Comment # 13 on bug 1188745 from Joerg Roedel

I looked through the failure log file, but this didn't give any insights.

What I found so far is that the device sends an interrupt request before any
IRQs have been configured by the device driver. The IOMMU is configured to
still block all interrupt requests from the device.

This is the only situation where this can happen (in fact, when IV=1b and
IntCtl=00b in the DTE for the device)

As there are no AMD IOMMU changes in the range, my guess is that some other
change between 5.14.2 and 5.14.4 influenced the behaviour of the Nvidia module,
so that it now sends an IRQ when it really shouldn't.

There are two options on how to proceed:

* Bisect upstream stable kernels between 5.13.2 and 5.13.4 to find the commit
which introduced the issue

* Work around the problem by booting with 'intremap=off' on the kernel command
line.

Please let me know what you prefer.