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