[PATCH] mm/slub.c: re-randomize random_seq if necessary

[PATCH] mm/slub.c: re-randomize random_seq if necessary

[Peng Wang]

calculate_sizes() could be called in several places
like (red_zone/poison/order/store_user)_store() while
random_seq remains unchanged.

If random_seq is not NULL in calculate_sizes(), re-randomize it.

Signed-off-by: Peng Wang <rocking@whu.edu.cn>
---
 mm/slub.c | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/mm/slub.c b/mm/slub.c
index 1e3d0ec4e200..2a9d18019545 100644
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -3583,6 +3583,15 @@ static int calculate_sizes(struct kmem_cache *s, int forced_order)
 	if (oo_objects(s->oo) > oo_objects(s->max))
 		s->max = s->oo;
 
+#ifdef CONFIG_SLAB_FREELIST_RANDOM
+	if (unlikely(s->random_seq)) {
+		kfree(s->random_seq);
+		s->random_seq = NULL;
+		if (init_cache_random_seq(s))
+			return 0;
+	}
+#endif
+
 	return !!oo_objects(s->oo);
 }
 
-- 
2.19.1

Re: [PATCH] mm/slub.c: re-randomize random_seq if necessary

[Matthew Wilcox]

On Wed, Jan 09, 2019 at 05:06:27PM +0800, Peng Wang wrote:
> calculate_sizes() could be called in several places
> like (red_zone/poison/order/store_user)_store() while
> random_seq remains unchanged.
> 
> If random_seq is not NULL in calculate_sizes(), re-randomize it.

Why do we want to re-randomise the slab at these points?

RE: [PATCH] mm/slub.c: re-randomize random_seq if necessary

[Peng Wang]

On Wednesday, January 9, 2019 8:14 PM, Matthew Wilcox wrote:
> On Wed, Jan 09, 2019 at 05:06:27PM +0800, Peng Wang wrote:
> > calculate_sizes() could be called in several places
> > like (red_zone/poison/order/store_user)_store() while
> > random_seq remains unchanged.
> >
> > If random_seq is not NULL in calculate_sizes(), re-randomize it.
> 
> Why do we want to re-randomise the slab at these points?

At these points, s->size might change,
but random_seq still use the old size and not updated.

When doing shuffle_freelist() in allocat_slab(),
old next object offset would be used. 

    idx = s->random_seq[*pos];

One possible case:

s->size gets smaller, then number of objects in a slab gets bigger.
The size of s->random_seq array should be bigger but not updated.
In next_freelist_entry(), *pos might exceed the s->random_seq.

When we get zero value from s->random_seq[*pos] twice after exceeding,
BUG_ON(object == fp) would be triggered in set_freepointer().