1 // SPDX-License-Identifier: GPL-2.0
3 * Slab allocator functions that are independent of the allocator strategy
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 <linux/debugfs.h>
21 #include <asm/cacheflush.h>
22 #include <asm/tlbflush.h>
24 #include <linux/memcontrol.h>
26 #define CREATE_TRACE_POINTS
27 #include <trace/events/kmem.h>
31 enum slab_state slab_state;
32 LIST_HEAD(slab_caches);
33 DEFINE_MUTEX(slab_mutex);
34 struct kmem_cache *kmem_cache;
36 #ifdef CONFIG_HARDENED_USERCOPY
37 bool usercopy_fallback __ro_after_init =
38 IS_ENABLED(CONFIG_HARDENED_USERCOPY_FALLBACK);
39 module_param(usercopy_fallback, bool, 0400);
40 MODULE_PARM_DESC(usercopy_fallback,
41 "WARN instead of reject usercopy whitelist violations");
44 static LIST_HEAD(slab_caches_to_rcu_destroy);
45 static void slab_caches_to_rcu_destroy_workfn(struct work_struct *work);
46 static DECLARE_WORK(slab_caches_to_rcu_destroy_work,
47 slab_caches_to_rcu_destroy_workfn);
50 * Set of flags that will prevent slab merging
52 #define SLAB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
53 SLAB_TRACE | SLAB_TYPESAFE_BY_RCU | SLAB_NOLEAKTRACE | \
54 SLAB_FAILSLAB | SLAB_KASAN)
56 #define SLAB_MERGE_SAME (SLAB_RECLAIM_ACCOUNT | SLAB_CACHE_DMA | \
57 SLAB_CACHE_DMA32 | SLAB_ACCOUNT)
60 * Merge control. If this is set then no merging of slab caches will occur.
62 static bool slab_nomerge = !IS_ENABLED(CONFIG_SLAB_MERGE_DEFAULT);
64 static int __init setup_slab_nomerge(char *str)
71 __setup_param("slub_nomerge", slub_nomerge, setup_slab_nomerge, 0);
74 __setup("slab_nomerge", setup_slab_nomerge);
77 * Determine the size of a slab object
79 unsigned int kmem_cache_size(struct kmem_cache *s)
81 return s->object_size;
83 EXPORT_SYMBOL(kmem_cache_size);
85 #ifdef CONFIG_DEBUG_VM
86 static int kmem_cache_sanity_check(const char *name, unsigned int size)
88 if (!name || in_interrupt() || size < sizeof(void *) ||
89 size > KMALLOC_MAX_SIZE) {
90 pr_err("kmem_cache_create(%s) integrity check failed\n", name);
94 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
98 static inline int kmem_cache_sanity_check(const char *name, unsigned int size)
104 void __kmem_cache_free_bulk(struct kmem_cache *s, size_t nr, void **p)
108 for (i = 0; i < nr; i++) {
110 kmem_cache_free(s, p[i]);
116 int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t nr,
121 for (i = 0; i < nr; i++) {
122 void *x = p[i] = kmem_cache_alloc(s, flags);
124 __kmem_cache_free_bulk(s, i, p);
131 #ifdef CONFIG_MEMCG_KMEM
133 LIST_HEAD(slab_root_caches);
134 static DEFINE_SPINLOCK(memcg_kmem_wq_lock);
136 static void kmemcg_cache_shutdown(struct percpu_ref *percpu_ref);
138 void slab_init_memcg_params(struct kmem_cache *s)
140 s->memcg_params.root_cache = NULL;
141 RCU_INIT_POINTER(s->memcg_params.memcg_caches, NULL);
142 INIT_LIST_HEAD(&s->memcg_params.children);
143 s->memcg_params.dying = false;
146 static int init_memcg_params(struct kmem_cache *s,
147 struct kmem_cache *root_cache)
149 struct memcg_cache_array *arr;
152 int ret = percpu_ref_init(&s->memcg_params.refcnt,
153 kmemcg_cache_shutdown,
158 s->memcg_params.root_cache = root_cache;
159 INIT_LIST_HEAD(&s->memcg_params.children_node);
160 INIT_LIST_HEAD(&s->memcg_params.kmem_caches_node);
164 slab_init_memcg_params(s);
166 if (!memcg_nr_cache_ids)
169 arr = kvzalloc(sizeof(struct memcg_cache_array) +
170 memcg_nr_cache_ids * sizeof(void *),
175 RCU_INIT_POINTER(s->memcg_params.memcg_caches, arr);
179 static void destroy_memcg_params(struct kmem_cache *s)
181 if (is_root_cache(s))
182 kvfree(rcu_access_pointer(s->memcg_params.memcg_caches));
184 percpu_ref_exit(&s->memcg_params.refcnt);
187 static void free_memcg_params(struct rcu_head *rcu)
189 struct memcg_cache_array *old;
191 old = container_of(rcu, struct memcg_cache_array, rcu);
195 static int update_memcg_params(struct kmem_cache *s, int new_array_size)
197 struct memcg_cache_array *old, *new;
199 new = kvzalloc(sizeof(struct memcg_cache_array) +
200 new_array_size * sizeof(void *), GFP_KERNEL);
204 old = rcu_dereference_protected(s->memcg_params.memcg_caches,
205 lockdep_is_held(&slab_mutex));
207 memcpy(new->entries, old->entries,
208 memcg_nr_cache_ids * sizeof(void *));
210 rcu_assign_pointer(s->memcg_params.memcg_caches, new);
212 call_rcu(&old->rcu, free_memcg_params);
216 int memcg_update_all_caches(int num_memcgs)
218 struct kmem_cache *s;
221 mutex_lock(&slab_mutex);
222 list_for_each_entry(s, &slab_root_caches, root_caches_node) {
223 ret = update_memcg_params(s, num_memcgs);
225 * Instead of freeing the memory, we'll just leave the caches
226 * up to this point in an updated state.
231 mutex_unlock(&slab_mutex);
235 void memcg_link_cache(struct kmem_cache *s, struct mem_cgroup *memcg)
237 if (is_root_cache(s)) {
238 list_add(&s->root_caches_node, &slab_root_caches);
240 css_get(&memcg->css);
241 s->memcg_params.memcg = memcg;
242 list_add(&s->memcg_params.children_node,
243 &s->memcg_params.root_cache->memcg_params.children);
244 list_add(&s->memcg_params.kmem_caches_node,
245 &s->memcg_params.memcg->kmem_caches);
249 static void memcg_unlink_cache(struct kmem_cache *s)
251 if (is_root_cache(s)) {
252 list_del(&s->root_caches_node);
254 list_del(&s->memcg_params.children_node);
255 list_del(&s->memcg_params.kmem_caches_node);
256 mem_cgroup_put(s->memcg_params.memcg);
257 WRITE_ONCE(s->memcg_params.memcg, NULL);
261 static inline int init_memcg_params(struct kmem_cache *s,
262 struct kmem_cache *root_cache)
267 static inline void destroy_memcg_params(struct kmem_cache *s)
271 static inline void memcg_unlink_cache(struct kmem_cache *s)
274 #endif /* CONFIG_MEMCG_KMEM */
277 * Figure out what the alignment of the objects will be given a set of
278 * flags, a user specified alignment and the size of the objects.
280 static unsigned int calculate_alignment(slab_flags_t flags,
281 unsigned int align, unsigned int size)
284 * If the user wants hardware cache aligned objects then follow that
285 * suggestion if the object is sufficiently large.
287 * The hardware cache alignment cannot override the specified
288 * alignment though. If that is greater then use it.
290 if (flags & SLAB_HWCACHE_ALIGN) {
293 ralign = cache_line_size();
294 while (size <= ralign / 2)
296 align = max(align, ralign);
299 if (align < ARCH_SLAB_MINALIGN)
300 align = ARCH_SLAB_MINALIGN;
302 return ALIGN(align, sizeof(void *));
306 * Find a mergeable slab cache
308 int slab_unmergeable(struct kmem_cache *s)
310 if (slab_nomerge || (s->flags & SLAB_NEVER_MERGE))
313 if (!is_root_cache(s))
323 * We may have set a slab to be unmergeable during bootstrap.
331 struct kmem_cache *find_mergeable(unsigned int size, unsigned int align,
332 slab_flags_t flags, const char *name, void (*ctor)(void *))
334 struct kmem_cache *s;
342 size = ALIGN(size, sizeof(void *));
343 align = calculate_alignment(flags, align, size);
344 size = ALIGN(size, align);
345 flags = kmem_cache_flags(size, flags, name, NULL);
347 if (flags & SLAB_NEVER_MERGE)
350 list_for_each_entry_reverse(s, &slab_root_caches, root_caches_node) {
351 if (slab_unmergeable(s))
357 if ((flags & SLAB_MERGE_SAME) != (s->flags & SLAB_MERGE_SAME))
360 * Check if alignment is compatible.
361 * Courtesy of Adrian Drzewiecki
363 if ((s->size & ~(align - 1)) != s->size)
366 if (s->size - size >= sizeof(void *))
369 if (IS_ENABLED(CONFIG_SLAB) && align &&
370 (align > s->align || s->align % align))
378 static struct kmem_cache *create_cache(const char *name,
379 unsigned int object_size, unsigned int align,
380 slab_flags_t flags, unsigned int useroffset,
381 unsigned int usersize, void (*ctor)(void *),
382 struct mem_cgroup *memcg, struct kmem_cache *root_cache)
384 struct kmem_cache *s;
387 if (WARN_ON(useroffset + usersize > object_size))
388 useroffset = usersize = 0;
391 s = kmem_cache_zalloc(kmem_cache, GFP_KERNEL);
396 s->size = s->object_size = object_size;
399 s->useroffset = useroffset;
400 s->usersize = usersize;
402 err = init_memcg_params(s, root_cache);
406 err = __kmem_cache_create(s, flags);
411 list_add(&s->list, &slab_caches);
412 memcg_link_cache(s, memcg);
419 destroy_memcg_params(s);
420 kmem_cache_free(kmem_cache, s);
425 * kmem_cache_create_usercopy - Create a cache with a region suitable
426 * for copying to userspace
427 * @name: A string which is used in /proc/slabinfo to identify this cache.
428 * @size: The size of objects to be created in this cache.
429 * @align: The required alignment for the objects.
431 * @useroffset: Usercopy region offset
432 * @usersize: Usercopy region size
433 * @ctor: A constructor for the objects.
435 * Cannot be called within a interrupt, but can be interrupted.
436 * The @ctor is run when new pages are allocated by the cache.
440 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
441 * to catch references to uninitialised memory.
443 * %SLAB_RED_ZONE - Insert `Red` zones around the allocated memory to check
444 * for buffer overruns.
446 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
447 * cacheline. This can be beneficial if you're counting cycles as closely
450 * Return: a pointer to the cache on success, NULL on failure.
453 kmem_cache_create_usercopy(const char *name,
454 unsigned int size, unsigned int align,
456 unsigned int useroffset, unsigned int usersize,
457 void (*ctor)(void *))
459 struct kmem_cache *s = NULL;
460 const char *cache_name;
465 memcg_get_cache_ids();
467 mutex_lock(&slab_mutex);
469 err = kmem_cache_sanity_check(name, size);
474 /* Refuse requests with allocator specific flags */
475 if (flags & ~SLAB_FLAGS_PERMITTED) {
481 * Some allocators will constraint the set of valid flags to a subset
482 * of all flags. We expect them to define CACHE_CREATE_MASK in this
483 * case, and we'll just provide them with a sanitized version of the
486 flags &= CACHE_CREATE_MASK;
488 /* Fail closed on bad usersize of useroffset values. */
489 if (WARN_ON(!usersize && useroffset) ||
490 WARN_ON(size < usersize || size - usersize < useroffset))
491 usersize = useroffset = 0;
494 s = __kmem_cache_alias(name, size, align, flags, ctor);
498 cache_name = kstrdup_const(name, GFP_KERNEL);
504 s = create_cache(cache_name, size,
505 calculate_alignment(flags, align, size),
506 flags, useroffset, usersize, ctor, NULL, NULL);
509 kfree_const(cache_name);
513 mutex_unlock(&slab_mutex);
515 memcg_put_cache_ids();
520 if (flags & SLAB_PANIC)
521 panic("kmem_cache_create: Failed to create slab '%s'. Error %d\n",
524 pr_warn("kmem_cache_create(%s) failed with error %d\n",
532 EXPORT_SYMBOL(kmem_cache_create_usercopy);
535 * kmem_cache_create - Create a cache.
536 * @name: A string which is used in /proc/slabinfo to identify this cache.
537 * @size: The size of objects to be created in this cache.
538 * @align: The required alignment for the objects.
540 * @ctor: A constructor for the objects.
542 * Cannot be called within a interrupt, but can be interrupted.
543 * The @ctor is run when new pages are allocated by the cache.
547 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
548 * to catch references to uninitialised memory.
550 * %SLAB_RED_ZONE - Insert `Red` zones around the allocated memory to check
551 * for buffer overruns.
553 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
554 * cacheline. This can be beneficial if you're counting cycles as closely
557 * Return: a pointer to the cache on success, NULL on failure.
560 kmem_cache_create(const char *name, unsigned int size, unsigned int align,
561 slab_flags_t flags, void (*ctor)(void *))
563 return kmem_cache_create_usercopy(name, size, align, flags, 0, 0,
566 EXPORT_SYMBOL(kmem_cache_create);
568 static void slab_caches_to_rcu_destroy_workfn(struct work_struct *work)
570 LIST_HEAD(to_destroy);
571 struct kmem_cache *s, *s2;
574 * On destruction, SLAB_TYPESAFE_BY_RCU kmem_caches are put on the
575 * @slab_caches_to_rcu_destroy list. The slab pages are freed
576 * through RCU and and the associated kmem_cache are dereferenced
577 * while freeing the pages, so the kmem_caches should be freed only
578 * after the pending RCU operations are finished. As rcu_barrier()
579 * is a pretty slow operation, we batch all pending destructions
582 mutex_lock(&slab_mutex);
583 list_splice_init(&slab_caches_to_rcu_destroy, &to_destroy);
584 mutex_unlock(&slab_mutex);
586 if (list_empty(&to_destroy))
591 list_for_each_entry_safe(s, s2, &to_destroy, list) {
592 #ifdef SLAB_SUPPORTS_SYSFS
593 sysfs_slab_release(s);
595 slab_kmem_cache_release(s);
600 static int shutdown_cache(struct kmem_cache *s)
602 /* free asan quarantined objects */
603 kasan_cache_shutdown(s);
605 if (__kmem_cache_shutdown(s) != 0)
608 memcg_unlink_cache(s);
611 if (s->flags & SLAB_TYPESAFE_BY_RCU) {
612 #ifdef SLAB_SUPPORTS_SYSFS
613 sysfs_slab_unlink(s);
615 list_add_tail(&s->list, &slab_caches_to_rcu_destroy);
616 schedule_work(&slab_caches_to_rcu_destroy_work);
618 #ifdef SLAB_SUPPORTS_SYSFS
619 sysfs_slab_unlink(s);
620 sysfs_slab_release(s);
622 slab_kmem_cache_release(s);
629 #ifdef CONFIG_MEMCG_KMEM
631 * memcg_create_kmem_cache - Create a cache for a memory cgroup.
632 * @memcg: The memory cgroup the new cache is for.
633 * @root_cache: The parent of the new cache.
635 * This function attempts to create a kmem cache that will serve allocation
636 * requests going from @memcg to @root_cache. The new cache inherits properties
639 void memcg_create_kmem_cache(struct mem_cgroup *memcg,
640 struct kmem_cache *root_cache)
642 static char memcg_name_buf[NAME_MAX + 1]; /* protected by slab_mutex */
643 struct cgroup_subsys_state *css = &memcg->css;
644 struct memcg_cache_array *arr;
645 struct kmem_cache *s = NULL;
652 mutex_lock(&slab_mutex);
655 * The memory cgroup could have been offlined while the cache
656 * creation work was pending.
658 if (memcg->kmem_state != KMEM_ONLINE)
661 idx = memcg_cache_id(memcg);
662 arr = rcu_dereference_protected(root_cache->memcg_params.memcg_caches,
663 lockdep_is_held(&slab_mutex));
666 * Since per-memcg caches are created asynchronously on first
667 * allocation (see memcg_kmem_get_cache()), several threads can try to
668 * create the same cache, but only one of them may succeed.
670 if (arr->entries[idx])
673 cgroup_name(css->cgroup, memcg_name_buf, sizeof(memcg_name_buf));
674 cache_name = kasprintf(GFP_KERNEL, "%s(%llu:%s)", root_cache->name,
675 css->serial_nr, memcg_name_buf);
679 s = create_cache(cache_name, root_cache->object_size,
681 root_cache->flags & CACHE_CREATE_MASK,
682 root_cache->useroffset, root_cache->usersize,
683 root_cache->ctor, memcg, root_cache);
685 * If we could not create a memcg cache, do not complain, because
686 * that's not critical at all as we can always proceed with the root
695 * Since readers won't lock (see memcg_kmem_get_cache()), we need a
696 * barrier here to ensure nobody will see the kmem_cache partially
700 arr->entries[idx] = s;
703 mutex_unlock(&slab_mutex);
709 static void kmemcg_workfn(struct work_struct *work)
711 struct kmem_cache *s = container_of(work, struct kmem_cache,
717 mutex_lock(&slab_mutex);
718 s->memcg_params.work_fn(s);
719 mutex_unlock(&slab_mutex);
725 static void kmemcg_rcufn(struct rcu_head *head)
727 struct kmem_cache *s = container_of(head, struct kmem_cache,
728 memcg_params.rcu_head);
731 * We need to grab blocking locks. Bounce to ->work. The
732 * work item shares the space with the RCU head and can't be
733 * initialized eariler.
735 INIT_WORK(&s->memcg_params.work, kmemcg_workfn);
736 queue_work(memcg_kmem_cache_wq, &s->memcg_params.work);
739 static void kmemcg_cache_shutdown_fn(struct kmem_cache *s)
741 WARN_ON(shutdown_cache(s));
744 static void kmemcg_cache_shutdown(struct percpu_ref *percpu_ref)
746 struct kmem_cache *s = container_of(percpu_ref, struct kmem_cache,
747 memcg_params.refcnt);
750 spin_lock_irqsave(&memcg_kmem_wq_lock, flags);
751 if (s->memcg_params.root_cache->memcg_params.dying)
754 s->memcg_params.work_fn = kmemcg_cache_shutdown_fn;
755 INIT_WORK(&s->memcg_params.work, kmemcg_workfn);
756 queue_work(memcg_kmem_cache_wq, &s->memcg_params.work);
759 spin_unlock_irqrestore(&memcg_kmem_wq_lock, flags);
762 static void kmemcg_cache_deactivate_after_rcu(struct kmem_cache *s)
764 __kmemcg_cache_deactivate_after_rcu(s);
765 percpu_ref_kill(&s->memcg_params.refcnt);
768 static void kmemcg_cache_deactivate(struct kmem_cache *s)
770 if (WARN_ON_ONCE(is_root_cache(s)))
773 __kmemcg_cache_deactivate(s);
774 s->flags |= SLAB_DEACTIVATED;
777 * memcg_kmem_wq_lock is used to synchronize memcg_params.dying
778 * flag and make sure that no new kmem_cache deactivation tasks
779 * are queued (see flush_memcg_workqueue() ).
781 spin_lock_irq(&memcg_kmem_wq_lock);
782 if (s->memcg_params.root_cache->memcg_params.dying)
785 s->memcg_params.work_fn = kmemcg_cache_deactivate_after_rcu;
786 call_rcu(&s->memcg_params.rcu_head, kmemcg_rcufn);
788 spin_unlock_irq(&memcg_kmem_wq_lock);
791 void memcg_deactivate_kmem_caches(struct mem_cgroup *memcg,
792 struct mem_cgroup *parent)
795 struct memcg_cache_array *arr;
796 struct kmem_cache *s, *c;
797 unsigned int nr_reparented;
799 idx = memcg_cache_id(memcg);
804 mutex_lock(&slab_mutex);
805 list_for_each_entry(s, &slab_root_caches, root_caches_node) {
806 arr = rcu_dereference_protected(s->memcg_params.memcg_caches,
807 lockdep_is_held(&slab_mutex));
808 c = arr->entries[idx];
812 kmemcg_cache_deactivate(c);
813 arr->entries[idx] = NULL;
816 list_for_each_entry(s, &memcg->kmem_caches,
817 memcg_params.kmem_caches_node) {
818 WRITE_ONCE(s->memcg_params.memcg, parent);
819 css_put(&memcg->css);
823 list_splice_init(&memcg->kmem_caches,
824 &parent->kmem_caches);
825 css_get_many(&parent->css, nr_reparented);
827 mutex_unlock(&slab_mutex);
833 static int shutdown_memcg_caches(struct kmem_cache *s)
835 struct memcg_cache_array *arr;
836 struct kmem_cache *c, *c2;
840 BUG_ON(!is_root_cache(s));
843 * First, shutdown active caches, i.e. caches that belong to online
846 arr = rcu_dereference_protected(s->memcg_params.memcg_caches,
847 lockdep_is_held(&slab_mutex));
848 for_each_memcg_cache_index(i) {
852 if (shutdown_cache(c))
854 * The cache still has objects. Move it to a temporary
855 * list so as not to try to destroy it for a second
856 * time while iterating over inactive caches below.
858 list_move(&c->memcg_params.children_node, &busy);
861 * The cache is empty and will be destroyed soon. Clear
862 * the pointer to it in the memcg_caches array so that
863 * it will never be accessed even if the root cache
866 arr->entries[i] = NULL;
870 * Second, shutdown all caches left from memory cgroups that are now
873 list_for_each_entry_safe(c, c2, &s->memcg_params.children,
874 memcg_params.children_node)
877 list_splice(&busy, &s->memcg_params.children);
880 * A cache being destroyed must be empty. In particular, this means
881 * that all per memcg caches attached to it must be empty too.
883 if (!list_empty(&s->memcg_params.children))
888 static void flush_memcg_workqueue(struct kmem_cache *s)
890 spin_lock_irq(&memcg_kmem_wq_lock);
891 s->memcg_params.dying = true;
892 spin_unlock_irq(&memcg_kmem_wq_lock);
895 * SLAB and SLUB deactivate the kmem_caches through call_rcu. Make
896 * sure all registered rcu callbacks have been invoked.
901 * SLAB and SLUB create memcg kmem_caches through workqueue and SLUB
902 * deactivates the memcg kmem_caches through workqueue. Make sure all
903 * previous workitems on workqueue are processed.
905 flush_workqueue(memcg_kmem_cache_wq);
908 static inline int shutdown_memcg_caches(struct kmem_cache *s)
913 static inline void flush_memcg_workqueue(struct kmem_cache *s)
916 #endif /* CONFIG_MEMCG_KMEM */
918 void slab_kmem_cache_release(struct kmem_cache *s)
920 __kmem_cache_release(s);
921 destroy_memcg_params(s);
922 kfree_const(s->name);
923 kmem_cache_free(kmem_cache, s);
926 void kmem_cache_destroy(struct kmem_cache *s)
933 flush_memcg_workqueue(s);
938 mutex_lock(&slab_mutex);
944 err = shutdown_memcg_caches(s);
946 err = shutdown_cache(s);
949 pr_err("kmem_cache_destroy %s: Slab cache still has objects\n",
954 mutex_unlock(&slab_mutex);
959 EXPORT_SYMBOL(kmem_cache_destroy);
962 * kmem_cache_shrink - Shrink a cache.
963 * @cachep: The cache to shrink.
965 * Releases as many slabs as possible for a cache.
966 * To help debugging, a zero exit status indicates all slabs were released.
968 * Return: %0 if all slabs were released, non-zero otherwise
970 int kmem_cache_shrink(struct kmem_cache *cachep)
976 kasan_cache_shrink(cachep);
977 ret = __kmem_cache_shrink(cachep);
982 EXPORT_SYMBOL(kmem_cache_shrink);
985 * kmem_cache_shrink_all - shrink a cache and all memcg caches for root cache
986 * @s: The cache pointer
988 void kmem_cache_shrink_all(struct kmem_cache *s)
990 struct kmem_cache *c;
992 if (!IS_ENABLED(CONFIG_MEMCG_KMEM) || !is_root_cache(s)) {
993 kmem_cache_shrink(s);
999 kasan_cache_shrink(s);
1000 __kmem_cache_shrink(s);
1003 * We have to take the slab_mutex to protect from the memcg list
1006 mutex_lock(&slab_mutex);
1007 for_each_memcg_cache(c, s) {
1009 * Don't need to shrink deactivated memcg caches.
1011 if (s->flags & SLAB_DEACTIVATED)
1013 kasan_cache_shrink(c);
1014 __kmem_cache_shrink(c);
1016 mutex_unlock(&slab_mutex);
1021 bool slab_is_available(void)
1023 return slab_state >= UP;
1027 /* Create a cache during boot when no slab services are available yet */
1028 void __init create_boot_cache(struct kmem_cache *s, const char *name,
1029 unsigned int size, slab_flags_t flags,
1030 unsigned int useroffset, unsigned int usersize)
1035 s->size = s->object_size = size;
1036 s->align = calculate_alignment(flags, ARCH_KMALLOC_MINALIGN, size);
1037 s->useroffset = useroffset;
1038 s->usersize = usersize;
1040 slab_init_memcg_params(s);
1042 err = __kmem_cache_create(s, flags);
1045 panic("Creation of kmalloc slab %s size=%u failed. Reason %d\n",
1048 s->refcount = -1; /* Exempt from merging for now */
1051 struct kmem_cache *__init create_kmalloc_cache(const char *name,
1052 unsigned int size, slab_flags_t flags,
1053 unsigned int useroffset, unsigned int usersize)
1055 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
1058 panic("Out of memory when creating slab %s\n", name);
1060 create_boot_cache(s, name, size, flags, useroffset, usersize);
1061 list_add(&s->list, &slab_caches);
1062 memcg_link_cache(s, NULL);
1068 kmalloc_caches[NR_KMALLOC_TYPES][KMALLOC_SHIFT_HIGH + 1] __ro_after_init =
1069 { /* initialization for https://bugs.llvm.org/show_bug.cgi?id=42570 */ };
1070 EXPORT_SYMBOL(kmalloc_caches);
1073 * Conversion table for small slabs sizes / 8 to the index in the
1074 * kmalloc array. This is necessary for slabs < 192 since we have non power
1075 * of two cache sizes there. The size of larger slabs can be determined using
1078 static u8 size_index[24] __ro_after_init = {
1105 static inline unsigned int size_index_elem(unsigned int bytes)
1107 return (bytes - 1) / 8;
1111 * Find the kmem_cache structure that serves a given size of
1114 struct kmem_cache *kmalloc_slab(size_t size, gfp_t flags)
1120 return ZERO_SIZE_PTR;
1122 index = size_index[size_index_elem(size)];
1124 if (WARN_ON_ONCE(size > KMALLOC_MAX_CACHE_SIZE))
1126 index = fls(size - 1);
1129 return kmalloc_caches[kmalloc_type(flags)][index];
1133 * kmalloc_info[] is to make slub_debug=,kmalloc-xx option work at boot time.
1134 * kmalloc_index() supports up to 2^26=64MB, so the final entry of the table is
1137 const struct kmalloc_info_struct kmalloc_info[] __initconst = {
1138 {NULL, 0}, {"kmalloc-96", 96},
1139 {"kmalloc-192", 192}, {"kmalloc-8", 8},
1140 {"kmalloc-16", 16}, {"kmalloc-32", 32},
1141 {"kmalloc-64", 64}, {"kmalloc-128", 128},
1142 {"kmalloc-256", 256}, {"kmalloc-512", 512},
1143 {"kmalloc-1k", 1024}, {"kmalloc-2k", 2048},
1144 {"kmalloc-4k", 4096}, {"kmalloc-8k", 8192},
1145 {"kmalloc-16k", 16384}, {"kmalloc-32k", 32768},
1146 {"kmalloc-64k", 65536}, {"kmalloc-128k", 131072},
1147 {"kmalloc-256k", 262144}, {"kmalloc-512k", 524288},
1148 {"kmalloc-1M", 1048576}, {"kmalloc-2M", 2097152},
1149 {"kmalloc-4M", 4194304}, {"kmalloc-8M", 8388608},
1150 {"kmalloc-16M", 16777216}, {"kmalloc-32M", 33554432},
1151 {"kmalloc-64M", 67108864}
1155 * Patch up the size_index table if we have strange large alignment
1156 * requirements for the kmalloc array. This is only the case for
1157 * MIPS it seems. The standard arches will not generate any code here.
1159 * Largest permitted alignment is 256 bytes due to the way we
1160 * handle the index determination for the smaller caches.
1162 * Make sure that nothing crazy happens if someone starts tinkering
1163 * around with ARCH_KMALLOC_MINALIGN
1165 void __init setup_kmalloc_cache_index_table(void)
1169 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
1170 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
1172 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
1173 unsigned int elem = size_index_elem(i);
1175 if (elem >= ARRAY_SIZE(size_index))
1177 size_index[elem] = KMALLOC_SHIFT_LOW;
1180 if (KMALLOC_MIN_SIZE >= 64) {
1182 * The 96 byte size cache is not used if the alignment
1185 for (i = 64 + 8; i <= 96; i += 8)
1186 size_index[size_index_elem(i)] = 7;
1190 if (KMALLOC_MIN_SIZE >= 128) {
1192 * The 192 byte sized cache is not used if the alignment
1193 * is 128 byte. Redirect kmalloc to use the 256 byte cache
1196 for (i = 128 + 8; i <= 192; i += 8)
1197 size_index[size_index_elem(i)] = 8;
1202 kmalloc_cache_name(const char *prefix, unsigned int size)
1205 static const char units[3] = "\0kM";
1208 while (size >= 1024 && (size % 1024 == 0)) {
1213 return kasprintf(GFP_NOWAIT, "%s-%u%c", prefix, size, units[idx]);
1217 new_kmalloc_cache(int idx, int type, slab_flags_t flags)
1221 if (type == KMALLOC_RECLAIM) {
1222 flags |= SLAB_RECLAIM_ACCOUNT;
1223 name = kmalloc_cache_name("kmalloc-rcl",
1224 kmalloc_info[idx].size);
1227 name = kmalloc_info[idx].name;
1230 kmalloc_caches[type][idx] = create_kmalloc_cache(name,
1231 kmalloc_info[idx].size, flags, 0,
1232 kmalloc_info[idx].size);
1236 * Create the kmalloc array. Some of the regular kmalloc arrays
1237 * may already have been created because they were needed to
1238 * enable allocations for slab creation.
1240 void __init create_kmalloc_caches(slab_flags_t flags)
1244 for (type = KMALLOC_NORMAL; type <= KMALLOC_RECLAIM; type++) {
1245 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) {
1246 if (!kmalloc_caches[type][i])
1247 new_kmalloc_cache(i, type, flags);
1250 * Caches that are not of the two-to-the-power-of size.
1251 * These have to be created immediately after the
1252 * earlier power of two caches
1254 if (KMALLOC_MIN_SIZE <= 32 && i == 6 &&
1255 !kmalloc_caches[type][1])
1256 new_kmalloc_cache(1, type, flags);
1257 if (KMALLOC_MIN_SIZE <= 64 && i == 7 &&
1258 !kmalloc_caches[type][2])
1259 new_kmalloc_cache(2, type, flags);
1263 /* Kmalloc array is now usable */
1266 #ifdef CONFIG_ZONE_DMA
1267 for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
1268 struct kmem_cache *s = kmalloc_caches[KMALLOC_NORMAL][i];
1271 unsigned int size = kmalloc_size(i);
1272 const char *n = kmalloc_cache_name("dma-kmalloc", size);
1275 kmalloc_caches[KMALLOC_DMA][i] = create_kmalloc_cache(
1276 n, size, SLAB_CACHE_DMA | flags, 0, 0);
1281 #endif /* !CONFIG_SLOB */
1284 * To avoid unnecessary overhead, we pass through large allocation requests
1285 * directly to the page allocator. We use __GFP_COMP, because we will need to
1286 * know the allocation order to free the pages properly in kfree.
1288 void *kmalloc_order(size_t size, gfp_t flags, unsigned int order)
1293 flags |= __GFP_COMP;
1294 page = alloc_pages(flags, order);
1295 ret = page ? page_address(page) : NULL;
1296 ret = kasan_kmalloc_large(ret, size, flags);
1297 /* As ret might get tagged, call kmemleak hook after KASAN. */
1298 kmemleak_alloc(ret, size, 1, flags);
1301 EXPORT_SYMBOL(kmalloc_order);
1303 #ifdef CONFIG_TRACING
1304 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
1306 void *ret = kmalloc_order(size, flags, order);
1307 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
1310 EXPORT_SYMBOL(kmalloc_order_trace);
1313 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1314 /* Randomize a generic freelist */
1315 static void freelist_randomize(struct rnd_state *state, unsigned int *list,
1321 for (i = 0; i < count; i++)
1324 /* Fisher-Yates shuffle */
1325 for (i = count - 1; i > 0; i--) {
1326 rand = prandom_u32_state(state);
1328 swap(list[i], list[rand]);
1332 /* Create a random sequence per cache */
1333 int cache_random_seq_create(struct kmem_cache *cachep, unsigned int count,
1336 struct rnd_state state;
1338 if (count < 2 || cachep->random_seq)
1341 cachep->random_seq = kcalloc(count, sizeof(unsigned int), gfp);
1342 if (!cachep->random_seq)
1345 /* Get best entropy at this stage of boot */
1346 prandom_seed_state(&state, get_random_long());
1348 freelist_randomize(&state, cachep->random_seq, count);
1352 /* Destroy the per-cache random freelist sequence */
1353 void cache_random_seq_destroy(struct kmem_cache *cachep)
1355 kfree(cachep->random_seq);
1356 cachep->random_seq = NULL;
1358 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1360 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
1362 #define SLABINFO_RIGHTS (0600)
1364 #define SLABINFO_RIGHTS (0400)
1367 static void print_slabinfo_header(struct seq_file *m)
1370 * Output format version, so at least we can change it
1371 * without _too_ many complaints.
1373 #ifdef CONFIG_DEBUG_SLAB
1374 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
1376 seq_puts(m, "slabinfo - version: 2.1\n");
1378 seq_puts(m, "# name <active_objs> <num_objs> <objsize> <objperslab> <pagesperslab>");
1379 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
1380 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
1381 #ifdef CONFIG_DEBUG_SLAB
1382 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> <error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
1383 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
1388 void *slab_start(struct seq_file *m, loff_t *pos)
1390 mutex_lock(&slab_mutex);
1391 return seq_list_start(&slab_root_caches, *pos);
1394 void *slab_next(struct seq_file *m, void *p, loff_t *pos)
1396 return seq_list_next(p, &slab_root_caches, pos);
1399 void slab_stop(struct seq_file *m, void *p)
1401 mutex_unlock(&slab_mutex);
1405 memcg_accumulate_slabinfo(struct kmem_cache *s, struct slabinfo *info)
1407 struct kmem_cache *c;
1408 struct slabinfo sinfo;
1410 if (!is_root_cache(s))
1413 for_each_memcg_cache(c, s) {
1414 memset(&sinfo, 0, sizeof(sinfo));
1415 get_slabinfo(c, &sinfo);
1417 info->active_slabs += sinfo.active_slabs;
1418 info->num_slabs += sinfo.num_slabs;
1419 info->shared_avail += sinfo.shared_avail;
1420 info->active_objs += sinfo.active_objs;
1421 info->num_objs += sinfo.num_objs;
1425 static void cache_show(struct kmem_cache *s, struct seq_file *m)
1427 struct slabinfo sinfo;
1429 memset(&sinfo, 0, sizeof(sinfo));
1430 get_slabinfo(s, &sinfo);
1432 memcg_accumulate_slabinfo(s, &sinfo);
1434 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
1435 cache_name(s), sinfo.active_objs, sinfo.num_objs, s->size,
1436 sinfo.objects_per_slab, (1 << sinfo.cache_order));
1438 seq_printf(m, " : tunables %4u %4u %4u",
1439 sinfo.limit, sinfo.batchcount, sinfo.shared);
1440 seq_printf(m, " : slabdata %6lu %6lu %6lu",
1441 sinfo.active_slabs, sinfo.num_slabs, sinfo.shared_avail);
1442 slabinfo_show_stats(m, s);
1446 static int slab_show(struct seq_file *m, void *p)
1448 struct kmem_cache *s = list_entry(p, struct kmem_cache, root_caches_node);
1450 if (p == slab_root_caches.next)
1451 print_slabinfo_header(m);
1456 void dump_unreclaimable_slab(void)
1458 struct kmem_cache *s, *s2;
1459 struct slabinfo sinfo;
1462 * Here acquiring slab_mutex is risky since we don't prefer to get
1463 * sleep in oom path. But, without mutex hold, it may introduce a
1465 * Use mutex_trylock to protect the list traverse, dump nothing
1466 * without acquiring the mutex.
1468 if (!mutex_trylock(&slab_mutex)) {
1469 pr_warn("excessive unreclaimable slab but cannot dump stats\n");
1473 pr_info("Unreclaimable slab info:\n");
1474 pr_info("Name Used Total\n");
1476 list_for_each_entry_safe(s, s2, &slab_caches, list) {
1477 if (!is_root_cache(s) || (s->flags & SLAB_RECLAIM_ACCOUNT))
1480 get_slabinfo(s, &sinfo);
1482 if (sinfo.num_objs > 0)
1483 pr_info("%-17s %10luKB %10luKB\n", cache_name(s),
1484 (sinfo.active_objs * s->size) / 1024,
1485 (sinfo.num_objs * s->size) / 1024);
1487 mutex_unlock(&slab_mutex);
1490 #if defined(CONFIG_MEMCG)
1491 void *memcg_slab_start(struct seq_file *m, loff_t *pos)
1493 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
1495 mutex_lock(&slab_mutex);
1496 return seq_list_start(&memcg->kmem_caches, *pos);
1499 void *memcg_slab_next(struct seq_file *m, void *p, loff_t *pos)
1501 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
1503 return seq_list_next(p, &memcg->kmem_caches, pos);
1506 void memcg_slab_stop(struct seq_file *m, void *p)
1508 mutex_unlock(&slab_mutex);
1511 int memcg_slab_show(struct seq_file *m, void *p)
1513 struct kmem_cache *s = list_entry(p, struct kmem_cache,
1514 memcg_params.kmem_caches_node);
1515 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
1517 if (p == memcg->kmem_caches.next)
1518 print_slabinfo_header(m);
1525 * slabinfo_op - iterator that generates /proc/slabinfo
1534 * num-pages-per-slab
1535 * + further values on SMP and with statistics enabled
1537 static const struct seq_operations slabinfo_op = {
1538 .start = slab_start,
1544 static int slabinfo_open(struct inode *inode, struct file *file)
1546 return seq_open(file, &slabinfo_op);
1549 static const struct file_operations proc_slabinfo_operations = {
1550 .open = slabinfo_open,
1552 .write = slabinfo_write,
1553 .llseek = seq_lseek,
1554 .release = seq_release,
1557 static int __init slab_proc_init(void)
1559 proc_create("slabinfo", SLABINFO_RIGHTS, NULL,
1560 &proc_slabinfo_operations);
1563 module_init(slab_proc_init);
1565 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_MEMCG_KMEM)
1567 * Display information about kmem caches that have child memcg caches.
1569 static int memcg_slabinfo_show(struct seq_file *m, void *unused)
1571 struct kmem_cache *s, *c;
1572 struct slabinfo sinfo;
1574 mutex_lock(&slab_mutex);
1575 seq_puts(m, "# <name> <css_id[:dead|deact]> <active_objs> <num_objs>");
1576 seq_puts(m, " <active_slabs> <num_slabs>\n");
1577 list_for_each_entry(s, &slab_root_caches, root_caches_node) {
1579 * Skip kmem caches that don't have any memcg children.
1581 if (list_empty(&s->memcg_params.children))
1584 memset(&sinfo, 0, sizeof(sinfo));
1585 get_slabinfo(s, &sinfo);
1586 seq_printf(m, "%-17s root %6lu %6lu %6lu %6lu\n",
1587 cache_name(s), sinfo.active_objs, sinfo.num_objs,
1588 sinfo.active_slabs, sinfo.num_slabs);
1590 for_each_memcg_cache(c, s) {
1591 struct cgroup_subsys_state *css;
1594 css = &c->memcg_params.memcg->css;
1595 if (!(css->flags & CSS_ONLINE))
1597 else if (c->flags & SLAB_DEACTIVATED)
1600 memset(&sinfo, 0, sizeof(sinfo));
1601 get_slabinfo(c, &sinfo);
1602 seq_printf(m, "%-17s %4d%-6s %6lu %6lu %6lu %6lu\n",
1603 cache_name(c), css->id, status,
1604 sinfo.active_objs, sinfo.num_objs,
1605 sinfo.active_slabs, sinfo.num_slabs);
1608 mutex_unlock(&slab_mutex);
1611 DEFINE_SHOW_ATTRIBUTE(memcg_slabinfo);
1613 static int __init memcg_slabinfo_init(void)
1615 debugfs_create_file("memcg_slabinfo", S_IFREG | S_IRUGO,
1616 NULL, NULL, &memcg_slabinfo_fops);
1620 late_initcall(memcg_slabinfo_init);
1621 #endif /* CONFIG_DEBUG_FS && CONFIG_MEMCG_KMEM */
1622 #endif /* CONFIG_SLAB || CONFIG_SLUB_DEBUG */
1624 static __always_inline void *__do_krealloc(const void *p, size_t new_size,
1633 if (ks >= new_size) {
1634 p = kasan_krealloc((void *)p, new_size, flags);
1638 ret = kmalloc_track_caller(new_size, flags);
1646 * __krealloc - like krealloc() but don't free @p.
1647 * @p: object to reallocate memory for.
1648 * @new_size: how many bytes of memory are required.
1649 * @flags: the type of memory to allocate.
1651 * This function is like krealloc() except it never frees the originally
1652 * allocated buffer. Use this if you don't want to free the buffer immediately
1653 * like, for example, with RCU.
1655 * Return: pointer to the allocated memory or %NULL in case of error
1657 void *__krealloc(const void *p, size_t new_size, gfp_t flags)
1659 if (unlikely(!new_size))
1660 return ZERO_SIZE_PTR;
1662 return __do_krealloc(p, new_size, flags);
1665 EXPORT_SYMBOL(__krealloc);
1668 * krealloc - reallocate memory. The contents will remain unchanged.
1669 * @p: object to reallocate memory for.
1670 * @new_size: how many bytes of memory are required.
1671 * @flags: the type of memory to allocate.
1673 * The contents of the object pointed to are preserved up to the
1674 * lesser of the new and old sizes. If @p is %NULL, krealloc()
1675 * behaves exactly like kmalloc(). If @new_size is 0 and @p is not a
1676 * %NULL pointer, the object pointed to is freed.
1678 * Return: pointer to the allocated memory or %NULL in case of error
1680 void *krealloc(const void *p, size_t new_size, gfp_t flags)
1684 if (unlikely(!new_size)) {
1686 return ZERO_SIZE_PTR;
1689 ret = __do_krealloc(p, new_size, flags);
1690 if (ret && kasan_reset_tag(p) != kasan_reset_tag(ret))
1695 EXPORT_SYMBOL(krealloc);
1698 * kzfree - like kfree but zero memory
1699 * @p: object to free memory of
1701 * The memory of the object @p points to is zeroed before freed.
1702 * If @p is %NULL, kzfree() does nothing.
1704 * Note: this function zeroes the whole allocated buffer which can be a good
1705 * deal bigger than the requested buffer size passed to kmalloc(). So be
1706 * careful when using this function in performance sensitive code.
1708 void kzfree(const void *p)
1711 void *mem = (void *)p;
1713 if (unlikely(ZERO_OR_NULL_PTR(mem)))
1719 EXPORT_SYMBOL(kzfree);
1722 * ksize - get the actual amount of memory allocated for a given object
1723 * @objp: Pointer to the object
1725 * kmalloc may internally round up allocations and return more memory
1726 * than requested. ksize() can be used to determine the actual amount of
1727 * memory allocated. The caller may use this additional memory, even though
1728 * a smaller amount of memory was initially specified with the kmalloc call.
1729 * The caller must guarantee that objp points to a valid object previously
1730 * allocated with either kmalloc() or kmem_cache_alloc(). The object
1731 * must not be freed during the duration of the call.
1733 * Return: size of the actual memory used by @objp in bytes
1735 size_t ksize(const void *objp)
1739 if (WARN_ON_ONCE(!objp))
1742 * We need to check that the pointed to object is valid, and only then
1743 * unpoison the shadow memory below. We use __kasan_check_read(), to
1744 * generate a more useful report at the time ksize() is called (rather
1745 * than later where behaviour is undefined due to potential
1746 * use-after-free or double-free).
1748 * If the pointed to memory is invalid we return 0, to avoid users of
1749 * ksize() writing to and potentially corrupting the memory region.
1751 * We want to perform the check before __ksize(), to avoid potentially
1752 * crashing in __ksize() due to accessing invalid metadata.
1754 if (unlikely(objp == ZERO_SIZE_PTR) || !__kasan_check_read(objp, 1))
1757 size = __ksize(objp);
1759 * We assume that ksize callers could use whole allocated area,
1760 * so we need to unpoison this area.
1762 kasan_unpoison_shadow(objp, size);
1765 EXPORT_SYMBOL(ksize);
1767 /* Tracepoints definitions. */
1768 EXPORT_TRACEPOINT_SYMBOL(kmalloc);
1769 EXPORT_TRACEPOINT_SYMBOL(kmem_cache_alloc);
1770 EXPORT_TRACEPOINT_SYMBOL(kmalloc_node);
1771 EXPORT_TRACEPOINT_SYMBOL(kmem_cache_alloc_node);
1772 EXPORT_TRACEPOINT_SYMBOL(kfree);
1773 EXPORT_TRACEPOINT_SYMBOL(kmem_cache_free);
1775 int should_failslab(struct kmem_cache *s, gfp_t gfpflags)
1777 if (__should_failslab(s, gfpflags))
1781 ALLOW_ERROR_INJECTION(should_failslab, ERRNO);