KVM Memory and Device Virtualization

Coverage: EPT/NPT (stage-2 page table) → MMU virtualization → virtio (paravirtualization) → VFIO (passthrough) → vhost (kernel-space acceleration) → virtio-net/virtio-blk detailed explanation Kernel version: 2.6 ~ 6.x

Memory Virtualization: EPT/NPT

Three-level address translation:
  GVA (guest virtual address)
    → guest page table (managed by guest OS)
    → GPA (guest physical address)
      → EPT/NPT (managed by KVM, "stage-2 page table")
      → HPA (host physical address)

EPT (Intel Extended Page Tables) / NPT (AMD Nested Page Tables):
  Hardware directly executes the 2D page walk for GVA→GPA→HPA (24 steps!)
  → No need for KVM to maintain shadow page tables
  → Supported by hardware on modern CPUs

EPT violation: guest modifies its own page table → EPT page table becomes stale → VM exit
  → KVM updates EPT mapping → VMRESUME

MMU Virtualization

// arch/x86/kvm/mmu/mmu.c
// KVM MMU maintains two types of page tables:
//   1. EPT page table (used by hardware): GPN → HPN
//   2. Shadow page table (only for older CPUs without EPT): GVA → HPA

// TDP (Two-Dimensional Paging) = EPT on Intel, NPT on AMD

// kvm_mmu_page_fault():
//   guest accesses GPA → EPT violation → handle_ept_violation()
//     → kvm_mmu_map() → establishes GPA→HPA mapping in EPT
//     → returns to guest

virtio: Paravirtualized I/O

// drivers/virtio/
// Design principle: guest knows it is in a VM → communicates using shared memory (virtqueue)
//          No need to emulate real hardware → extremely high I/O performance

// virtio device model:
//   Frontend (guest): virtio-blk, virtio-net, virtio-scsi, virtio-gpu
//   Backend (host): 
//     QEMU (software emulation)
//     vhost (kernel-space acceleration: vhost-net, vhost-scsi, vhost-vsock)
//     vDPA (hardware acceleration: NICs/smartNICs)

virtqueue: Shared Ring Buffer

// Core transmission mechanism of virtio:
//   1. split virtqueue: legacy format (descriptor table + available ring + used ring)
//   2. packed virtqueue (1.1): more compact, cache-line friendly

// guest submits I/O:
//   write descriptor + available ring → kick (write doorbell)
// host consumes:
//   fetch descriptor → process I/O → write used ring → interrupt (MSI-X)
// guest completes:
//   interrupt → read used ring → free buffer

vhost: Kernel-Space Acceleration

// drivers/vhost/net.c
// vhost moves the virtio data path to the kernel:
//   guest → virtqueue → vhost (host kernel) → tap device → protocol stack
//   Bypasses QEMU user space → latency reduced by ~50%

// vhost-net:   networking (most commonly used)
// vhost-scsi:  storage
// vhost-vsock: guest↔host communication (AF_UNIX across VM boundaries)

VFIO: Device Passthrough

// Exclusively assign a physical PCIe device to a guest:
//   1. Unbind device from host (echo dev > /sys/bus/pci/drivers/xxx/unbind)
//   2. Bind to vfio-pci (echo vendor device > /sys/bus/pci/drivers/vfio-pci/new_id)
//   3. QEMU: -device vfio-pci,host=01:00.0
//   4. IOMMU isolation: device DMA can only access this VM's memory

// Performance: close to bare metal (GPU, NVMe passthrough to guest)
// Limitations: Cannot live migrate (device state cannot be saved/restored)

References

• Source Code: arch/x86/kvm/mmu/, drivers/virtio/, drivers/vhost/, drivers/vfio/
• Kernel Documentation: Documentation/virt/kvm/, Documentation/driver-api/virtio/
• LWN: "Virtio without the QEMU middleman (vhost)", "VFIO and device assignment"

Keywords: EPT, NPT, stage-2 page table, virtio, virtqueue, vhost, VFIO, PCI passthrough