What were you trying to do that didn't work?

While profiling qemu-kvm I noticed that perf-top(1) lists virtqueue_split_pop() as one of the biggest CPU consumers. This function is called frequently during I/O and was zeroing many KBs of memory after the -ftrivial-auto-var-init=zero flag was enabled in RHEL 9.5.

Further investigation by Daniel Berrangé showed that there are more places in qemu-kvm where performance regressions could occur:

https://lore.kernel.org/qemu-devel/aEHn5XHU0-N67gWs@redhat.com/

What is the impact of this issue to you?

This is a performance regression. Earlier qemu-kvm releases were built without ftrivialauto-var-init=zero.

Avoiding zeroing memory in virtqueue_split_pop() improves fio randread bs=4k iodepth=64 performance from 304k to 332k IOPS (+9%).

Please provide the package NVR for which the bug is seen:

How reproducible is this bug?:

100%

Steps to reproduce

Run the following fio command-line:

BLKDEV=/dev/vda
fio --ioengine=io_uring \
            --ioscheduler=none \
            --numjobs=$(nproc) \
            --direct=1 \ 
            --runtime=30 \
            --ramp_time=5 \
            --clocksource=cpu \
            --norandommap \
            --cpus_allowed_policy=split \
            --rw=randread \
            --bs=4k \
            --iodepth=64 \ 
            --time_based=1 \
            --group_reporting \
            --name=job1 \
            --filename=$BLKDEV

Expected results

After the patch performance improves.

Actual results

Before the patch performance is lower.