Re: [PATCH v3] x86, efi: Calling __pa() with an ioremap'd address is invalid

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

On 11/07/2011 12:23 PM, Matthew Garrett wrote:
> On Mon, Nov 07, 2011 at 03:34:48PM +0000, Matt Fleming wrote:
> 
>> After the feedback from v1 I tried to unify the efi_ioremap()
>> implementations but ran into the issue detailed in the RH bug report
>> in the changelog. Unless we teach the x86 setup code that
>> EFI_RUNTIME_SERVICES_DATA regions should be part of the direct kernel
>> mapping table (even though they're marked as E820_RESERVED) I think
>> this patch makes the most sense.
> 
> Honestly it seems like there may well be an argument for that. We're 
> talking about executable code that the kernel will be calling - it seems 
> theoretically neater for it to be added to the direct mapping. We're 
> just heavily constrained by our collapsing of the EFI memory map onto 
> the rather less fine-grained E820 one and the lack of any obvious way to 
> extend that in an OS-specific manner. I guess we could expect the 
> bootloader to conform to the standard and then re-walk the EFI memory 
> map ourselves to fix things up, but eww...
> 

Yes, I think it makes a lot of sense.  If we need to introduce new
meta-types to deal with the fact that there are EFI types that don't map
to E820, then so be it... and this is *exactly* why we want the EFI
setup stub to be part of the kernel image and not off in a separate
bootloader, requiring a stable interface...

	-hpa