xNVMe VS tokio-uring

Basically io_uring grew a lot, the latest API offers incredible tools but there are very few examples in any language (xnvme and fio), and in particular I'm struggling to understand how to do it in Rust: where should unsafe code stop? should I simply expose the io_uring api as unsafe, or should I do more work in C and present the ring when ready to rust?

The foundation of xNVMe is libxnvme, a user space library for working with NVMe devices. It provides a C API for memory management, that is, for allocating physical / DMA transferable memory when needed. An NVMe command interface allowing you to submit and complete NVMe commands in a synchronous as well as an asynchronous manner. Similar to libnvme, however, with a focus on I/O performance and portability. Thus, xNVMe and the libxnvme library works not just on Linux, but also on FreeBSD, Windows, and MacOS.

While Mio will probably not implement uring in its current design, there's https://github.com/tokio-rs/tokio-uring if you want to use io_uring in Rust.

It's still in development, but the Tokio team seems intent on getting good io_uring support at least!

As the README states, the Rust implementation requires a kernel newer than the one that shipped with Ubuntu 20.04 so I think it'll be a while before we'll see significant development among major libraries.

That's what the pool is for: https://github.com/tokio-rs/tokio-uring/blob/master/src/buf/fixed/pool.rs

Tokio supports io_uring (https://github.com/tokio-rs/tokio-uring), so perhaps when it's mature and battle-tested, it'd be easier to transition to it if Cloudflare aren't using it already.

- Tokio suffers from a similar problem

Eg via tokio-uring.

I strongly recommend you to look into io-uring and use async executors that take advantages of it: - tokio-uring (not recommended as it is still undergoing development) - monoio - glommio

This is my thinking as well. Specifically, I realized that if you don’t use tasks, but rather futures and join, than structured concurrency just works out (at the cost of less efficient poll). In a single-threaded/thread-per-core runtime, tasks could have the same semantics as futures. Somewhat elaborated here: https://github.com/tokio-rs/tokio-uring/issues/81

There's a new API on Linux called io_uring that has performance benefits, but most executors don't use it yet, except executors meant specifically to harness the power of io_uring like tokio-uring and Glommio

Breaking it down a bit further- Rust's async is zero-cost, and there's no way to write faster equivalent code to the language construct in Rust (and presumably other LLVM languages). Tokio introduces abstractions over OS APIs (indirectly) and provides a runtime. The runtime isn't zero cost, but it is likely to be better optimized for "standard" situations than a homebrewed solution, and its primary competition is in the form of other large async runtimes. On the other hand, Tokio's IO routines are (AFAIK) about as well written as one can get with blocking OS APIs, and the only competitors in that space are projects like tokio-uring that use APIs more well suited for asynchronous usage.