[lexoliu]

With io-uring-cqe32 the ring is set up with IORING_SETUP_CQE32, but the driver only passes res and flags on to the op (create_entry, multishot path), so the extra 16 bytes are dropped. Ops that return data there can't be written as an OpCode on top of compio: IORING_OP_URING_CMD, and IORING_OP_RECV_ZC, whose ring has to be CQE32 in the first place (zcrx.c, v6.18).

This copies big_cqe into Extra on both paths and adds Extra::big_cqe(), which returns Unsupported when the feature is off. Without the feature nothing changes size.

The test submits a NOP with IORING_NOP_CQE32, which echoes sqe->off and sqe->addr into the big CQE (nop.c, v6.18), and checks both values come back through Extra. It passes on Linux 6.18 and skips on older kernels; the io-uring-sqe128,io-uring-cqe32 CI lane is the one that builds it.

Choosing the CQE size per ring, or IORING_SETUP_CQE_MIXED, would spare every other ring in the process the larger CQ. I left that out: the io-uring crate types its CQ by entry size and has no mixed-ring support yet.

[lexoliu]

I'm building skyzen (https://github.com/zen-rs/skyzen), an HTTP server framework that runs thread-per-core on compio, with its own HTTP/1.1 and HTTP/3 over io_uring. An early review would let us build on upstream compio rather than a fork. Happy to reshape anything.

Also, thank you for building such a great library❤️.

[Berrysoft]

Also test fusion please. You can use IourOpCode instead.