The start pass reads a module's look-again mark after letting go, and hands it on to takeDoneKey (issue 1945)

cmd/mesh-controller/composed_keys.go

@@ -355,11 +355,19 @@ const startTakeIns = 16
 // which tries again every doneKeyRetry, loudly, so a failing module never holds up the look at the others.
 func takeDoneKeyAtStart(ctx context.Context, open *stores, module string) {
 	mu, _ := doneKeysMu.LoadOrStore(module, &sync.Mutex{})
-	mu.(*sync.Mutex).Lock()
-	err := takeDoneKeyOnce(ctx, open, module)
+	if !mu.(*sync.Mutex).TryLock() {
+		// A take-in already follows the module: told to look again, it takes what the start would have (issue 1945).
+		doneKeysLookAgain.Store(module, true)
+		return
+	}
+	doneKeysLookAgain.Delete(module) // what this look reads covers every wake before it
+	err := doneKeyTakeOnce(ctx, open, module)
 	mu.(*sync.Mutex).Unlock()
-	if err != nil {
-		// A failure or a wait is followed by takeDoneKey, which says it and looks again (issue 1945).
+	// Read after letting go, as every holder of a module's lock does: a wake during this look only marked the module
+	// (the review of issue 1945).
+	_, again := doneKeysLookAgain.LoadAndDelete(module)
+	if err != nil || again {
+		// A failure, a wait or a wake meanwhile is followed by takeDoneKey, which says it and looks again.
 		inBackground(func() { takeDoneKey(ctx, open, module) })
 	}
 }