Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]


Groups > linux.kernel > #1403198 > unrolled thread

[RFC v1 2/2] mm: SLUB Freelist randomization

Started byThomas Garnier <thgarnie@google.com>
First post2016-05-18 20:00 +0200
Last post2016-05-24 07:20 +0200
Articles 10 — 4 participants

Back to article view | Back to linux.kernel

This discussion starts older than the indexed window; earlier articles aren't shown. The article labeled Started by below is the oldest one visible, not the original post.


Contents

  [RFC v1 2/2] mm: SLUB Freelist randomization Thomas Garnier <thgarnie@google.com> - 2016-05-18 20:00 +0200
    Re: [RFC v1 2/2] mm: SLUB Freelist randomization Christoph Lameter <cl@linux.com> - 2016-05-18 20:30 +0200
      Re: [RFC v1 2/2] mm: SLUB Freelist randomization Thomas Garnier <thgarnie@google.com> - 2016-05-18 20:40 +0200
        Re: [RFC v1 2/2] mm: SLUB Freelist randomization Christoph Lameter <cl@linux.com> - 2016-05-18 21:10 +0200
          Re: [RFC v1 2/2] mm: SLUB Freelist randomization Thomas Garnier <thgarnie@google.com> - 2016-05-18 21:20 +0200
            Re: [RFC v1 2/2] mm: SLUB Freelist randomization Joonsoo Kim <iamjoonsoo.kim@lge.com> - 2016-05-19 04:10 +0200
              Re: [RFC v1 2/2] mm: SLUB Freelist randomization Thomas Garnier <thgarnie@google.com> - 2016-05-19 22:30 +0200
                Re: [RFC v1 2/2] mm: SLUB Freelist randomization Joonsoo Kim <js1304@gmail.com> - 2016-05-20 04:20 +0200
                  Re: [RFC v1 2/2] mm: SLUB Freelist randomization Thomas Garnier <thgarnie@google.com> - 2016-05-20 18:30 +0200
                    Re: [RFC v1 2/2] mm: SLUB Freelist randomization Joonsoo Kim <iamjoonsoo.kim@lge.com> - 2016-05-24 07:20 +0200

#1403198 — [RFC v1 2/2] mm: SLUB Freelist randomization

FromThomas Garnier <thgarnie@google.com>
Date2016-05-18 20:00 +0200
Subject[RFC v1 2/2] mm: SLUB Freelist randomization
Message-ID<rAdVf-5eB-5@gated-at.bofh.it>
Implements Freelist randomization for the SLUB allocator. It was
previous implemented for the SLAB allocator. Both use the same
configuration option (CONFIG_SLAB_FREELIST_RANDOM).

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. This
security feature reduces the predictability of the kernel SLUB allocator
against heap overflows rendering attacks much less stable.

For example these attacks exploit the predictability of the heap:
 - Linux Kernel CAN SLUB overflow (https://goo.gl/oMNWkU)
 - Exploiting Linux Kernel Heap corruptions (http://goo.gl/EXLn95)

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.

Performance results highlighted no major changes:

slab_test, before:

Single thread testing
=====================
1. Kmalloc: Repeatedly allocate then free test
10000 times kmalloc(8) -> 67 cycles kfree -> 101 cycles
10000 times kmalloc(16) -> 68 cycles kfree -> 109 cycles
10000 times kmalloc(32) -> 76 cycles kfree -> 119 cycles
10000 times kmalloc(64) -> 88 cycles kfree -> 114 cycles
10000 times kmalloc(128) -> 100 cycles kfree -> 122 cycles
10000 times kmalloc(256) -> 128 cycles kfree -> 149 cycles
10000 times kmalloc(512) -> 108 cycles kfree -> 152 cycles
10000 times kmalloc(1024) -> 112 cycles kfree -> 158 cycles
10000 times kmalloc(2048) -> 161 cycles kfree -> 208 cycles
10000 times kmalloc(4096) -> 231 cycles kfree -> 239 cycles
10000 times kmalloc(8192) -> 341 cycles kfree -> 270 cycles
10000 times kmalloc(16384) -> 481 cycles kfree -> 323 cycles
2. Kmalloc: alloc/free test
10000 times kmalloc(8)/kfree -> 90 cycles
10000 times kmalloc(16)/kfree -> 89 cycles
10000 times kmalloc(32)/kfree -> 88 cycles
10000 times kmalloc(64)/kfree -> 88 cycles
10000 times kmalloc(128)/kfree -> 94 cycles
10000 times kmalloc(256)/kfree -> 87 cycles
10000 times kmalloc(512)/kfree -> 91 cycles
10000 times kmalloc(1024)/kfree -> 90 cycles
10000 times kmalloc(2048)/kfree -> 90 cycles
10000 times kmalloc(4096)/kfree -> 90 cycles
10000 times kmalloc(8192)/kfree -> 90 cycles
10000 times kmalloc(16384)/kfree -> 642 cycles

After:

Single thread testing
=====================
1. Kmalloc: Repeatedly allocate then free test
10000 times kmalloc(8) -> 60 cycles kfree -> 74 cycles
10000 times kmalloc(16) -> 63 cycles kfree -> 78 cycles
10000 times kmalloc(32) -> 72 cycles kfree -> 85 cycles
10000 times kmalloc(64) -> 91 cycles kfree -> 99 cycles
10000 times kmalloc(128) -> 112 cycles kfree -> 109 cycles
10000 times kmalloc(256) -> 127 cycles kfree -> 120 cycles
10000 times kmalloc(512) -> 125 cycles kfree -> 121 cycles
10000 times kmalloc(1024) -> 128 cycles kfree -> 125 cycles
10000 times kmalloc(2048) -> 167 cycles kfree -> 141 cycles
10000 times kmalloc(4096) -> 249 cycles kfree -> 174 cycles
10000 times kmalloc(8192) -> 377 cycles kfree -> 225 cycles
10000 times kmalloc(16384) -> 459 cycles kfree -> 247 cycles
2. Kmalloc: alloc/free test
10000 times kmalloc(8)/kfree -> 72 cycles
10000 times kmalloc(16)/kfree -> 74 cycles
10000 times kmalloc(32)/kfree -> 71 cycles
10000 times kmalloc(64)/kfree -> 75 cycles
10000 times kmalloc(128)/kfree -> 71 cycles
10000 times kmalloc(256)/kfree -> 72 cycles
10000 times kmalloc(512)/kfree -> 72 cycles
10000 times kmalloc(1024)/kfree -> 73 cycles
10000 times kmalloc(2048)/kfree -> 73 cycles
10000 times kmalloc(4096)/kfree -> 72 cycles
10000 times kmalloc(8192)/kfree -> 72 cycles
10000 times kmalloc(16384)/kfree -> 546 cycles

Kernbench, before:

Average Optimal load -j 12 Run (std deviation):
Elapsed Time 101.873 (1.16069)
User Time 1045.22 (1.60447)
System Time 88.969 (0.559195)
Percent CPU 1112.9 (13.8279)
Context Switches 189140 (2282.15)
Sleeps 99008.6 (768.091)

After:

Average Optimal load -j 12 Run (std deviation):
Elapsed Time 102.47 (0.562732)
User Time 1045.3 (1.34263)
System Time 88.311 (0.342554)
Percent CPU 1105.8 (6.49444)
Context Switches 189081 (2355.78)
Sleeps 99231.5 (800.358)

Signed-off-by: Thomas Garnier <thgarnie@google.com>
---
Based on next-20160517
---
 include/linux/slub_def.h |   8 +++
 init/Kconfig             |   4 +-
 mm/slub.c                | 135 ++++++++++++++++++++++++++++++++++++++++++++---
 3 files changed, 138 insertions(+), 9 deletions(-)

diff --git a/include/linux/slub_def.h b/include/linux/slub_def.h
index 665cd0c..22d487e 100644
--- a/include/linux/slub_def.h
+++ b/include/linux/slub_def.h
@@ -56,6 +56,9 @@ struct kmem_cache_order_objects {
 	unsigned long x;
 };
 
+/* Index used for freelist randomization */
+typedef unsigned int freelist_idx_t;
+
 /*
  * Slab cache management.
  */
@@ -99,6 +102,11 @@ struct kmem_cache {
 	 */
 	int remote_node_defrag_ratio;
 #endif
+
+#ifdef CONFIG_SLAB_FREELIST_RANDOM
+	freelist_idx_t *random_seq;
+#endif
+
 	struct kmem_cache_node *node[MAX_NUMNODES];
 };
 
diff --git a/init/Kconfig b/init/Kconfig
index 7b82f3f..4590629 100644
--- a/init/Kconfig
+++ b/init/Kconfig
@@ -1784,10 +1784,10 @@ endchoice
 
 config SLAB_FREELIST_RANDOM
 	default n
-	depends on SLAB
+	depends on SLAB || SLUB
 	bool "SLAB freelist randomization"
 	help
-	  Randomizes the freelist order used on creating new SLABs. This
+	  Randomizes the freelist order used on creating new pages. This
 	  security feature reduces the predictability of the kernel slab
 	  allocator against heap overflows.
 
diff --git a/mm/slub.c b/mm/slub.c
index 538c858..d89780a 100644
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -1405,6 +1405,111 @@ static inline struct page *alloc_slab_page(struct kmem_cache *s,
 	return page;
 }
 
+#ifdef CONFIG_SLAB_FREELIST_RANDOM
+/* Pre-initialize the random sequence cache */
+static int init_cache_random_seq(struct kmem_cache *s)
+{
+	int err;
+	unsigned long i, count = oo_objects(s->oo);
+
+	err = cache_random_seq_create(s, count, GFP_KERNEL);
+	if (err) {
+		pr_err("SLUB: Unable to initialize free list for %s\n",
+			s->name);
+		return err;
+	}
+
+	/* Transform to an offset on the set of pages */
+	if (s->random_seq) {
+		for (i = 0; i < count; i++)
+			s->random_seq[i] *= s->size;
+	}
+	return 0;
+}
+
+/* Initialize each random sequence freelist per cache */
+static void __init init_freelist_randomization(void)
+{
+	struct kmem_cache *s;
+
+	mutex_lock(&slab_mutex);
+
+	list_for_each_entry(s, &slab_caches, list)
+		init_cache_random_seq(s);
+
+	mutex_unlock(&slab_mutex);
+}
+
+/* Get the next entry on the pre-computed freelist randomized */
+static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
+				unsigned long *pos, void *start,
+				unsigned long page_limit,
+				unsigned long freelist_count)
+{
+	freelist_idx_t idx;
+
+	/*
+	 * If the target page allocation failed, the number of objects on the
+	 * page might be smaller than the usual size defined by the cache.
+	 */
+	do {
+		idx = s->random_seq[*pos];
+		*pos += 1;
+		if (*pos >= freelist_count)
+			*pos = 0;
+	} while (unlikely(idx >= page_limit));
+
+	return (char *)start + idx;
+}
+
+/* Shuffle the single linked freelist based on a random pre-computed sequence */
+static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
+{
+	void *start;
+	void *cur;
+	void *next;
+	unsigned long idx, pos, page_limit, freelist_count;
+
+	if (page->objects < 2 || !s->random_seq)
+		return false;
+
+	/* Use best entropy available to define a random start */
+	freelist_count = oo_objects(s->oo);
+	get_random_bytes_arch(&pos, sizeof(pos));
+	pos %= freelist_count;
+
+	page_limit = page->objects * s->size;
+	start = fixup_red_left(s, page_address(page));
+
+	/* First entry is used as the base of the freelist */
+	cur = next_freelist_entry(s, page, &pos, start, page_limit,
+				freelist_count);
+	page->freelist = cur;
+
+	for (idx = 1; idx < page->objects; idx++) {
+		setup_object(s, page, cur);
+		next = next_freelist_entry(s, page, &pos, start, page_limit,
+			freelist_count);
+		set_freepointer(s, cur, next);
+		cur = next;
+	}
+	setup_object(s, page, cur);
+	set_freepointer(s, cur, NULL);
+
+	return true;
+}
+#else
+static inline int init_cache_random_seq(struct kmem_cache *s)
+{
+	return 0;
+}
+static inline void init_freelist_randomization(void) { }
+static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
+{
+	return false;
+}
+#endif /* CONFIG_SLAB_FREELIST_RANDOM */
+
 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
 {
 	struct page *page;
@@ -1412,6 +1517,7 @@ static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
 	gfp_t alloc_gfp;
 	void *start, *p;
 	int idx, order;
+	bool shuffle;
 
 	flags &= gfp_allowed_mask;
 
@@ -1473,15 +1579,19 @@ static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
 
 	kasan_poison_slab(page);
 
-	for_each_object_idx(p, idx, s, start, page->objects) {
-		setup_object(s, page, p);
-		if (likely(idx < page->objects))
-			set_freepointer(s, p, p + s->size);
-		else
-			set_freepointer(s, p, NULL);
+	shuffle = shuffle_freelist(s, page);
+
+	if (!shuffle) {
+		for_each_object_idx(p, idx, s, start, page->objects) {
+			setup_object(s, page, p);
+			if (likely(idx < page->objects))
+				set_freepointer(s, p, p + s->size);
+			else
+				set_freepointer(s, p, NULL);
+		}
+		page->freelist = fixup_red_left(s, start);
 	}
 
-	page->freelist = fixup_red_left(s, start);
 	page->inuse = page->objects;
 	page->frozen = 1;
 
@@ -3207,6 +3317,7 @@ static void free_kmem_cache_nodes(struct kmem_cache *s)
 
 void __kmem_cache_release(struct kmem_cache *s)
 {
+	cache_random_seq_destroy(s);
 	free_percpu(s->cpu_slab);
 	free_kmem_cache_nodes(s);
 }
@@ -3431,6 +3542,13 @@ static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
 #ifdef CONFIG_NUMA
 	s->remote_node_defrag_ratio = 1000;
 #endif
+
+	/* Initialize the pre-computed randomized freelist if slab is up */
+	if (slab_state >= UP) {
+		if (init_cache_random_seq(s))
+			goto error;
+	}
+
 	if (!init_kmem_cache_nodes(s))
 		goto error;
 
@@ -3947,6 +4065,9 @@ void __init kmem_cache_init(void)
 	setup_kmalloc_cache_index_table();
 	create_kmalloc_caches(0);
 
+	/* Setup random freelists for each cache */
+	init_freelist_randomization();
+
 #ifdef CONFIG_SMP
 	register_cpu_notifier(&slab_notifier);
 #endif
-- 
2.8.0.rc3.226.g39d4020

[toc] | [next] | [standalone]


#1403212

FromChristoph Lameter <cl@linux.com>
Date2016-05-18 20:30 +0200
Message-ID<rAeoi-5Fp-9@gated-at.bofh.it>
In reply to#1403198
0.On Wed, 18 May 2016, Thomas Garnier wrote:

> slab_test, before:
> 10000 times kmalloc(8) -> 67 cycles kfree -> 101 cycles
> 10000 times kmalloc(16) -> 68 cycles kfree -> 109 cycles
> 10000 times kmalloc(32) -> 76 cycles kfree -> 119 cycles
> 10000 times kmalloc(64) -> 88 cycles kfree -> 114 cycles

> After:
> 10000 times kmalloc(8) -> 60 cycles kfree -> 74 cycles
> 10000 times kmalloc(16) -> 63 cycles kfree -> 78 cycles
> 10000 times kmalloc(32) -> 72 cycles kfree -> 85 cycles
> 10000 times kmalloc(64) -> 91 cycles kfree -> 99 cycles

Erm... The fastpath was not touched and the tests primarily exercise the
fastpath. This is likely some artifact of code placement by the compiler?

[toc] | [prev] | [next] | [standalone]


#1403216

FromThomas Garnier <thgarnie@google.com>
Date2016-05-18 20:40 +0200
Message-ID<rAexY-5ID-23@gated-at.bofh.it>
In reply to#1403212
Yes, I agree that it is not related to the changes.

On Wed, May 18, 2016 at 11:24 AM, Christoph Lameter <cl@linux.com> wrote:
> 0.On Wed, 18 May 2016, Thomas Garnier wrote:
>
>> slab_test, before:
>> 10000 times kmalloc(8) -> 67 cycles kfree -> 101 cycles
>> 10000 times kmalloc(16) -> 68 cycles kfree -> 109 cycles
>> 10000 times kmalloc(32) -> 76 cycles kfree -> 119 cycles
>> 10000 times kmalloc(64) -> 88 cycles kfree -> 114 cycles
>
>> After:
>> 10000 times kmalloc(8) -> 60 cycles kfree -> 74 cycles
>> 10000 times kmalloc(16) -> 63 cycles kfree -> 78 cycles
>> 10000 times kmalloc(32) -> 72 cycles kfree -> 85 cycles
>> 10000 times kmalloc(64) -> 91 cycles kfree -> 99 cycles
>
> Erm... The fastpath was not touched and the tests primarily exercise the
> fastpath. This is likely some artifact of code placement by the compiler?
>

[toc] | [prev] | [next] | [standalone]


#1403231

FromChristoph Lameter <cl@linux.com>
Date2016-05-18 21:10 +0200
Message-ID<rAf0Z-67j-5@gated-at.bofh.it>
In reply to#1403216
On Wed, 18 May 2016, Thomas Garnier wrote:

> Yes, I agree that it is not related to the changes.

Could you please provide meaningful test data?

[toc] | [prev] | [next] | [standalone]


#1403241

FromThomas Garnier <thgarnie@google.com>
Date2016-05-18 21:20 +0200
Message-ID<rAfaG-6av-17@gated-at.bofh.it>
In reply to#1403231
I thought the mix of slab_test & kernbench would show a diverse
picture on perf data. Is there another test that you think would be
useful?

Thanks,
Thomas

On Wed, May 18, 2016 at 12:02 PM, Christoph Lameter <cl@linux.com> wrote:
> On Wed, 18 May 2016, Thomas Garnier wrote:
>
>> Yes, I agree that it is not related to the changes.
>
> Could you please provide meaningful test data?

[toc] | [prev] | [next] | [standalone]


#1403393

FromJoonsoo Kim <iamjoonsoo.kim@lge.com>
Date2016-05-19 04:10 +0200
Message-ID<rAlzr-1So-5@gated-at.bofh.it>
In reply to#1403241
On Wed, May 18, 2016 at 12:12:13PM -0700, Thomas Garnier wrote:
> I thought the mix of slab_test & kernbench would show a diverse
> picture on perf data. Is there another test that you think would be
> useful?

Single thread testing on slab_test would be meaningful because it also
touch the slowpath. Problem is just unstable result of slab_test.

You can get more stable result of slab_test if you repeat same test
sometimes and get average result.

Please use following slab_test. It will do each operations 100000
times and repeat it 50 times.

https://github.com/JoonsooKim/linux/blob/slab_test_robust-next-20160509/mm/slab_test.c

I did a quick test for this patchset and get following result.

- Before (With patch and randomization is disabled by config)

Single thread testing
=====================
1. Kmalloc: Repeatedly allocate then free test
100000 times kmalloc(8) -> 42 cycles kfree -> 67 cycles
100000 times kmalloc(16) -> 43 cycles kfree -> 68 cycles
100000 times kmalloc(32) -> 47 cycles kfree -> 72 cycles
100000 times kmalloc(64) -> 54 cycles kfree -> 78 cycles
100000 times kmalloc(128) -> 75 cycles kfree -> 87 cycles
100000 times kmalloc(256) -> 84 cycles kfree -> 111 cycles
100000 times kmalloc(512) -> 82 cycles kfree -> 112 cycles
100000 times kmalloc(1024) -> 86 cycles kfree -> 113 cycles
100000 times kmalloc(2048) -> 113 cycles kfree -> 127 cycles
100000 times kmalloc(4096) -> 151 cycles kfree -> 154 cycles

- After (With patch and randomization is enabled by config)

Single thread testing
=====================
1. Kmalloc: Repeatedly allocate then free test
100000 times kmalloc(8) -> 51 cycles kfree -> 68 cycles
100000 times kmalloc(16) -> 57 cycles kfree -> 70 cycles
100000 times kmalloc(32) -> 70 cycles kfree -> 75 cycles
100000 times kmalloc(64) -> 95 cycles kfree -> 84 cycles
100000 times kmalloc(128) -> 142 cycles kfree -> 97 cycles
100000 times kmalloc(256) -> 150 cycles kfree -> 107 cycles
100000 times kmalloc(512) -> 151 cycles kfree -> 107 cycles
100000 times kmalloc(1024) -> 154 cycles kfree -> 110 cycles
100000 times kmalloc(2048) -> 230 cycles kfree -> 124 cycles
100000 times kmalloc(4096) -> 423 cycles kfree -> 165 cycles

It seems that performance decreases a lot but I don't care about it
because it is a security feature and I don't have a better idea.

Thanks.

[toc] | [prev] | [next] | [standalone]


#1403955

FromThomas Garnier <thgarnie@google.com>
Date2016-05-19 22:30 +0200
Message-ID<rACJX-4xR-7@gated-at.bofh.it>
In reply to#1403393
I ran the test given by Joonsoo and it gave me these minimum cycles
per size across 20 usage:

size,before,after
8,63.00,64.50 (102.38%)
16,64.50,65.00 (100.78%)
32,65.00,65.00 (100.00%)
64,66.00,65.00 (98.48%)
128,66.00,65.00 (98.48%)
256,64.00,64.00 (100.00%)
512,65.00,66.00 (101.54%)
1024,68.00,64.00 (94.12%)
2048,66.00,65.00 (98.48%)
4096,66.00,66.00 (100.00%)

I assume the difference is bigger if you don't have RDRAND support.

Christoph, Joonsoo: Do you think it would be valuable to add a CONFIG
to disable additional randomization per new page? It will remove
additional entropy but increase performance for machines without arch
specific randomization instructions.

Thanks,
Thomas


On Wed, May 18, 2016 at 7:07 PM, Joonsoo Kim <iamjoonsoo.kim@lge.com> wrote:
> On Wed, May 18, 2016 at 12:12:13PM -0700, Thomas Garnier wrote:
>> I thought the mix of slab_test & kernbench would show a diverse
>> picture on perf data. Is there another test that you think would be
>> useful?
>
> Single thread testing on slab_test would be meaningful because it also
> touch the slowpath. Problem is just unstable result of slab_test.
>
> You can get more stable result of slab_test if you repeat same test
> sometimes and get average result.
>
> Please use following slab_test. It will do each operations 100000
> times and repeat it 50 times.
>
> https://github.com/JoonsooKim/linux/blob/slab_test_robust-next-20160509/mm/slab_test.c
>
> I did a quick test for this patchset and get following result.
>
> - Before (With patch and randomization is disabled by config)
>
> Single thread testing
> =====================
> 1. Kmalloc: Repeatedly allocate then free test
> 100000 times kmalloc(8) -> 42 cycles kfree -> 67 cycles
> 100000 times kmalloc(16) -> 43 cycles kfree -> 68 cycles
> 100000 times kmalloc(32) -> 47 cycles kfree -> 72 cycles
> 100000 times kmalloc(64) -> 54 cycles kfree -> 78 cycles
> 100000 times kmalloc(128) -> 75 cycles kfree -> 87 cycles
> 100000 times kmalloc(256) -> 84 cycles kfree -> 111 cycles
> 100000 times kmalloc(512) -> 82 cycles kfree -> 112 cycles
> 100000 times kmalloc(1024) -> 86 cycles kfree -> 113 cycles
> 100000 times kmalloc(2048) -> 113 cycles kfree -> 127 cycles
> 100000 times kmalloc(4096) -> 151 cycles kfree -> 154 cycles
>
> - After (With patch and randomization is enabled by config)
>
> Single thread testing
> =====================
> 1. Kmalloc: Repeatedly allocate then free test
> 100000 times kmalloc(8) -> 51 cycles kfree -> 68 cycles
> 100000 times kmalloc(16) -> 57 cycles kfree -> 70 cycles
> 100000 times kmalloc(32) -> 70 cycles kfree -> 75 cycles
> 100000 times kmalloc(64) -> 95 cycles kfree -> 84 cycles
> 100000 times kmalloc(128) -> 142 cycles kfree -> 97 cycles
> 100000 times kmalloc(256) -> 150 cycles kfree -> 107 cycles
> 100000 times kmalloc(512) -> 151 cycles kfree -> 107 cycles
> 100000 times kmalloc(1024) -> 154 cycles kfree -> 110 cycles
> 100000 times kmalloc(2048) -> 230 cycles kfree -> 124 cycles
> 100000 times kmalloc(4096) -> 423 cycles kfree -> 165 cycles
>
> It seems that performance decreases a lot but I don't care about it
> because it is a security feature and I don't have a better idea.
>
> Thanks.

[toc] | [prev] | [next] | [standalone]


#1404070

FromJoonsoo Kim <js1304@gmail.com>
Date2016-05-20 04:20 +0200
Message-ID<rAIcF-83P-1@gated-at.bofh.it>
In reply to#1403955
2016-05-20 5:20 GMT+09:00 Thomas Garnier <thgarnie@google.com>:
> I ran the test given by Joonsoo and it gave me these minimum cycles
> per size across 20 usage:

I can't understand what you did here. Maybe, it's due to my poor Engling.
Please explain more. You did single thread test? Why minimum cycles
rather than average?

> size,before,after
> 8,63.00,64.50 (102.38%)
> 16,64.50,65.00 (100.78%)
> 32,65.00,65.00 (100.00%)
> 64,66.00,65.00 (98.48%)
> 128,66.00,65.00 (98.48%)
> 256,64.00,64.00 (100.00%)
> 512,65.00,66.00 (101.54%)
> 1024,68.00,64.00 (94.12%)
> 2048,66.00,65.00 (98.48%)
> 4096,66.00,66.00 (100.00%)

It looks like performance of all size classes are the same?

> I assume the difference is bigger if you don't have RDRAND support.

What does RDRAND means? Kconfig? How can I check if I have RDRAND?

> Christoph, Joonsoo: Do you think it would be valuable to add a CONFIG
> to disable additional randomization per new page? It will remove
> additional entropy but increase performance for machines without arch
> specific randomization instructions.

I don't think that it deserve another CONFIG. If performance is a matter,
I think that removing additional entropy is better until it is proved that
entropy is a problem.

Thanks.

[toc] | [prev] | [next] | [standalone]


#1404577

FromThomas Garnier <thgarnie@google.com>
Date2016-05-20 18:30 +0200
Message-ID<rAVtf-7Ld-3@gated-at.bofh.it>
In reply to#1404070
On Thu, May 19, 2016 at 7:15 PM, Joonsoo Kim <js1304@gmail.com> wrote:
> 2016-05-20 5:20 GMT+09:00 Thomas Garnier <thgarnie@google.com>:
>> I ran the test given by Joonsoo and it gave me these minimum cycles
>> per size across 20 usage:
>
> I can't understand what you did here. Maybe, it's due to my poor Engling.
> Please explain more. You did single thread test? Why minimum cycles
> rather than average?
>

I used your version of slab_test and ran it 20 times for each
versions. I compared
the minimum number of cycles as an optimal case for comparison. As you said
slab_test results can be unreliable. Comparing the average across multiple runs
always gave odd results.

>> size,before,after
>> 8,63.00,64.50 (102.38%)
>> 16,64.50,65.00 (100.78%)
>> 32,65.00,65.00 (100.00%)
>> 64,66.00,65.00 (98.48%)
>> 128,66.00,65.00 (98.48%)
>> 256,64.00,64.00 (100.00%)
>> 512,65.00,66.00 (101.54%)
>> 1024,68.00,64.00 (94.12%)
>> 2048,66.00,65.00 (98.48%)
>> 4096,66.00,66.00 (100.00%)
>
> It looks like performance of all size classes are the same?
>
>> I assume the difference is bigger if you don't have RDRAND support.
>
> What does RDRAND means? Kconfig? How can I check if I have RDRAND?
>

Sorry, I was referring to the usage of get_random_bytes_arch which
will be faster
if the test machine support specific instructions (like RDRAND).

>> Christoph, Joonsoo: Do you think it would be valuable to add a CONFIG
>> to disable additional randomization per new page? It will remove
>> additional entropy but increase performance for machines without arch
>> specific randomization instructions.
>
> I don't think that it deserve another CONFIG. If performance is a matter,
> I think that removing additional entropy is better until it is proved that
> entropy is a problem.
>

I will do more testing before the next RFC to decide the best approach.

> Thanks.

Thanks for the comments,
Thomas

[toc] | [prev] | [next] | [standalone]


#1405846

FromJoonsoo Kim <iamjoonsoo.kim@lge.com>
Date2016-05-24 07:20 +0200
Message-ID<rCcV3-7vi-1@gated-at.bofh.it>
In reply to#1404577
On Fri, May 20, 2016 at 09:24:35AM -0700, Thomas Garnier wrote:
> On Thu, May 19, 2016 at 7:15 PM, Joonsoo Kim <js1304@gmail.com> wrote:
> > 2016-05-20 5:20 GMT+09:00 Thomas Garnier <thgarnie@google.com>:
> >> I ran the test given by Joonsoo and it gave me these minimum cycles
> >> per size across 20 usage:
> >
> > I can't understand what you did here. Maybe, it's due to my poor Engling.
> > Please explain more. You did single thread test? Why minimum cycles
> > rather than average?
> >
> 
> I used your version of slab_test and ran it 20 times for each
> versions. I compared
> the minimum number of cycles as an optimal case for comparison. As you said
> slab_test results can be unreliable. Comparing the average across multiple runs
> always gave odd results.

Hmm... With my version, slab_test results seems to be reliable for me. You
can use average result in this case. Anyway, your minimum result looks
odd even if my version is used. Large sized test would go slowpath
more frequently so it should be worse.

> 
> >> size,before,after
> >> 8,63.00,64.50 (102.38%)
> >> 16,64.50,65.00 (100.78%)
> >> 32,65.00,65.00 (100.00%)
> >> 64,66.00,65.00 (98.48%)
> >> 128,66.00,65.00 (98.48%)
> >> 256,64.00,64.00 (100.00%)
> >> 512,65.00,66.00 (101.54%)
> >> 1024,68.00,64.00 (94.12%)
> >> 2048,66.00,65.00 (98.48%)
> >> 4096,66.00,66.00 (100.00%)
> >
> > It looks like performance of all size classes are the same?
> >
> >> I assume the difference is bigger if you don't have RDRAND support.
> >
> > What does RDRAND means? Kconfig? How can I check if I have RDRAND?
> >
> 
> Sorry, I was referring to the usage of get_random_bytes_arch which
> will be faster
> if the test machine support specific instructions (like RDRAND).

Thanks! I checked that my test bed (QEMU) doesn't support rdrand.
(/proc/cpuinfo)

> >> Christoph, Joonsoo: Do you think it would be valuable to add a CONFIG
> >> to disable additional randomization per new page? It will remove
> >> additional entropy but increase performance for machines without arch
> >> specific randomization instructions.
> >
> > I don't think that it deserve another CONFIG. If performance is a matter,
> > I think that removing additional entropy is better until it is proved that
> > entropy is a problem.
> >
> 
> I will do more testing before the next RFC to decide the best approach.

Okay.

Thanks.

[toc] | [prev] | [standalone]


Back to top | Article view | linux.kernel


csiph-web