Merge pull request #13778 from hashicorp/jbardin/GH-13617

improve resource.WaitForState and add refreshGracePeriod

helper/resource/state.go

@@ -2,10 +2,11 @@ package resource
 
 import (
 	"log"
-	"sync/atomic"
 	"time"
 )
 
+var refreshGracePeriod = 30 * time.Second
+
 // StateRefreshFunc is a function type used for StateChangeConf that is
 // responsible for refreshing the item being watched for a state change.
 //
@@ -62,58 +63,76 @@ func (conf *StateChangeConf) WaitForState() (interface{}, error) {
 		conf.ContinuousTargetOccurence = 1
 	}
 
-	// We can't safely read the result values if we timeout, so store them in
-	// an atomic.Value
 	type Result struct {
 		Result interface{}
 		State  string
 		Error  error
+		Done   bool
 	}
-	var lastResult atomic.Value
-	lastResult.Store(Result{})
 
-	doneCh := make(chan struct{})
+	// Read every result from the refresh loop, waiting for a positive result.Done.
+	resCh := make(chan Result, 1)
+	// cancellation channel for the refresh loop
+	cancelCh := make(chan struct{})
+
+	result := Result{}
+
 	go func() {
-		defer close(doneCh)
+		defer close(resCh)
 
-		// Wait for the delay
 		time.Sleep(conf.Delay)
 
-		wait := 100 * time.Millisecond
+		// start with 0 delay for the first loop
+		var wait time.Duration
 
 		for {
+			// store the last result
+			resCh <- result
+
+			// wait and watch for cancellation
+			select {
+			case <-cancelCh:
+				return
+			case <-time.After(wait):
+				// first round had no wait
+				if wait == 0 {
+					wait = 100 * time.Millisecond
+				}
+			}
+
 			res, currentState, err := conf.Refresh()
-			result := Result{
+			result = Result{
 				Result: res,
 				State:  currentState,
 				Error:  err,
 			}
-			lastResult.Store(result)
 
 			if err != nil {
+				resCh <- result
 				return
 			}
 
 			// If we're waiting for the absence of a thing, then return
 			if res == nil && len(conf.Target) == 0 {
-				targetOccurence += 1
+				targetOccurence++
 				if conf.ContinuousTargetOccurence == targetOccurence {
+					result.Done = true
+					resCh <- result
 					return
-				} else {
-					continue
 				}
+				continue
 			}
 
 			if res == nil {
 				// If we didn't find the resource, check if we have been
 				// not finding it for awhile, and if so, report an error.
-				notfoundTick += 1
+				notfoundTick++
 				if notfoundTick > conf.NotFoundChecks {
 					result.Error = &NotFoundError{
 						LastError: err,
 						Retries:   notfoundTick,
 					}
-					lastResult.Store(result)
+					resCh <- result
 					return
 				}
 			} else {
@@ -124,12 +143,13 @@ func (conf *StateChangeConf) WaitForState() (interface{}, error) {
 				for _, allowed := range conf.Target {
 					if currentState == allowed {
 						found = true
-						targetOccurence += 1
+						targetOccurence++
 						if conf.ContinuousTargetOccurence == targetOccurence {
+							result.Done = true
+							resCh <- result
 							return
-						} else {
-							continue
 						}
+						continue
 					}
 				}
 
@@ -147,11 +167,17 @@ func (conf *StateChangeConf) WaitForState() (interface{}, error) {
 						State:         result.State,
 						ExpectedState: conf.Target,
 					}
-					lastResult.Store(result)
+					resCh <- result
 					return
 				}
 			}
 
+			// Wait between refreshes using exponential backoff, except when
+			// waiting for the target state to reoccur.
+			if targetOccurence == 0 {
+				wait *= 2
+			}
+
 			// If a poll interval has been specified, choose that interval.
 			// Otherwise bound the default value.
 			if conf.PollInterval > 0 && conf.PollInterval < 180*time.Second {
@@ -165,27 +191,69 @@ func (conf *StateChangeConf) WaitForState() (interface{}, error) {
 			}
 
 			log.Printf("[TRACE] Waiting %s before next try", wait)
-			time.Sleep(wait)
-
-			// Wait between refreshes using exponential backoff, except when
-			// waiting for the target state to reoccur.
-			if targetOccurence == 0 {
-				wait *= 2
-			}
 		}
 	}()
 
+	// store the last value result from the refresh loop
+	lastResult := Result{}
+
+	timeout := time.After(conf.Timeout)
+	for {
 		select {
-	case <-doneCh:
-		r := lastResult.Load().(Result)
+		case r, ok := <-resCh:
+			// channel closed, so return the last result
+			if !ok {
+				return lastResult.Result, lastResult.Error
+			}
+
+			// we reached the intended state
+			if r.Done {
 				return r.Result, r.Error
-	case <-time.After(conf.Timeout):
-		r := lastResult.Load().(Result)
+			}
+
+			// still waiting, store the last result
+			lastResult = r
+
+		case <-timeout:
+			log.Printf("[WARN] WaitForState timeout after %s", conf.Timeout)
+			log.Printf("[WARN] WaitForState starting %s refresh grace period", refreshGracePeriod)
+
+			// cancel the goroutine and start our grace period timer
+			close(cancelCh)
+			timeout := time.After(refreshGracePeriod)
+
+			// we need a for loop and a label to break on, because we may have
+			// an extra response value to read, but still want to wait for the
+			// channel to close.
+		forSelect:
+			for {
+				select {
+				case r, ok := <-resCh:
+					if r.Done {
+						// the last refresh loop reached the desired state
+						return r.Result, r.Error
+					}
+
+					if !ok {
+						// the goroutine returned
+						break forSelect
+					}
+
+					// target state not reached, save the result for the
+					// TimeoutError and wait for the channel to close
+					lastResult = r
+				case <-timeout:
+					log.Println("[ERROR] WaitForState exceeded refresh grace period")
+					break forSelect
+				}
+			}
+
 			return nil, &TimeoutError{
-			LastError:     r.Error,
-			LastState:     r.State,
+				LastError:     lastResult.Error,
+				LastState:     lastResult.State,
 				Timeout:       conf.Timeout,
 				ExpectedState: conf.Target,
 			}
 		}
 	}
+}

helper/resource/state_test.go

@@ -2,6 +2,8 @@ package resource
 
 import (
 	"errors"
+	"strings"
+	"sync/atomic"
 	"testing"
 	"time"
 )
@@ -109,11 +111,18 @@ func TestWaitForState_inconsistent_positive(t *testing.T) {
 }
 
 func TestWaitForState_inconsistent_negative(t *testing.T) {
+	refreshCount := int64(0)
+	f := InconsistentStateRefreshFunc()
+	refresh := func() (interface{}, string, error) {
+		atomic.AddInt64(&refreshCount, 1)
+		return f()
+	}
+
 	conf := &StateChangeConf{
 		Pending:                   []string{"replicating"},
 		Target:                    []string{"done"},
-		Refresh:                   InconsistentStateRefreshFunc(),
-		Timeout:                   90 * time.Millisecond,
+		Refresh:                   refresh,
+		Timeout:                   85 * time.Millisecond,
 		PollInterval:              10 * time.Millisecond,
 		ContinuousTargetOccurence: 4,
 	}
@@ -123,13 +132,27 @@ func TestWaitForState_inconsistent_negative(t *testing.T) {
 	if err == nil {
 		t.Fatal("Expected timeout error. No error returned.")
 	}
-	expectedErr := "timeout while waiting for state to become 'done' (last state: 'done', timeout: 90ms)"
-	if err.Error() != expectedErr {
-		t.Fatalf("Errors don't match.\nExpected: %q\nGiven: %q\n", expectedErr, err.Error())
+
+	// we can't guarantee the exact number of refresh calls in the tests by
+	// timing them, but we want to make sure the test at least went through th
+	// required states.
+	if atomic.LoadInt64(&refreshCount) < 6 {
+		t.Fatal("refreshed called too few times")
+	}
+
+	expectedErr := "timeout while waiting for state to become 'done'"
+	if !strings.HasPrefix(err.Error(), expectedErr) {
+		t.Fatalf("error prefix doesn't match.\nExpected: %q\nGiven: %q\n", expectedErr, err.Error())
 	}
 }
 
 func TestWaitForState_timeout(t *testing.T) {
+	old := refreshGracePeriod
+	refreshGracePeriod = 5 * time.Millisecond
+	defer func() {
+		refreshGracePeriod = old
+	}()
+
 	conf := &StateChangeConf{
 		Pending: []string{"pending", "incomplete"},
 		Target:  []string{"running"},
@@ -148,6 +171,62 @@ func TestWaitForState_timeout(t *testing.T) {
 		t.Fatalf("Errors don't match.\nExpected: %q\nGiven: %q\n", expectedErr, err.Error())
 	}
 
+	if obj != nil {
+		t.Fatalf("should not return obj")
+	}
+}
+
+// Make sure a timeout actually cancels the refresh goroutine and waits for its
+// return.
+func TestWaitForState_cancel(t *testing.T) {
+	// make this refresh func block until we cancel it
+	cancel := make(chan struct{})
+	refresh := func() (interface{}, string, error) {
+		<-cancel
+		return nil, "pending", nil
+	}
+	conf := &StateChangeConf{
+		Pending:      []string{"pending", "incomplete"},
+		Target:       []string{"running"},
+		Refresh:      refresh,
+		Timeout:      10 * time.Millisecond,
+		PollInterval: 10 * time.Second,
+	}
+
+	var obj interface{}
+	var err error
+
+	waitDone := make(chan struct{})
+	go func() {
+		defer close(waitDone)
+		obj, err = conf.WaitForState()
+	}()
+
+	// make sure WaitForState is blocked
+	select {
+	case <-waitDone:
+		t.Fatal("WaitForState returned too early")
+	case <-time.After(10 * time.Millisecond):
+	}
+
+	// unlock the refresh function
+	close(cancel)
+	// make sure WaitForState returns
+	select {
+	case <-waitDone:
+	case <-time.After(time.Second):
+		t.Fatal("WaitForState didn't return after refresh finished")
+	}
+
+	if err == nil {
+		t.Fatal("Expected timeout error. No error returned.")
+	}
+
+	expectedErr := "timeout while waiting for state to become 'running'"
+	if !strings.HasPrefix(err.Error(), expectedErr) {
+		t.Fatalf("Errors don't match.\nExpected: %q\nGiven: %q\n", expectedErr, err.Error())
+	}
+
 	if obj != nil {
 		t.Fatalf("should not return obj")
 	}

helper/resource/wait_test.go

@@ -25,6 +25,21 @@ func TestRetry(t *testing.T) {
 	}
 }
 
+// make sure a slow StateRefreshFunc is allowed to complete after timeout
+func TestRetry_grace(t *testing.T) {
+	t.Parallel()
+
+	f := func() *RetryError {
+		time.Sleep(1 * time.Second)
+		return nil
+	}
+
+	err := Retry(10*time.Millisecond, f)
+	if err != nil {
+		t.Fatalf("err: %s", err)
+	}
+}
+
 func TestRetry_timeout(t *testing.T) {
 	t.Parallel()
 
@@ -39,14 +54,18 @@ func TestRetry_timeout(t *testing.T) {
 }
 
 func TestRetry_hang(t *testing.T) {
-	t.Parallel()
+	old := refreshGracePeriod
+	refreshGracePeriod = 50 * time.Millisecond
+	defer func() {
+		refreshGracePeriod = old
+	}()
 
 	f := func() *RetryError {
 		time.Sleep(2 * time.Second)
 		return nil
 	}
 
-	err := Retry(1*time.Second, f)
+	err := Retry(50*time.Millisecond, f)
 	if err == nil {
 		t.Fatal("should error")
 	}