Federated graph learning (FGL) has demonstrated great potential in collaboratively training graph neural network (GNN) models under the federated learning (FL) framework, benefiting from its advantage of privacy preservation. However, FGL suffers from the significant challenge of non-IID problem. In the real world, cross-source or even cross-domain graph data frequently exist among clients. To tackle this challenge, we propose a novel FGL framework, FedDAGC, from the new perspective of graph coarsening and consensus. Specifically, our method inserts a dynamic adaptive graph coarsening (DAGC) module into local GNN models to generate coarsened graphs as augmented inputs for subsequent model training. By sharing the intermediate parameters in DAGC instead of traditional GNN model parameters, FedDAGC enables clients to reach a consensus on coarsening directions, for extracting crucial patterns from local graphs based on common core knowledge while protecting privacy. Extensive experiments and analyses on various non-IID settings demonstrate the efficacy and superiority of FedDAGC.