diff --git a/cont.c b/cont.c
index 095cc86767..a43950c6cb 100644
--- a/cont.c
+++ b/cont.c
@@ -1288,6 +1288,7 @@ root_fiber_alloc(rb_thread_t *th)
     rb_fiber_t *fib;
     /* no need to allocate vm stack */
     fib = fiber_t_alloc(fiber_alloc(rb_cFiber));
+    fib->cont.saved_thread.ec.stack = NULL;
     fib->cont.type = ROOT_FIBER_CONTEXT;
 #if FIBER_USE_NATIVE
 #ifdef _WIN32
diff --git a/thread.c b/thread.c
index b7ee1d8d9b..0fa37b70f0 100644
--- a/thread.c
+++ b/thread.c
@@ -694,10 +694,8 @@ thread_start_func_2(rb_thread_t *th, VALUE *stack_start, VALUE *register_stack_s
 	rb_threadptr_unlock_all_locking_mutexes(th);
 	rb_check_deadlock(th->vm);
 
-	if (!th->root_fiber) {
-	    rb_thread_recycle_stack_release(th->ec.stack);
-	    th->ec.stack = 0;
-	}
+	rb_thread_recycle_stack_release(th->ec.stack);
+	th->ec.stack = NULL;
     }
     native_mutex_lock(&th->vm->thread_destruct_lock);
     /* make sure vm->running_thread never point me after this point.*/
diff --git a/vm.c b/vm.c
index 6815e84f76..266b251485 100644
--- a/vm.c
+++ b/vm.c
@@ -2330,7 +2330,7 @@ static int thread_recycle_stack_count = 0;
 static VALUE *
 thread_recycle_stack(size_t size)
 {
-    if (thread_recycle_stack_count) {
+    if (thread_recycle_stack_count > 0) {
 	/* TODO: check stack size if stack sizes are variable */
 	return thread_recycle_stack_slot[--thread_recycle_stack_count];
     }
@@ -2346,6 +2346,8 @@ thread_recycle_stack(size_t size)
 void
 rb_thread_recycle_stack_release(VALUE *stack)
 {
+    VM_ASSERT(stack != NULL);
+
 #if USE_THREAD_DATA_RECYCLE
     if (thread_recycle_stack_count < RECYCLE_MAX) {
 	thread_recycle_stack_slot[thread_recycle_stack_count++] = stack;
@@ -2429,8 +2431,9 @@ thread_free(void *ptr)
     rb_thread_t *th = ptr;
     RUBY_FREE_ENTER("thread");
 
-    if (!th->root_fiber) {
-	RUBY_FREE_UNLESS_NULL(th->ec.stack);
+    if (th->ec.stack != NULL) {
+	rb_thread_recycle_stack_release(th->ec.stack);
+	th->ec.stack = NULL;
     }
 
     if (th->locking_mutex != Qfalse) {