While l2cap_user callbacks are running, the whole hci_dev is locked. Even
if we would add more fine-grained locking to HCI core, it would still be
called from the non-reentrant rx work-queue and thus block the event
processing.

However, if we want to perform synchronous I/O during HID device
registration (eg., to perform device-detection), we need the HCI core
to be able to dispatch incoming data.

Therefore, we now move device-registration to a separate worker. The HCI
core can continue running and we add devices asynchronously in another
kernel thread. Device removal is synchronized and waits for the worker
to exit before calling the usual device removal functions.

If l2cap_user->remove is called before the thread registered the devices,
we set "terminate" to true and the thread will skip it. If
l2cap_user->remove is called after it, we notice this as the device
is no longer in HIDP_SESSION_PREPARING state and simply unregister the
device as we did before.
There is no new deadlock as we now call hidp_session_add_dev() with
one lock less held (the HCI lock) and it cannot itself call back into
HCI as it was called with the HCI-lock held before.

One might wonder whether this can block during device unregistration.
But we set "terminate" to true and wake the HIDP thread up _before_
unregistering the HID/input devices. Therefore, all pending HID I/O
operations are canceled. All further I/O attempts will fail with ENODEV
or EIO. So all latency we can get are few context-switches, but no
timeouts or blocking I/O waits!

This change also prepares for a long standing HID bug. All HID devices
that register power_supply devices need to be able to handle callbacks
during registration (a power_supply oddity that cannot easily be fixed).
So with this patch available, we can allow HID I/O during registration
by calling the recently introduced hid_device_io_start/stop helpers,
which currently are a no-op for bluetooth due to this locking.

Note that we cannot do the same for input devices. input-core doesn't
allow us to call input_event() asynchronously to input_register_device(),
which HID-core kindly allows (for good reasons).
Fixing input-core to allow this isn't as easy as it sounds and is,
beside simplifying HIDP, not really an improvement. Hence, we still
register input devices synchronously as we did before. Only HID devices
are registered asynchronously.
Signed-off-by: David Herrmann <dh.herrmann@gmail.com>
Acked-by: Jiri Kosina <jkosina@suse.cz>
Acked-by: Gustavo Padovan <gustavo.padovan@collabora.co.uk>
Tested-by: Daniel Nicoletti <dantti12@gmail.com>
Signed-off-by: Jiri Kosina <jkosina@suse.cz>