A good rule to follow in software engineering in general is "never write your own lock free data structures". I'm the one writing them so that my coworkers don't need to, though. I've used atomics to implement various flavors of queues in low level C for use on microcontrollers. A good single-producer single-consumer queue abstraction is helpful for communicating with interrupt service routines in peripheral drivers. Bespoke multiple-producer single or multiple-consumer queues are often useful for implementing logging abstractions that work from any context, including in interrupt handlers and critical sections.