Accurate parametrization of force fields (FFs) is of ultimate importance for computer simulations to be reliable and to possess a predictive power. In this work, we analyzed, in multi-microsecond simulations of a 40-base-pair DNA fragment, the performance of four force fields, namely, the two recent major updates of CHARMM and two from the AMBER family. We focused on a description of double-helix DNA flexibility and dynamics both at atomistic and at mesoscale level in coarse-grained (CG) simulations. In addition to the traditional analysis of different base-pair and base-step parameters, we extended our analysis to investigate the ability of the force field to parametrize a CG DNA model by structure-based bottom-up coarse-graining, computing DNA persistence length as a function of ionic strength. Our simulations unambiguously showed that the CHARMM36 force field is unable to preserve DNA's structural stability at over-microsecond time scale. Both versions of the AMBER FF, parmbsc0 and parmbsc1, showed good agreement with experiment, with some bias of parmbsc0 parameters for intermediate A/B form DNA structures. The CHARMM27 force field provides stable atomistic trajectories and overall (among the considered force fields) the best fit to experimentally determined DNA flexibility parameters both at atomistic and at mesoscale level.