The pad paired successfully. It bonded. The panel said Connected: yes. It just never sent a single button press.
The Situation
An Xbox Wireless Controller (the BLE one, model 1914) on a desktop running omarchy, with the internal Intel AX210 radio. The complaint was the vague kind: it disconnects at random. No pattern worth the name. Sometimes it held for an hour, sometimes it fell over while I was reading the Bluetooth panel instead of using the thing.
The obvious suspects were the usual ones — an out-of-tree driver, a flaky antenna, a tired battery. None of them were involved.
Three separate failures were stacked on top of each other, and each one produced a different kind of disconnection. That’s why it read as random: I was watching three unrelated bugs wearing one coat.
Act I: An agent that answers into /dev/null
The first thing worth reading is not the Bluetooth panel. It’s the journal.
bt-agent[1320]: Authorize this device pairing (yes/no)? Device: Xbox Wireless Controller
bt-agent[1320]: Authorize this device pairing (yes/no)? Device: Xbox Wireless Controller
Same line, every seven seconds, for as long as the pad was awake. That is not a signal problem. That is a prompt nobody can answer.
omarchy ships bt-agent.service as a user unit, running bt-agent -c NoInputNoOutput from bluez-tools 0.2.0 — a package whose last release predates the controller by several years. The -c NoInputNoOutput flag looks like it should auto-accept everything. It does not. In bluez-tools, that flag only sets the advertised capability; the callbacks stay interactive unless you pass -d. So RequestAuthorization prints a yes/no prompt and reads stdin.
Under systemd, stdin is not a terminal:
/proc/1320/fd/0 -> /dev/null
EOF. BlueZ gets org.bluez.Error.Rejected, the controller retries seven seconds later, and the two of them do this until the battery dies. It also ignores SIGTERM, which is why systemctl --user stop sat in deactivating until systemd lost patience and killed it.
The fix is a real agent. BlueZ’s Agent1 interface has nine methods and a gamepad only needs one behaviour, so it’s a short Gio script that returns success for everything and logs what it was asked:
agent object registered at /local/bt_auto_agent
RegisterAgent ok
RequestDefaultAgent ok
Installed as a drop-in override on the stock unit (ExecStart= cleared, then pointed at the script), so a distro update doesn’t put the broken one back.
That fixed pairing. It did not fix the controller.
Act II: The mitigation that wasn’t
While confirming the radio wasn’t being suspended, one line was worth a second look:
options usbcore autosuspend=-1 # /etc/modprobe.d/omarchy-usb-autosuspend.conf
Nice idea. It does nothing here, because usbcore is built into the kernel on this build:
$ modinfo -n usbcore
(builtin)
modprobe.d options are only honoured for loadable modules. The runtime value told the real story:
$ cat /sys/module/usbcore/parameters/autosuspend
2
$ cat /sys/module/btusb/parameters/enable_autosuspend
Y
/sys/bus/usb/devices/1-14/power/control = auto # the Bluetooth radio
Two seconds of idleness and the radio is allowed to suspend. The kernel log had already said so, twice:
Bluetooth: hci0: ACL packet for unknown connection handle 2048
The fix is a udev rule matching the device class rather than the USB path, so it survives the radio moving ports:
ACTION=="add", SUBSYSTEM=="usb", ATTR{bDeviceClass}=="e0", TEST=="power/control", ATTR{power/control}="on"
Still not the main bug. But it explains the drops that happened while everything else worked.
Act III: “Pairing successful” is a lie
With a working agent in place, pairing finally completed:
[CHG] Device ... Bonded: yes
[CHG] Device ... Paired: yes
Pairing successful
Then, four seconds later:
[SIGNAL] LE.Disconnected - org.bluez.Reason.Authentication, Connection terminated due to authentication failure
And again. And again. Meanwhile the agent log stayed completely silent — the agent was never asked anything, so this wasn’t a rejected pairing. BlueZ believed it already had a bond and was trying to use it.
The default for a peer-initiated re-pair is to refuse:
# /etc/bluetooth/main.conf
#JustWorksRepairing = never
Setting it to confirm made no difference whatsoever. Worth saying plainly, because I expected it to be the answer and it wasn’t.
Act IV: The key that never got sent
The pad’s LED was still blinking fast while the host insisted it was bonded, and it was advertising over 400 packets every ten seconds — a controller that does not believe it is paired.
The xpadneo documentation added a section for this one week before I hit it, titled Controller Keeps Blinking After Pairing:
The controller considers the bond incomplete because bluez does not distribute identity information unless LE privacy is enabled.
That was it. The adapter supports privacy and had it off:
current settings: powered bondable ssp br/edr le secure-conn wide-band-speech ...
No privacy in that list. macOS and Android hand out identity information during pairing as a matter of course, which is why the same pad behaves on a console and acts confused on Linux.
The fix is three lines plus two interval values:
[General]
Privacy = device
JustWorksRepairing = confirm
[LE]
MinConnectionInterval = 7
MaxConnectionInterval = 9
ConnectionLatency = 0
ConnectionSupervisionTimeout = 200
Restart bluetoothd, remove the half-finished pairing, pair and trust again. Immediately after:
current settings: ... le secure-conn privacy ...
Paired: yes Bonded: yes Trusted: yes Connected: yes
N: Name="Xbox Wireless Controller"
H: Handlers=kbd event30 js0
js0 is the part I actually cared about. Not just evdev — the legacy joystick node, which is what the older SDL1-era games look for.
Receipts
| Layer | Symptom | Cause | Fix |
|---|---|---|---|
| Userspace agent | pad rejected every 7s, never bonds | bt-agent prompts on /dev/null |
replace with auto-accept Gio agent |
| USB power | drops during play, ACL ... unknown handle |
usbcore builtin, autosuspend 2s |
udev rule pins radio to power/control=on |
| BlueZ policy | Reason.Authentication on every reconnect |
no LE privacy → no IRK distributed | Privacy = device + LE intervals |
Three different-looking failures, one Bluetooth panel cheerfully reporting Connected: yes through all of it.
Takeaway
A BLE HID device that pairs but sends no input is failing authentication, not transport — and the layer that reports “Connected” is the layer least able to tell you why. Read the daemon’s own log lines before you touch a driver.
Two caveats from the documentation, both real. Privacy = device can stop unrelated devices from pairing; if something new refuses, that’s the line to remove. And some controllers need JustWorksRepairing = always rather than confirm. The permanent cure is a controller firmware update, which requires Windows — so this config is the workaround for everyone who doesn’t keep a Windows box around.
What to run
# stop the broken agent, install an auto-accepting one
systemctl --user disable --now bt-agent
# (see override drop-in for ExecStart replacement)
# keep the radio out of USB autosuspend
sudo tee /etc/udev/rules.d/60-bt-no-autosuspend.rules >/dev/null <<'EOF'
ACTION=="add", SUBSYSTEM=="usb", ATTR{bDeviceClass}=="e0", TEST=="power/control", ATTR{power/control}="on"
EOF
sudo udevadm control --reload-rules
# hand out identity information during pairing
sudo sed -i -e 's/^#Privacy = off/Privacy = device/' \
-e 's/^#JustWorksRepairing = never/JustWorksRepairing = confirm/' \
/etc/bluetooth/main.conf
sudo systemctl restart bluetooth
bluetoothctl remove <CONTROLLER_MAC>
bluetoothctl pair <CONTROLLER_MAC>
bluetoothctl trust <CONTROLLER_MAC>
