1 // Copyright (C) The Arvados Authors. All rights reserved.
3 // SPDX-License-Identifier: AGPL-3.0
34 "git.arvados.org/arvados.git/lib/cloud"
35 "git.arvados.org/arvados.git/lib/cmd"
36 "git.arvados.org/arvados.git/lib/config"
37 "git.arvados.org/arvados.git/lib/crunchstat"
38 "git.arvados.org/arvados.git/sdk/go/arvados"
39 "git.arvados.org/arvados.git/sdk/go/arvadosclient"
40 "git.arvados.org/arvados.git/sdk/go/ctxlog"
41 "git.arvados.org/arvados.git/sdk/go/keepclient"
42 "git.arvados.org/arvados.git/sdk/go/manifest"
43 "golang.org/x/sys/unix"
48 var Command = command{}
50 // ConfigData contains environment variables and (when needed) cluster
51 // configuration, passed from dispatchcloud to crunch-run on stdin.
52 type ConfigData struct {
56 Cluster *arvados.Cluster
59 // IArvadosClient is the minimal Arvados API methods used by crunch-run.
60 type IArvadosClient interface {
61 Create(resourceType string, parameters arvadosclient.Dict, output interface{}) error
62 Get(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
63 Update(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
64 Call(method, resourceType, uuid, action string, parameters arvadosclient.Dict, output interface{}) error
65 CallRaw(method string, resourceType string, uuid string, action string, parameters arvadosclient.Dict) (reader io.ReadCloser, err error)
66 Discovery(key string) (interface{}, error)
69 // ErrCancelled is the error returned when the container is cancelled.
70 var ErrCancelled = errors.New("Cancelled")
72 // IKeepClient is the minimal Keep API methods used by crunch-run.
73 type IKeepClient interface {
74 BlockWrite(context.Context, arvados.BlockWriteOptions) (arvados.BlockWriteResponse, error)
75 ReadAt(locator string, p []byte, off int) (int, error)
76 ManifestFileReader(m manifest.Manifest, filename string) (arvados.File, error)
77 LocalLocator(locator string) (string, error)
79 SetStorageClasses(sc []string)
82 // NewLogWriter is a factory function to create a new log writer.
83 type NewLogWriter func(name string) (io.WriteCloser, error)
85 type RunArvMount func(cmdline []string, tok string) (*exec.Cmd, error)
87 type MkTempDir func(string, string) (string, error)
89 type PsProcess interface {
90 CmdlineSlice() ([]string, error)
93 // ContainerRunner is the main stateful struct used for a single execution of a
95 type ContainerRunner struct {
96 executor containerExecutor
97 executorStdin io.Closer
98 executorStdout io.Closer
99 executorStderr io.Closer
101 // Dispatcher client is initialized with the Dispatcher token.
102 // This is a privileged token used to manage container status
105 // We have both dispatcherClient and DispatcherArvClient
106 // because there are two different incompatible Arvados Go
107 // SDKs and we have to use both (hopefully this gets fixed in
109 dispatcherClient *arvados.Client
110 DispatcherArvClient IArvadosClient
111 DispatcherKeepClient IKeepClient
113 // Container client is initialized with the Container token
114 // This token controls the permissions of the container, and
115 // must be used for operations such as reading collections.
117 // Same comment as above applies to
118 // containerClient/ContainerArvClient.
119 containerClient *arvados.Client
120 ContainerArvClient IArvadosClient
121 ContainerKeepClient IKeepClient
123 Container arvados.Container
126 NewLogWriter NewLogWriter
127 CrunchLog *ThrottledLogger
130 LogCollection arvados.CollectionFileSystem
132 RunArvMount RunArvMount
137 Volumes map[string]struct{}
139 SigChan chan os.Signal
140 ArvMountExit chan error
141 SecretMounts map[string]arvados.Mount
142 MkArvClient func(token string) (IArvadosClient, IKeepClient, *arvados.Client, error)
145 costStartTime time.Time
148 keepstoreLogger io.WriteCloser
149 keepstoreLogbuf *bufThenWrite
150 statLogger io.WriteCloser
151 statReporter *crunchstat.Reporter
152 hoststatLogger io.WriteCloser
153 hoststatReporter *crunchstat.Reporter
154 statInterval time.Duration
156 // What we expect the container's cgroup parent to be.
157 expectCgroupParent string
158 // What we tell docker to use as the container's cgroup
159 // parent. Note: Ideally we would use the same field for both
160 // expectCgroupParent and setCgroupParent, and just make it
161 // default to "docker". However, when using docker < 1.10 with
162 // systemd, specifying a non-empty cgroup parent (even the
163 // default value "docker") hits a docker bug
164 // (https://github.com/docker/docker/issues/17126). Using two
165 // separate fields makes it possible to use the "expect cgroup
166 // parent to be X" feature even on sites where the "specify
167 // cgroup parent" feature breaks.
168 setCgroupParent string
170 cStateLock sync.Mutex
171 cCancelled bool // StopContainer() invoked
173 enableMemoryLimit bool
174 enableNetwork string // one of "default" or "always"
175 networkMode string // "none", "host", or "" -- passed through to executor
176 brokenNodeHook string // script to run if node appears to be broken
177 arvMountLog *ThrottledLogger
179 containerWatchdogInterval time.Duration
183 prices []cloud.InstancePrice
184 pricesLock sync.Mutex
187 // setupSignals sets up signal handling to gracefully terminate the
188 // underlying container and update state when receiving a TERM, INT or
190 func (runner *ContainerRunner) setupSignals() {
191 runner.SigChan = make(chan os.Signal, 1)
192 signal.Notify(runner.SigChan, syscall.SIGTERM)
193 signal.Notify(runner.SigChan, syscall.SIGINT)
194 signal.Notify(runner.SigChan, syscall.SIGQUIT)
196 go func(sig chan os.Signal) {
203 // stop the underlying container.
204 func (runner *ContainerRunner) stop(sig os.Signal) {
205 runner.cStateLock.Lock()
206 defer runner.cStateLock.Unlock()
208 runner.CrunchLog.Printf("caught signal: %v", sig)
210 runner.cCancelled = true
211 runner.CrunchLog.Printf("stopping container")
212 err := runner.executor.Stop()
214 runner.CrunchLog.Printf("error stopping container: %s", err)
218 var errorBlacklist = []string{
219 "(?ms).*[Cc]annot connect to the Docker daemon.*",
220 "(?ms).*oci runtime error.*starting container process.*container init.*mounting.*to rootfs.*no such file or directory.*",
221 "(?ms).*grpc: the connection is unavailable.*",
224 func (runner *ContainerRunner) runBrokenNodeHook() {
225 if runner.brokenNodeHook == "" {
226 path := filepath.Join(lockdir, brokenfile)
227 runner.CrunchLog.Printf("Writing %s to mark node as broken", path)
228 f, err := os.OpenFile(path, os.O_CREATE|os.O_RDWR, 0700)
230 runner.CrunchLog.Printf("Error writing %s: %s", path, err)
235 runner.CrunchLog.Printf("Running broken node hook %q", runner.brokenNodeHook)
237 c := exec.Command(runner.brokenNodeHook)
238 c.Stdout = runner.CrunchLog
239 c.Stderr = runner.CrunchLog
242 runner.CrunchLog.Printf("Error running broken node hook: %v", err)
247 func (runner *ContainerRunner) checkBrokenNode(goterr error) bool {
248 for _, d := range errorBlacklist {
249 if m, e := regexp.MatchString(d, goterr.Error()); m && e == nil {
250 runner.CrunchLog.Printf("Error suggests node is unable to run containers: %v", goterr)
251 runner.runBrokenNodeHook()
258 // LoadImage determines the docker image id from the container record and
259 // checks if it is available in the local Docker image store. If not, it loads
260 // the image from Keep.
261 func (runner *ContainerRunner) LoadImage() (string, error) {
262 runner.CrunchLog.Printf("Fetching Docker image from collection '%s'", runner.Container.ContainerImage)
264 d, err := os.Open(runner.ArvMountPoint + "/by_id/" + runner.Container.ContainerImage)
269 allfiles, err := d.Readdirnames(-1)
273 var tarfiles []string
274 for _, fnm := range allfiles {
275 if strings.HasSuffix(fnm, ".tar") {
276 tarfiles = append(tarfiles, fnm)
279 if len(tarfiles) == 0 {
280 return "", fmt.Errorf("image collection does not include a .tar image file")
282 if len(tarfiles) > 1 {
283 return "", fmt.Errorf("cannot choose from multiple tar files in image collection: %v", tarfiles)
285 imageID := tarfiles[0][:len(tarfiles[0])-4]
286 imageTarballPath := runner.ArvMountPoint + "/by_id/" + runner.Container.ContainerImage + "/" + imageID + ".tar"
287 runner.CrunchLog.Printf("Using Docker image id %q", imageID)
289 runner.CrunchLog.Print("Loading Docker image from keep")
290 err = runner.executor.LoadImage(imageID, imageTarballPath, runner.Container, runner.ArvMountPoint,
291 runner.containerClient)
299 func (runner *ContainerRunner) ArvMountCmd(cmdline []string, token string) (c *exec.Cmd, err error) {
300 c = exec.Command(cmdline[0], cmdline[1:]...)
302 // Copy our environment, but override ARVADOS_API_TOKEN with
303 // the container auth token.
305 for _, s := range os.Environ() {
306 if !strings.HasPrefix(s, "ARVADOS_API_TOKEN=") {
307 c.Env = append(c.Env, s)
310 c.Env = append(c.Env, "ARVADOS_API_TOKEN="+token)
312 w, err := runner.NewLogWriter("arv-mount")
316 runner.arvMountLog = NewThrottledLogger(w)
317 scanner := logScanner{
320 "Block not found error",
321 "Unhandled exception during FUSE operation",
323 ReportFunc: func(pattern, text string) {
324 runner.updateRuntimeStatus("arv-mount: "+pattern, text)
327 c.Stdout = runner.arvMountLog
328 c.Stderr = io.MultiWriter(runner.arvMountLog, os.Stderr, &scanner)
330 runner.CrunchLog.Printf("Running %v", c.Args)
337 statReadme := make(chan bool)
338 runner.ArvMountExit = make(chan error)
343 time.Sleep(100 * time.Millisecond)
344 _, err = os.Stat(fmt.Sprintf("%s/by_id/README", runner.ArvMountPoint))
356 runner.CrunchLog.Printf("Arv-mount exit error: %v", mnterr)
358 runner.ArvMountExit <- mnterr
359 close(runner.ArvMountExit)
365 case err := <-runner.ArvMountExit:
366 runner.ArvMount = nil
374 func (runner *ContainerRunner) SetupArvMountPoint(prefix string) (err error) {
375 if runner.ArvMountPoint == "" {
376 runner.ArvMountPoint, err = runner.MkTempDir(runner.parentTemp, prefix)
381 func copyfile(src string, dst string) (err error) {
382 srcfile, err := os.Open(src)
387 os.MkdirAll(path.Dir(dst), 0777)
389 dstfile, err := os.Create(dst)
393 _, err = io.Copy(dstfile, srcfile)
398 err = srcfile.Close()
399 err2 := dstfile.Close()
412 func (runner *ContainerRunner) SetupMounts() (map[string]bindmount, error) {
413 bindmounts := map[string]bindmount{}
414 err := runner.SetupArvMountPoint("keep")
416 return nil, fmt.Errorf("While creating keep mount temp dir: %v", err)
419 token, err := runner.ContainerToken()
421 return nil, fmt.Errorf("could not get container token: %s", err)
423 runner.CrunchLog.Printf("container token %q", token)
427 arvMountCmd := []string{
431 "--storage-classes", strings.Join(runner.Container.OutputStorageClasses, ","),
432 fmt.Sprintf("--crunchstat-interval=%v", runner.statInterval.Seconds())}
434 if _, isdocker := runner.executor.(*dockerExecutor); isdocker {
435 arvMountCmd = append(arvMountCmd, "--allow-other")
438 if runner.Container.RuntimeConstraints.KeepCacheDisk > 0 {
439 keepcachedir, err := runner.MkTempDir(runner.parentTemp, "keepcache")
441 return nil, fmt.Errorf("while creating keep cache temp dir: %v", err)
443 arvMountCmd = append(arvMountCmd, "--disk-cache", "--disk-cache-dir", keepcachedir, "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheDisk))
444 } else if runner.Container.RuntimeConstraints.KeepCacheRAM > 0 {
445 arvMountCmd = append(arvMountCmd, "--ram-cache", "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheRAM))
448 collectionPaths := []string{}
449 needCertMount := true
450 type copyFile struct {
454 var copyFiles []copyFile
457 for bind := range runner.Container.Mounts {
458 binds = append(binds, bind)
460 for bind := range runner.SecretMounts {
461 if _, ok := runner.Container.Mounts[bind]; ok {
462 return nil, fmt.Errorf("secret mount %q conflicts with regular mount", bind)
464 if runner.SecretMounts[bind].Kind != "json" &&
465 runner.SecretMounts[bind].Kind != "text" {
466 return nil, fmt.Errorf("secret mount %q type is %q but only 'json' and 'text' are permitted",
467 bind, runner.SecretMounts[bind].Kind)
469 binds = append(binds, bind)
473 for _, bind := range binds {
474 mnt, notSecret := runner.Container.Mounts[bind]
476 mnt = runner.SecretMounts[bind]
478 if bind == "stdout" || bind == "stderr" {
479 // Is it a "file" mount kind?
480 if mnt.Kind != "file" {
481 return nil, fmt.Errorf("unsupported mount kind '%s' for %s: only 'file' is supported", mnt.Kind, bind)
484 // Does path start with OutputPath?
485 prefix := runner.Container.OutputPath
486 if !strings.HasSuffix(prefix, "/") {
489 if !strings.HasPrefix(mnt.Path, prefix) {
490 return nil, fmt.Errorf("%s path does not start with OutputPath: %s, %s", strings.Title(bind), mnt.Path, prefix)
495 // Is it a "collection" mount kind?
496 if mnt.Kind != "collection" && mnt.Kind != "json" {
497 return nil, fmt.Errorf("unsupported mount kind '%s' for stdin: only 'collection' and 'json' are supported", mnt.Kind)
501 if bind == "/etc/arvados/ca-certificates.crt" {
502 needCertMount = false
505 if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && bind != runner.Container.OutputPath+"/" {
506 if mnt.Kind != "collection" && mnt.Kind != "text" && mnt.Kind != "json" {
507 return nil, fmt.Errorf("only mount points of kind 'collection', 'text' or 'json' are supported underneath the output_path for %q, was %q", bind, mnt.Kind)
512 case mnt.Kind == "collection" && bind != "stdin":
514 if mnt.UUID != "" && mnt.PortableDataHash != "" {
515 return nil, fmt.Errorf("cannot specify both 'uuid' and 'portable_data_hash' for a collection mount")
519 return nil, fmt.Errorf("writing to existing collections currently not permitted")
522 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.UUID)
523 } else if mnt.PortableDataHash != "" {
524 if mnt.Writable && !strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
525 return nil, fmt.Errorf("can never write to a collection specified by portable data hash")
527 idx := strings.Index(mnt.PortableDataHash, "/")
529 mnt.Path = path.Clean(mnt.PortableDataHash[idx:])
530 mnt.PortableDataHash = mnt.PortableDataHash[0:idx]
531 runner.Container.Mounts[bind] = mnt
533 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.PortableDataHash)
534 if mnt.Path != "" && mnt.Path != "." {
535 if strings.HasPrefix(mnt.Path, "./") {
536 mnt.Path = mnt.Path[2:]
537 } else if strings.HasPrefix(mnt.Path, "/") {
538 mnt.Path = mnt.Path[1:]
540 src += "/" + mnt.Path
543 src = fmt.Sprintf("%s/tmp%d", runner.ArvMountPoint, tmpcount)
544 arvMountCmd = append(arvMountCmd, "--mount-tmp", fmt.Sprintf("tmp%d", tmpcount))
548 if bind == runner.Container.OutputPath {
549 runner.HostOutputDir = src
550 bindmounts[bind] = bindmount{HostPath: src}
551 } else if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
552 copyFiles = append(copyFiles, copyFile{src, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
554 bindmounts[bind] = bindmount{HostPath: src}
557 bindmounts[bind] = bindmount{HostPath: src, ReadOnly: true}
559 collectionPaths = append(collectionPaths, src)
561 case mnt.Kind == "tmp":
563 tmpdir, err = runner.MkTempDir(runner.parentTemp, "tmp")
565 return nil, fmt.Errorf("while creating mount temp dir: %v", err)
567 st, staterr := os.Stat(tmpdir)
569 return nil, fmt.Errorf("while Stat on temp dir: %v", staterr)
571 err = os.Chmod(tmpdir, st.Mode()|os.ModeSetgid|0777)
573 return nil, fmt.Errorf("while Chmod temp dir: %v", err)
575 bindmounts[bind] = bindmount{HostPath: tmpdir}
576 if bind == runner.Container.OutputPath {
577 runner.HostOutputDir = tmpdir
580 case mnt.Kind == "json" || mnt.Kind == "text":
582 if mnt.Kind == "json" {
583 filedata, err = json.Marshal(mnt.Content)
585 return nil, fmt.Errorf("encoding json data: %v", err)
588 text, ok := mnt.Content.(string)
590 return nil, fmt.Errorf("content for mount %q must be a string", bind)
592 filedata = []byte(text)
595 tmpdir, err := runner.MkTempDir(runner.parentTemp, mnt.Kind)
597 return nil, fmt.Errorf("creating temp dir: %v", err)
599 tmpfn := filepath.Join(tmpdir, "mountdata."+mnt.Kind)
600 err = ioutil.WriteFile(tmpfn, filedata, 0444)
602 return nil, fmt.Errorf("writing temp file: %v", err)
604 if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && (notSecret || runner.Container.Mounts[runner.Container.OutputPath].Kind != "collection") {
605 // In most cases, if the container
606 // specifies a literal file inside the
607 // output path, we copy it into the
608 // output directory (either a mounted
609 // collection or a staging area on the
610 // host fs). If it's a secret, it will
611 // be skipped when copying output from
612 // staging to Keep later.
613 copyFiles = append(copyFiles, copyFile{tmpfn, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
615 // If a secret is outside OutputPath,
616 // we bind mount the secret file
617 // directly just like other mounts. We
618 // also use this strategy when a
619 // secret is inside OutputPath but
620 // OutputPath is a live collection, to
621 // avoid writing the secret to
622 // Keep. Attempting to remove a
623 // bind-mounted secret file from
624 // inside the container will return a
625 // "Device or resource busy" error
626 // that might not be handled well by
627 // the container, which is why we
628 // don't use this strategy when
629 // OutputPath is a staging directory.
630 bindmounts[bind] = bindmount{HostPath: tmpfn, ReadOnly: true}
633 case mnt.Kind == "git_tree":
634 tmpdir, err := runner.MkTempDir(runner.parentTemp, "git_tree")
636 return nil, fmt.Errorf("creating temp dir: %v", err)
638 err = gitMount(mnt).extractTree(runner.ContainerArvClient, tmpdir, token)
642 bindmounts[bind] = bindmount{HostPath: tmpdir, ReadOnly: true}
646 if runner.HostOutputDir == "" {
647 return nil, fmt.Errorf("output path does not correspond to a writable mount point")
650 if needCertMount && runner.Container.RuntimeConstraints.API {
651 for _, certfile := range arvadosclient.CertFiles {
652 _, err := os.Stat(certfile)
654 bindmounts["/etc/arvados/ca-certificates.crt"] = bindmount{HostPath: certfile, ReadOnly: true}
661 // If we are only mounting collections by pdh, make
662 // sure we don't subscribe to websocket events to
663 // avoid putting undesired load on the API server
664 arvMountCmd = append(arvMountCmd, "--mount-by-pdh", "by_id", "--disable-event-listening")
666 arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_id")
668 // the by_uuid mount point is used by singularity when writing
669 // out docker images converted to SIF
670 arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_uuid")
671 arvMountCmd = append(arvMountCmd, runner.ArvMountPoint)
673 runner.ArvMount, err = runner.RunArvMount(arvMountCmd, token)
675 return nil, fmt.Errorf("while trying to start arv-mount: %v", err)
677 if runner.hoststatReporter != nil && runner.ArvMount != nil {
678 runner.hoststatReporter.ReportPID("arv-mount", runner.ArvMount.Process.Pid)
681 for _, p := range collectionPaths {
684 return nil, fmt.Errorf("while checking that input files exist: %v", err)
688 for _, cp := range copyFiles {
689 st, err := os.Stat(cp.src)
691 return nil, fmt.Errorf("while staging writable file from %q to %q: %v", cp.src, cp.bind, err)
694 err = filepath.Walk(cp.src, func(walkpath string, walkinfo os.FileInfo, walkerr error) error {
698 target := path.Join(cp.bind, walkpath[len(cp.src):])
699 if walkinfo.Mode().IsRegular() {
700 copyerr := copyfile(walkpath, target)
704 return os.Chmod(target, walkinfo.Mode()|0777)
705 } else if walkinfo.Mode().IsDir() {
706 mkerr := os.MkdirAll(target, 0777)
710 return os.Chmod(target, walkinfo.Mode()|os.ModeSetgid|0777)
712 return fmt.Errorf("source %q is not a regular file or directory", cp.src)
715 } else if st.Mode().IsRegular() {
716 err = copyfile(cp.src, cp.bind)
718 err = os.Chmod(cp.bind, st.Mode()|0777)
722 return nil, fmt.Errorf("while staging writable file from %q to %q: %v", cp.src, cp.bind, err)
726 return bindmounts, nil
729 func (runner *ContainerRunner) stopHoststat() error {
730 if runner.hoststatReporter == nil {
733 runner.hoststatReporter.Stop()
734 err := runner.hoststatLogger.Close()
736 return fmt.Errorf("error closing hoststat logs: %v", err)
741 func (runner *ContainerRunner) startHoststat() error {
742 w, err := runner.NewLogWriter("hoststat")
746 runner.hoststatLogger = NewThrottledLogger(w)
747 runner.hoststatReporter = &crunchstat.Reporter{
748 Logger: log.New(runner.hoststatLogger, "", 0),
749 CgroupRoot: runner.cgroupRoot,
750 PollPeriod: runner.statInterval,
752 runner.hoststatReporter.Start()
753 runner.hoststatReporter.ReportPID("crunch-run", os.Getpid())
757 func (runner *ContainerRunner) startCrunchstat() error {
758 w, err := runner.NewLogWriter("crunchstat")
762 runner.statLogger = NewThrottledLogger(w)
763 runner.statReporter = &crunchstat.Reporter{
764 CID: runner.executor.CgroupID(),
765 Logger: log.New(runner.statLogger, "", 0),
766 CgroupParent: runner.expectCgroupParent,
767 CgroupRoot: runner.cgroupRoot,
768 PollPeriod: runner.statInterval,
769 TempDir: runner.parentTemp,
771 runner.statReporter.Start()
775 type infoCommand struct {
780 // LogHostInfo logs info about the current host, for debugging and
781 // accounting purposes. Although it's logged as "node-info", this is
782 // about the environment where crunch-run is actually running, which
783 // might differ from what's described in the node record (see
785 func (runner *ContainerRunner) LogHostInfo() (err error) {
786 w, err := runner.NewLogWriter("node-info")
791 commands := []infoCommand{
793 label: "Host Information",
794 cmd: []string{"uname", "-a"},
797 label: "CPU Information",
798 cmd: []string{"cat", "/proc/cpuinfo"},
801 label: "Memory Information",
802 cmd: []string{"cat", "/proc/meminfo"},
806 cmd: []string{"df", "-m", "/", os.TempDir()},
809 label: "Disk INodes",
810 cmd: []string{"df", "-i", "/", os.TempDir()},
814 // Run commands with informational output to be logged.
815 for _, command := range commands {
816 fmt.Fprintln(w, command.label)
817 cmd := exec.Command(command.cmd[0], command.cmd[1:]...)
820 if err := cmd.Run(); err != nil {
821 err = fmt.Errorf("While running command %q: %v", command.cmd, err)
830 return fmt.Errorf("While closing node-info logs: %v", err)
835 // LogContainerRecord gets and saves the raw JSON container record from the API server
836 func (runner *ContainerRunner) LogContainerRecord() error {
837 logged, err := runner.logAPIResponse("container", "containers", map[string]interface{}{"filters": [][]string{{"uuid", "=", runner.Container.UUID}}}, nil)
838 if !logged && err == nil {
839 err = fmt.Errorf("error: no container record found for %s", runner.Container.UUID)
844 // LogNodeRecord logs the current host's InstanceType config entry (or
845 // the arvados#node record, if running via crunch-dispatch-slurm).
846 func (runner *ContainerRunner) LogNodeRecord() error {
847 if it := os.Getenv("InstanceType"); it != "" {
848 // Dispatched via arvados-dispatch-cloud. Save
849 // InstanceType config fragment received from
850 // dispatcher on stdin.
851 w, err := runner.LogCollection.OpenFile("node.json", os.O_CREATE|os.O_WRONLY, 0666)
856 _, err = io.WriteString(w, it)
862 // Dispatched via crunch-dispatch-slurm. Look up
863 // apiserver's node record corresponding to
865 hostname := os.Getenv("SLURMD_NODENAME")
867 hostname, _ = os.Hostname()
869 _, err := runner.logAPIResponse("node", "nodes", map[string]interface{}{"filters": [][]string{{"hostname", "=", hostname}}}, func(resp interface{}) {
870 // The "info" field has admin-only info when
871 // obtained with a privileged token, and
872 // should not be logged.
873 node, ok := resp.(map[string]interface{})
881 func (runner *ContainerRunner) logAPIResponse(label, path string, params map[string]interface{}, munge func(interface{})) (logged bool, err error) {
882 writer, err := runner.LogCollection.OpenFile(label+".json", os.O_CREATE|os.O_WRONLY, 0666)
887 ArvClient: runner.DispatcherArvClient,
888 UUID: runner.Container.UUID,
889 loggingStream: label,
893 reader, err := runner.DispatcherArvClient.CallRaw("GET", path, "", "", arvadosclient.Dict(params))
895 return false, fmt.Errorf("error getting %s record: %v", label, err)
899 dec := json.NewDecoder(reader)
901 var resp map[string]interface{}
902 if err = dec.Decode(&resp); err != nil {
903 return false, fmt.Errorf("error decoding %s list response: %v", label, err)
905 items, ok := resp["items"].([]interface{})
907 return false, fmt.Errorf("error decoding %s list response: no \"items\" key in API list response", label)
908 } else if len(items) < 1 {
914 // Re-encode it using indentation to improve readability
915 enc := json.NewEncoder(w)
916 enc.SetIndent("", " ")
917 if err = enc.Encode(items[0]); err != nil {
918 return false, fmt.Errorf("error logging %s record: %v", label, err)
922 return false, fmt.Errorf("error closing %s.json in log collection: %v", label, err)
927 func (runner *ContainerRunner) getStdoutFile(mntPath string) (*os.File, error) {
928 stdoutPath := mntPath[len(runner.Container.OutputPath):]
929 index := strings.LastIndex(stdoutPath, "/")
931 subdirs := stdoutPath[:index]
933 st, err := os.Stat(runner.HostOutputDir)
935 return nil, fmt.Errorf("While Stat on temp dir: %v", err)
937 stdoutPath := filepath.Join(runner.HostOutputDir, subdirs)
938 err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777)
940 return nil, fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err)
944 stdoutFile, err := os.Create(filepath.Join(runner.HostOutputDir, stdoutPath))
946 return nil, fmt.Errorf("While creating file %q: %v", stdoutPath, err)
949 return stdoutFile, nil
952 // CreateContainer creates the docker container.
953 func (runner *ContainerRunner) CreateContainer(imageID string, bindmounts map[string]bindmount) error {
954 var stdin io.ReadCloser = ioutil.NopCloser(bytes.NewReader(nil))
955 if mnt, ok := runner.Container.Mounts["stdin"]; ok {
962 collID = mnt.PortableDataHash
964 path := runner.ArvMountPoint + "/by_id/" + collID + "/" + mnt.Path
965 f, err := os.Open(path)
971 j, err := json.Marshal(mnt.Content)
973 return fmt.Errorf("error encoding stdin json data: %v", err)
975 stdin = ioutil.NopCloser(bytes.NewReader(j))
977 return fmt.Errorf("stdin mount has unsupported kind %q", mnt.Kind)
981 var stdout, stderr io.WriteCloser
982 if mnt, ok := runner.Container.Mounts["stdout"]; ok {
983 f, err := runner.getStdoutFile(mnt.Path)
988 } else if w, err := runner.NewLogWriter("stdout"); err != nil {
991 stdout = NewThrottledLogger(w)
994 if mnt, ok := runner.Container.Mounts["stderr"]; ok {
995 f, err := runner.getStdoutFile(mnt.Path)
1000 } else if w, err := runner.NewLogWriter("stderr"); err != nil {
1003 stderr = NewThrottledLogger(w)
1006 env := runner.Container.Environment
1007 enableNetwork := runner.enableNetwork == "always"
1008 if runner.Container.RuntimeConstraints.API {
1009 enableNetwork = true
1010 tok, err := runner.ContainerToken()
1014 env = map[string]string{}
1015 for k, v := range runner.Container.Environment {
1018 env["ARVADOS_API_TOKEN"] = tok
1019 env["ARVADOS_API_HOST"] = os.Getenv("ARVADOS_API_HOST")
1020 env["ARVADOS_API_HOST_INSECURE"] = os.Getenv("ARVADOS_API_HOST_INSECURE")
1021 env["ARVADOS_KEEP_SERVICES"] = os.Getenv("ARVADOS_KEEP_SERVICES")
1023 workdir := runner.Container.Cwd
1025 // both "" and "." mean default
1028 ram := runner.Container.RuntimeConstraints.RAM
1029 if !runner.enableMemoryLimit {
1032 runner.executorStdin = stdin
1033 runner.executorStdout = stdout
1034 runner.executorStderr = stderr
1036 if runner.Container.RuntimeConstraints.CUDA.DeviceCount > 0 {
1037 nvidiaModprobe(runner.CrunchLog)
1040 return runner.executor.Create(containerSpec{
1042 VCPUs: runner.Container.RuntimeConstraints.VCPUs,
1044 WorkingDir: workdir,
1046 BindMounts: bindmounts,
1047 Command: runner.Container.Command,
1048 EnableNetwork: enableNetwork,
1049 CUDADeviceCount: runner.Container.RuntimeConstraints.CUDA.DeviceCount,
1050 NetworkMode: runner.networkMode,
1051 CgroupParent: runner.setCgroupParent,
1058 // StartContainer starts the docker container created by CreateContainer.
1059 func (runner *ContainerRunner) StartContainer() error {
1060 runner.CrunchLog.Printf("Starting container")
1061 runner.cStateLock.Lock()
1062 defer runner.cStateLock.Unlock()
1063 if runner.cCancelled {
1066 err := runner.executor.Start()
1069 if m, e := regexp.MatchString("(?ms).*(exec|System error).*(no such file or directory|file not found).*", err.Error()); m && e == nil {
1070 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])
1072 return fmt.Errorf("could not start container: %v%s", err, advice)
1077 // WaitFinish waits for the container to terminate, capture the exit code, and
1078 // close the stdout/stderr logging.
1079 func (runner *ContainerRunner) WaitFinish() error {
1080 runner.CrunchLog.Print("Waiting for container to finish")
1081 var timeout <-chan time.Time
1082 if s := runner.Container.SchedulingParameters.MaxRunTime; s > 0 {
1083 timeout = time.After(time.Duration(s) * time.Second)
1085 ctx, cancel := context.WithCancel(context.Background())
1090 runner.CrunchLog.Printf("maximum run time exceeded. Stopping container.")
1092 case <-runner.ArvMountExit:
1093 runner.CrunchLog.Printf("arv-mount exited while container is still running. Stopping container.")
1098 exitcode, err := runner.executor.Wait(ctx)
1100 runner.checkBrokenNode(err)
1103 runner.ExitCode = &exitcode
1106 if exitcode&0x80 != 0 {
1107 // Convert raw exit status (0x80 + signal number) to a
1108 // string to log after the code, like " (signal 101)"
1109 // or " (signal 9, killed)"
1110 sig := syscall.WaitStatus(exitcode).Signal()
1111 if name := unix.SignalName(sig); name != "" {
1112 extra = fmt.Sprintf(" (signal %d, %s)", sig, name)
1114 extra = fmt.Sprintf(" (signal %d)", sig)
1117 runner.CrunchLog.Printf("Container exited with status code %d%s", exitcode, extra)
1118 err = runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1119 "container": arvadosclient.Dict{"exit_code": exitcode},
1122 runner.CrunchLog.Printf("ignoring error updating exit_code: %s", err)
1126 if err = runner.executorStdin.Close(); err != nil {
1127 err = fmt.Errorf("error closing container stdin: %s", err)
1128 runner.CrunchLog.Printf("%s", err)
1131 if err = runner.executorStdout.Close(); err != nil {
1132 err = fmt.Errorf("error closing container stdout: %s", err)
1133 runner.CrunchLog.Printf("%s", err)
1134 if returnErr == nil {
1138 if err = runner.executorStderr.Close(); err != nil {
1139 err = fmt.Errorf("error closing container stderr: %s", err)
1140 runner.CrunchLog.Printf("%s", err)
1141 if returnErr == nil {
1146 if runner.statReporter != nil {
1147 runner.statReporter.Stop()
1148 err = runner.statLogger.Close()
1150 runner.CrunchLog.Printf("error closing crunchstat logs: %v", err)
1156 func (runner *ContainerRunner) updateLogs() {
1157 ticker := time.NewTicker(crunchLogUpdatePeriod / 360)
1160 sigusr1 := make(chan os.Signal, 1)
1161 signal.Notify(sigusr1, syscall.SIGUSR1)
1162 defer signal.Stop(sigusr1)
1164 saveAtTime := time.Now().Add(crunchLogUpdatePeriod)
1165 saveAtSize := crunchLogUpdateSize
1171 saveAtTime = time.Now()
1173 runner.logMtx.Lock()
1174 done := runner.LogsPDH != nil
1175 runner.logMtx.Unlock()
1179 size := runner.LogCollection.Size()
1180 if size == savedSize || (time.Now().Before(saveAtTime) && size < saveAtSize) {
1183 saveAtTime = time.Now().Add(crunchLogUpdatePeriod)
1184 saveAtSize = runner.LogCollection.Size() + crunchLogUpdateSize
1185 saved, err := runner.saveLogCollection(false)
1187 runner.CrunchLog.Printf("error updating log collection: %s", err)
1191 err = runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1192 "container": arvadosclient.Dict{"log": saved.PortableDataHash},
1195 runner.CrunchLog.Printf("error updating container log to %s: %s", saved.PortableDataHash, err)
1203 var spotInterruptionCheckInterval = 5 * time.Second
1204 var ec2MetadataBaseURL = "http://169.254.169.254"
1206 func (runner *ContainerRunner) checkSpotInterruptionNotices() {
1207 type ec2metadata struct {
1208 Action string `json:"action"`
1209 Time time.Time `json:"time"`
1211 runner.CrunchLog.Printf("Checking for spot interruptions every %v using instance metadata at %s", spotInterruptionCheckInterval, ec2MetadataBaseURL)
1212 var metadata ec2metadata
1213 check := func() error {
1214 ctx, cancel := context.WithDeadline(context.Background(), time.Now().Add(time.Minute))
1216 req, err := http.NewRequestWithContext(ctx, http.MethodPut, ec2MetadataBaseURL+"/latest/api/token", nil)
1220 req.Header.Set("X-aws-ec2-metadata-token-ttl-seconds", "21600")
1221 resp, err := http.DefaultClient.Do(req)
1225 defer resp.Body.Close()
1226 if resp.StatusCode != http.StatusOK {
1227 return fmt.Errorf("%s", resp.Status)
1229 token, err := ioutil.ReadAll(resp.Body)
1233 req, err = http.NewRequestWithContext(ctx, http.MethodGet, ec2MetadataBaseURL+"/latest/meta-data/spot/instance-action", nil)
1237 req.Header.Set("X-aws-ec2-metadata-token", strings.TrimSpace(string(token)))
1238 resp, err = http.DefaultClient.Do(req)
1242 defer resp.Body.Close()
1243 metadata = ec2metadata{}
1244 if resp.StatusCode == http.StatusNotFound {
1245 // "If Amazon EC2 is not preparing to stop or
1246 // terminate the instance, or if you
1247 // terminated the instance yourself,
1248 // instance-action is not present in the
1249 // instance metadata and you receive an HTTP
1250 // 404 error when you try to retrieve it."
1252 } else if resp.StatusCode != http.StatusOK {
1253 return fmt.Errorf("%s", resp.Status)
1255 err = json.NewDecoder(resp.Body).Decode(&metadata)
1262 var lastmetadata ec2metadata
1263 for range time.NewTicker(spotInterruptionCheckInterval).C {
1266 runner.CrunchLog.Printf("Error checking spot interruptions: %s", err)
1269 runner.CrunchLog.Printf("Giving up on checking spot interruptions after too many errors")
1274 if metadata != lastmetadata {
1275 lastmetadata = metadata
1276 text := fmt.Sprintf("Cloud provider indicates instance action %q scheduled for time %q", metadata.Action, metadata.Time.UTC().Format(time.RFC3339))
1277 runner.CrunchLog.Printf("%s", text)
1278 runner.updateRuntimeStatus("instance interruption", text)
1279 if proc, err := os.FindProcess(os.Getpid()); err == nil {
1280 // trigger updateLogs
1281 proc.Signal(syscall.SIGUSR1)
1287 func (runner *ContainerRunner) updateRuntimeStatus(warning, detail string) {
1288 err := runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1289 "container": arvadosclient.Dict{
1290 "runtime_status": arvadosclient.Dict{
1292 "warningDetail": detail,
1297 runner.CrunchLog.Printf("error updating container runtime_status: %s", err)
1301 // CaptureOutput saves data from the container's output directory if
1302 // needed, and updates the container output accordingly.
1303 func (runner *ContainerRunner) CaptureOutput(bindmounts map[string]bindmount) error {
1304 if runner.Container.RuntimeConstraints.API {
1305 // Output may have been set directly by the container, so
1306 // refresh the container record to check.
1307 err := runner.DispatcherArvClient.Get("containers", runner.Container.UUID,
1308 nil, &runner.Container)
1312 if runner.Container.Output != "" {
1313 // Container output is already set.
1314 runner.OutputPDH = &runner.Container.Output
1319 txt, err := (&copier{
1320 client: runner.containerClient,
1321 arvClient: runner.ContainerArvClient,
1322 keepClient: runner.ContainerKeepClient,
1323 hostOutputDir: runner.HostOutputDir,
1324 ctrOutputDir: runner.Container.OutputPath,
1325 bindmounts: bindmounts,
1326 mounts: runner.Container.Mounts,
1327 secretMounts: runner.SecretMounts,
1328 logger: runner.CrunchLog,
1333 if n := len(regexp.MustCompile(` [0-9a-f]+\+\S*\+R`).FindAllStringIndex(txt, -1)); n > 0 {
1334 runner.CrunchLog.Printf("Copying %d data blocks from remote input collections...", n)
1335 fs, err := (&arvados.Collection{ManifestText: txt}).FileSystem(runner.containerClient, runner.ContainerKeepClient)
1339 txt, err = fs.MarshalManifest(".")
1344 var resp arvados.Collection
1345 err = runner.ContainerArvClient.Create("collections", arvadosclient.Dict{
1346 "ensure_unique_name": true,
1347 "collection": arvadosclient.Dict{
1349 "name": "output for " + runner.Container.UUID,
1350 "manifest_text": txt,
1354 return fmt.Errorf("error creating output collection: %v", err)
1356 runner.OutputPDH = &resp.PortableDataHash
1360 func (runner *ContainerRunner) CleanupDirs() {
1361 if runner.ArvMount != nil {
1363 umount := exec.Command("arv-mount", fmt.Sprintf("--unmount-timeout=%d", delay), "--unmount", runner.ArvMountPoint)
1364 umount.Stdout = runner.CrunchLog
1365 umount.Stderr = runner.CrunchLog
1366 runner.CrunchLog.Printf("Running %v", umount.Args)
1367 umnterr := umount.Start()
1370 runner.CrunchLog.Printf("Error unmounting: %v", umnterr)
1371 runner.ArvMount.Process.Kill()
1373 // If arv-mount --unmount gets stuck for any reason, we
1374 // don't want to wait for it forever. Do Wait() in a goroutine
1375 // so it doesn't block crunch-run.
1376 umountExit := make(chan error)
1378 mnterr := umount.Wait()
1380 runner.CrunchLog.Printf("Error unmounting: %v", mnterr)
1382 umountExit <- mnterr
1385 for again := true; again; {
1391 case <-runner.ArvMountExit:
1393 case <-time.After(time.Duration((delay + 1) * int64(time.Second))):
1394 runner.CrunchLog.Printf("Timed out waiting for unmount")
1396 umount.Process.Kill()
1398 runner.ArvMount.Process.Kill()
1402 runner.ArvMount = nil
1405 if runner.ArvMountPoint != "" {
1406 if rmerr := os.Remove(runner.ArvMountPoint); rmerr != nil {
1407 runner.CrunchLog.Printf("While cleaning up arv-mount directory %s: %v", runner.ArvMountPoint, rmerr)
1409 runner.ArvMountPoint = ""
1412 if rmerr := os.RemoveAll(runner.parentTemp); rmerr != nil {
1413 runner.CrunchLog.Printf("While cleaning up temporary directory %s: %v", runner.parentTemp, rmerr)
1417 // CommitLogs posts the collection containing the final container logs.
1418 func (runner *ContainerRunner) CommitLogs() error {
1420 // Hold cStateLock to prevent races on CrunchLog (e.g., stop()).
1421 runner.cStateLock.Lock()
1422 defer runner.cStateLock.Unlock()
1424 runner.CrunchLog.Print(runner.finalState)
1426 if runner.arvMountLog != nil {
1427 runner.arvMountLog.Close()
1429 runner.CrunchLog.Close()
1431 // Closing CrunchLog above allows them to be committed to Keep at this
1432 // point, but re-open crunch log with ArvClient in case there are any
1433 // other further errors (such as failing to write the log to Keep!)
1434 // while shutting down
1435 runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{
1436 ArvClient: runner.DispatcherArvClient,
1437 UUID: runner.Container.UUID,
1438 loggingStream: "crunch-run",
1441 runner.CrunchLog.Immediate = log.New(os.Stderr, runner.Container.UUID+" ", 0)
1444 if runner.keepstoreLogger != nil {
1445 // Flush any buffered logs from our local keepstore
1446 // process. Discard anything logged after this point
1447 // -- it won't end up in the log collection, so
1448 // there's no point writing it to the collectionfs.
1449 runner.keepstoreLogbuf.SetWriter(io.Discard)
1450 runner.keepstoreLogger.Close()
1451 runner.keepstoreLogger = nil
1454 if runner.LogsPDH != nil {
1455 // If we have already assigned something to LogsPDH,
1456 // we must be closing the re-opened log, which won't
1457 // end up getting attached to the container record and
1458 // therefore doesn't need to be saved as a collection
1459 // -- it exists only to send logs to other channels.
1463 saved, err := runner.saveLogCollection(true)
1465 return fmt.Errorf("error saving log collection: %s", err)
1467 runner.logMtx.Lock()
1468 defer runner.logMtx.Unlock()
1469 runner.LogsPDH = &saved.PortableDataHash
1473 func (runner *ContainerRunner) saveLogCollection(final bool) (response arvados.Collection, err error) {
1474 runner.logMtx.Lock()
1475 defer runner.logMtx.Unlock()
1476 if runner.LogsPDH != nil {
1477 // Already finalized.
1480 updates := arvadosclient.Dict{
1481 "name": "logs for " + runner.Container.UUID,
1483 mt, err1 := runner.LogCollection.MarshalManifest(".")
1485 // Only send updated manifest text if there was no
1487 updates["manifest_text"] = mt
1490 // Even if flushing the manifest had an error, we still want
1491 // to update the log record, if possible, to push the trash_at
1492 // and delete_at times into the future. Details on bug
1495 updates["is_trashed"] = true
1497 exp := time.Now().Add(crunchLogUpdatePeriod * 24)
1498 updates["trash_at"] = exp
1499 updates["delete_at"] = exp
1501 reqBody := arvadosclient.Dict{"collection": updates}
1503 if runner.logUUID == "" {
1504 reqBody["ensure_unique_name"] = true
1505 err2 = runner.DispatcherArvClient.Create("collections", reqBody, &response)
1507 err2 = runner.DispatcherArvClient.Update("collections", runner.logUUID, reqBody, &response)
1510 runner.logUUID = response.UUID
1513 if err1 != nil || err2 != nil {
1514 err = fmt.Errorf("error recording logs: %q, %q", err1, err2)
1519 // UpdateContainerRunning updates the container state to "Running"
1520 func (runner *ContainerRunner) UpdateContainerRunning(logId string) error {
1521 runner.cStateLock.Lock()
1522 defer runner.cStateLock.Unlock()
1523 if runner.cCancelled {
1526 updates := arvadosclient.Dict{
1527 "gateway_address": runner.gateway.Address,
1531 updates["log"] = logId
1533 return runner.DispatcherArvClient.Update(
1535 runner.Container.UUID,
1536 arvadosclient.Dict{"container": updates},
1541 // ContainerToken returns the api_token the container (and any
1542 // arv-mount processes) are allowed to use.
1543 func (runner *ContainerRunner) ContainerToken() (string, error) {
1544 if runner.token != "" {
1545 return runner.token, nil
1548 var auth arvados.APIClientAuthorization
1549 err := runner.DispatcherArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth)
1553 runner.token = fmt.Sprintf("v2/%s/%s/%s", auth.UUID, auth.APIToken, runner.Container.UUID)
1554 return runner.token, nil
1557 // UpdateContainerFinal updates the container record state on API
1558 // server to "Complete" or "Cancelled"
1559 func (runner *ContainerRunner) UpdateContainerFinal() error {
1560 update := arvadosclient.Dict{}
1561 update["state"] = runner.finalState
1562 if runner.LogsPDH != nil {
1563 update["log"] = *runner.LogsPDH
1565 if runner.ExitCode != nil {
1566 update["exit_code"] = *runner.ExitCode
1568 update["exit_code"] = nil
1570 if runner.finalState == "Complete" && runner.OutputPDH != nil {
1571 update["output"] = *runner.OutputPDH
1573 update["cost"] = runner.calculateCost(time.Now())
1574 return runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{"container": update}, nil)
1577 // IsCancelled returns the value of Cancelled, with goroutine safety.
1578 func (runner *ContainerRunner) IsCancelled() bool {
1579 runner.cStateLock.Lock()
1580 defer runner.cStateLock.Unlock()
1581 return runner.cCancelled
1584 // NewArvLogWriter creates an ArvLogWriter
1585 func (runner *ContainerRunner) NewArvLogWriter(name string) (io.WriteCloser, error) {
1586 writer, err := runner.LogCollection.OpenFile(name+".txt", os.O_CREATE|os.O_WRONLY, 0666)
1590 return &ArvLogWriter{
1591 ArvClient: runner.DispatcherArvClient,
1592 UUID: runner.Container.UUID,
1593 loggingStream: name,
1594 writeCloser: writer,
1598 // Run the full container lifecycle.
1599 func (runner *ContainerRunner) Run() (err error) {
1600 runner.CrunchLog.Printf("crunch-run %s started", cmd.Version.String())
1601 runner.CrunchLog.Printf("%s", currentUserAndGroups())
1602 v, _ := exec.Command("arv-mount", "--version").CombinedOutput()
1603 runner.CrunchLog.Printf("Using FUSE mount: %s", v)
1604 runner.CrunchLog.Printf("Using container runtime: %s", runner.executor.Runtime())
1605 runner.CrunchLog.Printf("Executing container: %s", runner.Container.UUID)
1606 runner.costStartTime = time.Now()
1608 hostname, hosterr := os.Hostname()
1610 runner.CrunchLog.Printf("Error getting hostname '%v'", hosterr)
1612 runner.CrunchLog.Printf("Executing on host '%s'", hostname)
1615 sigusr2 := make(chan os.Signal, 1)
1616 signal.Notify(sigusr2, syscall.SIGUSR2)
1617 defer signal.Stop(sigusr2)
1625 runner.finalState = "Queued"
1628 runner.CleanupDirs()
1630 runner.CrunchLog.Printf("crunch-run finished")
1631 runner.CrunchLog.Close()
1634 err = runner.fetchContainerRecord()
1638 if runner.Container.State != "Locked" {
1639 return fmt.Errorf("dispatch error detected: container %q has state %q", runner.Container.UUID, runner.Container.State)
1642 var bindmounts map[string]bindmount
1644 // checkErr prints e (unless it's nil) and sets err to
1645 // e (unless err is already non-nil). Thus, if err
1646 // hasn't already been assigned when Run() returns,
1647 // this cleanup func will cause Run() to return the
1648 // first non-nil error that is passed to checkErr().
1649 checkErr := func(errorIn string, e error) {
1653 runner.CrunchLog.Printf("error in %s: %v", errorIn, e)
1657 if runner.finalState == "Complete" {
1658 // There was an error in the finalization.
1659 runner.finalState = "Cancelled"
1663 // Log the error encountered in Run(), if any
1664 checkErr("Run", err)
1666 if runner.finalState == "Queued" {
1667 runner.UpdateContainerFinal()
1671 if runner.IsCancelled() {
1672 runner.finalState = "Cancelled"
1673 // but don't return yet -- we still want to
1674 // capture partial output and write logs
1677 if bindmounts != nil {
1678 checkErr("CaptureOutput", runner.CaptureOutput(bindmounts))
1680 checkErr("stopHoststat", runner.stopHoststat())
1681 checkErr("CommitLogs", runner.CommitLogs())
1682 runner.CleanupDirs()
1683 checkErr("UpdateContainerFinal", runner.UpdateContainerFinal())
1686 runner.setupSignals()
1687 err = runner.startHoststat()
1691 if runner.keepstore != nil {
1692 runner.hoststatReporter.ReportPID("keepstore", runner.keepstore.Process.Pid)
1695 // set up FUSE mount and binds
1696 bindmounts, err = runner.SetupMounts()
1698 runner.finalState = "Cancelled"
1699 err = fmt.Errorf("While setting up mounts: %v", err)
1703 // check for and/or load image
1704 imageID, err := runner.LoadImage()
1706 if !runner.checkBrokenNode(err) {
1707 // Failed to load image but not due to a "broken node"
1708 // condition, probably user error.
1709 runner.finalState = "Cancelled"
1711 err = fmt.Errorf("While loading container image: %v", err)
1715 err = runner.CreateContainer(imageID, bindmounts)
1719 err = runner.LogHostInfo()
1723 err = runner.LogNodeRecord()
1727 err = runner.LogContainerRecord()
1732 if runner.IsCancelled() {
1736 logCollection, err := runner.saveLogCollection(false)
1739 logId = logCollection.PortableDataHash
1741 runner.CrunchLog.Printf("Error committing initial log collection: %v", err)
1743 err = runner.UpdateContainerRunning(logId)
1747 runner.finalState = "Cancelled"
1749 err = runner.startCrunchstat()
1754 err = runner.StartContainer()
1756 runner.checkBrokenNode(err)
1760 err = runner.WaitFinish()
1761 if err == nil && !runner.IsCancelled() {
1762 runner.finalState = "Complete"
1767 // Fetch the current container record (uuid = runner.Container.UUID)
1768 // into runner.Container.
1769 func (runner *ContainerRunner) fetchContainerRecord() error {
1770 reader, err := runner.DispatcherArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
1772 return fmt.Errorf("error fetching container record: %v", err)
1774 defer reader.Close()
1776 dec := json.NewDecoder(reader)
1778 err = dec.Decode(&runner.Container)
1780 return fmt.Errorf("error decoding container record: %v", err)
1784 SecretMounts map[string]arvados.Mount `json:"secret_mounts"`
1787 containerToken, err := runner.ContainerToken()
1789 return fmt.Errorf("error getting container token: %v", err)
1792 runner.ContainerArvClient, runner.ContainerKeepClient,
1793 runner.containerClient, err = runner.MkArvClient(containerToken)
1795 return fmt.Errorf("error creating container API client: %v", err)
1798 runner.ContainerKeepClient.SetStorageClasses(runner.Container.OutputStorageClasses)
1799 runner.DispatcherKeepClient.SetStorageClasses(runner.Container.OutputStorageClasses)
1801 err = runner.ContainerArvClient.Call("GET", "containers", runner.Container.UUID, "secret_mounts", nil, &sm)
1803 if apierr, ok := err.(arvadosclient.APIServerError); !ok || apierr.HttpStatusCode != 404 {
1804 return fmt.Errorf("error fetching secret_mounts: %v", err)
1806 // ok && apierr.HttpStatusCode == 404, which means
1807 // secret_mounts isn't supported by this API server.
1809 runner.SecretMounts = sm.SecretMounts
1814 // NewContainerRunner creates a new container runner.
1815 func NewContainerRunner(dispatcherClient *arvados.Client,
1816 dispatcherArvClient IArvadosClient,
1817 dispatcherKeepClient IKeepClient,
1818 containerUUID string) (*ContainerRunner, error) {
1820 cr := &ContainerRunner{
1821 dispatcherClient: dispatcherClient,
1822 DispatcherArvClient: dispatcherArvClient,
1823 DispatcherKeepClient: dispatcherKeepClient,
1825 cr.NewLogWriter = cr.NewArvLogWriter
1826 cr.RunArvMount = cr.ArvMountCmd
1827 cr.MkTempDir = ioutil.TempDir
1828 cr.MkArvClient = func(token string) (IArvadosClient, IKeepClient, *arvados.Client, error) {
1829 cl, err := arvadosclient.MakeArvadosClient()
1831 return nil, nil, nil, err
1834 kc, err := keepclient.MakeKeepClient(cl)
1836 return nil, nil, nil, err
1838 c2 := arvados.NewClientFromEnv()
1839 c2.AuthToken = token
1840 return cl, kc, c2, nil
1843 cr.LogCollection, err = (&arvados.Collection{}).FileSystem(cr.dispatcherClient, cr.DispatcherKeepClient)
1847 cr.Container.UUID = containerUUID
1848 w, err := cr.NewLogWriter("crunch-run")
1852 cr.CrunchLog = NewThrottledLogger(w)
1853 cr.CrunchLog.Immediate = log.New(os.Stderr, containerUUID+" ", 0)
1855 loadLogThrottleParams(dispatcherArvClient)
1861 func (command) RunCommand(prog string, args []string, stdin io.Reader, stdout, stderr io.Writer) int {
1862 log := log.New(stderr, "", 0)
1863 flags := flag.NewFlagSet(prog, flag.ContinueOnError)
1864 statInterval := flags.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting")
1865 cgroupRoot := flags.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree")
1866 cgroupParent := flags.String("cgroup-parent", "docker", "name of container's parent cgroup (ignored if -cgroup-parent-subsystem is used)")
1867 cgroupParentSubsystem := flags.String("cgroup-parent-subsystem", "", "use current cgroup for given subsystem as parent cgroup for container")
1868 caCertsPath := flags.String("ca-certs", "", "Path to TLS root certificates")
1869 detach := flags.Bool("detach", false, "Detach from parent process and run in the background")
1870 stdinConfig := flags.Bool("stdin-config", false, "Load config and environment variables from JSON message on stdin")
1871 configFile := flags.String("config", arvados.DefaultConfigFile, "filename of cluster config file to try loading if -stdin-config=false (default is $ARVADOS_CONFIG)")
1872 sleep := flags.Duration("sleep", 0, "Delay before starting (testing use only)")
1873 kill := flags.Int("kill", -1, "Send signal to an existing crunch-run process for given UUID")
1874 list := flags.Bool("list", false, "List UUIDs of existing crunch-run processes (and notify them to use price data passed on stdin)")
1875 enableMemoryLimit := flags.Bool("enable-memory-limit", true, "tell container runtime to limit container's memory usage")
1876 enableNetwork := flags.String("container-enable-networking", "default", "enable networking \"always\" (for all containers) or \"default\" (for containers that request it)")
1877 networkMode := flags.String("container-network-mode", "default", `Docker network mode for container (use any argument valid for docker --net)`)
1878 memprofile := flags.String("memprofile", "", "write memory profile to `file` after running container")
1879 runtimeEngine := flags.String("runtime-engine", "docker", "container runtime: docker or singularity")
1880 brokenNodeHook := flags.String("broken-node-hook", "", "script to run if node is detected to be broken (for example, Docker daemon is not running)")
1881 flags.Duration("check-containerd", 0, "Ignored. Exists for compatibility with older versions.")
1882 version := flags.Bool("version", false, "Write version information to stdout and exit 0.")
1884 ignoreDetachFlag := false
1885 if len(args) > 0 && args[0] == "-no-detach" {
1886 // This process was invoked by a parent process, which
1887 // has passed along its own arguments, including
1888 // -detach, after the leading -no-detach flag. Strip
1889 // the leading -no-detach flag (it's not recognized by
1890 // flags.Parse()) and ignore the -detach flag that
1893 ignoreDetachFlag = true
1896 if ok, code := cmd.ParseFlags(flags, prog, args, "container-uuid", stderr); !ok {
1898 } else if *version {
1899 fmt.Fprintln(stdout, prog, cmd.Version.String())
1901 } else if !*list && flags.NArg() != 1 {
1902 fmt.Fprintf(stderr, "missing required argument: container-uuid (try -help)\n")
1906 containerUUID := flags.Arg(0)
1909 case *detach && !ignoreDetachFlag:
1910 return Detach(containerUUID, prog, args, stdin, stdout, stderr)
1912 return KillProcess(containerUUID, syscall.Signal(*kill), stdout, stderr)
1914 return ListProcesses(stdin, stdout, stderr)
1917 if len(containerUUID) != 27 {
1918 log.Printf("usage: %s [options] UUID", prog)
1922 var keepstoreLogbuf bufThenWrite
1925 err := json.NewDecoder(stdin).Decode(&conf)
1927 log.Printf("decode stdin: %s", err)
1930 for k, v := range conf.Env {
1931 err = os.Setenv(k, v)
1933 log.Printf("setenv(%q): %s", k, err)
1937 if conf.Cluster != nil {
1938 // ClusterID is missing from the JSON
1939 // representation, but we need it to generate
1940 // a valid config file for keepstore, so we
1941 // fill it using the container UUID prefix.
1942 conf.Cluster.ClusterID = containerUUID[:5]
1945 conf = hpcConfData(containerUUID, *configFile, io.MultiWriter(&keepstoreLogbuf, stderr))
1948 log.Printf("crunch-run %s started", cmd.Version.String())
1951 if *caCertsPath != "" {
1952 arvadosclient.CertFiles = []string{*caCertsPath}
1955 keepstore, err := startLocalKeepstore(conf, io.MultiWriter(&keepstoreLogbuf, stderr))
1960 if keepstore != nil {
1961 defer keepstore.Process.Kill()
1964 api, err := arvadosclient.MakeArvadosClient()
1966 log.Printf("%s: %v", containerUUID, err)
1971 kc, err := keepclient.MakeKeepClient(api)
1973 log.Printf("%s: %v", containerUUID, err)
1976 kc.BlockCache = &keepclient.BlockCache{MaxBlocks: 2}
1979 cr, err := NewContainerRunner(arvados.NewClientFromEnv(), api, kc, containerUUID)
1985 cr.keepstore = keepstore
1986 if keepstore == nil {
1987 // Log explanation (if any) for why we're not running
1988 // a local keepstore.
1989 var buf bytes.Buffer
1990 keepstoreLogbuf.SetWriter(&buf)
1992 cr.CrunchLog.Printf("%s", strings.TrimSpace(buf.String()))
1994 } else if logWhat := conf.Cluster.Containers.LocalKeepLogsToContainerLog; logWhat == "none" {
1995 cr.CrunchLog.Printf("using local keepstore process (pid %d) at %s", keepstore.Process.Pid, os.Getenv("ARVADOS_KEEP_SERVICES"))
1996 keepstoreLogbuf.SetWriter(io.Discard)
1998 cr.CrunchLog.Printf("using local keepstore process (pid %d) at %s, writing logs to keepstore.txt in log collection", keepstore.Process.Pid, os.Getenv("ARVADOS_KEEP_SERVICES"))
1999 logwriter, err := cr.NewLogWriter("keepstore")
2004 cr.keepstoreLogger = NewThrottledLogger(logwriter)
2006 var writer io.WriteCloser = cr.keepstoreLogger
2007 if logWhat == "errors" {
2008 writer = &filterKeepstoreErrorsOnly{WriteCloser: writer}
2009 } else if logWhat != "all" {
2010 // should have been caught earlier by
2011 // dispatcher's config loader
2012 log.Printf("invalid value for Containers.LocalKeepLogsToContainerLog: %q", logWhat)
2015 err = keepstoreLogbuf.SetWriter(writer)
2020 cr.keepstoreLogbuf = &keepstoreLogbuf
2023 switch *runtimeEngine {
2025 cr.executor, err = newDockerExecutor(containerUUID, cr.CrunchLog.Printf, cr.containerWatchdogInterval)
2027 cr.executor, err = newSingularityExecutor(cr.CrunchLog.Printf)
2029 cr.CrunchLog.Printf("%s: unsupported RuntimeEngine %q", containerUUID, *runtimeEngine)
2030 cr.CrunchLog.Close()
2034 cr.CrunchLog.Printf("%s: %v", containerUUID, err)
2035 cr.checkBrokenNode(err)
2036 cr.CrunchLog.Close()
2039 defer cr.executor.Close()
2041 cr.brokenNodeHook = *brokenNodeHook
2043 gwAuthSecret := os.Getenv("GatewayAuthSecret")
2044 os.Unsetenv("GatewayAuthSecret")
2045 if gwAuthSecret == "" {
2046 // not safe to run a gateway service without an auth
2048 cr.CrunchLog.Printf("Not starting a gateway server (GatewayAuthSecret was not provided by dispatcher)")
2050 gwListen := os.Getenv("GatewayAddress")
2051 cr.gateway = Gateway{
2053 AuthSecret: gwAuthSecret,
2054 ContainerUUID: containerUUID,
2055 Target: cr.executor,
2059 // Direct connection won't work, so we use the
2060 // gateway_address field to indicate the
2061 // internalURL of the controller process that
2062 // has the current tunnel connection.
2063 cr.gateway.ArvadosClient = cr.dispatcherClient
2064 cr.gateway.UpdateTunnelURL = func(url string) {
2065 cr.gateway.Address = "tunnel " + url
2066 cr.DispatcherArvClient.Update("containers", containerUUID,
2067 arvadosclient.Dict{"container": arvadosclient.Dict{"gateway_address": cr.gateway.Address}}, nil)
2070 err = cr.gateway.Start()
2072 log.Printf("error starting gateway server: %s", err)
2077 parentTemp, tmperr := cr.MkTempDir("", "crunch-run."+containerUUID+".")
2079 log.Printf("%s: %v", containerUUID, tmperr)
2083 cr.parentTemp = parentTemp
2084 cr.statInterval = *statInterval
2085 cr.cgroupRoot = *cgroupRoot
2086 cr.expectCgroupParent = *cgroupParent
2087 cr.enableMemoryLimit = *enableMemoryLimit
2088 cr.enableNetwork = *enableNetwork
2089 cr.networkMode = *networkMode
2090 if *cgroupParentSubsystem != "" {
2091 p, err := findCgroup(*cgroupParentSubsystem)
2093 log.Printf("fatal: cgroup parent subsystem: %s", err)
2096 cr.setCgroupParent = p
2097 cr.expectCgroupParent = p
2100 if conf.EC2SpotCheck {
2101 go cr.checkSpotInterruptionNotices()
2106 if *memprofile != "" {
2107 f, err := os.Create(*memprofile)
2109 log.Printf("could not create memory profile: %s", err)
2111 runtime.GC() // get up-to-date statistics
2112 if err := pprof.WriteHeapProfile(f); err != nil {
2113 log.Printf("could not write memory profile: %s", err)
2115 closeerr := f.Close()
2116 if closeerr != nil {
2117 log.Printf("closing memprofile file: %s", err)
2122 log.Printf("%s: %v", containerUUID, runerr)
2128 // Try to load ConfigData in hpc (slurm/lsf) environment. This means
2129 // loading the cluster config from the specified file and (if that
2130 // works) getting the runtime_constraints container field from
2131 // controller to determine # VCPUs so we can calculate KeepBuffers.
2132 func hpcConfData(uuid string, configFile string, stderr io.Writer) ConfigData {
2134 conf.Cluster = loadClusterConfigFile(configFile, stderr)
2135 if conf.Cluster == nil {
2136 // skip loading the container record -- we won't be
2137 // able to start local keepstore anyway.
2140 arv, err := arvadosclient.MakeArvadosClient()
2142 fmt.Fprintf(stderr, "error setting up arvadosclient: %s\n", err)
2146 var ctr arvados.Container
2147 err = arv.Call("GET", "containers", uuid, "", arvadosclient.Dict{"select": []string{"runtime_constraints"}}, &ctr)
2149 fmt.Fprintf(stderr, "error getting container record: %s\n", err)
2152 if ctr.RuntimeConstraints.VCPUs > 0 {
2153 conf.KeepBuffers = ctr.RuntimeConstraints.VCPUs * conf.Cluster.Containers.LocalKeepBlobBuffersPerVCPU
2158 // Load cluster config file from given path. If an error occurs, log
2159 // the error to stderr and return nil.
2160 func loadClusterConfigFile(path string, stderr io.Writer) *arvados.Cluster {
2161 ldr := config.NewLoader(&bytes.Buffer{}, ctxlog.New(stderr, "plain", "info"))
2163 cfg, err := ldr.Load()
2165 fmt.Fprintf(stderr, "could not load config file %s: %s\n", path, err)
2168 cluster, err := cfg.GetCluster("")
2170 fmt.Fprintf(stderr, "could not use config file %s: %s\n", path, err)
2173 fmt.Fprintf(stderr, "loaded config file %s\n", path)
2177 func startLocalKeepstore(configData ConfigData, logbuf io.Writer) (*exec.Cmd, error) {
2178 if configData.KeepBuffers < 1 {
2179 fmt.Fprintf(logbuf, "not starting a local keepstore process because KeepBuffers=%v in config\n", configData.KeepBuffers)
2182 if configData.Cluster == nil {
2183 fmt.Fprint(logbuf, "not starting a local keepstore process because cluster config file was not loaded\n")
2186 for uuid, vol := range configData.Cluster.Volumes {
2187 if len(vol.AccessViaHosts) > 0 {
2188 fmt.Fprintf(logbuf, "not starting a local keepstore process because a volume (%s) uses AccessViaHosts\n", uuid)
2191 if !vol.ReadOnly && vol.Replication < configData.Cluster.Collections.DefaultReplication {
2192 fmt.Fprintf(logbuf, "not starting a local keepstore process because a writable volume (%s) has replication less than Collections.DefaultReplication (%d < %d)\n", uuid, vol.Replication, configData.Cluster.Collections.DefaultReplication)
2197 // Rather than have an alternate way to tell keepstore how
2198 // many buffers to use when starting it this way, we just
2199 // modify the cluster configuration that we feed it on stdin.
2200 configData.Cluster.API.MaxKeepBlobBuffers = configData.KeepBuffers
2202 localaddr := localKeepstoreAddr()
2203 ln, err := net.Listen("tcp", net.JoinHostPort(localaddr, "0"))
2207 _, port, err := net.SplitHostPort(ln.Addr().String())
2213 url := "http://" + net.JoinHostPort(localaddr, port)
2215 fmt.Fprintf(logbuf, "starting keepstore on %s\n", url)
2217 var confJSON bytes.Buffer
2218 err = json.NewEncoder(&confJSON).Encode(arvados.Config{
2219 Clusters: map[string]arvados.Cluster{
2220 configData.Cluster.ClusterID: *configData.Cluster,
2226 cmd := exec.Command("/proc/self/exe", "keepstore", "-config=-")
2227 if target, err := os.Readlink(cmd.Path); err == nil && strings.HasSuffix(target, ".test") {
2228 // If we're a 'go test' process, running
2229 // /proc/self/exe would start the test suite in a
2230 // child process, which is not what we want.
2231 cmd.Path, _ = exec.LookPath("go")
2232 cmd.Args = append([]string{"go", "run", "../../cmd/arvados-server"}, cmd.Args[1:]...)
2233 cmd.Env = os.Environ()
2235 cmd.Stdin = &confJSON
2238 cmd.Env = append(cmd.Env,
2240 "ARVADOS_SERVICE_INTERNAL_URL="+url)
2243 return nil, fmt.Errorf("error starting keepstore process: %w", err)
2250 ctx, cancel := context.WithDeadline(context.Background(), time.Now().Add(time.Second*10))
2252 poll := time.NewTicker(time.Second / 10)
2254 client := http.Client{}
2256 testReq, err := http.NewRequestWithContext(ctx, "GET", url+"/_health/ping", nil)
2257 testReq.Header.Set("Authorization", "Bearer "+configData.Cluster.ManagementToken)
2261 resp, err := client.Do(testReq)
2264 if resp.StatusCode == http.StatusOK {
2269 return nil, fmt.Errorf("keepstore child process exited")
2271 if ctx.Err() != nil {
2272 return nil, fmt.Errorf("timed out waiting for new keepstore process to report healthy")
2275 os.Setenv("ARVADOS_KEEP_SERVICES", url)
2279 // return current uid, gid, groups in a format suitable for logging:
2280 // "crunch-run process has uid=1234(arvados) gid=1234(arvados)
2281 // groups=1234(arvados),114(fuse)"
2282 func currentUserAndGroups() string {
2283 u, err := user.Current()
2285 return fmt.Sprintf("error getting current user ID: %s", err)
2287 s := fmt.Sprintf("crunch-run process has uid=%s(%s) gid=%s", u.Uid, u.Username, u.Gid)
2288 if g, err := user.LookupGroupId(u.Gid); err == nil {
2289 s += fmt.Sprintf("(%s)", g.Name)
2292 if gids, err := u.GroupIds(); err == nil {
2293 for i, gid := range gids {
2298 if g, err := user.LookupGroupId(gid); err == nil {
2299 s += fmt.Sprintf("(%s)", g.Name)
2306 // Return a suitable local interface address for a local keepstore
2307 // service. Currently this is the numerically lowest non-loopback ipv4
2308 // address assigned to a local interface that is not in any of the
2309 // link-local/vpn/loopback ranges 169.254/16, 100.64/10, or 127/8.
2310 func localKeepstoreAddr() string {
2312 // Ignore error (proceed with zero IPs)
2313 addrs, _ := processIPs(os.Getpid())
2314 for addr := range addrs {
2315 ip := net.ParseIP(addr)
2320 if ip.Mask(net.CIDRMask(8, 32)).Equal(net.IPv4(127, 0, 0, 0)) ||
2321 ip.Mask(net.CIDRMask(10, 32)).Equal(net.IPv4(100, 64, 0, 0)) ||
2322 ip.Mask(net.CIDRMask(16, 32)).Equal(net.IPv4(169, 254, 0, 0)) {
2326 ips = append(ips, ip)
2331 sort.Slice(ips, func(ii, jj int) bool {
2332 i, j := ips[ii], ips[jj]
2333 if len(i) != len(j) {
2334 return len(i) < len(j)
2343 return ips[0].String()
2346 func (cr *ContainerRunner) loadPrices() {
2347 buf, err := os.ReadFile(filepath.Join(lockdir, pricesfile))
2349 if !os.IsNotExist(err) {
2350 cr.CrunchLog.Printf("loadPrices: read: %s", err)
2354 var prices []cloud.InstancePrice
2355 err = json.Unmarshal(buf, &prices)
2357 cr.CrunchLog.Printf("loadPrices: decode: %s", err)
2360 cr.pricesLock.Lock()
2361 defer cr.pricesLock.Unlock()
2362 var lastKnown time.Time
2363 if len(cr.prices) > 0 {
2364 lastKnown = cr.prices[0].StartTime
2366 cr.prices = cloud.NormalizePriceHistory(append(prices, cr.prices...))
2367 for i := len(cr.prices) - 1; i >= 0; i-- {
2368 price := cr.prices[i]
2369 if price.StartTime.After(lastKnown) {
2370 cr.CrunchLog.Printf("Instance price changed to %#.3g at %s", price.Price, price.StartTime.UTC())
2375 func (cr *ContainerRunner) calculateCost(now time.Time) float64 {
2376 cr.pricesLock.Lock()
2377 defer cr.pricesLock.Unlock()
2379 // First, make a "prices" slice with the real data as far back
2380 // as it goes, and (if needed) a "since the beginning of time"
2381 // placeholder containing a reasonable guess about what the
2382 // price was between cr.costStartTime and the earliest real
2385 if len(prices) == 0 {
2386 // use price info in InstanceType record initially
2387 // provided by cloud dispatcher
2389 var it arvados.InstanceType
2390 if j := os.Getenv("InstanceType"); j != "" && json.Unmarshal([]byte(j), &it) == nil && it.Price > 0 {
2393 prices = []cloud.InstancePrice{{Price: p}}
2394 } else if prices[len(prices)-1].StartTime.After(cr.costStartTime) {
2395 // guess earlier pricing was the same as the earliest
2396 // price we know about
2397 filler := prices[len(prices)-1]
2398 filler.StartTime = time.Time{}
2399 prices = append(prices, filler)
2402 // Now that our history of price changes goes back at least as
2403 // far as cr.costStartTime, add up the costs for each
2407 for _, ip := range prices {
2408 spanStart := ip.StartTime
2409 if spanStart.After(now) {
2410 // pricing information from the future -- not
2411 // expected from AWS, but possible in
2412 // principle, and exercised by tests.
2416 if spanStart.Before(cr.costStartTime) {
2417 spanStart = cr.costStartTime
2420 cost += ip.Price * spanEnd.Sub(spanStart).Seconds() / 3600