[7.3.2] VFIO/GPU Passthrough Failure Causes Other Running VM to Become Unresponsive

Unraid 7.3.2
Kernel: 6.18.38-Unraid
QEMU: 10.2.3
libvirt: 12.2
CPU: Intel i9-12900K
Motherboard: ASRock Z690 Taichi
GPU passthrough: 2x GTX 1080 Ti

I have been experiencing repeated VM instability since upgrading into the 7.3.x branch. This system previously ran these Windows VMs/passthrough configuration for years without this type of issue.

Most recent failure:

• Three Windows VMs showed as running, but Game_SVR and VideoSVR_igpu were inaccessible.

• QMP for both VMs still reported running: true.

• Guest agent on Game_SVR and VideoSVR timed out. Util_SVR continued responding.

• Game_SVR uses GTX 1080 Ti at 02:00.0/02:00.1.

• The host logged thousands of DMAR/IOMMU faults from 02:00.0, including:
  PTE Read access is not set
PTE Write access is not set
non-zero reserved fields in PTE

• Game_SVR's QEMU process was using approximately 400% CPU while the host still had substantial idle CPU and available RAM.

The most significant test:

I force-stopped ONLY Game_SVR.

Immediately afterward:

• VFIO reset 02:00.0 and 02:00.1 successfully.

• VideoSVR_igpu's guest agent immediately began responding again.

• VideoSVR became accessible again through Parsec without being restarted.

• Blue Iris live camera feeds immediately resumed.

• Util_SVR remained healthy.

• No further DMAR faults appeared.

VideoSVR was never reset or restarted. Removing the malfunctioning Game_SVR/VFIO device restored the unrelated running VM.

I observed similar cross-VM behavior previously on 7.3.2 where resetting one VM caused another frozen VM to recover.

This does not appear to be host CPU, RAM, or storage exhaustion. ZFS pools were healthy and the host remained responsive.

I am attaching:

1. Diagnostics captured while the failure was active.

2. Diagnostics from the same boot immediately after stopping Game_SVR and recovering the other VMs.

There are other reports of GPU passthrough/VM regressions in the 7.3.x branch, but I have not found another report with this exact cross-VM/IOMMU behavior.

At this point I suspect a regression somewhere in the 7.3.2 kernel/KVM/VFIO/QEMU stack. I would appreciate guidance on any additional traces or diagnostics that would be useful if/when this reproduces again.

blackbox-diagnostics-20260910-2258-RECOVERED-after-Game-destroy.zip blackbox-diagnostics-20260910-1626-FROZEN-Game-running.zip