[PATCH] Generic ioremap_page_range: flush_cache_vmap

The existing implementation of ioremap_page_range(), which was taken
from i386, does this:

flush_cache_all();
/* modify page tables */
flush_tlb_all();

I think this is a bit defensive, so this patch changes the generic
implementation to do:

/* modify page tables */
flush_cache_vmap(start, end);

instead, which is similar to what vmalloc() does. This should still
be correct because we never modify existing PTEs. According to
James Bottomley:

The problem the flush_tlb_all() is trying to solve is to avoid stale tlb
entries in the ioremap area. We're just being conservative by flushing
on both map and unmap. Technically what vmalloc/vfree does (only flush
the tlb on unmap) is just fine because it means that the only tlb
entries in the remap area must belong to in-use mappings.

Signed-off-by: Haavard Skinnemoen <[e-mail]>
Cc: Richard Henderson <[e-mail]>
Cc: Ivan Kokshaysky <[e-mail]>
Cc: Russell King <[e-mail]>
Cc: Mikael Starvik <[e-mail]>
Cc: Andi Kleen <[e-mail]>
Cc: <[e-mail]>
Cc: Ralf Baechle <[e-mail]>
Cc: Kyle McMartin <[e-mail]>
Cc: Martin Schwidefsky <[e-mail]>
Cc: Paul Mundt <[e-mail]>
Signed-off-by: Andrew Morton <[e-mail]>
Signed-off-by: Linus Torvalds <[e-mail]>

lib/ioremap.c

···
 
 	BUG_ON(addr >= end);
 
-	flush_cache_all();
-
 	start = addr;
 	phys_addr -= addr;
 	pgd = pgd_offset_k(addr);
···
 			break;
 	} while (pgd++, addr = next, addr != end);
 
-	flush_tlb_all();
+	flush_cache_vmap(start, end);
 
 	return err;
 }