HID: wiimote: fix FF deadlock

The input core has an internal spinlock that is acquired during event
injection via input_event() and friends but also held during FF callbacks.
That means, there is no way to share a lock between event-injection and FF
handling. Unfortunately, this is what is required for wiimote state
tracking and what we do with state.lock and input->lock.

This deadlock can be triggered when using continuous data reporting and FF
on a wiimote device at the same time. I takes me at least 30m of
stress-testing to trigger it but users reported considerably shorter
times (http://bpaste.net/show/132504/) when using some gaming-console
emulators.

The real problem is that we have two copies of internal state, one in the
wiimote objects and the other in the input device. As the input-lock is
not supposed to be accessed from outside of input-core, we have no other
chance than offloading FF handling into a worker. This actually works
pretty nice and also allows to implictly merge fast rumble changes into a
single request.

Due to the 3-layered workers (rumble+queue+l2cap) this might reduce FF
responsiveness. Initial tests were fine so lets fix the race first and if
it turns out to be too slow we can always handle FF out-of-band and skip
the queue-worker.

Cc: <stable@vger.kernel.org> # 3.11+
Reported-by: Thomas Schneider
Signed-off-by: David Herrmann <dh.herrmann@gmail.com>
Signed-off-by: Jiri Kosina <jkosina@suse.cz>

drivers/hid/hid-wiimote.h

@@ -133,13 +133,15 @@ struct wiimote_state {
 	__u8 *cmd_read_buf;
 	__u8 cmd_read_size;
 
-	/* calibration data */
+	/* calibration/cache data */
 	__u16 calib_bboard[4][3];
+	__u8 cache_rumble;
 };
 
 struct wiimote_data {
 	struct hid_device *hdev;
 	struct input_dev *input;
+	struct work_struct rumble_worker;
 	struct led_classdev *leds[4];
 	struct input_dev *accel;
 	struct input_dev *ir;