It looks like if you use zfs as the root filesystem, the ZPL delete queue
on that filesystem does not run (or at least, not enough):
z_delete_head = {
...
z_drained = 0x1
z_draining = 0
z_thread_target = 0
z_thread_count = 0
z_znode_count = 0x13da2
...
}
Work Around
After the system boot up you can do the following to get the necessary delete thread running. This should flush the in-core delete queue.
# zfs set readonly=off <pool>/<fs_name>