Goal

• I want to take advantage of io_uring performance optimizations for storage in QEMU. Modify the ioeventfd code in QEMU to use IORING_OP_READ instead of IORING_OP_POLL_ADD + read(2) because it is faster to do it in a single read operation.
• A prototype demonstrated that there is a modest improvement in IOPS on <10 microsecond NVMe drives:

Operation BS QD IOPS   Change
randread  4k  1 112739 +1.7%
randread  4k 64 504866 +0.068%
randwrite 4k  1 110171 +5.7%
randwrite 4k 64 516398 +5.3%

(Full data here: https://gitlab.com/stefanha/virt-playbooks/-/commit/46161cd6c877759bd8ade826f953b7fd520abab7)

• Code for a prototype is available here:
  https://gitlab.com/stefanha/qemu/-/commits/io_uring-eventnotifier-reads?ref_type=heads

Acceptance criteria

A list of verification conditions, successful functional tests, or expected outcomes in order to declare this story/task successfully completed.

• fio randread and randwrite with blocksize 4 KB and iodepths 1 and 64 does not degrade on a virtio-blk disk backed by a host NVMe block device. The guest has 4 vCPUs and the virtio-blk disks are configured with 4 IOThreads.