Chunming,
I am not an expert on deactivation at all, but here is my suggestion. Since you have a block design, it should be easy to check whether the deactivation is real or something else that was messed up. Open a Graph window from AFNI graphical viewer. Seleact a voxel where you see negative percent signal change, and then overlay the orignal signal with the modeled response curve (Opt --> Trans 1D --> Dataset #2 ...; Opt --> Double Plot --> Overlay).
If you see both the original signal and the modeled response curve are below baseline, you might get real deactivation. Otherwise hopefully you can figure out what went wrong.
Cheers,
Gang