Knowledge graph reasoning plays a vital role in various applications and has garnered considerable attention. Despite the impressive performance achieved by recent path-based methods, they may face limitations due to constraints in message-passing neural networks, including missing paths and information over-squashing. In this paper, we revisit the application of transformers for knowledge graph reasoning and propose a novel approach called KnowFormer, which utilizes a transformer architecture to conduct reasoning on knowledge graphs from a message-passing perspective, rather than by encoding textual information like previous transformer-based methods. By leveraging the all-pair interaction of the attention mechanism, KnowFormer effectively addresses the constraints faced by path-based methods. To achieve this, we introduce a query prototype based attention definition, facilitating convenient construction and efficient optimization. Furthermore, we introduce two sub-modules to enable the construction of a structure-aware self-attention. Additionally, we present an efficient attention computation method to enhance scalability. Experimental results demonstrate the superior performance of KnowFormer compared to prominent baseline methods on both transductive and inductive benchmarks.