[PATCH] x86: ioremap_nocache fix

[PATCH] x86: ioremap_nocache fix

[Huang, Ying]

This patch fixes a bug of ioremap_nocache. ioremap_nocache() will call
__ioremap() with flags != 0 to do the real work, which will call
change_page_attr_addr() if phys_addr + size - 1 < (end_pfn_map << PAGE_SHIFT).
But some pages between 0 ~ end_pfn_map << PAGE_SHIFT are not mapped by 
identity map, this will make change_page_attr_addr failed.

This patch is based on latest x86 git and has been tested on x86_64 platform.

Signed-off-by: Huang Ying <ying.huang@intel.com>

---
 arch/x86/mm/ioremap_64.c |    7 +++++++
 1 file changed, 7 insertions(+)

--- a/arch/x86/mm/ioremap_64.c
+++ b/arch/x86/mm/ioremap_64.c
@@ -41,8 +41,15 @@ ioremap_change_attr(unsigned long phys_a
 	if (phys_addr + size - 1 < (end_pfn_map << PAGE_SHIFT)) {
 		unsigned long npages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 		unsigned long vaddr = (unsigned long) __va(phys_addr);
+		int level;
 
 		/*
+		 * If there is no identity map for this address,
+		 * change_page_attr_addr is unnecessary
+		 */
+		if (!lookup_address(vaddr, &level))
+			return err;
+		/*
  		 * Must use a address here and not struct page because the phys addr
 		 * can be a in hole between nodes and not have an memmap entry.
 		 */

Re: [PATCH] x86: ioremap_nocache fix

[Ingo Molnar]

* Huang, Ying <ying.huang@intel.com> wrote:

> This patch fixes a bug of ioremap_nocache. ioremap_nocache() will call 
> __ioremap() with flags != 0 to do the real work, which will call 
> change_page_attr_addr() if phys_addr + size - 1 < (end_pfn_map << 
> PAGE_SHIFT). But some pages between 0 ~ end_pfn_map << PAGE_SHIFT are 
> not mapped by identity map, this will make change_page_attr_addr 
> failed.

very interesting! Is this in response to a bug you've triggered?

i believe the scenario you outlne could trigger on 64-bit boxen, if they 
try to ioremap an area not covered by the direct ptes:

> @@ -41,8 +41,15 @@ ioremap_change_attr(unsigned long phys_a
>  	if (phys_addr + size - 1 < (end_pfn_map << PAGE_SHIFT)) {
>  		unsigned long npages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
>  		unsigned long vaddr = (unsigned long) __va(phys_addr);
> +		int level;
>  
>  		/*
> +		 * If there is no identity map for this address,
> +		 * change_page_attr_addr is unnecessary
> +		 */
> +		if (!lookup_address(vaddr, &level))
> +			return err;
> +		/*

and we'd incorrectly fail the ioremap() with -EINVAL, and return a hard 
error to the driver - and causing broken boxes, right?

Thomas has unified most of ioremap*.c as well, and we'll make sure this 
fix survives the unification. The 64-bit code limped along before by 
accident, because the (pre-cleanup) 64-bit __change_page_attr() code 
incorrectly returned 0 for invalid ranges:

        kpte = lookup_address(address);
        if (!kpte) return 0;

which masked the bug you fixed now. 32-bit always returned -EINVAL on 
invalid ranges.

	Ingo

Re: [PATCH] x86: ioremap_nocache fix

[huang ying]

On Jan 25, 2008 6:27 PM, Ingo Molnar <mingo@elte.hu> wrote:
>
> * Huang, Ying <ying.huang@intel.com> wrote:
>
> > This patch fixes a bug of ioremap_nocache. ioremap_nocache() will call
> > __ioremap() with flags != 0 to do the real work, which will call
> > change_page_attr_addr() if phys_addr + size - 1 < (end_pfn_map <<
> > PAGE_SHIFT). But some pages between 0 ~ end_pfn_map << PAGE_SHIFT are
> > not mapped by identity map, this will make change_page_attr_addr
> > failed.
>
> very interesting! Is this in response to a bug you've triggered?
>
> i believe the scenario you outlne could trigger on 64-bit boxen, if they
> try to ioremap an area not covered by the direct ptes:

Yes. The efifb driver try to ioremap 0x40000000 on my 64-bit box and
it failed because 0x40000000 < (end_pfn_map << PAGE_SHIFT) and it is
not covered by direct ptes.

> > @@ -41,8 +41,15 @@ ioremap_change_attr(unsigned long phys_a
> >       if (phys_addr + size - 1 < (end_pfn_map << PAGE_SHIFT)) {
> >               unsigned long npages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
> >               unsigned long vaddr = (unsigned long) __va(phys_addr);
> > +             int level;
> >
> >               /*
> > +              * If there is no identity map for this address,
> > +              * change_page_attr_addr is unnecessary
> > +              */
> > +             if (!lookup_address(vaddr, &level))
> > +                     return err;
> > +             /*
>
> and we'd incorrectly fail the ioremap() with -EINVAL, and return a hard
> error to the driver - and causing broken boxes, right?

Yes.

> Thomas has unified most of ioremap*.c as well, and we'll make sure this
> fix survives the unification. The 64-bit code limped along before by
> accident, because the (pre-cleanup) 64-bit __change_page_attr() code
> incorrectly returned 0 for invalid ranges:
>
>         kpte = lookup_address(address);
>         if (!kpte) return 0;
>
> which masked the bug you fixed now. 32-bit always returned -EINVAL on
> invalid ranges.

Best Regards,
Huang Ying