Re: Suspend-to-RAM on Sony Vaios

["Emmanuel Thomé" <Emmanuel.Thome-MZpvjPyXg2s@public.gmane.org>]

[-- Attachment #1: Type: text/plain, Size: 3002 bytes --]


On Mon, Nov 01, 2004 at 04:41:28PM +0100, Pavel Machek wrote:
> Patch looks good to me...

Great. Here's an updated version.

> Uff, this looks pretty ugly. But I guess it is i386 being ugly...

The present version does the indirection once and for all, it's more
readable. I've documented a bit the segment descriptor stuff, so that
it's not so mystifying. I can do most of it in C, using asm/processor.h
and asm/desc.h if you prefer.

Ole, while digging up what happens with your real mode / pmode switch, I
discovered that it turns out you're _not_ initializing the segment
descriptor for the data segment. It does not matter too much,  since the
wakeup code (once in real mode) hooks ds to the same place as cs anyway,
and sets up the stack accordingly, but is there a reason for doing so ?
I've initialized the descriptor, to make things more clear.

In fact, I hoped that this was a key explaining vgabios crashes: the es
segment point into nowhere (the wakeup code does not touch it), but maybe
the ROM code uses it (who knows). Unfortunately, movw %ax, %es does not
improve things.

I would say, still, that adding movw %ax, %es on top of wakeup_start
would make sense, perhaps some machine where s3_{,late_}{bios,mode}
almost work would benefit from it. Do you agree ?

> > +	if (((pci_dev->class ^ (PCI_CLASS_DISPLAY_VGA << 8)) & 0xffff00)==0)
> The use of xor here is certainly interesting. Is it possible to say
> 	pci_dev->class & 0xffff00 == PCI_CLASS_DISPLAY_VGA << 8

;-) This one got me laughing as well. I do prefer the (a & b) == c idiom
(now used), but I was amused by this xor in pci_match_one_device.

> > [ sleep_hacks.h ]
> Does it really deserve its own header file? Hmm and it should probably
> be common between x86-64 and i386, as we might need 64-bit version in
> future...

Well, the problem is that this stuff has to be included from the .S file
as well.

I've moved it into asm/acpi.h instead, which was the place where I
intended to put it at first. This implies protecting most of asm/acpi.h
with #ifndef __ASSEMBLY__ , but this makes sense for asm/ includes,
doesn't it ? Do you prefer this location ?

Attached are three files.

s3_late_bios.patch.gz is the current, working patch for i386.

s3_late_bios_radeon.patch.gz puts an acpi_vgapost call into the radeonfb
driver, just to illustrate how it can be inserted into custom pci resume
functions (oh, this will change in 2.6.10 with the removal of
pci_restore_state's second arg...).

s3_late_bios-x86_64-draft.patch.gz is untested. It lays the ground for
doing the same for x86_64 , but currently I don't have access to amd64
boxes where I can play with acpi (you wouldn't play such games on a
department level server, would you ?).


I've read on this ml that this stuff crashes S4 resume, understandably.
So now, calls to acpi_vgapost are protected by (pci_dev->dev.power_state
!= 4) checks, as seen elsewhere. Is .power_state==4 effectively used for
marking S4 sleep ?

Hoping this version is OK for you,

E.

[-- Attachment #2: s3_late_bios.patch.gz --]
[-- Type: application/x-gzip, Size: 3633 bytes --]

[-- Attachment #3: s3_late_bios_radeon.patch.gz --]
[-- Type: application/x-gzip, Size: 342 bytes --]

[-- Attachment #4: s3_late_bios-x86_64-draft.patch.gz --]
[-- Type: application/x-gzip, Size: 987 bytes --]