Large Language Models (LLMs) excel in various applications, from conversational agents to medical diagnostics. However, they often inherit societal biases from training data, risking discriminatory outcomes. This survey provides a structured review of fairness in LLMs, analyzing the origins of bias in data and model design. We summarize key bias evaluation metrics, including embedding-based and probability-based approaches, and examine mitigation techniques in pre-processing, in-processing, and post-processing strategies. We also highlight essential datasets and tools for fairness research and discuss current challenges, offering future directions to guide the development of fairer LLMs.