19166: Don't set up a reverse tunnel if we won't be advertising it.
[arvados.git] / lib / crunchrun / crunchrun.go
1 // Copyright (C) The Arvados Authors. All rights reserved.
2 //
3 // SPDX-License-Identifier: AGPL-3.0
4
5 package crunchrun
6
7 import (
8         "bytes"
9         "context"
10         "encoding/json"
11         "errors"
12         "flag"
13         "fmt"
14         "io"
15         "io/ioutil"
16         "log"
17         "net"
18         "net/http"
19         "os"
20         "os/exec"
21         "os/signal"
22         "os/user"
23         "path"
24         "path/filepath"
25         "regexp"
26         "runtime"
27         "runtime/pprof"
28         "sort"
29         "strings"
30         "sync"
31         "syscall"
32         "time"
33
34         "git.arvados.org/arvados.git/lib/cmd"
35         "git.arvados.org/arvados.git/lib/config"
36         "git.arvados.org/arvados.git/lib/crunchstat"
37         "git.arvados.org/arvados.git/sdk/go/arvados"
38         "git.arvados.org/arvados.git/sdk/go/arvadosclient"
39         "git.arvados.org/arvados.git/sdk/go/ctxlog"
40         "git.arvados.org/arvados.git/sdk/go/keepclient"
41         "git.arvados.org/arvados.git/sdk/go/manifest"
42         "golang.org/x/sys/unix"
43 )
44
45 type command struct{}
46
47 var Command = command{}
48
49 // ConfigData contains environment variables and (when needed) cluster
50 // configuration, passed from dispatchcloud to crunch-run on stdin.
51 type ConfigData struct {
52         Env         map[string]string
53         KeepBuffers int
54         Cluster     *arvados.Cluster
55 }
56
57 // IArvadosClient is the minimal Arvados API methods used by crunch-run.
58 type IArvadosClient interface {
59         Create(resourceType string, parameters arvadosclient.Dict, output interface{}) error
60         Get(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
61         Update(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
62         Call(method, resourceType, uuid, action string, parameters arvadosclient.Dict, output interface{}) error
63         CallRaw(method string, resourceType string, uuid string, action string, parameters arvadosclient.Dict) (reader io.ReadCloser, err error)
64         Discovery(key string) (interface{}, error)
65 }
66
67 // ErrCancelled is the error returned when the container is cancelled.
68 var ErrCancelled = errors.New("Cancelled")
69
70 // IKeepClient is the minimal Keep API methods used by crunch-run.
71 type IKeepClient interface {
72         BlockWrite(context.Context, arvados.BlockWriteOptions) (arvados.BlockWriteResponse, error)
73         ReadAt(locator string, p []byte, off int) (int, error)
74         ManifestFileReader(m manifest.Manifest, filename string) (arvados.File, error)
75         LocalLocator(locator string) (string, error)
76         ClearBlockCache()
77         SetStorageClasses(sc []string)
78 }
79
80 // NewLogWriter is a factory function to create a new log writer.
81 type NewLogWriter func(name string) (io.WriteCloser, error)
82
83 type RunArvMount func(cmdline []string, tok string) (*exec.Cmd, error)
84
85 type MkTempDir func(string, string) (string, error)
86
87 type PsProcess interface {
88         CmdlineSlice() ([]string, error)
89 }
90
91 // ContainerRunner is the main stateful struct used for a single execution of a
92 // container.
93 type ContainerRunner struct {
94         executor       containerExecutor
95         executorStdin  io.Closer
96         executorStdout io.Closer
97         executorStderr io.Closer
98
99         // Dispatcher client is initialized with the Dispatcher token.
100         // This is a privileged token used to manage container status
101         // and logs.
102         //
103         // We have both dispatcherClient and DispatcherArvClient
104         // because there are two different incompatible Arvados Go
105         // SDKs and we have to use both (hopefully this gets fixed in
106         // #14467)
107         dispatcherClient     *arvados.Client
108         DispatcherArvClient  IArvadosClient
109         DispatcherKeepClient IKeepClient
110
111         // Container client is initialized with the Container token
112         // This token controls the permissions of the container, and
113         // must be used for operations such as reading collections.
114         //
115         // Same comment as above applies to
116         // containerClient/ContainerArvClient.
117         containerClient     *arvados.Client
118         ContainerArvClient  IArvadosClient
119         ContainerKeepClient IKeepClient
120
121         Container     arvados.Container
122         token         string
123         ExitCode      *int
124         NewLogWriter  NewLogWriter
125         CrunchLog     *ThrottledLogger
126         logUUID       string
127         logMtx        sync.Mutex
128         LogCollection arvados.CollectionFileSystem
129         LogsPDH       *string
130         RunArvMount   RunArvMount
131         MkTempDir     MkTempDir
132         ArvMount      *exec.Cmd
133         ArvMountPoint string
134         HostOutputDir string
135         Volumes       map[string]struct{}
136         OutputPDH     *string
137         SigChan       chan os.Signal
138         ArvMountExit  chan error
139         SecretMounts  map[string]arvados.Mount
140         MkArvClient   func(token string) (IArvadosClient, IKeepClient, *arvados.Client, error)
141         finalState    string
142         parentTemp    string
143
144         keepstoreLogger  io.WriteCloser
145         keepstoreLogbuf  *bufThenWrite
146         statLogger       io.WriteCloser
147         statReporter     *crunchstat.Reporter
148         hoststatLogger   io.WriteCloser
149         hoststatReporter *crunchstat.Reporter
150         statInterval     time.Duration
151         cgroupRoot       string
152         // What we expect the container's cgroup parent to be.
153         expectCgroupParent string
154         // What we tell docker to use as the container's cgroup
155         // parent. Note: Ideally we would use the same field for both
156         // expectCgroupParent and setCgroupParent, and just make it
157         // default to "docker". However, when using docker < 1.10 with
158         // systemd, specifying a non-empty cgroup parent (even the
159         // default value "docker") hits a docker bug
160         // (https://github.com/docker/docker/issues/17126). Using two
161         // separate fields makes it possible to use the "expect cgroup
162         // parent to be X" feature even on sites where the "specify
163         // cgroup parent" feature breaks.
164         setCgroupParent string
165
166         cStateLock sync.Mutex
167         cCancelled bool // StopContainer() invoked
168
169         enableMemoryLimit bool
170         enableNetwork     string // one of "default" or "always"
171         networkMode       string // "none", "host", or "" -- passed through to executor
172         brokenNodeHook    string // script to run if node appears to be broken
173         arvMountLog       *ThrottledLogger
174
175         containerWatchdogInterval time.Duration
176
177         gateway Gateway
178 }
179
180 // setupSignals sets up signal handling to gracefully terminate the
181 // underlying container and update state when receiving a TERM, INT or
182 // QUIT signal.
183 func (runner *ContainerRunner) setupSignals() {
184         runner.SigChan = make(chan os.Signal, 1)
185         signal.Notify(runner.SigChan, syscall.SIGTERM)
186         signal.Notify(runner.SigChan, syscall.SIGINT)
187         signal.Notify(runner.SigChan, syscall.SIGQUIT)
188
189         go func(sig chan os.Signal) {
190                 for s := range sig {
191                         runner.stop(s)
192                 }
193         }(runner.SigChan)
194 }
195
196 // stop the underlying container.
197 func (runner *ContainerRunner) stop(sig os.Signal) {
198         runner.cStateLock.Lock()
199         defer runner.cStateLock.Unlock()
200         if sig != nil {
201                 runner.CrunchLog.Printf("caught signal: %v", sig)
202         }
203         runner.cCancelled = true
204         runner.CrunchLog.Printf("stopping container")
205         err := runner.executor.Stop()
206         if err != nil {
207                 runner.CrunchLog.Printf("error stopping container: %s", err)
208         }
209 }
210
211 var errorBlacklist = []string{
212         "(?ms).*[Cc]annot connect to the Docker daemon.*",
213         "(?ms).*oci runtime error.*starting container process.*container init.*mounting.*to rootfs.*no such file or directory.*",
214         "(?ms).*grpc: the connection is unavailable.*",
215 }
216
217 func (runner *ContainerRunner) runBrokenNodeHook() {
218         if runner.brokenNodeHook == "" {
219                 path := filepath.Join(lockdir, brokenfile)
220                 runner.CrunchLog.Printf("Writing %s to mark node as broken", path)
221                 f, err := os.OpenFile(path, os.O_CREATE|os.O_RDWR, 0700)
222                 if err != nil {
223                         runner.CrunchLog.Printf("Error writing %s: %s", path, err)
224                         return
225                 }
226                 f.Close()
227         } else {
228                 runner.CrunchLog.Printf("Running broken node hook %q", runner.brokenNodeHook)
229                 // run killme script
230                 c := exec.Command(runner.brokenNodeHook)
231                 c.Stdout = runner.CrunchLog
232                 c.Stderr = runner.CrunchLog
233                 err := c.Run()
234                 if err != nil {
235                         runner.CrunchLog.Printf("Error running broken node hook: %v", err)
236                 }
237         }
238 }
239
240 func (runner *ContainerRunner) checkBrokenNode(goterr error) bool {
241         for _, d := range errorBlacklist {
242                 if m, e := regexp.MatchString(d, goterr.Error()); m && e == nil {
243                         runner.CrunchLog.Printf("Error suggests node is unable to run containers: %v", goterr)
244                         runner.runBrokenNodeHook()
245                         return true
246                 }
247         }
248         return false
249 }
250
251 // LoadImage determines the docker image id from the container record and
252 // checks if it is available in the local Docker image store.  If not, it loads
253 // the image from Keep.
254 func (runner *ContainerRunner) LoadImage() (string, error) {
255         runner.CrunchLog.Printf("Fetching Docker image from collection '%s'", runner.Container.ContainerImage)
256
257         d, err := os.Open(runner.ArvMountPoint + "/by_id/" + runner.Container.ContainerImage)
258         if err != nil {
259                 return "", err
260         }
261         defer d.Close()
262         allfiles, err := d.Readdirnames(-1)
263         if err != nil {
264                 return "", err
265         }
266         var tarfiles []string
267         for _, fnm := range allfiles {
268                 if strings.HasSuffix(fnm, ".tar") {
269                         tarfiles = append(tarfiles, fnm)
270                 }
271         }
272         if len(tarfiles) == 0 {
273                 return "", fmt.Errorf("image collection does not include a .tar image file")
274         }
275         if len(tarfiles) > 1 {
276                 return "", fmt.Errorf("cannot choose from multiple tar files in image collection: %v", tarfiles)
277         }
278         imageID := tarfiles[0][:len(tarfiles[0])-4]
279         imageTarballPath := runner.ArvMountPoint + "/by_id/" + runner.Container.ContainerImage + "/" + imageID + ".tar"
280         runner.CrunchLog.Printf("Using Docker image id %q", imageID)
281
282         runner.CrunchLog.Print("Loading Docker image from keep")
283         err = runner.executor.LoadImage(imageID, imageTarballPath, runner.Container, runner.ArvMountPoint,
284                 runner.containerClient)
285         if err != nil {
286                 return "", err
287         }
288
289         return imageID, nil
290 }
291
292 func (runner *ContainerRunner) ArvMountCmd(cmdline []string, token string) (c *exec.Cmd, err error) {
293         c = exec.Command(cmdline[0], cmdline[1:]...)
294
295         // Copy our environment, but override ARVADOS_API_TOKEN with
296         // the container auth token.
297         c.Env = nil
298         for _, s := range os.Environ() {
299                 if !strings.HasPrefix(s, "ARVADOS_API_TOKEN=") {
300                         c.Env = append(c.Env, s)
301                 }
302         }
303         c.Env = append(c.Env, "ARVADOS_API_TOKEN="+token)
304
305         w, err := runner.NewLogWriter("arv-mount")
306         if err != nil {
307                 return nil, err
308         }
309         runner.arvMountLog = NewThrottledLogger(w)
310         scanner := logScanner{
311                 Patterns: []string{
312                         "Keep write error",
313                         "Block not found error",
314                         "Unhandled exception during FUSE operation",
315                 },
316                 ReportFunc: runner.reportArvMountWarning,
317         }
318         c.Stdout = runner.arvMountLog
319         c.Stderr = io.MultiWriter(runner.arvMountLog, os.Stderr, &scanner)
320
321         runner.CrunchLog.Printf("Running %v", c.Args)
322
323         err = c.Start()
324         if err != nil {
325                 return nil, err
326         }
327
328         statReadme := make(chan bool)
329         runner.ArvMountExit = make(chan error)
330
331         keepStatting := true
332         go func() {
333                 for keepStatting {
334                         time.Sleep(100 * time.Millisecond)
335                         _, err = os.Stat(fmt.Sprintf("%s/by_id/README", runner.ArvMountPoint))
336                         if err == nil {
337                                 keepStatting = false
338                                 statReadme <- true
339                         }
340                 }
341                 close(statReadme)
342         }()
343
344         go func() {
345                 mnterr := c.Wait()
346                 if mnterr != nil {
347                         runner.CrunchLog.Printf("Arv-mount exit error: %v", mnterr)
348                 }
349                 runner.ArvMountExit <- mnterr
350                 close(runner.ArvMountExit)
351         }()
352
353         select {
354         case <-statReadme:
355                 break
356         case err := <-runner.ArvMountExit:
357                 runner.ArvMount = nil
358                 keepStatting = false
359                 return nil, err
360         }
361
362         return c, nil
363 }
364
365 func (runner *ContainerRunner) SetupArvMountPoint(prefix string) (err error) {
366         if runner.ArvMountPoint == "" {
367                 runner.ArvMountPoint, err = runner.MkTempDir(runner.parentTemp, prefix)
368         }
369         return
370 }
371
372 func copyfile(src string, dst string) (err error) {
373         srcfile, err := os.Open(src)
374         if err != nil {
375                 return
376         }
377
378         os.MkdirAll(path.Dir(dst), 0777)
379
380         dstfile, err := os.Create(dst)
381         if err != nil {
382                 return
383         }
384         _, err = io.Copy(dstfile, srcfile)
385         if err != nil {
386                 return
387         }
388
389         err = srcfile.Close()
390         err2 := dstfile.Close()
391
392         if err != nil {
393                 return
394         }
395
396         if err2 != nil {
397                 return err2
398         }
399
400         return nil
401 }
402
403 func (runner *ContainerRunner) SetupMounts() (map[string]bindmount, error) {
404         bindmounts := map[string]bindmount{}
405         err := runner.SetupArvMountPoint("keep")
406         if err != nil {
407                 return nil, fmt.Errorf("While creating keep mount temp dir: %v", err)
408         }
409
410         token, err := runner.ContainerToken()
411         if err != nil {
412                 return nil, fmt.Errorf("could not get container token: %s", err)
413         }
414         runner.CrunchLog.Printf("container token %q", token)
415
416         pdhOnly := true
417         tmpcount := 0
418         arvMountCmd := []string{
419                 "arv-mount",
420                 "--foreground",
421                 "--read-write",
422                 "--storage-classes", strings.Join(runner.Container.OutputStorageClasses, ","),
423                 fmt.Sprintf("--crunchstat-interval=%v", runner.statInterval.Seconds())}
424
425         if _, isdocker := runner.executor.(*dockerExecutor); isdocker {
426                 arvMountCmd = append(arvMountCmd, "--allow-other")
427         }
428
429         if runner.Container.RuntimeConstraints.KeepCacheRAM > 0 {
430                 arvMountCmd = append(arvMountCmd, "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheRAM))
431         }
432
433         collectionPaths := []string{}
434         needCertMount := true
435         type copyFile struct {
436                 src  string
437                 bind string
438         }
439         var copyFiles []copyFile
440
441         var binds []string
442         for bind := range runner.Container.Mounts {
443                 binds = append(binds, bind)
444         }
445         for bind := range runner.SecretMounts {
446                 if _, ok := runner.Container.Mounts[bind]; ok {
447                         return nil, fmt.Errorf("secret mount %q conflicts with regular mount", bind)
448                 }
449                 if runner.SecretMounts[bind].Kind != "json" &&
450                         runner.SecretMounts[bind].Kind != "text" {
451                         return nil, fmt.Errorf("secret mount %q type is %q but only 'json' and 'text' are permitted",
452                                 bind, runner.SecretMounts[bind].Kind)
453                 }
454                 binds = append(binds, bind)
455         }
456         sort.Strings(binds)
457
458         for _, bind := range binds {
459                 mnt, notSecret := runner.Container.Mounts[bind]
460                 if !notSecret {
461                         mnt = runner.SecretMounts[bind]
462                 }
463                 if bind == "stdout" || bind == "stderr" {
464                         // Is it a "file" mount kind?
465                         if mnt.Kind != "file" {
466                                 return nil, fmt.Errorf("unsupported mount kind '%s' for %s: only 'file' is supported", mnt.Kind, bind)
467                         }
468
469                         // Does path start with OutputPath?
470                         prefix := runner.Container.OutputPath
471                         if !strings.HasSuffix(prefix, "/") {
472                                 prefix += "/"
473                         }
474                         if !strings.HasPrefix(mnt.Path, prefix) {
475                                 return nil, fmt.Errorf("%s path does not start with OutputPath: %s, %s", strings.Title(bind), mnt.Path, prefix)
476                         }
477                 }
478
479                 if bind == "stdin" {
480                         // Is it a "collection" mount kind?
481                         if mnt.Kind != "collection" && mnt.Kind != "json" {
482                                 return nil, fmt.Errorf("unsupported mount kind '%s' for stdin: only 'collection' and 'json' are supported", mnt.Kind)
483                         }
484                 }
485
486                 if bind == "/etc/arvados/ca-certificates.crt" {
487                         needCertMount = false
488                 }
489
490                 if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && bind != runner.Container.OutputPath+"/" {
491                         if mnt.Kind != "collection" && mnt.Kind != "text" && mnt.Kind != "json" {
492                                 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)
493                         }
494                 }
495
496                 switch {
497                 case mnt.Kind == "collection" && bind != "stdin":
498                         var src string
499                         if mnt.UUID != "" && mnt.PortableDataHash != "" {
500                                 return nil, fmt.Errorf("cannot specify both 'uuid' and 'portable_data_hash' for a collection mount")
501                         }
502                         if mnt.UUID != "" {
503                                 if mnt.Writable {
504                                         return nil, fmt.Errorf("writing to existing collections currently not permitted")
505                                 }
506                                 pdhOnly = false
507                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.UUID)
508                         } else if mnt.PortableDataHash != "" {
509                                 if mnt.Writable && !strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
510                                         return nil, fmt.Errorf("can never write to a collection specified by portable data hash")
511                                 }
512                                 idx := strings.Index(mnt.PortableDataHash, "/")
513                                 if idx > 0 {
514                                         mnt.Path = path.Clean(mnt.PortableDataHash[idx:])
515                                         mnt.PortableDataHash = mnt.PortableDataHash[0:idx]
516                                         runner.Container.Mounts[bind] = mnt
517                                 }
518                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.PortableDataHash)
519                                 if mnt.Path != "" && mnt.Path != "." {
520                                         if strings.HasPrefix(mnt.Path, "./") {
521                                                 mnt.Path = mnt.Path[2:]
522                                         } else if strings.HasPrefix(mnt.Path, "/") {
523                                                 mnt.Path = mnt.Path[1:]
524                                         }
525                                         src += "/" + mnt.Path
526                                 }
527                         } else {
528                                 src = fmt.Sprintf("%s/tmp%d", runner.ArvMountPoint, tmpcount)
529                                 arvMountCmd = append(arvMountCmd, "--mount-tmp", fmt.Sprintf("tmp%d", tmpcount))
530                                 tmpcount++
531                         }
532                         if mnt.Writable {
533                                 if bind == runner.Container.OutputPath {
534                                         runner.HostOutputDir = src
535                                         bindmounts[bind] = bindmount{HostPath: src}
536                                 } else if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
537                                         copyFiles = append(copyFiles, copyFile{src, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
538                                 } else {
539                                         bindmounts[bind] = bindmount{HostPath: src}
540                                 }
541                         } else {
542                                 bindmounts[bind] = bindmount{HostPath: src, ReadOnly: true}
543                         }
544                         collectionPaths = append(collectionPaths, src)
545
546                 case mnt.Kind == "tmp":
547                         var tmpdir string
548                         tmpdir, err = runner.MkTempDir(runner.parentTemp, "tmp")
549                         if err != nil {
550                                 return nil, fmt.Errorf("while creating mount temp dir: %v", err)
551                         }
552                         st, staterr := os.Stat(tmpdir)
553                         if staterr != nil {
554                                 return nil, fmt.Errorf("while Stat on temp dir: %v", staterr)
555                         }
556                         err = os.Chmod(tmpdir, st.Mode()|os.ModeSetgid|0777)
557                         if staterr != nil {
558                                 return nil, fmt.Errorf("while Chmod temp dir: %v", err)
559                         }
560                         bindmounts[bind] = bindmount{HostPath: tmpdir}
561                         if bind == runner.Container.OutputPath {
562                                 runner.HostOutputDir = tmpdir
563                         }
564
565                 case mnt.Kind == "json" || mnt.Kind == "text":
566                         var filedata []byte
567                         if mnt.Kind == "json" {
568                                 filedata, err = json.Marshal(mnt.Content)
569                                 if err != nil {
570                                         return nil, fmt.Errorf("encoding json data: %v", err)
571                                 }
572                         } else {
573                                 text, ok := mnt.Content.(string)
574                                 if !ok {
575                                         return nil, fmt.Errorf("content for mount %q must be a string", bind)
576                                 }
577                                 filedata = []byte(text)
578                         }
579
580                         tmpdir, err := runner.MkTempDir(runner.parentTemp, mnt.Kind)
581                         if err != nil {
582                                 return nil, fmt.Errorf("creating temp dir: %v", err)
583                         }
584                         tmpfn := filepath.Join(tmpdir, "mountdata."+mnt.Kind)
585                         err = ioutil.WriteFile(tmpfn, filedata, 0444)
586                         if err != nil {
587                                 return nil, fmt.Errorf("writing temp file: %v", err)
588                         }
589                         if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && (notSecret || runner.Container.Mounts[runner.Container.OutputPath].Kind != "collection") {
590                                 // In most cases, if the container
591                                 // specifies a literal file inside the
592                                 // output path, we copy it into the
593                                 // output directory (either a mounted
594                                 // collection or a staging area on the
595                                 // host fs). If it's a secret, it will
596                                 // be skipped when copying output from
597                                 // staging to Keep later.
598                                 copyFiles = append(copyFiles, copyFile{tmpfn, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
599                         } else {
600                                 // If a secret is outside OutputPath,
601                                 // we bind mount the secret file
602                                 // directly just like other mounts. We
603                                 // also use this strategy when a
604                                 // secret is inside OutputPath but
605                                 // OutputPath is a live collection, to
606                                 // avoid writing the secret to
607                                 // Keep. Attempting to remove a
608                                 // bind-mounted secret file from
609                                 // inside the container will return a
610                                 // "Device or resource busy" error
611                                 // that might not be handled well by
612                                 // the container, which is why we
613                                 // don't use this strategy when
614                                 // OutputPath is a staging directory.
615                                 bindmounts[bind] = bindmount{HostPath: tmpfn, ReadOnly: true}
616                         }
617
618                 case mnt.Kind == "git_tree":
619                         tmpdir, err := runner.MkTempDir(runner.parentTemp, "git_tree")
620                         if err != nil {
621                                 return nil, fmt.Errorf("creating temp dir: %v", err)
622                         }
623                         err = gitMount(mnt).extractTree(runner.ContainerArvClient, tmpdir, token)
624                         if err != nil {
625                                 return nil, err
626                         }
627                         bindmounts[bind] = bindmount{HostPath: tmpdir, ReadOnly: true}
628                 }
629         }
630
631         if runner.HostOutputDir == "" {
632                 return nil, fmt.Errorf("output path does not correspond to a writable mount point")
633         }
634
635         if needCertMount && runner.Container.RuntimeConstraints.API {
636                 for _, certfile := range arvadosclient.CertFiles {
637                         _, err := os.Stat(certfile)
638                         if err == nil {
639                                 bindmounts["/etc/arvados/ca-certificates.crt"] = bindmount{HostPath: certfile, ReadOnly: true}
640                                 break
641                         }
642                 }
643         }
644
645         if pdhOnly {
646                 // If we are only mounting collections by pdh, make
647                 // sure we don't subscribe to websocket events to
648                 // avoid putting undesired load on the API server
649                 arvMountCmd = append(arvMountCmd, "--mount-by-pdh", "by_id", "--disable-event-listening")
650         } else {
651                 arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_id")
652         }
653         // the by_uuid mount point is used by singularity when writing
654         // out docker images converted to SIF
655         arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_uuid")
656         arvMountCmd = append(arvMountCmd, runner.ArvMountPoint)
657
658         runner.ArvMount, err = runner.RunArvMount(arvMountCmd, token)
659         if err != nil {
660                 return nil, fmt.Errorf("while trying to start arv-mount: %v", err)
661         }
662
663         for _, p := range collectionPaths {
664                 _, err = os.Stat(p)
665                 if err != nil {
666                         return nil, fmt.Errorf("while checking that input files exist: %v", err)
667                 }
668         }
669
670         for _, cp := range copyFiles {
671                 st, err := os.Stat(cp.src)
672                 if err != nil {
673                         return nil, fmt.Errorf("while staging writable file from %q to %q: %v", cp.src, cp.bind, err)
674                 }
675                 if st.IsDir() {
676                         err = filepath.Walk(cp.src, func(walkpath string, walkinfo os.FileInfo, walkerr error) error {
677                                 if walkerr != nil {
678                                         return walkerr
679                                 }
680                                 target := path.Join(cp.bind, walkpath[len(cp.src):])
681                                 if walkinfo.Mode().IsRegular() {
682                                         copyerr := copyfile(walkpath, target)
683                                         if copyerr != nil {
684                                                 return copyerr
685                                         }
686                                         return os.Chmod(target, walkinfo.Mode()|0777)
687                                 } else if walkinfo.Mode().IsDir() {
688                                         mkerr := os.MkdirAll(target, 0777)
689                                         if mkerr != nil {
690                                                 return mkerr
691                                         }
692                                         return os.Chmod(target, walkinfo.Mode()|os.ModeSetgid|0777)
693                                 } else {
694                                         return fmt.Errorf("source %q is not a regular file or directory", cp.src)
695                                 }
696                         })
697                 } else if st.Mode().IsRegular() {
698                         err = copyfile(cp.src, cp.bind)
699                         if err == nil {
700                                 err = os.Chmod(cp.bind, st.Mode()|0777)
701                         }
702                 }
703                 if err != nil {
704                         return nil, fmt.Errorf("while staging writable file from %q to %q: %v", cp.src, cp.bind, err)
705                 }
706         }
707
708         return bindmounts, nil
709 }
710
711 func (runner *ContainerRunner) stopHoststat() error {
712         if runner.hoststatReporter == nil {
713                 return nil
714         }
715         runner.hoststatReporter.Stop()
716         err := runner.hoststatLogger.Close()
717         if err != nil {
718                 return fmt.Errorf("error closing hoststat logs: %v", err)
719         }
720         return nil
721 }
722
723 func (runner *ContainerRunner) startHoststat() error {
724         w, err := runner.NewLogWriter("hoststat")
725         if err != nil {
726                 return err
727         }
728         runner.hoststatLogger = NewThrottledLogger(w)
729         runner.hoststatReporter = &crunchstat.Reporter{
730                 Logger:     log.New(runner.hoststatLogger, "", 0),
731                 CgroupRoot: runner.cgroupRoot,
732                 PollPeriod: runner.statInterval,
733         }
734         runner.hoststatReporter.Start()
735         return nil
736 }
737
738 func (runner *ContainerRunner) startCrunchstat() error {
739         w, err := runner.NewLogWriter("crunchstat")
740         if err != nil {
741                 return err
742         }
743         runner.statLogger = NewThrottledLogger(w)
744         runner.statReporter = &crunchstat.Reporter{
745                 CID:          runner.executor.CgroupID(),
746                 Logger:       log.New(runner.statLogger, "", 0),
747                 CgroupParent: runner.expectCgroupParent,
748                 CgroupRoot:   runner.cgroupRoot,
749                 PollPeriod:   runner.statInterval,
750                 TempDir:      runner.parentTemp,
751         }
752         runner.statReporter.Start()
753         return nil
754 }
755
756 type infoCommand struct {
757         label string
758         cmd   []string
759 }
760
761 // LogHostInfo logs info about the current host, for debugging and
762 // accounting purposes. Although it's logged as "node-info", this is
763 // about the environment where crunch-run is actually running, which
764 // might differ from what's described in the node record (see
765 // LogNodeRecord).
766 func (runner *ContainerRunner) LogHostInfo() (err error) {
767         w, err := runner.NewLogWriter("node-info")
768         if err != nil {
769                 return
770         }
771
772         commands := []infoCommand{
773                 {
774                         label: "Host Information",
775                         cmd:   []string{"uname", "-a"},
776                 },
777                 {
778                         label: "CPU Information",
779                         cmd:   []string{"cat", "/proc/cpuinfo"},
780                 },
781                 {
782                         label: "Memory Information",
783                         cmd:   []string{"cat", "/proc/meminfo"},
784                 },
785                 {
786                         label: "Disk Space",
787                         cmd:   []string{"df", "-m", "/", os.TempDir()},
788                 },
789                 {
790                         label: "Disk INodes",
791                         cmd:   []string{"df", "-i", "/", os.TempDir()},
792                 },
793         }
794
795         // Run commands with informational output to be logged.
796         for _, command := range commands {
797                 fmt.Fprintln(w, command.label)
798                 cmd := exec.Command(command.cmd[0], command.cmd[1:]...)
799                 cmd.Stdout = w
800                 cmd.Stderr = w
801                 if err := cmd.Run(); err != nil {
802                         err = fmt.Errorf("While running command %q: %v", command.cmd, err)
803                         fmt.Fprintln(w, err)
804                         return err
805                 }
806                 fmt.Fprintln(w, "")
807         }
808
809         err = w.Close()
810         if err != nil {
811                 return fmt.Errorf("While closing node-info logs: %v", err)
812         }
813         return nil
814 }
815
816 // LogContainerRecord gets and saves the raw JSON container record from the API server
817 func (runner *ContainerRunner) LogContainerRecord() error {
818         logged, err := runner.logAPIResponse("container", "containers", map[string]interface{}{"filters": [][]string{{"uuid", "=", runner.Container.UUID}}}, nil)
819         if !logged && err == nil {
820                 err = fmt.Errorf("error: no container record found for %s", runner.Container.UUID)
821         }
822         return err
823 }
824
825 // LogNodeRecord logs the current host's InstanceType config entry (or
826 // the arvados#node record, if running via crunch-dispatch-slurm).
827 func (runner *ContainerRunner) LogNodeRecord() error {
828         if it := os.Getenv("InstanceType"); it != "" {
829                 // Dispatched via arvados-dispatch-cloud. Save
830                 // InstanceType config fragment received from
831                 // dispatcher on stdin.
832                 w, err := runner.LogCollection.OpenFile("node.json", os.O_CREATE|os.O_WRONLY, 0666)
833                 if err != nil {
834                         return err
835                 }
836                 defer w.Close()
837                 _, err = io.WriteString(w, it)
838                 if err != nil {
839                         return err
840                 }
841                 return w.Close()
842         }
843         // Dispatched via crunch-dispatch-slurm. Look up
844         // apiserver's node record corresponding to
845         // $SLURMD_NODENAME.
846         hostname := os.Getenv("SLURMD_NODENAME")
847         if hostname == "" {
848                 hostname, _ = os.Hostname()
849         }
850         _, err := runner.logAPIResponse("node", "nodes", map[string]interface{}{"filters": [][]string{{"hostname", "=", hostname}}}, func(resp interface{}) {
851                 // The "info" field has admin-only info when
852                 // obtained with a privileged token, and
853                 // should not be logged.
854                 node, ok := resp.(map[string]interface{})
855                 if ok {
856                         delete(node, "info")
857                 }
858         })
859         return err
860 }
861
862 func (runner *ContainerRunner) logAPIResponse(label, path string, params map[string]interface{}, munge func(interface{})) (logged bool, err error) {
863         writer, err := runner.LogCollection.OpenFile(label+".json", os.O_CREATE|os.O_WRONLY, 0666)
864         if err != nil {
865                 return false, err
866         }
867         w := &ArvLogWriter{
868                 ArvClient:     runner.DispatcherArvClient,
869                 UUID:          runner.Container.UUID,
870                 loggingStream: label,
871                 writeCloser:   writer,
872         }
873
874         reader, err := runner.DispatcherArvClient.CallRaw("GET", path, "", "", arvadosclient.Dict(params))
875         if err != nil {
876                 return false, fmt.Errorf("error getting %s record: %v", label, err)
877         }
878         defer reader.Close()
879
880         dec := json.NewDecoder(reader)
881         dec.UseNumber()
882         var resp map[string]interface{}
883         if err = dec.Decode(&resp); err != nil {
884                 return false, fmt.Errorf("error decoding %s list response: %v", label, err)
885         }
886         items, ok := resp["items"].([]interface{})
887         if !ok {
888                 return false, fmt.Errorf("error decoding %s list response: no \"items\" key in API list response", label)
889         } else if len(items) < 1 {
890                 return false, nil
891         }
892         if munge != nil {
893                 munge(items[0])
894         }
895         // Re-encode it using indentation to improve readability
896         enc := json.NewEncoder(w)
897         enc.SetIndent("", "    ")
898         if err = enc.Encode(items[0]); err != nil {
899                 return false, fmt.Errorf("error logging %s record: %v", label, err)
900         }
901         err = w.Close()
902         if err != nil {
903                 return false, fmt.Errorf("error closing %s.json in log collection: %v", label, err)
904         }
905         return true, nil
906 }
907
908 func (runner *ContainerRunner) getStdoutFile(mntPath string) (*os.File, error) {
909         stdoutPath := mntPath[len(runner.Container.OutputPath):]
910         index := strings.LastIndex(stdoutPath, "/")
911         if index > 0 {
912                 subdirs := stdoutPath[:index]
913                 if subdirs != "" {
914                         st, err := os.Stat(runner.HostOutputDir)
915                         if err != nil {
916                                 return nil, fmt.Errorf("While Stat on temp dir: %v", err)
917                         }
918                         stdoutPath := filepath.Join(runner.HostOutputDir, subdirs)
919                         err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777)
920                         if err != nil {
921                                 return nil, fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err)
922                         }
923                 }
924         }
925         stdoutFile, err := os.Create(filepath.Join(runner.HostOutputDir, stdoutPath))
926         if err != nil {
927                 return nil, fmt.Errorf("While creating file %q: %v", stdoutPath, err)
928         }
929
930         return stdoutFile, nil
931 }
932
933 // CreateContainer creates the docker container.
934 func (runner *ContainerRunner) CreateContainer(imageID string, bindmounts map[string]bindmount) error {
935         var stdin io.ReadCloser = ioutil.NopCloser(bytes.NewReader(nil))
936         if mnt, ok := runner.Container.Mounts["stdin"]; ok {
937                 switch mnt.Kind {
938                 case "collection":
939                         var collID string
940                         if mnt.UUID != "" {
941                                 collID = mnt.UUID
942                         } else {
943                                 collID = mnt.PortableDataHash
944                         }
945                         path := runner.ArvMountPoint + "/by_id/" + collID + "/" + mnt.Path
946                         f, err := os.Open(path)
947                         if err != nil {
948                                 return err
949                         }
950                         stdin = f
951                 case "json":
952                         j, err := json.Marshal(mnt.Content)
953                         if err != nil {
954                                 return fmt.Errorf("error encoding stdin json data: %v", err)
955                         }
956                         stdin = ioutil.NopCloser(bytes.NewReader(j))
957                 default:
958                         return fmt.Errorf("stdin mount has unsupported kind %q", mnt.Kind)
959                 }
960         }
961
962         var stdout, stderr io.WriteCloser
963         if mnt, ok := runner.Container.Mounts["stdout"]; ok {
964                 f, err := runner.getStdoutFile(mnt.Path)
965                 if err != nil {
966                         return err
967                 }
968                 stdout = f
969         } else if w, err := runner.NewLogWriter("stdout"); err != nil {
970                 return err
971         } else {
972                 stdout = NewThrottledLogger(w)
973         }
974
975         if mnt, ok := runner.Container.Mounts["stderr"]; ok {
976                 f, err := runner.getStdoutFile(mnt.Path)
977                 if err != nil {
978                         return err
979                 }
980                 stderr = f
981         } else if w, err := runner.NewLogWriter("stderr"); err != nil {
982                 return err
983         } else {
984                 stderr = NewThrottledLogger(w)
985         }
986
987         env := runner.Container.Environment
988         enableNetwork := runner.enableNetwork == "always"
989         if runner.Container.RuntimeConstraints.API {
990                 enableNetwork = true
991                 tok, err := runner.ContainerToken()
992                 if err != nil {
993                         return err
994                 }
995                 env = map[string]string{}
996                 for k, v := range runner.Container.Environment {
997                         env[k] = v
998                 }
999                 env["ARVADOS_API_TOKEN"] = tok
1000                 env["ARVADOS_API_HOST"] = os.Getenv("ARVADOS_API_HOST")
1001                 env["ARVADOS_API_HOST_INSECURE"] = os.Getenv("ARVADOS_API_HOST_INSECURE")
1002         }
1003         workdir := runner.Container.Cwd
1004         if workdir == "." {
1005                 // both "" and "." mean default
1006                 workdir = ""
1007         }
1008         ram := runner.Container.RuntimeConstraints.RAM
1009         if !runner.enableMemoryLimit {
1010                 ram = 0
1011         }
1012         runner.executorStdin = stdin
1013         runner.executorStdout = stdout
1014         runner.executorStderr = stderr
1015
1016         if runner.Container.RuntimeConstraints.CUDA.DeviceCount > 0 {
1017                 nvidiaModprobe(runner.CrunchLog)
1018         }
1019
1020         return runner.executor.Create(containerSpec{
1021                 Image:           imageID,
1022                 VCPUs:           runner.Container.RuntimeConstraints.VCPUs,
1023                 RAM:             ram,
1024                 WorkingDir:      workdir,
1025                 Env:             env,
1026                 BindMounts:      bindmounts,
1027                 Command:         runner.Container.Command,
1028                 EnableNetwork:   enableNetwork,
1029                 CUDADeviceCount: runner.Container.RuntimeConstraints.CUDA.DeviceCount,
1030                 NetworkMode:     runner.networkMode,
1031                 CgroupParent:    runner.setCgroupParent,
1032                 Stdin:           stdin,
1033                 Stdout:          stdout,
1034                 Stderr:          stderr,
1035         })
1036 }
1037
1038 // StartContainer starts the docker container created by CreateContainer.
1039 func (runner *ContainerRunner) StartContainer() error {
1040         runner.CrunchLog.Printf("Starting container")
1041         runner.cStateLock.Lock()
1042         defer runner.cStateLock.Unlock()
1043         if runner.cCancelled {
1044                 return ErrCancelled
1045         }
1046         err := runner.executor.Start()
1047         if err != nil {
1048                 var advice string
1049                 if m, e := regexp.MatchString("(?ms).*(exec|System error).*(no such file or directory|file not found).*", err.Error()); m && e == nil {
1050                         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])
1051                 }
1052                 return fmt.Errorf("could not start container: %v%s", err, advice)
1053         }
1054         return nil
1055 }
1056
1057 // WaitFinish waits for the container to terminate, capture the exit code, and
1058 // close the stdout/stderr logging.
1059 func (runner *ContainerRunner) WaitFinish() error {
1060         runner.CrunchLog.Print("Waiting for container to finish")
1061         var timeout <-chan time.Time
1062         if s := runner.Container.SchedulingParameters.MaxRunTime; s > 0 {
1063                 timeout = time.After(time.Duration(s) * time.Second)
1064         }
1065         ctx, cancel := context.WithCancel(context.Background())
1066         defer cancel()
1067         go func() {
1068                 select {
1069                 case <-timeout:
1070                         runner.CrunchLog.Printf("maximum run time exceeded. Stopping container.")
1071                         runner.stop(nil)
1072                 case <-runner.ArvMountExit:
1073                         runner.CrunchLog.Printf("arv-mount exited while container is still running. Stopping container.")
1074                         runner.stop(nil)
1075                 case <-ctx.Done():
1076                 }
1077         }()
1078         exitcode, err := runner.executor.Wait(ctx)
1079         if err != nil {
1080                 runner.checkBrokenNode(err)
1081                 return err
1082         }
1083         runner.ExitCode = &exitcode
1084
1085         extra := ""
1086         if exitcode&0x80 != 0 {
1087                 // Convert raw exit status (0x80 + signal number) to a
1088                 // string to log after the code, like " (signal 101)"
1089                 // or " (signal 9, killed)"
1090                 sig := syscall.WaitStatus(exitcode).Signal()
1091                 if name := unix.SignalName(sig); name != "" {
1092                         extra = fmt.Sprintf(" (signal %d, %s)", sig, name)
1093                 } else {
1094                         extra = fmt.Sprintf(" (signal %d)", sig)
1095                 }
1096         }
1097         runner.CrunchLog.Printf("Container exited with status code %d%s", exitcode, extra)
1098         err = runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1099                 "container": arvadosclient.Dict{"exit_code": exitcode},
1100         }, nil)
1101         if err != nil {
1102                 runner.CrunchLog.Printf("ignoring error updating exit_code: %s", err)
1103         }
1104
1105         var returnErr error
1106         if err = runner.executorStdin.Close(); err != nil {
1107                 err = fmt.Errorf("error closing container stdin: %s", err)
1108                 runner.CrunchLog.Printf("%s", err)
1109                 returnErr = err
1110         }
1111         if err = runner.executorStdout.Close(); err != nil {
1112                 err = fmt.Errorf("error closing container stdout: %s", err)
1113                 runner.CrunchLog.Printf("%s", err)
1114                 if returnErr == nil {
1115                         returnErr = err
1116                 }
1117         }
1118         if err = runner.executorStderr.Close(); err != nil {
1119                 err = fmt.Errorf("error closing container stderr: %s", err)
1120                 runner.CrunchLog.Printf("%s", err)
1121                 if returnErr == nil {
1122                         returnErr = err
1123                 }
1124         }
1125
1126         if runner.statReporter != nil {
1127                 runner.statReporter.Stop()
1128                 err = runner.statLogger.Close()
1129                 if err != nil {
1130                         runner.CrunchLog.Printf("error closing crunchstat logs: %v", err)
1131                 }
1132         }
1133         return returnErr
1134 }
1135
1136 func (runner *ContainerRunner) updateLogs() {
1137         ticker := time.NewTicker(crunchLogUpdatePeriod / 360)
1138         defer ticker.Stop()
1139
1140         sigusr1 := make(chan os.Signal, 1)
1141         signal.Notify(sigusr1, syscall.SIGUSR1)
1142         defer signal.Stop(sigusr1)
1143
1144         saveAtTime := time.Now().Add(crunchLogUpdatePeriod)
1145         saveAtSize := crunchLogUpdateSize
1146         var savedSize int64
1147         for {
1148                 select {
1149                 case <-ticker.C:
1150                 case <-sigusr1:
1151                         saveAtTime = time.Now()
1152                 }
1153                 runner.logMtx.Lock()
1154                 done := runner.LogsPDH != nil
1155                 runner.logMtx.Unlock()
1156                 if done {
1157                         return
1158                 }
1159                 size := runner.LogCollection.Size()
1160                 if size == savedSize || (time.Now().Before(saveAtTime) && size < saveAtSize) {
1161                         continue
1162                 }
1163                 saveAtTime = time.Now().Add(crunchLogUpdatePeriod)
1164                 saveAtSize = runner.LogCollection.Size() + crunchLogUpdateSize
1165                 saved, err := runner.saveLogCollection(false)
1166                 if err != nil {
1167                         runner.CrunchLog.Printf("error updating log collection: %s", err)
1168                         continue
1169                 }
1170
1171                 err = runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1172                         "container": arvadosclient.Dict{"log": saved.PortableDataHash},
1173                 }, nil)
1174                 if err != nil {
1175                         runner.CrunchLog.Printf("error updating container log to %s: %s", saved.PortableDataHash, err)
1176                         continue
1177                 }
1178
1179                 savedSize = size
1180         }
1181 }
1182
1183 func (runner *ContainerRunner) reportArvMountWarning(pattern, text string) {
1184         var updated arvados.Container
1185         err := runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1186                 "container": arvadosclient.Dict{
1187                         "runtime_status": arvadosclient.Dict{
1188                                 "warning":       "arv-mount: " + pattern,
1189                                 "warningDetail": text,
1190                         },
1191                 },
1192         }, &updated)
1193         if err != nil {
1194                 runner.CrunchLog.Printf("error updating container runtime_status: %s", err)
1195         }
1196 }
1197
1198 // CaptureOutput saves data from the container's output directory if
1199 // needed, and updates the container output accordingly.
1200 func (runner *ContainerRunner) CaptureOutput(bindmounts map[string]bindmount) error {
1201         if runner.Container.RuntimeConstraints.API {
1202                 // Output may have been set directly by the container, so
1203                 // refresh the container record to check.
1204                 err := runner.DispatcherArvClient.Get("containers", runner.Container.UUID,
1205                         nil, &runner.Container)
1206                 if err != nil {
1207                         return err
1208                 }
1209                 if runner.Container.Output != "" {
1210                         // Container output is already set.
1211                         runner.OutputPDH = &runner.Container.Output
1212                         return nil
1213                 }
1214         }
1215
1216         txt, err := (&copier{
1217                 client:        runner.containerClient,
1218                 arvClient:     runner.ContainerArvClient,
1219                 keepClient:    runner.ContainerKeepClient,
1220                 hostOutputDir: runner.HostOutputDir,
1221                 ctrOutputDir:  runner.Container.OutputPath,
1222                 bindmounts:    bindmounts,
1223                 mounts:        runner.Container.Mounts,
1224                 secretMounts:  runner.SecretMounts,
1225                 logger:        runner.CrunchLog,
1226         }).Copy()
1227         if err != nil {
1228                 return err
1229         }
1230         if n := len(regexp.MustCompile(` [0-9a-f]+\+\S*\+R`).FindAllStringIndex(txt, -1)); n > 0 {
1231                 runner.CrunchLog.Printf("Copying %d data blocks from remote input collections...", n)
1232                 fs, err := (&arvados.Collection{ManifestText: txt}).FileSystem(runner.containerClient, runner.ContainerKeepClient)
1233                 if err != nil {
1234                         return err
1235                 }
1236                 txt, err = fs.MarshalManifest(".")
1237                 if err != nil {
1238                         return err
1239                 }
1240         }
1241         var resp arvados.Collection
1242         err = runner.ContainerArvClient.Create("collections", arvadosclient.Dict{
1243                 "ensure_unique_name": true,
1244                 "collection": arvadosclient.Dict{
1245                         "is_trashed":    true,
1246                         "name":          "output for " + runner.Container.UUID,
1247                         "manifest_text": txt,
1248                 },
1249         }, &resp)
1250         if err != nil {
1251                 return fmt.Errorf("error creating output collection: %v", err)
1252         }
1253         runner.OutputPDH = &resp.PortableDataHash
1254         return nil
1255 }
1256
1257 func (runner *ContainerRunner) CleanupDirs() {
1258         if runner.ArvMount != nil {
1259                 var delay int64 = 8
1260                 umount := exec.Command("arv-mount", fmt.Sprintf("--unmount-timeout=%d", delay), "--unmount", runner.ArvMountPoint)
1261                 umount.Stdout = runner.CrunchLog
1262                 umount.Stderr = runner.CrunchLog
1263                 runner.CrunchLog.Printf("Running %v", umount.Args)
1264                 umnterr := umount.Start()
1265
1266                 if umnterr != nil {
1267                         runner.CrunchLog.Printf("Error unmounting: %v", umnterr)
1268                         runner.ArvMount.Process.Kill()
1269                 } else {
1270                         // If arv-mount --unmount gets stuck for any reason, we
1271                         // don't want to wait for it forever.  Do Wait() in a goroutine
1272                         // so it doesn't block crunch-run.
1273                         umountExit := make(chan error)
1274                         go func() {
1275                                 mnterr := umount.Wait()
1276                                 if mnterr != nil {
1277                                         runner.CrunchLog.Printf("Error unmounting: %v", mnterr)
1278                                 }
1279                                 umountExit <- mnterr
1280                         }()
1281
1282                         for again := true; again; {
1283                                 again = false
1284                                 select {
1285                                 case <-umountExit:
1286                                         umount = nil
1287                                         again = true
1288                                 case <-runner.ArvMountExit:
1289                                         break
1290                                 case <-time.After(time.Duration((delay + 1) * int64(time.Second))):
1291                                         runner.CrunchLog.Printf("Timed out waiting for unmount")
1292                                         if umount != nil {
1293                                                 umount.Process.Kill()
1294                                         }
1295                                         runner.ArvMount.Process.Kill()
1296                                 }
1297                         }
1298                 }
1299                 runner.ArvMount = nil
1300         }
1301
1302         if runner.ArvMountPoint != "" {
1303                 if rmerr := os.Remove(runner.ArvMountPoint); rmerr != nil {
1304                         runner.CrunchLog.Printf("While cleaning up arv-mount directory %s: %v", runner.ArvMountPoint, rmerr)
1305                 }
1306                 runner.ArvMountPoint = ""
1307         }
1308
1309         if rmerr := os.RemoveAll(runner.parentTemp); rmerr != nil {
1310                 runner.CrunchLog.Printf("While cleaning up temporary directory %s: %v", runner.parentTemp, rmerr)
1311         }
1312 }
1313
1314 // CommitLogs posts the collection containing the final container logs.
1315 func (runner *ContainerRunner) CommitLogs() error {
1316         func() {
1317                 // Hold cStateLock to prevent races on CrunchLog (e.g., stop()).
1318                 runner.cStateLock.Lock()
1319                 defer runner.cStateLock.Unlock()
1320
1321                 runner.CrunchLog.Print(runner.finalState)
1322
1323                 if runner.arvMountLog != nil {
1324                         runner.arvMountLog.Close()
1325                 }
1326                 runner.CrunchLog.Close()
1327
1328                 // Closing CrunchLog above allows them to be committed to Keep at this
1329                 // point, but re-open crunch log with ArvClient in case there are any
1330                 // other further errors (such as failing to write the log to Keep!)
1331                 // while shutting down
1332                 runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{
1333                         ArvClient:     runner.DispatcherArvClient,
1334                         UUID:          runner.Container.UUID,
1335                         loggingStream: "crunch-run",
1336                         writeCloser:   nil,
1337                 })
1338                 runner.CrunchLog.Immediate = log.New(os.Stderr, runner.Container.UUID+" ", 0)
1339         }()
1340
1341         if runner.keepstoreLogger != nil {
1342                 // Flush any buffered logs from our local keepstore
1343                 // process.  Discard anything logged after this point
1344                 // -- it won't end up in the log collection, so
1345                 // there's no point writing it to the collectionfs.
1346                 runner.keepstoreLogbuf.SetWriter(io.Discard)
1347                 runner.keepstoreLogger.Close()
1348                 runner.keepstoreLogger = nil
1349         }
1350
1351         if runner.LogsPDH != nil {
1352                 // If we have already assigned something to LogsPDH,
1353                 // we must be closing the re-opened log, which won't
1354                 // end up getting attached to the container record and
1355                 // therefore doesn't need to be saved as a collection
1356                 // -- it exists only to send logs to other channels.
1357                 return nil
1358         }
1359
1360         saved, err := runner.saveLogCollection(true)
1361         if err != nil {
1362                 return fmt.Errorf("error saving log collection: %s", err)
1363         }
1364         runner.logMtx.Lock()
1365         defer runner.logMtx.Unlock()
1366         runner.LogsPDH = &saved.PortableDataHash
1367         return nil
1368 }
1369
1370 func (runner *ContainerRunner) saveLogCollection(final bool) (response arvados.Collection, err error) {
1371         runner.logMtx.Lock()
1372         defer runner.logMtx.Unlock()
1373         if runner.LogsPDH != nil {
1374                 // Already finalized.
1375                 return
1376         }
1377         updates := arvadosclient.Dict{
1378                 "name": "logs for " + runner.Container.UUID,
1379         }
1380         mt, err1 := runner.LogCollection.MarshalManifest(".")
1381         if err1 == nil {
1382                 // Only send updated manifest text if there was no
1383                 // error.
1384                 updates["manifest_text"] = mt
1385         }
1386
1387         // Even if flushing the manifest had an error, we still want
1388         // to update the log record, if possible, to push the trash_at
1389         // and delete_at times into the future.  Details on bug
1390         // #17293.
1391         if final {
1392                 updates["is_trashed"] = true
1393         } else {
1394                 exp := time.Now().Add(crunchLogUpdatePeriod * 24)
1395                 updates["trash_at"] = exp
1396                 updates["delete_at"] = exp
1397         }
1398         reqBody := arvadosclient.Dict{"collection": updates}
1399         var err2 error
1400         if runner.logUUID == "" {
1401                 reqBody["ensure_unique_name"] = true
1402                 err2 = runner.DispatcherArvClient.Create("collections", reqBody, &response)
1403         } else {
1404                 err2 = runner.DispatcherArvClient.Update("collections", runner.logUUID, reqBody, &response)
1405         }
1406         if err2 == nil {
1407                 runner.logUUID = response.UUID
1408         }
1409
1410         if err1 != nil || err2 != nil {
1411                 err = fmt.Errorf("error recording logs: %q, %q", err1, err2)
1412         }
1413         return
1414 }
1415
1416 // UpdateContainerRunning updates the container state to "Running"
1417 func (runner *ContainerRunner) UpdateContainerRunning() error {
1418         runner.cStateLock.Lock()
1419         defer runner.cStateLock.Unlock()
1420         if runner.cCancelled {
1421                 return ErrCancelled
1422         }
1423         return runner.DispatcherArvClient.Update("containers", runner.Container.UUID,
1424                 arvadosclient.Dict{"container": arvadosclient.Dict{"state": "Running", "gateway_address": runner.gateway.Address}}, nil)
1425 }
1426
1427 // ContainerToken returns the api_token the container (and any
1428 // arv-mount processes) are allowed to use.
1429 func (runner *ContainerRunner) ContainerToken() (string, error) {
1430         if runner.token != "" {
1431                 return runner.token, nil
1432         }
1433
1434         var auth arvados.APIClientAuthorization
1435         err := runner.DispatcherArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth)
1436         if err != nil {
1437                 return "", err
1438         }
1439         runner.token = fmt.Sprintf("v2/%s/%s/%s", auth.UUID, auth.APIToken, runner.Container.UUID)
1440         return runner.token, nil
1441 }
1442
1443 // UpdateContainerFinal updates the container record state on API
1444 // server to "Complete" or "Cancelled"
1445 func (runner *ContainerRunner) UpdateContainerFinal() error {
1446         update := arvadosclient.Dict{}
1447         update["state"] = runner.finalState
1448         if runner.LogsPDH != nil {
1449                 update["log"] = *runner.LogsPDH
1450         }
1451         if runner.ExitCode != nil {
1452                 update["exit_code"] = *runner.ExitCode
1453         } else {
1454                 update["exit_code"] = nil
1455         }
1456         if runner.finalState == "Complete" && runner.OutputPDH != nil {
1457                 update["output"] = *runner.OutputPDH
1458         }
1459         return runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{"container": update}, nil)
1460 }
1461
1462 // IsCancelled returns the value of Cancelled, with goroutine safety.
1463 func (runner *ContainerRunner) IsCancelled() bool {
1464         runner.cStateLock.Lock()
1465         defer runner.cStateLock.Unlock()
1466         return runner.cCancelled
1467 }
1468
1469 // NewArvLogWriter creates an ArvLogWriter
1470 func (runner *ContainerRunner) NewArvLogWriter(name string) (io.WriteCloser, error) {
1471         writer, err := runner.LogCollection.OpenFile(name+".txt", os.O_CREATE|os.O_WRONLY, 0666)
1472         if err != nil {
1473                 return nil, err
1474         }
1475         return &ArvLogWriter{
1476                 ArvClient:     runner.DispatcherArvClient,
1477                 UUID:          runner.Container.UUID,
1478                 loggingStream: name,
1479                 writeCloser:   writer,
1480         }, nil
1481 }
1482
1483 // Run the full container lifecycle.
1484 func (runner *ContainerRunner) Run() (err error) {
1485         runner.CrunchLog.Printf("crunch-run %s started", cmd.Version.String())
1486         runner.CrunchLog.Printf("%s", currentUserAndGroups())
1487         v, _ := exec.Command("arv-mount", "--version").CombinedOutput()
1488         runner.CrunchLog.Printf("Using FUSE mount: %s", v)
1489         runner.CrunchLog.Printf("Using container runtime: %s", runner.executor.Runtime())
1490         runner.CrunchLog.Printf("Executing container: %s", runner.Container.UUID)
1491
1492         hostname, hosterr := os.Hostname()
1493         if hosterr != nil {
1494                 runner.CrunchLog.Printf("Error getting hostname '%v'", hosterr)
1495         } else {
1496                 runner.CrunchLog.Printf("Executing on host '%s'", hostname)
1497         }
1498
1499         runner.finalState = "Queued"
1500
1501         defer func() {
1502                 runner.CleanupDirs()
1503
1504                 runner.CrunchLog.Printf("crunch-run finished")
1505                 runner.CrunchLog.Close()
1506         }()
1507
1508         err = runner.fetchContainerRecord()
1509         if err != nil {
1510                 return
1511         }
1512         if runner.Container.State != "Locked" {
1513                 return fmt.Errorf("dispatch error detected: container %q has state %q", runner.Container.UUID, runner.Container.State)
1514         }
1515
1516         var bindmounts map[string]bindmount
1517         defer func() {
1518                 // checkErr prints e (unless it's nil) and sets err to
1519                 // e (unless err is already non-nil). Thus, if err
1520                 // hasn't already been assigned when Run() returns,
1521                 // this cleanup func will cause Run() to return the
1522                 // first non-nil error that is passed to checkErr().
1523                 checkErr := func(errorIn string, e error) {
1524                         if e == nil {
1525                                 return
1526                         }
1527                         runner.CrunchLog.Printf("error in %s: %v", errorIn, e)
1528                         if err == nil {
1529                                 err = e
1530                         }
1531                         if runner.finalState == "Complete" {
1532                                 // There was an error in the finalization.
1533                                 runner.finalState = "Cancelled"
1534                         }
1535                 }
1536
1537                 // Log the error encountered in Run(), if any
1538                 checkErr("Run", err)
1539
1540                 if runner.finalState == "Queued" {
1541                         runner.UpdateContainerFinal()
1542                         return
1543                 }
1544
1545                 if runner.IsCancelled() {
1546                         runner.finalState = "Cancelled"
1547                         // but don't return yet -- we still want to
1548                         // capture partial output and write logs
1549                 }
1550
1551                 if bindmounts != nil {
1552                         checkErr("CaptureOutput", runner.CaptureOutput(bindmounts))
1553                 }
1554                 checkErr("stopHoststat", runner.stopHoststat())
1555                 checkErr("CommitLogs", runner.CommitLogs())
1556                 runner.CleanupDirs()
1557                 checkErr("UpdateContainerFinal", runner.UpdateContainerFinal())
1558         }()
1559
1560         runner.setupSignals()
1561         err = runner.startHoststat()
1562         if err != nil {
1563                 return
1564         }
1565
1566         // set up FUSE mount and binds
1567         bindmounts, err = runner.SetupMounts()
1568         if err != nil {
1569                 runner.finalState = "Cancelled"
1570                 err = fmt.Errorf("While setting up mounts: %v", err)
1571                 return
1572         }
1573
1574         // check for and/or load image
1575         imageID, err := runner.LoadImage()
1576         if err != nil {
1577                 if !runner.checkBrokenNode(err) {
1578                         // Failed to load image but not due to a "broken node"
1579                         // condition, probably user error.
1580                         runner.finalState = "Cancelled"
1581                 }
1582                 err = fmt.Errorf("While loading container image: %v", err)
1583                 return
1584         }
1585
1586         err = runner.CreateContainer(imageID, bindmounts)
1587         if err != nil {
1588                 return
1589         }
1590         err = runner.LogHostInfo()
1591         if err != nil {
1592                 return
1593         }
1594         err = runner.LogNodeRecord()
1595         if err != nil {
1596                 return
1597         }
1598         err = runner.LogContainerRecord()
1599         if err != nil {
1600                 return
1601         }
1602
1603         if runner.IsCancelled() {
1604                 return
1605         }
1606
1607         err = runner.UpdateContainerRunning()
1608         if err != nil {
1609                 return
1610         }
1611         runner.finalState = "Cancelled"
1612
1613         err = runner.startCrunchstat()
1614         if err != nil {
1615                 return
1616         }
1617
1618         err = runner.StartContainer()
1619         if err != nil {
1620                 runner.checkBrokenNode(err)
1621                 return
1622         }
1623
1624         err = runner.WaitFinish()
1625         if err == nil && !runner.IsCancelled() {
1626                 runner.finalState = "Complete"
1627         }
1628         return
1629 }
1630
1631 // Fetch the current container record (uuid = runner.Container.UUID)
1632 // into runner.Container.
1633 func (runner *ContainerRunner) fetchContainerRecord() error {
1634         reader, err := runner.DispatcherArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
1635         if err != nil {
1636                 return fmt.Errorf("error fetching container record: %v", err)
1637         }
1638         defer reader.Close()
1639
1640         dec := json.NewDecoder(reader)
1641         dec.UseNumber()
1642         err = dec.Decode(&runner.Container)
1643         if err != nil {
1644                 return fmt.Errorf("error decoding container record: %v", err)
1645         }
1646
1647         var sm struct {
1648                 SecretMounts map[string]arvados.Mount `json:"secret_mounts"`
1649         }
1650
1651         containerToken, err := runner.ContainerToken()
1652         if err != nil {
1653                 return fmt.Errorf("error getting container token: %v", err)
1654         }
1655
1656         runner.ContainerArvClient, runner.ContainerKeepClient,
1657                 runner.containerClient, err = runner.MkArvClient(containerToken)
1658         if err != nil {
1659                 return fmt.Errorf("error creating container API client: %v", err)
1660         }
1661
1662         runner.ContainerKeepClient.SetStorageClasses(runner.Container.OutputStorageClasses)
1663         runner.DispatcherKeepClient.SetStorageClasses(runner.Container.OutputStorageClasses)
1664
1665         err = runner.ContainerArvClient.Call("GET", "containers", runner.Container.UUID, "secret_mounts", nil, &sm)
1666         if err != nil {
1667                 if apierr, ok := err.(arvadosclient.APIServerError); !ok || apierr.HttpStatusCode != 404 {
1668                         return fmt.Errorf("error fetching secret_mounts: %v", err)
1669                 }
1670                 // ok && apierr.HttpStatusCode == 404, which means
1671                 // secret_mounts isn't supported by this API server.
1672         }
1673         runner.SecretMounts = sm.SecretMounts
1674
1675         return nil
1676 }
1677
1678 // NewContainerRunner creates a new container runner.
1679 func NewContainerRunner(dispatcherClient *arvados.Client,
1680         dispatcherArvClient IArvadosClient,
1681         dispatcherKeepClient IKeepClient,
1682         containerUUID string) (*ContainerRunner, error) {
1683
1684         cr := &ContainerRunner{
1685                 dispatcherClient:     dispatcherClient,
1686                 DispatcherArvClient:  dispatcherArvClient,
1687                 DispatcherKeepClient: dispatcherKeepClient,
1688         }
1689         cr.NewLogWriter = cr.NewArvLogWriter
1690         cr.RunArvMount = cr.ArvMountCmd
1691         cr.MkTempDir = ioutil.TempDir
1692         cr.MkArvClient = func(token string) (IArvadosClient, IKeepClient, *arvados.Client, error) {
1693                 cl, err := arvadosclient.MakeArvadosClient()
1694                 if err != nil {
1695                         return nil, nil, nil, err
1696                 }
1697                 cl.ApiToken = token
1698                 kc, err := keepclient.MakeKeepClient(cl)
1699                 if err != nil {
1700                         return nil, nil, nil, err
1701                 }
1702                 c2 := arvados.NewClientFromEnv()
1703                 c2.AuthToken = token
1704                 return cl, kc, c2, nil
1705         }
1706         var err error
1707         cr.LogCollection, err = (&arvados.Collection{}).FileSystem(cr.dispatcherClient, cr.DispatcherKeepClient)
1708         if err != nil {
1709                 return nil, err
1710         }
1711         cr.Container.UUID = containerUUID
1712         w, err := cr.NewLogWriter("crunch-run")
1713         if err != nil {
1714                 return nil, err
1715         }
1716         cr.CrunchLog = NewThrottledLogger(w)
1717         cr.CrunchLog.Immediate = log.New(os.Stderr, containerUUID+" ", 0)
1718
1719         loadLogThrottleParams(dispatcherArvClient)
1720         go cr.updateLogs()
1721
1722         return cr, nil
1723 }
1724
1725 func (command) RunCommand(prog string, args []string, stdin io.Reader, stdout, stderr io.Writer) int {
1726         log := log.New(stderr, "", 0)
1727         flags := flag.NewFlagSet(prog, flag.ContinueOnError)
1728         statInterval := flags.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting")
1729         cgroupRoot := flags.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree")
1730         cgroupParent := flags.String("cgroup-parent", "docker", "name of container's parent cgroup (ignored if -cgroup-parent-subsystem is used)")
1731         cgroupParentSubsystem := flags.String("cgroup-parent-subsystem", "", "use current cgroup for given subsystem as parent cgroup for container")
1732         caCertsPath := flags.String("ca-certs", "", "Path to TLS root certificates")
1733         detach := flags.Bool("detach", false, "Detach from parent process and run in the background")
1734         stdinConfig := flags.Bool("stdin-config", false, "Load config and environment variables from JSON message on stdin")
1735         configFile := flags.String("config", arvados.DefaultConfigFile, "filename of cluster config file to try loading if -stdin-config=false (default is $ARVADOS_CONFIG)")
1736         sleep := flags.Duration("sleep", 0, "Delay before starting (testing use only)")
1737         kill := flags.Int("kill", -1, "Send signal to an existing crunch-run process for given UUID")
1738         list := flags.Bool("list", false, "List UUIDs of existing crunch-run processes")
1739         enableMemoryLimit := flags.Bool("enable-memory-limit", true, "tell container runtime to limit container's memory usage")
1740         enableNetwork := flags.String("container-enable-networking", "default", "enable networking \"always\" (for all containers) or \"default\" (for containers that request it)")
1741         networkMode := flags.String("container-network-mode", "default", `Docker network mode for container (use any argument valid for docker --net)`)
1742         memprofile := flags.String("memprofile", "", "write memory profile to `file` after running container")
1743         runtimeEngine := flags.String("runtime-engine", "docker", "container runtime: docker or singularity")
1744         brokenNodeHook := flags.String("broken-node-hook", "", "script to run if node is detected to be broken (for example, Docker daemon is not running)")
1745         flags.Duration("check-containerd", 0, "Ignored. Exists for compatibility with older versions.")
1746
1747         ignoreDetachFlag := false
1748         if len(args) > 0 && args[0] == "-no-detach" {
1749                 // This process was invoked by a parent process, which
1750                 // has passed along its own arguments, including
1751                 // -detach, after the leading -no-detach flag.  Strip
1752                 // the leading -no-detach flag (it's not recognized by
1753                 // flags.Parse()) and ignore the -detach flag that
1754                 // comes later.
1755                 args = args[1:]
1756                 ignoreDetachFlag = true
1757         }
1758
1759         if ok, code := cmd.ParseFlags(flags, prog, args, "container-uuid", stderr); !ok {
1760                 return code
1761         } else if !*list && flags.NArg() != 1 {
1762                 fmt.Fprintf(stderr, "missing required argument: container-uuid (try -help)\n")
1763                 return 2
1764         }
1765
1766         containerUUID := flags.Arg(0)
1767
1768         switch {
1769         case *detach && !ignoreDetachFlag:
1770                 return Detach(containerUUID, prog, args, os.Stdin, os.Stdout, os.Stderr)
1771         case *kill >= 0:
1772                 return KillProcess(containerUUID, syscall.Signal(*kill), os.Stdout, os.Stderr)
1773         case *list:
1774                 return ListProcesses(os.Stdout, os.Stderr)
1775         }
1776
1777         if len(containerUUID) != 27 {
1778                 log.Printf("usage: %s [options] UUID", prog)
1779                 return 1
1780         }
1781
1782         var keepstoreLogbuf bufThenWrite
1783         var conf ConfigData
1784         if *stdinConfig {
1785                 err := json.NewDecoder(stdin).Decode(&conf)
1786                 if err != nil {
1787                         log.Printf("decode stdin: %s", err)
1788                         return 1
1789                 }
1790                 for k, v := range conf.Env {
1791                         err = os.Setenv(k, v)
1792                         if err != nil {
1793                                 log.Printf("setenv(%q): %s", k, err)
1794                                 return 1
1795                         }
1796                 }
1797                 if conf.Cluster != nil {
1798                         // ClusterID is missing from the JSON
1799                         // representation, but we need it to generate
1800                         // a valid config file for keepstore, so we
1801                         // fill it using the container UUID prefix.
1802                         conf.Cluster.ClusterID = containerUUID[:5]
1803                 }
1804         } else {
1805                 conf = hpcConfData(containerUUID, *configFile, io.MultiWriter(&keepstoreLogbuf, stderr))
1806         }
1807
1808         log.Printf("crunch-run %s started", cmd.Version.String())
1809         time.Sleep(*sleep)
1810
1811         if *caCertsPath != "" {
1812                 arvadosclient.CertFiles = []string{*caCertsPath}
1813         }
1814
1815         keepstore, err := startLocalKeepstore(conf, io.MultiWriter(&keepstoreLogbuf, stderr))
1816         if err != nil {
1817                 log.Print(err)
1818                 return 1
1819         }
1820         if keepstore != nil {
1821                 defer keepstore.Process.Kill()
1822         }
1823
1824         api, err := arvadosclient.MakeArvadosClient()
1825         if err != nil {
1826                 log.Printf("%s: %v", containerUUID, err)
1827                 return 1
1828         }
1829         api.Retries = 8
1830
1831         kc, err := keepclient.MakeKeepClient(api)
1832         if err != nil {
1833                 log.Printf("%s: %v", containerUUID, err)
1834                 return 1
1835         }
1836         kc.BlockCache = &keepclient.BlockCache{MaxBlocks: 2}
1837         kc.Retries = 4
1838
1839         cr, err := NewContainerRunner(arvados.NewClientFromEnv(), api, kc, containerUUID)
1840         if err != nil {
1841                 log.Print(err)
1842                 return 1
1843         }
1844
1845         if keepstore == nil {
1846                 // Log explanation (if any) for why we're not running
1847                 // a local keepstore.
1848                 var buf bytes.Buffer
1849                 keepstoreLogbuf.SetWriter(&buf)
1850                 if buf.Len() > 0 {
1851                         cr.CrunchLog.Printf("%s", strings.TrimSpace(buf.String()))
1852                 }
1853         } else if logWhat := conf.Cluster.Containers.LocalKeepLogsToContainerLog; logWhat == "none" {
1854                 cr.CrunchLog.Printf("using local keepstore process (pid %d) at %s", keepstore.Process.Pid, os.Getenv("ARVADOS_KEEP_SERVICES"))
1855                 keepstoreLogbuf.SetWriter(io.Discard)
1856         } else {
1857                 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"))
1858                 logwriter, err := cr.NewLogWriter("keepstore")
1859                 if err != nil {
1860                         log.Print(err)
1861                         return 1
1862                 }
1863                 cr.keepstoreLogger = NewThrottledLogger(logwriter)
1864
1865                 var writer io.WriteCloser = cr.keepstoreLogger
1866                 if logWhat == "errors" {
1867                         writer = &filterKeepstoreErrorsOnly{WriteCloser: writer}
1868                 } else if logWhat != "all" {
1869                         // should have been caught earlier by
1870                         // dispatcher's config loader
1871                         log.Printf("invalid value for Containers.LocalKeepLogsToContainerLog: %q", logWhat)
1872                         return 1
1873                 }
1874                 err = keepstoreLogbuf.SetWriter(writer)
1875                 if err != nil {
1876                         log.Print(err)
1877                         return 1
1878                 }
1879                 cr.keepstoreLogbuf = &keepstoreLogbuf
1880         }
1881
1882         switch *runtimeEngine {
1883         case "docker":
1884                 cr.executor, err = newDockerExecutor(containerUUID, cr.CrunchLog.Printf, cr.containerWatchdogInterval)
1885         case "singularity":
1886                 cr.executor, err = newSingularityExecutor(cr.CrunchLog.Printf)
1887         default:
1888                 cr.CrunchLog.Printf("%s: unsupported RuntimeEngine %q", containerUUID, *runtimeEngine)
1889                 cr.CrunchLog.Close()
1890                 return 1
1891         }
1892         if err != nil {
1893                 cr.CrunchLog.Printf("%s: %v", containerUUID, err)
1894                 cr.checkBrokenNode(err)
1895                 cr.CrunchLog.Close()
1896                 return 1
1897         }
1898         defer cr.executor.Close()
1899
1900         cr.brokenNodeHook = *brokenNodeHook
1901
1902         gwAuthSecret := os.Getenv("GatewayAuthSecret")
1903         os.Unsetenv("GatewayAuthSecret")
1904         if gwAuthSecret == "" {
1905                 // not safe to run a gateway service without an auth
1906                 // secret
1907                 cr.CrunchLog.Printf("Not starting a gateway server (GatewayAuthSecret was not provided by dispatcher)")
1908         } else {
1909                 gwListen := os.Getenv("GatewayAddress")
1910                 cr.gateway = Gateway{
1911                         Address:       gwListen,
1912                         AuthSecret:    gwAuthSecret,
1913                         ContainerUUID: containerUUID,
1914                         Target:        cr.executor,
1915                         Log:           cr.CrunchLog,
1916                 }
1917                 if gwListen == "" {
1918                         // Direct connection won't work, so we use the
1919                         // gateway_address field to indicate the
1920                         // internalURL of the controller process that
1921                         // has the current tunnel connection.
1922                         cr.gateway.ArvadosClient = cr.dispatcherClient
1923                         cr.gateway.UpdateTunnelURL = func(url string) {
1924                                 cr.gateway.Address = "tunnel " + url
1925                                 cr.DispatcherArvClient.Update("containers", containerUUID,
1926                                         arvadosclient.Dict{"container": arvadosclient.Dict{"gateway_address": cr.gateway.Address}}, nil)
1927                         }
1928                 }
1929                 err = cr.gateway.Start()
1930                 if err != nil {
1931                         log.Printf("error starting gateway server: %s", err)
1932                         return 1
1933                 }
1934         }
1935
1936         parentTemp, tmperr := cr.MkTempDir("", "crunch-run."+containerUUID+".")
1937         if tmperr != nil {
1938                 log.Printf("%s: %v", containerUUID, tmperr)
1939                 return 1
1940         }
1941
1942         cr.parentTemp = parentTemp
1943         cr.statInterval = *statInterval
1944         cr.cgroupRoot = *cgroupRoot
1945         cr.expectCgroupParent = *cgroupParent
1946         cr.enableMemoryLimit = *enableMemoryLimit
1947         cr.enableNetwork = *enableNetwork
1948         cr.networkMode = *networkMode
1949         if *cgroupParentSubsystem != "" {
1950                 p, err := findCgroup(*cgroupParentSubsystem)
1951                 if err != nil {
1952                         log.Printf("fatal: cgroup parent subsystem: %s", err)
1953                         return 1
1954                 }
1955                 cr.setCgroupParent = p
1956                 cr.expectCgroupParent = p
1957         }
1958
1959         runerr := cr.Run()
1960
1961         if *memprofile != "" {
1962                 f, err := os.Create(*memprofile)
1963                 if err != nil {
1964                         log.Printf("could not create memory profile: %s", err)
1965                 }
1966                 runtime.GC() // get up-to-date statistics
1967                 if err := pprof.WriteHeapProfile(f); err != nil {
1968                         log.Printf("could not write memory profile: %s", err)
1969                 }
1970                 closeerr := f.Close()
1971                 if closeerr != nil {
1972                         log.Printf("closing memprofile file: %s", err)
1973                 }
1974         }
1975
1976         if runerr != nil {
1977                 log.Printf("%s: %v", containerUUID, runerr)
1978                 return 1
1979         }
1980         return 0
1981 }
1982
1983 // Try to load ConfigData in hpc (slurm/lsf) environment. This means
1984 // loading the cluster config from the specified file and (if that
1985 // works) getting the runtime_constraints container field from
1986 // controller to determine # VCPUs so we can calculate KeepBuffers.
1987 func hpcConfData(uuid string, configFile string, stderr io.Writer) ConfigData {
1988         var conf ConfigData
1989         conf.Cluster = loadClusterConfigFile(configFile, stderr)
1990         if conf.Cluster == nil {
1991                 // skip loading the container record -- we won't be
1992                 // able to start local keepstore anyway.
1993                 return conf
1994         }
1995         arv, err := arvadosclient.MakeArvadosClient()
1996         if err != nil {
1997                 fmt.Fprintf(stderr, "error setting up arvadosclient: %s\n", err)
1998                 return conf
1999         }
2000         arv.Retries = 8
2001         var ctr arvados.Container
2002         err = arv.Call("GET", "containers", uuid, "", arvadosclient.Dict{"select": []string{"runtime_constraints"}}, &ctr)
2003         if err != nil {
2004                 fmt.Fprintf(stderr, "error getting container record: %s\n", err)
2005                 return conf
2006         }
2007         if ctr.RuntimeConstraints.VCPUs > 0 {
2008                 conf.KeepBuffers = ctr.RuntimeConstraints.VCPUs * conf.Cluster.Containers.LocalKeepBlobBuffersPerVCPU
2009         }
2010         return conf
2011 }
2012
2013 // Load cluster config file from given path. If an error occurs, log
2014 // the error to stderr and return nil.
2015 func loadClusterConfigFile(path string, stderr io.Writer) *arvados.Cluster {
2016         ldr := config.NewLoader(&bytes.Buffer{}, ctxlog.New(stderr, "plain", "info"))
2017         ldr.Path = path
2018         cfg, err := ldr.Load()
2019         if err != nil {
2020                 fmt.Fprintf(stderr, "could not load config file %s: %s\n", path, err)
2021                 return nil
2022         }
2023         cluster, err := cfg.GetCluster("")
2024         if err != nil {
2025                 fmt.Fprintf(stderr, "could not use config file %s: %s\n", path, err)
2026                 return nil
2027         }
2028         fmt.Fprintf(stderr, "loaded config file %s\n", path)
2029         return cluster
2030 }
2031
2032 func startLocalKeepstore(configData ConfigData, logbuf io.Writer) (*exec.Cmd, error) {
2033         if configData.KeepBuffers < 1 {
2034                 fmt.Fprintf(logbuf, "not starting a local keepstore process because KeepBuffers=%v in config\n", configData.KeepBuffers)
2035                 return nil, nil
2036         }
2037         if configData.Cluster == nil {
2038                 fmt.Fprint(logbuf, "not starting a local keepstore process because cluster config file was not loaded\n")
2039                 return nil, nil
2040         }
2041         for uuid, vol := range configData.Cluster.Volumes {
2042                 if len(vol.AccessViaHosts) > 0 {
2043                         fmt.Fprintf(logbuf, "not starting a local keepstore process because a volume (%s) uses AccessViaHosts\n", uuid)
2044                         return nil, nil
2045                 }
2046                 if !vol.ReadOnly && vol.Replication < configData.Cluster.Collections.DefaultReplication {
2047                         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)
2048                         return nil, nil
2049                 }
2050         }
2051
2052         // Rather than have an alternate way to tell keepstore how
2053         // many buffers to use when starting it this way, we just
2054         // modify the cluster configuration that we feed it on stdin.
2055         configData.Cluster.API.MaxKeepBlobBuffers = configData.KeepBuffers
2056
2057         ln, err := net.Listen("tcp", "localhost:0")
2058         if err != nil {
2059                 return nil, err
2060         }
2061         _, port, err := net.SplitHostPort(ln.Addr().String())
2062         if err != nil {
2063                 ln.Close()
2064                 return nil, err
2065         }
2066         ln.Close()
2067         url := "http://localhost:" + port
2068
2069         fmt.Fprintf(logbuf, "starting keepstore on %s\n", url)
2070
2071         var confJSON bytes.Buffer
2072         err = json.NewEncoder(&confJSON).Encode(arvados.Config{
2073                 Clusters: map[string]arvados.Cluster{
2074                         configData.Cluster.ClusterID: *configData.Cluster,
2075                 },
2076         })
2077         if err != nil {
2078                 return nil, err
2079         }
2080         cmd := exec.Command("/proc/self/exe", "keepstore", "-config=-")
2081         if target, err := os.Readlink(cmd.Path); err == nil && strings.HasSuffix(target, ".test") {
2082                 // If we're a 'go test' process, running
2083                 // /proc/self/exe would start the test suite in a
2084                 // child process, which is not what we want.
2085                 cmd.Path, _ = exec.LookPath("go")
2086                 cmd.Args = append([]string{"go", "run", "../../cmd/arvados-server"}, cmd.Args[1:]...)
2087                 cmd.Env = os.Environ()
2088         }
2089         cmd.Stdin = &confJSON
2090         cmd.Stdout = logbuf
2091         cmd.Stderr = logbuf
2092         cmd.Env = append(cmd.Env,
2093                 "GOGC=10",
2094                 "ARVADOS_SERVICE_INTERNAL_URL="+url)
2095         err = cmd.Start()
2096         if err != nil {
2097                 return nil, fmt.Errorf("error starting keepstore process: %w", err)
2098         }
2099         cmdExited := false
2100         go func() {
2101                 cmd.Wait()
2102                 cmdExited = true
2103         }()
2104         ctx, cancel := context.WithDeadline(context.Background(), time.Now().Add(time.Second*10))
2105         defer cancel()
2106         poll := time.NewTicker(time.Second / 10)
2107         defer poll.Stop()
2108         client := http.Client{}
2109         for range poll.C {
2110                 testReq, err := http.NewRequestWithContext(ctx, "GET", url+"/_health/ping", nil)
2111                 testReq.Header.Set("Authorization", "Bearer "+configData.Cluster.ManagementToken)
2112                 if err != nil {
2113                         return nil, err
2114                 }
2115                 resp, err := client.Do(testReq)
2116                 if err == nil {
2117                         resp.Body.Close()
2118                         if resp.StatusCode == http.StatusOK {
2119                                 break
2120                         }
2121                 }
2122                 if cmdExited {
2123                         return nil, fmt.Errorf("keepstore child process exited")
2124                 }
2125                 if ctx.Err() != nil {
2126                         return nil, fmt.Errorf("timed out waiting for new keepstore process to report healthy")
2127                 }
2128         }
2129         os.Setenv("ARVADOS_KEEP_SERVICES", url)
2130         return cmd, nil
2131 }
2132
2133 // return current uid, gid, groups in a format suitable for logging:
2134 // "crunch-run process has uid=1234(arvados) gid=1234(arvados)
2135 // groups=1234(arvados),114(fuse)"
2136 func currentUserAndGroups() string {
2137         u, err := user.Current()
2138         if err != nil {
2139                 return fmt.Sprintf("error getting current user ID: %s", err)
2140         }
2141         s := fmt.Sprintf("crunch-run process has uid=%s(%s) gid=%s", u.Uid, u.Username, u.Gid)
2142         if g, err := user.LookupGroupId(u.Gid); err == nil {
2143                 s += fmt.Sprintf("(%s)", g.Name)
2144         }
2145         s += " groups="
2146         if gids, err := u.GroupIds(); err == nil {
2147                 for i, gid := range gids {
2148                         if i > 0 {
2149                                 s += ","
2150                         }
2151                         s += gid
2152                         if g, err := user.LookupGroupId(gid); err == nil {
2153                                 s += fmt.Sprintf("(%s)", g.Name)
2154                         }
2155                 }
2156         }
2157         return s
2158 }