Re: [PATCH v3 08/11] slub: Replace cmpxchg_double()

[Peter Zijlstra]

On Tue, May 30, 2023 at 04:22:32PM +0200, Peter Zijlstra wrote:

> Yet another alternative is using a struct type and an equality function,
> just for this.

The best I could come up with in the regard is the below. It builds on
HPPA64 and x86_64, but I've not ran it yet.

(also, the introduction of this_cpu_try_cmpxchg() should probably be
split out into its own patch)

--- a/include/asm-generic/percpu.h
+++ b/include/asm-generic/percpu.h
@@ -99,6 +99,15 @@ do {									\
 	__ret;								\
 })
 
+#define raw_cpu_generic_try_cmpxchg(pcp, ovalp, nval)			\
+({									\
+	typeof(pcp) __ret, __old = *(ovalp);				\
+	__ret = raw_cpu_cmpxchg(pcp, __old, nval);			\
+	if (!likely(__ret == __old))					\
+		*(ovalp) = __ret;					\
+	likely(__ret == __old);						\
+})
+
 #define __this_cpu_generic_read_nopreempt(pcp)				\
 ({									\
 	typeof(pcp) ___ret;						\
@@ -167,6 +176,15 @@ do {									\
 	__ret;								\
 })
 
+#define this_cpu_generic_try_cmpxchg(pcp, ovalp, nval)			\
+({									\
+	typeof(pcp) __ret, __old = *(ovalp);				\
+	__ret = this_cpu_cmpxchg(pcp, __old, nval);			\
+	if (!likely(__ret == __old))					\
+		*(ovalp) = __ret;					\
+	likely(__ret == __old);						\
+})
+
 #ifndef raw_cpu_read_1
 #define raw_cpu_read_1(pcp)		raw_cpu_generic_read(pcp)
 #endif
@@ -258,6 +276,36 @@ do {									\
 #define raw_cpu_xchg_8(pcp, nval)	raw_cpu_generic_xchg(pcp, nval)
 #endif
 
+#ifndef __SIZEOF_INT128__
+#define raw_cpu_generic_try_cmpxchg_memcmp(pcp, ovalp, nval)		\
+({									\
+	typeof(pcp) *__p = raw_cpu_ptr(&(pcp));				\
+	typeof(pcp) __ret, __old = *(ovalp);				\
+	bool __s;							\
+	__ret = *__p;							\
+	if (!__builtin_memcmp(&__ret, &__old, sizeof(pcp))) {		\
+		*__p = nval;						\
+		__s = true;						\
+	} else {							\
+		*(ovalp) = __ret;					\
+		__s = false;						\
+	}								\
+	__s;								\
+})
+
+#define raw_cpu_generic_cmpxchg_memcmp(pcp, oval, nval)			\
+({									\
+	typeof(pcp) __old = (oval);					\
+	raw_cpu_generic_try_cmpxchg_memcpy(pcp, &__old, nval);		\
+	__old;								\
+})
+
+#define raw_cpu_cmpxchg128(pcp, oval, nval) \
+	raw_cpu_generic_cmpxchg_memcmp(pcp, oval, nval)
+#define raw_cpu_try_cmpxchg128(pcp, ovalp, nval) \
+	raw_cpu_generic_try_cmpxchg_memcmp(pcp, ovalp, nval)
+#endif
+
 #ifndef raw_cpu_cmpxchg_1
 #define raw_cpu_cmpxchg_1(pcp, oval, nval) \
 	raw_cpu_generic_cmpxchg(pcp, oval, nval)
@@ -283,6 +331,31 @@ do {									\
 	raw_cpu_generic_cmpxchg(pcp, oval, nval)
 #endif
 
+#ifndef raw_cpu_try_cmpxchg_1
+#define raw_cpu_try_cmpxchg_1(pcp, ovalp, nval) \
+	raw_cpu_generic_try_cmpxchg(pcp, ovalp, nval)
+#endif
+#ifndef raw_cpu_try_cmpxchg_2
+#define raw_cpu_try_cmpxchg_2(pcp, ovalp, nval) \
+	raw_cpu_generic_try_cmpxchg(pcp, ovalp, nval)
+#endif
+#ifndef raw_cpu_try_cmpxchg_4
+#define raw_cpu_try_cmpxchg_4(pcp, ovalp, nval) \
+	raw_cpu_generic_try_cmpxchg(pcp, ovalp, nval)
+#endif
+#ifndef raw_cpu_try_cmpxchg_8
+#define raw_cpu_try_cmpxchg_8(pcp, ovalp, nval) \
+	raw_cpu_generic_try_cmpxchg(pcp, ovalp, nval)
+#endif
+#ifndef raw_cpu_try_cmpxchg64
+#define raw_cpu_try_cmpxchg64(pcp, ovalp, nval) \
+	raw_cpu_generic_try_cmpxchg(pcp, ovalp, nval)
+#endif
+#ifndef raw_cpu_try_cmpxchg128
+#define raw_cpu_try_cmpxchg128(pcp, ovalp, nval) \
+	raw_cpu_generic_try_cmpxchg(pcp, ovalp, nval)
+#endif
+
 #ifndef this_cpu_read_1
 #define this_cpu_read_1(pcp)		this_cpu_generic_read(pcp)
 #endif
@@ -374,6 +447,33 @@ do {									\
 #define this_cpu_xchg_8(pcp, nval)	this_cpu_generic_xchg(pcp, nval)
 #endif
 
+#ifndef __SIZEOF_INT128__
+#define this_cpu_generic_try_cmpxchg_memcmp(pcp, ovalp, nval)		\
+({									\
+ 	bool __ret;							\
+	unsigned long __flags;						\
+	raw_local_irq_save(__flags);					\
+	__ret = raw_cpu_generic_try_cmpxchg_memcmp(pcp, ovalp, nval);	\
+	raw_local_irq_restore(__flags);					\
+	__ret;								\
+})
+
+#define this_cpu_generic_cmpxchg_memcmp(pcp, oval, nval)		\
+({									\
+	typeof(pcp) __ret;						\
+	unsigned long __flags;						\
+	raw_local_irq_save(__flags);					\
+	__ret = raw_cpu_generic_cmpxchg_memcmp(pcp, oval, nval);	\
+	raw_local_irq_restore(__flags);					\
+	__ret;								\
+})
+
+#define this_cpu_cmpxchg128(pcp, oval, nval) \
+	this_cpu_generic_cmpxchg_memcmp(pcp, oval, nval)
+#define this_cpu_try_cmpxchg128(pcp, ovalp, nval) \
+	this_cpu_generic_try_cmpxchg_memcmp(pcp, ovalp, nval)
+#endif
+
 #ifndef this_cpu_cmpxchg_1
 #define this_cpu_cmpxchg_1(pcp, oval, nval) \
 	this_cpu_generic_cmpxchg(pcp, oval, nval)
@@ -399,4 +499,29 @@ do {									\
 	this_cpu_generic_cmpxchg(pcp, oval, nval)
 #endif
 
+#ifndef this_cpu_try_cmpxchg_1
+#define this_cpu_try_cmpxchg_1(pcp, ovalp, nval) \
+	this_cpu_generic_try_cmpxchg(pcp, ovalp, nval)
+#endif
+#ifndef this_cpu_try_cmpxchg_2
+#define this_cpu_try_cmpxchg_2(pcp, ovalp, nval) \
+	this_cpu_generic_try_cmpxchg(pcp, ovalp, nval)
+#endif
+#ifndef this_cpu_try_cmpxchg_4
+#define this_cpu_try_cmpxchg_4(pcp, ovalp, nval) \
+	this_cpu_generic_try_cmpxchg(pcp, ovalp, nval)
+#endif
+#ifndef this_cpu_try_cmpxchg_8
+#define this_cpu_try_cmpxchg_8(pcp, ovalp, nval) \
+	this_cpu_generic_try_cmpxchg(pcp, ovalp, nval)
+#endif
+#ifndef this_cpu_try_cmpxchg64
+#define this_cpu_try_cmpxchg64(pcp, ovalp, nval) \
+	this_cpu_generic_try_cmpxchg(pcp, ovalp, nval)
+#endif
+#ifndef this_cpu_try_cmpxchg128
+#define this_cpu_try_cmpxchg128(pcp, ovalp, nval) \
+	this_cpu_generic_try_cmpxchg(pcp, ovalp, nval)
+#endif
+
 #endif /* _ASM_GENERIC_PERCPU_H_ */
--- a/include/linux/types.h
+++ b/include/linux/types.h
@@ -13,6 +13,13 @@
 #ifdef __SIZEOF_INT128__
 typedef __s128 s128;
 typedef __u128 u128;
+#else
+#ifdef CONFIG_64BIT
+/* hack for this_cpu_cmpxchg128 */
+typedef struct {
+	u64 a, b;
+} u128 __attribute__((aligned(16)));
+#endif
 #endif
 
 typedef u32 __kernel_dev_t;
--- a/mm/slab.h
+++ b/mm/slab.h
@@ -11,14 +11,14 @@ void __init kmem_cache_init(void);
 # define system_has_freelist_aba()	system_has_cmpxchg128()
 # define try_cmpxchg_freelist		try_cmpxchg128
 # endif
-#define this_cpu_cmpxchg_freelist	this_cpu_cmpxchg128
+#define this_cpu_try_cmpxchg_freelist	this_cpu_try_cmpxchg128
 typedef u128 freelist_full_t;
 #else /* CONFIG_64BIT */
 # ifdef system_has_cmpxchg64
 # define system_has_freelist_aba()	system_has_cmpxchg64()
 # define try_cmpxchg_freelist		try_cmpxchg64
 # endif
-#define this_cpu_cmpxchg_freelist	this_cpu_cmpxchg64
+#define this_cpu_try_cmpxchg_freelist	this_cpu_try_cmpxchg64
 typedef u64 freelist_full_t;
 #endif /* CONFIG_64BIT */
 
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -3037,8 +3037,8 @@ __update_cpu_freelist_fast(struct kmem_c
 	freelist_aba_t old = { .freelist = freelist_old, .counter = tid };
 	freelist_aba_t new = { .freelist = freelist_new, .counter = next_tid(tid) };
 
-	return this_cpu_cmpxchg_freelist(s->cpu_slab->freelist_tid.full,
-					 old.full, new.full) == old.full;
+	return this_cpu_try_cmpxchg_freelist(s->cpu_slab->freelist_tid.full,
+					     &old.full, new.full);
 }
 
 /*