Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Slab allocator functions that are independent of the allocator strategy
4 *
5 * (C) 2012 Christoph Lameter <cl@linux.com>
6 */
7#include <linux/slab.h>
8
9#include <linux/mm.h>
10#include <linux/poison.h>
11#include <linux/interrupt.h>
12#include <linux/memory.h>
13#include <linux/cache.h>
14#include <linux/compiler.h>
15#include <linux/module.h>
16#include <linux/cpu.h>
17#include <linux/uaccess.h>
18#include <linux/seq_file.h>
19#include <linux/proc_fs.h>
20#include <linux/debugfs.h>
21#include <asm/cacheflush.h>
22#include <asm/tlbflush.h>
23#include <asm/page.h>
24#include <linux/memcontrol.h>
25
26#define CREATE_TRACE_POINTS
27#include <trace/events/kmem.h>
28
29#include "internal.h"
30
31#include "slab.h"
32
33enum slab_state slab_state;
34LIST_HEAD(slab_caches);
35DEFINE_MUTEX(slab_mutex);
36struct kmem_cache *kmem_cache;
37
38#ifdef CONFIG_HARDENED_USERCOPY
39bool usercopy_fallback __ro_after_init =
40 IS_ENABLED(CONFIG_HARDENED_USERCOPY_FALLBACK);
41module_param(usercopy_fallback, bool, 0400);
42MODULE_PARM_DESC(usercopy_fallback,
43 "WARN instead of reject usercopy whitelist violations");
44#endif
45
46static LIST_HEAD(slab_caches_to_rcu_destroy);
47static void slab_caches_to_rcu_destroy_workfn(struct work_struct *work);
48static DECLARE_WORK(slab_caches_to_rcu_destroy_work,
49 slab_caches_to_rcu_destroy_workfn);
50
51/*
52 * Set of flags that will prevent slab merging
53 */
54#define SLAB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
55 SLAB_TRACE | SLAB_TYPESAFE_BY_RCU | SLAB_NOLEAKTRACE | \
56 SLAB_FAILSLAB | SLAB_KASAN)
57
58#define SLAB_MERGE_SAME (SLAB_RECLAIM_ACCOUNT | SLAB_CACHE_DMA | \
59 SLAB_CACHE_DMA32 | SLAB_ACCOUNT)
60
61/*
62 * Merge control. If this is set then no merging of slab caches will occur.
63 */
64static bool slab_nomerge = !IS_ENABLED(CONFIG_SLAB_MERGE_DEFAULT);
65
66static int __init setup_slab_nomerge(char *str)
67{
68 slab_nomerge = true;
69 return 1;
70}
71
72#ifdef CONFIG_SLUB
73__setup_param("slub_nomerge", slub_nomerge, setup_slab_nomerge, 0);
74#endif
75
76__setup("slab_nomerge", setup_slab_nomerge);
77
78/*
79 * Determine the size of a slab object
80 */
81unsigned int kmem_cache_size(struct kmem_cache *s)
82{
83 return s->object_size;
84}
85EXPORT_SYMBOL(kmem_cache_size);
86
87#ifdef CONFIG_DEBUG_VM
88static int kmem_cache_sanity_check(const char *name, unsigned int size)
89{
90 if (!name || in_interrupt() || size < sizeof(void *) ||
91 size > KMALLOC_MAX_SIZE) {
92 pr_err("kmem_cache_create(%s) integrity check failed\n", name);
93 return -EINVAL;
94 }
95
96 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
97 return 0;
98}
99#else
100static inline int kmem_cache_sanity_check(const char *name, unsigned int size)
101{
102 return 0;
103}
104#endif
105
106void __kmem_cache_free_bulk(struct kmem_cache *s, size_t nr, void **p)
107{
108 size_t i;
109
110 for (i = 0; i < nr; i++) {
111 if (s)
112 kmem_cache_free(s, p[i]);
113 else
114 kfree(p[i]);
115 }
116}
117
118int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t nr,
119 void **p)
120{
121 size_t i;
122
123 for (i = 0; i < nr; i++) {
124 void *x = p[i] = kmem_cache_alloc(s, flags);
125 if (!x) {
126 __kmem_cache_free_bulk(s, i, p);
127 return 0;
128 }
129 }
130 return i;
131}
132
133/*
134 * Figure out what the alignment of the objects will be given a set of
135 * flags, a user specified alignment and the size of the objects.
136 */
137static unsigned int calculate_alignment(slab_flags_t flags,
138 unsigned int align, unsigned int size)
139{
140 /*
141 * If the user wants hardware cache aligned objects then follow that
142 * suggestion if the object is sufficiently large.
143 *
144 * The hardware cache alignment cannot override the specified
145 * alignment though. If that is greater then use it.
146 */
147 if (flags & SLAB_HWCACHE_ALIGN) {
148 unsigned int ralign;
149
150 ralign = cache_line_size();
151 while (size <= ralign / 2)
152 ralign /= 2;
153 align = max(align, ralign);
154 }
155
156 if (align < ARCH_SLAB_MINALIGN)
157 align = ARCH_SLAB_MINALIGN;
158
159 return ALIGN(align, sizeof(void *));
160}
161
162/*
163 * Find a mergeable slab cache
164 */
165int slab_unmergeable(struct kmem_cache *s)
166{
167 if (slab_nomerge || (s->flags & SLAB_NEVER_MERGE))
168 return 1;
169
170 if (s->ctor)
171 return 1;
172
173 if (s->usersize)
174 return 1;
175
176 /*
177 * We may have set a slab to be unmergeable during bootstrap.
178 */
179 if (s->refcount < 0)
180 return 1;
181
182 return 0;
183}
184
185struct kmem_cache *find_mergeable(unsigned int size, unsigned int align,
186 slab_flags_t flags, const char *name, void (*ctor)(void *))
187{
188 struct kmem_cache *s;
189
190 if (slab_nomerge)
191 return NULL;
192
193 if (ctor)
194 return NULL;
195
196 size = ALIGN(size, sizeof(void *));
197 align = calculate_alignment(flags, align, size);
198 size = ALIGN(size, align);
199 flags = kmem_cache_flags(size, flags, name, NULL);
200
201 if (flags & SLAB_NEVER_MERGE)
202 return NULL;
203
204 list_for_each_entry_reverse(s, &slab_caches, list) {
205 if (slab_unmergeable(s))
206 continue;
207
208 if (size > s->size)
209 continue;
210
211 if ((flags & SLAB_MERGE_SAME) != (s->flags & SLAB_MERGE_SAME))
212 continue;
213 /*
214 * Check if alignment is compatible.
215 * Courtesy of Adrian Drzewiecki
216 */
217 if ((s->size & ~(align - 1)) != s->size)
218 continue;
219
220 if (s->size - size >= sizeof(void *))
221 continue;
222
223 if (IS_ENABLED(CONFIG_SLAB) && align &&
224 (align > s->align || s->align % align))
225 continue;
226
227 return s;
228 }
229 return NULL;
230}
231
232static struct kmem_cache *create_cache(const char *name,
233 unsigned int object_size, unsigned int align,
234 slab_flags_t flags, unsigned int useroffset,
235 unsigned int usersize, void (*ctor)(void *),
236 struct kmem_cache *root_cache)
237{
238 struct kmem_cache *s;
239 int err;
240
241 if (WARN_ON(useroffset + usersize > object_size))
242 useroffset = usersize = 0;
243
244 err = -ENOMEM;
245 s = kmem_cache_zalloc(kmem_cache, GFP_KERNEL);
246 if (!s)
247 goto out;
248
249 s->name = name;
250 s->size = s->object_size = object_size;
251 s->align = align;
252 s->ctor = ctor;
253 s->useroffset = useroffset;
254 s->usersize = usersize;
255
256 err = __kmem_cache_create(s, flags);
257 if (err)
258 goto out_free_cache;
259
260 s->refcount = 1;
261 list_add(&s->list, &slab_caches);
262out:
263 if (err)
264 return ERR_PTR(err);
265 return s;
266
267out_free_cache:
268 kmem_cache_free(kmem_cache, s);
269 goto out;
270}
271
272/**
273 * kmem_cache_create_usercopy - Create a cache with a region suitable
274 * for copying to userspace
275 * @name: A string which is used in /proc/slabinfo to identify this cache.
276 * @size: The size of objects to be created in this cache.
277 * @align: The required alignment for the objects.
278 * @flags: SLAB flags
279 * @useroffset: Usercopy region offset
280 * @usersize: Usercopy region size
281 * @ctor: A constructor for the objects.
282 *
283 * Cannot be called within a interrupt, but can be interrupted.
284 * The @ctor is run when new pages are allocated by the cache.
285 *
286 * The flags are
287 *
288 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
289 * to catch references to uninitialised memory.
290 *
291 * %SLAB_RED_ZONE - Insert `Red` zones around the allocated memory to check
292 * for buffer overruns.
293 *
294 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
295 * cacheline. This can be beneficial if you're counting cycles as closely
296 * as davem.
297 *
298 * Return: a pointer to the cache on success, NULL on failure.
299 */
300struct kmem_cache *
301kmem_cache_create_usercopy(const char *name,
302 unsigned int size, unsigned int align,
303 slab_flags_t flags,
304 unsigned int useroffset, unsigned int usersize,
305 void (*ctor)(void *))
306{
307 struct kmem_cache *s = NULL;
308 const char *cache_name;
309 int err;
310
311 get_online_cpus();
312 get_online_mems();
313
314 mutex_lock(&slab_mutex);
315
316 err = kmem_cache_sanity_check(name, size);
317 if (err) {
318 goto out_unlock;
319 }
320
321 /* Refuse requests with allocator specific flags */
322 if (flags & ~SLAB_FLAGS_PERMITTED) {
323 err = -EINVAL;
324 goto out_unlock;
325 }
326
327 /*
328 * Some allocators will constraint the set of valid flags to a subset
329 * of all flags. We expect them to define CACHE_CREATE_MASK in this
330 * case, and we'll just provide them with a sanitized version of the
331 * passed flags.
332 */
333 flags &= CACHE_CREATE_MASK;
334
335 /* Fail closed on bad usersize of useroffset values. */
336 if (WARN_ON(!usersize && useroffset) ||
337 WARN_ON(size < usersize || size - usersize < useroffset))
338 usersize = useroffset = 0;
339
340 if (!usersize)
341 s = __kmem_cache_alias(name, size, align, flags, ctor);
342 if (s)
343 goto out_unlock;
344
345 cache_name = kstrdup_const(name, GFP_KERNEL);
346 if (!cache_name) {
347 err = -ENOMEM;
348 goto out_unlock;
349 }
350
351 s = create_cache(cache_name, size,
352 calculate_alignment(flags, align, size),
353 flags, useroffset, usersize, ctor, NULL);
354 if (IS_ERR(s)) {
355 err = PTR_ERR(s);
356 kfree_const(cache_name);
357 }
358
359out_unlock:
360 mutex_unlock(&slab_mutex);
361
362 put_online_mems();
363 put_online_cpus();
364
365 if (err) {
366 if (flags & SLAB_PANIC)
367 panic("kmem_cache_create: Failed to create slab '%s'. Error %d\n",
368 name, err);
369 else {
370 pr_warn("kmem_cache_create(%s) failed with error %d\n",
371 name, err);
372 dump_stack();
373 }
374 return NULL;
375 }
376 return s;
377}
378EXPORT_SYMBOL(kmem_cache_create_usercopy);
379
380/**
381 * kmem_cache_create - Create a cache.
382 * @name: A string which is used in /proc/slabinfo to identify this cache.
383 * @size: The size of objects to be created in this cache.
384 * @align: The required alignment for the objects.
385 * @flags: SLAB flags
386 * @ctor: A constructor for the objects.
387 *
388 * Cannot be called within a interrupt, but can be interrupted.
389 * The @ctor is run when new pages are allocated by the cache.
390 *
391 * The flags are
392 *
393 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
394 * to catch references to uninitialised memory.
395 *
396 * %SLAB_RED_ZONE - Insert `Red` zones around the allocated memory to check
397 * for buffer overruns.
398 *
399 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
400 * cacheline. This can be beneficial if you're counting cycles as closely
401 * as davem.
402 *
403 * Return: a pointer to the cache on success, NULL on failure.
404 */
405struct kmem_cache *
406kmem_cache_create(const char *name, unsigned int size, unsigned int align,
407 slab_flags_t flags, void (*ctor)(void *))
408{
409 return kmem_cache_create_usercopy(name, size, align, flags, 0, 0,
410 ctor);
411}
412EXPORT_SYMBOL(kmem_cache_create);
413
414static void slab_caches_to_rcu_destroy_workfn(struct work_struct *work)
415{
416 LIST_HEAD(to_destroy);
417 struct kmem_cache *s, *s2;
418
419 /*
420 * On destruction, SLAB_TYPESAFE_BY_RCU kmem_caches are put on the
421 * @slab_caches_to_rcu_destroy list. The slab pages are freed
422 * through RCU and the associated kmem_cache are dereferenced
423 * while freeing the pages, so the kmem_caches should be freed only
424 * after the pending RCU operations are finished. As rcu_barrier()
425 * is a pretty slow operation, we batch all pending destructions
426 * asynchronously.
427 */
428 mutex_lock(&slab_mutex);
429 list_splice_init(&slab_caches_to_rcu_destroy, &to_destroy);
430 mutex_unlock(&slab_mutex);
431
432 if (list_empty(&to_destroy))
433 return;
434
435 rcu_barrier();
436
437 list_for_each_entry_safe(s, s2, &to_destroy, list) {
438#ifdef SLAB_SUPPORTS_SYSFS
439 sysfs_slab_release(s);
440#else
441 slab_kmem_cache_release(s);
442#endif
443 }
444}
445
446static int shutdown_cache(struct kmem_cache *s)
447{
448 /* free asan quarantined objects */
449 kasan_cache_shutdown(s);
450
451 if (__kmem_cache_shutdown(s) != 0)
452 return -EBUSY;
453
454 list_del(&s->list);
455
456 if (s->flags & SLAB_TYPESAFE_BY_RCU) {
457#ifdef SLAB_SUPPORTS_SYSFS
458 sysfs_slab_unlink(s);
459#endif
460 list_add_tail(&s->list, &slab_caches_to_rcu_destroy);
461 schedule_work(&slab_caches_to_rcu_destroy_work);
462 } else {
463#ifdef SLAB_SUPPORTS_SYSFS
464 sysfs_slab_unlink(s);
465 sysfs_slab_release(s);
466#else
467 slab_kmem_cache_release(s);
468#endif
469 }
470
471 return 0;
472}
473
474void slab_kmem_cache_release(struct kmem_cache *s)
475{
476 __kmem_cache_release(s);
477 kfree_const(s->name);
478 kmem_cache_free(kmem_cache, s);
479}
480
481void kmem_cache_destroy(struct kmem_cache *s)
482{
483 int err;
484
485 if (unlikely(!s))
486 return;
487
488 get_online_cpus();
489 get_online_mems();
490
491 mutex_lock(&slab_mutex);
492
493 s->refcount--;
494 if (s->refcount)
495 goto out_unlock;
496
497 err = shutdown_cache(s);
498 if (err) {
499 pr_err("kmem_cache_destroy %s: Slab cache still has objects\n",
500 s->name);
501 dump_stack();
502 }
503out_unlock:
504 mutex_unlock(&slab_mutex);
505
506 put_online_mems();
507 put_online_cpus();
508}
509EXPORT_SYMBOL(kmem_cache_destroy);
510
511/**
512 * kmem_cache_shrink - Shrink a cache.
513 * @cachep: The cache to shrink.
514 *
515 * Releases as many slabs as possible for a cache.
516 * To help debugging, a zero exit status indicates all slabs were released.
517 *
518 * Return: %0 if all slabs were released, non-zero otherwise
519 */
520int kmem_cache_shrink(struct kmem_cache *cachep)
521{
522 int ret;
523
524 get_online_cpus();
525 get_online_mems();
526 kasan_cache_shrink(cachep);
527 ret = __kmem_cache_shrink(cachep);
528 put_online_mems();
529 put_online_cpus();
530 return ret;
531}
532EXPORT_SYMBOL(kmem_cache_shrink);
533
534bool slab_is_available(void)
535{
536 return slab_state >= UP;
537}
538
539#ifndef CONFIG_SLOB
540/* Create a cache during boot when no slab services are available yet */
541void __init create_boot_cache(struct kmem_cache *s, const char *name,
542 unsigned int size, slab_flags_t flags,
543 unsigned int useroffset, unsigned int usersize)
544{
545 int err;
546 unsigned int align = ARCH_KMALLOC_MINALIGN;
547
548 s->name = name;
549 s->size = s->object_size = size;
550
551 /*
552 * For power of two sizes, guarantee natural alignment for kmalloc
553 * caches, regardless of SL*B debugging options.
554 */
555 if (is_power_of_2(size))
556 align = max(align, size);
557 s->align = calculate_alignment(flags, align, size);
558
559 s->useroffset = useroffset;
560 s->usersize = usersize;
561
562 err = __kmem_cache_create(s, flags);
563
564 if (err)
565 panic("Creation of kmalloc slab %s size=%u failed. Reason %d\n",
566 name, size, err);
567
568 s->refcount = -1; /* Exempt from merging for now */
569}
570
571struct kmem_cache *__init create_kmalloc_cache(const char *name,
572 unsigned int size, slab_flags_t flags,
573 unsigned int useroffset, unsigned int usersize)
574{
575 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
576
577 if (!s)
578 panic("Out of memory when creating slab %s\n", name);
579
580 create_boot_cache(s, name, size, flags, useroffset, usersize);
581 list_add(&s->list, &slab_caches);
582 s->refcount = 1;
583 return s;
584}
585
586struct kmem_cache *
587kmalloc_caches[NR_KMALLOC_TYPES][KMALLOC_SHIFT_HIGH + 1] __ro_after_init =
588{ /* initialization for https://bugs.llvm.org/show_bug.cgi?id=42570 */ };
589EXPORT_SYMBOL(kmalloc_caches);
590
591/*
592 * Conversion table for small slabs sizes / 8 to the index in the
593 * kmalloc array. This is necessary for slabs < 192 since we have non power
594 * of two cache sizes there. The size of larger slabs can be determined using
595 * fls.
596 */
597static u8 size_index[24] __ro_after_init = {
598 3, /* 8 */
599 4, /* 16 */
600 5, /* 24 */
601 5, /* 32 */
602 6, /* 40 */
603 6, /* 48 */
604 6, /* 56 */
605 6, /* 64 */
606 1, /* 72 */
607 1, /* 80 */
608 1, /* 88 */
609 1, /* 96 */
610 7, /* 104 */
611 7, /* 112 */
612 7, /* 120 */
613 7, /* 128 */
614 2, /* 136 */
615 2, /* 144 */
616 2, /* 152 */
617 2, /* 160 */
618 2, /* 168 */
619 2, /* 176 */
620 2, /* 184 */
621 2 /* 192 */
622};
623
624static inline unsigned int size_index_elem(unsigned int bytes)
625{
626 return (bytes - 1) / 8;
627}
628
629/*
630 * Find the kmem_cache structure that serves a given size of
631 * allocation
632 */
633struct kmem_cache *kmalloc_slab(size_t size, gfp_t flags)
634{
635 unsigned int index;
636
637 if (size <= 192) {
638 if (!size)
639 return ZERO_SIZE_PTR;
640
641 index = size_index[size_index_elem(size)];
642 } else {
643 if (WARN_ON_ONCE(size > KMALLOC_MAX_CACHE_SIZE))
644 return NULL;
645 index = fls(size - 1);
646 }
647
648 return kmalloc_caches[kmalloc_type(flags)][index];
649}
650
651#ifdef CONFIG_ZONE_DMA
652#define INIT_KMALLOC_INFO(__size, __short_size) \
653{ \
654 .name[KMALLOC_NORMAL] = "kmalloc-" #__short_size, \
655 .name[KMALLOC_RECLAIM] = "kmalloc-rcl-" #__short_size, \
656 .name[KMALLOC_DMA] = "dma-kmalloc-" #__short_size, \
657 .size = __size, \
658}
659#else
660#define INIT_KMALLOC_INFO(__size, __short_size) \
661{ \
662 .name[KMALLOC_NORMAL] = "kmalloc-" #__short_size, \
663 .name[KMALLOC_RECLAIM] = "kmalloc-rcl-" #__short_size, \
664 .size = __size, \
665}
666#endif
667
668/*
669 * kmalloc_info[] is to make slub_debug=,kmalloc-xx option work at boot time.
670 * kmalloc_index() supports up to 2^26=64MB, so the final entry of the table is
671 * kmalloc-67108864.
672 */
673const struct kmalloc_info_struct kmalloc_info[] __initconst = {
674 INIT_KMALLOC_INFO(0, 0),
675 INIT_KMALLOC_INFO(96, 96),
676 INIT_KMALLOC_INFO(192, 192),
677 INIT_KMALLOC_INFO(8, 8),
678 INIT_KMALLOC_INFO(16, 16),
679 INIT_KMALLOC_INFO(32, 32),
680 INIT_KMALLOC_INFO(64, 64),
681 INIT_KMALLOC_INFO(128, 128),
682 INIT_KMALLOC_INFO(256, 256),
683 INIT_KMALLOC_INFO(512, 512),
684 INIT_KMALLOC_INFO(1024, 1k),
685 INIT_KMALLOC_INFO(2048, 2k),
686 INIT_KMALLOC_INFO(4096, 4k),
687 INIT_KMALLOC_INFO(8192, 8k),
688 INIT_KMALLOC_INFO(16384, 16k),
689 INIT_KMALLOC_INFO(32768, 32k),
690 INIT_KMALLOC_INFO(65536, 64k),
691 INIT_KMALLOC_INFO(131072, 128k),
692 INIT_KMALLOC_INFO(262144, 256k),
693 INIT_KMALLOC_INFO(524288, 512k),
694 INIT_KMALLOC_INFO(1048576, 1M),
695 INIT_KMALLOC_INFO(2097152, 2M),
696 INIT_KMALLOC_INFO(4194304, 4M),
697 INIT_KMALLOC_INFO(8388608, 8M),
698 INIT_KMALLOC_INFO(16777216, 16M),
699 INIT_KMALLOC_INFO(33554432, 32M),
700 INIT_KMALLOC_INFO(67108864, 64M)
701};
702
703/*
704 * Patch up the size_index table if we have strange large alignment
705 * requirements for the kmalloc array. This is only the case for
706 * MIPS it seems. The standard arches will not generate any code here.
707 *
708 * Largest permitted alignment is 256 bytes due to the way we
709 * handle the index determination for the smaller caches.
710 *
711 * Make sure that nothing crazy happens if someone starts tinkering
712 * around with ARCH_KMALLOC_MINALIGN
713 */
714void __init setup_kmalloc_cache_index_table(void)
715{
716 unsigned int i;
717
718 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
719 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
720
721 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
722 unsigned int elem = size_index_elem(i);
723
724 if (elem >= ARRAY_SIZE(size_index))
725 break;
726 size_index[elem] = KMALLOC_SHIFT_LOW;
727 }
728
729 if (KMALLOC_MIN_SIZE >= 64) {
730 /*
731 * The 96 byte size cache is not used if the alignment
732 * is 64 byte.
733 */
734 for (i = 64 + 8; i <= 96; i += 8)
735 size_index[size_index_elem(i)] = 7;
736
737 }
738
739 if (KMALLOC_MIN_SIZE >= 128) {
740 /*
741 * The 192 byte sized cache is not used if the alignment
742 * is 128 byte. Redirect kmalloc to use the 256 byte cache
743 * instead.
744 */
745 for (i = 128 + 8; i <= 192; i += 8)
746 size_index[size_index_elem(i)] = 8;
747 }
748}
749
750static void __init
751new_kmalloc_cache(int idx, enum kmalloc_cache_type type, slab_flags_t flags)
752{
753 if (type == KMALLOC_RECLAIM)
754 flags |= SLAB_RECLAIM_ACCOUNT;
755
756 kmalloc_caches[type][idx] = create_kmalloc_cache(
757 kmalloc_info[idx].name[type],
758 kmalloc_info[idx].size, flags, 0,
759 kmalloc_info[idx].size);
760}
761
762/*
763 * Create the kmalloc array. Some of the regular kmalloc arrays
764 * may already have been created because they were needed to
765 * enable allocations for slab creation.
766 */
767void __init create_kmalloc_caches(slab_flags_t flags)
768{
769 int i;
770 enum kmalloc_cache_type type;
771
772 for (type = KMALLOC_NORMAL; type <= KMALLOC_RECLAIM; type++) {
773 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) {
774 if (!kmalloc_caches[type][i])
775 new_kmalloc_cache(i, type, flags);
776
777 /*
778 * Caches that are not of the two-to-the-power-of size.
779 * These have to be created immediately after the
780 * earlier power of two caches
781 */
782 if (KMALLOC_MIN_SIZE <= 32 && i == 6 &&
783 !kmalloc_caches[type][1])
784 new_kmalloc_cache(1, type, flags);
785 if (KMALLOC_MIN_SIZE <= 64 && i == 7 &&
786 !kmalloc_caches[type][2])
787 new_kmalloc_cache(2, type, flags);
788 }
789 }
790
791 /* Kmalloc array is now usable */
792 slab_state = UP;
793
794#ifdef CONFIG_ZONE_DMA
795 for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
796 struct kmem_cache *s = kmalloc_caches[KMALLOC_NORMAL][i];
797
798 if (s) {
799 kmalloc_caches[KMALLOC_DMA][i] = create_kmalloc_cache(
800 kmalloc_info[i].name[KMALLOC_DMA],
801 kmalloc_info[i].size,
802 SLAB_CACHE_DMA | flags, 0,
803 kmalloc_info[i].size);
804 }
805 }
806#endif
807}
808#endif /* !CONFIG_SLOB */
809
810gfp_t kmalloc_fix_flags(gfp_t flags)
811{
812 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK;
813
814 flags &= ~GFP_SLAB_BUG_MASK;
815 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
816 invalid_mask, &invalid_mask, flags, &flags);
817 dump_stack();
818
819 return flags;
820}
821
822/*
823 * To avoid unnecessary overhead, we pass through large allocation requests
824 * directly to the page allocator. We use __GFP_COMP, because we will need to
825 * know the allocation order to free the pages properly in kfree.
826 */
827void *kmalloc_order(size_t size, gfp_t flags, unsigned int order)
828{
829 void *ret = NULL;
830 struct page *page;
831
832 if (unlikely(flags & GFP_SLAB_BUG_MASK))
833 flags = kmalloc_fix_flags(flags);
834
835 flags |= __GFP_COMP;
836 page = alloc_pages(flags, order);
837 if (likely(page)) {
838 ret = page_address(page);
839 mod_node_page_state(page_pgdat(page), NR_SLAB_UNRECLAIMABLE_B,
840 PAGE_SIZE << order);
841 }
842 ret = kasan_kmalloc_large(ret, size, flags);
843 /* As ret might get tagged, call kmemleak hook after KASAN. */
844 kmemleak_alloc(ret, size, 1, flags);
845 return ret;
846}
847EXPORT_SYMBOL(kmalloc_order);
848
849#ifdef CONFIG_TRACING
850void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
851{
852 void *ret = kmalloc_order(size, flags, order);
853 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
854 return ret;
855}
856EXPORT_SYMBOL(kmalloc_order_trace);
857#endif
858
859#ifdef CONFIG_SLAB_FREELIST_RANDOM
860/* Randomize a generic freelist */
861static void freelist_randomize(struct rnd_state *state, unsigned int *list,
862 unsigned int count)
863{
864 unsigned int rand;
865 unsigned int i;
866
867 for (i = 0; i < count; i++)
868 list[i] = i;
869
870 /* Fisher-Yates shuffle */
871 for (i = count - 1; i > 0; i--) {
872 rand = prandom_u32_state(state);
873 rand %= (i + 1);
874 swap(list[i], list[rand]);
875 }
876}
877
878/* Create a random sequence per cache */
879int cache_random_seq_create(struct kmem_cache *cachep, unsigned int count,
880 gfp_t gfp)
881{
882 struct rnd_state state;
883
884 if (count < 2 || cachep->random_seq)
885 return 0;
886
887 cachep->random_seq = kcalloc(count, sizeof(unsigned int), gfp);
888 if (!cachep->random_seq)
889 return -ENOMEM;
890
891 /* Get best entropy at this stage of boot */
892 prandom_seed_state(&state, get_random_long());
893
894 freelist_randomize(&state, cachep->random_seq, count);
895 return 0;
896}
897
898/* Destroy the per-cache random freelist sequence */
899void cache_random_seq_destroy(struct kmem_cache *cachep)
900{
901 kfree(cachep->random_seq);
902 cachep->random_seq = NULL;
903}
904#endif /* CONFIG_SLAB_FREELIST_RANDOM */
905
906#if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
907#ifdef CONFIG_SLAB
908#define SLABINFO_RIGHTS (0600)
909#else
910#define SLABINFO_RIGHTS (0400)
911#endif
912
913static void print_slabinfo_header(struct seq_file *m)
914{
915 /*
916 * Output format version, so at least we can change it
917 * without _too_ many complaints.
918 */
919#ifdef CONFIG_DEBUG_SLAB
920 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
921#else
922 seq_puts(m, "slabinfo - version: 2.1\n");
923#endif
924 seq_puts(m, "# name <active_objs> <num_objs> <objsize> <objperslab> <pagesperslab>");
925 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
926 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
927#ifdef CONFIG_DEBUG_SLAB
928 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> <error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
929 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
930#endif
931 seq_putc(m, '\n');
932}
933
934void *slab_start(struct seq_file *m, loff_t *pos)
935{
936 mutex_lock(&slab_mutex);
937 return seq_list_start(&slab_caches, *pos);
938}
939
940void *slab_next(struct seq_file *m, void *p, loff_t *pos)
941{
942 return seq_list_next(p, &slab_caches, pos);
943}
944
945void slab_stop(struct seq_file *m, void *p)
946{
947 mutex_unlock(&slab_mutex);
948}
949
950static void cache_show(struct kmem_cache *s, struct seq_file *m)
951{
952 struct slabinfo sinfo;
953
954 memset(&sinfo, 0, sizeof(sinfo));
955 get_slabinfo(s, &sinfo);
956
957 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
958 s->name, sinfo.active_objs, sinfo.num_objs, s->size,
959 sinfo.objects_per_slab, (1 << sinfo.cache_order));
960
961 seq_printf(m, " : tunables %4u %4u %4u",
962 sinfo.limit, sinfo.batchcount, sinfo.shared);
963 seq_printf(m, " : slabdata %6lu %6lu %6lu",
964 sinfo.active_slabs, sinfo.num_slabs, sinfo.shared_avail);
965 slabinfo_show_stats(m, s);
966 seq_putc(m, '\n');
967}
968
969static int slab_show(struct seq_file *m, void *p)
970{
971 struct kmem_cache *s = list_entry(p, struct kmem_cache, list);
972
973 if (p == slab_caches.next)
974 print_slabinfo_header(m);
975 cache_show(s, m);
976 return 0;
977}
978
979void dump_unreclaimable_slab(void)
980{
981 struct kmem_cache *s, *s2;
982 struct slabinfo sinfo;
983
984 /*
985 * Here acquiring slab_mutex is risky since we don't prefer to get
986 * sleep in oom path. But, without mutex hold, it may introduce a
987 * risk of crash.
988 * Use mutex_trylock to protect the list traverse, dump nothing
989 * without acquiring the mutex.
990 */
991 if (!mutex_trylock(&slab_mutex)) {
992 pr_warn("excessive unreclaimable slab but cannot dump stats\n");
993 return;
994 }
995
996 pr_info("Unreclaimable slab info:\n");
997 pr_info("Name Used Total\n");
998
999 list_for_each_entry_safe(s, s2, &slab_caches, list) {
1000 if (s->flags & SLAB_RECLAIM_ACCOUNT)
1001 continue;
1002
1003 get_slabinfo(s, &sinfo);
1004
1005 if (sinfo.num_objs > 0)
1006 pr_info("%-17s %10luKB %10luKB\n", s->name,
1007 (sinfo.active_objs * s->size) / 1024,
1008 (sinfo.num_objs * s->size) / 1024);
1009 }
1010 mutex_unlock(&slab_mutex);
1011}
1012
1013#if defined(CONFIG_MEMCG_KMEM)
1014int memcg_slab_show(struct seq_file *m, void *p)
1015{
1016 /*
1017 * Deprecated.
1018 * Please, take a look at tools/cgroup/slabinfo.py .
1019 */
1020 return 0;
1021}
1022#endif
1023
1024/*
1025 * slabinfo_op - iterator that generates /proc/slabinfo
1026 *
1027 * Output layout:
1028 * cache-name
1029 * num-active-objs
1030 * total-objs
1031 * object size
1032 * num-active-slabs
1033 * total-slabs
1034 * num-pages-per-slab
1035 * + further values on SMP and with statistics enabled
1036 */
1037static const struct seq_operations slabinfo_op = {
1038 .start = slab_start,
1039 .next = slab_next,
1040 .stop = slab_stop,
1041 .show = slab_show,
1042};
1043
1044static int slabinfo_open(struct inode *inode, struct file *file)
1045{
1046 return seq_open(file, &slabinfo_op);
1047}
1048
1049static const struct proc_ops slabinfo_proc_ops = {
1050 .proc_flags = PROC_ENTRY_PERMANENT,
1051 .proc_open = slabinfo_open,
1052 .proc_read = seq_read,
1053 .proc_write = slabinfo_write,
1054 .proc_lseek = seq_lseek,
1055 .proc_release = seq_release,
1056};
1057
1058static int __init slab_proc_init(void)
1059{
1060 proc_create("slabinfo", SLABINFO_RIGHTS, NULL, &slabinfo_proc_ops);
1061 return 0;
1062}
1063module_init(slab_proc_init);
1064
1065#endif /* CONFIG_SLAB || CONFIG_SLUB_DEBUG */
1066
1067static __always_inline void *__do_krealloc(const void *p, size_t new_size,
1068 gfp_t flags)
1069{
1070 void *ret;
1071 size_t ks;
1072
1073 ks = ksize(p);
1074
1075 if (ks >= new_size) {
1076 p = kasan_krealloc((void *)p, new_size, flags);
1077 return (void *)p;
1078 }
1079
1080 ret = kmalloc_track_caller(new_size, flags);
1081 if (ret && p)
1082 memcpy(ret, p, ks);
1083
1084 return ret;
1085}
1086
1087/**
1088 * krealloc - reallocate memory. The contents will remain unchanged.
1089 * @p: object to reallocate memory for.
1090 * @new_size: how many bytes of memory are required.
1091 * @flags: the type of memory to allocate.
1092 *
1093 * The contents of the object pointed to are preserved up to the
1094 * lesser of the new and old sizes. If @p is %NULL, krealloc()
1095 * behaves exactly like kmalloc(). If @new_size is 0 and @p is not a
1096 * %NULL pointer, the object pointed to is freed.
1097 *
1098 * Return: pointer to the allocated memory or %NULL in case of error
1099 */
1100void *krealloc(const void *p, size_t new_size, gfp_t flags)
1101{
1102 void *ret;
1103
1104 if (unlikely(!new_size)) {
1105 kfree(p);
1106 return ZERO_SIZE_PTR;
1107 }
1108
1109 ret = __do_krealloc(p, new_size, flags);
1110 if (ret && kasan_reset_tag(p) != kasan_reset_tag(ret))
1111 kfree(p);
1112
1113 return ret;
1114}
1115EXPORT_SYMBOL(krealloc);
1116
1117/**
1118 * kfree_sensitive - Clear sensitive information in memory before freeing
1119 * @p: object to free memory of
1120 *
1121 * The memory of the object @p points to is zeroed before freed.
1122 * If @p is %NULL, kfree_sensitive() does nothing.
1123 *
1124 * Note: this function zeroes the whole allocated buffer which can be a good
1125 * deal bigger than the requested buffer size passed to kmalloc(). So be
1126 * careful when using this function in performance sensitive code.
1127 */
1128void kfree_sensitive(const void *p)
1129{
1130 size_t ks;
1131 void *mem = (void *)p;
1132
1133 ks = ksize(mem);
1134 if (ks)
1135 memzero_explicit(mem, ks);
1136 kfree(mem);
1137}
1138EXPORT_SYMBOL(kfree_sensitive);
1139
1140/**
1141 * ksize - get the actual amount of memory allocated for a given object
1142 * @objp: Pointer to the object
1143 *
1144 * kmalloc may internally round up allocations and return more memory
1145 * than requested. ksize() can be used to determine the actual amount of
1146 * memory allocated. The caller may use this additional memory, even though
1147 * a smaller amount of memory was initially specified with the kmalloc call.
1148 * The caller must guarantee that objp points to a valid object previously
1149 * allocated with either kmalloc() or kmem_cache_alloc(). The object
1150 * must not be freed during the duration of the call.
1151 *
1152 * Return: size of the actual memory used by @objp in bytes
1153 */
1154size_t ksize(const void *objp)
1155{
1156 size_t size;
1157
1158 /*
1159 * We need to check that the pointed to object is valid, and only then
1160 * unpoison the shadow memory below. We use __kasan_check_read(), to
1161 * generate a more useful report at the time ksize() is called (rather
1162 * than later where behaviour is undefined due to potential
1163 * use-after-free or double-free).
1164 *
1165 * If the pointed to memory is invalid we return 0, to avoid users of
1166 * ksize() writing to and potentially corrupting the memory region.
1167 *
1168 * We want to perform the check before __ksize(), to avoid potentially
1169 * crashing in __ksize() due to accessing invalid metadata.
1170 */
1171 if (unlikely(ZERO_OR_NULL_PTR(objp)) || !__kasan_check_read(objp, 1))
1172 return 0;
1173
1174 size = __ksize(objp);
1175 /*
1176 * We assume that ksize callers could use whole allocated area,
1177 * so we need to unpoison this area.
1178 */
1179 kasan_unpoison_shadow(objp, size);
1180 return size;
1181}
1182EXPORT_SYMBOL(ksize);
1183
1184/* Tracepoints definitions. */
1185EXPORT_TRACEPOINT_SYMBOL(kmalloc);
1186EXPORT_TRACEPOINT_SYMBOL(kmem_cache_alloc);
1187EXPORT_TRACEPOINT_SYMBOL(kmalloc_node);
1188EXPORT_TRACEPOINT_SYMBOL(kmem_cache_alloc_node);
1189EXPORT_TRACEPOINT_SYMBOL(kfree);
1190EXPORT_TRACEPOINT_SYMBOL(kmem_cache_free);
1191
1192int should_failslab(struct kmem_cache *s, gfp_t gfpflags)
1193{
1194 if (__should_failslab(s, gfpflags))
1195 return -ENOMEM;
1196 return 0;
1197}
1198ALLOW_ERROR_INJECTION(should_failslab, ERRNO);
1/*
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>
14#include <linux/cpu.h>
15#include <linux/uaccess.h>
16#include <linux/seq_file.h>
17#include <linux/proc_fs.h>
18#include <asm/cacheflush.h>
19#include <asm/tlbflush.h>
20#include <asm/page.h>
21#include <linux/memcontrol.h>
22#include <trace/events/kmem.h>
23
24#include "slab.h"
25
26enum slab_state slab_state;
27LIST_HEAD(slab_caches);
28DEFINE_MUTEX(slab_mutex);
29struct kmem_cache *kmem_cache;
30
31#ifdef CONFIG_DEBUG_VM
32static int kmem_cache_sanity_check(const char *name, size_t size)
33{
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
58#if !defined(CONFIG_SLUB) || !defined(CONFIG_SLUB_DEBUG_ON)
59 if (!strcmp(s->name, name)) {
60 pr_err("%s (%s): Cache name already exists.\n",
61 __func__, name);
62 dump_stack();
63 s = NULL;
64 return -EINVAL;
65 }
66#endif
67 }
68
69 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
70 return 0;
71}
72#else
73static inline int kmem_cache_sanity_check(const char *name, size_t size)
74{
75 return 0;
76}
77#endif
78
79#ifdef CONFIG_MEMCG_KMEM
80int memcg_update_all_caches(int num_memcgs)
81{
82 struct kmem_cache *s;
83 int ret = 0;
84 mutex_lock(&slab_mutex);
85
86 list_for_each_entry(s, &slab_caches, list) {
87 if (!is_root_cache(s))
88 continue;
89
90 ret = memcg_update_cache_size(s, num_memcgs);
91 /*
92 * See comment in memcontrol.c, memcg_update_cache_size:
93 * Instead of freeing the memory, we'll just leave the caches
94 * up to this point in an updated state.
95 */
96 if (ret)
97 goto out;
98 }
99
100 memcg_update_array_size(num_memcgs);
101out:
102 mutex_unlock(&slab_mutex);
103 return ret;
104}
105#endif
106
107/*
108 * Figure out what the alignment of the objects will be given a set of
109 * flags, a user specified alignment and the size of the objects.
110 */
111unsigned long calculate_alignment(unsigned long flags,
112 unsigned long align, unsigned long size)
113{
114 /*
115 * If the user wants hardware cache aligned objects then follow that
116 * suggestion if the object is sufficiently large.
117 *
118 * The hardware cache alignment cannot override the specified
119 * alignment though. If that is greater then use it.
120 */
121 if (flags & SLAB_HWCACHE_ALIGN) {
122 unsigned long ralign = cache_line_size();
123 while (size <= ralign / 2)
124 ralign /= 2;
125 align = max(align, ralign);
126 }
127
128 if (align < ARCH_SLAB_MINALIGN)
129 align = ARCH_SLAB_MINALIGN;
130
131 return ALIGN(align, sizeof(void *));
132}
133
134static struct kmem_cache *
135do_kmem_cache_create(char *name, size_t object_size, size_t size, size_t align,
136 unsigned long flags, void (*ctor)(void *),
137 struct mem_cgroup *memcg, struct kmem_cache *root_cache)
138{
139 struct kmem_cache *s;
140 int err;
141
142 err = -ENOMEM;
143 s = kmem_cache_zalloc(kmem_cache, GFP_KERNEL);
144 if (!s)
145 goto out;
146
147 s->name = name;
148 s->object_size = object_size;
149 s->size = size;
150 s->align = align;
151 s->ctor = ctor;
152
153 err = memcg_alloc_cache_params(memcg, s, root_cache);
154 if (err)
155 goto out_free_cache;
156
157 err = __kmem_cache_create(s, flags);
158 if (err)
159 goto out_free_cache;
160
161 s->refcount = 1;
162 list_add(&s->list, &slab_caches);
163 memcg_register_cache(s);
164out:
165 if (err)
166 return ERR_PTR(err);
167 return s;
168
169out_free_cache:
170 memcg_free_cache_params(s);
171 kfree(s);
172 goto out;
173}
174
175/*
176 * kmem_cache_create - Create a cache.
177 * @name: A string which is used in /proc/slabinfo to identify this cache.
178 * @size: The size of objects to be created in this cache.
179 * @align: The required alignment for the objects.
180 * @flags: SLAB flags
181 * @ctor: A constructor for the objects.
182 *
183 * Returns a ptr to the cache on success, NULL on failure.
184 * Cannot be called within a interrupt, but can be interrupted.
185 * The @ctor is run when new pages are allocated by the cache.
186 *
187 * The flags are
188 *
189 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
190 * to catch references to uninitialised memory.
191 *
192 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
193 * for buffer overruns.
194 *
195 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
196 * cacheline. This can be beneficial if you're counting cycles as closely
197 * as davem.
198 */
199struct kmem_cache *
200kmem_cache_create(const char *name, size_t size, size_t align,
201 unsigned long flags, void (*ctor)(void *))
202{
203 struct kmem_cache *s;
204 char *cache_name;
205 int err;
206
207 get_online_cpus();
208 mutex_lock(&slab_mutex);
209
210 err = kmem_cache_sanity_check(name, size);
211 if (err)
212 goto out_unlock;
213
214 /*
215 * Some allocators will constraint the set of valid flags to a subset
216 * of all flags. We expect them to define CACHE_CREATE_MASK in this
217 * case, and we'll just provide them with a sanitized version of the
218 * passed flags.
219 */
220 flags &= CACHE_CREATE_MASK;
221
222 s = __kmem_cache_alias(name, size, align, flags, ctor);
223 if (s)
224 goto out_unlock;
225
226 cache_name = kstrdup(name, GFP_KERNEL);
227 if (!cache_name) {
228 err = -ENOMEM;
229 goto out_unlock;
230 }
231
232 s = do_kmem_cache_create(cache_name, size, size,
233 calculate_alignment(flags, align, size),
234 flags, ctor, NULL, NULL);
235 if (IS_ERR(s)) {
236 err = PTR_ERR(s);
237 kfree(cache_name);
238 }
239
240out_unlock:
241 mutex_unlock(&slab_mutex);
242 put_online_cpus();
243
244 if (err) {
245 if (flags & SLAB_PANIC)
246 panic("kmem_cache_create: Failed to create slab '%s'. Error %d\n",
247 name, err);
248 else {
249 printk(KERN_WARNING "kmem_cache_create(%s) failed with error %d",
250 name, err);
251 dump_stack();
252 }
253 return NULL;
254 }
255 return s;
256}
257EXPORT_SYMBOL(kmem_cache_create);
258
259#ifdef CONFIG_MEMCG_KMEM
260/*
261 * kmem_cache_create_memcg - Create a cache for a memory cgroup.
262 * @memcg: The memory cgroup the new cache is for.
263 * @root_cache: The parent of the new cache.
264 *
265 * This function attempts to create a kmem cache that will serve allocation
266 * requests going from @memcg to @root_cache. The new cache inherits properties
267 * from its parent.
268 */
269void kmem_cache_create_memcg(struct mem_cgroup *memcg, struct kmem_cache *root_cache)
270{
271 struct kmem_cache *s;
272 char *cache_name;
273
274 get_online_cpus();
275 mutex_lock(&slab_mutex);
276
277 /*
278 * Since per-memcg caches are created asynchronously on first
279 * allocation (see memcg_kmem_get_cache()), several threads can try to
280 * create the same cache, but only one of them may succeed.
281 */
282 if (cache_from_memcg_idx(root_cache, memcg_cache_id(memcg)))
283 goto out_unlock;
284
285 cache_name = memcg_create_cache_name(memcg, root_cache);
286 if (!cache_name)
287 goto out_unlock;
288
289 s = do_kmem_cache_create(cache_name, root_cache->object_size,
290 root_cache->size, root_cache->align,
291 root_cache->flags, root_cache->ctor,
292 memcg, root_cache);
293 if (IS_ERR(s)) {
294 kfree(cache_name);
295 goto out_unlock;
296 }
297
298 s->allocflags |= __GFP_KMEMCG;
299
300out_unlock:
301 mutex_unlock(&slab_mutex);
302 put_online_cpus();
303}
304
305static int kmem_cache_destroy_memcg_children(struct kmem_cache *s)
306{
307 int rc;
308
309 if (!s->memcg_params ||
310 !s->memcg_params->is_root_cache)
311 return 0;
312
313 mutex_unlock(&slab_mutex);
314 rc = __kmem_cache_destroy_memcg_children(s);
315 mutex_lock(&slab_mutex);
316
317 return rc;
318}
319#else
320static int kmem_cache_destroy_memcg_children(struct kmem_cache *s)
321{
322 return 0;
323}
324#endif /* CONFIG_MEMCG_KMEM */
325
326void slab_kmem_cache_release(struct kmem_cache *s)
327{
328 kfree(s->name);
329 kmem_cache_free(kmem_cache, s);
330}
331
332void kmem_cache_destroy(struct kmem_cache *s)
333{
334 get_online_cpus();
335 mutex_lock(&slab_mutex);
336
337 s->refcount--;
338 if (s->refcount)
339 goto out_unlock;
340
341 if (kmem_cache_destroy_memcg_children(s) != 0)
342 goto out_unlock;
343
344 list_del(&s->list);
345 memcg_unregister_cache(s);
346
347 if (__kmem_cache_shutdown(s) != 0) {
348 list_add(&s->list, &slab_caches);
349 memcg_register_cache(s);
350 printk(KERN_ERR "kmem_cache_destroy %s: "
351 "Slab cache still has objects\n", s->name);
352 dump_stack();
353 goto out_unlock;
354 }
355
356 mutex_unlock(&slab_mutex);
357 if (s->flags & SLAB_DESTROY_BY_RCU)
358 rcu_barrier();
359
360 memcg_free_cache_params(s);
361#ifdef SLAB_SUPPORTS_SYSFS
362 sysfs_slab_remove(s);
363#else
364 slab_kmem_cache_release(s);
365#endif
366 goto out_put_cpus;
367
368out_unlock:
369 mutex_unlock(&slab_mutex);
370out_put_cpus:
371 put_online_cpus();
372}
373EXPORT_SYMBOL(kmem_cache_destroy);
374
375int slab_is_available(void)
376{
377 return slab_state >= UP;
378}
379
380#ifndef CONFIG_SLOB
381/* Create a cache during boot when no slab services are available yet */
382void __init create_boot_cache(struct kmem_cache *s, const char *name, size_t size,
383 unsigned long flags)
384{
385 int err;
386
387 s->name = name;
388 s->size = s->object_size = size;
389 s->align = calculate_alignment(flags, ARCH_KMALLOC_MINALIGN, size);
390 err = __kmem_cache_create(s, flags);
391
392 if (err)
393 panic("Creation of kmalloc slab %s size=%zu failed. Reason %d\n",
394 name, size, err);
395
396 s->refcount = -1; /* Exempt from merging for now */
397}
398
399struct kmem_cache *__init create_kmalloc_cache(const char *name, size_t size,
400 unsigned long flags)
401{
402 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
403
404 if (!s)
405 panic("Out of memory when creating slab %s\n", name);
406
407 create_boot_cache(s, name, size, flags);
408 list_add(&s->list, &slab_caches);
409 s->refcount = 1;
410 return s;
411}
412
413struct kmem_cache *kmalloc_caches[KMALLOC_SHIFT_HIGH + 1];
414EXPORT_SYMBOL(kmalloc_caches);
415
416#ifdef CONFIG_ZONE_DMA
417struct kmem_cache *kmalloc_dma_caches[KMALLOC_SHIFT_HIGH + 1];
418EXPORT_SYMBOL(kmalloc_dma_caches);
419#endif
420
421/*
422 * Conversion table for small slabs sizes / 8 to the index in the
423 * kmalloc array. This is necessary for slabs < 192 since we have non power
424 * of two cache sizes there. The size of larger slabs can be determined using
425 * fls.
426 */
427static s8 size_index[24] = {
428 3, /* 8 */
429 4, /* 16 */
430 5, /* 24 */
431 5, /* 32 */
432 6, /* 40 */
433 6, /* 48 */
434 6, /* 56 */
435 6, /* 64 */
436 1, /* 72 */
437 1, /* 80 */
438 1, /* 88 */
439 1, /* 96 */
440 7, /* 104 */
441 7, /* 112 */
442 7, /* 120 */
443 7, /* 128 */
444 2, /* 136 */
445 2, /* 144 */
446 2, /* 152 */
447 2, /* 160 */
448 2, /* 168 */
449 2, /* 176 */
450 2, /* 184 */
451 2 /* 192 */
452};
453
454static inline int size_index_elem(size_t bytes)
455{
456 return (bytes - 1) / 8;
457}
458
459/*
460 * Find the kmem_cache structure that serves a given size of
461 * allocation
462 */
463struct kmem_cache *kmalloc_slab(size_t size, gfp_t flags)
464{
465 int index;
466
467 if (unlikely(size > KMALLOC_MAX_SIZE)) {
468 WARN_ON_ONCE(!(flags & __GFP_NOWARN));
469 return NULL;
470 }
471
472 if (size <= 192) {
473 if (!size)
474 return ZERO_SIZE_PTR;
475
476 index = size_index[size_index_elem(size)];
477 } else
478 index = fls(size - 1);
479
480#ifdef CONFIG_ZONE_DMA
481 if (unlikely((flags & GFP_DMA)))
482 return kmalloc_dma_caches[index];
483
484#endif
485 return kmalloc_caches[index];
486}
487
488/*
489 * Create the kmalloc array. Some of the regular kmalloc arrays
490 * may already have been created because they were needed to
491 * enable allocations for slab creation.
492 */
493void __init create_kmalloc_caches(unsigned long flags)
494{
495 int i;
496
497 /*
498 * Patch up the size_index table if we have strange large alignment
499 * requirements for the kmalloc array. This is only the case for
500 * MIPS it seems. The standard arches will not generate any code here.
501 *
502 * Largest permitted alignment is 256 bytes due to the way we
503 * handle the index determination for the smaller caches.
504 *
505 * Make sure that nothing crazy happens if someone starts tinkering
506 * around with ARCH_KMALLOC_MINALIGN
507 */
508 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
509 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
510
511 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
512 int elem = size_index_elem(i);
513
514 if (elem >= ARRAY_SIZE(size_index))
515 break;
516 size_index[elem] = KMALLOC_SHIFT_LOW;
517 }
518
519 if (KMALLOC_MIN_SIZE >= 64) {
520 /*
521 * The 96 byte size cache is not used if the alignment
522 * is 64 byte.
523 */
524 for (i = 64 + 8; i <= 96; i += 8)
525 size_index[size_index_elem(i)] = 7;
526
527 }
528
529 if (KMALLOC_MIN_SIZE >= 128) {
530 /*
531 * The 192 byte sized cache is not used if the alignment
532 * is 128 byte. Redirect kmalloc to use the 256 byte cache
533 * instead.
534 */
535 for (i = 128 + 8; i <= 192; i += 8)
536 size_index[size_index_elem(i)] = 8;
537 }
538 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) {
539 if (!kmalloc_caches[i]) {
540 kmalloc_caches[i] = create_kmalloc_cache(NULL,
541 1 << i, flags);
542 }
543
544 /*
545 * Caches that are not of the two-to-the-power-of size.
546 * These have to be created immediately after the
547 * earlier power of two caches
548 */
549 if (KMALLOC_MIN_SIZE <= 32 && !kmalloc_caches[1] && i == 6)
550 kmalloc_caches[1] = create_kmalloc_cache(NULL, 96, flags);
551
552 if (KMALLOC_MIN_SIZE <= 64 && !kmalloc_caches[2] && i == 7)
553 kmalloc_caches[2] = create_kmalloc_cache(NULL, 192, flags);
554 }
555
556 /* Kmalloc array is now usable */
557 slab_state = UP;
558
559 for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
560 struct kmem_cache *s = kmalloc_caches[i];
561 char *n;
562
563 if (s) {
564 n = kasprintf(GFP_NOWAIT, "kmalloc-%d", kmalloc_size(i));
565
566 BUG_ON(!n);
567 s->name = n;
568 }
569 }
570
571#ifdef CONFIG_ZONE_DMA
572 for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
573 struct kmem_cache *s = kmalloc_caches[i];
574
575 if (s) {
576 int size = kmalloc_size(i);
577 char *n = kasprintf(GFP_NOWAIT,
578 "dma-kmalloc-%d", size);
579
580 BUG_ON(!n);
581 kmalloc_dma_caches[i] = create_kmalloc_cache(n,
582 size, SLAB_CACHE_DMA | flags);
583 }
584 }
585#endif
586}
587#endif /* !CONFIG_SLOB */
588
589#ifdef CONFIG_TRACING
590void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
591{
592 void *ret = kmalloc_order(size, flags, order);
593 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
594 return ret;
595}
596EXPORT_SYMBOL(kmalloc_order_trace);
597#endif
598
599#ifdef CONFIG_SLABINFO
600
601#ifdef CONFIG_SLAB
602#define SLABINFO_RIGHTS (S_IWUSR | S_IRUSR)
603#else
604#define SLABINFO_RIGHTS S_IRUSR
605#endif
606
607void print_slabinfo_header(struct seq_file *m)
608{
609 /*
610 * Output format version, so at least we can change it
611 * without _too_ many complaints.
612 */
613#ifdef CONFIG_DEBUG_SLAB
614 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
615#else
616 seq_puts(m, "slabinfo - version: 2.1\n");
617#endif
618 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
619 "<objperslab> <pagesperslab>");
620 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
621 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
622#ifdef CONFIG_DEBUG_SLAB
623 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
624 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
625 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
626#endif
627 seq_putc(m, '\n');
628}
629
630static void *s_start(struct seq_file *m, loff_t *pos)
631{
632 loff_t n = *pos;
633
634 mutex_lock(&slab_mutex);
635 if (!n)
636 print_slabinfo_header(m);
637
638 return seq_list_start(&slab_caches, *pos);
639}
640
641void *slab_next(struct seq_file *m, void *p, loff_t *pos)
642{
643 return seq_list_next(p, &slab_caches, pos);
644}
645
646void slab_stop(struct seq_file *m, void *p)
647{
648 mutex_unlock(&slab_mutex);
649}
650
651static void
652memcg_accumulate_slabinfo(struct kmem_cache *s, struct slabinfo *info)
653{
654 struct kmem_cache *c;
655 struct slabinfo sinfo;
656 int i;
657
658 if (!is_root_cache(s))
659 return;
660
661 for_each_memcg_cache_index(i) {
662 c = cache_from_memcg_idx(s, i);
663 if (!c)
664 continue;
665
666 memset(&sinfo, 0, sizeof(sinfo));
667 get_slabinfo(c, &sinfo);
668
669 info->active_slabs += sinfo.active_slabs;
670 info->num_slabs += sinfo.num_slabs;
671 info->shared_avail += sinfo.shared_avail;
672 info->active_objs += sinfo.active_objs;
673 info->num_objs += sinfo.num_objs;
674 }
675}
676
677int cache_show(struct kmem_cache *s, struct seq_file *m)
678{
679 struct slabinfo sinfo;
680
681 memset(&sinfo, 0, sizeof(sinfo));
682 get_slabinfo(s, &sinfo);
683
684 memcg_accumulate_slabinfo(s, &sinfo);
685
686 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
687 cache_name(s), sinfo.active_objs, sinfo.num_objs, s->size,
688 sinfo.objects_per_slab, (1 << sinfo.cache_order));
689
690 seq_printf(m, " : tunables %4u %4u %4u",
691 sinfo.limit, sinfo.batchcount, sinfo.shared);
692 seq_printf(m, " : slabdata %6lu %6lu %6lu",
693 sinfo.active_slabs, sinfo.num_slabs, sinfo.shared_avail);
694 slabinfo_show_stats(m, s);
695 seq_putc(m, '\n');
696 return 0;
697}
698
699static int s_show(struct seq_file *m, void *p)
700{
701 struct kmem_cache *s = list_entry(p, struct kmem_cache, list);
702
703 if (!is_root_cache(s))
704 return 0;
705 return cache_show(s, m);
706}
707
708/*
709 * slabinfo_op - iterator that generates /proc/slabinfo
710 *
711 * Output layout:
712 * cache-name
713 * num-active-objs
714 * total-objs
715 * object size
716 * num-active-slabs
717 * total-slabs
718 * num-pages-per-slab
719 * + further values on SMP and with statistics enabled
720 */
721static const struct seq_operations slabinfo_op = {
722 .start = s_start,
723 .next = slab_next,
724 .stop = slab_stop,
725 .show = s_show,
726};
727
728static int slabinfo_open(struct inode *inode, struct file *file)
729{
730 return seq_open(file, &slabinfo_op);
731}
732
733static const struct file_operations proc_slabinfo_operations = {
734 .open = slabinfo_open,
735 .read = seq_read,
736 .write = slabinfo_write,
737 .llseek = seq_lseek,
738 .release = seq_release,
739};
740
741static int __init slab_proc_init(void)
742{
743 proc_create("slabinfo", SLABINFO_RIGHTS, NULL,
744 &proc_slabinfo_operations);
745 return 0;
746}
747module_init(slab_proc_init);
748#endif /* CONFIG_SLABINFO */