Guest → VFIO → IOMMU → PCIe device
The IOMMU limits device DMA to mapped memory. Intel commonly calls the platform feature VT-d; AMD calls it AMD-Vi. CPU virtualization alone does not create this DMA boundary.
Inspect, do not assume
The group is the minimum isolation unit.
find /sys/kernel/iommu_groups/ -type l \
-printf '%h: ' -exec basename {} \;
lspci -nnk -s 01:00.0
dmesg | grep -Ei 'DMAR|IOMMU|AMD-Vi'Every function in a group must remain in a state that preserves isolation. Multifunction GPUs often bring an audio function; bridges can place several endpoints in one group.
Failure modes
Four reasons “it appears in the list” is not enough.
Reset
Some devices cannot return to a clean state after guest shutdown. The next VM start may fail until the device or host resets.
Host ownership
The host must release the active driver. Passing the only display GPU can remove local recovery access.
Topology
Group membership, bridges and peer-to-peer DMA determine whether assignment preserves isolation.
Lifecycle
Snapshots, suspend and live migration are generally unavailable for a directly assigned physical device.
Choose the device model deliberately
| Model | Isolation unit | Mobility | Typical use |
|---|---|---|---|
| Full PCI passthrough | Physical function / IOMMU group | Low | Dedicated GPU or NIC |
| SR-IOV | Virtual function | Provider-dependent | High-throughput networking |
| mdev / vGPU | Mediated device slice | Vendor-dependent | Shared accelerator |
| Emulated / virtio | VMM device model | Highest | Portable general workloads |
What can still affect the host?
Ordinary guest DMA should be confined by the IOMMU. Remaining risks include device firmware defects, broken reset behavior, bus errors, incorrect platform tables and IOMMU implementation or configuration flaws. “Guest driver failure crashes the host” is too broad; investigate the actual hardware path.