Re: [kernel-hardening] [RFC][PATCH 3/3] mm/page_poisoning.c: Allow for zero poisoning

From: Laura Abbott
Date: Mon Jan 25 2016 - 20:33:37 EST

On 01/25/2016 02:05 PM, Kees Cook wrote:
On Mon, Jan 25, 2016 at 12:16 PM, Dave Hansen <dave.hansen@xxxxxxxxx> wrote:
Thanks for doing this! It all looks pretty straightforward.

On 01/25/2016 08:55 AM, Laura Abbott wrote:
By default, page poisoning uses a poison value (0xaa) on free. If this
is changed to 0, the page is not only sanitized but zeroing on alloc
with __GFP_ZERO can be skipped as well. The tradeoff is that detecting
corruption from the poisoning is harder to detect. This feature also
cannot be used with hibernation since pages are not guaranteed to be
zeroed after hibernation.

Ugh, that's a good point about hibernation. I'm not sure how widely it
gets used but it does look pretty widely enabled in distribution kernels.

Is this something that's fixable? It seems like we could have the
hibernation code run through and zero all the free lists. Or, we could
just disable the optimization at runtime when a hibernation is done.

We can also make hibernation run-time disabled when poisoning is used
(similar to how kASLR disables it).

I'll look into the approach kASLR uses to disable hibernation although
having the hibernation code zero the memory could be useful as well.
We can see if there are actual complaints.
Not that we _have_ to do any of this now, but if a runtime knob (like a
sysctl) could be fun too. I would be nice for folks to turn it on and
off if they wanted the added security of "real" poisoning vs. the
potential performance boost from this optimization.

+static inline bool should_zero(void)
+ !page_poisoning_enabled();

I wonder if calling this "free_pages_prezeroed()" would make things a
bit more clear when we use it in prep_new_page().

Yes that sounds much better

static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
int alloc_flags)
@@ -1401,7 +1407,7 @@ static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
kernel_map_pages(page, 1 << order, 1);
kasan_alloc_pages(page, order);

- if (gfp_flags & __GFP_ZERO)
+ if (should_zero() && gfp_flags & __GFP_ZERO)
for (i = 0; i < (1 << order); i++)
clear_highpage(page + i);

It's probably also worth pointing out that this can be a really nice
feature to have in virtual machines where memory is being deduplicated.
As it stands now, the free lists end up with gunk in them and tend not
to be easy to deduplicate. This patch would fix that.

Interesting, do you have any benchmarks I could test?

Oh, good point!