[bug report] nvme: fio workload hangs on 16m random read

[Keith Busch]

On Thu, Mar 06, 2025 at 09:46:37PM +0100, Andreas Hindborg wrote:
> Hi All,
> 
> My CI system has found an issue for the following fio workload:
> 
> root at debian:~# insmod /mnt/v6.14-rc5-build/drivers/nvme/host/nvme-core.ko
> root at debian:~# insmod /mnt/v6.14-rc5-build/drivers/nvme/host/nvme.ko poll_queues=1
> 
> "fio" "--group_reporting" "--name=default" "--filename=/dev/nvme0n1"
> "--readwrite=randread" "--blocksize=16777216" "--direct=1" "--numjobs=1"
> "--ioengine=io_uring" "--iodepth=1" "--hipri=1" "--time_based=1"
> "--runtime=3000"
> 
> fio will hang, some times immediately, some times after a few minutes.
> 
> I observed this on v6.14-rc1 and v6.14-rc5. It is not present in v6.13.
> 
> Can any of you reproduce this? I was able to reproduce this on real
> hardware and in qemu.

Could you try adding this patch to your kernel?

https://lore.kernel.org/io-uring/92b0a330-4782-45e9-8de7-3b90a94208c2@kernel.dk/