Distance between Relevant Information Pieces Causes Bias in Long-Context LLMs