[PATCH] x86/i386: Check PSE bit before using PAGE_KERNEL_LARGE.

[PATCH] x86/i386: Check PSE bit before using PAGE_KERNEL_LARGE.

[Konrad Rzeszutek Wilk]

During bootup we would unconditionally do this on any
machine that was built with CONFIG_NUMA=y on i386:

setup_arch
  \-initmem_init
      \-x86_numa_init (with dummy_init as callback)
          \- init_alloc_remap
               \- set_pmd_pfn (with PAGE_PSE)

without checking to see if the CPU supports PSE. This
patch adds that and also allows the init_alloc_remap function
to properly work by falling back on PTEs.

CC: stable@kernel.org
Tested-by: William Dauchy <wdauchy@gmail.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
 arch/x86/mm/pgtable_32.c |   25 ++++++++++++++++++++++++-
 1 files changed, 24 insertions(+), 1 deletions(-)

diff --git a/arch/x86/mm/pgtable_32.c b/arch/x86/mm/pgtable_32.c
index a69bcb8..9fd4abc 100644
--- a/arch/x86/mm/pgtable_32.c
+++ b/arch/x86/mm/pgtable_32.c
@@ -86,7 +86,30 @@ void set_pmd_pfn(unsigned long vaddr, unsigned long pfn, pgprot_t flags)
 	}
 	pud = pud_offset(pgd, vaddr);
 	pmd = pmd_offset(pud, vaddr);
-	set_pmd(pmd, pfn_pmd(pfn, flags));
+
+	if ((!cpu_has_pse) && (pgprot_val(flags) & _PAGE_PSE)) {
+		pte_t *pte;
+		int i;
+
+		pgprot_val(flags) &= ~_PAGE_PSE;
+
+		/*
+		 * This is run _after_ initial memory mapped so the
+		 * PTE page are allocated - but we check it just in case.
+		 */
+		if (pmd_none(*pmd)) {
+			printk(KERN_WARNING "set_pmd_pfn: pmd_none\n");
+			return;
+		}
+
+		pte = (pte_t *)pmd_page_vaddr(*pmd);
+		for (i = 0; i < PTRS_PER_PTE; i++) {
+			set_pte(pte, pfn_pte(pfn + i, flags));
+			pte++;
+		}
+	} else
+		set_pmd(pmd, pfn_pmd(pfn, flags));
+
 	/*
 	 * It's enough to flush this one mapping.
 	 * (PGE mappings get flushed as well)
-- 
1.7.7.6

Re: [PATCH] x86/i386: Check PSE bit before using PAGE_KERNEL_LARGE.

[H. Peter Anvin]

On 05/31/2012 12:40 PM, Konrad Rzeszutek Wilk wrote:
> During bootup we would unconditionally do this on any
> machine that was built with CONFIG_NUMA=y on i386:
> 
> setup_arch
>   \-initmem_init
>       \-x86_numa_init (with dummy_init as callback)
>           \- init_alloc_remap
>                \- set_pmd_pfn (with PAGE_PSE)
> 
> without checking to see if the CPU supports PSE. This
> patch adds that and also allows the init_alloc_remap function
> to properly work by falling back on PTEs.
> 

Well, the code looks like it is PAE-specific, and PAE implies PSE.

Xen breaks that, but that is a divergence of Xen from x86.

	-hpa

Re: [PATCH] x86/i386: Check PSE bit before using PAGE_KERNEL_LARGE.

[Konrad Rzeszutek Wilk]

On Thu, May 31, 2012 at 01:21:02PM -0700, H. Peter Anvin wrote:
> On 05/31/2012 12:40 PM, Konrad Rzeszutek Wilk wrote:
> > During bootup we would unconditionally do this on any
> > machine that was built with CONFIG_NUMA=y on i386:
> > 
> > setup_arch
> >   \-initmem_init
> >       \-x86_numa_init (with dummy_init as callback)
> >           \- init_alloc_remap
> >                \- set_pmd_pfn (with PAGE_PSE)
> > 
> > without checking to see if the CPU supports PSE. This
> > patch adds that and also allows the init_alloc_remap function
> > to properly work by falling back on PTEs.
> > 
> 
> Well, the code looks like it is PAE-specific, and PAE implies PSE.

I have to double check - but I think a kernel built with CONFIG_HIGHMEM64=y
and CONFIG_NUMA=y would boot on a Pentium II which can't do PAE.

But perhaps there are some other checks that would halt the kernel
before it even got there?

> 
> Xen breaks that, but that is a divergence of Xen from x86.

It certainly does <sigh>. And that is how I spotted this - b/c the
PMD would fail (with tons of mutlicalls warnings) - and the PTE's
would still point to the old PFNs.

Re: [PATCH] x86/i386: Check PSE bit before using PAGE_KERNEL_LARGE.

[H. Peter Anvin]

No, PAE kernels cannot boot on non-PAE hardware.

Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> wrote:

>On Thu, May 31, 2012 at 01:21:02PM -0700, H. Peter Anvin wrote:
>> On 05/31/2012 12:40 PM, Konrad Rzeszutek Wilk wrote:
>> > During bootup we would unconditionally do this on any
>> > machine that was built with CONFIG_NUMA=y on i386:
>> > 
>> > setup_arch
>> >   \-initmem_init
>> >       \-x86_numa_init (with dummy_init as callback)
>> >           \- init_alloc_remap
>> >                \- set_pmd_pfn (with PAGE_PSE)
>> > 
>> > without checking to see if the CPU supports PSE. This
>> > patch adds that and also allows the init_alloc_remap function
>> > to properly work by falling back on PTEs.
>> > 
>> 
>> Well, the code looks like it is PAE-specific, and PAE implies PSE.
>
>I have to double check - but I think a kernel built with
>CONFIG_HIGHMEM64=y
>and CONFIG_NUMA=y would boot on a Pentium II which can't do PAE.
>
>But perhaps there are some other checks that would halt the kernel
>before it even got there?
>
>> 
>> Xen breaks that, but that is a divergence of Xen from x86.
>
>It certainly does <sigh>. And that is how I spotted this - b/c the
>PMD would fail (with tons of mutlicalls warnings) - and the PTE's
>would still point to the old PFNs.

-- 
Sent from my mobile phone. Please excuse brevity and lack of formatting.