Re: [PATCH 1/1] lockdep: Kill held_lock->check and "int check" arg of __lock_acquire()

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Mon, Jan 13, 2014 at 06:28:33PM +0100, Peter Zijlstra wrote:
> On Mon, Jan 13, 2014 at 06:06:09PM +0100, Oleg Nesterov wrote:
> > On 01/12, Oleg Nesterov wrote:
> > >
> > > The "int check" argument of lock_acquire() and held_lock->check
> > > are misleading and unneeded. This is only used as a boolean, 2
> > > denotes "true", everything else is "false". And this boolean is
> > > always equal to prove_locking.
> > >
> > > The only exception is __lockdep_no_validate__ which should make
> > > this condition "false" in validate_chain().
> > 
> > And I missed mark_irqflags(),
> > 
> > > @@ -3136,7 +3130,7 @@ static int __lock_acquire(struct lockdep_map *lock, unsigned int subclass,
> > >  	hlock->holdtime_stamp = lockstat_clock();
> > >  #endif
> > >
> > > -	if (check == 2 && !mark_irqflags(curr, hlock))
> > > +	if (prove_locking && !mark_irqflags(curr, hlock))
> > >  		return 0;
> > 
> > This change is not right, at least it is not equivalent.
> > 
> > And I just realized that rcu_lock_acquire() does lock_acquire(check => 1).
> > Probably we can mark rcu_lock_map's as __lockdep_no_validate__.
> 
> Can't, RCU needs its own classes. Otherwise it cannot tell which version
> of the RCU read lock its holding at just that moment.

Just confirming this.  RCU uses this to detected mismatches between
the rcu_read_lock() group and the rcu_dereference() group.

							Thanx, Paul

> > Anything else I missed?
> 
> Nothing springs to mind, but then, I totally missed the RCU thing too.
> 
> At the very least we can reduce check to a single bit.
>