Imagine a shared counter being updated while several tasks are running. An atomic operation behaves like a single indivisible move: other threads see either the old value or the completed new value, never a half-finished update. section 8.6
Atomic operations are a simple synchronization technique for primitive data. Instead of acquiring a lock, changing a value, and releasing the lock, a supported hardware instruction completes the whole operation without interruption. This can reduce synchronization costs because competing tasks are not blocked. section 8.6
Use atomicity with care:
- Treat atomicity as an environment-dependent guarantee: hardware, the runtime, and the programming language must support it. section 8.7
- Do not assume every operation or data structure is atomic; use documented atomic types or structures when available. section 8.6
- Use locks when an operation is not atomic, and consider message passing or avoiding shared resources when synchronization can be designed away. section 8.7 section 9.10
A hardware-level increment
An instruction such as “add 0x9082a1b, $0x1” represents an increment of a memory location. Hardware guarantees that the addition either does not happen or reaches completion, with no visible intermediate state. section 8.6
add 0x9082a1b, $0x1
Inspect the assembly-style increment that illustrates an atomic update at a specific memory address. section 8.6
