To Forget or Not? Towards Practical Knowledge Unlearning for Large Language Models