sync: deflake TestWaitGroupMisuse2

Also runs 100X faster on average, because it takes so many
fewer attempts to trigger the failure.

Fixes #11443.

Change-Id: I8c39ee48bb3ff6c36fa63083e04076771b65a80d
Reviewed-on: https://go-review.googlesource.com/36841
Run-TryBot: Russ Cox <rsc@golang.org>
Reviewed-by: Dmitry Vyukov <dvyukov@google.com>

1 files changed, 16 insertions, 6 deletions

diff --git a/src/sync/waitgroup_test.go b/src/sync/waitgroup_test.go
index 8ec34fd343..c7c391ba23 100644
--- a/src/sync/waitgroup_test.go
+++ b/src/sync/waitgroup_test.go
@@ -70,11 +70,8 @@ func TestWaitGroupMisuse(t *testing.T) {
 
 func TestWaitGroupMisuse2(t *testing.T) {
 	knownRacy(t)
-	if testing.Short() {
-		t.Skip("skipping flaky test in short mode; see issue 11443")
-	}
-	if runtime.NumCPU() <= 2 {
-		t.Skip("NumCPU<=2, skipping: this test requires parallelism")
+	if runtime.NumCPU() <= 4 {
+		t.Skip("NumCPU<=4, skipping: this test requires parallelism")
 	}
 	defer func() {
 		err := recover()
@@ -86,24 +83,37 @@ func TestWaitGroupMisuse2(t *testing.T) {
 	}()
 	defer runtime.GOMAXPROCS(runtime.GOMAXPROCS(4))
 	done := make(chan interface{}, 2)
-	// The detection is opportunistically, so we want it to panic
+	// The detection is opportunistic, so we want it to panic
 	// at least in one run out of a million.
 	for i := 0; i < 1e6; i++ {
 		var wg WaitGroup
+		var here uint32
 		wg.Add(1)
 		go func() {
 			defer func() {
 				done <- recover()
 			}()
+			atomic.AddUint32(&here, 1)
+			for atomic.LoadUint32(&here) != 3 {
+				// spin
+			}
 			wg.Wait()
 		}()
 		go func() {
 			defer func() {
 				done <- recover()
 			}()
+			atomic.AddUint32(&here, 1)
+			for atomic.LoadUint32(&here) != 3 {
+				// spin
+			}
 			wg.Add(1) // This is the bad guy.
 			wg.Done()
 		}()
+		atomic.AddUint32(&here, 1)
+		for atomic.LoadUint32(&here) != 3 {
+			// spin
+		}
 		wg.Done()
 		for j := 0; j < 2; j++ {
 			if err := <-done; err != nil {