lib/boot/supervisor.go

   1 // Copyright (C) The Arvados Authors. All rights reserved.
   2 //
   3 // SPDX-License-Identifier: AGPL-3.0
   4
   5 package boot
   6
   7 import (
   8         "bytes"
   9         "context"
  10         "crypto/rand"
  11         "encoding/json"
  12         "errors"
  13         "fmt"
  14         "io"
  15         "io/ioutil"
  16         "net"
  17         "net/url"
  18         "os"
  19         "os/exec"
  20         "os/signal"
  21         "os/user"
  22         "path/filepath"
  23         "reflect"
  24         "strconv"
  25         "strings"
  26         "sync"
  27         "syscall"
  28         "time"
  29
  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"
  37 )
  38
  39 type Supervisor struct {
  40         // Config file location like "/etc/arvados/config.yml", or "-"
  41         // to read from Stdin (see below).
  42         ConfigPath string
  43         // Literal config file (useful for test suites). If non-empty,
  44         // this is used instead of ConfigPath.
  45         ConfigYAML string
  46         // Path to arvados source tree. Only used for dev/test
  47         // clusters.
  48         SourcePath string
  49         // Version number to build into binaries. Only used for
  50         // dev/test clusters.
  51         SourceVersion string
  52         // "production", "development", or "test".
  53         ClusterType string
  54         // Listening address for external services, and internal
  55         // services whose InternalURLs are not explicitly configured.
  56         // If blank, listen on the configured controller ExternalURL
  57         // host; if that is also blank, listen on all addresses
  58         // (0.0.0.0).
  59         ListenHost string
  60         // Default host:port for controller ExternalURL if not
  61         // explicitly configured in config file. If blank, use a
  62         // random port on ListenHost.
  63         ControllerAddr string
  64
  65         NoWorkbench1         bool
  66         NoWorkbench2         bool
  67         OwnTemporaryDatabase bool
  68         Stdin                io.Reader
  69         Stderr               io.Writer
  70
  71         logger   logrus.FieldLogger
  72         cluster  *arvados.Cluster       // nil if this is a multi-cluster supervisor
  73         children map[string]*Supervisor // nil if this is a single-cluster supervisor
  74
  75         ctx           context.Context
  76         cancel        context.CancelFunc
  77         done          chan struct{}      // closed when child procs/services have shut down
  78         err           error              // error that caused shutdown (valid when done is closed)
  79         healthChecker *health.Aggregator // nil if this is a multi-cluster supervisor, or still booting
  80         tasksReady    map[string]chan bool
  81         waitShutdown  sync.WaitGroup
  82
  83         bindir     string
  84         tempdir    string // in production mode, this is accessible only to root
  85         wwwtempdir string // in production mode, this is accessible only to www-data
  86         configfile string
  87         environ    []string // for child processes
  88 }
  89
  90 func (super *Supervisor) Clusters() map[string]*arvados.Cluster {
  91         m := map[string]*arvados.Cluster{}
  92         if super.cluster != nil {
  93                 m[super.cluster.ClusterID] = super.cluster
  94         }
  95         for id, super2 := range super.children {
  96                 m[id] = super2.Cluster("")
  97         }
  98         return m
  99 }
 100
 101 func (super *Supervisor) Cluster(id string) *arvados.Cluster {
 102         if super.children != nil {
 103                 return super.children[id].Cluster(id)
 104         } else {
 105                 return super.cluster
 106         }
 107 }
 108
 109 func (super *Supervisor) Start(ctx context.Context) {
 110         super.logger = ctxlog.FromContext(ctx)
 111         super.ctx, super.cancel = context.WithCancel(ctx)
 112         super.done = make(chan struct{})
 113
 114         sigch := make(chan os.Signal, 1)
 115         signal.Notify(sigch, syscall.SIGINT, syscall.SIGTERM, syscall.SIGHUP)
 116         go func() {
 117                 defer signal.Stop(sigch)
 118                 for {
 119                         select {
 120                         case <-ctx.Done():
 121                                 return
 122                         case sig := <-sigch:
 123                                 super.logger.WithField("signal", sig).Info("caught signal")
 124                                 if super.err == nil {
 125                                         if sig == syscall.SIGHUP {
 126                                                 super.err = errNeedConfigReload
 127                                         } else {
 128                                                 super.err = fmt.Errorf("caught signal %s", sig)
 129                                         }
 130                                 }
 131                                 super.cancel()
 132                         }
 133                 }
 134         }()
 135
 136         loaderStdin := super.Stdin
 137         if super.ConfigYAML != "" {
 138                 loaderStdin = bytes.NewBufferString(super.ConfigYAML)
 139         }
 140         loader := config.NewLoader(loaderStdin, super.logger)
 141         loader.SkipLegacy = true
 142         loader.SkipAPICalls = true
 143         loader.Path = super.ConfigPath
 144         if super.ConfigYAML != "" {
 145                 loader.Path = "-"
 146         }
 147         cfg, err := loader.Load()
 148         if err != nil {
 149                 super.err = err
 150                 close(super.done)
 151                 super.cancel()
 152                 return
 153         }
 154
 155         if super.ConfigPath != "" && super.ConfigPath != "-" && cfg.AutoReloadConfig {
 156                 go watchConfig(super.ctx, super.logger, super.ConfigPath, copyConfig(cfg), func() {
 157                         if super.err == nil {
 158                                 super.err = errNeedConfigReload
 159                         }
 160                         super.cancel()
 161                 })
 162         }
 163
 164         if len(cfg.Clusters) > 1 {
 165                 super.startFederation(cfg)
 166                 go func() {
 167                         defer super.cancel()
 168                         defer close(super.done)
 169                         for _, super2 := range super.children {
 170                                 err := super2.Wait()
 171                                 if super.err == nil {
 172                                         super.err = err
 173                                 }
 174                         }
 175                 }()
 176         } else {
 177                 go func() {
 178                         defer super.cancel()
 179                         defer close(super.done)
 180                         super.cluster, super.err = cfg.GetCluster("")
 181                         if super.err != nil {
 182                                 return
 183                         }
 184                         err := super.runCluster()
 185                         if err != nil {
 186                                 super.logger.WithError(err).Info("supervisor shut down")
 187                                 if super.err == nil {
 188                                         super.err = err
 189                                 }
 190                         }
 191                 }()
 192         }
 193 }
 194
 195 // Wait returns when all child processes and goroutines have exited.
 196 func (super *Supervisor) Wait() error {
 197         <-super.done
 198         return super.err
 199 }
 200
 201 // startFederation starts a child Supervisor for each cluster in the
 202 // given config. Each is a copy of the original/parent with the
 203 // original config reduced to a single cluster.
 204 func (super *Supervisor) startFederation(cfg *arvados.Config) {
 205         super.children = map[string]*Supervisor{}
 206         for id, cc := range cfg.Clusters {
 207                 yaml, err := json.Marshal(arvados.Config{Clusters: map[string]arvados.Cluster{id: cc}})
 208                 if err != nil {
 209                         panic(fmt.Sprintf("json.Marshal partial config: %s", err))
 210                 }
 211                 super2 := &Supervisor{
 212                         ConfigPath:           "-",
 213                         ConfigYAML:           string(yaml),
 214                         SourcePath:           super.SourcePath,
 215                         SourceVersion:        super.SourceVersion,
 216                         ClusterType:          super.ClusterType,
 217                         ListenHost:           super.ListenHost,
 218                         ControllerAddr:       super.ControllerAddr,
 219                         NoWorkbench1:         super.NoWorkbench1,
 220                         NoWorkbench2:         super.NoWorkbench2,
 221                         OwnTemporaryDatabase: super.OwnTemporaryDatabase,
 222                         Stdin:                super.Stdin,
 223                         Stderr:               super.Stderr,
 224                 }
 225                 if super2.ClusterType == "test" {
 226                         super2.Stderr = &service.LogPrefixer{
 227                                 Writer: super.Stderr,
 228                                 Prefix: []byte("[" + id + "] "),
 229                         }
 230                 }
 231                 super2.Start(super.ctx)
 232                 super.children[id] = super2
 233         }
 234 }
 235
 236 func (super *Supervisor) runCluster() error {
 237         cwd, err := os.Getwd()
 238         if err != nil {
 239                 return err
 240         }
 241         if super.ClusterType == "test" && super.SourcePath == "" {
 242                 // When invoked by test suite, default to current
 243                 // source tree
 244                 buf, err := exec.Command("git", "rev-parse", "--show-toplevel").CombinedOutput()
 245                 if err != nil {
 246                         return fmt.Errorf("git rev-parse: %w", err)
 247                 }
 248                 super.SourcePath = strings.TrimSuffix(string(buf), "\n")
 249         } else if !strings.HasPrefix(super.SourcePath, "/") {
 250                 super.SourcePath = filepath.Join(cwd, super.SourcePath)
 251         }
 252         super.SourcePath, err = filepath.EvalSymlinks(super.SourcePath)
 253         if err != nil {
 254                 return err
 255         }
 256
 257         if super.ListenHost == "" {
 258                 u := url.URL(super.cluster.Services.Controller.ExternalURL)
 259                 super.ListenHost = u.Hostname()
 260                 if super.ListenHost == "" {
 261                         super.ListenHost = "0.0.0.0"
 262                 }
 263         }
 264
 265         // Choose bin and temp dirs: /var/lib/arvados/... in
 266         // production, transient tempdir otherwise.
 267         if super.ClusterType == "production" {
 268                 // These dirs have already been created by
 269                 // "arvados-server install" (or by extracting a
 270                 // package).
 271                 super.tempdir = "/var/lib/arvados/tmp"
 272                 super.wwwtempdir = "/var/lib/arvados/wwwtmp"
 273                 super.bindir = "/var/lib/arvados/bin"
 274         } else {
 275                 super.tempdir, err = ioutil.TempDir("", "arvados-server-boot-")
 276                 if err != nil {
 277                         return err
 278                 }
 279                 defer os.RemoveAll(super.tempdir)
 280                 super.wwwtempdir = super.tempdir
 281                 super.bindir = filepath.Join(super.tempdir, "bin")
 282                 if err := os.Mkdir(super.bindir, 0755); err != nil {
 283                         return err
 284                 }
 285         }
 286
 287         // Fill in any missing config keys, and write the resulting
 288         // config in the temp dir for child services to use.
 289         err = super.autofillConfig()
 290         if err != nil {
 291                 return err
 292         }
 293         conffile, err := os.OpenFile(filepath.Join(super.wwwtempdir, "config.yml"), os.O_CREATE|os.O_WRONLY|os.O_TRUNC, 0644)
 294         if err != nil {
 295                 return err
 296         }
 297         defer conffile.Close()
 298         err = json.NewEncoder(conffile).Encode(arvados.Config{
 299                 Clusters: map[string]arvados.Cluster{
 300                         super.cluster.ClusterID: *super.cluster}})
 301         if err != nil {
 302                 return err
 303         }
 304         err = conffile.Close()
 305         if err != nil {
 306                 return err
 307         }
 308         super.configfile = conffile.Name()
 309
 310         super.environ = os.Environ()
 311         super.cleanEnv([]string{"ARVADOS_"})
 312         super.setEnv("ARVADOS_CONFIG", super.configfile)
 313         super.setEnv("RAILS_ENV", super.ClusterType)
 314         super.setEnv("TMPDIR", super.tempdir)
 315         super.prependEnv("PATH", "/var/lib/arvados/bin:")
 316         if super.ClusterType != "production" {
 317                 super.prependEnv("PATH", super.tempdir+"/bin:")
 318         }
 319         super.setEnv("ARVADOS_SERVER_ADDRESS", super.ListenHost)
 320
 321         // Now that we have the config, replace the bootstrap logger
 322         // with a new one according to the logging config.
 323         loglevel := super.cluster.SystemLogs.LogLevel
 324         if s := os.Getenv("ARVADOS_DEBUG"); s != "" && s != "0" {
 325                 loglevel = "debug"
 326         }
 327         super.logger = ctxlog.New(super.Stderr, super.cluster.SystemLogs.Format, loglevel).WithFields(logrus.Fields{
 328                 "PID": os.Getpid(),
 329         })
 330
 331         if super.SourceVersion == "" && super.ClusterType == "production" {
 332                 // don't need SourceVersion
 333         } else if super.SourceVersion == "" {
 334                 // Find current source tree version.
 335                 var buf bytes.Buffer
 336                 err = super.RunProgram(super.ctx, super.SourcePath, runOptions{output: &buf}, "git", "diff", "--shortstat")
 337                 if err != nil {
 338                         return err
 339                 }
 340                 dirty := buf.Len() > 0
 341                 buf.Reset()
 342                 err = super.RunProgram(super.ctx, super.SourcePath, runOptions{output: &buf}, "git", "log", "-n1", "--format=%H")
 343                 if err != nil {
 344                         return err
 345                 }
 346                 super.SourceVersion = strings.TrimSpace(buf.String())
 347                 if dirty {
 348                         super.SourceVersion += "+uncommitted"
 349                 }
 350         } else {
 351                 return errors.New("specifying a version to run is not yet supported")
 352         }
 353
 354         _, err = super.installGoProgram(super.ctx, "cmd/arvados-server")
 355         if err != nil {
 356                 return err
 357         }
 358         err = super.setupRubyEnv()
 359         if err != nil {
 360                 return err
 361         }
 362
 363         tasks := []supervisedTask{
 364                 createCertificates{},
 365                 runPostgreSQL{},
 366                 runNginx{},
 367                 railsDatabase{},
 368                 runServiceCommand{name: "controller", svc: super.cluster.Services.Controller, depends: []supervisedTask{railsDatabase{}}},
 369                 runServiceCommand{name: "health", svc: super.cluster.Services.Health},
 370                 runServiceCommand{name: "keepproxy", svc: super.cluster.Services.Keepproxy, depends: []supervisedTask{runPassenger{src: "services/api"}}},
 371                 runServiceCommand{name: "keepstore", svc: super.cluster.Services.Keepstore},
 372                 runServiceCommand{name: "keep-web", svc: super.cluster.Services.WebDAV},
 373                 runServiceCommand{name: "ws", svc: super.cluster.Services.Websocket, depends: []supervisedTask{railsDatabase{}}},
 374                 installPassenger{src: "services/api", varlibdir: "railsapi"},
 375                 runPassenger{src: "services/api", varlibdir: "railsapi", svc: super.cluster.Services.RailsAPI, depends: []supervisedTask{
 376                         createCertificates{},
 377                         installPassenger{src: "services/api", varlibdir: "railsapi"},
 378                         railsDatabase{},
 379                 }},
 380         }
 381         if !super.NoWorkbench1 {
 382                 return errors.New("workbench1 is no longer supported")
 383         }
 384         if !super.NoWorkbench2 {
 385                 tasks = append(tasks,
 386                         runWorkbench2{svc: super.cluster.Services.Workbench2},
 387                 )
 388         }
 389         if super.ClusterType != "test" {
 390                 tasks = append(tasks,
 391                         runServiceCommand{name: "keep-balance", svc: super.cluster.Services.Keepbalance},
 392                 )
 393         }
 394         if super.cluster.Containers.CloudVMs.Enable {
 395                 tasks = append(tasks,
 396                         runServiceCommand{name: "dispatch-cloud", svc: super.cluster.Services.DispatchCloud},
 397                 )
 398         }
 399         super.tasksReady = map[string]chan bool{}
 400         for _, task := range tasks {
 401                 super.tasksReady[task.String()] = make(chan bool)
 402         }
 403         for _, task := range tasks {
 404                 task := task
 405                 fail := func(err error) {
 406                         if super.ctx.Err() != nil {
 407                                 return
 408                         }
 409                         super.cancel()
 410                         super.logger.WithField("task", task.String()).WithError(err).Error("task failed")
 411                 }
 412                 go func() {
 413                         super.logger.WithField("task", task.String()).Info("starting")
 414                         err := task.Run(super.ctx, fail, super)
 415                         if err != nil {
 416                                 fail(err)
 417                                 return
 418                         }
 419                         close(super.tasksReady[task.String()])
 420                 }()
 421         }
 422         err = super.wait(super.ctx, tasks...)
 423         if err != nil {
 424                 return err
 425         }
 426         super.logger.Info("all startup tasks are complete; starting health checks")
 427         super.healthChecker = &health.Aggregator{Cluster: super.cluster}
 428         <-super.ctx.Done()
 429         super.logger.Info("shutting down")
 430         super.waitShutdown.Wait()
 431         return super.ctx.Err()
 432 }
 433
 434 func (super *Supervisor) wait(ctx context.Context, tasks ...supervisedTask) error {
 435         ticker := time.NewTicker(15 * time.Second)
 436         defer ticker.Stop()
 437         for _, task := range tasks {
 438                 ch, ok := super.tasksReady[task.String()]
 439                 if !ok {
 440                         return fmt.Errorf("no such task: %s", task)
 441                 }
 442                 super.logger.WithField("task", task.String()).Info("waiting")
 443                 for {
 444                         select {
 445                         case <-ch:
 446                                 super.logger.WithField("task", task.String()).Info("ready")
 447                         case <-ctx.Done():
 448                                 super.logger.WithField("task", task.String()).Info("task was never ready")
 449                                 return ctx.Err()
 450                         case <-ticker.C:
 451                                 super.logger.WithField("task", task.String()).Info("still waiting...")
 452                                 continue
 453                         }
 454                         break
 455                 }
 456         }
 457         return nil
 458 }
 459
 460 // Stop shuts down all child processes and goroutines, and returns
 461 // when all of them have exited.
 462 func (super *Supervisor) Stop() {
 463         super.cancel()
 464         <-super.done
 465 }
 466
 467 // WaitReady waits for the cluster(s) to be ready to handle requests,
 468 // then returns true. If startup fails, it returns false.
 469 func (super *Supervisor) WaitReady() bool {
 470         if super.children != nil {
 471                 for id, super2 := range super.children {
 472                         super.logger.Infof("waiting for %s to be ready", id)
 473                         if !super2.WaitReady() {
 474                                 super.logger.Infof("%s startup failed", id)
 475                                 super.Stop()
 476                                 return false
 477                         }
 478                         super.logger.Infof("%s is ready", id)
 479                 }
 480                 super.logger.Info("all clusters are ready")
 481                 return true
 482         }
 483         ticker := time.NewTicker(time.Second)
 484         defer ticker.Stop()
 485         for waiting := "all"; waiting != ""; {
 486                 select {
 487                 case <-ticker.C:
 488                 case <-super.ctx.Done():
 489                         super.Stop()
 490                         return false
 491                 }
 492                 if super.healthChecker == nil {
 493                         // not set up yet
 494                         continue
 495                 }
 496                 resp := super.healthChecker.ClusterHealth()
 497                 // The overall health check (resp.Health=="OK") might
 498                 // never pass due to missing components (like
 499                 // arvados-dispatch-cloud in a test cluster), so
 500                 // instead we wait for all configured components to
 501                 // pass.
 502                 waiting = ""
 503                 for target, check := range resp.Checks {
 504                         if check.Health != "OK" {
 505                                 waiting += " " + target
 506                         }
 507                 }
 508                 if waiting != "" {
 509                         super.logger.WithField("targets", waiting[1:]).Info("waiting")
 510                 }
 511         }
 512         return true
 513 }
 514
 515 func (super *Supervisor) prependEnv(key, prepend string) {
 516         for i, s := range super.environ {
 517                 if strings.HasPrefix(s, key+"=") {
 518                         super.environ[i] = key + "=" + prepend + s[len(key)+1:]
 519                         return
 520                 }
 521         }
 522         super.environ = append(super.environ, key+"="+prepend)
 523 }
 524
 525 func (super *Supervisor) cleanEnv(prefixes []string) {
 526         var cleaned []string
 527         for _, s := range super.environ {
 528                 drop := false
 529                 for _, p := range prefixes {
 530                         if strings.HasPrefix(s, p) {
 531                                 drop = true
 532                                 break
 533                         }
 534                 }
 535                 if !drop {
 536                         cleaned = append(cleaned, s)
 537                 }
 538         }
 539         super.environ = cleaned
 540 }
 541
 542 func (super *Supervisor) setEnv(key, val string) {
 543         for i, s := range super.environ {
 544                 if strings.HasPrefix(s, key+"=") {
 545                         super.environ[i] = key + "=" + val
 546                         return
 547                 }
 548         }
 549         super.environ = append(super.environ, key+"="+val)
 550 }
 551
 552 // Remove all but the first occurrence of each env var.
 553 func dedupEnv(in []string) []string {
 554         saw := map[string]bool{}
 555         var out []string
 556         for _, kv := range in {
 557                 if split := strings.Index(kv, "="); split < 1 {
 558                         panic("invalid environment var: " + kv)
 559                 } else if saw[kv[:split]] {
 560                         continue
 561                 } else {
 562                         saw[kv[:split]] = true
 563                         out = append(out, kv)
 564                 }
 565         }
 566         return out
 567 }
 568
 569 func (super *Supervisor) installGoProgram(ctx context.Context, srcpath string) (string, error) {
 570         _, basename := filepath.Split(srcpath)
 571         binfile := filepath.Join(super.bindir, basename)
 572         if super.ClusterType == "production" {
 573                 return binfile, nil
 574         }
 575         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)
 576         return binfile, err
 577 }
 578
 579 func (super *Supervisor) usingRVM() bool {
 580         return os.Getenv("rvm_path") != ""
 581 }
 582
 583 func (super *Supervisor) setupRubyEnv() error {
 584         if !super.usingRVM() {
 585                 // (If rvm is in use, assume the caller has everything
 586                 // set up as desired)
 587                 super.cleanEnv([]string{
 588                         "GEM_HOME=",
 589                         "GEM_PATH=",
 590                 })
 591                 gem := "gem"
 592                 if _, err := os.Stat("/var/lib/arvados/bin/gem"); err == nil || super.ClusterType == "production" {
 593                         gem = "/var/lib/arvados/bin/gem"
 594                 }
 595                 cmd := exec.Command(gem, "env", "gempath")
 596                 if super.ClusterType == "production" {
 597                         cmd.Args = append([]string{"sudo", "-u", "www-data", "-E", "HOME=/var/www"}, cmd.Args...)
 598                         path, err := exec.LookPath("sudo")
 599                         if err != nil {
 600                                 return fmt.Errorf("LookPath(\"sudo\"): %w", err)
 601                         }
 602                         cmd.Path = path
 603                 }
 604                 cmd.Stderr = super.Stderr
 605                 cmd.Env = super.environ
 606                 buf, err := cmd.Output() // /var/lib/arvados/.gem/ruby/2.5.0/bin:...
 607                 if err != nil || len(buf) == 0 {
 608                         return fmt.Errorf("gem env gempath: %w", err)
 609                 }
 610                 gempath := string(bytes.Split(buf, []byte{':'})[0])
 611                 super.prependEnv("PATH", gempath+"/bin:")
 612                 super.setEnv("GEM_HOME", gempath)
 613                 super.setEnv("GEM_PATH", gempath)
 614         }
 615         // Passenger install doesn't work unless $HOME is ~user
 616         u, err := user.Current()
 617         if err != nil {
 618                 return err
 619         }
 620         super.setEnv("HOME", u.HomeDir)
 621         return nil
 622 }
 623
 624 func (super *Supervisor) lookPath(prog string) string {
 625         for _, val := range super.environ {
 626                 if strings.HasPrefix(val, "PATH=") {
 627                         for _, dir := range filepath.SplitList(val[5:]) {
 628                                 path := filepath.Join(dir, prog)
 629                                 if fi, err := os.Stat(path); err == nil && fi.Mode()&0111 != 0 {
 630                                         return path
 631                                 }
 632                         }
 633                 }
 634         }
 635         return prog
 636 }
 637
 638 type runOptions struct {
 639         output io.Writer // attach stdout
 640         env    []string  // add/replace environment variables
 641         user   string    // run as specified user
 642         stdin  io.Reader
 643 }
 644
 645 // RunProgram runs prog with args, using dir as working directory. If ctx is
 646 // cancelled while the child is running, RunProgram terminates the child, waits
 647 // for it to exit, then returns.
 648 //
 649 // Child's environment will have our env vars, plus any given in env.
 650 //
 651 // Child's stdout will be written to output if non-nil, otherwise the
 652 // boot command's stderr.
 653 func (super *Supervisor) RunProgram(ctx context.Context, dir string, opts runOptions, prog string, args ...string) error {
 654         cmdline := fmt.Sprintf("%s", append([]string{prog}, args...))
 655         super.logger.WithField("command", cmdline).WithField("dir", dir).Info("executing")
 656
 657         logprefix := prog
 658         {
 659                 innerargs := args
 660                 if logprefix == "sudo" {
 661                         for i := 0; i < len(args); i++ {
 662                                 if args[i] == "-u" {
 663                                         i++
 664                                 } else if args[i] == "-E" || strings.Contains(args[i], "=") {
 665                                 } else {
 666                                         logprefix = args[i]
 667                                         innerargs = args[i+1:]
 668                                         break
 669                                 }
 670                         }
 671                 }
 672                 logprefix = strings.TrimPrefix(logprefix, "/var/lib/arvados/bin/")
 673                 logprefix = strings.TrimPrefix(logprefix, super.tempdir+"/bin/")
 674                 if logprefix == "bundle" && len(innerargs) > 2 && innerargs[0] == "exec" {
 675                         _, dirbase := filepath.Split(dir)
 676                         logprefix = innerargs[1] + "@" + dirbase
 677                 } else if logprefix == "arvados-server" && len(args) > 1 {
 678                         logprefix = args[0]
 679                 }
 680                 if !strings.HasPrefix(dir, "/") {
 681                         logprefix = dir + ": " + logprefix
 682                 }
 683         }
 684
 685         cmd := exec.Command(super.lookPath(prog), args...)
 686         cmd.Stdin = opts.stdin
 687         stdout, err := cmd.StdoutPipe()
 688         if err != nil {
 689                 return err
 690         }
 691         stderr, err := cmd.StderrPipe()
 692         if err != nil {
 693                 return err
 694         }
 695         logwriter := &service.LogPrefixer{Writer: super.Stderr, Prefix: []byte("[" + logprefix + "] ")}
 696         var copiers sync.WaitGroup
 697         copiers.Add(1)
 698         go func() {
 699                 io.Copy(logwriter, stderr)
 700                 copiers.Done()
 701         }()
 702         copiers.Add(1)
 703         go func() {
 704                 if opts.output == nil {
 705                         io.Copy(logwriter, stdout)
 706                 } else {
 707                         io.Copy(opts.output, stdout)
 708                 }
 709                 copiers.Done()
 710         }()
 711
 712         if strings.HasPrefix(dir, "/") {
 713                 cmd.Dir = dir
 714         } else {
 715                 cmd.Dir = filepath.Join(super.SourcePath, dir)
 716         }
 717         env := append([]string(nil), opts.env...)
 718         env = append(env, super.environ...)
 719         cmd.Env = dedupEnv(env)
 720
 721         if opts.user != "" {
 722                 // Note: We use this approach instead of "sudo"
 723                 // because in certain circumstances (we are pid 1 in a
 724                 // docker container, and our passenger child process
 725                 // changes to pgid 1) the intermediate sudo process
 726                 // notices we have the same pgid as our child and
 727                 // refuses to propagate signals from us to our child,
 728                 // so we can't signal/shutdown our passenger/rails
 729                 // apps. "chpst" or "setuidgid" would work, but these
 730                 // few lines avoid depending on runit/daemontools.
 731                 u, err := user.Lookup(opts.user)
 732                 if err != nil {
 733                         return fmt.Errorf("user.Lookup(%q): %w", opts.user, err)
 734                 }
 735                 uid, _ := strconv.Atoi(u.Uid)
 736                 gid, _ := strconv.Atoi(u.Gid)
 737                 cmd.SysProcAttr = &syscall.SysProcAttr{
 738                         Credential: &syscall.Credential{
 739                                 Uid: uint32(uid),
 740                                 Gid: uint32(gid),
 741                         },
 742                 }
 743         }
 744
 745         exited := false
 746         defer func() { exited = true }()
 747         go func() {
 748                 <-ctx.Done()
 749                 log := ctxlog.FromContext(ctx).WithFields(logrus.Fields{"dir": dir, "cmdline": cmdline})
 750                 for !exited {
 751                         if cmd.Process == nil {
 752                                 log.Debug("waiting for child process to start")
 753                                 time.Sleep(time.Second / 2)
 754                         } else {
 755                                 log.WithField("PID", cmd.Process.Pid).Debug("sending SIGTERM")
 756                                 cmd.Process.Signal(syscall.SIGTERM)
 757                                 time.Sleep(5 * time.Second)
 758                                 if !exited {
 759                                         stdout.Close()
 760                                         stderr.Close()
 761                                         log.WithField("PID", cmd.Process.Pid).Warn("still waiting for child process to exit 5s after SIGTERM")
 762                                 }
 763                         }
 764                 }
 765         }()
 766
 767         err = cmd.Start()
 768         if err != nil {
 769                 return err
 770         }
 771         copiers.Wait()
 772         err = cmd.Wait()
 773         if ctx.Err() != nil {
 774                 // Return "context canceled", instead of the "killed"
 775                 // error that was probably caused by the context being
 776                 // canceled.
 777                 return ctx.Err()
 778         } else if err != nil {
 779                 return fmt.Errorf("%s: error: %v", cmdline, err)
 780         }
 781         return nil
 782 }
 783
 784 func (super *Supervisor) autofillConfig() error {
 785         usedPort := map[string]bool{}
 786         nextPort := func(host string) (string, error) {
 787                 for {
 788                         port, err := availablePort(host)
 789                         if err != nil {
 790                                 port, err = availablePort(super.ListenHost)
 791                         }
 792                         if err != nil {
 793                                 return "", err
 794                         }
 795                         if usedPort[port] {
 796                                 continue
 797                         }
 798                         usedPort[port] = true
 799                         return port, nil
 800                 }
 801         }
 802         if super.cluster.Services.Controller.ExternalURL.Host == "" {
 803                 h, p, err := net.SplitHostPort(super.ControllerAddr)
 804                 if err != nil && super.ControllerAddr != "" {
 805                         return fmt.Errorf("SplitHostPort(ControllerAddr %q): %w", super.ControllerAddr, err)
 806                 }
 807                 if h == "" {
 808                         h = super.ListenHost
 809                 }
 810                 if p == "0" || p == "" {
 811                         p, err = nextPort(h)
 812                         if err != nil {
 813                                 return err
 814                         }
 815                 }
 816                 super.cluster.Services.Controller.ExternalURL = arvados.URL{Scheme: "https", Host: net.JoinHostPort(h, p), Path: "/"}
 817         }
 818         u := url.URL(super.cluster.Services.Controller.ExternalURL)
 819         defaultExtHost := u.Hostname()
 820         for _, svc := range []*arvados.Service{
 821                 &super.cluster.Services.Controller,
 822                 &super.cluster.Services.DispatchCloud,
 823                 &super.cluster.Services.Health,
 824                 &super.cluster.Services.Keepproxy,
 825                 &super.cluster.Services.Keepstore,
 826                 &super.cluster.Services.RailsAPI,
 827                 &super.cluster.Services.WebDAV,
 828                 &super.cluster.Services.WebDAVDownload,
 829                 &super.cluster.Services.Websocket,
 830                 &super.cluster.Services.Workbench1,
 831                 &super.cluster.Services.Workbench2,
 832         } {
 833                 if svc.ExternalURL.Host == "" {
 834                         port, err := nextPort(defaultExtHost)
 835                         if err != nil {
 836                                 return err
 837                         }
 838                         host := net.JoinHostPort(defaultExtHost, port)
 839                         if svc == &super.cluster.Services.Controller ||
 840                                 svc == &super.cluster.Services.Health ||
 841                                 svc == &super.cluster.Services.Keepproxy ||
 842                                 svc == &super.cluster.Services.WebDAV ||
 843                                 svc == &super.cluster.Services.WebDAVDownload ||
 844                                 svc == &super.cluster.Services.Workbench1 ||
 845                                 svc == &super.cluster.Services.Workbench2 {
 846                                 svc.ExternalURL = arvados.URL{Scheme: "https", Host: host, Path: "/"}
 847                         } else if svc == &super.cluster.Services.Websocket {
 848                                 svc.ExternalURL = arvados.URL{Scheme: "wss", Host: host, Path: "/websocket"}
 849                         }
 850                 }
 851                 if super.NoWorkbench1 && svc == &super.cluster.Services.Workbench1 ||
 852                         super.NoWorkbench2 && svc == &super.cluster.Services.Workbench2 ||
 853                         !super.cluster.Containers.CloudVMs.Enable && svc == &super.cluster.Services.DispatchCloud {
 854                         // When Workbench is disabled, it gets an
 855                         // ExternalURL (so we have a valid listening
 856                         // port to write in our Nginx config) but no
 857                         // InternalURLs (so health checker doesn't
 858                         // complain).
 859                         continue
 860                 }
 861                 if len(svc.InternalURLs) == 0 {
 862                         port, err := nextPort(super.ListenHost)
 863                         if err != nil {
 864                                 return err
 865                         }
 866                         host := net.JoinHostPort(super.ListenHost, port)
 867                         svc.InternalURLs = map[arvados.URL]arvados.ServiceInstance{
 868                                 {Scheme: "http", Host: host, Path: "/"}: {},
 869                         }
 870                 }
 871         }
 872         if super.ClusterType != "production" {
 873                 if super.cluster.SystemRootToken == "" {
 874                         super.cluster.SystemRootToken = randomHexString(64)
 875                 }
 876                 if super.cluster.ManagementToken == "" {
 877                         super.cluster.ManagementToken = randomHexString(64)
 878                 }
 879                 if super.cluster.Collections.BlobSigningKey == "" {
 880                         super.cluster.Collections.BlobSigningKey = randomHexString(64)
 881                 }
 882                 if super.cluster.Users.AnonymousUserToken == "" {
 883                         super.cluster.Users.AnonymousUserToken = randomHexString(64)
 884                 }
 885                 if super.cluster.Containers.DispatchPrivateKey == "" {
 886                         buf, err := ioutil.ReadFile(filepath.Join(super.SourcePath, "lib", "dispatchcloud", "test", "sshkey_dispatch"))
 887                         if err != nil {
 888                                 return err
 889                         }
 890                         super.cluster.Containers.DispatchPrivateKey = string(buf)
 891                 }
 892                 super.cluster.TLS.Insecure = true
 893         }
 894         if super.ClusterType == "test" {
 895                 // Add a second keepstore process.
 896                 port, err := nextPort(super.ListenHost)
 897                 if err != nil {
 898                         return err
 899                 }
 900                 host := net.JoinHostPort(super.ListenHost, port)
 901                 super.cluster.Services.Keepstore.InternalURLs[arvados.URL{Scheme: "http", Host: host, Path: "/"}] = arvados.ServiceInstance{}
 902
 903                 // Create a directory-backed volume for each keepstore
 904                 // process.
 905                 super.cluster.Volumes = map[string]arvados.Volume{}
 906                 for url := range super.cluster.Services.Keepstore.InternalURLs {
 907                         volnum := len(super.cluster.Volumes)
 908                         datadir := fmt.Sprintf("%s/keep%d.data", super.tempdir, volnum)
 909                         if _, err = os.Stat(datadir + "/."); err == nil {
 910                         } else if !os.IsNotExist(err) {
 911                                 return err
 912                         } else if err = os.Mkdir(datadir, 0755); err != nil {
 913                                 return err
 914                         }
 915                         super.cluster.Volumes[fmt.Sprintf(super.cluster.ClusterID+"-nyw5e-%015d", volnum)] = arvados.Volume{
 916                                 Driver:           "Directory",
 917                                 DriverParameters: json.RawMessage(fmt.Sprintf(`{"Root":%q}`, datadir)),
 918                                 AccessViaHosts: map[arvados.URL]arvados.VolumeAccess{
 919                                         url: {},
 920                                 },
 921                                 StorageClasses: map[string]bool{
 922                                         "default": true,
 923                                         "foo":     true,
 924                                         "bar":     true,
 925                                 },
 926                         }
 927                 }
 928                 super.cluster.StorageClasses = map[string]arvados.StorageClassConfig{
 929                         "default": {Default: true},
 930                         "foo":     {},
 931                         "bar":     {},
 932                 }
 933         }
 934         if super.OwnTemporaryDatabase {
 935                 port, err := nextPort("localhost")
 936                 if err != nil {
 937                         return err
 938                 }
 939                 super.cluster.PostgreSQL.Connection = arvados.PostgreSQLConnection{
 940                         "client_encoding": "utf8",
 941                         "host":            "localhost",
 942                         "port":            port,
 943                         "dbname":          "arvados_test",
 944                         "user":            "arvados",
 945                         "password":        "insecure_arvados_test",
 946                 }
 947         }
 948         return nil
 949 }
 950
 951 func addrIsLocal(addr string) (bool, error) {
 952         if h, _, err := net.SplitHostPort(addr); err != nil {
 953                 return false, err
 954         } else {
 955                 addr = net.JoinHostPort(h, "0")
 956         }
 957         listener, err := net.Listen("tcp", addr)
 958         if err == nil {
 959                 listener.Close()
 960                 return true, nil
 961         } else if strings.Contains(err.Error(), "cannot assign requested address") {
 962                 return false, nil
 963         } else {
 964                 return false, err
 965         }
 966 }
 967
 968 func randomHexString(chars int) string {
 969         b := make([]byte, chars/2)
 970         _, err := rand.Read(b)
 971         if err != nil {
 972                 panic(err)
 973         }
 974         return fmt.Sprintf("%x", b)
 975 }
 976
 977 func internalPort(svc arvados.Service) (host, port string, err error) {
 978         if len(svc.InternalURLs) > 1 {
 979                 return "", "", errors.New("internalPort() doesn't work with multiple InternalURLs")
 980         }
 981         for u := range svc.InternalURLs {
 982                 u := url.URL(u)
 983                 host, port = u.Hostname(), u.Port()
 984                 switch {
 985                 case port != "":
 986                 case u.Scheme == "https", u.Scheme == "ws":
 987                         port = "443"
 988                 default:
 989                         port = "80"
 990                 }
 991                 return
 992         }
 993         return "", "", fmt.Errorf("service has no InternalURLs")
 994 }
 995
 996 func externalPort(svc arvados.Service) (string, error) {
 997         u := url.URL(svc.ExternalURL)
 998         if p := u.Port(); p != "" {
 999                 return p, nil
1000         } else if u.Scheme == "https" || u.Scheme == "wss" {
1001                 return "443", nil
1002         } else {
1003                 return "80", nil
1004         }
1005 }
1006
1007 func availablePort(host string) (string, error) {
1008         ln, err := net.Listen("tcp", net.JoinHostPort(host, "0"))
1009         if err != nil {
1010                 return "", err
1011         }
1012         defer ln.Close()
1013         _, port, err := net.SplitHostPort(ln.Addr().String())
1014         if err != nil {
1015                 return "", err
1016         }
1017         return port, nil
1018 }
1019
1020 // Try to connect to addr until it works, then close ch. Give up if
1021 // ctx cancels.
1022 func waitForConnect(ctx context.Context, addr string) error {
1023         ctxlog.FromContext(ctx).WithField("addr", addr).Info("waitForConnect")
1024         dialer := net.Dialer{Timeout: time.Second}
1025         for ctx.Err() == nil {
1026                 conn, err := dialer.DialContext(ctx, "tcp", addr)
1027                 if err != nil {
1028                         time.Sleep(time.Second / 10)
1029                         continue
1030                 }
1031                 conn.Close()
1032                 return nil
1033         }
1034         return ctx.Err()
1035 }
1036
1037 func copyConfig(cfg *arvados.Config) *arvados.Config {
1038         pr, pw := io.Pipe()
1039         go func() {
1040                 err := json.NewEncoder(pw).Encode(cfg)
1041                 if err != nil {
1042                         panic(err)
1043                 }
1044                 pw.Close()
1045         }()
1046         cfg2 := new(arvados.Config)
1047         err := json.NewDecoder(pr).Decode(cfg2)
1048         if err != nil {
1049                 panic(err)
1050         }
1051         return cfg2
1052 }
1053
1054 func watchConfig(ctx context.Context, logger logrus.FieldLogger, cfgPath string, prevcfg *arvados.Config, fn func()) {
1055         watcher, err := fsnotify.NewWatcher()
1056         if err != nil {
1057                 logger.WithError(err).Error("fsnotify setup failed")
1058                 return
1059         }
1060         defer watcher.Close()
1061
1062         err = watcher.Add(cfgPath)
1063         if err != nil {
1064                 logger.WithError(err).Error("fsnotify watcher failed")
1065                 return
1066         }
1067
1068         for {
1069                 select {
1070                 case <-ctx.Done():
1071                         return
1072                 case err, ok := <-watcher.Errors:
1073                         if !ok {
1074                                 return
1075                         }
1076                         logger.WithError(err).Warn("fsnotify watcher reported error")
1077                 case _, ok := <-watcher.Events:
1078                         if !ok {
1079                                 return
1080                         }
1081                         for len(watcher.Events) > 0 {
1082                                 <-watcher.Events
1083                         }
1084                         loader := config.NewLoader(&bytes.Buffer{}, &logrus.Logger{Out: ioutil.Discard})
1085                         loader.Path = cfgPath
1086                         loader.SkipAPICalls = true
1087                         cfg, err := loader.Load()
1088                         if err != nil {
1089                                 logger.WithError(err).Warn("error reloading config file after change detected; ignoring new config for now")
1090                         } else if reflect.DeepEqual(cfg, prevcfg) {
1091                                 logger.Debug("config file changed but is still DeepEqual to the existing config")
1092                         } else {
1093                                 logger.Debug("config changed, notifying supervisor")
1094                                 fn()
1095                                 prevcfg = cfg
1096                         }
1097                 }
1098         }
1099 }