Tony Hutter

@MigeljanImeri I think there's a way we could do this and still get deadman detection to still work. We would basically do the "hybrid" method but add the bypass ZIOs...

> Call vdev_queue_length() to see if the non-bypass queue is empty. You make this a threshold too, like "allow bypass if the normal queue is 10% full or less". The...

Nice, I just retested and am only seeing a slight decrease as well: ``` IOPS fio, blocks=4k, seq read, direct=1 bypass=292k, previously 280k queue=140k, previously 150k fio, blocks=4k, randread, direct=1...

Could you reword the commit message to call out the `zfs_vdev_queue_bypass_pct` module param. Like maybe: ``` Allow bypassing the vdev queue on SSDs Allow bypassing the vdev queue on SSDs...

@MigeljanImeri looks like there's a NULL pointer dereference happening in the test suite: ``` [ 1641.645984] BUG: unable to handle kernel NULL pointer dereference at 0000000000000000 [ 1641.649605] PGD 0...

> don't compile with Centos last kernel-5.14.0-446.el9.x86_64 Which distro are you running specifically? zfs-2.1.15 builds on AlmaLinux 9.4 with the latest 5.14.0-427.16.1.el9_4.x86_64 kernel: ``` $ ./autogen.sh && ./configure && make...

Unfortunately we don't support CentOS stream

@smusich you should also be able to install the EL9 dkms packages as well. That should alleviate the issues when the RHEL/Alma/Rocky/Stream kernel version numbers are not 100% matched up.