cpuops: Use cmpxchg for xchg to avoid lock semantics

Use cmpxchg instead of xchg to realize this_cpu_xchg.

xchg will cause LOCK overhead since LOCK is always implied but cmpxchg
will not.

Baselines:

xchg()		= 18 cycles (no segment prefix, LOCK semantics)
__this_cpu_xchg = 1 cycle

(simulated using this_cpu_read/write, two prefixes. Looks like the
cpu can use loop optimization to get rid of most of the overhead)

Cycles before:

this_cpu_xchg	 = 37 cycles (segment prefix and LOCK (implied by xchg))

After:

this_cpu_xchg	= 11 cycle (using cmpxchg without lock semantics)

Signed-off-by: Christoph Lameter <cl@linux.com>
Signed-off-by: Tejun Heo <tj@kernel.org>

arch/x86/include/asm/percpu.h

@@ -263,8 +263,9 @@ do {									\
 })
 
 /*
- * Beware: xchg on x86 has an implied lock prefix. There will be the cost of
+ * xchg is implemented using cmpxchg without a lock prefix. xchg is
- * full lock semantics even though they are not needed.
+ * expensive due to the implied lock prefix.  The processor cannot prefetch
+ * cachelines if xchg is used.
  */
 #define percpu_xchg_op(var, nval)					\
 ({									\
@@ -272,25 +273,33 @@ do {									\
 	typeof(var) pxo_new__ = (nval);					\
 	switch (sizeof(var)) {						\
 	case 1:								\
-		asm("xchgb %2, "__percpu_arg(1)				\
+		asm("\n1:mov "__percpu_arg(1)",%%al"			\
+		    "\n\tcmpxchgb %2, "__percpu_arg(1)			\
+		    "\n\tjnz 1b"					\
 			    : "=a" (pxo_ret__), "+m" (var)		\
 			    : "q" (pxo_new__)				\
 			    : "memory");				\
 		break;							\
 	case 2:								\
-		asm("xchgw %2, "__percpu_arg(1)				\
+		asm("\n1:mov "__percpu_arg(1)",%%ax"			\
+		    "\n\tcmpxchgw %2, "__percpu_arg(1)			\
+		    "\n\tjnz 1b"					\
 			    : "=a" (pxo_ret__), "+m" (var)		\
 			    : "r" (pxo_new__)				\
 			    : "memory");				\
 		break;							\
 	case 4:								\
-		asm("xchgl %2, "__percpu_arg(1)				\
+		asm("\n1:mov "__percpu_arg(1)",%%eax"			\
+		    "\n\tcmpxchgl %2, "__percpu_arg(1)			\
+		    "\n\tjnz 1b"					\
 			    : "=a" (pxo_ret__), "+m" (var)		\
 			    : "r" (pxo_new__)				\
 			    : "memory");				\
 		break;							\
 	case 8:								\
-		asm("xchgq %2, "__percpu_arg(1)				\
+		asm("\n1:mov "__percpu_arg(1)",%%rax"			\
+		    "\n\tcmpxchgq %2, "__percpu_arg(1)			\
+		    "\n\tjnz 1b"					\
 			    : "=a" (pxo_ret__), "+m" (var)		\
 			    : "r" (pxo_new__)				\
 			    : "memory");				\