Support for IO_uring communication with Kernel

[darthShadow]

References:

• https://www.phoronix.com/news/Linux-6.14-FUSE
• https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=92cc9acff7194b1b9db078901f2a83182bb73202
• https://docs.kernel.org/next/filesystems/fuse-io-uring.html

[hanwen]

I did some prototyping for virtio (see https://github.com/hanwen/go-fuse/tree/virtiofs), which seems somewhat similar.

The doc explains you get one queue per CPU, which means we'd have to do something special so the go side respects this affinity. Our read routines also have global locks (Server.reqMu) which wil work poorly if split across multiple CPUs. Maybe we could use the singleReader pattern?

Combining the write of one response with the read of the next one seems intereesting, though.

[hanwen]

the implementation in libfuse hasn't landed yet, it seems.

[hanwen]

Also, can you say why you want this?

[darthShadow]

No real need/use, but better performance is always welcome 😄.

I believe this is the WIP tree for the libfuse change: https://github.com/bsbernd/libfuse/commits/uring/, but yeah, probably better to wait for it to get merged for an official reference.

[darthShadow]

Looks like it's been merged into mainline now:

libfuse/libfuse#1214
libfuse/libfuse#1177

[1037827920]

Hi! Is there an active plan to move this feature forward? I'm also quite interested in this feature.