Re: [PATCH] HID: wiimote: initialise rumble_worker once per device

From: Jiri Kosina

Date: Thu Nov 19 2026 - 15:46:40 EST


On Mon, 28 Sep 2026, Jinmo Yang wrote:

> wdata->rumble_worker is a single work_struct in struct wiimote_data, but
> two different modules called INIT_WORK() on it:
>
> wiimod_rumble_probe() devtype module, once per device
> wiimod_pro_probe() extension module, once per extension hotplug
>
> These have independent lifetimes. The devtype module is loaded once by
> wiimote_modules_load() and stays until wiimote_destroy(), while the
> extension module is loaded and unloaded any number of times by
> wiimote_ext_load() / wiimote_ext_unload(). wiimote_ext_load() calls the
> extension probe with no lock and no cancel_work_sync(), and for the
> common EXT_NONE -> EXT_PRO_CONTROLLER transition the preceding
> wiimote_ext_unload() is a no-op because wiimod_ext_table[] maps both
> EXT_NONE and EXT_UNKNOWN to wiimod_dummy, which has no .remove.
>
> So nothing stops the force-feedback path, which is still live in the
> base module, from having queued the work already:
>
> /* wiimod_rumble_play() */
> wdata->state.cache_rumble = value;
> schedule_work(&wdata->rumble_worker);
>
> INIT_WORK() on a queued work item is memory corruption. __INIT_WORK()
> does
>
> (_work)->data = (atomic_long_t) WORK_DATA_INIT();
> INIT_LIST_HEAD(&(_work)->entry);
>
> which clears the pool reference and the PENDING bit, and re-points
> ->entry at itself while the item is still linked into the worker pool
> list, leaving its neighbours pointing at a node that has left the list.
>
> Note that CONFIG_DEBUG_OBJECTS_WORK hides the impact rather than just
> reporting it: work_fixup_init() calls cancel_work_sync() for the caller,
> so a debug kernel prints a warning and then repairs the state. That is
> the kernel stating what the driver should have done. Production kernels
> have no such repair.
>
> Initialise the work once in wiimote_create(), which is where the object
> that contains it is created and which already initialises
> wdata->queue.worker, wdata->init_worker and wdata->timer, and drop both
> module-level INIT_WORK() calls. wiimote_create() runs as the first
> statement of wiimote_hid_probe(), before hid_parse() and hid_hw_start(),
> so it strictly precedes the existence of any input device and therefore
> of any path that could queue the work.
>
> This does not affect the deadlock fix in commit f50f9aabf32d ("HID:
> wiimote: fix FF deadlock"). That fix is the offload itself - caching
> the value and deferring to a worker instead of taking state.lock inside
> the FF callback - and this patch leaves wiimod_rumble_play(),
> wiimod_rumble_worker() and state.cache_rumble untouched. Only the
> placement of the initialisation changes.
>
> The two halves come from different actors, and both are available on a
> stock Android phone without root. The extension hotplug half needs only
> /dev/uhid, which is group 3011 and held by the shell. The shell cannot
> write to an evdev node - SELinux grants it read access only - so the
> force-feedback half comes from an ordinary app calling
> InputDevice.getVibrator().vibrate(), which reaches EVIOCSFF through
> system_server; VIBRATE is a normal permission and is granted
> automatically. A paired Bluetooth HID device can supply the hotplug
> half instead of /dev/uhid.
>
> The ordering matters: the device has to be created without an extension
> so that the devtype resolves to a GEN10/GEN20 that includes
> WIIMOD_RUMBLE, because WIIMOTE_DEV_PRO_CONTROLLER does not, and the
> extension has to be plugged afterwards.
>
> Measured before the change:
>
> Pixel 11, stock Android 17 user build, SELinux enforcing, no KASAN and
> no DEBUG_OBJECTS at runtime, kernel 6.12.81: reboots with
> ro.boot.bootreason=kernel_panic one to two seconds after the two paths
> start overlapping.
>
> Unable to handle kernel NULL pointer dereference at virtual address
> 0000000000000008
> CPU: 5 Comm: kworker/5:4
> Workqueue: wiimod_rumble_worker (events)
> pc : process_scheduled_works+0xd4/0x810
> Kernel panic - not syncing: Oops: Fatal exception
>
> The empty workqueue name in that line is itself the corruption:
> print_worker_info() reads wq->name with copy_from_kernel_nofault() and
> it failed, while another CPU in the same dump shows the normal form,
> "Workqueue: events wiimote_init_worker" - the task that called
> wiimod_pro_probe().
>
> x86_64 with CONFIG_KASAN_GENERIC and CONFIG_DEBUG_OBJECTS_WORK both
> enabled, 15 s in: KASAN says nothing, because INIT_WORK() writes to a
> valid field of a live object, and DEBUG_OBJECTS_WORK reports
>
> ODEBUG: init active (active state 0) object type: work_struct
> hint: wiimod_rumble_worker+0x0/0x70
> WARNING: lib/debugobjects.c:629 at debug_print_object
> Workqueue: events wiimote_init_worker
> __debug_object_init+0x1ff/0x3b0
> __init_work+0x51/0x60
> wiimod_pro_probe+0x28/0xba0
> wiimote_init_worker.cold+0xb6a/0xe88
>
> After the change the same reproducer drives 36774 extension hotplugs and
> 612156 force-feedback plays in 60 s on that x86_64 kernel with no ODEBUG,
> WARNING or KASAN output, and the rumble reports still arrive, so the
> force-feedback path is unaffected.
>
> Fixes: f50f9aabf32d ("HID: wiimote: fix FF deadlock")
> Cc: stable@xxxxxxxxxxxxxxx
> Signed-off-by: Jinmo Yang <jinmo44.yang@xxxxxxxxx>

Applied, thanks.

--
Jiri Kosina
SUSE Labs