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         Binds         []string
 110         Volumes       map[string]struct{}
 111         OutputPDH     *string
 112         SigChan       chan os.Signal
 113         ArvMountExit  chan error
 114         finalState    string
 115         parentTemp    string
 116
 117         statLogger       io.WriteCloser
 118         statReporter     *crunchstat.Reporter
 119         hoststatLogger   io.WriteCloser
 120         hoststatReporter *crunchstat.Reporter
 121         statInterval     time.Duration
 122         cgroupRoot       string
 123         // What we expect the container's cgroup parent to be.
 124         expectCgroupParent string
 125         // What we tell docker to use as the container's cgroup
 126         // parent. Note: Ideally we would use the same field for both
 127         // expectCgroupParent and setCgroupParent, and just make it
 128         // default to "docker". However, when using docker < 1.10 with
 129         // systemd, specifying a non-empty cgroup parent (even the
 130         // default value "docker") hits a docker bug
 131         // (https://github.com/docker/docker/issues/17126). Using two
 132         // separate fields makes it possible to use the "expect cgroup
 133         // parent to be X" feature even on sites where the "specify
 134         // cgroup parent" feature breaks.
 135         setCgroupParent string
 136
 137         cStateLock sync.Mutex
 138         cCancelled bool // StopContainer() invoked
 139
 140         enableNetwork string // one of "default" or "always"
 141         networkMode   string // passed through to HostConfig.NetworkMode
 142         arvMountLog   *ThrottledLogger
 143 }
 144
 145 // setupSignals sets up signal handling to gracefully terminate the underlying
 146 // Docker container and update state when receiving a TERM, INT or QUIT signal.
 147 func (runner *ContainerRunner) setupSignals() {
 148         runner.SigChan = make(chan os.Signal, 1)
 149         signal.Notify(runner.SigChan, syscall.SIGTERM)
 150         signal.Notify(runner.SigChan, syscall.SIGINT)
 151         signal.Notify(runner.SigChan, syscall.SIGQUIT)
 152
 153         go func(sig chan os.Signal) {
 154                 for s := range sig {
 155                         runner.CrunchLog.Printf("caught signal: %v", s)
 156                         runner.stop()
 157                 }
 158         }(runner.SigChan)
 159 }
 160
 161 // stop the underlying Docker container.
 162 func (runner *ContainerRunner) stop() {
 163         runner.cStateLock.Lock()
 164         defer runner.cStateLock.Unlock()
 165         if runner.ContainerID == "" {
 166                 return
 167         }
 168         runner.cCancelled = true
 169         runner.CrunchLog.Printf("removing container")
 170         err := runner.Docker.ContainerRemove(context.TODO(), runner.ContainerID, dockertypes.ContainerRemoveOptions{Force: true})
 171         if err != nil {
 172                 runner.CrunchLog.Printf("error removing container: %s", err)
 173         }
 174 }
 175
 176 func (runner *ContainerRunner) stopSignals() {
 177         if runner.SigChan != nil {
 178                 signal.Stop(runner.SigChan)
 179         }
 180 }
 181
 182 var errorBlacklist = []string{
 183         "(?ms).*[Cc]annot connect to the Docker daemon.*",
 184         "(?ms).*oci runtime error.*starting container process.*container init.*mounting.*to rootfs.*no such file or directory.*",
 185 }
 186 var brokenNodeHook *string = flag.String("broken-node-hook", "", "Script to run if node is detected to be broken (for example, Docker daemon is not running)")
 187
 188 func (runner *ContainerRunner) checkBrokenNode(goterr error) bool {
 189         for _, d := range errorBlacklist {
 190                 if m, e := regexp.MatchString(d, goterr.Error()); m && e == nil {
 191                         runner.CrunchLog.Printf("Error suggests node is unable to run containers: %v", goterr)
 192                         if *brokenNodeHook == "" {
 193                                 runner.CrunchLog.Printf("No broken node hook provided, cannot mark node as broken.")
 194                         } else {
 195                                 runner.CrunchLog.Printf("Running broken node hook %q", *brokenNodeHook)
 196                                 // run killme script
 197                                 c := exec.Command(*brokenNodeHook)
 198                                 c.Stdout = runner.CrunchLog
 199                                 c.Stderr = runner.CrunchLog
 200                                 err := c.Run()
 201                                 if err != nil {
 202                                         runner.CrunchLog.Printf("Error running broken node hook: %v", err)
 203                                 }
 204                         }
 205                         return true
 206                 }
 207         }
 208         return false
 209 }
 210
 211 // LoadImage determines the docker image id from the container record and
 212 // checks if it is available in the local Docker image store.  If not, it loads
 213 // the image from Keep.
 214 func (runner *ContainerRunner) LoadImage() (err error) {
 215
 216         runner.CrunchLog.Printf("Fetching Docker image from collection '%s'", runner.Container.ContainerImage)
 217
 218         var collection arvados.Collection
 219         err = runner.ArvClient.Get("collections", runner.Container.ContainerImage, nil, &collection)
 220         if err != nil {
 221                 return fmt.Errorf("While getting container image collection: %v", err)
 222         }
 223         manifest := manifest.Manifest{Text: collection.ManifestText}
 224         var img, imageID string
 225         for ms := range manifest.StreamIter() {
 226                 img = ms.FileStreamSegments[0].Name
 227                 if !strings.HasSuffix(img, ".tar") {
 228                         return fmt.Errorf("First file in the container image collection does not end in .tar")
 229                 }
 230                 imageID = img[:len(img)-4]
 231         }
 232
 233         runner.CrunchLog.Printf("Using Docker image id '%s'", imageID)
 234
 235         _, _, err = runner.Docker.ImageInspectWithRaw(context.TODO(), imageID)
 236         if err != nil {
 237                 runner.CrunchLog.Print("Loading Docker image from keep")
 238
 239                 var readCloser io.ReadCloser
 240                 readCloser, err = runner.Kc.ManifestFileReader(manifest, img)
 241                 if err != nil {
 242                         return fmt.Errorf("While creating ManifestFileReader for container image: %v", err)
 243                 }
 244
 245                 response, err := runner.Docker.ImageLoad(context.TODO(), readCloser, true)
 246                 if err != nil {
 247                         return fmt.Errorf("While loading container image into Docker: %v", err)
 248                 }
 249
 250                 defer response.Body.Close()
 251                 rbody, err := ioutil.ReadAll(response.Body)
 252                 if err != nil {
 253                         return fmt.Errorf("Reading response to image load: %v", err)
 254                 }
 255                 runner.CrunchLog.Printf("Docker response: %s", rbody)
 256         } else {
 257                 runner.CrunchLog.Print("Docker image is available")
 258         }
 259
 260         runner.ContainerConfig.Image = imageID
 261
 262         runner.Kc.ClearBlockCache()
 263
 264         return nil
 265 }
 266
 267 func (runner *ContainerRunner) ArvMountCmd(arvMountCmd []string, token string) (c *exec.Cmd, err error) {
 268         c = exec.Command("arv-mount", arvMountCmd...)
 269
 270         // Copy our environment, but override ARVADOS_API_TOKEN with
 271         // the container auth token.
 272         c.Env = nil
 273         for _, s := range os.Environ() {
 274                 if !strings.HasPrefix(s, "ARVADOS_API_TOKEN=") {
 275                         c.Env = append(c.Env, s)
 276                 }
 277         }
 278         c.Env = append(c.Env, "ARVADOS_API_TOKEN="+token)
 279
 280         runner.arvMountLog = NewThrottledLogger(runner.NewLogWriter("arv-mount"))
 281         c.Stdout = runner.arvMountLog
 282         c.Stderr = runner.arvMountLog
 283
 284         runner.CrunchLog.Printf("Running %v", c.Args)
 285
 286         err = c.Start()
 287         if err != nil {
 288                 return nil, err
 289         }
 290
 291         statReadme := make(chan bool)
 292         runner.ArvMountExit = make(chan error)
 293
 294         keepStatting := true
 295         go func() {
 296                 for keepStatting {
 297                         time.Sleep(100 * time.Millisecond)
 298                         _, err = os.Stat(fmt.Sprintf("%s/by_id/README", runner.ArvMountPoint))
 299                         if err == nil {
 300                                 keepStatting = false
 301                                 statReadme <- true
 302                         }
 303                 }
 304                 close(statReadme)
 305         }()
 306
 307         go func() {
 308                 mnterr := c.Wait()
 309                 if mnterr != nil {
 310                         runner.CrunchLog.Printf("Arv-mount exit error: %v", mnterr)
 311                 }
 312                 runner.ArvMountExit <- mnterr
 313                 close(runner.ArvMountExit)
 314         }()
 315
 316         select {
 317         case <-statReadme:
 318                 break
 319         case err := <-runner.ArvMountExit:
 320                 runner.ArvMount = nil
 321                 keepStatting = false
 322                 return nil, err
 323         }
 324
 325         return c, nil
 326 }
 327
 328 func (runner *ContainerRunner) SetupArvMountPoint(prefix string) (err error) {
 329         if runner.ArvMountPoint == "" {
 330                 runner.ArvMountPoint, err = runner.MkTempDir(runner.parentTemp, prefix)
 331         }
 332         return
 333 }
 334
 335 func copyfile(src string, dst string) (err error) {
 336         srcfile, err := os.Open(src)
 337         if err != nil {
 338                 return
 339         }
 340
 341         os.MkdirAll(path.Dir(dst), 0777)
 342
 343         dstfile, err := os.Create(dst)
 344         if err != nil {
 345                 return
 346         }
 347         _, err = io.Copy(dstfile, srcfile)
 348         if err != nil {
 349                 return
 350         }
 351
 352         err = srcfile.Close()
 353         err2 := dstfile.Close()
 354
 355         if err != nil {
 356                 return
 357         }
 358
 359         if err2 != nil {
 360                 return err2
 361         }
 362
 363         return nil
 364 }
 365
 366 func (runner *ContainerRunner) SetupMounts() (err error) {
 367         err = runner.SetupArvMountPoint("keep")
 368         if err != nil {
 369                 return fmt.Errorf("While creating keep mount temp dir: %v", err)
 370         }
 371
 372         token, err := runner.ContainerToken()
 373         if err != nil {
 374                 return fmt.Errorf("could not get container token: %s", err)
 375         }
 376
 377         pdhOnly := true
 378         tmpcount := 0
 379         arvMountCmd := []string{
 380                 "--foreground",
 381                 "--allow-other",
 382                 "--read-write",
 383                 fmt.Sprintf("--crunchstat-interval=%v", runner.statInterval.Seconds())}
 384
 385         if runner.Container.RuntimeConstraints.KeepCacheRAM > 0 {
 386                 arvMountCmd = append(arvMountCmd, "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheRAM))
 387         }
 388
 389         collectionPaths := []string{}
 390         runner.Binds = nil
 391         runner.Volumes = make(map[string]struct{})
 392         needCertMount := true
 393         type copyFile struct {
 394                 src  string
 395                 bind string
 396         }
 397         var copyFiles []copyFile
 398
 399         var binds []string
 400         for bind := range runner.Container.Mounts {
 401                 binds = append(binds, bind)
 402         }
 403         sort.Strings(binds)
 404
 405         for _, bind := range binds {
 406                 mnt := runner.Container.Mounts[bind]
 407                 if bind == "stdout" || bind == "stderr" {
 408                         // Is it a "file" mount kind?
 409                         if mnt.Kind != "file" {
 410                                 return fmt.Errorf("Unsupported mount kind '%s' for %s. Only 'file' is supported.", mnt.Kind, bind)
 411                         }
 412
 413                         // Does path start with OutputPath?
 414                         prefix := runner.Container.OutputPath
 415                         if !strings.HasSuffix(prefix, "/") {
 416                                 prefix += "/"
 417                         }
 418                         if !strings.HasPrefix(mnt.Path, prefix) {
 419                                 return fmt.Errorf("%s path does not start with OutputPath: %s, %s", strings.Title(bind), mnt.Path, prefix)
 420                         }
 421                 }
 422
 423                 if bind == "stdin" {
 424                         // Is it a "collection" mount kind?
 425                         if mnt.Kind != "collection" && mnt.Kind != "json" {
 426                                 return fmt.Errorf("Unsupported mount kind '%s' for stdin. Only 'collection' or 'json' are supported.", mnt.Kind)
 427                         }
 428                 }
 429
 430                 if bind == "/etc/arvados/ca-certificates.crt" {
 431                         needCertMount = false
 432                 }
 433
 434                 if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && bind != runner.Container.OutputPath+"/" {
 435                         if mnt.Kind != "collection" {
 436                                 return fmt.Errorf("Only mount points of kind 'collection' are supported underneath the output_path: %v", bind)
 437                         }
 438                 }
 439
 440                 switch {
 441                 case mnt.Kind == "collection" && bind != "stdin":
 442                         var src string
 443                         if mnt.UUID != "" && mnt.PortableDataHash != "" {
 444                                 return fmt.Errorf("Cannot specify both 'uuid' and 'portable_data_hash' for a collection mount")
 445                         }
 446                         if mnt.UUID != "" {
 447                                 if mnt.Writable {
 448                                         return fmt.Errorf("Writing to existing collections currently not permitted.")
 449                                 }
 450                                 pdhOnly = false
 451                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.UUID)
 452                         } else if mnt.PortableDataHash != "" {
 453                                 if mnt.Writable && !strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 454                                         return fmt.Errorf("Can never write to a collection specified by portable data hash")
 455                                 }
 456                                 idx := strings.Index(mnt.PortableDataHash, "/")
 457                                 if idx > 0 {
 458                                         mnt.Path = path.Clean(mnt.PortableDataHash[idx:])
 459                                         mnt.PortableDataHash = mnt.PortableDataHash[0:idx]
 460                                         runner.Container.Mounts[bind] = mnt
 461                                 }
 462                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.PortableDataHash)
 463                                 if mnt.Path != "" && mnt.Path != "." {
 464                                         if strings.HasPrefix(mnt.Path, "./") {
 465                                                 mnt.Path = mnt.Path[2:]
 466                                         } else if strings.HasPrefix(mnt.Path, "/") {
 467                                                 mnt.Path = mnt.Path[1:]
 468                                         }
 469                                         src += "/" + mnt.Path
 470                                 }
 471                         } else {
 472                                 src = fmt.Sprintf("%s/tmp%d", runner.ArvMountPoint, tmpcount)
 473                                 arvMountCmd = append(arvMountCmd, "--mount-tmp")
 474                                 arvMountCmd = append(arvMountCmd, fmt.Sprintf("tmp%d", tmpcount))
 475                                 tmpcount += 1
 476                         }
 477                         if mnt.Writable {
 478                                 if bind == runner.Container.OutputPath {
 479                                         runner.HostOutputDir = src
 480                                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", src, bind))
 481                                 } else if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 482                                         copyFiles = append(copyFiles, copyFile{src, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
 483                                 } else {
 484                                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", src, bind))
 485                                 }
 486                         } else {
 487                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", src, bind))
 488                         }
 489                         collectionPaths = append(collectionPaths, src)
 490
 491                 case mnt.Kind == "tmp":
 492                         var tmpdir string
 493                         tmpdir, err = runner.MkTempDir(runner.parentTemp, "tmp")
 494                         if err != nil {
 495                                 return fmt.Errorf("While creating mount temp dir: %v", err)
 496                         }
 497                         st, staterr := os.Stat(tmpdir)
 498                         if staterr != nil {
 499                                 return fmt.Errorf("While Stat on temp dir: %v", staterr)
 500                         }
 501                         err = os.Chmod(tmpdir, st.Mode()|os.ModeSetgid|0777)
 502                         if staterr != nil {
 503                                 return fmt.Errorf("While Chmod temp dir: %v", err)
 504                         }
 505                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", tmpdir, bind))
 506                         if bind == runner.Container.OutputPath {
 507                                 runner.HostOutputDir = tmpdir
 508                         }
 509
 510                 case mnt.Kind == "json":
 511                         jsondata, err := json.Marshal(mnt.Content)
 512                         if err != nil {
 513                                 return fmt.Errorf("encoding json data: %v", err)
 514                         }
 515                         // Create a tempdir with a single file
 516                         // (instead of just a tempfile): this way we
 517                         // can ensure the file is world-readable
 518                         // inside the container, without having to
 519                         // make it world-readable on the docker host.
 520                         tmpdir, err := runner.MkTempDir(runner.parentTemp, "json")
 521                         if err != nil {
 522                                 return fmt.Errorf("creating temp dir: %v", err)
 523                         }
 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(runner.parentTemp, "git_tree")
 533                         if err != nil {
 534                                 return fmt.Errorf("creating temp dir: %v", err)
 535                         }
 536                         err = gitMount(mnt).extractTree(runner.ArvClient, tmpdir, token)
 537                         if err != nil {
 538                                 return err
 539                         }
 540                         runner.Binds = append(runner.Binds, tmpdir+":"+bind+":ro")
 541                 }
 542         }
 543
 544         if runner.HostOutputDir == "" {
 545                 return fmt.Errorf("Output path does not correspond to a writable mount point")
 546         }
 547
 548         if wantAPI := runner.Container.RuntimeConstraints.API; needCertMount && wantAPI != nil && *wantAPI {
 549                 for _, certfile := range arvadosclient.CertFiles {
 550                         _, err := os.Stat(certfile)
 551                         if err == nil {
 552                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:/etc/arvados/ca-certificates.crt:ro", certfile))
 553                                 break
 554                         }
 555                 }
 556         }
 557
 558         if pdhOnly {
 559                 arvMountCmd = append(arvMountCmd, "--mount-by-pdh", "by_id")
 560         } else {
 561                 arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_id")
 562         }
 563         arvMountCmd = append(arvMountCmd, runner.ArvMountPoint)
 564
 565         runner.ArvMount, err = runner.RunArvMount(arvMountCmd, token)
 566         if err != nil {
 567                 return fmt.Errorf("While trying to start arv-mount: %v", err)
 568         }
 569
 570         for _, p := range collectionPaths {
 571                 _, err = os.Stat(p)
 572                 if err != nil {
 573                         return fmt.Errorf("While checking that input files exist: %v", err)
 574                 }
 575         }
 576
 577         for _, cp := range copyFiles {
 578                 st, err := os.Stat(cp.src)
 579                 if err != nil {
 580                         return fmt.Errorf("While staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 581                 }
 582                 if st.IsDir() {
 583                         err = filepath.Walk(cp.src, func(walkpath string, walkinfo os.FileInfo, walkerr error) error {
 584                                 if walkerr != nil {
 585                                         return walkerr
 586                                 }
 587                                 target := path.Join(cp.bind, walkpath[len(cp.src):])
 588                                 if walkinfo.Mode().IsRegular() {
 589                                         copyerr := copyfile(walkpath, target)
 590                                         if copyerr != nil {
 591                                                 return copyerr
 592                                         }
 593                                         return os.Chmod(target, walkinfo.Mode()|0777)
 594                                 } else if walkinfo.Mode().IsDir() {
 595                                         mkerr := os.MkdirAll(target, 0777)
 596                                         if mkerr != nil {
 597                                                 return mkerr
 598                                         }
 599                                         return os.Chmod(target, walkinfo.Mode()|os.ModeSetgid|0777)
 600                                 } else {
 601                                         return fmt.Errorf("Source %q is not a regular file or directory", cp.src)
 602                                 }
 603                         })
 604                 } else if st.Mode().IsRegular() {
 605                         err = copyfile(cp.src, cp.bind)
 606                         if err == nil {
 607                                 err = os.Chmod(cp.bind, st.Mode()|0777)
 608                         }
 609                 }
 610                 if err != nil {
 611                         return fmt.Errorf("While staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 612                 }
 613         }
 614
 615         return nil
 616 }
 617
 618 func (runner *ContainerRunner) ProcessDockerAttach(containerReader io.Reader) {
 619         // Handle docker log protocol
 620         // https://docs.docker.com/engine/reference/api/docker_remote_api_v1.15/#attach-to-a-container
 621         defer close(runner.loggingDone)
 622
 623         header := make([]byte, 8)
 624         var err error
 625         for err == nil {
 626                 _, err = io.ReadAtLeast(containerReader, header, 8)
 627                 if err != nil {
 628                         if err == io.EOF {
 629                                 err = nil
 630                         }
 631                         break
 632                 }
 633                 readsize := int64(header[7]) | (int64(header[6]) << 8) | (int64(header[5]) << 16) | (int64(header[4]) << 24)
 634                 if header[0] == 1 {
 635                         // stdout
 636                         _, err = io.CopyN(runner.Stdout, containerReader, readsize)
 637                 } else {
 638                         // stderr
 639                         _, err = io.CopyN(runner.Stderr, containerReader, readsize)
 640                 }
 641         }
 642
 643         if err != nil {
 644                 runner.CrunchLog.Printf("error reading docker logs: %v", err)
 645         }
 646
 647         err = runner.Stdout.Close()
 648         if err != nil {
 649                 runner.CrunchLog.Printf("error closing stdout logs: %v", err)
 650         }
 651
 652         err = runner.Stderr.Close()
 653         if err != nil {
 654                 runner.CrunchLog.Printf("error closing stderr logs: %v", err)
 655         }
 656
 657         if runner.statReporter != nil {
 658                 runner.statReporter.Stop()
 659                 err = runner.statLogger.Close()
 660                 if err != nil {
 661                         runner.CrunchLog.Printf("error closing crunchstat logs: %v", err)
 662                 }
 663         }
 664 }
 665
 666 func (runner *ContainerRunner) stopHoststat() error {
 667         if runner.hoststatReporter == nil {
 668                 return nil
 669         }
 670         runner.hoststatReporter.Stop()
 671         err := runner.hoststatLogger.Close()
 672         if err != nil {
 673                 return fmt.Errorf("error closing hoststat logs: %v", err)
 674         }
 675         return nil
 676 }
 677
 678 func (runner *ContainerRunner) startHoststat() {
 679         runner.hoststatLogger = NewThrottledLogger(runner.NewLogWriter("hoststat"))
 680         runner.hoststatReporter = &crunchstat.Reporter{
 681                 Logger:     log.New(runner.hoststatLogger, "", 0),
 682                 CgroupRoot: runner.cgroupRoot,
 683                 PollPeriod: runner.statInterval,
 684         }
 685         runner.hoststatReporter.Start()
 686 }
 687
 688 func (runner *ContainerRunner) startCrunchstat() {
 689         runner.statLogger = NewThrottledLogger(runner.NewLogWriter("crunchstat"))
 690         runner.statReporter = &crunchstat.Reporter{
 691                 CID:          runner.ContainerID,
 692                 Logger:       log.New(runner.statLogger, "", 0),
 693                 CgroupParent: runner.expectCgroupParent,
 694                 CgroupRoot:   runner.cgroupRoot,
 695                 PollPeriod:   runner.statInterval,
 696         }
 697         runner.statReporter.Start()
 698 }
 699
 700 type infoCommand struct {
 701         label string
 702         cmd   []string
 703 }
 704
 705 // LogHostInfo logs info about the current host, for debugging and
 706 // accounting purposes. Although it's logged as "node-info", this is
 707 // about the environment where crunch-run is actually running, which
 708 // might differ from what's described in the node record (see
 709 // LogNodeRecord).
 710 func (runner *ContainerRunner) LogHostInfo() (err error) {
 711         w := runner.NewLogWriter("node-info")
 712
 713         commands := []infoCommand{
 714                 {
 715                         label: "Host Information",
 716                         cmd:   []string{"uname", "-a"},
 717                 },
 718                 {
 719                         label: "CPU Information",
 720                         cmd:   []string{"cat", "/proc/cpuinfo"},
 721                 },
 722                 {
 723                         label: "Memory Information",
 724                         cmd:   []string{"cat", "/proc/meminfo"},
 725                 },
 726                 {
 727                         label: "Disk Space",
 728                         cmd:   []string{"df", "-m", "/", os.TempDir()},
 729                 },
 730                 {
 731                         label: "Disk INodes",
 732                         cmd:   []string{"df", "-i", "/", os.TempDir()},
 733                 },
 734         }
 735
 736         // Run commands with informational output to be logged.
 737         for _, command := range commands {
 738                 fmt.Fprintln(w, command.label)
 739                 cmd := exec.Command(command.cmd[0], command.cmd[1:]...)
 740                 cmd.Stdout = w
 741                 cmd.Stderr = w
 742                 if err := cmd.Run(); err != nil {
 743                         err = fmt.Errorf("While running command %q: %v", command.cmd, err)
 744                         fmt.Fprintln(w, err)
 745                         return err
 746                 }
 747                 fmt.Fprintln(w, "")
 748         }
 749
 750         err = w.Close()
 751         if err != nil {
 752                 return fmt.Errorf("While closing node-info logs: %v", err)
 753         }
 754         return nil
 755 }
 756
 757 // LogContainerRecord gets and saves the raw JSON container record from the API server
 758 func (runner *ContainerRunner) LogContainerRecord() error {
 759         logged, err := runner.logAPIResponse("container", "containers", map[string]interface{}{"filters": [][]string{{"uuid", "=", runner.Container.UUID}}}, nil)
 760         if !logged && err == nil {
 761                 err = fmt.Errorf("error: no container record found for %s", runner.Container.UUID)
 762         }
 763         return err
 764 }
 765
 766 // LogNodeRecord logs arvados#node record corresponding to the current host.
 767 func (runner *ContainerRunner) LogNodeRecord() error {
 768         hostname := os.Getenv("SLURMD_NODENAME")
 769         if hostname == "" {
 770                 hostname, _ = os.Hostname()
 771         }
 772         _, err := runner.logAPIResponse("node", "nodes", map[string]interface{}{"filters": [][]string{{"hostname", "=", hostname}}}, func(resp interface{}) {
 773                 // The "info" field has admin-only info when obtained
 774                 // with a privileged token, and should not be logged.
 775                 node, ok := resp.(map[string]interface{})
 776                 if ok {
 777                         delete(node, "info")
 778                 }
 779         })
 780         return err
 781 }
 782
 783 func (runner *ContainerRunner) logAPIResponse(label, path string, params map[string]interface{}, munge func(interface{})) (logged bool, err error) {
 784         w := &ArvLogWriter{
 785                 ArvClient:     runner.ArvClient,
 786                 UUID:          runner.Container.UUID,
 787                 loggingStream: label,
 788                 writeCloser:   runner.LogCollection.Open(label + ".json"),
 789         }
 790
 791         reader, err := runner.ArvClient.CallRaw("GET", path, "", "", arvadosclient.Dict(params))
 792         if err != nil {
 793                 return false, fmt.Errorf("error getting %s record: %v", label, err)
 794         }
 795         defer reader.Close()
 796
 797         dec := json.NewDecoder(reader)
 798         dec.UseNumber()
 799         var resp map[string]interface{}
 800         if err = dec.Decode(&resp); err != nil {
 801                 return false, fmt.Errorf("error decoding %s list response: %v", label, err)
 802         }
 803         items, ok := resp["items"].([]interface{})
 804         if !ok {
 805                 return false, fmt.Errorf("error decoding %s list response: no \"items\" key in API list response", label)
 806         } else if len(items) < 1 {
 807                 return false, nil
 808         }
 809         if munge != nil {
 810                 munge(items[0])
 811         }
 812         // Re-encode it using indentation to improve readability
 813         enc := json.NewEncoder(w)
 814         enc.SetIndent("", "    ")
 815         if err = enc.Encode(items[0]); err != nil {
 816                 return false, fmt.Errorf("error logging %s record: %v", label, err)
 817         }
 818         err = w.Close()
 819         if err != nil {
 820                 return false, fmt.Errorf("error closing %s.json in log collection: %v", label, err)
 821         }
 822         return true, nil
 823 }
 824
 825 // AttachStreams connects the docker container stdin, stdout and stderr logs
 826 // to the Arvados logger which logs to Keep and the API server logs table.
 827 func (runner *ContainerRunner) AttachStreams() (err error) {
 828
 829         runner.CrunchLog.Print("Attaching container streams")
 830
 831         // If stdin mount is provided, attach it to the docker container
 832         var stdinRdr arvados.File
 833         var stdinJson []byte
 834         if stdinMnt, ok := runner.Container.Mounts["stdin"]; ok {
 835                 if stdinMnt.Kind == "collection" {
 836                         var stdinColl arvados.Collection
 837                         collId := stdinMnt.UUID
 838                         if collId == "" {
 839                                 collId = stdinMnt.PortableDataHash
 840                         }
 841                         err = runner.ArvClient.Get("collections", collId, nil, &stdinColl)
 842                         if err != nil {
 843                                 return fmt.Errorf("While getting stding collection: %v", err)
 844                         }
 845
 846                         stdinRdr, err = runner.Kc.ManifestFileReader(manifest.Manifest{Text: stdinColl.ManifestText}, stdinMnt.Path)
 847                         if os.IsNotExist(err) {
 848                                 return fmt.Errorf("stdin collection path not found: %v", stdinMnt.Path)
 849                         } else if err != nil {
 850                                 return fmt.Errorf("While getting stdin collection path %v: %v", stdinMnt.Path, err)
 851                         }
 852                 } else if stdinMnt.Kind == "json" {
 853                         stdinJson, err = json.Marshal(stdinMnt.Content)
 854                         if err != nil {
 855                                 return fmt.Errorf("While encoding stdin json data: %v", err)
 856                         }
 857                 }
 858         }
 859
 860         stdinUsed := stdinRdr != nil || len(stdinJson) != 0
 861         response, err := runner.Docker.ContainerAttach(context.TODO(), runner.ContainerID,
 862                 dockertypes.ContainerAttachOptions{Stream: true, Stdin: stdinUsed, Stdout: true, Stderr: true})
 863         if err != nil {
 864                 return fmt.Errorf("While attaching container stdout/stderr streams: %v", err)
 865         }
 866
 867         runner.loggingDone = make(chan bool)
 868
 869         if stdoutMnt, ok := runner.Container.Mounts["stdout"]; ok {
 870                 stdoutFile, err := runner.getStdoutFile(stdoutMnt.Path)
 871                 if err != nil {
 872                         return err
 873                 }
 874                 runner.Stdout = stdoutFile
 875         } else {
 876                 runner.Stdout = NewThrottledLogger(runner.NewLogWriter("stdout"))
 877         }
 878
 879         if stderrMnt, ok := runner.Container.Mounts["stderr"]; ok {
 880                 stderrFile, err := runner.getStdoutFile(stderrMnt.Path)
 881                 if err != nil {
 882                         return err
 883                 }
 884                 runner.Stderr = stderrFile
 885         } else {
 886                 runner.Stderr = NewThrottledLogger(runner.NewLogWriter("stderr"))
 887         }
 888
 889         if stdinRdr != nil {
 890                 go func() {
 891                         _, err := io.Copy(response.Conn, stdinRdr)
 892                         if err != nil {
 893                                 runner.CrunchLog.Print("While writing stdin collection to docker container %q", err)
 894                                 runner.stop()
 895                         }
 896                         stdinRdr.Close()
 897                         response.CloseWrite()
 898                 }()
 899         } else if len(stdinJson) != 0 {
 900                 go func() {
 901                         _, err := io.Copy(response.Conn, bytes.NewReader(stdinJson))
 902                         if err != nil {
 903                                 runner.CrunchLog.Print("While writing stdin json to docker container %q", err)
 904                                 runner.stop()
 905                         }
 906                         response.CloseWrite()
 907                 }()
 908         }
 909
 910         go runner.ProcessDockerAttach(response.Reader)
 911
 912         return nil
 913 }
 914
 915 func (runner *ContainerRunner) getStdoutFile(mntPath string) (*os.File, error) {
 916         stdoutPath := mntPath[len(runner.Container.OutputPath):]
 917         index := strings.LastIndex(stdoutPath, "/")
 918         if index > 0 {
 919                 subdirs := stdoutPath[:index]
 920                 if subdirs != "" {
 921                         st, err := os.Stat(runner.HostOutputDir)
 922                         if err != nil {
 923                                 return nil, fmt.Errorf("While Stat on temp dir: %v", err)
 924                         }
 925                         stdoutPath := filepath.Join(runner.HostOutputDir, subdirs)
 926                         err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777)
 927                         if err != nil {
 928                                 return nil, fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err)
 929                         }
 930                 }
 931         }
 932         stdoutFile, err := os.Create(filepath.Join(runner.HostOutputDir, stdoutPath))
 933         if err != nil {
 934                 return nil, fmt.Errorf("While creating file %q: %v", stdoutPath, err)
 935         }
 936
 937         return stdoutFile, nil
 938 }
 939
 940 // CreateContainer creates the docker container.
 941 func (runner *ContainerRunner) CreateContainer() error {
 942         runner.CrunchLog.Print("Creating Docker container")
 943
 944         runner.ContainerConfig.Cmd = runner.Container.Command
 945         if runner.Container.Cwd != "." {
 946                 runner.ContainerConfig.WorkingDir = runner.Container.Cwd
 947         }
 948
 949         for k, v := range runner.Container.Environment {
 950                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env, k+"="+v)
 951         }
 952
 953         runner.ContainerConfig.Volumes = runner.Volumes
 954
 955         runner.HostConfig = dockercontainer.HostConfig{
 956                 Binds: runner.Binds,
 957                 LogConfig: dockercontainer.LogConfig{
 958                         Type: "none",
 959                 },
 960                 Resources: dockercontainer.Resources{
 961                         CgroupParent: runner.setCgroupParent,
 962                 },
 963         }
 964
 965         if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
 966                 tok, err := runner.ContainerToken()
 967                 if err != nil {
 968                         return err
 969                 }
 970                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env,
 971                         "ARVADOS_API_TOKEN="+tok,
 972                         "ARVADOS_API_HOST="+os.Getenv("ARVADOS_API_HOST"),
 973                         "ARVADOS_API_HOST_INSECURE="+os.Getenv("ARVADOS_API_HOST_INSECURE"),
 974                 )
 975                 runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
 976         } else {
 977                 if runner.enableNetwork == "always" {
 978                         runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
 979                 } else {
 980                         runner.HostConfig.NetworkMode = dockercontainer.NetworkMode("none")
 981                 }
 982         }
 983
 984         _, stdinUsed := runner.Container.Mounts["stdin"]
 985         runner.ContainerConfig.OpenStdin = stdinUsed
 986         runner.ContainerConfig.StdinOnce = stdinUsed
 987         runner.ContainerConfig.AttachStdin = stdinUsed
 988         runner.ContainerConfig.AttachStdout = true
 989         runner.ContainerConfig.AttachStderr = true
 990
 991         createdBody, err := runner.Docker.ContainerCreate(context.TODO(), &runner.ContainerConfig, &runner.HostConfig, nil, runner.Container.UUID)
 992         if err != nil {
 993                 return fmt.Errorf("While creating container: %v", err)
 994         }
 995
 996         runner.ContainerID = createdBody.ID
 997
 998         return runner.AttachStreams()
 999 }
1000
1001 // StartContainer starts the docker container created by CreateContainer.
1002 func (runner *ContainerRunner) StartContainer() error {
1003         runner.CrunchLog.Printf("Starting Docker container id '%s'", runner.ContainerID)
1004         runner.cStateLock.Lock()
1005         defer runner.cStateLock.Unlock()
1006         if runner.cCancelled {
1007                 return ErrCancelled
1008         }
1009         err := runner.Docker.ContainerStart(context.TODO(), runner.ContainerID,
1010                 dockertypes.ContainerStartOptions{})
1011         if err != nil {
1012                 var advice string
1013                 if m, e := regexp.MatchString("(?ms).*(exec|System error).*(no such file or directory|file not found).*", err.Error()); m && e == nil {
1014                         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])
1015                 }
1016                 return fmt.Errorf("could not start container: %v%s", err, advice)
1017         }
1018         return nil
1019 }
1020
1021 // WaitFinish waits for the container to terminate, capture the exit code, and
1022 // close the stdout/stderr logging.
1023 func (runner *ContainerRunner) WaitFinish() error {
1024         runner.CrunchLog.Print("Waiting for container to finish")
1025
1026         waitOk, waitErr := runner.Docker.ContainerWait(context.TODO(), runner.ContainerID, dockercontainer.WaitConditionNotRunning)
1027         arvMountExit := runner.ArvMountExit
1028         for {
1029                 select {
1030                 case waitBody := <-waitOk:
1031                         runner.CrunchLog.Printf("Container exited with code: %v", waitBody.StatusCode)
1032                         code := int(waitBody.StatusCode)
1033                         runner.ExitCode = &code
1034
1035                         // wait for stdout/stderr to complete
1036                         <-runner.loggingDone
1037                         return nil
1038
1039                 case err := <-waitErr:
1040                         return fmt.Errorf("container wait: %v", err)
1041
1042                 case <-arvMountExit:
1043                         runner.CrunchLog.Printf("arv-mount exited while container is still running.  Stopping container.")
1044                         runner.stop()
1045                         // arvMountExit will always be ready now that
1046                         // it's closed, but that doesn't interest us.
1047                         arvMountExit = nil
1048                 }
1049         }
1050 }
1051
1052 var ErrNotInOutputDir = fmt.Errorf("Must point to path within the output directory")
1053
1054 func (runner *ContainerRunner) derefOutputSymlink(path string, startinfo os.FileInfo) (tgt string, readlinktgt string, info os.FileInfo, err error) {
1055         // Follow symlinks if necessary
1056         info = startinfo
1057         tgt = path
1058         readlinktgt = ""
1059         nextlink := path
1060         for followed := 0; info.Mode()&os.ModeSymlink != 0; followed++ {
1061                 if followed >= limitFollowSymlinks {
1062                         // Got stuck in a loop or just a pathological number of links, give up.
1063                         err = fmt.Errorf("Followed more than %v symlinks from path %q", limitFollowSymlinks, path)
1064                         return
1065                 }
1066
1067                 readlinktgt, err = os.Readlink(nextlink)
1068                 if err != nil {
1069                         return
1070                 }
1071
1072                 tgt = readlinktgt
1073                 if !strings.HasPrefix(tgt, "/") {
1074                         // Relative symlink, resolve it to host path
1075                         tgt = filepath.Join(filepath.Dir(path), tgt)
1076                 }
1077                 if strings.HasPrefix(tgt, runner.Container.OutputPath+"/") && !strings.HasPrefix(tgt, runner.HostOutputDir+"/") {
1078                         // Absolute symlink to container output path, adjust it to host output path.
1079                         tgt = filepath.Join(runner.HostOutputDir, tgt[len(runner.Container.OutputPath):])
1080                 }
1081                 if !strings.HasPrefix(tgt, runner.HostOutputDir+"/") {
1082                         // After dereferencing, symlink target must either be
1083                         // within output directory, or must point to a
1084                         // collection mount.
1085                         err = ErrNotInOutputDir
1086                         return
1087                 }
1088
1089                 info, err = os.Lstat(tgt)
1090                 if err != nil {
1091                         // tgt
1092                         err = fmt.Errorf("Symlink in output %q points to invalid location %q: %v",
1093                                 path[len(runner.HostOutputDir):], readlinktgt, err)
1094                         return
1095                 }
1096
1097                 nextlink = tgt
1098         }
1099
1100         return
1101 }
1102
1103 var limitFollowSymlinks = 10
1104
1105 // UploadFile uploads files within the output directory, with special handling
1106 // for symlinks. If the symlink leads to a keep mount, copy the manifest text
1107 // from the keep mount into the output manifestText.  Ensure that whether
1108 // symlinks are relative or absolute, every symlink target (even targets that
1109 // are symlinks themselves) must point to a path in either the output directory
1110 // or a collection mount.
1111 //
1112 // Assumes initial value of "path" is absolute, and located within runner.HostOutputDir.
1113 func (runner *ContainerRunner) UploadOutputFile(
1114         path string,
1115         info os.FileInfo,
1116         infoerr error,
1117         binds []string,
1118         walkUpload *WalkUpload,
1119         relocateFrom string,
1120         relocateTo string,
1121         followed int) (manifestText string, err error) {
1122
1123         if infoerr != nil {
1124                 return "", infoerr
1125         }
1126
1127         if info.Mode().IsDir() {
1128                 // if empty, need to create a .keep file
1129                 dir, direrr := os.Open(path)
1130                 if direrr != nil {
1131                         return "", direrr
1132                 }
1133                 defer dir.Close()
1134                 names, eof := dir.Readdirnames(1)
1135                 if len(names) == 0 && eof == io.EOF && path != runner.HostOutputDir {
1136                         containerPath := runner.OutputPath + path[len(runner.HostOutputDir):]
1137                         for _, bind := range binds {
1138                                 mnt := runner.Container.Mounts[bind]
1139                                 // Check if there is a bind for this
1140                                 // directory, in which case assume we don't need .keep
1141                                 if (containerPath == bind || strings.HasPrefix(containerPath, bind+"/")) && mnt.PortableDataHash != "d41d8cd98f00b204e9800998ecf8427e+0" {
1142                                         return
1143                                 }
1144                         }
1145                         outputSuffix := path[len(runner.HostOutputDir)+1:]
1146                         return fmt.Sprintf("./%v d41d8cd98f00b204e9800998ecf8427e+0 0:0:.keep\n", outputSuffix), nil
1147                 }
1148                 return
1149         }
1150
1151         if followed >= limitFollowSymlinks {
1152                 // Got stuck in a loop or just a pathological number of
1153                 // directory links, give up.
1154                 err = fmt.Errorf("Followed more than %v symlinks from path %q", limitFollowSymlinks, path)
1155                 return
1156         }
1157
1158         // "path" is the actual path we are visiting
1159         // "tgt" is the target of "path" (a non-symlink) after following symlinks
1160         // "relocated" is the path in the output manifest where the file should be placed,
1161         // but has HostOutputDir as a prefix.
1162
1163         // The destination path in the output manifest may need to be
1164         // logically relocated to some other path in order to appear
1165         // in the correct location as a result of following a symlink.
1166         // Remove the relocateFrom prefix and replace it with
1167         // relocateTo.
1168         relocated := relocateTo + path[len(relocateFrom):]
1169
1170         tgt, readlinktgt, info, derefErr := runner.derefOutputSymlink(path, info)
1171         if derefErr != nil && derefErr != ErrNotInOutputDir {
1172                 return "", derefErr
1173         }
1174
1175         // go through mounts and try reverse map to collection reference
1176         for _, bind := range binds {
1177                 mnt := runner.Container.Mounts[bind]
1178                 if (tgt == bind || strings.HasPrefix(tgt, bind+"/")) && !mnt.Writable {
1179                         // get path relative to bind
1180                         targetSuffix := tgt[len(bind):]
1181
1182                         // Copy mount and adjust the path to add path relative to the bind
1183                         adjustedMount := mnt
1184                         adjustedMount.Path = filepath.Join(adjustedMount.Path, targetSuffix)
1185
1186                         // Terminates in this keep mount, so add the
1187                         // manifest text at appropriate location.
1188                         outputSuffix := relocated[len(runner.HostOutputDir):]
1189                         manifestText, err = runner.getCollectionManifestForPath(adjustedMount, outputSuffix)
1190                         return
1191                 }
1192         }
1193
1194         // If target is not a collection mount, it must be located within the
1195         // output directory, otherwise it is an error.
1196         if derefErr == ErrNotInOutputDir {
1197                 err = fmt.Errorf("Symlink in output %q points to invalid location %q, must point to path within the output directory.",
1198                         path[len(runner.HostOutputDir):], readlinktgt)
1199                 return
1200         }
1201
1202         if info.Mode().IsRegular() {
1203                 return "", walkUpload.UploadFile(relocated, tgt)
1204         }
1205
1206         if info.Mode().IsDir() {
1207                 // Symlink leads to directory.  Walk() doesn't follow
1208                 // directory symlinks, so we walk the target directory
1209                 // instead.  Within the walk, file paths are relocated
1210                 // so they appear under the original symlink path.
1211                 err = filepath.Walk(tgt, func(walkpath string, walkinfo os.FileInfo, walkerr error) error {
1212                         var m string
1213                         m, walkerr = runner.UploadOutputFile(walkpath, walkinfo, walkerr,
1214                                 binds, walkUpload, tgt, relocated, followed+1)
1215                         if walkerr == nil {
1216                                 manifestText = manifestText + m
1217                         }
1218                         return walkerr
1219                 })
1220                 return
1221         }
1222
1223         return
1224 }
1225
1226 // HandleOutput sets the output, unmounts the FUSE mount, and deletes temporary directories
1227 func (runner *ContainerRunner) CaptureOutput() error {
1228         if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
1229                 // Output may have been set directly by the container, so
1230                 // refresh the container record to check.
1231                 err := runner.ArvClient.Get("containers", runner.Container.UUID,
1232                         nil, &runner.Container)
1233                 if err != nil {
1234                         return err
1235                 }
1236                 if runner.Container.Output != "" {
1237                         // Container output is already set.
1238                         runner.OutputPDH = &runner.Container.Output
1239                         return nil
1240                 }
1241         }
1242
1243         if runner.HostOutputDir == "" {
1244                 return nil
1245         }
1246
1247         _, err := os.Stat(runner.HostOutputDir)
1248         if err != nil {
1249                 return fmt.Errorf("While checking host output path: %v", err)
1250         }
1251
1252         // Pre-populate output from the configured mount points
1253         var binds []string
1254         for bind, mnt := range runner.Container.Mounts {
1255                 if mnt.Kind == "collection" {
1256                         binds = append(binds, bind)
1257                 }
1258         }
1259         sort.Strings(binds)
1260
1261         var manifestText string
1262
1263         collectionMetafile := fmt.Sprintf("%s/.arvados#collection", runner.HostOutputDir)
1264         _, err = os.Stat(collectionMetafile)
1265         if err != nil {
1266                 // Regular directory
1267
1268                 cw := CollectionWriter{0, runner.Kc, nil, nil, sync.Mutex{}}
1269                 walkUpload := cw.BeginUpload(runner.HostOutputDir, runner.CrunchLog.Logger)
1270
1271                 var m string
1272                 err = filepath.Walk(runner.HostOutputDir, func(path string, info os.FileInfo, err error) error {
1273                         m, err = runner.UploadOutputFile(path, info, err, binds, walkUpload, "", "", 0)
1274                         if err == nil {
1275                                 manifestText = manifestText + m
1276                         }
1277                         return err
1278                 })
1279
1280                 cw.EndUpload(walkUpload)
1281
1282                 if err != nil {
1283                         return fmt.Errorf("While uploading output files: %v", err)
1284                 }
1285
1286                 m, err = cw.ManifestText()
1287                 manifestText = manifestText + m
1288                 if err != nil {
1289                         return fmt.Errorf("While uploading output files: %v", err)
1290                 }
1291         } else {
1292                 // FUSE mount directory
1293                 file, openerr := os.Open(collectionMetafile)
1294                 if openerr != nil {
1295                         return fmt.Errorf("While opening FUSE metafile: %v", err)
1296                 }
1297                 defer file.Close()
1298
1299                 var rec arvados.Collection
1300                 err = json.NewDecoder(file).Decode(&rec)
1301                 if err != nil {
1302                         return fmt.Errorf("While reading FUSE metafile: %v", err)
1303                 }
1304                 manifestText = rec.ManifestText
1305         }
1306
1307         for _, bind := range binds {
1308                 mnt := runner.Container.Mounts[bind]
1309
1310                 bindSuffix := strings.TrimPrefix(bind, runner.Container.OutputPath)
1311
1312                 if bindSuffix == bind || len(bindSuffix) <= 0 {
1313                         // either does not start with OutputPath or is OutputPath itself
1314                         continue
1315                 }
1316
1317                 if mnt.ExcludeFromOutput == true || mnt.Writable {
1318                         continue
1319                 }
1320
1321                 // append to manifest_text
1322                 m, err := runner.getCollectionManifestForPath(mnt, bindSuffix)
1323                 if err != nil {
1324                         return err
1325                 }
1326
1327                 manifestText = manifestText + m
1328         }
1329
1330         // Save output
1331         var response arvados.Collection
1332         manifest := manifest.Manifest{Text: manifestText}
1333         manifestText = manifest.Extract(".", ".").Text
1334         err = runner.ArvClient.Create("collections",
1335                 arvadosclient.Dict{
1336                         "ensure_unique_name": true,
1337                         "collection": arvadosclient.Dict{
1338                                 "is_trashed":    true,
1339                                 "name":          "output for " + runner.Container.UUID,
1340                                 "manifest_text": manifestText}},
1341                 &response)
1342         if err != nil {
1343                 return fmt.Errorf("While creating output collection: %v", err)
1344         }
1345         runner.OutputPDH = &response.PortableDataHash
1346         return nil
1347 }
1348
1349 var outputCollections = make(map[string]arvados.Collection)
1350
1351 // Fetch the collection for the mnt.PortableDataHash
1352 // Return the manifest_text fragment corresponding to the specified mnt.Path
1353 //  after making any required updates.
1354 //  Ex:
1355 //    If mnt.Path is not specified,
1356 //      return the entire manifest_text after replacing any "." with bindSuffix
1357 //    If mnt.Path corresponds to one stream,
1358 //      return the manifest_text for that stream after replacing that stream name with bindSuffix
1359 //    Otherwise, check if a filename in any one stream is being sought. Return the manifest_text
1360 //      for that stream after replacing stream name with bindSuffix minus the last word
1361 //      and the file name with last word of the bindSuffix
1362 //  Allowed path examples:
1363 //    "path":"/"
1364 //    "path":"/subdir1"
1365 //    "path":"/subdir1/subdir2"
1366 //    "path":"/subdir/filename" etc
1367 func (runner *ContainerRunner) getCollectionManifestForPath(mnt arvados.Mount, bindSuffix string) (string, error) {
1368         collection := outputCollections[mnt.PortableDataHash]
1369         if collection.PortableDataHash == "" {
1370                 err := runner.ArvClient.Get("collections", mnt.PortableDataHash, nil, &collection)
1371                 if err != nil {
1372                         return "", fmt.Errorf("While getting collection for %v: %v", mnt.PortableDataHash, err)
1373                 }
1374                 outputCollections[mnt.PortableDataHash] = collection
1375         }
1376
1377         if collection.ManifestText == "" {
1378                 runner.CrunchLog.Printf("No manifest text for collection %v", collection.PortableDataHash)
1379                 return "", nil
1380         }
1381
1382         mft := manifest.Manifest{Text: collection.ManifestText}
1383         extracted := mft.Extract(mnt.Path, bindSuffix)
1384         if extracted.Err != nil {
1385                 return "", fmt.Errorf("Error parsing manifest for %v: %v", mnt.PortableDataHash, extracted.Err.Error())
1386         }
1387         return extracted.Text, nil
1388 }
1389
1390 func (runner *ContainerRunner) CleanupDirs() {
1391         if runner.ArvMount != nil {
1392                 var delay int64 = 8
1393                 umount := exec.Command("arv-mount", fmt.Sprintf("--unmount-timeout=%d", delay), "--unmount", runner.ArvMountPoint)
1394                 umount.Stdout = runner.CrunchLog
1395                 umount.Stderr = runner.CrunchLog
1396                 runner.CrunchLog.Printf("Running %v", umount.Args)
1397                 umnterr := umount.Start()
1398
1399                 if umnterr != nil {
1400                         runner.CrunchLog.Printf("Error unmounting: %v", umnterr)
1401                 } else {
1402                         // If arv-mount --unmount gets stuck for any reason, we
1403                         // don't want to wait for it forever.  Do Wait() in a goroutine
1404                         // so it doesn't block crunch-run.
1405                         umountExit := make(chan error)
1406                         go func() {
1407                                 mnterr := umount.Wait()
1408                                 if mnterr != nil {
1409                                         runner.CrunchLog.Printf("Error unmounting: %v", mnterr)
1410                                 }
1411                                 umountExit <- mnterr
1412                         }()
1413
1414                         for again := true; again; {
1415                                 again = false
1416                                 select {
1417                                 case <-umountExit:
1418                                         umount = nil
1419                                         again = true
1420                                 case <-runner.ArvMountExit:
1421                                         break
1422                                 case <-time.After(time.Duration((delay + 1) * int64(time.Second))):
1423                                         runner.CrunchLog.Printf("Timed out waiting for unmount")
1424                                         if umount != nil {
1425                                                 umount.Process.Kill()
1426                                         }
1427                                         runner.ArvMount.Process.Kill()
1428                                 }
1429                         }
1430                 }
1431         }
1432
1433         if runner.ArvMountPoint != "" {
1434                 if rmerr := os.Remove(runner.ArvMountPoint); rmerr != nil {
1435                         runner.CrunchLog.Printf("While cleaning up arv-mount directory %s: %v", runner.ArvMountPoint, rmerr)
1436                 }
1437         }
1438
1439         if rmerr := os.RemoveAll(runner.parentTemp); rmerr != nil {
1440                 runner.CrunchLog.Printf("While cleaning up temporary directory %s: %v", runner.parentTemp, rmerr)
1441         }
1442 }
1443
1444 // CommitLogs posts the collection containing the final container logs.
1445 func (runner *ContainerRunner) CommitLogs() error {
1446         runner.CrunchLog.Print(runner.finalState)
1447
1448         if runner.arvMountLog != nil {
1449                 runner.arvMountLog.Close()
1450         }
1451         runner.CrunchLog.Close()
1452
1453         // Closing CrunchLog above allows them to be committed to Keep at this
1454         // point, but re-open crunch log with ArvClient in case there are any
1455         // other further errors (such as failing to write the log to Keep!)
1456         // while shutting down
1457         runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{ArvClient: runner.ArvClient,
1458                 UUID: runner.Container.UUID, loggingStream: "crunch-run", writeCloser: nil})
1459         runner.CrunchLog.Immediate = log.New(os.Stderr, runner.Container.UUID+" ", 0)
1460
1461         if runner.LogsPDH != nil {
1462                 // If we have already assigned something to LogsPDH,
1463                 // we must be closing the re-opened log, which won't
1464                 // end up getting attached to the container record and
1465                 // therefore doesn't need to be saved as a collection
1466                 // -- it exists only to send logs to other channels.
1467                 return nil
1468         }
1469
1470         mt, err := runner.LogCollection.ManifestText()
1471         if err != nil {
1472                 return fmt.Errorf("While creating log manifest: %v", err)
1473         }
1474
1475         var response arvados.Collection
1476         err = runner.ArvClient.Create("collections",
1477                 arvadosclient.Dict{
1478                         "ensure_unique_name": true,
1479                         "collection": arvadosclient.Dict{
1480                                 "is_trashed":    true,
1481                                 "name":          "logs for " + runner.Container.UUID,
1482                                 "manifest_text": mt}},
1483                 &response)
1484         if err != nil {
1485                 return fmt.Errorf("While creating log collection: %v", err)
1486         }
1487         runner.LogsPDH = &response.PortableDataHash
1488         return nil
1489 }
1490
1491 // UpdateContainerRunning updates the container state to "Running"
1492 func (runner *ContainerRunner) UpdateContainerRunning() error {
1493         runner.cStateLock.Lock()
1494         defer runner.cStateLock.Unlock()
1495         if runner.cCancelled {
1496                 return ErrCancelled
1497         }
1498         return runner.ArvClient.Update("containers", runner.Container.UUID,
1499                 arvadosclient.Dict{"container": arvadosclient.Dict{"state": "Running"}}, nil)
1500 }
1501
1502 // ContainerToken returns the api_token the container (and any
1503 // arv-mount processes) are allowed to use.
1504 func (runner *ContainerRunner) ContainerToken() (string, error) {
1505         if runner.token != "" {
1506                 return runner.token, nil
1507         }
1508
1509         var auth arvados.APIClientAuthorization
1510         err := runner.ArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth)
1511         if err != nil {
1512                 return "", err
1513         }
1514         runner.token = auth.APIToken
1515         return runner.token, nil
1516 }
1517
1518 // UpdateContainerComplete updates the container record state on API
1519 // server to "Complete" or "Cancelled"
1520 func (runner *ContainerRunner) UpdateContainerFinal() error {
1521         update := arvadosclient.Dict{}
1522         update["state"] = runner.finalState
1523         if runner.LogsPDH != nil {
1524                 update["log"] = *runner.LogsPDH
1525         }
1526         if runner.finalState == "Complete" {
1527                 if runner.ExitCode != nil {
1528                         update["exit_code"] = *runner.ExitCode
1529                 }
1530                 if runner.OutputPDH != nil {
1531                         update["output"] = *runner.OutputPDH
1532                 }
1533         }
1534         return runner.ArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{"container": update}, nil)
1535 }
1536
1537 // IsCancelled returns the value of Cancelled, with goroutine safety.
1538 func (runner *ContainerRunner) IsCancelled() bool {
1539         runner.cStateLock.Lock()
1540         defer runner.cStateLock.Unlock()
1541         return runner.cCancelled
1542 }
1543
1544 // NewArvLogWriter creates an ArvLogWriter
1545 func (runner *ContainerRunner) NewArvLogWriter(name string) io.WriteCloser {
1546         return &ArvLogWriter{
1547                 ArvClient:     runner.ArvClient,
1548                 UUID:          runner.Container.UUID,
1549                 loggingStream: name,
1550                 writeCloser:   runner.LogCollection.Open(name + ".txt")}
1551 }
1552
1553 // Run the full container lifecycle.
1554 func (runner *ContainerRunner) Run() (err error) {
1555         runner.CrunchLog.Printf("crunch-run %s started", version)
1556         runner.CrunchLog.Printf("Executing container '%s'", runner.Container.UUID)
1557
1558         hostname, hosterr := os.Hostname()
1559         if hosterr != nil {
1560                 runner.CrunchLog.Printf("Error getting hostname '%v'", hosterr)
1561         } else {
1562                 runner.CrunchLog.Printf("Executing on host '%s'", hostname)
1563         }
1564
1565         runner.finalState = "Queued"
1566
1567         defer func() {
1568                 runner.stopSignals()
1569                 runner.CleanupDirs()
1570
1571                 runner.CrunchLog.Printf("crunch-run finished")
1572                 runner.CrunchLog.Close()
1573         }()
1574
1575         defer func() {
1576                 // checkErr prints e (unless it's nil) and sets err to
1577                 // e (unless err is already non-nil). Thus, if err
1578                 // hasn't already been assigned when Run() returns,
1579                 // this cleanup func will cause Run() to return the
1580                 // first non-nil error that is passed to checkErr().
1581                 checkErr := func(e error) {
1582                         if e == nil {
1583                                 return
1584                         }
1585                         runner.CrunchLog.Print(e)
1586                         if err == nil {
1587                                 err = e
1588                         }
1589                         if runner.finalState == "Complete" {
1590                                 // There was an error in the finalization.
1591                                 runner.finalState = "Cancelled"
1592                         }
1593                 }
1594
1595                 // Log the error encountered in Run(), if any
1596                 checkErr(err)
1597
1598                 if runner.finalState == "Queued" {
1599                         runner.UpdateContainerFinal()
1600                         return
1601                 }
1602
1603                 if runner.IsCancelled() {
1604                         runner.finalState = "Cancelled"
1605                         // but don't return yet -- we still want to
1606                         // capture partial output and write logs
1607                 }
1608
1609                 checkErr(runner.CaptureOutput())
1610                 checkErr(runner.stopHoststat())
1611                 checkErr(runner.CommitLogs())
1612                 checkErr(runner.UpdateContainerFinal())
1613         }()
1614
1615         err = runner.fetchContainerRecord()
1616         if err != nil {
1617                 return
1618         }
1619         runner.setupSignals()
1620         runner.startHoststat()
1621
1622         // check for and/or load image
1623         err = runner.LoadImage()
1624         if err != nil {
1625                 if !runner.checkBrokenNode(err) {
1626                         // Failed to load image but not due to a "broken node"
1627                         // condition, probably user error.
1628                         runner.finalState = "Cancelled"
1629                 }
1630                 err = fmt.Errorf("While loading container image: %v", err)
1631                 return
1632         }
1633
1634         // set up FUSE mount and binds
1635         err = runner.SetupMounts()
1636         if err != nil {
1637                 runner.finalState = "Cancelled"
1638                 err = fmt.Errorf("While setting up mounts: %v", err)
1639                 return
1640         }
1641
1642         err = runner.CreateContainer()
1643         if err != nil {
1644                 return
1645         }
1646         err = runner.LogHostInfo()
1647         if err != nil {
1648                 return
1649         }
1650         err = runner.LogNodeRecord()
1651         if err != nil {
1652                 return
1653         }
1654         err = runner.LogContainerRecord()
1655         if err != nil {
1656                 return
1657         }
1658
1659         if runner.IsCancelled() {
1660                 return
1661         }
1662
1663         err = runner.UpdateContainerRunning()
1664         if err != nil {
1665                 return
1666         }
1667         runner.finalState = "Cancelled"
1668
1669         runner.startCrunchstat()
1670
1671         err = runner.StartContainer()
1672         if err != nil {
1673                 runner.checkBrokenNode(err)
1674                 return
1675         }
1676
1677         err = runner.WaitFinish()
1678         if err == nil && !runner.IsCancelled() {
1679                 runner.finalState = "Complete"
1680         }
1681         return
1682 }
1683
1684 // Fetch the current container record (uuid = runner.Container.UUID)
1685 // into runner.Container.
1686 func (runner *ContainerRunner) fetchContainerRecord() error {
1687         reader, err := runner.ArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
1688         if err != nil {
1689                 return fmt.Errorf("error fetching container record: %v", err)
1690         }
1691         defer reader.Close()
1692
1693         dec := json.NewDecoder(reader)
1694         dec.UseNumber()
1695         err = dec.Decode(&runner.Container)
1696         if err != nil {
1697                 return fmt.Errorf("error decoding container record: %v", err)
1698         }
1699         return nil
1700 }
1701
1702 // NewContainerRunner creates a new container runner.
1703 func NewContainerRunner(api IArvadosClient,
1704         kc IKeepClient,
1705         docker ThinDockerClient,
1706         containerUUID string) *ContainerRunner {
1707
1708         cr := &ContainerRunner{ArvClient: api, Kc: kc, Docker: docker}
1709         cr.NewLogWriter = cr.NewArvLogWriter
1710         cr.RunArvMount = cr.ArvMountCmd
1711         cr.MkTempDir = ioutil.TempDir
1712         cr.LogCollection = &CollectionWriter{0, kc, nil, nil, sync.Mutex{}}
1713         cr.Container.UUID = containerUUID
1714         cr.CrunchLog = NewThrottledLogger(cr.NewLogWriter("crunch-run"))
1715         cr.CrunchLog.Immediate = log.New(os.Stderr, containerUUID+" ", 0)
1716
1717         loadLogThrottleParams(api)
1718
1719         return cr
1720 }
1721
1722 func main() {
1723         statInterval := flag.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting")
1724         cgroupRoot := flag.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree")
1725         cgroupParent := flag.String("cgroup-parent", "docker", "name of container's parent cgroup (ignored if -cgroup-parent-subsystem is used)")
1726         cgroupParentSubsystem := flag.String("cgroup-parent-subsystem", "", "use current cgroup for given subsystem as parent cgroup for container")
1727         caCertsPath := flag.String("ca-certs", "", "Path to TLS root certificates")
1728         enableNetwork := flag.String("container-enable-networking", "default",
1729                 `Specify if networking should be enabled for container.  One of 'default', 'always':
1730         default: only enable networking if container requests it.
1731         always:  containers always have networking enabled
1732         `)
1733         networkMode := flag.String("container-network-mode", "default",
1734                 `Set networking mode for container.  Corresponds to Docker network mode (--net).
1735         `)
1736         memprofile := flag.String("memprofile", "", "write memory profile to `file` after running container")
1737         getVersion := flag.Bool("version", false, "Print version information and exit.")
1738         flag.Parse()
1739
1740         // Print version information if requested
1741         if *getVersion {
1742                 fmt.Printf("crunch-run %s\n", version)
1743                 return
1744         }
1745
1746         log.Printf("crunch-run %s started", version)
1747
1748         containerId := flag.Arg(0)
1749
1750         if *caCertsPath != "" {
1751                 arvadosclient.CertFiles = []string{*caCertsPath}
1752         }
1753
1754         api, err := arvadosclient.MakeArvadosClient()
1755         if err != nil {
1756                 log.Fatalf("%s: %v", containerId, err)
1757         }
1758         api.Retries = 8
1759
1760         kc, kcerr := keepclient.MakeKeepClient(api)
1761         if kcerr != nil {
1762                 log.Fatalf("%s: %v", containerId, kcerr)
1763         }
1764         kc.BlockCache = &keepclient.BlockCache{MaxBlocks: 2}
1765         kc.Retries = 4
1766
1767         // API version 1.21 corresponds to Docker 1.9, which is currently the
1768         // minimum version we want to support.
1769         docker, dockererr := dockerclient.NewClient(dockerclient.DefaultDockerHost, "1.21", nil, nil)
1770
1771         cr := NewContainerRunner(api, kc, docker, containerId)
1772         if dockererr != nil {
1773                 cr.CrunchLog.Printf("%s: %v", containerId, dockererr)
1774                 cr.checkBrokenNode(dockererr)
1775                 cr.CrunchLog.Close()
1776                 os.Exit(1)
1777         }
1778
1779         parentTemp, tmperr := cr.MkTempDir("", "crunch-run."+containerId+".")
1780         if tmperr != nil {
1781                 log.Fatalf("%s: %v", containerId, tmperr)
1782         }
1783
1784         cr.parentTemp = parentTemp
1785         cr.statInterval = *statInterval
1786         cr.cgroupRoot = *cgroupRoot
1787         cr.expectCgroupParent = *cgroupParent
1788         cr.enableNetwork = *enableNetwork
1789         cr.networkMode = *networkMode
1790         if *cgroupParentSubsystem != "" {
1791                 p := findCgroup(*cgroupParentSubsystem)
1792                 cr.setCgroupParent = p
1793                 cr.expectCgroupParent = p
1794         }
1795
1796         runerr := cr.Run()
1797
1798         if *memprofile != "" {
1799                 f, err := os.Create(*memprofile)
1800                 if err != nil {
1801                         log.Printf("could not create memory profile: ", err)
1802                 }
1803                 runtime.GC() // get up-to-date statistics
1804                 if err := pprof.WriteHeapProfile(f); err != nil {
1805                         log.Printf("could not write memory profile: ", err)
1806                 }
1807                 closeerr := f.Close()
1808                 if closeerr != nil {
1809                         log.Printf("closing memprofile file: ", err)
1810                 }
1811         }
1812
1813         if runerr != nil {
1814                 log.Fatalf("%s: %v", containerId, runerr)
1815         }
1816 }