3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same initializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'slab_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/slab.h>
91 #include <linux/poison.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/cpuset.h>
98 #include <linux/proc_fs.h>
99 #include <linux/seq_file.h>
100 #include <linux/notifier.h>
101 #include <linux/kallsyms.h>
102 #include <linux/cpu.h>
103 #include <linux/sysctl.h>
104 #include <linux/module.h>
105 #include <linux/rcupdate.h>
106 #include <linux/string.h>
107 #include <linux/uaccess.h>
108 #include <linux/nodemask.h>
109 #include <linux/kmemleak.h>
110 #include <linux/mempolicy.h>
111 #include <linux/mutex.h>
112 #include <linux/fault-inject.h>
113 #include <linux/rtmutex.h>
114 #include <linux/reciprocal_div.h>
115 #include <linux/debugobjects.h>
116 #include <linux/kmemcheck.h>
117 #include <linux/memory.h>
118 #include <linux/prefetch.h>
120 #include <net/sock.h>
122 #include <asm/cacheflush.h>
123 #include <asm/tlbflush.h>
124 #include <asm/page.h>
126 #include <trace/events/kmem.h>
128 #include "internal.h"
133 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
134 * 0 for faster, smaller code (especially in the critical paths).
136 * STATS - 1 to collect stats for /proc/slabinfo.
137 * 0 for faster, smaller code (especially in the critical paths).
139 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
142 #ifdef CONFIG_DEBUG_SLAB
145 #define FORCED_DEBUG 1
149 #define FORCED_DEBUG 0
152 /* Shouldn't this be in a header file somewhere? */
153 #define BYTES_PER_WORD sizeof(void *)
154 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
156 #ifndef ARCH_KMALLOC_FLAGS
157 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
160 #define FREELIST_BYTE_INDEX (((PAGE_SIZE >> BITS_PER_BYTE) \
161 <= SLAB_OBJ_MIN_SIZE) ? 1 : 0)
163 #if FREELIST_BYTE_INDEX
164 typedef unsigned char freelist_idx_t;
166 typedef unsigned short freelist_idx_t;
169 #define SLAB_OBJ_MAX_NUM ((1 << sizeof(freelist_idx_t) * BITS_PER_BYTE) - 1)
175 * - LIFO ordering, to hand out cache-warm objects from _alloc
176 * - reduce the number of linked list operations
177 * - reduce spinlock operations
179 * The limit is stored in the per-cpu structure to reduce the data cache
186 unsigned int batchcount;
187 unsigned int touched;
189 * Must have this definition in here for the proper
190 * alignment of array_cache. Also simplifies accessing
197 struct array_cache ac;
201 * Need this for bootstrapping a per node allocator.
203 #define NUM_INIT_LISTS (2 * MAX_NUMNODES)
204 static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
205 #define CACHE_CACHE 0
206 #define SIZE_NODE (MAX_NUMNODES)
208 static int drain_freelist(struct kmem_cache *cache,
209 struct kmem_cache_node *n, int tofree);
210 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
211 int node, struct list_head *list);
212 static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list);
213 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
214 static void cache_reap(struct work_struct *unused);
216 static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
218 static inline void fixup_slab_list(struct kmem_cache *cachep,
219 struct kmem_cache_node *n, struct page *page,
221 static int slab_early_init = 1;
223 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
225 static void kmem_cache_node_init(struct kmem_cache_node *parent)
227 INIT_LIST_HEAD(&parent->slabs_full);
228 INIT_LIST_HEAD(&parent->slabs_partial);
229 INIT_LIST_HEAD(&parent->slabs_free);
230 parent->shared = NULL;
231 parent->alien = NULL;
232 parent->colour_next = 0;
233 spin_lock_init(&parent->list_lock);
234 parent->free_objects = 0;
235 parent->free_touched = 0;
236 parent->num_slabs = 0;
239 #define MAKE_LIST(cachep, listp, slab, nodeid) \
241 INIT_LIST_HEAD(listp); \
242 list_splice(&get_node(cachep, nodeid)->slab, listp); \
245 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
247 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
248 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
249 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
252 #define CFLGS_OBJFREELIST_SLAB (0x40000000UL)
253 #define CFLGS_OFF_SLAB (0x80000000UL)
254 #define OBJFREELIST_SLAB(x) ((x)->flags & CFLGS_OBJFREELIST_SLAB)
255 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
257 #define BATCHREFILL_LIMIT 16
259 * Optimization question: fewer reaps means less probability for unnessary
260 * cpucache drain/refill cycles.
262 * OTOH the cpuarrays can contain lots of objects,
263 * which could lock up otherwise freeable slabs.
265 #define REAPTIMEOUT_AC (2*HZ)
266 #define REAPTIMEOUT_NODE (4*HZ)
269 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
270 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
271 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
272 #define STATS_INC_GROWN(x) ((x)->grown++)
273 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
274 #define STATS_SET_HIGH(x) \
276 if ((x)->num_active > (x)->high_mark) \
277 (x)->high_mark = (x)->num_active; \
279 #define STATS_INC_ERR(x) ((x)->errors++)
280 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
281 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
282 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
283 #define STATS_SET_FREEABLE(x, i) \
285 if ((x)->max_freeable < i) \
286 (x)->max_freeable = i; \
288 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
289 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
290 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
291 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
293 #define STATS_INC_ACTIVE(x) do { } while (0)
294 #define STATS_DEC_ACTIVE(x) do { } while (0)
295 #define STATS_INC_ALLOCED(x) do { } while (0)
296 #define STATS_INC_GROWN(x) do { } while (0)
297 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
298 #define STATS_SET_HIGH(x) do { } while (0)
299 #define STATS_INC_ERR(x) do { } while (0)
300 #define STATS_INC_NODEALLOCS(x) do { } while (0)
301 #define STATS_INC_NODEFREES(x) do { } while (0)
302 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
303 #define STATS_SET_FREEABLE(x, i) do { } while (0)
304 #define STATS_INC_ALLOCHIT(x) do { } while (0)
305 #define STATS_INC_ALLOCMISS(x) do { } while (0)
306 #define STATS_INC_FREEHIT(x) do { } while (0)
307 #define STATS_INC_FREEMISS(x) do { } while (0)
313 * memory layout of objects:
315 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
316 * the end of an object is aligned with the end of the real
317 * allocation. Catches writes behind the end of the allocation.
318 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
320 * cachep->obj_offset: The real object.
321 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
322 * cachep->size - 1* BYTES_PER_WORD: last caller address
323 * [BYTES_PER_WORD long]
325 static int obj_offset(struct kmem_cache *cachep)
327 return cachep->obj_offset;
330 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
332 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
333 return (unsigned long long*) (objp + obj_offset(cachep) -
334 sizeof(unsigned long long));
337 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
339 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
340 if (cachep->flags & SLAB_STORE_USER)
341 return (unsigned long long *)(objp + cachep->size -
342 sizeof(unsigned long long) -
344 return (unsigned long long *) (objp + cachep->size -
345 sizeof(unsigned long long));
348 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
350 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
351 return (void **)(objp + cachep->size - BYTES_PER_WORD);
356 #define obj_offset(x) 0
357 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
358 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
359 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
363 #ifdef CONFIG_DEBUG_SLAB_LEAK
365 static inline bool is_store_user_clean(struct kmem_cache *cachep)
367 return atomic_read(&cachep->store_user_clean) == 1;
370 static inline void set_store_user_clean(struct kmem_cache *cachep)
372 atomic_set(&cachep->store_user_clean, 1);
375 static inline void set_store_user_dirty(struct kmem_cache *cachep)
377 if (is_store_user_clean(cachep))
378 atomic_set(&cachep->store_user_clean, 0);
382 static inline void set_store_user_dirty(struct kmem_cache *cachep) {}
387 * Do not go above this order unless 0 objects fit into the slab or
388 * overridden on the command line.
390 #define SLAB_MAX_ORDER_HI 1
391 #define SLAB_MAX_ORDER_LO 0
392 static int slab_max_order = SLAB_MAX_ORDER_LO;
393 static bool slab_max_order_set __initdata;
395 static inline struct kmem_cache *virt_to_cache(const void *obj)
397 struct page *page = virt_to_head_page(obj);
398 return page->slab_cache;
401 static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
404 return page->s_mem + cache->size * idx;
408 * We want to avoid an expensive divide : (offset / cache->size)
409 * Using the fact that size is a constant for a particular cache,
410 * we can replace (offset / cache->size) by
411 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
413 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
414 const struct page *page, void *obj)
416 u32 offset = (obj - page->s_mem);
417 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
420 #define BOOT_CPUCACHE_ENTRIES 1
421 /* internal cache of cache description objs */
422 static struct kmem_cache kmem_cache_boot = {
424 .limit = BOOT_CPUCACHE_ENTRIES,
426 .size = sizeof(struct kmem_cache),
427 .name = "kmem_cache",
430 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
432 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
434 return this_cpu_ptr(cachep->cpu_cache);
438 * Calculate the number of objects and left-over bytes for a given buffer size.
440 static unsigned int cache_estimate(unsigned long gfporder, size_t buffer_size,
441 unsigned long flags, size_t *left_over)
444 size_t slab_size = PAGE_SIZE << gfporder;
447 * The slab management structure can be either off the slab or
448 * on it. For the latter case, the memory allocated for a
451 * - @buffer_size bytes for each object
452 * - One freelist_idx_t for each object
454 * We don't need to consider alignment of freelist because
455 * freelist will be at the end of slab page. The objects will be
456 * at the correct alignment.
458 * If the slab management structure is off the slab, then the
459 * alignment will already be calculated into the size. Because
460 * the slabs are all pages aligned, the objects will be at the
461 * correct alignment when allocated.
463 if (flags & (CFLGS_OBJFREELIST_SLAB | CFLGS_OFF_SLAB)) {
464 num = slab_size / buffer_size;
465 *left_over = slab_size % buffer_size;
467 num = slab_size / (buffer_size + sizeof(freelist_idx_t));
468 *left_over = slab_size %
469 (buffer_size + sizeof(freelist_idx_t));
476 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
478 static void __slab_error(const char *function, struct kmem_cache *cachep,
481 pr_err("slab error in %s(): cache `%s': %s\n",
482 function, cachep->name, msg);
484 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
489 * By default on NUMA we use alien caches to stage the freeing of
490 * objects allocated from other nodes. This causes massive memory
491 * inefficiencies when using fake NUMA setup to split memory into a
492 * large number of small nodes, so it can be disabled on the command
496 static int use_alien_caches __read_mostly = 1;
497 static int __init noaliencache_setup(char *s)
499 use_alien_caches = 0;
502 __setup("noaliencache", noaliencache_setup);
504 static int __init slab_max_order_setup(char *str)
506 get_option(&str, &slab_max_order);
507 slab_max_order = slab_max_order < 0 ? 0 :
508 min(slab_max_order, MAX_ORDER - 1);
509 slab_max_order_set = true;
513 __setup("slab_max_order=", slab_max_order_setup);
517 * Special reaping functions for NUMA systems called from cache_reap().
518 * These take care of doing round robin flushing of alien caches (containing
519 * objects freed on different nodes from which they were allocated) and the
520 * flushing of remote pcps by calling drain_node_pages.
522 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
524 static void init_reap_node(int cpu)
526 per_cpu(slab_reap_node, cpu) = next_node_in(cpu_to_mem(cpu),
530 static void next_reap_node(void)
532 int node = __this_cpu_read(slab_reap_node);
534 node = next_node_in(node, node_online_map);
535 __this_cpu_write(slab_reap_node, node);
539 #define init_reap_node(cpu) do { } while (0)
540 #define next_reap_node(void) do { } while (0)
544 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
545 * via the workqueue/eventd.
546 * Add the CPU number into the expiration time to minimize the possibility of
547 * the CPUs getting into lockstep and contending for the global cache chain
550 static void start_cpu_timer(int cpu)
552 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
555 * When this gets called from do_initcalls via cpucache_init(),
556 * init_workqueues() has already run, so keventd will be setup
559 if (keventd_up() && reap_work->work.func == NULL) {
561 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
562 schedule_delayed_work_on(cpu, reap_work,
563 __round_jiffies_relative(HZ, cpu));
567 static void init_arraycache(struct array_cache *ac, int limit, int batch)
570 * The array_cache structures contain pointers to free object.
571 * However, when such objects are allocated or transferred to another
572 * cache the pointers are not cleared and they could be counted as
573 * valid references during a kmemleak scan. Therefore, kmemleak must
574 * not scan such objects.
576 kmemleak_no_scan(ac);
580 ac->batchcount = batch;
585 static struct array_cache *alloc_arraycache(int node, int entries,
586 int batchcount, gfp_t gfp)
588 size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache);
589 struct array_cache *ac = NULL;
591 ac = kmalloc_node(memsize, gfp, node);
592 init_arraycache(ac, entries, batchcount);
596 static noinline void cache_free_pfmemalloc(struct kmem_cache *cachep,
597 struct page *page, void *objp)
599 struct kmem_cache_node *n;
603 page_node = page_to_nid(page);
604 n = get_node(cachep, page_node);
606 spin_lock(&n->list_lock);
607 free_block(cachep, &objp, 1, page_node, &list);
608 spin_unlock(&n->list_lock);
610 slabs_destroy(cachep, &list);
614 * Transfer objects in one arraycache to another.
615 * Locking must be handled by the caller.
617 * Return the number of entries transferred.
619 static int transfer_objects(struct array_cache *to,
620 struct array_cache *from, unsigned int max)
622 /* Figure out how many entries to transfer */
623 int nr = min3(from->avail, max, to->limit - to->avail);
628 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
638 #define drain_alien_cache(cachep, alien) do { } while (0)
639 #define reap_alien(cachep, n) do { } while (0)
641 static inline struct alien_cache **alloc_alien_cache(int node,
642 int limit, gfp_t gfp)
647 static inline void free_alien_cache(struct alien_cache **ac_ptr)
651 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
656 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
662 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
663 gfp_t flags, int nodeid)
668 static inline gfp_t gfp_exact_node(gfp_t flags)
670 return flags & ~__GFP_NOFAIL;
673 #else /* CONFIG_NUMA */
675 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
676 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
678 static struct alien_cache *__alloc_alien_cache(int node, int entries,
679 int batch, gfp_t gfp)
681 size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
682 struct alien_cache *alc = NULL;
684 alc = kmalloc_node(memsize, gfp, node);
686 init_arraycache(&alc->ac, entries, batch);
687 spin_lock_init(&alc->lock);
692 static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
694 struct alien_cache **alc_ptr;
695 size_t memsize = sizeof(void *) * nr_node_ids;
700 alc_ptr = kzalloc_node(memsize, gfp, node);
705 if (i == node || !node_online(i))
707 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
709 for (i--; i >= 0; i--)
718 static void free_alien_cache(struct alien_cache **alc_ptr)
729 static void __drain_alien_cache(struct kmem_cache *cachep,
730 struct array_cache *ac, int node,
731 struct list_head *list)
733 struct kmem_cache_node *n = get_node(cachep, node);
736 spin_lock(&n->list_lock);
738 * Stuff objects into the remote nodes shared array first.
739 * That way we could avoid the overhead of putting the objects
740 * into the free lists and getting them back later.
743 transfer_objects(n->shared, ac, ac->limit);
745 free_block(cachep, ac->entry, ac->avail, node, list);
747 spin_unlock(&n->list_lock);
752 * Called from cache_reap() to regularly drain alien caches round robin.
754 static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
756 int node = __this_cpu_read(slab_reap_node);
759 struct alien_cache *alc = n->alien[node];
760 struct array_cache *ac;
764 if (ac->avail && spin_trylock_irq(&alc->lock)) {
767 __drain_alien_cache(cachep, ac, node, &list);
768 spin_unlock_irq(&alc->lock);
769 slabs_destroy(cachep, &list);
775 static void drain_alien_cache(struct kmem_cache *cachep,
776 struct alien_cache **alien)
779 struct alien_cache *alc;
780 struct array_cache *ac;
783 for_each_online_node(i) {
789 spin_lock_irqsave(&alc->lock, flags);
790 __drain_alien_cache(cachep, ac, i, &list);
791 spin_unlock_irqrestore(&alc->lock, flags);
792 slabs_destroy(cachep, &list);
797 static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
798 int node, int page_node)
800 struct kmem_cache_node *n;
801 struct alien_cache *alien = NULL;
802 struct array_cache *ac;
805 n = get_node(cachep, node);
806 STATS_INC_NODEFREES(cachep);
807 if (n->alien && n->alien[page_node]) {
808 alien = n->alien[page_node];
810 spin_lock(&alien->lock);
811 if (unlikely(ac->avail == ac->limit)) {
812 STATS_INC_ACOVERFLOW(cachep);
813 __drain_alien_cache(cachep, ac, page_node, &list);
815 ac->entry[ac->avail++] = objp;
816 spin_unlock(&alien->lock);
817 slabs_destroy(cachep, &list);
819 n = get_node(cachep, page_node);
820 spin_lock(&n->list_lock);
821 free_block(cachep, &objp, 1, page_node, &list);
822 spin_unlock(&n->list_lock);
823 slabs_destroy(cachep, &list);
828 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
830 int page_node = page_to_nid(virt_to_page(objp));
831 int node = numa_mem_id();
833 * Make sure we are not freeing a object from another node to the array
836 if (likely(node == page_node))
839 return __cache_free_alien(cachep, objp, node, page_node);
843 * Construct gfp mask to allocate from a specific node but do not reclaim or
844 * warn about failures.
846 static inline gfp_t gfp_exact_node(gfp_t flags)
848 return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
852 static int init_cache_node(struct kmem_cache *cachep, int node, gfp_t gfp)
854 struct kmem_cache_node *n;
857 * Set up the kmem_cache_node for cpu before we can
858 * begin anything. Make sure some other cpu on this
859 * node has not already allocated this
861 n = get_node(cachep, node);
863 spin_lock_irq(&n->list_lock);
864 n->free_limit = (1 + nr_cpus_node(node)) * cachep->batchcount +
866 spin_unlock_irq(&n->list_lock);
871 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
875 kmem_cache_node_init(n);
876 n->next_reap = jiffies + REAPTIMEOUT_NODE +
877 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
880 (1 + nr_cpus_node(node)) * cachep->batchcount + cachep->num;
883 * The kmem_cache_nodes don't come and go as CPUs
884 * come and go. slab_mutex is sufficient
887 cachep->node[node] = n;
892 #if (defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)) || defined(CONFIG_SMP)
894 * Allocates and initializes node for a node on each slab cache, used for
895 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
896 * will be allocated off-node since memory is not yet online for the new node.
897 * When hotplugging memory or a cpu, existing node are not replaced if
900 * Must hold slab_mutex.
902 static int init_cache_node_node(int node)
905 struct kmem_cache *cachep;
907 list_for_each_entry(cachep, &slab_caches, list) {
908 ret = init_cache_node(cachep, node, GFP_KERNEL);
917 static int setup_kmem_cache_node(struct kmem_cache *cachep,
918 int node, gfp_t gfp, bool force_change)
921 struct kmem_cache_node *n;
922 struct array_cache *old_shared = NULL;
923 struct array_cache *new_shared = NULL;
924 struct alien_cache **new_alien = NULL;
927 if (use_alien_caches) {
928 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
933 if (cachep->shared) {
934 new_shared = alloc_arraycache(node,
935 cachep->shared * cachep->batchcount, 0xbaadf00d, gfp);
940 ret = init_cache_node(cachep, node, gfp);
944 n = get_node(cachep, node);
945 spin_lock_irq(&n->list_lock);
946 if (n->shared && force_change) {
947 free_block(cachep, n->shared->entry,
948 n->shared->avail, node, &list);
949 n->shared->avail = 0;
952 if (!n->shared || force_change) {
953 old_shared = n->shared;
954 n->shared = new_shared;
959 n->alien = new_alien;
963 spin_unlock_irq(&n->list_lock);
964 slabs_destroy(cachep, &list);
967 * To protect lockless access to n->shared during irq disabled context.
968 * If n->shared isn't NULL in irq disabled context, accessing to it is
969 * guaranteed to be valid until irq is re-enabled, because it will be
970 * freed after synchronize_sched().
972 if (old_shared && force_change)
978 free_alien_cache(new_alien);
985 static void cpuup_canceled(long cpu)
987 struct kmem_cache *cachep;
988 struct kmem_cache_node *n = NULL;
989 int node = cpu_to_mem(cpu);
990 const struct cpumask *mask = cpumask_of_node(node);
992 list_for_each_entry(cachep, &slab_caches, list) {
993 struct array_cache *nc;
994 struct array_cache *shared;
995 struct alien_cache **alien;
998 n = get_node(cachep, node);
1002 spin_lock_irq(&n->list_lock);
1004 /* Free limit for this kmem_cache_node */
1005 n->free_limit -= cachep->batchcount;
1007 /* cpu is dead; no one can alloc from it. */
1008 nc = per_cpu_ptr(cachep->cpu_cache, cpu);
1010 free_block(cachep, nc->entry, nc->avail, node, &list);
1014 if (!cpumask_empty(mask)) {
1015 spin_unlock_irq(&n->list_lock);
1021 free_block(cachep, shared->entry,
1022 shared->avail, node, &list);
1029 spin_unlock_irq(&n->list_lock);
1033 drain_alien_cache(cachep, alien);
1034 free_alien_cache(alien);
1038 slabs_destroy(cachep, &list);
1041 * In the previous loop, all the objects were freed to
1042 * the respective cache's slabs, now we can go ahead and
1043 * shrink each nodelist to its limit.
1045 list_for_each_entry(cachep, &slab_caches, list) {
1046 n = get_node(cachep, node);
1049 drain_freelist(cachep, n, INT_MAX);
1053 static int cpuup_prepare(long cpu)
1055 struct kmem_cache *cachep;
1056 int node = cpu_to_mem(cpu);
1060 * We need to do this right in the beginning since
1061 * alloc_arraycache's are going to use this list.
1062 * kmalloc_node allows us to add the slab to the right
1063 * kmem_cache_node and not this cpu's kmem_cache_node
1065 err = init_cache_node_node(node);
1070 * Now we can go ahead with allocating the shared arrays and
1073 list_for_each_entry(cachep, &slab_caches, list) {
1074 err = setup_kmem_cache_node(cachep, node, GFP_KERNEL, false);
1081 cpuup_canceled(cpu);
1085 int slab_prepare_cpu(unsigned int cpu)
1089 mutex_lock(&slab_mutex);
1090 err = cpuup_prepare(cpu);
1091 mutex_unlock(&slab_mutex);
1096 * This is called for a failed online attempt and for a successful
1099 * Even if all the cpus of a node are down, we don't free the
1100 * kmem_list3 of any cache. This to avoid a race between cpu_down, and
1101 * a kmalloc allocation from another cpu for memory from the node of
1102 * the cpu going down. The list3 structure is usually allocated from
1103 * kmem_cache_create() and gets destroyed at kmem_cache_destroy().
1105 int slab_dead_cpu(unsigned int cpu)
1107 mutex_lock(&slab_mutex);
1108 cpuup_canceled(cpu);
1109 mutex_unlock(&slab_mutex);
1114 static int slab_online_cpu(unsigned int cpu)
1116 start_cpu_timer(cpu);
1120 static int slab_offline_cpu(unsigned int cpu)
1123 * Shutdown cache reaper. Note that the slab_mutex is held so
1124 * that if cache_reap() is invoked it cannot do anything
1125 * expensive but will only modify reap_work and reschedule the
1128 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1129 /* Now the cache_reaper is guaranteed to be not running. */
1130 per_cpu(slab_reap_work, cpu).work.func = NULL;
1134 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1136 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1137 * Returns -EBUSY if all objects cannot be drained so that the node is not
1140 * Must hold slab_mutex.
1142 static int __meminit drain_cache_node_node(int node)
1144 struct kmem_cache *cachep;
1147 list_for_each_entry(cachep, &slab_caches, list) {
1148 struct kmem_cache_node *n;
1150 n = get_node(cachep, node);
1154 drain_freelist(cachep, n, INT_MAX);
1156 if (!list_empty(&n->slabs_full) ||
1157 !list_empty(&n->slabs_partial)) {
1165 static int __meminit slab_memory_callback(struct notifier_block *self,
1166 unsigned long action, void *arg)
1168 struct memory_notify *mnb = arg;
1172 nid = mnb->status_change_nid;
1177 case MEM_GOING_ONLINE:
1178 mutex_lock(&slab_mutex);
1179 ret = init_cache_node_node(nid);
1180 mutex_unlock(&slab_mutex);
1182 case MEM_GOING_OFFLINE:
1183 mutex_lock(&slab_mutex);
1184 ret = drain_cache_node_node(nid);
1185 mutex_unlock(&slab_mutex);
1189 case MEM_CANCEL_ONLINE:
1190 case MEM_CANCEL_OFFLINE:
1194 return notifier_from_errno(ret);
1196 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1199 * swap the static kmem_cache_node with kmalloced memory
1201 static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1204 struct kmem_cache_node *ptr;
1206 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1209 memcpy(ptr, list, sizeof(struct kmem_cache_node));
1211 * Do not assume that spinlocks can be initialized via memcpy:
1213 spin_lock_init(&ptr->list_lock);
1215 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1216 cachep->node[nodeid] = ptr;
1220 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1221 * size of kmem_cache_node.
1223 static void __init set_up_node(struct kmem_cache *cachep, int index)
1227 for_each_online_node(node) {
1228 cachep->node[node] = &init_kmem_cache_node[index + node];
1229 cachep->node[node]->next_reap = jiffies +
1231 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1236 * Initialisation. Called after the page allocator have been initialised and
1237 * before smp_init().
1239 void __init kmem_cache_init(void)
1243 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1244 sizeof(struct rcu_head));
1245 kmem_cache = &kmem_cache_boot;
1247 if (!IS_ENABLED(CONFIG_NUMA) || num_possible_nodes() == 1)
1248 use_alien_caches = 0;
1250 for (i = 0; i < NUM_INIT_LISTS; i++)
1251 kmem_cache_node_init(&init_kmem_cache_node[i]);
1254 * Fragmentation resistance on low memory - only use bigger
1255 * page orders on machines with more than 32MB of memory if
1256 * not overridden on the command line.
1258 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1259 slab_max_order = SLAB_MAX_ORDER_HI;
1261 /* Bootstrap is tricky, because several objects are allocated
1262 * from caches that do not exist yet:
1263 * 1) initialize the kmem_cache cache: it contains the struct
1264 * kmem_cache structures of all caches, except kmem_cache itself:
1265 * kmem_cache is statically allocated.
1266 * Initially an __init data area is used for the head array and the
1267 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1268 * array at the end of the bootstrap.
1269 * 2) Create the first kmalloc cache.
1270 * The struct kmem_cache for the new cache is allocated normally.
1271 * An __init data area is used for the head array.
1272 * 3) Create the remaining kmalloc caches, with minimally sized
1274 * 4) Replace the __init data head arrays for kmem_cache and the first
1275 * kmalloc cache with kmalloc allocated arrays.
1276 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1277 * the other cache's with kmalloc allocated memory.
1278 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1281 /* 1) create the kmem_cache */
1284 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1286 create_boot_cache(kmem_cache, "kmem_cache",
1287 offsetof(struct kmem_cache, node) +
1288 nr_node_ids * sizeof(struct kmem_cache_node *),
1289 SLAB_HWCACHE_ALIGN);
1290 list_add(&kmem_cache->list, &slab_caches);
1291 slab_state = PARTIAL;
1294 * Initialize the caches that provide memory for the kmem_cache_node
1295 * structures first. Without this, further allocations will bug.
1297 kmalloc_caches[INDEX_NODE] = create_kmalloc_cache("kmalloc-node",
1298 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
1299 slab_state = PARTIAL_NODE;
1300 setup_kmalloc_cache_index_table();
1302 slab_early_init = 0;
1304 /* 5) Replace the bootstrap kmem_cache_node */
1308 for_each_online_node(nid) {
1309 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
1311 init_list(kmalloc_caches[INDEX_NODE],
1312 &init_kmem_cache_node[SIZE_NODE + nid], nid);
1316 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1319 void __init kmem_cache_init_late(void)
1321 struct kmem_cache *cachep;
1325 /* 6) resize the head arrays to their final sizes */
1326 mutex_lock(&slab_mutex);
1327 list_for_each_entry(cachep, &slab_caches, list)
1328 if (enable_cpucache(cachep, GFP_NOWAIT))
1330 mutex_unlock(&slab_mutex);
1337 * Register a memory hotplug callback that initializes and frees
1340 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1344 * The reap timers are started later, with a module init call: That part
1345 * of the kernel is not yet operational.
1349 static int __init cpucache_init(void)
1354 * Register the timers that return unneeded pages to the page allocator
1356 ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "SLAB online",
1357 slab_online_cpu, slab_offline_cpu);
1364 __initcall(cpucache_init);
1366 static noinline void
1367 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1370 struct kmem_cache_node *n;
1372 unsigned long flags;
1374 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1375 DEFAULT_RATELIMIT_BURST);
1377 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1380 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1381 nodeid, gfpflags, &gfpflags);
1382 pr_warn(" cache: %s, object size: %d, order: %d\n",
1383 cachep->name, cachep->size, cachep->gfporder);
1385 for_each_kmem_cache_node(cachep, node, n) {
1386 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1387 unsigned long active_slabs = 0, num_slabs = 0;
1388 unsigned long num_slabs_partial = 0, num_slabs_free = 0;
1389 unsigned long num_slabs_full;
1391 spin_lock_irqsave(&n->list_lock, flags);
1392 num_slabs = n->num_slabs;
1393 list_for_each_entry(page, &n->slabs_partial, lru) {
1394 active_objs += page->active;
1395 num_slabs_partial++;
1397 list_for_each_entry(page, &n->slabs_free, lru)
1400 free_objects += n->free_objects;
1401 spin_unlock_irqrestore(&n->list_lock, flags);
1403 num_objs = num_slabs * cachep->num;
1404 active_slabs = num_slabs - num_slabs_free;
1405 num_slabs_full = num_slabs -
1406 (num_slabs_partial + num_slabs_free);
1407 active_objs += (num_slabs_full * cachep->num);
1409 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1410 node, active_slabs, num_slabs, active_objs, num_objs,
1417 * Interface to system's page allocator. No need to hold the
1418 * kmem_cache_node ->list_lock.
1420 * If we requested dmaable memory, we will get it. Even if we
1421 * did not request dmaable memory, we might get it, but that
1422 * would be relatively rare and ignorable.
1424 static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1430 flags |= cachep->allocflags;
1431 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1432 flags |= __GFP_RECLAIMABLE;
1434 page = __alloc_pages_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1436 slab_out_of_memory(cachep, flags, nodeid);
1440 if (memcg_charge_slab(page, flags, cachep->gfporder, cachep)) {
1441 __free_pages(page, cachep->gfporder);
1445 nr_pages = (1 << cachep->gfporder);
1446 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1447 add_zone_page_state(page_zone(page),
1448 NR_SLAB_RECLAIMABLE, nr_pages);
1450 add_zone_page_state(page_zone(page),
1451 NR_SLAB_UNRECLAIMABLE, nr_pages);
1453 __SetPageSlab(page);
1454 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1455 if (sk_memalloc_socks() && page_is_pfmemalloc(page))
1456 SetPageSlabPfmemalloc(page);
1458 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1459 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1462 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1464 kmemcheck_mark_unallocated_pages(page, nr_pages);
1471 * Interface to system's page release.
1473 static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
1475 int order = cachep->gfporder;
1476 unsigned long nr_freed = (1 << order);
1478 kmemcheck_free_shadow(page, order);
1480 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1481 sub_zone_page_state(page_zone(page),
1482 NR_SLAB_RECLAIMABLE, nr_freed);
1484 sub_zone_page_state(page_zone(page),
1485 NR_SLAB_UNRECLAIMABLE, nr_freed);
1487 BUG_ON(!PageSlab(page));
1488 __ClearPageSlabPfmemalloc(page);
1489 __ClearPageSlab(page);
1490 page_mapcount_reset(page);
1491 page->mapping = NULL;
1493 if (current->reclaim_state)
1494 current->reclaim_state->reclaimed_slab += nr_freed;
1495 memcg_uncharge_slab(page, order, cachep);
1496 __free_pages(page, order);
1499 static void kmem_rcu_free(struct rcu_head *head)
1501 struct kmem_cache *cachep;
1504 page = container_of(head, struct page, rcu_head);
1505 cachep = page->slab_cache;
1507 kmem_freepages(cachep, page);
1511 static bool is_debug_pagealloc_cache(struct kmem_cache *cachep)
1513 if (debug_pagealloc_enabled() && OFF_SLAB(cachep) &&
1514 (cachep->size % PAGE_SIZE) == 0)
1520 #ifdef CONFIG_DEBUG_PAGEALLOC
1521 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1522 unsigned long caller)
1524 int size = cachep->object_size;
1526 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1528 if (size < 5 * sizeof(unsigned long))
1531 *addr++ = 0x12345678;
1533 *addr++ = smp_processor_id();
1534 size -= 3 * sizeof(unsigned long);
1536 unsigned long *sptr = &caller;
1537 unsigned long svalue;
1539 while (!kstack_end(sptr)) {
1541 if (kernel_text_address(svalue)) {
1543 size -= sizeof(unsigned long);
1544 if (size <= sizeof(unsigned long))
1550 *addr++ = 0x87654321;
1553 static void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1554 int map, unsigned long caller)
1556 if (!is_debug_pagealloc_cache(cachep))
1560 store_stackinfo(cachep, objp, caller);
1562 kernel_map_pages(virt_to_page(objp), cachep->size / PAGE_SIZE, map);
1566 static inline void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1567 int map, unsigned long caller) {}
1571 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1573 int size = cachep->object_size;
1574 addr = &((char *)addr)[obj_offset(cachep)];
1576 memset(addr, val, size);
1577 *(unsigned char *)(addr + size - 1) = POISON_END;
1580 static void dump_line(char *data, int offset, int limit)
1583 unsigned char error = 0;
1586 pr_err("%03x: ", offset);
1587 for (i = 0; i < limit; i++) {
1588 if (data[offset + i] != POISON_FREE) {
1589 error = data[offset + i];
1593 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1594 &data[offset], limit, 1);
1596 if (bad_count == 1) {
1597 error ^= POISON_FREE;
1598 if (!(error & (error - 1))) {
1599 pr_err("Single bit error detected. Probably bad RAM.\n");
1601 pr_err("Run memtest86+ or a similar memory test tool.\n");
1603 pr_err("Run a memory test tool.\n");
1612 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1617 if (cachep->flags & SLAB_RED_ZONE) {
1618 pr_err("Redzone: 0x%llx/0x%llx\n",
1619 *dbg_redzone1(cachep, objp),
1620 *dbg_redzone2(cachep, objp));
1623 if (cachep->flags & SLAB_STORE_USER) {
1624 pr_err("Last user: [<%p>](%pSR)\n",
1625 *dbg_userword(cachep, objp),
1626 *dbg_userword(cachep, objp));
1628 realobj = (char *)objp + obj_offset(cachep);
1629 size = cachep->object_size;
1630 for (i = 0; i < size && lines; i += 16, lines--) {
1633 if (i + limit > size)
1635 dump_line(realobj, i, limit);
1639 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1645 if (is_debug_pagealloc_cache(cachep))
1648 realobj = (char *)objp + obj_offset(cachep);
1649 size = cachep->object_size;
1651 for (i = 0; i < size; i++) {
1652 char exp = POISON_FREE;
1655 if (realobj[i] != exp) {
1660 pr_err("Slab corruption (%s): %s start=%p, len=%d\n",
1661 print_tainted(), cachep->name,
1663 print_objinfo(cachep, objp, 0);
1665 /* Hexdump the affected line */
1668 if (i + limit > size)
1670 dump_line(realobj, i, limit);
1673 /* Limit to 5 lines */
1679 /* Print some data about the neighboring objects, if they
1682 struct page *page = virt_to_head_page(objp);
1685 objnr = obj_to_index(cachep, page, objp);
1687 objp = index_to_obj(cachep, page, objnr - 1);
1688 realobj = (char *)objp + obj_offset(cachep);
1689 pr_err("Prev obj: start=%p, len=%d\n", realobj, size);
1690 print_objinfo(cachep, objp, 2);
1692 if (objnr + 1 < cachep->num) {
1693 objp = index_to_obj(cachep, page, objnr + 1);
1694 realobj = (char *)objp + obj_offset(cachep);
1695 pr_err("Next obj: start=%p, len=%d\n", realobj, size);
1696 print_objinfo(cachep, objp, 2);
1703 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1708 if (OBJFREELIST_SLAB(cachep) && cachep->flags & SLAB_POISON) {
1709 poison_obj(cachep, page->freelist - obj_offset(cachep),
1713 for (i = 0; i < cachep->num; i++) {
1714 void *objp = index_to_obj(cachep, page, i);
1716 if (cachep->flags & SLAB_POISON) {
1717 check_poison_obj(cachep, objp);
1718 slab_kernel_map(cachep, objp, 1, 0);
1720 if (cachep->flags & SLAB_RED_ZONE) {
1721 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1722 slab_error(cachep, "start of a freed object was overwritten");
1723 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1724 slab_error(cachep, "end of a freed object was overwritten");
1729 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1736 * slab_destroy - destroy and release all objects in a slab
1737 * @cachep: cache pointer being destroyed
1738 * @page: page pointer being destroyed
1740 * Destroy all the objs in a slab page, and release the mem back to the system.
1741 * Before calling the slab page must have been unlinked from the cache. The
1742 * kmem_cache_node ->list_lock is not held/needed.
1744 static void slab_destroy(struct kmem_cache *cachep, struct page *page)
1748 freelist = page->freelist;
1749 slab_destroy_debugcheck(cachep, page);
1750 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
1751 call_rcu(&page->rcu_head, kmem_rcu_free);
1753 kmem_freepages(cachep, page);
1756 * From now on, we don't use freelist
1757 * although actual page can be freed in rcu context
1759 if (OFF_SLAB(cachep))
1760 kmem_cache_free(cachep->freelist_cache, freelist);
1763 static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1765 struct page *page, *n;
1767 list_for_each_entry_safe(page, n, list, lru) {
1768 list_del(&page->lru);
1769 slab_destroy(cachep, page);
1774 * calculate_slab_order - calculate size (page order) of slabs
1775 * @cachep: pointer to the cache that is being created
1776 * @size: size of objects to be created in this cache.
1777 * @flags: slab allocation flags
1779 * Also calculates the number of objects per slab.
1781 * This could be made much more intelligent. For now, try to avoid using
1782 * high order pages for slabs. When the gfp() functions are more friendly
1783 * towards high-order requests, this should be changed.
1785 static size_t calculate_slab_order(struct kmem_cache *cachep,
1786 size_t size, unsigned long flags)
1788 size_t left_over = 0;
1791 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
1795 num = cache_estimate(gfporder, size, flags, &remainder);
1799 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1800 if (num > SLAB_OBJ_MAX_NUM)
1803 if (flags & CFLGS_OFF_SLAB) {
1804 struct kmem_cache *freelist_cache;
1805 size_t freelist_size;
1807 freelist_size = num * sizeof(freelist_idx_t);
1808 freelist_cache = kmalloc_slab(freelist_size, 0u);
1809 if (!freelist_cache)
1813 * Needed to avoid possible looping condition
1814 * in cache_grow_begin()
1816 if (OFF_SLAB(freelist_cache))
1819 /* check if off slab has enough benefit */
1820 if (freelist_cache->size > cachep->size / 2)
1824 /* Found something acceptable - save it away */
1826 cachep->gfporder = gfporder;
1827 left_over = remainder;
1830 * A VFS-reclaimable slab tends to have most allocations
1831 * as GFP_NOFS and we really don't want to have to be allocating
1832 * higher-order pages when we are unable to shrink dcache.
1834 if (flags & SLAB_RECLAIM_ACCOUNT)
1838 * Large number of objects is good, but very large slabs are
1839 * currently bad for the gfp()s.
1841 if (gfporder >= slab_max_order)
1845 * Acceptable internal fragmentation?
1847 if (left_over * 8 <= (PAGE_SIZE << gfporder))
1853 static struct array_cache __percpu *alloc_kmem_cache_cpus(
1854 struct kmem_cache *cachep, int entries, int batchcount)
1858 struct array_cache __percpu *cpu_cache;
1860 size = sizeof(void *) * entries + sizeof(struct array_cache);
1861 cpu_cache = __alloc_percpu(size, sizeof(void *));
1866 for_each_possible_cpu(cpu) {
1867 init_arraycache(per_cpu_ptr(cpu_cache, cpu),
1868 entries, batchcount);
1874 static int __ref setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
1876 if (slab_state >= FULL)
1877 return enable_cpucache(cachep, gfp);
1879 cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
1880 if (!cachep->cpu_cache)
1883 if (slab_state == DOWN) {
1884 /* Creation of first cache (kmem_cache). */
1885 set_up_node(kmem_cache, CACHE_CACHE);
1886 } else if (slab_state == PARTIAL) {
1887 /* For kmem_cache_node */
1888 set_up_node(cachep, SIZE_NODE);
1892 for_each_online_node(node) {
1893 cachep->node[node] = kmalloc_node(
1894 sizeof(struct kmem_cache_node), gfp, node);
1895 BUG_ON(!cachep->node[node]);
1896 kmem_cache_node_init(cachep->node[node]);
1900 cachep->node[numa_mem_id()]->next_reap =
1901 jiffies + REAPTIMEOUT_NODE +
1902 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1904 cpu_cache_get(cachep)->avail = 0;
1905 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1906 cpu_cache_get(cachep)->batchcount = 1;
1907 cpu_cache_get(cachep)->touched = 0;
1908 cachep->batchcount = 1;
1909 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1913 unsigned long kmem_cache_flags(unsigned long object_size,
1914 unsigned long flags, const char *name,
1915 void (*ctor)(void *))
1921 __kmem_cache_alias(const char *name, size_t size, size_t align,
1922 unsigned long flags, void (*ctor)(void *))
1924 struct kmem_cache *cachep;
1926 cachep = find_mergeable(size, align, flags, name, ctor);
1931 * Adjust the object sizes so that we clear
1932 * the complete object on kzalloc.
1934 cachep->object_size = max_t(int, cachep->object_size, size);
1939 static bool set_objfreelist_slab_cache(struct kmem_cache *cachep,
1940 size_t size, unsigned long flags)
1946 if (cachep->ctor || flags & SLAB_DESTROY_BY_RCU)
1949 left = calculate_slab_order(cachep, size,
1950 flags | CFLGS_OBJFREELIST_SLAB);
1954 if (cachep->num * sizeof(freelist_idx_t) > cachep->object_size)
1957 cachep->colour = left / cachep->colour_off;
1962 static bool set_off_slab_cache(struct kmem_cache *cachep,
1963 size_t size, unsigned long flags)
1970 * Always use on-slab management when SLAB_NOLEAKTRACE
1971 * to avoid recursive calls into kmemleak.
1973 if (flags & SLAB_NOLEAKTRACE)
1977 * Size is large, assume best to place the slab management obj
1978 * off-slab (should allow better packing of objs).
1980 left = calculate_slab_order(cachep, size, flags | CFLGS_OFF_SLAB);
1985 * If the slab has been placed off-slab, and we have enough space then
1986 * move it on-slab. This is at the expense of any extra colouring.
1988 if (left >= cachep->num * sizeof(freelist_idx_t))
1991 cachep->colour = left / cachep->colour_off;
1996 static bool set_on_slab_cache(struct kmem_cache *cachep,
1997 size_t size, unsigned long flags)
2003 left = calculate_slab_order(cachep, size, flags);
2007 cachep->colour = left / cachep->colour_off;
2013 * __kmem_cache_create - Create a cache.
2014 * @cachep: cache management descriptor
2015 * @flags: SLAB flags
2017 * Returns a ptr to the cache on success, NULL on failure.
2018 * Cannot be called within a int, but can be interrupted.
2019 * The @ctor is run when new pages are allocated by the cache.
2023 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2024 * to catch references to uninitialised memory.
2026 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2027 * for buffer overruns.
2029 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2030 * cacheline. This can be beneficial if you're counting cycles as closely
2034 __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2036 size_t ralign = BYTES_PER_WORD;
2039 size_t size = cachep->size;
2044 * Enable redzoning and last user accounting, except for caches with
2045 * large objects, if the increased size would increase the object size
2046 * above the next power of two: caches with object sizes just above a
2047 * power of two have a significant amount of internal fragmentation.
2049 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2050 2 * sizeof(unsigned long long)))
2051 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2052 if (!(flags & SLAB_DESTROY_BY_RCU))
2053 flags |= SLAB_POISON;
2058 * Check that size is in terms of words. This is needed to avoid
2059 * unaligned accesses for some archs when redzoning is used, and makes
2060 * sure any on-slab bufctl's are also correctly aligned.
2062 if (size & (BYTES_PER_WORD - 1)) {
2063 size += (BYTES_PER_WORD - 1);
2064 size &= ~(BYTES_PER_WORD - 1);
2067 if (flags & SLAB_RED_ZONE) {
2068 ralign = REDZONE_ALIGN;
2069 /* If redzoning, ensure that the second redzone is suitably
2070 * aligned, by adjusting the object size accordingly. */
2071 size += REDZONE_ALIGN - 1;
2072 size &= ~(REDZONE_ALIGN - 1);
2075 /* 3) caller mandated alignment */
2076 if (ralign < cachep->align) {
2077 ralign = cachep->align;
2079 /* disable debug if necessary */
2080 if (ralign > __alignof__(unsigned long long))
2081 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2085 cachep->align = ralign;
2086 cachep->colour_off = cache_line_size();
2087 /* Offset must be a multiple of the alignment. */
2088 if (cachep->colour_off < cachep->align)
2089 cachep->colour_off = cachep->align;
2091 if (slab_is_available())
2099 * Both debugging options require word-alignment which is calculated
2102 if (flags & SLAB_RED_ZONE) {
2103 /* add space for red zone words */
2104 cachep->obj_offset += sizeof(unsigned long long);
2105 size += 2 * sizeof(unsigned long long);
2107 if (flags & SLAB_STORE_USER) {
2108 /* user store requires one word storage behind the end of
2109 * the real object. But if the second red zone needs to be
2110 * aligned to 64 bits, we must allow that much space.
2112 if (flags & SLAB_RED_ZONE)
2113 size += REDZONE_ALIGN;
2115 size += BYTES_PER_WORD;
2119 kasan_cache_create(cachep, &size, &flags);
2121 size = ALIGN(size, cachep->align);
2123 * We should restrict the number of objects in a slab to implement
2124 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2126 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2127 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2131 * To activate debug pagealloc, off-slab management is necessary
2132 * requirement. In early phase of initialization, small sized slab
2133 * doesn't get initialized so it would not be possible. So, we need
2134 * to check size >= 256. It guarantees that all necessary small
2135 * sized slab is initialized in current slab initialization sequence.
2137 if (debug_pagealloc_enabled() && (flags & SLAB_POISON) &&
2138 size >= 256 && cachep->object_size > cache_line_size()) {
2139 if (size < PAGE_SIZE || size % PAGE_SIZE == 0) {
2140 size_t tmp_size = ALIGN(size, PAGE_SIZE);
2142 if (set_off_slab_cache(cachep, tmp_size, flags)) {
2143 flags |= CFLGS_OFF_SLAB;
2144 cachep->obj_offset += tmp_size - size;
2152 if (set_objfreelist_slab_cache(cachep, size, flags)) {
2153 flags |= CFLGS_OBJFREELIST_SLAB;
2157 if (set_off_slab_cache(cachep, size, flags)) {
2158 flags |= CFLGS_OFF_SLAB;
2162 if (set_on_slab_cache(cachep, size, flags))
2168 cachep->freelist_size = cachep->num * sizeof(freelist_idx_t);
2169 cachep->flags = flags;
2170 cachep->allocflags = __GFP_COMP;
2171 if (flags & SLAB_CACHE_DMA)
2172 cachep->allocflags |= GFP_DMA;
2173 cachep->size = size;
2174 cachep->reciprocal_buffer_size = reciprocal_value(size);
2178 * If we're going to use the generic kernel_map_pages()
2179 * poisoning, then it's going to smash the contents of
2180 * the redzone and userword anyhow, so switch them off.
2182 if (IS_ENABLED(CONFIG_PAGE_POISONING) &&
2183 (cachep->flags & SLAB_POISON) &&
2184 is_debug_pagealloc_cache(cachep))
2185 cachep->flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2188 if (OFF_SLAB(cachep)) {
2189 cachep->freelist_cache =
2190 kmalloc_slab(cachep->freelist_size, 0u);
2193 err = setup_cpu_cache(cachep, gfp);
2195 __kmem_cache_release(cachep);
2203 static void check_irq_off(void)
2205 BUG_ON(!irqs_disabled());
2208 static void check_irq_on(void)
2210 BUG_ON(irqs_disabled());
2213 static void check_mutex_acquired(void)
2215 BUG_ON(!mutex_is_locked(&slab_mutex));
2218 static void check_spinlock_acquired(struct kmem_cache *cachep)
2222 assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
2226 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2230 assert_spin_locked(&get_node(cachep, node)->list_lock);
2235 #define check_irq_off() do { } while(0)
2236 #define check_irq_on() do { } while(0)
2237 #define check_mutex_acquired() do { } while(0)
2238 #define check_spinlock_acquired(x) do { } while(0)
2239 #define check_spinlock_acquired_node(x, y) do { } while(0)
2242 static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac,
2243 int node, bool free_all, struct list_head *list)
2247 if (!ac || !ac->avail)
2250 tofree = free_all ? ac->avail : (ac->limit + 4) / 5;
2251 if (tofree > ac->avail)
2252 tofree = (ac->avail + 1) / 2;
2254 free_block(cachep, ac->entry, tofree, node, list);
2255 ac->avail -= tofree;
2256 memmove(ac->entry, &(ac->entry[tofree]), sizeof(void *) * ac->avail);
2259 static void do_drain(void *arg)
2261 struct kmem_cache *cachep = arg;
2262 struct array_cache *ac;
2263 int node = numa_mem_id();
2264 struct kmem_cache_node *n;
2268 ac = cpu_cache_get(cachep);
2269 n = get_node(cachep, node);
2270 spin_lock(&n->list_lock);
2271 free_block(cachep, ac->entry, ac->avail, node, &list);
2272 spin_unlock(&n->list_lock);
2273 slabs_destroy(cachep, &list);
2277 static void drain_cpu_caches(struct kmem_cache *cachep)
2279 struct kmem_cache_node *n;
2283 on_each_cpu(do_drain, cachep, 1);
2285 for_each_kmem_cache_node(cachep, node, n)
2287 drain_alien_cache(cachep, n->alien);
2289 for_each_kmem_cache_node(cachep, node, n) {
2290 spin_lock_irq(&n->list_lock);
2291 drain_array_locked(cachep, n->shared, node, true, &list);
2292 spin_unlock_irq(&n->list_lock);
2294 slabs_destroy(cachep, &list);
2299 * Remove slabs from the list of free slabs.
2300 * Specify the number of slabs to drain in tofree.
2302 * Returns the actual number of slabs released.
2304 static int drain_freelist(struct kmem_cache *cache,
2305 struct kmem_cache_node *n, int tofree)
2307 struct list_head *p;
2312 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2314 spin_lock_irq(&n->list_lock);
2315 p = n->slabs_free.prev;
2316 if (p == &n->slabs_free) {
2317 spin_unlock_irq(&n->list_lock);
2321 page = list_entry(p, struct page, lru);
2322 list_del(&page->lru);
2325 * Safe to drop the lock. The slab is no longer linked
2328 n->free_objects -= cache->num;
2329 spin_unlock_irq(&n->list_lock);
2330 slab_destroy(cache, page);
2337 int __kmem_cache_shrink(struct kmem_cache *cachep)
2341 struct kmem_cache_node *n;
2343 drain_cpu_caches(cachep);
2346 for_each_kmem_cache_node(cachep, node, n) {
2347 drain_freelist(cachep, n, INT_MAX);
2349 ret += !list_empty(&n->slabs_full) ||
2350 !list_empty(&n->slabs_partial);
2352 return (ret ? 1 : 0);
2355 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2357 return __kmem_cache_shrink(cachep);
2360 void __kmem_cache_release(struct kmem_cache *cachep)
2363 struct kmem_cache_node *n;
2365 cache_random_seq_destroy(cachep);
2367 free_percpu(cachep->cpu_cache);
2369 /* NUMA: free the node structures */
2370 for_each_kmem_cache_node(cachep, i, n) {
2372 free_alien_cache(n->alien);
2374 cachep->node[i] = NULL;
2379 * Get the memory for a slab management obj.
2381 * For a slab cache when the slab descriptor is off-slab, the
2382 * slab descriptor can't come from the same cache which is being created,
2383 * Because if it is the case, that means we defer the creation of
2384 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2385 * And we eventually call down to __kmem_cache_create(), which
2386 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2387 * This is a "chicken-and-egg" problem.
2389 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2390 * which are all initialized during kmem_cache_init().
2392 static void *alloc_slabmgmt(struct kmem_cache *cachep,
2393 struct page *page, int colour_off,
2394 gfp_t local_flags, int nodeid)
2397 void *addr = page_address(page);
2399 page->s_mem = addr + colour_off;
2402 if (OBJFREELIST_SLAB(cachep))
2404 else if (OFF_SLAB(cachep)) {
2405 /* Slab management obj is off-slab. */
2406 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
2407 local_flags, nodeid);
2411 /* We will use last bytes at the slab for freelist */
2412 freelist = addr + (PAGE_SIZE << cachep->gfporder) -
2413 cachep->freelist_size;
2419 static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
2421 return ((freelist_idx_t *)page->freelist)[idx];
2424 static inline void set_free_obj(struct page *page,
2425 unsigned int idx, freelist_idx_t val)
2427 ((freelist_idx_t *)(page->freelist))[idx] = val;
2430 static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page)
2435 for (i = 0; i < cachep->num; i++) {
2436 void *objp = index_to_obj(cachep, page, i);
2438 if (cachep->flags & SLAB_STORE_USER)
2439 *dbg_userword(cachep, objp) = NULL;
2441 if (cachep->flags & SLAB_RED_ZONE) {
2442 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2443 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2446 * Constructors are not allowed to allocate memory from the same
2447 * cache which they are a constructor for. Otherwise, deadlock.
2448 * They must also be threaded.
2450 if (cachep->ctor && !(cachep->flags & SLAB_POISON)) {
2451 kasan_unpoison_object_data(cachep,
2452 objp + obj_offset(cachep));
2453 cachep->ctor(objp + obj_offset(cachep));
2454 kasan_poison_object_data(
2455 cachep, objp + obj_offset(cachep));
2458 if (cachep->flags & SLAB_RED_ZONE) {
2459 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2460 slab_error(cachep, "constructor overwrote the end of an object");
2461 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2462 slab_error(cachep, "constructor overwrote the start of an object");
2464 /* need to poison the objs? */
2465 if (cachep->flags & SLAB_POISON) {
2466 poison_obj(cachep, objp, POISON_FREE);
2467 slab_kernel_map(cachep, objp, 0, 0);
2473 #ifdef CONFIG_SLAB_FREELIST_RANDOM
2474 /* Hold information during a freelist initialization */
2475 union freelist_init_state {
2481 struct rnd_state rnd_state;
2485 * Initialize the state based on the randomization methode available.
2486 * return true if the pre-computed list is available, false otherwize.
2488 static bool freelist_state_initialize(union freelist_init_state *state,
2489 struct kmem_cache *cachep,
2495 /* Use best entropy available to define a random shift */
2496 rand = get_random_int();
2498 /* Use a random state if the pre-computed list is not available */
2499 if (!cachep->random_seq) {
2500 prandom_seed_state(&state->rnd_state, rand);
2503 state->list = cachep->random_seq;
2504 state->count = count;
2505 state->pos = rand % count;
2511 /* Get the next entry on the list and randomize it using a random shift */
2512 static freelist_idx_t next_random_slot(union freelist_init_state *state)
2514 if (state->pos >= state->count)
2516 return state->list[state->pos++];
2519 /* Swap two freelist entries */
2520 static void swap_free_obj(struct page *page, unsigned int a, unsigned int b)
2522 swap(((freelist_idx_t *)page->freelist)[a],
2523 ((freelist_idx_t *)page->freelist)[b]);
2527 * Shuffle the freelist initialization state based on pre-computed lists.
2528 * return true if the list was successfully shuffled, false otherwise.
2530 static bool shuffle_freelist(struct kmem_cache *cachep, struct page *page)
2532 unsigned int objfreelist = 0, i, rand, count = cachep->num;
2533 union freelist_init_state state;
2539 precomputed = freelist_state_initialize(&state, cachep, count);
2541 /* Take a random entry as the objfreelist */
2542 if (OBJFREELIST_SLAB(cachep)) {
2544 objfreelist = count - 1;
2546 objfreelist = next_random_slot(&state);
2547 page->freelist = index_to_obj(cachep, page, objfreelist) +
2553 * On early boot, generate the list dynamically.
2554 * Later use a pre-computed list for speed.
2557 for (i = 0; i < count; i++)
2558 set_free_obj(page, i, i);
2560 /* Fisher-Yates shuffle */
2561 for (i = count - 1; i > 0; i--) {
2562 rand = prandom_u32_state(&state.rnd_state);
2564 swap_free_obj(page, i, rand);
2567 for (i = 0; i < count; i++)
2568 set_free_obj(page, i, next_random_slot(&state));
2571 if (OBJFREELIST_SLAB(cachep))
2572 set_free_obj(page, cachep->num - 1, objfreelist);
2577 static inline bool shuffle_freelist(struct kmem_cache *cachep,
2582 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
2584 static void cache_init_objs(struct kmem_cache *cachep,
2591 cache_init_objs_debug(cachep, page);
2593 /* Try to randomize the freelist if enabled */
2594 shuffled = shuffle_freelist(cachep, page);
2596 if (!shuffled && OBJFREELIST_SLAB(cachep)) {
2597 page->freelist = index_to_obj(cachep, page, cachep->num - 1) +
2601 for (i = 0; i < cachep->num; i++) {
2602 objp = index_to_obj(cachep, page, i);
2603 kasan_init_slab_obj(cachep, objp);
2605 /* constructor could break poison info */
2606 if (DEBUG == 0 && cachep->ctor) {
2607 kasan_unpoison_object_data(cachep, objp);
2609 kasan_poison_object_data(cachep, objp);
2613 set_free_obj(page, i, i);
2617 static void *slab_get_obj(struct kmem_cache *cachep, struct page *page)
2621 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
2625 if (cachep->flags & SLAB_STORE_USER)
2626 set_store_user_dirty(cachep);
2632 static void slab_put_obj(struct kmem_cache *cachep,
2633 struct page *page, void *objp)
2635 unsigned int objnr = obj_to_index(cachep, page, objp);
2639 /* Verify double free bug */
2640 for (i = page->active; i < cachep->num; i++) {
2641 if (get_free_obj(page, i) == objnr) {
2642 pr_err("slab: double free detected in cache '%s', objp %p\n",
2643 cachep->name, objp);
2649 if (!page->freelist)
2650 page->freelist = objp + obj_offset(cachep);
2652 set_free_obj(page, page->active, objnr);
2656 * Map pages beginning at addr to the given cache and slab. This is required
2657 * for the slab allocator to be able to lookup the cache and slab of a
2658 * virtual address for kfree, ksize, and slab debugging.
2660 static void slab_map_pages(struct kmem_cache *cache, struct page *page,
2663 page->slab_cache = cache;
2664 page->freelist = freelist;
2668 * Grow (by 1) the number of slabs within a cache. This is called by
2669 * kmem_cache_alloc() when there are no active objs left in a cache.
2671 static struct page *cache_grow_begin(struct kmem_cache *cachep,
2672 gfp_t flags, int nodeid)
2678 struct kmem_cache_node *n;
2682 * Be lazy and only check for valid flags here, keeping it out of the
2683 * critical path in kmem_cache_alloc().
2685 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
2686 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK;
2687 flags &= ~GFP_SLAB_BUG_MASK;
2688 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
2689 invalid_mask, &invalid_mask, flags, &flags);
2692 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2695 if (gfpflags_allow_blocking(local_flags))
2699 * Get mem for the objs. Attempt to allocate a physical page from
2702 page = kmem_getpages(cachep, local_flags, nodeid);
2706 page_node = page_to_nid(page);
2707 n = get_node(cachep, page_node);
2709 /* Get colour for the slab, and cal the next value. */
2711 if (n->colour_next >= cachep->colour)
2714 offset = n->colour_next;
2715 if (offset >= cachep->colour)
2718 offset *= cachep->colour_off;
2720 /* Get slab management. */
2721 freelist = alloc_slabmgmt(cachep, page, offset,
2722 local_flags & ~GFP_CONSTRAINT_MASK, page_node);
2723 if (OFF_SLAB(cachep) && !freelist)
2726 slab_map_pages(cachep, page, freelist);
2728 kasan_poison_slab(page);
2729 cache_init_objs(cachep, page);
2731 if (gfpflags_allow_blocking(local_flags))
2732 local_irq_disable();
2737 kmem_freepages(cachep, page);
2739 if (gfpflags_allow_blocking(local_flags))
2740 local_irq_disable();
2744 static void cache_grow_end(struct kmem_cache *cachep, struct page *page)
2746 struct kmem_cache_node *n;
2754 INIT_LIST_HEAD(&page->lru);
2755 n = get_node(cachep, page_to_nid(page));
2757 spin_lock(&n->list_lock);
2759 list_add_tail(&page->lru, &(n->slabs_free));
2761 fixup_slab_list(cachep, n, page, &list);
2764 STATS_INC_GROWN(cachep);
2765 n->free_objects += cachep->num - page->active;
2766 spin_unlock(&n->list_lock);
2768 fixup_objfreelist_debug(cachep, &list);
2774 * Perform extra freeing checks:
2775 * - detect bad pointers.
2776 * - POISON/RED_ZONE checking
2778 static void kfree_debugcheck(const void *objp)
2780 if (!virt_addr_valid(objp)) {
2781 pr_err("kfree_debugcheck: out of range ptr %lxh\n",
2782 (unsigned long)objp);
2787 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2789 unsigned long long redzone1, redzone2;
2791 redzone1 = *dbg_redzone1(cache, obj);
2792 redzone2 = *dbg_redzone2(cache, obj);
2797 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2800 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2801 slab_error(cache, "double free detected");
2803 slab_error(cache, "memory outside object was overwritten");
2805 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2806 obj, redzone1, redzone2);
2809 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2810 unsigned long caller)
2815 BUG_ON(virt_to_cache(objp) != cachep);
2817 objp -= obj_offset(cachep);
2818 kfree_debugcheck(objp);
2819 page = virt_to_head_page(objp);
2821 if (cachep->flags & SLAB_RED_ZONE) {
2822 verify_redzone_free(cachep, objp);
2823 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2824 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2826 if (cachep->flags & SLAB_STORE_USER) {
2827 set_store_user_dirty(cachep);
2828 *dbg_userword(cachep, objp) = (void *)caller;
2831 objnr = obj_to_index(cachep, page, objp);
2833 BUG_ON(objnr >= cachep->num);
2834 BUG_ON(objp != index_to_obj(cachep, page, objnr));
2836 if (cachep->flags & SLAB_POISON) {
2837 poison_obj(cachep, objp, POISON_FREE);
2838 slab_kernel_map(cachep, objp, 0, caller);
2844 #define kfree_debugcheck(x) do { } while(0)
2845 #define cache_free_debugcheck(x,objp,z) (objp)
2848 static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
2856 objp = next - obj_offset(cachep);
2857 next = *(void **)next;
2858 poison_obj(cachep, objp, POISON_FREE);
2863 static inline void fixup_slab_list(struct kmem_cache *cachep,
2864 struct kmem_cache_node *n, struct page *page,
2867 /* move slabp to correct slabp list: */
2868 list_del(&page->lru);
2869 if (page->active == cachep->num) {
2870 list_add(&page->lru, &n->slabs_full);
2871 if (OBJFREELIST_SLAB(cachep)) {
2873 /* Poisoning will be done without holding the lock */
2874 if (cachep->flags & SLAB_POISON) {
2875 void **objp = page->freelist;
2881 page->freelist = NULL;
2884 list_add(&page->lru, &n->slabs_partial);
2887 /* Try to find non-pfmemalloc slab if needed */
2888 static noinline struct page *get_valid_first_slab(struct kmem_cache_node *n,
2889 struct page *page, bool pfmemalloc)
2897 if (!PageSlabPfmemalloc(page))
2900 /* No need to keep pfmemalloc slab if we have enough free objects */
2901 if (n->free_objects > n->free_limit) {
2902 ClearPageSlabPfmemalloc(page);
2906 /* Move pfmemalloc slab to the end of list to speed up next search */
2907 list_del(&page->lru);
2909 list_add_tail(&page->lru, &n->slabs_free);
2911 list_add_tail(&page->lru, &n->slabs_partial);
2913 list_for_each_entry(page, &n->slabs_partial, lru) {
2914 if (!PageSlabPfmemalloc(page))
2918 list_for_each_entry(page, &n->slabs_free, lru) {
2919 if (!PageSlabPfmemalloc(page))
2926 static struct page *get_first_slab(struct kmem_cache_node *n, bool pfmemalloc)
2930 page = list_first_entry_or_null(&n->slabs_partial,
2933 n->free_touched = 1;
2934 page = list_first_entry_or_null(&n->slabs_free,
2938 if (sk_memalloc_socks())
2939 return get_valid_first_slab(n, page, pfmemalloc);
2944 static noinline void *cache_alloc_pfmemalloc(struct kmem_cache *cachep,
2945 struct kmem_cache_node *n, gfp_t flags)
2951 if (!gfp_pfmemalloc_allowed(flags))
2954 spin_lock(&n->list_lock);
2955 page = get_first_slab(n, true);
2957 spin_unlock(&n->list_lock);
2961 obj = slab_get_obj(cachep, page);
2964 fixup_slab_list(cachep, n, page, &list);
2966 spin_unlock(&n->list_lock);
2967 fixup_objfreelist_debug(cachep, &list);
2973 * Slab list should be fixed up by fixup_slab_list() for existing slab
2974 * or cache_grow_end() for new slab
2976 static __always_inline int alloc_block(struct kmem_cache *cachep,
2977 struct array_cache *ac, struct page *page, int batchcount)
2980 * There must be at least one object available for
2983 BUG_ON(page->active >= cachep->num);
2985 while (page->active < cachep->num && batchcount--) {
2986 STATS_INC_ALLOCED(cachep);
2987 STATS_INC_ACTIVE(cachep);
2988 STATS_SET_HIGH(cachep);
2990 ac->entry[ac->avail++] = slab_get_obj(cachep, page);
2996 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2999 struct kmem_cache_node *n;
3000 struct array_cache *ac, *shared;
3006 node = numa_mem_id();
3008 ac = cpu_cache_get(cachep);
3009 batchcount = ac->batchcount;
3010 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
3012 * If there was little recent activity on this cache, then
3013 * perform only a partial refill. Otherwise we could generate
3016 batchcount = BATCHREFILL_LIMIT;
3018 n = get_node(cachep, node);
3020 BUG_ON(ac->avail > 0 || !n);
3021 shared = READ_ONCE(n->shared);
3022 if (!n->free_objects && (!shared || !shared->avail))
3025 spin_lock(&n->list_lock);
3026 shared = READ_ONCE(n->shared);
3028 /* See if we can refill from the shared array */
3029 if (shared && transfer_objects(ac, shared, batchcount)) {
3030 shared->touched = 1;
3034 while (batchcount > 0) {
3035 /* Get slab alloc is to come from. */
3036 page = get_first_slab(n, false);
3040 check_spinlock_acquired(cachep);
3042 batchcount = alloc_block(cachep, ac, page, batchcount);
3043 fixup_slab_list(cachep, n, page, &list);
3047 n->free_objects -= ac->avail;
3049 spin_unlock(&n->list_lock);
3050 fixup_objfreelist_debug(cachep, &list);
3053 if (unlikely(!ac->avail)) {
3054 /* Check if we can use obj in pfmemalloc slab */
3055 if (sk_memalloc_socks()) {
3056 void *obj = cache_alloc_pfmemalloc(cachep, n, flags);
3062 page = cache_grow_begin(cachep, gfp_exact_node(flags), node);
3065 * cache_grow_begin() can reenable interrupts,
3066 * then ac could change.
3068 ac = cpu_cache_get(cachep);
3069 if (!ac->avail && page)
3070 alloc_block(cachep, ac, page, batchcount);
3071 cache_grow_end(cachep, page);
3078 return ac->entry[--ac->avail];
3081 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3084 might_sleep_if(gfpflags_allow_blocking(flags));
3088 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3089 gfp_t flags, void *objp, unsigned long caller)
3093 if (cachep->flags & SLAB_POISON) {
3094 check_poison_obj(cachep, objp);
3095 slab_kernel_map(cachep, objp, 1, 0);
3096 poison_obj(cachep, objp, POISON_INUSE);
3098 if (cachep->flags & SLAB_STORE_USER)
3099 *dbg_userword(cachep, objp) = (void *)caller;
3101 if (cachep->flags & SLAB_RED_ZONE) {
3102 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3103 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3104 slab_error(cachep, "double free, or memory outside object was overwritten");
3105 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3106 objp, *dbg_redzone1(cachep, objp),
3107 *dbg_redzone2(cachep, objp));
3109 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3110 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3113 objp += obj_offset(cachep);
3114 if (cachep->ctor && cachep->flags & SLAB_POISON)
3116 if (ARCH_SLAB_MINALIGN &&
3117 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3118 pr_err("0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3119 objp, (int)ARCH_SLAB_MINALIGN);
3124 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3127 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3130 struct array_cache *ac;
3134 ac = cpu_cache_get(cachep);
3135 if (likely(ac->avail)) {
3137 objp = ac->entry[--ac->avail];
3139 STATS_INC_ALLOCHIT(cachep);
3143 STATS_INC_ALLOCMISS(cachep);
3144 objp = cache_alloc_refill(cachep, flags);
3146 * the 'ac' may be updated by cache_alloc_refill(),
3147 * and kmemleak_erase() requires its correct value.
3149 ac = cpu_cache_get(cachep);
3153 * To avoid a false negative, if an object that is in one of the
3154 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3155 * treat the array pointers as a reference to the object.
3158 kmemleak_erase(&ac->entry[ac->avail]);
3164 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
3166 * If we are in_interrupt, then process context, including cpusets and
3167 * mempolicy, may not apply and should not be used for allocation policy.
3169 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3171 int nid_alloc, nid_here;
3173 if (in_interrupt() || (flags & __GFP_THISNODE))
3175 nid_alloc = nid_here = numa_mem_id();
3176 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3177 nid_alloc = cpuset_slab_spread_node();
3178 else if (current->mempolicy)
3179 nid_alloc = mempolicy_slab_node();
3180 if (nid_alloc != nid_here)
3181 return ____cache_alloc_node(cachep, flags, nid_alloc);
3186 * Fallback function if there was no memory available and no objects on a
3187 * certain node and fall back is permitted. First we scan all the
3188 * available node for available objects. If that fails then we
3189 * perform an allocation without specifying a node. This allows the page
3190 * allocator to do its reclaim / fallback magic. We then insert the
3191 * slab into the proper nodelist and then allocate from it.
3193 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3195 struct zonelist *zonelist;
3198 enum zone_type high_zoneidx = gfp_zone(flags);
3202 unsigned int cpuset_mems_cookie;
3204 if (flags & __GFP_THISNODE)
3208 cpuset_mems_cookie = read_mems_allowed_begin();
3209 zonelist = node_zonelist(mempolicy_slab_node(), flags);
3213 * Look through allowed nodes for objects available
3214 * from existing per node queues.
3216 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3217 nid = zone_to_nid(zone);
3219 if (cpuset_zone_allowed(zone, flags) &&
3220 get_node(cache, nid) &&
3221 get_node(cache, nid)->free_objects) {
3222 obj = ____cache_alloc_node(cache,
3223 gfp_exact_node(flags), nid);
3231 * This allocation will be performed within the constraints
3232 * of the current cpuset / memory policy requirements.
3233 * We may trigger various forms of reclaim on the allowed
3234 * set and go into memory reserves if necessary.
3236 page = cache_grow_begin(cache, flags, numa_mem_id());
3237 cache_grow_end(cache, page);
3239 nid = page_to_nid(page);
3240 obj = ____cache_alloc_node(cache,
3241 gfp_exact_node(flags), nid);
3244 * Another processor may allocate the objects in
3245 * the slab since we are not holding any locks.
3252 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
3258 * A interface to enable slab creation on nodeid
3260 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3264 struct kmem_cache_node *n;
3268 VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
3269 n = get_node(cachep, nodeid);
3273 spin_lock(&n->list_lock);
3274 page = get_first_slab(n, false);
3278 check_spinlock_acquired_node(cachep, nodeid);
3280 STATS_INC_NODEALLOCS(cachep);
3281 STATS_INC_ACTIVE(cachep);
3282 STATS_SET_HIGH(cachep);
3284 BUG_ON(page->active == cachep->num);
3286 obj = slab_get_obj(cachep, page);
3289 fixup_slab_list(cachep, n, page, &list);
3291 spin_unlock(&n->list_lock);
3292 fixup_objfreelist_debug(cachep, &list);
3296 spin_unlock(&n->list_lock);
3297 page = cache_grow_begin(cachep, gfp_exact_node(flags), nodeid);
3299 /* This slab isn't counted yet so don't update free_objects */
3300 obj = slab_get_obj(cachep, page);
3302 cache_grow_end(cachep, page);
3304 return obj ? obj : fallback_alloc(cachep, flags);
3307 static __always_inline void *
3308 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3309 unsigned long caller)
3311 unsigned long save_flags;
3313 int slab_node = numa_mem_id();
3315 flags &= gfp_allowed_mask;
3316 cachep = slab_pre_alloc_hook(cachep, flags);
3317 if (unlikely(!cachep))
3320 cache_alloc_debugcheck_before(cachep, flags);
3321 local_irq_save(save_flags);
3323 if (nodeid == NUMA_NO_NODE)
3326 if (unlikely(!get_node(cachep, nodeid))) {
3327 /* Node not bootstrapped yet */
3328 ptr = fallback_alloc(cachep, flags);
3332 if (nodeid == slab_node) {
3334 * Use the locally cached objects if possible.
3335 * However ____cache_alloc does not allow fallback
3336 * to other nodes. It may fail while we still have
3337 * objects on other nodes available.
3339 ptr = ____cache_alloc(cachep, flags);
3343 /* ___cache_alloc_node can fall back to other nodes */
3344 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3346 local_irq_restore(save_flags);
3347 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3349 if (unlikely(flags & __GFP_ZERO) && ptr)
3350 memset(ptr, 0, cachep->object_size);
3352 slab_post_alloc_hook(cachep, flags, 1, &ptr);
3356 static __always_inline void *
3357 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3361 if (current->mempolicy || cpuset_do_slab_mem_spread()) {
3362 objp = alternate_node_alloc(cache, flags);
3366 objp = ____cache_alloc(cache, flags);
3369 * We may just have run out of memory on the local node.
3370 * ____cache_alloc_node() knows how to locate memory on other nodes
3373 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3380 static __always_inline void *
3381 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3383 return ____cache_alloc(cachep, flags);
3386 #endif /* CONFIG_NUMA */
3388 static __always_inline void *
3389 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3391 unsigned long save_flags;
3394 flags &= gfp_allowed_mask;
3395 cachep = slab_pre_alloc_hook(cachep, flags);
3396 if (unlikely(!cachep))
3399 cache_alloc_debugcheck_before(cachep, flags);
3400 local_irq_save(save_flags);
3401 objp = __do_cache_alloc(cachep, flags);
3402 local_irq_restore(save_flags);
3403 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3406 if (unlikely(flags & __GFP_ZERO) && objp)
3407 memset(objp, 0, cachep->object_size);
3409 slab_post_alloc_hook(cachep, flags, 1, &objp);
3414 * Caller needs to acquire correct kmem_cache_node's list_lock
3415 * @list: List of detached free slabs should be freed by caller
3417 static void free_block(struct kmem_cache *cachep, void **objpp,
3418 int nr_objects, int node, struct list_head *list)
3421 struct kmem_cache_node *n = get_node(cachep, node);
3424 n->free_objects += nr_objects;
3426 for (i = 0; i < nr_objects; i++) {
3432 page = virt_to_head_page(objp);
3433 list_del(&page->lru);
3434 check_spinlock_acquired_node(cachep, node);
3435 slab_put_obj(cachep, page, objp);
3436 STATS_DEC_ACTIVE(cachep);
3438 /* fixup slab chains */
3439 if (page->active == 0)
3440 list_add(&page->lru, &n->slabs_free);
3442 /* Unconditionally move a slab to the end of the
3443 * partial list on free - maximum time for the
3444 * other objects to be freed, too.
3446 list_add_tail(&page->lru, &n->slabs_partial);
3450 while (n->free_objects > n->free_limit && !list_empty(&n->slabs_free)) {
3451 n->free_objects -= cachep->num;
3453 page = list_last_entry(&n->slabs_free, struct page, lru);
3454 list_move(&page->lru, list);
3459 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3462 struct kmem_cache_node *n;
3463 int node = numa_mem_id();
3466 batchcount = ac->batchcount;
3469 n = get_node(cachep, node);
3470 spin_lock(&n->list_lock);
3472 struct array_cache *shared_array = n->shared;
3473 int max = shared_array->limit - shared_array->avail;
3475 if (batchcount > max)
3477 memcpy(&(shared_array->entry[shared_array->avail]),
3478 ac->entry, sizeof(void *) * batchcount);
3479 shared_array->avail += batchcount;
3484 free_block(cachep, ac->entry, batchcount, node, &list);
3491 list_for_each_entry(page, &n->slabs_free, lru) {
3492 BUG_ON(page->active);
3496 STATS_SET_FREEABLE(cachep, i);
3499 spin_unlock(&n->list_lock);
3500 slabs_destroy(cachep, &list);
3501 ac->avail -= batchcount;
3502 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3506 * Release an obj back to its cache. If the obj has a constructed state, it must
3507 * be in this state _before_ it is released. Called with disabled ints.
3509 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3510 unsigned long caller)
3512 /* Put the object into the quarantine, don't touch it for now. */
3513 if (kasan_slab_free(cachep, objp))
3516 ___cache_free(cachep, objp, caller);
3519 void ___cache_free(struct kmem_cache *cachep, void *objp,
3520 unsigned long caller)
3522 struct array_cache *ac = cpu_cache_get(cachep);
3525 kmemleak_free_recursive(objp, cachep->flags);
3526 objp = cache_free_debugcheck(cachep, objp, caller);
3528 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3531 * Skip calling cache_free_alien() when the platform is not numa.
3532 * This will avoid cache misses that happen while accessing slabp (which
3533 * is per page memory reference) to get nodeid. Instead use a global
3534 * variable to skip the call, which is mostly likely to be present in
3537 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3540 if (ac->avail < ac->limit) {
3541 STATS_INC_FREEHIT(cachep);
3543 STATS_INC_FREEMISS(cachep);
3544 cache_flusharray(cachep, ac);
3547 if (sk_memalloc_socks()) {
3548 struct page *page = virt_to_head_page(objp);
3550 if (unlikely(PageSlabPfmemalloc(page))) {
3551 cache_free_pfmemalloc(cachep, page, objp);
3556 ac->entry[ac->avail++] = objp;
3560 * kmem_cache_alloc - Allocate an object
3561 * @cachep: The cache to allocate from.
3562 * @flags: See kmalloc().
3564 * Allocate an object from this cache. The flags are only relevant
3565 * if the cache has no available objects.
3567 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3569 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3571 kasan_slab_alloc(cachep, ret, flags);
3572 trace_kmem_cache_alloc(_RET_IP_, ret,
3573 cachep->object_size, cachep->size, flags);
3577 EXPORT_SYMBOL(kmem_cache_alloc);
3579 static __always_inline void
3580 cache_alloc_debugcheck_after_bulk(struct kmem_cache *s, gfp_t flags,
3581 size_t size, void **p, unsigned long caller)
3585 for (i = 0; i < size; i++)
3586 p[i] = cache_alloc_debugcheck_after(s, flags, p[i], caller);
3589 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3594 s = slab_pre_alloc_hook(s, flags);
3598 cache_alloc_debugcheck_before(s, flags);
3600 local_irq_disable();
3601 for (i = 0; i < size; i++) {
3602 void *objp = __do_cache_alloc(s, flags);
3604 if (unlikely(!objp))
3610 cache_alloc_debugcheck_after_bulk(s, flags, size, p, _RET_IP_);
3612 /* Clear memory outside IRQ disabled section */
3613 if (unlikely(flags & __GFP_ZERO))
3614 for (i = 0; i < size; i++)
3615 memset(p[i], 0, s->object_size);
3617 slab_post_alloc_hook(s, flags, size, p);
3618 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3622 cache_alloc_debugcheck_after_bulk(s, flags, i, p, _RET_IP_);
3623 slab_post_alloc_hook(s, flags, i, p);
3624 __kmem_cache_free_bulk(s, i, p);
3627 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3629 #ifdef CONFIG_TRACING
3631 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3635 ret = slab_alloc(cachep, flags, _RET_IP_);
3637 kasan_kmalloc(cachep, ret, size, flags);
3638 trace_kmalloc(_RET_IP_, ret,
3639 size, cachep->size, flags);
3642 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3647 * kmem_cache_alloc_node - Allocate an object on the specified node
3648 * @cachep: The cache to allocate from.
3649 * @flags: See kmalloc().
3650 * @nodeid: node number of the target node.
3652 * Identical to kmem_cache_alloc but it will allocate memory on the given
3653 * node, which can improve the performance for cpu bound structures.
3655 * Fallback to other node is possible if __GFP_THISNODE is not set.
3657 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3659 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3661 kasan_slab_alloc(cachep, ret, flags);
3662 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3663 cachep->object_size, cachep->size,
3668 EXPORT_SYMBOL(kmem_cache_alloc_node);
3670 #ifdef CONFIG_TRACING
3671 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3678 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3680 kasan_kmalloc(cachep, ret, size, flags);
3681 trace_kmalloc_node(_RET_IP_, ret,
3686 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3689 static __always_inline void *
3690 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3692 struct kmem_cache *cachep;
3695 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3697 cachep = kmalloc_slab(size, flags);
3698 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3700 ret = kmem_cache_alloc_node_trace(cachep, flags, node, size);
3701 kasan_kmalloc(cachep, ret, size, flags);
3706 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3708 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3710 EXPORT_SYMBOL(__kmalloc_node);
3712 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3713 int node, unsigned long caller)
3715 return __do_kmalloc_node(size, flags, node, caller);
3717 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3718 #endif /* CONFIG_NUMA */
3721 * __do_kmalloc - allocate memory
3722 * @size: how many bytes of memory are required.
3723 * @flags: the type of memory to allocate (see kmalloc).
3724 * @caller: function caller for debug tracking of the caller
3726 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3727 unsigned long caller)
3729 struct kmem_cache *cachep;
3732 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3734 cachep = kmalloc_slab(size, flags);
3735 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3737 ret = slab_alloc(cachep, flags, caller);
3739 kasan_kmalloc(cachep, ret, size, flags);
3740 trace_kmalloc(caller, ret,
3741 size, cachep->size, flags);
3746 void *__kmalloc(size_t size, gfp_t flags)
3748 return __do_kmalloc(size, flags, _RET_IP_);
3750 EXPORT_SYMBOL(__kmalloc);
3752 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3754 return __do_kmalloc(size, flags, caller);
3756 EXPORT_SYMBOL(__kmalloc_track_caller);
3759 * kmem_cache_free - Deallocate an object
3760 * @cachep: The cache the allocation was from.
3761 * @objp: The previously allocated object.
3763 * Free an object which was previously allocated from this
3766 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3768 unsigned long flags;
3769 cachep = cache_from_obj(cachep, objp);
3773 local_irq_save(flags);
3774 debug_check_no_locks_freed(objp, cachep->object_size);
3775 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3776 debug_check_no_obj_freed(objp, cachep->object_size);
3777 __cache_free(cachep, objp, _RET_IP_);
3778 local_irq_restore(flags);
3780 trace_kmem_cache_free(_RET_IP_, objp);
3782 EXPORT_SYMBOL(kmem_cache_free);
3784 void kmem_cache_free_bulk(struct kmem_cache *orig_s, size_t size, void **p)
3786 struct kmem_cache *s;
3789 local_irq_disable();
3790 for (i = 0; i < size; i++) {
3793 if (!orig_s) /* called via kfree_bulk */
3794 s = virt_to_cache(objp);
3796 s = cache_from_obj(orig_s, objp);
3798 debug_check_no_locks_freed(objp, s->object_size);
3799 if (!(s->flags & SLAB_DEBUG_OBJECTS))
3800 debug_check_no_obj_freed(objp, s->object_size);
3802 __cache_free(s, objp, _RET_IP_);
3806 /* FIXME: add tracing */
3808 EXPORT_SYMBOL(kmem_cache_free_bulk);
3811 * kfree - free previously allocated memory
3812 * @objp: pointer returned by kmalloc.
3814 * If @objp is NULL, no operation is performed.
3816 * Don't free memory not originally allocated by kmalloc()
3817 * or you will run into trouble.
3819 void kfree(const void *objp)
3821 struct kmem_cache *c;
3822 unsigned long flags;
3824 trace_kfree(_RET_IP_, objp);
3826 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3828 local_irq_save(flags);
3829 kfree_debugcheck(objp);
3830 c = virt_to_cache(objp);
3831 debug_check_no_locks_freed(objp, c->object_size);
3833 debug_check_no_obj_freed(objp, c->object_size);
3834 __cache_free(c, (void *)objp, _RET_IP_);
3835 local_irq_restore(flags);
3837 EXPORT_SYMBOL(kfree);
3840 * This initializes kmem_cache_node or resizes various caches for all nodes.
3842 static int setup_kmem_cache_nodes(struct kmem_cache *cachep, gfp_t gfp)
3846 struct kmem_cache_node *n;
3848 for_each_online_node(node) {
3849 ret = setup_kmem_cache_node(cachep, node, gfp, true);
3858 if (!cachep->list.next) {
3859 /* Cache is not active yet. Roll back what we did */
3862 n = get_node(cachep, node);
3865 free_alien_cache(n->alien);
3867 cachep->node[node] = NULL;
3875 /* Always called with the slab_mutex held */
3876 static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
3877 int batchcount, int shared, gfp_t gfp)
3879 struct array_cache __percpu *cpu_cache, *prev;
3882 cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3886 prev = cachep->cpu_cache;
3887 cachep->cpu_cache = cpu_cache;
3888 kick_all_cpus_sync();
3891 cachep->batchcount = batchcount;
3892 cachep->limit = limit;
3893 cachep->shared = shared;
3898 for_each_online_cpu(cpu) {
3901 struct kmem_cache_node *n;
3902 struct array_cache *ac = per_cpu_ptr(prev, cpu);
3904 node = cpu_to_mem(cpu);
3905 n = get_node(cachep, node);
3906 spin_lock_irq(&n->list_lock);
3907 free_block(cachep, ac->entry, ac->avail, node, &list);
3908 spin_unlock_irq(&n->list_lock);
3909 slabs_destroy(cachep, &list);
3914 return setup_kmem_cache_nodes(cachep, gfp);
3917 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3918 int batchcount, int shared, gfp_t gfp)
3921 struct kmem_cache *c;
3923 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3925 if (slab_state < FULL)
3928 if ((ret < 0) || !is_root_cache(cachep))
3931 lockdep_assert_held(&slab_mutex);
3932 for_each_memcg_cache(c, cachep) {
3933 /* return value determined by the root cache only */
3934 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3940 /* Called with slab_mutex held always */
3941 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3948 err = cache_random_seq_create(cachep, cachep->num, gfp);
3952 if (!is_root_cache(cachep)) {
3953 struct kmem_cache *root = memcg_root_cache(cachep);
3954 limit = root->limit;
3955 shared = root->shared;
3956 batchcount = root->batchcount;
3959 if (limit && shared && batchcount)
3962 * The head array serves three purposes:
3963 * - create a LIFO ordering, i.e. return objects that are cache-warm
3964 * - reduce the number of spinlock operations.
3965 * - reduce the number of linked list operations on the slab and
3966 * bufctl chains: array operations are cheaper.
3967 * The numbers are guessed, we should auto-tune as described by
3970 if (cachep->size > 131072)
3972 else if (cachep->size > PAGE_SIZE)
3974 else if (cachep->size > 1024)
3976 else if (cachep->size > 256)
3982 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3983 * allocation behaviour: Most allocs on one cpu, most free operations
3984 * on another cpu. For these cases, an efficient object passing between
3985 * cpus is necessary. This is provided by a shared array. The array
3986 * replaces Bonwick's magazine layer.
3987 * On uniprocessor, it's functionally equivalent (but less efficient)
3988 * to a larger limit. Thus disabled by default.
3991 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
3996 * With debugging enabled, large batchcount lead to excessively long
3997 * periods with disabled local interrupts. Limit the batchcount
4002 batchcount = (limit + 1) / 2;
4004 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
4007 pr_err("enable_cpucache failed for %s, error %d\n",
4008 cachep->name, -err);
4013 * Drain an array if it contains any elements taking the node lock only if
4014 * necessary. Note that the node listlock also protects the array_cache
4015 * if drain_array() is used on the shared array.
4017 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
4018 struct array_cache *ac, int node)
4022 /* ac from n->shared can be freed if we don't hold the slab_mutex. */
4023 check_mutex_acquired();
4025 if (!ac || !ac->avail)
4033 spin_lock_irq(&n->list_lock);
4034 drain_array_locked(cachep, ac, node, false, &list);
4035 spin_unlock_irq(&n->list_lock);
4037 slabs_destroy(cachep, &list);
4041 * cache_reap - Reclaim memory from caches.
4042 * @w: work descriptor
4044 * Called from workqueue/eventd every few seconds.
4046 * - clear the per-cpu caches for this CPU.
4047 * - return freeable pages to the main free memory pool.
4049 * If we cannot acquire the cache chain mutex then just give up - we'll try
4050 * again on the next iteration.
4052 static void cache_reap(struct work_struct *w)
4054 struct kmem_cache *searchp;
4055 struct kmem_cache_node *n;
4056 int node = numa_mem_id();
4057 struct delayed_work *work = to_delayed_work(w);
4059 if (!mutex_trylock(&slab_mutex))
4060 /* Give up. Setup the next iteration. */
4063 list_for_each_entry(searchp, &slab_caches, list) {
4067 * We only take the node lock if absolutely necessary and we
4068 * have established with reasonable certainty that
4069 * we can do some work if the lock was obtained.
4071 n = get_node(searchp, node);
4073 reap_alien(searchp, n);
4075 drain_array(searchp, n, cpu_cache_get(searchp), node);
4078 * These are racy checks but it does not matter
4079 * if we skip one check or scan twice.
4081 if (time_after(n->next_reap, jiffies))
4084 n->next_reap = jiffies + REAPTIMEOUT_NODE;
4086 drain_array(searchp, n, n->shared, node);
4088 if (n->free_touched)
4089 n->free_touched = 0;
4093 freed = drain_freelist(searchp, n, (n->free_limit +
4094 5 * searchp->num - 1) / (5 * searchp->num));
4095 STATS_ADD_REAPED(searchp, freed);
4101 mutex_unlock(&slab_mutex);
4104 /* Set up the next iteration */
4105 schedule_delayed_work_on(smp_processor_id(), work,
4106 round_jiffies_relative(REAPTIMEOUT_AC));
4109 #ifdef CONFIG_SLABINFO
4110 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
4113 unsigned long active_objs;
4114 unsigned long num_objs;
4115 unsigned long active_slabs = 0;
4116 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4117 unsigned long num_slabs_partial = 0, num_slabs_free = 0;
4118 unsigned long num_slabs_full = 0;
4122 struct kmem_cache_node *n;
4126 for_each_kmem_cache_node(cachep, node, n) {
4129 spin_lock_irq(&n->list_lock);
4131 num_slabs += n->num_slabs;
4133 list_for_each_entry(page, &n->slabs_partial, lru) {
4134 if (page->active == cachep->num && !error)
4135 error = "slabs_partial accounting error";
4136 if (!page->active && !error)
4137 error = "slabs_partial accounting error";
4138 active_objs += page->active;
4139 num_slabs_partial++;
4142 list_for_each_entry(page, &n->slabs_free, lru) {
4143 if (page->active && !error)
4144 error = "slabs_free accounting error";
4148 free_objects += n->free_objects;
4150 shared_avail += n->shared->avail;
4152 spin_unlock_irq(&n->list_lock);
4154 num_objs = num_slabs * cachep->num;
4155 active_slabs = num_slabs - num_slabs_free;
4156 num_slabs_full = num_slabs - (num_slabs_partial + num_slabs_free);
4157 active_objs += (num_slabs_full * cachep->num);
4159 if (num_objs - active_objs != free_objects && !error)
4160 error = "free_objects accounting error";
4162 name = cachep->name;
4164 pr_err("slab: cache %s error: %s\n", name, error);
4166 sinfo->active_objs = active_objs;
4167 sinfo->num_objs = num_objs;
4168 sinfo->active_slabs = active_slabs;
4169 sinfo->num_slabs = num_slabs;
4170 sinfo->shared_avail = shared_avail;
4171 sinfo->limit = cachep->limit;
4172 sinfo->batchcount = cachep->batchcount;
4173 sinfo->shared = cachep->shared;
4174 sinfo->objects_per_slab = cachep->num;
4175 sinfo->cache_order = cachep->gfporder;
4178 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4182 unsigned long high = cachep->high_mark;
4183 unsigned long allocs = cachep->num_allocations;
4184 unsigned long grown = cachep->grown;
4185 unsigned long reaped = cachep->reaped;
4186 unsigned long errors = cachep->errors;
4187 unsigned long max_freeable = cachep->max_freeable;
4188 unsigned long node_allocs = cachep->node_allocs;
4189 unsigned long node_frees = cachep->node_frees;
4190 unsigned long overflows = cachep->node_overflow;
4192 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
4193 allocs, high, grown,
4194 reaped, errors, max_freeable, node_allocs,
4195 node_frees, overflows);
4199 unsigned long allochit = atomic_read(&cachep->allochit);
4200 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4201 unsigned long freehit = atomic_read(&cachep->freehit);
4202 unsigned long freemiss = atomic_read(&cachep->freemiss);
4204 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4205 allochit, allocmiss, freehit, freemiss);
4210 #define MAX_SLABINFO_WRITE 128
4212 * slabinfo_write - Tuning for the slab allocator
4214 * @buffer: user buffer
4215 * @count: data length
4218 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4219 size_t count, loff_t *ppos)
4221 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4222 int limit, batchcount, shared, res;
4223 struct kmem_cache *cachep;
4225 if (count > MAX_SLABINFO_WRITE)
4227 if (copy_from_user(&kbuf, buffer, count))
4229 kbuf[MAX_SLABINFO_WRITE] = '\0';
4231 tmp = strchr(kbuf, ' ');
4236 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4239 /* Find the cache in the chain of caches. */
4240 mutex_lock(&slab_mutex);
4242 list_for_each_entry(cachep, &slab_caches, list) {
4243 if (!strcmp(cachep->name, kbuf)) {
4244 if (limit < 1 || batchcount < 1 ||
4245 batchcount > limit || shared < 0) {
4248 res = do_tune_cpucache(cachep, limit,
4255 mutex_unlock(&slab_mutex);
4261 #ifdef CONFIG_DEBUG_SLAB_LEAK
4263 static inline int add_caller(unsigned long *n, unsigned long v)
4273 unsigned long *q = p + 2 * i;
4287 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4293 static void handle_slab(unsigned long *n, struct kmem_cache *c,
4302 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
4305 for (j = page->active; j < c->num; j++) {
4306 if (get_free_obj(page, j) == i) {
4316 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4317 * mapping is established when actual object allocation and
4318 * we could mistakenly access the unmapped object in the cpu
4321 if (probe_kernel_read(&v, dbg_userword(c, p), sizeof(v)))
4324 if (!add_caller(n, v))
4329 static void show_symbol(struct seq_file *m, unsigned long address)
4331 #ifdef CONFIG_KALLSYMS
4332 unsigned long offset, size;
4333 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4335 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4336 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4338 seq_printf(m, " [%s]", modname);
4342 seq_printf(m, "%p", (void *)address);
4345 static int leaks_show(struct seq_file *m, void *p)
4347 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4349 struct kmem_cache_node *n;
4351 unsigned long *x = m->private;
4355 if (!(cachep->flags & SLAB_STORE_USER))
4357 if (!(cachep->flags & SLAB_RED_ZONE))
4361 * Set store_user_clean and start to grab stored user information
4362 * for all objects on this cache. If some alloc/free requests comes
4363 * during the processing, information would be wrong so restart
4367 set_store_user_clean(cachep);
4368 drain_cpu_caches(cachep);
4372 for_each_kmem_cache_node(cachep, node, n) {
4375 spin_lock_irq(&n->list_lock);
4377 list_for_each_entry(page, &n->slabs_full, lru)
4378 handle_slab(x, cachep, page);
4379 list_for_each_entry(page, &n->slabs_partial, lru)
4380 handle_slab(x, cachep, page);
4381 spin_unlock_irq(&n->list_lock);
4383 } while (!is_store_user_clean(cachep));
4385 name = cachep->name;
4387 /* Increase the buffer size */
4388 mutex_unlock(&slab_mutex);
4389 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4391 /* Too bad, we are really out */
4393 mutex_lock(&slab_mutex);
4396 *(unsigned long *)m->private = x[0] * 2;
4398 mutex_lock(&slab_mutex);
4399 /* Now make sure this entry will be retried */
4403 for (i = 0; i < x[1]; i++) {
4404 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4405 show_symbol(m, x[2*i+2]);
4412 static const struct seq_operations slabstats_op = {
4413 .start = slab_start,
4419 static int slabstats_open(struct inode *inode, struct file *file)
4423 n = __seq_open_private(file, &slabstats_op, PAGE_SIZE);
4427 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4432 static const struct file_operations proc_slabstats_operations = {
4433 .open = slabstats_open,
4435 .llseek = seq_lseek,
4436 .release = seq_release_private,
4440 static int __init slab_proc_init(void)
4442 #ifdef CONFIG_DEBUG_SLAB_LEAK
4443 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4447 module_init(slab_proc_init);
4450 #ifdef CONFIG_HARDENED_USERCOPY
4452 * Rejects objects that are incorrectly sized.
4454 * Returns NULL if check passes, otherwise const char * to name of cache
4455 * to indicate an error.
4457 const char *__check_heap_object(const void *ptr, unsigned long n,
4460 struct kmem_cache *cachep;
4462 unsigned long offset;
4464 /* Find and validate object. */
4465 cachep = page->slab_cache;
4466 objnr = obj_to_index(cachep, page, (void *)ptr);
4467 BUG_ON(objnr >= cachep->num);
4469 /* Find offset within object. */
4470 offset = ptr - index_to_obj(cachep, page, objnr) - obj_offset(cachep);
4472 /* Allow address range falling entirely within object size. */
4473 if (offset <= cachep->object_size && n <= cachep->object_size - offset)
4476 return cachep->name;
4478 #endif /* CONFIG_HARDENED_USERCOPY */
4481 * ksize - get the actual amount of memory allocated for a given object
4482 * @objp: Pointer to the object
4484 * kmalloc may internally round up allocations and return more memory
4485 * than requested. ksize() can be used to determine the actual amount of
4486 * memory allocated. The caller may use this additional memory, even though
4487 * a smaller amount of memory was initially specified with the kmalloc call.
4488 * The caller must guarantee that objp points to a valid object previously
4489 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4490 * must not be freed during the duration of the call.
4492 size_t ksize(const void *objp)
4497 if (unlikely(objp == ZERO_SIZE_PTR))
4500 size = virt_to_cache(objp)->object_size;
4501 /* We assume that ksize callers could use the whole allocated area,
4502 * so we need to unpoison this area.
4504 kasan_unpoison_shadow(objp, size);
4508 EXPORT_SYMBOL(ksize);