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         if super.ClusterType == "test" {
 321                 super.setEnv("ARVADOS_USE_KEEP_ACCESSIBLE_API", "true")
 322         }
 323
 324         // Now that we have the config, replace the bootstrap logger
 325         // with a new one according to the logging config.
 326         loglevel := super.cluster.SystemLogs.LogLevel
 327         if s := os.Getenv("ARVADOS_DEBUG"); s != "" && s != "0" {
 328                 loglevel = "debug"
 329         }
 330         super.logger = ctxlog.New(super.Stderr, super.cluster.SystemLogs.Format, loglevel).WithFields(logrus.Fields{
 331                 "PID": os.Getpid(),
 332         })
 333
 334         if super.SourceVersion == "" && super.ClusterType == "production" {
 335                 // don't need SourceVersion
 336         } else if super.SourceVersion == "" {
 337                 // Find current source tree version.
 338                 var buf bytes.Buffer
 339                 err = super.RunProgram(super.ctx, super.SourcePath, runOptions{output: &buf}, "git", "diff", "--shortstat")
 340                 if err != nil {
 341                         return err
 342                 }
 343                 dirty := buf.Len() > 0
 344                 buf.Reset()
 345                 err = super.RunProgram(super.ctx, super.SourcePath, runOptions{output: &buf}, "git", "log", "-n1", "--format=%H")
 346                 if err != nil {
 347                         return err
 348                 }
 349                 super.SourceVersion = strings.TrimSpace(buf.String())
 350                 if dirty {
 351                         super.SourceVersion += "+uncommitted"
 352                 }
 353         } else {
 354                 return errors.New("specifying a version to run is not yet supported")
 355         }
 356
 357         _, err = super.installGoProgram(super.ctx, "cmd/arvados-server")
 358         if err != nil {
 359                 return err
 360         }
 361         err = super.setupRubyEnv()
 362         if err != nil {
 363                 return err
 364         }
 365
 366         tasks := []supervisedTask{
 367                 createCertificates{},
 368                 runPostgreSQL{},
 369                 runNginx{},
 370                 railsDatabase{},
 371                 runServiceCommand{name: "controller", svc: super.cluster.Services.Controller, depends: []supervisedTask{railsDatabase{}}},
 372                 runServiceCommand{name: "health", svc: super.cluster.Services.Health},
 373                 runServiceCommand{name: "keepproxy", svc: super.cluster.Services.Keepproxy, depends: []supervisedTask{runPassenger{src: "services/api"}}},
 374                 runServiceCommand{name: "keepstore", svc: super.cluster.Services.Keepstore},
 375                 runServiceCommand{name: "keep-web", svc: super.cluster.Services.WebDAV},
 376                 runServiceCommand{name: "ws", svc: super.cluster.Services.Websocket, depends: []supervisedTask{railsDatabase{}}},
 377                 installPassenger{src: "services/api", varlibdir: "railsapi"},
 378                 runPassenger{src: "services/api", varlibdir: "railsapi", svc: super.cluster.Services.RailsAPI, depends: []supervisedTask{
 379                         createCertificates{},
 380                         installPassenger{src: "services/api", varlibdir: "railsapi"},
 381                         railsDatabase{},
 382                 }},
 383         }
 384         if !super.NoWorkbench1 {
 385                 return errors.New("workbench1 is no longer supported")
 386         }
 387         if !super.NoWorkbench2 {
 388                 tasks = append(tasks,
 389                         runWorkbench2{svc: super.cluster.Services.Workbench2},
 390                 )
 391         }
 392         if super.ClusterType != "test" {
 393                 tasks = append(tasks,
 394                         runServiceCommand{name: "keep-balance", svc: super.cluster.Services.Keepbalance},
 395                 )
 396         }
 397         if super.cluster.Containers.CloudVMs.Enable {
 398                 tasks = append(tasks,
 399                         runServiceCommand{name: "dispatch-cloud", svc: super.cluster.Services.DispatchCloud},
 400                 )
 401         }
 402         super.tasksReady = map[string]chan bool{}
 403         for _, task := range tasks {
 404                 super.tasksReady[task.String()] = make(chan bool)
 405         }
 406         for _, task := range tasks {
 407                 task := task
 408                 fail := func(err error) {
 409                         if super.ctx.Err() != nil {
 410                                 return
 411                         }
 412                         super.cancel()
 413                         super.logger.WithField("task", task.String()).WithError(err).Error("task failed")
 414                 }
 415                 go func() {
 416                         super.logger.WithField("task", task.String()).Info("starting")
 417                         err := task.Run(super.ctx, fail, super)
 418                         if err != nil {
 419                                 fail(err)
 420                                 return
 421                         }
 422                         close(super.tasksReady[task.String()])
 423                 }()
 424         }
 425         err = super.wait(super.ctx, tasks...)
 426         if err != nil {
 427                 return err
 428         }
 429         super.logger.Info("all startup tasks are complete; starting health checks")
 430         super.healthChecker = &health.Aggregator{Cluster: super.cluster}
 431         <-super.ctx.Done()
 432         super.logger.Info("shutting down")
 433         super.waitShutdown.Wait()
 434         return super.ctx.Err()
 435 }
 436
 437 func (super *Supervisor) wait(ctx context.Context, tasks ...supervisedTask) error {
 438         ticker := time.NewTicker(15 * time.Second)
 439         defer ticker.Stop()
 440         for _, task := range tasks {
 441                 ch, ok := super.tasksReady[task.String()]
 442                 if !ok {
 443                         return fmt.Errorf("no such task: %s", task)
 444                 }
 445                 super.logger.WithField("task", task.String()).Info("waiting")
 446                 for {
 447                         select {
 448                         case <-ch:
 449                                 super.logger.WithField("task", task.String()).Info("ready")
 450                         case <-ctx.Done():
 451                                 super.logger.WithField("task", task.String()).Info("task was never ready")
 452                                 return ctx.Err()
 453                         case <-ticker.C:
 454                                 super.logger.WithField("task", task.String()).Info("still waiting...")
 455                                 continue
 456                         }
 457                         break
 458                 }
 459         }
 460         return nil
 461 }
 462
 463 // Stop shuts down all child processes and goroutines, and returns
 464 // when all of them have exited.
 465 func (super *Supervisor) Stop() {
 466         super.cancel()
 467         <-super.done
 468 }
 469
 470 // WaitReady waits for the cluster(s) to be ready to handle requests,
 471 // then returns true. If startup fails, it returns false.
 472 func (super *Supervisor) WaitReady() bool {
 473         if super.children != nil {
 474                 for id, super2 := range super.children {
 475                         super.logger.Infof("waiting for %s to be ready", id)
 476                         if !super2.WaitReady() {
 477                                 super.logger.Infof("%s startup failed", id)
 478                                 super.Stop()
 479                                 return false
 480                         }
 481                         super.logger.Infof("%s is ready", id)
 482                 }
 483                 super.logger.Info("all clusters are ready")
 484                 return true
 485         }
 486         ticker := time.NewTicker(time.Second)
 487         defer ticker.Stop()
 488         for waiting := "all"; waiting != ""; {
 489                 select {
 490                 case <-ticker.C:
 491                 case <-super.ctx.Done():
 492                         super.Stop()
 493                         return false
 494                 }
 495                 if super.healthChecker == nil {
 496                         // not set up yet
 497                         continue
 498                 }
 499                 resp := super.healthChecker.ClusterHealth()
 500                 // The overall health check (resp.Health=="OK") might
 501                 // never pass due to missing components (like
 502                 // arvados-dispatch-cloud in a test cluster), so
 503                 // instead we wait for all configured components to
 504                 // pass.
 505                 waiting = ""
 506                 for target, check := range resp.Checks {
 507                         if check.Health != "OK" {
 508                                 waiting += " " + target
 509                         }
 510                 }
 511                 if waiting != "" {
 512                         super.logger.WithField("targets", waiting[1:]).Info("waiting")
 513                 }
 514         }
 515         return true
 516 }
 517
 518 func (super *Supervisor) prependEnv(key, prepend string) {
 519         for i, s := range super.environ {
 520                 if strings.HasPrefix(s, key+"=") {
 521                         super.environ[i] = key + "=" + prepend + s[len(key)+1:]
 522                         return
 523                 }
 524         }
 525         super.environ = append(super.environ, key+"="+prepend)
 526 }
 527
 528 func (super *Supervisor) cleanEnv(prefixes []string) {
 529         var cleaned []string
 530         for _, s := range super.environ {
 531                 drop := false
 532                 for _, p := range prefixes {
 533                         if strings.HasPrefix(s, p) {
 534                                 drop = true
 535                                 break
 536                         }
 537                 }
 538                 if !drop {
 539                         cleaned = append(cleaned, s)
 540                 }
 541         }
 542         super.environ = cleaned
 543 }
 544
 545 func (super *Supervisor) setEnv(key, val string) {
 546         for i, s := range super.environ {
 547                 if strings.HasPrefix(s, key+"=") {
 548                         super.environ[i] = key + "=" + val
 549                         return
 550                 }
 551         }
 552         super.environ = append(super.environ, key+"="+val)
 553 }
 554
 555 // Remove all but the first occurrence of each env var.
 556 func dedupEnv(in []string) []string {
 557         saw := map[string]bool{}
 558         var out []string
 559         for _, kv := range in {
 560                 if split := strings.Index(kv, "="); split < 1 {
 561                         panic("invalid environment var: " + kv)
 562                 } else if saw[kv[:split]] {
 563                         continue
 564                 } else {
 565                         saw[kv[:split]] = true
 566                         out = append(out, kv)
 567                 }
 568         }
 569         return out
 570 }
 571
 572 func (super *Supervisor) installGoProgram(ctx context.Context, srcpath string) (string, error) {
 573         _, basename := filepath.Split(srcpath)
 574         binfile := filepath.Join(super.bindir, basename)
 575         if super.ClusterType == "production" {
 576                 return binfile, nil
 577         }
 578         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)
 579         return binfile, err
 580 }
 581
 582 func (super *Supervisor) usingRVM() bool {
 583         return os.Getenv("rvm_path") != ""
 584 }
 585
 586 func (super *Supervisor) setupRubyEnv() error {
 587         if !super.usingRVM() {
 588                 // (If rvm is in use, assume the caller has everything
 589                 // set up as desired)
 590                 super.cleanEnv([]string{
 591                         "GEM_HOME=",
 592                         "GEM_PATH=",
 593                 })
 594                 gem := "gem"
 595                 if _, err := os.Stat("/var/lib/arvados/bin/gem"); err == nil || super.ClusterType == "production" {
 596                         gem = "/var/lib/arvados/bin/gem"
 597                 }
 598                 cmd := exec.Command(gem, "env", "gempath")
 599                 if super.ClusterType == "production" {
 600                         cmd.Args = append([]string{"sudo", "-u", "www-data", "-E", "HOME=/var/www"}, cmd.Args...)
 601                         path, err := exec.LookPath("sudo")
 602                         if err != nil {
 603                                 return fmt.Errorf("LookPath(\"sudo\"): %w", err)
 604                         }
 605                         cmd.Path = path
 606                 }
 607                 cmd.Stderr = super.Stderr
 608                 cmd.Env = super.environ
 609                 buf, err := cmd.Output() // /var/lib/arvados/.gem/ruby/2.5.0/bin:...
 610                 if err != nil || len(buf) == 0 {
 611                         return fmt.Errorf("gem env gempath: %w", err)
 612                 }
 613                 gempath := string(bytes.Split(buf, []byte{':'})[0])
 614                 super.prependEnv("PATH", gempath+"/bin:")
 615                 super.setEnv("GEM_HOME", gempath)
 616                 super.setEnv("GEM_PATH", gempath)
 617         }
 618         // Passenger install doesn't work unless $HOME is ~user
 619         u, err := user.Current()
 620         if err != nil {
 621                 return err
 622         }
 623         super.setEnv("HOME", u.HomeDir)
 624         return nil
 625 }
 626
 627 func (super *Supervisor) lookPath(prog string) string {
 628         for _, val := range super.environ {
 629                 if strings.HasPrefix(val, "PATH=") {
 630                         for _, dir := range filepath.SplitList(val[5:]) {
 631                                 path := filepath.Join(dir, prog)
 632                                 if fi, err := os.Stat(path); err == nil && fi.Mode()&0111 != 0 {
 633                                         return path
 634                                 }
 635                         }
 636                 }
 637         }
 638         return prog
 639 }
 640
 641 type runOptions struct {
 642         output io.Writer // attach stdout
 643         env    []string  // add/replace environment variables
 644         user   string    // run as specified user
 645         stdin  io.Reader
 646 }
 647
 648 // RunProgram runs prog with args, using dir as working directory. If ctx is
 649 // cancelled while the child is running, RunProgram terminates the child, waits
 650 // for it to exit, then returns.
 651 //
 652 // Child's environment will have our env vars, plus any given in env.
 653 //
 654 // Child's stdout will be written to output if non-nil, otherwise the
 655 // boot command's stderr.
 656 func (super *Supervisor) RunProgram(ctx context.Context, dir string, opts runOptions, prog string, args ...string) error {
 657         cmdline := fmt.Sprintf("%s", append([]string{prog}, args...))
 658         super.logger.WithField("command", cmdline).WithField("dir", dir).Info("executing")
 659
 660         logprefix := prog
 661         {
 662                 innerargs := args
 663                 if logprefix == "sudo" {
 664                         for i := 0; i < len(args); i++ {
 665                                 if args[i] == "-u" {
 666                                         i++
 667                                 } else if args[i] == "-E" || strings.Contains(args[i], "=") {
 668                                 } else {
 669                                         logprefix = args[i]
 670                                         innerargs = args[i+1:]
 671                                         break
 672                                 }
 673                         }
 674                 }
 675                 logprefix = strings.TrimPrefix(logprefix, "/var/lib/arvados/bin/")
 676                 logprefix = strings.TrimPrefix(logprefix, super.tempdir+"/bin/")
 677                 if logprefix == "bundle" && len(innerargs) > 2 && innerargs[0] == "exec" {
 678                         _, dirbase := filepath.Split(dir)
 679                         logprefix = innerargs[1] + "@" + dirbase
 680                 } else if logprefix == "arvados-server" && len(args) > 1 {
 681                         logprefix = args[0]
 682                 }
 683                 if !strings.HasPrefix(dir, "/") {
 684                         logprefix = dir + ": " + logprefix
 685                 }
 686         }
 687
 688         cmd := exec.Command(super.lookPath(prog), args...)
 689         cmd.Stdin = opts.stdin
 690         stdout, err := cmd.StdoutPipe()
 691         if err != nil {
 692                 return err
 693         }
 694         stderr, err := cmd.StderrPipe()
 695         if err != nil {
 696                 return err
 697         }
 698         logwriter := &service.LogPrefixer{Writer: super.Stderr, Prefix: []byte("[" + logprefix + "] ")}
 699         var copiers sync.WaitGroup
 700         copiers.Add(1)
 701         go func() {
 702                 io.Copy(logwriter, stderr)
 703                 copiers.Done()
 704         }()
 705         copiers.Add(1)
 706         go func() {
 707                 if opts.output == nil {
 708                         io.Copy(logwriter, stdout)
 709                 } else {
 710                         io.Copy(opts.output, stdout)
 711                 }
 712                 copiers.Done()
 713         }()
 714
 715         if strings.HasPrefix(dir, "/") {
 716                 cmd.Dir = dir
 717         } else {
 718                 cmd.Dir = filepath.Join(super.SourcePath, dir)
 719         }
 720         env := append([]string(nil), opts.env...)
 721         env = append(env, super.environ...)
 722         cmd.Env = dedupEnv(env)
 723
 724         if opts.user != "" {
 725                 // Note: We use this approach instead of "sudo"
 726                 // because in certain circumstances (we are pid 1 in a
 727                 // docker container, and our passenger child process
 728                 // changes to pgid 1) the intermediate sudo process
 729                 // notices we have the same pgid as our child and
 730                 // refuses to propagate signals from us to our child,
 731                 // so we can't signal/shutdown our passenger/rails
 732                 // apps. "chpst" or "setuidgid" would work, but these
 733                 // few lines avoid depending on runit/daemontools.
 734                 u, err := user.Lookup(opts.user)
 735                 if err != nil {
 736                         return fmt.Errorf("user.Lookup(%q): %w", opts.user, err)
 737                 }
 738                 uid, _ := strconv.Atoi(u.Uid)
 739                 gid, _ := strconv.Atoi(u.Gid)
 740                 cmd.SysProcAttr = &syscall.SysProcAttr{
 741                         Credential: &syscall.Credential{
 742                                 Uid: uint32(uid),
 743                                 Gid: uint32(gid),
 744                         },
 745                 }
 746         }
 747
 748         exited := false
 749         defer func() { exited = true }()
 750         go func() {
 751                 <-ctx.Done()
 752                 log := ctxlog.FromContext(ctx).WithFields(logrus.Fields{"dir": dir, "cmdline": cmdline})
 753                 for !exited {
 754                         if cmd.Process == nil {
 755                                 log.Debug("waiting for child process to start")
 756                                 time.Sleep(time.Second / 2)
 757                         } else {
 758                                 log.WithField("PID", cmd.Process.Pid).Debug("sending SIGTERM")
 759                                 cmd.Process.Signal(syscall.SIGTERM)
 760                                 time.Sleep(5 * time.Second)
 761                                 if !exited {
 762                                         stdout.Close()
 763                                         stderr.Close()
 764                                         log.WithField("PID", cmd.Process.Pid).Warn("still waiting for child process to exit 5s after SIGTERM")
 765                                 }
 766                         }
 767                 }
 768         }()
 769
 770         err = cmd.Start()
 771         if err != nil {
 772                 return err
 773         }
 774         copiers.Wait()
 775         err = cmd.Wait()
 776         if ctx.Err() != nil {
 777                 // Return "context canceled", instead of the "killed"
 778                 // error that was probably caused by the context being
 779                 // canceled.
 780                 return ctx.Err()
 781         } else if err != nil {
 782                 return fmt.Errorf("%s: error: %v", cmdline, err)
 783         }
 784         return nil
 785 }
 786
 787 func (super *Supervisor) autofillConfig() error {
 788         usedPort := map[string]bool{}
 789         nextPort := func(host string) (string, error) {
 790                 for {
 791                         port, err := availablePort(host)
 792                         if err != nil {
 793                                 port, err = availablePort(super.ListenHost)
 794                         }
 795                         if err != nil {
 796                                 return "", err
 797                         }
 798                         if usedPort[port] {
 799                                 continue
 800                         }
 801                         usedPort[port] = true
 802                         return port, nil
 803                 }
 804         }
 805         if super.cluster.Services.Controller.ExternalURL.Host == "" {
 806                 h, p, err := net.SplitHostPort(super.ControllerAddr)
 807                 if err != nil && super.ControllerAddr != "" {
 808                         return fmt.Errorf("SplitHostPort(ControllerAddr %q): %w", super.ControllerAddr, err)
 809                 }
 810                 if h == "" {
 811                         h = super.ListenHost
 812                 }
 813                 if p == "0" || p == "" {
 814                         p, err = nextPort(h)
 815                         if err != nil {
 816                                 return err
 817                         }
 818                 }
 819                 super.cluster.Services.Controller.ExternalURL = arvados.URL{Scheme: "https", Host: net.JoinHostPort(h, p), Path: "/"}
 820         }
 821         u := url.URL(super.cluster.Services.Controller.ExternalURL)
 822         defaultExtHost := u.Hostname()
 823         for _, svc := range []*arvados.Service{
 824                 &super.cluster.Services.Controller,
 825                 &super.cluster.Services.DispatchCloud,
 826                 &super.cluster.Services.Health,
 827                 &super.cluster.Services.Keepproxy,
 828                 &super.cluster.Services.Keepstore,
 829                 &super.cluster.Services.RailsAPI,
 830                 &super.cluster.Services.WebDAV,
 831                 &super.cluster.Services.WebDAVDownload,
 832                 &super.cluster.Services.Websocket,
 833                 &super.cluster.Services.Workbench1,
 834                 &super.cluster.Services.Workbench2,
 835         } {
 836                 if svc.ExternalURL.Host == "" {
 837                         port, err := nextPort(defaultExtHost)
 838                         if err != nil {
 839                                 return err
 840                         }
 841                         host := net.JoinHostPort(defaultExtHost, port)
 842                         if svc == &super.cluster.Services.Controller ||
 843                                 svc == &super.cluster.Services.Health ||
 844                                 svc == &super.cluster.Services.Keepproxy ||
 845                                 svc == &super.cluster.Services.WebDAV ||
 846                                 svc == &super.cluster.Services.WebDAVDownload ||
 847                                 svc == &super.cluster.Services.Workbench1 ||
 848                                 svc == &super.cluster.Services.Workbench2 {
 849                                 svc.ExternalURL = arvados.URL{Scheme: "https", Host: host, Path: "/"}
 850                         } else if svc == &super.cluster.Services.Websocket {
 851                                 svc.ExternalURL = arvados.URL{Scheme: "wss", Host: host, Path: "/websocket"}
 852                         }
 853                 }
 854                 if super.NoWorkbench1 && svc == &super.cluster.Services.Workbench1 ||
 855                         super.NoWorkbench2 && svc == &super.cluster.Services.Workbench2 ||
 856                         !super.cluster.Containers.CloudVMs.Enable && svc == &super.cluster.Services.DispatchCloud {
 857                         // When Workbench is disabled, it gets an
 858                         // ExternalURL (so we have a valid listening
 859                         // port to write in our Nginx config) but no
 860                         // InternalURLs (so health checker doesn't
 861                         // complain).
 862                         continue
 863                 }
 864                 if len(svc.InternalURLs) == 0 {
 865                         port, err := nextPort(super.ListenHost)
 866                         if err != nil {
 867                                 return err
 868                         }
 869                         host := net.JoinHostPort(super.ListenHost, port)
 870                         svc.InternalURLs = map[arvados.URL]arvados.ServiceInstance{
 871                                 {Scheme: "http", Host: host, Path: "/"}: {},
 872                         }
 873                 }
 874         }
 875         if super.ClusterType != "production" {
 876                 if super.cluster.SystemRootToken == "" {
 877                         super.cluster.SystemRootToken = randomHexString(64)
 878                 }
 879                 if super.cluster.ManagementToken == "" {
 880                         super.cluster.ManagementToken = randomHexString(64)
 881                 }
 882                 if super.cluster.Collections.BlobSigningKey == "" {
 883                         super.cluster.Collections.BlobSigningKey = randomHexString(64)
 884                 }
 885                 if super.cluster.Users.AnonymousUserToken == "" {
 886                         super.cluster.Users.AnonymousUserToken = randomHexString(64)
 887                 }
 888                 if super.cluster.Containers.DispatchPrivateKey == "" {
 889                         buf, err := ioutil.ReadFile(filepath.Join(super.SourcePath, "lib", "dispatchcloud", "test", "sshkey_dispatch"))
 890                         if err != nil {
 891                                 return err
 892                         }
 893                         super.cluster.Containers.DispatchPrivateKey = string(buf)
 894                 }
 895                 super.cluster.TLS.Insecure = true
 896         }
 897         if super.ClusterType == "test" {
 898                 // Add a second keepstore process.
 899                 port, err := nextPort(super.ListenHost)
 900                 if err != nil {
 901                         return err
 902                 }
 903                 host := net.JoinHostPort(super.ListenHost, port)
 904                 super.cluster.Services.Keepstore.InternalURLs[arvados.URL{Scheme: "http", Host: host, Path: "/"}] = arvados.ServiceInstance{}
 905
 906                 // Create a directory-backed volume for each keepstore
 907                 // process.
 908                 super.cluster.Volumes = map[string]arvados.Volume{}
 909                 for url := range super.cluster.Services.Keepstore.InternalURLs {
 910                         volnum := len(super.cluster.Volumes)
 911                         datadir := fmt.Sprintf("%s/keep%d.data", super.tempdir, volnum)
 912                         if _, err = os.Stat(datadir + "/."); err == nil {
 913                         } else if !os.IsNotExist(err) {
 914                                 return err
 915                         } else if err = os.Mkdir(datadir, 0755); err != nil {
 916                                 return err
 917                         }
 918                         super.cluster.Volumes[fmt.Sprintf(super.cluster.ClusterID+"-nyw5e-%015d", volnum)] = arvados.Volume{
 919                                 Driver:           "Directory",
 920                                 DriverParameters: json.RawMessage(fmt.Sprintf(`{"Root":%q}`, datadir)),
 921                                 AccessViaHosts: map[arvados.URL]arvados.VolumeAccess{
 922                                         url: {},
 923                                 },
 924                                 StorageClasses: map[string]bool{
 925                                         "default": true,
 926                                         "foo":     true,
 927                                         "bar":     true,
 928                                 },
 929                         }
 930                 }
 931                 super.cluster.StorageClasses = map[string]arvados.StorageClassConfig{
 932                         "default": {Default: true},
 933                         "foo":     {},
 934                         "bar":     {},
 935                 }
 936         }
 937         if super.OwnTemporaryDatabase {
 938                 port, err := nextPort("localhost")
 939                 if err != nil {
 940                         return err
 941                 }
 942                 super.cluster.PostgreSQL.Connection = arvados.PostgreSQLConnection{
 943                         "client_encoding": "utf8",
 944                         "host":            "localhost",
 945                         "port":            port,
 946                         "dbname":          "arvados_test",
 947                         "user":            "arvados",
 948                         "password":        "insecure_arvados_test",
 949                 }
 950         }
 951         return nil
 952 }
 953
 954 func addrIsLocal(addr string) (bool, error) {
 955         if h, _, err := net.SplitHostPort(addr); err != nil {
 956                 return false, err
 957         } else {
 958                 addr = net.JoinHostPort(h, "0")
 959         }
 960         listener, err := net.Listen("tcp", addr)
 961         if err == nil {
 962                 listener.Close()
 963                 return true, nil
 964         } else if strings.Contains(err.Error(), "cannot assign requested address") {
 965                 return false, nil
 966         } else {
 967                 return false, err
 968         }
 969 }
 970
 971 func randomHexString(chars int) string {
 972         b := make([]byte, chars/2)
 973         _, err := rand.Read(b)
 974         if err != nil {
 975                 panic(err)
 976         }
 977         return fmt.Sprintf("%x", b)
 978 }
 979
 980 func internalPort(svc arvados.Service) (host, port string, err error) {
 981         if len(svc.InternalURLs) > 1 {
 982                 return "", "", errors.New("internalPort() doesn't work with multiple InternalURLs")
 983         }
 984         for u := range svc.InternalURLs {
 985                 u := url.URL(u)
 986                 host, port = u.Hostname(), u.Port()
 987                 switch {
 988                 case port != "":
 989                 case u.Scheme == "https", u.Scheme == "ws":
 990                         port = "443"
 991                 default:
 992                         port = "80"
 993                 }
 994                 return
 995         }
 996         return "", "", fmt.Errorf("service has no InternalURLs")
 997 }
 998
 999 func externalPort(svc arvados.Service) (string, error) {
1000         u := url.URL(svc.ExternalURL)
1001         if p := u.Port(); p != "" {
1002                 return p, nil
1003         } else if u.Scheme == "https" || u.Scheme == "wss" {
1004                 return "443", nil
1005         } else {
1006                 return "80", nil
1007         }
1008 }
1009
1010 func availablePort(host string) (string, error) {
1011         ln, err := net.Listen("tcp", net.JoinHostPort(host, "0"))
1012         if err != nil {
1013                 return "", err
1014         }
1015         defer ln.Close()
1016         _, port, err := net.SplitHostPort(ln.Addr().String())
1017         if err != nil {
1018                 return "", err
1019         }
1020         return port, nil
1021 }
1022
1023 // Try to connect to addr until it works, then close ch. Give up if
1024 // ctx cancels.
1025 func waitForConnect(ctx context.Context, addr string) error {
1026         ctxlog.FromContext(ctx).WithField("addr", addr).Info("waitForConnect")
1027         dialer := net.Dialer{Timeout: time.Second}
1028         for ctx.Err() == nil {
1029                 conn, err := dialer.DialContext(ctx, "tcp", addr)
1030                 if err != nil {
1031                         time.Sleep(time.Second / 10)
1032                         continue
1033                 }
1034                 conn.Close()
1035                 return nil
1036         }
1037         return ctx.Err()
1038 }
1039
1040 func copyConfig(cfg *arvados.Config) *arvados.Config {
1041         pr, pw := io.Pipe()
1042         go func() {
1043                 err := json.NewEncoder(pw).Encode(cfg)
1044                 if err != nil {
1045                         panic(err)
1046                 }
1047                 pw.Close()
1048         }()
1049         cfg2 := new(arvados.Config)
1050         err := json.NewDecoder(pr).Decode(cfg2)
1051         if err != nil {
1052                 panic(err)
1053         }
1054         return cfg2
1055 }
1056
1057 func watchConfig(ctx context.Context, logger logrus.FieldLogger, cfgPath string, prevcfg *arvados.Config, fn func()) {
1058         watcher, err := fsnotify.NewWatcher()
1059         if err != nil {
1060                 logger.WithError(err).Error("fsnotify setup failed")
1061                 return
1062         }
1063         defer watcher.Close()
1064
1065         err = watcher.Add(cfgPath)
1066         if err != nil {
1067                 logger.WithError(err).Error("fsnotify watcher failed")
1068                 return
1069         }
1070
1071         for {
1072                 select {
1073                 case <-ctx.Done():
1074                         return
1075                 case err, ok := <-watcher.Errors:
1076                         if !ok {
1077                                 return
1078                         }
1079                         logger.WithError(err).Warn("fsnotify watcher reported error")
1080                 case _, ok := <-watcher.Events:
1081                         if !ok {
1082                                 return
1083                         }
1084                         for len(watcher.Events) > 0 {
1085                                 <-watcher.Events
1086                         }
1087                         loader := config.NewLoader(&bytes.Buffer{}, &logrus.Logger{Out: ioutil.Discard})
1088                         loader.Path = cfgPath
1089                         loader.SkipAPICalls = true
1090                         cfg, err := loader.Load()
1091                         if err != nil {
1092                                 logger.WithError(err).Warn("error reloading config file after change detected; ignoring new config for now")
1093                         } else if reflect.DeepEqual(cfg, prevcfg) {
1094                                 logger.Debug("config file changed but is still DeepEqual to the existing config")
1095                         } else {
1096                                 logger.Debug("config changed, notifying supervisor")
1097                                 fn()
1098                                 prevcfg = cfg
1099                         }
1100                 }
1101         }
1102 }