In the Internet of Vehicles (IoV), personalized federated learning (PFL) can generate personalized models tailored to different local data distributions, thereby improving driving decisions. Existing approaches often perform weighted aggregation on similar vehicles, where the weights are determined by the entire model parameters or loss values. Such model-wise aggre-gation results in inaccurate combination weights, as it overlooks differences in layer-level functionality and imbalances in inter-user accuracy. Additionally, the openness of IoV poses significant threats to system security. In this paper, we propose a blockchain-enabled layer-wise PFL (BLPFL) algorithm to improve system security, model accuracy, and user fairness. Specifically, to preserve personalized knowledge, each local model is split into a globally shared feature extractor and a locally private decision head. To secure the sharing of extractors, Directed Acyclic Graph (DAG)-based blockchain technology is employed. Furthermore, a fine-grained personalized model aggregation mechanism is designed to enhance the model accuracy and user fairness, which explores vehicle similarity at layer granularity and considers special vehicles with severely skewed label distributions. Finally, simulation results demonstrate the effectiveness of the proposed approach in highly heterogeneous scenarios.