Re: [PATCH 2.6.15] ia64: use i386 dmi_scan.c

[Andi Kleen <ak@suse.de>]

On Wednesday 18 January 2006 01:17, Bjorn Helgaas wrote:
> On Friday 13 January 2006 17:24, Bjorn Helgaas wrote:
> > ... the
> > DMI stuff crashes HP sx2000 (and probably sx1000) boxes, probably
> > because of some memory attribute problem.  So I'll have more
> > feedback after I debug that ;-)
>
> It *is* a memory attribute problem.  The current code always calls
> ioremap() on efi.smbios.  The first problem is that this is a
> physical address on x86, but a virtual address on ia64.
>
> The second problem is that we don't check the supported attributes
> for the SMBIOS table.  On HP sx1000/sx2000, these tables are in system
> memory, which doesn't support uncacheable access, so iorem
> ap() does the wrong thing.

At least on x86-64/i386 the ioremap is actually cached unless a MTRR
changes it, but it normally doesn't here. If one wants to force uncached
access one has to use ioremap_uncached(). You're saying IA64 ioremap
forces uncached access? That seems weird.

> +	if (efi_enabled) {
> +		if (efi_mem_attributes(base & EFI_MEMORY_WB)) {
> +			iomem = 0;
> +			buf = (u8 *) phys_to_virt(base);
> +		} else if (efi_mem_attributes(base & EFI_MEMORY_UC))
> +			buf = dmi_ioremap(base, len);
> +		else
> +			buf = NULL;

I would expect your ioremap to already do such a lookup. That is at least
how MTRRs on i386/x86-64 work.  If it does not how about you fix
ioremap()? Or provide a suitable ia64 dmi_ioremap, but it's likely
better to do it generally.

Regarding physical vs virtual efi.smbios -- it sounds nasty to have such
a difference in the EFI support for different architectures. Matthew, do
you have a suggestion how this can be unified?

-Andi