fix(linux): load uinput rather than rely on another daemon

[AalmanSadath]

Summary

The Linux packages ship 70-openlogi.rules, which grants /dev/uinput with
TAG+="uaccess" — an ACL logind applies in response to a device event. On a
machine that has never loaded the uinput module there is no such device, so
postinstall's udevadm trigger --attr-match=name=uinput matches nothing. It is
a no-op, and its || true hides that. OPTIONS+="static_node=uinput" does create
the node, but a static node gets default permissions and no ACL: opening it would
autoload the module, except opening is exactly what the missing permission forbids.

Nothing in the package breaks that cycle, so today it is broken by whatever else
happens to open /dev/uinput first. On a typical desktop something does —
bluetoothd opens it seconds after boot, and Steam and logiops do too — and
OpenLogi works. Where nothing does, the agent starts, reports the device,
battery and DPI normally over HID++, and button remapping silently never works,
with only a journal warning to say so.

systemd's own rules show the shape of the fix: snd/seq and snd/timer carry a
real GROUP= in 50-udev-default.rules alongside TAG+="uaccess" in
70-uaccess.rules, precisely because uaccess cannot cover the pre-module node.

Changes

• packaging/linux/modules-load/openlogi.conf (new) — loads uinput at boot.
  Package content rather than something the script writes, so uninstalling removes
  it.
• packaging/linux/nfpm.yaml — ships the above to /etc/modules-load.d/ at
  0644.
• packaging/linux/nfpm-scripts/postinstall.sh — modprobe uinput before the
  udevadm block, so the session that installed the package works too and not
  only the next boot.

No Rust changes; .deb, .rpm and .pkg.tar.zst all pick this up from the
shared nfpm config.

Reproducing

Any machine, ~30 seconds. Unloading the module and dropping the ACL leaves the
static node at root:root 0600 with no ACL — byte-for-byte the state a
never-loaded machine boots into.

systemctl --user stop openlogi-agent.service
sudo systemctl stop bluetooth          # or whatever holds /dev/uinput; check with: sudo lsof /dev/uinput
sudo modprobe -r uinput
sudo setfacl -x "u:$USER" /dev/uinput

test -w /dev/uinput && echo "WRITABLE — precondition not met" || echo "NOT WRITABLE — ready"

systemctl --user start openlogi-agent.service
journalctl --user -u openlogi-agent -n 5 --no-pager

Before this change:

WARN openlogi_agent_core::hook_runtime: could not install OS input hook —
  events will not be captured error=Linux input error: Permission denied (os error 13)

openlogi list still shows the mouse and its battery throughout, which is what
makes this easy to miss.

Restore with sudo systemctl start bluetooth && sudo modprobe uinput && sudo udevadm trigger.

Check the mouse's /dev/input/event* node is readable first. If it is not, the
hook fails earlier with no mouse device found under /dev/input and never
reaches /dev/uinput — a different problem.

Testing

Fedora 44, x86_64, on real hardware (MX Master 3S over Bluetooth).

Bug reproduced on two machines, both running released packages, neither with
this change:

• openlogi-v0.7.4-linux-amd64.rpm, with logiops and Solaar removed
• a second machine running the 0.7.3 package, with the mouse's event node
  confirmed readable first so the failure could only be the uinput side

Both produced Permission denied (os error 13) and created no virtual input
device. In its untouched state the second machine does not fail — bluetoothd
had held /dev/uinput since 14 s after boot.

Fix verified as shipped, not just by hand:

cargo run -p xtask -- linux package
rpm -qlp target/release/openlogi-0.7.4-1.x86_64.rpm | grep modules-load
#   /etc/modules-load.d/openlogi.conf
rpm -qp --dump target/release/openlogi-0.7.4-1.x86_64.rpm | grep modules-load
#   mode 0100644 root root
rpm -qp --scripts target/release/openlogi-0.7.4-1.x86_64.rpm | grep modprobe
#   modprobe uinput || true

sudo dnf install -y target/release/openlogi-0.7.4-1.x86_64.rpm
#   -> uinput module loaded, /dev/uinput writable, hook installs,
#      "OpenLogi virtual mouse" node created
sudo dnf remove -y openlogi
#   -> /etc/modules-load.d/openlogi.conf removed, no .rpmsave left behind

The .deb was checked the same way (ar p … data.tar.gz | tar tzf -) and carries
both the config file and the modprobe line.

CI:

cargo xtask ci
#   7 passed, 0 failed, 2 skipped

Includes the shell job (shellcheck + shfmt -d), which gates
postinstall.sh. Not run: tests (macos) (wrong host) and cargo-deny
(not installed locally) — neither is affected by a packaging-only change, but
neither was executed.

[greptile-apps]

Greptile Summary

The PR ensures Linux installations load uinput, allowing the existing udev rule to grant access before OpenLogi creates its virtual input device.

• Ships a modules-load configuration in Linux packages and the from-source installer.
• Loads uinput during package and from-source installation for immediate use.
• Enables uinput through the native NixOS module option.
• Waits for triggered udev events and removes the boot configuration during from-source uninstall.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
packaging/linux/install.sh	Installs the boot-time module configuration, loads uinput for the current session, and waits for queued udev events.
packaging/linux/modules-load/openlogi.conf	Adds the declarative boot-time request to load the uinput kernel module.
packaging/linux/nfpm-scripts/postinstall.sh	Loads uinput before retriggering udev rules during package installation.
packaging/linux/nfpm.yaml	Includes the new modules-load configuration in all Linux packages produced from the shared nfpm descriptor.
packaging/linux/nixos-module.nix	Uses the native NixOS kernel-module option to load uinput at boot.
packaging/linux/uninstall.sh	Removes the from-source modules-load configuration without unloading a potentially shared live module.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[Linux installation] --> B{Installation path}
  B -->|deb/rpm| C[Ship openlogi.conf]
  B -->|From source| D[Install openlogi.conf]
  B -->|NixOS module| E[Set boot.kernelModules]
  C --> F[modprobe uinput during postinstall]
  D --> G[modprobe uinput during install]
  E --> H[Load uinput during boot]
  F --> I[Trigger udev rules]
  G --> I
  H --> I
  I --> J[Apply uaccess ACL]
  J --> K[OpenLogi creates virtual input device]

Loading

_{Reviews (4): Last reviewed commit: "fix(linux): wait for udev in the from-so..." | Re-trigger Greptile}

[AalmanSadath]

Greptile flagged that install.sh reloaded and triggered udev but never settled, so the script could exit before the uaccess ACLs were applied. 34142fc adds the udevadm settle that postinstall.sh already had, so the two paths now agree.

The push dismissed your approval, so it would need re-approval.