Date:	Thu, 10 Apr 2008 15:32:06 -0700
From:	"Paul E. McKenney" <paulmck@...ux.vnet.ibm.com>
To:	Johannes Berg <johannes@...solutions.net>
Cc:	Linux Kernel list <linux-kernel@...r.kernel.org>,
	linux-sparse <linux-sparse@...r.kernel.org>
Subject: Re: Using sparse to catch invalid RCU dereferences?

On Wed, Apr 09, 2008 at 10:09:46PM +0200, Johannes Berg wrote:
> 
> > It might be.  There are a number of places where it is legal to access
> > RCU-protected pointers directly, and all of these would need to be
> > changed.  For example, in the example above, one could do:
> > 
> > 	foo = NULL;
> 
> Ok, that I understand, but sparse always treats NULL specially anyway.

But "int foo = 0;" would need the memory barrier -- index 0 of some
RCU-protected array.

> > I recently tried to modify rcu_assign_pointer() to issue the memory
> > memory barrier only when the pointer was non-NULL, but this ended badly.
> 
> Hm? I thought that's in the current tree.

It was for a bit.  Build failures in odd (but very real) circumstances.

> > Probably because I am not the greatest gcc expert around...  We ended
> > up having to define an rcu_assign_index() to handle the possibility of
> > assigning a zero-value array index, but my attempts to do type-checking
> > backfired, and I eventually gave it up.  Again, someone a bit more clued
> > in to gcc than I am could probably pull it off.
> 
> Ah, ok.
> 
> > In addition, it is legal to omit rcu_dereference() and rcu_assign_pointer()
> > when holding the update-side lock.
> 
> That I don't understand. Well, I do understand that omitting
> rcu_dereference() is ok, but it seems to me that the memory and compiler
> barrier in rcu_assign_pointer() is actually needed.

You are right -- I was confused.  The case where you can omit the
rcu_assign_pointer() would be when building a multiple-element data
structure that is then published as a unit.  For example:

	p = kmalloc(sizeof(*p), GFP_KERNEL);
	q = kmalloc(sizeof(*p), GFP_KERNEL);
	p->next = q; /* don't need rcu_assign_pointer() here. */
	q->next = NULL;  /* or here. */
	/* initialize other fields of p and q. */
	rcu_assign_pointer(global_pointer, p);

The assignment to p->next doesn't have to be rcu_assign_pointer() because
other CPUs are unable to access the data structure -- only the final
assignment that publishes the whole group need be rcu_assign_pointer().
On the other hand, the cost of the extra memory barrier would be
insignificant in most cases.

> I've been playing a bit, see below for my play rcupdate.h and test.c
> test program.
> 
> Unfortunately, sparse doesn't have the ability to declare
> "__attribute__((force_bitwise)) typeof(p)" or even
> "__attribute__((force)) typeof(p)" which makes this force more than
> necessary and causes it to not catch when incompatible pointers are
> used. gcc notices that because I only do a cast at all for sparse, but
> that doesn't help, since e.g. list_for_each_entry_rcu() requires that
> the correct type is returned. So without sparse supporting the latter
> notation, we don't stand a chance.

"<feff>"???

> Also, I wouldn't know how to declare that an array or so needs
> rcu-access to the members.

Hmmm...  Can you apply the address-space attribute to the array itself?
I suppose one could convert the array to a pointer, but yecch!

						Thanx, Paul

> johannes
> 
> 
> rcupdate.h:
> 
> #define USE_BITWISE
> 
> #ifdef __CHECKER__
> #ifdef USE_BITWISE
> #define __rcu __attribute__((bitwise))
> #define __force_rcu_cast(p) (*((__attribute__((force)) void **)&(p)))
> // would like instead:
> //#define __force_rcu_cast(p) ((__attribute__((force_bitwise)) typeof(p)) (p))
> #else /* not bitwise */
> #define __rcu __attribute__((address_space(3)))
> #define __force_rcu_cast(p) (*((__attribute__((force)) void **)&(p)))
> // would like instead:
> //#define __force_rcu_cast(p) ((__attribute__((force_address_space)) typeof(p)) (p))
> #endif
> 
> #else /* not checker */
> #define __rcu
> #define __force_rcu_cast(p) (p)
> #endif
> 
> #define ACCESS_ONCE(x) (*(volatile typeof(x) *)&(x))
> 
> #define rcu_dereference(p)     ({ \
> 				typeof(p) _________p1 = ACCESS_ONCE(p); \
> 				smp_read_barrier_depends(); \
> 				__force_rcu_cast(_________p1); \
> 				})
> 
> /**
>  * rcu_fetch - fetch an RCU-protected pointer in the update-locked
>  * critical section.
>  *
>  * This macro exists for documentation and code checking purposes.
>  */
> #define rcu_fetch(p)     __force_rcu_cast(p);
> 
> #define rcu_assign_pointer(p, v) \
> 	({ \
> 		if (!__builtin_constant_p(v) || \
> 		    ((v) != NULL)) \
> 			smp_wmb(); \
> 		__force_rcu_cast(p) = (v); \
> 	})
> 
> 
> test.c:
> 
> #include <stdlib.h>
> #include "rcupdate.h"
> 
> /* my rcu protected variables */
> static unsigned int __rcu *prot;
> static unsigned int __rcu *prot_same;
> static unsigned char __rcu *prot2;
> 
> // dummies
> static smp_read_barrier_depends(void) {}
> static smp_wmb(void) {}
> 
> int main(void)
> {
> 	unsigned int *tmp;
> 
> 	// no warnings from sparse due to forced cast
> 	rcu_assign_pointer(prot, tmp);
> 	// but gcc warns
> 	rcu_assign_pointer(prot2, tmp);
> 
> 	// no warnings
> 	rcu_assign_pointer(prot, NULL);
> 	rcu_assign_pointer(prot2, NULL);
> 
> 	// no warnings
> 	prot = NULL;
> 	prot2 = NULL;
> 
> 	// no warnings from sparse due to forced cast
> 	tmp = rcu_dereference(prot);
> 	// but gcc warns
> 	tmp = rcu_dereference(prot2);
> 
> 	/* now within locked section rcu_dereference isn't required */
> 
> 	// no warnings from sparse due to forced cast
> 	tmp = rcu_fetch(prot);
> 	// but gcc warns
> 	tmp = rcu_fetch(prot2);
> 
> 	/* not caught with address_space, but is caught with bitwise */
> 	prot = prot_same;
> }
> 


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists