diff --git a/drivers/android/binder.c b/drivers/android/binder.c
index 2899534b4727..b02a0b33ca7e 100644
--- a/drivers/android/binder.c
+++ b/drivers/android/binder.c
@@ -836,7 +836,7 @@ binder_enqueue_work_ilocked(struct binder_work *work,
 }
 
 /**
- * binder_enqueue_thread_work_ilocked_nowake() - Add thread work
+ * binder_enqueue_deferred_thread_work_ilocked() - Add deferred thread work
  * @thread:       thread to queue work to
  * @work:         struct binder_work to add to list
  *
@@ -847,8 +847,8 @@ binder_enqueue_work_ilocked(struct binder_work *work,
  * Requires the proc->inner_lock to be held.
  */
 static void
-binder_enqueue_thread_work_ilocked_nowake(struct binder_thread *thread,
-					  struct binder_work *work)
+binder_enqueue_deferred_thread_work_ilocked(struct binder_thread *thread,
+					    struct binder_work *work)
 {
 	binder_enqueue_work_ilocked(work, &thread->todo);
 }
@@ -3333,7 +3333,14 @@ static void binder_transaction(struct binder_proc *proc,
 	} else if (!(t->flags & TF_ONE_WAY)) {
 		BUG_ON(t->buffer->async_transaction != 0);
 		binder_inner_proc_lock(proc);
-		binder_enqueue_thread_work_ilocked_nowake(thread, tcomplete);
+		/*
+		 * Defer the TRANSACTION_COMPLETE, so we don't return to
+		 * userspace immediately; this allows the target process to
+		 * immediately start processing this transaction, reducing
+		 * latency. We will then return the TRANSACTION_COMPLETE when
+		 * the target replies (or there is an error).
+		 */
+		binder_enqueue_deferred_thread_work_ilocked(thread, tcomplete);
 		t->need_reply = 1;
 		t->from_parent = thread->transaction_stack;
 		thread->transaction_stack = t;