Re: [PATCH 2/4] xen/lowlevel: Implement pvop call for load_idt(sidt).

[Konrad Rzeszutek Wilk]

On Wed, Oct 17, 2012 at 04:51:17PM -0700, H. Peter Anvin wrote:
> On 10/17/2012 06:49 AM, Konrad Rzeszutek Wilk wrote:
> >In the past it used to point to 'sidt' (native_store_idt) operation
> >which is a non-privileged operation. This resulted in the
> >'struct desc_ptr' value containing the address of Xen's IDT table,
> >instead of the IDT table that Linux thinks its using. The end result
> >is that doing:
> >
> >   store_idt(&desc);
> >   load_idt(&desc);
> >
> >would blow up b/c xen_load_idt would try to parse the IDT contents
> >(desc) and de-reference a virtual address that is outside Linux's
> >__va (it is in Xen's virtual address).
> >
> >With this patch we are providing the last written IDT address.
> >
> 
> OK... this seems like another excellent set of pvops calls that
> should be nukable to Kingdom Come.  There is no reason, ever, to
> read the IDT and GDT from the kernel... the kernel already knows
> what they should be!

The code that uses these is "__save_processor_state" and 
"__restore_processor_state". To test the viability of removing
them, I did a very simple patch (see attached) and found
out that omitting those calls on AMD machines at least (hadn't
tried Intel yet) crashes the machine.

Interestingly enough, skipping the cr3, cr2, and cr0 loads in
arch/x86/kernel/acpi/wakeup_64.S worked fine!?