This modification adds support to benchmark ZNS finish operations in io_uring_cmd (io_uring passthrough) mode.
By specifying finish=1 on a trim workload, fio will finish the zone instead of trimming it (resetting).
It requires that a zone be open and written to (i.e., a 4KiB write before a finish), and that threads are used.
It requires that io_uring_cmd is set.
sudo nvme zns reset-zone /dev/nvme0n2 -aExample job for running finish benchmark (replacd ${DEV} with the device name, and ${SIZE} with the size in zones of the number of zones to finish). The job is configured with 1 write of 4K then 1 finish on the ZNS zone. A single write is required to open the zone and write a single block (or write as much data as bs is set to), as less allocated zones have a higher finish latency (e.g., 10% filled zone has higher latency than 60% filled zone).
Fio will issue the configured block size bs as a single I/O and then immediately issues a finish. The finish currently only tracks the number of I/Os and the bandwidth (latency is not being tracked currently).
The finish are submitted as nvme commands with ioctl through io_uring passthrough (bypassing the I/O scheduler). Hence, the finish is done synchronous with only a single finish at a time.
If the desired behavior is to finish zones concurrently, using concurrent jobs at offset_increment jobs with each a finish workload can issue concurrent finish commands.
[global]
name=finish
filename=/dev/${DEV}
ioengine=io_uring_cmd
sqthread_poll=1
size=${SIZE}z
bs=4K
group_reporting
zonemode=zbd
thread=1
lat_percentiles=1
[finish]
rw=write
iodepth=1
numjobs=1
bs=4K
finish=1``