Loading...
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
23#define CREATE_TRACE_POINTS
24#include <trace/events/kmem.h>
25
26#include "slab.h"
27
28enum slab_state slab_state;
29LIST_HEAD(slab_caches);
30DEFINE_MUTEX(slab_mutex);
31struct kmem_cache *kmem_cache;
32
33/*
34 * 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 | \
38 SLAB_FAILSLAB | SLAB_KASAN)
39
40#define SLAB_MERGE_SAME (SLAB_RECLAIM_ACCOUNT | SLAB_CACHE_DMA | \
41 SLAB_NOTRACK | SLAB_ACCOUNT)
42
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/*
62 * 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
70#ifdef CONFIG_DEBUG_VM
71static int kmem_cache_sanity_check(const char *name, size_t size)
72{
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 }
96 }
97
98 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
99 return 0;
100}
101#else
102static inline int kmem_cache_sanity_check(const char *name, size_t size)
103{
104 return 0;
105}
106#endif
107
108void __kmem_cache_free_bulk(struct kmem_cache *s, size_t nr, void **p)
109{
110 size_t i;
111
112 for (i = 0; i < nr; i++) {
113 if (s)
114 kmem_cache_free(s, p[i]);
115 else
116 kfree(p[i]);
117 }
118}
119
120int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t nr,
121 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);
129 return 0;
130 }
131 }
132 return i;
133}
134
135#if defined(CONFIG_MEMCG) && !defined(CONFIG_SLOB)
136void slab_init_memcg_params(struct kmem_cache *s)
137{
138 s->memcg_params.is_root_cache = true;
139 INIT_LIST_HEAD(&s->memcg_params.list);
140 RCU_INIT_POINTER(s->memcg_params.memcg_caches, NULL);
141}
142
143static 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;
147
148 if (memcg) {
149 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 }
154
155 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);
167 return 0;
168}
169
170static void destroy_memcg_params(struct kmem_cache *s)
171{
172 if (is_root_cache(s))
173 kfree(rcu_access_pointer(s->memcg_params.memcg_caches));
174}
175
176static int update_memcg_params(struct kmem_cache *s, int new_array_size)
177{
178 struct memcg_cache_array *old, *new;
179
180 if (!is_root_cache(s))
181 return 0;
182
183 new = kzalloc(sizeof(struct memcg_cache_array) +
184 new_array_size * sizeof(void *), GFP_KERNEL);
185 if (!new)
186 return -ENOMEM;
187
188 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 *));
193
194 rcu_assign_pointer(s->memcg_params.memcg_caches, new);
195 if (old)
196 kfree_rcu(old, rcu);
197 return 0;
198}
199
200int memcg_update_all_caches(int num_memcgs)
201{
202 struct kmem_cache *s;
203 int ret = 0;
204
205 mutex_lock(&slab_mutex);
206 list_for_each_entry(s, &slab_caches, list) {
207 ret = update_memcg_params(s, num_memcgs);
208 /*
209 * Instead of freeing the memory, we'll just leave the caches
210 * up to this point in an updated state.
211 */
212 if (ret)
213 break;
214 }
215 mutex_unlock(&slab_mutex);
216 return ret;
217}
218#else
219static inline int init_memcg_params(struct kmem_cache *s,
220 struct mem_cgroup *memcg, struct kmem_cache *root_cache)
221{
222 return 0;
223}
224
225static inline void destroy_memcg_params(struct kmem_cache *s)
226{
227}
228#endif /* CONFIG_MEMCG && !CONFIG_SLOB */
229
230/*
231 * 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
269 list_for_each_entry_reverse(s, &slab_caches, list) {
270 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
288 if (IS_ENABLED(CONFIG_SLAB) && align &&
289 (align > s->align || s->align % align))
290 continue;
291
292 return s;
293 }
294 return NULL;
295}
296
297/*
298 * 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
324static 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)
328{
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
343 err = init_memcg_params(s, memcg, root_cache);
344 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);
353out:
354 if (err)
355 return ERR_PTR(err);
356 return s;
357
358out_free_cache:
359 destroy_memcg_params(s);
360 kmem_cache_free(kmem_cache, s);
361 goto out;
362}
363
364/*
365 * 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 */
388struct kmem_cache *
389kmem_cache_create(const char *name, size_t size, size_t align,
390 unsigned long flags, void (*ctor)(void *))
391{
392 struct kmem_cache *s = NULL;
393 const char *cache_name;
394 int err;
395
396 get_online_cpus();
397 get_online_mems();
398 memcg_get_cache_ids();
399
400 mutex_lock(&slab_mutex);
401
402 err = kmem_cache_sanity_check(name, size);
403 if (err) {
404 goto out_unlock;
405 }
406
407 /*
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;
414
415 s = __kmem_cache_alias(name, size, align, flags, ctor);
416 if (s)
417 goto out_unlock;
418
419 cache_name = kstrdup_const(name, GFP_KERNEL);
420 if (!cache_name) {
421 err = -ENOMEM;
422 goto out_unlock;
423 }
424
425 s = create_cache(cache_name, size, size,
426 calculate_alignment(flags, align, size),
427 flags, ctor, NULL, NULL);
428 if (IS_ERR(s)) {
429 err = PTR_ERR(s);
430 kfree_const(cache_name);
431 }
432
433out_unlock:
434 mutex_unlock(&slab_mutex);
435
436 memcg_put_cache_ids();
437 put_online_mems();
438 put_online_cpus();
439
440 if (err) {
441 if (flags & SLAB_PANIC)
442 panic("kmem_cache_create: Failed to create slab '%s'. Error %d\n",
443 name, err);
444 else {
445 pr_warn("kmem_cache_create(%s) failed with error %d\n",
446 name, err);
447 dump_stack();
448 }
449 return NULL;
450 }
451 return s;
452}
453EXPORT_SYMBOL(kmem_cache_create);
454
455static int shutdown_cache(struct kmem_cache *s,
456 struct list_head *release, bool *need_rcu_barrier)
457{
458 if (__kmem_cache_shutdown(s) != 0)
459 return -EBUSY;
460
461 if (s->flags & SLAB_DESTROY_BY_RCU)
462 *need_rcu_barrier = true;
463
464 list_move(&s->list, release);
465 return 0;
466}
467
468static void release_caches(struct list_head *release, bool need_rcu_barrier)
469{
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
484#if defined(CONFIG_MEMCG) && !defined(CONFIG_SLOB)
485/*
486 * memcg_create_kmem_cache - Create a cache for a memory cgroup.
487 * @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 */
494void memcg_create_kmem_cache(struct mem_cgroup *memcg,
495 struct kmem_cache *root_cache)
496{
497 static char memcg_name_buf[NAME_MAX + 1]; /* protected by slab_mutex */
498 struct cgroup_subsys_state *css = &memcg->css;
499 struct memcg_cache_array *arr;
500 struct kmem_cache *s = NULL;
501 char *cache_name;
502 int idx;
503
504 get_online_cpus();
505 get_online_mems();
506
507 mutex_lock(&slab_mutex);
508
509 /*
510 * The memory cgroup could have been offlined while the cache
511 * creation work was pending.
512 */
513 if (memcg->kmem_state != KMEM_ONLINE)
514 goto out_unlock;
515
516 idx = memcg_cache_id(memcg);
517 arr = rcu_dereference_protected(root_cache->memcg_params.memcg_caches,
518 lockdep_is_held(&slab_mutex));
519
520 /*
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 */
525 if (arr->entries[idx])
526 goto out_unlock;
527
528 cgroup_name(css->cgroup, memcg_name_buf, sizeof(memcg_name_buf));
529 cache_name = kasprintf(GFP_KERNEL, "%s(%d:%s)", root_cache->name,
530 css->id, memcg_name_buf);
531 if (!cache_name)
532 goto out_unlock;
533
534 s = create_cache(cache_name, root_cache->object_size,
535 root_cache->size, root_cache->align,
536 root_cache->flags, root_cache->ctor,
537 memcg, root_cache);
538 /*
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 */
543 if (IS_ERR(s)) {
544 kfree(cache_name);
545 goto out_unlock;
546 }
547
548 list_add(&s->memcg_params.list, &root_cache->memcg_params.list);
549
550 /*
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();
556 arr->entries[idx] = s;
557
558out_unlock:
559 mutex_unlock(&slab_mutex);
560
561 put_online_mems();
562 put_online_cpus();
563}
564
565void memcg_deactivate_kmem_caches(struct mem_cgroup *memcg)
566{
567 int idx;
568 struct memcg_cache_array *arr;
569 struct kmem_cache *s, *c;
570
571 idx = memcg_cache_id(memcg);
572
573 get_online_cpus();
574 get_online_mems();
575
576 mutex_lock(&slab_mutex);
577 list_for_each_entry(s, &slab_caches, list) {
578 if (!is_root_cache(s))
579 continue;
580
581 arr = rcu_dereference_protected(s->memcg_params.memcg_caches,
582 lockdep_is_held(&slab_mutex));
583 c = arr->entries[idx];
584 if (!c)
585 continue;
586
587 __kmem_cache_shrink(c, true);
588 arr->entries[idx] = NULL;
589 }
590 mutex_unlock(&slab_mutex);
591
592 put_online_mems();
593 put_online_cpus();
594}
595
596static int __shutdown_memcg_cache(struct kmem_cache *s,
597 struct list_head *release, bool *need_rcu_barrier)
598{
599 BUG_ON(is_root_cache(s));
600
601 if (shutdown_cache(s, release, need_rcu_barrier))
602 return -EBUSY;
603
604 list_del(&s->memcg_params.list);
605 return 0;
606}
607
608void memcg_destroy_kmem_caches(struct mem_cgroup *memcg)
609{
610 LIST_HEAD(release);
611 bool need_rcu_barrier = false;
612 struct kmem_cache *s, *s2;
613
614 get_online_cpus();
615 get_online_mems();
616
617 mutex_lock(&slab_mutex);
618 list_for_each_entry_safe(s, s2, &slab_caches, list) {
619 if (is_root_cache(s) || s->memcg_params.memcg != memcg)
620 continue;
621 /*
622 * The cgroup is about to be freed and therefore has no charges
623 * left. Hence, all its caches must be empty by now.
624 */
625 BUG_ON(__shutdown_memcg_cache(s, &release, &need_rcu_barrier));
626 }
627 mutex_unlock(&slab_mutex);
628
629 put_online_mems();
630 put_online_cpus();
631
632 release_caches(&release, need_rcu_barrier);
633}
634
635static int shutdown_memcg_caches(struct kmem_cache *s,
636 struct list_head *release, bool *need_rcu_barrier)
637{
638 struct memcg_cache_array *arr;
639 struct kmem_cache *c, *c2;
640 LIST_HEAD(busy);
641 int i;
642
643 BUG_ON(!is_root_cache(s));
644
645 /*
646 * First, shutdown active caches, i.e. caches that belong to online
647 * memory cgroups.
648 */
649 arr = rcu_dereference_protected(s->memcg_params.memcg_caches,
650 lockdep_is_held(&slab_mutex));
651 for_each_memcg_cache_index(i) {
652 c = arr->entries[i];
653 if (!c)
654 continue;
655 if (__shutdown_memcg_cache(c, release, need_rcu_barrier))
656 /*
657 * The cache still has objects. Move it to a temporary
658 * list so as not to try to destroy it for a second
659 * time while iterating over inactive caches below.
660 */
661 list_move(&c->memcg_params.list, &busy);
662 else
663 /*
664 * The cache is empty and will be destroyed soon. Clear
665 * the pointer to it in the memcg_caches array so that
666 * it will never be accessed even if the root cache
667 * stays alive.
668 */
669 arr->entries[i] = NULL;
670 }
671
672 /*
673 * Second, shutdown all caches left from memory cgroups that are now
674 * offline.
675 */
676 list_for_each_entry_safe(c, c2, &s->memcg_params.list,
677 memcg_params.list)
678 __shutdown_memcg_cache(c, release, need_rcu_barrier);
679
680 list_splice(&busy, &s->memcg_params.list);
681
682 /*
683 * A cache being destroyed must be empty. In particular, this means
684 * that all per memcg caches attached to it must be empty too.
685 */
686 if (!list_empty(&s->memcg_params.list))
687 return -EBUSY;
688 return 0;
689}
690#else
691static inline int shutdown_memcg_caches(struct kmem_cache *s,
692 struct list_head *release, bool *need_rcu_barrier)
693{
694 return 0;
695}
696#endif /* CONFIG_MEMCG && !CONFIG_SLOB */
697
698void slab_kmem_cache_release(struct kmem_cache *s)
699{
700 __kmem_cache_release(s);
701 destroy_memcg_params(s);
702 kfree_const(s->name);
703 kmem_cache_free(kmem_cache, s);
704}
705
706void kmem_cache_destroy(struct kmem_cache *s)
707{
708 LIST_HEAD(release);
709 bool need_rcu_barrier = false;
710 int err;
711
712 if (unlikely(!s))
713 return;
714
715 get_online_cpus();
716 get_online_mems();
717
718 mutex_lock(&slab_mutex);
719
720 s->refcount--;
721 if (s->refcount)
722 goto out_unlock;
723
724 err = shutdown_memcg_caches(s, &release, &need_rcu_barrier);
725 if (!err)
726 err = shutdown_cache(s, &release, &need_rcu_barrier);
727
728 if (err) {
729 pr_err("kmem_cache_destroy %s: Slab cache still has objects\n",
730 s->name);
731 dump_stack();
732 }
733out_unlock:
734 mutex_unlock(&slab_mutex);
735
736 put_online_mems();
737 put_online_cpus();
738
739 release_caches(&release, need_rcu_barrier);
740}
741EXPORT_SYMBOL(kmem_cache_destroy);
742
743/**
744 * kmem_cache_shrink - Shrink a cache.
745 * @cachep: The cache to shrink.
746 *
747 * Releases as many slabs as possible for a cache.
748 * To help debugging, a zero exit status indicates all slabs were released.
749 */
750int kmem_cache_shrink(struct kmem_cache *cachep)
751{
752 int ret;
753
754 get_online_cpus();
755 get_online_mems();
756 ret = __kmem_cache_shrink(cachep, false);
757 put_online_mems();
758 put_online_cpus();
759 return ret;
760}
761EXPORT_SYMBOL(kmem_cache_shrink);
762
763bool slab_is_available(void)
764{
765 return slab_state >= UP;
766}
767
768#ifndef CONFIG_SLOB
769/* Create a cache during boot when no slab services are available yet */
770void __init create_boot_cache(struct kmem_cache *s, const char *name, size_t size,
771 unsigned long flags)
772{
773 int err;
774
775 s->name = name;
776 s->size = s->object_size = size;
777 s->align = calculate_alignment(flags, ARCH_KMALLOC_MINALIGN, size);
778
779 slab_init_memcg_params(s);
780
781 err = __kmem_cache_create(s, flags);
782
783 if (err)
784 panic("Creation of kmalloc slab %s size=%zu failed. Reason %d\n",
785 name, size, err);
786
787 s->refcount = -1; /* Exempt from merging for now */
788}
789
790struct kmem_cache *__init create_kmalloc_cache(const char *name, size_t size,
791 unsigned long flags)
792{
793 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
794
795 if (!s)
796 panic("Out of memory when creating slab %s\n", name);
797
798 create_boot_cache(s, name, size, flags);
799 list_add(&s->list, &slab_caches);
800 s->refcount = 1;
801 return s;
802}
803
804struct kmem_cache *kmalloc_caches[KMALLOC_SHIFT_HIGH + 1];
805EXPORT_SYMBOL(kmalloc_caches);
806
807#ifdef CONFIG_ZONE_DMA
808struct kmem_cache *kmalloc_dma_caches[KMALLOC_SHIFT_HIGH + 1];
809EXPORT_SYMBOL(kmalloc_dma_caches);
810#endif
811
812/*
813 * Conversion table for small slabs sizes / 8 to the index in the
814 * kmalloc array. This is necessary for slabs < 192 since we have non power
815 * of two cache sizes there. The size of larger slabs can be determined using
816 * fls.
817 */
818static s8 size_index[24] = {
819 3, /* 8 */
820 4, /* 16 */
821 5, /* 24 */
822 5, /* 32 */
823 6, /* 40 */
824 6, /* 48 */
825 6, /* 56 */
826 6, /* 64 */
827 1, /* 72 */
828 1, /* 80 */
829 1, /* 88 */
830 1, /* 96 */
831 7, /* 104 */
832 7, /* 112 */
833 7, /* 120 */
834 7, /* 128 */
835 2, /* 136 */
836 2, /* 144 */
837 2, /* 152 */
838 2, /* 160 */
839 2, /* 168 */
840 2, /* 176 */
841 2, /* 184 */
842 2 /* 192 */
843};
844
845static inline int size_index_elem(size_t bytes)
846{
847 return (bytes - 1) / 8;
848}
849
850/*
851 * Find the kmem_cache structure that serves a given size of
852 * allocation
853 */
854struct kmem_cache *kmalloc_slab(size_t size, gfp_t flags)
855{
856 int index;
857
858 if (unlikely(size > KMALLOC_MAX_SIZE)) {
859 WARN_ON_ONCE(!(flags & __GFP_NOWARN));
860 return NULL;
861 }
862
863 if (size <= 192) {
864 if (!size)
865 return ZERO_SIZE_PTR;
866
867 index = size_index[size_index_elem(size)];
868 } else
869 index = fls(size - 1);
870
871#ifdef CONFIG_ZONE_DMA
872 if (unlikely((flags & GFP_DMA)))
873 return kmalloc_dma_caches[index];
874
875#endif
876 return kmalloc_caches[index];
877}
878
879/*
880 * kmalloc_info[] is to make slub_debug=,kmalloc-xx option work at boot time.
881 * kmalloc_index() supports up to 2^26=64MB, so the final entry of the table is
882 * kmalloc-67108864.
883 */
884static struct {
885 const char *name;
886 unsigned long size;
887} const kmalloc_info[] __initconst = {
888 {NULL, 0}, {"kmalloc-96", 96},
889 {"kmalloc-192", 192}, {"kmalloc-8", 8},
890 {"kmalloc-16", 16}, {"kmalloc-32", 32},
891 {"kmalloc-64", 64}, {"kmalloc-128", 128},
892 {"kmalloc-256", 256}, {"kmalloc-512", 512},
893 {"kmalloc-1024", 1024}, {"kmalloc-2048", 2048},
894 {"kmalloc-4096", 4096}, {"kmalloc-8192", 8192},
895 {"kmalloc-16384", 16384}, {"kmalloc-32768", 32768},
896 {"kmalloc-65536", 65536}, {"kmalloc-131072", 131072},
897 {"kmalloc-262144", 262144}, {"kmalloc-524288", 524288},
898 {"kmalloc-1048576", 1048576}, {"kmalloc-2097152", 2097152},
899 {"kmalloc-4194304", 4194304}, {"kmalloc-8388608", 8388608},
900 {"kmalloc-16777216", 16777216}, {"kmalloc-33554432", 33554432},
901 {"kmalloc-67108864", 67108864}
902};
903
904/*
905 * Patch up the size_index table if we have strange large alignment
906 * requirements for the kmalloc array. This is only the case for
907 * MIPS it seems. The standard arches will not generate any code here.
908 *
909 * Largest permitted alignment is 256 bytes due to the way we
910 * handle the index determination for the smaller caches.
911 *
912 * Make sure that nothing crazy happens if someone starts tinkering
913 * around with ARCH_KMALLOC_MINALIGN
914 */
915void __init setup_kmalloc_cache_index_table(void)
916{
917 int i;
918
919 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
920 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
921
922 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
923 int elem = size_index_elem(i);
924
925 if (elem >= ARRAY_SIZE(size_index))
926 break;
927 size_index[elem] = KMALLOC_SHIFT_LOW;
928 }
929
930 if (KMALLOC_MIN_SIZE >= 64) {
931 /*
932 * The 96 byte size cache is not used if the alignment
933 * is 64 byte.
934 */
935 for (i = 64 + 8; i <= 96; i += 8)
936 size_index[size_index_elem(i)] = 7;
937
938 }
939
940 if (KMALLOC_MIN_SIZE >= 128) {
941 /*
942 * The 192 byte sized cache is not used if the alignment
943 * is 128 byte. Redirect kmalloc to use the 256 byte cache
944 * instead.
945 */
946 for (i = 128 + 8; i <= 192; i += 8)
947 size_index[size_index_elem(i)] = 8;
948 }
949}
950
951static void __init new_kmalloc_cache(int idx, unsigned long flags)
952{
953 kmalloc_caches[idx] = create_kmalloc_cache(kmalloc_info[idx].name,
954 kmalloc_info[idx].size, flags);
955}
956
957/*
958 * Create the kmalloc array. Some of the regular kmalloc arrays
959 * may already have been created because they were needed to
960 * enable allocations for slab creation.
961 */
962void __init create_kmalloc_caches(unsigned long flags)
963{
964 int i;
965
966 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) {
967 if (!kmalloc_caches[i])
968 new_kmalloc_cache(i, flags);
969
970 /*
971 * Caches that are not of the two-to-the-power-of size.
972 * These have to be created immediately after the
973 * earlier power of two caches
974 */
975 if (KMALLOC_MIN_SIZE <= 32 && !kmalloc_caches[1] && i == 6)
976 new_kmalloc_cache(1, flags);
977 if (KMALLOC_MIN_SIZE <= 64 && !kmalloc_caches[2] && i == 7)
978 new_kmalloc_cache(2, flags);
979 }
980
981 /* Kmalloc array is now usable */
982 slab_state = UP;
983
984#ifdef CONFIG_ZONE_DMA
985 for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
986 struct kmem_cache *s = kmalloc_caches[i];
987
988 if (s) {
989 int size = kmalloc_size(i);
990 char *n = kasprintf(GFP_NOWAIT,
991 "dma-kmalloc-%d", size);
992
993 BUG_ON(!n);
994 kmalloc_dma_caches[i] = create_kmalloc_cache(n,
995 size, SLAB_CACHE_DMA | flags);
996 }
997 }
998#endif
999}
1000#endif /* !CONFIG_SLOB */
1001
1002/*
1003 * To avoid unnecessary overhead, we pass through large allocation requests
1004 * directly to the page allocator. We use __GFP_COMP, because we will need to
1005 * know the allocation order to free the pages properly in kfree.
1006 */
1007void *kmalloc_order(size_t size, gfp_t flags, unsigned int order)
1008{
1009 void *ret;
1010 struct page *page;
1011
1012 flags |= __GFP_COMP;
1013 page = alloc_kmem_pages(flags, order);
1014 ret = page ? page_address(page) : NULL;
1015 kmemleak_alloc(ret, size, 1, flags);
1016 kasan_kmalloc_large(ret, size, flags);
1017 return ret;
1018}
1019EXPORT_SYMBOL(kmalloc_order);
1020
1021#ifdef CONFIG_TRACING
1022void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
1023{
1024 void *ret = kmalloc_order(size, flags, order);
1025 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
1026 return ret;
1027}
1028EXPORT_SYMBOL(kmalloc_order_trace);
1029#endif
1030
1031#ifdef CONFIG_SLABINFO
1032
1033#ifdef CONFIG_SLAB
1034#define SLABINFO_RIGHTS (S_IWUSR | S_IRUSR)
1035#else
1036#define SLABINFO_RIGHTS S_IRUSR
1037#endif
1038
1039static void print_slabinfo_header(struct seq_file *m)
1040{
1041 /*
1042 * Output format version, so at least we can change it
1043 * without _too_ many complaints.
1044 */
1045#ifdef CONFIG_DEBUG_SLAB
1046 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
1047#else
1048 seq_puts(m, "slabinfo - version: 2.1\n");
1049#endif
1050 seq_puts(m, "# name <active_objs> <num_objs> <objsize> <objperslab> <pagesperslab>");
1051 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
1052 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
1053#ifdef CONFIG_DEBUG_SLAB
1054 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> <error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
1055 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
1056#endif
1057 seq_putc(m, '\n');
1058}
1059
1060void *slab_start(struct seq_file *m, loff_t *pos)
1061{
1062 mutex_lock(&slab_mutex);
1063 return seq_list_start(&slab_caches, *pos);
1064}
1065
1066void *slab_next(struct seq_file *m, void *p, loff_t *pos)
1067{
1068 return seq_list_next(p, &slab_caches, pos);
1069}
1070
1071void slab_stop(struct seq_file *m, void *p)
1072{
1073 mutex_unlock(&slab_mutex);
1074}
1075
1076static void
1077memcg_accumulate_slabinfo(struct kmem_cache *s, struct slabinfo *info)
1078{
1079 struct kmem_cache *c;
1080 struct slabinfo sinfo;
1081
1082 if (!is_root_cache(s))
1083 return;
1084
1085 for_each_memcg_cache(c, s) {
1086 memset(&sinfo, 0, sizeof(sinfo));
1087 get_slabinfo(c, &sinfo);
1088
1089 info->active_slabs += sinfo.active_slabs;
1090 info->num_slabs += sinfo.num_slabs;
1091 info->shared_avail += sinfo.shared_avail;
1092 info->active_objs += sinfo.active_objs;
1093 info->num_objs += sinfo.num_objs;
1094 }
1095}
1096
1097static void cache_show(struct kmem_cache *s, struct seq_file *m)
1098{
1099 struct slabinfo sinfo;
1100
1101 memset(&sinfo, 0, sizeof(sinfo));
1102 get_slabinfo(s, &sinfo);
1103
1104 memcg_accumulate_slabinfo(s, &sinfo);
1105
1106 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
1107 cache_name(s), sinfo.active_objs, sinfo.num_objs, s->size,
1108 sinfo.objects_per_slab, (1 << sinfo.cache_order));
1109
1110 seq_printf(m, " : tunables %4u %4u %4u",
1111 sinfo.limit, sinfo.batchcount, sinfo.shared);
1112 seq_printf(m, " : slabdata %6lu %6lu %6lu",
1113 sinfo.active_slabs, sinfo.num_slabs, sinfo.shared_avail);
1114 slabinfo_show_stats(m, s);
1115 seq_putc(m, '\n');
1116}
1117
1118static int slab_show(struct seq_file *m, void *p)
1119{
1120 struct kmem_cache *s = list_entry(p, struct kmem_cache, list);
1121
1122 if (p == slab_caches.next)
1123 print_slabinfo_header(m);
1124 if (is_root_cache(s))
1125 cache_show(s, m);
1126 return 0;
1127}
1128
1129#if defined(CONFIG_MEMCG) && !defined(CONFIG_SLOB)
1130int memcg_slab_show(struct seq_file *m, void *p)
1131{
1132 struct kmem_cache *s = list_entry(p, struct kmem_cache, list);
1133 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
1134
1135 if (p == slab_caches.next)
1136 print_slabinfo_header(m);
1137 if (!is_root_cache(s) && s->memcg_params.memcg == memcg)
1138 cache_show(s, m);
1139 return 0;
1140}
1141#endif
1142
1143/*
1144 * slabinfo_op - iterator that generates /proc/slabinfo
1145 *
1146 * Output layout:
1147 * cache-name
1148 * num-active-objs
1149 * total-objs
1150 * object size
1151 * num-active-slabs
1152 * total-slabs
1153 * num-pages-per-slab
1154 * + further values on SMP and with statistics enabled
1155 */
1156static const struct seq_operations slabinfo_op = {
1157 .start = slab_start,
1158 .next = slab_next,
1159 .stop = slab_stop,
1160 .show = slab_show,
1161};
1162
1163static int slabinfo_open(struct inode *inode, struct file *file)
1164{
1165 return seq_open(file, &slabinfo_op);
1166}
1167
1168static const struct file_operations proc_slabinfo_operations = {
1169 .open = slabinfo_open,
1170 .read = seq_read,
1171 .write = slabinfo_write,
1172 .llseek = seq_lseek,
1173 .release = seq_release,
1174};
1175
1176static int __init slab_proc_init(void)
1177{
1178 proc_create("slabinfo", SLABINFO_RIGHTS, NULL,
1179 &proc_slabinfo_operations);
1180 return 0;
1181}
1182module_init(slab_proc_init);
1183#endif /* CONFIG_SLABINFO */
1184
1185static __always_inline void *__do_krealloc(const void *p, size_t new_size,
1186 gfp_t flags)
1187{
1188 void *ret;
1189 size_t ks = 0;
1190
1191 if (p)
1192 ks = ksize(p);
1193
1194 if (ks >= new_size) {
1195 kasan_krealloc((void *)p, new_size, flags);
1196 return (void *)p;
1197 }
1198
1199 ret = kmalloc_track_caller(new_size, flags);
1200 if (ret && p)
1201 memcpy(ret, p, ks);
1202
1203 return ret;
1204}
1205
1206/**
1207 * __krealloc - like krealloc() but don't free @p.
1208 * @p: object to reallocate memory for.
1209 * @new_size: how many bytes of memory are required.
1210 * @flags: the type of memory to allocate.
1211 *
1212 * This function is like krealloc() except it never frees the originally
1213 * allocated buffer. Use this if you don't want to free the buffer immediately
1214 * like, for example, with RCU.
1215 */
1216void *__krealloc(const void *p, size_t new_size, gfp_t flags)
1217{
1218 if (unlikely(!new_size))
1219 return ZERO_SIZE_PTR;
1220
1221 return __do_krealloc(p, new_size, flags);
1222
1223}
1224EXPORT_SYMBOL(__krealloc);
1225
1226/**
1227 * krealloc - reallocate memory. The contents will remain unchanged.
1228 * @p: object to reallocate memory for.
1229 * @new_size: how many bytes of memory are required.
1230 * @flags: the type of memory to allocate.
1231 *
1232 * The contents of the object pointed to are preserved up to the
1233 * lesser of the new and old sizes. If @p is %NULL, krealloc()
1234 * behaves exactly like kmalloc(). If @new_size is 0 and @p is not a
1235 * %NULL pointer, the object pointed to is freed.
1236 */
1237void *krealloc(const void *p, size_t new_size, gfp_t flags)
1238{
1239 void *ret;
1240
1241 if (unlikely(!new_size)) {
1242 kfree(p);
1243 return ZERO_SIZE_PTR;
1244 }
1245
1246 ret = __do_krealloc(p, new_size, flags);
1247 if (ret && p != ret)
1248 kfree(p);
1249
1250 return ret;
1251}
1252EXPORT_SYMBOL(krealloc);
1253
1254/**
1255 * kzfree - like kfree but zero memory
1256 * @p: object to free memory of
1257 *
1258 * The memory of the object @p points to is zeroed before freed.
1259 * If @p is %NULL, kzfree() does nothing.
1260 *
1261 * Note: this function zeroes the whole allocated buffer which can be a good
1262 * deal bigger than the requested buffer size passed to kmalloc(). So be
1263 * careful when using this function in performance sensitive code.
1264 */
1265void kzfree(const void *p)
1266{
1267 size_t ks;
1268 void *mem = (void *)p;
1269
1270 if (unlikely(ZERO_OR_NULL_PTR(mem)))
1271 return;
1272 ks = ksize(mem);
1273 memset(mem, 0, ks);
1274 kfree(mem);
1275}
1276EXPORT_SYMBOL(kzfree);
1277
1278/* Tracepoints definitions. */
1279EXPORT_TRACEPOINT_SYMBOL(kmalloc);
1280EXPORT_TRACEPOINT_SYMBOL(kmem_cache_alloc);
1281EXPORT_TRACEPOINT_SYMBOL(kmalloc_node);
1282EXPORT_TRACEPOINT_SYMBOL(kmem_cache_alloc_node);
1283EXPORT_TRACEPOINT_SYMBOL(kfree);
1284EXPORT_TRACEPOINT_SYMBOL(kmem_cache_free);
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 */