On Sun, 2008-10-19 at 00:32 +0200, Ingo Molnar wrote:

> the _real_ remapping in a graphics aperture happens on the GPU level 
> anyway, you manage an in-RAM GPU pagetable that just works like an 
> IOMMU, correct?

Yes, a one-level linear MMU which uses BIOS-reserved memory.

So, at least for a prototype, on 64-bit we can just use ioremap_wc and
hold the mapping while the driver is open? Is there any huge benefit to
using the kernel mapping?

> so on 32-bit we have the INVLPG TLB overhead and preemption restrictions 
> - but we knew that. We'd have to allow atomic_kmap() on non-highmem as 
> well but that's fair.

Yes, the non-highmem case is currently in fairly bad shape.

> Mind sending patches for this? :-)

I've got Venki lined up to do this work for me; once we're happy enough
with the API. In particular, the non-highmem 32-bit case seems a bit
tricky.

Also, does anyone have a better set of names for this stuff?
io_reserve_pci_mapping seems fairly ugly to me.

-- 
keith.packard@intel.com