Re: [PATCH 8/8] x86, cleanups: Simplify sync_core() in the case of no CPUID

[Borislav Petkov <bp@alien8.de>]

On Wed, Nov 28, 2012 at 04:14:43PM -0800, H. Peter Anvin wrote:
> Wrong barrier semantics.

Let me try to understand what you mean by that :)

Now, your version's asm output looks like this:

----
	.loc 2 95 0
	movl	$139, %esi	#, tmp140
	xorl	%eax, %eax	# tmp141
	movl	%esi, %ecx	# tmp140,
	movl	%eax, %edx	# tmp141,
#APP
# 95 "/home/boris/kernel/linux-2.6/arch/x86/include/asm/msr.h" 1
	wrmsr
# 0 "" 2
#NO_APP
	.loc 3 694 0
	movb	$1, %al	#,
#APP
# 694 "/home/boris/kernel/linux-2.6/arch/x86/include/asm/processor.h" 1
	cpuid
# 0 "" 2
.LVL15:
#NO_APP
----

This is the sync_core() call from early_init_intel(). Now look like 139
remains in %ecx after the WRMSR and CPUID actually gets called with
RAX=1 and RCX=139 (btw, 139 is MSR_IA32_UCODE_REV). Even if this works
I'd say, we don't want to have any stray values in RCX when doing CPUID,
no?

Now here's the version with the change I proposed:

----
	.loc 2 95 0
	xorl	%esi, %esi	# tmp144
	movl	$139, %edi	#, tmp143
	movl	%edi, %ecx	# tmp143,
	movl	%esi, %eax	# tmp144,
	movl	%esi, %edx	# tmp144,
#APP
# 95 "/home/boris/kernel/linux-2.6/arch/x86/include/asm/msr.h" 1
	wrmsr
# 0 "" 2
.LVL15:
#NO_APP
	.loc 3 199 0
	movb	$1, %al	#,
	movl	%esi, %ecx	# tmp144, ecx
#APP
# 199 "/home/boris/kernel/linux-2.6/arch/x86/include/asm/processor.h" 1
	cpuid
# 0 "" 2
.LVL16:
#NO_APP
----

RCX gets correctly cleaned to 0 and *then* we call CPUID.

And the asm output is the same except that RCX gets correctly cleaned up
before calling CPUID.

So what am I missing?

Thanks.

-- 
Regards/Gruss,
    Boris.