In this paper, we perform an empirical study on commercial 5G. Unlike previous works studying 5G in early deployments, we focus on fully deployed 5G networks that already serve massive numbers of users, and in particular, we perform controlled experiments to analyze TCP and multipath TCP (MPTCP) performances and behaviors over 5G. Our results show that 5G faces a dilemma in maintaining in-network buffer for TCP traffic: on one hand, TCP packets are less likely to accumulate within the network, thanks to the high data rate of 5G radio; but on the other hand, with TCP performance highly sensitive to signal quality, when signal becomes poor abruptly in mobile environment, the in-network buffer easily overflows, leading to packet loss bursts. We also confirm that employing MPTCP can aggregate capacities on 5G and 4G paths, but under the current deployment model where 5G and 4G base stations co-locate and share the same core network, MPTCP does not necessarily enhance communication reliability. In addition to active measurement experiments, we leverage a speedtest service to evaluate commercial 5G from a nationwide perspective. By exploiting over 13 k crowdsourced speedtest results reported from 197 cities in China, we find that there exist considerable regional differences on 5G performances across the nation, and with the 4G co-locating deployment model, the 5G networks in some populous metropolitan areas have inferior performances. Our study provides insights for people to understand commercial 5G and to optimize upper-layer protocols and applications over 5G.