As reported by Antonio, there are cases where the ep->lock
can be taken twice, triggering a deadlock.
The typical sequence is :
 irq_handler
   \
    -> gadget.complete()
       \
        -> pxa27x_udc.pxa_ep_queue() : ep->lock is taken
           \
            -> gadget.complete()
               \
                -> pxa27x_udc.pxa_ep_queue() : ep->lock is taken
                                               ==> *deadlock*
The patch fixes this by :
 - releasing the lock each time gadget.complete() is called
 - adding a check in handle_ep() to detect a recursive call,
   in which case the function becomes on no-op.

The patch is still not good enough for ep0. For this unique
endpoint, another well thought over patch will be needed.
Reported-by: Antonio Ospite <ospite@studenti.unina.it>
Tested-by: Antonio Ospite <ospite@studenti.unina.it>
Signed-off-by: Robert Jarzmik <robert.jarzmik@free.fr>
Cc: David Brownell <dbrownell@users.sourceforge.net>
Cc: Eric Miao <eric.y.miao@gmail.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>