Disclaimer: I'm not a Rust programmer.

Having written a few libs for working with io_uring (in Ruby), cancellation is indeed tricky, with regards to keeping track of buffers. This is an area where working with fibers (i.e. stackful coroutines) is beneficial. If you keep metadata (and even buffers) for the ongoing I/O op on the stack, there's much less book-keeping involved. Managing the I/O op lifetime, especially cleaning up, becomes much simpler, as long as you make sure to not return before receiving a CQE, even after having issued a cancellation SQE.