Given that the caches do all of this work to provide coherence, do programs (or the OS itself) have to worry about anything when they access shared data? The answer, unfortunately, is yes, and is documented in great detail in the second piece of this course on the topic of concurrency. While we won’t get into the details here, we’ll sketch/review some of the basic ideas here (assuming you’re familiar with concurrency).
When accessing (and in particular, updating) shared data items or structures across CPUs, mutual exclusion primitives (such as locks) should likely be used to guarantee correctness (other approaches, such as building lock-free data structures, are complex and only used on occasion; see the chapter on deadlock in the piece on concurrency for details). For example, assume we have a shared queue being accessed on multiple CPUs concurrently. Without locks, adding or removing elements from the queue concurrently will not work as expected, even with the underlying coherence protocols; one needs locks to atomically update the data structure to its new state.
Example
To make this more concrete, imagine this code sequence, which is used to remove an element from a shared linked list, as we see in the snippet below:
Get hands-on with 1200+ tech skills courses.