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         statInterval time.Duration
 120         cgroupRoot   string
 121         // What we expect the container's cgroup parent to be.
 122         expectCgroupParent string
 123         // What we tell docker to use as the container's cgroup
 124         // parent. Note: Ideally we would use the same field for both
 125         // expectCgroupParent and setCgroupParent, and just make it
 126         // default to "docker". However, when using docker < 1.10 with
 127         // systemd, specifying a non-empty cgroup parent (even the
 128         // default value "docker") hits a docker bug
 129         // (https://github.com/docker/docker/issues/17126). Using two
 130         // separate fields makes it possible to use the "expect cgroup
 131         // parent to be X" feature even on sites where the "specify
 132         // cgroup parent" feature breaks.
 133         setCgroupParent string
 134
 135         cStateLock sync.Mutex
 136         cCancelled bool // StopContainer() invoked
 137
 138         enableNetwork string // one of "default" or "always"
 139         networkMode   string // passed through to HostConfig.NetworkMode
 140         arvMountLog   *ThrottledLogger
 141 }
 142
 143 // setupSignals sets up signal handling to gracefully terminate the underlying
 144 // Docker container and update state when receiving a TERM, INT or QUIT signal.
 145 func (runner *ContainerRunner) setupSignals() {
 146         runner.SigChan = make(chan os.Signal, 1)
 147         signal.Notify(runner.SigChan, syscall.SIGTERM)
 148         signal.Notify(runner.SigChan, syscall.SIGINT)
 149         signal.Notify(runner.SigChan, syscall.SIGQUIT)
 150
 151         go func(sig chan os.Signal) {
 152                 for s := range sig {
 153                         runner.CrunchLog.Printf("caught signal: %v", s)
 154                         runner.stop()
 155                 }
 156         }(runner.SigChan)
 157 }
 158
 159 // stop the underlying Docker container.
 160 func (runner *ContainerRunner) stop() {
 161         runner.cStateLock.Lock()
 162         defer runner.cStateLock.Unlock()
 163         if runner.ContainerID == "" {
 164                 return
 165         }
 166         runner.cCancelled = true
 167         runner.CrunchLog.Printf("removing container")
 168         err := runner.Docker.ContainerRemove(context.TODO(), runner.ContainerID, dockertypes.ContainerRemoveOptions{Force: true})
 169         if err != nil {
 170                 runner.CrunchLog.Printf("error removing container: %s", err)
 171         }
 172 }
 173
 174 func (runner *ContainerRunner) stopSignals() {
 175         if runner.SigChan != nil {
 176                 signal.Stop(runner.SigChan)
 177         }
 178 }
 179
 180 var errorBlacklist = []string{
 181         "(?ms).*[Cc]annot connect to the Docker daemon.*",
 182         "(?ms).*oci runtime error.*starting container process.*container init.*mounting.*to rootfs.*no such file or directory.*",
 183 }
 184 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)")
 185
 186 func (runner *ContainerRunner) checkBrokenNode(goterr error) bool {
 187         for _, d := range errorBlacklist {
 188                 if m, e := regexp.MatchString(d, goterr.Error()); m && e == nil {
 189                         runner.CrunchLog.Printf("Error suggests node is unable to run containers: %v", goterr)
 190                         if *brokenNodeHook == "" {
 191                                 runner.CrunchLog.Printf("No broken node hook provided, cannot mark node as broken.")
 192                         } else {
 193                                 runner.CrunchLog.Printf("Running broken node hook %q", *brokenNodeHook)
 194                                 // run killme script
 195                                 c := exec.Command(*brokenNodeHook)
 196                                 c.Stdout = runner.CrunchLog
 197                                 c.Stderr = runner.CrunchLog
 198                                 err := c.Run()
 199                                 if err != nil {
 200                                         runner.CrunchLog.Printf("Error running broken node hook: %v", err)
 201                                 }
 202                         }
 203                         return true
 204                 }
 205         }
 206         return false
 207 }
 208
 209 // LoadImage determines the docker image id from the container record and
 210 // checks if it is available in the local Docker image store.  If not, it loads
 211 // the image from Keep.
 212 func (runner *ContainerRunner) LoadImage() (err error) {
 213
 214         runner.CrunchLog.Printf("Fetching Docker image from collection '%s'", runner.Container.ContainerImage)
 215
 216         var collection arvados.Collection
 217         err = runner.ArvClient.Get("collections", runner.Container.ContainerImage, nil, &collection)
 218         if err != nil {
 219                 return fmt.Errorf("While getting container image collection: %v", err)
 220         }
 221         manifest := manifest.Manifest{Text: collection.ManifestText}
 222         var img, imageID string
 223         for ms := range manifest.StreamIter() {
 224                 img = ms.FileStreamSegments[0].Name
 225                 if !strings.HasSuffix(img, ".tar") {
 226                         return fmt.Errorf("First file in the container image collection does not end in .tar")
 227                 }
 228                 imageID = img[:len(img)-4]
 229         }
 230
 231         runner.CrunchLog.Printf("Using Docker image id '%s'", imageID)
 232
 233         _, _, err = runner.Docker.ImageInspectWithRaw(context.TODO(), imageID)
 234         if err != nil {
 235                 runner.CrunchLog.Print("Loading Docker image from keep")
 236
 237                 var readCloser io.ReadCloser
 238                 readCloser, err = runner.Kc.ManifestFileReader(manifest, img)
 239                 if err != nil {
 240                         return fmt.Errorf("While creating ManifestFileReader for container image: %v", err)
 241                 }
 242
 243                 response, err := runner.Docker.ImageLoad(context.TODO(), readCloser, true)
 244                 if err != nil {
 245                         return fmt.Errorf("While loading container image into Docker: %v", err)
 246                 }
 247
 248                 defer response.Body.Close()
 249                 rbody, err := ioutil.ReadAll(response.Body)
 250                 if err != nil {
 251                         return fmt.Errorf("Reading response to image load: %v", err)
 252                 }
 253                 runner.CrunchLog.Printf("Docker response: %s", rbody)
 254         } else {
 255                 runner.CrunchLog.Print("Docker image is available")
 256         }
 257
 258         runner.ContainerConfig.Image = imageID
 259
 260         runner.Kc.ClearBlockCache()
 261
 262         return nil
 263 }
 264
 265 func (runner *ContainerRunner) ArvMountCmd(arvMountCmd []string, token string) (c *exec.Cmd, err error) {
 266         c = exec.Command("arv-mount", arvMountCmd...)
 267
 268         // Copy our environment, but override ARVADOS_API_TOKEN with
 269         // the container auth token.
 270         c.Env = nil
 271         for _, s := range os.Environ() {
 272                 if !strings.HasPrefix(s, "ARVADOS_API_TOKEN=") {
 273                         c.Env = append(c.Env, s)
 274                 }
 275         }
 276         c.Env = append(c.Env, "ARVADOS_API_TOKEN="+token)
 277
 278         runner.arvMountLog = NewThrottledLogger(runner.NewLogWriter("arv-mount"))
 279         c.Stdout = runner.arvMountLog
 280         c.Stderr = runner.arvMountLog
 281
 282         runner.CrunchLog.Printf("Running %v", c.Args)
 283
 284         err = c.Start()
 285         if err != nil {
 286                 return nil, err
 287         }
 288
 289         statReadme := make(chan bool)
 290         runner.ArvMountExit = make(chan error)
 291
 292         keepStatting := true
 293         go func() {
 294                 for keepStatting {
 295                         time.Sleep(100 * time.Millisecond)
 296                         _, err = os.Stat(fmt.Sprintf("%s/by_id/README", runner.ArvMountPoint))
 297                         if err == nil {
 298                                 keepStatting = false
 299                                 statReadme <- true
 300                         }
 301                 }
 302                 close(statReadme)
 303         }()
 304
 305         go func() {
 306                 mnterr := c.Wait()
 307                 if mnterr != nil {
 308                         runner.CrunchLog.Printf("Arv-mount exit error: %v", mnterr)
 309                 }
 310                 runner.ArvMountExit <- mnterr
 311                 close(runner.ArvMountExit)
 312         }()
 313
 314         select {
 315         case <-statReadme:
 316                 break
 317         case err := <-runner.ArvMountExit:
 318                 runner.ArvMount = nil
 319                 keepStatting = false
 320                 return nil, err
 321         }
 322
 323         return c, nil
 324 }
 325
 326 func (runner *ContainerRunner) SetupArvMountPoint(prefix string) (err error) {
 327         if runner.ArvMountPoint == "" {
 328                 runner.ArvMountPoint, err = runner.MkTempDir("", prefix)
 329         }
 330         return
 331 }
 332
 333 func copyfile(src string, dst string) (err error) {
 334         srcfile, err := os.Open(src)
 335         if err != nil {
 336                 return
 337         }
 338
 339         os.MkdirAll(path.Dir(dst), 0770)
 340
 341         dstfile, err := os.Create(dst)
 342         if err != nil {
 343                 return
 344         }
 345         _, err = io.Copy(dstfile, srcfile)
 346         if err != nil {
 347                 return
 348         }
 349
 350         err = srcfile.Close()
 351         err2 := dstfile.Close()
 352
 353         if err != nil {
 354                 return
 355         }
 356
 357         if err2 != nil {
 358                 return err2
 359         }
 360
 361         return nil
 362 }
 363
 364 func (runner *ContainerRunner) SetupMounts() (err error) {
 365         err = runner.SetupArvMountPoint("keep")
 366         if err != nil {
 367                 return fmt.Errorf("While creating keep mount temp dir: %v", err)
 368         }
 369
 370         token, err := runner.ContainerToken()
 371         if err != nil {
 372                 return fmt.Errorf("could not get container token: %s", err)
 373         }
 374
 375         pdhOnly := true
 376         tmpcount := 0
 377         arvMountCmd := []string{
 378                 "--foreground",
 379                 "--allow-other",
 380                 "--read-write",
 381                 fmt.Sprintf("--crunchstat-interval=%v", runner.statInterval.Seconds())}
 382
 383         if runner.Container.RuntimeConstraints.KeepCacheRAM > 0 {
 384                 arvMountCmd = append(arvMountCmd, "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheRAM))
 385         }
 386
 387         collectionPaths := []string{}
 388         runner.Binds = nil
 389         runner.Volumes = make(map[string]struct{})
 390         needCertMount := true
 391         type copyFile struct {
 392                 src  string
 393                 bind string
 394         }
 395         var copyFiles []copyFile
 396
 397         var binds []string
 398         for bind := range runner.Container.Mounts {
 399                 binds = append(binds, bind)
 400         }
 401         sort.Strings(binds)
 402
 403         for _, bind := range binds {
 404                 mnt := runner.Container.Mounts[bind]
 405                 if bind == "stdout" || bind == "stderr" {
 406                         // Is it a "file" mount kind?
 407                         if mnt.Kind != "file" {
 408                                 return fmt.Errorf("Unsupported mount kind '%s' for %s. Only 'file' is supported.", mnt.Kind, bind)
 409                         }
 410
 411                         // Does path start with OutputPath?
 412                         prefix := runner.Container.OutputPath
 413                         if !strings.HasSuffix(prefix, "/") {
 414                                 prefix += "/"
 415                         }
 416                         if !strings.HasPrefix(mnt.Path, prefix) {
 417                                 return fmt.Errorf("%s path does not start with OutputPath: %s, %s", strings.Title(bind), mnt.Path, prefix)
 418                         }
 419                 }
 420
 421                 if bind == "stdin" {
 422                         // Is it a "collection" mount kind?
 423                         if mnt.Kind != "collection" && mnt.Kind != "json" {
 424                                 return fmt.Errorf("Unsupported mount kind '%s' for stdin. Only 'collection' or 'json' are supported.", mnt.Kind)
 425                         }
 426                 }
 427
 428                 if bind == "/etc/arvados/ca-certificates.crt" {
 429                         needCertMount = false
 430                 }
 431
 432                 if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && bind != runner.Container.OutputPath+"/" {
 433                         if mnt.Kind != "collection" {
 434                                 return fmt.Errorf("Only mount points of kind 'collection' are supported underneath the output_path: %v", bind)
 435                         }
 436                 }
 437
 438                 switch {
 439                 case mnt.Kind == "collection" && bind != "stdin":
 440                         var src string
 441                         if mnt.UUID != "" && mnt.PortableDataHash != "" {
 442                                 return fmt.Errorf("Cannot specify both 'uuid' and 'portable_data_hash' for a collection mount")
 443                         }
 444                         if mnt.UUID != "" {
 445                                 if mnt.Writable {
 446                                         return fmt.Errorf("Writing to existing collections currently not permitted.")
 447                                 }
 448                                 pdhOnly = false
 449                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.UUID)
 450                         } else if mnt.PortableDataHash != "" {
 451                                 if mnt.Writable && !strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 452                                         return fmt.Errorf("Can never write to a collection specified by portable data hash")
 453                                 }
 454                                 idx := strings.Index(mnt.PortableDataHash, "/")
 455                                 if idx > 0 {
 456                                         mnt.Path = path.Clean(mnt.PortableDataHash[idx:])
 457                                         mnt.PortableDataHash = mnt.PortableDataHash[0:idx]
 458                                         runner.Container.Mounts[bind] = mnt
 459                                 }
 460                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.PortableDataHash)
 461                                 if mnt.Path != "" && mnt.Path != "." {
 462                                         if strings.HasPrefix(mnt.Path, "./") {
 463                                                 mnt.Path = mnt.Path[2:]
 464                                         } else if strings.HasPrefix(mnt.Path, "/") {
 465                                                 mnt.Path = mnt.Path[1:]
 466                                         }
 467                                         src += "/" + mnt.Path
 468                                 }
 469                         } else {
 470                                 src = fmt.Sprintf("%s/tmp%d", runner.ArvMountPoint, tmpcount)
 471                                 arvMountCmd = append(arvMountCmd, "--mount-tmp")
 472                                 arvMountCmd = append(arvMountCmd, fmt.Sprintf("tmp%d", tmpcount))
 473                                 tmpcount += 1
 474                         }
 475                         if mnt.Writable {
 476                                 if bind == runner.Container.OutputPath {
 477                                         runner.HostOutputDir = src
 478                                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", src, bind))
 479                                 } else if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 480                                         copyFiles = append(copyFiles, copyFile{src, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
 481                                 } else {
 482                                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", src, bind))
 483                                 }
 484                         } else {
 485                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", src, bind))
 486                         }
 487                         collectionPaths = append(collectionPaths, src)
 488
 489                 case mnt.Kind == "tmp":
 490                         var tmpdir string
 491                         tmpdir, err = runner.MkTempDir("", "")
 492                         if err != nil {
 493                                 return fmt.Errorf("While creating mount temp dir: %v", err)
 494                         }
 495                         st, staterr := os.Stat(tmpdir)
 496                         if staterr != nil {
 497                                 return fmt.Errorf("While Stat on temp dir: %v", staterr)
 498                         }
 499                         err = os.Chmod(tmpdir, st.Mode()|os.ModeSetgid|0777)
 500                         if staterr != nil {
 501                                 return fmt.Errorf("While Chmod temp dir: %v", err)
 502                         }
 503                         runner.CleanupTempDir = append(runner.CleanupTempDir, tmpdir)
 504                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", tmpdir, bind))
 505                         if bind == runner.Container.OutputPath {
 506                                 runner.HostOutputDir = tmpdir
 507                         }
 508
 509                 case mnt.Kind == "json":
 510                         jsondata, err := json.Marshal(mnt.Content)
 511                         if err != nil {
 512                                 return fmt.Errorf("encoding json data: %v", err)
 513                         }
 514                         // Create a tempdir with a single file
 515                         // (instead of just a tempfile): this way we
 516                         // can ensure the file is world-readable
 517                         // inside the container, without having to
 518                         // make it world-readable on the docker host.
 519                         tmpdir, err := runner.MkTempDir("", "")
 520                         if err != nil {
 521                                 return fmt.Errorf("creating temp dir: %v", err)
 522                         }
 523                         runner.CleanupTempDir = append(runner.CleanupTempDir, tmpdir)
 524                         tmpfn := filepath.Join(tmpdir, "mountdata.json")
 525                         err = ioutil.WriteFile(tmpfn, jsondata, 0644)
 526                         if err != nil {
 527                                 return fmt.Errorf("writing temp file: %v", err)
 528                         }
 529                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", tmpfn, bind))
 530
 531                 case mnt.Kind == "git_tree":
 532                         tmpdir, err := runner.MkTempDir("", "")
 533                         if err != nil {
 534                                 return fmt.Errorf("creating temp dir: %v", err)
 535                         }
 536                         runner.CleanupTempDir = append(runner.CleanupTempDir, tmpdir)
 537                         err = gitMount(mnt).extractTree(runner.ArvClient, tmpdir, token)
 538                         if err != nil {
 539                                 return err
 540                         }
 541                         runner.Binds = append(runner.Binds, tmpdir+":"+bind+":ro")
 542                 }
 543         }
 544
 545         if runner.HostOutputDir == "" {
 546                 return fmt.Errorf("Output path does not correspond to a writable mount point")
 547         }
 548
 549         if wantAPI := runner.Container.RuntimeConstraints.API; needCertMount && wantAPI != nil && *wantAPI {
 550                 for _, certfile := range arvadosclient.CertFiles {
 551                         _, err := os.Stat(certfile)
 552                         if err == nil {
 553                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:/etc/arvados/ca-certificates.crt:ro", certfile))
 554                                 break
 555                         }
 556                 }
 557         }
 558
 559         if pdhOnly {
 560                 arvMountCmd = append(arvMountCmd, "--mount-by-pdh", "by_id")
 561         } else {
 562                 arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_id")
 563         }
 564         arvMountCmd = append(arvMountCmd, runner.ArvMountPoint)
 565
 566         runner.ArvMount, err = runner.RunArvMount(arvMountCmd, token)
 567         if err != nil {
 568                 return fmt.Errorf("While trying to start arv-mount: %v", err)
 569         }
 570
 571         for _, p := range collectionPaths {
 572                 _, err = os.Stat(p)
 573                 if err != nil {
 574                         return fmt.Errorf("While checking that input files exist: %v", err)
 575                 }
 576         }
 577
 578         for _, cp := range copyFiles {
 579                 dir, err := os.Stat(cp.src)
 580                 if err != nil {
 581                         return fmt.Errorf("While staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 582                 }
 583                 if dir.IsDir() {
 584                         err = filepath.Walk(cp.src, func(walkpath string, walkinfo os.FileInfo, walkerr error) error {
 585                                 if walkerr != nil {
 586                                         return walkerr
 587                                 }
 588                                 if walkinfo.Mode().IsRegular() {
 589                                         return copyfile(walkpath, path.Join(cp.bind, walkpath[len(cp.src):]))
 590                                 } else if walkinfo.Mode().IsDir() {
 591                                         // will be visited by Walk()
 592                                         return nil
 593                                 } else {
 594                                         return fmt.Errorf("Source %q is not a regular file or directory", cp.src)
 595                                 }
 596                         })
 597                 } else {
 598                         err = copyfile(cp.src, cp.bind)
 599                 }
 600                 if err != nil {
 601                         return fmt.Errorf("While staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 602                 }
 603         }
 604
 605         return nil
 606 }
 607
 608 func (runner *ContainerRunner) ProcessDockerAttach(containerReader io.Reader) {
 609         // Handle docker log protocol
 610         // https://docs.docker.com/engine/reference/api/docker_remote_api_v1.15/#attach-to-a-container
 611
 612         header := make([]byte, 8)
 613         for {
 614                 _, readerr := io.ReadAtLeast(containerReader, header, 8)
 615
 616                 if readerr == nil {
 617                         readsize := int64(header[7]) | (int64(header[6]) << 8) | (int64(header[5]) << 16) | (int64(header[4]) << 24)
 618                         if header[0] == 1 {
 619                                 // stdout
 620                                 _, readerr = io.CopyN(runner.Stdout, containerReader, readsize)
 621                         } else {
 622                                 // stderr
 623                                 _, readerr = io.CopyN(runner.Stderr, containerReader, readsize)
 624                         }
 625                 }
 626
 627                 if readerr != nil {
 628                         if readerr != io.EOF {
 629                                 runner.CrunchLog.Printf("While reading docker logs: %v", readerr)
 630                         }
 631
 632                         closeerr := runner.Stdout.Close()
 633                         if closeerr != nil {
 634                                 runner.CrunchLog.Printf("While closing stdout logs: %v", closeerr)
 635                         }
 636
 637                         closeerr = runner.Stderr.Close()
 638                         if closeerr != nil {
 639                                 runner.CrunchLog.Printf("While closing stderr logs: %v", closeerr)
 640                         }
 641
 642                         if runner.statReporter != nil {
 643                                 runner.statReporter.Stop()
 644                                 closeerr = runner.statLogger.Close()
 645                                 if closeerr != nil {
 646                                         runner.CrunchLog.Printf("While closing crunchstat logs: %v", closeerr)
 647                                 }
 648                         }
 649
 650                         close(runner.loggingDone)
 651                         return
 652                 }
 653         }
 654 }
 655
 656 func (runner *ContainerRunner) StartCrunchstat() {
 657         runner.statLogger = NewThrottledLogger(runner.NewLogWriter("crunchstat"))
 658         runner.statReporter = &crunchstat.Reporter{
 659                 CID:          runner.ContainerID,
 660                 Logger:       log.New(runner.statLogger, "", 0),
 661                 CgroupParent: runner.expectCgroupParent,
 662                 CgroupRoot:   runner.cgroupRoot,
 663                 PollPeriod:   runner.statInterval,
 664         }
 665         runner.statReporter.Start()
 666 }
 667
 668 type infoCommand struct {
 669         label string
 670         cmd   []string
 671 }
 672
 673 // LogHostInfo logs info about the current host, for debugging and
 674 // accounting purposes. Although it's logged as "node-info", this is
 675 // about the environment where crunch-run is actually running, which
 676 // might differ from what's described in the node record (see
 677 // LogNodeRecord).
 678 func (runner *ContainerRunner) LogHostInfo() (err error) {
 679         w := runner.NewLogWriter("node-info")
 680
 681         commands := []infoCommand{
 682                 {
 683                         label: "Host Information",
 684                         cmd:   []string{"uname", "-a"},
 685                 },
 686                 {
 687                         label: "CPU Information",
 688                         cmd:   []string{"cat", "/proc/cpuinfo"},
 689                 },
 690                 {
 691                         label: "Memory Information",
 692                         cmd:   []string{"cat", "/proc/meminfo"},
 693                 },
 694                 {
 695                         label: "Disk Space",
 696                         cmd:   []string{"df", "-m", "/", os.TempDir()},
 697                 },
 698                 {
 699                         label: "Disk INodes",
 700                         cmd:   []string{"df", "-i", "/", os.TempDir()},
 701                 },
 702         }
 703
 704         // Run commands with informational output to be logged.
 705         for _, command := range commands {
 706                 fmt.Fprintln(w, command.label)
 707                 cmd := exec.Command(command.cmd[0], command.cmd[1:]...)
 708                 cmd.Stdout = w
 709                 cmd.Stderr = w
 710                 if err := cmd.Run(); err != nil {
 711                         err = fmt.Errorf("While running command %q: %v", command.cmd, err)
 712                         fmt.Fprintln(w, err)
 713                         return err
 714                 }
 715                 fmt.Fprintln(w, "")
 716         }
 717
 718         err = w.Close()
 719         if err != nil {
 720                 return fmt.Errorf("While closing node-info logs: %v", err)
 721         }
 722         return nil
 723 }
 724
 725 // LogContainerRecord gets and saves the raw JSON container record from the API server
 726 func (runner *ContainerRunner) LogContainerRecord() error {
 727         logged, err := runner.logAPIResponse("container", "containers", map[string]interface{}{"filters": [][]string{{"uuid", "=", runner.Container.UUID}}}, nil)
 728         if !logged && err == nil {
 729                 err = fmt.Errorf("error: no container record found for %s", runner.Container.UUID)
 730         }
 731         return err
 732 }
 733
 734 // LogNodeRecord logs arvados#node record corresponding to the current host.
 735 func (runner *ContainerRunner) LogNodeRecord() error {
 736         hostname := os.Getenv("SLURMD_NODENAME")
 737         if hostname == "" {
 738                 hostname, _ = os.Hostname()
 739         }
 740         _, err := runner.logAPIResponse("node", "nodes", map[string]interface{}{"filters": [][]string{{"hostname", "=", hostname}}}, func(resp interface{}) {
 741                 // The "info" field has admin-only info when obtained
 742                 // with a privileged token, and should not be logged.
 743                 node, ok := resp.(map[string]interface{})
 744                 if ok {
 745                         delete(node, "info")
 746                 }
 747         })
 748         return err
 749 }
 750
 751 func (runner *ContainerRunner) logAPIResponse(label, path string, params map[string]interface{}, munge func(interface{})) (logged bool, err error) {
 752         w := &ArvLogWriter{
 753                 ArvClient:     runner.ArvClient,
 754                 UUID:          runner.Container.UUID,
 755                 loggingStream: label,
 756                 writeCloser:   runner.LogCollection.Open(label + ".json"),
 757         }
 758
 759         reader, err := runner.ArvClient.CallRaw("GET", path, "", "", arvadosclient.Dict(params))
 760         if err != nil {
 761                 return false, fmt.Errorf("error getting %s record: %v", label, err)
 762         }
 763         defer reader.Close()
 764
 765         dec := json.NewDecoder(reader)
 766         dec.UseNumber()
 767         var resp map[string]interface{}
 768         if err = dec.Decode(&resp); err != nil {
 769                 return false, fmt.Errorf("error decoding %s list response: %v", label, err)
 770         }
 771         items, ok := resp["items"].([]interface{})
 772         if !ok {
 773                 return false, fmt.Errorf("error decoding %s list response: no \"items\" key in API list response", label)
 774         } else if len(items) < 1 {
 775                 return false, nil
 776         }
 777         if munge != nil {
 778                 munge(items[0])
 779         }
 780         // Re-encode it using indentation to improve readability
 781         enc := json.NewEncoder(w)
 782         enc.SetIndent("", "    ")
 783         if err = enc.Encode(items[0]); err != nil {
 784                 return false, fmt.Errorf("error logging %s record: %v", label, err)
 785         }
 786         err = w.Close()
 787         if err != nil {
 788                 return false, fmt.Errorf("error closing %s.json in log collection: %v", label, err)
 789         }
 790         return true, nil
 791 }
 792
 793 // AttachStreams connects the docker container stdin, stdout and stderr logs
 794 // to the Arvados logger which logs to Keep and the API server logs table.
 795 func (runner *ContainerRunner) AttachStreams() (err error) {
 796
 797         runner.CrunchLog.Print("Attaching container streams")
 798
 799         // If stdin mount is provided, attach it to the docker container
 800         var stdinRdr arvados.File
 801         var stdinJson []byte
 802         if stdinMnt, ok := runner.Container.Mounts["stdin"]; ok {
 803                 if stdinMnt.Kind == "collection" {
 804                         var stdinColl arvados.Collection
 805                         collId := stdinMnt.UUID
 806                         if collId == "" {
 807                                 collId = stdinMnt.PortableDataHash
 808                         }
 809                         err = runner.ArvClient.Get("collections", collId, nil, &stdinColl)
 810                         if err != nil {
 811                                 return fmt.Errorf("While getting stding collection: %v", err)
 812                         }
 813
 814                         stdinRdr, err = runner.Kc.ManifestFileReader(manifest.Manifest{Text: stdinColl.ManifestText}, stdinMnt.Path)
 815                         if os.IsNotExist(err) {
 816                                 return fmt.Errorf("stdin collection path not found: %v", stdinMnt.Path)
 817                         } else if err != nil {
 818                                 return fmt.Errorf("While getting stdin collection path %v: %v", stdinMnt.Path, err)
 819                         }
 820                 } else if stdinMnt.Kind == "json" {
 821                         stdinJson, err = json.Marshal(stdinMnt.Content)
 822                         if err != nil {
 823                                 return fmt.Errorf("While encoding stdin json data: %v", err)
 824                         }
 825                 }
 826         }
 827
 828         stdinUsed := stdinRdr != nil || len(stdinJson) != 0
 829         response, err := runner.Docker.ContainerAttach(context.TODO(), runner.ContainerID,
 830                 dockertypes.ContainerAttachOptions{Stream: true, Stdin: stdinUsed, Stdout: true, Stderr: true})
 831         if err != nil {
 832                 return fmt.Errorf("While attaching container stdout/stderr streams: %v", err)
 833         }
 834
 835         runner.loggingDone = make(chan bool)
 836
 837         if stdoutMnt, ok := runner.Container.Mounts["stdout"]; ok {
 838                 stdoutFile, err := runner.getStdoutFile(stdoutMnt.Path)
 839                 if err != nil {
 840                         return err
 841                 }
 842                 runner.Stdout = stdoutFile
 843         } else {
 844                 runner.Stdout = NewThrottledLogger(runner.NewLogWriter("stdout"))
 845         }
 846
 847         if stderrMnt, ok := runner.Container.Mounts["stderr"]; ok {
 848                 stderrFile, err := runner.getStdoutFile(stderrMnt.Path)
 849                 if err != nil {
 850                         return err
 851                 }
 852                 runner.Stderr = stderrFile
 853         } else {
 854                 runner.Stderr = NewThrottledLogger(runner.NewLogWriter("stderr"))
 855         }
 856
 857         if stdinRdr != nil {
 858                 go func() {
 859                         _, err := io.Copy(response.Conn, stdinRdr)
 860                         if err != nil {
 861                                 runner.CrunchLog.Print("While writing stdin collection to docker container %q", err)
 862                                 runner.stop()
 863                         }
 864                         stdinRdr.Close()
 865                         response.CloseWrite()
 866                 }()
 867         } else if len(stdinJson) != 0 {
 868                 go func() {
 869                         _, err := io.Copy(response.Conn, bytes.NewReader(stdinJson))
 870                         if err != nil {
 871                                 runner.CrunchLog.Print("While writing stdin json to docker container %q", err)
 872                                 runner.stop()
 873                         }
 874                         response.CloseWrite()
 875                 }()
 876         }
 877
 878         go runner.ProcessDockerAttach(response.Reader)
 879
 880         return nil
 881 }
 882
 883 func (runner *ContainerRunner) getStdoutFile(mntPath string) (*os.File, error) {
 884         stdoutPath := mntPath[len(runner.Container.OutputPath):]
 885         index := strings.LastIndex(stdoutPath, "/")
 886         if index > 0 {
 887                 subdirs := stdoutPath[:index]
 888                 if subdirs != "" {
 889                         st, err := os.Stat(runner.HostOutputDir)
 890                         if err != nil {
 891                                 return nil, fmt.Errorf("While Stat on temp dir: %v", err)
 892                         }
 893                         stdoutPath := filepath.Join(runner.HostOutputDir, subdirs)
 894                         err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777)
 895                         if err != nil {
 896                                 return nil, fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err)
 897                         }
 898                 }
 899         }
 900         stdoutFile, err := os.Create(filepath.Join(runner.HostOutputDir, stdoutPath))
 901         if err != nil {
 902                 return nil, fmt.Errorf("While creating file %q: %v", stdoutPath, err)
 903         }
 904
 905         return stdoutFile, nil
 906 }
 907
 908 // CreateContainer creates the docker container.
 909 func (runner *ContainerRunner) CreateContainer() error {
 910         runner.CrunchLog.Print("Creating Docker container")
 911
 912         runner.ContainerConfig.Cmd = runner.Container.Command
 913         if runner.Container.Cwd != "." {
 914                 runner.ContainerConfig.WorkingDir = runner.Container.Cwd
 915         }
 916
 917         for k, v := range runner.Container.Environment {
 918                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env, k+"="+v)
 919         }
 920
 921         runner.ContainerConfig.Volumes = runner.Volumes
 922
 923         runner.HostConfig = dockercontainer.HostConfig{
 924                 Binds: runner.Binds,
 925                 LogConfig: dockercontainer.LogConfig{
 926                         Type: "none",
 927                 },
 928                 Resources: dockercontainer.Resources{
 929                         CgroupParent: runner.setCgroupParent,
 930                 },
 931         }
 932
 933         if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
 934                 tok, err := runner.ContainerToken()
 935                 if err != nil {
 936                         return err
 937                 }
 938                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env,
 939                         "ARVADOS_API_TOKEN="+tok,
 940                         "ARVADOS_API_HOST="+os.Getenv("ARVADOS_API_HOST"),
 941                         "ARVADOS_API_HOST_INSECURE="+os.Getenv("ARVADOS_API_HOST_INSECURE"),
 942                 )
 943                 runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
 944         } else {
 945                 if runner.enableNetwork == "always" {
 946                         runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
 947                 } else {
 948                         runner.HostConfig.NetworkMode = dockercontainer.NetworkMode("none")
 949                 }
 950         }
 951
 952         _, stdinUsed := runner.Container.Mounts["stdin"]
 953         runner.ContainerConfig.OpenStdin = stdinUsed
 954         runner.ContainerConfig.StdinOnce = stdinUsed
 955         runner.ContainerConfig.AttachStdin = stdinUsed
 956         runner.ContainerConfig.AttachStdout = true
 957         runner.ContainerConfig.AttachStderr = true
 958
 959         createdBody, err := runner.Docker.ContainerCreate(context.TODO(), &runner.ContainerConfig, &runner.HostConfig, nil, runner.Container.UUID)
 960         if err != nil {
 961                 return fmt.Errorf("While creating container: %v", err)
 962         }
 963
 964         runner.ContainerID = createdBody.ID
 965
 966         return runner.AttachStreams()
 967 }
 968
 969 // StartContainer starts the docker container created by CreateContainer.
 970 func (runner *ContainerRunner) StartContainer() error {
 971         runner.CrunchLog.Printf("Starting Docker container id '%s'", runner.ContainerID)
 972         runner.cStateLock.Lock()
 973         defer runner.cStateLock.Unlock()
 974         if runner.cCancelled {
 975                 return ErrCancelled
 976         }
 977         err := runner.Docker.ContainerStart(context.TODO(), runner.ContainerID,
 978                 dockertypes.ContainerStartOptions{})
 979         if err != nil {
 980                 var advice string
 981                 if m, e := regexp.MatchString("(?ms).*(exec|System error).*(no such file or directory|file not found).*", err.Error()); m && e == nil {
 982                         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])
 983                 }
 984                 return fmt.Errorf("could not start container: %v%s", err, advice)
 985         }
 986         return nil
 987 }
 988
 989 // WaitFinish waits for the container to terminate, capture the exit code, and
 990 // close the stdout/stderr logging.
 991 func (runner *ContainerRunner) WaitFinish() error {
 992         runner.CrunchLog.Print("Waiting for container to finish")
 993
 994         waitOk, waitErr := runner.Docker.ContainerWait(context.TODO(), runner.ContainerID, dockercontainer.WaitConditionNotRunning)
 995         arvMountExit := runner.ArvMountExit
 996         for {
 997                 select {
 998                 case waitBody := <-waitOk:
 999                         runner.CrunchLog.Printf("Container exited with code: %v", waitBody.StatusCode)
1000                         code := int(waitBody.StatusCode)
1001                         runner.ExitCode = &code
1002
1003                         // wait for stdout/stderr to complete
1004                         <-runner.loggingDone
1005                         return nil
1006
1007                 case err := <-waitErr:
1008                         return fmt.Errorf("container wait: %v", err)
1009
1010                 case <-arvMountExit:
1011                         runner.CrunchLog.Printf("arv-mount exited while container is still running.  Stopping container.")
1012                         runner.stop()
1013                         // arvMountExit will always be ready now that
1014                         // it's closed, but that doesn't interest us.
1015                         arvMountExit = nil
1016                 }
1017         }
1018 }
1019
1020 var ErrNotInOutputDir = fmt.Errorf("Must point to path within the output directory")
1021
1022 func (runner *ContainerRunner) derefOutputSymlink(path string, startinfo os.FileInfo) (tgt string, readlinktgt string, info os.FileInfo, err error) {
1023         // Follow symlinks if necessary
1024         info = startinfo
1025         tgt = path
1026         readlinktgt = ""
1027         nextlink := path
1028         for followed := 0; info.Mode()&os.ModeSymlink != 0; followed++ {
1029                 if followed >= limitFollowSymlinks {
1030                         // Got stuck in a loop or just a pathological number of links, give up.
1031                         err = fmt.Errorf("Followed more than %v symlinks from path %q", limitFollowSymlinks, path)
1032                         return
1033                 }
1034
1035                 readlinktgt, err = os.Readlink(nextlink)
1036                 if err != nil {
1037                         return
1038                 }
1039
1040                 tgt = readlinktgt
1041                 if !strings.HasPrefix(tgt, "/") {
1042                         // Relative symlink, resolve it to host path
1043                         tgt = filepath.Join(filepath.Dir(path), tgt)
1044                 }
1045                 if strings.HasPrefix(tgt, runner.Container.OutputPath+"/") && !strings.HasPrefix(tgt, runner.HostOutputDir+"/") {
1046                         // Absolute symlink to container output path, adjust it to host output path.
1047                         tgt = filepath.Join(runner.HostOutputDir, tgt[len(runner.Container.OutputPath):])
1048                 }
1049                 if !strings.HasPrefix(tgt, runner.HostOutputDir+"/") {
1050                         // After dereferencing, symlink target must either be
1051                         // within output directory, or must point to a
1052                         // collection mount.
1053                         err = ErrNotInOutputDir
1054                         return
1055                 }
1056
1057                 info, err = os.Lstat(tgt)
1058                 if err != nil {
1059                         // tgt
1060                         err = fmt.Errorf("Symlink in output %q points to invalid location %q: %v",
1061                                 path[len(runner.HostOutputDir):], readlinktgt, err)
1062                         return
1063                 }
1064
1065                 nextlink = tgt
1066         }
1067
1068         return
1069 }
1070
1071 var limitFollowSymlinks = 10
1072
1073 // UploadFile uploads files within the output directory, with special handling
1074 // for symlinks. If the symlink leads to a keep mount, copy the manifest text
1075 // from the keep mount into the output manifestText.  Ensure that whether
1076 // symlinks are relative or absolute, every symlink target (even targets that
1077 // are symlinks themselves) must point to a path in either the output directory
1078 // or a collection mount.
1079 //
1080 // Assumes initial value of "path" is absolute, and located within runner.HostOutputDir.
1081 func (runner *ContainerRunner) UploadOutputFile(
1082         path string,
1083         info os.FileInfo,
1084         infoerr error,
1085         binds []string,
1086         walkUpload *WalkUpload,
1087         relocateFrom string,
1088         relocateTo string,
1089         followed int) (manifestText string, err error) {
1090
1091         if infoerr != nil {
1092                 return "", infoerr
1093         }
1094
1095         if info.Mode().IsDir() {
1096                 // if empty, need to create a .keep file
1097                 dir, direrr := os.Open(path)
1098                 if direrr != nil {
1099                         return "", direrr
1100                 }
1101                 defer dir.Close()
1102                 names, eof := dir.Readdirnames(1)
1103                 if len(names) == 0 && eof == io.EOF && path != runner.HostOutputDir {
1104                         containerPath := runner.OutputPath + path[len(runner.HostOutputDir):]
1105                         for _, bind := range binds {
1106                                 mnt := runner.Container.Mounts[bind]
1107                                 // Check if there is a bind for this
1108                                 // directory, in which case assume we don't need .keep
1109                                 if (containerPath == bind || strings.HasPrefix(containerPath, bind+"/")) && mnt.PortableDataHash != "d41d8cd98f00b204e9800998ecf8427e+0" {
1110                                         return
1111                                 }
1112                         }
1113                         outputSuffix := path[len(runner.HostOutputDir)+1:]
1114                         return fmt.Sprintf("./%v d41d8cd98f00b204e9800998ecf8427e+0 0:0:.keep\n", outputSuffix), nil
1115                 }
1116                 return
1117         }
1118
1119         if followed >= limitFollowSymlinks {
1120                 // Got stuck in a loop or just a pathological number of
1121                 // directory links, give up.
1122                 err = fmt.Errorf("Followed more than %v symlinks from path %q", limitFollowSymlinks, path)
1123                 return
1124         }
1125
1126         // "path" is the actual path we are visiting
1127         // "tgt" is the target of "path" (a non-symlink) after following symlinks
1128         // "relocated" is the path in the output manifest where the file should be placed,
1129         // but has HostOutputDir as a prefix.
1130
1131         // The destination path in the output manifest may need to be
1132         // logically relocated to some other path in order to appear
1133         // in the correct location as a result of following a symlink.
1134         // Remove the relocateFrom prefix and replace it with
1135         // relocateTo.
1136         relocated := relocateTo + path[len(relocateFrom):]
1137
1138         tgt, readlinktgt, info, derefErr := runner.derefOutputSymlink(path, info)
1139         if derefErr != nil && derefErr != ErrNotInOutputDir {
1140                 return "", derefErr
1141         }
1142
1143         // go through mounts and try reverse map to collection reference
1144         for _, bind := range binds {
1145                 mnt := runner.Container.Mounts[bind]
1146                 if (tgt == bind || strings.HasPrefix(tgt, bind+"/")) && !mnt.Writable {
1147                         // get path relative to bind
1148                         targetSuffix := tgt[len(bind):]
1149
1150                         // Copy mount and adjust the path to add path relative to the bind
1151                         adjustedMount := mnt
1152                         adjustedMount.Path = filepath.Join(adjustedMount.Path, targetSuffix)
1153
1154                         // Terminates in this keep mount, so add the
1155                         // manifest text at appropriate location.
1156                         outputSuffix := relocated[len(runner.HostOutputDir):]
1157                         manifestText, err = runner.getCollectionManifestForPath(adjustedMount, outputSuffix)
1158                         return
1159                 }
1160         }
1161
1162         // If target is not a collection mount, it must be located within the
1163         // output directory, otherwise it is an error.
1164         if derefErr == ErrNotInOutputDir {
1165                 err = fmt.Errorf("Symlink in output %q points to invalid location %q, must point to path within the output directory.",
1166                         path[len(runner.HostOutputDir):], readlinktgt)
1167                 return
1168         }
1169
1170         if info.Mode().IsRegular() {
1171                 return "", walkUpload.UploadFile(relocated, tgt)
1172         }
1173
1174         if info.Mode().IsDir() {
1175                 // Symlink leads to directory.  Walk() doesn't follow
1176                 // directory symlinks, so we walk the target directory
1177                 // instead.  Within the walk, file paths are relocated
1178                 // so they appear under the original symlink path.
1179                 err = filepath.Walk(tgt, func(walkpath string, walkinfo os.FileInfo, walkerr error) error {
1180                         var m string
1181                         m, walkerr = runner.UploadOutputFile(walkpath, walkinfo, walkerr,
1182                                 binds, walkUpload, tgt, relocated, followed+1)
1183                         if walkerr == nil {
1184                                 manifestText = manifestText + m
1185                         }
1186                         return walkerr
1187                 })
1188                 return
1189         }
1190
1191         return
1192 }
1193
1194 // HandleOutput sets the output, unmounts the FUSE mount, and deletes temporary directories
1195 func (runner *ContainerRunner) CaptureOutput() error {
1196         if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
1197                 // Output may have been set directly by the container, so
1198                 // refresh the container record to check.
1199                 err := runner.ArvClient.Get("containers", runner.Container.UUID,
1200                         nil, &runner.Container)
1201                 if err != nil {
1202                         return err
1203                 }
1204                 if runner.Container.Output != "" {
1205                         // Container output is already set.
1206                         runner.OutputPDH = &runner.Container.Output
1207                         return nil
1208                 }
1209         }
1210
1211         if runner.HostOutputDir == "" {
1212                 return nil
1213         }
1214
1215         _, err := os.Stat(runner.HostOutputDir)
1216         if err != nil {
1217                 return fmt.Errorf("While checking host output path: %v", err)
1218         }
1219
1220         // Pre-populate output from the configured mount points
1221         var binds []string
1222         for bind, mnt := range runner.Container.Mounts {
1223                 if mnt.Kind == "collection" {
1224                         binds = append(binds, bind)
1225                 }
1226         }
1227         sort.Strings(binds)
1228
1229         var manifestText string
1230
1231         collectionMetafile := fmt.Sprintf("%s/.arvados#collection", runner.HostOutputDir)
1232         _, err = os.Stat(collectionMetafile)
1233         if err != nil {
1234                 // Regular directory
1235
1236                 cw := CollectionWriter{0, runner.Kc, nil, nil, sync.Mutex{}}
1237                 walkUpload := cw.BeginUpload(runner.HostOutputDir, runner.CrunchLog.Logger)
1238
1239                 var m string
1240                 err = filepath.Walk(runner.HostOutputDir, func(path string, info os.FileInfo, err error) error {
1241                         m, err = runner.UploadOutputFile(path, info, err, binds, walkUpload, "", "", 0)
1242                         if err == nil {
1243                                 manifestText = manifestText + m
1244                         }
1245                         return err
1246                 })
1247
1248                 cw.EndUpload(walkUpload)
1249
1250                 if err != nil {
1251                         return fmt.Errorf("While uploading output files: %v", err)
1252                 }
1253
1254                 m, err = cw.ManifestText()
1255                 manifestText = manifestText + m
1256                 if err != nil {
1257                         return fmt.Errorf("While uploading output files: %v", err)
1258                 }
1259         } else {
1260                 // FUSE mount directory
1261                 file, openerr := os.Open(collectionMetafile)
1262                 if openerr != nil {
1263                         return fmt.Errorf("While opening FUSE metafile: %v", err)
1264                 }
1265                 defer file.Close()
1266
1267                 var rec arvados.Collection
1268                 err = json.NewDecoder(file).Decode(&rec)
1269                 if err != nil {
1270                         return fmt.Errorf("While reading FUSE metafile: %v", err)
1271                 }
1272                 manifestText = rec.ManifestText
1273         }
1274
1275         for _, bind := range binds {
1276                 mnt := runner.Container.Mounts[bind]
1277
1278                 bindSuffix := strings.TrimPrefix(bind, runner.Container.OutputPath)
1279
1280                 if bindSuffix == bind || len(bindSuffix) <= 0 {
1281                         // either does not start with OutputPath or is OutputPath itself
1282                         continue
1283                 }
1284
1285                 if mnt.ExcludeFromOutput == true || mnt.Writable {
1286                         continue
1287                 }
1288
1289                 // append to manifest_text
1290                 m, err := runner.getCollectionManifestForPath(mnt, bindSuffix)
1291                 if err != nil {
1292                         return err
1293                 }
1294
1295                 manifestText = manifestText + m
1296         }
1297
1298         // Save output
1299         var response arvados.Collection
1300         manifest := manifest.Manifest{Text: manifestText}
1301         manifestText = manifest.Extract(".", ".").Text
1302         err = runner.ArvClient.Create("collections",
1303                 arvadosclient.Dict{
1304                         "ensure_unique_name": true,
1305                         "collection": arvadosclient.Dict{
1306                                 "is_trashed":    true,
1307                                 "name":          "output for " + runner.Container.UUID,
1308                                 "manifest_text": manifestText}},
1309                 &response)
1310         if err != nil {
1311                 return fmt.Errorf("While creating output collection: %v", err)
1312         }
1313         runner.OutputPDH = &response.PortableDataHash
1314         return nil
1315 }
1316
1317 var outputCollections = make(map[string]arvados.Collection)
1318
1319 // Fetch the collection for the mnt.PortableDataHash
1320 // Return the manifest_text fragment corresponding to the specified mnt.Path
1321 //  after making any required updates.
1322 //  Ex:
1323 //    If mnt.Path is not specified,
1324 //      return the entire manifest_text after replacing any "." with bindSuffix
1325 //    If mnt.Path corresponds to one stream,
1326 //      return the manifest_text for that stream after replacing that stream name with bindSuffix
1327 //    Otherwise, check if a filename in any one stream is being sought. Return the manifest_text
1328 //      for that stream after replacing stream name with bindSuffix minus the last word
1329 //      and the file name with last word of the bindSuffix
1330 //  Allowed path examples:
1331 //    "path":"/"
1332 //    "path":"/subdir1"
1333 //    "path":"/subdir1/subdir2"
1334 //    "path":"/subdir/filename" etc
1335 func (runner *ContainerRunner) getCollectionManifestForPath(mnt arvados.Mount, bindSuffix string) (string, error) {
1336         collection := outputCollections[mnt.PortableDataHash]
1337         if collection.PortableDataHash == "" {
1338                 err := runner.ArvClient.Get("collections", mnt.PortableDataHash, nil, &collection)
1339                 if err != nil {
1340                         return "", fmt.Errorf("While getting collection for %v: %v", mnt.PortableDataHash, err)
1341                 }
1342                 outputCollections[mnt.PortableDataHash] = collection
1343         }
1344
1345         if collection.ManifestText == "" {
1346                 runner.CrunchLog.Printf("No manifest text for collection %v", collection.PortableDataHash)
1347                 return "", nil
1348         }
1349
1350         mft := manifest.Manifest{Text: collection.ManifestText}
1351         extracted := mft.Extract(mnt.Path, bindSuffix)
1352         if extracted.Err != nil {
1353                 return "", fmt.Errorf("Error parsing manifest for %v: %v", mnt.PortableDataHash, extracted.Err.Error())
1354         }
1355         return extracted.Text, nil
1356 }
1357
1358 func (runner *ContainerRunner) CleanupDirs() {
1359         if runner.ArvMount != nil {
1360                 var delay int64 = 8
1361                 umount := exec.Command("arv-mount", fmt.Sprintf("--unmount-timeout=%d", delay), "--unmount", runner.ArvMountPoint)
1362                 umount.Stdout = runner.CrunchLog
1363                 umount.Stderr = runner.CrunchLog
1364                 runner.CrunchLog.Printf("Running %v", umount.Args)
1365                 umnterr := umount.Start()
1366
1367                 if umnterr != nil {
1368                         runner.CrunchLog.Printf("Error unmounting: %v", umnterr)
1369                 } else {
1370                         // If arv-mount --unmount gets stuck for any reason, we
1371                         // don't want to wait for it forever.  Do Wait() in a goroutine
1372                         // so it doesn't block crunch-run.
1373                         umountExit := make(chan error)
1374                         go func() {
1375                                 mnterr := umount.Wait()
1376                                 if mnterr != nil {
1377                                         runner.CrunchLog.Printf("Error unmounting: %v", mnterr)
1378                                 }
1379                                 umountExit <- mnterr
1380                         }()
1381
1382                         for again := true; again; {
1383                                 again = false
1384                                 select {
1385                                 case <-umountExit:
1386                                         umount = nil
1387                                         again = true
1388                                 case <-runner.ArvMountExit:
1389                                         break
1390                                 case <-time.After(time.Duration((delay + 1) * int64(time.Second))):
1391                                         runner.CrunchLog.Printf("Timed out waiting for unmount")
1392                                         if umount != nil {
1393                                                 umount.Process.Kill()
1394                                         }
1395                                         runner.ArvMount.Process.Kill()
1396                                 }
1397                         }
1398                 }
1399         }
1400
1401         if runner.ArvMountPoint != "" {
1402                 if rmerr := os.Remove(runner.ArvMountPoint); rmerr != nil {
1403                         runner.CrunchLog.Printf("While cleaning up arv-mount directory %s: %v", runner.ArvMountPoint, rmerr)
1404                 }
1405         }
1406
1407         for _, tmpdir := range runner.CleanupTempDir {
1408                 if rmerr := os.RemoveAll(tmpdir); rmerr != nil {
1409                         runner.CrunchLog.Printf("While cleaning up temporary directory %s: %v", tmpdir, rmerr)
1410                 }
1411         }
1412 }
1413
1414 // CommitLogs posts the collection containing the final container logs.
1415 func (runner *ContainerRunner) CommitLogs() error {
1416         runner.CrunchLog.Print(runner.finalState)
1417
1418         if runner.arvMountLog != nil {
1419                 runner.arvMountLog.Close()
1420         }
1421         runner.CrunchLog.Close()
1422
1423         // Closing CrunchLog above allows them to be committed to Keep at this
1424         // point, but re-open crunch log with ArvClient in case there are any
1425         // other further errors (such as failing to write the log to Keep!)
1426         // while shutting down
1427         runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{ArvClient: runner.ArvClient,
1428                 UUID: runner.Container.UUID, loggingStream: "crunch-run", writeCloser: nil})
1429         runner.CrunchLog.Immediate = log.New(os.Stderr, runner.Container.UUID+" ", 0)
1430
1431         if runner.LogsPDH != nil {
1432                 // If we have already assigned something to LogsPDH,
1433                 // we must be closing the re-opened log, which won't
1434                 // end up getting attached to the container record and
1435                 // therefore doesn't need to be saved as a collection
1436                 // -- it exists only to send logs to other channels.
1437                 return nil
1438         }
1439
1440         mt, err := runner.LogCollection.ManifestText()
1441         if err != nil {
1442                 return fmt.Errorf("While creating log manifest: %v", err)
1443         }
1444
1445         var response arvados.Collection
1446         err = runner.ArvClient.Create("collections",
1447                 arvadosclient.Dict{
1448                         "ensure_unique_name": true,
1449                         "collection": arvadosclient.Dict{
1450                                 "is_trashed":    true,
1451                                 "name":          "logs for " + runner.Container.UUID,
1452                                 "manifest_text": mt}},
1453                 &response)
1454         if err != nil {
1455                 return fmt.Errorf("While creating log collection: %v", err)
1456         }
1457         runner.LogsPDH = &response.PortableDataHash
1458         return nil
1459 }
1460
1461 // UpdateContainerRunning updates the container state to "Running"
1462 func (runner *ContainerRunner) UpdateContainerRunning() error {
1463         runner.cStateLock.Lock()
1464         defer runner.cStateLock.Unlock()
1465         if runner.cCancelled {
1466                 return ErrCancelled
1467         }
1468         return runner.ArvClient.Update("containers", runner.Container.UUID,
1469                 arvadosclient.Dict{"container": arvadosclient.Dict{"state": "Running"}}, nil)
1470 }
1471
1472 // ContainerToken returns the api_token the container (and any
1473 // arv-mount processes) are allowed to use.
1474 func (runner *ContainerRunner) ContainerToken() (string, error) {
1475         if runner.token != "" {
1476                 return runner.token, nil
1477         }
1478
1479         var auth arvados.APIClientAuthorization
1480         err := runner.ArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth)
1481         if err != nil {
1482                 return "", err
1483         }
1484         runner.token = auth.APIToken
1485         return runner.token, nil
1486 }
1487
1488 // UpdateContainerComplete updates the container record state on API
1489 // server to "Complete" or "Cancelled"
1490 func (runner *ContainerRunner) UpdateContainerFinal() error {
1491         update := arvadosclient.Dict{}
1492         update["state"] = runner.finalState
1493         if runner.LogsPDH != nil {
1494                 update["log"] = *runner.LogsPDH
1495         }
1496         if runner.finalState == "Complete" {
1497                 if runner.ExitCode != nil {
1498                         update["exit_code"] = *runner.ExitCode
1499                 }
1500                 if runner.OutputPDH != nil {
1501                         update["output"] = *runner.OutputPDH
1502                 }
1503         }
1504         return runner.ArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{"container": update}, nil)
1505 }
1506
1507 // IsCancelled returns the value of Cancelled, with goroutine safety.
1508 func (runner *ContainerRunner) IsCancelled() bool {
1509         runner.cStateLock.Lock()
1510         defer runner.cStateLock.Unlock()
1511         return runner.cCancelled
1512 }
1513
1514 // NewArvLogWriter creates an ArvLogWriter
1515 func (runner *ContainerRunner) NewArvLogWriter(name string) io.WriteCloser {
1516         return &ArvLogWriter{
1517                 ArvClient:     runner.ArvClient,
1518                 UUID:          runner.Container.UUID,
1519                 loggingStream: name,
1520                 writeCloser:   runner.LogCollection.Open(name + ".txt")}
1521 }
1522
1523 // Run the full container lifecycle.
1524 func (runner *ContainerRunner) Run() (err error) {
1525         runner.CrunchLog.Printf("crunch-run %s started", version)
1526         runner.CrunchLog.Printf("Executing container '%s'", runner.Container.UUID)
1527
1528         hostname, hosterr := os.Hostname()
1529         if hosterr != nil {
1530                 runner.CrunchLog.Printf("Error getting hostname '%v'", hosterr)
1531         } else {
1532                 runner.CrunchLog.Printf("Executing on host '%s'", hostname)
1533         }
1534
1535         runner.finalState = "Queued"
1536
1537         defer func() {
1538                 runner.stopSignals()
1539                 runner.CleanupDirs()
1540
1541                 runner.CrunchLog.Printf("crunch-run finished")
1542                 runner.CrunchLog.Close()
1543         }()
1544
1545         defer func() {
1546                 // checkErr prints e (unless it's nil) and sets err to
1547                 // e (unless err is already non-nil). Thus, if err
1548                 // hasn't already been assigned when Run() returns,
1549                 // this cleanup func will cause Run() to return the
1550                 // first non-nil error that is passed to checkErr().
1551                 checkErr := func(e error) {
1552                         if e == nil {
1553                                 return
1554                         }
1555                         runner.CrunchLog.Print(e)
1556                         if err == nil {
1557                                 err = e
1558                         }
1559                         if runner.finalState == "Complete" {
1560                                 // There was an error in the finalization.
1561                                 runner.finalState = "Cancelled"
1562                         }
1563                 }
1564
1565                 // Log the error encountered in Run(), if any
1566                 checkErr(err)
1567
1568                 if runner.finalState == "Queued" {
1569                         runner.UpdateContainerFinal()
1570                         return
1571                 }
1572
1573                 if runner.IsCancelled() {
1574                         runner.finalState = "Cancelled"
1575                         // but don't return yet -- we still want to
1576                         // capture partial output and write logs
1577                 }
1578
1579                 checkErr(runner.CaptureOutput())
1580                 checkErr(runner.CommitLogs())
1581                 checkErr(runner.UpdateContainerFinal())
1582         }()
1583
1584         err = runner.fetchContainerRecord()
1585         if err != nil {
1586                 return
1587         }
1588
1589         // setup signal handling
1590         runner.setupSignals()
1591
1592         // check for and/or load image
1593         err = runner.LoadImage()
1594         if err != nil {
1595                 if !runner.checkBrokenNode(err) {
1596                         // Failed to load image but not due to a "broken node"
1597                         // condition, probably user error.
1598                         runner.finalState = "Cancelled"
1599                 }
1600                 err = fmt.Errorf("While loading container image: %v", err)
1601                 return
1602         }
1603
1604         // set up FUSE mount and binds
1605         err = runner.SetupMounts()
1606         if err != nil {
1607                 runner.finalState = "Cancelled"
1608                 err = fmt.Errorf("While setting up mounts: %v", err)
1609                 return
1610         }
1611
1612         err = runner.CreateContainer()
1613         if err != nil {
1614                 return
1615         }
1616         err = runner.LogHostInfo()
1617         if err != nil {
1618                 return
1619         }
1620         err = runner.LogNodeRecord()
1621         if err != nil {
1622                 return
1623         }
1624         err = runner.LogContainerRecord()
1625         if err != nil {
1626                 return
1627         }
1628
1629         if runner.IsCancelled() {
1630                 return
1631         }
1632
1633         err = runner.UpdateContainerRunning()
1634         if err != nil {
1635                 return
1636         }
1637         runner.finalState = "Cancelled"
1638
1639         runner.StartCrunchstat()
1640
1641         err = runner.StartContainer()
1642         if err != nil {
1643                 runner.checkBrokenNode(err)
1644                 return
1645         }
1646
1647         err = runner.WaitFinish()
1648         if err == nil && !runner.IsCancelled() {
1649                 runner.finalState = "Complete"
1650         }
1651         return
1652 }
1653
1654 // Fetch the current container record (uuid = runner.Container.UUID)
1655 // into runner.Container.
1656 func (runner *ContainerRunner) fetchContainerRecord() error {
1657         reader, err := runner.ArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
1658         if err != nil {
1659                 return fmt.Errorf("error fetching container record: %v", err)
1660         }
1661         defer reader.Close()
1662
1663         dec := json.NewDecoder(reader)
1664         dec.UseNumber()
1665         err = dec.Decode(&runner.Container)
1666         if err != nil {
1667                 return fmt.Errorf("error decoding container record: %v", err)
1668         }
1669         return nil
1670 }
1671
1672 // NewContainerRunner creates a new container runner.
1673 func NewContainerRunner(api IArvadosClient,
1674         kc IKeepClient,
1675         docker ThinDockerClient,
1676         containerUUID string) *ContainerRunner {
1677
1678         cr := &ContainerRunner{ArvClient: api, Kc: kc, Docker: docker}
1679         cr.NewLogWriter = cr.NewArvLogWriter
1680         cr.RunArvMount = cr.ArvMountCmd
1681         cr.MkTempDir = ioutil.TempDir
1682         cr.LogCollection = &CollectionWriter{0, kc, nil, nil, sync.Mutex{}}
1683         cr.Container.UUID = containerUUID
1684         cr.CrunchLog = NewThrottledLogger(cr.NewLogWriter("crunch-run"))
1685         cr.CrunchLog.Immediate = log.New(os.Stderr, containerUUID+" ", 0)
1686
1687         loadLogThrottleParams(api)
1688
1689         return cr
1690 }
1691
1692 func main() {
1693         statInterval := flag.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting")
1694         cgroupRoot := flag.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree")
1695         cgroupParent := flag.String("cgroup-parent", "docker", "name of container's parent cgroup (ignored if -cgroup-parent-subsystem is used)")
1696         cgroupParentSubsystem := flag.String("cgroup-parent-subsystem", "", "use current cgroup for given subsystem as parent cgroup for container")
1697         caCertsPath := flag.String("ca-certs", "", "Path to TLS root certificates")
1698         enableNetwork := flag.String("container-enable-networking", "default",
1699                 `Specify if networking should be enabled for container.  One of 'default', 'always':
1700         default: only enable networking if container requests it.
1701         always:  containers always have networking enabled
1702         `)
1703         networkMode := flag.String("container-network-mode", "default",
1704                 `Set networking mode for container.  Corresponds to Docker network mode (--net).
1705         `)
1706         memprofile := flag.String("memprofile", "", "write memory profile to `file` after running container")
1707         getVersion := flag.Bool("version", false, "Print version information and exit.")
1708         flag.Parse()
1709
1710         // Print version information if requested
1711         if *getVersion {
1712                 fmt.Printf("crunch-run %s\n", version)
1713                 return
1714         }
1715
1716         log.Printf("crunch-run %s started", version)
1717
1718         containerId := flag.Arg(0)
1719
1720         if *caCertsPath != "" {
1721                 arvadosclient.CertFiles = []string{*caCertsPath}
1722         }
1723
1724         api, err := arvadosclient.MakeArvadosClient()
1725         if err != nil {
1726                 log.Fatalf("%s: %v", containerId, err)
1727         }
1728         api.Retries = 8
1729
1730         kc, kcerr := keepclient.MakeKeepClient(api)
1731         if kcerr != nil {
1732                 log.Fatalf("%s: %v", containerId, kcerr)
1733         }
1734         kc.BlockCache = &keepclient.BlockCache{MaxBlocks: 2}
1735         kc.Retries = 4
1736
1737         // API version 1.21 corresponds to Docker 1.9, which is currently the
1738         // minimum version we want to support.
1739         docker, dockererr := dockerclient.NewClient(dockerclient.DefaultDockerHost, "1.21", nil, nil)
1740
1741         cr := NewContainerRunner(api, kc, docker, containerId)
1742         if dockererr != nil {
1743                 cr.CrunchLog.Printf("%s: %v", containerId, dockererr)
1744                 cr.checkBrokenNode(dockererr)
1745                 cr.CrunchLog.Close()
1746                 os.Exit(1)
1747         }
1748
1749         cr.statInterval = *statInterval
1750         cr.cgroupRoot = *cgroupRoot
1751         cr.expectCgroupParent = *cgroupParent
1752         cr.enableNetwork = *enableNetwork
1753         cr.networkMode = *networkMode
1754         if *cgroupParentSubsystem != "" {
1755                 p := findCgroup(*cgroupParentSubsystem)
1756                 cr.setCgroupParent = p
1757                 cr.expectCgroupParent = p
1758         }
1759
1760         runerr := cr.Run()
1761
1762         if *memprofile != "" {
1763                 f, err := os.Create(*memprofile)
1764                 if err != nil {
1765                         log.Printf("could not create memory profile: ", err)
1766                 }
1767                 runtime.GC() // get up-to-date statistics
1768                 if err := pprof.WriteHeapProfile(f); err != nil {
1769                         log.Printf("could not write memory profile: ", err)
1770                 }
1771                 closeerr := f.Close()
1772                 if closeerr != nil {
1773                         log.Printf("closing memprofile file: ", err)
1774                 }
1775         }
1776
1777         if runerr != nil {
1778                 log.Fatalf("%s: %v", containerId, runerr)
1779         }
1780 }