PCIe Device Passthrough: NIC Name Instability and MAC Pinning

My Proxmox node rebooted, and suddenly the host was unreachable via SSH. I had to plug in a physical monitor and keyboard only to find that my primary network interface, which had been enp4s0 for months, had decided to rename itself to enp5s0.

Because my /etc/network/interfaces file was explicitly tied to enp4s0, the bridge didn't come up, the IP wasn't assigned, and I was locked out of my own hardware.

What I expected

I expected the Linux kernel to consistently enumerate my PCIe devices. In a static hardware environment where nothing has moved, the PCI bus address should be deterministic. If the NIC is plugged into the same slot and the BIOS hasn't changed, enp4s0 should stay enp4s0 forever. This is the "happy path" most documentation assumes.

What actually happened

The reality is that PCIe enumeration is not always a constant. I'm using a mix of onboard NICs and a PCIe expansion card. I also have a GPU passed through to a VM.

The surprise here is how the kernel's predictable network interface naming (systemd-udevd) interacts with the PCIe topology. When I added a new PCIe device and tweaked some BIOS settings for IOMMU, the way the kernel mapped the physical slots to the virtual naming changed. A slight shift in how the PCIe switch reported the devices caused the index to jump.

This isn't just a "one-time fluke." If you're running a multi-node cluster or using GPUs that might move addresses (something I've documented before in GPU PCI Address Instability), you'll find that the kernel is surprisingly flexible with where it puts things.

The root cause is that enp4s0 is a name derived from the PCI location. If the location changes—even by one digit—the name changes. If your network config depends on that name, your system is one reboot away from a blackout.

The Fix: MAC Pinning

The only way to stop this is to stop relying on the PCI slot location and start relying on the hardware's unique identifier: the MAC address.

I decided to use systemd .link files. This allows me to tell the kernel: "I don't care where this device is on the PCIe bus; if it has this MAC address, call it eth0."

1. Identify the MAC address

First, I had to find the actual MAC of the problematic NIC while I had console access.

ip link show

I looked for the interface that was currently named enp5s0 (the "wrong" name) and copied the link/ether value.

2. Create the .link file

I created a custom link file in /etc/systemd/network/. I chose the name 10-lan.link to ensure it loads early in the boot process.

# /etc/systemd/network/10-lan.link
[Match]
MACAddress=00:11:22:33:44:55

[Link]
Name=eth0

(Note: I've anonymized the MAC address above. Use your actual hardware MAC here.)

3. Update the network configuration

Once the interface is pinned to eth0, I had to update the Proxmox network configuration to match. I edited /etc/network/interfaces to replace the volatile enp4s0 with the stable eth0.

# Example snippet from /etc/network/interfaces
auto eth0
iface eth0 inet manual

auto vmbr0
iface vmbr0 inet static
    address 10.0.0.x/24
    gateway 10.0.0.1
    bridge-ports eth0
    bridge-stp off
    bridge-fd 0

4. Apply and verify

I ran systemd-networkd-restart (or just rebooted, since I was already at the console) and verified the name with ip a. The NIC was now consistently eth0, regardless of whether the PCIe bus shifted.

Why this matters

If you're just running a single VM on a desktop, this is a minor annoyance. But if you're building a production-grade homelab, this is a critical failure point.

You'll hit this specifically in these scenarios:

• Adding/Removing PCIe Hardware: Adding a new NVMe drive or a GPU can shift the enumeration of other devices on the same root complex.
• BIOS Updates: A BIOS update often resets PCIe lane bifurcation or IOMMU settings, which can completely reorder how the kernel sees your NICs.
• Using PCIe Switches: Some high-end motherboards or riser cables use PCIe switches that can report different topologies depending on the power state of the devices.

The Tradeoff

The tradeoff here is that you're moving away from the "modern" predictable naming convention back to the "old" ethX style. Some people find eth0 ugly or outdated, but in a headless server environment, "ugly" is better than "unreachable."

I've also seen people try to fix this using udev rules in /etc/udev/rules.d/. While that works, .link files are the native systemd way to handle this and are generally cleaner to maintain.

Lessons Learned

The biggest lesson here is that documentation for Proxmox and Debian assumes your hardware topology is a constant. It isn't.

When you're doing complex things like PCIe passthrough—which I've detailed in my GPU Passthrough Gotcha Guide—you are intentionally messing with the PCI bus. You're telling the host kernel to ignore certain devices so the VM can claim them. This volatility is a side effect of that power.

If you are passing through NICs or GPUs, do not trust the default interface names. Pin your critical management interfaces to their MAC addresses immediately. It takes five minutes to set up and saves you from a midnight trip to the server rack because a reboot decided your network card now lives at enp6s0.

For those of you managing larger fleets or complex AI agent infrastructure, this kind of hardware-level stability is the foundation. You can't build a reliable multi-agent AI pipeline if the underlying Kubernetes worker nodes are randomly losing their network identity.

Next time you're configuring a new node, don't just copy the enpXsX name from the GUI. Take the extra step to pin it. Your future self will thank you when the next BIOS update doesn't break your entire cluster.