Re: [rft] s2ram wakeup moves to .c, could fix few machines

["H. Peter Anvin" <hpa@zytor.com>]

Rafael J. Wysocki wrote:
> On Thursday, 7 of February 2008, H. Peter Anvin wrote:
>> Rafael J. Wysocki wrote:
>>> Index: linux-2.6/arch/x86/kernel/acpi/realmode/wakeup.S
>>> ===================================================================
>>> --- /dev/null
>>> +++ linux-2.6/arch/x86/kernel/acpi/realmode/wakeup.S
>>> @@ -0,0 +1,122 @@
>>> +/*
>>> + * ACPI wakeup real mode startup stub
>>> + */
>>> +#include <asm/segment.h>
>>> +#include <asm/msr-index.h>
>>> +#include <asm/page_64.h>
>>> +#include <asm/pgtable_64.h>
>>> +
>>> +	.code16
>>> +	.section ".header", "a"
>>> +
>>> +/* This should match the structure in wakeup.h */
>>> +		.globl	wakeup_header
>>> +wakeup_header:
>>> +video_mode:	.short	0	/* Video mode number */
>>> +pmode_return:	.byte	0x66, 0xea	/* ljmpl */
>>> +		.long	0	/* offset goes here */
>>> +		.short	__KERNEL_CS
>> Missing a .short pad here... Pavel fixed that at some point, I thought.
> 
> Hm, the struct in wakeup.h doesn't contain it too.  Why exactly is it
> necessary?

Err, I guess it's only necessary if _jmp2 is declared u32.  I generally 
prefer to keep fields naturally aligned even though x86 doesn't require 
it, it's a bit of paranoia on my part.

> 
> Yeah.  Pavel, what's at 0x3f00, btw?
> 

To be fair, this might have come from my early hack, I don't know for sure.

> In fact, I'd prefer to remove .signature from the header and use the
> end_signature only, so that I can use struct wakeup_header for addressing
> the header fields in the assembly too.

I *thought* that's what I had originally.  I'm a bit confused, or it 
might have been something Pavel changed.

>> Furthermore, by doing so, you're also padding the binary out to its maximum
>> length, so you might as well just remove the .bss-clearing stuff.
> 
> Do you mean placing the header at the end will fill the area between it and the
> code with zeros, so the .bss clearing is not necessary?

Yes.  It could, of course, also be cleared by a simple memset() in the 
setup code in the kernel, as opposed by the initial assembly code. 
That's probably the preferred option.

	-hpa