The cable was seated. The settings panel had nothing to configure. The NIC was sitting on the PCI bus the whole time, and the driver had already refused it.

The Situation

We wanted a Quest 3 to see a Linux desktop the way Virtual Desktop used to. WiVRn was the runtime. The line that stopped the afternoon was “ethernet is plugged in, but there is no way to set it or use it.” That sounds like a missing toggle. It was not.

There Was Nothing to Configure

ip -br link listed two interfaces: lo and the Wi-Fi device. No eno, no eth, no USB dongle pretending to be one. NetworkManager agreed. Its device list was Wi-Fi, loopback, and a disconnected peer-to-peer Wi-Fi device. You cannot set an address on a port the kernel has not created.

The hardware was not missing. lspci had it:

00:1f.6 Ethernet controller: Intel Corporation Ethernet Connection (17) I219-V (rev 11)
        Kernel modules: e1000e

Onboard, Intel, the usual laptop-and-board chip. The e1000e module was already loaded. Its use count was 0. /sys/bus/pci/devices/0000:00:1f.6/driver did not exist. The module knew the device ID and had declined to bind.

The kernel log, from boot, was two lines:

e1000e 0000:00:1f.6: The NVM Checksum Is Not Valid
e1000e 0000:00:1f.6: probe with driver e1000e failed with error -5

Error -5 is -EIO. Probe failed, so no netdev, so no settings row, so “no way to set it.” The cable never got a vote.

This was an MSI H610 board. Same symptom on a lot of OEM I219-V parts: the device enumerates, e1000e prints the checksum line, and then it leaves.

The Driver That Refuses to Bind

The check lives in the probe path, not in NetworkManager. In Linux 7.2.8 the loop in drivers/net/ethernet/intel/e1000e/netdev.c tries the checksum three times, then gives up:

for (i = 0;; i++) {
    if (e1000_validate_nvm_checksum(&adapter->hw) >= 0)
        break;
    if (i == 2) {
        dev_err(&pdev->dev, "The NVM Checksum Is Not Valid\n");
        err = -EIO;
        goto err_eeprom;
    }
}

goto err_eeprom is the whole failure from the outside. Probe unwinds. The PCI device stays unbound. ip link cannot invent an interface the driver refused to register.

The checksum is the old Intel rule: add the NVM words, expect 0xBABA. OEM images often do not sum that way. Windows ignores it. Linux does not.

There is a second, nastier detail in ich8lan.c. On older PCH parts, a clear “checksum valid” bit makes the driver repair the NVM: set the bit, recompute, write it back. On Tiger Lake and everything after, that write is how you lock the chip up. Kernel commit 4051f68318ca (“e1000e: Do not take care about recovery NVM checksum”, merged for 5.14) stopped the repair on platforms where the NVM is effectively read-only. The commit message is blunt. Attempting to update the NVM causes a lockup. Emit an error instead.

What that commit did not do is let the device work. On Alder Lake and newer (e1000_pch_adp, which is this I219-V), the valid-bit path skips the write and falls through to e1000e_validate_nvm_checksum_generic(). The generic sum fails. Probe still returns -EIO. You get the safe half of the fix and none of the useful half.

“Just correct the checksum” is the recipe the kernel authors already backed away from. We did not run it.

Do Not Repair the Checksum

The running kernel was 7.2.8-9. Headers were installed, and /lib/modules/$(uname -r)/build was a real tree. It was not a source tree. drivers/net/ethernet/intel/ there had Kconfig stubs and no e1000e/*.c.

The matching sources are the v7.2.8 tag. We pulled that directory, missed regs.h on the first compile, and built out of tree against the running headers:

make -C /lib/modules/$(uname -r)/build M=$PWD modules

The patch is only that loop. On the third failure, warn and break. Do not set -EIO. Do not goto. Probe still reads the MAC and rejects a garbage address. A bad OEM checksum is common. A garbage MAC means the NVM read itself failed, and you should stop.

if (i == 2) {
    dev_warn(&pdev->dev,
             "The NVM Checksum Is Not Valid, continuing without rewrite\n");
    break;
}

Nothing in that diff writes a word to the EEPROM. e1000_write_nvm is untouched. That was the constraint.

/usr/lib/depmod.d/search.conf says search updates extramodules built-in, so a module in /lib/modules/$(uname -r)/updates/ wins over the in-tree one. Remove the stock module, modprobe e1000e, done. A kernel upgrade will not carry that file. The port disappears again until the object is rebuilt.

The Interface Shows Up

The new probe log:

e1000e 0000:00:1f.6: The NVM Checksum Is Not Valid, continuing without rewrite
e1000e 0000:00:1f.6 eth0: Intel(R) PRO/1000 Network Connection
e1000e 0000:00:1f.6 eno1: renamed from eth0

ip -br link finally had eno1. First glance it was DOWN and NO-CARRIER. The driver had registered the device. Link had not come up yet. A minute later NetworkManager had activated “Wired connection 1” on eno1, and Wi-Fi was DOWN. The default route moved to the wire without anyone asking. That is what you want for a headset on the same switch. It also invalidates the address the headset was already using.

WiVRn Streams Games, Not Monitors

WiVRn 26.9 was already up. Avahi was already running. The Quest had paired. The server log shows the session actually focused:

INFO [operator()] Session state changed: XR_SESSION_STATE_FOCUSED

Hand tracking swapped in. A recenter landed. Then:

INFO [operator()] Session state changed: XR_SESSION_STATE_IDLE
ERROR [run_worker] Exception in worker thread: Network is unreachable

The worker died because the route it had been using went away when Wi-Fi dropped. Reconnecting means the Quest has to use the wired address, not the one from ten minutes ago. We restarted the user unit so it was listening on 9757 again. ss had shown the port gone while the process was still alive.

Even with the link healthy, WiVRn does not mirror a desktop. It is an OpenXR runtime. The in-headset launcher lists Steam VR games and .desktop files that carry X-WiVRn-VR. Both lists were empty. Steam was not installed. There was no OpenXR demo. “I connected, now how do I stream something” has no button.

Virtual Desktop does not fill the gap. Its streamer is a Windows program. The Quest Virtual Desktop app has nothing on this machine to talk to.

WayVR Is the Other Half

The program that puts Wayland screens in the headset, on top of WiVRn or Monado, is WayVR. It used to be called WlxOverlay-S. Current release at the time of this session: 26.8.0.

The raw wayvr binary from that release is not self-contained. ldd reported two missing libraries:

libopenxr_loader.so.1 => not found
libopenvr_api.so => not found

The AppImage from the same release is. WayVR-v26.8.0-x86_64.AppImage --help printed usage and exited 0. --version printed wayvr 26.8.0.

Start WayVR on the PC and grant the screen share in the order the notification names. Wrong order means the laser drives the cursor on a different display. Settings has “Clear PipeWire tokens” for that. In the headset, stay in the WiVRn app. WayVR is not a second store app. Double-tap B or Y on the left controller to show or hide the screens. Blue laser is left click, orange is right click, the stick scrolls.

Do not let WiVRn auto-start it. WayVR’s notes say that launch path wraps the overlay in the wrong environment and the screens come up blank. Start it yourself. wayvr --wait is the documented way to have it waiting before the headset connects. A kernel upgrade still drops the patched e1000e. The AppImage does not fix that.

Takeaway

If lspci shows an Ethernet controller and ip link does not, stop looking at NetworkManager. Read the probe line. The NVM Checksum Is Not Valid followed by probe with driver e1000e failed with error -5 means the driver refused to bind. On I219-V and anything else in that PCH family, do not “repair” the checksum. The kernel already learned that write can lock the part. Skip the failed check, keep the MAC validity test, and put the module in updates/ so it wins over the in-tree one.

And a VR runtime is not a desktop streamer. WiVRn gets the headset into an OpenXR session. WayVR is what puts the monitors in that session. They are two programs, and the Quest app for the first one will not speak to Virtual Desktop’s streamer.

What to run

# the NIC is present but unbound?
lspci -k -s 00:1f.6
journalctl -k -b | grep -E 'NVM Checksum|e1000e.*probe'

# after the patched module is in updates/
sudo modprobe -r e1000e
sudo modprobe e1000e
ip -br link

# headset stream, not a desktop mirror
systemctl --user restart wivrn
ss -lnt | grep 9757

# screens, once the headset is connected to the current address
WayVR.AppImage --openxr