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         "github.com/shirou/gopsutil/process"
  36         "golang.org/x/net/context"
  37
  38         dockertypes "github.com/docker/docker/api/types"
  39         dockercontainer "github.com/docker/docker/api/types/container"
  40         dockernetwork "github.com/docker/docker/api/types/network"
  41         dockerclient "github.com/docker/docker/client"
  42 )
  43
  44 var version = "dev"
  45
  46 // IArvadosClient is the minimal Arvados API methods used by crunch-run.
  47 type IArvadosClient interface {
  48         Create(resourceType string, parameters arvadosclient.Dict, output interface{}) error
  49         Get(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  50         Update(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  51         Call(method, resourceType, uuid, action string, parameters arvadosclient.Dict, output interface{}) error
  52         CallRaw(method string, resourceType string, uuid string, action string, parameters arvadosclient.Dict) (reader io.ReadCloser, err error)
  53         Discovery(key string) (interface{}, error)
  54 }
  55
  56 // ErrCancelled is the error returned when the container is cancelled.
  57 var ErrCancelled = errors.New("Cancelled")
  58
  59 // IKeepClient is the minimal Keep API methods used by crunch-run.
  60 type IKeepClient interface {
  61         PutB(buf []byte) (string, int, error)
  62         ReadAt(locator string, p []byte, off int) (int, error)
  63         ManifestFileReader(m manifest.Manifest, filename string) (arvados.File, error)
  64         ClearBlockCache()
  65 }
  66
  67 // NewLogWriter is a factory function to create a new log writer.
  68 type NewLogWriter func(name string) (io.WriteCloser, error)
  69
  70 type RunArvMount func(args []string, tok string) (*exec.Cmd, error)
  71
  72 type MkTempDir func(string, string) (string, error)
  73
  74 // ThinDockerClient is the minimal Docker client interface used by crunch-run.
  75 type ThinDockerClient interface {
  76         ContainerAttach(ctx context.Context, container string, options dockertypes.ContainerAttachOptions) (dockertypes.HijackedResponse, error)
  77         ContainerCreate(ctx context.Context, config *dockercontainer.Config, hostConfig *dockercontainer.HostConfig,
  78                 networkingConfig *dockernetwork.NetworkingConfig, containerName string) (dockercontainer.ContainerCreateCreatedBody, error)
  79         ContainerStart(ctx context.Context, container string, options dockertypes.ContainerStartOptions) error
  80         ContainerRemove(ctx context.Context, container string, options dockertypes.ContainerRemoveOptions) error
  81         ContainerWait(ctx context.Context, container string, condition dockercontainer.WaitCondition) (<-chan dockercontainer.ContainerWaitOKBody, <-chan error)
  82         ImageInspectWithRaw(ctx context.Context, image string) (dockertypes.ImageInspect, []byte, error)
  83         ImageLoad(ctx context.Context, input io.Reader, quiet bool) (dockertypes.ImageLoadResponse, error)
  84         ImageRemove(ctx context.Context, image string, options dockertypes.ImageRemoveOptions) ([]dockertypes.ImageDeleteResponseItem, error)
  85 }
  86
  87 // ContainerRunner is the main stateful struct used for a single execution of a
  88 // container.
  89 type ContainerRunner struct {
  90         Docker    ThinDockerClient
  91         client    *arvados.Client
  92         ArvClient IArvadosClient
  93         Kc        IKeepClient
  94         arvados.Container
  95         ContainerConfig dockercontainer.Config
  96         dockercontainer.HostConfig
  97         token       string
  98         ContainerID string
  99         ExitCode    *int
 100         NewLogWriter
 101         loggingDone   chan bool
 102         CrunchLog     *ThrottledLogger
 103         Stdout        io.WriteCloser
 104         Stderr        io.WriteCloser
 105         LogCollection arvados.CollectionFileSystem
 106         LogsPDH       *string
 107         RunArvMount
 108         MkTempDir
 109         ArvMount      *exec.Cmd
 110         ArvMountPoint string
 111         HostOutputDir string
 112         Binds         []string
 113         Volumes       map[string]struct{}
 114         OutputPDH     *string
 115         SigChan       chan os.Signal
 116         ArvMountExit  chan error
 117         SecretMounts  map[string]arvados.Mount
 118         MkArvClient   func(token string) (IArvadosClient, error)
 119         finalState    string
 120         parentTemp    string
 121
 122         statLogger       io.WriteCloser
 123         statReporter     *crunchstat.Reporter
 124         hoststatLogger   io.WriteCloser
 125         hoststatReporter *crunchstat.Reporter
 126         statInterval     time.Duration
 127         cgroupRoot       string
 128         // What we expect the container's cgroup parent to be.
 129         expectCgroupParent string
 130         // What we tell docker to use as the container's cgroup
 131         // parent. Note: Ideally we would use the same field for both
 132         // expectCgroupParent and setCgroupParent, and just make it
 133         // default to "docker". However, when using docker < 1.10 with
 134         // systemd, specifying a non-empty cgroup parent (even the
 135         // default value "docker") hits a docker bug
 136         // (https://github.com/docker/docker/issues/17126). Using two
 137         // separate fields makes it possible to use the "expect cgroup
 138         // parent to be X" feature even on sites where the "specify
 139         // cgroup parent" feature breaks.
 140         setCgroupParent string
 141
 142         cStateLock sync.Mutex
 143         cCancelled bool // StopContainer() invoked
 144
 145         enableNetwork   string // one of "default" or "always"
 146         networkMode     string // passed through to HostConfig.NetworkMode
 147         arvMountLog     *ThrottledLogger
 148         checkContainerd bool
 149 }
 150
 151 // setupSignals sets up signal handling to gracefully terminate the underlying
 152 // Docker container and update state when receiving a TERM, INT or QUIT signal.
 153 func (runner *ContainerRunner) setupSignals() {
 154         runner.SigChan = make(chan os.Signal, 1)
 155         signal.Notify(runner.SigChan, syscall.SIGTERM)
 156         signal.Notify(runner.SigChan, syscall.SIGINT)
 157         signal.Notify(runner.SigChan, syscall.SIGQUIT)
 158
 159         go func(sig chan os.Signal) {
 160                 for s := range sig {
 161                         runner.stop(s)
 162                 }
 163         }(runner.SigChan)
 164 }
 165
 166 // stop the underlying Docker container.
 167 func (runner *ContainerRunner) stop(sig os.Signal) {
 168         runner.cStateLock.Lock()
 169         defer runner.cStateLock.Unlock()
 170         if sig != nil {
 171                 runner.CrunchLog.Printf("caught signal: %v", sig)
 172         }
 173         if runner.ContainerID == "" {
 174                 return
 175         }
 176         runner.cCancelled = true
 177         runner.CrunchLog.Printf("removing container")
 178         err := runner.Docker.ContainerRemove(context.TODO(), runner.ContainerID, dockertypes.ContainerRemoveOptions{Force: true})
 179         if err != nil {
 180                 runner.CrunchLog.Printf("error removing container: %s", err)
 181         }
 182 }
 183
 184 var errorBlacklist = []string{
 185         "(?ms).*[Cc]annot connect to the Docker daemon.*",
 186         "(?ms).*oci runtime error.*starting container process.*container init.*mounting.*to rootfs.*no such file or directory.*",
 187 }
 188 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)")
 189
 190 func (runner *ContainerRunner) runBrokenNodeHook() {
 191         if *brokenNodeHook == "" {
 192                 runner.CrunchLog.Printf("No broken node hook provided, cannot mark node as broken.")
 193         } else {
 194                 runner.CrunchLog.Printf("Running broken node hook %q", *brokenNodeHook)
 195                 // run killme script
 196                 c := exec.Command(*brokenNodeHook)
 197                 c.Stdout = runner.CrunchLog
 198                 c.Stderr = runner.CrunchLog
 199                 err := c.Run()
 200                 if err != nil {
 201                         runner.CrunchLog.Printf("Error running broken node hook: %v", err)
 202                 }
 203         }
 204 }
 205
 206 func (runner *ContainerRunner) checkBrokenNode(goterr error) bool {
 207         for _, d := range errorBlacklist {
 208                 if m, e := regexp.MatchString(d, goterr.Error()); m && e == nil {
 209                         runner.CrunchLog.Printf("Error suggests node is unable to run containers: %v", goterr)
 210                         runner.runBrokenNodeHook()
 211                         return true
 212                 }
 213         }
 214         return false
 215 }
 216
 217 // LoadImage determines the docker image id from the container record and
 218 // checks if it is available in the local Docker image store.  If not, it loads
 219 // the image from Keep.
 220 func (runner *ContainerRunner) LoadImage() (err error) {
 221
 222         runner.CrunchLog.Printf("Fetching Docker image from collection '%s'", runner.Container.ContainerImage)
 223
 224         var collection arvados.Collection
 225         err = runner.ArvClient.Get("collections", runner.Container.ContainerImage, nil, &collection)
 226         if err != nil {
 227                 return fmt.Errorf("While getting container image collection: %v", err)
 228         }
 229         manifest := manifest.Manifest{Text: collection.ManifestText}
 230         var img, imageID string
 231         for ms := range manifest.StreamIter() {
 232                 img = ms.FileStreamSegments[0].Name
 233                 if !strings.HasSuffix(img, ".tar") {
 234                         return fmt.Errorf("First file in the container image collection does not end in .tar")
 235                 }
 236                 imageID = img[:len(img)-4]
 237         }
 238
 239         runner.CrunchLog.Printf("Using Docker image id '%s'", imageID)
 240
 241         _, _, err = runner.Docker.ImageInspectWithRaw(context.TODO(), imageID)
 242         if err != nil {
 243                 runner.CrunchLog.Print("Loading Docker image from keep")
 244
 245                 var readCloser io.ReadCloser
 246                 readCloser, err = runner.Kc.ManifestFileReader(manifest, img)
 247                 if err != nil {
 248                         return fmt.Errorf("While creating ManifestFileReader for container image: %v", err)
 249                 }
 250
 251                 response, err := runner.Docker.ImageLoad(context.TODO(), readCloser, true)
 252                 if err != nil {
 253                         return fmt.Errorf("While loading container image into Docker: %v", err)
 254                 }
 255
 256                 defer response.Body.Close()
 257                 rbody, err := ioutil.ReadAll(response.Body)
 258                 if err != nil {
 259                         return fmt.Errorf("Reading response to image load: %v", err)
 260                 }
 261                 runner.CrunchLog.Printf("Docker response: %s", rbody)
 262         } else {
 263                 runner.CrunchLog.Print("Docker image is available")
 264         }
 265
 266         runner.ContainerConfig.Image = imageID
 267
 268         runner.Kc.ClearBlockCache()
 269
 270         return nil
 271 }
 272
 273 func (runner *ContainerRunner) ArvMountCmd(arvMountCmd []string, token string) (c *exec.Cmd, err error) {
 274         c = exec.Command("arv-mount", arvMountCmd...)
 275
 276         // Copy our environment, but override ARVADOS_API_TOKEN with
 277         // the container auth token.
 278         c.Env = nil
 279         for _, s := range os.Environ() {
 280                 if !strings.HasPrefix(s, "ARVADOS_API_TOKEN=") {
 281                         c.Env = append(c.Env, s)
 282                 }
 283         }
 284         c.Env = append(c.Env, "ARVADOS_API_TOKEN="+token)
 285
 286         w, err := runner.NewLogWriter("arv-mount")
 287         if err != nil {
 288                 return nil, err
 289         }
 290         runner.arvMountLog = NewThrottledLogger(w)
 291         c.Stdout = runner.arvMountLog
 292         c.Stderr = runner.arvMountLog
 293
 294         runner.CrunchLog.Printf("Running %v", c.Args)
 295
 296         err = c.Start()
 297         if err != nil {
 298                 return nil, err
 299         }
 300
 301         statReadme := make(chan bool)
 302         runner.ArvMountExit = make(chan error)
 303
 304         keepStatting := true
 305         go func() {
 306                 for keepStatting {
 307                         time.Sleep(100 * time.Millisecond)
 308                         _, err = os.Stat(fmt.Sprintf("%s/by_id/README", runner.ArvMountPoint))
 309                         if err == nil {
 310                                 keepStatting = false
 311                                 statReadme <- true
 312                         }
 313                 }
 314                 close(statReadme)
 315         }()
 316
 317         go func() {
 318                 mnterr := c.Wait()
 319                 if mnterr != nil {
 320                         runner.CrunchLog.Printf("Arv-mount exit error: %v", mnterr)
 321                 }
 322                 runner.ArvMountExit <- mnterr
 323                 close(runner.ArvMountExit)
 324         }()
 325
 326         select {
 327         case <-statReadme:
 328                 break
 329         case err := <-runner.ArvMountExit:
 330                 runner.ArvMount = nil
 331                 keepStatting = false
 332                 return nil, err
 333         }
 334
 335         return c, nil
 336 }
 337
 338 func (runner *ContainerRunner) SetupArvMountPoint(prefix string) (err error) {
 339         if runner.ArvMountPoint == "" {
 340                 runner.ArvMountPoint, err = runner.MkTempDir(runner.parentTemp, prefix)
 341         }
 342         return
 343 }
 344
 345 func copyfile(src string, dst string) (err error) {
 346         srcfile, err := os.Open(src)
 347         if err != nil {
 348                 return
 349         }
 350
 351         os.MkdirAll(path.Dir(dst), 0777)
 352
 353         dstfile, err := os.Create(dst)
 354         if err != nil {
 355                 return
 356         }
 357         _, err = io.Copy(dstfile, srcfile)
 358         if err != nil {
 359                 return
 360         }
 361
 362         err = srcfile.Close()
 363         err2 := dstfile.Close()
 364
 365         if err != nil {
 366                 return
 367         }
 368
 369         if err2 != nil {
 370                 return err2
 371         }
 372
 373         return nil
 374 }
 375
 376 func (runner *ContainerRunner) SetupMounts() (err error) {
 377         err = runner.SetupArvMountPoint("keep")
 378         if err != nil {
 379                 return fmt.Errorf("While creating keep mount temp dir: %v", err)
 380         }
 381
 382         token, err := runner.ContainerToken()
 383         if err != nil {
 384                 return fmt.Errorf("could not get container token: %s", err)
 385         }
 386
 387         pdhOnly := true
 388         tmpcount := 0
 389         arvMountCmd := []string{
 390                 "--foreground",
 391                 "--allow-other",
 392                 "--read-write",
 393                 fmt.Sprintf("--crunchstat-interval=%v", runner.statInterval.Seconds())}
 394
 395         if runner.Container.RuntimeConstraints.KeepCacheRAM > 0 {
 396                 arvMountCmd = append(arvMountCmd, "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheRAM))
 397         }
 398
 399         collectionPaths := []string{}
 400         runner.Binds = nil
 401         runner.Volumes = make(map[string]struct{})
 402         needCertMount := true
 403         type copyFile struct {
 404                 src  string
 405                 bind string
 406         }
 407         var copyFiles []copyFile
 408
 409         var binds []string
 410         for bind := range runner.Container.Mounts {
 411                 binds = append(binds, bind)
 412         }
 413         for bind := range runner.SecretMounts {
 414                 if _, ok := runner.Container.Mounts[bind]; ok {
 415                         return fmt.Errorf("Secret mount %q conflicts with regular mount", bind)
 416                 }
 417                 if runner.SecretMounts[bind].Kind != "json" &&
 418                         runner.SecretMounts[bind].Kind != "text" {
 419                         return fmt.Errorf("Secret mount %q type is %q but only 'json' and 'text' are permitted.",
 420                                 bind, runner.SecretMounts[bind].Kind)
 421                 }
 422                 binds = append(binds, bind)
 423         }
 424         sort.Strings(binds)
 425
 426         for _, bind := range binds {
 427                 mnt, ok := runner.Container.Mounts[bind]
 428                 if !ok {
 429                         mnt = runner.SecretMounts[bind]
 430                 }
 431                 if bind == "stdout" || bind == "stderr" {
 432                         // Is it a "file" mount kind?
 433                         if mnt.Kind != "file" {
 434                                 return fmt.Errorf("Unsupported mount kind '%s' for %s. Only 'file' is supported.", mnt.Kind, bind)
 435                         }
 436
 437                         // Does path start with OutputPath?
 438                         prefix := runner.Container.OutputPath
 439                         if !strings.HasSuffix(prefix, "/") {
 440                                 prefix += "/"
 441                         }
 442                         if !strings.HasPrefix(mnt.Path, prefix) {
 443                                 return fmt.Errorf("%s path does not start with OutputPath: %s, %s", strings.Title(bind), mnt.Path, prefix)
 444                         }
 445                 }
 446
 447                 if bind == "stdin" {
 448                         // Is it a "collection" mount kind?
 449                         if mnt.Kind != "collection" && mnt.Kind != "json" {
 450                                 return fmt.Errorf("Unsupported mount kind '%s' for stdin. Only 'collection' or 'json' are supported.", mnt.Kind)
 451                         }
 452                 }
 453
 454                 if bind == "/etc/arvados/ca-certificates.crt" {
 455                         needCertMount = false
 456                 }
 457
 458                 if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && bind != runner.Container.OutputPath+"/" {
 459                         if mnt.Kind != "collection" && mnt.Kind != "text" && mnt.Kind != "json" {
 460                                 return fmt.Errorf("Only mount points of kind 'collection', 'text' or 'json' are supported underneath the output_path for %q, was %q", bind, mnt.Kind)
 461                         }
 462                 }
 463
 464                 switch {
 465                 case mnt.Kind == "collection" && bind != "stdin":
 466                         var src string
 467                         if mnt.UUID != "" && mnt.PortableDataHash != "" {
 468                                 return fmt.Errorf("Cannot specify both 'uuid' and 'portable_data_hash' for a collection mount")
 469                         }
 470                         if mnt.UUID != "" {
 471                                 if mnt.Writable {
 472                                         return fmt.Errorf("Writing to existing collections currently not permitted.")
 473                                 }
 474                                 pdhOnly = false
 475                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.UUID)
 476                         } else if mnt.PortableDataHash != "" {
 477                                 if mnt.Writable && !strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 478                                         return fmt.Errorf("Can never write to a collection specified by portable data hash")
 479                                 }
 480                                 idx := strings.Index(mnt.PortableDataHash, "/")
 481                                 if idx > 0 {
 482                                         mnt.Path = path.Clean(mnt.PortableDataHash[idx:])
 483                                         mnt.PortableDataHash = mnt.PortableDataHash[0:idx]
 484                                         runner.Container.Mounts[bind] = mnt
 485                                 }
 486                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.PortableDataHash)
 487                                 if mnt.Path != "" && mnt.Path != "." {
 488                                         if strings.HasPrefix(mnt.Path, "./") {
 489                                                 mnt.Path = mnt.Path[2:]
 490                                         } else if strings.HasPrefix(mnt.Path, "/") {
 491                                                 mnt.Path = mnt.Path[1:]
 492                                         }
 493                                         src += "/" + mnt.Path
 494                                 }
 495                         } else {
 496                                 src = fmt.Sprintf("%s/tmp%d", runner.ArvMountPoint, tmpcount)
 497                                 arvMountCmd = append(arvMountCmd, "--mount-tmp")
 498                                 arvMountCmd = append(arvMountCmd, fmt.Sprintf("tmp%d", tmpcount))
 499                                 tmpcount += 1
 500                         }
 501                         if mnt.Writable {
 502                                 if bind == runner.Container.OutputPath {
 503                                         runner.HostOutputDir = src
 504                                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", src, bind))
 505                                 } else if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 506                                         copyFiles = append(copyFiles, copyFile{src, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
 507                                 } else {
 508                                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", src, bind))
 509                                 }
 510                         } else {
 511                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", src, bind))
 512                         }
 513                         collectionPaths = append(collectionPaths, src)
 514
 515                 case mnt.Kind == "tmp":
 516                         var tmpdir string
 517                         tmpdir, err = runner.MkTempDir(runner.parentTemp, "tmp")
 518                         if err != nil {
 519                                 return fmt.Errorf("While creating mount temp dir: %v", err)
 520                         }
 521                         st, staterr := os.Stat(tmpdir)
 522                         if staterr != nil {
 523                                 return fmt.Errorf("While Stat on temp dir: %v", staterr)
 524                         }
 525                         err = os.Chmod(tmpdir, st.Mode()|os.ModeSetgid|0777)
 526                         if staterr != nil {
 527                                 return fmt.Errorf("While Chmod temp dir: %v", err)
 528                         }
 529                         runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", tmpdir, bind))
 530                         if bind == runner.Container.OutputPath {
 531                                 runner.HostOutputDir = tmpdir
 532                         }
 533
 534                 case mnt.Kind == "json" || mnt.Kind == "text":
 535                         var filedata []byte
 536                         if mnt.Kind == "json" {
 537                                 filedata, err = json.Marshal(mnt.Content)
 538                                 if err != nil {
 539                                         return fmt.Errorf("encoding json data: %v", err)
 540                                 }
 541                         } else {
 542                                 text, ok := mnt.Content.(string)
 543                                 if !ok {
 544                                         return fmt.Errorf("content for mount %q must be a string", bind)
 545                                 }
 546                                 filedata = []byte(text)
 547                         }
 548
 549                         tmpdir, err := runner.MkTempDir(runner.parentTemp, mnt.Kind)
 550                         if err != nil {
 551                                 return fmt.Errorf("creating temp dir: %v", err)
 552                         }
 553                         tmpfn := filepath.Join(tmpdir, "mountdata."+mnt.Kind)
 554                         err = ioutil.WriteFile(tmpfn, filedata, 0444)
 555                         if err != nil {
 556                                 return fmt.Errorf("writing temp file: %v", err)
 557                         }
 558                         if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 559                                 copyFiles = append(copyFiles, copyFile{tmpfn, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
 560                         } else {
 561                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", tmpfn, bind))
 562                         }
 563
 564                 case mnt.Kind == "git_tree":
 565                         tmpdir, err := runner.MkTempDir(runner.parentTemp, "git_tree")
 566                         if err != nil {
 567                                 return fmt.Errorf("creating temp dir: %v", err)
 568                         }
 569                         err = gitMount(mnt).extractTree(runner.ArvClient, tmpdir, token)
 570                         if err != nil {
 571                                 return err
 572                         }
 573                         runner.Binds = append(runner.Binds, tmpdir+":"+bind+":ro")
 574                 }
 575         }
 576
 577         if runner.HostOutputDir == "" {
 578                 return fmt.Errorf("Output path does not correspond to a writable mount point")
 579         }
 580
 581         if wantAPI := runner.Container.RuntimeConstraints.API; needCertMount && wantAPI != nil && *wantAPI {
 582                 for _, certfile := range arvadosclient.CertFiles {
 583                         _, err := os.Stat(certfile)
 584                         if err == nil {
 585                                 runner.Binds = append(runner.Binds, fmt.Sprintf("%s:/etc/arvados/ca-certificates.crt:ro", certfile))
 586                                 break
 587                         }
 588                 }
 589         }
 590
 591         if pdhOnly {
 592                 arvMountCmd = append(arvMountCmd, "--mount-by-pdh", "by_id")
 593         } else {
 594                 arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_id")
 595         }
 596         arvMountCmd = append(arvMountCmd, runner.ArvMountPoint)
 597
 598         runner.ArvMount, err = runner.RunArvMount(arvMountCmd, token)
 599         if err != nil {
 600                 return fmt.Errorf("While trying to start arv-mount: %v", err)
 601         }
 602
 603         for _, p := range collectionPaths {
 604                 _, err = os.Stat(p)
 605                 if err != nil {
 606                         return fmt.Errorf("While checking that input files exist: %v", err)
 607                 }
 608         }
 609
 610         for _, cp := range copyFiles {
 611                 st, err := os.Stat(cp.src)
 612                 if err != nil {
 613                         return fmt.Errorf("While staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 614                 }
 615                 if st.IsDir() {
 616                         err = filepath.Walk(cp.src, func(walkpath string, walkinfo os.FileInfo, walkerr error) error {
 617                                 if walkerr != nil {
 618                                         return walkerr
 619                                 }
 620                                 target := path.Join(cp.bind, walkpath[len(cp.src):])
 621                                 if walkinfo.Mode().IsRegular() {
 622                                         copyerr := copyfile(walkpath, target)
 623                                         if copyerr != nil {
 624                                                 return copyerr
 625                                         }
 626                                         return os.Chmod(target, walkinfo.Mode()|0777)
 627                                 } else if walkinfo.Mode().IsDir() {
 628                                         mkerr := os.MkdirAll(target, 0777)
 629                                         if mkerr != nil {
 630                                                 return mkerr
 631                                         }
 632                                         return os.Chmod(target, walkinfo.Mode()|os.ModeSetgid|0777)
 633                                 } else {
 634                                         return fmt.Errorf("Source %q is not a regular file or directory", cp.src)
 635                                 }
 636                         })
 637                 } else if st.Mode().IsRegular() {
 638                         err = copyfile(cp.src, cp.bind)
 639                         if err == nil {
 640                                 err = os.Chmod(cp.bind, st.Mode()|0777)
 641                         }
 642                 }
 643                 if err != nil {
 644                         return fmt.Errorf("While staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 645                 }
 646         }
 647
 648         return nil
 649 }
 650
 651 func (runner *ContainerRunner) ProcessDockerAttach(containerReader io.Reader) {
 652         // Handle docker log protocol
 653         // https://docs.docker.com/engine/reference/api/docker_remote_api_v1.15/#attach-to-a-container
 654         defer close(runner.loggingDone)
 655
 656         header := make([]byte, 8)
 657         var err error
 658         for err == nil {
 659                 _, err = io.ReadAtLeast(containerReader, header, 8)
 660                 if err != nil {
 661                         if err == io.EOF {
 662                                 err = nil
 663                         }
 664                         break
 665                 }
 666                 readsize := int64(header[7]) | (int64(header[6]) << 8) | (int64(header[5]) << 16) | (int64(header[4]) << 24)
 667                 if header[0] == 1 {
 668                         // stdout
 669                         _, err = io.CopyN(runner.Stdout, containerReader, readsize)
 670                 } else {
 671                         // stderr
 672                         _, err = io.CopyN(runner.Stderr, containerReader, readsize)
 673                 }
 674         }
 675
 676         if err != nil {
 677                 runner.CrunchLog.Printf("error reading docker logs: %v", err)
 678         }
 679
 680         err = runner.Stdout.Close()
 681         if err != nil {
 682                 runner.CrunchLog.Printf("error closing stdout logs: %v", err)
 683         }
 684
 685         err = runner.Stderr.Close()
 686         if err != nil {
 687                 runner.CrunchLog.Printf("error closing stderr logs: %v", err)
 688         }
 689
 690         if runner.statReporter != nil {
 691                 runner.statReporter.Stop()
 692                 err = runner.statLogger.Close()
 693                 if err != nil {
 694                         runner.CrunchLog.Printf("error closing crunchstat logs: %v", err)
 695                 }
 696         }
 697 }
 698
 699 func (runner *ContainerRunner) stopHoststat() error {
 700         if runner.hoststatReporter == nil {
 701                 return nil
 702         }
 703         runner.hoststatReporter.Stop()
 704         err := runner.hoststatLogger.Close()
 705         if err != nil {
 706                 return fmt.Errorf("error closing hoststat logs: %v", err)
 707         }
 708         return nil
 709 }
 710
 711 func (runner *ContainerRunner) startHoststat() error {
 712         w, err := runner.NewLogWriter("hoststat")
 713         if err != nil {
 714                 return err
 715         }
 716         runner.hoststatLogger = NewThrottledLogger(w)
 717         runner.hoststatReporter = &crunchstat.Reporter{
 718                 Logger:     log.New(runner.hoststatLogger, "", 0),
 719                 CgroupRoot: runner.cgroupRoot,
 720                 PollPeriod: runner.statInterval,
 721         }
 722         runner.hoststatReporter.Start()
 723         return nil
 724 }
 725
 726 func (runner *ContainerRunner) startCrunchstat() error {
 727         w, err := runner.NewLogWriter("crunchstat")
 728         if err != nil {
 729                 return err
 730         }
 731         runner.statLogger = NewThrottledLogger(w)
 732         runner.statReporter = &crunchstat.Reporter{
 733                 CID:          runner.ContainerID,
 734                 Logger:       log.New(runner.statLogger, "", 0),
 735                 CgroupParent: runner.expectCgroupParent,
 736                 CgroupRoot:   runner.cgroupRoot,
 737                 PollPeriod:   runner.statInterval,
 738         }
 739         runner.statReporter.Start()
 740         return nil
 741 }
 742
 743 type infoCommand struct {
 744         label string
 745         cmd   []string
 746 }
 747
 748 // LogHostInfo logs info about the current host, for debugging and
 749 // accounting purposes. Although it's logged as "node-info", this is
 750 // about the environment where crunch-run is actually running, which
 751 // might differ from what's described in the node record (see
 752 // LogNodeRecord).
 753 func (runner *ContainerRunner) LogHostInfo() (err error) {
 754         w, err := runner.NewLogWriter("node-info")
 755         if err != nil {
 756                 return
 757         }
 758
 759         commands := []infoCommand{
 760                 {
 761                         label: "Host Information",
 762                         cmd:   []string{"uname", "-a"},
 763                 },
 764                 {
 765                         label: "CPU Information",
 766                         cmd:   []string{"cat", "/proc/cpuinfo"},
 767                 },
 768                 {
 769                         label: "Memory Information",
 770                         cmd:   []string{"cat", "/proc/meminfo"},
 771                 },
 772                 {
 773                         label: "Disk Space",
 774                         cmd:   []string{"df", "-m", "/", os.TempDir()},
 775                 },
 776                 {
 777                         label: "Disk INodes",
 778                         cmd:   []string{"df", "-i", "/", os.TempDir()},
 779                 },
 780         }
 781
 782         // Run commands with informational output to be logged.
 783         for _, command := range commands {
 784                 fmt.Fprintln(w, command.label)
 785                 cmd := exec.Command(command.cmd[0], command.cmd[1:]...)
 786                 cmd.Stdout = w
 787                 cmd.Stderr = w
 788                 if err := cmd.Run(); err != nil {
 789                         err = fmt.Errorf("While running command %q: %v", command.cmd, err)
 790                         fmt.Fprintln(w, err)
 791                         return err
 792                 }
 793                 fmt.Fprintln(w, "")
 794         }
 795
 796         err = w.Close()
 797         if err != nil {
 798                 return fmt.Errorf("While closing node-info logs: %v", err)
 799         }
 800         return nil
 801 }
 802
 803 // LogContainerRecord gets and saves the raw JSON container record from the API server
 804 func (runner *ContainerRunner) LogContainerRecord() error {
 805         logged, err := runner.logAPIResponse("container", "containers", map[string]interface{}{"filters": [][]string{{"uuid", "=", runner.Container.UUID}}}, nil)
 806         if !logged && err == nil {
 807                 err = fmt.Errorf("error: no container record found for %s", runner.Container.UUID)
 808         }
 809         return err
 810 }
 811
 812 // LogNodeRecord logs arvados#node record corresponding to the current host.
 813 func (runner *ContainerRunner) LogNodeRecord() error {
 814         hostname := os.Getenv("SLURMD_NODENAME")
 815         if hostname == "" {
 816                 hostname, _ = os.Hostname()
 817         }
 818         _, err := runner.logAPIResponse("node", "nodes", map[string]interface{}{"filters": [][]string{{"hostname", "=", hostname}}}, func(resp interface{}) {
 819                 // The "info" field has admin-only info when obtained
 820                 // with a privileged token, and should not be logged.
 821                 node, ok := resp.(map[string]interface{})
 822                 if ok {
 823                         delete(node, "info")
 824                 }
 825         })
 826         return err
 827 }
 828
 829 func (runner *ContainerRunner) logAPIResponse(label, path string, params map[string]interface{}, munge func(interface{})) (logged bool, err error) {
 830         writer, err := runner.LogCollection.OpenFile(label+".json", os.O_CREATE|os.O_WRONLY, 0666)
 831         if err != nil {
 832                 return false, err
 833         }
 834         w := &ArvLogWriter{
 835                 ArvClient:     runner.ArvClient,
 836                 UUID:          runner.Container.UUID,
 837                 loggingStream: label,
 838                 writeCloser:   writer,
 839         }
 840
 841         reader, err := runner.ArvClient.CallRaw("GET", path, "", "", arvadosclient.Dict(params))
 842         if err != nil {
 843                 return false, fmt.Errorf("error getting %s record: %v", label, err)
 844         }
 845         defer reader.Close()
 846
 847         dec := json.NewDecoder(reader)
 848         dec.UseNumber()
 849         var resp map[string]interface{}
 850         if err = dec.Decode(&resp); err != nil {
 851                 return false, fmt.Errorf("error decoding %s list response: %v", label, err)
 852         }
 853         items, ok := resp["items"].([]interface{})
 854         if !ok {
 855                 return false, fmt.Errorf("error decoding %s list response: no \"items\" key in API list response", label)
 856         } else if len(items) < 1 {
 857                 return false, nil
 858         }
 859         if munge != nil {
 860                 munge(items[0])
 861         }
 862         // Re-encode it using indentation to improve readability
 863         enc := json.NewEncoder(w)
 864         enc.SetIndent("", "    ")
 865         if err = enc.Encode(items[0]); err != nil {
 866                 return false, fmt.Errorf("error logging %s record: %v", label, err)
 867         }
 868         err = w.Close()
 869         if err != nil {
 870                 return false, fmt.Errorf("error closing %s.json in log collection: %v", label, err)
 871         }
 872         return true, nil
 873 }
 874
 875 // AttachStreams connects the docker container stdin, stdout and stderr logs
 876 // to the Arvados logger which logs to Keep and the API server logs table.
 877 func (runner *ContainerRunner) AttachStreams() (err error) {
 878
 879         runner.CrunchLog.Print("Attaching container streams")
 880
 881         // If stdin mount is provided, attach it to the docker container
 882         var stdinRdr arvados.File
 883         var stdinJson []byte
 884         if stdinMnt, ok := runner.Container.Mounts["stdin"]; ok {
 885                 if stdinMnt.Kind == "collection" {
 886                         var stdinColl arvados.Collection
 887                         collId := stdinMnt.UUID
 888                         if collId == "" {
 889                                 collId = stdinMnt.PortableDataHash
 890                         }
 891                         err = runner.ArvClient.Get("collections", collId, nil, &stdinColl)
 892                         if err != nil {
 893                                 return fmt.Errorf("While getting stding collection: %v", err)
 894                         }
 895
 896                         stdinRdr, err = runner.Kc.ManifestFileReader(manifest.Manifest{Text: stdinColl.ManifestText}, stdinMnt.Path)
 897                         if os.IsNotExist(err) {
 898                                 return fmt.Errorf("stdin collection path not found: %v", stdinMnt.Path)
 899                         } else if err != nil {
 900                                 return fmt.Errorf("While getting stdin collection path %v: %v", stdinMnt.Path, err)
 901                         }
 902                 } else if stdinMnt.Kind == "json" {
 903                         stdinJson, err = json.Marshal(stdinMnt.Content)
 904                         if err != nil {
 905                                 return fmt.Errorf("While encoding stdin json data: %v", err)
 906                         }
 907                 }
 908         }
 909
 910         stdinUsed := stdinRdr != nil || len(stdinJson) != 0
 911         response, err := runner.Docker.ContainerAttach(context.TODO(), runner.ContainerID,
 912                 dockertypes.ContainerAttachOptions{Stream: true, Stdin: stdinUsed, Stdout: true, Stderr: true})
 913         if err != nil {
 914                 return fmt.Errorf("While attaching container stdout/stderr streams: %v", err)
 915         }
 916
 917         runner.loggingDone = make(chan bool)
 918
 919         if stdoutMnt, ok := runner.Container.Mounts["stdout"]; ok {
 920                 stdoutFile, err := runner.getStdoutFile(stdoutMnt.Path)
 921                 if err != nil {
 922                         return err
 923                 }
 924                 runner.Stdout = stdoutFile
 925         } else if w, err := runner.NewLogWriter("stdout"); err != nil {
 926                 return err
 927         } else {
 928                 runner.Stdout = NewThrottledLogger(w)
 929         }
 930
 931         if stderrMnt, ok := runner.Container.Mounts["stderr"]; ok {
 932                 stderrFile, err := runner.getStdoutFile(stderrMnt.Path)
 933                 if err != nil {
 934                         return err
 935                 }
 936                 runner.Stderr = stderrFile
 937         } else if w, err := runner.NewLogWriter("stderr"); err != nil {
 938                 return err
 939         } else {
 940                 runner.Stderr = NewThrottledLogger(w)
 941         }
 942
 943         if stdinRdr != nil {
 944                 go func() {
 945                         _, err := io.Copy(response.Conn, stdinRdr)
 946                         if err != nil {
 947                                 runner.CrunchLog.Print("While writing stdin collection to docker container %q", err)
 948                                 runner.stop(nil)
 949                         }
 950                         stdinRdr.Close()
 951                         response.CloseWrite()
 952                 }()
 953         } else if len(stdinJson) != 0 {
 954                 go func() {
 955                         _, err := io.Copy(response.Conn, bytes.NewReader(stdinJson))
 956                         if err != nil {
 957                                 runner.CrunchLog.Print("While writing stdin json to docker container %q", err)
 958                                 runner.stop(nil)
 959                         }
 960                         response.CloseWrite()
 961                 }()
 962         }
 963
 964         go runner.ProcessDockerAttach(response.Reader)
 965
 966         return nil
 967 }
 968
 969 func (runner *ContainerRunner) getStdoutFile(mntPath string) (*os.File, error) {
 970         stdoutPath := mntPath[len(runner.Container.OutputPath):]
 971         index := strings.LastIndex(stdoutPath, "/")
 972         if index > 0 {
 973                 subdirs := stdoutPath[:index]
 974                 if subdirs != "" {
 975                         st, err := os.Stat(runner.HostOutputDir)
 976                         if err != nil {
 977                                 return nil, fmt.Errorf("While Stat on temp dir: %v", err)
 978                         }
 979                         stdoutPath := filepath.Join(runner.HostOutputDir, subdirs)
 980                         err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777)
 981                         if err != nil {
 982                                 return nil, fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err)
 983                         }
 984                 }
 985         }
 986         stdoutFile, err := os.Create(filepath.Join(runner.HostOutputDir, stdoutPath))
 987         if err != nil {
 988                 return nil, fmt.Errorf("While creating file %q: %v", stdoutPath, err)
 989         }
 990
 991         return stdoutFile, nil
 992 }
 993
 994 // CreateContainer creates the docker container.
 995 func (runner *ContainerRunner) CreateContainer() error {
 996         runner.CrunchLog.Print("Creating Docker container")
 997
 998         runner.ContainerConfig.Cmd = runner.Container.Command
 999         if runner.Container.Cwd != "." {
1000                 runner.ContainerConfig.WorkingDir = runner.Container.Cwd
1001         }
1002
1003         for k, v := range runner.Container.Environment {
1004                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env, k+"="+v)
1005         }
1006
1007         runner.ContainerConfig.Volumes = runner.Volumes
1008
1009         maxRAM := int64(runner.Container.RuntimeConstraints.RAM)
1010         runner.HostConfig = dockercontainer.HostConfig{
1011                 Binds: runner.Binds,
1012                 LogConfig: dockercontainer.LogConfig{
1013                         Type: "none",
1014                 },
1015                 Resources: dockercontainer.Resources{
1016                         CgroupParent: runner.setCgroupParent,
1017                         NanoCPUs:     int64(runner.Container.RuntimeConstraints.VCPUs) * 1000000000,
1018                         Memory:       maxRAM, // RAM
1019                         MemorySwap:   maxRAM, // RAM+swap
1020                         KernelMemory: maxRAM, // kernel portion
1021                 },
1022         }
1023
1024         if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
1025                 tok, err := runner.ContainerToken()
1026                 if err != nil {
1027                         return err
1028                 }
1029                 runner.ContainerConfig.Env = append(runner.ContainerConfig.Env,
1030                         "ARVADOS_API_TOKEN="+tok,
1031                         "ARVADOS_API_HOST="+os.Getenv("ARVADOS_API_HOST"),
1032                         "ARVADOS_API_HOST_INSECURE="+os.Getenv("ARVADOS_API_HOST_INSECURE"),
1033                 )
1034                 runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
1035         } else {
1036                 if runner.enableNetwork == "always" {
1037                         runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
1038                 } else {
1039                         runner.HostConfig.NetworkMode = dockercontainer.NetworkMode("none")
1040                 }
1041         }
1042
1043         _, stdinUsed := runner.Container.Mounts["stdin"]
1044         runner.ContainerConfig.OpenStdin = stdinUsed
1045         runner.ContainerConfig.StdinOnce = stdinUsed
1046         runner.ContainerConfig.AttachStdin = stdinUsed
1047         runner.ContainerConfig.AttachStdout = true
1048         runner.ContainerConfig.AttachStderr = true
1049
1050         createdBody, err := runner.Docker.ContainerCreate(context.TODO(), &runner.ContainerConfig, &runner.HostConfig, nil, runner.Container.UUID)
1051         if err != nil {
1052                 return fmt.Errorf("While creating container: %v", err)
1053         }
1054
1055         runner.ContainerID = createdBody.ID
1056
1057         return runner.AttachStreams()
1058 }
1059
1060 // StartContainer starts the docker container created by CreateContainer.
1061 func (runner *ContainerRunner) StartContainer() error {
1062         runner.CrunchLog.Printf("Starting Docker container id '%s'", runner.ContainerID)
1063         runner.cStateLock.Lock()
1064         defer runner.cStateLock.Unlock()
1065         if runner.cCancelled {
1066                 return ErrCancelled
1067         }
1068         err := runner.Docker.ContainerStart(context.TODO(), runner.ContainerID,
1069                 dockertypes.ContainerStartOptions{})
1070         if err != nil {
1071                 var advice string
1072                 if m, e := regexp.MatchString("(?ms).*(exec|System error).*(no such file or directory|file not found).*", err.Error()); m && e == nil {
1073                         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])
1074                 }
1075                 return fmt.Errorf("could not start container: %v%s", err, advice)
1076         }
1077         return nil
1078 }
1079
1080 // checkContainerd checks if "containerd" is present in the process list.
1081 func (runner *ContainerRunner) CheckContainerd() error {
1082         if !runner.checkContainerd {
1083                 return nil
1084         }
1085         p, _ := process.Processes()
1086         for _, i := range p {
1087                 e, _ := i.CmdlineSlice()
1088                 if len(e) > 0 {
1089                         if strings.Index(e[0], "containerd") > -1 {
1090                                 return nil
1091                         }
1092                 }
1093         }
1094
1095         // Not found
1096         runner.runBrokenNodeHook()
1097         runner.stop(nil)
1098         return fmt.Errorf("'containerd' not found in process list.")
1099 }
1100
1101 // WaitFinish waits for the container to terminate, capture the exit code, and
1102 // close the stdout/stderr logging.
1103 func (runner *ContainerRunner) WaitFinish() error {
1104         var runTimeExceeded <-chan time.Time
1105         runner.CrunchLog.Print("Waiting for container to finish")
1106
1107         waitOk, waitErr := runner.Docker.ContainerWait(context.TODO(), runner.ContainerID, dockercontainer.WaitConditionNotRunning)
1108         arvMountExit := runner.ArvMountExit
1109         if timeout := runner.Container.SchedulingParameters.MaxRunTime; timeout > 0 {
1110                 runTimeExceeded = time.After(time.Duration(timeout) * time.Second)
1111         }
1112
1113         containerdGone := make(chan error)
1114         defer func() {
1115                 close(containerdGone)
1116         }()
1117         go func() {
1118                 ticker := time.NewTicker(time.Duration(60 * time.Second))
1119                 defer ticker.Stop()
1120                 for {
1121                         select {
1122                         case <-ticker.C:
1123                                 if ck := runner.CheckContainerd(); ck != nil {
1124                                         containerdGone <- ck
1125                                         return
1126                                 }
1127                         case <-containerdGone:
1128                                 break
1129                         }
1130                 }
1131         }()
1132
1133         for {
1134                 select {
1135                 case waitBody := <-waitOk:
1136                         runner.CrunchLog.Printf("Container exited with code: %v", waitBody.StatusCode)
1137                         code := int(waitBody.StatusCode)
1138                         runner.ExitCode = &code
1139
1140                         // wait for stdout/stderr to complete
1141                         <-runner.loggingDone
1142                         return nil
1143
1144                 case err := <-waitErr:
1145                         return fmt.Errorf("container wait: %v", err)
1146
1147                 case <-arvMountExit:
1148                         runner.CrunchLog.Printf("arv-mount exited while container is still running.  Stopping container.")
1149                         runner.stop(nil)
1150                         // arvMountExit will always be ready now that
1151                         // it's closed, but that doesn't interest us.
1152                         arvMountExit = nil
1153
1154                 case <-runTimeExceeded:
1155                         runner.CrunchLog.Printf("maximum run time exceeded. Stopping container.")
1156                         runner.stop(nil)
1157                         runTimeExceeded = nil
1158
1159                 case err := <-containerdGone:
1160                         return err
1161                 }
1162         }
1163 }
1164
1165 // CaptureOutput saves data from the container's output directory if
1166 // needed, and updates the container output accordingly.
1167 func (runner *ContainerRunner) CaptureOutput() error {
1168         if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
1169                 // Output may have been set directly by the container, so
1170                 // refresh the container record to check.
1171                 err := runner.ArvClient.Get("containers", runner.Container.UUID,
1172                         nil, &runner.Container)
1173                 if err != nil {
1174                         return err
1175                 }
1176                 if runner.Container.Output != "" {
1177                         // Container output is already set.
1178                         runner.OutputPDH = &runner.Container.Output
1179                         return nil
1180                 }
1181         }
1182
1183         txt, err := (&copier{
1184                 client:        runner.client,
1185                 arvClient:     runner.ArvClient,
1186                 keepClient:    runner.Kc,
1187                 hostOutputDir: runner.HostOutputDir,
1188                 ctrOutputDir:  runner.Container.OutputPath,
1189                 binds:         runner.Binds,
1190                 mounts:        runner.Container.Mounts,
1191                 secretMounts:  runner.SecretMounts,
1192                 logger:        runner.CrunchLog,
1193         }).Copy()
1194         if err != nil {
1195                 return err
1196         }
1197         var resp arvados.Collection
1198         err = runner.ArvClient.Create("collections", arvadosclient.Dict{
1199                 "ensure_unique_name": true,
1200                 "collection": arvadosclient.Dict{
1201                         "is_trashed":    true,
1202                         "name":          "output for " + runner.Container.UUID,
1203                         "manifest_text": txt,
1204                 },
1205         }, &resp)
1206         if err != nil {
1207                 return fmt.Errorf("error creating output collection: %v", err)
1208         }
1209         runner.OutputPDH = &resp.PortableDataHash
1210         return nil
1211 }
1212
1213 func (runner *ContainerRunner) CleanupDirs() {
1214         if runner.ArvMount != nil {
1215                 var delay int64 = 8
1216                 umount := exec.Command("arv-mount", fmt.Sprintf("--unmount-timeout=%d", delay), "--unmount", runner.ArvMountPoint)
1217                 umount.Stdout = runner.CrunchLog
1218                 umount.Stderr = runner.CrunchLog
1219                 runner.CrunchLog.Printf("Running %v", umount.Args)
1220                 umnterr := umount.Start()
1221
1222                 if umnterr != nil {
1223                         runner.CrunchLog.Printf("Error unmounting: %v", umnterr)
1224                 } else {
1225                         // If arv-mount --unmount gets stuck for any reason, we
1226                         // don't want to wait for it forever.  Do Wait() in a goroutine
1227                         // so it doesn't block crunch-run.
1228                         umountExit := make(chan error)
1229                         go func() {
1230                                 mnterr := umount.Wait()
1231                                 if mnterr != nil {
1232                                         runner.CrunchLog.Printf("Error unmounting: %v", mnterr)
1233                                 }
1234                                 umountExit <- mnterr
1235                         }()
1236
1237                         for again := true; again; {
1238                                 again = false
1239                                 select {
1240                                 case <-umountExit:
1241                                         umount = nil
1242                                         again = true
1243                                 case <-runner.ArvMountExit:
1244                                         break
1245                                 case <-time.After(time.Duration((delay + 1) * int64(time.Second))):
1246                                         runner.CrunchLog.Printf("Timed out waiting for unmount")
1247                                         if umount != nil {
1248                                                 umount.Process.Kill()
1249                                         }
1250                                         runner.ArvMount.Process.Kill()
1251                                 }
1252                         }
1253                 }
1254         }
1255
1256         if runner.ArvMountPoint != "" {
1257                 if rmerr := os.Remove(runner.ArvMountPoint); rmerr != nil {
1258                         runner.CrunchLog.Printf("While cleaning up arv-mount directory %s: %v", runner.ArvMountPoint, rmerr)
1259                 }
1260         }
1261
1262         if rmerr := os.RemoveAll(runner.parentTemp); rmerr != nil {
1263                 runner.CrunchLog.Printf("While cleaning up temporary directory %s: %v", runner.parentTemp, rmerr)
1264         }
1265 }
1266
1267 // CommitLogs posts the collection containing the final container logs.
1268 func (runner *ContainerRunner) CommitLogs() error {
1269         func() {
1270                 // Hold cStateLock to prevent races on CrunchLog (e.g., stop()).
1271                 runner.cStateLock.Lock()
1272                 defer runner.cStateLock.Unlock()
1273
1274                 runner.CrunchLog.Print(runner.finalState)
1275
1276                 if runner.arvMountLog != nil {
1277                         runner.arvMountLog.Close()
1278                 }
1279                 runner.CrunchLog.Close()
1280
1281                 // Closing CrunchLog above allows them to be committed to Keep at this
1282                 // point, but re-open crunch log with ArvClient in case there are any
1283                 // other further errors (such as failing to write the log to Keep!)
1284                 // while shutting down
1285                 runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{
1286                         ArvClient:     runner.ArvClient,
1287                         UUID:          runner.Container.UUID,
1288                         loggingStream: "crunch-run",
1289                         writeCloser:   nil,
1290                 })
1291                 runner.CrunchLog.Immediate = log.New(os.Stderr, runner.Container.UUID+" ", 0)
1292         }()
1293
1294         if runner.LogsPDH != nil {
1295                 // If we have already assigned something to LogsPDH,
1296                 // we must be closing the re-opened log, which won't
1297                 // end up getting attached to the container record and
1298                 // therefore doesn't need to be saved as a collection
1299                 // -- it exists only to send logs to other channels.
1300                 return nil
1301         }
1302
1303         mt, err := runner.LogCollection.MarshalManifest(".")
1304         if err != nil {
1305                 return fmt.Errorf("While creating log manifest: %v", err)
1306         }
1307
1308         var response arvados.Collection
1309         err = runner.ArvClient.Create("collections",
1310                 arvadosclient.Dict{
1311                         "ensure_unique_name": true,
1312                         "collection": arvadosclient.Dict{
1313                                 "is_trashed":    true,
1314                                 "name":          "logs for " + runner.Container.UUID,
1315                                 "manifest_text": mt}},
1316                 &response)
1317         if err != nil {
1318                 return fmt.Errorf("While creating log collection: %v", err)
1319         }
1320         runner.LogsPDH = &response.PortableDataHash
1321         return nil
1322 }
1323
1324 // UpdateContainerRunning updates the container state to "Running"
1325 func (runner *ContainerRunner) UpdateContainerRunning() error {
1326         runner.cStateLock.Lock()
1327         defer runner.cStateLock.Unlock()
1328         if runner.cCancelled {
1329                 return ErrCancelled
1330         }
1331         return runner.ArvClient.Update("containers", runner.Container.UUID,
1332                 arvadosclient.Dict{"container": arvadosclient.Dict{"state": "Running"}}, nil)
1333 }
1334
1335 // ContainerToken returns the api_token the container (and any
1336 // arv-mount processes) are allowed to use.
1337 func (runner *ContainerRunner) ContainerToken() (string, error) {
1338         if runner.token != "" {
1339                 return runner.token, nil
1340         }
1341
1342         var auth arvados.APIClientAuthorization
1343         err := runner.ArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth)
1344         if err != nil {
1345                 return "", err
1346         }
1347         runner.token = auth.APIToken
1348         return runner.token, nil
1349 }
1350
1351 // UpdateContainerComplete updates the container record state on API
1352 // server to "Complete" or "Cancelled"
1353 func (runner *ContainerRunner) UpdateContainerFinal() error {
1354         update := arvadosclient.Dict{}
1355         update["state"] = runner.finalState
1356         if runner.LogsPDH != nil {
1357                 update["log"] = *runner.LogsPDH
1358         }
1359         if runner.finalState == "Complete" {
1360                 if runner.ExitCode != nil {
1361                         update["exit_code"] = *runner.ExitCode
1362                 }
1363                 if runner.OutputPDH != nil {
1364                         update["output"] = *runner.OutputPDH
1365                 }
1366         }
1367         return runner.ArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{"container": update}, nil)
1368 }
1369
1370 // IsCancelled returns the value of Cancelled, with goroutine safety.
1371 func (runner *ContainerRunner) IsCancelled() bool {
1372         runner.cStateLock.Lock()
1373         defer runner.cStateLock.Unlock()
1374         return runner.cCancelled
1375 }
1376
1377 // NewArvLogWriter creates an ArvLogWriter
1378 func (runner *ContainerRunner) NewArvLogWriter(name string) (io.WriteCloser, error) {
1379         writer, err := runner.LogCollection.OpenFile(name+".txt", os.O_CREATE|os.O_WRONLY, 0666)
1380         if err != nil {
1381                 return nil, err
1382         }
1383         return &ArvLogWriter{
1384                 ArvClient:     runner.ArvClient,
1385                 UUID:          runner.Container.UUID,
1386                 loggingStream: name,
1387                 writeCloser:   writer,
1388         }, nil
1389 }
1390
1391 // Run the full container lifecycle.
1392 func (runner *ContainerRunner) Run() (err error) {
1393         runner.CrunchLog.Printf("crunch-run %s started", version)
1394         runner.CrunchLog.Printf("Executing container '%s'", runner.Container.UUID)
1395
1396         hostname, hosterr := os.Hostname()
1397         if hosterr != nil {
1398                 runner.CrunchLog.Printf("Error getting hostname '%v'", hosterr)
1399         } else {
1400                 runner.CrunchLog.Printf("Executing on host '%s'", hostname)
1401         }
1402
1403         runner.finalState = "Queued"
1404
1405         defer func() {
1406                 runner.CleanupDirs()
1407
1408                 runner.CrunchLog.Printf("crunch-run finished")
1409                 runner.CrunchLog.Close()
1410         }()
1411
1412         defer func() {
1413                 // checkErr prints e (unless it's nil) and sets err to
1414                 // e (unless err is already non-nil). Thus, if err
1415                 // hasn't already been assigned when Run() returns,
1416                 // this cleanup func will cause Run() to return the
1417                 // first non-nil error that is passed to checkErr().
1418                 checkErr := func(e error) {
1419                         if e == nil {
1420                                 return
1421                         }
1422                         runner.CrunchLog.Print(e)
1423                         if err == nil {
1424                                 err = e
1425                         }
1426                         if runner.finalState == "Complete" {
1427                                 // There was an error in the finalization.
1428                                 runner.finalState = "Cancelled"
1429                         }
1430                 }
1431
1432                 // Log the error encountered in Run(), if any
1433                 checkErr(err)
1434
1435                 if runner.finalState == "Queued" {
1436                         runner.UpdateContainerFinal()
1437                         return
1438                 }
1439
1440                 if runner.IsCancelled() {
1441                         runner.finalState = "Cancelled"
1442                         // but don't return yet -- we still want to
1443                         // capture partial output and write logs
1444                 }
1445
1446                 checkErr(runner.CaptureOutput())
1447                 checkErr(runner.stopHoststat())
1448                 checkErr(runner.CommitLogs())
1449                 checkErr(runner.UpdateContainerFinal())
1450         }()
1451
1452         err = runner.fetchContainerRecord()
1453         if err != nil {
1454                 return
1455         }
1456         runner.setupSignals()
1457         err = runner.startHoststat()
1458         if err != nil {
1459                 return
1460         }
1461
1462         // Sanity check that containerd is running.
1463         err = runner.CheckContainerd()
1464         if err != nil {
1465                 return
1466         }
1467
1468         // check for and/or load image
1469         err = runner.LoadImage()
1470         if err != nil {
1471                 if !runner.checkBrokenNode(err) {
1472                         // Failed to load image but not due to a "broken node"
1473                         // condition, probably user error.
1474                         runner.finalState = "Cancelled"
1475                 }
1476                 err = fmt.Errorf("While loading container image: %v", err)
1477                 return
1478         }
1479
1480         // set up FUSE mount and binds
1481         err = runner.SetupMounts()
1482         if err != nil {
1483                 runner.finalState = "Cancelled"
1484                 err = fmt.Errorf("While setting up mounts: %v", err)
1485                 return
1486         }
1487
1488         err = runner.CreateContainer()
1489         if err != nil {
1490                 return
1491         }
1492         err = runner.LogHostInfo()
1493         if err != nil {
1494                 return
1495         }
1496         err = runner.LogNodeRecord()
1497         if err != nil {
1498                 return
1499         }
1500         err = runner.LogContainerRecord()
1501         if err != nil {
1502                 return
1503         }
1504
1505         if runner.IsCancelled() {
1506                 return
1507         }
1508
1509         err = runner.UpdateContainerRunning()
1510         if err != nil {
1511                 return
1512         }
1513         runner.finalState = "Cancelled"
1514
1515         err = runner.startCrunchstat()
1516         if err != nil {
1517                 return
1518         }
1519
1520         err = runner.StartContainer()
1521         if err != nil {
1522                 runner.checkBrokenNode(err)
1523                 return
1524         }
1525
1526         err = runner.WaitFinish()
1527         if err == nil && !runner.IsCancelled() {
1528                 runner.finalState = "Complete"
1529         }
1530         return
1531 }
1532
1533 // Fetch the current container record (uuid = runner.Container.UUID)
1534 // into runner.Container.
1535 func (runner *ContainerRunner) fetchContainerRecord() error {
1536         reader, err := runner.ArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
1537         if err != nil {
1538                 return fmt.Errorf("error fetching container record: %v", err)
1539         }
1540         defer reader.Close()
1541
1542         dec := json.NewDecoder(reader)
1543         dec.UseNumber()
1544         err = dec.Decode(&runner.Container)
1545         if err != nil {
1546                 return fmt.Errorf("error decoding container record: %v", err)
1547         }
1548
1549         var sm struct {
1550                 SecretMounts map[string]arvados.Mount `json:"secret_mounts"`
1551         }
1552
1553         containerToken, err := runner.ContainerToken()
1554         if err != nil {
1555                 return fmt.Errorf("error getting container token: %v", err)
1556         }
1557
1558         containerClient, err := runner.MkArvClient(containerToken)
1559         if err != nil {
1560                 return fmt.Errorf("error creating container API client: %v", err)
1561         }
1562
1563         err = containerClient.Call("GET", "containers", runner.Container.UUID, "secret_mounts", nil, &sm)
1564         if err != nil {
1565                 if apierr, ok := err.(arvadosclient.APIServerError); !ok || apierr.HttpStatusCode != 404 {
1566                         return fmt.Errorf("error fetching secret_mounts: %v", err)
1567                 }
1568                 // ok && apierr.HttpStatusCode == 404, which means
1569                 // secret_mounts isn't supported by this API server.
1570         }
1571         runner.SecretMounts = sm.SecretMounts
1572
1573         return nil
1574 }
1575
1576 // NewContainerRunner creates a new container runner.
1577 func NewContainerRunner(client *arvados.Client, api IArvadosClient, kc IKeepClient, docker ThinDockerClient, containerUUID string) (*ContainerRunner, error) {
1578         cr := &ContainerRunner{
1579                 client:    client,
1580                 ArvClient: api,
1581                 Kc:        kc,
1582                 Docker:    docker,
1583         }
1584         cr.NewLogWriter = cr.NewArvLogWriter
1585         cr.RunArvMount = cr.ArvMountCmd
1586         cr.MkTempDir = ioutil.TempDir
1587         cr.MkArvClient = func(token string) (IArvadosClient, error) {
1588                 cl, err := arvadosclient.MakeArvadosClient()
1589                 if err != nil {
1590                         return nil, err
1591                 }
1592                 cl.ApiToken = token
1593                 return cl, nil
1594         }
1595         var err error
1596         cr.LogCollection, err = (&arvados.Collection{}).FileSystem(cr.client, cr.Kc)
1597         if err != nil {
1598                 return nil, err
1599         }
1600         cr.Container.UUID = containerUUID
1601         w, err := cr.NewLogWriter("crunch-run")
1602         if err != nil {
1603                 return nil, err
1604         }
1605         cr.CrunchLog = NewThrottledLogger(w)
1606         cr.CrunchLog.Immediate = log.New(os.Stderr, containerUUID+" ", 0)
1607
1608         loadLogThrottleParams(api)
1609
1610         return cr, nil
1611 }
1612
1613 func main() {
1614         statInterval := flag.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting")
1615         cgroupRoot := flag.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree")
1616         cgroupParent := flag.String("cgroup-parent", "docker", "name of container's parent cgroup (ignored if -cgroup-parent-subsystem is used)")
1617         cgroupParentSubsystem := flag.String("cgroup-parent-subsystem", "", "use current cgroup for given subsystem as parent cgroup for container")
1618         caCertsPath := flag.String("ca-certs", "", "Path to TLS root certificates")
1619         enableNetwork := flag.String("container-enable-networking", "default",
1620                 `Specify if networking should be enabled for container.  One of 'default', 'always':
1621         default: only enable networking if container requests it.
1622         always:  containers always have networking enabled
1623         `)
1624         networkMode := flag.String("container-network-mode", "default",
1625                 `Set networking mode for container.  Corresponds to Docker network mode (--net).
1626         `)
1627         memprofile := flag.String("memprofile", "", "write memory profile to `file` after running container")
1628         getVersion := flag.Bool("version", false, "Print version information and exit.")
1629         checkContainerd := flag.Bool("check-containerd", false, "Periodically check if (docker-)containerd is running, cancel if missing.")
1630         flag.Parse()
1631
1632         // Print version information if requested
1633         if *getVersion {
1634                 fmt.Printf("crunch-run %s\n", version)
1635                 return
1636         }
1637
1638         log.Printf("crunch-run %s started", version)
1639
1640         containerId := flag.Arg(0)
1641
1642         if *caCertsPath != "" {
1643                 arvadosclient.CertFiles = []string{*caCertsPath}
1644         }
1645
1646         api, err := arvadosclient.MakeArvadosClient()
1647         if err != nil {
1648                 log.Fatalf("%s: %v", containerId, err)
1649         }
1650         api.Retries = 8
1651
1652         kc, kcerr := keepclient.MakeKeepClient(api)
1653         if kcerr != nil {
1654                 log.Fatalf("%s: %v", containerId, kcerr)
1655         }
1656         kc.BlockCache = &keepclient.BlockCache{MaxBlocks: 2}
1657         kc.Retries = 4
1658
1659         // API version 1.21 corresponds to Docker 1.9, which is currently the
1660         // minimum version we want to support.
1661         docker, dockererr := dockerclient.NewClient(dockerclient.DefaultDockerHost, "1.21", nil, nil)
1662
1663         cr, err := NewContainerRunner(arvados.NewClientFromEnv(), api, kc, docker, containerId)
1664         if err != nil {
1665                 log.Fatal(err)
1666         }
1667         if dockererr != nil {
1668                 cr.CrunchLog.Printf("%s: %v", containerId, dockererr)
1669                 cr.checkBrokenNode(dockererr)
1670                 cr.CrunchLog.Close()
1671                 os.Exit(1)
1672         }
1673
1674         parentTemp, tmperr := cr.MkTempDir("", "crunch-run."+containerId+".")
1675         if tmperr != nil {
1676                 log.Fatalf("%s: %v", containerId, tmperr)
1677         }
1678
1679         cr.parentTemp = parentTemp
1680         cr.statInterval = *statInterval
1681         cr.cgroupRoot = *cgroupRoot
1682         cr.expectCgroupParent = *cgroupParent
1683         cr.enableNetwork = *enableNetwork
1684         cr.networkMode = *networkMode
1685         cr.checkContainerd = *checkContainerd
1686         if *cgroupParentSubsystem != "" {
1687                 p := findCgroup(*cgroupParentSubsystem)
1688                 cr.setCgroupParent = p
1689                 cr.expectCgroupParent = p
1690         }
1691
1692         runerr := cr.Run()
1693
1694         if *memprofile != "" {
1695                 f, err := os.Create(*memprofile)
1696                 if err != nil {
1697                         log.Printf("could not create memory profile: %s", err)
1698                 }
1699                 runtime.GC() // get up-to-date statistics
1700                 if err := pprof.WriteHeapProfile(f); err != nil {
1701                         log.Printf("could not write memory profile: %s", err)
1702                 }
1703                 closeerr := f.Close()
1704                 if closeerr != nil {
1705                         log.Printf("closing memprofile file: %s", err)
1706                 }
1707         }
1708
1709         if runerr != nil {
1710                 log.Fatalf("%s: %v", containerId, runerr)
1711         }
1712 }