// LAB LOG  ·  2026-07-14

NVIDIA RTX PRO 5000 "Blackwell" GPU Passthrough on Proxmox VE 9

Three bugs, three fixes. Getting the kvm: segfault at 40, the guest NVRM: BAR0 is 0M, and RmInitAdapter failed (0x22:0x56:897) out of the way — all on one card, in sequence.

TL;DR

Passing an NVIDIA RTX PRO 5000 Blackwell (48 GB, PCI 10de:2bb3) through to a VM on Proxmox VE 9 failed three times in a row, each with a different root cause. If you are hitting any of these on a Blackwell card (RTX PRO 5000/6000, RTX 50-series), you probably need all three fixes:

1. QEMU segfaults on VM start (kvm: segfault at 40, NULL deref right after "reset done") → QEMU wrongly applies the legacy GeForce BAR0 quirk to a professional card. Fix: x-no-geforce-quirks=on on the vfio-pci device.
2. Guest boots but the GPU is dead (NVRM: BAR0 is 0M @ 0x0, nvidia-smi: No devices were found) → the card's 64 MB 32-bit BAR0 can't be assigned because the eGPU/OCuLink PCIe root port only exposes a 1 MB 32-bit window. Fix: pci=realloc on the host kernel cmdline.
3. Driver loads but won't init (NVRM: requires use of the NVIDIA open kernel modules) → Blackwell requires the open kernel modules, not the proprietary ones. Fix: install the -open DKMS variant.

Environment

The exact same config worked fine with an RTX 4080 in the same slot. Everything below is Blackwell-specific.

Bug 1 — QEMU segfaults on VM start (kvm: segfault at 40)

Symptom

# qm start 120
kvm: ... segfault at 40 ip ... sp ... error 4 in qemu-system-x86_64

The host dmesg shows a NULL-pointer dereference in the KVM/QEMU process, occurring immediately after the VFIO reset ("reset done"). Reproduces identically across pve-qemu-kvm 10.0.2, 10.2.1 and 11.0.2, and on kernel 6.17.9. The card itself binds and resets cleanly at the kernel/vfio-pci level — this is purely a QEMU crash.

Dead ends (so you don't waste time)

• The warning VFIO dma-buf not supported in kernel: PCI BAR IOMMU mappings may fail is a red herring — the segfault happens with or without it.
• BAR1 size is not the cause. The card exposes a 64 GB prefetchable BAR1 (Resizable BAR). Shrinking it to 256 MB (echo 8 > /sys/bus/pci/devices/0000:c7:00.0/resource1_resize, after unbinding from vfio-pci) did not stop the crash.
• The szoran kernel patch for Firmware has requested this device have a 1:1 IOMMU mapping, rejecting… does not apply here — that dmesg rejection never appears, and our cmdline has no iommu=pt.

Root cause

A gdb backtrace on a minimal QEMU reproducer pins the crash to:

vfio_probe_nvidia_bar0_quirk   (hw/vfio/pci-quirks.c:928)
  -> memory_region_add_subregion  with a NULL `mirror` MemoryRegion

QEMU applies its legacy GeForce BAR0 quirk to the card and dereferences NULL. The quirk is meant for consumer GeForce cards; on a professional Blackwell card the code path hits a NULL region instead of degrading gracefully.

Fix

Disable the GeForce quirks on the vfio-pci device. The QEMU property description literally reads "for NVIDIA Quadro/GRID/Tesla" — i.e. professional cards:

-device vfio-pci,host=0000:c7:00.0,...,x-no-geforce-quirks=on

The minimal reproducer exits 0 (instead of 139) with this flag, and the VM starts for the first time.

Making it persistent in Proxmox — Proxmox builds the QEMU command line itself, so patch the generator. In /usr/share/perl5/PVE/QemuServer/PCI.pm, inside the primary-function block (if ($j == 0), around line 715), append the property:

$devicestr .= ',x-no-geforce-quirks=on';

Back up the file first. Verify it took effect:

qm showcmd <vmid> | grep geforce

⚠️ This patch does not survive an apt upgrade of qemu-server. Re-apply it after any Proxmox package update that touches PCI.pm. (A cleaner long-term option is a systemd hook, or waiting for a Proxmox/QEMU release that guards this quirk for professional cards.)

Bug 2 — Guest sees the GPU but it's dead (NVRM: BAR0 is 0M)

Symptom

VM now boots. In the guest the card is visible (lspci shows 01:00.0 10de:2bb3) and the 580 driver loads, but:

NVRM: GPU 0000:01:00.0: BAR0 is 0M @ 0x0 (PCI:0000:01:00.0)
NVRM: The system BIOS may have misconfigured your GPU.
nvidia-smi: No devices were found

On the host, the boot log reveals the real problem:

pci 0000:c7:00.0: BAR 0 [mem size 0x04000000]: can't assign; no space
pci 0000:c7:00.0: BAR 0 [mem size 0x04000000]: failed to assign
...
pci 0000:03.1: bridge window [mem 0xdd200000-0xdd2fffff 64bit] ... [size=1M]

Root cause

BAR0 is 64 MB, 32-bit non-prefetchable, so it must live below 4 GB. The PCIe root port the eGPU/OCuLink hangs off of (00:03.1 here) only advertises a 1 MB 32-bit memory window. 64 MB doesn't fit → BAR0 gets size 0 → vfio-pci can't pass it through → the guest sees no Region 0 → BAR0 is 0M.

Hotplug remove + rescan does not help (the window isn't enlarged without reallocation). BAR1/ReBAR resizing is irrelevant here — BAR0 is a separate, 32-bit BAR.

Fix

Let the kernel reallocate PCI bridge windows so the 1 MB window can grow to ≥64 MB. Add pci=realloc to the host kernel cmdline in /etc/default/grub:

GRUB_CMDLINE_LINUX_DEFAULT="quiet amd_iommu=on pci=realloc"

update-grub && reboot

Verify after reboot:

cat /proc/cmdline | grep pci=realloc
dmesg | grep 'c7:00.0.*BAR 0'      # must now say "assigned", e.g. 0x80000000-0x83ffffff
cat /sys/bus/pci/devices/0000:c7:00.0/resource | head -1   # no longer 0x0 0x0 0x0

If pci=realloc alone isn't enough, try adding pci=nocrs (ignores ACPI _CRS window limits — slightly riskier), and check Above 4G Decoding / Resizable BAR in the host BIOS.

Bug 3 — Driver loads but won't initialize (open kernel modules required)

Symptom

BAR0 is now assigned and passed through, but the guest still fails:

NVRM: The NVIDIA GPU 0000:01:00.0 (PCI ID: 10de:2bb3)
NVRM: installed in this system requires use of the NVIDIA open kernel modules.
NVRM: GPU 0000:01:00.0: RmInitAdapter failed! (0x22:0x56:897)
NVRM: GPU 0000:01:00.0: rm_init_adapter failed, device minor number 0

Root cause

Blackwell (and newer) GPUs are only supported by the NVIDIA open kernel modules. The proprietary/closed kernel module (license: NVIDIA) will load but cannot initialize the card. You need the -open build (license: Dual MIT/GPL).

Fix (Ubuntu 24.04, packaged driver)

Swap the closed DKMS module for the open one. This removes the closed kernel packages but keeps the 580 userspace libraries:

apt install nvidia-dkms-580-server-open

Then reload the stack without rebooting the VM (nothing is using the GPU yet, since it never initialized):

systemctl stop nvidia-persistenced
rmmod nvidia_drm nvidia_uvm nvidia_modeset nvidia
modprobe nvidia
systemctl start nvidia-persistenced

Verify the open module is loaded:

modinfo nvidia | grep license   # => license: Dual MIT/GPL  (open)  — NOT "NVIDIA" (closed)
nvidia-smi                      # => NVIDIA RTX PRO 5000 Blackwell, 48 GB

If you install the driver via the .run installer instead, use --kernel-module-type=open. Make sure future DKMS rebuilds / driver upgrades keep the open flavor.

Final verification

$ nvidia-smi
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 580.159.03     Driver Version: 580.159.03     CUDA Version: 13.0  |
|  0  NVIDIA RTX PRO 5000 Blac... | 00000000:01:00.0 Off | 0MiB / 48935MiB | ..|
+-----------------------------------------------------------------------------+

Order of operations

1. Patch PCI.pm with x-no-geforce-quirks=on → VM starts without segfault.
2. Add pci=realloc to GRUB, update-grub, reboot host → BAR0 gets assigned.
3. In the guest: install nvidia-dkms-<ver>-server-open, reload modules → nvidia-smi works.

Bugs 1 and 2 are host-side; bug 3 is guest-side. Bug 2 requires a host reboot; the others don't.

Notes / caveats

• The PCI.pm patch is fragile across Proxmox package upgrades — re-check qm showcmd after updates.
• pre-enrolled-keys=0 (Secure Boot off in the VM's OVMF) avoids the separate MOK-signing rabbit hole for the DKMS module.
• Tested with pve-qemu-kvm 11.0.2-1 / kernel 6.17.9-1-pve. Newer Proxmox/QEMU may guard the GeForce quirk upstream, making fix #1 unnecessary — check before patching.

References

• Proxmox forum thread #181195 (RTX 5080 Blackwell, PVE9 / kernel 6.17)
• github.com/szoran53/proxmox-kernel-6.17-gpu-passthrough (a different Blackwell issue — the 1:1 IOMMU reject — not the one fixed here)
• NVIDIA Developer Forum #339038 (PRO 6000 / 5090 host crash reports)
• QEMU source: hw/vfio/pci-quirks.c — vfio_probe_nvidia_bar0_quirk

Written up after getting an RTX PRO 5000 Blackwell fully operational under Proxmox VE 9 in an OCuLink eGPU setup. Shared in case it saves someone else the three-bug detour. — Sunwalkers Studios