Re: [PATCH 1/2] lockdep: check the depth of subclass

[Hitoshi Mitake <mitake@dcl.info.waseda.ac.jp>]

On 10/13/10 16:33, Peter Zijlstra wrote:
 > On Wed, 2010-10-13 at 11:26 +0900, Hitoshi Mitake wrote:
 >>   >>  @@ -639,6 +639,21 @@ look_up_lock_class(struct lockdep_map 
*lock, unsigned int subclass)
 >>   >>      }
 >>   >>    #endif
 >>   >>
 >>   >>  +   if (unlikely(subclass>= MAX_LOCKDEP_SUBCLASSES)) {
 >>   >>  +           /*
 >>   >>  +            * This check should be done not only in 
__lock_acquire()
 >>   >>  +            * but also here. Because register_lock_class() is 
also called
 >>   >>  +            * by lock_set_class(). Callers of 
lock_set_class() can
 >>   >>  +            * pass invalid value as subclass.
 >>   >>  +            */
 >>   >>  +
 >>   >>  +           debug_locks_off();
 >>   >>  +           printk(KERN_ERR "BUG: looking up invalid subclass: 
%u\n", subclass);
 >>   >>  +           printk(KERN_ERR "turning off the locking 
correctness validator.\n");
 >>   >>  +           dump_stack();
 >>   >>  +           return NULL;
 >>   >>  +   }
 >>   >
 >>   >  Would we catch all cases if we moved this check from 
__lock_acquire()
 >>   >  into register_lock_class()? It would result in only a single 
instance of
 >>   >  this logic.
 >>   >
 >>
 >> I think that __lock_acquire() should also check the value of subclass.
 >> Because it access to the lock->class_cache as array
 >> before calling look_up_lock_class() after applying this patch.
 >>
 >> So if the check isn't done in __lock_acquire(),
 >> the invalid addresses might be interpreted as the addresses of
 >> struct lock_class.
 >
 >
 > But __lock_acquire() does:
 >
 >    if (subclass<  NR_LOCKDEP_CACHING_CLASSES)
 >      class = lock->class_cache[subclass];
 >
 >    if (!class)
 >      class = register_lock_class();
 >
 > So by moving the: subclass>= MAX_LOCKDEP_SUBCLASSES, check into
 > register_lock_class() it would still trigger for __lock_acquire().
 > Because NR_LOCKDEP_CACHING_CLASSES<= MAX_LOCKDEP_SUBCLASSES, and thus
 > for subclass>= MAX_LOCKDEP_SUBCLASSES we'll always call into
 > register_lock_class() and trigger the failure there, no?
 >
 >
 >

Ahh, sorry, you are right.
So current checking subclass >= MAX_LOCKDEP_SUBCLASSES is redundant,
I'll remove this checking and resend second version later.

Thanks for your advice!