1 // SPDX-License-Identifier: GPL-2.0
3 * Slab allocator functions that are independent of the allocator strategy
5 * (C) 2012 Christoph Lameter <cl@linux.com>
7 #include <linux/slab.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 <asm/cacheflush.h>
21 #include <asm/tlbflush.h>
23 #include <linux/memcontrol.h>
25 #define CREATE_TRACE_POINTS
26 #include <trace/events/kmem.h>
30 enum slab_state slab_state;
31 LIST_HEAD(slab_caches);
32 DEFINE_MUTEX(slab_mutex);
33 struct kmem_cache *kmem_cache;
35 #ifdef CONFIG_HARDENED_USERCOPY
36 bool usercopy_fallback __ro_after_init =
37 IS_ENABLED(CONFIG_HARDENED_USERCOPY_FALLBACK);
38 module_param(usercopy_fallback, bool, 0400);
39 MODULE_PARM_DESC(usercopy_fallback,
40 "WARN instead of reject usercopy whitelist violations");
43 static LIST_HEAD(slab_caches_to_rcu_destroy);
44 static void slab_caches_to_rcu_destroy_workfn(struct work_struct *work);
45 static DECLARE_WORK(slab_caches_to_rcu_destroy_work,
46 slab_caches_to_rcu_destroy_workfn);
49 * Set of flags that will prevent slab merging
51 #define SLAB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
52 SLAB_TRACE | SLAB_TYPESAFE_BY_RCU | SLAB_NOLEAKTRACE | \
53 SLAB_FAILSLAB | SLAB_KASAN)
55 #define SLAB_MERGE_SAME (SLAB_RECLAIM_ACCOUNT | SLAB_CACHE_DMA | \
59 * Merge control. If this is set then no merging of slab caches will occur.
61 static bool slab_nomerge = !IS_ENABLED(CONFIG_SLAB_MERGE_DEFAULT);
63 static int __init setup_slab_nomerge(char *str)
70 __setup_param("slub_nomerge", slub_nomerge, setup_slab_nomerge, 0);
73 __setup("slab_nomerge", setup_slab_nomerge);
76 * Determine the size of a slab object
78 unsigned int kmem_cache_size(struct kmem_cache *s)
80 return s->object_size;
82 EXPORT_SYMBOL(kmem_cache_size);
84 #ifdef CONFIG_DEBUG_VM
85 static int kmem_cache_sanity_check(const char *name, unsigned int size)
87 if (!name || in_interrupt() || size < sizeof(void *) ||
88 size > KMALLOC_MAX_SIZE) {
89 pr_err("kmem_cache_create(%s) integrity check failed\n", name);
93 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
97 static inline int kmem_cache_sanity_check(const char *name, unsigned int size)
103 void __kmem_cache_free_bulk(struct kmem_cache *s, size_t nr, void **p)
107 for (i = 0; i < nr; i++) {
109 kmem_cache_free(s, p[i]);
115 int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t nr,
120 for (i = 0; i < nr; i++) {
121 void *x = p[i] = kmem_cache_alloc(s, flags);
123 __kmem_cache_free_bulk(s, i, p);
130 #if defined(CONFIG_MEMCG) && !defined(CONFIG_SLOB)
132 LIST_HEAD(slab_root_caches);
134 void slab_init_memcg_params(struct kmem_cache *s)
136 s->memcg_params.root_cache = NULL;
137 RCU_INIT_POINTER(s->memcg_params.memcg_caches, NULL);
138 INIT_LIST_HEAD(&s->memcg_params.children);
139 s->memcg_params.dying = false;
142 static int init_memcg_params(struct kmem_cache *s,
143 struct mem_cgroup *memcg, struct kmem_cache *root_cache)
145 struct memcg_cache_array *arr;
148 s->memcg_params.root_cache = root_cache;
149 s->memcg_params.memcg = memcg;
150 INIT_LIST_HEAD(&s->memcg_params.children_node);
151 INIT_LIST_HEAD(&s->memcg_params.kmem_caches_node);
155 slab_init_memcg_params(s);
157 if (!memcg_nr_cache_ids)
160 arr = kvzalloc(sizeof(struct memcg_cache_array) +
161 memcg_nr_cache_ids * sizeof(void *),
166 RCU_INIT_POINTER(s->memcg_params.memcg_caches, arr);
170 static void destroy_memcg_params(struct kmem_cache *s)
172 if (is_root_cache(s))
173 kvfree(rcu_access_pointer(s->memcg_params.memcg_caches));
176 static void free_memcg_params(struct rcu_head *rcu)
178 struct memcg_cache_array *old;
180 old = container_of(rcu, struct memcg_cache_array, rcu);
184 static int update_memcg_params(struct kmem_cache *s, int new_array_size)
186 struct memcg_cache_array *old, *new;
188 new = kvzalloc(sizeof(struct memcg_cache_array) +
189 new_array_size * sizeof(void *), GFP_KERNEL);
193 old = rcu_dereference_protected(s->memcg_params.memcg_caches,
194 lockdep_is_held(&slab_mutex));
196 memcpy(new->entries, old->entries,
197 memcg_nr_cache_ids * sizeof(void *));
199 rcu_assign_pointer(s->memcg_params.memcg_caches, new);
201 call_rcu(&old->rcu, free_memcg_params);
205 int memcg_update_all_caches(int num_memcgs)
207 struct kmem_cache *s;
210 mutex_lock(&slab_mutex);
211 list_for_each_entry(s, &slab_root_caches, root_caches_node) {
212 ret = update_memcg_params(s, num_memcgs);
214 * Instead of freeing the memory, we'll just leave the caches
215 * up to this point in an updated state.
220 mutex_unlock(&slab_mutex);
224 void memcg_link_cache(struct kmem_cache *s)
226 if (is_root_cache(s)) {
227 list_add(&s->root_caches_node, &slab_root_caches);
229 list_add(&s->memcg_params.children_node,
230 &s->memcg_params.root_cache->memcg_params.children);
231 list_add(&s->memcg_params.kmem_caches_node,
232 &s->memcg_params.memcg->kmem_caches);
236 static void memcg_unlink_cache(struct kmem_cache *s)
238 if (is_root_cache(s)) {
239 list_del(&s->root_caches_node);
241 list_del(&s->memcg_params.children_node);
242 list_del(&s->memcg_params.kmem_caches_node);
246 static inline int init_memcg_params(struct kmem_cache *s,
247 struct mem_cgroup *memcg, struct kmem_cache *root_cache)
252 static inline void destroy_memcg_params(struct kmem_cache *s)
256 static inline void memcg_unlink_cache(struct kmem_cache *s)
259 #endif /* CONFIG_MEMCG && !CONFIG_SLOB */
262 * Figure out what the alignment of the objects will be given a set of
263 * flags, a user specified alignment and the size of the objects.
265 static unsigned int calculate_alignment(slab_flags_t flags,
266 unsigned int align, unsigned int size)
269 * If the user wants hardware cache aligned objects then follow that
270 * suggestion if the object is sufficiently large.
272 * The hardware cache alignment cannot override the specified
273 * alignment though. If that is greater then use it.
275 if (flags & SLAB_HWCACHE_ALIGN) {
278 ralign = cache_line_size();
279 while (size <= ralign / 2)
281 align = max(align, ralign);
284 if (align < ARCH_SLAB_MINALIGN)
285 align = ARCH_SLAB_MINALIGN;
287 return ALIGN(align, sizeof(void *));
291 * Find a mergeable slab cache
293 int slab_unmergeable(struct kmem_cache *s)
295 if (slab_nomerge || (s->flags & SLAB_NEVER_MERGE))
298 if (!is_root_cache(s))
308 * We may have set a slab to be unmergeable during bootstrap.
316 struct kmem_cache *find_mergeable(unsigned int size, unsigned int align,
317 slab_flags_t flags, const char *name, void (*ctor)(void *))
319 struct kmem_cache *s;
327 size = ALIGN(size, sizeof(void *));
328 align = calculate_alignment(flags, align, size);
329 size = ALIGN(size, align);
330 flags = kmem_cache_flags(size, flags, name, NULL);
332 if (flags & SLAB_NEVER_MERGE)
335 list_for_each_entry_reverse(s, &slab_root_caches, root_caches_node) {
336 if (slab_unmergeable(s))
342 if ((flags & SLAB_MERGE_SAME) != (s->flags & SLAB_MERGE_SAME))
345 * Check if alignment is compatible.
346 * Courtesy of Adrian Drzewiecki
348 if ((s->size & ~(align - 1)) != s->size)
351 if (s->size - size >= sizeof(void *))
354 if (IS_ENABLED(CONFIG_SLAB) && align &&
355 (align > s->align || s->align % align))
363 static struct kmem_cache *create_cache(const char *name,
364 unsigned int object_size, unsigned int align,
365 slab_flags_t flags, unsigned int useroffset,
366 unsigned int usersize, void (*ctor)(void *),
367 struct mem_cgroup *memcg, struct kmem_cache *root_cache)
369 struct kmem_cache *s;
372 if (WARN_ON(useroffset + usersize > object_size))
373 useroffset = usersize = 0;
376 s = kmem_cache_zalloc(kmem_cache, GFP_KERNEL);
381 s->size = s->object_size = object_size;
384 s->useroffset = useroffset;
385 s->usersize = usersize;
387 err = init_memcg_params(s, memcg, root_cache);
391 err = __kmem_cache_create(s, flags);
396 list_add(&s->list, &slab_caches);
404 destroy_memcg_params(s);
405 kmem_cache_free(kmem_cache, s);
410 * kmem_cache_create_usercopy - Create a cache.
411 * @name: A string which is used in /proc/slabinfo to identify this cache.
412 * @size: The size of objects to be created in this cache.
413 * @align: The required alignment for the objects.
415 * @useroffset: Usercopy region offset
416 * @usersize: Usercopy region size
417 * @ctor: A constructor for the objects.
419 * Returns a ptr to the cache on success, NULL on failure.
420 * Cannot be called within a interrupt, but can be interrupted.
421 * The @ctor is run when new pages are allocated by the cache.
425 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
426 * to catch references to uninitialised memory.
428 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
429 * for buffer overruns.
431 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
432 * cacheline. This can be beneficial if you're counting cycles as closely
436 kmem_cache_create_usercopy(const char *name,
437 unsigned int size, unsigned int align,
439 unsigned int useroffset, unsigned int usersize,
440 void (*ctor)(void *))
442 struct kmem_cache *s = NULL;
443 const char *cache_name;
448 memcg_get_cache_ids();
450 mutex_lock(&slab_mutex);
452 err = kmem_cache_sanity_check(name, size);
457 /* Refuse requests with allocator specific flags */
458 if (flags & ~SLAB_FLAGS_PERMITTED) {
464 * Some allocators will constraint the set of valid flags to a subset
465 * of all flags. We expect them to define CACHE_CREATE_MASK in this
466 * case, and we'll just provide them with a sanitized version of the
469 flags &= CACHE_CREATE_MASK;
471 /* Fail closed on bad usersize of useroffset values. */
472 if (WARN_ON(!usersize && useroffset) ||
473 WARN_ON(size < usersize || size - usersize < useroffset))
474 usersize = useroffset = 0;
477 s = __kmem_cache_alias(name, size, align, flags, ctor);
481 cache_name = kstrdup_const(name, GFP_KERNEL);
487 s = create_cache(cache_name, size,
488 calculate_alignment(flags, align, size),
489 flags, useroffset, usersize, ctor, NULL, NULL);
492 kfree_const(cache_name);
496 mutex_unlock(&slab_mutex);
498 memcg_put_cache_ids();
503 if (flags & SLAB_PANIC)
504 panic("kmem_cache_create: Failed to create slab '%s'. Error %d\n",
507 pr_warn("kmem_cache_create(%s) failed with error %d\n",
515 EXPORT_SYMBOL(kmem_cache_create_usercopy);
518 kmem_cache_create(const char *name, unsigned int size, unsigned int align,
519 slab_flags_t flags, void (*ctor)(void *))
521 return kmem_cache_create_usercopy(name, size, align, flags, 0, 0,
524 EXPORT_SYMBOL(kmem_cache_create);
526 static void slab_caches_to_rcu_destroy_workfn(struct work_struct *work)
528 LIST_HEAD(to_destroy);
529 struct kmem_cache *s, *s2;
532 * On destruction, SLAB_TYPESAFE_BY_RCU kmem_caches are put on the
533 * @slab_caches_to_rcu_destroy list. The slab pages are freed
534 * through RCU and and the associated kmem_cache are dereferenced
535 * while freeing the pages, so the kmem_caches should be freed only
536 * after the pending RCU operations are finished. As rcu_barrier()
537 * is a pretty slow operation, we batch all pending destructions
540 mutex_lock(&slab_mutex);
541 list_splice_init(&slab_caches_to_rcu_destroy, &to_destroy);
542 mutex_unlock(&slab_mutex);
544 if (list_empty(&to_destroy))
549 list_for_each_entry_safe(s, s2, &to_destroy, list) {
550 #ifdef SLAB_SUPPORTS_SYSFS
551 sysfs_slab_release(s);
553 slab_kmem_cache_release(s);
558 static int shutdown_cache(struct kmem_cache *s)
560 /* free asan quarantined objects */
561 kasan_cache_shutdown(s);
563 if (__kmem_cache_shutdown(s) != 0)
566 memcg_unlink_cache(s);
569 if (s->flags & SLAB_TYPESAFE_BY_RCU) {
570 #ifdef SLAB_SUPPORTS_SYSFS
571 sysfs_slab_unlink(s);
573 list_add_tail(&s->list, &slab_caches_to_rcu_destroy);
574 schedule_work(&slab_caches_to_rcu_destroy_work);
576 #ifdef SLAB_SUPPORTS_SYSFS
577 sysfs_slab_unlink(s);
578 sysfs_slab_release(s);
580 slab_kmem_cache_release(s);
587 #if defined(CONFIG_MEMCG) && !defined(CONFIG_SLOB)
589 * memcg_create_kmem_cache - Create a cache for a memory cgroup.
590 * @memcg: The memory cgroup the new cache is for.
591 * @root_cache: The parent of the new cache.
593 * This function attempts to create a kmem cache that will serve allocation
594 * requests going from @memcg to @root_cache. The new cache inherits properties
597 void memcg_create_kmem_cache(struct mem_cgroup *memcg,
598 struct kmem_cache *root_cache)
600 static char memcg_name_buf[NAME_MAX + 1]; /* protected by slab_mutex */
601 struct cgroup_subsys_state *css = &memcg->css;
602 struct memcg_cache_array *arr;
603 struct kmem_cache *s = NULL;
610 mutex_lock(&slab_mutex);
613 * The memory cgroup could have been offlined while the cache
614 * creation work was pending.
616 if (memcg->kmem_state != KMEM_ONLINE || root_cache->memcg_params.dying)
619 idx = memcg_cache_id(memcg);
620 arr = rcu_dereference_protected(root_cache->memcg_params.memcg_caches,
621 lockdep_is_held(&slab_mutex));
624 * Since per-memcg caches are created asynchronously on first
625 * allocation (see memcg_kmem_get_cache()), several threads can try to
626 * create the same cache, but only one of them may succeed.
628 if (arr->entries[idx])
631 cgroup_name(css->cgroup, memcg_name_buf, sizeof(memcg_name_buf));
632 cache_name = kasprintf(GFP_KERNEL, "%s(%llu:%s)", root_cache->name,
633 css->serial_nr, memcg_name_buf);
637 s = create_cache(cache_name, root_cache->object_size,
639 root_cache->flags & CACHE_CREATE_MASK,
640 root_cache->useroffset, root_cache->usersize,
641 root_cache->ctor, memcg, root_cache);
643 * If we could not create a memcg cache, do not complain, because
644 * that's not critical at all as we can always proceed with the root
653 * Since readers won't lock (see cache_from_memcg_idx()), we need a
654 * barrier here to ensure nobody will see the kmem_cache partially
658 arr->entries[idx] = s;
661 mutex_unlock(&slab_mutex);
667 static void kmemcg_deactivate_workfn(struct work_struct *work)
669 struct kmem_cache *s = container_of(work, struct kmem_cache,
670 memcg_params.deact_work);
675 mutex_lock(&slab_mutex);
677 s->memcg_params.deact_fn(s);
679 mutex_unlock(&slab_mutex);
684 /* done, put the ref from slab_deactivate_memcg_cache_rcu_sched() */
685 css_put(&s->memcg_params.memcg->css);
688 static void kmemcg_deactivate_rcufn(struct rcu_head *head)
690 struct kmem_cache *s = container_of(head, struct kmem_cache,
691 memcg_params.deact_rcu_head);
694 * We need to grab blocking locks. Bounce to ->deact_work. The
695 * work item shares the space with the RCU head and can't be
696 * initialized eariler.
698 INIT_WORK(&s->memcg_params.deact_work, kmemcg_deactivate_workfn);
699 queue_work(memcg_kmem_cache_wq, &s->memcg_params.deact_work);
703 * slab_deactivate_memcg_cache_rcu_sched - schedule deactivation after a
704 * sched RCU grace period
705 * @s: target kmem_cache
706 * @deact_fn: deactivation function to call
708 * Schedule @deact_fn to be invoked with online cpus, mems and slab_mutex
709 * held after a sched RCU grace period. The slab is guaranteed to stay
710 * alive until @deact_fn is finished. This is to be used from
711 * __kmemcg_cache_deactivate().
713 void slab_deactivate_memcg_cache_rcu_sched(struct kmem_cache *s,
714 void (*deact_fn)(struct kmem_cache *))
716 if (WARN_ON_ONCE(is_root_cache(s)) ||
717 WARN_ON_ONCE(s->memcg_params.deact_fn))
720 if (s->memcg_params.root_cache->memcg_params.dying)
723 /* pin memcg so that @s doesn't get destroyed in the middle */
724 css_get(&s->memcg_params.memcg->css);
726 s->memcg_params.deact_fn = deact_fn;
727 call_rcu_sched(&s->memcg_params.deact_rcu_head, kmemcg_deactivate_rcufn);
730 void memcg_deactivate_kmem_caches(struct mem_cgroup *memcg)
733 struct memcg_cache_array *arr;
734 struct kmem_cache *s, *c;
736 idx = memcg_cache_id(memcg);
741 mutex_lock(&slab_mutex);
742 list_for_each_entry(s, &slab_root_caches, root_caches_node) {
743 arr = rcu_dereference_protected(s->memcg_params.memcg_caches,
744 lockdep_is_held(&slab_mutex));
745 c = arr->entries[idx];
749 __kmemcg_cache_deactivate(c);
750 arr->entries[idx] = NULL;
752 mutex_unlock(&slab_mutex);
758 void memcg_destroy_kmem_caches(struct mem_cgroup *memcg)
760 struct kmem_cache *s, *s2;
765 mutex_lock(&slab_mutex);
766 list_for_each_entry_safe(s, s2, &memcg->kmem_caches,
767 memcg_params.kmem_caches_node) {
769 * The cgroup is about to be freed and therefore has no charges
770 * left. Hence, all its caches must be empty by now.
772 BUG_ON(shutdown_cache(s));
774 mutex_unlock(&slab_mutex);
780 static int shutdown_memcg_caches(struct kmem_cache *s)
782 struct memcg_cache_array *arr;
783 struct kmem_cache *c, *c2;
787 BUG_ON(!is_root_cache(s));
790 * First, shutdown active caches, i.e. caches that belong to online
793 arr = rcu_dereference_protected(s->memcg_params.memcg_caches,
794 lockdep_is_held(&slab_mutex));
795 for_each_memcg_cache_index(i) {
799 if (shutdown_cache(c))
801 * The cache still has objects. Move it to a temporary
802 * list so as not to try to destroy it for a second
803 * time while iterating over inactive caches below.
805 list_move(&c->memcg_params.children_node, &busy);
808 * The cache is empty and will be destroyed soon. Clear
809 * the pointer to it in the memcg_caches array so that
810 * it will never be accessed even if the root cache
813 arr->entries[i] = NULL;
817 * Second, shutdown all caches left from memory cgroups that are now
820 list_for_each_entry_safe(c, c2, &s->memcg_params.children,
821 memcg_params.children_node)
824 list_splice(&busy, &s->memcg_params.children);
827 * A cache being destroyed must be empty. In particular, this means
828 * that all per memcg caches attached to it must be empty too.
830 if (!list_empty(&s->memcg_params.children))
835 static void flush_memcg_workqueue(struct kmem_cache *s)
837 mutex_lock(&slab_mutex);
838 s->memcg_params.dying = true;
839 mutex_unlock(&slab_mutex);
842 * SLUB deactivates the kmem_caches through call_rcu_sched. Make
843 * sure all registered rcu callbacks have been invoked.
845 if (IS_ENABLED(CONFIG_SLUB))
849 * SLAB and SLUB create memcg kmem_caches through workqueue and SLUB
850 * deactivates the memcg kmem_caches through workqueue. Make sure all
851 * previous workitems on workqueue are processed.
853 flush_workqueue(memcg_kmem_cache_wq);
856 static inline int shutdown_memcg_caches(struct kmem_cache *s)
861 static inline void flush_memcg_workqueue(struct kmem_cache *s)
864 #endif /* CONFIG_MEMCG && !CONFIG_SLOB */
866 void slab_kmem_cache_release(struct kmem_cache *s)
868 __kmem_cache_release(s);
869 destroy_memcg_params(s);
870 kfree_const(s->name);
871 kmem_cache_free(kmem_cache, s);
874 void kmem_cache_destroy(struct kmem_cache *s)
881 flush_memcg_workqueue(s);
886 mutex_lock(&slab_mutex);
892 err = shutdown_memcg_caches(s);
894 err = shutdown_cache(s);
897 pr_err("kmem_cache_destroy %s: Slab cache still has objects\n",
902 mutex_unlock(&slab_mutex);
907 EXPORT_SYMBOL(kmem_cache_destroy);
910 * kmem_cache_shrink - Shrink a cache.
911 * @cachep: The cache to shrink.
913 * Releases as many slabs as possible for a cache.
914 * To help debugging, a zero exit status indicates all slabs were released.
916 int kmem_cache_shrink(struct kmem_cache *cachep)
922 kasan_cache_shrink(cachep);
923 ret = __kmem_cache_shrink(cachep);
928 EXPORT_SYMBOL(kmem_cache_shrink);
930 bool slab_is_available(void)
932 return slab_state >= UP;
936 /* Create a cache during boot when no slab services are available yet */
937 void __init create_boot_cache(struct kmem_cache *s, const char *name,
938 unsigned int size, slab_flags_t flags,
939 unsigned int useroffset, unsigned int usersize)
944 s->size = s->object_size = size;
945 s->align = calculate_alignment(flags, ARCH_KMALLOC_MINALIGN, size);
946 s->useroffset = useroffset;
947 s->usersize = usersize;
949 slab_init_memcg_params(s);
951 err = __kmem_cache_create(s, flags);
954 panic("Creation of kmalloc slab %s size=%u failed. Reason %d\n",
957 s->refcount = -1; /* Exempt from merging for now */
960 struct kmem_cache *__init create_kmalloc_cache(const char *name,
961 unsigned int size, slab_flags_t flags,
962 unsigned int useroffset, unsigned int usersize)
964 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
967 panic("Out of memory when creating slab %s\n", name);
969 create_boot_cache(s, name, size, flags, useroffset, usersize);
970 list_add(&s->list, &slab_caches);
976 struct kmem_cache *kmalloc_caches[KMALLOC_SHIFT_HIGH + 1] __ro_after_init;
977 EXPORT_SYMBOL(kmalloc_caches);
979 #ifdef CONFIG_ZONE_DMA
980 struct kmem_cache *kmalloc_dma_caches[KMALLOC_SHIFT_HIGH + 1] __ro_after_init;
981 EXPORT_SYMBOL(kmalloc_dma_caches);
985 * Conversion table for small slabs sizes / 8 to the index in the
986 * kmalloc array. This is necessary for slabs < 192 since we have non power
987 * of two cache sizes there. The size of larger slabs can be determined using
990 static u8 size_index[24] __ro_after_init = {
1017 static inline unsigned int size_index_elem(unsigned int bytes)
1019 return (bytes - 1) / 8;
1023 * Find the kmem_cache structure that serves a given size of
1026 struct kmem_cache *kmalloc_slab(size_t size, gfp_t flags)
1030 if (unlikely(size > KMALLOC_MAX_SIZE)) {
1031 WARN_ON_ONCE(!(flags & __GFP_NOWARN));
1037 return ZERO_SIZE_PTR;
1039 index = size_index[size_index_elem(size)];
1041 index = fls(size - 1);
1043 #ifdef CONFIG_ZONE_DMA
1044 if (unlikely((flags & GFP_DMA)))
1045 return kmalloc_dma_caches[index];
1048 return kmalloc_caches[index];
1052 * kmalloc_info[] is to make slub_debug=,kmalloc-xx option work at boot time.
1053 * kmalloc_index() supports up to 2^26=64MB, so the final entry of the table is
1056 const struct kmalloc_info_struct kmalloc_info[] __initconst = {
1057 {NULL, 0}, {"kmalloc-96", 96},
1058 {"kmalloc-192", 192}, {"kmalloc-8", 8},
1059 {"kmalloc-16", 16}, {"kmalloc-32", 32},
1060 {"kmalloc-64", 64}, {"kmalloc-128", 128},
1061 {"kmalloc-256", 256}, {"kmalloc-512", 512},
1062 {"kmalloc-1024", 1024}, {"kmalloc-2048", 2048},
1063 {"kmalloc-4096", 4096}, {"kmalloc-8192", 8192},
1064 {"kmalloc-16384", 16384}, {"kmalloc-32768", 32768},
1065 {"kmalloc-65536", 65536}, {"kmalloc-131072", 131072},
1066 {"kmalloc-262144", 262144}, {"kmalloc-524288", 524288},
1067 {"kmalloc-1048576", 1048576}, {"kmalloc-2097152", 2097152},
1068 {"kmalloc-4194304", 4194304}, {"kmalloc-8388608", 8388608},
1069 {"kmalloc-16777216", 16777216}, {"kmalloc-33554432", 33554432},
1070 {"kmalloc-67108864", 67108864}
1074 * Patch up the size_index table if we have strange large alignment
1075 * requirements for the kmalloc array. This is only the case for
1076 * MIPS it seems. The standard arches will not generate any code here.
1078 * Largest permitted alignment is 256 bytes due to the way we
1079 * handle the index determination for the smaller caches.
1081 * Make sure that nothing crazy happens if someone starts tinkering
1082 * around with ARCH_KMALLOC_MINALIGN
1084 void __init setup_kmalloc_cache_index_table(void)
1088 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
1089 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
1091 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
1092 unsigned int elem = size_index_elem(i);
1094 if (elem >= ARRAY_SIZE(size_index))
1096 size_index[elem] = KMALLOC_SHIFT_LOW;
1099 if (KMALLOC_MIN_SIZE >= 64) {
1101 * The 96 byte size cache is not used if the alignment
1104 for (i = 64 + 8; i <= 96; i += 8)
1105 size_index[size_index_elem(i)] = 7;
1109 if (KMALLOC_MIN_SIZE >= 128) {
1111 * The 192 byte sized cache is not used if the alignment
1112 * is 128 byte. Redirect kmalloc to use the 256 byte cache
1115 for (i = 128 + 8; i <= 192; i += 8)
1116 size_index[size_index_elem(i)] = 8;
1120 static void __init new_kmalloc_cache(int idx, slab_flags_t flags)
1122 kmalloc_caches[idx] = create_kmalloc_cache(kmalloc_info[idx].name,
1123 kmalloc_info[idx].size, flags, 0,
1124 kmalloc_info[idx].size);
1128 * Create the kmalloc array. Some of the regular kmalloc arrays
1129 * may already have been created because they were needed to
1130 * enable allocations for slab creation.
1132 void __init create_kmalloc_caches(slab_flags_t flags)
1136 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) {
1137 if (!kmalloc_caches[i])
1138 new_kmalloc_cache(i, flags);
1141 * Caches that are not of the two-to-the-power-of size.
1142 * These have to be created immediately after the
1143 * earlier power of two caches
1145 if (KMALLOC_MIN_SIZE <= 32 && !kmalloc_caches[1] && i == 6)
1146 new_kmalloc_cache(1, flags);
1147 if (KMALLOC_MIN_SIZE <= 64 && !kmalloc_caches[2] && i == 7)
1148 new_kmalloc_cache(2, flags);
1151 /* Kmalloc array is now usable */
1154 #ifdef CONFIG_ZONE_DMA
1155 for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
1156 struct kmem_cache *s = kmalloc_caches[i];
1159 unsigned int size = kmalloc_size(i);
1160 char *n = kasprintf(GFP_NOWAIT,
1161 "dma-kmalloc-%u", size);
1164 kmalloc_dma_caches[i] = create_kmalloc_cache(n,
1165 size, SLAB_CACHE_DMA | flags, 0, 0);
1170 #endif /* !CONFIG_SLOB */
1173 * To avoid unnecessary overhead, we pass through large allocation requests
1174 * directly to the page allocator. We use __GFP_COMP, because we will need to
1175 * know the allocation order to free the pages properly in kfree.
1177 void *kmalloc_order(size_t size, gfp_t flags, unsigned int order)
1182 flags |= __GFP_COMP;
1183 page = alloc_pages(flags, order);
1184 ret = page ? page_address(page) : NULL;
1185 kmemleak_alloc(ret, size, 1, flags);
1186 kasan_kmalloc_large(ret, size, flags);
1189 EXPORT_SYMBOL(kmalloc_order);
1191 #ifdef CONFIG_TRACING
1192 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
1194 void *ret = kmalloc_order(size, flags, order);
1195 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
1198 EXPORT_SYMBOL(kmalloc_order_trace);
1201 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1202 /* Randomize a generic freelist */
1203 static void freelist_randomize(struct rnd_state *state, unsigned int *list,
1209 for (i = 0; i < count; i++)
1212 /* Fisher-Yates shuffle */
1213 for (i = count - 1; i > 0; i--) {
1214 rand = prandom_u32_state(state);
1216 swap(list[i], list[rand]);
1220 /* Create a random sequence per cache */
1221 int cache_random_seq_create(struct kmem_cache *cachep, unsigned int count,
1224 struct rnd_state state;
1226 if (count < 2 || cachep->random_seq)
1229 cachep->random_seq = kcalloc(count, sizeof(unsigned int), gfp);
1230 if (!cachep->random_seq)
1233 /* Get best entropy at this stage of boot */
1234 prandom_seed_state(&state, get_random_long());
1236 freelist_randomize(&state, cachep->random_seq, count);
1240 /* Destroy the per-cache random freelist sequence */
1241 void cache_random_seq_destroy(struct kmem_cache *cachep)
1243 kfree(cachep->random_seq);
1244 cachep->random_seq = NULL;
1246 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1248 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
1250 #define SLABINFO_RIGHTS (0600)
1252 #define SLABINFO_RIGHTS (0400)
1255 static void print_slabinfo_header(struct seq_file *m)
1258 * Output format version, so at least we can change it
1259 * without _too_ many complaints.
1261 #ifdef CONFIG_DEBUG_SLAB
1262 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
1264 seq_puts(m, "slabinfo - version: 2.1\n");
1266 seq_puts(m, "# name <active_objs> <num_objs> <objsize> <objperslab> <pagesperslab>");
1267 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
1268 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
1269 #ifdef CONFIG_DEBUG_SLAB
1270 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> <error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
1271 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
1276 void *slab_start(struct seq_file *m, loff_t *pos)
1278 mutex_lock(&slab_mutex);
1279 return seq_list_start(&slab_root_caches, *pos);
1282 void *slab_next(struct seq_file *m, void *p, loff_t *pos)
1284 return seq_list_next(p, &slab_root_caches, pos);
1287 void slab_stop(struct seq_file *m, void *p)
1289 mutex_unlock(&slab_mutex);
1293 memcg_accumulate_slabinfo(struct kmem_cache *s, struct slabinfo *info)
1295 struct kmem_cache *c;
1296 struct slabinfo sinfo;
1298 if (!is_root_cache(s))
1301 for_each_memcg_cache(c, s) {
1302 memset(&sinfo, 0, sizeof(sinfo));
1303 get_slabinfo(c, &sinfo);
1305 info->active_slabs += sinfo.active_slabs;
1306 info->num_slabs += sinfo.num_slabs;
1307 info->shared_avail += sinfo.shared_avail;
1308 info->active_objs += sinfo.active_objs;
1309 info->num_objs += sinfo.num_objs;
1313 static void cache_show(struct kmem_cache *s, struct seq_file *m)
1315 struct slabinfo sinfo;
1317 memset(&sinfo, 0, sizeof(sinfo));
1318 get_slabinfo(s, &sinfo);
1320 memcg_accumulate_slabinfo(s, &sinfo);
1322 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
1323 cache_name(s), sinfo.active_objs, sinfo.num_objs, s->size,
1324 sinfo.objects_per_slab, (1 << sinfo.cache_order));
1326 seq_printf(m, " : tunables %4u %4u %4u",
1327 sinfo.limit, sinfo.batchcount, sinfo.shared);
1328 seq_printf(m, " : slabdata %6lu %6lu %6lu",
1329 sinfo.active_slabs, sinfo.num_slabs, sinfo.shared_avail);
1330 slabinfo_show_stats(m, s);
1334 static int slab_show(struct seq_file *m, void *p)
1336 struct kmem_cache *s = list_entry(p, struct kmem_cache, root_caches_node);
1338 if (p == slab_root_caches.next)
1339 print_slabinfo_header(m);
1344 void dump_unreclaimable_slab(void)
1346 struct kmem_cache *s, *s2;
1347 struct slabinfo sinfo;
1350 * Here acquiring slab_mutex is risky since we don't prefer to get
1351 * sleep in oom path. But, without mutex hold, it may introduce a
1353 * Use mutex_trylock to protect the list traverse, dump nothing
1354 * without acquiring the mutex.
1356 if (!mutex_trylock(&slab_mutex)) {
1357 pr_warn("excessive unreclaimable slab but cannot dump stats\n");
1361 pr_info("Unreclaimable slab info:\n");
1362 pr_info("Name Used Total\n");
1364 list_for_each_entry_safe(s, s2, &slab_caches, list) {
1365 if (!is_root_cache(s) || (s->flags & SLAB_RECLAIM_ACCOUNT))
1368 get_slabinfo(s, &sinfo);
1370 if (sinfo.num_objs > 0)
1371 pr_info("%-17s %10luKB %10luKB\n", cache_name(s),
1372 (sinfo.active_objs * s->size) / 1024,
1373 (sinfo.num_objs * s->size) / 1024);
1375 mutex_unlock(&slab_mutex);
1378 #if defined(CONFIG_MEMCG)
1379 void *memcg_slab_start(struct seq_file *m, loff_t *pos)
1381 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
1383 mutex_lock(&slab_mutex);
1384 return seq_list_start(&memcg->kmem_caches, *pos);
1387 void *memcg_slab_next(struct seq_file *m, void *p, loff_t *pos)
1389 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
1391 return seq_list_next(p, &memcg->kmem_caches, pos);
1394 void memcg_slab_stop(struct seq_file *m, void *p)
1396 mutex_unlock(&slab_mutex);
1399 int memcg_slab_show(struct seq_file *m, void *p)
1401 struct kmem_cache *s = list_entry(p, struct kmem_cache,
1402 memcg_params.kmem_caches_node);
1403 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
1405 if (p == memcg->kmem_caches.next)
1406 print_slabinfo_header(m);
1413 * slabinfo_op - iterator that generates /proc/slabinfo
1422 * num-pages-per-slab
1423 * + further values on SMP and with statistics enabled
1425 static const struct seq_operations slabinfo_op = {
1426 .start = slab_start,
1432 static int slabinfo_open(struct inode *inode, struct file *file)
1434 return seq_open(file, &slabinfo_op);
1437 static const struct file_operations proc_slabinfo_operations = {
1438 .open = slabinfo_open,
1440 .write = slabinfo_write,
1441 .llseek = seq_lseek,
1442 .release = seq_release,
1445 static int __init slab_proc_init(void)
1447 proc_create("slabinfo", SLABINFO_RIGHTS, NULL,
1448 &proc_slabinfo_operations);
1451 module_init(slab_proc_init);
1452 #endif /* CONFIG_SLAB || CONFIG_SLUB_DEBUG */
1454 static __always_inline void *__do_krealloc(const void *p, size_t new_size,
1463 if (ks >= new_size) {
1464 kasan_krealloc((void *)p, new_size, flags);
1468 ret = kmalloc_track_caller(new_size, flags);
1476 * __krealloc - like krealloc() but don't free @p.
1477 * @p: object to reallocate memory for.
1478 * @new_size: how many bytes of memory are required.
1479 * @flags: the type of memory to allocate.
1481 * This function is like krealloc() except it never frees the originally
1482 * allocated buffer. Use this if you don't want to free the buffer immediately
1483 * like, for example, with RCU.
1485 void *__krealloc(const void *p, size_t new_size, gfp_t flags)
1487 if (unlikely(!new_size))
1488 return ZERO_SIZE_PTR;
1490 return __do_krealloc(p, new_size, flags);
1493 EXPORT_SYMBOL(__krealloc);
1496 * krealloc - reallocate memory. The contents will remain unchanged.
1497 * @p: object to reallocate memory for.
1498 * @new_size: how many bytes of memory are required.
1499 * @flags: the type of memory to allocate.
1501 * The contents of the object pointed to are preserved up to the
1502 * lesser of the new and old sizes. If @p is %NULL, krealloc()
1503 * behaves exactly like kmalloc(). If @new_size is 0 and @p is not a
1504 * %NULL pointer, the object pointed to is freed.
1506 void *krealloc(const void *p, size_t new_size, gfp_t flags)
1510 if (unlikely(!new_size)) {
1512 return ZERO_SIZE_PTR;
1515 ret = __do_krealloc(p, new_size, flags);
1516 if (ret && p != ret)
1521 EXPORT_SYMBOL(krealloc);
1524 * kzfree - like kfree but zero memory
1525 * @p: object to free memory of
1527 * The memory of the object @p points to is zeroed before freed.
1528 * If @p is %NULL, kzfree() does nothing.
1530 * Note: this function zeroes the whole allocated buffer which can be a good
1531 * deal bigger than the requested buffer size passed to kmalloc(). So be
1532 * careful when using this function in performance sensitive code.
1534 void kzfree(const void *p)
1537 void *mem = (void *)p;
1539 if (unlikely(ZERO_OR_NULL_PTR(mem)))
1545 EXPORT_SYMBOL(kzfree);
1547 /* Tracepoints definitions. */
1548 EXPORT_TRACEPOINT_SYMBOL(kmalloc);
1549 EXPORT_TRACEPOINT_SYMBOL(kmem_cache_alloc);
1550 EXPORT_TRACEPOINT_SYMBOL(kmalloc_node);
1551 EXPORT_TRACEPOINT_SYMBOL(kmem_cache_alloc_node);
1552 EXPORT_TRACEPOINT_SYMBOL(kfree);
1553 EXPORT_TRACEPOINT_SYMBOL(kmem_cache_free);
1555 int should_failslab(struct kmem_cache *s, gfp_t gfpflags)
1557 if (__should_failslab(s, gfpflags))
1561 ALLOW_ERROR_INJECTION(should_failslab, ERRNO);