Re: [PATCH 1/2] remove rcu_assign_pointer(NULL) penalty with type/macro safety

[Stephen Hemminger <shemminger@vyatta.com>]

On Wed, 13 Feb 2008 16:14:04 -0800
"Paul E. McKenney" <paulmck@linux.vnet.ibm.com> wrote:

> On Wed, Feb 13, 2008 at 03:51:58PM -0800, Stephen Hemminger wrote:
> > On Wed, 13 Feb 2008 15:37:44 -0800
> > "Paul E. McKenney" <paulmck@linux.vnet.ibm.com> wrote:
> > > On Wed, Feb 13, 2008 at 02:42:33PM -0800, Stephen Hemminger wrote:
> > > > On Wed, 13 Feb 2008 14:41:34 -0800
> > > > "Paul E. McKenney" <paulmck@linux.vnet.ibm.com> wrote:
> > > > 
> > > > > On Wed, Feb 13, 2008 at 02:35:37PM -0800, Stephen Hemminger wrote:
> > > > > > On Wed, 13 Feb 2008 14:00:24 -0800
> > > > > > "Paul E. McKenney" <paulmck@linux.vnet.ibm.com> wrote:
> > > > > > 
> > > > > > > Hello!
> > > > > > > 
> > > > > > > This is an updated version of the patch posted last November:
> > > > > > > 
> > > > > > > 	http://archives.free.net.ph/message/20071201.003721.cd6ff17c.en.html
> > > > > > > 
> > > > > > > This new version permits arguments with side effects, for example:
> > > > > > > 
> > > > > > > 	rcu_assign_pointer(global_p, p++);
> > > > > > > 
> > > > > > > and also verifies that the arguments are pointers, while still avoiding
> > > > > > > the unnecessary memory barrier when assigning NULL to a pointer.
> > > > > > > This memory-barrier avoidance means that rcu_assign_pointer() is now only
> > > > > > > permitted for pointers (not array indexes), and so this version emits a
> > > > > > > compiler warning if the first argument is not a pointer.  I built a "make
> > > > > > > allyesconfig" version on an x86 system, and received no such warnings.
> > > > > > > If RCU is ever applied to array indexes, then the second patch in this
> > > > > > > series should be applied, and the resulting rcu_assign_index() be used.
> > > > > > > 
> > > > > > > Given the rather surprising history of subtlely broken implementations of
> > > > > > > rcu_assign_pointer(), I took the precaution of generating a full set of
> > > > > > > test cases and verified that memory barriers and compiler warnings were
> > > > > > > emitted when required.  I guess it is the simple things that get you...
> > > > > > > 
> > > > > > > Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
> > > > > > > ---
> > > > > > > 
> > > > > > >  rcupdate.h |   16 ++++++++++++----
> > > > > > >  1 file changed, 12 insertions(+), 4 deletions(-)
> > > > > > > 
> > > > > > > diff -urpNa -X dontdiff linux-2.6.24/include/linux/rcupdate.h linux-2.6.24-rap/include/linux/rcupdate.h
> > > > > > > --- linux-2.6.24/include/linux/rcupdate.h	2008-01-24 14:58:37.000000000 -0800
> > > > > > > +++ linux-2.6.24-rap/include/linux/rcupdate.h	2008-02-13 13:36:47.000000000 -0800
> > > > > > > @@ -270,12 +270,20 @@ extern struct lockdep_map rcu_lock_map;
> > > > > > >   * structure after the pointer assignment.  More importantly, this
> > > > > > >   * call documents which pointers will be dereferenced by RCU read-side
> > > > > > >   * code.
> > > > > > > + *
> > > > > > > + * Throws a compiler warning for non-pointer arguments.
> > > > > > > + *
> > > > > > > + * Does not insert a memory barrier for a NULL pointer.
> > > > > > >   */
> > > > > > >  
> > > > > > > -#define rcu_assign_pointer(p, v)	({ \
> > > > > > > -						smp_wmb(); \
> > > > > > > -						(p) = (v); \
> > > > > > > -					})
> > > > > > > +#define rcu_assign_pointer(p, v)	\
> > > > > > > +	({ \
> > > > > > > +		typeof(*p) *_________p1 = (v); \
> > > > > > > +		\
> > > > > > > +		if (!__builtin_constant_p(v) || (_________p1 != NULL)) \
> > > > > > > +			smp_wmb(); \
> > > > > > > +		(p) = _________p1; \
> > > > > > > +	})
> > > > > > >  
> > > > > > >  /**
> > > > > > >   * synchronize_sched - block until all CPUs have exited any non-preemptive
> > > > > > 
> > > > > > Will this still work if p is unsigned long?
> > > > > 
> > > > > Hello, Steve,
> > > > > 
> > > > > If p is unsigned long, then use rcu_assign_index() from the next patch in
> > > > > the set.  Looks like Andrew has applied it to -mm -- so please make sure
> > > > > that he is aware if you do use it.
> > > > 
> > > > Make sure fib_trie still works and doesn't get warnings.
> > > 
> > > Ah.  It does take a bit to get fib_trie into one's build -- allyesconfig
> > > doesn't cut it.  Please accept my apologies for my confusion!!!
> > > 
> > > Once fib_trie is configured, I do indeed get:
> > > 
> > >     net/ipv4/fib_trie.c: In function ‘node_set_parent’:
> > >     net/ipv4/fib_trie.c:182: warning: comparison between pointer and integer
> > > 
> > > So, given that node->parent is an unsigned long, I changed node_set_parent()
> > > to the following:
> > > 
> > > static inline void node_set_parent(struct node *node, struct tnode *ptr)
> > > {
> > > 	rcu_assign_index(node->parent, (unsigned long)ptr | NODE_TYPE(node));
> > > }
> > > 
> > > This removes the warnings.  I am a little ambivalent about this, as
> > > this is really a pointer in disguise rather than an array index, but
> > > patch below.  I suppose that another option would be to make node->parent
> > > be a void* and provide appropriate accessor functions/macros.
> > > 
> > > Thoughts?
> > > 
> > > Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
> > 
> > Maybe cast both sides to void * in this case:
> > 
> > static inline void node_set_parent(struct node *node, struct tnode *ptr)
> > {
> >  	rcu_assign_pointer((void *) node->parent, (void *)((unsigned long)ptr | NODE_TYPE(node)));
> > }
> 
> That gets me the following:
> 
> 	net/ipv4/fib_trie.c: In function ‘node_set_parent’:
> 	net/ipv4/fib_trie.c:182: error: invalid lvalue in assignment
> 
> However, as with much in computing, an extra level of indirection fixes
> things.  Your call as to whether or not the cure is preferable to the
> disease.  ;-)
> 
> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
> ---
> 
>  fib_trie.c |    4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff -urpNa -X dontdiff linux-2.6.25-rc1/net/ipv4/fib_trie.c linux-2.6.25-rc1-fib_trie-warn.compile/net/ipv4/fib_trie.c
> --- linux-2.6.25-rc1/net/ipv4/fib_trie.c	2008-02-13 14:38:12.000000000 -0800
> +++ linux-2.6.25-rc1-fib_trie-warn.compile/net/ipv4/fib_trie.c	2008-02-13 16:10:07.000000000 -0800
> @@ -179,8 +179,8 @@ static inline struct tnode *node_parent_
>  
>  static inline void node_set_parent(struct node *node, struct tnode *ptr)
>  {
> -	rcu_assign_pointer(node->parent,
> -			   (unsigned long)ptr | NODE_TYPE(node));
> +	rcu_assign_pointer((*(void **)&node->parent),
> +			   (void *)((unsigned long)ptr | NODE_TYPE(node)));
>  }

That is heading towards ugly...  Maybe not using the macro at all (for this case) would be best:

static inline void node_set_parent(struct node *node, struct tnode *ptr)
{
	smp_wmb();
	node->parent = (unsigned long)ptr | NODE_TYPE(node);
}