services/crunch-run/crunchrun.go

   1 // Copyright (C) The Arvados Authors. All rights reserved.
   2 //
   3 // SPDX-License-Identifier: AGPL-3.0
   4
   5 package main
   6
   7 import (
   8         "bytes"
   9         "encoding/json"
  10         "errors"
  11         "flag"
  12         "fmt"
  13         "io"
  14         "io/ioutil"
  15         "log"
  16         "os"
  17         "os/exec"
  18         "os/signal"
  19         "path"
  20         "path/filepath"
  21         "regexp"
  22         "runtime"
  23         "runtime/pprof"
  24         "sort"
  25         "strings"
  26         "sync"
  27         "syscall"
  28         "time"
  29
  30         "git.curoverse.com/arvados.git/lib/crunchstat"
  31         "git.curoverse.com/arvados.git/sdk/go/arvados"
  32         "git.curoverse.com/arvados.git/sdk/go/arvadosclient"
  33         "git.curoverse.com/arvados.git/sdk/go/keepclient"
  34         "git.curoverse.com/arvados.git/sdk/go/manifest"
  35         "golang.org/x/net/context"
  36
  37         dockertypes "github.com/docker/docker/api/types"
  38         dockercontainer "github.com/docker/docker/api/types/container"
  39         dockernetwork "github.com/docker/docker/api/types/network"
  40         dockerclient "github.com/docker/docker/client"
  41 )
  42
  43 var version = "dev"
  44
  45 // IArvadosClient is the minimal Arvados API methods used by crunch-run.
  46 type IArvadosClient interface {
  47         Create(resourceType string, parameters arvadosclient.Dict, output interface{}) error
  48         Get(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  49         Update(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  50         Call(method, resourceType, uuid, action string, parameters arvadosclient.Dict, output interface{}) error
  51         CallRaw(method string, resourceType string, uuid string, action string, parameters arvadosclient.Dict) (reader io.ReadCloser, err error)
  52         Discovery(key string) (interface{}, error)
  53 }
  54
  55 // ErrCancelled is the error returned when the container is cancelled.
  56 var ErrCancelled = errors.New("Cancelled")
  57
  58 // IKeepClient is the minimal Keep API methods used by crunch-run.
  59 type IKeepClient interface {
  60         PutHB(hash string, buf []byte) (string, int, error)
  61         ManifestFileReader(m manifest.Manifest, filename string) (arvados.File, error)
  62         ClearBlockCache()
  63 }
  64
  65 // NewLogWriter is a factory function to create a new log writer.
  66 type NewLogWriter func(name string) io.WriteCloser
  67
  68 type RunArvMount func(args []string, tok string) (*exec.Cmd, error)
  69
  70 type MkTempDir func(string, string) (string, error)
  71
  72 // ThinDockerClient is the minimal Docker client interface used by crunch-run.
  73 type ThinDockerClient interface {
  74         ContainerAttach(ctx context.Context, container string, options dockertypes.ContainerAttachOptions) (dockertypes.HijackedResponse, error)
  75         ContainerCreate(ctx context.Context, config *dockercontainer.Config, hostConfig *dockercontainer.HostConfig,
  76                 networkingConfig *dockernetwork.NetworkingConfig, containerName string) (dockercontainer.ContainerCreateCreatedBody, error)
  77         ContainerStart(ctx context.Context, container string, options dockertypes.ContainerStartOptions) error
  78         ContainerRemove(ctx context.Context, container string, options dockertypes.ContainerRemoveOptions) error
  79         ContainerWait(ctx context.Context, container string, condition dockercontainer.WaitCondition) (<-chan dockercontainer.ContainerWaitOKBody, <-chan error)
  80         ImageInspectWithRaw(ctx context.Context, image string) (dockertypes.ImageInspect, []byte, error)
  81         ImageLoad(ctx context.Context, input io.Reader, quiet bool) (dockertypes.ImageLoadResponse, error)
  82         ImageRemove(ctx context.Context, image string, options dockertypes.ImageRemoveOptions) ([]dockertypes.ImageDeleteResponseItem, error)
  83 }
  84
  85 // ContainerRunner is the main stateful struct used for a single execution of a
  86 // container.
  87 type ContainerRunner struct {
  88         Docker    ThinDockerClient
  89         ArvClient IArvadosClient
  90         Kc        IKeepClient
  91         arvados.Container
  92         ContainerConfig dockercontainer.Config
  93         dockercontainer.HostConfig
  94         token       string
  95         ContainerID string
  96         ExitCode    *int
  97         NewLogWriter
  98         loggingDone   chan bool
  99         CrunchLog     *ThrottledLogger
 100         Stdout        io.WriteCloser
 101         Stderr        io.WriteCloser
 102         LogCollection *CollectionWriter
 103         LogsPDH       *string
 104         RunArvMount
 105         MkTempDir
 106         ArvMount       *exec.Cmd
 107         ArvMountPoint  string
 108         HostOutputDir  string
 109         CleanupTempDir []string
 110         Binds          []string
 111         Volumes        map[string]struct{}
 112         OutputPDH      *string
 113         SigChan        chan os.Signal
 114         ArvMountExit   chan error
 115         finalState     string
 116
 117         statLogger       io.WriteCloser
 118         statReporter     *crunchstat.Reporter
 119         hoststatLogger   io.WriteCloser
 120         hoststatReporter *crunchstat.Reporter
 121         statInterval     time.Duration
 122         cgroupRoot       string
 123         // What we expect the container's cgroup parent to be.
 124         expectCgroupParent string
 125         // What we tell docker to use as the container's cgroup
 126         // parent. Note: Ideally we would use the same field for both
 127         // expectCgroupParent and setCgroupParent, and just make it
 128         // default to "docker". However, when using docker < 1.10 with
 129         // systemd, specifying a non-empty cgroup parent (even the
 130         // default value "docker") hits a docker bug
 131         // (https://github.com/docker/docker/issues/17126). Using two
 132         // separate fields makes it possible to use the "expect cgroup
 133         // parent to be X" feature even on sites where the "specify
 134         // cgroup parent" feature breaks.
 135         setCgroupParent string
 136
 137         cStateLock sync.Mutex
 138         cCancelled bool // StopContainer() invoked
 139
 140         enableNetwork string // one of "default" or "always"
 141         networkMode   string // passed through to HostConfig.NetworkMode
 142         arvMountLog   *ThrottledLogger
 143 }
 144
 145 // setupSignals sets up signal handling to gracefully terminate the underlying
 146 // Docker container and update state when receiving a TERM, INT or QUIT signal.
 147 func (runner *ContainerRunner) setupSignals() {
 148         runner.SigChan = make(chan os.Signal, 1)
 149         signal.Notify(runner.SigChan, syscall.SIGTERM)
 150         signal.Notify(runner.SigChan, syscall.SIGINT)
 151         signal.Notify(runner.SigChan, syscall.SIGQUIT)
 152
 153         go func(sig chan os.Signal) {
 154                 for s := range sig {
 155                         runner.CrunchLog.Printf("caught signal: %v", s)
 156                         runner.stop()
 157                 }
 158         }(runner.SigChan)
 159 }
 160
 161 // stop the underlying Docker container.
 162 func (runner *ContainerRunner) stop() {
 163         runner.cStateLock.Lock()
 164         defer runner.cStateLock.Unlock()
 165         if runner.ContainerID == "" {
 166                 return
 167         }
 168         runner.cCancelled = true
 169         runner.CrunchLog.Printf("removing container")
 170         err := runner.Docker.ContainerRemove(context.TODO(), runner.ContainerID, dockertypes.ContainerRemoveOptions{Force: true})
 171         if err != nil {
 172                 runner.CrunchLog.Printf("error removing container: %s", err)
 173         }
 174 }
 175
 176 func (runner *ContainerRunner) stopSignals() {
 177         if runner.SigChan != nil {
 178                 signal.Stop(runner.SigChan)
 179         }
 180 }
 181
 182 var errorBlacklist = []string{
 183         "(?ms).*[Cc]annot connect to the Docker daemon.*",
 184         "(?ms).*oci runtime error.*starting container process.*container init.*mounting.*to rootfs.*no such file or directory.*",
 185 }
 186 var brokenNodeHook *string = flag.String("broken-node-hook", "", "Script to run if node is detected to be broken (for example, Docker daemon is not running)")
 187
 188 func (runner *ContainerRunner) checkBrokenNode(goterr error) bool {
 189         for _, d := range errorBlacklist {
 190                 if m, e := regexp.MatchString(d, goterr.Error()); m && e == nil {
 191                         runner.CrunchLog.Printf("Error suggests node is unable to run containers: %v", goterr)
 192                         if *brokenNodeHook == "" {
 193                                 runner.CrunchLog.Printf("No broken node hook provided, cannot mark node as broken.")
 194                         } else {
 195                                 runner.CrunchLog.Printf("Running broken node hook %q", *brokenNodeHook)
 196                                 // run killme script
 197                                 c := exec.Command(*brokenNodeHook)
 198                                 c.Stdout = runner.CrunchLog
 199                                 c.Stderr = runner.CrunchLog
 200                                 err := c.Run()
 201                                 if err != nil {
 202                                         runner.CrunchLog.Printf("Error running broken node hook: %v", err)
 203                                 }
 204                         }
 205                         return true
 206                 }
 207         }
 208         return false
 209 }
 210
 211 // LoadImage determines the docker image id from the container record and
 212 // checks if it is available in the local Docker image store.  If not, it loads
 213 // the image from Keep.
 214 func (runner *ContainerRunner) LoadImage() (err error) {
 215
 216         runner.CrunchLog.Printf("Fetching Docker image from collection '%s'", runner.Container.ContainerImage)
 217
 218         var collection arvados.Collection
 219         err = runner.ArvClient.Get("collections", runner.Container.ContainerImage, nil, &collection)
 220         if err != nil {
 221                 return fmt.Errorf("While getting container image collection: %v", err)
 222         }
 223         manifest := manifest.Manifest{Text: collection.ManifestText}
 224         var img, imageID string
 225         for ms := range manifest.StreamIter() {
 226                 img = ms.FileStreamSegments[0].Name
 227                 if !strings.HasSuffix(img, ".tar") {
 228                         return fmt.Errorf("First file in the container image collection does not end in .tar")
 229                 }
 230                 imageID = img[:len(img)-4]
 231         }
 232
 233         runner.CrunchLog.Printf("Using Docker image id '%s'", imageID)
 234
 235         _, _, err = runner.Docker.ImageInspectWithRaw(context.TODO(), imageID)
 236         if err != nil {
 237                 runner.CrunchLog.Print("Loading Docker image from keep")
 238
 239                 var readCloser io.ReadCloser
 240                 readCloser, err = runner.Kc.ManifestFileReader(manifest, img)
 241                 if err != nil {
 242                         return fmt.Errorf("While creating ManifestFileReader for container image: %v", err)
 243                 }
 244
 245                 response, err := runner.Docker.ImageLoad(context.TODO(), readCloser, true)
 246                 if err != nil {
 247                         return fmt.Errorf("While loading container image into Docker: %v", err)
 248                 }
 249
 250                 defer response.Body.Close()
 251                 rbody, err := ioutil.ReadAll(response.Body)
 252                 if err != nil {
 253                         return fmt.Errorf("Reading response to image load: %v", err)
 254                 }
 255                 runner.CrunchLog.Printf("Docker response: %s", rbody)
 256         } else {
 257                 runner.CrunchLog.Print("Docker image is available")
 258         }
 259
 260         runner.ContainerConfig.Image = imageID
 261
 262         runner.Kc.ClearBlockCache()
 263
 264         return nil
 265 }
 266
 267 func (runner *ContainerRunner) ArvMountCmd(arvMountCmd []string, token string) (c *exec.Cmd, err error) {
 268         c = exec.Command("arv-mount", arvMountCmd...)
 269
 270         // Copy our environment, but override ARVADOS_API_TOKEN with
 271         // the container auth token.
 272         c.Env = nil
 273         for _, s := range os.Environ() {
 274                 if !strings.HasPrefix(s, "ARVADOS_API_TOKEN=") {
 275                         c.Env = append(c.Env, s)
 276                 }
 277         }
 278         c.Env = append(c.Env, "ARVADOS_API_TOKEN="+token)
 279
 280         runner.arvMountLog = NewThrottledLogger(runner.NewLogWriter("arv-mount"))
 281         c.Stdout = runner.arvMountLog
 282         c.Stderr = runner.arvMountLog
 283
 284         runner.CrunchLog.Printf("Running %v", c.Args)
 285
 286         err = c.Start()
 287         if err != nil {
 288                 return nil, err
 289         }
 290
 291         statReadme := make(chan bool)
 292         runner.ArvMountExit = make(chan error)
 293
 294         keepStatting := true
 295         go func() {
 296                 for keepStatting {
 297                         time.Sleep(100 * time.Millisecond)
 298                         _, err = os.Stat(fmt.Sprintf("%s/by_id/README", runner.ArvMountPoint))
 299                         if err == nil {
 300                                 keepStatting = false
 301                                 statReadme <- true
 302                         }
 303                 }
 304                 close(statReadme)
 305         }()
 306
 307         go func() {
 308                 mnterr := c.Wait()
 309                 if mnterr != nil {
 310                         runner.CrunchLog.Printf("Arv-mount exit error: %v", mnterr)
 311                 }
 312                 runner.ArvMountExit <- mnterr
 313                 close(runner.ArvMountExit)
 314         }()
 315
 316         select {
 317         case <-statReadme:
 318                 break
 319         case err := <-runner.ArvMountExit:
 320                 runner.ArvMount = nil
 321                 keepStatting = false
 322                 return nil, err
 323         }
 324
 325         return c, nil
 326 }
 327
 328 func (runner *ContainerRunner) SetupArvMountPoint(prefix string) (err error) {
 329         if runner.ArvMountPoint == "" {
 330                 runner.ArvMountPoint, err = runner.MkTempDir("", prefix)
 331         }
 332         return
 333 }
 334
 335 func (runner *ContainerRunner) SetupMounts() (err error) {
 336         err = runner.SetupArvMountPoint("keep")
 337         if err != nil {
 338                 return fmt.Errorf("While creating keep mount temp dir: %v", err)
 339         }
 340
 341         token, err := runner.ContainerToken()
 342         if err != nil {
 343                 return fmt.Errorf("could not get container token: %s", err)
 344         }
 345
 346         pdhOnly := true
 347         tmpcount := 0
 348         arvMountCmd := []string{
 349                 "--foreground",
 350                 "--allow-other",
 351                 "--read-write",
 352                 fmt.Sprintf("--crunchstat-interval=%v", runner.statInterval.Seconds())}
 353
 354         if runner.Container.RuntimeConstraints.KeepCacheRAM > 0 {
 355                 arvMountCmd = append(arvMountCmd, "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheRAM))
 356         }
 357
 358         collectionPaths := []string{}
 359         runner.Binds = nil
 360         runner.Volumes = make(map[string]struct{})
 361         needCertMount := true
 362
 363         var binds []string
 364         for bind := range runner.Container.Mounts {
 365                 binds = append(binds, bind)
 366         }
 367         sort.Strings(binds)
 368
 369         for _, bind := range binds {
 370                 mnt := runner.Container.Mounts[bind]
 371                 if bind == "stdout" || bind == "stderr" {
 372                         // Is it a "file" mount kind?
 373                         if mnt.Kind != "file" {
 374                                 return fmt.Errorf("Unsupported mount kind '%s' for %s. Only 'file' is supported.", mnt.Kind, bind)
 375                         }
 376
 377                         // Does path start with OutputPath?
 378                         prefix := runner.Container.OutputPath
 379                         if !strings.HasSuffix(prefix, "/") {
 380                                 prefix += "/"
 381                         }
 382                         if !strings.HasPrefix(mnt.Path, prefix) {
 383                                 return fmt.Errorf("%s path does not start with OutputPath: %s, %s", strings.Title(bind), mnt.Path, prefix)
 384                         }
 385                 }
 386
 387                 if bind == "stdin" {
 388                         // Is it a "collection" mount kind?
 389                         if mnt.Kind != "collection" && mnt.Kind != "json" {
 390                                 return fmt.Errorf("Unsupported mount kind '%s' for stdin. Only 'collection' or 'json' are supported.", mnt.Kind)
 391                         }
 392                 }
 393
 394                 if bind == "/etc/arvados/ca-certificates.crt" {
 395                         needCertMount = false
 396                 }
 397
 398                 if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && bind != runner.Container.OutputPath+"/" {
 399                         if mnt.Kind != "collection" {
 400                                 return fmt.Errorf("Only mount points of kind 'collection' are supported underneath the output_path: %v", bind)
 401                         }
 402                 }
 403
 404                 switch {
 405                 case mnt.Kind == "collection" && bind != "stdin":
 406                         var src string
 407                         if mnt.UUID != "" && mnt.PortableDataHash != "" {
 408                                 return fmt.Errorf("Cannot specify both 'uuid' and 'portable_data_hash' for a collection mount")
 409                         }
 410                         if mnt.UUID != "" {
 411                                 if mnt.Writable {
 412                                         return fmt.Errorf("Writing to existing collections currently not permitted.")
 413                                 }
 414                                 pdhOnly = false
 415                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.UUID)
 416                         } else if mnt.PortableDataHash != "" {
 417                                 if mnt.Writable {
 418                                         return fmt.Errorf("Can never write to a collection specified by portable data hash")
 419                                 }
 420                                 idx := strings.Index(mnt.PortableDataHash, "/")
 421                                 if idx > 0 {
 422                                         mnt.Path = path.Clean(mnt.PortableDataHash[idx:])
 423                                         mnt.PortableDataHash = mnt.PortableDataHash[0:idx]
 424                                         runner.Container.Mounts[bind] = mnt
 425                                 }
 426                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.PortableDataHash)
 427                                 if mnt.Path != "" && mnt.Path != "." {
 428                                         if strings.HasPrefix(mnt.Path, "./") {
 429                                                 mnt.Path = mnt.Path[2:]
 430                                         } else if strings.HasPrefix(mnt.Path, "/") {
 431                                                 mnt.Path = mnt.Path[1:]
 432                                         }
 433                                         src += "/" + mnt.Path
 434                                 }
 435                         } else {
 436                                 src = fmt.Sprintf("%s/tmp%d", runner.ArvMountPoint, tmpcount)
 437                                 arvMountCmd = append(arvMountCmd, "--mount-tmp")
 438                                 arvMountCmd = append(arvMountCmd, fmt.Sprintf("tmp%d", tmpcount))
 439                                 tmpcount += 1
 440                         }
 441                         if mnt.Writable {
 442                                 if bind == runner.Container.OutputPath {
 443                                         runner.HostOutputDir = src
 444                                 } else if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 445                                         return fmt.Errorf("Writable mount points are not permitted underneath the output_path: %v", bind)
 446                                 }
 447                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", src, bind))
 448                         } else {
 449                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", src, bind))
 450                         }
 451                         collectionPaths = append(collectionPaths, src)
 452
 453                 case mnt.Kind == "tmp":
 454                         var tmpdir string
 455                         tmpdir, err = runner.MkTempDir("", "")
 456                         if err != nil {
 457                                 return fmt.Errorf("While creating mount temp dir: %v", err)
 458                         }
 459                         st, staterr := os.Stat(tmpdir)
 460                         if staterr != nil {
 461                                 return fmt.Errorf("While Stat on temp dir: %v", staterr)
 462                         }
 463                         err = os.Chmod(tmpdir, st.Mode()|os.ModeSetgid|0777)
 464                         if staterr != nil {
 465                                 return fmt.Errorf("While Chmod temp dir: %v", err)
 466                         }
 467                         runner.CleanupTempDir = append(runner.CleanupTempDir, tmpdir)
 468                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", tmpdir, bind))
 469                         if bind == runner.Container.OutputPath {
 470                                 runner.HostOutputDir = tmpdir
 471                         }
 472
 473                 case mnt.Kind == "json":
 474                         jsondata, err := json.Marshal(mnt.Content)
 475                         if err != nil {
 476                                 return fmt.Errorf("encoding json data: %v", err)
 477                         }
 478                         // Create a tempdir with a single file
 479                         // (instead of just a tempfile): this way we
 480                         // can ensure the file is world-readable
 481                         // inside the container, without having to
 482                         // make it world-readable on the docker host.
 483                         tmpdir, err := runner.MkTempDir("", "")
 484                         if err != nil {
 485                                 return fmt.Errorf("creating temp dir: %v", err)
 486                         }
 487                         runner.CleanupTempDir = append(runner.CleanupTempDir, tmpdir)
 488                         tmpfn := filepath.Join(tmpdir, "mountdata.json")
 489                         err = ioutil.WriteFile(tmpfn, jsondata, 0644)
 490                         if err != nil {
 491                                 return fmt.Errorf("writing temp file: %v", err)
 492                         }
 493                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", tmpfn, bind))
 494
 495                 case mnt.Kind == "git_tree":
 496                         tmpdir, err := runner.MkTempDir("", "")
 497                         if err != nil {
 498                                 return fmt.Errorf("creating temp dir: %v", err)
 499                         }
 500                         runner.CleanupTempDir = append(runner.CleanupTempDir, tmpdir)
 501                         err = gitMount(mnt).extractTree(runner.ArvClient, tmpdir, token)
 502                         if err != nil {
 503                                 return err
 504                         }
 505                         runner.Binds = append(runner.Binds, tmpdir+":"+bind+":ro")
 506                 }
 507         }
 508
 509         if runner.HostOutputDir == "" {
 510                 return fmt.Errorf("Output path does not correspond to a writable mount point")
 511         }
 512
 513         if wantAPI := runner.Container.RuntimeConstraints.API; needCertMount && wantAPI != nil && *wantAPI {
 514                 for _, certfile := range arvadosclient.CertFiles {
 515                         _, err := os.Stat(certfile)
 516                         if err == nil {
 517                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:/etc/arvados/ca-certificates.crt:ro", certfile))
 518                                 break
 519                         }
 520                 }
 521         }
 522
 523         if pdhOnly {
 524                 arvMountCmd = append(arvMountCmd, "--mount-by-pdh", "by_id")
 525         } else {
 526                 arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_id")
 527         }
 528         arvMountCmd = append(arvMountCmd, runner.ArvMountPoint)
 529
 530         runner.ArvMount, err = runner.RunArvMount(arvMountCmd, token)
 531         if err != nil {
 532                 return fmt.Errorf("While trying to start arv-mount: %v", err)
 533         }
 534
 535         for _, p := range collectionPaths {
 536                 _, err = os.Stat(p)
 537                 if err != nil {
 538                         return fmt.Errorf("While checking that input files exist: %v", err)
 539                 }
 540         }
 541
 542         return nil
 543 }
 544
 545 func (runner *ContainerRunner) ProcessDockerAttach(containerReader io.Reader) {
 546         // Handle docker log protocol
 547         // https://docs.docker.com/engine/reference/api/docker_remote_api_v1.15/#attach-to-a-container
 548         defer close(runner.loggingDone)
 549
 550         header := make([]byte, 8)
 551         var err error
 552         for err == nil {
 553                 _, err = io.ReadAtLeast(containerReader, header, 8)
 554                 if err != nil {
 555                         if err == io.EOF {
 556                                 err = nil
 557                         }
 558                         break
 559                 }
 560                 readsize := int64(header[7]) | (int64(header[6]) << 8) | (int64(header[5]) << 16) | (int64(header[4]) << 24)
 561                 if header[0] == 1 {
 562                         // stdout
 563                         _, err = io.CopyN(runner.Stdout, containerReader, readsize)
 564                 } else {
 565                         // stderr
 566                         _, err = io.CopyN(runner.Stderr, containerReader, readsize)
 567                 }
 568         }
 569
 570         if err != nil {
 571                 runner.CrunchLog.Printf("error reading docker logs: %v", err)
 572         }
 573
 574         err = runner.Stdout.Close()
 575         if err != nil {
 576                 runner.CrunchLog.Printf("error closing stdout logs: %v", err)
 577         }
 578
 579         err = runner.Stderr.Close()
 580         if err != nil {
 581                 runner.CrunchLog.Printf("error closing stderr logs: %v", err)
 582         }
 583
 584         if runner.statReporter != nil {
 585                 runner.statReporter.Stop()
 586                 err = runner.statLogger.Close()
 587                 if err != nil {
 588                         runner.CrunchLog.Printf("error closing crunchstat logs: %v", err)
 589                 }
 590         }
 591 }
 592
 593 func (runner *ContainerRunner) stopHoststat() error {
 594         if runner.hoststatReporter == nil {
 595                 return nil
 596         }
 597         runner.hoststatReporter.Stop()
 598         err := runner.hoststatLogger.Close()
 599         if err != nil {
 600                 return fmt.Errorf("error closing hoststat logs: %v", err)
 601         }
 602         return nil
 603 }
 604
 605 func (runner *ContainerRunner) startHoststat() {
 606         runner.hoststatLogger = NewThrottledLogger(runner.NewLogWriter("hoststat"))
 607         runner.hoststatReporter = &crunchstat.Reporter{
 608                 Logger:     log.New(runner.hoststatLogger, "", 0),
 609                 CgroupRoot: runner.cgroupRoot,
 610                 PollPeriod: runner.statInterval,
 611         }
 612         runner.hoststatReporter.Start()
 613 }
 614
 615 func (runner *ContainerRunner) startCrunchstat() {
 616         runner.statLogger = NewThrottledLogger(runner.NewLogWriter("crunchstat"))
 617         runner.statReporter = &crunchstat.Reporter{
 618                 CID:          runner.ContainerID,
 619                 Logger:       log.New(runner.statLogger, "", 0),
 620                 CgroupParent: runner.expectCgroupParent,
 621                 CgroupRoot:   runner.cgroupRoot,
 622                 PollPeriod:   runner.statInterval,
 623         }
 624         runner.statReporter.Start()
 625 }
 626
 627 type infoCommand struct {
 628         label string
 629         cmd   []string
 630 }
 631
 632 // LogHostInfo logs info about the current host, for debugging and
 633 // accounting purposes. Although it's logged as "node-info", this is
 634 // about the environment where crunch-run is actually running, which
 635 // might differ from what's described in the node record (see
 636 // LogNodeRecord).
 637 func (runner *ContainerRunner) LogHostInfo() (err error) {
 638         w := runner.NewLogWriter("node-info")
 639
 640         commands := []infoCommand{
 641                 {
 642                         label: "Host Information",
 643                         cmd:   []string{"uname", "-a"},
 644                 },
 645                 {
 646                         label: "CPU Information",
 647                         cmd:   []string{"cat", "/proc/cpuinfo"},
 648                 },
 649                 {
 650                         label: "Memory Information",
 651                         cmd:   []string{"cat", "/proc/meminfo"},
 652                 },
 653                 {
 654                         label: "Disk Space",
 655                         cmd:   []string{"df", "-m", "/", os.TempDir()},
 656                 },
 657                 {
 658                         label: "Disk INodes",
 659                         cmd:   []string{"df", "-i", "/", os.TempDir()},
 660                 },
 661         }
 662
 663         // Run commands with informational output to be logged.
 664         for _, command := range commands {
 665                 fmt.Fprintln(w, command.label)
 666                 cmd := exec.Command(command.cmd[0], command.cmd[1:]...)
 667                 cmd.Stdout = w
 668                 cmd.Stderr = w
 669                 if err := cmd.Run(); err != nil {
 670                         err = fmt.Errorf("While running command %q: %v", command.cmd, err)
 671                         fmt.Fprintln(w, err)
 672                         return err
 673                 }
 674                 fmt.Fprintln(w, "")
 675         }
 676
 677         err = w.Close()
 678         if err != nil {
 679                 return fmt.Errorf("While closing node-info logs: %v", err)
 680         }
 681         return nil
 682 }
 683
 684 // LogContainerRecord gets and saves the raw JSON container record from the API server
 685 func (runner *ContainerRunner) LogContainerRecord() error {
 686         logged, err := runner.logAPIResponse("container", "containers", map[string]interface{}{"filters": [][]string{{"uuid", "=", runner.Container.UUID}}}, nil)
 687         if !logged && err == nil {
 688                 err = fmt.Errorf("error: no container record found for %s", runner.Container.UUID)
 689         }
 690         return err
 691 }
 692
 693 // LogNodeRecord logs arvados#node record corresponding to the current host.
 694 func (runner *ContainerRunner) LogNodeRecord() error {
 695         hostname := os.Getenv("SLURMD_NODENAME")
 696         if hostname == "" {
 697                 hostname, _ = os.Hostname()
 698         }
 699         _, err := runner.logAPIResponse("node", "nodes", map[string]interface{}{"filters": [][]string{{"hostname", "=", hostname}}}, func(resp interface{}) {
 700                 // The "info" field has admin-only info when obtained
 701                 // with a privileged token, and should not be logged.
 702                 node, ok := resp.(map[string]interface{})
 703                 if ok {
 704                         delete(node, "info")
 705                 }
 706         })
 707         return err
 708 }
 709
 710 func (runner *ContainerRunner) logAPIResponse(label, path string, params map[string]interface{}, munge func(interface{})) (logged bool, err error) {
 711         w := &ArvLogWriter{
 712                 ArvClient:     runner.ArvClient,
 713                 UUID:          runner.Container.UUID,
 714                 loggingStream: label,
 715                 writeCloser:   runner.LogCollection.Open(label + ".json"),
 716         }
 717
 718         reader, err := runner.ArvClient.CallRaw("GET", path, "", "", arvadosclient.Dict(params))
 719         if err != nil {
 720                 return false, fmt.Errorf("error getting %s record: %v", label, err)
 721         }
 722         defer reader.Close()
 723
 724         dec := json.NewDecoder(reader)
 725         dec.UseNumber()
 726         var resp map[string]interface{}
 727         if err = dec.Decode(&resp); err != nil {
 728                 return false, fmt.Errorf("error decoding %s list response: %v", label, err)
 729         }
 730         items, ok := resp["items"].([]interface{})
 731         if !ok {
 732                 return false, fmt.Errorf("error decoding %s list response: no \"items\" key in API list response", label)
 733         } else if len(items) < 1 {
 734                 return false, nil
 735         }
 736         if munge != nil {
 737                 munge(items[0])
 738         }
 739         // Re-encode it using indentation to improve readability
 740         enc := json.NewEncoder(w)
 741         enc.SetIndent("", "    ")
 742         if err = enc.Encode(items[0]); err != nil {
 743                 return false, fmt.Errorf("error logging %s record: %v", label, err)
 744         }
 745         err = w.Close()
 746         if err != nil {
 747                 return false, fmt.Errorf("error closing %s.json in log collection: %v", label, err)
 748         }
 749         return true, nil
 750 }
 751
 752 // AttachStreams connects the docker container stdin, stdout and stderr logs
 753 // to the Arvados logger which logs to Keep and the API server logs table.
 754 func (runner *ContainerRunner) AttachStreams() (err error) {
 755
 756         runner.CrunchLog.Print("Attaching container streams")
 757
 758         // If stdin mount is provided, attach it to the docker container
 759         var stdinRdr arvados.File
 760         var stdinJson []byte
 761         if stdinMnt, ok := runner.Container.Mounts["stdin"]; ok {
 762                 if stdinMnt.Kind == "collection" {
 763                         var stdinColl arvados.Collection
 764                         collId := stdinMnt.UUID
 765                         if collId == "" {
 766                                 collId = stdinMnt.PortableDataHash
 767                         }
 768                         err = runner.ArvClient.Get("collections", collId, nil, &stdinColl)
 769                         if err != nil {
 770                                 return fmt.Errorf("While getting stding collection: %v", err)
 771                         }
 772
 773                         stdinRdr, err = runner.Kc.ManifestFileReader(manifest.Manifest{Text: stdinColl.ManifestText}, stdinMnt.Path)
 774                         if os.IsNotExist(err) {
 775                                 return fmt.Errorf("stdin collection path not found: %v", stdinMnt.Path)
 776                         } else if err != nil {
 777                                 return fmt.Errorf("While getting stdin collection path %v: %v", stdinMnt.Path, err)
 778                         }
 779                 } else if stdinMnt.Kind == "json" {
 780                         stdinJson, err = json.Marshal(stdinMnt.Content)
 781                         if err != nil {
 782                                 return fmt.Errorf("While encoding stdin json data: %v", err)
 783                         }
 784                 }
 785         }
 786
 787         stdinUsed := stdinRdr != nil || len(stdinJson) != 0
 788         response, err := runner.Docker.ContainerAttach(context.TODO(), runner.ContainerID,
 789                 dockertypes.ContainerAttachOptions{Stream: true, Stdin: stdinUsed, Stdout: true, Stderr: true})
 790         if err != nil {
 791                 return fmt.Errorf("While attaching container stdout/stderr streams: %v", err)
 792         }
 793
 794         runner.loggingDone = make(chan bool)
 795
 796         if stdoutMnt, ok := runner.Container.Mounts["stdout"]; ok {
 797                 stdoutFile, err := runner.getStdoutFile(stdoutMnt.Path)
 798                 if err != nil {
 799                         return err
 800                 }
 801                 runner.Stdout = stdoutFile
 802         } else {
 803                 runner.Stdout = NewThrottledLogger(runner.NewLogWriter("stdout"))
 804         }
 805
 806         if stderrMnt, ok := runner.Container.Mounts["stderr"]; ok {
 807                 stderrFile, err := runner.getStdoutFile(stderrMnt.Path)
 808                 if err != nil {
 809                         return err
 810                 }
 811                 runner.Stderr = stderrFile
 812         } else {
 813                 runner.Stderr = NewThrottledLogger(runner.NewLogWriter("stderr"))
 814         }
 815
 816         if stdinRdr != nil {
 817                 go func() {
 818                         _, err := io.Copy(response.Conn, stdinRdr)
 819                         if err != nil {
 820                                 runner.CrunchLog.Print("While writing stdin collection to docker container %q", err)
 821                                 runner.stop()
 822                         }
 823                         stdinRdr.Close()
 824                         response.CloseWrite()
 825                 }()
 826         } else if len(stdinJson) != 0 {
 827                 go func() {
 828                         _, err := io.Copy(response.Conn, bytes.NewReader(stdinJson))
 829                         if err != nil {
 830                                 runner.CrunchLog.Print("While writing stdin json to docker container %q", err)
 831                                 runner.stop()
 832                         }
 833                         response.CloseWrite()
 834                 }()
 835         }
 836
 837         go runner.ProcessDockerAttach(response.Reader)
 838
 839         return nil
 840 }
 841
 842 func (runner *ContainerRunner) getStdoutFile(mntPath string) (*os.File, error) {
 843         stdoutPath := mntPath[len(runner.Container.OutputPath):]
 844         index := strings.LastIndex(stdoutPath, "/")
 845         if index > 0 {
 846                 subdirs := stdoutPath[:index]
 847                 if subdirs != "" {
 848                         st, err := os.Stat(runner.HostOutputDir)
 849                         if err != nil {
 850                                 return nil, fmt.Errorf("While Stat on temp dir: %v", err)
 851                         }
 852                         stdoutPath := filepath.Join(runner.HostOutputDir, subdirs)
 853                         err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777)
 854                         if err != nil {
 855                                 return nil, fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err)
 856                         }
 857                 }
 858         }
 859         stdoutFile, err := os.Create(filepath.Join(runner.HostOutputDir, stdoutPath))
 860         if err != nil {
 861                 return nil, fmt.Errorf("While creating file %q: %v", stdoutPath, err)
 862         }
 863
 864         return stdoutFile, nil
 865 }
 866
 867 // CreateContainer creates the docker container.
 868 func (runner *ContainerRunner) CreateContainer() error {
 869         runner.CrunchLog.Print("Creating Docker container")
 870
 871         runner.ContainerConfig.Cmd = runner.Container.Command
 872         if runner.Container.Cwd != "." {
 873                 runner.ContainerConfig.WorkingDir = runner.Container.Cwd
 874         }
 875
 876         for k, v := range runner.Container.Environment {
 877                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env, k+"="+v)
 878         }
 879
 880         runner.ContainerConfig.Volumes = runner.Volumes
 881
 882         runner.HostConfig = dockercontainer.HostConfig{
 883                 Binds: runner.Binds,
 884                 LogConfig: dockercontainer.LogConfig{
 885                         Type: "none",
 886                 },
 887                 Resources: dockercontainer.Resources{
 888                         CgroupParent: runner.setCgroupParent,
 889                 },
 890         }
 891
 892         if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
 893                 tok, err := runner.ContainerToken()
 894                 if err != nil {
 895                         return err
 896                 }
 897                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env,
 898                         "ARVADOS_API_TOKEN="+tok,
 899                         "ARVADOS_API_HOST="+os.Getenv("ARVADOS_API_HOST"),
 900                         "ARVADOS_API_HOST_INSECURE="+os.Getenv("ARVADOS_API_HOST_INSECURE"),
 901                 )
 902                 runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
 903         } else {
 904                 if runner.enableNetwork == "always" {
 905                         runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
 906                 } else {
 907                         runner.HostConfig.NetworkMode = dockercontainer.NetworkMode("none")
 908                 }
 909         }
 910
 911         _, stdinUsed := runner.Container.Mounts["stdin"]
 912         runner.ContainerConfig.OpenStdin = stdinUsed
 913         runner.ContainerConfig.StdinOnce = stdinUsed
 914         runner.ContainerConfig.AttachStdin = stdinUsed
 915         runner.ContainerConfig.AttachStdout = true
 916         runner.ContainerConfig.AttachStderr = true
 917
 918         createdBody, err := runner.Docker.ContainerCreate(context.TODO(), &runner.ContainerConfig, &runner.HostConfig, nil, runner.Container.UUID)
 919         if err != nil {
 920                 return fmt.Errorf("While creating container: %v", err)
 921         }
 922
 923         runner.ContainerID = createdBody.ID
 924
 925         return runner.AttachStreams()
 926 }
 927
 928 // StartContainer starts the docker container created by CreateContainer.
 929 func (runner *ContainerRunner) StartContainer() error {
 930         runner.CrunchLog.Printf("Starting Docker container id '%s'", runner.ContainerID)
 931         runner.cStateLock.Lock()
 932         defer runner.cStateLock.Unlock()
 933         if runner.cCancelled {
 934                 return ErrCancelled
 935         }
 936         err := runner.Docker.ContainerStart(context.TODO(), runner.ContainerID,
 937                 dockertypes.ContainerStartOptions{})
 938         if err != nil {
 939                 var advice string
 940                 if m, e := regexp.MatchString("(?ms).*(exec|System error).*(no such file or directory|file not found).*", err.Error()); m && e == nil {
 941                         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])
 942                 }
 943                 return fmt.Errorf("could not start container: %v%s", err, advice)
 944         }
 945         return nil
 946 }
 947
 948 // WaitFinish waits for the container to terminate, capture the exit code, and
 949 // close the stdout/stderr logging.
 950 func (runner *ContainerRunner) WaitFinish() error {
 951         runner.CrunchLog.Print("Waiting for container to finish")
 952
 953         waitOk, waitErr := runner.Docker.ContainerWait(context.TODO(), runner.ContainerID, dockercontainer.WaitConditionNotRunning)
 954         arvMountExit := runner.ArvMountExit
 955         for {
 956                 select {
 957                 case waitBody := <-waitOk:
 958                         runner.CrunchLog.Printf("Container exited with code: %v", waitBody.StatusCode)
 959                         code := int(waitBody.StatusCode)
 960                         runner.ExitCode = &code
 961
 962                         // wait for stdout/stderr to complete
 963                         <-runner.loggingDone
 964                         return nil
 965
 966                 case err := <-waitErr:
 967                         return fmt.Errorf("container wait: %v", err)
 968
 969                 case <-arvMountExit:
 970                         runner.CrunchLog.Printf("arv-mount exited while container is still running.  Stopping container.")
 971                         runner.stop()
 972                         // arvMountExit will always be ready now that
 973                         // it's closed, but that doesn't interest us.
 974                         arvMountExit = nil
 975                 }
 976         }
 977 }
 978
 979 var ErrNotInOutputDir = fmt.Errorf("Must point to path within the output directory")
 980
 981 func (runner *ContainerRunner) derefOutputSymlink(path string, startinfo os.FileInfo) (tgt string, readlinktgt string, info os.FileInfo, err error) {
 982         // Follow symlinks if necessary
 983         info = startinfo
 984         tgt = path
 985         readlinktgt = ""
 986         nextlink := path
 987         for followed := 0; info.Mode()&os.ModeSymlink != 0; followed++ {
 988                 if followed >= limitFollowSymlinks {
 989                         // Got stuck in a loop or just a pathological number of links, give up.
 990                         err = fmt.Errorf("Followed more than %v symlinks from path %q", limitFollowSymlinks, path)
 991                         return
 992                 }
 993
 994                 readlinktgt, err = os.Readlink(nextlink)
 995                 if err != nil {
 996                         return
 997                 }
 998
 999                 tgt = readlinktgt
1000                 if !strings.HasPrefix(tgt, "/") {
1001                         // Relative symlink, resolve it to host path
1002                         tgt = filepath.Join(filepath.Dir(path), tgt)
1003                 }
1004                 if strings.HasPrefix(tgt, runner.Container.OutputPath+"/") && !strings.HasPrefix(tgt, runner.HostOutputDir+"/") {
1005                         // Absolute symlink to container output path, adjust it to host output path.
1006                         tgt = filepath.Join(runner.HostOutputDir, tgt[len(runner.Container.OutputPath):])
1007                 }
1008                 if !strings.HasPrefix(tgt, runner.HostOutputDir+"/") {
1009                         // After dereferencing, symlink target must either be
1010                         // within output directory, or must point to a
1011                         // collection mount.
1012                         err = ErrNotInOutputDir
1013                         return
1014                 }
1015
1016                 info, err = os.Lstat(tgt)
1017                 if err != nil {
1018                         // tgt
1019                         err = fmt.Errorf("Symlink in output %q points to invalid location %q: %v",
1020                                 path[len(runner.HostOutputDir):], readlinktgt, err)
1021                         return
1022                 }
1023
1024                 nextlink = tgt
1025         }
1026
1027         return
1028 }
1029
1030 var limitFollowSymlinks = 10
1031
1032 // UploadFile uploads files within the output directory, with special handling
1033 // for symlinks. If the symlink leads to a keep mount, copy the manifest text
1034 // from the keep mount into the output manifestText.  Ensure that whether
1035 // symlinks are relative or absolute, every symlink target (even targets that
1036 // are symlinks themselves) must point to a path in either the output directory
1037 // or a collection mount.
1038 //
1039 // Assumes initial value of "path" is absolute, and located within runner.HostOutputDir.
1040 func (runner *ContainerRunner) UploadOutputFile(
1041         path string,
1042         info os.FileInfo,
1043         infoerr error,
1044         binds []string,
1045         walkUpload *WalkUpload,
1046         relocateFrom string,
1047         relocateTo string,
1048         followed int) (manifestText string, err error) {
1049
1050         if infoerr != nil {
1051                 return "", infoerr
1052         }
1053
1054         if info.Mode().IsDir() {
1055                 // if empty, need to create a .keep file
1056                 dir, direrr := os.Open(path)
1057                 if direrr != nil {
1058                         return "", direrr
1059                 }
1060                 defer dir.Close()
1061                 names, eof := dir.Readdirnames(1)
1062                 if len(names) == 0 && eof == io.EOF && path != runner.HostOutputDir {
1063                         containerPath := runner.OutputPath + path[len(runner.HostOutputDir):]
1064                         for _, bind := range binds {
1065                                 mnt := runner.Container.Mounts[bind]
1066                                 // Check if there is a bind for this
1067                                 // directory, in which case assume we don't need .keep
1068                                 if (containerPath == bind || strings.HasPrefix(containerPath, bind+"/")) && mnt.PortableDataHash != "d41d8cd98f00b204e9800998ecf8427e+0" {
1069                                         return
1070                                 }
1071                         }
1072                         outputSuffix := path[len(runner.HostOutputDir)+1:]
1073                         return fmt.Sprintf("./%v d41d8cd98f00b204e9800998ecf8427e+0 0:0:.keep\n", outputSuffix), nil
1074                 }
1075                 return
1076         }
1077
1078         if followed >= limitFollowSymlinks {
1079                 // Got stuck in a loop or just a pathological number of
1080                 // directory links, give up.
1081                 err = fmt.Errorf("Followed more than %v symlinks from path %q", limitFollowSymlinks, path)
1082                 return
1083         }
1084
1085         // "path" is the actual path we are visiting
1086         // "tgt" is the target of "path" (a non-symlink) after following symlinks
1087         // "relocated" is the path in the output manifest where the file should be placed,
1088         // but has HostOutputDir as a prefix.
1089
1090         // The destination path in the output manifest may need to be
1091         // logically relocated to some other path in order to appear
1092         // in the correct location as a result of following a symlink.
1093         // Remove the relocateFrom prefix and replace it with
1094         // relocateTo.
1095         relocated := relocateTo + path[len(relocateFrom):]
1096
1097         tgt, readlinktgt, info, derefErr := runner.derefOutputSymlink(path, info)
1098         if derefErr != nil && derefErr != ErrNotInOutputDir {
1099                 return "", derefErr
1100         }
1101
1102         // go through mounts and try reverse map to collection reference
1103         for _, bind := range binds {
1104                 mnt := runner.Container.Mounts[bind]
1105                 if tgt == bind || strings.HasPrefix(tgt, bind+"/") {
1106                         // get path relative to bind
1107                         targetSuffix := tgt[len(bind):]
1108
1109                         // Copy mount and adjust the path to add path relative to the bind
1110                         adjustedMount := mnt
1111                         adjustedMount.Path = filepath.Join(adjustedMount.Path, targetSuffix)
1112
1113                         // Terminates in this keep mount, so add the
1114                         // manifest text at appropriate location.
1115                         outputSuffix := relocated[len(runner.HostOutputDir):]
1116                         manifestText, err = runner.getCollectionManifestForPath(adjustedMount, outputSuffix)
1117                         return
1118                 }
1119         }
1120
1121         // If target is not a collection mount, it must be located within the
1122         // output directory, otherwise it is an error.
1123         if derefErr == ErrNotInOutputDir {
1124                 err = fmt.Errorf("Symlink in output %q points to invalid location %q, must point to path within the output directory.",
1125                         path[len(runner.HostOutputDir):], readlinktgt)
1126                 return
1127         }
1128
1129         if info.Mode().IsRegular() {
1130                 return "", walkUpload.UploadFile(relocated, tgt)
1131         }
1132
1133         if info.Mode().IsDir() {
1134                 // Symlink leads to directory.  Walk() doesn't follow
1135                 // directory symlinks, so we walk the target directory
1136                 // instead.  Within the walk, file paths are relocated
1137                 // so they appear under the original symlink path.
1138                 err = filepath.Walk(tgt, func(walkpath string, walkinfo os.FileInfo, walkerr error) error {
1139                         var m string
1140                         m, walkerr = runner.UploadOutputFile(walkpath, walkinfo, walkerr,
1141                                 binds, walkUpload, tgt, relocated, followed+1)
1142                         if walkerr == nil {
1143                                 manifestText = manifestText + m
1144                         }
1145                         return walkerr
1146                 })
1147                 return
1148         }
1149
1150         return
1151 }
1152
1153 // HandleOutput sets the output, unmounts the FUSE mount, and deletes temporary directories
1154 func (runner *ContainerRunner) CaptureOutput() error {
1155         if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
1156                 // Output may have been set directly by the container, so
1157                 // refresh the container record to check.
1158                 err := runner.ArvClient.Get("containers", runner.Container.UUID,
1159                         nil, &runner.Container)
1160                 if err != nil {
1161                         return err
1162                 }
1163                 if runner.Container.Output != "" {
1164                         // Container output is already set.
1165                         runner.OutputPDH = &runner.Container.Output
1166                         return nil
1167                 }
1168         }
1169
1170         if runner.HostOutputDir == "" {
1171                 return nil
1172         }
1173
1174         _, err := os.Stat(runner.HostOutputDir)
1175         if err != nil {
1176                 return fmt.Errorf("While checking host output path: %v", err)
1177         }
1178
1179         // Pre-populate output from the configured mount points
1180         var binds []string
1181         for bind, mnt := range runner.Container.Mounts {
1182                 if mnt.Kind == "collection" {
1183                         binds = append(binds, bind)
1184                 }
1185         }
1186         sort.Strings(binds)
1187
1188         var manifestText string
1189
1190         collectionMetafile := fmt.Sprintf("%s/.arvados#collection", runner.HostOutputDir)
1191         _, err = os.Stat(collectionMetafile)
1192         if err != nil {
1193                 // Regular directory
1194
1195                 cw := CollectionWriter{0, runner.Kc, nil, nil, sync.Mutex{}}
1196                 walkUpload := cw.BeginUpload(runner.HostOutputDir, runner.CrunchLog.Logger)
1197
1198                 var m string
1199                 err = filepath.Walk(runner.HostOutputDir, func(path string, info os.FileInfo, err error) error {
1200                         m, err = runner.UploadOutputFile(path, info, err, binds, walkUpload, "", "", 0)
1201                         if err == nil {
1202                                 manifestText = manifestText + m
1203                         }
1204                         return err
1205                 })
1206
1207                 cw.EndUpload(walkUpload)
1208
1209                 if err != nil {
1210                         return fmt.Errorf("While uploading output files: %v", err)
1211                 }
1212
1213                 m, err = cw.ManifestText()
1214                 manifestText = manifestText + m
1215                 if err != nil {
1216                         return fmt.Errorf("While uploading output files: %v", err)
1217                 }
1218         } else {
1219                 // FUSE mount directory
1220                 file, openerr := os.Open(collectionMetafile)
1221                 if openerr != nil {
1222                         return fmt.Errorf("While opening FUSE metafile: %v", err)
1223                 }
1224                 defer file.Close()
1225
1226                 var rec arvados.Collection
1227                 err = json.NewDecoder(file).Decode(&rec)
1228                 if err != nil {
1229                         return fmt.Errorf("While reading FUSE metafile: %v", err)
1230                 }
1231                 manifestText = rec.ManifestText
1232         }
1233
1234         for _, bind := range binds {
1235                 mnt := runner.Container.Mounts[bind]
1236
1237                 bindSuffix := strings.TrimPrefix(bind, runner.Container.OutputPath)
1238
1239                 if bindSuffix == bind || len(bindSuffix) <= 0 {
1240                         // either does not start with OutputPath or is OutputPath itself
1241                         continue
1242                 }
1243
1244                 if mnt.ExcludeFromOutput == true {
1245                         continue
1246                 }
1247
1248                 // append to manifest_text
1249                 m, err := runner.getCollectionManifestForPath(mnt, bindSuffix)
1250                 if err != nil {
1251                         return err
1252                 }
1253
1254                 manifestText = manifestText + m
1255         }
1256
1257         // Save output
1258         var response arvados.Collection
1259         manifest := manifest.Manifest{Text: manifestText}
1260         manifestText = manifest.Extract(".", ".").Text
1261         err = runner.ArvClient.Create("collections",
1262                 arvadosclient.Dict{
1263                         "ensure_unique_name": true,
1264                         "collection": arvadosclient.Dict{
1265                                 "is_trashed":    true,
1266                                 "name":          "output for " + runner.Container.UUID,
1267                                 "manifest_text": manifestText}},
1268                 &response)
1269         if err != nil {
1270                 return fmt.Errorf("While creating output collection: %v", err)
1271         }
1272         runner.OutputPDH = &response.PortableDataHash
1273         return nil
1274 }
1275
1276 var outputCollections = make(map[string]arvados.Collection)
1277
1278 // Fetch the collection for the mnt.PortableDataHash
1279 // Return the manifest_text fragment corresponding to the specified mnt.Path
1280 //  after making any required updates.
1281 //  Ex:
1282 //    If mnt.Path is not specified,
1283 //      return the entire manifest_text after replacing any "." with bindSuffix
1284 //    If mnt.Path corresponds to one stream,
1285 //      return the manifest_text for that stream after replacing that stream name with bindSuffix
1286 //    Otherwise, check if a filename in any one stream is being sought. Return the manifest_text
1287 //      for that stream after replacing stream name with bindSuffix minus the last word
1288 //      and the file name with last word of the bindSuffix
1289 //  Allowed path examples:
1290 //    "path":"/"
1291 //    "path":"/subdir1"
1292 //    "path":"/subdir1/subdir2"
1293 //    "path":"/subdir/filename" etc
1294 func (runner *ContainerRunner) getCollectionManifestForPath(mnt arvados.Mount, bindSuffix string) (string, error) {
1295         collection := outputCollections[mnt.PortableDataHash]
1296         if collection.PortableDataHash == "" {
1297                 err := runner.ArvClient.Get("collections", mnt.PortableDataHash, nil, &collection)
1298                 if err != nil {
1299                         return "", fmt.Errorf("While getting collection for %v: %v", mnt.PortableDataHash, err)
1300                 }
1301                 outputCollections[mnt.PortableDataHash] = collection
1302         }
1303
1304         if collection.ManifestText == "" {
1305                 runner.CrunchLog.Printf("No manifest text for collection %v", collection.PortableDataHash)
1306                 return "", nil
1307         }
1308
1309         mft := manifest.Manifest{Text: collection.ManifestText}
1310         extracted := mft.Extract(mnt.Path, bindSuffix)
1311         if extracted.Err != nil {
1312                 return "", fmt.Errorf("Error parsing manifest for %v: %v", mnt.PortableDataHash, extracted.Err.Error())
1313         }
1314         return extracted.Text, nil
1315 }
1316
1317 func (runner *ContainerRunner) CleanupDirs() {
1318         if runner.ArvMount != nil {
1319                 var delay int64 = 8
1320                 umount := exec.Command("arv-mount", fmt.Sprintf("--unmount-timeout=%d", delay), "--unmount", runner.ArvMountPoint)
1321                 umount.Stdout = runner.CrunchLog
1322                 umount.Stderr = runner.CrunchLog
1323                 runner.CrunchLog.Printf("Running %v", umount.Args)
1324                 umnterr := umount.Start()
1325
1326                 if umnterr != nil {
1327                         runner.CrunchLog.Printf("Error unmounting: %v", umnterr)
1328                 } else {
1329                         // If arv-mount --unmount gets stuck for any reason, we
1330                         // don't want to wait for it forever.  Do Wait() in a goroutine
1331                         // so it doesn't block crunch-run.
1332                         umountExit := make(chan error)
1333                         go func() {
1334                                 mnterr := umount.Wait()
1335                                 if mnterr != nil {
1336                                         runner.CrunchLog.Printf("Error unmounting: %v", mnterr)
1337                                 }
1338                                 umountExit <- mnterr
1339                         }()
1340
1341                         for again := true; again; {
1342                                 again = false
1343                                 select {
1344                                 case <-umountExit:
1345                                         umount = nil
1346                                         again = true
1347                                 case <-runner.ArvMountExit:
1348                                         break
1349                                 case <-time.After(time.Duration((delay + 1) * int64(time.Second))):
1350                                         runner.CrunchLog.Printf("Timed out waiting for unmount")
1351                                         if umount != nil {
1352                                                 umount.Process.Kill()
1353                                         }
1354                                         runner.ArvMount.Process.Kill()
1355                                 }
1356                         }
1357                 }
1358         }
1359
1360         if runner.ArvMountPoint != "" {
1361                 if rmerr := os.Remove(runner.ArvMountPoint); rmerr != nil {
1362                         runner.CrunchLog.Printf("While cleaning up arv-mount directory %s: %v", runner.ArvMountPoint, rmerr)
1363                 }
1364         }
1365
1366         for _, tmpdir := range runner.CleanupTempDir {
1367                 if rmerr := os.RemoveAll(tmpdir); rmerr != nil {
1368                         runner.CrunchLog.Printf("While cleaning up temporary directory %s: %v", tmpdir, rmerr)
1369                 }
1370         }
1371 }
1372
1373 // CommitLogs posts the collection containing the final container logs.
1374 func (runner *ContainerRunner) CommitLogs() error {
1375         runner.CrunchLog.Print(runner.finalState)
1376
1377         if runner.arvMountLog != nil {
1378                 runner.arvMountLog.Close()
1379         }
1380         runner.CrunchLog.Close()
1381
1382         // Closing CrunchLog above allows them to be committed to Keep at this
1383         // point, but re-open crunch log with ArvClient in case there are any
1384         // other further errors (such as failing to write the log to Keep!)
1385         // while shutting down
1386         runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{ArvClient: runner.ArvClient,
1387                 UUID: runner.Container.UUID, loggingStream: "crunch-run", writeCloser: nil})
1388         runner.CrunchLog.Immediate = log.New(os.Stderr, runner.Container.UUID+" ", 0)
1389
1390         if runner.LogsPDH != nil {
1391                 // If we have already assigned something to LogsPDH,
1392                 // we must be closing the re-opened log, which won't
1393                 // end up getting attached to the container record and
1394                 // therefore doesn't need to be saved as a collection
1395                 // -- it exists only to send logs to other channels.
1396                 return nil
1397         }
1398
1399         mt, err := runner.LogCollection.ManifestText()
1400         if err != nil {
1401                 return fmt.Errorf("While creating log manifest: %v", err)
1402         }
1403
1404         var response arvados.Collection
1405         err = runner.ArvClient.Create("collections",
1406                 arvadosclient.Dict{
1407                         "ensure_unique_name": true,
1408                         "collection": arvadosclient.Dict{
1409                                 "is_trashed":    true,
1410                                 "name":          "logs for " + runner.Container.UUID,
1411                                 "manifest_text": mt}},
1412                 &response)
1413         if err != nil {
1414                 return fmt.Errorf("While creating log collection: %v", err)
1415         }
1416         runner.LogsPDH = &response.PortableDataHash
1417         return nil
1418 }
1419
1420 // UpdateContainerRunning updates the container state to "Running"
1421 func (runner *ContainerRunner) UpdateContainerRunning() error {
1422         runner.cStateLock.Lock()
1423         defer runner.cStateLock.Unlock()
1424         if runner.cCancelled {
1425                 return ErrCancelled
1426         }
1427         return runner.ArvClient.Update("containers", runner.Container.UUID,
1428                 arvadosclient.Dict{"container": arvadosclient.Dict{"state": "Running"}}, nil)
1429 }
1430
1431 // ContainerToken returns the api_token the container (and any
1432 // arv-mount processes) are allowed to use.
1433 func (runner *ContainerRunner) ContainerToken() (string, error) {
1434         if runner.token != "" {
1435                 return runner.token, nil
1436         }
1437
1438         var auth arvados.APIClientAuthorization
1439         err := runner.ArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth)
1440         if err != nil {
1441                 return "", err
1442         }
1443         runner.token = auth.APIToken
1444         return runner.token, nil
1445 }
1446
1447 // UpdateContainerComplete updates the container record state on API
1448 // server to "Complete" or "Cancelled"
1449 func (runner *ContainerRunner) UpdateContainerFinal() error {
1450         update := arvadosclient.Dict{}
1451         update["state"] = runner.finalState
1452         if runner.LogsPDH != nil {
1453                 update["log"] = *runner.LogsPDH
1454         }
1455         if runner.finalState == "Complete" {
1456                 if runner.ExitCode != nil {
1457                         update["exit_code"] = *runner.ExitCode
1458                 }
1459                 if runner.OutputPDH != nil {
1460                         update["output"] = *runner.OutputPDH
1461                 }
1462         }
1463         return runner.ArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{"container": update}, nil)
1464 }
1465
1466 // IsCancelled returns the value of Cancelled, with goroutine safety.
1467 func (runner *ContainerRunner) IsCancelled() bool {
1468         runner.cStateLock.Lock()
1469         defer runner.cStateLock.Unlock()
1470         return runner.cCancelled
1471 }
1472
1473 // NewArvLogWriter creates an ArvLogWriter
1474 func (runner *ContainerRunner) NewArvLogWriter(name string) io.WriteCloser {
1475         return &ArvLogWriter{
1476                 ArvClient:     runner.ArvClient,
1477                 UUID:          runner.Container.UUID,
1478                 loggingStream: name,
1479                 writeCloser:   runner.LogCollection.Open(name + ".txt")}
1480 }
1481
1482 // Run the full container lifecycle.
1483 func (runner *ContainerRunner) Run() (err error) {
1484         runner.CrunchLog.Printf("crunch-run %s started", version)
1485         runner.CrunchLog.Printf("Executing container '%s'", runner.Container.UUID)
1486
1487         hostname, hosterr := os.Hostname()
1488         if hosterr != nil {
1489                 runner.CrunchLog.Printf("Error getting hostname '%v'", hosterr)
1490         } else {
1491                 runner.CrunchLog.Printf("Executing on host '%s'", hostname)
1492         }
1493
1494         runner.finalState = "Queued"
1495
1496         defer func() {
1497                 runner.stopSignals()
1498                 runner.CleanupDirs()
1499
1500                 runner.CrunchLog.Printf("crunch-run finished")
1501                 runner.CrunchLog.Close()
1502         }()
1503
1504         defer func() {
1505                 // checkErr prints e (unless it's nil) and sets err to
1506                 // e (unless err is already non-nil). Thus, if err
1507                 // hasn't already been assigned when Run() returns,
1508                 // this cleanup func will cause Run() to return the
1509                 // first non-nil error that is passed to checkErr().
1510                 checkErr := func(e error) {
1511                         if e == nil {
1512                                 return
1513                         }
1514                         runner.CrunchLog.Print(e)
1515                         if err == nil {
1516                                 err = e
1517                         }
1518                         if runner.finalState == "Complete" {
1519                                 // There was an error in the finalization.
1520                                 runner.finalState = "Cancelled"
1521                         }
1522                 }
1523
1524                 // Log the error encountered in Run(), if any
1525                 checkErr(err)
1526
1527                 if runner.finalState == "Queued" {
1528                         runner.UpdateContainerFinal()
1529                         return
1530                 }
1531
1532                 if runner.IsCancelled() {
1533                         runner.finalState = "Cancelled"
1534                         // but don't return yet -- we still want to
1535                         // capture partial output and write logs
1536                 }
1537
1538                 checkErr(runner.CaptureOutput())
1539                 checkErr(runner.stopHoststat())
1540                 checkErr(runner.CommitLogs())
1541                 checkErr(runner.UpdateContainerFinal())
1542         }()
1543
1544         err = runner.fetchContainerRecord()
1545         if err != nil {
1546                 return
1547         }
1548         runner.setupSignals()
1549         runner.startHoststat()
1550
1551         // check for and/or load image
1552         err = runner.LoadImage()
1553         if err != nil {
1554                 if !runner.checkBrokenNode(err) {
1555                         // Failed to load image but not due to a "broken node"
1556                         // condition, probably user error.
1557                         runner.finalState = "Cancelled"
1558                 }
1559                 err = fmt.Errorf("While loading container image: %v", err)
1560                 return
1561         }
1562
1563         // set up FUSE mount and binds
1564         err = runner.SetupMounts()
1565         if err != nil {
1566                 runner.finalState = "Cancelled"
1567                 err = fmt.Errorf("While setting up mounts: %v", err)
1568                 return
1569         }
1570
1571         err = runner.CreateContainer()
1572         if err != nil {
1573                 return
1574         }
1575         err = runner.LogHostInfo()
1576         if err != nil {
1577                 return
1578         }
1579         err = runner.LogNodeRecord()
1580         if err != nil {
1581                 return
1582         }
1583         err = runner.LogContainerRecord()
1584         if err != nil {
1585                 return
1586         }
1587
1588         if runner.IsCancelled() {
1589                 return
1590         }
1591
1592         err = runner.UpdateContainerRunning()
1593         if err != nil {
1594                 return
1595         }
1596         runner.finalState = "Cancelled"
1597
1598         runner.startCrunchstat()
1599
1600         err = runner.StartContainer()
1601         if err != nil {
1602                 runner.checkBrokenNode(err)
1603                 return
1604         }
1605
1606         err = runner.WaitFinish()
1607         if err == nil && !runner.IsCancelled() {
1608                 runner.finalState = "Complete"
1609         }
1610         return
1611 }
1612
1613 // Fetch the current container record (uuid = runner.Container.UUID)
1614 // into runner.Container.
1615 func (runner *ContainerRunner) fetchContainerRecord() error {
1616         reader, err := runner.ArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
1617         if err != nil {
1618                 return fmt.Errorf("error fetching container record: %v", err)
1619         }
1620         defer reader.Close()
1621
1622         dec := json.NewDecoder(reader)
1623         dec.UseNumber()
1624         err = dec.Decode(&runner.Container)
1625         if err != nil {
1626                 return fmt.Errorf("error decoding container record: %v", err)
1627         }
1628         return nil
1629 }
1630
1631 // NewContainerRunner creates a new container runner.
1632 func NewContainerRunner(api IArvadosClient,
1633         kc IKeepClient,
1634         docker ThinDockerClient,
1635         containerUUID string) *ContainerRunner {
1636
1637         cr := &ContainerRunner{ArvClient: api, Kc: kc, Docker: docker}
1638         cr.NewLogWriter = cr.NewArvLogWriter
1639         cr.RunArvMount = cr.ArvMountCmd
1640         cr.MkTempDir = ioutil.TempDir
1641         cr.LogCollection = &CollectionWriter{0, kc, nil, nil, sync.Mutex{}}
1642         cr.Container.UUID = containerUUID
1643         cr.CrunchLog = NewThrottledLogger(cr.NewLogWriter("crunch-run"))
1644         cr.CrunchLog.Immediate = log.New(os.Stderr, containerUUID+" ", 0)
1645
1646         loadLogThrottleParams(api)
1647
1648         return cr
1649 }
1650
1651 func main() {
1652         statInterval := flag.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting")
1653         cgroupRoot := flag.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree")
1654         cgroupParent := flag.String("cgroup-parent", "docker", "name of container's parent cgroup (ignored if -cgroup-parent-subsystem is used)")
1655         cgroupParentSubsystem := flag.String("cgroup-parent-subsystem", "", "use current cgroup for given subsystem as parent cgroup for container")
1656         caCertsPath := flag.String("ca-certs", "", "Path to TLS root certificates")
1657         enableNetwork := flag.String("container-enable-networking", "default",
1658                 `Specify if networking should be enabled for container.  One of 'default', 'always':
1659         default: only enable networking if container requests it.
1660         always:  containers always have networking enabled
1661         `)
1662         networkMode := flag.String("container-network-mode", "default",
1663                 `Set networking mode for container.  Corresponds to Docker network mode (--net).
1664         `)
1665         memprofile := flag.String("memprofile", "", "write memory profile to `file` after running container")
1666         getVersion := flag.Bool("version", false, "Print version information and exit.")
1667         flag.Parse()
1668
1669         // Print version information if requested
1670         if *getVersion {
1671                 fmt.Printf("crunch-run %s\n", version)
1672                 return
1673         }
1674
1675         log.Printf("crunch-run %s started", version)
1676
1677         containerId := flag.Arg(0)
1678
1679         if *caCertsPath != "" {
1680                 arvadosclient.CertFiles = []string{*caCertsPath}
1681         }
1682
1683         api, err := arvadosclient.MakeArvadosClient()
1684         if err != nil {
1685                 log.Fatalf("%s: %v", containerId, err)
1686         }
1687         api.Retries = 8
1688
1689         kc, kcerr := keepclient.MakeKeepClient(api)
1690         if kcerr != nil {
1691                 log.Fatalf("%s: %v", containerId, kcerr)
1692         }
1693         kc.BlockCache = &keepclient.BlockCache{MaxBlocks: 2}
1694         kc.Retries = 4
1695
1696         // API version 1.21 corresponds to Docker 1.9, which is currently the
1697         // minimum version we want to support.
1698         docker, dockererr := dockerclient.NewClient(dockerclient.DefaultDockerHost, "1.21", nil, nil)
1699
1700         cr := NewContainerRunner(api, kc, docker, containerId)
1701         if dockererr != nil {
1702                 cr.CrunchLog.Printf("%s: %v", containerId, dockererr)
1703                 cr.checkBrokenNode(dockererr)
1704                 cr.CrunchLog.Close()
1705                 os.Exit(1)
1706         }
1707
1708         cr.statInterval = *statInterval
1709         cr.cgroupRoot = *cgroupRoot
1710         cr.expectCgroupParent = *cgroupParent
1711         cr.enableNetwork = *enableNetwork
1712         cr.networkMode = *networkMode
1713         if *cgroupParentSubsystem != "" {
1714                 p := findCgroup(*cgroupParentSubsystem)
1715                 cr.setCgroupParent = p
1716                 cr.expectCgroupParent = p
1717         }
1718
1719         runerr := cr.Run()
1720
1721         if *memprofile != "" {
1722                 f, err := os.Create(*memprofile)
1723                 if err != nil {
1724                         log.Printf("could not create memory profile: ", err)
1725                 }
1726                 runtime.GC() // get up-to-date statistics
1727                 if err := pprof.WriteHeapProfile(f); err != nil {
1728                         log.Printf("could not write memory profile: ", err)
1729                 }
1730                 closeerr := f.Close()
1731                 if closeerr != nil {
1732                         log.Printf("closing memprofile file: ", err)
1733                 }
1734         }
1735
1736         if runerr != nil {
1737                 log.Fatalf("%s: %v", containerId, runerr)
1738         }
1739 }