X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/7337b18bf7b6996a7fe4df0aba5356a03bda452d..8685bdc41012f1623cc02b573e27439fdf314799:/lib/dispatchcloud/test/stub_driver.go diff --git a/lib/dispatchcloud/test/stub_driver.go b/lib/dispatchcloud/test/stub_driver.go index f738e20664..1b31a71a26 100644 --- a/lib/dispatchcloud/test/stub_driver.go +++ b/lib/dispatchcloud/test/stub_driver.go @@ -6,18 +6,19 @@ package test import ( "crypto/rand" + "encoding/json" "errors" "fmt" "io" + "io/ioutil" math_rand "math/rand" "regexp" "strings" "sync" "time" - "git.curoverse.com/arvados.git/lib/cloud" - "git.curoverse.com/arvados.git/sdk/go/arvados" - "github.com/mitchellh/mapstructure" + "git.arvados.org/arvados.git/lib/cloud" + "git.arvados.org/arvados.git/sdk/go/arvados" "github.com/sirupsen/logrus" "golang.org/x/crypto/ssh" ) @@ -33,6 +34,11 @@ type StubDriver struct { // VM's error rate and other behaviors. SetupVM func(*StubVM) + // Bugf, if set, is called if a bug is detected in the caller + // or stub. Typically set to (*check.C)Errorf. If unset, + // logger.Warnf is called instead. + Bugf func(string, ...interface{}) + // StubVM's fake crunch-run uses this Queue to read and update // container state. Queue *Queue @@ -41,19 +47,36 @@ type StubDriver struct { // Destroy. 0=always succeed, 1=always fail. ErrorRateDestroy float64 + // If Create() or Instances() is called too frequently, return + // rate-limiting errors. + MinTimeBetweenCreateCalls time.Duration + MinTimeBetweenInstancesCalls time.Duration + + // If true, Create and Destroy calls block until Release() is + // called. + HoldCloudOps bool + instanceSets []*StubInstanceSet + holdCloudOps chan bool } // InstanceSet returns a new *StubInstanceSet. -func (sd *StubDriver) InstanceSet(params map[string]interface{}, id cloud.InstanceSetID, - logger logrus.FieldLogger) (cloud.InstanceSet, error) { - +func (sd *StubDriver) InstanceSet(params json.RawMessage, id cloud.InstanceSetID, _ cloud.SharedResourceTags, logger logrus.FieldLogger) (cloud.InstanceSet, error) { + if sd.holdCloudOps == nil { + sd.holdCloudOps = make(chan bool) + } sis := StubInstanceSet{ driver: sd, + logger: logger, servers: map[cloud.InstanceID]*StubVM{}, } sd.instanceSets = append(sd.instanceSets, &sis) - return &sis, mapstructure.Decode(params, &sis) + + var err error + if params != nil { + err = json.Unmarshal(params, &sis) + } + return &sis, err } // InstanceSets returns all instances that have been created by the @@ -63,31 +86,57 @@ func (sd *StubDriver) InstanceSets() []*StubInstanceSet { return sd.instanceSets } +// ReleaseCloudOps releases n pending Create/Destroy calls. If there +// are fewer than n blocked calls pending, it waits for the rest to +// arrive. +func (sd *StubDriver) ReleaseCloudOps(n int) { + for i := 0; i < n; i++ { + <-sd.holdCloudOps + } +} + type StubInstanceSet struct { driver *StubDriver + logger logrus.FieldLogger servers map[cloud.InstanceID]*StubVM mtx sync.RWMutex stopped bool + + allowCreateCall time.Time + allowInstancesCall time.Time + lastInstanceID int } -func (sis *StubInstanceSet) Create(it arvados.InstanceType, image cloud.ImageID, tags cloud.InstanceTags, authKey ssh.PublicKey) (cloud.Instance, error) { +func (sis *StubInstanceSet) Create(it arvados.InstanceType, image cloud.ImageID, tags cloud.InstanceTags, cmd cloud.InitCommand, authKey ssh.PublicKey) (cloud.Instance, error) { + if sis.driver.HoldCloudOps { + sis.driver.holdCloudOps <- true + } sis.mtx.Lock() defer sis.mtx.Unlock() if sis.stopped { return nil, errors.New("StubInstanceSet: Create called after Stop") } + if sis.allowCreateCall.After(time.Now()) { + return nil, RateLimitError{sis.allowCreateCall} + } + sis.allowCreateCall = time.Now().Add(sis.driver.MinTimeBetweenCreateCalls) ak := sis.driver.AuthorizedKeys if authKey != nil { ak = append([]ssh.PublicKey{authKey}, ak...) } + sis.lastInstanceID++ svm := &StubVM{ sis: sis, - id: cloud.InstanceID(fmt.Sprintf("stub-%s-%x", it.ProviderType, math_rand.Int63())), + id: cloud.InstanceID(fmt.Sprintf("inst%d,%s", sis.lastInstanceID, it.ProviderType)), tags: copyTags(tags), providerType: it.ProviderType, + initCommand: cmd, + running: map[string]stubProcess{}, + killing: map[string]bool{}, } svm.SSHService = SSHService{ HostKey: sis.driver.HostKey, + AuthorizedUser: "root", AuthorizedKeys: ak, Exec: svm.Exec, } @@ -101,6 +150,10 @@ func (sis *StubInstanceSet) Create(it arvados.InstanceType, image cloud.ImageID, func (sis *StubInstanceSet) Instances(cloud.InstanceTags) ([]cloud.Instance, error) { sis.mtx.RLock() defer sis.mtx.RUnlock() + if sis.allowInstancesCall.After(time.Now()) { + return nil, RateLimitError{sis.allowInstancesCall} + } + sis.allowInstancesCall = time.Now().Add(sis.driver.MinTimeBetweenInstancesCalls) var r []cloud.Instance for _, ss := range sis.servers { r = append(r, ss.Instance()) @@ -117,6 +170,11 @@ func (sis *StubInstanceSet) Stop() { sis.stopped = true } +type RateLimitError struct{ Retry time.Time } + +func (e RateLimitError) Error() string { return fmt.Sprintf("rate limited until %s", e.Retry) } +func (e RateLimitError) EarliestRetry() time.Time { return e.Retry } + // StubVM is a fake server that runs an SSH service. It represents a // VM running in a fake cloud. // @@ -125,22 +183,39 @@ func (sis *StubInstanceSet) Stop() { // running (and might change IP addresses, shut down, etc.) without // updating any stubInstances that have been returned to callers. type StubVM struct { - Boot time.Time - Broken time.Time - CrunchRunMissing bool - CrunchRunCrashRate float64 - CrunchRunDetachDelay time.Duration - ExecuteContainer func(arvados.Container) int + Boot time.Time + Broken time.Time + ReportBroken time.Time + CrunchRunMissing bool + CrunchRunCrashRate float64 + CrunchRunDetachDelay time.Duration + ArvMountMaxExitLag time.Duration + ArvMountDeadlockRate float64 + ExecuteContainer func(arvados.Container) int + CrashRunningContainer func(arvados.Container) + ExtraCrunchRunArgs string // extra args expected after "crunch-run --detach --stdin-env " sis *StubInstanceSet id cloud.InstanceID tags cloud.InstanceTags + initCommand cloud.InitCommand providerType string SSHService SSHService - running map[string]bool + running map[string]stubProcess + killing map[string]bool + lastPID int64 + deadlocked string sync.Mutex } +type stubProcess struct { + pid int64 + + // crunch-run has exited, but arv-mount process (or something) + // still holds lock in /var/run/ + exited bool +} + func (svm *StubVM) Instance() stubInstance { svm.Lock() defer svm.Unlock() @@ -157,7 +232,12 @@ func (svm *StubVM) Instance() stubInstance { } } -func (svm *StubVM) Exec(command string, stdin io.Reader, stdout, stderr io.Writer) uint32 { +func (svm *StubVM) Exec(env map[string]string, command string, stdin io.Reader, stdout, stderr io.Writer) uint32 { + stdinData, err := ioutil.ReadAll(stdin) + if err != nil { + fmt.Fprintf(stderr, "error reading stdin: %s\n", err) + return 1 + } queue := svm.sis.driver.Queue uuid := regexp.MustCompile(`.{5}-dz642-.{15}`).FindString(command) if eta := svm.Boot.Sub(time.Now()); eta > 0 { @@ -172,74 +252,136 @@ func (svm *StubVM) Exec(command string, stdin io.Reader, stdout, stderr io.Write fmt.Fprint(stderr, "crunch-run: command not found\n") return 1 } - if strings.HasPrefix(command, "crunch-run --detach ") { - svm.Lock() - if svm.running == nil { - svm.running = map[string]bool{} + if strings.HasPrefix(command, "crunch-run --detach --stdin-env "+svm.ExtraCrunchRunArgs) { + var stdinKV map[string]string + err := json.Unmarshal(stdinData, &stdinKV) + if err != nil { + fmt.Fprintf(stderr, "unmarshal stdin: %s (stdin was: %q)\n", err, stdinData) + return 1 } - svm.running[uuid] = true + for _, name := range []string{"ARVADOS_API_HOST", "ARVADOS_API_TOKEN"} { + if stdinKV[name] == "" { + fmt.Fprintf(stderr, "%s env var missing from stdin %q\n", name, stdinData) + return 1 + } + } + svm.Lock() + svm.lastPID++ + pid := svm.lastPID + svm.running[uuid] = stubProcess{pid: pid} svm.Unlock() time.Sleep(svm.CrunchRunDetachDelay) fmt.Fprintf(stderr, "starting %s\n", uuid) - logger := logrus.WithField("ContainerUUID", uuid) + logger := svm.sis.logger.WithFields(logrus.Fields{ + "Instance": svm.id, + "ContainerUUID": uuid, + "PID": pid, + }) logger.Printf("[test] starting crunch-run stub") go func() { + var ctr arvados.Container + var started, completed bool + defer func() { + logger.Print("[test] exiting crunch-run stub") + svm.Lock() + defer svm.Unlock() + if svm.running[uuid].pid != pid { + bugf := svm.sis.driver.Bugf + if bugf == nil { + bugf = logger.Warnf + } + bugf("[test] StubDriver bug or caller bug: pid %d exiting, running[%s].pid==%d", pid, uuid, svm.running[uuid].pid) + return + } + if !completed { + logger.WithField("State", ctr.State).Print("[test] crashing crunch-run stub") + if started && svm.CrashRunningContainer != nil { + svm.CrashRunningContainer(ctr) + } + } + sproc := svm.running[uuid] + sproc.exited = true + svm.running[uuid] = sproc + svm.Unlock() + time.Sleep(svm.ArvMountMaxExitLag * time.Duration(math_rand.Float64())) + svm.Lock() + if math_rand.Float64() >= svm.ArvMountDeadlockRate { + delete(svm.running, uuid) + } + }() + crashluck := math_rand.Float64() + wantCrash := crashluck < svm.CrunchRunCrashRate + wantCrashEarly := crashluck < svm.CrunchRunCrashRate/2 + ctr, ok := queue.Get(uuid) if !ok { logger.Print("[test] container not in queue") return } - if crashluck > svm.CrunchRunCrashRate/2 { - time.Sleep(time.Duration(math_rand.Float64()*20) * time.Millisecond) - ctr.State = arvados.ContainerStateRunning - queue.Notify(ctr) - } time.Sleep(time.Duration(math_rand.Float64()*20) * time.Millisecond) + svm.Lock() - _, running := svm.running[uuid] + killed := svm.killing[uuid] svm.Unlock() - if !running { - logger.Print("[test] container was killed") + if killed || wantCrashEarly { + return + } + + ctr.State = arvados.ContainerStateRunning + started = queue.Notify(ctr) + if !started { + ctr, _ = queue.Get(uuid) + logger.Print("[test] erroring out because state=Running update was rejected") + return + } + + if wantCrash { + logger.WithField("State", ctr.State).Print("[test] crashing crunch-run stub") return } if svm.ExecuteContainer != nil { ctr.ExitCode = svm.ExecuteContainer(ctr) } - // TODO: Check whether the stub instance has - // been destroyed, and if so, don't call - // queue.Notify. Then "container finished - // twice" can be classified as a bug. - if crashluck < svm.CrunchRunCrashRate { - logger.Print("[test] crashing crunch-run stub") - } else { - ctr.State = arvados.ContainerStateComplete - queue.Notify(ctr) - } - logger.Print("[test] exiting crunch-run stub") - svm.Lock() - defer svm.Unlock() - delete(svm.running, uuid) + logger.WithField("ExitCode", ctr.ExitCode).Print("[test] completing container") + ctr.State = arvados.ContainerStateComplete + completed = queue.Notify(ctr) }() return 0 } if command == "crunch-run --list" { svm.Lock() defer svm.Unlock() - for uuid := range svm.running { - fmt.Fprintf(stdout, "%s\n", uuid) + for uuid, sproc := range svm.running { + if sproc.exited { + fmt.Fprintf(stdout, "%s stale\n", uuid) + } else { + fmt.Fprintf(stdout, "%s\n", uuid) + } } + if !svm.ReportBroken.IsZero() && svm.ReportBroken.Before(time.Now()) { + fmt.Fprintln(stdout, "broken") + } + fmt.Fprintln(stdout, svm.deadlocked) return 0 } if strings.HasPrefix(command, "crunch-run --kill ") { svm.Lock() - defer svm.Unlock() - if svm.running[uuid] { - delete(svm.running, uuid) - } else { - fmt.Fprintf(stderr, "%s: container is not running\n", uuid) + sproc, running := svm.running[uuid] + if running && !sproc.exited { + svm.killing[uuid] = true + svm.Unlock() + time.Sleep(time.Duration(math_rand.Float64()*2) * time.Millisecond) + svm.Lock() + sproc, running = svm.running[uuid] } + svm.Unlock() + if running && !sproc.exited { + fmt.Fprintf(stderr, "%s: container is running\n", uuid) + return 1 + } + fmt.Fprintf(stderr, "%s: container is not running\n", uuid) return 0 } if command == "true" { @@ -263,12 +405,19 @@ func (si stubInstance) Address() string { return si.addr } +func (si stubInstance) RemoteUser() string { + return si.svm.SSHService.AuthorizedUser +} + func (si stubInstance) Destroy() error { + sis := si.svm.sis + if sis.driver.HoldCloudOps { + sis.driver.holdCloudOps <- true + } if math_rand.Float64() < si.svm.sis.driver.ErrorRateDestroy { return errors.New("instance could not be destroyed") } si.svm.SSHService.Close() - sis := si.svm.sis sis.mtx.Lock() defer sis.mtx.Unlock() delete(sis.servers, si.svm.id) @@ -291,7 +440,9 @@ func (si stubInstance) SetTags(tags cloud.InstanceTags) error { } func (si stubInstance) Tags() cloud.InstanceTags { - return si.tags + // Return a copy to ensure a caller can't change our saved + // tags just by writing to the returned map. + return copyTags(si.tags) } func (si stubInstance) String() string {