Lenovo ThinkStation P360 Ultra - VM GPU Passthrough deadlock on shutdown/stop (vfio_container_dma_map = -22 / BARs not supported)

Hi everyone,

I am experiencing a frustrating system deadlock (WebUI freeze, Error 504) whenever I attempt to gracefully shut down or stop a Windows 11 VM with a dedicated GPU passed through. The only way to keep Unraid stable is by using "Force Stop", but then the GPU stays bound to VFIO until manually re-attached.

Here are my exact system specifications and hardware topology gathered via lshw and Unraid System Devices:

• Host Computer: Lenovo ThinkStation P360 Ultra (Desktop Computer / Motherboard SKU: LENOVO_MT_30G0)

• BIOS Version: S0FKT34A (Date: 04/29/2025) with BIOS display set to "Auto"

• CPU: 12th Gen Intel Core i7-12700 (Alder Lake-S)

• RAM: 64 GB DDR5 (2x34GB SK Hynix SODIMM 5600 MHz)

• Primary/Passed-through GPU: NVIDIA Corporation TU117GL [T400 4GB / T400E] (rev a1) – Note: The physical monitor cable is plugged into this card.

• Integrated GPU: Intel Corporation Alder Lake-S GT1 [UHD Graphics 770]

• Additional PCIe Hardware: Intel Corporation I350 Gigabit Network Connection (4-port server card installed in the second slot via riser, currently idle/no cables connected).

• Storage: 2x 1TB Samsung PM9A1 NVMe SSDs + external TerraMaster D5-300C 5-bay USB enclosure via Thunderbolt/USB 3.2.

The Problem & Logs

When the VM boots, everything works perfectly. However, during a normal Windows shutdown or an ACPI "Stop" from the Unraid panel, the host logs immediately spam the following lines before the entire WebUI locks up:

char device redirected to /dev/pts/1 (label charserial0)

qemu-system-x86_64: warning: vfio_container_dma_map(...) = -22 (Invalid argument)

0000:70:00.0: PCI peer-to-peer transactions on BARs are not supported.

What I have already tried (and failed):

1. BIOS Tweaks: Disabled ASPM (Active State Power Management) completely to avoid PCIe power-state issues.

2. Unraid VM Settings (Startup): Enabled VFIO Allow Unsafe Interrupts and turned On all framebuffer blocks (EFI Framebuffer, Simple Framebuffer, Block All Framebuffers).

3. VM XML Patch: Modified the CPU section to include the <maxphysaddr mode='passthrough' limit='39'/> parameter, which is known to solve DMA mapping issues on 12th/13th Gen Intel CPUs. After rebuilding the NVRAM template, the VM started cleanly, but the host still locked up upon standard shutdown.

4. PCIe Conflict Suspect: I suspect a hardware limitation or interrupt conflict on the Lenovo motherboard/riser card between the NVIDIA T400 and the multi-port Intel I350 network card, as they share the CPU's primary PCIe lanes.

Since "Force Stop" works without crashing the host, I am currently using a custom User Script to manually re-attach the GPU back to the unRAID core/dockers after forcing the VM down. However, I would love to find a native fix for proper ACPI power-downs.

Has anyone encountered this specific behavior on Lenovo mini PCs or stasions with Intel 12th Gen CPUs sharing lanes with an un-utilized server NIC? Any advice on how to bypass this kernel/VFIO DMA mapping restriction would be highly appreciated!

Thank you in advance!