         _, err := os.Stat(runner.HostOutputDir)
1140         if err != nil {
1141                 return fmt.Errorf("While checking host output path: %v", err)
1142         }
1143
1144         // Pre-populate output from the configured mount points
1145         var binds []string
1146         for bind, mnt := range runner.Container.Mounts {
1147                 if mnt.Kind == "collection" {
1148                         binds = append(binds, bind)
1149                 }
1150         }
1151         sort.Strings(binds)
1152
1153         var manifestText string
1154
1155         collectionMetafile := fmt.Sprintf("%s/.arvados#collection", runner.HostOutputDir)
1156         _, err = os.Stat(collectionMetafile)
1157         if err != nil {
1158                 // Regular directory
1159
1160                 cw := CollectionWriter{0, runner.Kc, nil, nil, sync.Mutex{}}
1161                 walkUpload := cw.BeginUpload(runner.HostOutputDir, runner.CrunchLog.Logger)
1162
1163                 var m string
1164                 err = filepath.Walk(runner.HostOutputDir, func(path string, info os.FileInfo, err error) error {
1165                         m, err = runner.UploadOutputFile(path, info, err, binds, walkUpload, "", "", 0)
1166                         if err == nil {
1167                                 manifestText = manifestText + m
1168                         }
1169                         return err
1170                 })
1171
1172                 cw.EndUpload(walkUpload)
1173
1174                 if err != nil {
1175                         return fmt.Errorf("While uploading output files: %v", err)
1176                 }
1177
1178                 m, err = cw.ManifestText()
1179                 manifestText = manifestText + m
1180                 if err != nil {
1181                         return fmt.Errorf("While uploading output files: %v", err)
1182                 }
1183         } else {
1184                 // FUSE mount directory
1185                 file, openerr := os.Open(collectionMetafile)
1186                 if openerr != nil {
1187                         return fmt.Errorf("While opening FUSE metafile: %v", err)
1188                 }
1189                 defer file.Close()
1190
1191                 var rec arvados.Collection
1192                 err = json.NewDecoder(file).Decode(&rec)
1193                 if err != nil {
1194                         return fmt.Errorf("While reading FUSE metafile: %v", err)
1195                 }
1196                 manifestText = rec.ManifestText
1197         }
1198
1199         for _, bind := range binds {
1200                 mnt := runner.Container.Mounts[bind]
1201
1202                 bindSuffix := strings.TrimPrefix(bind, runner.Container.OutputPath)
1203
1204                 if bindSuffix == bind || len(bindSuffix) <= 0 {
1205                         // either does not start with OutputPath or is OutputPath itself
1206                         continue
1207                 }
1208
1209                 if mnt.ExcludeFromOutput == true {
1210                         continue
1211                 }
1212
1213                 // append to manifest_text
1214                 m, err := runner.getCollectionManifestForPath(mnt, bindSuffix)
1215                 if err != nil {
1216                         return err
1217                 }
1218
1219                 manifestText = manifestText + m
1220         }
1221
1222         // Save output
1223         var response arvados.Collection
1224         manifest := manifest.Manifest{Text: manifestText}
1225         manifestText = manifest.Extract(".", ".").Text
1226         err = runner.ArvClient.Create("collections",
1227                 arvadosclient.Dict{
1228                         "ensure_unique_name": true,
1229                         "collection": arvadosclient.Dict{
1230                                 "is_trashed":    true,
1231                                 "name":          "output for " + runner.Container.UUID,
1232                                 "manifest_text": manifestText}},
1233                 &response)
1234         if err != nil {
1235                 return fmt.Errorf("While creating output collection: %v", err)
1236         }
1237         runner.OutputPDH = &response.PortableDataHash
1238         return nil
1239 }
1240
1241 var outputCollections = make(map[string]arvados.Collection)
1242
1243 // Fetch the collection for the mnt.PortableDataHash
1244 // Return the manifest_text fragment corresponding to the specified mnt.Path
1245 //  after making any required updates.
1246 //  Ex:
1247 //    If mnt.Path is not specified,
1248 //      return the entire manifest_text after replacing any "." with bindSuffix
1249 //    If mnt.Path corresponds to one stream,
1250 //      return the manifest_text for that stream after replacing that stream name with bindSuffix
1251 //    Otherwise, check if a filename in any one stream is being sought. Return the manifest_text
1252 //      for that stream after replacing stream name with bindSuffix minus the last word
1253 //      and the file name with last word of the bindSuffix
1254 //  Allowed path examples:
1255 //    "path":"/"
1256 //    "path":"/subdir1"
1257 //    "path":"/subdir1/subdir2"
1258 //    "path":"/subdir/filename" etc
1259 func (runner *ContainerRunner) getCollectionManifestForPath(mnt arvados.Mount, bindSuffix string) (string, error) {
1260         collection := outputCollections[mnt.PortableDataHash]
1261         if collection.PortableDataHash == "" {
1262                 err := runner.ArvClient.Get("collections", mnt.PortableDataHash, nil, &collection)
1263                 if err != nil {
1264                         return "", fmt.Errorf("While getting collection for %v: %v", mnt.PortableDataHash, err)
1265                 }
1266                 outputCollections[mnt.PortableDataHash] = collection
1267         }
1268
1269         if collection.ManifestText == "" {
1270                 runner.CrunchLog.Printf("No manifest text for collection %v", collection.PortableDataHash)
1271                 return "", nil
1272         }
1273
1274         mft := manifest.Manifest{Text: collection.ManifestText}
1275         extracted := mft.Extract(mnt.Path, bindSuffix)
1276         if extracted.Err != nil {
1277                 return "", fmt.Errorf("Error parsing manifest for %v: %v", mnt.PortableDataHash, extracted.Err.Error())
1278         }
1279         return extracted.Text, nil
1280 }
1281
1282 func (runner *ContainerRunner) CleanupDirs() {
1283         if runner.ArvMount != nil {
1284                 var delay int64 = 8
1285                 umount := exec.Command("arv-mount", fmt.Sprintf("--unmount-timeout=%d", delay), "--unmount", runner.ArvMountPoint)
1286                 umount.Stdout = runner.CrunchLog
1287                 umount.Stderr = runner.CrunchLog
1288                 runner.CrunchLog.Printf("Running %v", umount.Args)
1289                 umnterr := umount.Start()
1290
1291                 if umnterr != nil {
1292                         runner.CrunchLog.Printf("Error unmounting: %v", umnterr)
1293                 } else {
1294                         // If arv-mount --unmount gets stuck for any reason, we
1295                         // don't want to wait for it forever.  Do Wait() in a goroutine
1296                         // so it doesn't block crunch-run.
1297                         umountExit := make(chan error)
1298                         go func() {
1299                                 mnterr := umount.Wait()
1300                                 if mnterr != nil {
1301                                         runner.CrunchLog.Printf("Error unmounting: %v", mnterr)
1302                                 }
1303                                 umountExit <- mnterr
1304                         }()
1305
1306                         for again := true; again; {
1307                                 again = false
1308                                 select {
1309                                 case <-umountExit:
1310                                         umount = nil
1311                                         again = true
1312                                 case <-runner.ArvMountExit:
1313                                         break
1314                                 case <-time.After(time.Duration((delay + 1) * int64(time.Second))):
1315                                         runner.CrunchLog.Printf("Timed out waiting for unmount")
1316                                         if umount != nil {
1317                                                 umount.Process.Kill()
1318                                         }
1319                                         runner.ArvMount.Process.Kill()
1320                                 }
1321                         }
1322                 }
1323         }
1324
1325         if runner.ArvMountPoint != "" {
1326                 if rmerr := os.Remove(runner.ArvMountPoint); rmerr != nil {
1327                         runner.CrunchLog.Printf("While cleaning up arv-mount directory %s: %v", runner.ArvMountPoint, rmerr)
1328                 }
1329         }
1330
1331         for _, tmpdir := range runner.CleanupTempDir {
1332                 if rmerr := os.RemoveAll(tmpdir); rmerr != nil {
1333                         runner.CrunchLog.Printf("While cleaning up temporary directory %s: %v", tmpdir, rmerr)
1334                 }
1335         }
1336 }
1337
1338 // CommitLogs posts the collection containing the final container logs.
1339 func (runner *ContainerRunner) CommitLogs() error {
1340         runner.CrunchLog.Print(runner.finalState)
1341
1342         if runner.arvMountLog != nil {
1343                 runner.arvMountLog.Close()
1344         }
1345         runner.CrunchLog.Close()
1346
1347         // Closing CrunchLog above allows them to be committed to Keep at this
1348         // point, but re-open crunch log with ArvClient in case there are any
1349         // other further errors (such as failing to write the log to Keep!)
1350         // while shutting down
1351         runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{ArvClient: runner.ArvClient,
1352                 UUID: runner.Container.UUID, loggingStream: "crunch-run", writeCloser: nil})
1353         runner.CrunchLog.Immediate = log.New(os.Stderr, runner.Container.UUID+" ", 0)
1354
1355         if runner.LogsPDH != nil {
1356                 // If we have already assigned something to LogsPDH,
1357                 // we must be closing the re-opened log, which won't
1358                 // end up getting attached to the container record and
1359                 // therefore doesn't need to be saved as a collection
1360                 // -- it exists only to send logs to other channels.
1361                 return nil
1362         }
1363
1364         mt, err := runner.LogCollection.ManifestText()
1365         if err != nil {
1366                 return fmt.Errorf("While creating log manifest: %v", err)
1367         }
1368
1369         var response arvados.Collection
1370         err = runner.ArvClient.Create("collections",
1371                 arvadosclient.Dict{
1372                         "ensure_unique_name": true,
1373                         "collection": arvadosclient.Dict{
1374                                 "is_trashed":    true,
1375                                 "name":          "logs for " + runner.Container.UUID,
1376                                 "manifest_text": mt}},
1377                 &response)
1378         if err != nil {
1379                 return fmt.Errorf("While creating log collection: %v", err)
1380         }
1381         runner.LogsPDH = &response.PortableDataHash
1382         return nil
1383 }
1384
1385 // UpdateContainerRunning updates the container state to "Running"
1386 func (runner *ContainerRunner) UpdateContainerRunning() error {
1387         runner.cStateLock.Lock()
1388         defer runner.cStateLock.Unlock()
1389         if runner.cCancelled {
1390                 return ErrCancelled
1391         }
1392         return runner.ArvClient.Update("containers", runner.Container.UUID,
1393                 arvadosclient.Dict{"container": arvadosclient.Dict{"state": "Running"}}, nil)
1394 }
1395
1396 // ContainerToken returns the api_token the container (and any
1397 // arv-mount processes) are allowed to use.
1398 func (runner *ContainerRunner) ContainerToken() (string, error) {
1399         if runner.token != "" {
1400                 return runner.token, nil
1401         }
1402
1403         var auth arvados.APIClientAuthorization
1404         err := runner.ArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth)
1405         if err != nil {
1406                 return "", err
1407         }
1408         runner.token = auth.APIToken
1409         return runner.token, nil
1410 }
1411
1412 // UpdateContainerComplete updates the container record state on API
1413 // server to "Complete" or "Cancelled"
1414 func (runner *ContainerRunner) UpdateContainerFinal() error {
1415         update := arvadosclient.Dict{}
1416         update["state"] = runner.finalState
1417         if runner.LogsPDH != nil {
1418                 update["log"] = *runner.LogsPDH
1419         }
1420         if runner.finalState == "Complete" {
1421                 if runner.ExitCode != nil {
1422                         update["exit_code"] = *runner.ExitCode
1423                 }
1424                 if runner.OutputPDH != nil {
1425                         update["output"] = *runner.OutputPDH
1426                 }
1427         }
1428         return runner.ArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{"container": update}, nil)
1429 }
1430
1431 // IsCancelled returns the value of Cancelled, with goroutine safety.
1432 func (runner *ContainerRunner) IsCancelled() bool {
1433         runner.cStateLock.Lock()
1434         defer runner.cStateLock.Unlock()
1435         return runner.cCancelled
1436 }
1437
1438 // NewArvLogWriter creates an ArvLogWriter
1439 func (runner *ContainerRunner) NewArvLogWriter(name string) io.WriteCloser {
1440         return &ArvLogWriter{
1441                 ArvClient:     runner.ArvClient,
1442                 UUID:          runner.Container.UUID,
1443                 loggingStream: name,
1444                 writeCloser:   runner.LogCollection.Open(name + ".txt")}
1445 }
1446
1447 // Run the full container lifecycle.
1448 func (runner *ContainerRunner) Run() (err error) {
1449         runner.CrunchLog.Printf("crunch-run %s started", version)
1450         runner.CrunchLog.Printf("Executing container '%s'", runner.Container.UUID)
1451
1452         hostname, hosterr := os.Hostname()
1453         if hosterr != nil {
1454                 runner.CrunchLog.Printf("Error getting hostname '%v'", hosterr)
1455         } else {
1456                 runner.CrunchLog.Printf("Executing on host '%s'", hostname)
1457         }
1458
1459         runner.finalState = "Queued"
1460
1461         defer func() {
1462                 runner.stopSignals()
1463                 runner.CleanupDirs()
1464
1465                 runner.CrunchLog.Printf("crunch-run finished")
1466                 runner.CrunchLog.Close()
1467         }()
1468
1469         defer func() {
1470                 // checkErr prints e (unless it's nil) and sets err to
1471                 // e (unless err is already non-nil). Thus, if err
1472                 // hasn't already been assigned when Run() returns,
1473                 // this cleanup func will cause Run() to return the
1474                 // first non-nil error that is passed to checkErr().
1475                 checkErr := func(e error) {
1476                         if e == nil {
1477                                 return
1478                         }
1479                         runner.CrunchLog.Print(e)
1480                         if err == nil {
1481                                 err = e
1482                         }
1483                         if runner.finalState == "Complete" {
1484                                 // There was an error in the finalization.
1485                                 runner.finalState = "Cancelled"
1486                         }
1487                 }
1488
1489                 // Log the error encountered in Run(), if any
1490                 checkErr(err)
1491
1492                 if runner.finalState == "Queued" {
1493                         runner.UpdateContainerFinal()
1494                         return
1495                 }
1496
1497                 if runner.IsCancelled() {
1498                         runner.finalState = "Cancelled"
1499                         // but don't return yet -- we still want to
1500                         // capture partial output and write logs
1501                 }
1502
1503                 checkErr(runner.CaptureOutput())
1504                 checkErr(runner.CommitLogs())
1505                 checkErr(runner.UpdateContainerFinal())
1506         }()
1507
1508         err = runner.fetchContainerRecord()
1509         if err != nil {
1510                 return
1511         }
1512
1513         // setup signal handling
1514         runner.setupSignals()
1515
1516         // check for and/or load image
1517         err = runner.LoadImage()
1518         if err != nil {
1519                 if !runner.checkBrokenNode(err) {
1520                         // Failed to load image but not due to a "broken node"
1521                         // condition, probably user error.
1522                         runner.finalState = "Cancelled"
1523                 }
1524                 err = fmt.Errorf("While loading container image: %v", err)
1525                 return
1526         }
1527
1528         // set up FUSE mount and binds
1529         err = runner.SetupMounts()
1530         if err != nil {
1531                 runner.finalState = "Cancelled"
1532                 err = fmt.Errorf("While setting up mounts: %v", err)
1533                 return
1534         }
1535
1536         err = runner.CreateContainer()
1537         if err != nil {
1538                 return
1539         }
1540
1541         // Gather and record node information
1542         err = runner.LogNodeInfo()
1543         if err != nil {
1544                 return
1545         }
1546         // Save container.json record on log collection
1547         err = runner.LogContainerRecord()
1548         if err != nil {
1549                 return
1550         }
1551
1552         if runner.IsCancelled() {
1553                 return
1554         }
1555
1556         err = runner.UpdateContainerRunning()
1557         if err != nil {
1558                 return
1559         }
1560         runner.finalState = "Cancelled"
1561
1562         runner.StartCrunchstat()
1563
1564         err = runner.StartContainer()
1565         if err != nil {
1566                 runner.checkBrokenNode(err)
1567                 return
1568         }
1569
1570         err = runner.WaitFinish()
1571         if err == nil {
1572                 runner.finalState = "Complete"
1573         }
1574         return
1575 }
1576
1577 // Fetch the current container record (uuid = runner.Container.UUID)
1578 // into runner.Container.
1579 func (runner *ContainerRunner) fetchContainerRecord() error {
1580         reader, err := runner.ArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
1581         if err != nil {
1582                 return fmt.Errorf("error fetching container record: %v", err)
1583         }
1584         defer reader.Close()
1585
1586         dec := json.NewDecoder(reader)
1587         dec.UseNumber()
1588         err = dec.Decode(&runner.Container)
1589         if err != nil {
1590                 return fmt.Errorf("error decoding container record: %v", err)
1591         }
1592         return nil
1593 }
1594
1595 // NewContainerRunner creates a new container runner.
1596 func NewContainerRunner(api IArvadosClient,
1597         kc IKeepClient,
1598         docker ThinDockerClient,
1599         containerUUID string) *ContainerRunner {
1600
1601         cr := &ContainerRunner{ArvClient: api, Kc: kc, Docker: docker}
1602         cr.NewLogWriter = cr.NewArvLogWriter
1603         cr.RunArvMount = cr.ArvMountCmd
1604         cr.MkTempDir = ioutil.TempDir
1605         cr.LogCollection = &CollectionWriter{0, kc, nil, nil, sync.Mutex{}}
1606         cr.Container.UUID = containerUUID
1607         cr.CrunchLog = NewThrottledLogger(cr.NewLogWriter("crunch-run"))
1608         cr.CrunchLog.Immediate = log.New(os.Stderr, containerUUID+" ", 0)
1609
1610         loadLogThrottleParams(api)
1611
1612         return cr
1613 }
1614
1615 func main() {
1616         statInterval := flag.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting")
1617         cgroupRoot := flag.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree")
1618         cgroupParent := flag.String("cgroup-parent", "docker", "name of container's parent cgroup (ignored if -cgroup-parent-subsystem is used)")
1619         cgroupParentSubsystem := flag.String("cgroup-parent-subsystem", "", "use current cgroup for given subsystem as parent cgroup for container")
1620         caCertsPath := flag.String("ca-certs", "", "Path to TLS root certificates")
1621         enableNetwork := flag.String("container-enable-networking", "default",
1622                 `Specify if networking should be enabled for container.  One of 'default', 'always':
1623         default: only enable networking if container requests it.
1624         always:  containers always have networking enabled
1625         `)
1626         networkMode := flag.String("container-network-mode", "default",
1627                 `Set networking mode for container.  Corresponds to Docker network mode (--net).
1628         `)
1629         memprofile := flag.String("memprofile", "", "write memory profile to `file` after running container")
1630         getVersion := flag.Bool("version", false, "Print version information and exit.")
1631         flag.Parse()
1632
1633         // Print version information if requested
1634         if *getVersion {
1635                 fmt.Printf("crunch-run %s\n", version)
1636                 return
1637         }
1638
1639         log.Printf("crunch-run %s started", version)
1640
1641         containerId := flag.Arg(0)
1642
1643         if *caCertsPath != "" {
1644                 arvadosclient.CertFiles = []string{*caCertsPath}
1645         }
1646
1647         api, err := arvadosclient.MakeArvadosClient()
1648         if err != nil {
1649                 log.Fatalf("%s: %v", containerId, err)
1650         }
1651         api.Retries = 8
1652
1653         kc, kcerr := keepclient.MakeKeepClient(api)
1654         if kcerr != nil {
1655                 log.Fatalf("%s: %v", containerId, kcerr)
1656         }
1657         kc.BlockCache = &keepclient.BlockCache{MaxBlocks: 2}
1658         kc.Retries = 4
1659
1660         // API version 1.21 corresponds to Docker 1.9, which is currently the
1661         // minimum version we want to support.
1662         docker, dockererr := dockerclient.NewClient(dockerclient.DefaultDockerHost, "1.21", nil, nil)
1663         dockerClientProxy := ThinDockerClientProxy{Docker: docker}
1664
1665         cr := NewContainerRunner(api, kc, dockerClientProxy, containerId)
1666
1667         if dockererr != nil {
1668                 cr.CrunchLog.Printf("%s: %v", containerId, dockererr)
1669                 cr.checkBrokenNode(dockererr)
1670                 cr.CrunchLog.Close()
1671                 os.Exit(1)
1672         }
1673
1674         cr.statInterval = *statInterval
1675         cr.cgroupRoot = *cgroupRoot
1676         cr.expectCgroupParent = *cgroupParent
1677         cr.enableNetwork = *enableNetwork
1678         cr.networkMode = *networkMode
1679         if *cgroupParentSubsystem != "" {
1680                 p := findCgroup(*cgroupParentSubsystem)
1681                 cr.setCgroupParent = p
1682                 cr.expectCgroupParent = p
1683         }
1684
1685         runerr := cr.Run()
1686
1687         if *memprofile != "" {
1688                 f, err := os.Create(*memprofile)
1689                 if err != nil {
1690                         log.Printf("could not create memory profile: ", err)
1691                 }
1692                 runtime.GC() // get up-to-date statistics
1693                 if err := pprof.WriteHeapProfile(f); err != nil {
1694                         log.Printf("could not write memory profile: ", err)
1695                 }
1696                 closeerr := f.Close()
1697                 if closeerr != nil {
1698                         log.Printf("closing memprofile file: ", err)
1699                 }
1700         }
1701
1702         if runerr != nil {
1703                 log.Fatalf("%s: %v", containerId, runerr)
1704         }
1705 }