rep nop question

[Manfred Spraul <manfred@colorfullife.com>]

[-- Attachment #1: Type: text/plain, Size: 675 bytes --]

The Intel documentation recommends that spinlocks should use

loop:
	rep;nop;
	cmp $0,lock_var
	jne loop

ftp://download.intel.com/design/perftool/cbts/appnotes/sse2/w_spinlock.pdf

but the linux spinlock implementation uses

loop:
	cmp $0, lock_var
	rep; nop;
	jne loop.

Why?
According to the Intel documentation, rep nop is a predefined delay that
slows down the cpu to the speed of the memory bus. The linux
implementation will delay again _after_ the lock was released.

What about the attached patch? It also adds 'rep;nop' into the rw
spinlock implementation.

It boots with my Pentium III, but the cpu doesn't support 'rep nop'
(i.e. returns immediately)

--
	Manfred

[-- Attachment #2: patch-rep-nop --]
[-- Type: text/plain, Size: 1013 bytes --]

// $Header$
// Kernel Version:
//  VERSION = 2
//  PATCHLEVEL = 4
//  SUBLEVEL = 3
//  EXTRAVERSION = -ac3
--- 2.4/arch/i386/kernel/semaphore.c	Sun Nov 19 02:31:25 2000
+++ build-2.4/arch/i386/kernel/semaphore.c	Sat Apr  7 21:26:35 2001
@@ -430,7 +430,8 @@
 .globl	__write_lock_failed
 __write_lock_failed:
 	" LOCK "addl	$" RW_LOCK_BIAS_STR ",(%eax)
-1:	cmpl	$" RW_LOCK_BIAS_STR ",(%eax)
+1:	rep; nop
+	cmpl	$" RW_LOCK_BIAS_STR ",(%eax)
 	jne	1b
 
 	" LOCK "subl	$" RW_LOCK_BIAS_STR ",(%eax)
@@ -442,7 +443,8 @@
 .globl	__read_lock_failed
 __read_lock_failed:
 	lock ; incl	(%eax)
-1:	cmpl	$1,(%eax)
+1:	rep; nop
+	cmpl	$1,(%eax)
 	js	1b
 
 	lock ; decl	(%eax)
--- 2.4/include/asm-i386/spinlock.h	Sat Apr  7 22:02:23 2001
+++ build-2.4/include/asm-i386/spinlock.h	Sat Apr  7 22:01:43 2001
@@ -58,10 +58,10 @@
 	"js 2f\n" \
 	".section .text.lock,\"ax\"\n" \
 	"2:\t" \
-	"cmpb $0,%0\n\t" \
 	"rep;nop\n\t" \
-	"jle 2b\n\t" \
-	"jmp 1b\n" \
+	"cmpb $0,%0\n\t" \
+	"jg 1b\n\t" \
+	"jmp 2b\n" \
 	".previous"
 
 /*