1 // Copyright (C) The Arvados Authors. All rights reserved.
3 // SPDX-License-Identifier: AGPL-3.0
30 "git.arvados.org/arvados.git/lib/config"
31 "git.arvados.org/arvados.git/lib/service"
32 "git.arvados.org/arvados.git/sdk/go/arvados"
33 "git.arvados.org/arvados.git/sdk/go/ctxlog"
34 "git.arvados.org/arvados.git/sdk/go/health"
35 "github.com/fsnotify/fsnotify"
36 "github.com/sirupsen/logrus"
39 type Supervisor struct {
40 SourcePath string // e.g., /home/username/src/arvados
41 SourceVersion string // e.g., acbd1324...
42 ClusterType string // e.g., production
43 ListenHost string // e.g., localhost
44 ControllerAddr string // e.g., 127.0.0.1:8000
45 OwnTemporaryDatabase bool
48 logger logrus.FieldLogger
49 cluster *arvados.Cluster
52 cancel context.CancelFunc
53 done chan struct{} // closed when child procs/services have shut down
54 err error // error that caused shutdown (valid when done is closed)
55 healthChecker *health.Aggregator
56 tasksReady map[string]chan bool
57 waitShutdown sync.WaitGroup
63 environ []string // for child processes
66 func (super *Supervisor) Start(ctx context.Context, cfg *arvados.Config, cfgPath string) {
67 super.ctx, super.cancel = context.WithCancel(ctx)
68 super.done = make(chan struct{})
71 defer close(super.done)
73 sigch := make(chan os.Signal)
74 signal.Notify(sigch, syscall.SIGINT, syscall.SIGTERM)
75 defer signal.Stop(sigch)
77 for sig := range sigch {
78 super.logger.WithField("signal", sig).Info("caught signal")
80 super.err = fmt.Errorf("caught signal %s", sig)
86 hupch := make(chan os.Signal)
87 signal.Notify(hupch, syscall.SIGHUP)
88 defer signal.Stop(hupch)
90 for sig := range hupch {
91 super.logger.WithField("signal", sig).Info("caught signal")
93 super.err = errNeedConfigReload
99 if cfgPath != "" && cfgPath != "-" && cfg.AutoReloadConfig {
100 go watchConfig(super.ctx, super.logger, cfgPath, copyConfig(cfg), func() {
101 if super.err == nil {
102 super.err = errNeedConfigReload
108 err := super.run(cfg)
110 super.logger.WithError(err).Warn("supervisor shut down")
111 if super.err == nil {
118 func (super *Supervisor) Wait() error {
123 func (super *Supervisor) run(cfg *arvados.Config) error {
126 cwd, err := os.Getwd()
130 if !strings.HasPrefix(super.SourcePath, "/") {
131 super.SourcePath = filepath.Join(cwd, super.SourcePath)
133 super.SourcePath, err = filepath.EvalSymlinks(super.SourcePath)
138 // Choose bin and temp dirs: /var/lib/arvados/... in
139 // production, transient tempdir otherwise.
140 if super.ClusterType == "production" {
141 // These dirs have already been created by
142 // "arvados-server install" (or by extracting a
144 super.tempdir = "/var/lib/arvados/tmp"
145 super.wwwtempdir = "/var/lib/arvados/wwwtmp"
146 super.bindir = "/var/lib/arvados/bin"
148 super.tempdir, err = ioutil.TempDir("", "arvados-server-boot-")
152 defer os.RemoveAll(super.tempdir)
153 super.wwwtempdir = super.tempdir
154 super.bindir = filepath.Join(super.tempdir, "bin")
155 if err := os.Mkdir(super.bindir, 0755); err != nil {
160 // Fill in any missing config keys, and write the resulting
161 // config in the temp dir for child services to use.
162 err = super.autofillConfig(cfg)
166 conffile, err := os.OpenFile(filepath.Join(super.wwwtempdir, "config.yml"), os.O_CREATE|os.O_WRONLY, 0644)
170 defer conffile.Close()
171 err = json.NewEncoder(conffile).Encode(cfg)
175 err = conffile.Close()
179 super.configfile = conffile.Name()
181 super.environ = os.Environ()
182 super.cleanEnv([]string{"ARVADOS_"})
183 super.setEnv("ARVADOS_CONFIG", super.configfile)
184 super.setEnv("RAILS_ENV", super.ClusterType)
185 super.setEnv("TMPDIR", super.tempdir)
186 super.prependEnv("PATH", "/var/lib/arvados/bin:")
187 if super.ClusterType != "production" {
188 super.prependEnv("PATH", super.tempdir+"/bin:")
191 super.cluster, err = cfg.GetCluster("")
195 // Now that we have the config, replace the bootstrap logger
196 // with a new one according to the logging config.
197 loglevel := super.cluster.SystemLogs.LogLevel
198 if s := os.Getenv("ARVADOS_DEBUG"); s != "" && s != "0" {
201 super.logger = ctxlog.New(super.Stderr, super.cluster.SystemLogs.Format, loglevel).WithFields(logrus.Fields{
205 if super.SourceVersion == "" && super.ClusterType == "production" {
206 // don't need SourceVersion
207 } else if super.SourceVersion == "" {
208 // Find current source tree version.
210 err = super.RunProgram(super.ctx, ".", runOptions{output: &buf}, "git", "diff", "--shortstat")
214 dirty := buf.Len() > 0
216 err = super.RunProgram(super.ctx, ".", runOptions{output: &buf}, "git", "log", "-n1", "--format=%H")
220 super.SourceVersion = strings.TrimSpace(buf.String())
222 super.SourceVersion += "+uncommitted"
225 return errors.New("specifying a version to run is not yet supported")
228 _, err = super.installGoProgram(super.ctx, "cmd/arvados-server")
232 err = super.setupRubyEnv()
237 tasks := []supervisedTask{
238 createCertificates{},
241 runServiceCommand{name: "controller", svc: super.cluster.Services.Controller, depends: []supervisedTask{seedDatabase{}}},
242 runGoProgram{src: "services/arv-git-httpd", svc: super.cluster.Services.GitHTTP},
243 runGoProgram{src: "services/health", svc: super.cluster.Services.Health},
244 runGoProgram{src: "services/keepproxy", svc: super.cluster.Services.Keepproxy, depends: []supervisedTask{runPassenger{src: "services/api"}}},
245 runGoProgram{src: "services/keepstore", svc: super.cluster.Services.Keepstore},
246 runGoProgram{src: "services/keep-web", svc: super.cluster.Services.WebDAV},
247 runServiceCommand{name: "ws", svc: super.cluster.Services.Websocket, depends: []supervisedTask{seedDatabase{}}},
248 installPassenger{src: "services/api"},
249 runPassenger{src: "services/api", varlibdir: "railsapi", svc: super.cluster.Services.RailsAPI, depends: []supervisedTask{createCertificates{}, seedDatabase{}, installPassenger{src: "services/api"}}},
250 installPassenger{src: "apps/workbench", depends: []supervisedTask{seedDatabase{}}}, // dependency ensures workbench doesn't delay api install/startup
251 runPassenger{src: "apps/workbench", varlibdir: "workbench1", svc: super.cluster.Services.Workbench1, depends: []supervisedTask{installPassenger{src: "apps/workbench"}}},
254 if super.ClusterType != "test" {
255 tasks = append(tasks,
256 runServiceCommand{name: "dispatch-cloud", svc: super.cluster.Services.DispatchCloud},
257 runGoProgram{src: "services/keep-balance", svc: super.cluster.Services.Keepbalance},
260 super.tasksReady = map[string]chan bool{}
261 for _, task := range tasks {
262 super.tasksReady[task.String()] = make(chan bool)
264 for _, task := range tasks {
266 fail := func(err error) {
267 if super.ctx.Err() != nil {
271 super.logger.WithField("task", task.String()).WithError(err).Error("task failed")
274 super.logger.WithField("task", task.String()).Info("starting")
275 err := task.Run(super.ctx, fail, super)
280 close(super.tasksReady[task.String()])
283 err = super.wait(super.ctx, tasks...)
287 super.logger.Info("all startup tasks are complete; starting health checks")
288 super.healthChecker = &health.Aggregator{Cluster: super.cluster}
290 super.logger.Info("shutting down")
291 super.waitShutdown.Wait()
292 return super.ctx.Err()
295 func (super *Supervisor) wait(ctx context.Context, tasks ...supervisedTask) error {
296 for _, task := range tasks {
297 ch, ok := super.tasksReady[task.String()]
299 return fmt.Errorf("no such task: %s", task)
301 super.logger.WithField("task", task.String()).Info("waiting")
304 super.logger.WithField("task", task.String()).Info("ready")
306 super.logger.WithField("task", task.String()).Info("task was never ready")
313 func (super *Supervisor) Stop() {
318 func (super *Supervisor) WaitReady() (*arvados.URL, bool) {
319 ticker := time.NewTicker(time.Second)
321 for waiting := "all"; waiting != ""; {
324 case <-super.ctx.Done():
327 if super.healthChecker == nil {
331 resp := super.healthChecker.ClusterHealth()
332 // The overall health check (resp.Health=="OK") might
333 // never pass due to missing components (like
334 // arvados-dispatch-cloud in a test cluster), so
335 // instead we wait for all configured components to
338 for target, check := range resp.Checks {
339 if check.Health != "OK" {
340 waiting += " " + target
344 super.logger.WithField("targets", waiting[1:]).Info("waiting")
347 u := super.cluster.Services.Controller.ExternalURL
351 func (super *Supervisor) prependEnv(key, prepend string) {
352 for i, s := range super.environ {
353 if strings.HasPrefix(s, key+"=") {
354 super.environ[i] = key + "=" + prepend + s[len(key)+1:]
358 super.environ = append(super.environ, key+"="+prepend)
361 func (super *Supervisor) cleanEnv(prefixes []string) {
363 for _, s := range super.environ {
365 for _, p := range prefixes {
366 if strings.HasPrefix(s, p) {
372 cleaned = append(cleaned, s)
375 super.environ = cleaned
378 func (super *Supervisor) setEnv(key, val string) {
379 for i, s := range super.environ {
380 if strings.HasPrefix(s, key+"=") {
381 super.environ[i] = key + "=" + val
385 super.environ = append(super.environ, key+"="+val)
388 // Remove all but the first occurrence of each env var.
389 func dedupEnv(in []string) []string {
390 saw := map[string]bool{}
392 for _, kv := range in {
393 if split := strings.Index(kv, "="); split < 1 {
394 panic("invalid environment var: " + kv)
395 } else if saw[kv[:split]] {
398 saw[kv[:split]] = true
399 out = append(out, kv)
405 func (super *Supervisor) installGoProgram(ctx context.Context, srcpath string) (string, error) {
406 _, basename := filepath.Split(srcpath)
407 binfile := filepath.Join(super.bindir, basename)
408 if super.ClusterType == "production" {
411 err := super.RunProgram(ctx, filepath.Join(super.SourcePath, srcpath), runOptions{env: []string{"GOBIN=" + super.bindir}}, "go", "install", "-ldflags", "-X git.arvados.org/arvados.git/lib/cmd.version="+super.SourceVersion+" -X main.version="+super.SourceVersion)
415 func (super *Supervisor) usingRVM() bool {
416 return os.Getenv("rvm_path") != ""
419 func (super *Supervisor) setupRubyEnv() error {
420 if !super.usingRVM() {
421 // (If rvm is in use, assume the caller has everything
422 // set up as desired)
423 super.cleanEnv([]string{
428 if _, err := os.Stat("/var/lib/arvados/bin/gem"); err == nil || super.ClusterType == "production" {
429 gem = "/var/lib/arvados/bin/gem"
431 cmd := exec.Command(gem, "env", "gempath")
432 if super.ClusterType == "production" {
433 cmd.Args = append([]string{"sudo", "-u", "www-data", "-E", "HOME=/var/www"}, cmd.Args...)
434 path, err := exec.LookPath("sudo")
436 return fmt.Errorf("LookPath(\"sudo\"): %w", err)
440 cmd.Stderr = super.Stderr
441 cmd.Env = super.environ
442 buf, err := cmd.Output() // /var/lib/arvados/.gem/ruby/2.5.0/bin:...
443 if err != nil || len(buf) == 0 {
444 return fmt.Errorf("gem env gempath: %v", err)
446 gempath := string(bytes.Split(buf, []byte{':'})[0])
447 super.prependEnv("PATH", gempath+"/bin:")
448 super.setEnv("GEM_HOME", gempath)
449 super.setEnv("GEM_PATH", gempath)
451 // Passenger install doesn't work unless $HOME is ~user
452 u, err := user.Current()
456 super.setEnv("HOME", u.HomeDir)
460 func (super *Supervisor) lookPath(prog string) string {
461 for _, val := range super.environ {
462 if strings.HasPrefix(val, "PATH=") {
463 for _, dir := range filepath.SplitList(val[5:]) {
464 path := filepath.Join(dir, prog)
465 if fi, err := os.Stat(path); err == nil && fi.Mode()&0111 != 0 {
474 type runOptions struct {
475 output io.Writer // attach stdout
476 env []string // add/replace environment variables
477 user string // run as specified user
480 // RunProgram runs prog with args, using dir as working directory. If ctx is
481 // cancelled while the child is running, RunProgram terminates the child, waits
482 // for it to exit, then returns.
484 // Child's environment will have our env vars, plus any given in env.
486 // Child's stdout will be written to output if non-nil, otherwise the
487 // boot command's stderr.
488 func (super *Supervisor) RunProgram(ctx context.Context, dir string, opts runOptions, prog string, args ...string) error {
489 cmdline := fmt.Sprintf("%s", append([]string{prog}, args...))
490 super.logger.WithField("command", cmdline).WithField("dir", dir).Info("executing")
495 if logprefix == "sudo" {
496 for i := 0; i < len(args); i++ {
499 } else if args[i] == "-E" || strings.Contains(args[i], "=") {
502 innerargs = args[i+1:]
507 logprefix = strings.TrimPrefix(logprefix, "/var/lib/arvados/bin/")
508 logprefix = strings.TrimPrefix(logprefix, super.tempdir+"/bin/")
509 if logprefix == "bundle" && len(innerargs) > 2 && innerargs[0] == "exec" {
510 _, dirbase := filepath.Split(dir)
511 logprefix = innerargs[1] + "@" + dirbase
512 } else if logprefix == "arvados-server" && len(args) > 1 {
515 if !strings.HasPrefix(dir, "/") {
516 logprefix = dir + ": " + logprefix
520 cmd := exec.Command(super.lookPath(prog), args...)
521 stdout, err := cmd.StdoutPipe()
525 stderr, err := cmd.StderrPipe()
529 logwriter := &service.LogPrefixer{Writer: super.Stderr, Prefix: []byte("[" + logprefix + "] ")}
530 var copiers sync.WaitGroup
533 io.Copy(logwriter, stderr)
538 if opts.output == nil {
539 io.Copy(logwriter, stdout)
541 io.Copy(opts.output, stdout)
546 if strings.HasPrefix(dir, "/") {
549 cmd.Dir = filepath.Join(super.SourcePath, dir)
551 env := append([]string(nil), opts.env...)
552 env = append(env, super.environ...)
553 cmd.Env = dedupEnv(env)
556 // Note: We use this approach instead of "sudo"
557 // because in certain circumstances (we are pid 1 in a
558 // docker container, and our passenger child process
559 // changes to pgid 1) the intermediate sudo process
560 // notices we have the same pgid as our child and
561 // refuses to propagate signals from us to our child,
562 // so we can't signal/shutdown our passenger/rails
563 // apps. "chpst" or "setuidgid" would work, but these
564 // few lines avoid depending on runit/daemontools.
565 u, err := user.Lookup(opts.user)
567 return fmt.Errorf("user.Lookup(%q): %w", opts.user, err)
569 uid, _ := strconv.Atoi(u.Uid)
570 gid, _ := strconv.Atoi(u.Gid)
571 cmd.SysProcAttr = &syscall.SysProcAttr{
572 Credential: &syscall.Credential{
580 defer func() { exited = true }()
583 log := ctxlog.FromContext(ctx).WithFields(logrus.Fields{"dir": dir, "cmdline": cmdline})
585 if cmd.Process == nil {
586 log.Debug("waiting for child process to start")
587 time.Sleep(time.Second / 2)
589 log.WithField("PID", cmd.Process.Pid).Debug("sending SIGTERM")
590 cmd.Process.Signal(syscall.SIGTERM)
591 time.Sleep(5 * time.Second)
595 log.WithField("PID", cmd.Process.Pid).Warn("still waiting for child process to exit 5s after SIGTERM")
607 if ctx.Err() != nil {
608 // Return "context canceled", instead of the "killed"
609 // error that was probably caused by the context being
612 } else if err != nil {
613 return fmt.Errorf("%s: error: %v", cmdline, err)
618 func (super *Supervisor) autofillConfig(cfg *arvados.Config) error {
619 cluster, err := cfg.GetCluster("")
623 usedPort := map[string]bool{}
624 nextPort := func(host string) string {
626 port, err := availablePort(host)
633 usedPort[port] = true
637 if cluster.Services.Controller.ExternalURL.Host == "" {
638 h, p, err := net.SplitHostPort(super.ControllerAddr)
648 cluster.Services.Controller.ExternalURL = arvados.URL{Scheme: "https", Host: net.JoinHostPort(h, p), Path: "/"}
650 for _, svc := range []*arvados.Service{
651 &cluster.Services.Controller,
652 &cluster.Services.DispatchCloud,
653 &cluster.Services.GitHTTP,
654 &cluster.Services.Health,
655 &cluster.Services.Keepproxy,
656 &cluster.Services.Keepstore,
657 &cluster.Services.RailsAPI,
658 &cluster.Services.WebDAV,
659 &cluster.Services.WebDAVDownload,
660 &cluster.Services.Websocket,
661 &cluster.Services.Workbench1,
663 if svc == &cluster.Services.DispatchCloud && super.ClusterType == "test" {
666 if svc.ExternalURL.Host == "" {
667 if svc == &cluster.Services.Controller ||
668 svc == &cluster.Services.GitHTTP ||
669 svc == &cluster.Services.Health ||
670 svc == &cluster.Services.Keepproxy ||
671 svc == &cluster.Services.WebDAV ||
672 svc == &cluster.Services.WebDAVDownload ||
673 svc == &cluster.Services.Workbench1 {
674 svc.ExternalURL = arvados.URL{Scheme: "https", Host: fmt.Sprintf("%s:%s", super.ListenHost, nextPort(super.ListenHost)), Path: "/"}
675 } else if svc == &cluster.Services.Websocket {
676 svc.ExternalURL = arvados.URL{Scheme: "wss", Host: fmt.Sprintf("%s:%s", super.ListenHost, nextPort(super.ListenHost)), Path: "/websocket"}
679 if len(svc.InternalURLs) == 0 {
680 svc.InternalURLs = map[arvados.URL]arvados.ServiceInstance{
681 {Scheme: "http", Host: fmt.Sprintf("%s:%s", super.ListenHost, nextPort(super.ListenHost)), Path: "/"}: {},
685 if super.ClusterType != "production" {
686 if cluster.SystemRootToken == "" {
687 cluster.SystemRootToken = randomHexString(64)
689 if cluster.ManagementToken == "" {
690 cluster.ManagementToken = randomHexString(64)
692 if cluster.Collections.BlobSigningKey == "" {
693 cluster.Collections.BlobSigningKey = randomHexString(64)
695 if cluster.Users.AnonymousUserToken == "" {
696 cluster.Users.AnonymousUserToken = randomHexString(64)
698 if cluster.Containers.DispatchPrivateKey == "" {
699 buf, err := ioutil.ReadFile(filepath.Join(super.SourcePath, "lib", "dispatchcloud", "test", "sshkey_dispatch"))
703 cluster.Containers.DispatchPrivateKey = string(buf)
705 cluster.TLS.Insecure = true
707 if super.ClusterType == "test" {
708 // Add a second keepstore process.
709 cluster.Services.Keepstore.InternalURLs[arvados.URL{Scheme: "http", Host: fmt.Sprintf("%s:%s", super.ListenHost, nextPort(super.ListenHost)), Path: "/"}] = arvados.ServiceInstance{}
711 // Create a directory-backed volume for each keepstore
713 cluster.Volumes = map[string]arvados.Volume{}
714 for url := range cluster.Services.Keepstore.InternalURLs {
715 volnum := len(cluster.Volumes)
716 datadir := fmt.Sprintf("%s/keep%d.data", super.tempdir, volnum)
717 if _, err = os.Stat(datadir + "/."); err == nil {
718 } else if !os.IsNotExist(err) {
720 } else if err = os.Mkdir(datadir, 0755); err != nil {
723 cluster.Volumes[fmt.Sprintf(cluster.ClusterID+"-nyw5e-%015d", volnum)] = arvados.Volume{
725 DriverParameters: json.RawMessage(fmt.Sprintf(`{"Root":%q}`, datadir)),
726 AccessViaHosts: map[arvados.URL]arvados.VolumeAccess{
732 if super.OwnTemporaryDatabase {
733 cluster.PostgreSQL.Connection = arvados.PostgreSQLConnection{
734 "client_encoding": "utf8",
736 "port": nextPort(super.ListenHost),
737 "dbname": "arvados_test",
739 "password": "insecure_arvados_test",
743 cfg.Clusters[cluster.ClusterID] = *cluster
747 func addrIsLocal(addr string) (bool, error) {
749 listener, err := net.Listen("tcp", addr)
753 } else if strings.Contains(err.Error(), "cannot assign requested address") {
760 func randomHexString(chars int) string {
761 b := make([]byte, chars/2)
762 _, err := rand.Read(b)
766 return fmt.Sprintf("%x", b)
769 func internalPort(svc arvados.Service) (string, error) {
770 if len(svc.InternalURLs) > 1 {
771 return "", errors.New("internalPort() doesn't work with multiple InternalURLs")
773 for u := range svc.InternalURLs {
775 if p := u.Port(); p != "" {
777 } else if u.Scheme == "https" || u.Scheme == "ws" {
783 return "", fmt.Errorf("service has no InternalURLs")
786 func externalPort(svc arvados.Service) (string, error) {
787 u := url.URL(svc.ExternalURL)
788 if p := u.Port(); p != "" {
790 } else if u.Scheme == "https" || u.Scheme == "wss" {
797 func availablePort(host string) (string, error) {
798 ln, err := net.Listen("tcp", net.JoinHostPort(host, "0"))
803 _, port, err := net.SplitHostPort(ln.Addr().String())
810 // Try to connect to addr until it works, then close ch. Give up if
812 func waitForConnect(ctx context.Context, addr string) error {
813 dialer := net.Dialer{Timeout: time.Second}
814 for ctx.Err() == nil {
815 conn, err := dialer.DialContext(ctx, "tcp", addr)
817 time.Sleep(time.Second / 10)
826 func copyConfig(cfg *arvados.Config) *arvados.Config {
829 err := json.NewEncoder(pw).Encode(cfg)
835 cfg2 := new(arvados.Config)
836 err := json.NewDecoder(pr).Decode(cfg2)
843 func watchConfig(ctx context.Context, logger logrus.FieldLogger, cfgPath string, prevcfg *arvados.Config, fn func()) {
844 watcher, err := fsnotify.NewWatcher()
846 logger.WithError(err).Error("fsnotify setup failed")
849 defer watcher.Close()
851 err = watcher.Add(cfgPath)
853 logger.WithError(err).Error("fsnotify watcher failed")
861 case err, ok := <-watcher.Errors:
865 logger.WithError(err).Warn("fsnotify watcher reported error")
866 case _, ok := <-watcher.Events:
870 for len(watcher.Events) > 0 {
873 loader := config.NewLoader(&bytes.Buffer{}, &logrus.Logger{Out: ioutil.Discard})
874 loader.Path = cfgPath
875 loader.SkipAPICalls = true
876 cfg, err := loader.Load()
878 logger.WithError(err).Warn("error reloading config file after change detected; ignoring new config for now")
879 } else if reflect.DeepEqual(cfg, prevcfg) {
880 logger.Debug("config file changed but is still DeepEqual to the existing config")
882 logger.Debug("config changed, notifying supervisor")