i386 and PAE: pud_present()

i386 and PAE: pud_present()

[Magnus Damm]

Hi guys,

In file include/asm-i386/pgtable-3level.h:

On i386 with PAE enabled, shouldn't pud_present() return (pud_val(pud)
& _PAGE_PRESENT) instead of constant 1?

Today pud_present() returns constant 1 regardless of PAE or not. This
looks wrong to me, but maybe I'm misunderstanding how to fold the page
tables... =)

Thanks,

/ magnus

Re: i386 and PAE: pud_present()

[Nick Piggin]

Magnus Damm wrote:
> Hi guys,
> 
> In file include/asm-i386/pgtable-3level.h:
> 
> On i386 with PAE enabled, shouldn't pud_present() return (pud_val(pud)
> & _PAGE_PRESENT) instead of constant 1?
> 
> Today pud_present() returns constant 1 regardless of PAE or not. This
> looks wrong to me, but maybe I'm misunderstanding how to fold the page
> tables... =)

Take a look a little further down the page for the comment.

In i386 + PAE, pud is always present.

-- 
SUSE Labs, Novell Inc.
Send instant messages to your online friends http://au.messenger.yahoo.com 

Re: i386 and PAE: pud_present()

[Andi Kleen]

On Friday 28 April 2006 09:54, Nick Piggin wrote:
> Magnus Damm wrote:
> > Hi guys,
> > 
> > In file include/asm-i386/pgtable-3level.h:
> > 
> > On i386 with PAE enabled, shouldn't pud_present() return (pud_val(pud)
> > & _PAGE_PRESENT) instead of constant 1?
> > 
> > Today pud_present() returns constant 1 regardless of PAE or not. This
> > looks wrong to me, but maybe I'm misunderstanding how to fold the page
> > tables... =)
> 
> Take a look a little further down the page for the comment.
> 
> In i386 + PAE, pud is always present.

I think his problem is that the PGD is always present too (in pgtables-nopud.h)
Indeed looks strange.

-Andi

Re: i386 and PAE: pud_present()

[Dave McCracken]

--On Friday, April 28, 2006 10:27:21 +0200 Andi Kleen <ak@suse.de> wrote:

>> Take a look a little further down the page for the comment.
>> 
>> In i386 + PAE, pud is always present.
> 
> I think his problem is that the PGD is always present too (in
> pgtables-nopud.h) Indeed looks strange.

The PGD is always fully populated on i386 if PAE is enabled.  All three of
the pmd pages are allocated at page table creation time and persist till
the page table is deleted.

Dave McCracken

Re: i386 and PAE: pud_present()

[Andi Kleen]

On Sunday 30 April 2006 08:07, Dave McCracken wrote:
> 
> --On Friday, April 28, 2006 10:27:21 +0200 Andi Kleen <ak@suse.de> wrote:
> 
> >> Take a look a little further down the page for the comment.
> >> 
> >> In i386 + PAE, pud is always present.
> > 
> > I think his problem is that the PGD is always present too (in
> > pgtables-nopud.h) Indeed looks strange.
> 
> The PGD is always fully populated on i386 if PAE is enabled.  All three of
> the pmd pages are allocated at page table creation time and persist till
> the page table is deleted.

At least with the new flexmmap it sounds like a waste of memory.

-Andi

Re: i386 and PAE: pud_present()

[Magnus Damm]

On 4/30/06, Dave McCracken <dmccr@us.ibm.com> wrote:
>
> --On Friday, April 28, 2006 10:27:21 +0200 Andi Kleen <ak@suse.de> wrote:
>
> >> Take a look a little further down the page for the comment.
> >>
> >> In i386 + PAE, pud is always present.
> >
> > I think his problem is that the PGD is always present too (in
> > pgtables-nopud.h) Indeed looks strange.
>
> The PGD is always fully populated on i386 if PAE is enabled.  All three of
> the pmd pages are allocated at page table creation time and persist till
> the page table is deleted.

The following code snippet is from some kexec patch of mine. The
function is used to build a new set of page tables which are used when
jumping to the new kernel.

The code should be pretty archtecture independent - the same code
works on x86 and x86_64. And x86/PAE with a workaround.

#ifdef CONFIG_X86_PAE
#undef pud_present
#define pud_present(pud) (pud_val(pud) & _PAGE_PRESENT)
#endif

#define pa_page(page) __pa(page_address(page))

static int create_mapping(struct page *root, struct page **pages,
			  unsigned long va, unsigned long pa)
{
	pgd_t *pgd;
	pud_t *pud;
	pmd_t *pmd;
	pte_t *pte;
	int k = 0;

	pgd = (pgd_t *)page_address(root) + pgd_index(va);
	if (!pgd_present(*pgd))
		set_pgd(pgd, __pgd(pa_page(pages[k++]) | _KERNPG_TABLE));

	pud = pud_offset(pgd, va);
	if (!pud_present(*pud))
		set_pud(pud, __pud(pa_page(pages[k++]) | _KERNPG_TABLE));

	pmd = pmd_offset(pud, va);
	if (!pmd_present(*pmd))
		set_pmd(pmd, __pmd(pa_page(pages[k++]) | _KERNPG_TABLE));

	pte = (pte_t *)page_address(pmd_page(*pmd)) + pte_index(va);
	set_pte(pte, __pte(pa | _PAGE_KERNEL_EXEC));

	return k;
}

Any comments?

/ magnus