Hello everyone,
I am experiencing a VFIO PCI passthrough issue on my Beelink ME Pro, which is based on the AMD H255 platform with a Ryzen 7 255 CPU.
My goal is to pass through two ASMedia SATA controllers directly to a virtual machine running fnOS (a NAS operating system) under Proxmox VE.
I would appreciate some help determining whether this is a BIOS/AGESA firmware issue or a Linux AMD IOMMU driver issue.
- Hardware and software environment
Mini PC: Beelink ME Pro
Platform: AMD H255
CPU: AMD Ryzen 7 255
Hypervisor: Proxmox VE 9.2.11
Proxmox package version: proxmox-ve 9.2.0
Kernel initially tested: 7.0.2-6-pve
Kernel also tested: 7.0.14-26-pve
QEMU: 11.0.3-2
IOMMU: AMD-Vi enabled
Kernel command line:
amd_iommu=on iommu=pt
The two SATA controllers are:
Controller 1:
PCI address: 0000:03:00.0
Device ID: 1b21:1062
Subsystem ID: 1b21:2116
IOMMU Group: 19
Controller 2:
PCI address: 0000:05:00.0
Device ID: 1b21:1062
Subsystem ID: 1b21:2116
IOMMU Group: 21
Both are identified as ASMedia SATA controllers.
- The actual problem
Both SATA controllers can be successfully bound to vfio-pci.
However, when I start the VM with the controller at 0000:03:00.0 passed through, QEMU fails to initialize the VFIO container.
The exact error is:
kvm: -device vfio-pci,host=0000:03:00.0,id=hostpci2,bus=ich9-pcie-port-3,addr=0×0:
vfio 0000:03:00.0: failed to setup container for group 19:
Failed to set group container: Invalid argument
TASK ERROR: start failed: QEMU exited with code 1
The kernel log reports:
vfio-pci 0000:03:00.0: resetting
vfio-pci 0000:03:00.0: reset done
vfio-pci 0000:05:00.0: resetting
vfio-pci 0000:05:00.0: reset done
vfio-pci 0000:03:00.0: Firmware has requested this device have a 1:1 IOMMU mapping, rejecting configuring the device without a 1:1 mapping. Contact your platform vendor.
vfio-pci 0000:03:00.0: Firmware has requested this device have a 1:1 IOMMU mapping, rejecting configuring the device without a 1:1 mapping. Contact your platform vendor.
The important detail is that the error specifically mentions a firmware-requested 1:1 IOMMU mapping.
- IOMMU group isolation is not the apparent problem
I checked the IOMMU group membership:
find /sys/kernel/iommu_groups/19/devices/ -maxdepth 1 -type l -printf ‘%f\n’
The output contains only:
0000:03:00.0
Therefore, Group 19 contains only the ASMedia SATA controller. There are no other PCI devices in this group that could explain the failure through ordinary IOMMU group isolation restrictions.
The controller is correctly bound to vfio-pci:
SATA controller [0106]: ASMedia Technology Inc. Device [1b21:1062] (rev 02)
Subsystem: ASMedia Technology Inc. Device [1b21:2116]
Kernel driver in use: vfio-pci
Kernel modules: ahci
- Reserved IOMMU regions
I inspected the reserved regions exposed by the AMD IOMMU driver:
cat /sys/kernel/iommu_groups/19/reserved_regions
The output is:
0×0000000097fed000 0×0000000097feefff direct
0×00000000981f2000 0×00000000981f3fff direct
0×00000000fee00000 0×00000000feefffff msi
0×000000fd00000000 0×000000ffffffffff reserved
Group 21 reports exactly the same reserved regions.
The two direct regions are particularly interesting:
0×97fed000 - 0×97feefff
0×981f2000 - 0×981f3fff
These regions appear to correspond to the exclusion ranges defined in the system’s ACPI IVRS table.
- ACPI IVRS / IVMD investigation
I dumped and disassembled the ACPI IVRS table using ACPICA tools:
acpidump -n IVRS -b
iasl -d ivrs.dat
The IVRS table is 536 bytes long and contains:
- Two legacy IVHD entries (types 0×10 and 0×11).
- One extended/mixed-format IVHD entry (type 0×40).
- Four IVMD entries.
The four IVMD entries have the following properties:
Flags: 0×08
Unity: 0
Readable: 0
Writeable: 0
Exclusion Range: 1
Their address ranges are:
0×981F3000, length 1
0×981F2000, length 1
0×97FEE000, length 1
0×97FED000, length 1
The entries are marked as Exclusion Range, not Unity Mapping.
After inspecting the Linux AMD IOMMU driver behavior, I suspect that these IVMD exclusion entries may be converted into read/write unity-mapped regions, which are subsequently exposed to VFIO as IOMMU_RESV_DIRECT.
This would explain why VFIO refuses to configure the device without a 1:1 mapping.
I have not conclusively established why the firmware provides these particular exclusion ranges or why they cause this problem for the ASMedia controller.
- Kernel upgrade did not solve the issue
I initially encountered the problem on kernel 7.0.2-6-pve.
I then upgraded to kernel 7.0.14-26-pve.
The problem remains exactly the same.
After upgrading, the reserved regions still contain the same direct entries, and the kernel still reports:
Firmware has requested this device have a 1:1 IOMMU mapping, rejecting configuring the device without a 1:1 mapping.
Therefore, simply upgrading to the newer available PVE kernel did not resolve the issue.
What I have already checked
AMD IOMMU is enabled.
iommu=pt is enabled.
AMD-Vi interrupt remapping is enabled.
Both SATA controllers are bound to vfio-pci.
The affected IOMMU group contains only the SATA controller.
The controller reset operation completes successfully.
The failure persists across two PVE kernel versions.
The affected IOMMU groups report the same direct reserved regions.
Questions for the community
Is this likely to be a BIOS/AGESA firmware issue on the Beelink ME Pro AMD H255 platform?
Why does the firmware define these IVMD entries with IVMD_FLAG_EXCL_RANGE (flags = 0×08), and why do they end up being exposed as IOMMU_RESV_DIRECT regions?
Is the Linux AMD IOMMU driver incorrectly interpreting these exclusion entries as read/write unity mappings?
Is there a safe workaround that allows VFIO to pass through the ASMedia SATA controller without requiring an identity mapping for the entire device?
Has anyone encountered a similar issue on an AMD Ryzen 7 255 / H255 system or another recent AMD platform?
Would a BIOS/AGESA update be necessary, or is this primarily a Linux AMD IOMMU driver issue?
Potential upstream Linux kernel fix
While investigating this issue, I found an AMD IOMMU patch series by Suravee Suthikulpanit:
[PATCH v2 0/4] AMD IOMMU: IVMD handling cleanup and boot-time override
Posted on April 24, 2026:
https://lists.openwall.net/linux-kernel/2026/04/24/576
The series contains four patches:
- Remove unused global exclusion range fields and initialization.
- Stop converting IVMD exclusion ranges into unity-mapped ranges.
- Clean up IVMD handling and derive reserved regions directly from the original IVMD flags.
- Introduce an ivmd= kernel boot option to supplement firmware-provided IVMD entries.
Patch 2 is particularly relevant because it removes the special case that forces an exclusion range to become a read/write unity mapping.
Patch 3 changes the reserved-region handling so that an entry with IVMD_FLAG_EXCL is exposed as IOMMU_RESV_RESERVED rather than IOMMU_RESV_DIRECT.
This appears closely related to the behavior observed on my system.
I have not yet applied these patches or built a custom kernel, so I cannot confirm whether they resolve the issue.
Could someone familiar with the AMD IOMMU subsystem confirm whether this patch series is intended to address this exact VFIO failure, and whether a newer revision has already been merged into the IOMMU maintainer tree or upstream Linux?
Thank you for your help.