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