In this paper, we propose a new technique, called a delayed locking technique, to improve the real-time performance of embedded Linux. The proposed technique employs the rule that entering a critical section is allowed only if the operation does not disturb the future execution of the real-time application. To execute this rule, we introduce the concepts of timer interrupt prediction and lock hold time acquisition. In addition, we designed and implemented a new high-resolution timer that is simple, yet efficient. We implemented the prototype on Linux 2.4.18. Experimental results show that the worst-case OS latency of real-time process is reduced to 23% of the original one, at the expense of slow-down of the non-real-time process by 20%. Though we focus only on embedded Linux, our technique is useful for all kinds of real-time operating systems in which the critical section is significantly long.