locking/atomic/x86: Introduce arch_try_cmpxchg64_local()

Introduce arch_try_cmpxchg64_local() for 64-bit and 32-bit targets
to improve code using cmpxchg64_local().  On 64-bit targets, the
generated assembly improves from:

    3e28:	31 c0                	xor    %eax,%eax
    3e2a:	4d 0f b1 7d 00       	cmpxchg %r15,0x0(%r13)
    3e2f:	48 85 c0             	test   %rax,%rax
    3e32:	0f 85 9f 00 00 00    	jne    3ed7 <...>

to:

    3e28:	31 c0                	xor    %eax,%eax
    3e2a:	4d 0f b1 7d 00       	cmpxchg %r15,0x0(%r13)
    3e2f:	0f 85 9f 00 00 00    	jne    3ed4 <...>

where a TEST instruction after CMPXCHG is saved.  The improvements
for 32-bit targets are even more noticeable, because double-word
compare after CMPXCHG8B gets eliminated.
Signed-off-by: Uros Bizjak <ubizjak@gmail.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Waiman Long <longman@redhat.com>
Link: https://lore.kernel.org/r/20240414161257.49145-1-ubizjak@gmail.com

arch/x86/include/asm/cmpxchg_32.h

@@ -64,6 +64,11 @@ static __always_inline bool __try_cmpxchg64(volatile u64 *ptr, u64 *oldp, u64 ne
 	return __arch_try_cmpxchg64(ptr, oldp, new, LOCK_PREFIX);
 }
 
+static __always_inline bool __try_cmpxchg64_local(volatile u64 *ptr, u64 *oldp, u64 new)
+{
+	return __arch_try_cmpxchg64(ptr, oldp, new,);
+}
+
 #ifdef CONFIG_X86_CMPXCHG64
 
 #define arch_cmpxchg64 __cmpxchg64
@@ -72,6 +77,8 @@ static __always_inline bool __try_cmpxchg64(volatile u64 *ptr, u64 *oldp, u64 ne
 
 #define arch_try_cmpxchg64 __try_cmpxchg64
 
+#define arch_try_cmpxchg64_local __try_cmpxchg64_local
+
 #else
 
 /*
@@ -150,6 +157,33 @@ static __always_inline bool arch_try_cmpxchg64(volatile u64 *ptr, u64 *oldp, u64
 }
 #define arch_try_cmpxchg64 arch_try_cmpxchg64
 
+#define __arch_try_cmpxchg64_emu_local(_ptr, _oldp, _new)		\
+({									\
+	union __u64_halves o = { .full = *(_oldp), },			\
+			   n = { .full = (_new), };			\
+	bool ret;							\
+									\
+	asm volatile(ALTERNATIVE("call cmpxchg8b_emu",			\
+				 "cmpxchg8b %[ptr]", X86_FEATURE_CX8)	\
+		     CC_SET(e)						\
+		     : CC_OUT(e) (ret),					\
+		       [ptr] "+m" (*(_ptr)),				\
+		       "+a" (o.low), "+d" (o.high)			\
+		     : "b" (n.low), "c" (n.high), "S" (_ptr)		\
+		     : "memory");					\
+									\
+	if (unlikely(!ret))						\
+		*(_oldp) = o.full;					\
+									\
+	likely(ret);							\
+})
+
+static __always_inline bool arch_try_cmpxchg64_local(volatile u64 *ptr, u64 *oldp, u64 new)
+{
+	return __arch_try_cmpxchg64_emu_local(ptr, oldp, new);
+}
+#define arch_try_cmpxchg64_local arch_try_cmpxchg64_local
+
 #endif
 
 #define system_has_cmpxchg64()		boot_cpu_has(X86_FEATURE_CX8)

arch/x86/include/asm/cmpxchg_64.h

@@ -20,6 +20,12 @@
 	arch_try_cmpxchg((ptr), (po), (n));				\
 })
 
+#define arch_try_cmpxchg64_local(ptr, po, n)				\
+({									\
+	BUILD_BUG_ON(sizeof(*(ptr)) != 8);				\
+	arch_try_cmpxchg_local((ptr), (po), (n));			\
+})
+
 union __u128_halves {
 	u128 full;
 	struct {