services/crunch-dispatch-local/crunch-dispatch-local.go

   1 // Copyright (C) The Arvados Authors. All rights reserved.
   2 //
   3 // SPDX-License-Identifier: AGPL-3.0
   4
   5 package main
   6
   7 // Dispatcher service for Crunch that runs containers locally.
   8
   9 import (
  10         "context"
  11         "flag"
  12         "fmt"
  13         "os"
  14         "os/exec"
  15         "os/signal"
  16         "sync"
  17         "syscall"
  18         "time"
  19
  20         "git.arvados.org/arvados.git/lib/config"
  21         "git.arvados.org/arvados.git/sdk/go/arvados"
  22         "git.arvados.org/arvados.git/sdk/go/arvadosclient"
  23         "git.arvados.org/arvados.git/sdk/go/dispatch"
  24         "github.com/sirupsen/logrus"
  25 )
  26
  27 var version = "dev"
  28
  29 func main() {
  30         err := doMain()
  31         if err != nil {
  32                 logrus.Fatalf("%q", err)
  33         }
  34 }
  35
  36 var (
  37         runningCmds      map[string]*exec.Cmd
  38         runningCmdsMutex sync.Mutex
  39         waitGroup        sync.WaitGroup
  40         crunchRunCommand *string
  41 )
  42
  43 func doMain() error {
  44         logger := logrus.StandardLogger()
  45         if os.Getenv("DEBUG") != "" {
  46                 logger.SetLevel(logrus.DebugLevel)
  47         }
  48         logger.Formatter = &logrus.JSONFormatter{
  49                 TimestampFormat: "2006-01-02T15:04:05.000000000Z07:00",
  50         }
  51
  52         flags := flag.NewFlagSet("crunch-dispatch-local", flag.ExitOnError)
  53
  54         pollInterval := flags.Int(
  55                 "poll-interval",
  56                 10,
  57                 "Interval in seconds to poll for queued containers")
  58
  59         crunchRunCommand = flags.String(
  60                 "crunch-run-command",
  61                 "/usr/bin/crunch-run",
  62                 "Crunch command to run container")
  63
  64         getVersion := flags.Bool(
  65                 "version",
  66                 false,
  67                 "Print version information and exit.")
  68
  69         // Parse args; omit the first arg which is the command name
  70         err := flags.Parse(os.Args[1:])
  71         if err == flag.ErrHelp {
  72                 return nil
  73         } else if err != nil {
  74                 return err
  75         } else if flags.NArg() != 0 {
  76                 return fmt.Errorf("unrecognized command line arguments: %v", flags.Args())
  77         }
  78
  79         // Print version information if requested
  80         if *getVersion {
  81                 fmt.Printf("crunch-dispatch-local %s\n", version)
  82                 return nil
  83         }
  84
  85         loader := config.NewLoader(nil, logger)
  86         cfg, err := loader.Load()
  87         cluster, err := cfg.GetCluster("")
  88         if err != nil {
  89                 return fmt.Errorf("config error: %s", err)
  90         }
  91
  92         logger.Printf("crunch-dispatch-local %s started", version)
  93
  94         runningCmds = make(map[string]*exec.Cmd)
  95
  96         var client arvados.Client
  97         client.APIHost = cluster.Services.Controller.ExternalURL.Host
  98         client.AuthToken = cluster.SystemRootToken
  99         client.Insecure = cluster.TLS.Insecure
 100
 101         if client.APIHost != "" || client.AuthToken != "" {
 102                 // Copy real configs into env vars so [a]
 103                 // MakeArvadosClient() uses them, and [b] they get
 104                 // propagated to crunch-run via SLURM.
 105                 os.Setenv("ARVADOS_API_HOST", client.APIHost)
 106                 os.Setenv("ARVADOS_API_TOKEN", client.AuthToken)
 107                 os.Setenv("ARVADOS_API_HOST_INSECURE", "")
 108                 if client.Insecure {
 109                         os.Setenv("ARVADOS_API_HOST_INSECURE", "1")
 110                 }
 111                 os.Setenv("ARVADOS_EXTERNAL_CLIENT", "")
 112         } else {
 113                 logger.Warnf("Client credentials missing from config, so falling back on environment variables (deprecated).")
 114         }
 115
 116         arv, err := arvadosclient.MakeArvadosClient()
 117         if err != nil {
 118                 logger.Errorf("error making Arvados client: %v", err)
 119                 return err
 120         }
 121         arv.Retries = 25
 122
 123         ctx, cancel := context.WithCancel(context.Background())
 124
 125         dispatcher := dispatch.Dispatcher{
 126                 Logger:       logger,
 127                 Arv:          arv,
 128                 RunContainer: (&LocalRun{startFunc, make(chan bool, 8), ctx, cluster}).run,
 129                 PollPeriod:   time.Duration(*pollInterval) * time.Second,
 130         }
 131
 132         err = dispatcher.Run(ctx)
 133         if err != nil {
 134                 return err
 135         }
 136
 137         c := make(chan os.Signal, 1)
 138         signal.Notify(c, os.Interrupt, syscall.SIGTERM, syscall.SIGQUIT)
 139         sig := <-c
 140         logger.Printf("Received %s, shutting down", sig)
 141         signal.Stop(c)
 142
 143         cancel()
 144
 145         runningCmdsMutex.Lock()
 146         // Finished dispatching; interrupt any crunch jobs that are still running
 147         for _, cmd := range runningCmds {
 148                 cmd.Process.Signal(os.Interrupt)
 149         }
 150         runningCmdsMutex.Unlock()
 151
 152         // Wait for all running crunch jobs to complete / terminate
 153         waitGroup.Wait()
 154
 155         return nil
 156 }
 157
 158 func startFunc(container arvados.Container, cmd *exec.Cmd) error {
 159         return cmd.Start()
 160 }
 161
 162 type LocalRun struct {
 163         startCmd         func(container arvados.Container, cmd *exec.Cmd) error
 164         concurrencyLimit chan bool
 165         ctx              context.Context
 166         cluster          *arvados.Cluster
 167 }
 168
 169 // Run a container.
 170 //
 171 // If the container is Locked, start a new crunch-run process and wait until
 172 // crunch-run completes.  If the priority is set to zero, set an interrupt
 173 // signal to the crunch-run process.
 174 //
 175 // If the container is in any other state, or is not Complete/Cancelled after
 176 // crunch-run terminates, mark the container as Cancelled.
 177 func (lr *LocalRun) run(dispatcher *dispatch.Dispatcher,
 178         container arvados.Container,
 179         status <-chan arvados.Container) error {
 180
 181         uuid := container.UUID
 182
 183         if container.State == dispatch.Locked {
 184
 185                 select {
 186                 case lr.concurrencyLimit <- true:
 187                         break
 188                 case <-lr.ctx.Done():
 189                         return lr.ctx.Err()
 190                 }
 191
 192                 defer func() { <-lr.concurrencyLimit }()
 193
 194                 select {
 195                 case c := <-status:
 196                         // Check for state updates after possibly
 197                         // waiting to be ready-to-run
 198                         if c.Priority == 0 {
 199                                 goto Finish
 200                         }
 201                 default:
 202                         break
 203                 }
 204
 205                 waitGroup.Add(1)
 206                 defer waitGroup.Done()
 207
 208                 cmd := exec.Command(*crunchRunCommand, "--runtime-engine="+lr.cluster.Containers.RuntimeEngine, uuid)
 209                 cmd.Stdin = nil
 210                 cmd.Stderr = os.Stderr
 211                 cmd.Stdout = os.Stderr
 212
 213                 dispatcher.Logger.Printf("starting container %v", uuid)
 214
 215                 // Add this crunch job to the list of runningCmds only if we
 216                 // succeed in starting crunch-run.
 217
 218                 runningCmdsMutex.Lock()
 219                 if err := lr.startCmd(container, cmd); err != nil {
 220                         runningCmdsMutex.Unlock()
 221                         dispatcher.Logger.Warnf("error starting %q for %s: %s", *crunchRunCommand, uuid, err)
 222                         dispatcher.UpdateState(uuid, dispatch.Cancelled)
 223                 } else {
 224                         runningCmds[uuid] = cmd
 225                         runningCmdsMutex.Unlock()
 226
 227                         // Need to wait for crunch-run to exit
 228                         done := make(chan struct{})
 229
 230                         go func() {
 231                                 if _, err := cmd.Process.Wait(); err != nil {
 232                                         dispatcher.Logger.Warnf("error while waiting for crunch job to finish for %v: %q", uuid, err)
 233                                 }
 234                                 dispatcher.Logger.Debugf("sending done")
 235                                 done <- struct{}{}
 236                         }()
 237
 238                 Loop:
 239                         for {
 240                                 select {
 241                                 case <-done:
 242                                         break Loop
 243                                 case c := <-status:
 244                                         // Interrupt the child process if priority changes to 0
 245                                         if (c.State == dispatch.Locked || c.State == dispatch.Running) && c.Priority == 0 {
 246                                                 dispatcher.Logger.Printf("sending SIGINT to pid %d to cancel container %v", cmd.Process.Pid, uuid)
 247                                                 cmd.Process.Signal(os.Interrupt)
 248                                         }
 249                                 }
 250                         }
 251                         close(done)
 252
 253                         dispatcher.Logger.Printf("finished container run for %v", uuid)
 254
 255                         // Remove the crunch job from runningCmds
 256                         runningCmdsMutex.Lock()
 257                         delete(runningCmds, uuid)
 258                         runningCmdsMutex.Unlock()
 259                 }
 260         }
 261
 262 Finish:
 263
 264         // If the container is not finalized, then change it to "Cancelled".
 265         err := dispatcher.Arv.Get("containers", uuid, nil, &container)
 266         if err != nil {
 267                 dispatcher.Logger.Warnf("error getting final container state: %v", err)
 268         }
 269         if container.State == dispatch.Locked || container.State == dispatch.Running {
 270                 dispatcher.Logger.Warnf("after %q process termination, container state for %v is %q; updating it to %q",
 271                         *crunchRunCommand, uuid, container.State, dispatch.Cancelled)
 272                 dispatcher.UpdateState(uuid, dispatch.Cancelled)
 273         }
 274
 275         // drain any subsequent status changes
 276         for range status {
 277         }
 278
 279         dispatcher.Logger.Printf("finalized container %v", uuid)
 280         return nil
 281 }