Re: 2.6.11-rc2-mm1 strange messages

[Andrew Morton <akpm@osdl.org>]

Norbert Preining <preining@logic.at> wrote:
>
> On Don, 24 Feb 2005, Andrew Morton wrote:
> > What does the stack backtrace from iounmap-debugging.patch say?
> 
> iounmap: bad address c00fffd9
>  [<c03f8430>] trap_init+0x30/0x190
>  [<c03f2697>] start_kernel+0x47/0x1c0

ah hah.

trap_init() does:

	void __iomem *p = ioremap(0x0FFFD9, 4);

which returns phys_to_virt(0x0FFFD9) = 0xc00fffd9

then trap_init() does:

	iounmap(p);

and iounmap() does

	if ((void __force *) addr <= high_memory) 
		return; 

which doesn't work, because 0xc00fffd9 is outside 0 ... high_memory.

A quick fix is to delete that iounmap() call from trap_init(), because we
"know" how ioremap() works.  Or, better, simply use phys_to_virt(0x0FFFD9)
in trap_init().

Although a better fix might be to make __iounmap() behave symmetrically:

	if ((long)addr >= phys_to_virt(0xA0000) &&
			(long)addr < phys_to_virt(0x100000))
		return;

but that's not quite right, because we're assuming that the range to be
unmapped is wholly within the PCI/ISA region.  Without a VMA there just
isn't enough info to determine that.

Does anyone have any preferences?