Why multiple caches need coordination
When several cores cache copies of a memory location, a write by one core must be coordinated with those copies. Cache coherence concerns the consistency of observations of that location. Implementations can use line states, invalidation, and ownership transfers, although the exact protocol is processor-specific. The mechanism does not require every write to reach main memory immediately. A cache can hold modified data while the overall system still maintains the coherence rules. Distinguish the abstract guarantee from a particular assumption about where the newest physical copy resides.
An invalidation example
Consider an illustrative invalidation-based protocol. Cores A and B both hold a readable copy of one line. Before A modifies it, A obtains the required write permission and B's old copy is invalidated. A later B access cannot simply use that stale cached copy. The example explains the purpose of coordination without claiming that all processors use exactly these states. Notice that the management unit can be an entire cache line, even when the application mutation concerns only one small variable inside it.
How false sharing appears
Two threads can update different counters and still compete if those counters occupy one cache line. Repeated ownership transfers can then impose a cost even though the application-level variables are distinct. This is false sharing: the shared hardware management unit, rather than a logically shared variable, creates interference. Layout changes or padding can help an observed case, but increase memory usage and depend on actual alignment and line size. Measure the write pattern before adding padding everywhere, especially when the larger footprint could worsen capacity behavior elsewhere.
Coherence does not replace synchronization
Coordinating writes to one location is not the same as ordering a payload store before a flag store at another location. A processor with coherent caches can still require the synchronization prescribed by the programming-language memory model. Locks or correctly paired release and acquire operations provide guarantees that cache coherence alone does not establish. It is therefore unsafe to infer that a data race is harmless because the hardware is coherent. Evaluate the costs of sharing separately from the proof that the application publishes and accesses its data correctly.