Can FD use io_uring for faster performance?

[ScottRosenberg2]

I'm not sure if it actually would help, but could this use io_uring for faster performance, by using less syscalls?

There are some rather impressive benchmarks of io performance using io_uring, so I was curious if it could help in this case.

[tavianator]

There is no io_uring operation for reading directories yet. I have experimented with a setup that uses io_uring to open multiple dirs at once and distributes getdents() to a thread pool, but io_uring doesn't seem to help over just doing open() in the same thread pool.

[tavianator]

I see a new version of the getdents() patch has been posted: https://lore.kernel.org/io-uring/[email protected]/T/

[ScottRosenberg2]

From what I understand, io_uring can be used in a threadpool, but you basically have to have an instance of the circular buffer for each thread