lib/dispatchcloud/test/stub_driver.go

   1 // Copyright (C) The Arvados Authors. All rights reserved.
   2 //
   3 // SPDX-License-Identifier: AGPL-3.0
   4
   5 package test
   6
   7 import (
   8         "crypto/rand"
   9         "encoding/json"
  10         "errors"
  11         "fmt"
  12         "io"
  13         "io/ioutil"
  14         math_rand "math/rand"
  15         "regexp"
  16         "strings"
  17         "sync"
  18         "time"
  19
  20         "git.curoverse.com/arvados.git/lib/cloud"
  21         "git.curoverse.com/arvados.git/sdk/go/arvados"
  22         "github.com/sirupsen/logrus"
  23         "golang.org/x/crypto/ssh"
  24 )
  25
  26 // A StubDriver implements cloud.Driver by setting up local SSH
  27 // servers that do fake command executions.
  28 type StubDriver struct {
  29         HostKey        ssh.Signer
  30         AuthorizedKeys []ssh.PublicKey
  31
  32         // SetupVM, if set, is called upon creation of each new
  33         // StubVM. This is the caller's opportunity to customize the
  34         // VM's error rate and other behaviors.
  35         SetupVM func(*StubVM)
  36
  37         // StubVM's fake crunch-run uses this Queue to read and update
  38         // container state.
  39         Queue *Queue
  40
  41         // Frequency of artificially introduced errors on calls to
  42         // Destroy. 0=always succeed, 1=always fail.
  43         ErrorRateDestroy float64
  44
  45         // If Create() or Instances() is called too frequently, return
  46         // rate-limiting errors.
  47         MinTimeBetweenCreateCalls    time.Duration
  48         MinTimeBetweenInstancesCalls time.Duration
  49
  50         // If true, Create and Destroy calls block until Release() is
  51         // called.
  52         HoldCloudOps bool
  53
  54         instanceSets []*StubInstanceSet
  55         holdCloudOps chan bool
  56 }
  57
  58 // InstanceSet returns a new *StubInstanceSet.
  59 func (sd *StubDriver) InstanceSet(params json.RawMessage, id cloud.InstanceSetID, logger logrus.FieldLogger) (cloud.InstanceSet, error) {
  60         if sd.holdCloudOps == nil {
  61                 sd.holdCloudOps = make(chan bool)
  62         }
  63         sis := StubInstanceSet{
  64                 driver:  sd,
  65                 logger:  logger,
  66                 servers: map[cloud.InstanceID]*StubVM{},
  67         }
  68         sd.instanceSets = append(sd.instanceSets, &sis)
  69
  70         var err error
  71         if params != nil {
  72                 err = json.Unmarshal(params, &sis)
  73         }
  74         return &sis, err
  75 }
  76
  77 // InstanceSets returns all instances that have been created by the
  78 // driver. This can be used to test a component that uses the driver
  79 // but doesn't expose the InstanceSets it has created.
  80 func (sd *StubDriver) InstanceSets() []*StubInstanceSet {
  81         return sd.instanceSets
  82 }
  83
  84 // ReleaseCloudOps releases n pending Create/Destroy calls. If there
  85 // are fewer than n blocked calls pending, it waits for the rest to
  86 // arrive.
  87 func (sd *StubDriver) ReleaseCloudOps(n int) {
  88         for i := 0; i < n; i++ {
  89                 <-sd.holdCloudOps
  90         }
  91 }
  92
  93 type StubInstanceSet struct {
  94         driver  *StubDriver
  95         logger  logrus.FieldLogger
  96         servers map[cloud.InstanceID]*StubVM
  97         mtx     sync.RWMutex
  98         stopped bool
  99
 100         allowCreateCall    time.Time
 101         allowInstancesCall time.Time
 102 }
 103
 104 func (sis *StubInstanceSet) Create(it arvados.InstanceType, image cloud.ImageID, tags cloud.InstanceTags, cmd cloud.InitCommand, authKey ssh.PublicKey) (cloud.Instance, error) {
 105         if sis.driver.HoldCloudOps {
 106                 sis.driver.holdCloudOps <- true
 107         }
 108         sis.mtx.Lock()
 109         defer sis.mtx.Unlock()
 110         if sis.stopped {
 111                 return nil, errors.New("StubInstanceSet: Create called after Stop")
 112         }
 113         if sis.allowCreateCall.After(time.Now()) {
 114                 return nil, RateLimitError{sis.allowCreateCall}
 115         } else {
 116                 sis.allowCreateCall = time.Now().Add(sis.driver.MinTimeBetweenCreateCalls)
 117         }
 118
 119         ak := sis.driver.AuthorizedKeys
 120         if authKey != nil {
 121                 ak = append([]ssh.PublicKey{authKey}, ak...)
 122         }
 123         svm := &StubVM{
 124                 sis:          sis,
 125                 id:           cloud.InstanceID(fmt.Sprintf("stub-%s-%x", it.ProviderType, math_rand.Int63())),
 126                 tags:         copyTags(tags),
 127                 providerType: it.ProviderType,
 128                 initCommand:  cmd,
 129                 running:      map[string]int64{},
 130                 killing:      map[string]bool{},
 131         }
 132         svm.SSHService = SSHService{
 133                 HostKey:        sis.driver.HostKey,
 134                 AuthorizedUser: "root",
 135                 AuthorizedKeys: ak,
 136                 Exec:           svm.Exec,
 137         }
 138         if setup := sis.driver.SetupVM; setup != nil {
 139                 setup(svm)
 140         }
 141         sis.servers[svm.id] = svm
 142         return svm.Instance(), nil
 143 }
 144
 145 func (sis *StubInstanceSet) Instances(cloud.InstanceTags) ([]cloud.Instance, error) {
 146         sis.mtx.RLock()
 147         defer sis.mtx.RUnlock()
 148         if sis.allowInstancesCall.After(time.Now()) {
 149                 return nil, RateLimitError{sis.allowInstancesCall}
 150         } else {
 151                 sis.allowInstancesCall = time.Now().Add(sis.driver.MinTimeBetweenInstancesCalls)
 152         }
 153         var r []cloud.Instance
 154         for _, ss := range sis.servers {
 155                 r = append(r, ss.Instance())
 156         }
 157         return r, nil
 158 }
 159
 160 func (sis *StubInstanceSet) Stop() {
 161         sis.mtx.Lock()
 162         defer sis.mtx.Unlock()
 163         if sis.stopped {
 164                 panic("Stop called twice")
 165         }
 166         sis.stopped = true
 167 }
 168
 169 type RateLimitError struct{ Retry time.Time }
 170
 171 func (e RateLimitError) Error() string            { return fmt.Sprintf("rate limited until %s", e.Retry) }
 172 func (e RateLimitError) EarliestRetry() time.Time { return e.Retry }
 173
 174 // StubVM is a fake server that runs an SSH service. It represents a
 175 // VM running in a fake cloud.
 176 //
 177 // Note this is distinct from a stubInstance, which is a snapshot of
 178 // the VM's metadata. Like a VM in a real cloud, a StubVM keeps
 179 // running (and might change IP addresses, shut down, etc.)  without
 180 // updating any stubInstances that have been returned to callers.
 181 type StubVM struct {
 182         Boot                  time.Time
 183         Broken                time.Time
 184         CrunchRunMissing      bool
 185         CrunchRunCrashRate    float64
 186         CrunchRunDetachDelay  time.Duration
 187         ExecuteContainer      func(arvados.Container) int
 188         CrashRunningContainer func(arvados.Container)
 189
 190         sis          *StubInstanceSet
 191         id           cloud.InstanceID
 192         tags         cloud.InstanceTags
 193         initCommand  cloud.InitCommand
 194         providerType string
 195         SSHService   SSHService
 196         running      map[string]int64
 197         killing      map[string]bool
 198         lastPID      int64
 199         sync.Mutex
 200 }
 201
 202 func (svm *StubVM) Instance() stubInstance {
 203         svm.Lock()
 204         defer svm.Unlock()
 205         return stubInstance{
 206                 svm:  svm,
 207                 addr: svm.SSHService.Address(),
 208                 // We deliberately return a cached/stale copy of the
 209                 // real tags here, so that (Instance)Tags() sometimes
 210                 // returns old data after a call to
 211                 // (Instance)SetTags().  This is permitted by the
 212                 // driver interface, and this might help remind
 213                 // callers that they need to tolerate it.
 214                 tags: copyTags(svm.tags),
 215         }
 216 }
 217
 218 func (svm *StubVM) Exec(env map[string]string, command string, stdin io.Reader, stdout, stderr io.Writer) uint32 {
 219         stdinData, err := ioutil.ReadAll(stdin)
 220         if err != nil {
 221                 fmt.Fprintf(stderr, "error reading stdin: %s\n", err)
 222                 return 1
 223         }
 224         queue := svm.sis.driver.Queue
 225         uuid := regexp.MustCompile(`.{5}-dz642-.{15}`).FindString(command)
 226         if eta := svm.Boot.Sub(time.Now()); eta > 0 {
 227                 fmt.Fprintf(stderr, "stub is booting, ETA %s\n", eta)
 228                 return 1
 229         }
 230         if !svm.Broken.IsZero() && svm.Broken.Before(time.Now()) {
 231                 fmt.Fprintf(stderr, "cannot fork\n")
 232                 return 2
 233         }
 234         if svm.CrunchRunMissing && strings.Contains(command, "crunch-run") {
 235                 fmt.Fprint(stderr, "crunch-run: command not found\n")
 236                 return 1
 237         }
 238         if strings.HasPrefix(command, "crunch-run --detach --stdin-env ") {
 239                 var stdinKV map[string]string
 240                 err := json.Unmarshal(stdinData, &stdinKV)
 241                 if err != nil {
 242                         fmt.Fprintf(stderr, "unmarshal stdin: %s (stdin was: %q)\n", err, stdinData)
 243                         return 1
 244                 }
 245                 for _, name := range []string{"ARVADOS_API_HOST", "ARVADOS_API_TOKEN"} {
 246                         if stdinKV[name] == "" {
 247                                 fmt.Fprintf(stderr, "%s env var missing from stdin %q\n", name, stdin)
 248                                 return 1
 249                         }
 250                 }
 251                 svm.Lock()
 252                 svm.lastPID++
 253                 pid := svm.lastPID
 254                 svm.running[uuid] = pid
 255                 svm.Unlock()
 256                 time.Sleep(svm.CrunchRunDetachDelay)
 257                 fmt.Fprintf(stderr, "starting %s\n", uuid)
 258                 logger := svm.sis.logger.WithFields(logrus.Fields{
 259                         "Instance":      svm.id,
 260                         "ContainerUUID": uuid,
 261                         "PID":           pid,
 262                 })
 263                 logger.Printf("[test] starting crunch-run stub")
 264                 go func() {
 265                         crashluck := math_rand.Float64()
 266                         ctr, ok := queue.Get(uuid)
 267                         if !ok {
 268                                 logger.Print("[test] container not in queue")
 269                                 return
 270                         }
 271
 272                         defer func() {
 273                                 if ctr.State == arvados.ContainerStateRunning && svm.CrashRunningContainer != nil {
 274                                         svm.CrashRunningContainer(ctr)
 275                                 }
 276                         }()
 277
 278                         if crashluck > svm.CrunchRunCrashRate/2 {
 279                                 time.Sleep(time.Duration(math_rand.Float64()*20) * time.Millisecond)
 280                                 ctr.State = arvados.ContainerStateRunning
 281                                 if !queue.Notify(ctr) {
 282                                         ctr, _ = queue.Get(uuid)
 283                                         logger.Print("[test] erroring out because state=Running update was rejected")
 284                                         return
 285                                 }
 286                         }
 287
 288                         time.Sleep(time.Duration(math_rand.Float64()*20) * time.Millisecond)
 289
 290                         svm.Lock()
 291                         defer svm.Unlock()
 292                         if svm.running[uuid] != pid {
 293                                 logger.Print("[test] container was killed")
 294                                 return
 295                         }
 296                         delete(svm.running, uuid)
 297
 298                         if crashluck < svm.CrunchRunCrashRate {
 299                                 logger.WithField("State", ctr.State).Print("[test] crashing crunch-run stub")
 300                         } else {
 301                                 if svm.ExecuteContainer != nil {
 302                                         ctr.ExitCode = svm.ExecuteContainer(ctr)
 303                                 }
 304                                 logger.WithField("ExitCode", ctr.ExitCode).Print("[test] exiting crunch-run stub")
 305                                 ctr.State = arvados.ContainerStateComplete
 306                                 go queue.Notify(ctr)
 307                         }
 308                 }()
 309                 return 0
 310         }
 311         if command == "crunch-run --list" {
 312                 svm.Lock()
 313                 defer svm.Unlock()
 314                 for uuid := range svm.running {
 315                         fmt.Fprintf(stdout, "%s\n", uuid)
 316                 }
 317                 return 0
 318         }
 319         if strings.HasPrefix(command, "crunch-run --kill ") {
 320                 svm.Lock()
 321                 pid, running := svm.running[uuid]
 322                 if running && !svm.killing[uuid] {
 323                         svm.killing[uuid] = true
 324                         go func() {
 325                                 time.Sleep(time.Duration(math_rand.Float64()*30) * time.Millisecond)
 326                                 svm.Lock()
 327                                 defer svm.Unlock()
 328                                 if svm.running[uuid] == pid {
 329                                         // Kill only if the running entry
 330                                         // hasn't since been killed and
 331                                         // replaced with a different one.
 332                                         delete(svm.running, uuid)
 333                                 }
 334                                 delete(svm.killing, uuid)
 335                         }()
 336                         svm.Unlock()
 337                         time.Sleep(time.Duration(math_rand.Float64()*2) * time.Millisecond)
 338                         svm.Lock()
 339                         _, running = svm.running[uuid]
 340                 }
 341                 svm.Unlock()
 342                 if running {
 343                         fmt.Fprintf(stderr, "%s: container is running\n", uuid)
 344                         return 1
 345                 } else {
 346                         fmt.Fprintf(stderr, "%s: container is not running\n", uuid)
 347                         return 0
 348                 }
 349         }
 350         if command == "true" {
 351                 return 0
 352         }
 353         fmt.Fprintf(stderr, "%q: command not found", command)
 354         return 1
 355 }
 356
 357 type stubInstance struct {
 358         svm  *StubVM
 359         addr string
 360         tags cloud.InstanceTags
 361 }
 362
 363 func (si stubInstance) ID() cloud.InstanceID {
 364         return si.svm.id
 365 }
 366
 367 func (si stubInstance) Address() string {
 368         return si.addr
 369 }
 370
 371 func (si stubInstance) RemoteUser() string {
 372         return si.svm.SSHService.AuthorizedUser
 373 }
 374
 375 func (si stubInstance) Destroy() error {
 376         sis := si.svm.sis
 377         if sis.driver.HoldCloudOps {
 378                 sis.driver.holdCloudOps <- true
 379         }
 380         if math_rand.Float64() < si.svm.sis.driver.ErrorRateDestroy {
 381                 return errors.New("instance could not be destroyed")
 382         }
 383         si.svm.SSHService.Close()
 384         sis.mtx.Lock()
 385         defer sis.mtx.Unlock()
 386         delete(sis.servers, si.svm.id)
 387         return nil
 388 }
 389
 390 func (si stubInstance) ProviderType() string {
 391         return si.svm.providerType
 392 }
 393
 394 func (si stubInstance) SetTags(tags cloud.InstanceTags) error {
 395         tags = copyTags(tags)
 396         svm := si.svm
 397         go func() {
 398                 svm.Lock()
 399                 defer svm.Unlock()
 400                 svm.tags = tags
 401         }()
 402         return nil
 403 }
 404
 405 func (si stubInstance) Tags() cloud.InstanceTags {
 406         // Return a copy to ensure a caller can't change our saved
 407         // tags just by writing to the returned map.
 408         return copyTags(si.tags)
 409 }
 410
 411 func (si stubInstance) String() string {
 412         return string(si.svm.id)
 413 }
 414
 415 func (si stubInstance) VerifyHostKey(key ssh.PublicKey, client *ssh.Client) error {
 416         buf := make([]byte, 512)
 417         _, err := io.ReadFull(rand.Reader, buf)
 418         if err != nil {
 419                 return err
 420         }
 421         sig, err := si.svm.sis.driver.HostKey.Sign(rand.Reader, buf)
 422         if err != nil {
 423                 return err
 424         }
 425         return key.Verify(buf, sig)
 426 }
 427
 428 func copyTags(src cloud.InstanceTags) cloud.InstanceTags {
 429         dst := cloud.InstanceTags{}
 430         for k, v := range src {
 431                 dst[k] = v
 432         }
 433         return dst
 434 }