Re: 2.6.17-rc1-mm1

["Luck, Tony" <tony.luck@intel.com>]

On Wed, Apr 05, 2006 at 10:15:34AM -0600, Bjorn Helgaas wrote:
> Huh.  Intel firmware used to just not mention the VGA framebuffer
> (0xa0000-0xc0000) at all in the EFI memory map.  I think that was
> clearly a bug.  So maybe they fixed that by marking it WB (and
> hopefully UC as well).

Nope ... not fixed (at least not in the f/w that I'm running). The
VGA buffer is still simply not mentioned in the EFI memory map.

The problem looks to come from this code in vgacon.c:

	vga_vram_base = VGA_MAP_MEM(vga_vram_base);
	vga_vram_end = VGA_MAP_MEM(vga_vram_end);
	vga_vram_size = vga_vram_end - vga_vram_base;

vga_vram_base is 0xb8000, and this call gets a UC return of
c0000000000b8000.  But vga_vram_end is 0xc0000 ... which is
the address of the start of a block of memory that is both
WB and UC capable.  So ioremap() gives us e0000000000c0000
(which means that vga_vram_size is 2000000000008000, surely
the biggest, baddest video card in the history of the world!).

Perhaps the right fix is to subtract 1 from vga_vram_end and pass
that into VGA_MAP_MEM(), and then add the 1 byte back when computing
the size?  But I don't know whether that might do something bad on
some other architecture that uses vgacon.c.  If this is not
acceptable, then we can fall back and use the Nanhai/Bjorn fix
of using ioremap_nocache().

Signed-off-by: Tony Luck <tony.luck@intel.com>

---

diff --git a/drivers/video/console/vgacon.c b/drivers/video/console/vgacon.c
index d5a04b6..4ca9877 100644
--- a/drivers/video/console/vgacon.c
+++ b/drivers/video/console/vgacon.c
@@ -484,8 +484,8 @@ #endif
 	}
 
 	vga_vram_base = VGA_MAP_MEM(vga_vram_base);
-	vga_vram_end = VGA_MAP_MEM(vga_vram_end);
-	vga_vram_size = vga_vram_end - vga_vram_base;
+	vga_vram_end = VGA_MAP_MEM(vga_vram_end - 1);
+	vga_vram_size = vga_vram_end - vga_vram_base + 1;
 
 	/*
 	 *      Find out if there is a graphics card present.