At their core, they are distinct only regarding when data is copied from kernel to user space. Readiness based copies when the application does the read() after

At their core, they are distinct only regarding when data is copied from kernel to user space. Readiness based copies when the application does the read() after being notified of "readiness", completion based copies (or DMAs directly from disk controller and maps into user space) before the application is notified of completion. The former is easier to use (no buffer ownership headaches), the latter is faster. Though there may be buffer management headaches in the kernel with readiness - what to do about ready reads the aren't collected? (Probably just throw them away if uncollected at the next AIO poll...)