MTRR vs. PAT and Intel Graphics

[Keith Whitwell]

Bruce Perens wrote:
> Keith and Co.,
> 
> In the aftermath of my investigation of i965G performance dropping 
> severely on my Gigabyte motherboard when a particular fan setting was 
> configured, I found that this is ultimately attributable to the fact 
> that Windows uses the more versatile Page Attribute Tables while Linux 
> is still using MTRRs alone for the same purpose. And for this reason we 
> are probably not going to get much help on the issue from BIOS and 
> motherboard manufacturers. PCI Express systems seem to have made the 
> situation worse in that memory-mapped I/O regions are in general less 
> contiguous. nVidia, Jeff Hartmann, and others have been trying to get 
> PAT into the kernel since at least 2004, the latest patch to do so 
> appears to be a year old and didn't quite meet approval. More 
> information is at http://technocrat.net/d/2007/6/1/20838
> 
> Since this problem can severely effect Intel graphics performance on 
> Linux systems while leaving Windows unscathed, I think it would be 
> desirable for Intel to involve itself in this issue until PAT is 
> successfully incorporated into mainstream kernels. Keith, can you help 
> with this?

I spoke with Thomas about this patch at one point.  His belief was that 
PAT support can be more simply supported by directly manipulating page 
attributes from the drm/ttm memory manager - this could be as little as 
a few more functions worth of code or perhaps a new small file's worth.

If you're not familiar with it, this is a good overview:

http://www.tungstengraphics.com/mm.pdf

Keith