[tip: locking/core] locking/atomic/x86: Introduce arch_sync_try_cmpxchg()

[tip: locking/core] locking/atomic/x86: Introduce arch_sync_try_cmpxchg()

[tip-bot2 for Uros Bizjak]

The following commit has been merged into the locking/core branch of tip:

Commit-ID:     636d6a8b850a2bb8462b5f743c9db9aa4d735cda
Gitweb:        https://git.kernel.org/tip/636d6a8b850a2bb8462b5f743c9db9aa4d735cda
Author:        Uros Bizjak <ubizjak@gmail.com>
AuthorDate:    Mon, 25 Sep 2023 16:55:48 +02:00
Committer:     Ingo Molnar <mingo@kernel.org>
CommitterDate: Mon, 09 Oct 2023 18:14:25 +02:00

locking/atomic/x86: Introduce arch_sync_try_cmpxchg()

Introduce the arch_sync_try_cmpxchg() macro to improve code using
sync_try_cmpxchg() locking primitive. The new definitions use existing
__raw_try_cmpxchg() macros, but use its own "lock; " prefix.

The new macros improve assembly of the cmpxchg loop in
evtchn_fifo_unmask() from drivers/xen/events/events_fifo.c from:

 57a:	85 c0                	test   %eax,%eax
 57c:	78 52                	js     5d0 <...>
 57e:	89 c1                	mov    %eax,%ecx
 580:	25 ff ff ff af       	and    $0xafffffff,%eax
 585:	c7 04 24 00 00 00 00 	movl   $0x0,(%rsp)
 58c:	81 e1 ff ff ff ef    	and    $0xefffffff,%ecx
 592:	89 4c 24 04          	mov    %ecx,0x4(%rsp)
 596:	89 44 24 08          	mov    %eax,0x8(%rsp)
 59a:	8b 74 24 08          	mov    0x8(%rsp),%esi
 59e:	8b 44 24 04          	mov    0x4(%rsp),%eax
 5a2:	f0 0f b1 32          	lock cmpxchg %esi,(%rdx)
 5a6:	89 04 24             	mov    %eax,(%rsp)
 5a9:	8b 04 24             	mov    (%rsp),%eax
 5ac:	39 c1                	cmp    %eax,%ecx
 5ae:	74 07                	je     5b7 <...>
 5b0:	a9 00 00 00 40       	test   $0x40000000,%eax
 5b5:	75 c3                	jne    57a <...>
 <...>

to:

 578:	a9 00 00 00 40       	test   $0x40000000,%eax
 57d:	74 2b                	je     5aa <...>
 57f:	85 c0                	test   %eax,%eax
 581:	78 40                	js     5c3 <...>
 583:	89 c1                	mov    %eax,%ecx
 585:	25 ff ff ff af       	and    $0xafffffff,%eax
 58a:	81 e1 ff ff ff ef    	and    $0xefffffff,%ecx
 590:	89 4c 24 04          	mov    %ecx,0x4(%rsp)
 594:	89 44 24 08          	mov    %eax,0x8(%rsp)
 598:	8b 4c 24 08          	mov    0x8(%rsp),%ecx
 59c:	8b 44 24 04          	mov    0x4(%rsp),%eax
 5a0:	f0 0f b1 0a          	lock cmpxchg %ecx,(%rdx)
 5a4:	89 44 24 04          	mov    %eax,0x4(%rsp)
 5a8:	75 30                	jne    5da <...>
 <...>
 5da:	8b 44 24 04          	mov    0x4(%rsp),%eax
 5de:	eb 98                	jmp    578 <...>

The new code removes move instructions from 585: 5a6: and 5a9:
and the compare from 5ac:. Additionally, the compiler assumes that
cmpxchg success is more probable and optimizes code flow accordingly.

Signed-off-by: Uros Bizjak <ubizjak@gmail.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: linux-kernel@vger.kernel.org
---
 arch/x86/include/asm/cmpxchg.h | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/arch/x86/include/asm/cmpxchg.h b/arch/x86/include/asm/cmpxchg.h
index d536365..5612648 100644
--- a/arch/x86/include/asm/cmpxchg.h
+++ b/arch/x86/include/asm/cmpxchg.h
@@ -221,12 +221,18 @@ extern void __add_wrong_size(void)
 #define __try_cmpxchg(ptr, pold, new, size)				\
 	__raw_try_cmpxchg((ptr), (pold), (new), (size), LOCK_PREFIX)
 
+#define __sync_try_cmpxchg(ptr, pold, new, size)			\
+	__raw_try_cmpxchg((ptr), (pold), (new), (size), "lock; ")
+
 #define __try_cmpxchg_local(ptr, pold, new, size)			\
 	__raw_try_cmpxchg((ptr), (pold), (new), (size), "")
 
 #define arch_try_cmpxchg(ptr, pold, new) 				\
 	__try_cmpxchg((ptr), (pold), (new), sizeof(*(ptr)))
 
+#define arch_sync_try_cmpxchg(ptr, pold, new) 				\
+	__sync_try_cmpxchg((ptr), (pold), (new), sizeof(*(ptr)))
+
 #define arch_try_cmpxchg_local(ptr, pold, new)				\
 	__try_cmpxchg_local((ptr), (pold), (new), sizeof(*(ptr)))