Graph rewiring can improve graph neural network accuracy by mitigating topological pathologies such as oversmoothing and oversquashing, but those same changes may amplify fairness issues. This work studies the fairness-accuracy tradeoff in graph rewiring and introduces topological bias as a fairness metric for evaluating relational bias separately from feature bias. It then proposes a fairness constraint that can be incorporated into classical rewiring methods to reduce topological bias. Experiments show that fairness-constrained rewiring can balance downstream accuracy with fairness in graph learning tasks.