Skip to contentSkip to conversation

csky: cmpxchg: Optimize with acquire & release

PATCH Superseded [email protected] 2 participants 2 messages Apr 6, 2022 – Apr 11, 2022

Subsystems (2)
Generic include/asm header files C-sky architecture

v1 of 4 Superseded Latest is v4 →

lore
Share & tools
lore

[email protected]

author maintainer
#0 lore raw

PATCH 1/1 csky: cmpxchg: Optimize with acquire & release

From: Guo Ren <[email protected]>

Optimize arch_xchg|cmpxchg|cmpxchg_local with ASM acquire|release instructions instead of previous C based.

Signed-off-by Guo Ren <[email protected]>
Signed-off-by Guo Ren <[email protected]>
---
 arch/csky/include/asm/barrier.h |   8 +-
 arch/csky/include/asm/cmpxchg.h | 173 ++++++++++++++++++++++++++++++--
 2 files changed, 172 insertions(+), 9 deletions(-)
arch/csky/include/asm/barrier.h +5 -3
index f4045dd53e17..a075f17d02dd 100644
@@ -37,6 +37,9 @@
  * bar.brar
  * bar.bwaw
  */
+#define ACQUIRE_FENCE		".long 0x8427c000\n"
+#define RELEASE_FENCE		".long 0x842ec000\n"
+
 #define __bar_brw()	asm volatile (".long 0x842cc000\n":::"memory")
 #define __bar_br()	asm volatile (".long 0x8424c000\n":::"memory")
 #define __bar_bw()	asm volatile (".long 0x8428c000\n":::"memory")

@@ -44,10 +47,10 @@
 #define __bar_ar()	asm volatile (".long 0x8421c000\n":::"memory")
 #define __bar_aw()	asm volatile (".long 0x8422c000\n":::"memory")
 #define __bar_brwarw()	asm volatile (".long 0x842fc000\n":::"memory")
-#define __bar_brarw()	asm volatile (".long 0x8427c000\n":::"memory")
+#define __bar_brarw()	asm volatile (ACQUIRE_FENCE:::"memory")
 #define __bar_bwarw()	asm volatile (".long 0x842bc000\n":::"memory")
 #define __bar_brwar()	asm volatile (".long 0x842dc000\n":::"memory")
-#define __bar_brwaw()	asm volatile (".long 0x842ec000\n":::"memory")
+#define __bar_brwaw()	asm volatile (RELEASE_FENCE:::"memory")
 #define __bar_brar()	asm volatile (".long 0x8425c000\n":::"memory")
 #define __bar_brar()	asm volatile (".long 0x8425c000\n":::"memory")
 #define __bar_bwaw()	asm volatile (".long 0x842ac000\n":::"memory")

@@ -56,7 +59,6 @@
 #define __smp_rmb()	__bar_brar()
 #define __smp_wmb()	__bar_bwaw()
 
-#define ACQUIRE_FENCE		".long 0x8427c000\n"
 #define __smp_acquire_fence()	__bar_brarw()
 #define __smp_release_fence()	__bar_brwaw()
arch/csky/include/asm/cmpxchg.h +167 -6

196 lines not on this page · raw message

Signature · 1 line
2.25.1
#1 lore raw

Hi Guo,

Quoted 22 lines · On Wed, Apr 06, 2022 at 08:40:56PM +0800, [email protected] wrote:

From: Guo Ren <[email protected]>

Optimize arch_xchg|cmpxchg|cmpxchg_local with ASM acquire|release instructions instead of previous C based.

Signed-off-by Guo Ren <[email protected]>
Signed-off-by Guo Ren <[email protected]>
---
 arch/csky/include/asm/barrier.h |   8 +-
 arch/csky/include/asm/cmpxchg.h | 173 ++++++++++++++++++++++++++++++--
 2 files changed, 172 insertions(+), 9 deletions(-)
arch/csky/include/asm/barrier.h +2 -0
index f4045dd53e17..a075f17d02dd 100644
@@ -37,6 +37,9 @@
  * bar.brar
  * bar.bwaw
  */
+#define ACQUIRE_FENCE		".long 0x8427c000\n"
+#define RELEASE_FENCE		".long 0x842ec000\n"

[...]

Quoted 8 lines
arch/csky/include/asm/barrier.h +8 -0
+#define __xchg(new, ptr, size)					\
+({								\
+	__typeof__(ptr) __ptr = (ptr);				\
+	__typeof__(new) __new = (new);				\
+	__typeof__(*(ptr)) __ret;				\
+	unsigned long tmp;					\
+	switch (size) {						\
+	case 4:							\

arch/csky/include/asm/barrier.h

Quoted 10 lines
arch/csky/include/asm/barrier.h +10 -0
+		asm volatile (					\
+		"1:	ldex.w		%0, (%3) \n"		\
+		ACQUIRE_FENCE					\
+		"	mov		%1, %2   \n"		\
+		RELEASE_FENCE					\
+		"	stex.w		%1, (%3) \n"		\
+		"	bez		%1, 1b   \n"		\
+			: "=&r" (__ret), "=&r" (tmp)		\
+			: "r" (__new), "r"(__ptr)		\
+			:);					\

As with the riscv case, I believe this suffers the problem described in:

8e86f0b409a44193 ("arm64: atomics: fix use of acquire + release for full barrier semantics")

... and does not provide FULL ordering semantics.

[...]

Quoted 12 lines
arch/csky/include/asm/barrier.h +12 -0
+#define __cmpxchg(ptr, old, new, size)				\
+({								\
+	__typeof__(ptr) __ptr = (ptr);				\
+	__typeof__(new) __new = (new);				\
+	__typeof__(new) __tmp;					\
+	__typeof__(old) __old = (old);				\
+	__typeof__(*(ptr)) __ret;				\
+	switch (size) {						\
+	case 4:							\
+		asm volatile (					\
+		"1:	ldex.w		%0, (%3) \n"		\
+		ACQUIRE_FENCE					\

arch/csky/include/asm/barrier.h

Quoted 10 lines
arch/csky/include/asm/barrier.h +10 -0
+		"	cmpne		%0, %4   \n"		\
+		"	bt		2f       \n"		\
+		"	mov		%1, %2   \n"		\
+		RELEASE_FENCE					\
+		"	stex.w		%1, (%3) \n"		\
+		"	bez		%1, 1b   \n"		\
+		"2:				 \n"		\
+			: "=&r" (__ret), "=&r" (__tmp)		\
+			: "r" (__new), "r"(__ptr), "r"(__old)	\
+			:);					\

Likewise.

Thanks,
Mark.