[ttanev]

Tl; dr: this setup is probably very close to the GPU pass-through (with a twist) many folks are using for GPU-heavy use-cases in a VM already, so not sure if it could be classified as unusual.

At work, we use VMs with NIC virtual function interfaces (https://www.kernel.org/doc/html/latest/driver-api/vfio.html) from the hypervisor for first line of validation for our product (full disclosure: part of the quality and automation team at StorPool Storage), so that we can have an environment both close to the production systems with network hardware acceleration enabled in the VM (i.e. kernel bypass) and easy to reproduce and re-create. There are some limitations and setup quirks with different orchestrations, but they are not live-migrated anyway, so not really an issue.

At home the main usage is for isolated environments (not unusual), and recently to be able to access a ZFS in an older Freenas/Freebsd zpool drive (the Ubuntu ZFSoL could not detect it for some reason, haven't looked into it more). I was a bit surprised that passing through the whole storage controller with the sole SATA disk behind it (the host root OS is on an NVMe) is actually slower than attaching to the VM as virtio raw disk (i.e. /dev/sdN).