Re: [PATCH] x86: fix PSE pagetable construction

[ebiederm@xmission.com (Eric W. Biederman)]

Andi Kleen <ak@suse.de> writes:

> On Saturday 28 April 2007 19:23:55 Eric W. Biederman wrote:
>> Andi Kleen <ak@suse.de> writes:
>> 
>> >> My preference would be for whoever had:
>> >> paravirt_ops-hooks-to-set-up-initial-pagetable.patch
>> >
>> > That would be very messy because there are many followup dependencies.
>> >
>> > But once we have a good fix aggreed by everybody I can merge it into the 
>> > initial patch then.
>> 
>> I don't see Xen patches merged yet in your tree yet.
>
> Hmm, I thought you were talking about paravirt-initial-pagetable ?

Yes that is the patch with problems.

> That one is in my tree already. The Xen patches that are still outstanding
> iirc are all addons patches, as they don't impact existing code paths anymore.

Ok.

Right now Xen is the only paravirtualization solution that wants to make
page tables read only.

The current Xen methodology is to setup the identity mapping with the
initial page table set to read-only, before calling paging_init().
kernel_physical_mapping_init was modified to skip pages which are
already in the page table.

However in other cases this causes us to fail to setup important bits
like PSE and NX where we want them on the page tables.  So for the
leaf pages we need to continue setting things at least on real hardware.

The current plan was that Jeremy was going to go and modify the
Xen set_pte early in boot, to preserve to mask off _PAGE_RW in
early boot for pages that don't currently have _PAGE_RW set.

So the fix is two parts.
- Part one unconditionally set the page table entries for
  4K and pse pages.
- Fix the Xen set_pte.

Short of making the identity mapping setup paravirt specific 
I don't know what we can do.

Eric