As andrew pointed out, it is bad idea to run_task_queue() when
spinlock is held. This fixes it (only could be triggered by user
pressing Magic-D). There's minor problem left where Magic-D could be
delayed/ignored under high load. I'll either fix that or kill Magic-D
support (with acpi, it is very easy to trigger swsusp by echo 4 >
/proc/acpi/sleep, so magic key is probably not neccessary any more).