Re: [PATCH tip/core/urgent 1/4] rcu: add rcu_access_pointer and rcu_dereference_protected

[Josh Triplett <josh@joshtriplett.org>]

On Fri, Apr 09, 2010 at 09:57:35AM -0700, Paul E. McKenney wrote:
> On Fri, Apr 09, 2010 at 11:24:18AM +0200, Eric Dumazet wrote:
> > Le vendredi 09 avril 2010 à 10:10 +0100, David Howells a écrit :
> > > Paul E. McKenney <paulmck@linux.vnet.ibm.com> wrote:
> > > 
> > > > +#define rcu_access_pointer(p) ((void *)ACCESS_ONCE(p))
> > > > ...
> > > > +#define rcu_access_pointer(p)		((void *)ACCESS_ONCE(p))
> > > 
> > > There's no difference between your two versions of rcu_access_pointer(), so
> > > you could move that whole construct outside of the #ifdef'ed section.
> > > 
> > > Other than that:
> > > 
> > > Acked-by: David Howells <dhowells@redhat.com>
> > 
> > I also prefer to keep type checking if possible
> > 
> > #define rcu_access_pointer(p)		(ACCESS_ONCE(p))
> 
> Yikes!  I was obsessing on preventing people from dereferencing the
> pointer, but you are quite right, they might be comparing it to a
> constant pointer to check for array bounds or for use of a statically
> allocated emergency-pool item.

If you want to prevent people from deferencing the pointer directly, or
from assigning it to some other pointer, you could use Sparse's
__attribute__((noderef)) and
__attribute__((address_space(...)).

See the definitions of __user and __iomem in include/linux/compiler.h
for instance.  (And, looking at that file, now __percpu.  Nice!)

You'd then have to explicitly use __force at the point where you
legitimately dereference it.

- Josh Triplett