> Date: Fri, 23 Nov 2012 12:23:19 +0100
> From: Martin Pieuchot <[email protected]>
> 
> Ok, a bit of explanation first.
> 
> On macppc because the AGP chips do not translate pages, the kernel and
> the applications have access to the AGP memory regions through standard
> mappings. Because these regions are mean to share commands and data and
> the AGP bridges are not cache-coherent, they are mapped NOCACHE by the
> appleagp driver.
> 
> However in the case of the dri/drm infrastructure these memory regions
> are first allocated and mapped by the agp driver then mmap'ed by the
> userland graphic driver.

Martin, I probably should have asked this earlier, but how is this
memory mapped by the userland graphics driver?  Is that done by
opening /dev/agp and then calling mmap(2) on the filedescriptor?  If
so, perhaps a better approach would be to have the agp mmap function
return a PMAP_NC bit in the lower bits of the physical address.  This
is what sparc and sparc64 do.

Reply via email to