locking/x86: Implement local_xchg() using CMPXCHG without the LOCK prefix

Implement local_xchg() using the CMPXCHG instruction without the LOCK prefix.
XCHG is expensive due to the implied LOCK prefix.  The processor
cannot prefetch cachelines if XCHG is used.

Signed-off-by: Uros Bizjak <[email protected]>
Signed-off-by: Ingo Molnar <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Cc: Waiman Long <[email protected]>
Cc: Will Deacon <[email protected]>
Cc: Thomas Gleixner <[email protected]>
Cc: Linus Torvalds <[email protected]>
Cc: Paul E. McKenney <[email protected]>
Cc: "H. Peter Anvin" <[email protected]>
Link: https://lore.kernel.org/r/[email protected]

1 files changed, 14 insertions, 2 deletions

diff --git a/arch/x86/include/asm/local.h b/arch/x86/include/asm/local.h
index 73dba8b94443..59aa966dc212 100644
--- a/arch/x86/include/asm/local.h
+++ b/arch/x86/include/asm/local.h
@@ -131,8 +131,20 @@ static inline bool local_try_cmpxchg(local_t *l, long *old, long new)
 				 (typeof(l->a.counter) *) old, new);
 }
 
-/* Always has a lock prefix */
-#define local_xchg(l, n) (xchg(&((l)->a.counter), (n)))
+/*
+ * Implement local_xchg using CMPXCHG instruction without the LOCK prefix.
+ * XCHG is expensive due to the implied LOCK prefix.  The processor
+ * cannot prefetch cachelines if XCHG is used.
+ */
+static __always_inline long
+local_xchg(local_t *l, long n)
+{
+	long c = local_read(l);
+
+	do { } while (!local_try_cmpxchg(l, &c, n));
+
+	return c;
+}
 
 /**
  * local_add_unless - add unless the number is already a given value