Shared-state versus message-passing

One of the most basic distinctions made in concurrency theory is between shared-state concurrency and message-passing concurrency. I'm finally getting around to reading John Reppy's book on Concurrent ML, and he has a succinct characterization of the two:

Shared-memory languages rely on imperative features ... for interprocess communication, and provide separate synchronization primitives. Message-passing languages provide a single unified mechanism for both synchronization and communication.