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 | \
56 SLAB_CACHE_DMA32 | SLAB_ACCOUNT)
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 #ifdef CONFIG_MEMCG_KMEM
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_KMEM */
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 with a region suitable
411 * for copying to userspace
412 * @name: A string which is used in /proc/slabinfo to identify this cache.
413 * @size: The size of objects to be created in this cache.
414 * @align: The required alignment for the objects.
416 * @useroffset: Usercopy region offset
417 * @usersize: Usercopy region size
418 * @ctor: A constructor for the objects.
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
435 * Return: a pointer to the cache on success, NULL on failure.
438 kmem_cache_create_usercopy(const char *name,
439 unsigned int size, unsigned int align,
441 unsigned int useroffset, unsigned int usersize,
442 void (*ctor)(void *))
444 struct kmem_cache *s = NULL;
445 const char *cache_name;
450 memcg_get_cache_ids();
452 mutex_lock(&slab_mutex);
454 err = kmem_cache_sanity_check(name, size);
459 /* Refuse requests with allocator specific flags */
460 if (flags & ~SLAB_FLAGS_PERMITTED) {
466 * Some allocators will constraint the set of valid flags to a subset
467 * of all flags. We expect them to define CACHE_CREATE_MASK in this
468 * case, and we'll just provide them with a sanitized version of the
471 flags &= CACHE_CREATE_MASK;
473 /* Fail closed on bad usersize of useroffset values. */
474 if (WARN_ON(!usersize && useroffset) ||
475 WARN_ON(size < usersize || size - usersize < useroffset))
476 usersize = useroffset = 0;
479 s = __kmem_cache_alias(name, size, align, flags, ctor);
483 cache_name = kstrdup_const(name, GFP_KERNEL);
489 s = create_cache(cache_name, size,
490 calculate_alignment(flags, align, size),
491 flags, useroffset, usersize, ctor, NULL, NULL);
494 kfree_const(cache_name);
498 mutex_unlock(&slab_mutex);
500 memcg_put_cache_ids();
505 if (flags & SLAB_PANIC)
506 panic("kmem_cache_create: Failed to create slab '%s'. Error %d\n",
509 pr_warn("kmem_cache_create(%s) failed with error %d\n",
517 EXPORT_SYMBOL(kmem_cache_create_usercopy);
520 * kmem_cache_create - Create a cache.
521 * @name: A string which is used in /proc/slabinfo to identify this cache.
522 * @size: The size of objects to be created in this cache.
523 * @align: The required alignment for the objects.
525 * @ctor: A constructor for the objects.
527 * Cannot be called within a interrupt, but can be interrupted.
528 * The @ctor is run when new pages are allocated by the cache.
532 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
533 * to catch references to uninitialised memory.
535 * %SLAB_RED_ZONE - Insert `Red` zones around the allocated memory to check
536 * for buffer overruns.
538 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
539 * cacheline. This can be beneficial if you're counting cycles as closely
542 * Return: a pointer to the cache on success, NULL on failure.
545 kmem_cache_create(const char *name, unsigned int size, unsigned int align,
546 slab_flags_t flags, void (*ctor)(void *))
548 return kmem_cache_create_usercopy(name, size, align, flags, 0, 0,
551 EXPORT_SYMBOL(kmem_cache_create);
553 static void slab_caches_to_rcu_destroy_workfn(struct work_struct *work)
555 LIST_HEAD(to_destroy);
556 struct kmem_cache *s, *s2;
559 * On destruction, SLAB_TYPESAFE_BY_RCU kmem_caches are put on the
560 * @slab_caches_to_rcu_destroy list. The slab pages are freed
561 * through RCU and and the associated kmem_cache are dereferenced
562 * while freeing the pages, so the kmem_caches should be freed only
563 * after the pending RCU operations are finished. As rcu_barrier()
564 * is a pretty slow operation, we batch all pending destructions
567 mutex_lock(&slab_mutex);
568 list_splice_init(&slab_caches_to_rcu_destroy, &to_destroy);
569 mutex_unlock(&slab_mutex);
571 if (list_empty(&to_destroy))
576 list_for_each_entry_safe(s, s2, &to_destroy, list) {
577 #ifdef SLAB_SUPPORTS_SYSFS
578 sysfs_slab_release(s);
580 slab_kmem_cache_release(s);
585 static int shutdown_cache(struct kmem_cache *s)
587 /* free asan quarantined objects */
588 kasan_cache_shutdown(s);
590 if (__kmem_cache_shutdown(s) != 0)
593 memcg_unlink_cache(s);
596 if (s->flags & SLAB_TYPESAFE_BY_RCU) {
597 #ifdef SLAB_SUPPORTS_SYSFS
598 sysfs_slab_unlink(s);
600 list_add_tail(&s->list, &slab_caches_to_rcu_destroy);
601 schedule_work(&slab_caches_to_rcu_destroy_work);
603 #ifdef SLAB_SUPPORTS_SYSFS
604 sysfs_slab_unlink(s);
605 sysfs_slab_release(s);
607 slab_kmem_cache_release(s);
614 #ifdef CONFIG_MEMCG_KMEM
616 * memcg_create_kmem_cache - Create a cache for a memory cgroup.
617 * @memcg: The memory cgroup the new cache is for.
618 * @root_cache: The parent of the new cache.
620 * This function attempts to create a kmem cache that will serve allocation
621 * requests going from @memcg to @root_cache. The new cache inherits properties
624 void memcg_create_kmem_cache(struct mem_cgroup *memcg,
625 struct kmem_cache *root_cache)
627 static char memcg_name_buf[NAME_MAX + 1]; /* protected by slab_mutex */
628 struct cgroup_subsys_state *css = &memcg->css;
629 struct memcg_cache_array *arr;
630 struct kmem_cache *s = NULL;
637 mutex_lock(&slab_mutex);
640 * The memory cgroup could have been offlined while the cache
641 * creation work was pending.
643 if (memcg->kmem_state != KMEM_ONLINE || root_cache->memcg_params.dying)
646 idx = memcg_cache_id(memcg);
647 arr = rcu_dereference_protected(root_cache->memcg_params.memcg_caches,
648 lockdep_is_held(&slab_mutex));
651 * Since per-memcg caches are created asynchronously on first
652 * allocation (see memcg_kmem_get_cache()), several threads can try to
653 * create the same cache, but only one of them may succeed.
655 if (arr->entries[idx])
658 cgroup_name(css->cgroup, memcg_name_buf, sizeof(memcg_name_buf));
659 cache_name = kasprintf(GFP_KERNEL, "%s(%llu:%s)", root_cache->name,
660 css->serial_nr, memcg_name_buf);
664 s = create_cache(cache_name, root_cache->object_size,
666 root_cache->flags & CACHE_CREATE_MASK,
667 root_cache->useroffset, root_cache->usersize,
668 root_cache->ctor, memcg, root_cache);
670 * If we could not create a memcg cache, do not complain, because
671 * that's not critical at all as we can always proceed with the root
680 * Since readers won't lock (see cache_from_memcg_idx()), we need a
681 * barrier here to ensure nobody will see the kmem_cache partially
685 arr->entries[idx] = s;
688 mutex_unlock(&slab_mutex);
694 static void kmemcg_deactivate_workfn(struct work_struct *work)
696 struct kmem_cache *s = container_of(work, struct kmem_cache,
697 memcg_params.deact_work);
702 mutex_lock(&slab_mutex);
704 s->memcg_params.deact_fn(s);
706 mutex_unlock(&slab_mutex);
711 /* done, put the ref from slab_deactivate_memcg_cache_rcu_sched() */
712 css_put(&s->memcg_params.memcg->css);
715 static void kmemcg_deactivate_rcufn(struct rcu_head *head)
717 struct kmem_cache *s = container_of(head, struct kmem_cache,
718 memcg_params.deact_rcu_head);
721 * We need to grab blocking locks. Bounce to ->deact_work. The
722 * work item shares the space with the RCU head and can't be
723 * initialized eariler.
725 INIT_WORK(&s->memcg_params.deact_work, kmemcg_deactivate_workfn);
726 queue_work(memcg_kmem_cache_wq, &s->memcg_params.deact_work);
730 * slab_deactivate_memcg_cache_rcu_sched - schedule deactivation after a
731 * sched RCU grace period
732 * @s: target kmem_cache
733 * @deact_fn: deactivation function to call
735 * Schedule @deact_fn to be invoked with online cpus, mems and slab_mutex
736 * held after a sched RCU grace period. The slab is guaranteed to stay
737 * alive until @deact_fn is finished. This is to be used from
738 * __kmemcg_cache_deactivate().
740 void slab_deactivate_memcg_cache_rcu_sched(struct kmem_cache *s,
741 void (*deact_fn)(struct kmem_cache *))
743 if (WARN_ON_ONCE(is_root_cache(s)) ||
744 WARN_ON_ONCE(s->memcg_params.deact_fn))
747 if (s->memcg_params.root_cache->memcg_params.dying)
750 /* pin memcg so that @s doesn't get destroyed in the middle */
751 css_get(&s->memcg_params.memcg->css);
753 s->memcg_params.deact_fn = deact_fn;
754 call_rcu(&s->memcg_params.deact_rcu_head, kmemcg_deactivate_rcufn);
757 void memcg_deactivate_kmem_caches(struct mem_cgroup *memcg)
760 struct memcg_cache_array *arr;
761 struct kmem_cache *s, *c;
763 idx = memcg_cache_id(memcg);
768 mutex_lock(&slab_mutex);
769 list_for_each_entry(s, &slab_root_caches, root_caches_node) {
770 arr = rcu_dereference_protected(s->memcg_params.memcg_caches,
771 lockdep_is_held(&slab_mutex));
772 c = arr->entries[idx];
776 __kmemcg_cache_deactivate(c);
777 arr->entries[idx] = NULL;
779 mutex_unlock(&slab_mutex);
785 void memcg_destroy_kmem_caches(struct mem_cgroup *memcg)
787 struct kmem_cache *s, *s2;
792 mutex_lock(&slab_mutex);
793 list_for_each_entry_safe(s, s2, &memcg->kmem_caches,
794 memcg_params.kmem_caches_node) {
796 * The cgroup is about to be freed and therefore has no charges
797 * left. Hence, all its caches must be empty by now.
799 BUG_ON(shutdown_cache(s));
801 mutex_unlock(&slab_mutex);
807 static int shutdown_memcg_caches(struct kmem_cache *s)
809 struct memcg_cache_array *arr;
810 struct kmem_cache *c, *c2;
814 BUG_ON(!is_root_cache(s));
817 * First, shutdown active caches, i.e. caches that belong to online
820 arr = rcu_dereference_protected(s->memcg_params.memcg_caches,
821 lockdep_is_held(&slab_mutex));
822 for_each_memcg_cache_index(i) {
826 if (shutdown_cache(c))
828 * The cache still has objects. Move it to a temporary
829 * list so as not to try to destroy it for a second
830 * time while iterating over inactive caches below.
832 list_move(&c->memcg_params.children_node, &busy);
835 * The cache is empty and will be destroyed soon. Clear
836 * the pointer to it in the memcg_caches array so that
837 * it will never be accessed even if the root cache
840 arr->entries[i] = NULL;
844 * Second, shutdown all caches left from memory cgroups that are now
847 list_for_each_entry_safe(c, c2, &s->memcg_params.children,
848 memcg_params.children_node)
851 list_splice(&busy, &s->memcg_params.children);
854 * A cache being destroyed must be empty. In particular, this means
855 * that all per memcg caches attached to it must be empty too.
857 if (!list_empty(&s->memcg_params.children))
862 static void flush_memcg_workqueue(struct kmem_cache *s)
864 mutex_lock(&slab_mutex);
865 s->memcg_params.dying = true;
866 mutex_unlock(&slab_mutex);
869 * SLUB deactivates the kmem_caches through call_rcu. Make
870 * sure all registered rcu callbacks have been invoked.
872 if (IS_ENABLED(CONFIG_SLUB))
876 * SLAB and SLUB create memcg kmem_caches through workqueue and SLUB
877 * deactivates the memcg kmem_caches through workqueue. Make sure all
878 * previous workitems on workqueue are processed.
880 flush_workqueue(memcg_kmem_cache_wq);
883 static inline int shutdown_memcg_caches(struct kmem_cache *s)
888 static inline void flush_memcg_workqueue(struct kmem_cache *s)
891 #endif /* CONFIG_MEMCG_KMEM */
893 void slab_kmem_cache_release(struct kmem_cache *s)
895 __kmem_cache_release(s);
896 destroy_memcg_params(s);
897 kfree_const(s->name);
898 kmem_cache_free(kmem_cache, s);
901 void kmem_cache_destroy(struct kmem_cache *s)
908 flush_memcg_workqueue(s);
913 mutex_lock(&slab_mutex);
919 err = shutdown_memcg_caches(s);
921 err = shutdown_cache(s);
924 pr_err("kmem_cache_destroy %s: Slab cache still has objects\n",
929 mutex_unlock(&slab_mutex);
934 EXPORT_SYMBOL(kmem_cache_destroy);
937 * kmem_cache_shrink - Shrink a cache.
938 * @cachep: The cache to shrink.
940 * Releases as many slabs as possible for a cache.
941 * To help debugging, a zero exit status indicates all slabs were released.
943 * Return: %0 if all slabs were released, non-zero otherwise
945 int kmem_cache_shrink(struct kmem_cache *cachep)
951 kasan_cache_shrink(cachep);
952 ret = __kmem_cache_shrink(cachep);
957 EXPORT_SYMBOL(kmem_cache_shrink);
959 bool slab_is_available(void)
961 return slab_state >= UP;
965 /* Create a cache during boot when no slab services are available yet */
966 void __init create_boot_cache(struct kmem_cache *s, const char *name,
967 unsigned int size, slab_flags_t flags,
968 unsigned int useroffset, unsigned int usersize)
973 s->size = s->object_size = size;
974 s->align = calculate_alignment(flags, ARCH_KMALLOC_MINALIGN, size);
975 s->useroffset = useroffset;
976 s->usersize = usersize;
978 slab_init_memcg_params(s);
980 err = __kmem_cache_create(s, flags);
983 panic("Creation of kmalloc slab %s size=%u failed. Reason %d\n",
986 s->refcount = -1; /* Exempt from merging for now */
989 struct kmem_cache *__init create_kmalloc_cache(const char *name,
990 unsigned int size, slab_flags_t flags,
991 unsigned int useroffset, unsigned int usersize)
993 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
996 panic("Out of memory when creating slab %s\n", name);
998 create_boot_cache(s, name, size, flags, useroffset, usersize);
999 list_add(&s->list, &slab_caches);
1000 memcg_link_cache(s);
1006 kmalloc_caches[NR_KMALLOC_TYPES][KMALLOC_SHIFT_HIGH + 1] __ro_after_init;
1007 EXPORT_SYMBOL(kmalloc_caches);
1010 * Conversion table for small slabs sizes / 8 to the index in the
1011 * kmalloc array. This is necessary for slabs < 192 since we have non power
1012 * of two cache sizes there. The size of larger slabs can be determined using
1015 static u8 size_index[24] __ro_after_init = {
1042 static inline unsigned int size_index_elem(unsigned int bytes)
1044 return (bytes - 1) / 8;
1048 * Find the kmem_cache structure that serves a given size of
1051 struct kmem_cache *kmalloc_slab(size_t size, gfp_t flags)
1057 return ZERO_SIZE_PTR;
1059 index = size_index[size_index_elem(size)];
1061 if (WARN_ON_ONCE(size > KMALLOC_MAX_CACHE_SIZE))
1063 index = fls(size - 1);
1066 return kmalloc_caches[kmalloc_type(flags)][index];
1070 * kmalloc_info[] is to make slub_debug=,kmalloc-xx option work at boot time.
1071 * kmalloc_index() supports up to 2^26=64MB, so the final entry of the table is
1074 const struct kmalloc_info_struct kmalloc_info[] __initconst = {
1075 {NULL, 0}, {"kmalloc-96", 96},
1076 {"kmalloc-192", 192}, {"kmalloc-8", 8},
1077 {"kmalloc-16", 16}, {"kmalloc-32", 32},
1078 {"kmalloc-64", 64}, {"kmalloc-128", 128},
1079 {"kmalloc-256", 256}, {"kmalloc-512", 512},
1080 {"kmalloc-1k", 1024}, {"kmalloc-2k", 2048},
1081 {"kmalloc-4k", 4096}, {"kmalloc-8k", 8192},
1082 {"kmalloc-16k", 16384}, {"kmalloc-32k", 32768},
1083 {"kmalloc-64k", 65536}, {"kmalloc-128k", 131072},
1084 {"kmalloc-256k", 262144}, {"kmalloc-512k", 524288},
1085 {"kmalloc-1M", 1048576}, {"kmalloc-2M", 2097152},
1086 {"kmalloc-4M", 4194304}, {"kmalloc-8M", 8388608},
1087 {"kmalloc-16M", 16777216}, {"kmalloc-32M", 33554432},
1088 {"kmalloc-64M", 67108864}
1092 * Patch up the size_index table if we have strange large alignment
1093 * requirements for the kmalloc array. This is only the case for
1094 * MIPS it seems. The standard arches will not generate any code here.
1096 * Largest permitted alignment is 256 bytes due to the way we
1097 * handle the index determination for the smaller caches.
1099 * Make sure that nothing crazy happens if someone starts tinkering
1100 * around with ARCH_KMALLOC_MINALIGN
1102 void __init setup_kmalloc_cache_index_table(void)
1106 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
1107 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
1109 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
1110 unsigned int elem = size_index_elem(i);
1112 if (elem >= ARRAY_SIZE(size_index))
1114 size_index[elem] = KMALLOC_SHIFT_LOW;
1117 if (KMALLOC_MIN_SIZE >= 64) {
1119 * The 96 byte size cache is not used if the alignment
1122 for (i = 64 + 8; i <= 96; i += 8)
1123 size_index[size_index_elem(i)] = 7;
1127 if (KMALLOC_MIN_SIZE >= 128) {
1129 * The 192 byte sized cache is not used if the alignment
1130 * is 128 byte. Redirect kmalloc to use the 256 byte cache
1133 for (i = 128 + 8; i <= 192; i += 8)
1134 size_index[size_index_elem(i)] = 8;
1139 kmalloc_cache_name(const char *prefix, unsigned int size)
1142 static const char units[3] = "\0kM";
1145 while (size >= 1024 && (size % 1024 == 0)) {
1150 return kasprintf(GFP_NOWAIT, "%s-%u%c", prefix, size, units[idx]);
1154 new_kmalloc_cache(int idx, int type, slab_flags_t flags)
1158 if (type == KMALLOC_RECLAIM) {
1159 flags |= SLAB_RECLAIM_ACCOUNT;
1160 name = kmalloc_cache_name("kmalloc-rcl",
1161 kmalloc_info[idx].size);
1164 name = kmalloc_info[idx].name;
1167 kmalloc_caches[type][idx] = create_kmalloc_cache(name,
1168 kmalloc_info[idx].size, flags, 0,
1169 kmalloc_info[idx].size);
1173 * Create the kmalloc array. Some of the regular kmalloc arrays
1174 * may already have been created because they were needed to
1175 * enable allocations for slab creation.
1177 void __init create_kmalloc_caches(slab_flags_t flags)
1181 for (type = KMALLOC_NORMAL; type <= KMALLOC_RECLAIM; type++) {
1182 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) {
1183 if (!kmalloc_caches[type][i])
1184 new_kmalloc_cache(i, type, flags);
1187 * Caches that are not of the two-to-the-power-of size.
1188 * These have to be created immediately after the
1189 * earlier power of two caches
1191 if (KMALLOC_MIN_SIZE <= 32 && i == 6 &&
1192 !kmalloc_caches[type][1])
1193 new_kmalloc_cache(1, type, flags);
1194 if (KMALLOC_MIN_SIZE <= 64 && i == 7 &&
1195 !kmalloc_caches[type][2])
1196 new_kmalloc_cache(2, type, flags);
1200 /* Kmalloc array is now usable */
1203 #ifdef CONFIG_ZONE_DMA
1204 for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
1205 struct kmem_cache *s = kmalloc_caches[KMALLOC_NORMAL][i];
1208 unsigned int size = kmalloc_size(i);
1209 const char *n = kmalloc_cache_name("dma-kmalloc", size);
1212 kmalloc_caches[KMALLOC_DMA][i] = create_kmalloc_cache(
1213 n, size, SLAB_CACHE_DMA | flags, 0, 0);
1218 #endif /* !CONFIG_SLOB */
1221 * To avoid unnecessary overhead, we pass through large allocation requests
1222 * directly to the page allocator. We use __GFP_COMP, because we will need to
1223 * know the allocation order to free the pages properly in kfree.
1225 void *kmalloc_order(size_t size, gfp_t flags, unsigned int order)
1230 flags |= __GFP_COMP;
1231 page = alloc_pages(flags, order);
1232 ret = page ? page_address(page) : NULL;
1233 ret = kasan_kmalloc_large(ret, size, flags);
1234 /* As ret might get tagged, call kmemleak hook after KASAN. */
1235 kmemleak_alloc(ret, size, 1, flags);
1238 EXPORT_SYMBOL(kmalloc_order);
1240 #ifdef CONFIG_TRACING
1241 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
1243 void *ret = kmalloc_order(size, flags, order);
1244 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
1247 EXPORT_SYMBOL(kmalloc_order_trace);
1250 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1251 /* Randomize a generic freelist */
1252 static void freelist_randomize(struct rnd_state *state, unsigned int *list,
1258 for (i = 0; i < count; i++)
1261 /* Fisher-Yates shuffle */
1262 for (i = count - 1; i > 0; i--) {
1263 rand = prandom_u32_state(state);
1265 swap(list[i], list[rand]);
1269 /* Create a random sequence per cache */
1270 int cache_random_seq_create(struct kmem_cache *cachep, unsigned int count,
1273 struct rnd_state state;
1275 if (count < 2 || cachep->random_seq)
1278 cachep->random_seq = kcalloc(count, sizeof(unsigned int), gfp);
1279 if (!cachep->random_seq)
1282 /* Get best entropy at this stage of boot */
1283 prandom_seed_state(&state, get_random_long());
1285 freelist_randomize(&state, cachep->random_seq, count);
1289 /* Destroy the per-cache random freelist sequence */
1290 void cache_random_seq_destroy(struct kmem_cache *cachep)
1292 kfree(cachep->random_seq);
1293 cachep->random_seq = NULL;
1295 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1297 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
1299 #define SLABINFO_RIGHTS (0600)
1301 #define SLABINFO_RIGHTS (0400)
1304 static void print_slabinfo_header(struct seq_file *m)
1307 * Output format version, so at least we can change it
1308 * without _too_ many complaints.
1310 #ifdef CONFIG_DEBUG_SLAB
1311 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
1313 seq_puts(m, "slabinfo - version: 2.1\n");
1315 seq_puts(m, "# name <active_objs> <num_objs> <objsize> <objperslab> <pagesperslab>");
1316 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
1317 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
1318 #ifdef CONFIG_DEBUG_SLAB
1319 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> <error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
1320 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
1325 void *slab_start(struct seq_file *m, loff_t *pos)
1327 mutex_lock(&slab_mutex);
1328 return seq_list_start(&slab_root_caches, *pos);
1331 void *slab_next(struct seq_file *m, void *p, loff_t *pos)
1333 return seq_list_next(p, &slab_root_caches, pos);
1336 void slab_stop(struct seq_file *m, void *p)
1338 mutex_unlock(&slab_mutex);
1342 memcg_accumulate_slabinfo(struct kmem_cache *s, struct slabinfo *info)
1344 struct kmem_cache *c;
1345 struct slabinfo sinfo;
1347 if (!is_root_cache(s))
1350 for_each_memcg_cache(c, s) {
1351 memset(&sinfo, 0, sizeof(sinfo));
1352 get_slabinfo(c, &sinfo);
1354 info->active_slabs += sinfo.active_slabs;
1355 info->num_slabs += sinfo.num_slabs;
1356 info->shared_avail += sinfo.shared_avail;
1357 info->active_objs += sinfo.active_objs;
1358 info->num_objs += sinfo.num_objs;
1362 static void cache_show(struct kmem_cache *s, struct seq_file *m)
1364 struct slabinfo sinfo;
1366 memset(&sinfo, 0, sizeof(sinfo));
1367 get_slabinfo(s, &sinfo);
1369 memcg_accumulate_slabinfo(s, &sinfo);
1371 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
1372 cache_name(s), sinfo.active_objs, sinfo.num_objs, s->size,
1373 sinfo.objects_per_slab, (1 << sinfo.cache_order));
1375 seq_printf(m, " : tunables %4u %4u %4u",
1376 sinfo.limit, sinfo.batchcount, sinfo.shared);
1377 seq_printf(m, " : slabdata %6lu %6lu %6lu",
1378 sinfo.active_slabs, sinfo.num_slabs, sinfo.shared_avail);
1379 slabinfo_show_stats(m, s);
1383 static int slab_show(struct seq_file *m, void *p)
1385 struct kmem_cache *s = list_entry(p, struct kmem_cache, root_caches_node);
1387 if (p == slab_root_caches.next)
1388 print_slabinfo_header(m);
1393 void dump_unreclaimable_slab(void)
1395 struct kmem_cache *s, *s2;
1396 struct slabinfo sinfo;
1399 * Here acquiring slab_mutex is risky since we don't prefer to get
1400 * sleep in oom path. But, without mutex hold, it may introduce a
1402 * Use mutex_trylock to protect the list traverse, dump nothing
1403 * without acquiring the mutex.
1405 if (!mutex_trylock(&slab_mutex)) {
1406 pr_warn("excessive unreclaimable slab but cannot dump stats\n");
1410 pr_info("Unreclaimable slab info:\n");
1411 pr_info("Name Used Total\n");
1413 list_for_each_entry_safe(s, s2, &slab_caches, list) {
1414 if (!is_root_cache(s) || (s->flags & SLAB_RECLAIM_ACCOUNT))
1417 get_slabinfo(s, &sinfo);
1419 if (sinfo.num_objs > 0)
1420 pr_info("%-17s %10luKB %10luKB\n", cache_name(s),
1421 (sinfo.active_objs * s->size) / 1024,
1422 (sinfo.num_objs * s->size) / 1024);
1424 mutex_unlock(&slab_mutex);
1427 #if defined(CONFIG_MEMCG)
1428 void *memcg_slab_start(struct seq_file *m, loff_t *pos)
1430 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
1432 mutex_lock(&slab_mutex);
1433 return seq_list_start(&memcg->kmem_caches, *pos);
1436 void *memcg_slab_next(struct seq_file *m, void *p, loff_t *pos)
1438 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
1440 return seq_list_next(p, &memcg->kmem_caches, pos);
1443 void memcg_slab_stop(struct seq_file *m, void *p)
1445 mutex_unlock(&slab_mutex);
1448 int memcg_slab_show(struct seq_file *m, void *p)
1450 struct kmem_cache *s = list_entry(p, struct kmem_cache,
1451 memcg_params.kmem_caches_node);
1452 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
1454 if (p == memcg->kmem_caches.next)
1455 print_slabinfo_header(m);
1462 * slabinfo_op - iterator that generates /proc/slabinfo
1471 * num-pages-per-slab
1472 * + further values on SMP and with statistics enabled
1474 static const struct seq_operations slabinfo_op = {
1475 .start = slab_start,
1481 static int slabinfo_open(struct inode *inode, struct file *file)
1483 return seq_open(file, &slabinfo_op);
1486 static const struct file_operations proc_slabinfo_operations = {
1487 .open = slabinfo_open,
1489 .write = slabinfo_write,
1490 .llseek = seq_lseek,
1491 .release = seq_release,
1494 static int __init slab_proc_init(void)
1496 proc_create("slabinfo", SLABINFO_RIGHTS, NULL,
1497 &proc_slabinfo_operations);
1500 module_init(slab_proc_init);
1501 #endif /* CONFIG_SLAB || CONFIG_SLUB_DEBUG */
1503 static __always_inline void *__do_krealloc(const void *p, size_t new_size,
1512 if (ks >= new_size) {
1513 p = kasan_krealloc((void *)p, new_size, flags);
1517 ret = kmalloc_track_caller(new_size, flags);
1525 * __krealloc - like krealloc() but don't free @p.
1526 * @p: object to reallocate memory for.
1527 * @new_size: how many bytes of memory are required.
1528 * @flags: the type of memory to allocate.
1530 * This function is like krealloc() except it never frees the originally
1531 * allocated buffer. Use this if you don't want to free the buffer immediately
1532 * like, for example, with RCU.
1534 * Return: pointer to the allocated memory or %NULL in case of error
1536 void *__krealloc(const void *p, size_t new_size, gfp_t flags)
1538 if (unlikely(!new_size))
1539 return ZERO_SIZE_PTR;
1541 return __do_krealloc(p, new_size, flags);
1544 EXPORT_SYMBOL(__krealloc);
1547 * krealloc - reallocate memory. The contents will remain unchanged.
1548 * @p: object to reallocate memory for.
1549 * @new_size: how many bytes of memory are required.
1550 * @flags: the type of memory to allocate.
1552 * The contents of the object pointed to are preserved up to the
1553 * lesser of the new and old sizes. If @p is %NULL, krealloc()
1554 * behaves exactly like kmalloc(). If @new_size is 0 and @p is not a
1555 * %NULL pointer, the object pointed to is freed.
1557 * Return: pointer to the allocated memory or %NULL in case of error
1559 void *krealloc(const void *p, size_t new_size, gfp_t flags)
1563 if (unlikely(!new_size)) {
1565 return ZERO_SIZE_PTR;
1568 ret = __do_krealloc(p, new_size, flags);
1569 if (ret && kasan_reset_tag(p) != kasan_reset_tag(ret))
1574 EXPORT_SYMBOL(krealloc);
1577 * kzfree - like kfree but zero memory
1578 * @p: object to free memory of
1580 * The memory of the object @p points to is zeroed before freed.
1581 * If @p is %NULL, kzfree() does nothing.
1583 * Note: this function zeroes the whole allocated buffer which can be a good
1584 * deal bigger than the requested buffer size passed to kmalloc(). So be
1585 * careful when using this function in performance sensitive code.
1587 void kzfree(const void *p)
1590 void *mem = (void *)p;
1592 if (unlikely(ZERO_OR_NULL_PTR(mem)))
1598 EXPORT_SYMBOL(kzfree);
1600 /* Tracepoints definitions. */
1601 EXPORT_TRACEPOINT_SYMBOL(kmalloc);
1602 EXPORT_TRACEPOINT_SYMBOL(kmem_cache_alloc);
1603 EXPORT_TRACEPOINT_SYMBOL(kmalloc_node);
1604 EXPORT_TRACEPOINT_SYMBOL(kmem_cache_alloc_node);
1605 EXPORT_TRACEPOINT_SYMBOL(kfree);
1606 EXPORT_TRACEPOINT_SYMBOL(kmem_cache_free);
1608 int should_failslab(struct kmem_cache *s, gfp_t gfpflags)
1610 if (__should_failslab(s, gfpflags))
1614 ALLOW_ERROR_INJECTION(should_failslab, ERRNO);