[RFC v1] mm: SLAB freelist randomization

[Thomas Garnier]

Provide an optional config (CONFIG_FREELIST_RANDOM) to randomize the
SLAB freelist. This security feature reduces the predictability of
the kernel slab allocator against heap overflows.

Randomized lists are pre-computed using a Fisher-Yates shuffle and
re-used on slab creation for performance.
---
Based on next-20160405
---
init/Kconfig | 9 ++++
mm/slab.c | 155 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
2 files changed, 164 insertions(+)

diff --git a/init/Kconfig b/init/Kconfig
index 0dfd09d..ee35418 100644
--- a/init/Kconfig
+++ b/init/Kconfig
@@ -1742,6 +1742,15 @@ config SLOB

endchoice

+config FREELIST_RANDOM
+ default n
+ depends on SLAB
+ bool "SLAB freelist randomization"
+ help
+ Randomizes the freelist order used on creating new SLABs. This
+ security feature reduces the predictability of the kernel slab
+ allocator against heap overflows.
+
config SLUB_CPU_PARTIAL
default y
depends on SLUB && SMP
diff --git a/mm/slab.c b/mm/slab.c
index b70aabf..6f0d7be 100644
--- a/mm/slab.c
+++ b/mm/slab.c
@@ -1229,6 +1229,59 @@ static void __init set_up_node(struct kmem_cache *cachep, int index)
}
}

+#ifdef CONFIG_FREELIST_RANDOM
+/*
+ * Master lists are pre-computed random lists
+ * Lists of different sizes are used to optimize performance on different
+ * SLAB object sizes per pages.
+ */
+static freelist_idx_t master_list_2[2];
+static freelist_idx_t master_list_4[4];
+static freelist_idx_t master_list_8[8];
+static freelist_idx_t master_list_16[16];
+static freelist_idx_t master_list_32[32];
+static freelist_idx_t master_list_64[64];
+static freelist_idx_t master_list_128[128];
+static freelist_idx_t master_list_256[256];
+static struct m_list {
+ size_t count;
+ freelist_idx_t *list;
+} master_lists[] = {
+ { ARRAY_SIZE(master_list_2), master_list_2 },
+ { ARRAY_SIZE(master_list_4), master_list_4 },
+ { ARRAY_SIZE(master_list_8), master_list_8 },
+ { ARRAY_SIZE(master_list_16), master_list_16 },
+ { ARRAY_SIZE(master_list_32), master_list_32 },
+ { ARRAY_SIZE(master_list_64), master_list_64 },
+ { ARRAY_SIZE(master_list_128), master_list_128 },
+ { ARRAY_SIZE(master_list_256), master_list_256 },
+};
+
+void __init freelist_random_init(void)
+{
+ unsigned int seed;
+ size_t z, i, rand;
+ struct rnd_state slab_rand;
+
+ get_random_bytes_arch(&seed, sizeof(seed));
+ prandom_seed_state(&slab_rand, seed);
+
+ for (z = 0; z < ARRAY_SIZE(master_lists); z++) {
+ for (i = 0; i < master_lists[z].count; i++)
+ master_lists[z].list[i] = i;
+
+ /* Fisher-Yates shuffle */
+ for (i = master_lists[z].count - 1; i > 0; i--) {
+ rand = prandom_u32_state(&slab_rand);
+ rand %= (i + 1);
+ swap(master_lists[z].list[i],
+ master_lists[z].list[rand]);
+ }
+ }
+}
+#endif /* CONFIG_FREELIST_RANDOM */
+
+
/*
* Initialisation. Called after the page allocator have been initialised and
* before smp_init().
@@ -1255,6 +1308,10 @@ void __init kmem_cache_init(void)
if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
slab_max_order = SLAB_MAX_ORDER_HI;

+#ifdef CONFIG_FREELIST_RANDOM
+ freelist_random_init();
+#endif /* CONFIG_FREELIST_RANDOM */
+
/* Bootstrap is tricky, because several objects are allocated
* from caches that do not exist yet:
* 1) initialize the kmem_cache cache: it contains the struct
@@ -2442,6 +2499,98 @@ static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page)
#endif
}

+#ifdef CONFIG_FREELIST_RANDOM
+enum master_type {
+ match,
+ less,
+ more
+};
+
+struct random_mng {
+ unsigned int padding;
+ unsigned int pos;
+ unsigned int count;
+ struct m_list master_list;
+ unsigned int master_count;
+ enum master_type type;
+};
+
+static void random_mng_initialize(struct random_mng *mng, unsigned int count)
+{
+ unsigned int idx;
+ const unsigned int last_idx = ARRAY_SIZE(master_lists) - 1;
+
+ memset(mng, 0, sizeof(*mng));
+ mng->count = count;
+ mng->pos = 0;
+ /* count is >= 2 */
+ idx = ilog2(count) - 1;
+ if (idx >= last_idx)
+ idx = last_idx;
+ else if (roundup_pow_of_two(idx + 1) != count)
+ idx++;
+ mng->master_list = master_lists[idx];
+ if (mng->master_list.count == mng->count)
+ mng->type = match;
+ else if (mng->master_list.count > mng->count)
+ mng->type = more;
+ else
+ mng->type = less;
+}
+
+static freelist_idx_t get_next_entry(struct random_mng *mng)
+{
+ if (mng->type == less && mng->pos == mng->master_list.count) {
+ mng->padding += mng->pos;
+ mng->pos = 0;
+ }
+ BUG_ON(mng->pos >= mng->master_list.count);
+ return mng->master_list.list[mng->pos++];
+}
+
+static freelist_idx_t next_random_slot(struct random_mng *mng)
+{
+ freelist_idx_t cur, entry;
+
+ entry = get_next_entry(mng);
+
+ if (mng->type != match) {
+ while ((entry + mng->padding) >= mng->count)
+ entry = get_next_entry(mng);
+ cur = entry + mng->padding;
+ BUG_ON(cur >= mng->count);
+ } else {
+ cur = entry;
+ }
+
+ return cur;
+}
+
+static void shuffle_freelist(struct kmem_cache *cachep, struct page *page,
+ unsigned int count)
+{
+ unsigned int i;
+ struct random_mng mng;
+
+ if (count < 2) {
+ for (i = 0; i < count; i++)
+ set_free_obj(page, i, i);
+ return;
+ }
+
+ /* Last chunk is used already in this case */
+ if (OBJFREELIST_SLAB(cachep))
+ count--;
+
+ random_mng_initialize(&mng, count);
+ for (i = 0; i < count; i++)
+ set_free_obj(page, i, next_random_slot(&mng));
+
+ if (OBJFREELIST_SLAB(cachep))
+ set_free_obj(page, i, i);
+}
+#endif /* CONFIG_FREELIST_RANDOM */
+
static void cache_init_objs(struct kmem_cache *cachep,
struct page *page)
{
@@ -2464,8 +2613,14 @@ static void cache_init_objs(struct kmem_cache *cachep,
kasan_poison_object_data(cachep, objp);
}

+#ifndef CONFIG_FREELIST_RANDOM
set_free_obj(page, i, i);
+#endif /* CONFIG_FREELIST_RANDOM */
}
+
+#ifdef CONFIG_FREELIST_RANDOM
+ shuffle_freelist(cachep, page, cachep->num);
+#endif /* CONFIG_FREELIST_RANDOM */
}

static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
--
2.8.0.rc3.226.g39d4020

[Greg KH]

On Wed, Apr 06, 2016 at 12:35:48PM -0700, Thomas Garnier wrote:
> Provide an optional config (CONFIG_FREELIST_RANDOM) to randomize the
> SLAB freelist. This security feature reduces the predictability of
> the kernel slab allocator against heap overflows.
>
> Randomized lists are pre-computed using a Fisher-Yates shuffle and
> re-used on slab creation for performance.
> ---
> Based on next-20160405
> ---

No signed-off-by:?

[Thomas Garnier]

Yes, sorry about that. It will be in the next RFC or PATCH.

[Kees Cook]

On Wed, Apr 6, 2016 at 12:35 PM, Thomas Garnier <thga...@google.com> wrote:
> Provide an optional config (CONFIG_FREELIST_RANDOM) to randomize the
> SLAB freelist.

It may be useful to describe _how_ it randomizes it (i.e. a high-level
description of what needed changing).

> This security feature reduces the predictability of
> the kernel slab allocator against heap overflows.

I would add "... rendering attacks much less stable." And if you can
find a specific example exploit that is foiled by this, I would refer
to it.

> Randomized lists are pre-computed using a Fisher-Yates shuffle and

Should the use of Fisher-Yates (over other things) be justified?

> re-used on slab creation for performance.

I'd like to see some benchmark results for this so the Kconfig can
include the performance characteristics. I recommend using hackbench
and kernel build times with a before/after comparison.

> ---
> Based on next-20160405
> ---
> init/Kconfig | 9 ++++
> mm/slab.c | 155 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
> 2 files changed, 164 insertions(+)
>
> diff --git a/init/Kconfig b/init/Kconfig
> index 0dfd09d..ee35418 100644
> --- a/init/Kconfig
> +++ b/init/Kconfig
> @@ -1742,6 +1742,15 @@ config SLOB
>
> endchoice
>
> +config FREELIST_RANDOM

I think I would name this "SLAB_FREELIST_RANDOM" since it's
SLAB-specific, unless you think it could be extended to the other
allocators in the future too? (If so, I'd mention the naming choice in
the commit log.)

For below...

#else
static inline freelist_random_init(void) { }

> +#endif /* CONFIG_FREELIST_RANDOM */
> +
> +
> /*
> * Initialisation. Called after the page allocator have been initialised and
> * before smp_init().
> @@ -1255,6 +1308,10 @@ void __init kmem_cache_init(void)
> if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
> slab_max_order = SLAB_MAX_ORDER_HI;
>
> +#ifdef CONFIG_FREELIST_RANDOM
> + freelist_random_init();
> +#endif /* CONFIG_FREELIST_RANDOM */

Rather than these embedded ifdefs, I would create stub function at the
top, as above.

Same thing here...

#else
static inline void set_free_obj(...) { }
static inline void shuffle_freelist(struct kmem_cache *cachep,
struct page *page, unsigned int count) { }

> +#endif /* CONFIG_FREELIST_RANDOM */
> +
> static void cache_init_objs(struct kmem_cache *cachep,
> struct page *page)
> {
> @@ -2464,8 +2613,14 @@ static void cache_init_objs(struct kmem_cache *cachep,
> kasan_poison_object_data(cachep, objp);
> }
>
> +#ifndef CONFIG_FREELIST_RANDOM
> set_free_obj(page, i, i);
> +#endif /* CONFIG_FREELIST_RANDOM */

For this one, I'd use:

if (config_enabled(CONFIG_FREELIST_RANDOM))
set_free_obj(page, i, i);

> }
> +
> +#ifdef CONFIG_FREELIST_RANDOM
> + shuffle_freelist(cachep, page, cachep->num);
> +#endif /* CONFIG_FREELIST_RANDOM */

This one can drop the ifdef in favor of using the stub function too.

> }
>
> static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
> --
> 2.8.0.rc3.226.g39d4020
>

Exciting!

-Kees

--
Kees Cook
Chrome OS & Brillo Security

[Thomas Garnier]

Thanks for the feedback Kees. I am preparing another RFC version.

For the config, I plan on creating an equivalent option for SLUB. Both
can benefit from randomizing their freelist order.

Thomas

[Yves-Alexis Perez]

On mer., 2016-04-06 at 14:45 -0700, Kees Cook wrote:
> > This security feature reduces the predictability of
> > the kernel slab allocator against heap overflows.
>
> I would add "... rendering attacks much less stable." And if you can
> find a specific example exploit that is foiled by this, I would refer
> to it.

One good example might (or might not) be the keyring issue from earlier this
year (CVE-2016-0728):

http://perception-point.io/2016/01/14/analysis-and-exploitation-of-a-linux-ker
nel-vulnerability-cve-2016-0728/

Regards,
--
Yves-Alexis

[Thomas Garnier]

That's a use after free. The randomization of the freelist should not
have much effect on that. I was going to quote this exploit that is
applicable to SLAB as well:
https://jon.oberheide.org/blog/2010/09/10/linux-kernel-can-slub-overflow

Regards.
Thomas

[Jesper Dangaard Brouer]

On Wed, 6 Apr 2016 14:45:30 -0700 Kees Cook <kees...@chromium.org> wrote:

> On Wed, Apr 6, 2016 at 12:35 PM, Thomas Garnier <thga...@google.com> wrote:

[...]

> > re-used on slab creation for performance.
>
> I'd like to see some benchmark results for this so the Kconfig can
> include the performance characteristics. I recommend using hackbench
> and kernel build times with a before/after comparison.
>

It looks like it only happens on init, right? (Thus must bench tools
might not be the right choice).

My slab tools for benchmarking the fastpath is here:
https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/mm/slab_bulk_test01.c

And I also carry a version of Christoph's slab bench tool:
https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/mm/slab_test.c

--
Best regards,
Jesper Dangaard Brouer
MSc.CS, Principal Kernel Engineer at Red Hat
Author of http://www.iptv-analyzer.org
LinkedIn: http://www.linkedin.com/in/brouer

[Kees Cook]

On Thu, Apr 7, 2016 at 2:14 PM, Jesper Dangaard Brouer
<bro...@redhat.com> wrote:
>
> On Wed, 6 Apr 2016 14:45:30 -0700 Kees Cook <kees...@chromium.org> wrote:
>
>> On Wed, Apr 6, 2016 at 12:35 PM, Thomas Garnier <thga...@google.com> wrote:
> [...]
>> > re-used on slab creation for performance.
>>
>> I'd like to see some benchmark results for this so the Kconfig can
>> include the performance characteristics. I recommend using hackbench
>> and kernel build times with a before/after comparison.
>>
>
> It looks like it only happens on init, right? (Thus must bench tools
> might not be the right choice).

Oh! Yes, you're right. I entirely missed that detail. :) 0-cost
randomization! Sounds good to me. :)

-Kees

>
> My slab tools for benchmarking the fastpath is here:
> https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/mm/slab_bulk_test01.c
>
> And I also carry a version of Christoph's slab bench tool:
> https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/mm/slab_test.c
>
> --
> Best regards,
> Jesper Dangaard Brouer
> MSc.CS, Principal Kernel Engineer at Red Hat
> Author of http://www.iptv-analyzer.org
> LinkedIn: http://www.linkedin.com/in/brouer

[Thomas Garnier]

Provide an optional config (CONFIG_FREELIST_RANDOM) to randomize the

SLAB freelist. The list is randomized during initialization of a new set
of pages. The order on different freelist sizes is pre-computed at boot
for performance. This security feature reduces the predictability of the
kernel SLAB allocator against heap overflows rendering attacks much less
stable.

For example this attack against SLUB (also applicable against SLAB)
would be affected:
https://jon.oberheide.org/blog/2010/09/10/linux-kernel-can-slub-overflow/

The config option name is not specific to the SLAB as this approach will
be extended to other allocators like SLUB.

Performance results highlighted no major changes:

Netperf average on 10 runs:

threads,base,change
16,576943.10,585905.90 (101.55%)
32,564082.00,569741.20 (101.00%)
48,558334.30,561851.20 (100.63%)
64,552025.20,556448.30 (100.80%)
80,552294.40,551743.10 (99.90%)
96,552435.30,547529.20 (99.11%)
112,551320.60,550183.20 (99.79%)
128,549138.30,550542.70 (100.26%)
144,549344.50,544529.10 (99.12%)
160,550360.80,539929.30 (98.10%)

slab_test 1 run on boot. After is faster except for odd result on size
2048.

Before:

Single thread testing
=====================
1. Kmalloc: Repeatedly allocate then free test
10000 times kmalloc(8) -> 137 cycles kfree -> 126 cycles
10000 times kmalloc(16) -> 118 cycles kfree -> 119 cycles
10000 times kmalloc(32) -> 112 cycles kfree -> 119 cycles
10000 times kmalloc(64) -> 126 cycles kfree -> 123 cycles
10000 times kmalloc(128) -> 135 cycles kfree -> 131 cycles
10000 times kmalloc(256) -> 165 cycles kfree -> 104 cycles
10000 times kmalloc(512) -> 174 cycles kfree -> 126 cycles
10000 times kmalloc(1024) -> 242 cycles kfree -> 160 cycles
10000 times kmalloc(2048) -> 478 cycles kfree -> 239 cycles
10000 times kmalloc(4096) -> 747 cycles kfree -> 364 cycles
10000 times kmalloc(8192) -> 774 cycles kfree -> 404 cycles
10000 times kmalloc(16384) -> 849 cycles kfree -> 430 cycles
2. Kmalloc: alloc/free test
10000 times kmalloc(8)/kfree -> 118 cycles
10000 times kmalloc(16)/kfree -> 118 cycles
10000 times kmalloc(32)/kfree -> 118 cycles
10000 times kmalloc(64)/kfree -> 121 cycles
10000 times kmalloc(128)/kfree -> 118 cycles
10000 times kmalloc(256)/kfree -> 115 cycles
10000 times kmalloc(512)/kfree -> 115 cycles
10000 times kmalloc(1024)/kfree -> 115 cycles
10000 times kmalloc(2048)/kfree -> 115 cycles
10000 times kmalloc(4096)/kfree -> 115 cycles
10000 times kmalloc(8192)/kfree -> 115 cycles
10000 times kmalloc(16384)/kfree -> 115 cycles

After:

Single thread testing
=====================
1. Kmalloc: Repeatedly allocate then free test
10000 times kmalloc(8) -> 99 cycles kfree -> 84 cycles
10000 times kmalloc(16) -> 88 cycles kfree -> 83 cycles
10000 times kmalloc(32) -> 90 cycles kfree -> 81 cycles
10000 times kmalloc(64) -> 107 cycles kfree -> 97 cycles
10000 times kmalloc(128) -> 134 cycles kfree -> 89 cycles
10000 times kmalloc(256) -> 145 cycles kfree -> 97 cycles
10000 times kmalloc(512) -> 177 cycles kfree -> 116 cycles
10000 times kmalloc(1024) -> 223 cycles kfree -> 151 cycles
10000 times kmalloc(2048) -> 1429 cycles kfree -> 221 cycles
10000 times kmalloc(4096) -> 720 cycles kfree -> 348 cycles
10000 times kmalloc(8192) -> 788 cycles kfree -> 393 cycles
10000 times kmalloc(16384) -> 867 cycles kfree -> 433 cycles
2. Kmalloc: alloc/free test
10000 times kmalloc(8)/kfree -> 115 cycles
10000 times kmalloc(16)/kfree -> 115 cycles
10000 times kmalloc(32)/kfree -> 115 cycles
10000 times kmalloc(64)/kfree -> 120 cycles
10000 times kmalloc(128)/kfree -> 127 cycles
10000 times kmalloc(256)/kfree -> 119 cycles
10000 times kmalloc(512)/kfree -> 112 cycles
10000 times kmalloc(1024)/kfree -> 112 cycles
10000 times kmalloc(2048)/kfree -> 112 cycles
10000 times kmalloc(4096)/kfree -> 112 cycles
10000 times kmalloc(8192)/kfree -> 112 cycles
10000 times kmalloc(16384)/kfree -> 112 cycles

slab_bulk results (look about the same)

Before:

DEBUG: cpu:10
Type:for_loop Per elem: 1 cycles(tsc) 0.318 ns (step:0) - (measurement
period time:0.031823553 sec time_interval:31823553) - (invoke
count:100000000 tsc_interval:111112332)
Type:kmem fastpath reuse Per elem: 104 cycles(tsc) 29.808 ns (step:0)
- (measurement period time:0.298081100 sec time_interval:298081100)
- (invoke count:10000000 tsc_interval:1040760500)
Type:kmem bulk_fallback Per elem: 125 cycles(tsc) 35.872 ns (step:1)
- (measurement period time:0.358726418 sec time_interval:358726418)
- (invoke count:10000000 tsc_interval:1252505764)
Type:kmem bulk_quick_reuse Per elem: 55 cycles(tsc) 15.850 ns (step:1)
- (measurement period time:0.158508913 sec time_interval:158508913)
- (invoke count:10000000 tsc_interval:553439128)
Type:kmem bulk_fallback Per elem: 112 cycles(tsc) 32.326 ns (step:2)
- (measurement period time:0.323265281 sec time_interval:323265281)
- (invoke count:10000000 tsc_interval:1128692072)
Type:kmem bulk_quick_reuse Per elem: 34 cycles(tsc) 9.867 ns (step:2)
- (measurement period time:0.098670493 sec time_interval:98670493)
- (invoke count:10000000 tsc_interval:344510914)
Type:kmem bulk_fallback Per elem: 107 cycles(tsc) 30.907 ns (step:3)
- (measurement period time:0.309076362 sec time_interval:309076362)
- (invoke count:9999999 tsc_interval:1079150859)
Type:kmem bulk_quick_reuse Per elem: 28 cycles(tsc) 8.045 ns (step:3)
- (measurement period time:0.080459150 sec time_interval:80459150)
- (invoke count:9999999 tsc_interval:280925570)
Type:kmem bulk_fallback Per elem: 105 cycles(tsc) 30.156 ns (step:4)
- (measurement period time:0.301569211 sec time_interval:301569211)
- (invoke count:10000000 tsc_interval:1052939565)
Type:kmem bulk_quick_reuse Per elem: 25 cycles(tsc) 7.368 ns (step:4)
- (measurement period time:0.073680499 sec time_interval:73680499)
- (invoke count:10000000 tsc_interval:257257775)
Type:kmem bulk_fallback Per elem: 103 cycles(tsc) 29.717 ns (step:8)
- (measurement period time:0.297170419 sec time_interval:297170419)
- (invoke count:10000000 tsc_interval:1037580931)
Type:kmem bulk_quick_reuse Per elem: 22 cycles(tsc) 6.446 ns (step:8)
- (measurement period time:0.064465569 sec time_interval:64465569)
- (invoke count:10000000 tsc_interval:225083219)
Type:kmem bulk_fallback Per elem: 102 cycles(tsc) 29.435 ns (step:16)
- (measurement period time:0.294353584 sec time_interval:294353584)
- (invoke count:10000000 tsc_interval:1027745957)
Type:kmem bulk_quick_reuse Per elem: 21 cycles(tsc) 6.052 ns (step:16)
- (measurement period time:0.060526862 sec time_interval:60526862)
- (invoke count:10000000 tsc_interval:211331314)
Type:kmem bulk_fallback Per elem: 127 cycles(tsc) 36.440 ns (step:30)
- (measurement period time:0.364403518 sec time_interval:364403518)
- (invoke count:9999990 tsc_interval:1272325901)
Type:kmem bulk_quick_reuse Per elem: 32 cycles(tsc) 9.213 ns (step:30)
- (measurement period time:0.092130623 sec time_interval:92130623)
- (invoke count:9999990 tsc_interval:321676961)
Type:kmem bulk_fallback Per elem: 129 cycles(tsc) 36.985 ns (step:32)
- (measurement period time:0.369859273 sec time_interval:369859273)
- (invoke count:10000000 tsc_interval:1291376818)
Type:kmem bulk_quick_reuse Per elem: 31 cycles(tsc) 9.083 ns (step:32)
- (measurement period time:0.090834101 sec time_interval:90834101)
- (invoke count:10000000 tsc_interval:317150093)
Type:kmem bulk_fallback Per elem: 129 cycles(tsc) 37.057 ns (step:34)
- (measurement period time:0.370577150 sec time_interval:370577150)
- (invoke count:9999978 tsc_interval:1293883110)
Type:kmem bulk_quick_reuse Per elem: 32 cycles(tsc) 9.182 ns (step:34)
- (measurement period time:0.091828683 sec time_interval:91828683)
- (invoke count:9999978 tsc_interval:320622702)
Type:kmem bulk_fallback Per elem: 126 cycles(tsc) 36.244 ns (step:48)
- (measurement period time:0.362448363 sec time_interval:362448363)
- (invoke count:9999984 tsc_interval:1265501472)
Type:kmem bulk_quick_reuse Per elem: 20 cycles(tsc) 6.012 ns (step:48)
- (measurement period time:0.060121234 sec time_interval:60121234)
- (invoke count:9999984 tsc_interval:209914922)
Type:kmem bulk_fallback Per elem: 108 cycles(tsc) 31.123 ns (step:64)
- (measurement period time:0.311231972 sec time_interval:311231972)
- (invoke count:10000000 tsc_interval:1086677115)
Type:kmem bulk_quick_reuse Per elem: 20 cycles(tsc) 6.014 ns (step:64)
- (measurement period time:0.060142595 sec time_interval:60142595)
- (invoke count:10000000 tsc_interval:209989505)
Type:kmem bulk_fallback Per elem: 107 cycles(tsc) 30.676 ns (step:128)
- (measurement period time:0.306766510 sec time_interval:306766510)
- (invoke count:10000000 tsc_interval:1071085978)
Type:kmem bulk_quick_reuse Per elem: 23 cycles(tsc) 6.696 ns (step:128)
- (measurement period time:0.066960903 sec time_interval:66960903)
- (invoke count:10000000 tsc_interval:233795952)
Type:kmem bulk_fallback Per elem: 106 cycles(tsc) 30.614 ns (step:158)
- (measurement period time:0.306148249 sec time_interval:306148249)
- (invoke count:9999978 tsc_interval:1068927458)
Type:kmem bulk_quick_reuse Per elem: 24 cycles(tsc) 7.077 ns (step:158)
- (measurement period time:0.070774353 sec time_interval:70774353)
- (invoke count:9999978 tsc_interval:247110695)
Type:kmem bulk_fallback Per elem: 107 cycles(tsc) 30.914 ns (step:250)
- (measurement period time:0.309144377 sec time_interval:309144377)
- (invoke count:10000000 tsc_interval:1079388538)
Type:kmem bulk_quick_reuse Per elem: 25 cycles(tsc) 7.283 ns (step:250)
- (measurement period time:0.072836404 sec time_interval:72836404)
- (invoke count:10000000 tsc_interval:254309986)

After:

DEBUG: cpu:1
Type:for_loop Per elem: 1 cycles(tsc) 0.289 ns (step:0) - (measurement
period time:0.028953054 sec time_interval:28953054) - (invoke
count:100000000 tsc_interval:101090400)
Type:kmem fastpath reuse Per elem: 104 cycles(tsc) 29.800 ns (step:0)
- (measurement period time:0.298003253 sec time_interval:298003253)
- (invoke count:10000000 tsc_interval:1040491972)
Type:kmem bulk_fallback Per elem: 125 cycles(tsc) 35.892 ns (step:1)
- (measurement period time:0.358924780 sec time_interval:358924780)
- (invoke count:10000000 tsc_interval:1253202488)
Type:kmem bulk_quick_reuse Per elem: 55 cycles(tsc) 15.780 ns (step:1)
- (measurement period time:0.157804420 sec time_interval:157804420)
- (invoke count:10000000 tsc_interval:550981196)
Type:kmem bulk_fallback Per elem: 112 cycles(tsc) 32.220 ns (step:2)
- (measurement period time:0.322202576 sec time_interval:322202576)
- (invoke count:10000000 tsc_interval:1124985278)
Type:kmem bulk_quick_reuse Per elem: 34 cycles(tsc) 9.834 ns (step:2)
- (measurement period time:0.098344277 sec time_interval:98344277)
- (invoke count:10000000 tsc_interval:343373130)
Type:kmem bulk_fallback Per elem: 108 cycles(tsc) 30.946 ns (step:3)
- (measurement period time:0.309467507 sec time_interval:309467507)
- (invoke count:9999999 tsc_interval:1080519939)
Type:kmem bulk_quick_reuse Per elem: 28 cycles(tsc) 8.183 ns (step:3)
- (measurement period time:0.081831772 sec time_interval:81831772)
- (invoke count:9999999 tsc_interval:285718619)
Type:kmem bulk_fallback Per elem: 107 cycles(tsc) 30.832 ns (step:4)
- (measurement period time:0.308327840 sec time_interval:308327840)
- (invoke count:10000000 tsc_interval:1076540739)
Type:kmem bulk_quick_reuse Per elem: 26 cycles(tsc) 7.512 ns (step:4)
- (measurement period time:0.075123119 sec time_interval:75123119)
- (invoke count:10000000 tsc_interval:262295498)
Type:kmem bulk_fallback Per elem: 105 cycles(tsc) 30.291 ns (step:8)
- (measurement period time:0.302919692 sec time_interval:302919692)
- (invoke count:10000000 tsc_interval:1057657960)
Type:kmem bulk_quick_reuse Per elem: 22 cycles(tsc) 6.578 ns (step:8)
- (measurement period time:0.065788230 sec time_interval:65788230)
- (invoke count:10000000 tsc_interval:229700901)
Type:kmem bulk_fallback Per elem: 104 cycles(tsc) 29.805 ns (step:16)
- (measurement period time:0.298055380 sec time_interval:298055380)
- (invoke count:10000000 tsc_interval:1040674120)
Type:kmem bulk_quick_reuse Per elem: 21 cycles(tsc) 6.144 ns (step:16)
- (measurement period time:0.061447185 sec time_interval:61447185)
- (invoke count:10000000 tsc_interval:214545335)
Type:kmem bulk_fallback Per elem: 104 cycles(tsc) 29.837 ns (step:30)
- (measurement period time:0.298372940 sec time_interval:298372940)
- (invoke count:9999990 tsc_interval:1041782971)
Type:kmem bulk_quick_reuse Per elem: 21 cycles(tsc) 6.031 ns (step:30)
- (measurement period time:0.060319478 sec time_interval:60319478)
- (invoke count:9999990 tsc_interval:210607930)
Type:kmem bulk_fallback Per elem: 104 cycles(tsc) 29.967 ns (step:32)
- (measurement period time:0.299670182 sec time_interval:299670182)
- (invoke count:10000000 tsc_interval:1046312308)
Type:kmem bulk_quick_reuse Per elem: 21 cycles(tsc) 6.027 ns (step:32)
- (measurement period time:0.060277128 sec time_interval:60277128)
- (invoke count:10000000 tsc_interval:210460013)
Type:kmem bulk_fallback Per elem: 104 cycles(tsc) 29.989 ns (step:34)
- (measurement period time:0.299891491 sec time_interval:299891491)
- (invoke count:9999978 tsc_interval:1047083288)
Type:kmem bulk_quick_reuse Per elem: 20 cycles(tsc) 5.954 ns (step:34)
- (measurement period time:0.059547431 sec time_interval:59547431)
- (invoke count:9999978 tsc_interval:207912186)
Type:kmem bulk_fallback Per elem: 103 cycles(tsc) 29.767 ns (step:48)
- (measurement period time:0.297677464 sec time_interval:297677464)
- (invoke count:9999984 tsc_interval:1039354626)
Type:kmem bulk_quick_reuse Per elem: 20 cycles(tsc) 6.001 ns (step:48)
- (measurement period time:0.060014156 sec time_interval:60014156)
- (invoke count:9999984 tsc_interval:209541572)
Type:kmem bulk_fallback Per elem: 104 cycles(tsc) 29.879 ns (step:64)
- (measurement period time:0.298799724 sec time_interval:298799724)
- (invoke count:10000000 tsc_interval:1043273056)
Type:kmem bulk_quick_reuse Per elem: 20 cycles(tsc) 5.917 ns (step:64)
- (measurement period time:0.059172278 sec time_interval:59172278)
- (invoke count:10000000 tsc_interval:206602418)
Type:kmem bulk_fallback Per elem: 105 cycles(tsc) 30.261 ns (step:128)
- (measurement period time:0.302610710 sec time_interval:302610710)
- (invoke count:10000000 tsc_interval:1056579291)
Type:kmem bulk_quick_reuse Per elem: 22 cycles(tsc) 6.431 ns (step:128)
- (measurement period time:0.064314751 sec time_interval:64314751)
- (invoke count:10000000 tsc_interval:224557590)
Type:kmem bulk_fallback Per elem: 108 cycles(tsc) 31.027 ns (step:158)
- (measurement period time:0.310276416 sec time_interval:310276416)
- (invoke count:9999978 tsc_interval:1083341310)
Type:kmem bulk_quick_reuse Per elem: 24 cycles(tsc) 6.989 ns (step:158)
- (measurement period time:0.069891439 sec time_interval:69891439)
- (invoke count:9999978 tsc_interval:244028721)
Type:kmem bulk_fallback Per elem: 107 cycles(tsc) 30.833 ns (step:250)
- (measurement period time:0.308335100 sec time_interval:308335100)
- (invoke count:10000000 tsc_interval:1076566255)
Type:kmem bulk_quick_reuse Per elem: 24 cycles(tsc) 6.947 ns (step:250)
- (measurement period time:0.069477012 sec time_interval:69477012)
- (invoke count:10000000 tsc_interval:242581824)

Signed-off-by: Thomas Garnier <thga...@google.com>

---
Based on next-20160405
---
init/Kconfig | 9 ++++

mm/slab.c | 158 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++-
2 files changed, 166 insertions(+), 1 deletion(-)

diff --git a/init/Kconfig b/init/Kconfig
index 0dfd09d..ee35418 100644
--- a/init/Kconfig
+++ b/init/Kconfig
@@ -1742,6 +1742,15 @@ config SLOB

endchoice

+config FREELIST_RANDOM

+ default n
+ depends on SLAB
+ bool "SLAB freelist randomization"
+ help
+ Randomizes the freelist order used on creating new SLABs. This
+ security feature reduces the predictability of the kernel slab
+ allocator against heap overflows.
+
config SLUB_CPU_PARTIAL
default y
depends on SLUB && SMP
diff --git a/mm/slab.c b/mm/slab.c

index b70aabf..5d8bde2 100644
--- a/mm/slab.c
+++ b/mm/slab.c
@@ -1229,6 +1229,61 @@ static void __init set_up_node(struct kmem_cache *cachep, int index)

+static void __init freelist_random_init(void)

+{
+ unsigned int seed;
+ size_t z, i, rand;
+ struct rnd_state slab_rand;
+
+ get_random_bytes_arch(&seed, sizeof(seed));
+ prandom_seed_state(&slab_rand, seed);
+
+ for (z = 0; z < ARRAY_SIZE(master_lists); z++) {
+ for (i = 0; i < master_lists[z].count; i++)
+ master_lists[z].list[i] = i;
+
+ /* Fisher-Yates shuffle */
+ for (i = master_lists[z].count - 1; i > 0; i--) {
+ rand = prandom_u32_state(&slab_rand);
+ rand %= (i + 1);
+ swap(master_lists[z].list[i],
+ master_lists[z].list[rand]);
+ }
+ }
+}

+#else
+static inline void __init freelist_random_init(void) { }

+#endif /* CONFIG_FREELIST_RANDOM */
+
+
/*
* Initialisation. Called after the page allocator have been initialised and
* before smp_init().

@@ -1255,6 +1310,8 @@ void __init kmem_cache_init(void)

if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
slab_max_order = SLAB_MAX_ORDER_HI;

+ freelist_random_init();

+
/* Bootstrap is tricky, because several objects are allocated
* from caches that do not exist yet:
* 1) initialize the kmem_cache cache: it contains the struct

@@ -2442,6 +2499,101 @@ static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page)

+#else
+static inline void shuffle_freelist(struct kmem_cache *cachep,
+ struct page *page, unsigned int count) { }

+#endif /* CONFIG_FREELIST_RANDOM */
+
static void cache_init_objs(struct kmem_cache *cachep,
struct page *page)
{

@@ -2464,8 +2616,12 @@ static void cache_init_objs(struct kmem_cache *cachep,
kasan_poison_object_data(cachep, objp);
}

- set_free_obj(page, i, i);
+ /* If enabled, initialization is done in shuffle_freelist */
+ if (!config_enabled(CONFIG_FREELIST_RANDOM))

+ set_free_obj(page, i, i);
}
+

+ shuffle_freelist(cachep, page, cachep->num);
}

[Thomas Garnier]

Provide an optional config (CONFIG_FREELIST_RANDOM) to randomize the
SLAB freelist. The list is randomized during initialization of a new set
of pages. The order on different freelist sizes is pre-computed at boot
for performance. This security feature reduces the predictability of the
kernel SLAB allocator against heap overflows rendering attacks much less
stable.

For example this attack against SLUB (also applicable against SLAB)
would be affected:
https://jon.oberheide.org/blog/2010/09/10/linux-kernel-can-slub-overflow/

Also, since v4.6 the freelist was moved at the end of the SLAB. It means
a controllable heap is opened to new attacks not yet publicly discussed.
A kernel heap overflow can be transformed to multiple use-after-free.
This feature makes this type of attack harder too.

Signed-off-by: Thomas Garnier <thga...@google.com>
---

Based on next-20160414

[Andrew Morton]

On Fri, 15 Apr 2016 10:25:59 -0700 Thomas Garnier <thga...@google.com> wrote:

> Provide an optional config (CONFIG_FREELIST_RANDOM) to randomize the
> SLAB freelist. The list is randomized during initialization of a new set
> of pages. The order on different freelist sizes is pre-computed at boot
> for performance. This security feature reduces the predictability of the
> kernel SLAB allocator against heap overflows rendering attacks much less
> stable.
>
> For example this attack against SLUB (also applicable against SLAB)
> would be affected:
> https://jon.oberheide.org/blog/2010/09/10/linux-kernel-can-slub-overflow/
>
> Also, since v4.6 the freelist was moved at the end of the SLAB. It means
> a controllable heap is opened to new attacks not yet publicly discussed.
> A kernel heap overflow can be transformed to multiple use-after-free.
> This feature makes this type of attack harder too.
>
> The config option name is not specific to the SLAB as this approach will
> be extended to other allocators like SLUB.
>
> Performance results highlighted no major changes:
>

> ...

>
> --- a/mm/slab.c
> +++ b/mm/slab.c
> @@ -1229,6 +1229,61 @@ static void __init set_up_node(struct kmem_cache *cachep, int index)
> }
> }
>
> +#ifdef CONFIG_FREELIST_RANDOM
> +/*
> + * Master lists are pre-computed random lists
> + * Lists of different sizes are used to optimize performance on different
> + * SLAB object sizes per pages.

"object sizes per pages" doesn't make sense. "object-per-page counts"?
"object sizes"?

Using get_random_bytes_arch() seems a rather poor decision. There are
the caveats described at the get_random_bytes_arch() definition site,
and the minor issue that get_random_bytes_arch() only actually works on
x86 and powerpc!

This is run-once __init code, so rather than adding the kernel's very
first get_random_bytes_arch() call site(!), why not stick with good old
get_random_bytes()?

If there's something I'm missing, please at least place a very good
comment here explaining the reasoning.

> + prandom_seed_state(&slab_rand, seed);
> +
> + for (z = 0; z < ARRAY_SIZE(master_lists); z++) {
> + for (i = 0; i < master_lists[z].count; i++)
> + master_lists[z].list[i] = i;
> +
> + /* Fisher-Yates shuffle */
> + for (i = master_lists[z].count - 1; i > 0; i--) {
> + rand = prandom_u32_state(&slab_rand);
> + rand %= (i + 1);
> + swap(master_lists[z].list[i],
> + master_lists[z].list[rand]);
> + }
> + }
> +}
> +#else
> +static inline void __init freelist_random_init(void) { }
> +#endif /* CONFIG_FREELIST_RANDOM */
> +
> +
> /*
> * Initialisation. Called after the page allocator have been initialised and
> * before smp_init().
>

> ...

>
> @@ -2442,6 +2499,101 @@ static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page)
> #endif
> }
>
> +#ifdef CONFIG_FREELIST_RANDOM
> +enum master_type {
> + match,
> + less,
> + more
> +};
> +
> +struct random_mng {

I can't work out what "mng" means in this code.

> + unsigned int padding;
> + unsigned int pos;
> + unsigned int count;
> + struct m_list master_list;
> + unsigned int master_count;
> + enum master_type type;
> +};

It would be useful to document the above struct. Skilfully documenting
the data structures is key to making the code understandable.

> +static void random_mng_initialize(struct random_mng *mng, unsigned int count)
> +{
> + unsigned int idx;
> + const unsigned int last_idx = ARRAY_SIZE(master_lists) - 1;
> +
> + memset(mng, 0, sizeof(*mng));
> + mng->count = count;
> + mng->pos = 0;
> + /* count is >= 2 */
> + idx = ilog2(count) - 1;

slab.c should now include log2.h.

Sorry, but the code is really too light on comments. Each of the above
functions would benefit from a description of what they do and, most
importantly, why they do it.

Please address these things and let's wait for the slab maintainers
(and perhaps Kees?) to weigh in.

[Joe Perches]

On Fri, 2016-04-15 at 15:00 -0700, Andrew Morton wrote:
> On Fri, 15 Apr 2016 10:25:59 -0700 Thomas Garnier <thga...@google.com> wrote:
> > Provide an optional config (CONFIG_FREELIST_RANDOM) to randomize the
> > SLAB freelist. The list is randomized during initialization of a new set
> > of pages. The order on different freelist sizes is pre-computed at boot
> > for performance. This security feature reduces the predictability of the
> > kernel SLAB allocator against heap overflows rendering attacks much less
> > stable.

trivia:

> > @@ -1229,6 +1229,61 @@ static void __init set_up_node(struct kmem_cache *cachep, int index)

[]

> > + */
> > +static freelist_idx_t master_list_2[2];
> > +static freelist_idx_t master_list_4[4];
> > +static freelist_idx_t master_list_8[8];
> > +static freelist_idx_t master_list_16[16];
> > +static freelist_idx_t master_list_32[32];
> > +static freelist_idx_t master_list_64[64];
> > +static freelist_idx_t master_list_128[128];
> > +static freelist_idx_t master_list_256[256];
> > +static struct m_list {
> > + size_t count;
> > + freelist_idx_t *list;
> > +} master_lists[] = {
> > + { ARRAY_SIZE(master_list_2), master_list_2 },
> > + { ARRAY_SIZE(master_list_4), master_list_4 },
> > + { ARRAY_SIZE(master_list_8), master_list_8 },
> > + { ARRAY_SIZE(master_list_16), master_list_16 },
> > + { ARRAY_SIZE(master_list_32), master_list_32 },
> > + { ARRAY_SIZE(master_list_64), master_list_64 },
> > + { ARRAY_SIZE(master_list_128), master_list_128 },
> > + { ARRAY_SIZE(master_list_256), master_list_256 },
> > +};

static const struct m_list?

[Thomas Garnier]

Thanks for the comments. I will address them in a v2 early next week.

If anyone has other comments, please let me know.

Thomas

[Thomas Garnier]

I will send the next version today. Note that I get_random_bytes_arch
is used because at that stage we have 0 bits of entropy. It seemed
like a better idea to use the arch version that will fallback on
get_random_bytes sub API in the worse case.

[Thomas Garnier]

Provides an optional config (CONFIG_FREELIST_RANDOM) to randomize the

SLAB freelist. The list is randomized during initialization of a new set
of pages. The order on different freelist sizes is pre-computed at boot
for performance. This security feature reduces the predictability of the
kernel SLAB allocator against heap overflows rendering attacks much less
stable.

For example this attack against SLUB (also applicable against SLAB)
would be affected:
https://jon.oberheide.org/blog/2010/09/10/linux-kernel-can-slub-overflow/

Also, since v4.6 the freelist was moved at the end of the SLAB. It means
a controllable heap is opened to new attacks not yet publicly discussed.
A kernel heap overflow can be transformed to multiple use-after-free.
This feature makes this type of attack harder too.

To generate entropy, we use get_random_bytes_arch because 0 bits of
entropy is available at that boot stage. In the worse case this function
will fallback to the get_random_bytes sub API.

The config option name is not specific to the SLAB as this approach will
be extended to other allocators like SLUB.

Performance results highlighted no major changes:

Based on next-20160418
---
init/Kconfig | 9 ++++
mm/slab.c | 166 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++-
2 files changed, 174 insertions(+), 1 deletion(-)

diff --git a/init/Kconfig b/init/Kconfig
index 0dfd09d..ee35418 100644
--- a/init/Kconfig
+++ b/init/Kconfig
@@ -1742,6 +1742,15 @@ config SLOB

endchoice

+config FREELIST_RANDOM
+ default n
+ depends on SLAB
+ bool "SLAB freelist randomization"
+ help
+ Randomizes the freelist order used on creating new SLABs. This
+ security feature reduces the predictability of the kernel slab
+ allocator against heap overflows.
+
config SLUB_CPU_PARTIAL
default y
depends on SLUB && SMP
diff --git a/mm/slab.c b/mm/slab.c

index b70aabf..8371d80 100644
--- a/mm/slab.c
+++ b/mm/slab.c
@@ -116,6 +116,7 @@
#include <linux/kmemcheck.h>
#include <linux/memory.h>
#include <linux/prefetch.h>
+#include <linux/log2.h>

#include <net/sock.h>

@@ -1229,6 +1230,62 @@ static void __init set_up_node(struct kmem_cache *cachep, int index)

}
}

+#ifdef CONFIG_FREELIST_RANDOM
+/*
+ * Master lists are pre-computed random lists

+ * Lists of different sizes are used to optimize performance on SLABS with
+ * different object counts.

+ */
+static freelist_idx_t master_list_2[2];
+static freelist_idx_t master_list_4[4];
+static freelist_idx_t master_list_8[8];
+static freelist_idx_t master_list_16[16];
+static freelist_idx_t master_list_32[32];
+static freelist_idx_t master_list_64[64];
+static freelist_idx_t master_list_128[128];
+static freelist_idx_t master_list_256[256];

+const static struct m_list {

+ size_t count;
+ freelist_idx_t *list;
+} master_lists[] = {
+ { ARRAY_SIZE(master_list_2), master_list_2 },
+ { ARRAY_SIZE(master_list_4), master_list_4 },
+ { ARRAY_SIZE(master_list_8), master_list_8 },
+ { ARRAY_SIZE(master_list_16), master_list_16 },
+ { ARRAY_SIZE(master_list_32), master_list_32 },
+ { ARRAY_SIZE(master_list_64), master_list_64 },
+ { ARRAY_SIZE(master_list_128), master_list_128 },
+ { ARRAY_SIZE(master_list_256), master_list_256 },
+};

+
+/* Pre-compute the Freelist master lists at boot */

+static void __init freelist_random_init(void)
+{
+ unsigned int seed;
+ size_t z, i, rand;
+ struct rnd_state slab_rand;
+
+ get_random_bytes_arch(&seed, sizeof(seed));

+ prandom_seed_state(&slab_rand, seed);
+
+ for (z = 0; z < ARRAY_SIZE(master_lists); z++) {
+ for (i = 0; i < master_lists[z].count; i++)
+ master_lists[z].list[i] = i;
+
+ /* Fisher-Yates shuffle */
+ for (i = master_lists[z].count - 1; i > 0; i--) {
+ rand = prandom_u32_state(&slab_rand);
+ rand %= (i + 1);
+ swap(master_lists[z].list[i],
+ master_lists[z].list[rand]);
+ }
+ }
+}
+#else
+static inline void __init freelist_random_init(void) { }
+#endif /* CONFIG_FREELIST_RANDOM */
+
+
/*
* Initialisation. Called after the page allocator have been initialised and
* before smp_init().

@@ -1255,6 +1312,8 @@ void __init kmem_cache_init(void)

if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
slab_max_order = SLAB_MAX_ORDER_HI;

+ freelist_random_init();
+
/* Bootstrap is tricky, because several objects are allocated
* from caches that do not exist yet:
* 1) initialize the kmem_cache cache: it contains the struct

@@ -2442,6 +2501,107 @@ static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page)
#endif
}

+#ifdef CONFIG_FREELIST_RANDOM
+/* Identify if the target freelist matches the pre-computed list */

+enum master_type {
+ match,
+ less,
+ more
+};
+

+/* Hold information during a freelist initialization */
+struct freelist_init_state {

+ unsigned int padding;
+ unsigned int pos;
+ unsigned int count;
+ struct m_list master_list;
+ unsigned int master_count;
+ enum master_type type;
+};

+
+/* Select the right pre-computed master list and initialize state */
+static void freelist_state_initialize(struct freelist_init_state *state,

+ unsigned int count)
+{

+ unsigned int idx;
+ const unsigned int last_idx = ARRAY_SIZE(master_lists) - 1;
+

+ memset(state, 0, sizeof(*state));
+ state->count = count;
+ state->pos = 0;
+ /* count is always >= 2 */

+ idx = ilog2(count) - 1;

+ if (idx >= last_idx)
+ idx = last_idx;
+ else if (roundup_pow_of_two(idx + 1) != count)
+ idx++;

+ state->master_list = master_lists[idx];
+ if (state->master_list.count == state->count)
+ state->type = match;
+ else if (state->master_list.count > state->count)
+ state->type = more;
+ else
+ state->type = less;
+}
+
+/* Get the next entry on the master list depending on the target list size */
+static freelist_idx_t get_next_entry(struct freelist_init_state *state)
+{
+ if (state->type == less && state->pos == state->master_list.count) {
+ state->padding += state->pos;
+ state->pos = 0;
+ }
+ BUG_ON(state->pos >= state->master_list.count);
+ return state->master_list.list[state->pos++];
+}
+
+static freelist_idx_t next_random_slot(struct freelist_init_state *state)

+{
+ freelist_idx_t cur, entry;
+

+ entry = get_next_entry(state);
+
+ if (state->type != match) {
+ while ((entry + state->padding) >= state->count)
+ entry = get_next_entry(state);
+ cur = entry + state->padding;
+ BUG_ON(cur >= state->count);

+ } else {
+ cur = entry;
+ }
+
+ return cur;
+}
+

+/* Shuffle the freelist initialization state based on pre-computed lists */

+static void shuffle_freelist(struct kmem_cache *cachep, struct page *page,
+ unsigned int count)
+{
+ unsigned int i;

+ struct freelist_init_state state;

+
+ if (count < 2) {
+ for (i = 0; i < count; i++)
+ set_free_obj(page, i, i);
+ return;
+ }
+
+ /* Last chunk is used already in this case */
+ if (OBJFREELIST_SLAB(cachep))
+ count--;
+

+ freelist_state_initialize(&state, count);

+ for (i = 0; i < count; i++)

+ set_free_obj(page, i, next_random_slot(&state));

+
+ if (OBJFREELIST_SLAB(cachep))
+ set_free_obj(page, i, i);
+}

+#else
+static inline void shuffle_freelist(struct kmem_cache *cachep,
+ struct page *page, unsigned int count) { }
+#endif /* CONFIG_FREELIST_RANDOM */
+
static void cache_init_objs(struct kmem_cache *cachep,
struct page *page)
{

@@ -2464,8 +2624,12 @@ static void cache_init_objs(struct kmem_cache *cachep,

kasan_poison_object_data(cachep, objp);
}

- set_free_obj(page, i, i);
+ /* If enabled, initialization is done in shuffle_freelist */
+ if (!config_enabled(CONFIG_FREELIST_RANDOM))
+ set_free_obj(page, i, i);
}
+
+ shuffle_freelist(cachep, page, cachep->num);
}

static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)

--
2.8.0.rc3.226.g39d4020

[Thomas Garnier]

[Laura Abbott]

On 04/18/2016 08:59 AM, Thomas Garnier wrote:
> I will send the next version today. Note that I get_random_bytes_arch
> is used because at that stage we have 0 bits of entropy. It seemed
> like a better idea to use the arch version that will fallback on
> get_random_bytes sub API in the worse case.
>

This is unfortunate for ARM/ARM64. Those platforms don't have a standard
method for getting random numbers so until additional entropy is added
get_random_bytes will always return the same seed and indeed I always
see the same shuffle on a quick test of arm64. For KASLR, the workaround
was to require the bootloader to pass in entropy. It might be good to
either document this or require this only be used with CONFIG_ARCH_RANDOM.

[Thomas Garnier]

I agree, if we had a generic way to pass entropy across boots on all
architecture that would be amazing. I will let the SLAB maintainers to
decide on requiring CONFIG_ARCH_RANDOM or documenting it.

[Joonsoo Kim]

On Mon, Apr 18, 2016 at 10:14:39AM -0700, Thomas Garnier wrote:
> Provides an optional config (CONFIG_FREELIST_RANDOM) to randomize the
> SLAB freelist. The list is randomized during initialization of a new set
> of pages. The order on different freelist sizes is pre-computed at boot
> for performance. This security feature reduces the predictability of the
> kernel SLAB allocator against heap overflows rendering attacks much less
> stable.

I'm not familiar on security but it doesn't look much secure than
before. Is there any other way to generate different sequence of freelist
for each new set of pages? Current approach using pre-computed array will
generate same sequence of freelist for all new set of pages having same size
class. Is it sufficient?

> For example this attack against SLUB (also applicable against SLAB)
> would be affected:
> https://jon.oberheide.org/blog/2010/09/10/linux-kernel-can-slub-overflow/
>
> Also, since v4.6 the freelist was moved at the end of the SLAB. It means
> a controllable heap is opened to new attacks not yet publicly discussed.
> A kernel heap overflow can be transformed to multiple use-after-free.
> This feature makes this type of attack harder too.
>
> To generate entropy, we use get_random_bytes_arch because 0 bits of
> entropy is available at that boot stage. In the worse case this function
> will fallback to the get_random_bytes sub API.
>
> The config option name is not specific to the SLAB as this approach will
> be extended to other allocators like SLUB.

If this feature will be applied to the SLUB, it's better to put common
code to mm/slab_common.c.

>
> Performance results highlighted no major changes:
>
> Netperf average on 10 runs:
>
> threads,base,change
> 16,576943.10,585905.90 (101.55%)
> 32,564082.00,569741.20 (101.00%)
> 48,558334.30,561851.20 (100.63%)
> 64,552025.20,556448.30 (100.80%)
> 80,552294.40,551743.10 (99.90%)
> 96,552435.30,547529.20 (99.11%)
> 112,551320.60,550183.20 (99.79%)
> 128,549138.30,550542.70 (100.26%)
> 144,549344.50,544529.10 (99.12%)
> 160,550360.80,539929.30 (98.10%)
>
> slab_test 1 run on boot. After is faster except for odd result on size
> 2048.

Hmm... It's odd result. It adds more logic and it should
decrease performance. I guess it would be experimental error but
do you have any analysis about this result?

If it is for optimization, it would be one option to have separate
random list for each kmem_cache. It would consume more memory but it
would be marginal. And, it provides more un-predictability and it can
give better performance because we don't need state->type (more, less)
and special handling related for it.

Using pos = 0 here looks not good in terms of security. In this case,
every new page having same size class have same sequence of freelist since boot.

How about using random value to set pos? It provides some more randomness
with minimal overhead.

Please consider last object of OBJFREELIST_SLAB cache, too.

freelist_state_init()
last_obj = next_randome_slot()
page->freelist = XXX
for (i = 0; i < count - 1; i++)
set_free_obj()
set_free_obj(last_obj);

Thanks.

> +}
> +#else
> +static inline void shuffle_freelist(struct kmem_cache *cachep,
> + struct page *page, unsigned int count) { }
> +#endif /* CONFIG_FREELIST_RANDOM */
> +
> static void cache_init_objs(struct kmem_cache *cachep,
> struct page *page)
> {
> @@ -2464,8 +2624,12 @@ static void cache_init_objs(struct kmem_cache *cachep,
> kasan_poison_object_data(cachep, objp);
> }
>
> - set_free_obj(page, i, i);
> + /* If enabled, initialization is done in shuffle_freelist */
> + if (!config_enabled(CONFIG_FREELIST_RANDOM))
> + set_free_obj(page, i, i);
> }
> +
> + shuffle_freelist(cachep, page, cachep->num);
> }
>
> static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
> --
> 2.8.0.rc3.226.g39d4020
>

> --
> To unsubscribe, send a message with 'unsubscribe linux-mm' in
> the body to majo...@kvack.org. For more info on Linux MM,
> see: http://www.linux-mm.org/ .
> Don't email: <a href=mailto:"do...@kvack.org"> em...@kvack.org </a>

[Thomas Garnier]

On Tue, Apr 19, 2016 at 12:15 AM, Joonsoo Kim <iamjoon...@lge.com> wrote:
> On Mon, Apr 18, 2016 at 10:14:39AM -0700, Thomas Garnier wrote:
>> Provides an optional config (CONFIG_FREELIST_RANDOM) to randomize the
>> SLAB freelist. The list is randomized during initialization of a new set
>> of pages. The order on different freelist sizes is pre-computed at boot
>> for performance. This security feature reduces the predictability of the
>> kernel SLAB allocator against heap overflows rendering attacks much less
>> stable.
>
> I'm not familiar on security but it doesn't look much secure than
> before. Is there any other way to generate different sequence of freelist
> for each new set of pages? Current approach using pre-computed array will
> generate same sequence of freelist for all new set of pages having same size
> class. Is it sufficient?
>

I think it is sufficient. There is a tradeoff for performance. We could randomly
pick an object from the freelist every time (on slab_get_obj) but I
think it will
have significant impact (at least 3%).

>> For example this attack against SLUB (also applicable against SLAB)
>> would be affected:
>> https://jon.oberheide.org/blog/2010/09/10/linux-kernel-can-slub-overflow/
>>
>> Also, since v4.6 the freelist was moved at the end of the SLAB. It means
>> a controllable heap is opened to new attacks not yet publicly discussed.
>> A kernel heap overflow can be transformed to multiple use-after-free.
>> This feature makes this type of attack harder too.
>>
>> To generate entropy, we use get_random_bytes_arch because 0 bits of
>> entropy is available at that boot stage. In the worse case this function
>> will fallback to the get_random_bytes sub API.
>>
>> The config option name is not specific to the SLAB as this approach will
>> be extended to other allocators like SLUB.
>
> If this feature will be applied to the SLUB, it's better to put common
> code to mm/slab_common.c.
>

I think it might be moved there once we implement the SLUB counterpart
but it is too early to define which part will be common.

>>
>> Performance results highlighted no major changes:
>>
>> Netperf average on 10 runs:
>>
>> threads,base,change
>> 16,576943.10,585905.90 (101.55%)
>> 32,564082.00,569741.20 (101.00%)
>> 48,558334.30,561851.20 (100.63%)
>> 64,552025.20,556448.30 (100.80%)
>> 80,552294.40,551743.10 (99.90%)
>> 96,552435.30,547529.20 (99.11%)
>> 112,551320.60,550183.20 (99.79%)
>> 128,549138.30,550542.70 (100.26%)
>> 144,549344.50,544529.10 (99.12%)
>> 160,550360.80,539929.30 (98.10%)
>>
>> slab_test 1 run on boot. After is faster except for odd result on size
>> 2048.
>
> Hmm... It's odd result. It adds more logic and it should
> decrease performance. I guess it would be experimental error but
> do you have any analysis about this result?
>

I don't. I am glad to redo the test. I found that slab_test has very different
result based on the heap state at the time of the test. If I run the
test multiple
times, I have really various results on with or without the mitigation (on
dedicated hardware).

I am not sur because major caches are created early at boot time. We still have
the same entropy problem and we are wasting a bit more memory. It will be faster
on usage though but not sure it will be significant.

I think it is a good idea. I will add that for the next iteration.

The current implementation take the last chunk by default before the
freelist is initialized. Do you want it to be randomized as well?

[Joonsoo Kim]

I think that entropy problem is another issue. It should be considered
separately. If it is solved, making per-computed array for each
kmem_cache will provide more un-predictability. If someone who succeed to
exploit some kmem_cache with 128 object per slab want to exploit
another kmem_cache with 128 object per slab, this separate pre-computed array
will be helpful.

> on usage though but not sure it will be significant.

I also think it's not significant. But, besides performance effect,
code doesn't look very attractive and extendable. In case of SLUB,
there is setup_slub_max_order option and object per slab could be larger
than 256. To deal with it, we need to add many more static definition
and it looks not good to me. Please use dynamic allocated memory
instead of static array definition.

Yes.

Thanks.

[Thomas Garnier]

You don't need to. We wrap the list used (if you look at get_next_entry
we reset at pos 0 when we arrive to the list size).

I do think that the design will be better with a dedicated list per cache. Given
you seem fine with the memory differences, performance can only get better...

I will refactor for that on the next iteration.

[Thomas Garnier]

Provides an optional config (CONFIG_FREELIST_RANDOM) to randomize the
SLAB freelist. The list is randomized during initialization of a new set
of pages. The order on different freelist sizes is pre-computed at boot

for performance. Each kmem_cache has its own randomized freelist except
early on boot where global lists are used. This security feature reduces

the predictability of the kernel SLAB allocator against heap overflows
rendering attacks much less stable.

For example this attack against SLUB (also applicable against SLAB)
would be affected:
https://jon.oberheide.org/blog/2010/09/10/linux-kernel-can-slub-overflow/

Also, since v4.6 the freelist was moved at the end of the SLAB. It means
a controllable heap is opened to new attacks not yet publicly discussed.
A kernel heap overflow can be transformed to multiple use-after-free.
This feature makes this type of attack harder too.

To generate entropy, we use get_random_bytes_arch because 0 bits of

entropy is available in the boot stage. In the worse case this function
will fallback to the get_random_bytes sub API. We also generate a shift
random number to shift pre-computed freelist for each new set of pages.

The config option name is not specific to the SLAB as this approach will
be extended to other allocators like SLUB.

Performance results highlighted no major changes:

slab_test 1 run on boot. Difference only seen on the 2048 size test
being the worse case scenario covered by freelist randomization. New
slab pages are constantly being created on the 10000 allocations.
Variance should be mainly due to getting new pages every few
allocations.

Before:

Single thread testing
=====================
1. Kmalloc: Repeatedly allocate then free test

10000 times kmalloc(8) -> 99 cycles kfree -> 112 cycles
10000 times kmalloc(16) -> 109 cycles kfree -> 140 cycles
10000 times kmalloc(32) -> 129 cycles kfree -> 137 cycles
10000 times kmalloc(64) -> 141 cycles kfree -> 141 cycles
10000 times kmalloc(128) -> 152 cycles kfree -> 148 cycles
10000 times kmalloc(256) -> 195 cycles kfree -> 167 cycles
10000 times kmalloc(512) -> 257 cycles kfree -> 199 cycles
10000 times kmalloc(1024) -> 393 cycles kfree -> 251 cycles
10000 times kmalloc(2048) -> 649 cycles kfree -> 228 cycles
10000 times kmalloc(4096) -> 806 cycles kfree -> 370 cycles
10000 times kmalloc(8192) -> 814 cycles kfree -> 411 cycles
10000 times kmalloc(16384) -> 892 cycles kfree -> 455 cycles
2. Kmalloc: alloc/free test
10000 times kmalloc(8)/kfree -> 121 cycles
10000 times kmalloc(16)/kfree -> 121 cycles
10000 times kmalloc(32)/kfree -> 121 cycles

10000 times kmalloc(64)/kfree -> 121 cycles

10000 times kmalloc(128)/kfree -> 121 cycles

10000 times kmalloc(256)/kfree -> 119 cycles

10000 times kmalloc(512)/kfree -> 119 cycles
10000 times kmalloc(1024)/kfree -> 119 cycles
10000 times kmalloc(2048)/kfree -> 119 cycles
10000 times kmalloc(4096)/kfree -> 121 cycles
10000 times kmalloc(8192)/kfree -> 119 cycles
10000 times kmalloc(16384)/kfree -> 119 cycles

After:

Single thread testing
=====================
1. Kmalloc: Repeatedly allocate then free test

10000 times kmalloc(8) -> 130 cycles kfree -> 86 cycles
10000 times kmalloc(16) -> 118 cycles kfree -> 86 cycles
10000 times kmalloc(32) -> 121 cycles kfree -> 85 cycles
10000 times kmalloc(64) -> 176 cycles kfree -> 102 cycles
10000 times kmalloc(128) -> 178 cycles kfree -> 100 cycles
10000 times kmalloc(256) -> 205 cycles kfree -> 109 cycles
10000 times kmalloc(512) -> 262 cycles kfree -> 136 cycles
10000 times kmalloc(1024) -> 342 cycles kfree -> 157 cycles
10000 times kmalloc(2048) -> 701 cycles kfree -> 238 cycles
10000 times kmalloc(4096) -> 803 cycles kfree -> 364 cycles
10000 times kmalloc(8192) -> 835 cycles kfree -> 404 cycles
10000 times kmalloc(16384) -> 896 cycles kfree -> 441 cycles
2. Kmalloc: alloc/free test
10000 times kmalloc(8)/kfree -> 121 cycles
10000 times kmalloc(16)/kfree -> 121 cycles
10000 times kmalloc(32)/kfree -> 123 cycles
10000 times kmalloc(64)/kfree -> 142 cycles
10000 times kmalloc(128)/kfree -> 121 cycles

10000 times kmalloc(256)/kfree -> 119 cycles

10000 times kmalloc(512)/kfree -> 119 cycles
10000 times kmalloc(1024)/kfree -> 119 cycles
10000 times kmalloc(2048)/kfree -> 119 cycles
10000 times kmalloc(4096)/kfree -> 119 cycles
10000 times kmalloc(8192)/kfree -> 119 cycles
10000 times kmalloc(16384)/kfree -> 119 cycles

Signed-off-by: Thomas Garnier <thga...@google.com>
---

Based on next-20160422
---
include/linux/slab_def.h | 4 +
init/Kconfig | 9 ++
mm/slab.c | 213 ++++++++++++++++++++++++++++++++++++++++++++++-
3 files changed, 224 insertions(+), 2 deletions(-)

diff --git a/include/linux/slab_def.h b/include/linux/slab_def.h
index 9edbbf3..182ec26 100644
--- a/include/linux/slab_def.h
+++ b/include/linux/slab_def.h
@@ -80,6 +80,10 @@ struct kmem_cache {
struct kasan_cache kasan_info;
#endif

+#ifdef CONFIG_FREELIST_RANDOM
+ void *random_seq;
+#endif
+
struct kmem_cache_node *node[MAX_NUMNODES];
};

diff --git a/init/Kconfig b/init/Kconfig
index 0c66640..73453d0 100644

--- a/init/Kconfig
+++ b/init/Kconfig
@@ -1742,6 +1742,15 @@ config SLOB

endchoice

+config FREELIST_RANDOM
+ default n
+ depends on SLAB
+ bool "SLAB freelist randomization"
+ help
+ Randomizes the freelist order used on creating new SLABs. This
+ security feature reduces the predictability of the kernel slab
+ allocator against heap overflows.
+
config SLUB_CPU_PARTIAL
default y
depends on SLUB && SMP
diff --git a/mm/slab.c b/mm/slab.c

index b82ee6b..89eb617 100644

--- a/mm/slab.c
+++ b/mm/slab.c
@@ -116,6 +116,7 @@
#include <linux/kmemcheck.h>
#include <linux/memory.h>
#include <linux/prefetch.h>
+#include <linux/log2.h>

#include <net/sock.h>

@@ -1230,6 +1231,100 @@ static void __init set_up_node(struct kmem_cache *cachep, int index)
}
}

+#ifdef CONFIG_FREELIST_RANDOM
+static void freelist_randomize(struct rnd_state *state, freelist_idx_t *list,
+ size_t count)
+{
+ size_t i;
+ unsigned int rand;
+

+ for (i = 0; i < count; i++)

+ list[i] = i;

+
+ /* Fisher-Yates shuffle */

+ for (i = count - 1; i > 0; i--) {
+ rand = prandom_u32_state(state);

+ rand %= (i + 1);

+ swap(list[i], list[rand]);
+ }
+}
+
+/* Create a random sequence per cache */
+static void cache_random_seq_create(struct kmem_cache *cachep)
+{
+ unsigned int seed, count = cachep->num;
+ struct rnd_state state;

+
+ if (count < 2)

+ return;
+
+ cachep->random_seq = kcalloc(count, sizeof(freelist_idx_t), GFP_KERNEL);
+ BUG_ON(cachep->random_seq == NULL);
+
+ /* Get best entropy at this stage */
+ get_random_bytes_arch(&seed, sizeof(seed));
+ prandom_seed_state(&state, seed);
+
+ freelist_randomize(&state, cachep->random_seq, count);
+}
+
+/* Destroy the per-cache random freelist sequence */
+static void cache_random_seq_destroy(struct kmem_cache *cachep)
+{
+ kfree(cachep->random_seq);
+ cachep->random_seq = NULL;
+}
+
+/*
+ * Global static list are used when pre-computed cache list are not yet
+ * available. Lists of different sizes are created to optimize performance on
+ * SLABS with different object counts.
+ */
+static freelist_idx_t freelist_random_seq_2[2];
+static freelist_idx_t freelist_random_seq_4[4];
+static freelist_idx_t freelist_random_seq_8[8];
+static freelist_idx_t freelist_random_seq_16[16];
+static freelist_idx_t freelist_random_seq_32[32];
+static freelist_idx_t freelist_random_seq_64[64];
+static freelist_idx_t freelist_random_seq_128[128];
+static freelist_idx_t freelist_random_seq_256[256];

+const static struct m_list {
+ size_t count;
+ freelist_idx_t *list;

+} freelist_random_seqs[] = {
+ { ARRAY_SIZE(freelist_random_seq_2), freelist_random_seq_2 },
+ { ARRAY_SIZE(freelist_random_seq_4), freelist_random_seq_4 },
+ { ARRAY_SIZE(freelist_random_seq_8), freelist_random_seq_8 },
+ { ARRAY_SIZE(freelist_random_seq_16), freelist_random_seq_16 },
+ { ARRAY_SIZE(freelist_random_seq_32), freelist_random_seq_32 },
+ { ARRAY_SIZE(freelist_random_seq_64), freelist_random_seq_64 },
+ { ARRAY_SIZE(freelist_random_seq_128), freelist_random_seq_128 },
+ { ARRAY_SIZE(freelist_random_seq_256), freelist_random_seq_256 },
+};
+
+/* Pre-compute the global pre-computed lists early at boot */

+static void __init freelist_random_init(void)
+{
+ unsigned int seed;

+ size_t i;
+ struct rnd_state state;
+
+ /* Get best entropy available at this stage */
+ get_random_bytes_arch(&seed, sizeof(seed));
+ prandom_seed_state(&state, seed);
+
+ for (i = 0; i < ARRAY_SIZE(freelist_random_seqs); i++) {
+ freelist_randomize(&state, freelist_random_seqs[i].list,
+ freelist_random_seqs[i].count);

+ }
+}
+#else
+static inline void __init freelist_random_init(void) { }

+static inline void cache_random_seq_create(struct kmem_cache *cachep) { }
+static inline void cache_random_seq_destroy(struct kmem_cache *cachep) { }

+#endif /* CONFIG_FREELIST_RANDOM */
+
+
/*
* Initialisation. Called after the page allocator have been initialised and
* before smp_init().

@@ -1256,6 +1351,8 @@ void __init kmem_cache_init(void)

if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
slab_max_order = SLAB_MAX_ORDER_HI;

+ freelist_random_init();
+
/* Bootstrap is tricky, because several objects are allocated
* from caches that do not exist yet:
* 1) initialize the kmem_cache cache: it contains the struct

@@ -2337,6 +2434,8 @@ void __kmem_cache_release(struct kmem_cache *cachep)
int i;
struct kmem_cache_node *n;

+ cache_random_seq_destroy(cachep);
+
free_percpu(cachep->cpu_cache);

/* NUMA: free the node structures */
@@ -2443,15 +2542,122 @@ static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page)
#endif
}

+#ifdef CONFIG_FREELIST_RANDOM

+/* Hold information during a freelist initialization */
+struct freelist_init_state {
+ unsigned int padding;
+ unsigned int pos;
+ unsigned int count;

+ unsigned int rand;
+ struct m_list freelist_random_seq;
+};
+
+/* Select the right pre-computed list and initialize state */

+static void freelist_state_initialize(struct freelist_init_state *state,

+ struct kmem_cache *cachep,

+ unsigned int count)
+{
+ unsigned int idx;

+ const unsigned int last_idx = ARRAY_SIZE(freelist_random_seqs) - 1;

+
+ memset(state, 0, sizeof(*state));
+ state->count = count;
+ state->pos = 0;

+
+ /* Use best entropy available to define a random shift */
+ get_random_bytes_arch(&state->rand, sizeof(state->rand));
+
+ if (cachep->random_seq) {
+ state->freelist_random_seq.list = cachep->random_seq;
+ state->freelist_random_seq.count = count;
+ } else {

+ /* count is always >= 2 */
+ idx = ilog2(count) - 1;
+ if (idx >= last_idx)
+ idx = last_idx;
+ else if (roundup_pow_of_two(idx + 1) != count)
+ idx++;

+ state->freelist_random_seq = freelist_random_seqs[idx];
+ }
+}
+
+/* Get the next entry on the list depending on the target list size */

+static freelist_idx_t get_next_entry(struct freelist_init_state *state)
+{

+ freelist_idx_t ret;
+
+ if (state->pos == state->freelist_random_seq.count) {

+ state->padding += state->pos;
+ state->pos = 0;
+ }
+

+ /* Randomize the entry using the random shift */
+ ret = state->freelist_random_seq.list[state->pos++];
+ ret = (ret + state->rand) % state->freelist_random_seq.count;
+ return ret;

+}
+
+static freelist_idx_t next_random_slot(struct freelist_init_state *state)
+{

+ freelist_idx_t entry;
+
+ do {
+ entry = get_next_entry(state);

+ } while ((entry + state->padding) >= state->count);
+

+ return entry + state->padding;
+}
+
+/*
+ * Shuffle the freelist initialization state based on pre-computed lists.
+ * return true if the list was successfully shuffled, false otherwise.
+ */
+static bool shuffle_freelist(struct kmem_cache *cachep, struct page *page)
+{
+ unsigned int objfreelist, i, count = cachep->num;

+ struct freelist_init_state state;
+
+ if (count < 2)

+ return false;
+
+ objfreelist = 0;
+ freelist_state_initialize(&state, cachep, count);
+
+ /* Take the first random entry as the objfreelist */
+ if (OBJFREELIST_SLAB(cachep)) {
+ objfreelist = next_random_slot(&state);
+ page->freelist = index_to_obj(cachep, page, objfreelist) +
+ obj_offset(cachep);
+ count--;
+ }

+ for (i = 0; i < count; i++)
+ set_free_obj(page, i, next_random_slot(&state));
+
+ if (OBJFREELIST_SLAB(cachep))

+ set_free_obj(page, i, objfreelist);
+ return true;
+}
+#else
+static inline bool shuffle_freelist(struct kmem_cache *cachep,
+ struct page *page)
+{
+ return false;
+}

+#endif /* CONFIG_FREELIST_RANDOM */
+
static void cache_init_objs(struct kmem_cache *cachep,
struct page *page)
{

int i;
void *objp;
+ bool shuffled;

cache_init_objs_debug(cachep, page);

- if (OBJFREELIST_SLAB(cachep)) {
+ /* Try to randomize the freelist if enabled */
+ shuffled = shuffle_freelist(cachep, page);
+
+ if (!shuffled && OBJFREELIST_SLAB(cachep)) {
page->freelist = index_to_obj(cachep, page, cachep->num - 1) +
obj_offset(cachep);
}
@@ -2465,7 +2671,8 @@ static void cache_init_objs(struct kmem_cache *cachep,

kasan_poison_object_data(cachep, objp);
}

- set_free_obj(page, i, i);

+ if (!shuffled)
+ set_free_obj(page, i, i);
}
}

@@ -3815,6 +4022,8 @@ static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
int shared = 0;
int batchcount = 0;

+ cache_random_seq_create(cachep);
+
if (!is_root_cache(cachep)) {
struct kmem_cache *root = memcg_root_cache(cachep);
limit = root->limit;
--
2.8.0.rc3.226.g39d4020

[Andrew Morton]

On Mon, 18 Apr 2016 12:52:30 -0700 Thomas Garnier <thga...@google.com> wrote:

> I agree, if we had a generic way to pass entropy across boots on all
> architecture that would be amazing. I will let the SLAB maintainers to
> decide on requiring CONFIG_ARCH_RANDOM or documenting it.

In our world, requiring that sort of attention from maintainers
requires a pretty active level of pinging, poking and harrassing ;)

I do think that if you stick with get_random_bytes_arch() then it need
a comment explaining why.

And I (still) don't think that get_random_bytes_arch() actually does
what you want - if CONFIG_ARCH_RANDOM isn't implemented then
get_random_bytes_arch() just fails. IOW your statement "the arch

version that will fallback on get_random_bytes sub API in the worse

case" is a misconception? There is no fallback.

[Andrew Morton]

> ...

>
> --- a/include/linux/slab_def.h
> +++ b/include/linux/slab_def.h
> @@ -80,6 +80,10 @@ struct kmem_cache {
> struct kasan_cache kasan_info;
> #endif
>
> +#ifdef CONFIG_FREELIST_RANDOM

CONFIG_FREELIST_RANDOM bugs me a bit - "freelist" is so vague.
CONFIG_SLAB_FREELIST_RANDOM would be better. I mean, what Kconfig
identifier could be used for implementing randomisation in
slub/slob/etc once CONFIG_FREELIST_RANDOM is used up?

Yikes, that's a bit rude. Is there no way of recovering from this? If
the answer to that is really really "no" then I guess we should put a
__GFP_NOFAIL in there. Add a comment explaining why (apologetically -
__GFP_NOFAIL is unpopular!) and remove the now-unneeded BUG_ON.

> + /* Get best entropy at this stage */
> + get_random_bytes_arch(&seed, sizeof(seed));

See concerns in other email - isn't this a no-op if CONFIG_ARCH_RANDOM=n?

[Thomas Garnier]

On your previous email. (trying to stay in one thread). I added a
comment on this
version to explain that we need best entropy at this boot stage.

>
> Yikes, that's a bit rude. Is there no way of recovering from this? If
> the answer to that is really really "no" then I guess we should put a
> __GFP_NOFAIL in there. Add a comment explaining why (apologetically -
> __GFP_NOFAIL is unpopular!) and remove the now-unneeded BUG_ON.
>
>

We can always use the static. I will update on next iteration to remove the

[Thomas Garnier]

The arch_* functions will return 0 which will break the loop in
get_random_bytes_arch and make it uses extract_entropy (as does
get_random_bytes).
(cf http://lxr.free-electrons.com/source/drivers/char/random.c#L1335)

I might be missing something.

[Andrew Morton]

On Mon, 25 Apr 2016 14:14:33 -0700 Thomas Garnier <thga...@google.com> wrote:

> >>> + /* Get best entropy at this stage */
> >>> + get_random_bytes_arch(&seed, sizeof(seed));
> >>
> >> See concerns in other email - isn't this a no-op if CONFIG_ARCH_RANDOM=n?
> >>
>
> The arch_* functions will return 0 which will break the loop in
> get_random_bytes_arch and make it uses extract_entropy (as does
> get_random_bytes).
> (cf http://lxr.free-electrons.com/source/drivers/char/random.c#L1335)
>

oop, sorry, I misread the code.

(and the get_random_bytes_arch() comment "This function will use the
architecture-specific hardware random number generator if it is
available" is misleading, so there)

[Thomas Garnier]

No problem, better double check it. I agree it is misleading.

Thomas

[Joonsoo Kim]

Hello,

Please make function return int and propagate error to the cache creator.

I'd like to remove this global static list even if we can't get random
sequence in early boot-up process. In this stage that kernel is not
yet initialized, malicious user cannot do anything so random sequence
doesn't give any more security. After kernel initialization, we will
use per cache random sequence so problem suface is really small. If you
want to randomize freelist sequence even in this case, you can manually
permute the sequence with calling prandom_u32_state(). But, I don't
think it is necessary.

Thanks.

[Thomas Garnier]

Make sense. I think it is still valuable to randomize earlier pages. I
will adapt the code, test and send patch v4.

Thanks for the quick feedback,
Thomas

[Christoph Lameter]

On Mon, 25 Apr 2016, Thomas Garnier wrote:

> To generate entropy, we use get_random_bytes_arch because 0 bits of
> entropy is available in the boot stage. In the worse case this function
> will fallback to the get_random_bytes sub API. We also generate a shift
> random number to shift pre-computed freelist for each new set of pages.
>
> The config option name is not specific to the SLAB as this approach will
> be extended to other allocators like SLUB.
>
> Performance results highlighted no major changes:

Ok. alloc/free tests are not affected since this exercises the per cpu
objects. And the other ones as well since most of the overhead occurs on
slab page initialization.

> Before:

> 10000 times kmalloc(1024) -> 393 cycles kfree -> 251 cycles
> 10000 times kmalloc(2048) -> 649 cycles kfree -> 228 cycles
> 10000 times kmalloc(4096) -> 806 cycles kfree -> 370 cycles
> 10000 times kmalloc(8192) -> 814 cycles kfree -> 411 cycles
> 10000 times kmalloc(16384) -> 892 cycles kfree -> 455 cycles
>

> After:

> 10000 times kmalloc(1024) -> 342 cycles kfree -> 157 cycles
> 10000 times kmalloc(2048) -> 701 cycles kfree -> 238 cycles
> 10000 times kmalloc(4096) -> 803 cycles kfree -> 364 cycles
> 10000 times kmalloc(8192) -> 835 cycles kfree -> 404 cycles
> 10000 times kmalloc(16384) -> 896 cycles kfree -> 441 cycles

And there is some slight regression with the larger objects. Not sure if
we are really hitting the slab page initialization too much there either.
Pretty minimal in synthetic tests. Can you run something like hackbench
too?

Otherwise this looks ok.

Acked-by: Christoph Lameter <c...@linux.com>

[Thomas Garnier]

Provides an optional config (CONFIG_FREELIST_RANDOM) to randomize the
SLAB freelist. The list is randomized during initialization of a new set
of pages. The order on different freelist sizes is pre-computed at boot

for performance. Each kmem_cache has its own randomized freelist. Before
pre-computed lists are available freelists are generated
dynamically. This security feature reduces the predictability of the

kernel SLAB allocator against heap overflows rendering attacks much less
stable.

For example this attack against SLUB (also applicable against SLAB)
would be affected:
https://jon.oberheide.org/blog/2010/09/10/linux-kernel-can-slub-overflow/

Also, since v4.6 the freelist was moved at the end of the SLAB. It means
a controllable heap is opened to new attacks not yet publicly discussed.
A kernel heap overflow can be transformed to multiple use-after-free.
This feature makes this type of attack harder too.

To generate entropy, we use get_random_bytes_arch because 0 bits of
entropy is available in the boot stage. In the worse case this function
will fallback to the get_random_bytes sub API. We also generate a shift
random number to shift pre-computed freelist for each new set of pages.

The config option name is not specific to the SLAB as this approach will
be extended to other allocators like SLUB.

Performance results highlighted no major changes:

Hackbench (running 90 10 times):

Before average: 0.0698
After average: 0.0663 (-5.01%)

slab_test 1 run on boot. Difference only seen on the 2048 size test
being the worse case scenario covered by freelist randomization. New
slab pages are constantly being created on the 10000 allocations.
Variance should be mainly due to getting new pages every few
allocations.

Before:

Single thread testing
=====================
1. Kmalloc: Repeatedly allocate then free test
10000 times kmalloc(8) -> 99 cycles kfree -> 112 cycles
10000 times kmalloc(16) -> 109 cycles kfree -> 140 cycles
10000 times kmalloc(32) -> 129 cycles kfree -> 137 cycles
10000 times kmalloc(64) -> 141 cycles kfree -> 141 cycles
10000 times kmalloc(128) -> 152 cycles kfree -> 148 cycles
10000 times kmalloc(256) -> 195 cycles kfree -> 167 cycles
10000 times kmalloc(512) -> 257 cycles kfree -> 199 cycles

10000 times kmalloc(1024) -> 393 cycles kfree -> 251 cycles
10000 times kmalloc(2048) -> 649 cycles kfree -> 228 cycles
10000 times kmalloc(4096) -> 806 cycles kfree -> 370 cycles
10000 times kmalloc(8192) -> 814 cycles kfree -> 411 cycles
10000 times kmalloc(16384) -> 892 cycles kfree -> 455 cycles

2. Kmalloc: alloc/free test
10000 times kmalloc(8)/kfree -> 121 cycles
10000 times kmalloc(16)/kfree -> 121 cycles
10000 times kmalloc(32)/kfree -> 121 cycles
10000 times kmalloc(64)/kfree -> 121 cycles
10000 times kmalloc(128)/kfree -> 121 cycles
10000 times kmalloc(256)/kfree -> 119 cycles
10000 times kmalloc(512)/kfree -> 119 cycles
10000 times kmalloc(1024)/kfree -> 119 cycles
10000 times kmalloc(2048)/kfree -> 119 cycles
10000 times kmalloc(4096)/kfree -> 121 cycles
10000 times kmalloc(8192)/kfree -> 119 cycles
10000 times kmalloc(16384)/kfree -> 119 cycles

After:

Single thread testing
=====================
1. Kmalloc: Repeatedly allocate then free test
10000 times kmalloc(8) -> 130 cycles kfree -> 86 cycles
10000 times kmalloc(16) -> 118 cycles kfree -> 86 cycles
10000 times kmalloc(32) -> 121 cycles kfree -> 85 cycles
10000 times kmalloc(64) -> 176 cycles kfree -> 102 cycles
10000 times kmalloc(128) -> 178 cycles kfree -> 100 cycles
10000 times kmalloc(256) -> 205 cycles kfree -> 109 cycles
10000 times kmalloc(512) -> 262 cycles kfree -> 136 cycles

10000 times kmalloc(1024) -> 342 cycles kfree -> 157 cycles
10000 times kmalloc(2048) -> 701 cycles kfree -> 238 cycles
10000 times kmalloc(4096) -> 803 cycles kfree -> 364 cycles
10000 times kmalloc(8192) -> 835 cycles kfree -> 404 cycles
10000 times kmalloc(16384) -> 896 cycles kfree -> 441 cycles

2. Kmalloc: alloc/free test
10000 times kmalloc(8)/kfree -> 121 cycles
10000 times kmalloc(16)/kfree -> 121 cycles
10000 times kmalloc(32)/kfree -> 123 cycles
10000 times kmalloc(64)/kfree -> 142 cycles
10000 times kmalloc(128)/kfree -> 121 cycles
10000 times kmalloc(256)/kfree -> 119 cycles
10000 times kmalloc(512)/kfree -> 119 cycles
10000 times kmalloc(1024)/kfree -> 119 cycles
10000 times kmalloc(2048)/kfree -> 119 cycles
10000 times kmalloc(4096)/kfree -> 119 cycles
10000 times kmalloc(8192)/kfree -> 119 cycles
10000 times kmalloc(16384)/kfree -> 119 cycles

Signed-off-by: Thomas Garnier <thga...@google.com>

Acked-by: Christoph Lameter <c...@linux.com>

---
Based on next-20160422
---

include/linux/slab_def.h | 4 ++
init/Kconfig | 9 +++
mm/slab.c | 167 ++++++++++++++++++++++++++++++++++++++++++++++-
3 files changed, 178 insertions(+), 2 deletions(-)

index b82ee6b..0ed728a 100644
--- a/mm/slab.c
+++ b/mm/slab.c
@@ -1230,6 +1230,61 @@ static void __init set_up_node(struct kmem_cache *cachep, int index)

}
}

+#ifdef CONFIG_FREELIST_RANDOM
+static void freelist_randomize(struct rnd_state *state, freelist_idx_t *list,
+ size_t count)
+{
+ size_t i;
+ unsigned int rand;
+
+ for (i = 0; i < count; i++)
+ list[i] = i;
+
+ /* Fisher-Yates shuffle */
+ for (i = count - 1; i > 0; i--) {
+ rand = prandom_u32_state(state);
+ rand %= (i + 1);
+ swap(list[i], list[rand]);
+ }
+}
+
+/* Create a random sequence per cache */

+static int cache_random_seq_create(struct kmem_cache *cachep)

+{
+ unsigned int seed, count = cachep->num;
+ struct rnd_state state;
+
+ if (count < 2)

+ return 0;
+
+ /* If it fails, we will just use the global lists */

+ cachep->random_seq = kcalloc(count, sizeof(freelist_idx_t), GFP_KERNEL);

+ if (!cachep->random_seq)
+ return -ENOMEM;

+
+ /* Get best entropy at this stage */
+ get_random_bytes_arch(&seed, sizeof(seed));
+ prandom_seed_state(&state, seed);
+
+ freelist_randomize(&state, cachep->random_seq, count);

+ return 0;

+}
+
+/* Destroy the per-cache random freelist sequence */
+static void cache_random_seq_destroy(struct kmem_cache *cachep)
+{
+ kfree(cachep->random_seq);
+ cachep->random_seq = NULL;
+}

+#else
+static inline int cache_random_seq_create(struct kmem_cache *cachep)
+{
+ return 0;
+}

+static inline void cache_random_seq_destroy(struct kmem_cache *cachep) { }
+#endif /* CONFIG_FREELIST_RANDOM */
+
+
/*
* Initialisation. Called after the page allocator have been initialised and
* before smp_init().

@@ -2337,6 +2392,8 @@ void __kmem_cache_release(struct kmem_cache *cachep)

int i;
struct kmem_cache_node *n;

+ cache_random_seq_destroy(cachep);
+
free_percpu(cachep->cpu_cache);

/* NUMA: free the node structures */

@@ -2443,15 +2500,115 @@ static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page)

#endif
}

+#ifdef CONFIG_FREELIST_RANDOM
+/* Hold information during a freelist initialization */

+union freelist_init_state {
+ struct {
+ unsigned int pos;
+ freelist_idx_t *list;

+ unsigned int count;
+ unsigned int rand;

+ };
+ struct rnd_state rnd_state;
+};
+
+/*
+ * Initialize the state based on the randomization methode available.
+ * return true if the pre-computed list is available, false otherwize.
+ */
+static bool freelist_state_initialize(union freelist_init_state *state,

+ struct kmem_cache *cachep,
+ unsigned int count)
+{

+ bool ret;

+ unsigned int rand;
+

+ /* Use best entropy available to define a random shift */

+ get_random_bytes_arch(&rand, sizeof(rand));
+
+ /* Use a random state if the pre-computed list is not available */
+ if (!cachep->random_seq) {
+ prandom_seed_state(&state->rnd_state, rand);
+ ret = false;
+ } else {
+ state->list = cachep->random_seq;

+ state->count = count;
+ state->pos = 0;

+ state->rand = rand;
+ ret = true;
+ }

+ return ret;
+}
+

+/* Get the next entry on the list and randomize it using a random shift */
+static freelist_idx_t next_random_slot(union freelist_init_state *state)
+{
+ return (state->list[state->pos++] + state->rand) % state->count;

+}
+
+/*
+ * Shuffle the freelist initialization state based on pre-computed lists.
+ * return true if the list was successfully shuffled, false otherwise.
+ */
+static bool shuffle_freelist(struct kmem_cache *cachep, struct page *page)
+{

+ unsigned int objfreelist = 0, i, count = cachep->num;
+ union freelist_init_state state;
+ bool precomputed;

+
+ if (count < 2)
+ return false;
+

+ precomputed = freelist_state_initialize(&state, cachep, count);
+
+ /* Take a random entry as the objfreelist */
+ if (OBJFREELIST_SLAB(cachep)) {
+ if (!precomputed)
+ objfreelist = count - 1;
+ else

+ objfreelist = next_random_slot(&state);
+ page->freelist = index_to_obj(cachep, page, objfreelist) +
+ obj_offset(cachep);
+ count--;
+ }
+

+ /*
+ * On early boot, generate the list dynamically.
+ * Later use a pre-computed list for speed.
+ */
+ if (!precomputed) {
+ freelist_randomize(&state.rnd_state, page->freelist, count);
+ } else {

+ for (i = 0; i < count; i++)
+ set_free_obj(page, i, next_random_slot(&state));
+ }
+
+ if (OBJFREELIST_SLAB(cachep))

+ set_free_obj(page, cachep->num - 1, objfreelist);
+

+ return true;
+}
+#else
+static inline bool shuffle_freelist(struct kmem_cache *cachep,
+ struct page *page)
+{
+ return false;
+}
+#endif /* CONFIG_FREELIST_RANDOM */
+
static void cache_init_objs(struct kmem_cache *cachep,
struct page *page)
{
int i;
void *objp;
+ bool shuffled;

cache_init_objs_debug(cachep, page);

- if (OBJFREELIST_SLAB(cachep)) {
+ /* Try to randomize the freelist if enabled */
+ shuffled = shuffle_freelist(cachep, page);
+
+ if (!shuffled && OBJFREELIST_SLAB(cachep)) {
page->freelist = index_to_obj(cachep, page, cachep->num - 1) +
obj_offset(cachep);
}

@@ -2465,7 +2622,8 @@ static void cache_init_objs(struct kmem_cache *cachep,

kasan_poison_object_data(cachep, objp);
}

- set_free_obj(page, i, i);
+ if (!shuffled)
+ set_free_obj(page, i, i);
}
}

@@ -3815,6 +3973,10 @@ static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)

int shared = 0;
int batchcount = 0;

+ err = cache_random_seq_create(cachep);
+ if (err)
+ goto end;

+
if (!is_root_cache(cachep)) {
struct kmem_cache *root = memcg_root_cache(cachep);
limit = root->limit;

@@ -3868,6 +4030,7 @@ static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
batchcount = (limit + 1) / 2;
skip_setup:
err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
+end:
if (err)
pr_err("enable_cpucache failed for %s, error %d\n",
cachep->name, -err);
--
2.8.0.rc3.226.g39d4020

[Andrew Morton]

> ...

>
> --- a/include/linux/slab_def.h
> +++ b/include/linux/slab_def.h
> @@ -80,6 +80,10 @@ struct kmem_cache {
> struct kasan_cache kasan_info;
> #endif
>
> +#ifdef CONFIG_FREELIST_RANDOM
> + void *random_seq;
> +#endif
> +
> struct kmem_cache_node *node[MAX_NUMNODES];
> };
>
> diff --git a/init/Kconfig b/init/Kconfig
> index 0c66640..73453d0 100644
> --- a/init/Kconfig
> +++ b/init/Kconfig
> @@ -1742,6 +1742,15 @@ config SLOB
>
> endchoice
>
> +config FREELIST_RANDOM
> + default n
> + depends on SLAB
> + bool "SLAB freelist randomization"
> + help
> + Randomizes the freelist order used on creating new SLABs. This
> + security feature reduces the predictability of the kernel slab
> + allocator against heap overflows.

Against the v2 patch I didst observe:

: CONFIG_FREELIST_RANDOM bugs me a bit - "freelist" is so vague.

: CONFIG_SLAB_FREELIST_RANDOM would be better. I mean, what Kconfig
: identifier could be used for implementing randomisation in
: slub/slob/etc once CONFIG_FREELIST_RANDOM is used up?

but this pearl appeared to pass unnoticed.

OK, no BUG. If this happens, kmem_cache_init_late() will go BUG
instead ;)

Questions for slab maintainers:

What's going on with the gfp_flags in there? kmem_cache_init_late()
passes GFP_NOWAIT into enable_cpucache().

a) why the heck does it do that? It's __init code!

b) if there's a legit reason then your new cache_random_seq_create()
should be getting its gfp_t from its caller, rather than blindly
assuming GFP_KERNEL.

c) kmem_cache_init_late() goes BUG on ENOMEM. Generally that's OK in
__init code: we assume infinite memory during bootup. But it's really
quite weird to use GFP_NOWAIT and then to go BUG if GFP_NOWAIT had its
predictable outcome (ie: failure).

Finally, all callers of enable_cpucache() (and hence of
cache_random_seq_create()) are __init, so we're unnecessarily bloating
up vmlinux. Could someone please take a look at this as a separate
thing?

> + /* Get best entropy at this stage */
> + get_random_bytes_arch(&seed, sizeof(seed));
> + prandom_seed_state(&state, seed);
> +
> + freelist_randomize(&state, cachep->random_seq, count);
> + return 0;
> +}
> +
>

> ...
>

[Thomas Garnier]

It was discussed a bit before. The intent is to have a similar feature
for other kernel heap (I know it is possible for SLUB). That's why I
think it make sense to have a similar config name used for all
allocators.

Yes, as Christophe asked.

[Joonsoo Kim]

Until some boot-up point, we should not enable interrupt.
In slab subsystem, If we use __GFP_DIRECT_RECLAIM, it will cause to
enable interrupt when allocating new slab page. GFP_NOWAIT is to
prevent that situation.

Anyway, I audit the code and kmem_cache_init_late() could use
__GFP_DIRECT_RECLAIM because it is called after interrupt is enabled
which means that that's safe time to manipulate interrupt. (See
kmem_cache_init_late() in start_kernel()).

>
> b) if there's a legit reason then your new cache_random_seq_create()
> should be getting its gfp_t from its caller, rather than blindly
> assuming GFP_KERNEL.

In any case, ignoring provided gfp argument isn't good practice.

> c) kmem_cache_init_late() goes BUG on ENOMEM. Generally that's OK in
> __init code: we assume infinite memory during bootup. But it's really
> quite weird to use GFP_NOWAIT and then to go BUG if GFP_NOWAIT had its
> predictable outcome (ie: failure).

I don't think BUG() here is weird code. It just means that if we can't
initialize slab subsystem properly, machine cannot run properly so
BUG().

> Finally, all callers of enable_cpucache() (and hence of
> cache_random_seq_create()) are __init, so we're unnecessarily bloating
> up vmlinux. Could someone please take a look at this as a separate
> thing?

That's not true. It is called whenever new kmem_cache is created.
I don't know concrete reason why setup_cpu_cache() is defined with
__init_refok tag but looks like it needs to be fixed.

I will look at it soon.

Thanks.

[Christoph Lameter]

On Tue, 26 Apr 2016, Andrew Morton wrote:

> : CONFIG_FREELIST_RANDOM bugs me a bit - "freelist" is so vague.
> : CONFIG_SLAB_FREELIST_RANDOM would be better. I mean, what Kconfig
> : identifier could be used for implementing randomisation in
> : slub/slob/etc once CONFIG_FREELIST_RANDOM is used up?
>
> but this pearl appeared to pass unnoticed.

Ok. lets add SLAB here and then use this option for the other allocators
as well.

> > + /* If it fails, we will just use the global lists */
> > + cachep->random_seq = kcalloc(count, sizeof(freelist_idx_t), GFP_KERNEL);
> > + if (!cachep->random_seq)
> > + return -ENOMEM;
>
> OK, no BUG. If this happens, kmem_cache_init_late() will go BUG
> instead ;)
>
> Questions for slab maintainers:
>
> What's going on with the gfp_flags in there? kmem_cache_init_late()
> passes GFP_NOWAIT into enable_cpucache().
>
> a) why the heck does it do that? It's __init code!

enable_cpucache() was called when a slab cache was reconfigured by writing to /proc/slabinfo.
That was changed awhile back when the memcg changes were made ot slab. So
now its ok to be made init code.

> Finally, all callers of enable_cpucache() (and hence of
> cache_random_seq_create()) are __init, so we're unnecessarily bloating
> up vmlinux. Could someone please take a look at this as a separate
> thing?

Hmmm. Well if that is the case then lots of stuff could be straightened
out. Joonsoo?

[Christoph Lameter]

On Tue, 26 Apr 2016, Thomas Garnier wrote:

> It was discussed a bit before. The intent is to have a similar feature
> for other kernel heap (I know it is possible for SLUB). That's why I
> think it make sense to have a similar config name used for all
> allocators.

Please use CONFIG_SLAB_FREELIST_RANDOM to signify that it is for all slab
allocators. Not SLAB specific.

[Thomas Garnier]

Provides an optional config (CONFIG_SLAB_FREELIST_RANDOM) to randomize

index 9edbbf3..8694f7a 100644

--- a/include/linux/slab_def.h
+++ b/include/linux/slab_def.h
@@ -80,6 +80,10 @@ struct kmem_cache {
struct kasan_cache kasan_info;
#endif

+#ifdef CONFIG_SLAB_FREELIST_RANDOM

+ void *random_seq;
+#endif
+
struct kmem_cache_node *node[MAX_NUMNODES];
};

diff --git a/init/Kconfig b/init/Kconfig

index 0c66640..ccd615a 100644

--- a/init/Kconfig
+++ b/init/Kconfig
@@ -1742,6 +1742,15 @@ config SLOB

endchoice

+config SLAB_FREELIST_RANDOM

+ default n
+ depends on SLAB
+ bool "SLAB freelist randomization"
+ help
+ Randomizes the freelist order used on creating new SLABs. This
+ security feature reduces the predictability of the kernel slab
+ allocator against heap overflows.

+

config SLUB_CPU_PARTIAL
default y
depends on SLUB && SMP
diff --git a/mm/slab.c b/mm/slab.c

index b82ee6b..e797388 100644

--- a/mm/slab.c
+++ b/mm/slab.c
@@ -1230,6 +1230,61 @@ static void __init set_up_node(struct kmem_cache *cachep, int index)
}
}

+#ifdef CONFIG_SLAB_FREELIST_RANDOM

+ /* If it fails, we will just use the global lists */
+ cachep->random_seq = kcalloc(count, sizeof(freelist_idx_t), GFP_KERNEL);
+ if (!cachep->random_seq)
+ return -ENOMEM;

+

+ /* Get best entropy at this stage */
+ get_random_bytes_arch(&seed, sizeof(seed));
+ prandom_seed_state(&state, seed);
+
+ freelist_randomize(&state, cachep->random_seq, count);

+ return 0;
+}
+

+/* Destroy the per-cache random freelist sequence */
+static void cache_random_seq_destroy(struct kmem_cache *cachep)
+{
+ kfree(cachep->random_seq);
+ cachep->random_seq = NULL;
+}
+#else

+static inline int cache_random_seq_create(struct kmem_cache *cachep)
+{

+ return 0;
+}
+static inline void cache_random_seq_destroy(struct kmem_cache *cachep) { }

+#endif /* CONFIG_SLAB_FREELIST_RANDOM */

+
+
/*
* Initialisation. Called after the page allocator have been initialised and
* before smp_init().
@@ -2337,6 +2392,8 @@ void __kmem_cache_release(struct kmem_cache *cachep)
int i;
struct kmem_cache_node *n;

+ cache_random_seq_destroy(cachep);
+
free_percpu(cachep->cpu_cache);

/* NUMA: free the node structures */
@@ -2443,15 +2500,115 @@ static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page)
#endif
}

+#ifdef CONFIG_SLAB_FREELIST_RANDOM

+/* Hold information during a freelist initialization */
+union freelist_init_state {
+ struct {
+ unsigned int pos;
+ freelist_idx_t *list;
+ unsigned int count;

+ unsigned int rand;
+ };

+ struct rnd_state rnd_state;
+};
+
+/*
+ * Initialize the state based on the randomization methode available.
+ * return true if the pre-computed list is available, false otherwize.
+ */
+static bool freelist_state_initialize(union freelist_init_state *state,
+ struct kmem_cache *cachep,
+ unsigned int count)
+{
+ bool ret;

+ unsigned int rand;
+

+ /* Use best entropy available to define a random shift */
+ get_random_bytes_arch(&rand, sizeof(rand));
+
+ /* Use a random state if the pre-computed list is not available */

+ if (!cachep->random_seq) {

+
+ if (count < 2)

+ return false;
+
+ precomputed = freelist_state_initialize(&state, cachep, count);
+
+ /* Take a random entry as the objfreelist */
+ if (OBJFREELIST_SLAB(cachep)) {
+ if (!precomputed)
+ objfreelist = count - 1;
+ else
+ objfreelist = next_random_slot(&state);
+ page->freelist = index_to_obj(cachep, page, objfreelist) +
+ obj_offset(cachep);
+ count--;
+ }
+
+ /*
+ * On early boot, generate the list dynamically.
+ * Later use a pre-computed list for speed.
+ */
+ if (!precomputed) {
+ freelist_randomize(&state.rnd_state, page->freelist, count);
+ } else {

+ for (i = 0; i < count; i++)

+ set_free_obj(page, i, next_random_slot(&state));
+ }
+
+ if (OBJFREELIST_SLAB(cachep))
+ set_free_obj(page, cachep->num - 1, objfreelist);
+
+ return true;
+}
+#else
+static inline bool shuffle_freelist(struct kmem_cache *cachep,
+ struct page *page)
+{
+ return false;
+}

+#endif /* CONFIG_SLAB_FREELIST_RANDOM */

[Andrew Morton]

On Wed, 27 Apr 2016 10:20:59 -0700 Thomas Garnier <thga...@google.com> wrote:

> Provides an optional config (CONFIG_SLAB_FREELIST_RANDOM) to randomize
> the SLAB freelist.

Forgot this bit?

From: Andrew Morton <ak...@linux-foundation.org>
Subject: mm-slab-freelist-randomization-v5-fix

propagate gfp_t into cache_random_seq_create()

Cc: Christoph Lameter <c...@linux.com>
Cc: David Rientjes <rien...@google.com>
Cc: Greg Thelen <gth...@google.com>
Cc: Joonsoo Kim <iamjoon...@lge.com>
Cc: Kees Cook <kees...@chromium.org>
Cc: Laura Abbott <lab...@fedoraproject.org>
Cc: Pekka Enberg <pen...@kernel.org>
Cc: Thomas Garnier <thga...@google.com>
Signed-off-by: Andrew Morton <ak...@linux-foundation.org>

--- a/mm/slab.c~mm-slab-freelist-randomization-v5-fix
+++ a/mm/slab.c
@@ -1262,7 +1262,7 @@ static void freelist_randomize(struct rn

}

/* Create a random sequence per cache */

-static int cache_random_seq_create(struct kmem_cache *cachep)
+static int cache_random_seq_create(struct kmem_cache *cachep, gfp_t gfp)
{

unsigned int seed, count = cachep->num;

struct rnd_state state;
@@ -1271,7 +1271,7 @@ static int cache_random_seq_create(struc
return 0;

/* If it fails, we will just use the global lists */

- cachep->random_seq = kcalloc(count, sizeof(freelist_idx_t), GFP_KERNEL);
+ cachep->random_seq = kcalloc(count, sizeof(freelist_idx_t), gfp);
if (!cachep->random_seq)
return -ENOMEM;

@@ -1290,7 +1290,7 @@ static void cache_random_seq_destroy(str
cachep->random_seq = NULL;
}
#else
-static inline int cache_random_seq_create(struct kmem_cache *cachep)
+static inline int cache_random_seq_create(struct kmem_cache *cachep, gfp_t gfp)
{
return 0;
}
@@ -3999,7 +3999,7 @@ static int enable_cpucache(struct kmem_c

int shared = 0;
int batchcount = 0;

- err = cache_random_seq_create(cachep);
+ err = cache_random_seq_create(cachep, gfp);
if (err)
goto end;

_

[Thomas Garnier]

On Wed, Apr 27, 2016 at 12:16 PM, Andrew Morton
<ak...@linux-foundation.org> wrote:
> On Wed, 27 Apr 2016 10:20:59 -0700 Thomas Garnier <thga...@google.com> wrote:
>
>> Provides an optional config (CONFIG_SLAB_FREELIST_RANDOM) to randomize
>> the SLAB freelist.
>
> Forgot this bit?
>

I thought I would change it when we support other kernel heaps.

[Joonsoo Kim]

As I mentioned in other thread, enable_cpucache() can be called
whenever kmem_cache is created. It should not be __init.

Thanks.