Lock Granularity Trade-off

A concurrent hash table uses a single mutex for all operations. A colleague suggests switching to per-bucket locks. What is the trade-off?