radeon memory not detected correctly with DRI (pci quirk?)

[Benjamin Herrenschmidt]

On Sat, 2006-03-18 at 22:23 -0500, Brian Hinz wrote:

> seemed to be the cause of the problem (although I couldn't see why).  By 
> patching a 2.6.15 kernel to use the agp_bridge->driver->mask_memory, rather 
> than the newer virt_to_gart I was able to start X (7.0 modular) with DRI. 
> Afterwards I noticed that when if I use a framebuffer like radeonfb or 
> vesafb-tng, I get MTRR overlaps.  /proc/mtrr shows that the overlaps are due 
> to DRI (incorrectly) detecting 128Mb of video memory, while radeonfb finds 
> the correct amount (16Mb).  lspci also reports 128Mb however, so it appears 
> that this is really an issue with the PCI registers being incorrectly 
> initialized.  Just to be sure, I hacked the radeon_cp.c file from the DRI 
> code from CVS and replaced the call to pci_get_resource_len with the numeric 
> value of 16Mb.  As expected, the mtrr overlaps are gone.  So what's the 
> proper way to handle this if the hardware or the bios is really incorrect? 
> Add an entry into the kernel quirks.c?  If those registers are set by the 
> bios, can it be overridden with a custom DSDT block (I did try re-compiling 
> the kernel with pci access mode=bios, no difference though)?  I haven't 
> tried re-compiling the kernel with virt_to_gart and the hardcoded value of 
> the fb size in the DRI code together, I don't really understand the kernel 
> agp code well enough to see if that's really where the problem lies.  If 
> anyone can shed some light on this or offer advice on how to proceed it 
> would be appreciated. Also, the machine is a Transmeta efficeon based 
> notebook (sharp mm20) with a mobility M6 LY chip, and I'm using cvs versions 
> of both the xf86-video-ati and drm.  A complete description of the problem 
> can be found at:
> http://bugzilla.kernel.org/show_bug.cgi?id=4513

I'll reply on that bug report.

Ben.