Re: [Qemu-devel] [PULL for-1.8 1/2] pc: disable pci-info

[Michael S. Tsirkin]

On Tue, Nov 26, 2013 at 04:42:16PM +0100, Gerd Hoffmann wrote:
>   Hi,
> 
> > This doesn't clamp the w32.begin value into [0x80000000, 0xe0000000],
> > which seems wrong.
> 
> Why?  In a 1G guest you can map pci bars at 0x40000000 just fine.
> 
> _CRS in acpi should declare the area where you can map pci bars, and
> that happens to be end-of-ram -> ioapci base.
> 
> Firmware can choose to use a smaller area.  Both seabios and ovmf will
> not map anyting below 0x80000000.  That is just fine.  As long as all
> pci bars get mapped inside the region declared in _CRS things will work.
> 
> While thinking about it:  There is one possible reason to not do it:
> Guest bugs.  IIRC windows doesn't like the 64bit pci window being larger
> than 2G.
>  Possibly that is an issue with the 32bit window too.  If that
> is the case we should indeed not use w32.begin values smaller than
> 0x8000000.  Igor, any clue?

I think the issues are all around 64 bit.
It's not hard to test.

> > Guys, I'm confused and annoyed out of my brains by the divergence here.
> > In my perception Michael, Igor, and Gerd are all proposing different
> > things, and my ideas are either rejected or ignored (even though they
> > occasionally overlap with ideas from others).
> 
> Hmm, as far as I can see there is an agreement that it is qemu's fault,
> the acpi tables need fixing, and ovmf should not need changes any
> changes.
> 
> Mimicing seabios/qemu logic in ovmf can be used as temporary stopgap
> while details are sorted on the qemu side.
> 
> cheers,
>   Gerd
>