Since v->counter is both read and written, it should be an output as well
as an input for the asm.  The current code only gets away with this because
counter is volatile.  Also, according to Documents/atomic_ops.txt,
atomic_add_return should provide a memory barrier, in particular a compiler
barrier, so the asm should be marked as clobbering memory.

Test case:

#include <stdio.h>

typedef struct { int counter; } atomic_t; /* NB: no "volatile" */

#define ATOMIC_INIT(i)	{ (i) }

#define atomic_read(v)		((v)->counter)

static __inline__ int atomic_add_return(int i, atomic_t *v)
{
	int __i = i;

	__asm__ __volatile__(
		"lock; xaddl %0, %1;"
		:"=r"(i)
		:"m"(v->counter), "0"(i));
/*	__asm__ __volatile__(
		"lock; xaddl %0, %1"
		:"+r" (i), "+m" (v->counter)
		: : "memory"); */
	return i + __i;
}

int main (void) {
	atomic_t a = ATOMIC_INIT(0);
	int x;

	x = atomic_add_return (1, &a);
	if ((x!=1) || (atomic_read(&a)!=1))
		printf("fail: %i, %i\n", x, atomic_read(&a));
}
Signed-off-by: Duncan Sands <baldrick@free.fr>
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: Andi Kleen <ak@suse.de>
Acked-by: David Howells <dhowells@redhat.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>