Re: RW and executable hole in page tables on x86_64

[Thomas Gleixner <tglx@linutronix.de>]

On Fri, 14 Nov 2014, Kees Cook wrote:
> Continuing a thread from a year ago...
> 
> On Fri, Oct 25, 2013 at 03:59:23PM +0100, H. Peter Anvin wrote:
> > On 10/25/2013 02:34 PM, Kees Cook wrote:
> > > Hi,
> > > 
> > > I've noticed there's a chunk of kernel memory still marked RW and x. See
> > > 0xffffffff82956000 below...
> > > 
> > > ---[ High Kernel Mapping ]---
> > > 0xffffffff80000000-0xffffffff81000000          16M                           pmd
> > > 0xffffffff81000000-0xffffffff81a00000          10M     ro         PSE GLB x  pmd
> > > 0xffffffff81a00000-0xffffffff81e00000           4M     ro         PSE GLB NX pmd
> > > 0xffffffff81e00000-0xffffffff82200000           4M     RW             GLB NX pte
> > > 0xffffffff82200000-0xffffffff82800000           6M     RW         PSE GLB NX pmd
> > > 0xffffffff82800000-0xffffffff82956000        1368K     RW             GLB NX pte
> > > 0xffffffff82956000-0xffffffff82a00000         680K     RW             GLB x  pte
> > > 0xffffffff82a00000-0xffffffffa0000000         470M                           pmd
> > > 
> > > HPA looked at it for a bit, but it wasn't obvious what was going on. It's
> > > after the end of bss. I do note that the two adjacent regions add up to
> > > 2MiB. Is this some kind of leftover mapping? What is this region? Is there
> > > a sensible place to clean it up?
> > > 
> > 
> > It looks to be what is left after the 2 MB page for bss is broken up.
> > It doesn't mean it isn't broken, though.
> 
> It looks like the problem still exists:
> 
> ---[ High Kernel Mapping ]---
> 0xffffffff80000000-0xffffffff9ca00000     458M                   pmd
> 0xffffffff9ca00000-0xffffffff9d200000       8M   ro   PSE GLB x  pmd
> 0xffffffff9d200000-0xffffffff9d3f3000    1996K   ro       GLB x  pte
> 0xffffffff9d3f3000-0xffffffff9d400000      52K   ro           x  pte
> 0xffffffff9d400000-0xffffffff9d600000       2M   ro   PSE GLB NX pmd
> 0xffffffff9d600000-0xffffffff9d7e8000    1952K   ro       GLB NX pte
> 0xffffffff9d7e8000-0xffffffff9d800000      96K   ro           NX pte
> 0xffffffff9d800000-0xffffffff9d8ff000    1020K   RW       GLB NX pte
> 0xffffffff9d8ff000-0xffffffff9da2d000    1208K   RW           NX pte
> 0xffffffff9da2d000-0xffffffff9dc00000    1868K   RW       GLB NX pte
> 0xffffffff9dc00000-0xffffffff9e600000      10M   RW   PSE GLB NX pmd
> 0xffffffff9e600000-0xffffffff9e7f5000    2004K   RW       GLB NX pte
> 0xffffffff9e7f5000-0xffffffff9e800000      44K   RW       GLB x  pte
> 0xffffffff9e800000-0xffffffffc0000000     536M                   pmd
> 
> Still seems to be the bss getting broken up. What is this left-over
> memory used for? Any pointers to where it happens? I'd really like to
> kill this area.

mark_rodata_ro()

	set_memory_nx(rodata_start, (all_end - rodata_start) >> PAGE_SHIFT);

all_end is _end which is the end of the the __brk section.

That looks indeed like a 2MB page which is split by set_memory_nx().

That reminder of the 2MB is a leftover from cleanup_highmap()

 * We limit the mappings to the region from _text to _brk_end.  _brk_end
 * is rounded up to the 2MB boundary.

So what you see is the reminder between _brk_end and the 2MB boundary.

Now the simple solution is to round up all_end in mark_rodata_ro() to
the 2MB boundary, but we should probably get rid of the mapping
completely.

Something like:

   end = roundup((unsigned long)all_end, PMD_SIZE) - 1;	  

   free_init_pages(all_end + 1, end)

should do the trick.

Thanks,

	tglx