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         parentTemp     string
 117
 118         statLogger       io.WriteCloser
 119         statReporter     *crunchstat.Reporter
 120         hoststatLogger   io.WriteCloser
 121         hoststatReporter *crunchstat.Reporter
 122         statInterval     time.Duration
 123         cgroupRoot       string
 124         // What we expect the container's cgroup parent to be.
 125         expectCgroupParent string
 126         // What we tell docker to use as the container's cgroup
 127         // parent. Note: Ideally we would use the same field for both
 128         // expectCgroupParent and setCgroupParent, and just make it
 129         // default to "docker". However, when using docker < 1.10 with
 130         // systemd, specifying a non-empty cgroup parent (even the
 131         // default value "docker") hits a docker bug
 132         // (https://github.com/docker/docker/issues/17126). Using two
 133         // separate fields makes it possible to use the "expect cgroup
 134         // parent to be X" feature even on sites where the "specify
 135         // cgroup parent" feature breaks.
 136         setCgroupParent string
 137
 138         cStateLock sync.Mutex
 139         cCancelled bool // StopContainer() invoked
 140
 141         enableNetwork string // one of "default" or "always"
 142         networkMode   string // passed through to HostConfig.NetworkMode
 143         arvMountLog   *ThrottledLogger
 144 }
 145
 146 // setupSignals sets up signal handling to gracefully terminate the underlying
 147 // Docker container and update state when receiving a TERM, INT or QUIT signal.
 148 func (runner *ContainerRunner) setupSignals() {
 149         runner.SigChan = make(chan os.Signal, 1)
 150         signal.Notify(runner.SigChan, syscall.SIGTERM)
 151         signal.Notify(runner.SigChan, syscall.SIGINT)
 152         signal.Notify(runner.SigChan, syscall.SIGQUIT)
 153
 154         go func(sig chan os.Signal) {
 155                 for s := range sig {
 156                         runner.CrunchLog.Printf("caught signal: %v", s)
 157                         runner.stop()
 158                 }
 159         }(runner.SigChan)
 160 }
 161
 162 // stop the underlying Docker container.
 163 func (runner *ContainerRunner) stop() {
 164         runner.cStateLock.Lock()
 165         defer runner.cStateLock.Unlock()
 166         if runner.ContainerID == "" {
 167                 return
 168         }
 169         runner.cCancelled = true
 170         runner.CrunchLog.Printf("removing container")
 171         err := runner.Docker.ContainerRemove(context.TODO(), runner.ContainerID, dockertypes.ContainerRemoveOptions{Force: true})
 172         if err != nil {
 173                 runner.CrunchLog.Printf("error removing container: %s", err)
 174         }
 175 }
 176
 177 func (runner *ContainerRunner) stopSignals() {
 178         if runner.SigChan != nil {
 179                 signal.Stop(runner.SigChan)
 180         }
 181 }
 182
 183 var errorBlacklist = []string{
 184         "(?ms).*[Cc]annot connect to the Docker daemon.*",
 185         "(?ms).*oci runtime error.*starting container process.*container init.*mounting.*to rootfs.*no such file or directory.*",
 186 }
 187 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)")
 188
 189 func (runner *ContainerRunner) checkBrokenNode(goterr error) bool {
 190         for _, d := range errorBlacklist {
 191                 if m, e := regexp.MatchString(d, goterr.Error()); m && e == nil {
 192                         runner.CrunchLog.Printf("Error suggests node is unable to run containers: %v", goterr)
 193                         if *brokenNodeHook == "" {
 194                                 runner.CrunchLog.Printf("No broken node hook provided, cannot mark node as broken.")
 195                         } else {
 196                                 runner.CrunchLog.Printf("Running broken node hook %q", *brokenNodeHook)
 197                                 // run killme script
 198                                 c := exec.Command(*brokenNodeHook)
 199                                 c.Stdout = runner.CrunchLog
 200                                 c.Stderr = runner.CrunchLog
 201                                 err := c.Run()
 202                                 if err != nil {
 203                                         runner.CrunchLog.Printf("Error running broken node hook: %v", err)
 204                                 }
 205                         }
 206                         return true
 207                 }
 208         }
 209         return false
 210 }
 211
 212 // LoadImage determines the docker image id from the container record and
 213 // checks if it is available in the local Docker image store.  If not, it loads
 214 // the image from Keep.
 215 func (runner *ContainerRunner) LoadImage() (err error) {
 216
 217         runner.CrunchLog.Printf("Fetching Docker image from collection '%s'", runner.Container.ContainerImage)
 218
 219         var collection arvados.Collection
 220         err = runner.ArvClient.Get("collections", runner.Container.ContainerImage, nil, &collection)
 221         if err != nil {
 222                 return fmt.Errorf("While getting container image collection: %v", err)
 223         }
 224         manifest := manifest.Manifest{Text: collection.ManifestText}
 225         var img, imageID string
 226         for ms := range manifest.StreamIter() {
 227                 img = ms.FileStreamSegments[0].Name
 228                 if !strings.HasSuffix(img, ".tar") {
 229                         return fmt.Errorf("First file in the container image collection does not end in .tar")
 230                 }
 231                 imageID = img[:len(img)-4]
 232         }
 233
 234         runner.CrunchLog.Printf("Using Docker image id '%s'", imageID)
 235
 236         _, _, err = runner.Docker.ImageInspectWithRaw(context.TODO(), imageID)
 237         if err != nil {
 238                 runner.CrunchLog.Print("Loading Docker image from keep")
 239
 240                 var readCloser io.ReadCloser
 241                 readCloser, err = runner.Kc.ManifestFileReader(manifest, img)
 242                 if err != nil {
 243                         return fmt.Errorf("While creating ManifestFileReader for container image: %v", err)
 244                 }
 245
 246                 response, err := runner.Docker.ImageLoad(context.TODO(), readCloser, true)
 247                 if err != nil {
 248                         return fmt.Errorf("While loading container image into Docker: %v", err)
 249                 }
 250
 251                 defer response.Body.Close()
 252                 rbody, err := ioutil.ReadAll(response.Body)
 253                 if err != nil {
 254                         return fmt.Errorf("Reading response to image load: %v", err)
 255                 }
 256                 runner.CrunchLog.Printf("Docker response: %s", rbody)
 257         } else {
 258                 runner.CrunchLog.Print("Docker image is available")
 259         }
 260
 261         runner.ContainerConfig.Image = imageID
 262
 263         runner.Kc.ClearBlockCache()
 264
 265         return nil
 266 }
 267
 268 func (runner *ContainerRunner) ArvMountCmd(arvMountCmd []string, token string) (c *exec.Cmd, err error) {
 269         c = exec.Command("arv-mount", arvMountCmd...)
 270
 271         // Copy our environment, but override ARVADOS_API_TOKEN with
 272         // the container auth token.
 273         c.Env = nil
 274         for _, s := range os.Environ() {
 275                 if !strings.HasPrefix(s, "ARVADOS_API_TOKEN=") {
 276                         c.Env = append(c.Env, s)
 277                 }
 278         }
 279         c.Env = append(c.Env, "ARVADOS_API_TOKEN="+token)
 280
 281         runner.arvMountLog = NewThrottledLogger(runner.NewLogWriter("arv-mount"))
 282         c.Stdout = runner.arvMountLog
 283         c.Stderr = runner.arvMountLog
 284
 285         runner.CrunchLog.Printf("Running %v", c.Args)
 286
 287         err = c.Start()
 288         if err != nil {
 289                 return nil, err
 290         }
 291
 292         statReadme := make(chan bool)
 293         runner.ArvMountExit = make(chan error)
 294
 295         keepStatting := true
 296         go func() {
 297                 for keepStatting {
 298                         time.Sleep(100 * time.Millisecond)
 299                         _, err = os.Stat(fmt.Sprintf("%s/by_id/README", runner.ArvMountPoint))
 300                         if err == nil {
 301                                 keepStatting = false
 302                                 statReadme <- true
 303                         }
 304                 }
 305                 close(statReadme)
 306         }()
 307
 308         go func() {
 309                 mnterr := c.Wait()
 310                 if mnterr != nil {
 311                         runner.CrunchLog.Printf("Arv-mount exit error: %v", mnterr)
 312                 }
 313                 runner.ArvMountExit <- mnterr
 314                 close(runner.ArvMountExit)
 315         }()
 316
 317         select {
 318         case <-statReadme:
 319                 break
 320         case err := <-runner.ArvMountExit:
 321                 runner.ArvMount = nil
 322                 keepStatting = false
 323                 return nil, err
 324         }
 325
 326         return c, nil
 327 }
 328
 329 func (runner *ContainerRunner) SetupArvMountPoint(prefix string) (err error) {
 330         if runner.ArvMountPoint == "" {
 331                 runner.ArvMountPoint, err = runner.MkTempDir(runner.parentTemp, prefix)
 332         }
 333         return
 334 }
 335
 336 func copyfile(src string, dst string) (err error) {
 337         srcfile, err := os.Open(src)
 338         if err != nil {
 339                 return
 340         }
 341
 342         os.MkdirAll(path.Dir(dst), 0777)
 343
 344         dstfile, err := os.Create(dst)
 345         if err != nil {
 346                 return
 347         }
 348         _, err = io.Copy(dstfile, srcfile)
 349         if err != nil {
 350                 return
 351         }
 352
 353         err = srcfile.Close()
 354         err2 := dstfile.Close()
 355
 356         if err != nil {
 357                 return
 358         }
 359
 360         if err2 != nil {
 361                 return err2
 362         }
 363
 364         return nil
 365 }
 366
 367 func (runner *ContainerRunner) SetupMounts() (err error) {
 368         err = runner.SetupArvMountPoint("keep")
 369         if err != nil {
 370                 return fmt.Errorf("While creating keep mount temp dir: %v", err)
 371         }
 372
 373         token, err := runner.ContainerToken()
 374         if err != nil {
 375                 return fmt.Errorf("could not get container token: %s", err)
 376         }
 377
 378         pdhOnly := true
 379         tmpcount := 0
 380         arvMountCmd := []string{
 381                 "--foreground",
 382                 "--allow-other",
 383                 "--read-write",
 384                 fmt.Sprintf("--crunchstat-interval=%v", runner.statInterval.Seconds())}
 385
 386         if runner.Container.RuntimeConstraints.KeepCacheRAM > 0 {
 387                 arvMountCmd = append(arvMountCmd, "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheRAM))
 388         }
 389
 390         collectionPaths := []string{}
 391         runner.Binds = nil
 392         runner.Volumes = make(map[string]struct{})
 393         needCertMount := true
 394         type copyFile struct {
 395                 src  string
 396                 bind string
 397         }
 398         var copyFiles []copyFile
 399
 400         var binds []string
 401         for bind := range runner.Container.Mounts {
 402                 binds = append(binds, bind)
 403         }
 404         sort.Strings(binds)
 405
 406         for _, bind := range binds {
 407                 mnt := runner.Container.Mounts[bind]
 408                 if bind == "stdout" || bind == "stderr" {
 409                         // Is it a "file" mount kind?
 410                         if mnt.Kind != "file" {
 411                                 return fmt.Errorf("Unsupported mount kind '%s' for %s. Only 'file' is supported.", mnt.Kind, bind)
 412                         }
 413
 414                         // Does path start with OutputPath?
 415                         prefix := runner.Container.OutputPath
 416                         if !strings.HasSuffix(prefix, "/") {
 417                                 prefix += "/"
 418                         }
 419                         if !strings.HasPrefix(mnt.Path, prefix) {
 420                                 return fmt.Errorf("%s path does not start with OutputPath: %s, %s", strings.Title(bind), mnt.Path, prefix)
 421                         }
 422                 }
 423
 424                 if bind == "stdin" {
 425                         // Is it a "collection" mount kind?
 426                         if mnt.Kind != "collection" && mnt.Kind != "json" {
 427                                 return fmt.Errorf("Unsupported mount kind '%s' for stdin. Only 'collection' or 'json' are supported.", mnt.Kind)
 428                         }
 429                 }
 430
 431                 if bind == "/etc/arvados/ca-certificates.crt" {
 432                         needCertMount = false
 433                 }
 434
 435                 if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && bind != runner.Container.OutputPath+"/" {
 436                         if mnt.Kind != "collection" {
 437                                 return fmt.Errorf("Only mount points of kind 'collection' are supported underneath the output_path: %v", bind)
 438                         }
 439                 }
 440
 441                 switch {
 442                 case mnt.Kind == "collection" && bind != "stdin":
 443                         var src string
 444                         if mnt.UUID != "" && mnt.PortableDataHash != "" {
 445                                 return fmt.Errorf("Cannot specify both 'uuid' and 'portable_data_hash' for a collection mount")
 446                         }
 447                         if mnt.UUID != "" {
 448                                 if mnt.Writable {
 449                                         return fmt.Errorf("Writing to existing collections currently not permitted.")
 450                                 }
 451                                 pdhOnly = false
 452                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.UUID)
 453                         } else if mnt.PortableDataHash != "" {
 454                                 if mnt.Writable && !strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 455                                         return fmt.Errorf("Can never write to a collection specified by portable data hash")
 456                                 }
 457                                 idx := strings.Index(mnt.PortableDataHash, "/")
 458                                 if idx > 0 {
 459                                         mnt.Path = path.Clean(mnt.PortableDataHash[idx:])
 460                                         mnt.PortableDataHash = mnt.PortableDataHash[0:idx]
 461                                         runner.Container.Mounts[bind] = mnt
 462                                 }
 463                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.PortableDataHash)
 464                                 if mnt.Path != "" && mnt.Path != "." {
 465                                         if strings.HasPrefix(mnt.Path, "./") {
 466                                                 mnt.Path = mnt.Path[2:]
 467                                         } else if strings.HasPrefix(mnt.Path, "/") {
 468                                                 mnt.Path = mnt.Path[1:]
 469                                         }
 470                                         src += "/" + mnt.Path
 471                                 }
 472                         } else {
 473                                 src = fmt.Sprintf("%s/tmp%d", runner.ArvMountPoint, tmpcount)
 474                                 arvMountCmd = append(arvMountCmd, "--mount-tmp")
 475                                 arvMountCmd = append(arvMountCmd, fmt.Sprintf("tmp%d", tmpcount))
 476                                 tmpcount += 1
 477                         }
 478                         if mnt.Writable {
 479                                 if bind == runner.Container.OutputPath {
 480                                         runner.HostOutputDir = src
 481                                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", src, bind))
 482                                 } else if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 483                                         copyFiles = append(copyFiles, copyFile{src, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
 484                                 } else {
 485                                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", src, bind))
 486                                 }
 487                         } else {
 488                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", src, bind))
 489                         }
 490                         collectionPaths = append(collectionPaths, src)
 491
 492                 case mnt.Kind == "tmp":
 493                         var tmpdir string
 494                         tmpdir, err = runner.MkTempDir(runner.parentTemp, "tmp")
 495                         if err != nil {
 496                                 return fmt.Errorf("While creating mount temp dir: %v", err)
 497                         }
 498                         st, staterr := os.Stat(tmpdir)
 499                         if staterr != nil {
 500                                 return fmt.Errorf("While Stat on temp dir: %v", staterr)
 501                         }
 502                         err = os.Chmod(tmpdir, st.Mode()|os.ModeSetgid|0777)
 503                         if staterr != nil {
 504                                 return fmt.Errorf("While Chmod temp dir: %v", err)
 505                         }
 506                         runner.CleanupTempDir = append(runner.CleanupTempDir, tmpdir)
 507                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", tmpdir, bind))
 508                         if bind == runner.Container.OutputPath {
 509                                 runner.HostOutputDir = tmpdir
 510                         }
 511
 512                 case mnt.Kind == "json":
 513                         jsondata, err := json.Marshal(mnt.Content)
 514                         if err != nil {
 515                                 return fmt.Errorf("encoding json data: %v", err)
 516                         }
 517                         // Create a tempdir with a single file
 518                         // (instead of just a tempfile): this way we
 519                         // can ensure the file is world-readable
 520                         // inside the container, without having to
 521                         // make it world-readable on the docker host.
 522                         tmpdir, err := runner.MkTempDir(runner.parentTemp, "json")
 523                         if err != nil {
 524                                 return fmt.Errorf("creating temp dir: %v", err)
 525                         }
 526                         runner.CleanupTempDir = append(runner.CleanupTempDir, tmpdir)
 527                         tmpfn := filepath.Join(tmpdir, "mountdata.json")
 528                         err = ioutil.WriteFile(tmpfn, jsondata, 0644)
 529                         if err != nil {
 530                                 return fmt.Errorf("writing temp file: %v", err)
 531                         }
 532                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", tmpfn, bind))
 533
 534                 case mnt.Kind == "git_tree":
 535                         tmpdir, err := runner.MkTempDir(runner.parentTemp, "git_tree")
 536                         if err != nil {
 537                                 return fmt.Errorf("creating temp dir: %v", err)
 538                         }
 539                         runner.CleanupTempDir = append(runner.CleanupTempDir, tmpdir)
 540                         err = gitMount(mnt).extractTree(runner.ArvClient, tmpdir, token)
 541                         if err != nil {
 542                                 return err
 543                         }
 544                         runner.Binds = append(runner.Binds, tmpdir+":"+bind+":ro")
 545                 }
 546         }
 547
 548         if runner.HostOutputDir == "" {
 549                 return fmt.Errorf("Output path does not correspond to a writable mount point")
 550         }
 551
 552         if wantAPI := runner.Container.RuntimeConstraints.API; needCertMount && wantAPI != nil && *wantAPI {
 553                 for _, certfile := range arvadosclient.CertFiles {
 554                         _, err := os.Stat(certfile)
 555                         if err == nil {
 556                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:/etc/arvados/ca-certificates.crt:ro", certfile))
 557                                 break
 558                         }
 559                 }
 560         }
 561
 562         if pdhOnly {
 563                 arvMountCmd = append(arvMountCmd, "--mount-by-pdh", "by_id")
 564         } else {
 565                 arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_id")
 566         }
 567         arvMountCmd = append(arvMountCmd, runner.ArvMountPoint)
 568
 569         runner.ArvMount, err = runner.RunArvMount(arvMountCmd, token)
 570         if err != nil {
 571                 return fmt.Errorf("While trying to start arv-mount: %v", err)
 572         }
 573
 574         for _, p := range collectionPaths {
 575                 _, err = os.Stat(p)
 576                 if err != nil {
 577                         return fmt.Errorf("While checking that input files exist: %v", err)
 578                 }
 579         }
 580
 581         for _, cp := range copyFiles {
 582                 dir, err := os.Stat(cp.src)
 583                 if err != nil {
 584                         return fmt.Errorf("While staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 585                 }
 586                 if dir.IsDir() {
 587                         err = filepath.Walk(cp.src, func(walkpath string, walkinfo os.FileInfo, walkerr error) error {
 588                                 if walkerr != nil {
 589                                         return walkerr
 590                                 }
 591                                 if walkinfo.Mode().IsRegular() {
 592                                         return copyfile(walkpath, path.Join(cp.bind, walkpath[len(cp.src):]))
 593                                 } else if walkinfo.Mode().IsDir() {
 594                                         return os.MkdirAll(path.Join(cp.bind, walkpath[len(cp.src):]), 0777)
 595                                 } else {
 596                                         return fmt.Errorf("Source %q is not a regular file or directory", cp.src)
 597                                 }
 598                         })
 599                 } else {
 600                         err = copyfile(cp.src, cp.bind)
 601                 }
 602                 if err != nil {
 603                         return fmt.Errorf("While staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 604                 }
 605         }
 606
 607         return nil
 608 }
 609
 610 func (runner *ContainerRunner) ProcessDockerAttach(containerReader io.Reader) {
 611         // Handle docker log protocol
 612         // https://docs.docker.com/engine/reference/api/docker_remote_api_v1.15/#attach-to-a-container
 613         defer close(runner.loggingDone)
 614
 615         header := make([]byte, 8)
 616         var err error
 617         for err == nil {
 618                 _, err = io.ReadAtLeast(containerReader, header, 8)
 619                 if err != nil {
 620                         if err == io.EOF {
 621                                 err = nil
 622                         }
 623                         break
 624                 }
 625                 readsize := int64(header[7]) | (int64(header[6]) << 8) | (int64(header[5]) << 16) | (int64(header[4]) << 24)
 626                 if header[0] == 1 {
 627                         // stdout
 628                         _, err = io.CopyN(runner.Stdout, containerReader, readsize)
 629                 } else {
 630                         // stderr
 631                         _, err = io.CopyN(runner.Stderr, containerReader, readsize)
 632                 }
 633         }
 634
 635         if err != nil {
 636                 runner.CrunchLog.Printf("error reading docker logs: %v", err)
 637         }
 638
 639         err = runner.Stdout.Close()
 640         if err != nil {
 641                 runner.CrunchLog.Printf("error closing stdout logs: %v", err)
 642         }
 643
 644         err = runner.Stderr.Close()
 645         if err != nil {
 646                 runner.CrunchLog.Printf("error closing stderr logs: %v", err)
 647         }
 648
 649         if runner.statReporter != nil {
 650                 runner.statReporter.Stop()
 651                 err = runner.statLogger.Close()
 652                 if err != nil {
 653                         runner.CrunchLog.Printf("error closing crunchstat logs: %v", err)
 654                 }
 655         }
 656 }
 657
 658 func (runner *ContainerRunner) stopHoststat() error {
 659         if runner.hoststatReporter == nil {
 660                 return nil
 661         }
 662         runner.hoststatReporter.Stop()
 663         err := runner.hoststatLogger.Close()
 664         if err != nil {
 665                 return fmt.Errorf("error closing hoststat logs: %v", err)
 666         }
 667         return nil
 668 }
 669
 670 func (runner *ContainerRunner) startHoststat() {
 671         runner.hoststatLogger = NewThrottledLogger(runner.NewLogWriter("hoststat"))
 672         runner.hoststatReporter = &crunchstat.Reporter{
 673                 Logger:     log.New(runner.hoststatLogger, "", 0),
 674                 CgroupRoot: runner.cgroupRoot,
 675                 PollPeriod: runner.statInterval,
 676         }
 677         runner.hoststatReporter.Start()
 678 }
 679
 680 func (runner *ContainerRunner) startCrunchstat() {
 681         runner.statLogger = NewThrottledLogger(runner.NewLogWriter("crunchstat"))
 682         runner.statReporter = &crunchstat.Reporter{
 683                 CID:          runner.ContainerID,
 684                 Logger:       log.New(runner.statLogger, "", 0),
 685                 CgroupParent: runner.expectCgroupParent,
 686                 CgroupRoot:   runner.cgroupRoot,
 687                 PollPeriod:   runner.statInterval,
 688         }
 689         runner.statReporter.Start()
 690 }
 691
 692 type infoCommand struct {
 693         label string
 694         cmd   []string
 695 }
 696
 697 // LogHostInfo logs info about the current host, for debugging and
 698 // accounting purposes. Although it's logged as "node-info", this is
 699 // about the environment where crunch-run is actually running, which
 700 // might differ from what's described in the node record (see
 701 // LogNodeRecord).
 702 func (runner *ContainerRunner) LogHostInfo() (err error) {
 703         w := runner.NewLogWriter("node-info")
 704
 705         commands := []infoCommand{
 706                 {
 707                         label: "Host Information",
 708                         cmd:   []string{"uname", "-a"},
 709                 },
 710                 {
 711                         label: "CPU Information",
 712                         cmd:   []string{"cat", "/proc/cpuinfo"},
 713                 },
 714                 {
 715                         label: "Memory Information",
 716                         cmd:   []string{"cat", "/proc/meminfo"},
 717                 },
 718                 {
 719                         label: "Disk Space",
 720                         cmd:   []string{"df", "-m", "/", os.TempDir()},
 721                 },
 722                 {
 723                         label: "Disk INodes",
 724                         cmd:   []string{"df", "-i", "/", os.TempDir()},
 725                 },
 726         }
 727
 728         // Run commands with informational output to be logged.
 729         for _, command := range commands {
 730                 fmt.Fprintln(w, command.label)
 731                 cmd := exec.Command(command.cmd[0], command.cmd[1:]...)
 732                 cmd.Stdout = w
 733                 cmd.Stderr = w
 734                 if err := cmd.Run(); err != nil {
 735                         err = fmt.Errorf("While running command %q: %v", command.cmd, err)
 736                         fmt.Fprintln(w, err)
 737                         return err
 738                 }
 739                 fmt.Fprintln(w, "")
 740         }
 741
 742         err = w.Close()
 743         if err != nil {
 744                 return fmt.Errorf("While closing node-info logs: %v", err)
 745         }
 746         return nil
 747 }
 748
 749 // LogContainerRecord gets and saves the raw JSON container record from the API server
 750 func (runner *ContainerRunner) LogContainerRecord() error {
 751         logged, err := runner.logAPIResponse("container", "containers", map[string]interface{}{"filters": [][]string{{"uuid", "=", runner.Container.UUID}}}, nil)
 752         if !logged && err == nil {
 753                 err = fmt.Errorf("error: no container record found for %s", runner.Container.UUID)
 754         }
 755         return err
 756 }
 757
 758 // LogNodeRecord logs arvados#node record corresponding to the current host.
 759 func (runner *ContainerRunner) LogNodeRecord() error {
 760         hostname := os.Getenv("SLURMD_NODENAME")
 761         if hostname == "" {
 762                 hostname, _ = os.Hostname()
 763         }
 764         _, err := runner.logAPIResponse("node", "nodes", map[string]interface{}{"filters": [][]string{{"hostname", "=", hostname}}}, func(resp interface{}) {
 765                 // The "info" field has admin-only info when obtained
 766                 // with a privileged token, and should not be logged.
 767                 node, ok := resp.(map[string]interface{})
 768                 if ok {
 769                         delete(node, "info")
 770                 }
 771         })
 772         return err
 773 }
 774
 775 func (runner *ContainerRunner) logAPIResponse(label, path string, params map[string]interface{}, munge func(interface{})) (logged bool, err error) {
 776         w := &ArvLogWriter{
 777                 ArvClient:     runner.ArvClient,
 778                 UUID:          runner.Container.UUID,
 779                 loggingStream: label,
 780                 writeCloser:   runner.LogCollection.Open(label + ".json"),
 781         }
 782
 783         reader, err := runner.ArvClient.CallRaw("GET", path, "", "", arvadosclient.Dict(params))
 784         if err != nil {
 785                 return false, fmt.Errorf("error getting %s record: %v", label, err)
 786         }
 787         defer reader.Close()
 788
 789         dec := json.NewDecoder(reader)
 790         dec.UseNumber()
 791         var resp map[string]interface{}
 792         if err = dec.Decode(&resp); err != nil {
 793                 return false, fmt.Errorf("error decoding %s list response: %v", label, err)
 794         }
 795         items, ok := resp["items"].([]interface{})
 796         if !ok {
 797                 return false, fmt.Errorf("error decoding %s list response: no \"items\" key in API list response", label)
 798         } else if len(items) < 1 {
 799                 return false, nil
 800         }
 801         if munge != nil {
 802                 munge(items[0])
 803         }
 804         // Re-encode it using indentation to improve readability
 805         enc := json.NewEncoder(w)
 806         enc.SetIndent("", "    ")
 807         if err = enc.Encode(items[0]); err != nil {
 808                 return false, fmt.Errorf("error logging %s record: %v", label, err)
 809         }
 810         err = w.Close()
 811         if err != nil {
 812                 return false, fmt.Errorf("error closing %s.json in log collection: %v", label, err)
 813         }
 814         return true, nil
 815 }
 816
 817 // AttachStreams connects the docker container stdin, stdout and stderr logs
 818 // to the Arvados logger which logs to Keep and the API server logs table.
 819 func (runner *ContainerRunner) AttachStreams() (err error) {
 820
 821         runner.CrunchLog.Print("Attaching container streams")
 822
 823         // If stdin mount is provided, attach it to the docker container
 824         var stdinRdr arvados.File
 825         var stdinJson []byte
 826         if stdinMnt, ok := runner.Container.Mounts["stdin"]; ok {
 827                 if stdinMnt.Kind == "collection" {
 828                         var stdinColl arvados.Collection
 829                         collId := stdinMnt.UUID
 830                         if collId == "" {
 831                                 collId = stdinMnt.PortableDataHash
 832                         }
 833                         err = runner.ArvClient.Get("collections", collId, nil, &stdinColl)
 834                         if err != nil {
 835                                 return fmt.Errorf("While getting stding collection: %v", err)
 836                         }
 837
 838                         stdinRdr, err = runner.Kc.ManifestFileReader(manifest.Manifest{Text: stdinColl.ManifestText}, stdinMnt.Path)
 839                         if os.IsNotExist(err) {
 840                                 return fmt.Errorf("stdin collection path not found: %v", stdinMnt.Path)
 841                         } else if err != nil {
 842                                 return fmt.Errorf("While getting stdin collection path %v: %v", stdinMnt.Path, err)
 843                         }
 844                 } else if stdinMnt.Kind == "json" {
 845                         stdinJson, err = json.Marshal(stdinMnt.Content)
 846                         if err != nil {
 847                                 return fmt.Errorf("While encoding stdin json data: %v", err)
 848                         }
 849                 }
 850         }
 851
 852         stdinUsed := stdinRdr != nil || len(stdinJson) != 0
 853         response, err := runner.Docker.ContainerAttach(context.TODO(), runner.ContainerID,
 854                 dockertypes.ContainerAttachOptions{Stream: true, Stdin: stdinUsed, Stdout: true, Stderr: true})
 855         if err != nil {
 856                 return fmt.Errorf("While attaching container stdout/stderr streams: %v", err)
 857         }
 858
 859         runner.loggingDone = make(chan bool)
 860
 861         if stdoutMnt, ok := runner.Container.Mounts["stdout"]; ok {
 862                 stdoutFile, err := runner.getStdoutFile(stdoutMnt.Path)
 863                 if err != nil {
 864                         return err
 865                 }
 866                 runner.Stdout = stdoutFile
 867         } else {
 868                 runner.Stdout = NewThrottledLogger(runner.NewLogWriter("stdout"))
 869         }
 870
 871         if stderrMnt, ok := runner.Container.Mounts["stderr"]; ok {
 872                 stderrFile, err := runner.getStdoutFile(stderrMnt.Path)
 873                 if err != nil {
 874                         return err
 875                 }
 876                 runner.Stderr = stderrFile
 877         } else {
 878                 runner.Stderr = NewThrottledLogger(runner.NewLogWriter("stderr"))
 879         }
 880
 881         if stdinRdr != nil {
 882                 go func() {
 883                         _, err := io.Copy(response.Conn, stdinRdr)
 884                         if err != nil {
 885                                 runner.CrunchLog.Print("While writing stdin collection to docker container %q", err)
 886                                 runner.stop()
 887                         }
 888                         stdinRdr.Close()
 889                         response.CloseWrite()
 890                 }()
 891         } else if len(stdinJson) != 0 {
 892                 go func() {
 893                         _, err := io.Copy(response.Conn, bytes.NewReader(stdinJson))
 894                         if err != nil {
 895                                 runner.CrunchLog.Print("While writing stdin json to docker container %q", err)
 896                                 runner.stop()
 897                         }
 898                         response.CloseWrite()
 899                 }()
 900         }
 901
 902         go runner.ProcessDockerAttach(response.Reader)
 903
 904         return nil
 905 }
 906
 907 func (runner *ContainerRunner) getStdoutFile(mntPath string) (*os.File, error) {
 908         stdoutPath := mntPath[len(runner.Container.OutputPath):]
 909         index := strings.LastIndex(stdoutPath, "/")
 910         if index > 0 {
 911                 subdirs := stdoutPath[:index]
 912                 if subdirs != "" {
 913                         st, err := os.Stat(runner.HostOutputDir)
 914                         if err != nil {
 915                                 return nil, fmt.Errorf("While Stat on temp dir: %v", err)
 916                         }
 917                         stdoutPath := filepath.Join(runner.HostOutputDir, subdirs)
 918                         err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777)
 919                         if err != nil {
 920                                 return nil, fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err)
 921                         }
 922                 }
 923         }
 924         stdoutFile, err := os.Create(filepath.Join(runner.HostOutputDir, stdoutPath))
 925         if err != nil {
 926                 return nil, fmt.Errorf("While creating file %q: %v", stdoutPath, err)
 927         }
 928
 929         return stdoutFile, nil
 930 }
 931
 932 // CreateContainer creates the docker container.
 933 func (runner *ContainerRunner) CreateContainer() error {
 934         runner.CrunchLog.Print("Creating Docker container")
 935
 936         runner.ContainerConfig.Cmd = runner.Container.Command
 937         if runner.Container.Cwd != "." {
 938                 runner.ContainerConfig.WorkingDir = runner.Container.Cwd
 939         }
 940
 941         for k, v := range runner.Container.Environment {
 942                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env, k+"="+v)
 943         }
 944
 945         runner.ContainerConfig.Volumes = runner.Volumes
 946
 947         runner.HostConfig = dockercontainer.HostConfig{
 948                 Binds: runner.Binds,
 949                 LogConfig: dockercontainer.LogConfig{
 950                         Type: "none",
 951                 },
 952                 Resources: dockercontainer.Resources{
 953                         CgroupParent: runner.setCgroupParent,
 954                 },
 955         }
 956
 957         if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
 958                 tok, err := runner.ContainerToken()
 959                 if err != nil {
 960                         return err
 961                 }
 962                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env,
 963                         "ARVADOS_API_TOKEN="+tok,
 964                         "ARVADOS_API_HOST="+os.Getenv("ARVADOS_API_HOST"),
 965                         "ARVADOS_API_HOST_INSECURE="+os.Getenv("ARVADOS_API_HOST_INSECURE"),
 966                 )
 967                 runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
 968         } else {
 969                 if runner.enableNetwork == "always" {
 970                         runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
 971                 } else {
 972                         runner.HostConfig.NetworkMode = dockercontainer.NetworkMode("none")
 973                 }
 974         }
 975
 976         _, stdinUsed := runner.Container.Mounts["stdin"]
 977         runner.ContainerConfig.OpenStdin = stdinUsed
 978         runner.ContainerConfig.StdinOnce = stdinUsed
 979         runner.ContainerConfig.AttachStdin = stdinUsed
 980         runner.ContainerConfig.AttachStdout = true
 981         runner.ContainerConfig.AttachStderr = true
 982
 983         createdBody, err := runner.Docker.ContainerCreate(context.TODO(), &runner.ContainerConfig, &runner.HostConfig, nil, runner.Container.UUID)
 984         if err != nil {
 985                 return fmt.Errorf("While creating container: %v", err)
 986         }
 987
 988         runner.ContainerID = createdBody.ID
 989
 990         return runner.AttachStreams()
 991 }
 992
 993 // StartContainer starts the docker container created by CreateContainer.
 994 func (runner *ContainerRunner) StartContainer() error {
 995         runner.CrunchLog.Printf("Starting Docker container id '%s'", runner.ContainerID)
 996         runner.cStateLock.Lock()
 997         defer runner.cStateLock.Unlock()
 998         if runner.cCancelled {
 999                 return ErrCancelled
1000         }
1001         err := runner.Docker.ContainerStart(context.TODO(), runner.ContainerID,
1002                 dockertypes.ContainerStartOptions{})
1003         if err != nil {
1004                 var advice string
1005                 if m, e := regexp.MatchString("(?ms).*(exec|System error).*(no such file or directory|file not found).*", err.Error()); m && e == nil {
1006                         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])
1007                 }
1008                 return fmt.Errorf("could not start container: %v%s", err, advice)
1009         }
1010         return nil
1011 }
1012
1013 // WaitFinish waits for the container to terminate, capture the exit code, and
1014 // close the stdout/stderr logging.
1015 func (runner *ContainerRunner) WaitFinish() error {
1016         runner.CrunchLog.Print("Waiting for container to finish")
1017
1018         waitOk, waitErr := runner.Docker.ContainerWait(context.TODO(), runner.ContainerID, dockercontainer.WaitConditionNotRunning)
1019         arvMountExit := runner.ArvMountExit
1020         for {
1021                 select {
1022                 case waitBody := <-waitOk:
1023                         runner.CrunchLog.Printf("Container exited with code: %v", waitBody.StatusCode)
1024                         code := int(waitBody.StatusCode)
1025                         runner.ExitCode = &code
1026
1027                         // wait for stdout/stderr to complete
1028                         <-runner.loggingDone
1029                         return nil
1030
1031                 case err := <-waitErr:
1032                         return fmt.Errorf("container wait: %v", err)
1033
1034                 case <-arvMountExit:
1035                         runner.CrunchLog.Printf("arv-mount exited while container is still running.  Stopping container.")
1036                         runner.stop()
1037                         // arvMountExit will always be ready now that
1038                         // it's closed, but that doesn't interest us.
1039                         arvMountExit = nil
1040                 }
1041         }
1042 }
1043
1044 var ErrNotInOutputDir = fmt.Errorf("Must point to path within the output directory")
1045
1046 func (runner *ContainerRunner) derefOutputSymlink(path string, startinfo os.FileInfo) (tgt string, readlinktgt string, info os.FileInfo, err error) {
1047         // Follow symlinks if necessary
1048         info = startinfo
1049         tgt = path
1050         readlinktgt = ""
1051         nextlink := path
1052         for followed := 0; info.Mode()&os.ModeSymlink != 0; followed++ {
1053                 if followed >= limitFollowSymlinks {
1054                         // Got stuck in a loop or just a pathological number of links, give up.
1055                         err = fmt.Errorf("Followed more than %v symlinks from path %q", limitFollowSymlinks, path)
1056                         return
1057                 }
1058
1059                 readlinktgt, err = os.Readlink(nextlink)
1060                 if err != nil {
1061                         return
1062                 }
1063
1064                 tgt = readlinktgt
1065                 if !strings.HasPrefix(tgt, "/") {
1066                         // Relative symlink, resolve it to host path
1067                         tgt = filepath.Join(filepath.Dir(path), tgt)
1068                 }
1069                 if strings.HasPrefix(tgt, runner.Container.OutputPath+"/") && !strings.HasPrefix(tgt, runner.HostOutputDir+"/") {
1070                         // Absolute symlink to container output path, adjust it to host output path.
1071                         tgt = filepath.Join(runner.HostOutputDir, tgt[len(runner.Container.OutputPath):])
1072                 }
1073                 if !strings.HasPrefix(tgt, runner.HostOutputDir+"/") {
1074                         // After dereferencing, symlink target must either be
1075                         // within output directory, or must point to a
1076                         // collection mount.
1077                         err = ErrNotInOutputDir
1078                         return
1079                 }
1080
1081                 info, err = os.Lstat(tgt)
1082                 if err != nil {
1083                         // tgt
1084                         err = fmt.Errorf("Symlink in output %q points to invalid location %q: %v",
1085                                 path[len(runner.HostOutputDir):], readlinktgt, err)
1086                         return
1087                 }
1088
1089                 nextlink = tgt
1090         }
1091
1092         return
1093 }
1094
1095 var limitFollowSymlinks = 10
1096
1097 // UploadFile uploads files within the output directory, with special handling
1098 // for symlinks. If the symlink leads to a keep mount, copy the manifest text
1099 // from the keep mount into the output manifestText.  Ensure that whether
1100 // symlinks are relative or absolute, every symlink target (even targets that
1101 // are symlinks themselves) must point to a path in either the output directory
1102 // or a collection mount.
1103 //
1104 // Assumes initial value of "path" is absolute, and located within runner.HostOutputDir.
1105 func (runner *ContainerRunner) UploadOutputFile(
1106         path string,
1107         info os.FileInfo,
1108         infoerr error,
1109         binds []string,
1110         walkUpload *WalkUpload,
1111         relocateFrom string,
1112         relocateTo string,
1113         followed int) (manifestText string, err error) {
1114
1115         if infoerr != nil {
1116                 return "", infoerr
1117         }
1118
1119         if info.Mode().IsDir() {
1120                 // if empty, need to create a .keep file
1121                 dir, direrr := os.Open(path)
1122                 if direrr != nil {
1123                         return "", direrr
1124                 }
1125                 defer dir.Close()
1126                 names, eof := dir.Readdirnames(1)
1127                 if len(names) == 0 && eof == io.EOF && path != runner.HostOutputDir {
1128                         containerPath := runner.OutputPath + path[len(runner.HostOutputDir):]
1129                         for _, bind := range binds {
1130                                 mnt := runner.Container.Mounts[bind]
1131                                 // Check if there is a bind for this
1132                                 // directory, in which case assume we don't need .keep
1133                                 if (containerPath == bind || strings.HasPrefix(containerPath, bind+"/")) && mnt.PortableDataHash != "d41d8cd98f00b204e9800998ecf8427e+0" {
1134                                         return
1135                                 }
1136                         }
1137                         outputSuffix := path[len(runner.HostOutputDir)+1:]
1138                         return fmt.Sprintf("./%v d41d8cd98f00b204e9800998ecf8427e+0 0:0:.keep\n", outputSuffix), nil
1139                 }
1140                 return
1141         }
1142
1143         if followed >= limitFollowSymlinks {
1144                 // Got stuck in a loop or just a pathological number of
1145                 // directory links, give up.
1146                 err = fmt.Errorf("Followed more than %v symlinks from path %q", limitFollowSymlinks, path)
1147                 return
1148         }
1149
1150         // "path" is the actual path we are visiting
1151         // "tgt" is the target of "path" (a non-symlink) after following symlinks
1152         // "relocated" is the path in the output manifest where the file should be placed,
1153         // but has HostOutputDir as a prefix.
1154
1155         // The destination path in the output manifest may need to be
1156         // logically relocated to some other path in order to appear
1157         // in the correct location as a result of following a symlink.
1158         // Remove the relocateFrom prefix and replace it with
1159         // relocateTo.
1160         relocated := relocateTo + path[len(relocateFrom):]
1161
1162         tgt, readlinktgt, info, derefErr := runner.derefOutputSymlink(path, info)
1163         if derefErr != nil && derefErr != ErrNotInOutputDir {
1164                 return "", derefErr
1165         }
1166
1167         // go through mounts and try reverse map to collection reference
1168         for _, bind := range binds {
1169                 mnt := runner.Container.Mounts[bind]
1170                 if (tgt == bind || strings.HasPrefix(tgt, bind+"/")) && !mnt.Writable {
1171                         // get path relative to bind
1172                         targetSuffix := tgt[len(bind):]
1173
1174                         // Copy mount and adjust the path to add path relative to the bind
1175                         adjustedMount := mnt
1176                         adjustedMount.Path = filepath.Join(adjustedMount.Path, targetSuffix)
1177
1178                         // Terminates in this keep mount, so add the
1179                         // manifest text at appropriate location.
1180                         outputSuffix := relocated[len(runner.HostOutputDir):]
1181                         manifestText, err = runner.getCollectionManifestForPath(adjustedMount, outputSuffix)
1182                         return
1183                 }
1184         }
1185
1186         // If target is not a collection mount, it must be located within the
1187         // output directory, otherwise it is an error.
1188         if derefErr == ErrNotInOutputDir {
1189                 err = fmt.Errorf("Symlink in output %q points to invalid location %q, must point to path within the output directory.",
1190                         path[len(runner.HostOutputDir):], readlinktgt)
1191                 return
1192         }
1193
1194         if info.Mode().IsRegular() {
1195                 return "", walkUpload.UploadFile(relocated, tgt)
1196         }
1197
1198         if info.Mode().IsDir() {
1199                 // Symlink leads to directory.  Walk() doesn't follow
1200                 // directory symlinks, so we walk the target directory
1201                 // instead.  Within the walk, file paths are relocated
1202                 // so they appear under the original symlink path.
1203                 err = filepath.Walk(tgt, func(walkpath string, walkinfo os.FileInfo, walkerr error) error {
1204                         var m string
1205                         m, walkerr = runner.UploadOutputFile(walkpath, walkinfo, walkerr,
1206                                 binds, walkUpload, tgt, relocated, followed+1)
1207                         if walkerr == nil {
1208                                 manifestText = manifestText + m
1209                         }
1210                         return walkerr
1211                 })
1212                 return
1213         }
1214
1215         return
1216 }
1217
1218 // HandleOutput sets the output, unmounts the FUSE mount, and deletes temporary directories
1219 func (runner *ContainerRunner) CaptureOutput() error {
1220         if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
1221                 // Output may have been set directly by the container, so
1222                 // refresh the container record to check.
1223                 err := runner.ArvClient.Get("containers", runner.Container.UUID,
1224                         nil, &runner.Container)
1225                 if err != nil {
1226                         return err
1227                 }
1228                 if runner.Container.Output != "" {
1229                         // Container output is already set.
1230                         runner.OutputPDH = &runner.Container.Output
1231                         return nil
1232                 }
1233         }
1234
1235         if runner.HostOutputDir == "" {
1236                 return nil
1237         }
1238
1239         _, err := os.Stat(runner.HostOutputDir)
1240         if err != nil {
1241                 return fmt.Errorf("While checking host output path: %v", err)
1242         }
1243
1244         // Pre-populate output from the configured mount points
1245         var binds []string
1246         for bind, mnt := range runner.Container.Mounts {
1247                 if mnt.Kind == "collection" {
1248                         binds = append(binds, bind)
1249                 }
1250         }
1251         sort.Strings(binds)
1252
1253         var manifestText string
1254
1255         collectionMetafile := fmt.Sprintf("%s/.arvados#collection", runner.HostOutputDir)
1256         _, err = os.Stat(collectionMetafile)
1257         if err != nil {
1258                 // Regular directory
1259
1260                 cw := CollectionWriter{0, runner.Kc, nil, nil, sync.Mutex{}}
1261                 walkUpload := cw.BeginUpload(runner.HostOutputDir, runner.CrunchLog.Logger)
1262
1263                 var m string
1264                 err = filepath.Walk(runner.HostOutputDir, func(path string, info os.FileInfo, err error) error {
1265                         m, err = runner.UploadOutputFile(path, info, err, binds, walkUpload, "", "", 0)
1266                         if err == nil {
1267                                 manifestText = manifestText + m
1268                         }
1269                         return err
1270                 })
1271
1272                 cw.EndUpload(walkUpload)
1273
1274                 if err != nil {
1275                         return fmt.Errorf("While uploading output files: %v", err)
1276                 }
1277
1278                 m, err = cw.ManifestText()
1279                 manifestText = manifestText + m
1280                 if err != nil {
1281                         return fmt.Errorf("While uploading output files: %v", err)
1282                 }
1283         } else {
1284                 // FUSE mount directory
1285                 file, openerr := os.Open(collectionMetafile)
1286                 if openerr != nil {
1287                         return fmt.Errorf("While opening FUSE metafile: %v", err)
1288                 }
1289                 defer file.Close()
1290
1291                 var rec arvados.Collection
1292                 err = json.NewDecoder(file).Decode(&rec)
1293                 if err != nil {
1294                         return fmt.Errorf("While reading FUSE metafile: %v", err)
1295                 }
1296                 manifestText = rec.ManifestText
1297         }
1298
1299         for _, bind := range binds {
1300                 mnt := runner.Container.Mounts[bind]
1301
1302                 bindSuffix := strings.TrimPrefix(bind, runner.Container.OutputPath)
1303
1304                 if bindSuffix == bind || len(bindSuffix) <= 0 {
1305                         // either does not start with OutputPath or is OutputPath itself
1306                         continue
1307                 }
1308
1309                 if mnt.ExcludeFromOutput == true || mnt.Writable {
1310                         continue
1311                 }
1312
1313                 // append to manifest_text
1314                 m, err := runner.getCollectionManifestForPath(mnt, bindSuffix)
1315                 if err != nil {
1316                         return err
1317                 }
1318
1319                 manifestText = manifestText + m
1320         }
1321
1322         // Save output
1323         var response arvados.Collection
1324         manifest := manifest.Manifest{Text: manifestText}
1325         manifestText = manifest.Extract(".", ".").Text
1326         err = runner.ArvClient.Create("collections",
1327                 arvadosclient.Dict{
1328                         "ensure_unique_name": true,
1329                         "collection": arvadosclient.Dict{
1330                                 "is_trashed":    true,
1331                                 "name":          "output for " + runner.Container.UUID,
1332                                 "manifest_text": manifestText}},
1333                 &response)
1334         if err != nil {
1335                 return fmt.Errorf("While creating output collection: %v", err)
1336         }
1337         runner.OutputPDH = &response.PortableDataHash
1338         return nil
1339 }
1340
1341 var outputCollections = make(map[string]arvados.Collection)
1342
1343 // Fetch the collection for the mnt.PortableDataHash
1344 // Return the manifest_text fragment corresponding to the specified mnt.Path
1345 //  after making any required updates.
1346 //  Ex:
1347 //    If mnt.Path is not specified,
1348 //      return the entire manifest_text after replacing any "." with bindSuffix
1349 //    If mnt.Path corresponds to one stream,
1350 //      return the manifest_text for that stream after replacing that stream name with bindSuffix
1351 //    Otherwise, check if a filename in any one stream is being sought. Return the manifest_text
1352 //      for that stream after replacing stream name with bindSuffix minus the last word
1353 //      and the file name with last word of the bindSuffix
1354 //  Allowed path examples:
1355 //    "path":"/"
1356 //    "path":"/subdir1"
1357 //    "path":"/subdir1/subdir2"
1358 //    "path":"/subdir/filename" etc
1359 func (runner *ContainerRunner) getCollectionManifestForPath(mnt arvados.Mount, bindSuffix string) (string, error) {
1360         collection := outputCollections[mnt.PortableDataHash]
1361         if collection.PortableDataHash == "" {
1362                 err := runner.ArvClient.Get("collections", mnt.PortableDataHash, nil, &collection)
1363                 if err != nil {
1364                         return "", fmt.Errorf("While getting collection for %v: %v", mnt.PortableDataHash, err)
1365                 }
1366                 outputCollections[mnt.PortableDataHash] = collection
1367         }
1368
1369         if collection.ManifestText == "" {
1370                 runner.CrunchLog.Printf("No manifest text for collection %v", collection.PortableDataHash)
1371                 return "", nil
1372         }
1373
1374         mft := manifest.Manifest{Text: collection.ManifestText}
1375         extracted := mft.Extract(mnt.Path, bindSuffix)
1376         if extracted.Err != nil {
1377                 return "", fmt.Errorf("Error parsing manifest for %v: %v", mnt.PortableDataHash, extracted.Err.Error())
1378         }
1379         return extracted.Text, nil
1380 }
1381
1382 func (runner *ContainerRunner) CleanupDirs() {
1383         if runner.ArvMount != nil {
1384                 var delay int64 = 8
1385                 umount := exec.Command("arv-mount", fmt.Sprintf("--unmount-timeout=%d", delay), "--unmount", runner.ArvMountPoint)
1386                 umount.Stdout = runner.CrunchLog
1387                 umount.Stderr = runner.CrunchLog
1388                 runner.CrunchLog.Printf("Running %v", umount.Args)
1389                 umnterr := umount.Start()
1390
1391                 if umnterr != nil {
1392                         runner.CrunchLog.Printf("Error unmounting: %v", umnterr)
1393                 } else {
1394                         // If arv-mount --unmount gets stuck for any reason, we
1395                         // don't want to wait for it forever.  Do Wait() in a goroutine
1396                         // so it doesn't block crunch-run.
1397                         umountExit := make(chan error)
1398                         go func() {
1399                                 mnterr := umount.Wait()
1400                                 if mnterr != nil {
1401                                         runner.CrunchLog.Printf("Error unmounting: %v", mnterr)
1402                                 }
1403                                 umountExit <- mnterr
1404                         }()
1405
1406                         for again := true; again; {
1407                                 again = false
1408                                 select {
1409                                 case <-umountExit:
1410                                         umount = nil
1411                                         again = true
1412                                 case <-runner.ArvMountExit:
1413                                         break
1414                                 case <-time.After(time.Duration((delay + 1) * int64(time.Second))):
1415                                         runner.CrunchLog.Printf("Timed out waiting for unmount")
1416                                         if umount != nil {
1417                                                 umount.Process.Kill()
1418                                         }
1419                                         runner.ArvMount.Process.Kill()
1420                                 }
1421                         }
1422                 }
1423         }
1424
1425         if runner.ArvMountPoint != "" {
1426                 if rmerr := os.Remove(runner.ArvMountPoint); rmerr != nil {
1427                         runner.CrunchLog.Printf("While cleaning up arv-mount directory %s: %v", runner.ArvMountPoint, rmerr)
1428                 }
1429         }
1430
1431         for _, tmpdir := range runner.CleanupTempDir {
1432                 if rmerr := os.RemoveAll(tmpdir); rmerr != nil {
1433                         runner.CrunchLog.Printf("While cleaning up temporary directory %s: %v", tmpdir, rmerr)
1434                 }
1435         }
1436
1437         if rmerr := os.RemoveAll(runner.parentTemp); rmerr != nil {
1438                 runner.CrunchLog.Printf("While cleaning up temporary directory %s: %v", runner.parentTemp, rmerr)
1439         }
1440 }
1441
1442 // CommitLogs posts the collection containing the final container logs.
1443 func (runner *ContainerRunner) CommitLogs() error {
1444         runner.CrunchLog.Print(runner.finalState)
1445
1446         if runner.arvMountLog != nil {
1447                 runner.arvMountLog.Close()
1448         }
1449         runner.CrunchLog.Close()
1450
1451         // Closing CrunchLog above allows them to be committed to Keep at this
1452         // point, but re-open crunch log with ArvClient in case there are any
1453         // other further errors (such as failing to write the log to Keep!)
1454         // while shutting down
1455         runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{ArvClient: runner.ArvClient,
1456                 UUID: runner.Container.UUID, loggingStream: "crunch-run", writeCloser: nil})
1457         runner.CrunchLog.Immediate = log.New(os.Stderr, runner.Container.UUID+" ", 0)
1458
1459         if runner.LogsPDH != nil {
1460                 // If we have already assigned something to LogsPDH,
1461                 // we must be closing the re-opened log, which won't
1462                 // end up getting attached to the container record and
1463                 // therefore doesn't need to be saved as a collection
1464                 // -- it exists only to send logs to other channels.
1465                 return nil
1466         }
1467
1468         mt, err := runner.LogCollection.ManifestText()
1469         if err != nil {
1470                 return fmt.Errorf("While creating log manifest: %v", err)
1471         }
1472
1473         var response arvados.Collection
1474         err = runner.ArvClient.Create("collections",
1475                 arvadosclient.Dict{
1476                         "ensure_unique_name": true,
1477                         "collection": arvadosclient.Dict{
1478                                 "is_trashed":    true,
1479                                 "name":          "logs for " + runner.Container.UUID,
1480                                 "manifest_text": mt}},
1481                 &response)
1482         if err != nil {
1483                 return fmt.Errorf("While creating log collection: %v", err)
1484         }
1485         runner.LogsPDH = &response.PortableDataHash
1486         return nil
1487 }
1488
1489 // UpdateContainerRunning updates the container state to "Running"
1490 func (runner *ContainerRunner) UpdateContainerRunning() error {
1491         runner.cStateLock.Lock()
1492         defer runner.cStateLock.Unlock()
1493         if runner.cCancelled {
1494                 return ErrCancelled
1495         }
1496         return runner.ArvClient.Update("containers", runner.Container.UUID,
1497                 arvadosclient.Dict{"container": arvadosclient.Dict{"state": "Running"}}, nil)
1498 }
1499
1500 // ContainerToken returns the api_token the container (and any
1501 // arv-mount processes) are allowed to use.
1502 func (runner *ContainerRunner) ContainerToken() (string, error) {
1503         if runner.token != "" {
1504                 return runner.token, nil
1505         }
1506
1507         var auth arvados.APIClientAuthorization
1508         err := runner.ArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth)
1509         if err != nil {
1510                 return "", err
1511         }
1512         runner.token = auth.APIToken
1513         return runner.token, nil
1514 }
1515
1516 // UpdateContainerComplete updates the container record state on API
1517 // server to "Complete" or "Cancelled"
1518 func (runner *ContainerRunner) UpdateContainerFinal() error {
1519         update := arvadosclient.Dict{}
1520         update["state"] = runner.finalState
1521         if runner.LogsPDH != nil {
1522                 update["log"] = *runner.LogsPDH
1523         }
1524         if runner.finalState == "Complete" {
1525                 if runner.ExitCode != nil {
1526                         update["exit_code"] = *runner.ExitCode
1527                 }
1528                 if runner.OutputPDH != nil {
1529                         update["output"] = *runner.OutputPDH
1530                 }
1531         }
1532         return runner.ArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{"container": update}, nil)
1533 }
1534
1535 // IsCancelled returns the value of Cancelled, with goroutine safety.
1536 func (runner *ContainerRunner) IsCancelled() bool {
1537         runner.cStateLock.Lock()
1538         defer runner.cStateLock.Unlock()
1539         return runner.cCancelled
1540 }
1541
1542 // NewArvLogWriter creates an ArvLogWriter
1543 func (runner *ContainerRunner) NewArvLogWriter(name string) io.WriteCloser {
1544         return &ArvLogWriter{
1545                 ArvClient:     runner.ArvClient,
1546                 UUID:          runner.Container.UUID,
1547                 loggingStream: name,
1548                 writeCloser:   runner.LogCollection.Open(name + ".txt")}
1549 }
1550
1551 // Run the full container lifecycle.
1552 func (runner *ContainerRunner) Run() (err error) {
1553         runner.CrunchLog.Printf("crunch-run %s started", version)
1554         runner.CrunchLog.Printf("Executing container '%s'", runner.Container.UUID)
1555
1556         hostname, hosterr := os.Hostname()
1557         if hosterr != nil {
1558                 runner.CrunchLog.Printf("Error getting hostname '%v'", hosterr)
1559         } else {
1560                 runner.CrunchLog.Printf("Executing on host '%s'", hostname)
1561         }
1562
1563         runner.finalState = "Queued"
1564
1565         defer func() {
1566                 runner.stopSignals()
1567                 runner.CleanupDirs()
1568
1569                 runner.CrunchLog.Printf("crunch-run finished")
1570                 runner.CrunchLog.Close()
1571         }()
1572
1573         defer func() {
1574                 // checkErr prints e (unless it's nil) and sets err to
1575                 // e (unless err is already non-nil). Thus, if err
1576                 // hasn't already been assigned when Run() returns,
1577                 // this cleanup func will cause Run() to return the
1578                 // first non-nil error that is passed to checkErr().
1579                 checkErr := func(e error) {
1580                         if e == nil {
1581                                 return
1582                         }
1583                         runner.CrunchLog.Print(e)
1584                         if err == nil {
1585                                 err = e
1586                         }
1587                         if runner.finalState == "Complete" {
1588                                 // There was an error in the finalization.
1589                                 runner.finalState = "Cancelled"
1590                         }
1591                 }
1592
1593                 // Log the error encountered in Run(), if any
1594                 checkErr(err)
1595
1596                 if runner.finalState == "Queued" {
1597                         runner.UpdateContainerFinal()
1598                         return
1599                 }
1600
1601                 if runner.IsCancelled() {
1602                         runner.finalState = "Cancelled"
1603                         // but don't return yet -- we still want to
1604                         // capture partial output and write logs
1605                 }
1606
1607                 checkErr(runner.CaptureOutput())
1608                 checkErr(runner.stopHoststat())
1609                 checkErr(runner.CommitLogs())
1610                 checkErr(runner.UpdateContainerFinal())
1611         }()
1612
1613         err = runner.fetchContainerRecord()
1614         if err != nil {
1615                 return
1616         }
1617         runner.setupSignals()
1618         runner.startHoststat()
1619
1620         // check for and/or load image
1621         err = runner.LoadImage()
1622         if err != nil {
1623                 if !runner.checkBrokenNode(err) {
1624                         // Failed to load image but not due to a "broken node"
1625                         // condition, probably user error.
1626                         runner.finalState = "Cancelled"
1627                 }
1628                 err = fmt.Errorf("While loading container image: %v", err)
1629                 return
1630         }
1631
1632         // set up FUSE mount and binds
1633         err = runner.SetupMounts()
1634         if err != nil {
1635                 runner.finalState = "Cancelled"
1636                 err = fmt.Errorf("While setting up mounts: %v", err)
1637                 return
1638         }
1639
1640         err = runner.CreateContainer()
1641         if err != nil {
1642                 return
1643         }
1644         err = runner.LogHostInfo()
1645         if err != nil {
1646                 return
1647         }
1648         err = runner.LogNodeRecord()
1649         if err != nil {
1650                 return
1651         }
1652         err = runner.LogContainerRecord()
1653         if err != nil {
1654                 return
1655         }
1656
1657         if runner.IsCancelled() {
1658                 return
1659         }
1660
1661         err = runner.UpdateContainerRunning()
1662         if err != nil {
1663                 return
1664         }
1665         runner.finalState = "Cancelled"
1666
1667         runner.startCrunchstat()
1668
1669         err = runner.StartContainer()
1670         if err != nil {
1671                 runner.checkBrokenNode(err)
1672                 return
1673         }
1674
1675         err = runner.WaitFinish()
1676         if err == nil && !runner.IsCancelled() {
1677                 runner.finalState = "Complete"
1678         }
1679         return
1680 }
1681
1682 // Fetch the current container record (uuid = runner.Container.UUID)
1683 // into runner.Container.
1684 func (runner *ContainerRunner) fetchContainerRecord() error {
1685         reader, err := runner.ArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
1686         if err != nil {
1687                 return fmt.Errorf("error fetching container record: %v", err)
1688         }
1689         defer reader.Close()
1690
1691         dec := json.NewDecoder(reader)
1692         dec.UseNumber()
1693         err = dec.Decode(&runner.Container)
1694         if err != nil {
1695                 return fmt.Errorf("error decoding container record: %v", err)
1696         }
1697         return nil
1698 }
1699
1700 // NewContainerRunner creates a new container runner.
1701 func NewContainerRunner(api IArvadosClient,
1702         kc IKeepClient,
1703         docker ThinDockerClient,
1704         containerUUID string) *ContainerRunner {
1705
1706         cr := &ContainerRunner{ArvClient: api, Kc: kc, Docker: docker}
1707         cr.NewLogWriter = cr.NewArvLogWriter
1708         cr.RunArvMount = cr.ArvMountCmd
1709         cr.MkTempDir = ioutil.TempDir
1710         cr.LogCollection = &CollectionWriter{0, kc, nil, nil, sync.Mutex{}}
1711         cr.Container.UUID = containerUUID
1712         cr.CrunchLog = NewThrottledLogger(cr.NewLogWriter("crunch-run"))
1713         cr.CrunchLog.Immediate = log.New(os.Stderr, containerUUID+" ", 0)
1714
1715         loadLogThrottleParams(api)
1716
1717         return cr
1718 }
1719
1720 func main() {
1721         statInterval := flag.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting")
1722         cgroupRoot := flag.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree")
1723         cgroupParent := flag.String("cgroup-parent", "docker", "name of container's parent cgroup (ignored if -cgroup-parent-subsystem is used)")
1724         cgroupParentSubsystem := flag.String("cgroup-parent-subsystem", "", "use current cgroup for given subsystem as parent cgroup for container")
1725         caCertsPath := flag.String("ca-certs", "", "Path to TLS root certificates")
1726         enableNetwork := flag.String("container-enable-networking", "default",
1727                 `Specify if networking should be enabled for container.  One of 'default', 'always':
1728         default: only enable networking if container requests it.
1729         always:  containers always have networking enabled
1730         `)
1731         networkMode := flag.String("container-network-mode", "default",
1732                 `Set networking mode for container.  Corresponds to Docker network mode (--net).
1733         `)
1734         memprofile := flag.String("memprofile", "", "write memory profile to `file` after running container")
1735         getVersion := flag.Bool("version", false, "Print version information and exit.")
1736         flag.Parse()
1737
1738         // Print version information if requested
1739         if *getVersion {
1740                 fmt.Printf("crunch-run %s\n", version)
1741                 return
1742         }
1743
1744         log.Printf("crunch-run %s started", version)
1745
1746         containerId := flag.Arg(0)
1747
1748         if *caCertsPath != "" {
1749                 arvadosclient.CertFiles = []string{*caCertsPath}
1750         }
1751
1752         api, err := arvadosclient.MakeArvadosClient()
1753         if err != nil {
1754                 log.Fatalf("%s: %v", containerId, err)
1755         }
1756         api.Retries = 8
1757
1758         kc, kcerr := keepclient.MakeKeepClient(api)
1759         if kcerr != nil {
1760                 log.Fatalf("%s: %v", containerId, kcerr)
1761         }
1762         kc.BlockCache = &keepclient.BlockCache{MaxBlocks: 2}
1763         kc.Retries = 4
1764
1765         // API version 1.21 corresponds to Docker 1.9, which is currently the
1766         // minimum version we want to support.
1767         docker, dockererr := dockerclient.NewClient(dockerclient.DefaultDockerHost, "1.21", nil, nil)
1768
1769         cr := NewContainerRunner(api, kc, docker, containerId)
1770         if dockererr != nil {
1771                 cr.CrunchLog.Printf("%s: %v", containerId, dockererr)
1772                 cr.checkBrokenNode(dockererr)
1773                 cr.CrunchLog.Close()
1774                 os.Exit(1)
1775         }
1776
1777         parentTemp, tmperr := cr.MkTempDir("", "crunch-run")
1778         if tmperr != nil {
1779                 log.Fatalf("%s: %v", containerId, tmperr)
1780         }
1781
1782         cr.parentTemp = parentTemp
1783         cr.statInterval = *statInterval
1784         cr.cgroupRoot = *cgroupRoot
1785         cr.expectCgroupParent = *cgroupParent
1786         cr.enableNetwork = *enableNetwork
1787         cr.networkMode = *networkMode
1788         if *cgroupParentSubsystem != "" {
1789                 p := findCgroup(*cgroupParentSubsystem)
1790                 cr.setCgroupParent = p
1791                 cr.expectCgroupParent = p
1792         }
1793
1794         runerr := cr.Run()
1795
1796         if *memprofile != "" {
1797                 f, err := os.Create(*memprofile)
1798                 if err != nil {
1799                         log.Printf("could not create memory profile: ", err)
1800                 }
1801                 runtime.GC() // get up-to-date statistics
1802                 if err := pprof.WriteHeapProfile(f); err != nil {
1803                         log.Printf("could not write memory profile: ", err)
1804                 }
1805                 closeerr := f.Close()
1806                 if closeerr != nil {
1807                         log.Printf("closing memprofile file: ", err)
1808                 }
1809         }
1810
1811         if runerr != nil {
1812                 log.Fatalf("%s: %v", containerId, runerr)
1813         }
1814 }