Re: [RFC patch 17/18] MIPS : Trace clock

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Fri, 2008-11-07 at 00:23 -0500, Mathieu Desnoyers wrote:
> Note for Peter Zijlstra :
> You should probably have a look at lockdep.c raw_spinlock_t lockdep_lock usage.
> I suspect it may be used with preemption enabled in graph_lock(). (not sure
> though, but it's worth double-checking.

Are you worried about the graph_lock() instance in
lookup_chain_cache() ?

That is the locking for validate_chain,

__lock_acquire()
 validate_chain()
   look_up_chain_cache()
     graph_lock()
   check_prevs_add()
     check_prev_add()
       graph_unlock()
       graph_lock()
   graph_unlock()

which is all done without modifying IRQ state.

However, __lock_acquire() is only called with IRQs disabled:

lock_acquire()
  raw_local_irq_save()
  __lock_acquire()

lock_release()
  raw_local_irq_save()
  __lock_release()
    lock_release_nested()
      __lock_acquire()
    lock_release_non_nested()
      __lock_acquire()

lock_set_subclass()
  raw_local_irq_save()
  __lock_set_subclass()
    __lock_acquire()

So I think we're good.