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.curoverse.com/arvados.git/sdk/go/arvados"
  21         "git.curoverse.com/arvados.git/sdk/go/arvadosclient"
  22         "git.curoverse.com/arvados.git/sdk/go/dispatch"
  23         "github.com/sirupsen/logrus"
  24 )
  25
  26 var version = "dev"
  27
  28 func main() {
  29         err := doMain()
  30         if err != nil {
  31                 logrus.Fatalf("%q", err)
  32         }
  33 }
  34
  35 var (
  36         runningCmds      map[string]*exec.Cmd
  37         runningCmdsMutex sync.Mutex
  38         waitGroup        sync.WaitGroup
  39         crunchRunCommand *string
  40 )
  41
  42 func doMain() error {
  43         logger := logrus.StandardLogger()
  44         if os.Getenv("DEBUG") != "" {
  45                 logger.SetLevel(logrus.DebugLevel)
  46         }
  47         logger.Formatter = &logrus.JSONFormatter{
  48                 TimestampFormat: "2006-01-02T15:04:05.000000000Z07:00",
  49         }
  50
  51         flags := flag.NewFlagSet("crunch-dispatch-local", flag.ExitOnError)
  52
  53         pollInterval := flags.Int(
  54                 "poll-interval",
  55                 10,
  56                 "Interval in seconds to poll for queued containers")
  57
  58         crunchRunCommand = flags.String(
  59                 "crunch-run-command",
  60                 "/usr/bin/crunch-run",
  61                 "Crunch command to run container")
  62
  63         getVersion := flags.Bool(
  64                 "version",
  65                 false,
  66                 "Print version information and exit.")
  67
  68         // Parse args; omit the first arg which is the command name
  69         flags.Parse(os.Args[1:])
  70
  71         // Print version information if requested
  72         if *getVersion {
  73                 fmt.Printf("crunch-dispatch-local %s\n", version)
  74                 return nil
  75         }
  76
  77         logger.Printf("crunch-dispatch-local %s started", version)
  78
  79         runningCmds = make(map[string]*exec.Cmd)
  80
  81         arv, err := arvadosclient.MakeArvadosClient()
  82         if err != nil {
  83                 logger.Errorf("error making Arvados client: %v", err)
  84                 return err
  85         }
  86         arv.Retries = 25
  87
  88         dispatcher := dispatch.Dispatcher{
  89                 Logger:       logger,
  90                 Arv:          arv,
  91                 RunContainer: run,
  92                 PollPeriod:   time.Duration(*pollInterval) * time.Second,
  93         }
  94
  95         ctx, cancel := context.WithCancel(context.Background())
  96         err = dispatcher.Run(ctx)
  97         if err != nil {
  98                 return err
  99         }
 100
 101         c := make(chan os.Signal, 1)
 102         signal.Notify(c, os.Interrupt, syscall.SIGTERM, syscall.SIGQUIT)
 103         sig := <-c
 104         logger.Printf("Received %s, shutting down", sig)
 105         signal.Stop(c)
 106
 107         cancel()
 108
 109         runningCmdsMutex.Lock()
 110         // Finished dispatching; interrupt any crunch jobs that are still running
 111         for _, cmd := range runningCmds {
 112                 cmd.Process.Signal(os.Interrupt)
 113         }
 114         runningCmdsMutex.Unlock()
 115
 116         // Wait for all running crunch jobs to complete / terminate
 117         waitGroup.Wait()
 118
 119         return nil
 120 }
 121
 122 func startFunc(container arvados.Container, cmd *exec.Cmd) error {
 123         return cmd.Start()
 124 }
 125
 126 var startCmd = startFunc
 127
 128 // Run a container.
 129 //
 130 // If the container is Locked, start a new crunch-run process and wait until
 131 // crunch-run completes.  If the priority is set to zero, set an interrupt
 132 // signal to the crunch-run process.
 133 //
 134 // If the container is in any other state, or is not Complete/Cancelled after
 135 // crunch-run terminates, mark the container as Cancelled.
 136 func run(dispatcher *dispatch.Dispatcher,
 137         container arvados.Container,
 138         status <-chan arvados.Container) {
 139
 140         uuid := container.UUID
 141
 142         if container.State == dispatch.Locked {
 143                 waitGroup.Add(1)
 144
 145                 cmd := exec.Command(*crunchRunCommand, uuid)
 146                 cmd.Stdin = nil
 147                 cmd.Stderr = os.Stderr
 148                 cmd.Stdout = os.Stderr
 149
 150                 dispatcher.Logger.Printf("starting container %v", uuid)
 151
 152                 // Add this crunch job to the list of runningCmds only if we
 153                 // succeed in starting crunch-run.
 154
 155                 runningCmdsMutex.Lock()
 156                 if err := startCmd(container, cmd); err != nil {
 157                         runningCmdsMutex.Unlock()
 158                         dispatcher.Logger.Warnf("error starting %q for %s: %s", *crunchRunCommand, uuid, err)
 159                         dispatcher.UpdateState(uuid, dispatch.Cancelled)
 160                 } else {
 161                         runningCmds[uuid] = cmd
 162                         runningCmdsMutex.Unlock()
 163
 164                         // Need to wait for crunch-run to exit
 165                         done := make(chan struct{})
 166
 167                         go func() {
 168                                 if _, err := cmd.Process.Wait(); err != nil {
 169                                         dispatcher.Logger.Warnf("error while waiting for crunch job to finish for %v: %q", uuid, err)
 170                                 }
 171                                 dispatcher.Logger.Debugf("sending done")
 172                                 done <- struct{}{}
 173                         }()
 174
 175                 Loop:
 176                         for {
 177                                 select {
 178                                 case <-done:
 179                                         break Loop
 180                                 case c := <-status:
 181                                         // Interrupt the child process if priority changes to 0
 182                                         if (c.State == dispatch.Locked || c.State == dispatch.Running) && c.Priority == 0 {
 183                                                 dispatcher.Logger.Printf("sending SIGINT to pid %d to cancel container %v", cmd.Process.Pid, uuid)
 184                                                 cmd.Process.Signal(os.Interrupt)
 185                                         }
 186                                 }
 187                         }
 188                         close(done)
 189
 190                         dispatcher.Logger.Printf("finished container run for %v", uuid)
 191
 192                         // Remove the crunch job from runningCmds
 193                         runningCmdsMutex.Lock()
 194                         delete(runningCmds, uuid)
 195                         runningCmdsMutex.Unlock()
 196                 }
 197                 waitGroup.Done()
 198         }
 199
 200         // If the container is not finalized, then change it to "Cancelled".
 201         err := dispatcher.Arv.Get("containers", uuid, nil, &container)
 202         if err != nil {
 203                 dispatcher.Logger.Warnf("error getting final container state: %v", err)
 204         }
 205         if container.State == dispatch.Locked || container.State == dispatch.Running {
 206                 dispatcher.Logger.Warnf("after %q process termination, container state for %v is %q; updating it to %q",
 207                         *crunchRunCommand, uuid, container.State, dispatch.Cancelled)
 208                 dispatcher.UpdateState(uuid, dispatch.Cancelled)
 209         }
 210
 211         // drain any subsequent status changes
 212         for range status {
 213         }
 214
 215         dispatcher.Logger.Printf("finalized container %v", uuid)
 216 }