The point of this patch is to remove a chipset dependency from DRM so it 
can cleanly support both the Intel 460GX and an HP IA64 chipset.  I would 
normally send this to Jeff Hartmann, but he seems to have disappeared, 
leaving agpgart without an obvious maintainer.

DRM needs to know the physical addresses of pages bound into the AGP 
aperture, and today that information is communicated from agpgart to DRM 
in the memory[] array of the agp_memory structure.  The problem is that 
the values in the array are not simple physical addresses.  Rather, they 
are the physical addresses mangled to be suitable for direct insertion 
into the GATT (the hardware-visible translation table).

Hence, DRM needs a way to unmangle these values to get the physical 
address back.  Today, agpgart exports a "page_mask" value, so DRM does the 
equivalent of this:

    paddr = agpmem->memory->memory[offset] & dev->agp->page_mask;

This isn't sufficient for the 460GX, because its mangling function 
requires a shift as well as a mask.  The 460GX code (which is still in the 
ia64 patch, not in the mainstream kernel) looks like this:

    paddr = (agpmem->memory->memory[offset] & 0xffffff) << 12;

So DRM currently has this chipset-specific code in it to unmangle the 
values in memory[], while there's no reason I know of for the mangled 
values to ever be exported from agpgart.

The attached patch against 2.4.16 remedies this by leaving the plain 
physical addresses in memory[].  The chipset-specific mangling is then 
done at the point where it is needed, when those values are inserted into 
the GATT.

It's a little tricky because the DRM code is distributed with XFree86 as 
well as with the kernel.  All the combinations (old DRM + new kernel; new 
DRM + old kernel; new DRM + new kernel) should still work, but DRM has to 
check the kernel version it's compiled against, so there are three 
KERNEL_VERSION(...)  instances in the patch that need to be adjusted to 
reflect the first version containing these changes.

Any feedback is welcome, especially if you can test it.  I've been able to 
test 460GX & ATI Radeon, HP & ATI Radeon, and Serverworks HE & MGA G400, 
but obviously that's not exhaustive.

Bjorn Helgaas
--
Linux Systems Operation R&D
Hewlett-Packard Company
Fort Collins, CO