Virtual Function I/O

Presented by Simone

Virtualization Primer

Hypervisors let you run multiple OSes on one machine
- KVM and QEMU on Linux
- Xen, VMWare, HyperV on other platforms
Virtual Machines
- Guest operating system
- Almost native performance because CPU instructions are run directly
- Other hardware is emulated
VMs don’t have direct hardware access, which is a problem
- QEMU must emulate GPU, network, etc
- Paravirtualization helps with some of these
  - Instead of emulating, a guest driver passes some things through
  - Usually needs guest OS support
  - Still plenty of limites
- Dual booting is annoying

Gives a VM control over PCIE devices
Uses IOMMU (most devices in the past 10 years have this)
Remaps memory addresses and isolates devices
- Prevents device as PCIe directly to guest
- Low overhead
- Little host interference ### Capabilities Generally allows faster access and removes overhead and quirks from the host OS and passthrough
GPU Access
Storage/NVMe
Network devices (good for high throughput, bypasses host network stack)
USB and audio

Assigns an entire PCIe device to a guest
- Binds the device to vfio-pci on host side
- Unbinds from normal host driver
Host can’t use device anymore
Works for almost every device

Single-root I/O Virtualization
Shares PCIe deives between multiple guests
Host retains device access
More efficient sharing of resources
Used heavily in cloud computing
Virtual Function:
- Single PCIe devices splits into multiple VFs
- Can be assigned to guest
Only some hardware is capable
- Mostly enterprise grade devices

Can write to a device file to add VFs
Adds more PCI devices in lspci output
Intel GPU drivers are NOT READY for this, was not able to load a driver despite seeing the VF