QEMU-KVM, GPU Passthrough, Windows 10 - Crashing

I've been happily running a couple of Windows 10 VMs on QEMU-KVM with GPU Passthrough for the last few years with little issue. A few days ago however (21 May 2020) both my VMs started crashing on start-up.

If I boot Windows in safe mode - it works. If I boot with the Windows GPU drivers disabled - it works. As soon as install or enable the driver it crashes (blank screen, VM uses 100% CPU, can't toggle num lock on keyboard) and after a few seconds the VM reboots.

Anyone had similar issues or know how I can further troubleshoot?

Hardware

• MB: GA-X99P-SLI
• CPU: i7-6850K
• RAM: 32GB
• Host GPU: Nvidia Geforce 9400 GT
• Guest GPUs x2: Nvidia GTX GeForce 1060 6GB
• Disk: Samsung 950 PRO NVMe 512GB

Software

• Ubuntu Server 18.04/20.04
• qemu-kvm 2.11/4.2
• Windows 10 x64 1909

Things I've Tried

• Although unlikely, as both VMs started crashing at the same time, I rolled back several Windows updates
• Tried old drivers, windows update drivers, and the latest Nvidia drivers
• This coincided with an apt upgrade so I tried rolling back but I couldn't (easily) as I'd cleaned my cache (d'oh!!) and the packages weren't available
• There's no errors or warnings in Windows events
• There's no errors or warnings in /var/log/libvirt/qemu
• All GPUs run fine together on bare metal
• Rebuilt everything with latest LTS version of Ubuntu (20.04) and related qemu-kvm (4.2)
• Switched VMs from i440 to Q35

2 Answers

OK, so thanks to this post I've managed to fix the issue, at least in a test environment:

The steps I took were:

1. Set <kvm><hidden state='off'/></kvm> in my VM config - this disabled the GPU drivers so I could boot into windows
2. Set the MSISupported registry key to "1" for my GPU
3. Re-set <kvm><hidden state='on'/></kvm> in my VM config

--

I'll rebuild my VMs properly next weekend (I've been running one Win 10 instance on bare metal) and, all things going well, accept this as the answer.

--

Edit: I rebuilt my VMs and the above initially fixed the issue. I run multiple different GPUs across my VMs however and, every time I changed GPU, it reset the registry key. I ended up replacing all my GPUs with AMDs which worked better (still have a reset bug, but at least they worked).

Adding this fixes the problem

 <hyperv> <vendor_id state="on" value="whatever"/> </hyperv>

Had the same problem.