Tony Hutter

@robn I was actually able to get this PR to trigger deadman events even with `queue_io=off`. I set the ZFS params to: ``` zfs_deadman_checktime_ms=100 zfs_deadman_ziotime_ms=500 zfs_deadman_enabled=1 ``` I then created...

I'm going to see if I can starve out a scrub and MMP write with `queue_io=off`.

@robn - some random thoughts: One alternative is to pull it in as a vdev property and then keep it in master until 2.4.x. So don't backport it to 2.3.x...

I think you raise some good points on the vdev prop vs module param discussion. I'll have to think it over a little more. > What if it didn't affect...

@MigeljanImeri I've been trying to test/benchmark this PR, but in the process I've uncovered some unrelated bugs in ZFS (https://github.com/openzfs/zfs/issues/16707, https://github.com/openzfs/zfs/issues/16695) that are slowing down my testing. Can you describe...

I just tested with a single NVMe pool. Sure enough, turning off queuing does improve IOPS with O_DIRECT: -- IOPS -- fio, blocks=4k, seq read, direct=1 queue on:150k queue off:...

I did get over 500k IOPS with direct=0, but I'm assuming that's due to caching.

I'll try to run my "scrub while FIOing" test again, but this time using the NVMe drive, since there seems to be weird stuff going on with brd + ZFS.

I'm having trouble seeing the benefits of the "Direct IOs can bypass the queue" special case. What does that get us over just having `queue_io` control all IOs? If the...

> I do think this is a per-vdev option. I just don't think it's a property, because those are saved, which implies a certain amount of confidence that we have...