Message-ID: <20160420080821.GB7071@js1304-P5Q-DELUXE>
Date:	Wed, 20 Apr 2016 17:08:21 +0900
From:	Joonsoo Kim <iamjoonsoo.kim@....com>
To:	Thomas Garnier <thgarnie@...gle.com>
Cc:	Christoph Lameter <cl@...ux.com>,
	Pekka Enberg <penberg@...nel.org>,
	David Rientjes <rientjes@...gle.com>,
	Andrew Morton <akpm@...ux-foundation.org>,
	Kees Cook <keescook@...omium.org>,
	Greg Thelen <gthelen@...gle.com>,
	Laura Abbott <labbott@...oraproject.org>,
	kernel-hardening@...ts.openwall.com,
	LKML <linux-kernel@...r.kernel.org>,
	Linux-MM <linux-mm@...ck.org>
Subject: Re: [PATCH v2] mm: SLAB freelist randomization

On Tue, Apr 19, 2016 at 09:44:54AM -0700, Thomas Garnier wrote:
> On Tue, Apr 19, 2016 at 12:15 AM, Joonsoo Kim <iamjoonsoo.kim@....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).
> 
> >>
> >> 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
> >>
> >> Signed-off-by: Thomas Garnier <thgarnie@...gle.com>
> >> ---
> >> 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.
> >> + */
> >
> > 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.
> >
> 
> 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

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.

> 
> >> +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;
> >
> > 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.
> >
> 
> I think it is a good idea. I will add that for the next iteration.
> 
> >> +     /* 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);
> >
> > 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.
> >
> 
> The current implementation take the last chunk by default before the
> freelist is initialized. Do you want it to be randomized as well?

Yes.

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

Powered by blists - more mailing lists