[Feature]: Safe IO on io_uring (research project)

[alewycky-tenstorrent]

Is your feature request related to a problem? Please describe.

The throughput of safe IO matters. Can io_uring help? Since 5.19 or so there has been support for uring commands roughly equivalent to ioctls.

Describe the Solution You'd Like

All the IO operations (small and block) supported through io_uring.

Describe Alternatives You've Considered

This is optional and should only be published if it improves performance. The ioctl path will remain available.

Why is this Feature Important?

It may vastly improve the throughput of IO ioctls.

Use Cases

Practical uses depend on how good the performance is. Effective use requires the ability to queue up multiple IO operations and wait for them to complete.

[sempervictus]

To my understanding, io_uring is not a very safe interface itself on various kernel versions and differs in its implementation/backporting state - keeping up with the churn may be a headache.

A number of the storage/network stacks for HPC world avoid ctx switch expense by staying in userspace using DPDK/SPDK/VPP/etc to manage raw device handles or inversely the "keep everything in ring0" plays such as SCST which try to perform zero-copy operations with access to physical page tables (RDMA being another approach to that effect).