The goal of document-level coreference resolution is to recognize all the mentions pointing to the same entity in a document, and its main challenge lies in the computational complexity and explicit memory requirement problems faced when dealing with long texts. Existing methods reduce the computational pressure by pruning strategy on one hand, and on the other hand, adopt distance-based heuristic strategy to make some preferential selective assumptions on the distance information between mentions. Although these methods improve the computational efficiency to some extent, they fail to effectively solve the balance problem between information loss and efficiency. For this reason, this paper proposes a distance feature-based coreference resolution model DisCoref. DisCoref analyzes the role of distance features in depth from three key aspects, namely, mention prediction, coreference representation, and coreference prediction, and thus achieves an effective balance between information loss and efficiency. First, distance feature is introduced in mention prediction to help the model identify unreasonable candidate mentions across sentences. Next, joint semantic information, entity type information and sentence information are used for coreference representation to strengthen the distance perception of coreference. Finally, based on the relative distance between mentions, the coreference prediction is optimized in terms of soft and hard constraints based on distance to improve the clustering robustness. The experimental results prove that DisCoref has obvious superiority, and offers new insights into understanding coreference resolution tasks and advances document-level knowledge extraction.