[Kernel Bug] AMD IVMD exclusion ranges incorrectly mapped as IOMMU_RESV_DIRECT, breaking VFIO passthrough (7.0.14-30-pve)

user7500

New Member
Jul 16, 2025
2
0
1
Hi Proxmox team,

I'm reporting a reproducible AMD IOMMU/VFIO regression affecting GPU passthrough on Proxmox VE 9.

GPU passthrough was previously working with an AMD Ryzen 7 7700X and MSI BIOS 7E12v1J. The issue appeared after simultaneously upgrading to a Ryzen 9 9950X and BIOS 7E12v1L3. Since both changes were made together, the exact trigger cannot be isolated. A change in the firmware-generated ACPI IVRS table is suspected.

The issue persists across multiple official Proxmox kernels, including 7.0.14-30-pve.

Hardware:

CPU: AMD Ryzen 9 9950X (previously Ryzen 7 7700X)

Motherboard: MSI MAG X670E TOMAHAWK WIFI

BIOS: 7E12v1L3 (previously 7E12v1J)

GPU: NVIDIA RTX 3060 (10de:2504)

GPU audio: 10de:228e

Secure Boot enabled, AMD IOMMU active

Both GPU functions correctly bound to vfio-pci

1. Exact error

Starting the VM produces:

kvm: -device vfio-pci,host=0000:01:00.0,id=hostpci0.0,bus=ich9-pcie-port-1,addr=0x0.0,multifunction=on: vfio 0000:01:00.0: failed to setup container for group 12: Failed to set group container: Invalid argument

start failed: QEMU exited with code 1

Kernel log:

vfio-pci 0000:01:00.0: Firmware has requested this device have a 1:1 IOMMU mapping, rejecting configuring the device without a 1:1 mapping. Contact your platform vendor.

2. IOMMU reserved regions

Command:

cat /sys/kernel/iommu_groups/12/reserved_regions

Output:

0x0000000085a18000 0x0000000085a19fff direct

0x0000000097f06000 0x0000000097f06fff direct

0x00000000fee00000 0x00000000feefffff msi

0x000000fd00000000 0x000000ffffffffff reserved

Analysis of the firmware ACPI IVRS table identifies three IVMD entries (type=0x22, flags=0x08):

IVMD base

Length

0x85a18000

0x1000

0x85a19000

0x1000

0x97f06000

0x1000

These correspond exactly to the two direct regions exposed by the kernel.

The suspected cause is the AMD IOMMU driver's handling of IVMD exclusion ranges, converting them into read/write unity mappings (IOMMU_RESV_DIRECT) instead of treating them as reserved/unmapped regions (IOMMU_RESV_RESERVED).

3. AMD upstream fix

AMD has published a patch series addressing this specific behavior:

[PATCH v4 0/3] iommu/amd: IVMD handling cleanup and ivmd= boot override (November 17, 2026)

https://lore-kernel.gnuweeb.org/linux-iommu/20261117070654.3699-1-sarunkod@amd.com/

Most relevant patch:

[PATCH v4 2/3] iommu/amd: Do not convert IVMD exclusion range to unity range

https://lkml.iu.edu/hypermail/linux/kernel/2611.2/04204.html

This appears to address the IVMD classification problem observed here without disabling VFIO security checks or compromising IOMMU isolation.

A similar issue and a successful test of an earlier AMD patch revision have also been reported:

https://bugzilla.redhat.com/show_bug.cgi?id=2559096

4. Tested Proxmox kernels

Kernel -
Result

6.17.13-21-pve -
Affected

7.0.14-20-pve -
Affected

7.0.14-26-pve -
Affected

7.0.14-30-pve -
Affected (November 20, 2026)

The latest kernel was installed from the official Proxmox repository using the signed kernel package. The issue remains unchanged.

5. Request

Could the Proxmox kernel maintainers please review AMD's November 17 IVMD patch series and consider backporting the relevant fixes into an official Proxmox VE 9 kernel?

The goal is to restore VFIO passthrough while preserving Secure Boot, DMA remapping, and IOMMU isolation, without requiring users to maintain custom kernels.

I'm happy to provide the original ACPI IVRS binary or additional diagnostics upon request, and can test a patched Proxmox kernel build.

Thanks!
 
Last edited: