On Wed, May 07, 2008 at 09:02:17AM +0200, Ingo Molnar wrote:
quoted text
> 
> * Venki Pallipadi <venkatesh.pallipadi@intel.com> wrote:
> 
> > There is a hole in mprotect, which lets the user to change the page 
> > cache type bits by-passing the kernel reserve_memtype and free_memtype 
> > wrappers. Fix the hole by not letting mprotect change the PAT bits.
> > 
> > Some versions of X used the mprotect hole to change caching type from 
> > UC to WB, so that it can then use mtrr to program WC for that region 
> > [1]. Change the mmap of pci space through /sys or /proc interfaces 
> > from UC to UC_MINUS. With this change, X will not need to use mprotect 
> > hole to get WC type.
> > 
> > [1] lkml.org/lkml/2008/4/16/369
> > 
> > Signed-off-by: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
> > Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
> > 
> > ---
> >  arch/x86/pci/i386.c       |    4 +---
> >  include/asm-x86/pgtable.h |    5 ++++-
> >  include/linux/mm.h        |    1 +
> >  mm/mmap.c                 |   13 +++++++++++++
> >  mm/mprotect.c             |    4 +++-
> >  5 files changed, 22 insertions(+), 5 deletions(-)
> 
> hm, that's one dangerous looking patch. (Cc:-ed more MM folks. I've 
> attached the patch below for reference.)
> 
> the purpose of the fix itself seems to make some sense - we dont want 
> mprotect() change the PAT bits in the pte from what they got populated 
> with at fault or mmap time.
> 
> the pte_modify() change looks correct at first sight.
> 
> The _PAGE_PROT_PRESERVE_BITS solution looks a bit ugly (although we do 
> have a couple of other similar #ifndefs in the MM already).
> 
> at minimum we should add vm_get_page_prot_preserve() as an inline 
> function to mm.h if _PAGE_PROT_PRESERVE_BITS is not defined, and make it 
> call vm_get_page_prot(). Also, vm_get_page_prot() in mm/mmap.c should 
> probably be marked inline so that we'll have only a single function call 
> [vm_get_page_prot() is trivial].

I did check that with _PAGE_PROT_PRESERVE_BITS defined to zero,
compiler optimizes vm_get_page_prot_preserve to generate same code as
vm_get_page_prot with no function call (on x86 64). So, things should be OK
from overhead perspective. But, from the code cleanliness aspect,
PRESERVE_BITS looks unclean and needs some cleanup. The reason I posted the
patch as is, was to get the confirmation on the original thread, whether this
indeed fixes those PAT related error messages.
 
quoted text
> but i'm wondering why similar issues never came up on other 
> architectures - i thought it would be rather common to have immutable 
> pte details. So maybe i'm missing something here ...

Probably other architectures does not depend on preserving things in
vma->vm_page_prot once ptes are set correctly. With PAT, we use
vm_page_prot to keep track of PAT attributes for vmas from parent to
child across fork.

pte_modify() part will be required in all archs that wants to preserve
some bits in pte in a mprotect call.

Thanks,
Venki

--
unsubscribe notice
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/