Prior work on multiprocessor real-time locking protocols has shown how to support fine-grained lock nesting with asymptotically optimal worst-case priority-inversion blocking (pi-blocking) bounds. However, contention for each resource has heretofore been considered an unconstrained variable. This paper presents the first fine-grained multiprocessor realtime locking protocol with contention-sensitive worst-case pi-blocking bounds. Contention-sensitive pi-blocking is made possible by incorporating knowledge of maximum criticalsection lengths-which must be known a priori for analysis anyway-into the lock logic.