blob: 5d2f24fbafc5535997f282e6751476afdf28f83e [file] [log] [blame]
Christoph Lameter039363f2012-07-06 15:25:10 -05001/*
2 * Slab allocator functions that are independent of the allocator strategy
3 *
4 * (C) 2012 Christoph Lameter <cl@linux.com>
5 */
6#include <linux/slab.h>
7
8#include <linux/mm.h>
9#include <linux/poison.h>
10#include <linux/interrupt.h>
11#include <linux/memory.h>
12#include <linux/compiler.h>
13#include <linux/module.h>
Christoph Lameter20cea962012-07-06 15:25:13 -050014#include <linux/cpu.h>
15#include <linux/uaccess.h>
Glauber Costab7454ad2012-10-19 18:20:25 +040016#include <linux/seq_file.h>
17#include <linux/proc_fs.h>
Christoph Lameter039363f2012-07-06 15:25:10 -050018#include <asm/cacheflush.h>
19#include <asm/tlbflush.h>
20#include <asm/page.h>
Glauber Costa2633d7a2012-12-18 14:22:34 -080021#include <linux/memcontrol.h>
Andrey Ryabinin928cec92014-08-06 16:04:44 -070022
23#define CREATE_TRACE_POINTS
Christoph Lameterf1b6eb62013-09-04 16:35:34 +000024#include <trace/events/kmem.h>
Christoph Lameter039363f2012-07-06 15:25:10 -050025
Christoph Lameter97d06602012-07-06 15:25:11 -050026#include "slab.h"
27
28enum slab_state slab_state;
Christoph Lameter18004c52012-07-06 15:25:12 -050029LIST_HEAD(slab_caches);
30DEFINE_MUTEX(slab_mutex);
Christoph Lameter9b030cb2012-09-05 00:20:33 +000031struct kmem_cache *kmem_cache;
Christoph Lameter97d06602012-07-06 15:25:11 -050032
Joonsoo Kim07f361b2014-10-09 15:26:00 -070033/*
Joonsoo Kim423c9292014-10-09 15:26:22 -070034 * Set of flags that will prevent slab merging
35 */
36#define SLAB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
37 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -070038 SLAB_FAILSLAB | SLAB_KASAN)
Joonsoo Kim423c9292014-10-09 15:26:22 -070039
Vladimir Davydov230e9fc2016-01-14 15:18:15 -080040#define SLAB_MERGE_SAME (SLAB_RECLAIM_ACCOUNT | SLAB_CACHE_DMA | \
41 SLAB_NOTRACK | SLAB_ACCOUNT)
Joonsoo Kim423c9292014-10-09 15:26:22 -070042
43/*
44 * Merge control. If this is set then no merging of slab caches will occur.
45 * (Could be removed. This was introduced to pacify the merge skeptics.)
46 */
47static int slab_nomerge;
48
49static int __init setup_slab_nomerge(char *str)
50{
51 slab_nomerge = 1;
52 return 1;
53}
54
55#ifdef CONFIG_SLUB
56__setup_param("slub_nomerge", slub_nomerge, setup_slab_nomerge, 0);
57#endif
58
59__setup("slab_nomerge", setup_slab_nomerge);
60
61/*
Joonsoo Kim07f361b2014-10-09 15:26:00 -070062 * Determine the size of a slab object
63 */
64unsigned int kmem_cache_size(struct kmem_cache *s)
65{
66 return s->object_size;
67}
68EXPORT_SYMBOL(kmem_cache_size);
69
Shuah Khan77be4b12012-08-16 00:09:46 -070070#ifdef CONFIG_DEBUG_VM
Vladimir Davydov794b1242014-04-07 15:39:26 -070071static int kmem_cache_sanity_check(const char *name, size_t size)
Shuah Khan77be4b12012-08-16 00:09:46 -070072{
73 struct kmem_cache *s = NULL;
74
75 if (!name || in_interrupt() || size < sizeof(void *) ||
76 size > KMALLOC_MAX_SIZE) {
77 pr_err("kmem_cache_create(%s) integrity check failed\n", name);
78 return -EINVAL;
79 }
80
81 list_for_each_entry(s, &slab_caches, list) {
82 char tmp;
83 int res;
84
85 /*
86 * This happens when the module gets unloaded and doesn't
87 * destroy its slab cache and no-one else reuses the vmalloc
88 * area of the module. Print a warning.
89 */
90 res = probe_kernel_address(s->name, tmp);
91 if (res) {
92 pr_err("Slab cache with size %d has lost its name\n",
93 s->object_size);
94 continue;
95 }
Shuah Khan77be4b12012-08-16 00:09:46 -070096 }
97
98 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
99 return 0;
100}
101#else
Vladimir Davydov794b1242014-04-07 15:39:26 -0700102static inline int kmem_cache_sanity_check(const char *name, size_t size)
Shuah Khan77be4b12012-08-16 00:09:46 -0700103{
104 return 0;
105}
106#endif
107
Christoph Lameter484748f2015-09-04 15:45:34 -0700108void __kmem_cache_free_bulk(struct kmem_cache *s, size_t nr, void **p)
109{
110 size_t i;
111
Jesper Dangaard Brouerca257192016-03-15 14:54:00 -0700112 for (i = 0; i < nr; i++) {
113 if (s)
114 kmem_cache_free(s, p[i]);
115 else
116 kfree(p[i]);
117 }
Christoph Lameter484748f2015-09-04 15:45:34 -0700118}
119
Jesper Dangaard Brouer865762a2015-11-20 15:57:58 -0800120int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t nr,
Christoph Lameter484748f2015-09-04 15:45:34 -0700121 void **p)
122{
123 size_t i;
124
125 for (i = 0; i < nr; i++) {
126 void *x = p[i] = kmem_cache_alloc(s, flags);
127 if (!x) {
128 __kmem_cache_free_bulk(s, i, p);
Jesper Dangaard Brouer865762a2015-11-20 15:57:58 -0800129 return 0;
Christoph Lameter484748f2015-09-04 15:45:34 -0700130 }
131 }
Jesper Dangaard Brouer865762a2015-11-20 15:57:58 -0800132 return i;
Christoph Lameter484748f2015-09-04 15:45:34 -0700133}
134
Johannes Weiner127424c2016-01-20 15:02:32 -0800135#if defined(CONFIG_MEMCG) && !defined(CONFIG_SLOB)
Vladimir Davydovf7ce3192015-02-12 14:59:20 -0800136void slab_init_memcg_params(struct kmem_cache *s)
Vladimir Davydov33a690c2014-10-09 15:28:43 -0700137{
Vladimir Davydovf7ce3192015-02-12 14:59:20 -0800138 s->memcg_params.is_root_cache = true;
Vladimir Davydov426589f2015-02-12 14:59:23 -0800139 INIT_LIST_HEAD(&s->memcg_params.list);
Vladimir Davydovf7ce3192015-02-12 14:59:20 -0800140 RCU_INIT_POINTER(s->memcg_params.memcg_caches, NULL);
141}
Vladimir Davydov33a690c2014-10-09 15:28:43 -0700142
Vladimir Davydovf7ce3192015-02-12 14:59:20 -0800143static int init_memcg_params(struct kmem_cache *s,
144 struct mem_cgroup *memcg, struct kmem_cache *root_cache)
145{
146 struct memcg_cache_array *arr;
Vladimir Davydov33a690c2014-10-09 15:28:43 -0700147
148 if (memcg) {
Vladimir Davydovf7ce3192015-02-12 14:59:20 -0800149 s->memcg_params.is_root_cache = false;
150 s->memcg_params.memcg = memcg;
151 s->memcg_params.root_cache = root_cache;
152 return 0;
153 }
Vladimir Davydov33a690c2014-10-09 15:28:43 -0700154
Vladimir Davydovf7ce3192015-02-12 14:59:20 -0800155 slab_init_memcg_params(s);
156
157 if (!memcg_nr_cache_ids)
158 return 0;
159
160 arr = kzalloc(sizeof(struct memcg_cache_array) +
161 memcg_nr_cache_ids * sizeof(void *),
162 GFP_KERNEL);
163 if (!arr)
164 return -ENOMEM;
165
166 RCU_INIT_POINTER(s->memcg_params.memcg_caches, arr);
Vladimir Davydov33a690c2014-10-09 15:28:43 -0700167 return 0;
168}
169
Vladimir Davydovf7ce3192015-02-12 14:59:20 -0800170static void destroy_memcg_params(struct kmem_cache *s)
Vladimir Davydov33a690c2014-10-09 15:28:43 -0700171{
Vladimir Davydovf7ce3192015-02-12 14:59:20 -0800172 if (is_root_cache(s))
173 kfree(rcu_access_pointer(s->memcg_params.memcg_caches));
Vladimir Davydov33a690c2014-10-09 15:28:43 -0700174}
175
Vladimir Davydovf7ce3192015-02-12 14:59:20 -0800176static int update_memcg_params(struct kmem_cache *s, int new_array_size)
Vladimir Davydov6f817f42014-10-09 15:28:47 -0700177{
Vladimir Davydovf7ce3192015-02-12 14:59:20 -0800178 struct memcg_cache_array *old, *new;
Vladimir Davydov6f817f42014-10-09 15:28:47 -0700179
Vladimir Davydovf7ce3192015-02-12 14:59:20 -0800180 if (!is_root_cache(s))
181 return 0;
Vladimir Davydov6f817f42014-10-09 15:28:47 -0700182
Vladimir Davydovf7ce3192015-02-12 14:59:20 -0800183 new = kzalloc(sizeof(struct memcg_cache_array) +
184 new_array_size * sizeof(void *), GFP_KERNEL);
185 if (!new)
Vladimir Davydov6f817f42014-10-09 15:28:47 -0700186 return -ENOMEM;
187
Vladimir Davydovf7ce3192015-02-12 14:59:20 -0800188 old = rcu_dereference_protected(s->memcg_params.memcg_caches,
189 lockdep_is_held(&slab_mutex));
190 if (old)
191 memcpy(new->entries, old->entries,
192 memcg_nr_cache_ids * sizeof(void *));
Vladimir Davydov6f817f42014-10-09 15:28:47 -0700193
Vladimir Davydovf7ce3192015-02-12 14:59:20 -0800194 rcu_assign_pointer(s->memcg_params.memcg_caches, new);
195 if (old)
196 kfree_rcu(old, rcu);
Vladimir Davydov6f817f42014-10-09 15:28:47 -0700197 return 0;
198}
199
Glauber Costa55007d82012-12-18 14:22:38 -0800200int memcg_update_all_caches(int num_memcgs)
201{
202 struct kmem_cache *s;
203 int ret = 0;
Glauber Costa55007d82012-12-18 14:22:38 -0800204
Vladimir Davydov05257a12015-02-12 14:59:01 -0800205 mutex_lock(&slab_mutex);
Glauber Costa55007d82012-12-18 14:22:38 -0800206 list_for_each_entry(s, &slab_caches, list) {
Vladimir Davydovf7ce3192015-02-12 14:59:20 -0800207 ret = update_memcg_params(s, num_memcgs);
Glauber Costa55007d82012-12-18 14:22:38 -0800208 /*
Glauber Costa55007d82012-12-18 14:22:38 -0800209 * Instead of freeing the memory, we'll just leave the caches
210 * up to this point in an updated state.
211 */
212 if (ret)
Vladimir Davydov05257a12015-02-12 14:59:01 -0800213 break;
Glauber Costa55007d82012-12-18 14:22:38 -0800214 }
Glauber Costa55007d82012-12-18 14:22:38 -0800215 mutex_unlock(&slab_mutex);
216 return ret;
217}
Vladimir Davydov33a690c2014-10-09 15:28:43 -0700218#else
Vladimir Davydovf7ce3192015-02-12 14:59:20 -0800219static inline int init_memcg_params(struct kmem_cache *s,
220 struct mem_cgroup *memcg, struct kmem_cache *root_cache)
Vladimir Davydov33a690c2014-10-09 15:28:43 -0700221{
222 return 0;
223}
224
Vladimir Davydovf7ce3192015-02-12 14:59:20 -0800225static inline void destroy_memcg_params(struct kmem_cache *s)
Vladimir Davydov33a690c2014-10-09 15:28:43 -0700226{
227}
Johannes Weiner127424c2016-01-20 15:02:32 -0800228#endif /* CONFIG_MEMCG && !CONFIG_SLOB */
Glauber Costa55007d82012-12-18 14:22:38 -0800229
Christoph Lameter039363f2012-07-06 15:25:10 -0500230/*
Joonsoo Kim423c9292014-10-09 15:26:22 -0700231 * Find a mergeable slab cache
232 */
233int slab_unmergeable(struct kmem_cache *s)
234{
235 if (slab_nomerge || (s->flags & SLAB_NEVER_MERGE))
236 return 1;
237
238 if (!is_root_cache(s))
239 return 1;
240
241 if (s->ctor)
242 return 1;
243
244 /*
245 * We may have set a slab to be unmergeable during bootstrap.
246 */
247 if (s->refcount < 0)
248 return 1;
249
250 return 0;
251}
252
253struct kmem_cache *find_mergeable(size_t size, size_t align,
254 unsigned long flags, const char *name, void (*ctor)(void *))
255{
256 struct kmem_cache *s;
257
258 if (slab_nomerge || (flags & SLAB_NEVER_MERGE))
259 return NULL;
260
261 if (ctor)
262 return NULL;
263
264 size = ALIGN(size, sizeof(void *));
265 align = calculate_alignment(flags, align, size);
266 size = ALIGN(size, align);
267 flags = kmem_cache_flags(size, flags, name, NULL);
268
Joonsoo Kim54362052014-12-10 15:42:18 -0800269 list_for_each_entry_reverse(s, &slab_caches, list) {
Joonsoo Kim423c9292014-10-09 15:26:22 -0700270 if (slab_unmergeable(s))
271 continue;
272
273 if (size > s->size)
274 continue;
275
276 if ((flags & SLAB_MERGE_SAME) != (s->flags & SLAB_MERGE_SAME))
277 continue;
278 /*
279 * Check if alignment is compatible.
280 * Courtesy of Adrian Drzewiecki
281 */
282 if ((s->size & ~(align - 1)) != s->size)
283 continue;
284
285 if (s->size - size >= sizeof(void *))
286 continue;
287
Joonsoo Kim95069ac82014-11-13 15:19:25 -0800288 if (IS_ENABLED(CONFIG_SLAB) && align &&
289 (align > s->align || s->align % align))
290 continue;
291
Joonsoo Kim423c9292014-10-09 15:26:22 -0700292 return s;
293 }
294 return NULL;
295}
296
297/*
Christoph Lameter45906852012-11-28 16:23:16 +0000298 * Figure out what the alignment of the objects will be given a set of
299 * flags, a user specified alignment and the size of the objects.
300 */
301unsigned long calculate_alignment(unsigned long flags,
302 unsigned long align, unsigned long size)
303{
304 /*
305 * If the user wants hardware cache aligned objects then follow that
306 * suggestion if the object is sufficiently large.
307 *
308 * The hardware cache alignment cannot override the specified
309 * alignment though. If that is greater then use it.
310 */
311 if (flags & SLAB_HWCACHE_ALIGN) {
312 unsigned long ralign = cache_line_size();
313 while (size <= ralign / 2)
314 ralign /= 2;
315 align = max(align, ralign);
316 }
317
318 if (align < ARCH_SLAB_MINALIGN)
319 align = ARCH_SLAB_MINALIGN;
320
321 return ALIGN(align, sizeof(void *));
322}
323
Vladimir Davydovc9a77a72015-11-05 18:45:08 -0800324static struct kmem_cache *create_cache(const char *name,
325 size_t object_size, size_t size, size_t align,
326 unsigned long flags, void (*ctor)(void *),
327 struct mem_cgroup *memcg, struct kmem_cache *root_cache)
Vladimir Davydov794b1242014-04-07 15:39:26 -0700328{
329 struct kmem_cache *s;
330 int err;
331
332 err = -ENOMEM;
333 s = kmem_cache_zalloc(kmem_cache, GFP_KERNEL);
334 if (!s)
335 goto out;
336
337 s->name = name;
338 s->object_size = object_size;
339 s->size = size;
340 s->align = align;
341 s->ctor = ctor;
342
Vladimir Davydovf7ce3192015-02-12 14:59:20 -0800343 err = init_memcg_params(s, memcg, root_cache);
Vladimir Davydov794b1242014-04-07 15:39:26 -0700344 if (err)
345 goto out_free_cache;
346
347 err = __kmem_cache_create(s, flags);
348 if (err)
349 goto out_free_cache;
350
351 s->refcount = 1;
352 list_add(&s->list, &slab_caches);
Vladimir Davydov794b1242014-04-07 15:39:26 -0700353out:
354 if (err)
355 return ERR_PTR(err);
356 return s;
357
358out_free_cache:
Vladimir Davydovf7ce3192015-02-12 14:59:20 -0800359 destroy_memcg_params(s);
Vaishali Thakkar7c4da062015-02-10 14:09:40 -0800360 kmem_cache_free(kmem_cache, s);
Vladimir Davydov794b1242014-04-07 15:39:26 -0700361 goto out;
362}
Christoph Lameter45906852012-11-28 16:23:16 +0000363
364/*
Christoph Lameter039363f2012-07-06 15:25:10 -0500365 * kmem_cache_create - Create a cache.
366 * @name: A string which is used in /proc/slabinfo to identify this cache.
367 * @size: The size of objects to be created in this cache.
368 * @align: The required alignment for the objects.
369 * @flags: SLAB flags
370 * @ctor: A constructor for the objects.
371 *
372 * Returns a ptr to the cache on success, NULL on failure.
373 * Cannot be called within a interrupt, but can be interrupted.
374 * The @ctor is run when new pages are allocated by the cache.
375 *
376 * The flags are
377 *
378 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
379 * to catch references to uninitialised memory.
380 *
381 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
382 * for buffer overruns.
383 *
384 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
385 * cacheline. This can be beneficial if you're counting cycles as closely
386 * as davem.
387 */
Glauber Costa2633d7a2012-12-18 14:22:34 -0800388struct kmem_cache *
Vladimir Davydov794b1242014-04-07 15:39:26 -0700389kmem_cache_create(const char *name, size_t size, size_t align,
390 unsigned long flags, void (*ctor)(void *))
Christoph Lameter039363f2012-07-06 15:25:10 -0500391{
Alexandru Moise40911a72015-11-05 18:45:43 -0800392 struct kmem_cache *s = NULL;
Andrzej Hajda3dec16e2015-02-13 14:36:38 -0800393 const char *cache_name;
Vladimir Davydov3965fc32014-01-23 15:52:55 -0800394 int err;
Christoph Lameter039363f2012-07-06 15:25:10 -0500395
Pekka Enbergb9205362012-08-16 10:12:18 +0300396 get_online_cpus();
Vladimir Davydov03afc0e2014-06-04 16:07:20 -0700397 get_online_mems();
Vladimir Davydov05257a12015-02-12 14:59:01 -0800398 memcg_get_cache_ids();
Vladimir Davydov03afc0e2014-06-04 16:07:20 -0700399
Pekka Enbergb9205362012-08-16 10:12:18 +0300400 mutex_lock(&slab_mutex);
Christoph Lameter686d5502012-09-05 00:20:33 +0000401
Vladimir Davydov794b1242014-04-07 15:39:26 -0700402 err = kmem_cache_sanity_check(name, size);
Andrew Morton3aa24f52014-10-09 15:25:58 -0700403 if (err) {
Vladimir Davydov3965fc32014-01-23 15:52:55 -0800404 goto out_unlock;
Andrew Morton3aa24f52014-10-09 15:25:58 -0700405 }
Christoph Lameter686d5502012-09-05 00:20:33 +0000406
Glauber Costad8843922012-10-17 15:36:51 +0400407 /*
408 * Some allocators will constraint the set of valid flags to a subset
409 * of all flags. We expect them to define CACHE_CREATE_MASK in this
410 * case, and we'll just provide them with a sanitized version of the
411 * passed flags.
412 */
413 flags &= CACHE_CREATE_MASK;
Christoph Lameter686d5502012-09-05 00:20:33 +0000414
Vladimir Davydov794b1242014-04-07 15:39:26 -0700415 s = __kmem_cache_alias(name, size, align, flags, ctor);
416 if (s)
Vladimir Davydov3965fc32014-01-23 15:52:55 -0800417 goto out_unlock;
Glauber Costa2633d7a2012-12-18 14:22:34 -0800418
Andrzej Hajda3dec16e2015-02-13 14:36:38 -0800419 cache_name = kstrdup_const(name, GFP_KERNEL);
Vladimir Davydov794b1242014-04-07 15:39:26 -0700420 if (!cache_name) {
421 err = -ENOMEM;
422 goto out_unlock;
423 }
Glauber Costa2633d7a2012-12-18 14:22:34 -0800424
Vladimir Davydovc9a77a72015-11-05 18:45:08 -0800425 s = create_cache(cache_name, size, size,
426 calculate_alignment(flags, align, size),
427 flags, ctor, NULL, NULL);
Vladimir Davydov794b1242014-04-07 15:39:26 -0700428 if (IS_ERR(s)) {
429 err = PTR_ERR(s);
Andrzej Hajda3dec16e2015-02-13 14:36:38 -0800430 kfree_const(cache_name);
Vladimir Davydov794b1242014-04-07 15:39:26 -0700431 }
Vladimir Davydov3965fc32014-01-23 15:52:55 -0800432
433out_unlock:
Christoph Lameter20cea962012-07-06 15:25:13 -0500434 mutex_unlock(&slab_mutex);
Vladimir Davydov03afc0e2014-06-04 16:07:20 -0700435
Vladimir Davydov05257a12015-02-12 14:59:01 -0800436 memcg_put_cache_ids();
Vladimir Davydov03afc0e2014-06-04 16:07:20 -0700437 put_online_mems();
Christoph Lameter20cea962012-07-06 15:25:13 -0500438 put_online_cpus();
439
Dave Jonesba3253c72014-01-29 14:05:48 -0800440 if (err) {
Christoph Lameter686d5502012-09-05 00:20:33 +0000441 if (flags & SLAB_PANIC)
442 panic("kmem_cache_create: Failed to create slab '%s'. Error %d\n",
443 name, err);
444 else {
Joe Perches11705322016-03-17 14:19:50 -0700445 pr_warn("kmem_cache_create(%s) failed with error %d\n",
Christoph Lameter686d5502012-09-05 00:20:33 +0000446 name, err);
447 dump_stack();
448 }
Christoph Lameter686d5502012-09-05 00:20:33 +0000449 return NULL;
450 }
Christoph Lameter039363f2012-07-06 15:25:10 -0500451 return s;
Glauber Costa2633d7a2012-12-18 14:22:34 -0800452}
Christoph Lameter039363f2012-07-06 15:25:10 -0500453EXPORT_SYMBOL(kmem_cache_create);
Christoph Lameter97d06602012-07-06 15:25:11 -0500454
Vladimir Davydovc9a77a72015-11-05 18:45:08 -0800455static int shutdown_cache(struct kmem_cache *s,
Vladimir Davydovd5b3cf72015-02-10 14:11:47 -0800456 struct list_head *release, bool *need_rcu_barrier)
457{
Vladimir Davydovcd918c52015-11-05 18:45:14 -0800458 if (__kmem_cache_shutdown(s) != 0)
Vladimir Davydovd5b3cf72015-02-10 14:11:47 -0800459 return -EBUSY;
Vladimir Davydovd5b3cf72015-02-10 14:11:47 -0800460
461 if (s->flags & SLAB_DESTROY_BY_RCU)
462 *need_rcu_barrier = true;
463
Vladimir Davydovd5b3cf72015-02-10 14:11:47 -0800464 list_move(&s->list, release);
465 return 0;
466}
467
Vladimir Davydovc9a77a72015-11-05 18:45:08 -0800468static void release_caches(struct list_head *release, bool need_rcu_barrier)
Vladimir Davydovd5b3cf72015-02-10 14:11:47 -0800469{
470 struct kmem_cache *s, *s2;
471
472 if (need_rcu_barrier)
473 rcu_barrier();
474
475 list_for_each_entry_safe(s, s2, release, list) {
476#ifdef SLAB_SUPPORTS_SYSFS
477 sysfs_slab_remove(s);
478#else
479 slab_kmem_cache_release(s);
480#endif
481 }
482}
483
Johannes Weiner127424c2016-01-20 15:02:32 -0800484#if defined(CONFIG_MEMCG) && !defined(CONFIG_SLOB)
Vladimir Davydov794b1242014-04-07 15:39:26 -0700485/*
Vladimir Davydov776ed0f2014-06-04 16:10:02 -0700486 * memcg_create_kmem_cache - Create a cache for a memory cgroup.
Vladimir Davydov794b1242014-04-07 15:39:26 -0700487 * @memcg: The memory cgroup the new cache is for.
488 * @root_cache: The parent of the new cache.
489 *
490 * This function attempts to create a kmem cache that will serve allocation
491 * requests going from @memcg to @root_cache. The new cache inherits properties
492 * from its parent.
493 */
Vladimir Davydovd5b3cf72015-02-10 14:11:47 -0800494void memcg_create_kmem_cache(struct mem_cgroup *memcg,
495 struct kmem_cache *root_cache)
Vladimir Davydov794b1242014-04-07 15:39:26 -0700496{
Vladimir Davydov3e0350a2015-02-10 14:11:44 -0800497 static char memcg_name_buf[NAME_MAX + 1]; /* protected by slab_mutex */
Michal Hocko33398cf2015-09-08 15:01:02 -0700498 struct cgroup_subsys_state *css = &memcg->css;
Vladimir Davydovf7ce3192015-02-12 14:59:20 -0800499 struct memcg_cache_array *arr;
Vladimir Davydovbd673142014-06-04 16:07:40 -0700500 struct kmem_cache *s = NULL;
Vladimir Davydov794b1242014-04-07 15:39:26 -0700501 char *cache_name;
Vladimir Davydovf7ce3192015-02-12 14:59:20 -0800502 int idx;
Vladimir Davydov794b1242014-04-07 15:39:26 -0700503
504 get_online_cpus();
Vladimir Davydov03afc0e2014-06-04 16:07:20 -0700505 get_online_mems();
506
Vladimir Davydov794b1242014-04-07 15:39:26 -0700507 mutex_lock(&slab_mutex);
508
Vladimir Davydov2a4db7e2015-02-12 14:59:32 -0800509 /*
Johannes Weiner567e9ab2016-01-20 15:02:24 -0800510 * The memory cgroup could have been offlined while the cache
Vladimir Davydov2a4db7e2015-02-12 14:59:32 -0800511 * creation work was pending.
512 */
Vladimir Davydovb6ecd2d2016-03-17 14:18:33 -0700513 if (memcg->kmem_state != KMEM_ONLINE)
Vladimir Davydov2a4db7e2015-02-12 14:59:32 -0800514 goto out_unlock;
515
Vladimir Davydovf7ce3192015-02-12 14:59:20 -0800516 idx = memcg_cache_id(memcg);
517 arr = rcu_dereference_protected(root_cache->memcg_params.memcg_caches,
518 lockdep_is_held(&slab_mutex));
519
Vladimir Davydovd5b3cf72015-02-10 14:11:47 -0800520 /*
521 * Since per-memcg caches are created asynchronously on first
522 * allocation (see memcg_kmem_get_cache()), several threads can try to
523 * create the same cache, but only one of them may succeed.
524 */
Vladimir Davydovf7ce3192015-02-12 14:59:20 -0800525 if (arr->entries[idx])
Vladimir Davydovd5b3cf72015-02-10 14:11:47 -0800526 goto out_unlock;
527
Vladimir Davydovf1008362015-02-12 14:59:29 -0800528 cgroup_name(css->cgroup, memcg_name_buf, sizeof(memcg_name_buf));
Johannes Weiner73f576c2016-07-20 15:44:57 -0700529 cache_name = kasprintf(GFP_KERNEL, "%s(%llu:%s)", root_cache->name,
530 css->serial_nr, memcg_name_buf);
Vladimir Davydov794b1242014-04-07 15:39:26 -0700531 if (!cache_name)
532 goto out_unlock;
533
Vladimir Davydovc9a77a72015-11-05 18:45:08 -0800534 s = create_cache(cache_name, root_cache->object_size,
535 root_cache->size, root_cache->align,
Greg Thelenf773e362016-11-10 10:46:41 -0800536 root_cache->flags & CACHE_CREATE_MASK,
537 root_cache->ctor, memcg, root_cache);
Vladimir Davydovd5b3cf72015-02-10 14:11:47 -0800538 /*
539 * If we could not create a memcg cache, do not complain, because
540 * that's not critical at all as we can always proceed with the root
541 * cache.
542 */
Vladimir Davydovbd673142014-06-04 16:07:40 -0700543 if (IS_ERR(s)) {
Vladimir Davydov794b1242014-04-07 15:39:26 -0700544 kfree(cache_name);
Vladimir Davydovd5b3cf72015-02-10 14:11:47 -0800545 goto out_unlock;
Vladimir Davydovbd673142014-06-04 16:07:40 -0700546 }
Vladimir Davydov794b1242014-04-07 15:39:26 -0700547
Vladimir Davydov426589f2015-02-12 14:59:23 -0800548 list_add(&s->memcg_params.list, &root_cache->memcg_params.list);
549
Vladimir Davydovd5b3cf72015-02-10 14:11:47 -0800550 /*
551 * Since readers won't lock (see cache_from_memcg_idx()), we need a
552 * barrier here to ensure nobody will see the kmem_cache partially
553 * initialized.
554 */
555 smp_wmb();
Vladimir Davydovf7ce3192015-02-12 14:59:20 -0800556 arr->entries[idx] = s;
Vladimir Davydovd5b3cf72015-02-10 14:11:47 -0800557
Vladimir Davydov794b1242014-04-07 15:39:26 -0700558out_unlock:
559 mutex_unlock(&slab_mutex);
Vladimir Davydov03afc0e2014-06-04 16:07:20 -0700560
561 put_online_mems();
Vladimir Davydov794b1242014-04-07 15:39:26 -0700562 put_online_cpus();
563}
Vladimir Davydovb8529902014-04-07 15:39:28 -0700564
Vladimir Davydov2a4db7e2015-02-12 14:59:32 -0800565void memcg_deactivate_kmem_caches(struct mem_cgroup *memcg)
566{
567 int idx;
568 struct memcg_cache_array *arr;
Vladimir Davydovd6e0b7f2015-02-12 14:59:47 -0800569 struct kmem_cache *s, *c;
Vladimir Davydov2a4db7e2015-02-12 14:59:32 -0800570
571 idx = memcg_cache_id(memcg);
572
Vladimir Davydovd6e0b7f2015-02-12 14:59:47 -0800573 get_online_cpus();
574 get_online_mems();
575
Vladimir Davydovbc01eb92017-03-17 00:48:31 +0000576#ifdef CONFIG_SLUB
577 /*
578 * In case of SLUB, we need to disable empty slab caching to
579 * avoid pinning the offline memory cgroup by freeable kmem
580 * pages charged to it. SLAB doesn't need this, as it
581 * periodically purges unused slabs.
582 */
583 mutex_lock(&slab_mutex);
584 list_for_each_entry(s, &slab_caches, list) {
585 c = is_root_cache(s) ? cache_from_memcg_idx(s, idx) : NULL;
586 if (c) {
587 c->cpu_partial = 0;
588 c->min_partial = 0;
589 }
590 }
591 mutex_unlock(&slab_mutex);
592 /*
593 * kmem_cache->cpu_partial is checked locklessly (see
594 * put_cpu_partial()). Make sure the change is visible.
595 */
596 synchronize_sched();
597#endif
598
Vladimir Davydov2a4db7e2015-02-12 14:59:32 -0800599 mutex_lock(&slab_mutex);
600 list_for_each_entry(s, &slab_caches, list) {
601 if (!is_root_cache(s))
602 continue;
603
604 arr = rcu_dereference_protected(s->memcg_params.memcg_caches,
605 lockdep_is_held(&slab_mutex));
Vladimir Davydovd6e0b7f2015-02-12 14:59:47 -0800606 c = arr->entries[idx];
607 if (!c)
608 continue;
609
Vladimir Davydovbc01eb92017-03-17 00:48:31 +0000610 __kmem_cache_shrink(c);
Vladimir Davydov2a4db7e2015-02-12 14:59:32 -0800611 arr->entries[idx] = NULL;
612 }
613 mutex_unlock(&slab_mutex);
Vladimir Davydovd6e0b7f2015-02-12 14:59:47 -0800614
615 put_online_mems();
616 put_online_cpus();
Vladimir Davydov2a4db7e2015-02-12 14:59:32 -0800617}
618
Vladimir Davydovd60fdcc2015-11-05 18:45:11 -0800619static int __shutdown_memcg_cache(struct kmem_cache *s,
620 struct list_head *release, bool *need_rcu_barrier)
621{
622 BUG_ON(is_root_cache(s));
623
624 if (shutdown_cache(s, release, need_rcu_barrier))
625 return -EBUSY;
626
627 list_del(&s->memcg_params.list);
628 return 0;
629}
630
Vladimir Davydovd5b3cf72015-02-10 14:11:47 -0800631void memcg_destroy_kmem_caches(struct mem_cgroup *memcg)
Vladimir Davydovb8529902014-04-07 15:39:28 -0700632{
Vladimir Davydovd5b3cf72015-02-10 14:11:47 -0800633 LIST_HEAD(release);
634 bool need_rcu_barrier = false;
635 struct kmem_cache *s, *s2;
Vladimir Davydovb8529902014-04-07 15:39:28 -0700636
Vladimir Davydovd5b3cf72015-02-10 14:11:47 -0800637 get_online_cpus();
638 get_online_mems();
Vladimir Davydovb8529902014-04-07 15:39:28 -0700639
Vladimir Davydovb8529902014-04-07 15:39:28 -0700640 mutex_lock(&slab_mutex);
Vladimir Davydovd5b3cf72015-02-10 14:11:47 -0800641 list_for_each_entry_safe(s, s2, &slab_caches, list) {
Vladimir Davydovf7ce3192015-02-12 14:59:20 -0800642 if (is_root_cache(s) || s->memcg_params.memcg != memcg)
Vladimir Davydovd5b3cf72015-02-10 14:11:47 -0800643 continue;
644 /*
645 * The cgroup is about to be freed and therefore has no charges
646 * left. Hence, all its caches must be empty by now.
647 */
Vladimir Davydovd60fdcc2015-11-05 18:45:11 -0800648 BUG_ON(__shutdown_memcg_cache(s, &release, &need_rcu_barrier));
Vladimir Davydovd5b3cf72015-02-10 14:11:47 -0800649 }
650 mutex_unlock(&slab_mutex);
Vladimir Davydovb8529902014-04-07 15:39:28 -0700651
Vladimir Davydovd5b3cf72015-02-10 14:11:47 -0800652 put_online_mems();
653 put_online_cpus();
654
Vladimir Davydovc9a77a72015-11-05 18:45:08 -0800655 release_caches(&release, need_rcu_barrier);
Vladimir Davydovb8529902014-04-07 15:39:28 -0700656}
Vladimir Davydovd60fdcc2015-11-05 18:45:11 -0800657
658static int shutdown_memcg_caches(struct kmem_cache *s,
659 struct list_head *release, bool *need_rcu_barrier)
660{
661 struct memcg_cache_array *arr;
662 struct kmem_cache *c, *c2;
663 LIST_HEAD(busy);
664 int i;
665
666 BUG_ON(!is_root_cache(s));
667
668 /*
669 * First, shutdown active caches, i.e. caches that belong to online
670 * memory cgroups.
671 */
672 arr = rcu_dereference_protected(s->memcg_params.memcg_caches,
673 lockdep_is_held(&slab_mutex));
674 for_each_memcg_cache_index(i) {
675 c = arr->entries[i];
676 if (!c)
677 continue;
678 if (__shutdown_memcg_cache(c, release, need_rcu_barrier))
679 /*
680 * The cache still has objects. Move it to a temporary
681 * list so as not to try to destroy it for a second
682 * time while iterating over inactive caches below.
683 */
684 list_move(&c->memcg_params.list, &busy);
685 else
686 /*
687 * The cache is empty and will be destroyed soon. Clear
688 * the pointer to it in the memcg_caches array so that
689 * it will never be accessed even if the root cache
690 * stays alive.
691 */
692 arr->entries[i] = NULL;
693 }
694
695 /*
696 * Second, shutdown all caches left from memory cgroups that are now
697 * offline.
698 */
699 list_for_each_entry_safe(c, c2, &s->memcg_params.list,
700 memcg_params.list)
701 __shutdown_memcg_cache(c, release, need_rcu_barrier);
702
703 list_splice(&busy, &s->memcg_params.list);
704
705 /*
706 * A cache being destroyed must be empty. In particular, this means
707 * that all per memcg caches attached to it must be empty too.
708 */
709 if (!list_empty(&s->memcg_params.list))
710 return -EBUSY;
711 return 0;
712}
713#else
714static inline int shutdown_memcg_caches(struct kmem_cache *s,
715 struct list_head *release, bool *need_rcu_barrier)
716{
717 return 0;
718}
Johannes Weiner127424c2016-01-20 15:02:32 -0800719#endif /* CONFIG_MEMCG && !CONFIG_SLOB */
Vladimir Davydov794b1242014-04-07 15:39:26 -0700720
Christoph Lameter41a21282014-05-06 12:50:08 -0700721void slab_kmem_cache_release(struct kmem_cache *s)
722{
Dmitry Safonov52b4b952016-02-17 13:11:37 -0800723 __kmem_cache_release(s);
Vladimir Davydovf7ce3192015-02-12 14:59:20 -0800724 destroy_memcg_params(s);
Andrzej Hajda3dec16e2015-02-13 14:36:38 -0800725 kfree_const(s->name);
Christoph Lameter41a21282014-05-06 12:50:08 -0700726 kmem_cache_free(kmem_cache, s);
727}
728
Christoph Lameter945cf2b2012-09-04 23:18:33 +0000729void kmem_cache_destroy(struct kmem_cache *s)
730{
Vladimir Davydovd5b3cf72015-02-10 14:11:47 -0800731 LIST_HEAD(release);
732 bool need_rcu_barrier = false;
Vladimir Davydovd60fdcc2015-11-05 18:45:11 -0800733 int err;
Vladimir Davydovd5b3cf72015-02-10 14:11:47 -0800734
Sergey Senozhatsky3942d292015-09-08 15:00:50 -0700735 if (unlikely(!s))
736 return;
737
Christoph Lameter945cf2b2012-09-04 23:18:33 +0000738 get_online_cpus();
Vladimir Davydov03afc0e2014-06-04 16:07:20 -0700739 get_online_mems();
740
Alexander Potapenko55834c52016-05-20 16:59:11 -0700741 kasan_cache_destroy(s);
Christoph Lameter945cf2b2012-09-04 23:18:33 +0000742 mutex_lock(&slab_mutex);
Vladimir Davydovb8529902014-04-07 15:39:28 -0700743
Christoph Lameter945cf2b2012-09-04 23:18:33 +0000744 s->refcount--;
Vladimir Davydovb8529902014-04-07 15:39:28 -0700745 if (s->refcount)
746 goto out_unlock;
Christoph Lameter945cf2b2012-09-04 23:18:33 +0000747
Vladimir Davydovd60fdcc2015-11-05 18:45:11 -0800748 err = shutdown_memcg_caches(s, &release, &need_rcu_barrier);
749 if (!err)
Vladimir Davydovcd918c52015-11-05 18:45:14 -0800750 err = shutdown_cache(s, &release, &need_rcu_barrier);
Vladimir Davydovb8529902014-04-07 15:39:28 -0700751
Vladimir Davydovcd918c52015-11-05 18:45:14 -0800752 if (err) {
Joe Perches756a025f02016-03-17 14:19:47 -0700753 pr_err("kmem_cache_destroy %s: Slab cache still has objects\n",
754 s->name);
Vladimir Davydovcd918c52015-11-05 18:45:14 -0800755 dump_stack();
756 }
Vladimir Davydovb8529902014-04-07 15:39:28 -0700757out_unlock:
758 mutex_unlock(&slab_mutex);
Vladimir Davydovd5b3cf72015-02-10 14:11:47 -0800759
Vladimir Davydov03afc0e2014-06-04 16:07:20 -0700760 put_online_mems();
Christoph Lameter945cf2b2012-09-04 23:18:33 +0000761 put_online_cpus();
Vladimir Davydovd5b3cf72015-02-10 14:11:47 -0800762
Vladimir Davydovc9a77a72015-11-05 18:45:08 -0800763 release_caches(&release, need_rcu_barrier);
Christoph Lameter945cf2b2012-09-04 23:18:33 +0000764}
765EXPORT_SYMBOL(kmem_cache_destroy);
766
Vladimir Davydov03afc0e2014-06-04 16:07:20 -0700767/**
768 * kmem_cache_shrink - Shrink a cache.
769 * @cachep: The cache to shrink.
770 *
771 * Releases as many slabs as possible for a cache.
772 * To help debugging, a zero exit status indicates all slabs were released.
773 */
774int kmem_cache_shrink(struct kmem_cache *cachep)
775{
776 int ret;
777
778 get_online_cpus();
779 get_online_mems();
Alexander Potapenko55834c52016-05-20 16:59:11 -0700780 kasan_cache_shrink(cachep);
Vladimir Davydovbc01eb92017-03-17 00:48:31 +0000781 ret = __kmem_cache_shrink(cachep);
Vladimir Davydov03afc0e2014-06-04 16:07:20 -0700782 put_online_mems();
783 put_online_cpus();
784 return ret;
785}
786EXPORT_SYMBOL(kmem_cache_shrink);
787
Denis Kirjanovfda90122015-11-05 18:44:59 -0800788bool slab_is_available(void)
Christoph Lameter97d06602012-07-06 15:25:11 -0500789{
790 return slab_state >= UP;
791}
Glauber Costab7454ad2012-10-19 18:20:25 +0400792
Christoph Lameter45530c42012-11-28 16:23:07 +0000793#ifndef CONFIG_SLOB
794/* Create a cache during boot when no slab services are available yet */
795void __init create_boot_cache(struct kmem_cache *s, const char *name, size_t size,
796 unsigned long flags)
797{
798 int err;
799
800 s->name = name;
801 s->size = s->object_size = size;
Christoph Lameter45906852012-11-28 16:23:16 +0000802 s->align = calculate_alignment(flags, ARCH_KMALLOC_MINALIGN, size);
Vladimir Davydovf7ce3192015-02-12 14:59:20 -0800803
804 slab_init_memcg_params(s);
805
Christoph Lameter45530c42012-11-28 16:23:07 +0000806 err = __kmem_cache_create(s, flags);
807
808 if (err)
Christoph Lameter31ba7342013-01-10 19:00:53 +0000809 panic("Creation of kmalloc slab %s size=%zu failed. Reason %d\n",
Christoph Lameter45530c42012-11-28 16:23:07 +0000810 name, size, err);
811
812 s->refcount = -1; /* Exempt from merging for now */
813}
814
815struct kmem_cache *__init create_kmalloc_cache(const char *name, size_t size,
816 unsigned long flags)
817{
818 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
819
820 if (!s)
821 panic("Out of memory when creating slab %s\n", name);
822
823 create_boot_cache(s, name, size, flags);
824 list_add(&s->list, &slab_caches);
825 s->refcount = 1;
826 return s;
827}
828
Christoph Lameter9425c582013-01-10 19:12:17 +0000829struct kmem_cache *kmalloc_caches[KMALLOC_SHIFT_HIGH + 1];
830EXPORT_SYMBOL(kmalloc_caches);
831
832#ifdef CONFIG_ZONE_DMA
833struct kmem_cache *kmalloc_dma_caches[KMALLOC_SHIFT_HIGH + 1];
834EXPORT_SYMBOL(kmalloc_dma_caches);
835#endif
836
Christoph Lameterf97d5f62013-01-10 19:12:17 +0000837/*
Christoph Lameter2c59dd62013-01-10 19:14:19 +0000838 * Conversion table for small slabs sizes / 8 to the index in the
839 * kmalloc array. This is necessary for slabs < 192 since we have non power
840 * of two cache sizes there. The size of larger slabs can be determined using
841 * fls.
842 */
843static s8 size_index[24] = {
844 3, /* 8 */
845 4, /* 16 */
846 5, /* 24 */
847 5, /* 32 */
848 6, /* 40 */
849 6, /* 48 */
850 6, /* 56 */
851 6, /* 64 */
852 1, /* 72 */
853 1, /* 80 */
854 1, /* 88 */
855 1, /* 96 */
856 7, /* 104 */
857 7, /* 112 */
858 7, /* 120 */
859 7, /* 128 */
860 2, /* 136 */
861 2, /* 144 */
862 2, /* 152 */
863 2, /* 160 */
864 2, /* 168 */
865 2, /* 176 */
866 2, /* 184 */
867 2 /* 192 */
868};
869
870static inline int size_index_elem(size_t bytes)
871{
872 return (bytes - 1) / 8;
873}
874
875/*
876 * Find the kmem_cache structure that serves a given size of
877 * allocation
878 */
879struct kmem_cache *kmalloc_slab(size_t size, gfp_t flags)
880{
881 int index;
882
Joonsoo Kim9de1bc82013-08-02 11:02:42 +0900883 if (unlikely(size > KMALLOC_MAX_SIZE)) {
Sasha Levin907985f2013-06-10 15:18:00 -0400884 WARN_ON_ONCE(!(flags & __GFP_NOWARN));
Christoph Lameter6286ae92013-05-03 15:43:18 +0000885 return NULL;
Sasha Levin907985f2013-06-10 15:18:00 -0400886 }
Christoph Lameter6286ae92013-05-03 15:43:18 +0000887
Christoph Lameter2c59dd62013-01-10 19:14:19 +0000888 if (size <= 192) {
889 if (!size)
890 return ZERO_SIZE_PTR;
891
892 index = size_index[size_index_elem(size)];
893 } else
894 index = fls(size - 1);
895
896#ifdef CONFIG_ZONE_DMA
Joonsoo Kimb1e05412013-02-04 23:46:46 +0900897 if (unlikely((flags & GFP_DMA)))
Christoph Lameter2c59dd62013-01-10 19:14:19 +0000898 return kmalloc_dma_caches[index];
899
900#endif
901 return kmalloc_caches[index];
902}
903
904/*
Gavin Guo4066c332015-06-24 16:55:54 -0700905 * kmalloc_info[] is to make slub_debug=,kmalloc-xx option work at boot time.
906 * kmalloc_index() supports up to 2^26=64MB, so the final entry of the table is
907 * kmalloc-67108864.
908 */
909static struct {
910 const char *name;
911 unsigned long size;
912} const kmalloc_info[] __initconst = {
913 {NULL, 0}, {"kmalloc-96", 96},
914 {"kmalloc-192", 192}, {"kmalloc-8", 8},
915 {"kmalloc-16", 16}, {"kmalloc-32", 32},
916 {"kmalloc-64", 64}, {"kmalloc-128", 128},
917 {"kmalloc-256", 256}, {"kmalloc-512", 512},
918 {"kmalloc-1024", 1024}, {"kmalloc-2048", 2048},
919 {"kmalloc-4096", 4096}, {"kmalloc-8192", 8192},
920 {"kmalloc-16384", 16384}, {"kmalloc-32768", 32768},
921 {"kmalloc-65536", 65536}, {"kmalloc-131072", 131072},
922 {"kmalloc-262144", 262144}, {"kmalloc-524288", 524288},
923 {"kmalloc-1048576", 1048576}, {"kmalloc-2097152", 2097152},
924 {"kmalloc-4194304", 4194304}, {"kmalloc-8388608", 8388608},
925 {"kmalloc-16777216", 16777216}, {"kmalloc-33554432", 33554432},
926 {"kmalloc-67108864", 67108864}
927};
928
929/*
Daniel Sanders34cc6992015-06-24 16:55:57 -0700930 * Patch up the size_index table if we have strange large alignment
931 * requirements for the kmalloc array. This is only the case for
932 * MIPS it seems. The standard arches will not generate any code here.
933 *
934 * Largest permitted alignment is 256 bytes due to the way we
935 * handle the index determination for the smaller caches.
936 *
937 * Make sure that nothing crazy happens if someone starts tinkering
938 * around with ARCH_KMALLOC_MINALIGN
Christoph Lameterf97d5f62013-01-10 19:12:17 +0000939 */
Daniel Sanders34cc6992015-06-24 16:55:57 -0700940void __init setup_kmalloc_cache_index_table(void)
Christoph Lameterf97d5f62013-01-10 19:12:17 +0000941{
942 int i;
943
Christoph Lameter2c59dd62013-01-10 19:14:19 +0000944 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
945 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
946
947 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
948 int elem = size_index_elem(i);
949
950 if (elem >= ARRAY_SIZE(size_index))
951 break;
952 size_index[elem] = KMALLOC_SHIFT_LOW;
953 }
954
955 if (KMALLOC_MIN_SIZE >= 64) {
956 /*
957 * The 96 byte size cache is not used if the alignment
958 * is 64 byte.
959 */
960 for (i = 64 + 8; i <= 96; i += 8)
961 size_index[size_index_elem(i)] = 7;
962
963 }
964
965 if (KMALLOC_MIN_SIZE >= 128) {
966 /*
967 * The 192 byte sized cache is not used if the alignment
968 * is 128 byte. Redirect kmalloc to use the 256 byte cache
969 * instead.
970 */
971 for (i = 128 + 8; i <= 192; i += 8)
972 size_index[size_index_elem(i)] = 8;
973 }
Daniel Sanders34cc6992015-06-24 16:55:57 -0700974}
975
Christoph Lameterae6f2462015-06-30 09:01:11 -0500976static void __init new_kmalloc_cache(int idx, unsigned long flags)
Christoph Lametera9730fc2015-06-29 09:28:08 -0500977{
978 kmalloc_caches[idx] = create_kmalloc_cache(kmalloc_info[idx].name,
979 kmalloc_info[idx].size, flags);
980}
981
Daniel Sanders34cc6992015-06-24 16:55:57 -0700982/*
983 * Create the kmalloc array. Some of the regular kmalloc arrays
984 * may already have been created because they were needed to
985 * enable allocations for slab creation.
986 */
987void __init create_kmalloc_caches(unsigned long flags)
988{
989 int i;
990
Christoph Lametera9730fc2015-06-29 09:28:08 -0500991 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) {
992 if (!kmalloc_caches[i])
993 new_kmalloc_cache(i, flags);
Chris Mason956e46e2013-05-08 15:56:28 -0400994
995 /*
Christoph Lametera9730fc2015-06-29 09:28:08 -0500996 * Caches that are not of the two-to-the-power-of size.
997 * These have to be created immediately after the
998 * earlier power of two caches
Chris Mason956e46e2013-05-08 15:56:28 -0400999 */
Christoph Lametera9730fc2015-06-29 09:28:08 -05001000 if (KMALLOC_MIN_SIZE <= 32 && !kmalloc_caches[1] && i == 6)
1001 new_kmalloc_cache(1, flags);
1002 if (KMALLOC_MIN_SIZE <= 64 && !kmalloc_caches[2] && i == 7)
1003 new_kmalloc_cache(2, flags);
Christoph Lameter8a965b32013-05-03 18:04:18 +00001004 }
1005
Christoph Lameterf97d5f62013-01-10 19:12:17 +00001006 /* Kmalloc array is now usable */
1007 slab_state = UP;
1008
Christoph Lameterf97d5f62013-01-10 19:12:17 +00001009#ifdef CONFIG_ZONE_DMA
1010 for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
1011 struct kmem_cache *s = kmalloc_caches[i];
1012
1013 if (s) {
1014 int size = kmalloc_size(i);
1015 char *n = kasprintf(GFP_NOWAIT,
1016 "dma-kmalloc-%d", size);
1017
1018 BUG_ON(!n);
1019 kmalloc_dma_caches[i] = create_kmalloc_cache(n,
1020 size, SLAB_CACHE_DMA | flags);
1021 }
1022 }
1023#endif
1024}
Christoph Lameter45530c42012-11-28 16:23:07 +00001025#endif /* !CONFIG_SLOB */
1026
Vladimir Davydovcea371f2014-06-04 16:07:04 -07001027/*
1028 * To avoid unnecessary overhead, we pass through large allocation requests
1029 * directly to the page allocator. We use __GFP_COMP, because we will need to
1030 * know the allocation order to free the pages properly in kfree.
1031 */
Vladimir Davydov52383432014-06-04 16:06:39 -07001032void *kmalloc_order(size_t size, gfp_t flags, unsigned int order)
1033{
1034 void *ret;
1035 struct page *page;
1036
1037 flags |= __GFP_COMP;
Vladimir Davydov49491482016-07-26 15:24:24 -07001038 page = alloc_pages(flags, order);
Vladimir Davydov52383432014-06-04 16:06:39 -07001039 ret = page ? page_address(page) : NULL;
1040 kmemleak_alloc(ret, size, 1, flags);
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07001041 kasan_kmalloc_large(ret, size, flags);
Vladimir Davydov52383432014-06-04 16:06:39 -07001042 return ret;
1043}
1044EXPORT_SYMBOL(kmalloc_order);
1045
Christoph Lameterf1b6eb62013-09-04 16:35:34 +00001046#ifdef CONFIG_TRACING
1047void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
1048{
1049 void *ret = kmalloc_order(size, flags, order);
1050 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
1051 return ret;
1052}
1053EXPORT_SYMBOL(kmalloc_order_trace);
1054#endif
Christoph Lameter45530c42012-11-28 16:23:07 +00001055
Thomas Garnier7c00fce2016-07-26 15:21:56 -07001056#ifdef CONFIG_SLAB_FREELIST_RANDOM
1057/* Randomize a generic freelist */
1058static void freelist_randomize(struct rnd_state *state, unsigned int *list,
1059 size_t count)
1060{
1061 size_t i;
1062 unsigned int rand;
1063
1064 for (i = 0; i < count; i++)
1065 list[i] = i;
1066
1067 /* Fisher-Yates shuffle */
1068 for (i = count - 1; i > 0; i--) {
1069 rand = prandom_u32_state(state);
1070 rand %= (i + 1);
1071 swap(list[i], list[rand]);
1072 }
1073}
1074
1075/* Create a random sequence per cache */
1076int cache_random_seq_create(struct kmem_cache *cachep, unsigned int count,
1077 gfp_t gfp)
1078{
1079 struct rnd_state state;
1080
1081 if (count < 2 || cachep->random_seq)
1082 return 0;
1083
1084 cachep->random_seq = kcalloc(count, sizeof(unsigned int), gfp);
1085 if (!cachep->random_seq)
1086 return -ENOMEM;
1087
1088 /* Get best entropy at this stage of boot */
1089 prandom_seed_state(&state, get_random_long());
1090
1091 freelist_randomize(&state, cachep->random_seq, count);
1092 return 0;
1093}
1094
1095/* Destroy the per-cache random freelist sequence */
1096void cache_random_seq_destroy(struct kmem_cache *cachep)
1097{
1098 kfree(cachep->random_seq);
1099 cachep->random_seq = NULL;
1100}
1101#endif /* CONFIG_SLAB_FREELIST_RANDOM */
1102
Glauber Costab7454ad2012-10-19 18:20:25 +04001103#ifdef CONFIG_SLABINFO
Wanpeng Lie9b4db22013-07-04 08:33:24 +08001104
1105#ifdef CONFIG_SLAB
1106#define SLABINFO_RIGHTS (S_IWUSR | S_IRUSR)
1107#else
1108#define SLABINFO_RIGHTS S_IRUSR
1109#endif
1110
Vladimir Davydovb0475012014-12-10 15:44:19 -08001111static void print_slabinfo_header(struct seq_file *m)
Glauber Costabcee6e22012-10-19 18:20:26 +04001112{
1113 /*
1114 * Output format version, so at least we can change it
1115 * without _too_ many complaints.
1116 */
1117#ifdef CONFIG_DEBUG_SLAB
1118 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
1119#else
1120 seq_puts(m, "slabinfo - version: 2.1\n");
1121#endif
Joe Perches756a025f02016-03-17 14:19:47 -07001122 seq_puts(m, "# name <active_objs> <num_objs> <objsize> <objperslab> <pagesperslab>");
Glauber Costabcee6e22012-10-19 18:20:26 +04001123 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
1124 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
1125#ifdef CONFIG_DEBUG_SLAB
Joe Perches756a025f02016-03-17 14:19:47 -07001126 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> <error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Glauber Costabcee6e22012-10-19 18:20:26 +04001127 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
1128#endif
1129 seq_putc(m, '\n');
1130}
1131
Vladimir Davydov1df3b262014-12-10 15:42:16 -08001132void *slab_start(struct seq_file *m, loff_t *pos)
Glauber Costab7454ad2012-10-19 18:20:25 +04001133{
Glauber Costab7454ad2012-10-19 18:20:25 +04001134 mutex_lock(&slab_mutex);
Glauber Costab7454ad2012-10-19 18:20:25 +04001135 return seq_list_start(&slab_caches, *pos);
1136}
1137
Wanpeng Li276a2432013-07-08 08:08:28 +08001138void *slab_next(struct seq_file *m, void *p, loff_t *pos)
Glauber Costab7454ad2012-10-19 18:20:25 +04001139{
1140 return seq_list_next(p, &slab_caches, pos);
1141}
1142
Wanpeng Li276a2432013-07-08 08:08:28 +08001143void slab_stop(struct seq_file *m, void *p)
Glauber Costab7454ad2012-10-19 18:20:25 +04001144{
1145 mutex_unlock(&slab_mutex);
1146}
1147
Glauber Costa749c5412012-12-18 14:23:01 -08001148static void
1149memcg_accumulate_slabinfo(struct kmem_cache *s, struct slabinfo *info)
Glauber Costab7454ad2012-10-19 18:20:25 +04001150{
Glauber Costa749c5412012-12-18 14:23:01 -08001151 struct kmem_cache *c;
1152 struct slabinfo sinfo;
Glauber Costa749c5412012-12-18 14:23:01 -08001153
1154 if (!is_root_cache(s))
1155 return;
1156
Vladimir Davydov426589f2015-02-12 14:59:23 -08001157 for_each_memcg_cache(c, s) {
Glauber Costa749c5412012-12-18 14:23:01 -08001158 memset(&sinfo, 0, sizeof(sinfo));
1159 get_slabinfo(c, &sinfo);
1160
1161 info->active_slabs += sinfo.active_slabs;
1162 info->num_slabs += sinfo.num_slabs;
1163 info->shared_avail += sinfo.shared_avail;
1164 info->active_objs += sinfo.active_objs;
1165 info->num_objs += sinfo.num_objs;
1166 }
1167}
1168
Vladimir Davydovb0475012014-12-10 15:44:19 -08001169static void cache_show(struct kmem_cache *s, struct seq_file *m)
Glauber Costa749c5412012-12-18 14:23:01 -08001170{
Glauber Costa0d7561c2012-10-19 18:20:27 +04001171 struct slabinfo sinfo;
1172
1173 memset(&sinfo, 0, sizeof(sinfo));
1174 get_slabinfo(s, &sinfo);
1175
Glauber Costa749c5412012-12-18 14:23:01 -08001176 memcg_accumulate_slabinfo(s, &sinfo);
1177
Glauber Costa0d7561c2012-10-19 18:20:27 +04001178 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Glauber Costa749c5412012-12-18 14:23:01 -08001179 cache_name(s), sinfo.active_objs, sinfo.num_objs, s->size,
Glauber Costa0d7561c2012-10-19 18:20:27 +04001180 sinfo.objects_per_slab, (1 << sinfo.cache_order));
1181
1182 seq_printf(m, " : tunables %4u %4u %4u",
1183 sinfo.limit, sinfo.batchcount, sinfo.shared);
1184 seq_printf(m, " : slabdata %6lu %6lu %6lu",
1185 sinfo.active_slabs, sinfo.num_slabs, sinfo.shared_avail);
1186 slabinfo_show_stats(m, s);
1187 seq_putc(m, '\n');
Glauber Costab7454ad2012-10-19 18:20:25 +04001188}
1189
Vladimir Davydov1df3b262014-12-10 15:42:16 -08001190static int slab_show(struct seq_file *m, void *p)
Glauber Costa749c5412012-12-18 14:23:01 -08001191{
1192 struct kmem_cache *s = list_entry(p, struct kmem_cache, list);
1193
Vladimir Davydov1df3b262014-12-10 15:42:16 -08001194 if (p == slab_caches.next)
1195 print_slabinfo_header(m);
Vladimir Davydovb0475012014-12-10 15:44:19 -08001196 if (is_root_cache(s))
1197 cache_show(s, m);
1198 return 0;
Glauber Costa749c5412012-12-18 14:23:01 -08001199}
1200
Johannes Weiner127424c2016-01-20 15:02:32 -08001201#if defined(CONFIG_MEMCG) && !defined(CONFIG_SLOB)
Vladimir Davydovb0475012014-12-10 15:44:19 -08001202int memcg_slab_show(struct seq_file *m, void *p)
1203{
1204 struct kmem_cache *s = list_entry(p, struct kmem_cache, list);
1205 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
1206
1207 if (p == slab_caches.next)
1208 print_slabinfo_header(m);
Vladimir Davydovf7ce3192015-02-12 14:59:20 -08001209 if (!is_root_cache(s) && s->memcg_params.memcg == memcg)
Vladimir Davydovb0475012014-12-10 15:44:19 -08001210 cache_show(s, m);
1211 return 0;
1212}
1213#endif
1214
Glauber Costab7454ad2012-10-19 18:20:25 +04001215/*
1216 * slabinfo_op - iterator that generates /proc/slabinfo
1217 *
1218 * Output layout:
1219 * cache-name
1220 * num-active-objs
1221 * total-objs
1222 * object size
1223 * num-active-slabs
1224 * total-slabs
1225 * num-pages-per-slab
1226 * + further values on SMP and with statistics enabled
1227 */
1228static const struct seq_operations slabinfo_op = {
Vladimir Davydov1df3b262014-12-10 15:42:16 -08001229 .start = slab_start,
Wanpeng Li276a2432013-07-08 08:08:28 +08001230 .next = slab_next,
1231 .stop = slab_stop,
Vladimir Davydov1df3b262014-12-10 15:42:16 -08001232 .show = slab_show,
Glauber Costab7454ad2012-10-19 18:20:25 +04001233};
1234
1235static int slabinfo_open(struct inode *inode, struct file *file)
1236{
1237 return seq_open(file, &slabinfo_op);
1238}
1239
1240static const struct file_operations proc_slabinfo_operations = {
1241 .open = slabinfo_open,
1242 .read = seq_read,
1243 .write = slabinfo_write,
1244 .llseek = seq_lseek,
1245 .release = seq_release,
1246};
1247
1248static int __init slab_proc_init(void)
1249{
Wanpeng Lie9b4db22013-07-04 08:33:24 +08001250 proc_create("slabinfo", SLABINFO_RIGHTS, NULL,
1251 &proc_slabinfo_operations);
Glauber Costab7454ad2012-10-19 18:20:25 +04001252 return 0;
1253}
1254module_init(slab_proc_init);
1255#endif /* CONFIG_SLABINFO */
Andrey Ryabinin928cec92014-08-06 16:04:44 -07001256
1257static __always_inline void *__do_krealloc(const void *p, size_t new_size,
1258 gfp_t flags)
1259{
1260 void *ret;
1261 size_t ks = 0;
1262
1263 if (p)
1264 ks = ksize(p);
1265
Andrey Ryabinin0316bec2015-02-13 14:39:42 -08001266 if (ks >= new_size) {
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07001267 kasan_krealloc((void *)p, new_size, flags);
Andrey Ryabinin928cec92014-08-06 16:04:44 -07001268 return (void *)p;
Andrey Ryabinin0316bec2015-02-13 14:39:42 -08001269 }
Andrey Ryabinin928cec92014-08-06 16:04:44 -07001270
1271 ret = kmalloc_track_caller(new_size, flags);
1272 if (ret && p)
1273 memcpy(ret, p, ks);
1274
1275 return ret;
1276}
1277
1278/**
1279 * __krealloc - like krealloc() but don't free @p.
1280 * @p: object to reallocate memory for.
1281 * @new_size: how many bytes of memory are required.
1282 * @flags: the type of memory to allocate.
1283 *
1284 * This function is like krealloc() except it never frees the originally
1285 * allocated buffer. Use this if you don't want to free the buffer immediately
1286 * like, for example, with RCU.
1287 */
1288void *__krealloc(const void *p, size_t new_size, gfp_t flags)
1289{
1290 if (unlikely(!new_size))
1291 return ZERO_SIZE_PTR;
1292
1293 return __do_krealloc(p, new_size, flags);
1294
1295}
1296EXPORT_SYMBOL(__krealloc);
1297
1298/**
1299 * krealloc - reallocate memory. The contents will remain unchanged.
1300 * @p: object to reallocate memory for.
1301 * @new_size: how many bytes of memory are required.
1302 * @flags: the type of memory to allocate.
1303 *
1304 * The contents of the object pointed to are preserved up to the
1305 * lesser of the new and old sizes. If @p is %NULL, krealloc()
1306 * behaves exactly like kmalloc(). If @new_size is 0 and @p is not a
1307 * %NULL pointer, the object pointed to is freed.
1308 */
1309void *krealloc(const void *p, size_t new_size, gfp_t flags)
1310{
1311 void *ret;
1312
1313 if (unlikely(!new_size)) {
1314 kfree(p);
1315 return ZERO_SIZE_PTR;
1316 }
1317
1318 ret = __do_krealloc(p, new_size, flags);
1319 if (ret && p != ret)
1320 kfree(p);
1321
1322 return ret;
1323}
1324EXPORT_SYMBOL(krealloc);
1325
1326/**
1327 * kzfree - like kfree but zero memory
1328 * @p: object to free memory of
1329 *
1330 * The memory of the object @p points to is zeroed before freed.
1331 * If @p is %NULL, kzfree() does nothing.
1332 *
1333 * Note: this function zeroes the whole allocated buffer which can be a good
1334 * deal bigger than the requested buffer size passed to kmalloc(). So be
1335 * careful when using this function in performance sensitive code.
1336 */
1337void kzfree(const void *p)
1338{
1339 size_t ks;
1340 void *mem = (void *)p;
1341
1342 if (unlikely(ZERO_OR_NULL_PTR(mem)))
1343 return;
1344 ks = ksize(mem);
1345 memset(mem, 0, ks);
1346 kfree(mem);
1347}
1348EXPORT_SYMBOL(kzfree);
1349
1350/* Tracepoints definitions. */
1351EXPORT_TRACEPOINT_SYMBOL(kmalloc);
1352EXPORT_TRACEPOINT_SYMBOL(kmem_cache_alloc);
1353EXPORT_TRACEPOINT_SYMBOL(kmalloc_node);
1354EXPORT_TRACEPOINT_SYMBOL(kmem_cache_alloc_node);
1355EXPORT_TRACEPOINT_SYMBOL(kfree);
1356EXPORT_TRACEPOINT_SYMBOL(kmem_cache_free);