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         "regexp"
  23         "runtime"
  24         "runtime/pprof"
  25         "sort"
  26         "strings"
  27         "sync"
  28         "syscall"
  29         "time"
  30
  31         "git.curoverse.com/arvados.git/lib/crunchstat"
  32         "git.curoverse.com/arvados.git/sdk/go/arvados"
  33         "git.curoverse.com/arvados.git/sdk/go/arvadosclient"
  34         "git.curoverse.com/arvados.git/sdk/go/keepclient"
  35         "git.curoverse.com/arvados.git/sdk/go/manifest"
  36
  37         dockertypes "github.com/docker/docker/api/types"
  38         dockercontainer "github.com/docker/docker/api/types/container"
  39         dockernetwork "github.com/docker/docker/api/types/network"
  40         dockerclient "github.com/docker/docker/client"
  41 )
  42
  43 var version = "dev"
  44
  45 // IArvadosClient is the minimal Arvados API methods used by crunch-run.
  46 type IArvadosClient interface {
  47         Create(resourceType string, parameters arvadosclient.Dict, output interface{}) error
  48         Get(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  49         Update(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  50         Call(method, resourceType, uuid, action string, parameters arvadosclient.Dict, output interface{}) error
  51         CallRaw(method string, resourceType string, uuid string, action string, parameters arvadosclient.Dict) (reader io.ReadCloser, err error)
  52         Discovery(key string) (interface{}, error)
  53 }
  54
  55 // ErrCancelled is the error returned when the container is cancelled.
  56 var ErrCancelled = errors.New("Cancelled")
  57
  58 // IKeepClient is the minimal Keep API methods used by crunch-run.
  59 type IKeepClient interface {
  60         PutHB(hash string, buf []byte) (string, int, error)
  61         ManifestFileReader(m manifest.Manifest, filename string) (arvados.File, error)
  62         ClearBlockCache()
  63 }
  64
  65 // NewLogWriter is a factory function to create a new log writer.
  66 type NewLogWriter func(name string) io.WriteCloser
  67
  68 type RunArvMount func(args []string, tok string) (*exec.Cmd, error)
  69
  70 type MkTempDir func(string, string) (string, error)
  71
  72 // ThinDockerClient is the minimal Docker client interface used by crunch-run.
  73 type ThinDockerClient interface {
  74         ContainerAttach(ctx context.Context, container string, options dockertypes.ContainerAttachOptions) (dockertypes.HijackedResponse, error)
  75         ContainerCreate(ctx context.Context, config *dockercontainer.Config, hostConfig *dockercontainer.HostConfig,
  76                 networkingConfig *dockernetwork.NetworkingConfig, containerName string) (dockercontainer.ContainerCreateCreatedBody, error)
  77         ContainerStart(ctx context.Context, container string, options dockertypes.ContainerStartOptions) error
  78         ContainerStop(ctx context.Context, container string, timeout *time.Duration) error
  79         ContainerWait(ctx context.Context, container string, condition dockercontainer.WaitCondition) (<-chan dockercontainer.ContainerWaitOKBody, <-chan error)
  80         ImageInspectWithRaw(ctx context.Context, image string) (dockertypes.ImageInspect, []byte, error)
  81         ImageLoad(ctx context.Context, input io.Reader, quiet bool) (dockertypes.ImageLoadResponse, error)
  82         ImageRemove(ctx context.Context, image string, options dockertypes.ImageRemoveOptions) ([]dockertypes.ImageDeleteResponseItem, error)
  83 }
  84
  85 // ThinDockerClientProxy is a proxy implementation of ThinDockerClient
  86 // that executes the docker requests on dockerclient.Client
  87 type ThinDockerClientProxy struct {
  88         Docker *dockerclient.Client
  89 }
  90
  91 // ContainerAttach invokes dockerclient.Client.ContainerAttach
  92 func (proxy ThinDockerClientProxy) ContainerAttach(ctx context.Context, container string, options dockertypes.ContainerAttachOptions) (dockertypes.HijackedResponse, error) {
  93         return proxy.Docker.ContainerAttach(ctx, container, options)
  94 }
  95
  96 // ContainerCreate invokes dockerclient.Client.ContainerCreate
  97 func (proxy ThinDockerClientProxy) ContainerCreate(ctx context.Context, config *dockercontainer.Config, hostConfig *dockercontainer.HostConfig,
  98         networkingConfig *dockernetwork.NetworkingConfig, containerName string) (dockercontainer.ContainerCreateCreatedBody, error) {
  99         return proxy.Docker.ContainerCreate(ctx, config, hostConfig, networkingConfig, containerName)
 100 }
 101
 102 // ContainerStart invokes dockerclient.Client.ContainerStart
 103 func (proxy ThinDockerClientProxy) ContainerStart(ctx context.Context, container string, options dockertypes.ContainerStartOptions) error {
 104         return proxy.Docker.ContainerStart(ctx, container, options)
 105 }
 106
 107 // ContainerStop invokes dockerclient.Client.ContainerStop
 108 func (proxy ThinDockerClientProxy) ContainerStop(ctx context.Context, container string, timeout *time.Duration) error {
 109         return proxy.Docker.ContainerStop(ctx, container, timeout)
 110 }
 111
 112 // ContainerWait invokes dockerclient.Client.ContainerWait
 113 func (proxy ThinDockerClientProxy) ContainerWait(ctx context.Context, container string, condition dockercontainer.WaitCondition) (<-chan dockercontainer.ContainerWaitOKBody, <-chan error) {
 114         return proxy.Docker.ContainerWait(ctx, container, condition)
 115 }
 116
 117 // ImageInspectWithRaw invokes dockerclient.Client.ImageInspectWithRaw
 118 func (proxy ThinDockerClientProxy) ImageInspectWithRaw(ctx context.Context, image string) (dockertypes.ImageInspect, []byte, error) {
 119         return proxy.Docker.ImageInspectWithRaw(ctx, image)
 120 }
 121
 122 // ImageLoad invokes dockerclient.Client.ImageLoad
 123 func (proxy ThinDockerClientProxy) ImageLoad(ctx context.Context, input io.Reader, quiet bool) (dockertypes.ImageLoadResponse, error) {
 124         return proxy.Docker.ImageLoad(ctx, input, quiet)
 125 }
 126
 127 // ImageRemove invokes dockerclient.Client.ImageRemove
 128 func (proxy ThinDockerClientProxy) ImageRemove(ctx context.Context, image string, options dockertypes.ImageRemoveOptions) ([]dockertypes.ImageDeleteResponseItem, error) {
 129         return proxy.Docker.ImageRemove(ctx, image, options)
 130 }
 131
 132 // ContainerRunner is the main stateful struct used for a single execution of a
 133 // container.
 134 type ContainerRunner struct {
 135         Docker    ThinDockerClient
 136         ArvClient IArvadosClient
 137         Kc        IKeepClient
 138         arvados.Container
 139         ContainerConfig dockercontainer.Config
 140         dockercontainer.HostConfig
 141         token       string
 142         ContainerID string
 143         ExitCode    *int
 144         NewLogWriter
 145         loggingDone   chan bool
 146         CrunchLog     *ThrottledLogger
 147         Stdout        io.WriteCloser
 148         Stderr        io.WriteCloser
 149         LogCollection *CollectionWriter
 150         LogsPDH       *string
 151         RunArvMount
 152         MkTempDir
 153         ArvMount       *exec.Cmd
 154         ArvMountPoint  string
 155         HostOutputDir  string
 156         CleanupTempDir []string
 157         Binds          []string
 158         Volumes        map[string]struct{}
 159         OutputPDH      *string
 160         SigChan        chan os.Signal
 161         ArvMountExit   chan error
 162         finalState     string
 163
 164         statLogger   io.WriteCloser
 165         statReporter *crunchstat.Reporter
 166         statInterval time.Duration
 167         cgroupRoot   string
 168         // What we expect the container's cgroup parent to be.
 169         expectCgroupParent string
 170         // What we tell docker to use as the container's cgroup
 171         // parent. Note: Ideally we would use the same field for both
 172         // expectCgroupParent and setCgroupParent, and just make it
 173         // default to "docker". However, when using docker < 1.10 with
 174         // systemd, specifying a non-empty cgroup parent (even the
 175         // default value "docker") hits a docker bug
 176         // (https://github.com/docker/docker/issues/17126). Using two
 177         // separate fields makes it possible to use the "expect cgroup
 178         // parent to be X" feature even on sites where the "specify
 179         // cgroup parent" feature breaks.
 180         setCgroupParent string
 181
 182         cStateLock sync.Mutex
 183         cStarted   bool // StartContainer() succeeded
 184         cCancelled bool // StopContainer() invoked
 185
 186         enableNetwork string // one of "default" or "always"
 187         networkMode   string // passed through to HostConfig.NetworkMode
 188         arvMountLog   *ThrottledLogger
 189 }
 190
 191 // setupSignals sets up signal handling to gracefully terminate the underlying
 192 // Docker container and update state when receiving a TERM, INT or QUIT signal.
 193 func (runner *ContainerRunner) setupSignals() {
 194         runner.SigChan = make(chan os.Signal, 1)
 195         signal.Notify(runner.SigChan, syscall.SIGTERM)
 196         signal.Notify(runner.SigChan, syscall.SIGINT)
 197         signal.Notify(runner.SigChan, syscall.SIGQUIT)
 198
 199         go func(sig chan os.Signal) {
 200                 s := <-sig
 201                 if s != nil {
 202                         runner.CrunchLog.Printf("Caught signal %v", s)
 203                 }
 204                 runner.stop()
 205         }(runner.SigChan)
 206 }
 207
 208 // stop the underlying Docker container.
 209 func (runner *ContainerRunner) stop() {
 210         runner.cStateLock.Lock()
 211         defer runner.cStateLock.Unlock()
 212         if runner.cCancelled {
 213                 return
 214         }
 215         runner.cCancelled = true
 216         if runner.cStarted {
 217                 timeout := time.Duration(10)
 218                 err := runner.Docker.ContainerStop(context.TODO(), runner.ContainerID, &(timeout))
 219                 if err != nil {
 220                         runner.CrunchLog.Printf("StopContainer failed: %s", err)
 221                 }
 222                 // Suppress multiple calls to stop()
 223                 runner.cStarted = false
 224         }
 225 }
 226
 227 func (runner *ContainerRunner) stopSignals() {
 228         if runner.SigChan != nil {
 229                 signal.Stop(runner.SigChan)
 230                 close(runner.SigChan)
 231         }
 232 }
 233
 234 var errorBlacklist = []string{
 235         "(?ms).*[Cc]annot connect to the Docker daemon.*",
 236         "(?ms).*oci runtime error.*starting container process.*container init.*mounting.*to rootfs.*no such file or directory.*",
 237 }
 238 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)")
 239
 240 func (runner *ContainerRunner) checkBrokenNode(goterr error) bool {
 241         for _, d := range errorBlacklist {
 242                 if m, e := regexp.MatchString(d, goterr.Error()); m && e == nil {
 243                         runner.CrunchLog.Printf("Error suggests node is unable to run containers: %v", goterr)
 244                         if *brokenNodeHook == "" {
 245                                 runner.CrunchLog.Printf("No broken node hook provided, cannot mark node as broken.")
 246                         } else {
 247                                 runner.CrunchLog.Printf("Running broken node hook %q", *brokenNodeHook)
 248                                 // run killme script
 249                                 c := exec.Command(*brokenNodeHook)
 250                                 c.Stdout = runner.CrunchLog
 251                                 c.Stderr = runner.CrunchLog
 252                                 err := c.Run()
 253                                 if err != nil {
 254                                         runner.CrunchLog.Printf("Error running broken node hook: %v", err)
 255                                 }
 256                         }
 257                         return true
 258                 }
 259         }
 260         return false
 261 }
 262
 263 // LoadImage determines the docker image id from the container record and
 264 // checks if it is available in the local Docker image store.  If not, it loads
 265 // the image from Keep.
 266 func (runner *ContainerRunner) LoadImage() (err error) {
 267
 268         runner.CrunchLog.Printf("Fetching Docker image from collection '%s'", runner.Container.ContainerImage)
 269
 270         var collection arvados.Collection
 271         err = runner.ArvClient.Get("collections", runner.Container.ContainerImage, nil, &collection)
 272         if err != nil {
 273                 return fmt.Errorf("While getting container image collection: %v", err)
 274         }
 275         manifest := manifest.Manifest{Text: collection.ManifestText}
 276         var img, imageID string
 277         for ms := range manifest.StreamIter() {
 278                 img = ms.FileStreamSegments[0].Name
 279                 if !strings.HasSuffix(img, ".tar") {
 280                         return fmt.Errorf("First file in the container image collection does not end in .tar")
 281                 }
 282                 imageID = img[:len(img)-4]
 283         }
 284
 285         runner.CrunchLog.Printf("Using Docker image id '%s'", imageID)
 286
 287         _, _, err = runner.Docker.ImageInspectWithRaw(context.TODO(), imageID)
 288         if err != nil {
 289                 runner.CrunchLog.Print("Loading Docker image from keep")
 290
 291                 var readCloser io.ReadCloser
 292                 readCloser, err = runner.Kc.ManifestFileReader(manifest, img)
 293                 if err != nil {
 294                         return fmt.Errorf("While creating ManifestFileReader for container image: %v", err)
 295                 }
 296
 297                 response, err := runner.Docker.ImageLoad(context.TODO(), readCloser, true)
 298                 if err != nil {
 299                         return fmt.Errorf("While loading container image into Docker: %v", err)
 300                 }
 301
 302                 defer response.Body.Close()
 303                 rbody, err := ioutil.ReadAll(response.Body)
 304                 if err != nil {
 305                         return fmt.Errorf("Reading response to image load: %v", err)
 306                 }
 307                 runner.CrunchLog.Printf("Docker response: %s", rbody)
 308         } else {
 309                 runner.CrunchLog.Print("Docker image is available")
 310         }
 311
 312         runner.ContainerConfig.Image = imageID
 313
 314         runner.Kc.ClearBlockCache()
 315
 316         return nil
 317 }
 318
 319 func (runner *ContainerRunner) ArvMountCmd(arvMountCmd []string, token string) (c *exec.Cmd, err error) {
 320         c = exec.Command("arv-mount", arvMountCmd...)
 321
 322         // Copy our environment, but override ARVADOS_API_TOKEN with
 323         // the container auth token.
 324         c.Env = nil
 325         for _, s := range os.Environ() {
 326                 if !strings.HasPrefix(s, "ARVADOS_API_TOKEN=") {
 327                         c.Env = append(c.Env, s)
 328                 }
 329         }
 330         c.Env = append(c.Env, "ARVADOS_API_TOKEN="+token)
 331
 332         runner.arvMountLog = NewThrottledLogger(runner.NewLogWriter("arv-mount"))
 333         c.Stdout = runner.arvMountLog
 334         c.Stderr = runner.arvMountLog
 335
 336         runner.CrunchLog.Printf("Running %v", c.Args)
 337
 338         err = c.Start()
 339         if err != nil {
 340                 return nil, err
 341         }
 342
 343         statReadme := make(chan bool)
 344         runner.ArvMountExit = make(chan error)
 345
 346         keepStatting := true
 347         go func() {
 348                 for keepStatting {
 349                         time.Sleep(100 * time.Millisecond)
 350                         _, err = os.Stat(fmt.Sprintf("%s/by_id/README", runner.ArvMountPoint))
 351                         if err == nil {
 352                                 keepStatting = false
 353                                 statReadme <- true
 354                         }
 355                 }
 356                 close(statReadme)
 357         }()
 358
 359         go func() {
 360                 mnterr := c.Wait()
 361                 if mnterr != nil {
 362                         runner.CrunchLog.Printf("Arv-mount exit error: %v", mnterr)
 363                 }
 364                 runner.ArvMountExit <- mnterr
 365                 close(runner.ArvMountExit)
 366         }()
 367
 368         select {
 369         case <-statReadme:
 370                 break
 371         case err := <-runner.ArvMountExit:
 372                 runner.ArvMount = nil
 373                 keepStatting = false
 374                 return nil, err
 375         }
 376
 377         return c, nil
 378 }
 379
 380 func (runner *ContainerRunner) SetupArvMountPoint(prefix string) (err error) {
 381         if runner.ArvMountPoint == "" {
 382                 runner.ArvMountPoint, err = runner.MkTempDir("", prefix)
 383         }
 384         return
 385 }
 386
 387 func (runner *ContainerRunner) SetupMounts() (err error) {
 388         err = runner.SetupArvMountPoint("keep")
 389         if err != nil {
 390                 return fmt.Errorf("While creating keep mount temp dir: %v", err)
 391         }
 392
 393         pdhOnly := true
 394         tmpcount := 0
 395         arvMountCmd := []string{
 396                 "--foreground",
 397                 "--allow-other",
 398                 "--read-write",
 399                 fmt.Sprintf("--crunchstat-interval=%v", runner.statInterval.Seconds())}
 400
 401         if runner.Container.RuntimeConstraints.KeepCacheRAM > 0 {
 402                 arvMountCmd = append(arvMountCmd, "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheRAM))
 403         }
 404
 405         collectionPaths := []string{}
 406         runner.Binds = nil
 407         runner.Volumes = make(map[string]struct{})
 408         needCertMount := true
 409
 410         var binds []string
 411         for bind := range runner.Container.Mounts {
 412                 binds = append(binds, bind)
 413         }
 414         sort.Strings(binds)
 415
 416         for _, bind := range binds {
 417                 mnt := runner.Container.Mounts[bind]
 418                 if bind == "stdout" || bind == "stderr" {
 419                         // Is it a "file" mount kind?
 420                         if mnt.Kind != "file" {
 421                                 return fmt.Errorf("Unsupported mount kind '%s' for %s. Only 'file' is supported.", mnt.Kind, bind)
 422                         }
 423
 424                         // Does path start with OutputPath?
 425                         prefix := runner.Container.OutputPath
 426                         if !strings.HasSuffix(prefix, "/") {
 427                                 prefix += "/"
 428                         }
 429                         if !strings.HasPrefix(mnt.Path, prefix) {
 430                                 return fmt.Errorf("%s path does not start with OutputPath: %s, %s", strings.Title(bind), mnt.Path, prefix)
 431                         }
 432                 }
 433
 434                 if bind == "stdin" {
 435                         // Is it a "collection" mount kind?
 436                         if mnt.Kind != "collection" && mnt.Kind != "json" {
 437                                 return fmt.Errorf("Unsupported mount kind '%s' for stdin. Only 'collection' or 'json' are supported.", mnt.Kind)
 438                         }
 439                 }
 440
 441                 if bind == "/etc/arvados/ca-certificates.crt" {
 442                         needCertMount = false
 443                 }
 444
 445                 if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && bind != runner.Container.OutputPath+"/" {
 446                         if mnt.Kind != "collection" {
 447                                 return fmt.Errorf("Only mount points of kind 'collection' are supported underneath the output_path: %v", bind)
 448                         }
 449                 }
 450
 451                 switch {
 452                 case mnt.Kind == "collection" && bind != "stdin":
 453                         var src string
 454                         if mnt.UUID != "" && mnt.PortableDataHash != "" {
 455                                 return fmt.Errorf("Cannot specify both 'uuid' and 'portable_data_hash' for a collection mount")
 456                         }
 457                         if mnt.UUID != "" {
 458                                 if mnt.Writable {
 459                                         return fmt.Errorf("Writing to existing collections currently not permitted.")
 460                                 }
 461                                 pdhOnly = false
 462                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.UUID)
 463                         } else if mnt.PortableDataHash != "" {
 464                                 if mnt.Writable {
 465                                         return fmt.Errorf("Can never write to a collection specified by portable data hash")
 466                                 }
 467                                 idx := strings.Index(mnt.PortableDataHash, "/")
 468                                 if idx > 0 {
 469                                         mnt.Path = path.Clean(mnt.PortableDataHash[idx:])
 470                                         mnt.PortableDataHash = mnt.PortableDataHash[0:idx]
 471                                         runner.Container.Mounts[bind] = mnt
 472                                 }
 473                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.PortableDataHash)
 474                                 if mnt.Path != "" && mnt.Path != "." {
 475                                         if strings.HasPrefix(mnt.Path, "./") {
 476                                                 mnt.Path = mnt.Path[2:]
 477                                         } else if strings.HasPrefix(mnt.Path, "/") {
 478                                                 mnt.Path = mnt.Path[1:]
 479                                         }
 480                                         src += "/" + mnt.Path
 481                                 }
 482                         } else {
 483                                 src = fmt.Sprintf("%s/tmp%d", runner.ArvMountPoint, tmpcount)
 484                                 arvMountCmd = append(arvMountCmd, "--mount-tmp")
 485                                 arvMountCmd = append(arvMountCmd, fmt.Sprintf("tmp%d", tmpcount))
 486                                 tmpcount += 1
 487                         }
 488                         if mnt.Writable {
 489                                 if bind == runner.Container.OutputPath {
 490                                         runner.HostOutputDir = src
 491                                 } else if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 492                                         return fmt.Errorf("Writable mount points are not permitted underneath the output_path: %v", bind)
 493                                 }
 494                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", src, bind))
 495                         } else {
 496                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", src, bind))
 497                         }
 498                         collectionPaths = append(collectionPaths, src)
 499
 500                 case mnt.Kind == "tmp":
 501                         var tmpdir string
 502                         tmpdir, err = runner.MkTempDir("", "")
 503                         if err != nil {
 504                                 return fmt.Errorf("While creating mount temp dir: %v", err)
 505                         }
 506                         st, staterr := os.Stat(tmpdir)
 507                         if staterr != nil {
 508                                 return fmt.Errorf("While Stat on temp dir: %v", staterr)
 509                         }
 510                         err = os.Chmod(tmpdir, st.Mode()|os.ModeSetgid|0777)
 511                         if staterr != nil {
 512                                 return fmt.Errorf("While Chmod temp dir: %v", err)
 513                         }
 514                         runner.CleanupTempDir = append(runner.CleanupTempDir, tmpdir)
 515                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", tmpdir, bind))
 516                         if bind == runner.Container.OutputPath {
 517                                 runner.HostOutputDir = tmpdir
 518                         }
 519
 520                 case mnt.Kind == "json":
 521                         jsondata, err := json.Marshal(mnt.Content)
 522                         if err != nil {
 523                                 return fmt.Errorf("encoding json data: %v", err)
 524                         }
 525                         // Create a tempdir with a single file
 526                         // (instead of just a tempfile): this way we
 527                         // can ensure the file is world-readable
 528                         // inside the container, without having to
 529                         // make it world-readable on the docker host.
 530                         tmpdir, err := runner.MkTempDir("", "")
 531                         if err != nil {
 532                                 return fmt.Errorf("creating temp dir: %v", err)
 533                         }
 534                         runner.CleanupTempDir = append(runner.CleanupTempDir, tmpdir)
 535                         tmpfn := filepath.Join(tmpdir, "mountdata.json")
 536                         err = ioutil.WriteFile(tmpfn, jsondata, 0644)
 537                         if err != nil {
 538                                 return fmt.Errorf("writing temp file: %v", err)
 539                         }
 540                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", tmpfn, bind))
 541                 }
 542         }
 543
 544         if runner.HostOutputDir == "" {
 545                 return fmt.Errorf("Output path does not correspond to a writable mount point")
 546         }
 547
 548         if wantAPI := runner.Container.RuntimeConstraints.API; needCertMount && wantAPI != nil && *wantAPI {
 549                 for _, certfile := range arvadosclient.CertFiles {
 550                         _, err := os.Stat(certfile)
 551                         if err == nil {
 552                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:/etc/arvados/ca-certificates.crt:ro", certfile))
 553                                 break
 554                         }
 555                 }
 556         }
 557
 558         if pdhOnly {
 559                 arvMountCmd = append(arvMountCmd, "--mount-by-pdh", "by_id")
 560         } else {
 561                 arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_id")
 562         }
 563         arvMountCmd = append(arvMountCmd, runner.ArvMountPoint)
 564
 565         token, err := runner.ContainerToken()
 566         if err != nil {
 567                 return fmt.Errorf("could not get container token: %s", err)
 568         }
 569
 570         runner.ArvMount, err = runner.RunArvMount(arvMountCmd, token)
 571         if err != nil {
 572                 return fmt.Errorf("While trying to start arv-mount: %v", err)
 573         }
 574
 575         for _, p := range collectionPaths {
 576                 _, err = os.Stat(p)
 577                 if err != nil {
 578                         return fmt.Errorf("While checking that input files exist: %v", err)
 579                 }
 580         }
 581
 582         return nil
 583 }
 584
 585 func (runner *ContainerRunner) ProcessDockerAttach(containerReader io.Reader) {
 586         // Handle docker log protocol
 587         // https://docs.docker.com/engine/reference/api/docker_remote_api_v1.15/#attach-to-a-container
 588
 589         header := make([]byte, 8)
 590         for {
 591                 _, readerr := io.ReadAtLeast(containerReader, header, 8)
 592
 593                 if readerr == nil {
 594                         readsize := int64(header[7]) | (int64(header[6]) << 8) | (int64(header[5]) << 16) | (int64(header[4]) << 24)
 595                         if header[0] == 1 {
 596                                 // stdout
 597                                 _, readerr = io.CopyN(runner.Stdout, containerReader, readsize)
 598                         } else {
 599                                 // stderr
 600                                 _, readerr = io.CopyN(runner.Stderr, containerReader, readsize)
 601                         }
 602                 }
 603
 604                 if readerr != nil {
 605                         if readerr != io.EOF {
 606                                 runner.CrunchLog.Printf("While reading docker logs: %v", readerr)
 607                         }
 608
 609                         closeerr := runner.Stdout.Close()
 610                         if closeerr != nil {
 611                                 runner.CrunchLog.Printf("While closing stdout logs: %v", closeerr)
 612                         }
 613
 614                         closeerr = runner.Stderr.Close()
 615                         if closeerr != nil {
 616                                 runner.CrunchLog.Printf("While closing stderr logs: %v", closeerr)
 617                         }
 618
 619                         if runner.statReporter != nil {
 620                                 runner.statReporter.Stop()
 621                                 closeerr = runner.statLogger.Close()
 622                                 if closeerr != nil {
 623                                         runner.CrunchLog.Printf("While closing crunchstat logs: %v", closeerr)
 624                                 }
 625                         }
 626
 627                         runner.loggingDone <- true
 628                         close(runner.loggingDone)
 629                         return
 630                 }
 631         }
 632 }
 633
 634 func (runner *ContainerRunner) StartCrunchstat() {
 635         runner.statLogger = NewThrottledLogger(runner.NewLogWriter("crunchstat"))
 636         runner.statReporter = &crunchstat.Reporter{
 637                 CID:          runner.ContainerID,
 638                 Logger:       log.New(runner.statLogger, "", 0),
 639                 CgroupParent: runner.expectCgroupParent,
 640                 CgroupRoot:   runner.cgroupRoot,
 641                 PollPeriod:   runner.statInterval,
 642         }
 643         runner.statReporter.Start()
 644 }
 645
 646 type infoCommand struct {
 647         label string
 648         cmd   []string
 649 }
 650
 651 // LogNodeInfo gathers node information and store it on the log for debugging
 652 // purposes.
 653 func (runner *ContainerRunner) LogNodeInfo() (err error) {
 654         w := runner.NewLogWriter("node-info")
 655         logger := log.New(w, "node-info", 0)
 656
 657         commands := []infoCommand{
 658                 {
 659                         label: "Host Information",
 660                         cmd:   []string{"uname", "-a"},
 661                 },
 662                 {
 663                         label: "CPU Information",
 664                         cmd:   []string{"cat", "/proc/cpuinfo"},
 665                 },
 666                 {
 667                         label: "Memory Information",
 668                         cmd:   []string{"cat", "/proc/meminfo"},
 669                 },
 670                 {
 671                         label: "Disk Space",
 672                         cmd:   []string{"df", "-m", "/", os.TempDir()},
 673                 },
 674                 {
 675                         label: "Disk INodes",
 676                         cmd:   []string{"df", "-i", "/", os.TempDir()},
 677                 },
 678         }
 679
 680         // Run commands with informational output to be logged.
 681         var out []byte
 682         for _, command := range commands {
 683                 out, err = exec.Command(command.cmd[0], command.cmd[1:]...).CombinedOutput()
 684                 if err != nil {
 685                         return fmt.Errorf("While running command %q: %v",
 686                                 command.cmd, err)
 687                 }
 688                 logger.Println(command.label)
 689                 for _, line := range strings.Split(string(out), "\n") {
 690                         logger.Println(" ", line)
 691                 }
 692         }
 693
 694         err = w.Close()
 695         if err != nil {
 696                 return fmt.Errorf("While closing node-info logs: %v", err)
 697         }
 698         return nil
 699 }
 700
 701 // LogContainerRecord gets and saves the raw JSON container record from the API server
 702 func (runner *ContainerRunner) LogContainerRecord() (err error) {
 703         w := &ArvLogWriter{
 704                 ArvClient:     runner.ArvClient,
 705                 UUID:          runner.Container.UUID,
 706                 loggingStream: "container",
 707                 writeCloser:   runner.LogCollection.Open("container.json"),
 708         }
 709
 710         // Get Container record JSON from the API Server
 711         reader, err := runner.ArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
 712         if err != nil {
 713                 return fmt.Errorf("While retrieving container record from the API server: %v", err)
 714         }
 715         defer reader.Close()
 716
 717         dec := json.NewDecoder(reader)
 718         dec.UseNumber()
 719         var cr map[string]interface{}
 720         if err = dec.Decode(&cr); err != nil {
 721                 return fmt.Errorf("While decoding the container record JSON response: %v", err)
 722         }
 723         // Re-encode it using indentation to improve readability
 724         enc := json.NewEncoder(w)
 725         enc.SetIndent("", "    ")
 726         if err = enc.Encode(cr); err != nil {
 727                 return fmt.Errorf("While logging the JSON container record: %v", err)
 728         }
 729         err = w.Close()
 730         if err != nil {
 731                 return fmt.Errorf("While closing container.json log: %v", err)
 732         }
 733         return nil
 734 }
 735
 736 // AttachStreams connects the docker container stdin, stdout and stderr logs
 737 // to the Arvados logger which logs to Keep and the API server logs table.
 738 func (runner *ContainerRunner) AttachStreams() (err error) {
 739
 740         runner.CrunchLog.Print("Attaching container streams")
 741
 742         // If stdin mount is provided, attach it to the docker container
 743         var stdinRdr arvados.File
 744         var stdinJson []byte
 745         if stdinMnt, ok := runner.Container.Mounts["stdin"]; ok {
 746                 if stdinMnt.Kind == "collection" {
 747                         var stdinColl arvados.Collection
 748                         collId := stdinMnt.UUID
 749                         if collId == "" {
 750                                 collId = stdinMnt.PortableDataHash
 751                         }
 752                         err = runner.ArvClient.Get("collections", collId, nil, &stdinColl)
 753                         if err != nil {
 754                                 return fmt.Errorf("While getting stding collection: %v", err)
 755                         }
 756
 757                         stdinRdr, err = runner.Kc.ManifestFileReader(manifest.Manifest{Text: stdinColl.ManifestText}, stdinMnt.Path)
 758                         if os.IsNotExist(err) {
 759                                 return fmt.Errorf("stdin collection path not found: %v", stdinMnt.Path)
 760                         } else if err != nil {
 761                                 return fmt.Errorf("While getting stdin collection path %v: %v", stdinMnt.Path, err)
 762                         }
 763                 } else if stdinMnt.Kind == "json" {
 764                         stdinJson, err = json.Marshal(stdinMnt.Content)
 765                         if err != nil {
 766                                 return fmt.Errorf("While encoding stdin json data: %v", err)
 767                         }
 768                 }
 769         }
 770
 771         stdinUsed := stdinRdr != nil || len(stdinJson) != 0
 772         response, err := runner.Docker.ContainerAttach(context.TODO(), runner.ContainerID,
 773                 dockertypes.ContainerAttachOptions{Stream: true, Stdin: stdinUsed, Stdout: true, Stderr: true})
 774         if err != nil {
 775                 return fmt.Errorf("While attaching container stdout/stderr streams: %v", err)
 776         }
 777
 778         runner.loggingDone = make(chan bool)
 779
 780         if stdoutMnt, ok := runner.Container.Mounts["stdout"]; ok {
 781                 stdoutFile, err := runner.getStdoutFile(stdoutMnt.Path)
 782                 if err != nil {
 783                         return err
 784                 }
 785                 runner.Stdout = stdoutFile
 786         } else {
 787                 runner.Stdout = NewThrottledLogger(runner.NewLogWriter("stdout"))
 788         }
 789
 790         if stderrMnt, ok := runner.Container.Mounts["stderr"]; ok {
 791                 stderrFile, err := runner.getStdoutFile(stderrMnt.Path)
 792                 if err != nil {
 793                         return err
 794                 }
 795                 runner.Stderr = stderrFile
 796         } else {
 797                 runner.Stderr = NewThrottledLogger(runner.NewLogWriter("stderr"))
 798         }
 799
 800         if stdinRdr != nil {
 801                 go func() {
 802                         _, err := io.Copy(response.Conn, stdinRdr)
 803                         if err != nil {
 804                                 runner.CrunchLog.Print("While writing stdin collection to docker container %q", err)
 805                                 runner.stop()
 806                         }
 807                         stdinRdr.Close()
 808                         response.CloseWrite()
 809                 }()
 810         } else if len(stdinJson) != 0 {
 811                 go func() {
 812                         _, err := io.Copy(response.Conn, bytes.NewReader(stdinJson))
 813                         if err != nil {
 814                                 runner.CrunchLog.Print("While writing stdin json to docker container %q", err)
 815                                 runner.stop()
 816                         }
 817                         response.CloseWrite()
 818                 }()
 819         }
 820
 821         go runner.ProcessDockerAttach(response.Reader)
 822
 823         return nil
 824 }
 825
 826 func (runner *ContainerRunner) getStdoutFile(mntPath string) (*os.File, error) {
 827         stdoutPath := mntPath[len(runner.Container.OutputPath):]
 828         index := strings.LastIndex(stdoutPath, "/")
 829         if index > 0 {
 830                 subdirs := stdoutPath[:index]
 831                 if subdirs != "" {
 832                         st, err := os.Stat(runner.HostOutputDir)
 833                         if err != nil {
 834                                 return nil, fmt.Errorf("While Stat on temp dir: %v", err)
 835                         }
 836                         stdoutPath := filepath.Join(runner.HostOutputDir, subdirs)
 837                         err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777)
 838                         if err != nil {
 839                                 return nil, fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err)
 840                         }
 841                 }
 842         }
 843         stdoutFile, err := os.Create(filepath.Join(runner.HostOutputDir, stdoutPath))
 844         if err != nil {
 845                 return nil, fmt.Errorf("While creating file %q: %v", stdoutPath, err)
 846         }
 847
 848         return stdoutFile, nil
 849 }
 850
 851 // CreateContainer creates the docker container.
 852 func (runner *ContainerRunner) CreateContainer() error {
 853         runner.CrunchLog.Print("Creating Docker container")
 854
 855         runner.ContainerConfig.Cmd = runner.Container.Command
 856         if runner.Container.Cwd != "." {
 857                 runner.ContainerConfig.WorkingDir = runner.Container.Cwd
 858         }
 859
 860         for k, v := range runner.Container.Environment {
 861                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env, k+"="+v)
 862         }
 863
 864         runner.ContainerConfig.Volumes = runner.Volumes
 865
 866         runner.HostConfig = dockercontainer.HostConfig{
 867                 Binds: runner.Binds,
 868                 LogConfig: dockercontainer.LogConfig{
 869                         Type: "none",
 870                 },
 871                 Resources: dockercontainer.Resources{
 872                         CgroupParent: runner.setCgroupParent,
 873                 },
 874         }
 875
 876         if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
 877                 tok, err := runner.ContainerToken()
 878                 if err != nil {
 879                         return err
 880                 }
 881                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env,
 882                         "ARVADOS_API_TOKEN="+tok,
 883                         "ARVADOS_API_HOST="+os.Getenv("ARVADOS_API_HOST"),
 884                         "ARVADOS_API_HOST_INSECURE="+os.Getenv("ARVADOS_API_HOST_INSECURE"),
 885                 )
 886                 runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
 887         } else {
 888                 if runner.enableNetwork == "always" {
 889                         runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
 890                 } else {
 891                         runner.HostConfig.NetworkMode = dockercontainer.NetworkMode("none")
 892                 }
 893         }
 894
 895         _, stdinUsed := runner.Container.Mounts["stdin"]
 896         runner.ContainerConfig.OpenStdin = stdinUsed
 897         runner.ContainerConfig.StdinOnce = stdinUsed
 898         runner.ContainerConfig.AttachStdin = stdinUsed
 899         runner.ContainerConfig.AttachStdout = true
 900         runner.ContainerConfig.AttachStderr = true
 901
 902         createdBody, err := runner.Docker.ContainerCreate(context.TODO(), &runner.ContainerConfig, &runner.HostConfig, nil, runner.Container.UUID)
 903         if err != nil {
 904                 return fmt.Errorf("While creating container: %v", err)
 905         }
 906
 907         runner.ContainerID = createdBody.ID
 908
 909         return runner.AttachStreams()
 910 }
 911
 912 // StartContainer starts the docker container created by CreateContainer.
 913 func (runner *ContainerRunner) StartContainer() error {
 914         runner.CrunchLog.Printf("Starting Docker container id '%s'", runner.ContainerID)
 915         runner.cStateLock.Lock()
 916         defer runner.cStateLock.Unlock()
 917         if runner.cCancelled {
 918                 return ErrCancelled
 919         }
 920         err := runner.Docker.ContainerStart(context.TODO(), runner.ContainerID,
 921                 dockertypes.ContainerStartOptions{})
 922         if err != nil {
 923                 var advice string
 924                 if m, e := regexp.MatchString("(?ms).*(exec|System error).*(no such file or directory|file not found).*", err.Error()); m && e == nil {
 925                         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])
 926                 }
 927                 return fmt.Errorf("could not start container: %v%s", err, advice)
 928         }
 929         runner.cStarted = true
 930         return nil
 931 }
 932
 933 // WaitFinish waits for the container to terminate, capture the exit code, and
 934 // close the stdout/stderr logging.
 935 func (runner *ContainerRunner) WaitFinish() (err error) {
 936         runner.CrunchLog.Print("Waiting for container to finish")
 937
 938         waitOk, waitErr := runner.Docker.ContainerWait(context.TODO(), runner.ContainerID, "not-running")
 939
 940         go func() {
 941                 <-runner.ArvMountExit
 942                 if runner.cStarted {
 943                         runner.CrunchLog.Printf("arv-mount exited while container is still running.  Stopping container.")
 944                         runner.stop()
 945                 }
 946         }()
 947
 948         var waitBody dockercontainer.ContainerWaitOKBody
 949         select {
 950         case waitBody = <-waitOk:
 951         case err = <-waitErr:
 952         }
 953
 954         // Container isn't running any more
 955         runner.cStarted = false
 956
 957         if err != nil {
 958                 return fmt.Errorf("container wait: %v", err)
 959         }
 960
 961         runner.CrunchLog.Printf("Container exited with code: %v", waitBody.StatusCode)
 962         code := int(waitBody.StatusCode)
 963         runner.ExitCode = &code
 964
 965         // wait for stdout/stderr to complete
 966         <-runner.loggingDone
 967
 968         return nil
 969 }
 970
 971 var ErrNotInOutputDir = fmt.Errorf("Must point to path within the output directory")
 972
 973 func (runner *ContainerRunner) derefOutputSymlink(path string, startinfo os.FileInfo) (tgt string, readlinktgt string, info os.FileInfo, err error) {
 974         // Follow symlinks if necessary
 975         info = startinfo
 976         tgt = path
 977         readlinktgt = ""
 978         nextlink := path
 979         for followed := 0; info.Mode()&os.ModeSymlink != 0; followed++ {
 980                 if followed >= limitFollowSymlinks {
 981                         // Got stuck in a loop or just a pathological number of links, give up.
 982                         err = fmt.Errorf("Followed more than %v symlinks from path %q", limitFollowSymlinks, path)
 983                         return
 984                 }
 985
 986                 readlinktgt, err = os.Readlink(nextlink)
 987                 if err != nil {
 988                         return
 989                 }
 990
 991                 tgt = readlinktgt
 992                 if !strings.HasPrefix(tgt, "/") {
 993                         // Relative symlink, resolve it to host path
 994                         tgt = filepath.Join(filepath.Dir(path), tgt)
 995                 }
 996                 if strings.HasPrefix(tgt, runner.Container.OutputPath+"/") && !strings.HasPrefix(tgt, runner.HostOutputDir+"/") {
 997                         // Absolute symlink to container output path, adjust it to host output path.
 998                         tgt = filepath.Join(runner.HostOutputDir, tgt[len(runner.Container.OutputPath):])
 999                 }
1000                 if !strings.HasPrefix(tgt, runner.HostOutputDir+"/") {
1001                         // After dereferencing, symlink target must either be
1002                         // within output directory, or must point to a
1003                         // collection mount.
1004                         err = ErrNotInOutputDir
1005                         return
1006                 }
1007
1008                 info, err = os.Lstat(tgt)
1009                 if err != nil {
1010                         // tgt
1011                         err = fmt.Errorf("Symlink in output %q points to invalid location %q: %v",
1012                                 path[len(runner.HostOutputDir):], readlinktgt, err)
1013                         return
1014                 }
1015
1016                 nextlink = tgt
1017         }
1018
1019         return
1020 }
1021
1022 var limitFollowSymlinks = 10
1023
1024 // UploadFile uploads files within the output directory, with special handling
1025 // for symlinks. If the symlink leads to a keep mount, copy the manifest text
1026 // from the keep mount into the output manifestText.  Ensure that whether
1027 // symlinks are relative or absolute, every symlink target (even targets that
1028 // are symlinks themselves) must point to a path in either the output directory
1029 // or a collection mount.
1030 //
1031 // Assumes initial value of "path" is absolute, and located within runner.HostOutputDir.
1032 func (runner *ContainerRunner) UploadOutputFile(
1033         path string,
1034         info os.FileInfo,
1035         infoerr error,
1036         binds []string,
1037         walkUpload *WalkUpload,
1038         relocateFrom string,
1039         relocateTo string,
1040         followed int) (manifestText string, err error) {
1041
1042         if info.Mode().IsDir() {
1043                 return
1044         }
1045
1046         if infoerr != nil {
1047                 return "", infoerr
1048         }
1049
1050         if followed >= limitFollowSymlinks {
1051                 // Got stuck in a loop or just a pathological number of
1052                 // directory links, give up.
1053                 err = fmt.Errorf("Followed more than %v symlinks from path %q", limitFollowSymlinks, path)
1054                 return
1055         }
1056
1057         // When following symlinks, the source path may need to be logically
1058         // relocated to some other path within the output collection.  Remove
1059         // the relocateFrom prefix and replace it with relocateTo.
1060         relocated := relocateTo + path[len(relocateFrom):]
1061
1062         tgt, readlinktgt, info, derefErr := runner.derefOutputSymlink(path, info)
1063         if derefErr != nil && derefErr != ErrNotInOutputDir {
1064                 return "", derefErr
1065         }
1066
1067         // go through mounts and try reverse map to collection reference
1068         for _, bind := range binds {
1069                 mnt := runner.Container.Mounts[bind]
1070                 if tgt == bind || strings.HasPrefix(tgt, bind+"/") {
1071                         // get path relative to bind
1072                         targetSuffix := tgt[len(bind):]
1073
1074                         // Copy mount and adjust the path to add path relative to the bind
1075                         adjustedMount := mnt
1076                         adjustedMount.Path = filepath.Join(adjustedMount.Path, targetSuffix)
1077
1078                         // Terminates in this keep mount, so add the
1079                         // manifest text at appropriate location.
1080                         outputSuffix := path[len(runner.HostOutputDir):]
1081                         manifestText, err = runner.getCollectionManifestForPath(adjustedMount, outputSuffix)
1082                         return
1083                 }
1084         }
1085
1086         // If target is not a collection mount, it must be located within the
1087         // output directory, otherwise it is an error.
1088         if derefErr == ErrNotInOutputDir {
1089                 err = fmt.Errorf("Symlink in output %q points to invalid location %q, must point to path within the output directory.",
1090                         path[len(runner.HostOutputDir):], readlinktgt)
1091                 return
1092         }
1093
1094         if info.Mode().IsRegular() {
1095                 return "", walkUpload.UploadFile(relocated, tgt)
1096         }
1097
1098         if info.Mode().IsDir() {
1099                 // Symlink leads to directory.  Walk() doesn't follow
1100                 // directory symlinks, so we walk the target directory
1101                 // instead.  Within the walk, file paths are relocated
1102                 // so they appear under the original symlink path.
1103                 err = filepath.Walk(tgt, func(walkpath string, walkinfo os.FileInfo, walkerr error) error {
1104                         var m string
1105                         m, walkerr = runner.UploadOutputFile(walkpath, walkinfo, walkerr,
1106                                 binds, walkUpload, tgt, relocated, followed+1)
1107                         if walkerr == nil {
1108                                 manifestText = manifestText + m
1109                         }
1110                         return walkerr
1111                 })
1112                 return
1113         }
1114
1115         return
1116 }
1117
1118 // HandleOutput sets the output, unmounts the FUSE mount, and deletes temporary directories
1119 func (runner *ContainerRunner) CaptureOutput() error {
1120         if runner.finalState != "Complete" {
1121                 return nil
1122         }
1123
1124         if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
1125                 // Output may have been set directly by the container, so
1126                 // refresh the container record to check.
1127                 err := runner.ArvClient.Get("containers", runner.Container.UUID,
1128                         nil, &runner.Container)
1129                 if err != nil {
1130                         return err
1131                 }
1132                 if runner.Container.Output != "" {
1133                         // Container output is already set.
1134                         runner.OutputPDH = &runner.Container.Output
1135                         return nil
1136                 }
1137         }
1138
1139         if runner.HostOutputDir == "" {
1140                 return nil
1141         }
1142
1143         _, err := os.Stat(runner.HostOutputDir)
1144         if err != nil {
1145                 return fmt.Errorf("While checking host output path: %v", err)
1146         }
1147
1148         // Pre-populate output from the configured mount points
1149         var binds []string
1150         for bind, mnt := range runner.Container.Mounts {
1151                 if mnt.Kind == "collection" {
1152                         binds = append(binds, bind)
1153                 }
1154         }
1155         sort.Strings(binds)
1156
1157         var manifestText string
1158
1159         collectionMetafile := fmt.Sprintf("%s/.arvados#collection", runner.HostOutputDir)
1160         _, err = os.Stat(collectionMetafile)
1161         if err != nil {
1162                 // Regular directory
1163
1164                 cw := CollectionWriter{0, runner.Kc, nil, nil, sync.Mutex{}}
1165                 walkUpload := cw.BeginUpload(runner.HostOutputDir, runner.CrunchLog.Logger)
1166
1167                 var m string
1168                 err = filepath.Walk(runner.HostOutputDir, func(path string, info os.FileInfo, err error) error {
1169                         m, err = runner.UploadOutputFile(path, info, err, binds, walkUpload, "", "", 0)
1170                         if err == nil {
1171                                 manifestText = manifestText + m
1172                         }
1173                         return err
1174                 })
1175
1176                 cw.EndUpload(walkUpload)
1177
1178                 if err != nil {
1179                         return fmt.Errorf("While uploading output files: %v", err)
1180                 }
1181
1182                 m, err = cw.ManifestText()
1183                 manifestText = manifestText + m
1184                 if err != nil {
1185                         return fmt.Errorf("While uploading output files: %v", err)
1186                 }
1187         } else {
1188                 // FUSE mount directory
1189                 file, openerr := os.Open(collectionMetafile)
1190                 if openerr != nil {
1191                         return fmt.Errorf("While opening FUSE metafile: %v", err)
1192                 }
1193                 defer file.Close()
1194
1195                 var rec arvados.Collection
1196                 err = json.NewDecoder(file).Decode(&rec)
1197                 if err != nil {
1198                         return fmt.Errorf("While reading FUSE metafile: %v", err)
1199                 }
1200                 manifestText = rec.ManifestText
1201         }
1202
1203         for _, bind := range binds {
1204                 mnt := runner.Container.Mounts[bind]
1205
1206                 bindSuffix := strings.TrimPrefix(bind, runner.Container.OutputPath)
1207
1208                 if bindSuffix == bind || len(bindSuffix) <= 0 {
1209                         // either does not start with OutputPath or is OutputPath itself
1210                         continue
1211                 }
1212
1213                 if mnt.ExcludeFromOutput == true {
1214                         continue
1215                 }
1216
1217                 // append to manifest_text
1218                 m, err := runner.getCollectionManifestForPath(mnt, bindSuffix)
1219                 if err != nil {
1220                         return err
1221                 }
1222
1223                 manifestText = manifestText + m
1224         }
1225
1226         // Save output
1227         var response arvados.Collection
1228         manifest := manifest.Manifest{Text: manifestText}
1229         manifestText = manifest.Extract(".", ".").Text
1230         err = runner.ArvClient.Create("collections",
1231                 arvadosclient.Dict{
1232                         "ensure_unique_name": true,
1233                         "collection": arvadosclient.Dict{
1234                                 "is_trashed":    true,
1235                                 "name":          "output for " + runner.Container.UUID,
1236                                 "manifest_text": manifestText}},
1237                 &response)
1238         if err != nil {
1239                 return fmt.Errorf("While creating output collection: %v", err)
1240         }
1241         runner.OutputPDH = &response.PortableDataHash
1242         return nil
1243 }
1244
1245 var outputCollections = make(map[string]arvados.Collection)
1246
1247 // Fetch the collection for the mnt.PortableDataHash
1248 // Return the manifest_text fragment corresponding to the specified mnt.Path
1249 //  after making any required updates.
1250 //  Ex:
1251 //    If mnt.Path is not specified,
1252 //      return the entire manifest_text after replacing any "." with bindSuffix
1253 //    If mnt.Path corresponds to one stream,
1254 //      return the manifest_text for that stream after replacing that stream name with bindSuffix
1255 //    Otherwise, check if a filename in any one stream is being sought. Return the manifest_text
1256 //      for that stream after replacing stream name with bindSuffix minus the last word
1257 //      and the file name with last word of the bindSuffix
1258 //  Allowed path examples:
1259 //    "path":"/"
1260 //    "path":"/subdir1"
1261 //    "path":"/subdir1/subdir2"
1262 //    "path":"/subdir/filename" etc
1263 func (runner *ContainerRunner) getCollectionManifestForPath(mnt arvados.Mount, bindSuffix string) (string, error) {
1264         collection := outputCollections[mnt.PortableDataHash]
1265         if collection.PortableDataHash == "" {
1266                 err := runner.ArvClient.Get("collections", mnt.PortableDataHash, nil, &collection)
1267                 if err != nil {
1268                         return "", fmt.Errorf("While getting collection for %v: %v", mnt.PortableDataHash, err)
1269                 }
1270                 outputCollections[mnt.PortableDataHash] = collection
1271         }
1272
1273         if collection.ManifestText == "" {
1274                 runner.CrunchLog.Printf("No manifest text for collection %v", collection.PortableDataHash)
1275                 return "", nil
1276         }
1277
1278         mft := manifest.Manifest{Text: collection.ManifestText}
1279         extracted := mft.Extract(mnt.Path, bindSuffix)
1280         if extracted.Err != nil {
1281                 return "", fmt.Errorf("Error parsing manifest for %v: %v", mnt.PortableDataHash, extracted.Err.Error())
1282         }
1283         return extracted.Text, nil
1284 }
1285
1286 func (runner *ContainerRunner) CleanupDirs() {
1287         if runner.ArvMount != nil {
1288                 var delay int64 = 8
1289                 umount := exec.Command("arv-mount", fmt.Sprintf("--unmount-timeout=%d", delay), "--unmount", runner.ArvMountPoint)
1290                 umount.Stdout = runner.CrunchLog
1291                 umount.Stderr = runner.CrunchLog
1292                 runner.CrunchLog.Printf("Running %v", umount.Args)
1293                 umnterr := umount.Start()
1294
1295                 if umnterr != nil {
1296                         runner.CrunchLog.Printf("Error unmounting: %v", umnterr)
1297                 } else {
1298                         // If arv-mount --unmount gets stuck for any reason, we
1299                         // don't want to wait for it forever.  Do Wait() in a goroutine
1300                         // so it doesn't block crunch-run.
1301                         umountExit := make(chan error)
1302                         go func() {
1303                                 mnterr := umount.Wait()
1304                                 if mnterr != nil {
1305                                         runner.CrunchLog.Printf("Error unmounting: %v", mnterr)
1306                                 }
1307                                 umountExit <- mnterr
1308                         }()
1309
1310                         for again := true; again; {
1311                                 again = false
1312                                 select {
1313                                 case <-umountExit:
1314                                         umount = nil
1315                                         again = true
1316                                 case <-runner.ArvMountExit:
1317                                         break
1318                                 case <-time.After(time.Duration((delay + 1) * int64(time.Second))):
1319                                         runner.CrunchLog.Printf("Timed out waiting for unmount")
1320                                         if umount != nil {
1321                                                 umount.Process.Kill()
1322                                         }
1323                                         runner.ArvMount.Process.Kill()
1324                                 }
1325                         }
1326                 }
1327         }
1328
1329         if runner.ArvMountPoint != "" {
1330                 if rmerr := os.Remove(runner.ArvMountPoint); rmerr != nil {
1331                         runner.CrunchLog.Printf("While cleaning up arv-mount directory %s: %v", runner.ArvMountPoint, rmerr)
1332                 }
1333         }
1334
1335         for _, tmpdir := range runner.CleanupTempDir {
1336                 if rmerr := os.RemoveAll(tmpdir); rmerr != nil {
1337                         runner.CrunchLog.Printf("While cleaning up temporary directory %s: %v", tmpdir, rmerr)
1338                 }
1339         }
1340 }
1341
1342 // CommitLogs posts the collection containing the final container logs.
1343 func (runner *ContainerRunner) CommitLogs() error {
1344         runner.CrunchLog.Print(runner.finalState)
1345
1346         if runner.arvMountLog != nil {
1347                 runner.arvMountLog.Close()
1348         }
1349         runner.CrunchLog.Close()
1350
1351         // Closing CrunchLog above allows them to be committed to Keep at this
1352         // point, but re-open crunch log with ArvClient in case there are any
1353         // other further errors (such as failing to write the log to Keep!)
1354         // while shutting down
1355         runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{ArvClient: runner.ArvClient,
1356                 UUID: runner.Container.UUID, loggingStream: "crunch-run", writeCloser: nil})
1357         runner.CrunchLog.Immediate = log.New(os.Stderr, runner.Container.UUID+" ", 0)
1358
1359         if runner.LogsPDH != nil {
1360                 // If we have already assigned something to LogsPDH,
1361                 // we must be closing the re-opened log, which won't
1362                 // end up getting attached to the container record and
1363                 // therefore doesn't need to be saved as a collection
1364                 // -- it exists only to send logs to other channels.
1365                 return nil
1366         }
1367
1368         mt, err := runner.LogCollection.ManifestText()
1369         if err != nil {
1370                 return fmt.Errorf("While creating log manifest: %v", err)
1371         }
1372
1373         var response arvados.Collection
1374         err = runner.ArvClient.Create("collections",
1375                 arvadosclient.Dict{
1376                         "ensure_unique_name": true,
1377                         "collection": arvadosclient.Dict{
1378                                 "is_trashed":    true,
1379                                 "name":          "logs for " + runner.Container.UUID,
1380                                 "manifest_text": mt}},
1381                 &response)
1382         if err != nil {
1383                 return fmt.Errorf("While creating log collection: %v", err)
1384         }
1385         runner.LogsPDH = &response.PortableDataHash
1386         return nil
1387 }
1388
1389 // UpdateContainerRunning updates the container state to "Running"
1390 func (runner *ContainerRunner) UpdateContainerRunning() error {
1391         runner.cStateLock.Lock()
1392         defer runner.cStateLock.Unlock()
1393         if runner.cCancelled {
1394                 return ErrCancelled
1395         }
1396         return runner.ArvClient.Update("containers", runner.Container.UUID,
1397                 arvadosclient.Dict{"container": arvadosclient.Dict{"state": "Running"}}, nil)
1398 }
1399
1400 // ContainerToken returns the api_token the container (and any
1401 // arv-mount processes) are allowed to use.
1402 func (runner *ContainerRunner) ContainerToken() (string, error) {
1403         if runner.token != "" {
1404                 return runner.token, nil
1405         }
1406
1407         var auth arvados.APIClientAuthorization
1408         err := runner.ArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth)
1409         if err != nil {
1410                 return "", err
1411         }
1412         runner.token = auth.APIToken
1413         return runner.token, nil
1414 }
1415
1416 // UpdateContainerComplete updates the container record state on API
1417 // server to "Complete" or "Cancelled"
1418 func (runner *ContainerRunner) UpdateContainerFinal() error {
1419         update := arvadosclient.Dict{}
1420         update["state"] = runner.finalState
1421         if runner.LogsPDH != nil {
1422                 update["log"] = *runner.LogsPDH
1423         }
1424         if runner.finalState == "Complete" {
1425                 if runner.ExitCode != nil {
1426                         update["exit_code"] = *runner.ExitCode
1427                 }
1428                 if runner.OutputPDH != nil {
1429                         update["output"] = *runner.OutputPDH
1430                 }
1431         }
1432         return runner.ArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{"container": update}, nil)
1433 }
1434
1435 // IsCancelled returns the value of Cancelled, with goroutine safety.
1436 func (runner *ContainerRunner) IsCancelled() bool {
1437         runner.cStateLock.Lock()
1438         defer runner.cStateLock.Unlock()
1439         return runner.cCancelled
1440 }
1441
1442 // NewArvLogWriter creates an ArvLogWriter
1443 func (runner *ContainerRunner) NewArvLogWriter(name string) io.WriteCloser {
1444         return &ArvLogWriter{
1445                 ArvClient:     runner.ArvClient,
1446                 UUID:          runner.Container.UUID,
1447                 loggingStream: name,
1448                 writeCloser:   runner.LogCollection.Open(name + ".txt")}
1449 }
1450
1451 // Run the full container lifecycle.
1452 func (runner *ContainerRunner) Run() (err error) {
1453         runner.CrunchLog.Printf("crunch-run %s started", version)
1454         runner.CrunchLog.Printf("Executing container '%s'", runner.Container.UUID)
1455
1456         hostname, hosterr := os.Hostname()
1457         if hosterr != nil {
1458                 runner.CrunchLog.Printf("Error getting hostname '%v'", hosterr)
1459         } else {
1460                 runner.CrunchLog.Printf("Executing on host '%s'", hostname)
1461         }
1462
1463         runner.finalState = "Queued"
1464
1465         defer func() {
1466                 runner.stopSignals()
1467                 runner.CleanupDirs()
1468
1469                 runner.CrunchLog.Printf("crunch-run finished")
1470                 runner.CrunchLog.Close()
1471         }()
1472
1473         defer func() {
1474                 // checkErr prints e (unless it's nil) and sets err to
1475                 // e (unless err is already non-nil). Thus, if err
1476                 // hasn't already been assigned when Run() returns,
1477                 // this cleanup func will cause Run() to return the
1478                 // first non-nil error that is passed to checkErr().
1479                 checkErr := func(e error) {
1480                         if e == nil {
1481                                 return
1482                         }
1483                         runner.CrunchLog.Print(e)
1484                         if err == nil {
1485                                 err = e
1486                         }
1487                         if runner.finalState == "Complete" {
1488                                 // There was an error in the finalization.
1489                                 runner.finalState = "Cancelled"
1490                         }
1491                 }
1492
1493                 // Log the error encountered in Run(), if any
1494                 checkErr(err)
1495
1496                 if runner.finalState == "Queued" {
1497                         runner.UpdateContainerFinal()
1498                         return
1499                 }
1500
1501                 if runner.IsCancelled() {
1502                         runner.finalState = "Cancelled"
1503                         // but don't return yet -- we still want to
1504                         // capture partial output and write logs
1505                 }
1506
1507                 checkErr(runner.CaptureOutput())
1508                 checkErr(runner.CommitLogs())
1509                 checkErr(runner.UpdateContainerFinal())
1510         }()
1511
1512         err = runner.fetchContainerRecord()
1513         if err != nil {
1514                 return
1515         }
1516
1517         // setup signal handling
1518         runner.setupSignals()
1519
1520         // check for and/or load image
1521         err = runner.LoadImage()
1522         if err != nil {
1523                 if !runner.checkBrokenNode(err) {
1524                         // Failed to load image but not due to a "broken node"
1525                         // condition, probably user error.
1526                         runner.finalState = "Cancelled"
1527                 }
1528                 err = fmt.Errorf("While loading container image: %v", err)
1529                 return
1530         }
1531
1532         // set up FUSE mount and binds
1533         err = runner.SetupMounts()
1534         if err != nil {
1535                 runner.finalState = "Cancelled"
1536                 err = fmt.Errorf("While setting up mounts: %v", err)
1537                 return
1538         }
1539
1540         err = runner.CreateContainer()
1541         if err != nil {
1542                 return
1543         }
1544
1545         // Gather and record node information
1546         err = runner.LogNodeInfo()
1547         if err != nil {
1548                 return
1549         }
1550         // Save container.json record on log collection
1551         err = runner.LogContainerRecord()
1552         if err != nil {
1553                 return
1554         }
1555
1556         if runner.IsCancelled() {
1557                 return
1558         }
1559
1560         err = runner.UpdateContainerRunning()
1561         if err != nil {
1562                 return
1563         }
1564         runner.finalState = "Cancelled"
1565
1566         runner.StartCrunchstat()
1567
1568         err = runner.StartContainer()
1569         if err != nil {
1570                 runner.checkBrokenNode(err)
1571                 return
1572         }
1573
1574         err = runner.WaitFinish()
1575         if err == nil {
1576                 runner.finalState = "Complete"
1577         }
1578         return
1579 }
1580
1581 // Fetch the current container record (uuid = runner.Container.UUID)
1582 // into runner.Container.
1583 func (runner *ContainerRunner) fetchContainerRecord() error {
1584         reader, err := runner.ArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
1585         if err != nil {
1586                 return fmt.Errorf("error fetching container record: %v", err)
1587         }
1588         defer reader.Close()
1589
1590         dec := json.NewDecoder(reader)
1591         dec.UseNumber()
1592         err = dec.Decode(&runner.Container)
1593         if err != nil {
1594                 return fmt.Errorf("error decoding container record: %v", err)
1595         }
1596         return nil
1597 }
1598
1599 // NewContainerRunner creates a new container runner.
1600 func NewContainerRunner(api IArvadosClient,
1601         kc IKeepClient,
1602         docker ThinDockerClient,
1603         containerUUID string) *ContainerRunner {
1604
1605         cr := &ContainerRunner{ArvClient: api, Kc: kc, Docker: docker}
1606         cr.NewLogWriter = cr.NewArvLogWriter
1607         cr.RunArvMount = cr.ArvMountCmd
1608         cr.MkTempDir = ioutil.TempDir
1609         cr.LogCollection = &CollectionWriter{0, kc, nil, nil, sync.Mutex{}}
1610         cr.Container.UUID = containerUUID
1611         cr.CrunchLog = NewThrottledLogger(cr.NewLogWriter("crunch-run"))
1612         cr.CrunchLog.Immediate = log.New(os.Stderr, containerUUID+" ", 0)
1613
1614         loadLogThrottleParams(api)
1615
1616         return cr
1617 }
1618
1619 func main() {
1620         statInterval := flag.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting")
1621         cgroupRoot := flag.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree")
1622         cgroupParent := flag.String("cgroup-parent", "docker", "name of container's parent cgroup (ignored if -cgroup-parent-subsystem is used)")
1623         cgroupParentSubsystem := flag.String("cgroup-parent-subsystem", "", "use current cgroup for given subsystem as parent cgroup for container")
1624         caCertsPath := flag.String("ca-certs", "", "Path to TLS root certificates")
1625         enableNetwork := flag.String("container-enable-networking", "default",
1626                 `Specify if networking should be enabled for container.  One of 'default', 'always':
1627         default: only enable networking if container requests it.
1628         always:  containers always have networking enabled
1629         `)
1630         networkMode := flag.String("container-network-mode", "default",
1631                 `Set networking mode for container.  Corresponds to Docker network mode (--net).
1632         `)
1633         memprofile := flag.String("memprofile", "", "write memory profile to `file` after running container")
1634         getVersion := flag.Bool("version", false, "Print version information and exit.")
1635         flag.Parse()
1636
1637         // Print version information if requested
1638         if *getVersion {
1639                 fmt.Printf("crunch-run %s\n", version)
1640                 return
1641         }
1642
1643         log.Printf("crunch-run %s started", version)
1644
1645         containerId := flag.Arg(0)
1646
1647         if *caCertsPath != "" {
1648                 arvadosclient.CertFiles = []string{*caCertsPath}
1649         }
1650
1651         api, err := arvadosclient.MakeArvadosClient()
1652         if err != nil {
1653                 log.Fatalf("%s: %v", containerId, err)
1654         }
1655         api.Retries = 8
1656
1657         kc, kcerr := keepclient.MakeKeepClient(api)
1658         if kcerr != nil {
1659                 log.Fatalf("%s: %v", containerId, kcerr)
1660         }
1661         kc.BlockCache = &keepclient.BlockCache{MaxBlocks: 2}
1662         kc.Retries = 4
1663
1664         // API version 1.21 corresponds to Docker 1.9, which is currently the
1665         // minimum version we want to support.
1666         docker, dockererr := dockerclient.NewClient(dockerclient.DefaultDockerHost, "1.21", nil, nil)
1667         dockerClientProxy := ThinDockerClientProxy{Docker: docker}
1668
1669         cr := NewContainerRunner(api, kc, dockerClientProxy, containerId)
1670
1671         if dockererr != nil {
1672                 cr.CrunchLog.Printf("%s: %v", containerId, dockererr)
1673                 cr.checkBrokenNode(dockererr)
1674                 cr.CrunchLog.Close()
1675                 os.Exit(1)
1676         }
1677
1678         cr.statInterval = *statInterval
1679         cr.cgroupRoot = *cgroupRoot
1680         cr.expectCgroupParent = *cgroupParent
1681         cr.enableNetwork = *enableNetwork
1682         cr.networkMode = *networkMode
1683         if *cgroupParentSubsystem != "" {
1684                 p := findCgroup(*cgroupParentSubsystem)
1685                 cr.setCgroupParent = p
1686                 cr.expectCgroupParent = p
1687         }
1688
1689         runerr := cr.Run()
1690
1691         if *memprofile != "" {
1692                 f, err := os.Create(*memprofile)
1693                 if err != nil {
1694                         log.Printf("could not create memory profile: ", err)
1695                 }
1696                 runtime.GC() // get up-to-date statistics
1697                 if err := pprof.WriteHeapProfile(f); err != nil {
1698                         log.Printf("could not write memory profile: ", err)
1699                 }
1700                 closeerr := f.Close()
1701                 if closeerr != nil {
1702                         log.Printf("closing memprofile file: ", err)
1703                 }
1704         }
1705
1706         if runerr != nil {
1707                 log.Fatalf("%s: %v", containerId, runerr)
1708         }
1709 }