Re: [PATCH RFC] function probe_roms accessing improper addresses

[Matthew Garrett <mjg59@srcf.ucam.org>]

On Tue, Oct 23, 2012 at 10:05:01AM -0600, Randy Wright wrote:

> On the first: one way to be compliant with such a requirement would be 
> to design systems that implement softfail in this particular region.   
> What about a system that hardfails, but on which the resulting machine
> check can be handled by the kernel machine check handler?  Would
> appropriate re-ordering of the kernel initialization code to support
> such systems be acceptable?

Good question. I don't maintain that code, so I can't really answer 
it...

> Also, let me mention a possible amendment to your first idea: what if
> the mandate that probing be supported were qualified by some attribute
> that could be indicated in the UEFI environment?  For example: instead
> of just a hole in the UEFI memory map, what if this range was
> specifically present and typed as EfiUnusableMemory?  Another idea for
> UEFI systems - but one requiring a UEFI specification change - might be
> adding a UEFI variable that if present, indicates any area not
> explicitly included and typed in the UEFI memory map (including the
> legacy adapter region) must be explicitly avoided by an OS.

Yeah, I think if it were marked unusable we could probably justify 
staying away from it.

> > 2) Don't call probe_roms() by default, but leave it up to the graphics 
> > drivers. If they can get the rom by any other means then don't call it.
> 
> One the second idea: there are a quite a lot of video drivers in the kernel
> source tree.  Do you have a suggestion for how to evaluate which ones 
> rely on the setup performed by probe_roms?

Realistically - intel, radeon and nouveau. Basically, anything that 
calls pci_map_rom() and is under drivers/gpu/drm. I'll look into a patch 
that does that.

-- 
Matthew Garrett | mjg59@srcf.ucam.org