io-uring VS rio

Depending on whether the system calls for accessing the network stack will become the bottleneck, you may also need to use io_uring with kernel side polling for this (briefly explained in this paper on page 15) in addition to pre-allocating your own buffers and avoiding copying memory where you can. You mention you are using tokio, it has a library for io_uring and there is also this vanilla io_uring library crate, but it's a bit more low level. Both are for linux only and I'm not familiar enough with Windows to say if there exists something similar in Rust for Windows systems (which has I/O Rings), maybe someone else can comment on that?

Hmm, nothing really jumps out for me in the pastes you have here, apart from this. What kernel are you running? Have you tried LTS?

Also, not to be too bad about a reputation fallacy, but I found the author to be flippant and disrespectful when good-faith unsoundness was pointed out in his crates: https://github.com/spacejam/rio/issues/30

In my previous post on linear types I spent quite a bit of time motivating linear types. For example the ergonomic rio io_uring library could be made sound if it could guarantee destructors are run. Or performing FFI with async C++ could be made more efficient if it could rely directly on destructors rather than having to involve an intermediate runtime for each call.

It's like cargo should have a way to mark a dependency as unsafe. That way, you could have a safe mmap crate as an unsafe dependency. Or something like rio which is deliberately unsound (but is fine if you abide by its rules through the entirety of the program)

for completeness there is also rio, but:

rio still has known soundness issues– its Completion futures block the thread when dropped (!!!), and can allow for use-after-free bugs if leaked. See https://github.com/spacejam/rio/issues/30 for details.

Here are some posts about the design. https://without.boats/blog/io-uring/ https://github.com/spacejam/rio/issues/30 https://github.com/axboe/liburing/issues/109

The author of sled[1], an embedded database in Rust which has a number of promising features, has also written parts of rio[2], an underlying pure Rust io_uring library, which is intended to become the core write path for sled. rio has support for files but also has a demo for TCP (on Linux 5.5 and later) and O_DIRECT.

I tested rio recently as I had a Brilliant but Bad Idea™ involving file access and was pleasantly surprised by the API, as I have been with sled's.

I'm excited for the experimentation in the Rust ecosystem and for such low level crates to handle the complex io_uring tasks (relatively) safely!

[1]: https://github.com/spacejam/sled

[2]: https://github.com/spacejam/rio