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                 dir, 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 dir.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                                 if walkinfo.Mode().IsRegular() {
 588                                         return copyfile(walkpath, path.Join(cp.bind, walkpath[len(cp.src):]))
 589                                 } else if walkinfo.Mode().IsDir() {
 590                                         return os.MkdirAll(path.Join(cp.bind, walkpath[len(cp.src):]), 0777)
 591                                 } else {
 592                                         return fmt.Errorf("Source %q is not a regular file or directory", cp.src)
 593                                 }
 594                         })
 595                 } else {
 596                         err = copyfile(cp.src, cp.bind)
 597                 }
 598                 if err != nil {
 599                         return fmt.Errorf("While staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 600                 }
 601         }
 602
 603         return nil
 604 }
 605
 606 func (runner *ContainerRunner) ProcessDockerAttach(containerReader io.Reader) {
 607         // Handle docker log protocol
 608         // https://docs.docker.com/engine/reference/api/docker_remote_api_v1.15/#attach-to-a-container
 609         defer close(runner.loggingDone)
 610
 611         header := make([]byte, 8)
 612         var err error
 613         for err == nil {
 614                 _, err = io.ReadAtLeast(containerReader, header, 8)
 615                 if err != nil {
 616                         if err == io.EOF {
 617                                 err = nil
 618                         }
 619                         break
 620                 }
 621                 readsize := int64(header[7]) | (int64(header[6]) << 8) | (int64(header[5]) << 16) | (int64(header[4]) << 24)
 622                 if header[0] == 1 {
 623                         // stdout
 624                         _, err = io.CopyN(runner.Stdout, containerReader, readsize)
 625                 } else {
 626                         // stderr
 627                         _, err = io.CopyN(runner.Stderr, containerReader, readsize)
 628                 }
 629         }
 630
 631         if err != nil {
 632                 runner.CrunchLog.Printf("error reading docker logs: %v", err)
 633         }
 634
 635         err = runner.Stdout.Close()
 636         if err != nil {
 637                 runner.CrunchLog.Printf("error closing stdout logs: %v", err)
 638         }
 639
 640         err = runner.Stderr.Close()
 641         if err != nil {
 642                 runner.CrunchLog.Printf("error closing stderr logs: %v", err)
 643         }
 644
 645         if runner.statReporter != nil {
 646                 runner.statReporter.Stop()
 647                 err = runner.statLogger.Close()
 648                 if err != nil {
 649                         runner.CrunchLog.Printf("error closing crunchstat logs: %v", err)
 650                 }
 651         }
 652 }
 653
 654 func (runner *ContainerRunner) stopHoststat() error {
 655         if runner.hoststatReporter == nil {
 656                 return nil
 657         }
 658         runner.hoststatReporter.Stop()
 659         err := runner.hoststatLogger.Close()
 660         if err != nil {
 661                 return fmt.Errorf("error closing hoststat logs: %v", err)
 662         }
 663         return nil
 664 }
 665
 666 func (runner *ContainerRunner) startHoststat() {
 667         runner.hoststatLogger = NewThrottledLogger(runner.NewLogWriter("hoststat"))
 668         runner.hoststatReporter = &crunchstat.Reporter{
 669                 Logger:     log.New(runner.hoststatLogger, "", 0),
 670                 CgroupRoot: runner.cgroupRoot,
 671                 PollPeriod: runner.statInterval,
 672         }
 673         runner.hoststatReporter.Start()
 674 }
 675
 676 func (runner *ContainerRunner) startCrunchstat() {
 677         runner.statLogger = NewThrottledLogger(runner.NewLogWriter("crunchstat"))
 678         runner.statReporter = &crunchstat.Reporter{
 679                 CID:          runner.ContainerID,
 680                 Logger:       log.New(runner.statLogger, "", 0),
 681                 CgroupParent: runner.expectCgroupParent,
 682                 CgroupRoot:   runner.cgroupRoot,
 683                 PollPeriod:   runner.statInterval,
 684         }
 685         runner.statReporter.Start()
 686 }
 687
 688 type infoCommand struct {
 689         label string
 690         cmd   []string
 691 }
 692
 693 // LogHostInfo logs info about the current host, for debugging and
 694 // accounting purposes. Although it's logged as "node-info", this is
 695 // about the environment where crunch-run is actually running, which
 696 // might differ from what's described in the node record (see
 697 // LogNodeRecord).
 698 func (runner *ContainerRunner) LogHostInfo() (err error) {
 699         w := runner.NewLogWriter("node-info")
 700
 701         commands := []infoCommand{
 702                 {
 703                         label: "Host Information",
 704                         cmd:   []string{"uname", "-a"},
 705                 },
 706                 {
 707                         label: "CPU Information",
 708                         cmd:   []string{"cat", "/proc/cpuinfo"},
 709                 },
 710                 {
 711                         label: "Memory Information",
 712                         cmd:   []string{"cat", "/proc/meminfo"},
 713                 },
 714                 {
 715                         label: "Disk Space",
 716                         cmd:   []string{"df", "-m", "/", os.TempDir()},
 717                 },
 718                 {
 719                         label: "Disk INodes",
 720                         cmd:   []string{"df", "-i", "/", os.TempDir()},
 721                 },
 722         }
 723
 724         // Run commands with informational output to be logged.
 725         for _, command := range commands {
 726                 fmt.Fprintln(w, command.label)
 727                 cmd := exec.Command(command.cmd[0], command.cmd[1:]...)
 728                 cmd.Stdout = w
 729                 cmd.Stderr = w
 730                 if err := cmd.Run(); err != nil {
 731                         err = fmt.Errorf("While running command %q: %v", command.cmd, err)
 732                         fmt.Fprintln(w, err)
 733                         return err
 734                 }
 735                 fmt.Fprintln(w, "")
 736         }
 737
 738         err = w.Close()
 739         if err != nil {
 740                 return fmt.Errorf("While closing node-info logs: %v", err)
 741         }
 742         return nil
 743 }
 744
 745 // LogContainerRecord gets and saves the raw JSON container record from the API server
 746 func (runner *ContainerRunner) LogContainerRecord() error {
 747         logged, err := runner.logAPIResponse("container", "containers", map[string]interface{}{"filters": [][]string{{"uuid", "=", runner.Container.UUID}}}, nil)
 748         if !logged && err == nil {
 749                 err = fmt.Errorf("error: no container record found for %s", runner.Container.UUID)
 750         }
 751         return err
 752 }
 753
 754 // LogNodeRecord logs arvados#node record corresponding to the current host.
 755 func (runner *ContainerRunner) LogNodeRecord() error {
 756         hostname := os.Getenv("SLURMD_NODENAME")
 757         if hostname == "" {
 758                 hostname, _ = os.Hostname()
 759         }
 760         _, err := runner.logAPIResponse("node", "nodes", map[string]interface{}{"filters": [][]string{{"hostname", "=", hostname}}}, func(resp interface{}) {
 761                 // The "info" field has admin-only info when obtained
 762                 // with a privileged token, and should not be logged.
 763                 node, ok := resp.(map[string]interface{})
 764                 if ok {
 765                         delete(node, "info")
 766                 }
 767         })
 768         return err
 769 }
 770
 771 func (runner *ContainerRunner) logAPIResponse(label, path string, params map[string]interface{}, munge func(interface{})) (logged bool, err error) {
 772         w := &ArvLogWriter{
 773                 ArvClient:     runner.ArvClient,
 774                 UUID:          runner.Container.UUID,
 775                 loggingStream: label,
 776                 writeCloser:   runner.LogCollection.Open(label + ".json"),
 777         }
 778
 779         reader, err := runner.ArvClient.CallRaw("GET", path, "", "", arvadosclient.Dict(params))
 780         if err != nil {
 781                 return false, fmt.Errorf("error getting %s record: %v", label, err)
 782         }
 783         defer reader.Close()
 784
 785         dec := json.NewDecoder(reader)
 786         dec.UseNumber()
 787         var resp map[string]interface{}
 788         if err = dec.Decode(&resp); err != nil {
 789                 return false, fmt.Errorf("error decoding %s list response: %v", label, err)
 790         }
 791         items, ok := resp["items"].([]interface{})
 792         if !ok {
 793                 return false, fmt.Errorf("error decoding %s list response: no \"items\" key in API list response", label)
 794         } else if len(items) < 1 {
 795                 return false, nil
 796         }
 797         if munge != nil {
 798                 munge(items[0])
 799         }
 800         // Re-encode it using indentation to improve readability
 801         enc := json.NewEncoder(w)
 802         enc.SetIndent("", "    ")
 803         if err = enc.Encode(items[0]); err != nil {
 804                 return false, fmt.Errorf("error logging %s record: %v", label, err)
 805         }
 806         err = w.Close()
 807         if err != nil {
 808                 return false, fmt.Errorf("error closing %s.json in log collection: %v", label, err)
 809         }
 810         return true, nil
 811 }
 812
 813 // AttachStreams connects the docker container stdin, stdout and stderr logs
 814 // to the Arvados logger which logs to Keep and the API server logs table.
 815 func (runner *ContainerRunner) AttachStreams() (err error) {
 816
 817         runner.CrunchLog.Print("Attaching container streams")
 818
 819         // If stdin mount is provided, attach it to the docker container
 820         var stdinRdr arvados.File
 821         var stdinJson []byte
 822         if stdinMnt, ok := runner.Container.Mounts["stdin"]; ok {
 823                 if stdinMnt.Kind == "collection" {
 824                         var stdinColl arvados.Collection
 825                         collId := stdinMnt.UUID
 826                         if collId == "" {
 827                                 collId = stdinMnt.PortableDataHash
 828                         }
 829                         err = runner.ArvClient.Get("collections", collId, nil, &stdinColl)
 830                         if err != nil {
 831                                 return fmt.Errorf("While getting stding collection: %v", err)
 832                         }
 833
 834                         stdinRdr, err = runner.Kc.ManifestFileReader(manifest.Manifest{Text: stdinColl.ManifestText}, stdinMnt.Path)
 835                         if os.IsNotExist(err) {
 836                                 return fmt.Errorf("stdin collection path not found: %v", stdinMnt.Path)
 837                         } else if err != nil {
 838                                 return fmt.Errorf("While getting stdin collection path %v: %v", stdinMnt.Path, err)
 839                         }
 840                 } else if stdinMnt.Kind == "json" {
 841                         stdinJson, err = json.Marshal(stdinMnt.Content)
 842                         if err != nil {
 843                                 return fmt.Errorf("While encoding stdin json data: %v", err)
 844                         }
 845                 }
 846         }
 847
 848         stdinUsed := stdinRdr != nil || len(stdinJson) != 0
 849         response, err := runner.Docker.ContainerAttach(context.TODO(), runner.ContainerID,
 850                 dockertypes.ContainerAttachOptions{Stream: true, Stdin: stdinUsed, Stdout: true, Stderr: true})
 851         if err != nil {
 852                 return fmt.Errorf("While attaching container stdout/stderr streams: %v", err)
 853         }
 854
 855         runner.loggingDone = make(chan bool)
 856
 857         if stdoutMnt, ok := runner.Container.Mounts["stdout"]; ok {
 858                 stdoutFile, err := runner.getStdoutFile(stdoutMnt.Path)
 859                 if err != nil {
 860                         return err
 861                 }
 862                 runner.Stdout = stdoutFile
 863         } else {
 864                 runner.Stdout = NewThrottledLogger(runner.NewLogWriter("stdout"))
 865         }
 866
 867         if stderrMnt, ok := runner.Container.Mounts["stderr"]; ok {
 868                 stderrFile, err := runner.getStdoutFile(stderrMnt.Path)
 869                 if err != nil {
 870                         return err
 871                 }
 872                 runner.Stderr = stderrFile
 873         } else {
 874                 runner.Stderr = NewThrottledLogger(runner.NewLogWriter("stderr"))
 875         }
 876
 877         if stdinRdr != nil {
 878                 go func() {
 879                         _, err := io.Copy(response.Conn, stdinRdr)
 880                         if err != nil {
 881                                 runner.CrunchLog.Print("While writing stdin collection to docker container %q", err)
 882                                 runner.stop()
 883                         }
 884                         stdinRdr.Close()
 885                         response.CloseWrite()
 886                 }()
 887         } else if len(stdinJson) != 0 {
 888                 go func() {
 889                         _, err := io.Copy(response.Conn, bytes.NewReader(stdinJson))
 890                         if err != nil {
 891                                 runner.CrunchLog.Print("While writing stdin json to docker container %q", err)
 892                                 runner.stop()
 893                         }
 894                         response.CloseWrite()
 895                 }()
 896         }
 897
 898         go runner.ProcessDockerAttach(response.Reader)
 899
 900         return nil
 901 }
 902
 903 func (runner *ContainerRunner) getStdoutFile(mntPath string) (*os.File, error) {
 904         stdoutPath := mntPath[len(runner.Container.OutputPath):]
 905         index := strings.LastIndex(stdoutPath, "/")
 906         if index > 0 {
 907                 subdirs := stdoutPath[:index]
 908                 if subdirs != "" {
 909                         st, err := os.Stat(runner.HostOutputDir)
 910                         if err != nil {
 911                                 return nil, fmt.Errorf("While Stat on temp dir: %v", err)
 912                         }
 913                         stdoutPath := filepath.Join(runner.HostOutputDir, subdirs)
 914                         err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777)
 915                         if err != nil {
 916                                 return nil, fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err)
 917                         }
 918                 }
 919         }
 920         stdoutFile, err := os.Create(filepath.Join(runner.HostOutputDir, stdoutPath))
 921         if err != nil {
 922                 return nil, fmt.Errorf("While creating file %q: %v", stdoutPath, err)
 923         }
 924
 925         return stdoutFile, nil
 926 }
 927
 928 // CreateContainer creates the docker container.
 929 func (runner *ContainerRunner) CreateContainer() error {
 930         runner.CrunchLog.Print("Creating Docker container")
 931
 932         runner.ContainerConfig.Cmd = runner.Container.Command
 933         if runner.Container.Cwd != "." {
 934                 runner.ContainerConfig.WorkingDir = runner.Container.Cwd
 935         }
 936
 937         for k, v := range runner.Container.Environment {
 938                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env, k+"="+v)
 939         }
 940
 941         runner.ContainerConfig.Volumes = runner.Volumes
 942
 943         runner.HostConfig = dockercontainer.HostConfig{
 944                 Binds: runner.Binds,
 945                 LogConfig: dockercontainer.LogConfig{
 946                         Type: "none",
 947                 },
 948                 Resources: dockercontainer.Resources{
 949                         CgroupParent: runner.setCgroupParent,
 950                 },
 951         }
 952
 953         if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
 954                 tok, err := runner.ContainerToken()
 955                 if err != nil {
 956                         return err
 957                 }
 958                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env,
 959                         "ARVADOS_API_TOKEN="+tok,
 960                         "ARVADOS_API_HOST="+os.Getenv("ARVADOS_API_HOST"),
 961                         "ARVADOS_API_HOST_INSECURE="+os.Getenv("ARVADOS_API_HOST_INSECURE"),
 962                 )
 963                 runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
 964         } else {
 965                 if runner.enableNetwork == "always" {
 966                         runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
 967                 } else {
 968                         runner.HostConfig.NetworkMode = dockercontainer.NetworkMode("none")
 969                 }
 970         }
 971
 972         _, stdinUsed := runner.Container.Mounts["stdin"]
 973         runner.ContainerConfig.OpenStdin = stdinUsed
 974         runner.ContainerConfig.StdinOnce = stdinUsed
 975         runner.ContainerConfig.AttachStdin = stdinUsed
 976         runner.ContainerConfig.AttachStdout = true
 977         runner.ContainerConfig.AttachStderr = true
 978
 979         createdBody, err := runner.Docker.ContainerCreate(context.TODO(), &runner.ContainerConfig, &runner.HostConfig, nil, runner.Container.UUID)
 980         if err != nil {
 981                 return fmt.Errorf("While creating container: %v", err)
 982         }
 983
 984         runner.ContainerID = createdBody.ID
 985
 986         return runner.AttachStreams()
 987 }
 988
 989 // StartContainer starts the docker container created by CreateContainer.
 990 func (runner *ContainerRunner) StartContainer() error {
 991         runner.CrunchLog.Printf("Starting Docker container id '%s'", runner.ContainerID)
 992         runner.cStateLock.Lock()
 993         defer runner.cStateLock.Unlock()
 994         if runner.cCancelled {
 995                 return ErrCancelled
 996         }
 997         err := runner.Docker.ContainerStart(context.TODO(), runner.ContainerID,
 998                 dockertypes.ContainerStartOptions{})
 999         if err != nil {
1000                 var advice string
1001                 if m, e := regexp.MatchString("(?ms).*(exec|System error).*(no such file or directory|file not found).*", err.Error()); m && e == nil {
1002                         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])
1003                 }
1004                 return fmt.Errorf("could not start container: %v%s", err, advice)
1005         }
1006         return nil
1007 }
1008
1009 // WaitFinish waits for the container to terminate, capture the exit code, and
1010 // close the stdout/stderr logging.
1011 func (runner *ContainerRunner) WaitFinish() error {
1012         runner.CrunchLog.Print("Waiting for container to finish")
1013
1014         waitOk, waitErr := runner.Docker.ContainerWait(context.TODO(), runner.ContainerID, dockercontainer.WaitConditionNotRunning)
1015         arvMountExit := runner.ArvMountExit
1016         for {
1017                 select {
1018                 case waitBody := <-waitOk:
1019                         runner.CrunchLog.Printf("Container exited with code: %v", waitBody.StatusCode)
1020                         code := int(waitBody.StatusCode)
1021                         runner.ExitCode = &code
1022
1023                         // wait for stdout/stderr to complete
1024                         <-runner.loggingDone
1025                         return nil
1026
1027                 case err := <-waitErr:
1028                         return fmt.Errorf("container wait: %v", err)
1029
1030                 case <-arvMountExit:
1031                         runner.CrunchLog.Printf("arv-mount exited while container is still running.  Stopping container.")
1032                         runner.stop()
1033                         // arvMountExit will always be ready now that
1034                         // it's closed, but that doesn't interest us.
1035                         arvMountExit = nil
1036                 }
1037         }
1038 }
1039
1040 var ErrNotInOutputDir = fmt.Errorf("Must point to path within the output directory")
1041
1042 func (runner *ContainerRunner) derefOutputSymlink(path string, startinfo os.FileInfo) (tgt string, readlinktgt string, info os.FileInfo, err error) {
1043         // Follow symlinks if necessary
1044         info = startinfo
1045         tgt = path
1046         readlinktgt = ""
1047         nextlink := path
1048         for followed := 0; info.Mode()&os.ModeSymlink != 0; followed++ {
1049                 if followed >= limitFollowSymlinks {
1050                         // Got stuck in a loop or just a pathological number of links, give up.
1051                         err = fmt.Errorf("Followed more than %v symlinks from path %q", limitFollowSymlinks, path)
1052                         return
1053                 }
1054
1055                 readlinktgt, err = os.Readlink(nextlink)
1056                 if err != nil {
1057                         return
1058                 }
1059
1060                 tgt = readlinktgt
1061                 if !strings.HasPrefix(tgt, "/") {
1062                         // Relative symlink, resolve it to host path
1063                         tgt = filepath.Join(filepath.Dir(path), tgt)
1064                 }
1065                 if strings.HasPrefix(tgt, runner.Container.OutputPath+"/") && !strings.HasPrefix(tgt, runner.HostOutputDir+"/") {
1066                         // Absolute symlink to container output path, adjust it to host output path.
1067                         tgt = filepath.Join(runner.HostOutputDir, tgt[len(runner.Container.OutputPath):])
1068                 }
1069                 if !strings.HasPrefix(tgt, runner.HostOutputDir+"/") {
1070                         // After dereferencing, symlink target must either be
1071                         // within output directory, or must point to a
1072                         // collection mount.
1073                         err = ErrNotInOutputDir
1074                         return
1075                 }
1076
1077                 info, err = os.Lstat(tgt)
1078                 if err != nil {
1079                         // tgt
1080                         err = fmt.Errorf("Symlink in output %q points to invalid location %q: %v",
1081                                 path[len(runner.HostOutputDir):], readlinktgt, err)
1082                         return
1083                 }
1084
1085                 nextlink = tgt
1086         }
1087
1088         return
1089 }
1090
1091 var limitFollowSymlinks = 10
1092
1093 // UploadFile uploads files within the output directory, with special handling
1094 // for symlinks. If the symlink leads to a keep mount, copy the manifest text
1095 // from the keep mount into the output manifestText.  Ensure that whether
1096 // symlinks are relative or absolute, every symlink target (even targets that
1097 // are symlinks themselves) must point to a path in either the output directory
1098 // or a collection mount.
1099 //
1100 // Assumes initial value of "path" is absolute, and located within runner.HostOutputDir.
1101 func (runner *ContainerRunner) UploadOutputFile(
1102         path string,
1103         info os.FileInfo,
1104         infoerr error,
1105         binds []string,
1106         walkUpload *WalkUpload,
1107         relocateFrom string,
1108         relocateTo string,
1109         followed int) (manifestText string, err error) {
1110
1111         if infoerr != nil {
1112                 return "", infoerr
1113         }
1114
1115         if info.Mode().IsDir() {
1116                 // if empty, need to create a .keep file
1117                 dir, direrr := os.Open(path)
1118                 if direrr != nil {
1119                         return "", direrr
1120                 }
1121                 defer dir.Close()
1122                 names, eof := dir.Readdirnames(1)
1123                 if len(names) == 0 && eof == io.EOF && path != runner.HostOutputDir {
1124                         containerPath := runner.OutputPath + path[len(runner.HostOutputDir):]
1125                         for _, bind := range binds {
1126                                 mnt := runner.Container.Mounts[bind]
1127                                 // Check if there is a bind for this
1128                                 // directory, in which case assume we don't need .keep
1129                                 if (containerPath == bind || strings.HasPrefix(containerPath, bind+"/")) && mnt.PortableDataHash != "d41d8cd98f00b204e9800998ecf8427e+0" {
1130                                         return
1131                                 }
1132                         }
1133                         outputSuffix := path[len(runner.HostOutputDir)+1:]
1134                         return fmt.Sprintf("./%v d41d8cd98f00b204e9800998ecf8427e+0 0:0:.keep\n", outputSuffix), nil
1135                 }
1136                 return
1137         }
1138
1139         if followed >= limitFollowSymlinks {
1140                 // Got stuck in a loop or just a pathological number of
1141                 // directory links, give up.
1142                 err = fmt.Errorf("Followed more than %v symlinks from path %q", limitFollowSymlinks, path)
1143                 return
1144         }
1145
1146         // "path" is the actual path we are visiting
1147         // "tgt" is the target of "path" (a non-symlink) after following symlinks
1148         // "relocated" is the path in the output manifest where the file should be placed,
1149         // but has HostOutputDir as a prefix.
1150
1151         // The destination path in the output manifest may need to be
1152         // logically relocated to some other path in order to appear
1153         // in the correct location as a result of following a symlink.
1154         // Remove the relocateFrom prefix and replace it with
1155         // relocateTo.
1156         relocated := relocateTo + path[len(relocateFrom):]
1157
1158         tgt, readlinktgt, info, derefErr := runner.derefOutputSymlink(path, info)
1159         if derefErr != nil && derefErr != ErrNotInOutputDir {
1160                 return "", derefErr
1161         }
1162
1163         // go through mounts and try reverse map to collection reference
1164         for _, bind := range binds {
1165                 mnt := runner.Container.Mounts[bind]
1166                 if (tgt == bind || strings.HasPrefix(tgt, bind+"/")) && !mnt.Writable {
1167                         // get path relative to bind
1168                         targetSuffix := tgt[len(bind):]
1169
1170                         // Copy mount and adjust the path to add path relative to the bind
1171                         adjustedMount := mnt
1172                         adjustedMount.Path = filepath.Join(adjustedMount.Path, targetSuffix)
1173
1174                         // Terminates in this keep mount, so add the
1175                         // manifest text at appropriate location.
1176                         outputSuffix := relocated[len(runner.HostOutputDir):]
1177                         manifestText, err = runner.getCollectionManifestForPath(adjustedMount, outputSuffix)
1178                         return
1179                 }
1180         }
1181
1182         // If target is not a collection mount, it must be located within the
1183         // output directory, otherwise it is an error.
1184         if derefErr == ErrNotInOutputDir {
1185                 err = fmt.Errorf("Symlink in output %q points to invalid location %q, must point to path within the output directory.",
1186                         path[len(runner.HostOutputDir):], readlinktgt)
1187                 return
1188         }
1189
1190         if info.Mode().IsRegular() {
1191                 return "", walkUpload.UploadFile(relocated, tgt)
1192         }
1193
1194         if info.Mode().IsDir() {
1195                 // Symlink leads to directory.  Walk() doesn't follow
1196                 // directory symlinks, so we walk the target directory
1197                 // instead.  Within the walk, file paths are relocated
1198                 // so they appear under the original symlink path.
1199                 err = filepath.Walk(tgt, func(walkpath string, walkinfo os.FileInfo, walkerr error) error {
1200                         var m string
1201                         m, walkerr = runner.UploadOutputFile(walkpath, walkinfo, walkerr,
1202                                 binds, walkUpload, tgt, relocated, followed+1)
1203                         if walkerr == nil {
1204                                 manifestText = manifestText + m
1205                         }
1206                         return walkerr
1207                 })
1208                 return
1209         }
1210
1211         return
1212 }
1213
1214 // HandleOutput sets the output, unmounts the FUSE mount, and deletes temporary directories
1215 func (runner *ContainerRunner) CaptureOutput() error {
1216         if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
1217                 // Output may have been set directly by the container, so
1218                 // refresh the container record to check.
1219                 err := runner.ArvClient.Get("containers", runner.Container.UUID,
1220                         nil, &runner.Container)
1221                 if err != nil {
1222                         return err
1223                 }
1224                 if runner.Container.Output != "" {
1225                         // Container output is already set.
1226                         runner.OutputPDH = &runner.Container.Output
1227                         return nil
1228                 }
1229         }
1230
1231         if runner.HostOutputDir == "" {
1232                 return nil
1233         }
1234
1235         _, err := os.Stat(runner.HostOutputDir)
1236         if err != nil {
1237                 return fmt.Errorf("While checking host output path: %v", err)
1238         }
1239
1240         // Pre-populate output from the configured mount points
1241         var binds []string
1242         for bind, mnt := range runner.Container.Mounts {
1243                 if mnt.Kind == "collection" {
1244                         binds = append(binds, bind)
1245                 }
1246         }
1247         sort.Strings(binds)
1248
1249         var manifestText string
1250
1251         collectionMetafile := fmt.Sprintf("%s/.arvados#collection", runner.HostOutputDir)
1252         _, err = os.Stat(collectionMetafile)
1253         if err != nil {
1254                 // Regular directory
1255
1256                 cw := CollectionWriter{0, runner.Kc, nil, nil, sync.Mutex{}}
1257                 walkUpload := cw.BeginUpload(runner.HostOutputDir, runner.CrunchLog.Logger)
1258
1259                 var m string
1260                 err = filepath.Walk(runner.HostOutputDir, func(path string, info os.FileInfo, err error) error {
1261                         m, err = runner.UploadOutputFile(path, info, err, binds, walkUpload, "", "", 0)
1262                         if err == nil {
1263                                 manifestText = manifestText + m
1264                         }
1265                         return err
1266                 })
1267
1268                 cw.EndUpload(walkUpload)
1269
1270                 if err != nil {
1271                         return fmt.Errorf("While uploading output files: %v", err)
1272                 }
1273
1274                 m, err = cw.ManifestText()
1275                 manifestText = manifestText + m
1276                 if err != nil {
1277                         return fmt.Errorf("While uploading output files: %v", err)
1278                 }
1279         } else {
1280                 // FUSE mount directory
1281                 file, openerr := os.Open(collectionMetafile)
1282                 if openerr != nil {
1283                         return fmt.Errorf("While opening FUSE metafile: %v", err)
1284                 }
1285                 defer file.Close()
1286
1287                 var rec arvados.Collection
1288                 err = json.NewDecoder(file).Decode(&rec)
1289                 if err != nil {
1290                         return fmt.Errorf("While reading FUSE metafile: %v", err)
1291                 }
1292                 manifestText = rec.ManifestText
1293         }
1294
1295         for _, bind := range binds {
1296                 mnt := runner.Container.Mounts[bind]
1297
1298                 bindSuffix := strings.TrimPrefix(bind, runner.Container.OutputPath)
1299
1300                 if bindSuffix == bind || len(bindSuffix) <= 0 {
1301                         // either does not start with OutputPath or is OutputPath itself
1302                         continue
1303                 }
1304
1305                 if mnt.ExcludeFromOutput == true || mnt.Writable {
1306                         continue
1307                 }
1308
1309                 // append to manifest_text
1310                 m, err := runner.getCollectionManifestForPath(mnt, bindSuffix)
1311                 if err != nil {
1312                         return err
1313                 }
1314
1315                 manifestText = manifestText + m
1316         }
1317
1318         // Save output
1319         var response arvados.Collection
1320         manifest := manifest.Manifest{Text: manifestText}
1321         manifestText = manifest.Extract(".", ".").Text
1322         err = runner.ArvClient.Create("collections",
1323                 arvadosclient.Dict{
1324                         "ensure_unique_name": true,
1325                         "collection": arvadosclient.Dict{
1326                                 "is_trashed":    true,
1327                                 "name":          "output for " + runner.Container.UUID,
1328                                 "manifest_text": manifestText}},
1329                 &response)
1330         if err != nil {
1331                 return fmt.Errorf("While creating output collection: %v", err)
1332         }
1333         runner.OutputPDH = &response.PortableDataHash
1334         return nil
1335 }
1336
1337 var outputCollections = make(map[string]arvados.Collection)
1338
1339 // Fetch the collection for the mnt.PortableDataHash
1340 // Return the manifest_text fragment corresponding to the specified mnt.Path
1341 //  after making any required updates.
1342 //  Ex:
1343 //    If mnt.Path is not specified,
1344 //      return the entire manifest_text after replacing any "." with bindSuffix
1345 //    If mnt.Path corresponds to one stream,
1346 //      return the manifest_text for that stream after replacing that stream name with bindSuffix
1347 //    Otherwise, check if a filename in any one stream is being sought. Return the manifest_text
1348 //      for that stream after replacing stream name with bindSuffix minus the last word
1349 //      and the file name with last word of the bindSuffix
1350 //  Allowed path examples:
1351 //    "path":"/"
1352 //    "path":"/subdir1"
1353 //    "path":"/subdir1/subdir2"
1354 //    "path":"/subdir/filename" etc
1355 func (runner *ContainerRunner) getCollectionManifestForPath(mnt arvados.Mount, bindSuffix string) (string, error) {
1356         collection := outputCollections[mnt.PortableDataHash]
1357         if collection.PortableDataHash == "" {
1358                 err := runner.ArvClient.Get("collections", mnt.PortableDataHash, nil, &collection)
1359                 if err != nil {
1360                         return "", fmt.Errorf("While getting collection for %v: %v", mnt.PortableDataHash, err)
1361                 }
1362                 outputCollections[mnt.PortableDataHash] = collection
1363         }
1364
1365         if collection.ManifestText == "" {
1366                 runner.CrunchLog.Printf("No manifest text for collection %v", collection.PortableDataHash)
1367                 return "", nil
1368         }
1369
1370         mft := manifest.Manifest{Text: collection.ManifestText}
1371         extracted := mft.Extract(mnt.Path, bindSuffix)
1372         if extracted.Err != nil {
1373                 return "", fmt.Errorf("Error parsing manifest for %v: %v", mnt.PortableDataHash, extracted.Err.Error())
1374         }
1375         return extracted.Text, nil
1376 }
1377
1378 func (runner *ContainerRunner) CleanupDirs() {
1379         if runner.ArvMount != nil {
1380                 var delay int64 = 8
1381                 umount := exec.Command("arv-mount", fmt.Sprintf("--unmount-timeout=%d", delay), "--unmount", runner.ArvMountPoint)
1382                 umount.Stdout = runner.CrunchLog
1383                 umount.Stderr = runner.CrunchLog
1384                 runner.CrunchLog.Printf("Running %v", umount.Args)
1385                 umnterr := umount.Start()
1386
1387                 if umnterr != nil {
1388                         runner.CrunchLog.Printf("Error unmounting: %v", umnterr)
1389                 } else {
1390                         // If arv-mount --unmount gets stuck for any reason, we
1391                         // don't want to wait for it forever.  Do Wait() in a goroutine
1392                         // so it doesn't block crunch-run.
1393                         umountExit := make(chan error)
1394                         go func() {
1395                                 mnterr := umount.Wait()
1396                                 if mnterr != nil {
1397                                         runner.CrunchLog.Printf("Error unmounting: %v", mnterr)
1398                                 }
1399                                 umountExit <- mnterr
1400                         }()
1401
1402                         for again := true; again; {
1403                                 again = false
1404                                 select {
1405                                 case <-umountExit:
1406                                         umount = nil
1407                                         again = true
1408                                 case <-runner.ArvMountExit:
1409                                         break
1410                                 case <-time.After(time.Duration((delay + 1) * int64(time.Second))):
1411                                         runner.CrunchLog.Printf("Timed out waiting for unmount")
1412                                         if umount != nil {
1413                                                 umount.Process.Kill()
1414                                         }
1415                                         runner.ArvMount.Process.Kill()
1416                                 }
1417                         }
1418                 }
1419         }
1420
1421         if runner.ArvMountPoint != "" {
1422                 if rmerr := os.Remove(runner.ArvMountPoint); rmerr != nil {
1423                         runner.CrunchLog.Printf("While cleaning up arv-mount directory %s: %v", runner.ArvMountPoint, rmerr)
1424                 }
1425         }
1426
1427         if rmerr := os.RemoveAll(runner.parentTemp); rmerr != nil {
1428                 runner.CrunchLog.Printf("While cleaning up temporary directory %s: %v", runner.parentTemp, rmerr)
1429         }
1430 }
1431
1432 // CommitLogs posts the collection containing the final container logs.
1433 func (runner *ContainerRunner) CommitLogs() error {
1434         runner.CrunchLog.Print(runner.finalState)
1435
1436         if runner.arvMountLog != nil {
1437                 runner.arvMountLog.Close()
1438         }
1439         runner.CrunchLog.Close()
1440
1441         // Closing CrunchLog above allows them to be committed to Keep at this
1442         // point, but re-open crunch log with ArvClient in case there are any
1443         // other further errors (such as failing to write the log to Keep!)
1444         // while shutting down
1445         runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{ArvClient: runner.ArvClient,
1446                 UUID: runner.Container.UUID, loggingStream: "crunch-run", writeCloser: nil})
1447         runner.CrunchLog.Immediate = log.New(os.Stderr, runner.Container.UUID+" ", 0)
1448
1449         if runner.LogsPDH != nil {
1450                 // If we have already assigned something to LogsPDH,
1451                 // we must be closing the re-opened log, which won't
1452                 // end up getting attached to the container record and
1453                 // therefore doesn't need to be saved as a collection
1454                 // -- it exists only to send logs to other channels.
1455                 return nil
1456         }
1457
1458         mt, err := runner.LogCollection.ManifestText()
1459         if err != nil {
1460                 return fmt.Errorf("While creating log manifest: %v", err)
1461         }
1462
1463         var response arvados.Collection
1464         err = runner.ArvClient.Create("collections",
1465                 arvadosclient.Dict{
1466                         "ensure_unique_name": true,
1467                         "collection": arvadosclient.Dict{
1468                                 "is_trashed":    true,
1469                                 "name":          "logs for " + runner.Container.UUID,
1470                                 "manifest_text": mt}},
1471                 &response)
1472         if err != nil {
1473                 return fmt.Errorf("While creating log collection: %v", err)
1474         }
1475         runner.LogsPDH = &response.PortableDataHash
1476         return nil
1477 }
1478
1479 // UpdateContainerRunning updates the container state to "Running"
1480 func (runner *ContainerRunner) UpdateContainerRunning() error {
1481         runner.cStateLock.Lock()
1482         defer runner.cStateLock.Unlock()
1483         if runner.cCancelled {
1484                 return ErrCancelled
1485         }
1486         return runner.ArvClient.Update("containers", runner.Container.UUID,
1487                 arvadosclient.Dict{"container": arvadosclient.Dict{"state": "Running"}}, nil)
1488 }
1489
1490 // ContainerToken returns the api_token the container (and any
1491 // arv-mount processes) are allowed to use.
1492 func (runner *ContainerRunner) ContainerToken() (string, error) {
1493         if runner.token != "" {
1494                 return runner.token, nil
1495         }
1496
1497         var auth arvados.APIClientAuthorization
1498         err := runner.ArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth)
1499         if err != nil {
1500                 return "", err
1501         }
1502         runner.token = auth.APIToken
1503         return runner.token, nil
1504 }
1505
1506 // UpdateContainerComplete updates the container record state on API
1507 // server to "Complete" or "Cancelled"
1508 func (runner *ContainerRunner) UpdateContainerFinal() error {
1509         update := arvadosclient.Dict{}
1510         update["state"] = runner.finalState
1511         if runner.LogsPDH != nil {
1512                 update["log"] = *runner.LogsPDH
1513         }
1514         if runner.finalState == "Complete" {
1515                 if runner.ExitCode != nil {
1516                         update["exit_code"] = *runner.ExitCode
1517                 }
1518                 if runner.OutputPDH != nil {
1519                         update["output"] = *runner.OutputPDH
1520                 }
1521         }
1522         return runner.ArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{"container": update}, nil)
1523 }
1524
1525 // IsCancelled returns the value of Cancelled, with goroutine safety.
1526 func (runner *ContainerRunner) IsCancelled() bool {
1527         runner.cStateLock.Lock()
1528         defer runner.cStateLock.Unlock()
1529         return runner.cCancelled
1530 }
1531
1532 // NewArvLogWriter creates an ArvLogWriter
1533 func (runner *ContainerRunner) NewArvLogWriter(name string) io.WriteCloser {
1534         return &ArvLogWriter{
1535                 ArvClient:     runner.ArvClient,
1536                 UUID:          runner.Container.UUID,
1537                 loggingStream: name,
1538                 writeCloser:   runner.LogCollection.Open(name + ".txt")}
1539 }
1540
1541 // Run the full container lifecycle.
1542 func (runner *ContainerRunner) Run() (err error) {
1543         runner.CrunchLog.Printf("crunch-run %s started", version)
1544         runner.CrunchLog.Printf("Executing container '%s'", runner.Container.UUID)
1545
1546         hostname, hosterr := os.Hostname()
1547         if hosterr != nil {
1548                 runner.CrunchLog.Printf("Error getting hostname '%v'", hosterr)
1549         } else {
1550                 runner.CrunchLog.Printf("Executing on host '%s'", hostname)
1551         }
1552
1553         runner.finalState = "Queued"
1554
1555         defer func() {
1556                 runner.stopSignals()
1557                 runner.CleanupDirs()
1558
1559                 runner.CrunchLog.Printf("crunch-run finished")
1560                 runner.CrunchLog.Close()
1561         }()
1562
1563         defer func() {
1564                 // checkErr prints e (unless it's nil) and sets err to
1565                 // e (unless err is already non-nil). Thus, if err
1566                 // hasn't already been assigned when Run() returns,
1567                 // this cleanup func will cause Run() to return the
1568                 // first non-nil error that is passed to checkErr().
1569                 checkErr := func(e error) {
1570                         if e == nil {
1571                                 return
1572                         }
1573                         runner.CrunchLog.Print(e)
1574                         if err == nil {
1575                                 err = e
1576                         }
1577                         if runner.finalState == "Complete" {
1578                                 // There was an error in the finalization.
1579                                 runner.finalState = "Cancelled"
1580                         }
1581                 }
1582
1583                 // Log the error encountered in Run(), if any
1584                 checkErr(err)
1585
1586                 if runner.finalState == "Queued" {
1587                         runner.UpdateContainerFinal()
1588                         return
1589                 }
1590
1591                 if runner.IsCancelled() {
1592                         runner.finalState = "Cancelled"
1593                         // but don't return yet -- we still want to
1594                         // capture partial output and write logs
1595                 }
1596
1597                 checkErr(runner.CaptureOutput())
1598                 checkErr(runner.stopHoststat())
1599                 checkErr(runner.CommitLogs())
1600                 checkErr(runner.UpdateContainerFinal())
1601         }()
1602
1603         err = runner.fetchContainerRecord()
1604         if err != nil {
1605                 return
1606         }
1607         runner.setupSignals()
1608         runner.startHoststat()
1609
1610         // check for and/or load image
1611         err = runner.LoadImage()
1612         if err != nil {
1613                 if !runner.checkBrokenNode(err) {
1614                         // Failed to load image but not due to a "broken node"
1615                         // condition, probably user error.
1616                         runner.finalState = "Cancelled"
1617                 }
1618                 err = fmt.Errorf("While loading container image: %v", err)
1619                 return
1620         }
1621
1622         // set up FUSE mount and binds
1623         err = runner.SetupMounts()
1624         if err != nil {
1625                 runner.finalState = "Cancelled"
1626                 err = fmt.Errorf("While setting up mounts: %v", err)
1627                 return
1628         }
1629
1630         err = runner.CreateContainer()
1631         if err != nil {
1632                 return
1633         }
1634         err = runner.LogHostInfo()
1635         if err != nil {
1636                 return
1637         }
1638         err = runner.LogNodeRecord()
1639         if err != nil {
1640                 return
1641         }
1642         err = runner.LogContainerRecord()
1643         if err != nil {
1644                 return
1645         }
1646
1647         if runner.IsCancelled() {
1648                 return
1649         }
1650
1651         err = runner.UpdateContainerRunning()
1652         if err != nil {
1653                 return
1654         }
1655         runner.finalState = "Cancelled"
1656
1657         runner.startCrunchstat()
1658
1659         err = runner.StartContainer()
1660         if err != nil {
1661                 runner.checkBrokenNode(err)
1662                 return
1663         }
1664
1665         err = runner.WaitFinish()
1666         if err == nil && !runner.IsCancelled() {
1667                 runner.finalState = "Complete"
1668         }
1669         return
1670 }
1671
1672 // Fetch the current container record (uuid = runner.Container.UUID)
1673 // into runner.Container.
1674 func (runner *ContainerRunner) fetchContainerRecord() error {
1675         reader, err := runner.ArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
1676         if err != nil {
1677                 return fmt.Errorf("error fetching container record: %v", err)
1678         }
1679         defer reader.Close()
1680
1681         dec := json.NewDecoder(reader)
1682         dec.UseNumber()
1683         err = dec.Decode(&runner.Container)
1684         if err != nil {
1685                 return fmt.Errorf("error decoding container record: %v", err)
1686         }
1687         return nil
1688 }
1689
1690 // NewContainerRunner creates a new container runner.
1691 func NewContainerRunner(api IArvadosClient,
1692         kc IKeepClient,
1693         docker ThinDockerClient,
1694         containerUUID string) *ContainerRunner {
1695
1696         cr := &ContainerRunner{ArvClient: api, Kc: kc, Docker: docker}
1697         cr.NewLogWriter = cr.NewArvLogWriter
1698         cr.RunArvMount = cr.ArvMountCmd
1699         cr.MkTempDir = ioutil.TempDir
1700         cr.LogCollection = &CollectionWriter{0, kc, nil, nil, sync.Mutex{}}
1701         cr.Container.UUID = containerUUID
1702         cr.CrunchLog = NewThrottledLogger(cr.NewLogWriter("crunch-run"))
1703         cr.CrunchLog.Immediate = log.New(os.Stderr, containerUUID+" ", 0)
1704
1705         loadLogThrottleParams(api)
1706
1707         return cr
1708 }
1709
1710 func main() {
1711         statInterval := flag.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting")
1712         cgroupRoot := flag.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree")
1713         cgroupParent := flag.String("cgroup-parent", "docker", "name of container's parent cgroup (ignored if -cgroup-parent-subsystem is used)")
1714         cgroupParentSubsystem := flag.String("cgroup-parent-subsystem", "", "use current cgroup for given subsystem as parent cgroup for container")
1715         caCertsPath := flag.String("ca-certs", "", "Path to TLS root certificates")
1716         enableNetwork := flag.String("container-enable-networking", "default",
1717                 `Specify if networking should be enabled for container.  One of 'default', 'always':
1718         default: only enable networking if container requests it.
1719         always:  containers always have networking enabled
1720         `)
1721         networkMode := flag.String("container-network-mode", "default",
1722                 `Set networking mode for container.  Corresponds to Docker network mode (--net).
1723         `)
1724         memprofile := flag.String("memprofile", "", "write memory profile to `file` after running container")
1725         getVersion := flag.Bool("version", false, "Print version information and exit.")
1726         flag.Parse()
1727
1728         // Print version information if requested
1729         if *getVersion {
1730                 fmt.Printf("crunch-run %s\n", version)
1731                 return
1732         }
1733
1734         log.Printf("crunch-run %s started", version)
1735
1736         containerId := flag.Arg(0)
1737
1738         if *caCertsPath != "" {
1739                 arvadosclient.CertFiles = []string{*caCertsPath}
1740         }
1741
1742         api, err := arvadosclient.MakeArvadosClient()
1743         if err != nil {
1744                 log.Fatalf("%s: %v", containerId, err)
1745         }
1746         api.Retries = 8
1747
1748         kc, kcerr := keepclient.MakeKeepClient(api)
1749         if kcerr != nil {
1750                 log.Fatalf("%s: %v", containerId, kcerr)
1751         }
1752         kc.BlockCache = &keepclient.BlockCache{MaxBlocks: 2}
1753         kc.Retries = 4
1754
1755         // API version 1.21 corresponds to Docker 1.9, which is currently the
1756         // minimum version we want to support.
1757         docker, dockererr := dockerclient.NewClient(dockerclient.DefaultDockerHost, "1.21", nil, nil)
1758
1759         cr := NewContainerRunner(api, kc, docker, containerId)
1760         if dockererr != nil {
1761                 cr.CrunchLog.Printf("%s: %v", containerId, dockererr)
1762                 cr.checkBrokenNode(dockererr)
1763                 cr.CrunchLog.Close()
1764                 os.Exit(1)
1765         }
1766
1767         parentTemp, tmperr := cr.MkTempDir("", "crunch-run."+containerId+".")
1768         if tmperr != nil {
1769                 log.Fatalf("%s: %v", containerId, tmperr)
1770         }
1771
1772         cr.parentTemp = parentTemp
1773         cr.statInterval = *statInterval
1774         cr.cgroupRoot = *cgroupRoot
1775         cr.expectCgroupParent = *cgroupParent
1776         cr.enableNetwork = *enableNetwork
1777         cr.networkMode = *networkMode
1778         if *cgroupParentSubsystem != "" {
1779                 p := findCgroup(*cgroupParentSubsystem)
1780                 cr.setCgroupParent = p
1781                 cr.expectCgroupParent = p
1782         }
1783
1784         runerr := cr.Run()
1785
1786         if *memprofile != "" {
1787                 f, err := os.Create(*memprofile)
1788                 if err != nil {
1789                         log.Printf("could not create memory profile: ", err)
1790                 }
1791                 runtime.GC() // get up-to-date statistics
1792                 if err := pprof.WriteHeapProfile(f); err != nil {
1793                         log.Printf("could not write memory profile: ", err)
1794                 }
1795                 closeerr := f.Close()
1796                 if closeerr != nil {
1797                         log.Printf("closing memprofile file: ", err)
1798                 }
1799         }
1800
1801         if runerr != nil {
1802                 log.Fatalf("%s: %v", containerId, runerr)
1803         }
1804 }