fix historic ioremap() abuse in AGP

Several AGP drivers right now use ioremap_nocache() on kernel ram in order
to turn a page of regular memory uncached.

There are two problems with this:

    1) This is a total nightmare for the ioremap() implementation to keep
       various mappings of the same page coherent.

    2) It's a total nightmare for the AGP code since it adds a ton of
       complexity in terms of keeping track of 2 different pointers to
       the same thing, in terms of error handling etc etc.

This patch fixes this by making the AGP drivers use the new
set_memory_XX APIs instead.

Note: amd-k7-agp.c is built on Alpha too, and generic.c is built
on ia64 as well, which do not yet have the set_memory_*() APIs,
so for them some we have a few ugly #ifdefs - hopefully they'll
be fixed soon.

Signed-off-by: Arjan van de Ven <arjan@linux.intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Dave Airlie <airlied@linux.ie>

arch/x86/mm/ioremap.c

@@ -126,6 +126,8 @@ static void __iomem *__ioremap(unsigned long phys_addr, unsigned long size,
 			return NULL;
 	}
 
+	WARN_ON_ONCE(page_is_ram(pfn));
+
 	switch (mode) {
 	case IOR_MODE_UNCACHED:
 	default: