blob: 2b0ebb6d071db78b2a4756a4ae476789d838eab5 [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>
Christoph Lameter039363f2012-07-06 15:25:10 -050022
Christoph Lameter97d06602012-07-06 15:25:11 -050023#include "slab.h"
24
25enum slab_state slab_state;
Christoph Lameter18004c52012-07-06 15:25:12 -050026LIST_HEAD(slab_caches);
27DEFINE_MUTEX(slab_mutex);
Christoph Lameter9b030cb2012-09-05 00:20:33 +000028struct kmem_cache *kmem_cache;
Christoph Lameter97d06602012-07-06 15:25:11 -050029
Shuah Khan77be4b12012-08-16 00:09:46 -070030#ifdef CONFIG_DEBUG_VM
Glauber Costa2633d7a2012-12-18 14:22:34 -080031static int kmem_cache_sanity_check(struct mem_cgroup *memcg, const char *name,
32 size_t size)
Shuah Khan77be4b12012-08-16 00:09:46 -070033{
34 struct kmem_cache *s = NULL;
35
36 if (!name || in_interrupt() || size < sizeof(void *) ||
37 size > KMALLOC_MAX_SIZE) {
38 pr_err("kmem_cache_create(%s) integrity check failed\n", name);
39 return -EINVAL;
40 }
41
42 list_for_each_entry(s, &slab_caches, list) {
43 char tmp;
44 int res;
45
46 /*
47 * This happens when the module gets unloaded and doesn't
48 * destroy its slab cache and no-one else reuses the vmalloc
49 * area of the module. Print a warning.
50 */
51 res = probe_kernel_address(s->name, tmp);
52 if (res) {
53 pr_err("Slab cache with size %d has lost its name\n",
54 s->object_size);
55 continue;
56 }
57
Glauber Costa2633d7a2012-12-18 14:22:34 -080058 /*
59 * For simplicity, we won't check this in the list of memcg
60 * caches. We have control over memcg naming, and if there
61 * aren't duplicates in the global list, there won't be any
62 * duplicates in the memcg lists as well.
63 */
64 if (!memcg && !strcmp(s->name, name)) {
Shuah Khan77be4b12012-08-16 00:09:46 -070065 pr_err("%s (%s): Cache name already exists.\n",
66 __func__, name);
67 dump_stack();
68 s = NULL;
69 return -EINVAL;
70 }
71 }
72
73 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
74 return 0;
75}
76#else
Glauber Costa2633d7a2012-12-18 14:22:34 -080077static inline int kmem_cache_sanity_check(struct mem_cgroup *memcg,
78 const char *name, size_t size)
Shuah Khan77be4b12012-08-16 00:09:46 -070079{
80 return 0;
81}
82#endif
83
Glauber Costa55007d82012-12-18 14:22:38 -080084#ifdef CONFIG_MEMCG_KMEM
85int memcg_update_all_caches(int num_memcgs)
86{
87 struct kmem_cache *s;
88 int ret = 0;
89 mutex_lock(&slab_mutex);
90
91 list_for_each_entry(s, &slab_caches, list) {
92 if (!is_root_cache(s))
93 continue;
94
95 ret = memcg_update_cache_size(s, num_memcgs);
96 /*
97 * See comment in memcontrol.c, memcg_update_cache_size:
98 * Instead of freeing the memory, we'll just leave the caches
99 * up to this point in an updated state.
100 */
101 if (ret)
102 goto out;
103 }
104
105 memcg_update_array_size(num_memcgs);
106out:
107 mutex_unlock(&slab_mutex);
108 return ret;
109}
110#endif
111
Christoph Lameter039363f2012-07-06 15:25:10 -0500112/*
Christoph Lameter45906852012-11-28 16:23:16 +0000113 * Figure out what the alignment of the objects will be given a set of
114 * flags, a user specified alignment and the size of the objects.
115 */
116unsigned long calculate_alignment(unsigned long flags,
117 unsigned long align, unsigned long size)
118{
119 /*
120 * If the user wants hardware cache aligned objects then follow that
121 * suggestion if the object is sufficiently large.
122 *
123 * The hardware cache alignment cannot override the specified
124 * alignment though. If that is greater then use it.
125 */
126 if (flags & SLAB_HWCACHE_ALIGN) {
127 unsigned long ralign = cache_line_size();
128 while (size <= ralign / 2)
129 ralign /= 2;
130 align = max(align, ralign);
131 }
132
133 if (align < ARCH_SLAB_MINALIGN)
134 align = ARCH_SLAB_MINALIGN;
135
136 return ALIGN(align, sizeof(void *));
137}
138
139
140/*
Christoph Lameter039363f2012-07-06 15:25:10 -0500141 * kmem_cache_create - Create a cache.
142 * @name: A string which is used in /proc/slabinfo to identify this cache.
143 * @size: The size of objects to be created in this cache.
144 * @align: The required alignment for the objects.
145 * @flags: SLAB flags
146 * @ctor: A constructor for the objects.
147 *
148 * Returns a ptr to the cache on success, NULL on failure.
149 * Cannot be called within a interrupt, but can be interrupted.
150 * The @ctor is run when new pages are allocated by the cache.
151 *
152 * The flags are
153 *
154 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
155 * to catch references to uninitialised memory.
156 *
157 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
158 * for buffer overruns.
159 *
160 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
161 * cacheline. This can be beneficial if you're counting cycles as closely
162 * as davem.
163 */
164
Glauber Costa2633d7a2012-12-18 14:22:34 -0800165struct kmem_cache *
166kmem_cache_create_memcg(struct mem_cgroup *memcg, const char *name, size_t size,
Glauber Costa943a4512012-12-18 14:23:03 -0800167 size_t align, unsigned long flags, void (*ctor)(void *),
168 struct kmem_cache *parent_cache)
Christoph Lameter039363f2012-07-06 15:25:10 -0500169{
170 struct kmem_cache *s = NULL;
Christoph Lameter686d5502012-09-05 00:20:33 +0000171 int err = 0;
Christoph Lameter039363f2012-07-06 15:25:10 -0500172
Pekka Enbergb9205362012-08-16 10:12:18 +0300173 get_online_cpus();
174 mutex_lock(&slab_mutex);
Christoph Lameter686d5502012-09-05 00:20:33 +0000175
Glauber Costa2633d7a2012-12-18 14:22:34 -0800176 if (!kmem_cache_sanity_check(memcg, name, size) == 0)
Christoph Lameter686d5502012-09-05 00:20:33 +0000177 goto out_locked;
178
Glauber Costad8843922012-10-17 15:36:51 +0400179 /*
180 * Some allocators will constraint the set of valid flags to a subset
181 * of all flags. We expect them to define CACHE_CREATE_MASK in this
182 * case, and we'll just provide them with a sanitized version of the
183 * passed flags.
184 */
185 flags &= CACHE_CREATE_MASK;
Christoph Lameter686d5502012-09-05 00:20:33 +0000186
Glauber Costa2633d7a2012-12-18 14:22:34 -0800187 s = __kmem_cache_alias(memcg, name, size, align, flags, ctor);
Christoph Lametercbb79692012-09-05 00:18:32 +0000188 if (s)
189 goto out_locked;
190
Christoph Lameter278b1bb2012-09-05 00:20:34 +0000191 s = kmem_cache_zalloc(kmem_cache, GFP_KERNEL);
Christoph Lameterdb265ec2012-09-04 23:18:33 +0000192 if (s) {
Christoph Lameter8a13a4c2012-09-04 23:18:33 +0000193 s->object_size = s->size = size;
Christoph Lameter45906852012-11-28 16:23:16 +0000194 s->align = calculate_alignment(flags, align, size);
Christoph Lameter8a13a4c2012-09-04 23:18:33 +0000195 s->ctor = ctor;
Glauber Costa2633d7a2012-12-18 14:22:34 -0800196
Glauber Costa943a4512012-12-18 14:23:03 -0800197 if (memcg_register_cache(memcg, s, parent_cache)) {
Glauber Costa2633d7a2012-12-18 14:22:34 -0800198 kmem_cache_free(kmem_cache, s);
199 err = -ENOMEM;
200 goto out_locked;
201 }
202
Christoph Lameter8a13a4c2012-09-04 23:18:33 +0000203 s->name = kstrdup(name, GFP_KERNEL);
204 if (!s->name) {
205 kmem_cache_free(kmem_cache, s);
206 err = -ENOMEM;
207 goto out_locked;
208 }
209
210 err = __kmem_cache_create(s, flags);
Christoph Lametercce89f42012-09-04 23:38:33 +0000211 if (!err) {
Christoph Lametercce89f42012-09-04 23:38:33 +0000212 s->refcount = 1;
Christoph Lameterdb265ec2012-09-04 23:18:33 +0000213 list_add(&s->list, &slab_caches);
Glauber Costa2633d7a2012-12-18 14:22:34 -0800214 memcg_cache_list_add(memcg, s);
Christoph Lametercce89f42012-09-04 23:38:33 +0000215 } else {
Christoph Lameter8a13a4c2012-09-04 23:18:33 +0000216 kfree(s->name);
Christoph Lameter278b1bb2012-09-05 00:20:34 +0000217 kmem_cache_free(kmem_cache, s);
218 }
Christoph Lameter8a13a4c2012-09-04 23:18:33 +0000219 } else
Christoph Lameter278b1bb2012-09-05 00:20:34 +0000220 err = -ENOMEM;
Christoph Lameter7c9adf52012-09-04 23:38:33 +0000221
Christoph Lameter686d5502012-09-05 00:20:33 +0000222out_locked:
Christoph Lameter20cea962012-07-06 15:25:13 -0500223 mutex_unlock(&slab_mutex);
224 put_online_cpus();
225
Christoph Lameter686d5502012-09-05 00:20:33 +0000226 if (err) {
227
228 if (flags & SLAB_PANIC)
229 panic("kmem_cache_create: Failed to create slab '%s'. Error %d\n",
230 name, err);
231 else {
232 printk(KERN_WARNING "kmem_cache_create(%s) failed with error %d",
233 name, err);
234 dump_stack();
235 }
236
237 return NULL;
238 }
Christoph Lameter039363f2012-07-06 15:25:10 -0500239
240 return s;
241}
Glauber Costa2633d7a2012-12-18 14:22:34 -0800242
243struct kmem_cache *
244kmem_cache_create(const char *name, size_t size, size_t align,
245 unsigned long flags, void (*ctor)(void *))
246{
Glauber Costa943a4512012-12-18 14:23:03 -0800247 return kmem_cache_create_memcg(NULL, name, size, align, flags, ctor, NULL);
Glauber Costa2633d7a2012-12-18 14:22:34 -0800248}
Christoph Lameter039363f2012-07-06 15:25:10 -0500249EXPORT_SYMBOL(kmem_cache_create);
Christoph Lameter97d06602012-07-06 15:25:11 -0500250
Christoph Lameter945cf2b2012-09-04 23:18:33 +0000251void kmem_cache_destroy(struct kmem_cache *s)
252{
Glauber Costa7cf27982012-12-18 14:22:55 -0800253 /* Destroy all the children caches if we aren't a memcg cache */
254 kmem_cache_destroy_memcg_children(s);
255
Christoph Lameter945cf2b2012-09-04 23:18:33 +0000256 get_online_cpus();
257 mutex_lock(&slab_mutex);
258 s->refcount--;
259 if (!s->refcount) {
260 list_del(&s->list);
261
262 if (!__kmem_cache_shutdown(s)) {
Jiri Kosina210ed9d2012-10-08 09:26:01 +0200263 mutex_unlock(&slab_mutex);
Christoph Lameter945cf2b2012-09-04 23:18:33 +0000264 if (s->flags & SLAB_DESTROY_BY_RCU)
265 rcu_barrier();
266
Glauber Costa2633d7a2012-12-18 14:22:34 -0800267 memcg_release_cache(s);
Christoph Lameterdb265ec2012-09-04 23:18:33 +0000268 kfree(s->name);
Christoph Lameter8f4c7652012-09-05 00:18:32 +0000269 kmem_cache_free(kmem_cache, s);
Christoph Lameter945cf2b2012-09-04 23:18:33 +0000270 } else {
271 list_add(&s->list, &slab_caches);
Jiri Kosina210ed9d2012-10-08 09:26:01 +0200272 mutex_unlock(&slab_mutex);
Christoph Lameter945cf2b2012-09-04 23:18:33 +0000273 printk(KERN_ERR "kmem_cache_destroy %s: Slab cache still has objects\n",
274 s->name);
275 dump_stack();
276 }
Jiri Kosina210ed9d2012-10-08 09:26:01 +0200277 } else {
278 mutex_unlock(&slab_mutex);
Christoph Lameter945cf2b2012-09-04 23:18:33 +0000279 }
Christoph Lameter945cf2b2012-09-04 23:18:33 +0000280 put_online_cpus();
281}
282EXPORT_SYMBOL(kmem_cache_destroy);
283
Christoph Lameter97d06602012-07-06 15:25:11 -0500284int slab_is_available(void)
285{
286 return slab_state >= UP;
287}
Glauber Costab7454ad2012-10-19 18:20:25 +0400288
Christoph Lameter45530c42012-11-28 16:23:07 +0000289#ifndef CONFIG_SLOB
290/* Create a cache during boot when no slab services are available yet */
291void __init create_boot_cache(struct kmem_cache *s, const char *name, size_t size,
292 unsigned long flags)
293{
294 int err;
295
296 s->name = name;
297 s->size = s->object_size = size;
Christoph Lameter45906852012-11-28 16:23:16 +0000298 s->align = calculate_alignment(flags, ARCH_KMALLOC_MINALIGN, size);
Christoph Lameter45530c42012-11-28 16:23:07 +0000299 err = __kmem_cache_create(s, flags);
300
301 if (err)
Christoph Lameter31ba7342013-01-10 19:00:53 +0000302 panic("Creation of kmalloc slab %s size=%zu failed. Reason %d\n",
Christoph Lameter45530c42012-11-28 16:23:07 +0000303 name, size, err);
304
305 s->refcount = -1; /* Exempt from merging for now */
306}
307
308struct kmem_cache *__init create_kmalloc_cache(const char *name, size_t size,
309 unsigned long flags)
310{
311 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
312
313 if (!s)
314 panic("Out of memory when creating slab %s\n", name);
315
316 create_boot_cache(s, name, size, flags);
317 list_add(&s->list, &slab_caches);
318 s->refcount = 1;
319 return s;
320}
321
Christoph Lameter9425c582013-01-10 19:12:17 +0000322struct kmem_cache *kmalloc_caches[KMALLOC_SHIFT_HIGH + 1];
323EXPORT_SYMBOL(kmalloc_caches);
324
325#ifdef CONFIG_ZONE_DMA
326struct kmem_cache *kmalloc_dma_caches[KMALLOC_SHIFT_HIGH + 1];
327EXPORT_SYMBOL(kmalloc_dma_caches);
328#endif
329
Christoph Lameterf97d5f632013-01-10 19:12:17 +0000330/*
331 * Create the kmalloc array. Some of the regular kmalloc arrays
332 * may already have been created because they were needed to
333 * enable allocations for slab creation.
334 */
335void __init create_kmalloc_caches(unsigned long flags)
336{
337 int i;
338
339 /* Caches that are not of the two-to-the-power-of size */
340 if (KMALLOC_MIN_SIZE <= 32 && !kmalloc_caches[1])
341 kmalloc_caches[1] = create_kmalloc_cache(NULL, 96, flags);
342
343 if (KMALLOC_MIN_SIZE <= 64 && !kmalloc_caches[2])
344 kmalloc_caches[2] = create_kmalloc_cache(NULL, 192, flags);
345
346 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++)
347 if (!kmalloc_caches[i])
348 kmalloc_caches[i] = create_kmalloc_cache(NULL,
349 1 << i, flags);
350
351 /* Kmalloc array is now usable */
352 slab_state = UP;
353
354 for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
355 struct kmem_cache *s = kmalloc_caches[i];
356 char *n;
357
358 if (s) {
359 n = kasprintf(GFP_NOWAIT, "kmalloc-%d", kmalloc_size(i));
360
361 BUG_ON(!n);
362 s->name = n;
363 }
364 }
365
366#ifdef CONFIG_ZONE_DMA
367 for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
368 struct kmem_cache *s = kmalloc_caches[i];
369
370 if (s) {
371 int size = kmalloc_size(i);
372 char *n = kasprintf(GFP_NOWAIT,
373 "dma-kmalloc-%d", size);
374
375 BUG_ON(!n);
376 kmalloc_dma_caches[i] = create_kmalloc_cache(n,
377 size, SLAB_CACHE_DMA | flags);
378 }
379 }
380#endif
381}
382
383
Christoph Lameter45530c42012-11-28 16:23:07 +0000384#endif /* !CONFIG_SLOB */
385
386
Glauber Costab7454ad2012-10-19 18:20:25 +0400387#ifdef CONFIG_SLABINFO
Glauber Costa749c5412012-12-18 14:23:01 -0800388void print_slabinfo_header(struct seq_file *m)
Glauber Costabcee6e22012-10-19 18:20:26 +0400389{
390 /*
391 * Output format version, so at least we can change it
392 * without _too_ many complaints.
393 */
394#ifdef CONFIG_DEBUG_SLAB
395 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
396#else
397 seq_puts(m, "slabinfo - version: 2.1\n");
398#endif
399 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
400 "<objperslab> <pagesperslab>");
401 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
402 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
403#ifdef CONFIG_DEBUG_SLAB
404 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
405 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
406 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
407#endif
408 seq_putc(m, '\n');
409}
410
Glauber Costab7454ad2012-10-19 18:20:25 +0400411static void *s_start(struct seq_file *m, loff_t *pos)
412{
413 loff_t n = *pos;
414
415 mutex_lock(&slab_mutex);
416 if (!n)
417 print_slabinfo_header(m);
418
419 return seq_list_start(&slab_caches, *pos);
420}
421
422static void *s_next(struct seq_file *m, void *p, loff_t *pos)
423{
424 return seq_list_next(p, &slab_caches, pos);
425}
426
427static void s_stop(struct seq_file *m, void *p)
428{
429 mutex_unlock(&slab_mutex);
430}
431
Glauber Costa749c5412012-12-18 14:23:01 -0800432static void
433memcg_accumulate_slabinfo(struct kmem_cache *s, struct slabinfo *info)
Glauber Costab7454ad2012-10-19 18:20:25 +0400434{
Glauber Costa749c5412012-12-18 14:23:01 -0800435 struct kmem_cache *c;
436 struct slabinfo sinfo;
437 int i;
438
439 if (!is_root_cache(s))
440 return;
441
442 for_each_memcg_cache_index(i) {
443 c = cache_from_memcg(s, i);
444 if (!c)
445 continue;
446
447 memset(&sinfo, 0, sizeof(sinfo));
448 get_slabinfo(c, &sinfo);
449
450 info->active_slabs += sinfo.active_slabs;
451 info->num_slabs += sinfo.num_slabs;
452 info->shared_avail += sinfo.shared_avail;
453 info->active_objs += sinfo.active_objs;
454 info->num_objs += sinfo.num_objs;
455 }
456}
457
458int cache_show(struct kmem_cache *s, struct seq_file *m)
459{
Glauber Costa0d7561c2012-10-19 18:20:27 +0400460 struct slabinfo sinfo;
461
462 memset(&sinfo, 0, sizeof(sinfo));
463 get_slabinfo(s, &sinfo);
464
Glauber Costa749c5412012-12-18 14:23:01 -0800465 memcg_accumulate_slabinfo(s, &sinfo);
466
Glauber Costa0d7561c2012-10-19 18:20:27 +0400467 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Glauber Costa749c5412012-12-18 14:23:01 -0800468 cache_name(s), sinfo.active_objs, sinfo.num_objs, s->size,
Glauber Costa0d7561c2012-10-19 18:20:27 +0400469 sinfo.objects_per_slab, (1 << sinfo.cache_order));
470
471 seq_printf(m, " : tunables %4u %4u %4u",
472 sinfo.limit, sinfo.batchcount, sinfo.shared);
473 seq_printf(m, " : slabdata %6lu %6lu %6lu",
474 sinfo.active_slabs, sinfo.num_slabs, sinfo.shared_avail);
475 slabinfo_show_stats(m, s);
476 seq_putc(m, '\n');
477 return 0;
Glauber Costab7454ad2012-10-19 18:20:25 +0400478}
479
Glauber Costa749c5412012-12-18 14:23:01 -0800480static int s_show(struct seq_file *m, void *p)
481{
482 struct kmem_cache *s = list_entry(p, struct kmem_cache, list);
483
484 if (!is_root_cache(s))
485 return 0;
486 return cache_show(s, m);
487}
488
Glauber Costab7454ad2012-10-19 18:20:25 +0400489/*
490 * slabinfo_op - iterator that generates /proc/slabinfo
491 *
492 * Output layout:
493 * cache-name
494 * num-active-objs
495 * total-objs
496 * object size
497 * num-active-slabs
498 * total-slabs
499 * num-pages-per-slab
500 * + further values on SMP and with statistics enabled
501 */
502static const struct seq_operations slabinfo_op = {
503 .start = s_start,
504 .next = s_next,
505 .stop = s_stop,
506 .show = s_show,
507};
508
509static int slabinfo_open(struct inode *inode, struct file *file)
510{
511 return seq_open(file, &slabinfo_op);
512}
513
514static const struct file_operations proc_slabinfo_operations = {
515 .open = slabinfo_open,
516 .read = seq_read,
517 .write = slabinfo_write,
518 .llseek = seq_lseek,
519 .release = seq_release,
520};
521
522static int __init slab_proc_init(void)
523{
524 proc_create("slabinfo", S_IRUSR, NULL, &proc_slabinfo_operations);
525 return 0;
526}
527module_init(slab_proc_init);
528#endif /* CONFIG_SLABINFO */