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;
238 #define MAKE_LIST(cachep, listp, slab, nodeid) \
240 INIT_LIST_HEAD(listp); \
241 list_splice(&get_node(cachep, nodeid)->slab, listp); \
244 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
246 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
247 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
248 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
251 #define CFLGS_OBJFREELIST_SLAB (0x40000000UL)
252 #define CFLGS_OFF_SLAB (0x80000000UL)
253 #define OBJFREELIST_SLAB(x) ((x)->flags & CFLGS_OBJFREELIST_SLAB)
254 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
256 #define BATCHREFILL_LIMIT 16
258 * Optimization question: fewer reaps means less probability for unnessary
259 * cpucache drain/refill cycles.
261 * OTOH the cpuarrays can contain lots of objects,
262 * which could lock up otherwise freeable slabs.
264 #define REAPTIMEOUT_AC (2*HZ)
265 #define REAPTIMEOUT_NODE (4*HZ)
268 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
269 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
270 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
271 #define STATS_INC_GROWN(x) ((x)->grown++)
272 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
273 #define STATS_SET_HIGH(x) \
275 if ((x)->num_active > (x)->high_mark) \
276 (x)->high_mark = (x)->num_active; \
278 #define STATS_INC_ERR(x) ((x)->errors++)
279 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
280 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
281 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
282 #define STATS_SET_FREEABLE(x, i) \
284 if ((x)->max_freeable < i) \
285 (x)->max_freeable = i; \
287 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
288 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
289 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
290 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
292 #define STATS_INC_ACTIVE(x) do { } while (0)
293 #define STATS_DEC_ACTIVE(x) do { } while (0)
294 #define STATS_INC_ALLOCED(x) do { } while (0)
295 #define STATS_INC_GROWN(x) do { } while (0)
296 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
297 #define STATS_SET_HIGH(x) do { } while (0)
298 #define STATS_INC_ERR(x) do { } while (0)
299 #define STATS_INC_NODEALLOCS(x) do { } while (0)
300 #define STATS_INC_NODEFREES(x) do { } while (0)
301 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
302 #define STATS_SET_FREEABLE(x, i) do { } while (0)
303 #define STATS_INC_ALLOCHIT(x) do { } while (0)
304 #define STATS_INC_ALLOCMISS(x) do { } while (0)
305 #define STATS_INC_FREEHIT(x) do { } while (0)
306 #define STATS_INC_FREEMISS(x) do { } while (0)
312 * memory layout of objects:
314 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
315 * the end of an object is aligned with the end of the real
316 * allocation. Catches writes behind the end of the allocation.
317 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
319 * cachep->obj_offset: The real object.
320 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
321 * cachep->size - 1* BYTES_PER_WORD: last caller address
322 * [BYTES_PER_WORD long]
324 static int obj_offset(struct kmem_cache *cachep)
326 return cachep->obj_offset;
329 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
331 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
332 return (unsigned long long*) (objp + obj_offset(cachep) -
333 sizeof(unsigned long long));
336 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
338 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
339 if (cachep->flags & SLAB_STORE_USER)
340 return (unsigned long long *)(objp + cachep->size -
341 sizeof(unsigned long long) -
343 return (unsigned long long *) (objp + cachep->size -
344 sizeof(unsigned long long));
347 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
349 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
350 return (void **)(objp + cachep->size - BYTES_PER_WORD);
355 #define obj_offset(x) 0
356 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
357 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
358 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
362 #ifdef CONFIG_DEBUG_SLAB_LEAK
364 static inline bool is_store_user_clean(struct kmem_cache *cachep)
366 return atomic_read(&cachep->store_user_clean) == 1;
369 static inline void set_store_user_clean(struct kmem_cache *cachep)
371 atomic_set(&cachep->store_user_clean, 1);
374 static inline void set_store_user_dirty(struct kmem_cache *cachep)
376 if (is_store_user_clean(cachep))
377 atomic_set(&cachep->store_user_clean, 0);
381 static inline void set_store_user_dirty(struct kmem_cache *cachep) {}
386 * Do not go above this order unless 0 objects fit into the slab or
387 * overridden on the command line.
389 #define SLAB_MAX_ORDER_HI 1
390 #define SLAB_MAX_ORDER_LO 0
391 static int slab_max_order = SLAB_MAX_ORDER_LO;
392 static bool slab_max_order_set __initdata;
394 static inline struct kmem_cache *virt_to_cache(const void *obj)
396 struct page *page = virt_to_head_page(obj);
397 return page->slab_cache;
400 static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
403 return page->s_mem + cache->size * idx;
407 * We want to avoid an expensive divide : (offset / cache->size)
408 * Using the fact that size is a constant for a particular cache,
409 * we can replace (offset / cache->size) by
410 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
412 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
413 const struct page *page, void *obj)
415 u32 offset = (obj - page->s_mem);
416 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
419 #define BOOT_CPUCACHE_ENTRIES 1
420 /* internal cache of cache description objs */
421 static struct kmem_cache kmem_cache_boot = {
423 .limit = BOOT_CPUCACHE_ENTRIES,
425 .size = sizeof(struct kmem_cache),
426 .name = "kmem_cache",
429 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
431 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
433 return this_cpu_ptr(cachep->cpu_cache);
437 * Calculate the number of objects and left-over bytes for a given buffer size.
439 static unsigned int cache_estimate(unsigned long gfporder, size_t buffer_size,
440 unsigned long flags, size_t *left_over)
443 size_t slab_size = PAGE_SIZE << gfporder;
446 * The slab management structure can be either off the slab or
447 * on it. For the latter case, the memory allocated for a
450 * - @buffer_size bytes for each object
451 * - One freelist_idx_t for each object
453 * We don't need to consider alignment of freelist because
454 * freelist will be at the end of slab page. The objects will be
455 * at the correct alignment.
457 * If the slab management structure is off the slab, then the
458 * alignment will already be calculated into the size. Because
459 * the slabs are all pages aligned, the objects will be at the
460 * correct alignment when allocated.
462 if (flags & (CFLGS_OBJFREELIST_SLAB | CFLGS_OFF_SLAB)) {
463 num = slab_size / buffer_size;
464 *left_over = slab_size % buffer_size;
466 num = slab_size / (buffer_size + sizeof(freelist_idx_t));
467 *left_over = slab_size %
468 (buffer_size + sizeof(freelist_idx_t));
475 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
477 static void __slab_error(const char *function, struct kmem_cache *cachep,
480 pr_err("slab error in %s(): cache `%s': %s\n",
481 function, cachep->name, msg);
483 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
488 * By default on NUMA we use alien caches to stage the freeing of
489 * objects allocated from other nodes. This causes massive memory
490 * inefficiencies when using fake NUMA setup to split memory into a
491 * large number of small nodes, so it can be disabled on the command
495 static int use_alien_caches __read_mostly = 1;
496 static int __init noaliencache_setup(char *s)
498 use_alien_caches = 0;
501 __setup("noaliencache", noaliencache_setup);
503 static int __init slab_max_order_setup(char *str)
505 get_option(&str, &slab_max_order);
506 slab_max_order = slab_max_order < 0 ? 0 :
507 min(slab_max_order, MAX_ORDER - 1);
508 slab_max_order_set = true;
512 __setup("slab_max_order=", slab_max_order_setup);
516 * Special reaping functions for NUMA systems called from cache_reap().
517 * These take care of doing round robin flushing of alien caches (containing
518 * objects freed on different nodes from which they were allocated) and the
519 * flushing of remote pcps by calling drain_node_pages.
521 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
523 static void init_reap_node(int cpu)
527 node = next_node(cpu_to_mem(cpu), node_online_map);
528 if (node == MAX_NUMNODES)
529 node = first_node(node_online_map);
531 per_cpu(slab_reap_node, cpu) = node;
534 static void next_reap_node(void)
536 int node = __this_cpu_read(slab_reap_node);
538 node = next_node(node, node_online_map);
539 if (unlikely(node >= MAX_NUMNODES))
540 node = first_node(node_online_map);
541 __this_cpu_write(slab_reap_node, node);
545 #define init_reap_node(cpu) do { } while (0)
546 #define next_reap_node(void) do { } while (0)
550 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
551 * via the workqueue/eventd.
552 * Add the CPU number into the expiration time to minimize the possibility of
553 * the CPUs getting into lockstep and contending for the global cache chain
556 static void start_cpu_timer(int cpu)
558 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
561 * When this gets called from do_initcalls via cpucache_init(),
562 * init_workqueues() has already run, so keventd will be setup
565 if (keventd_up() && reap_work->work.func == NULL) {
567 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
568 schedule_delayed_work_on(cpu, reap_work,
569 __round_jiffies_relative(HZ, cpu));
573 static void init_arraycache(struct array_cache *ac, int limit, int batch)
576 * The array_cache structures contain pointers to free object.
577 * However, when such objects are allocated or transferred to another
578 * cache the pointers are not cleared and they could be counted as
579 * valid references during a kmemleak scan. Therefore, kmemleak must
580 * not scan such objects.
582 kmemleak_no_scan(ac);
586 ac->batchcount = batch;
591 static struct array_cache *alloc_arraycache(int node, int entries,
592 int batchcount, gfp_t gfp)
594 size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache);
595 struct array_cache *ac = NULL;
597 ac = kmalloc_node(memsize, gfp, node);
598 init_arraycache(ac, entries, batchcount);
602 static noinline void cache_free_pfmemalloc(struct kmem_cache *cachep,
603 struct page *page, void *objp)
605 struct kmem_cache_node *n;
609 page_node = page_to_nid(page);
610 n = get_node(cachep, page_node);
612 spin_lock(&n->list_lock);
613 free_block(cachep, &objp, 1, page_node, &list);
614 spin_unlock(&n->list_lock);
616 slabs_destroy(cachep, &list);
620 * Transfer objects in one arraycache to another.
621 * Locking must be handled by the caller.
623 * Return the number of entries transferred.
625 static int transfer_objects(struct array_cache *to,
626 struct array_cache *from, unsigned int max)
628 /* Figure out how many entries to transfer */
629 int nr = min3(from->avail, max, to->limit - to->avail);
634 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
644 #define drain_alien_cache(cachep, alien) do { } while (0)
645 #define reap_alien(cachep, n) do { } while (0)
647 static inline struct alien_cache **alloc_alien_cache(int node,
648 int limit, gfp_t gfp)
653 static inline void free_alien_cache(struct alien_cache **ac_ptr)
657 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
662 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
668 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
669 gfp_t flags, int nodeid)
674 static inline gfp_t gfp_exact_node(gfp_t flags)
676 return flags & ~__GFP_NOFAIL;
679 #else /* CONFIG_NUMA */
681 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
682 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
684 static struct alien_cache *__alloc_alien_cache(int node, int entries,
685 int batch, gfp_t gfp)
687 size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
688 struct alien_cache *alc = NULL;
690 alc = kmalloc_node(memsize, gfp, node);
691 init_arraycache(&alc->ac, entries, batch);
692 spin_lock_init(&alc->lock);
696 static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
698 struct alien_cache **alc_ptr;
699 size_t memsize = sizeof(void *) * nr_node_ids;
704 alc_ptr = kzalloc_node(memsize, gfp, node);
709 if (i == node || !node_online(i))
711 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
713 for (i--; i >= 0; i--)
722 static void free_alien_cache(struct alien_cache **alc_ptr)
733 static void __drain_alien_cache(struct kmem_cache *cachep,
734 struct array_cache *ac, int node,
735 struct list_head *list)
737 struct kmem_cache_node *n = get_node(cachep, node);
740 spin_lock(&n->list_lock);
742 * Stuff objects into the remote nodes shared array first.
743 * That way we could avoid the overhead of putting the objects
744 * into the free lists and getting them back later.
747 transfer_objects(n->shared, ac, ac->limit);
749 free_block(cachep, ac->entry, ac->avail, node, list);
751 spin_unlock(&n->list_lock);
756 * Called from cache_reap() to regularly drain alien caches round robin.
758 static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
760 int node = __this_cpu_read(slab_reap_node);
763 struct alien_cache *alc = n->alien[node];
764 struct array_cache *ac;
768 if (ac->avail && spin_trylock_irq(&alc->lock)) {
771 __drain_alien_cache(cachep, ac, node, &list);
772 spin_unlock_irq(&alc->lock);
773 slabs_destroy(cachep, &list);
779 static void drain_alien_cache(struct kmem_cache *cachep,
780 struct alien_cache **alien)
783 struct alien_cache *alc;
784 struct array_cache *ac;
787 for_each_online_node(i) {
793 spin_lock_irqsave(&alc->lock, flags);
794 __drain_alien_cache(cachep, ac, i, &list);
795 spin_unlock_irqrestore(&alc->lock, flags);
796 slabs_destroy(cachep, &list);
801 static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
802 int node, int page_node)
804 struct kmem_cache_node *n;
805 struct alien_cache *alien = NULL;
806 struct array_cache *ac;
809 n = get_node(cachep, node);
810 STATS_INC_NODEFREES(cachep);
811 if (n->alien && n->alien[page_node]) {
812 alien = n->alien[page_node];
814 spin_lock(&alien->lock);
815 if (unlikely(ac->avail == ac->limit)) {
816 STATS_INC_ACOVERFLOW(cachep);
817 __drain_alien_cache(cachep, ac, page_node, &list);
819 ac->entry[ac->avail++] = objp;
820 spin_unlock(&alien->lock);
821 slabs_destroy(cachep, &list);
823 n = get_node(cachep, page_node);
824 spin_lock(&n->list_lock);
825 free_block(cachep, &objp, 1, page_node, &list);
826 spin_unlock(&n->list_lock);
827 slabs_destroy(cachep, &list);
832 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
834 int page_node = page_to_nid(virt_to_page(objp));
835 int node = numa_mem_id();
837 * Make sure we are not freeing a object from another node to the array
840 if (likely(node == page_node))
843 return __cache_free_alien(cachep, objp, node, page_node);
847 * Construct gfp mask to allocate from a specific node but do not reclaim or
848 * warn about failures.
850 static inline gfp_t gfp_exact_node(gfp_t flags)
852 return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
856 static int init_cache_node(struct kmem_cache *cachep, int node, gfp_t gfp)
858 struct kmem_cache_node *n;
861 * Set up the kmem_cache_node for cpu before we can
862 * begin anything. Make sure some other cpu on this
863 * node has not already allocated this
865 n = get_node(cachep, node);
867 spin_lock_irq(&n->list_lock);
868 n->free_limit = (1 + nr_cpus_node(node)) * cachep->batchcount +
870 spin_unlock_irq(&n->list_lock);
875 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
879 kmem_cache_node_init(n);
880 n->next_reap = jiffies + REAPTIMEOUT_NODE +
881 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
884 (1 + nr_cpus_node(node)) * cachep->batchcount + cachep->num;
887 * The kmem_cache_nodes don't come and go as CPUs
888 * come and go. slab_mutex is sufficient
891 cachep->node[node] = n;
897 * Allocates and initializes node for a node on each slab cache, used for
898 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
899 * will be allocated off-node since memory is not yet online for the new node.
900 * When hotplugging memory or a cpu, existing node are not replaced if
903 * Must hold slab_mutex.
905 static int init_cache_node_node(int node)
908 struct kmem_cache *cachep;
910 list_for_each_entry(cachep, &slab_caches, list) {
911 ret = init_cache_node(cachep, node, GFP_KERNEL);
919 static int setup_kmem_cache_node(struct kmem_cache *cachep,
920 int node, gfp_t gfp, bool force_change)
923 struct kmem_cache_node *n;
924 struct array_cache *old_shared = NULL;
925 struct array_cache *new_shared = NULL;
926 struct alien_cache **new_alien = NULL;
929 if (use_alien_caches) {
930 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
935 if (cachep->shared) {
936 new_shared = alloc_arraycache(node,
937 cachep->shared * cachep->batchcount, 0xbaadf00d, gfp);
942 ret = init_cache_node(cachep, node, gfp);
946 n = get_node(cachep, node);
947 spin_lock_irq(&n->list_lock);
948 if (n->shared && force_change) {
949 free_block(cachep, n->shared->entry,
950 n->shared->avail, node, &list);
951 n->shared->avail = 0;
954 if (!n->shared || force_change) {
955 old_shared = n->shared;
956 n->shared = new_shared;
961 n->alien = new_alien;
965 spin_unlock_irq(&n->list_lock);
966 slabs_destroy(cachep, &list);
969 * To protect lockless access to n->shared during irq disabled context.
970 * If n->shared isn't NULL in irq disabled context, accessing to it is
971 * guaranteed to be valid until irq is re-enabled, because it will be
972 * freed after synchronize_sched().
980 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 static int cpuup_callback(struct notifier_block *nfb,
1086 unsigned long action, void *hcpu)
1088 long cpu = (long)hcpu;
1092 case CPU_UP_PREPARE:
1093 case CPU_UP_PREPARE_FROZEN:
1094 mutex_lock(&slab_mutex);
1095 err = cpuup_prepare(cpu);
1096 mutex_unlock(&slab_mutex);
1099 case CPU_ONLINE_FROZEN:
1100 start_cpu_timer(cpu);
1102 #ifdef CONFIG_HOTPLUG_CPU
1103 case CPU_DOWN_PREPARE:
1104 case CPU_DOWN_PREPARE_FROZEN:
1106 * Shutdown cache reaper. Note that the slab_mutex is
1107 * held so that if cache_reap() is invoked it cannot do
1108 * anything expensive but will only modify reap_work
1109 * and reschedule the timer.
1111 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1112 /* Now the cache_reaper is guaranteed to be not running. */
1113 per_cpu(slab_reap_work, cpu).work.func = NULL;
1115 case CPU_DOWN_FAILED:
1116 case CPU_DOWN_FAILED_FROZEN:
1117 start_cpu_timer(cpu);
1120 case CPU_DEAD_FROZEN:
1122 * Even if all the cpus of a node are down, we don't free the
1123 * kmem_cache_node of any cache. This to avoid a race between
1124 * cpu_down, and a kmalloc allocation from another cpu for
1125 * memory from the node of the cpu going down. The node
1126 * structure is usually allocated from kmem_cache_create() and
1127 * gets destroyed at kmem_cache_destroy().
1131 case CPU_UP_CANCELED:
1132 case CPU_UP_CANCELED_FROZEN:
1133 mutex_lock(&slab_mutex);
1134 cpuup_canceled(cpu);
1135 mutex_unlock(&slab_mutex);
1138 return notifier_from_errno(err);
1141 static struct notifier_block cpucache_notifier = {
1142 &cpuup_callback, NULL, 0
1145 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1147 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1148 * Returns -EBUSY if all objects cannot be drained so that the node is not
1151 * Must hold slab_mutex.
1153 static int __meminit drain_cache_node_node(int node)
1155 struct kmem_cache *cachep;
1158 list_for_each_entry(cachep, &slab_caches, list) {
1159 struct kmem_cache_node *n;
1161 n = get_node(cachep, node);
1165 drain_freelist(cachep, n, INT_MAX);
1167 if (!list_empty(&n->slabs_full) ||
1168 !list_empty(&n->slabs_partial)) {
1176 static int __meminit slab_memory_callback(struct notifier_block *self,
1177 unsigned long action, void *arg)
1179 struct memory_notify *mnb = arg;
1183 nid = mnb->status_change_nid;
1188 case MEM_GOING_ONLINE:
1189 mutex_lock(&slab_mutex);
1190 ret = init_cache_node_node(nid);
1191 mutex_unlock(&slab_mutex);
1193 case MEM_GOING_OFFLINE:
1194 mutex_lock(&slab_mutex);
1195 ret = drain_cache_node_node(nid);
1196 mutex_unlock(&slab_mutex);
1200 case MEM_CANCEL_ONLINE:
1201 case MEM_CANCEL_OFFLINE:
1205 return notifier_from_errno(ret);
1207 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1210 * swap the static kmem_cache_node with kmalloced memory
1212 static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1215 struct kmem_cache_node *ptr;
1217 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1220 memcpy(ptr, list, sizeof(struct kmem_cache_node));
1222 * Do not assume that spinlocks can be initialized via memcpy:
1224 spin_lock_init(&ptr->list_lock);
1226 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1227 cachep->node[nodeid] = ptr;
1231 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1232 * size of kmem_cache_node.
1234 static void __init set_up_node(struct kmem_cache *cachep, int index)
1238 for_each_online_node(node) {
1239 cachep->node[node] = &init_kmem_cache_node[index + node];
1240 cachep->node[node]->next_reap = jiffies +
1242 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1246 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1247 static void freelist_randomize(struct rnd_state *state, freelist_idx_t *list,
1253 for (i = 0; i < count; i++)
1256 /* Fisher-Yates shuffle */
1257 for (i = count - 1; i > 0; i--) {
1258 rand = prandom_u32_state(state);
1260 swap(list[i], list[rand]);
1264 /* Create a random sequence per cache */
1265 static int cache_random_seq_create(struct kmem_cache *cachep, gfp_t gfp)
1267 unsigned int seed, count = cachep->num;
1268 struct rnd_state state;
1273 /* If it fails, we will just use the global lists */
1274 cachep->random_seq = kcalloc(count, sizeof(freelist_idx_t), gfp);
1275 if (!cachep->random_seq)
1278 /* Get best entropy at this stage */
1279 get_random_bytes_arch(&seed, sizeof(seed));
1280 prandom_seed_state(&state, seed);
1282 freelist_randomize(&state, cachep->random_seq, count);
1286 /* Destroy the per-cache random freelist sequence */
1287 static void cache_random_seq_destroy(struct kmem_cache *cachep)
1289 kfree(cachep->random_seq);
1290 cachep->random_seq = NULL;
1293 static inline int cache_random_seq_create(struct kmem_cache *cachep, gfp_t gfp)
1297 static inline void cache_random_seq_destroy(struct kmem_cache *cachep) { }
1298 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1302 * Initialisation. Called after the page allocator have been initialised and
1303 * before smp_init().
1305 void __init kmem_cache_init(void)
1309 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1310 sizeof(struct rcu_head));
1311 kmem_cache = &kmem_cache_boot;
1313 if (!IS_ENABLED(CONFIG_NUMA) || num_possible_nodes() == 1)
1314 use_alien_caches = 0;
1316 for (i = 0; i < NUM_INIT_LISTS; i++)
1317 kmem_cache_node_init(&init_kmem_cache_node[i]);
1320 * Fragmentation resistance on low memory - only use bigger
1321 * page orders on machines with more than 32MB of memory if
1322 * not overridden on the command line.
1324 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1325 slab_max_order = SLAB_MAX_ORDER_HI;
1327 /* Bootstrap is tricky, because several objects are allocated
1328 * from caches that do not exist yet:
1329 * 1) initialize the kmem_cache cache: it contains the struct
1330 * kmem_cache structures of all caches, except kmem_cache itself:
1331 * kmem_cache is statically allocated.
1332 * Initially an __init data area is used for the head array and the
1333 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1334 * array at the end of the bootstrap.
1335 * 2) Create the first kmalloc cache.
1336 * The struct kmem_cache for the new cache is allocated normally.
1337 * An __init data area is used for the head array.
1338 * 3) Create the remaining kmalloc caches, with minimally sized
1340 * 4) Replace the __init data head arrays for kmem_cache and the first
1341 * kmalloc cache with kmalloc allocated arrays.
1342 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1343 * the other cache's with kmalloc allocated memory.
1344 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1347 /* 1) create the kmem_cache */
1350 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1352 create_boot_cache(kmem_cache, "kmem_cache",
1353 offsetof(struct kmem_cache, node) +
1354 nr_node_ids * sizeof(struct kmem_cache_node *),
1355 SLAB_HWCACHE_ALIGN);
1356 list_add(&kmem_cache->list, &slab_caches);
1357 slab_state = PARTIAL;
1360 * Initialize the caches that provide memory for the kmem_cache_node
1361 * structures first. Without this, further allocations will bug.
1363 kmalloc_caches[INDEX_NODE] = create_kmalloc_cache("kmalloc-node",
1364 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
1365 slab_state = PARTIAL_NODE;
1366 setup_kmalloc_cache_index_table();
1368 slab_early_init = 0;
1370 /* 5) Replace the bootstrap kmem_cache_node */
1374 for_each_online_node(nid) {
1375 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
1377 init_list(kmalloc_caches[INDEX_NODE],
1378 &init_kmem_cache_node[SIZE_NODE + nid], nid);
1382 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1385 void __init kmem_cache_init_late(void)
1387 struct kmem_cache *cachep;
1391 /* 6) resize the head arrays to their final sizes */
1392 mutex_lock(&slab_mutex);
1393 list_for_each_entry(cachep, &slab_caches, list)
1394 if (enable_cpucache(cachep, GFP_NOWAIT))
1396 mutex_unlock(&slab_mutex);
1402 * Register a cpu startup notifier callback that initializes
1403 * cpu_cache_get for all new cpus
1405 register_cpu_notifier(&cpucache_notifier);
1409 * Register a memory hotplug callback that initializes and frees
1412 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1416 * The reap timers are started later, with a module init call: That part
1417 * of the kernel is not yet operational.
1421 static int __init cpucache_init(void)
1426 * Register the timers that return unneeded pages to the page allocator
1428 for_each_online_cpu(cpu)
1429 start_cpu_timer(cpu);
1435 __initcall(cpucache_init);
1437 static noinline void
1438 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1441 struct kmem_cache_node *n;
1443 unsigned long flags;
1445 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1446 DEFAULT_RATELIMIT_BURST);
1448 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1451 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1452 nodeid, gfpflags, &gfpflags);
1453 pr_warn(" cache: %s, object size: %d, order: %d\n",
1454 cachep->name, cachep->size, cachep->gfporder);
1456 for_each_kmem_cache_node(cachep, node, n) {
1457 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1458 unsigned long active_slabs = 0, num_slabs = 0;
1460 spin_lock_irqsave(&n->list_lock, flags);
1461 list_for_each_entry(page, &n->slabs_full, lru) {
1462 active_objs += cachep->num;
1465 list_for_each_entry(page, &n->slabs_partial, lru) {
1466 active_objs += page->active;
1469 list_for_each_entry(page, &n->slabs_free, lru)
1472 free_objects += n->free_objects;
1473 spin_unlock_irqrestore(&n->list_lock, flags);
1475 num_slabs += active_slabs;
1476 num_objs = num_slabs * cachep->num;
1477 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1478 node, active_slabs, num_slabs, active_objs, num_objs,
1485 * Interface to system's page allocator. No need to hold the
1486 * kmem_cache_node ->list_lock.
1488 * If we requested dmaable memory, we will get it. Even if we
1489 * did not request dmaable memory, we might get it, but that
1490 * would be relatively rare and ignorable.
1492 static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1498 flags |= cachep->allocflags;
1499 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1500 flags |= __GFP_RECLAIMABLE;
1502 page = __alloc_pages_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1504 slab_out_of_memory(cachep, flags, nodeid);
1508 if (memcg_charge_slab(page, flags, cachep->gfporder, cachep)) {
1509 __free_pages(page, cachep->gfporder);
1513 nr_pages = (1 << cachep->gfporder);
1514 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1515 add_zone_page_state(page_zone(page),
1516 NR_SLAB_RECLAIMABLE, nr_pages);
1518 add_zone_page_state(page_zone(page),
1519 NR_SLAB_UNRECLAIMABLE, nr_pages);
1521 __SetPageSlab(page);
1522 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1523 if (sk_memalloc_socks() && page_is_pfmemalloc(page))
1524 SetPageSlabPfmemalloc(page);
1526 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1527 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1530 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1532 kmemcheck_mark_unallocated_pages(page, nr_pages);
1539 * Interface to system's page release.
1541 static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
1543 int order = cachep->gfporder;
1544 unsigned long nr_freed = (1 << order);
1546 kmemcheck_free_shadow(page, order);
1548 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1549 sub_zone_page_state(page_zone(page),
1550 NR_SLAB_RECLAIMABLE, nr_freed);
1552 sub_zone_page_state(page_zone(page),
1553 NR_SLAB_UNRECLAIMABLE, nr_freed);
1555 BUG_ON(!PageSlab(page));
1556 __ClearPageSlabPfmemalloc(page);
1557 __ClearPageSlab(page);
1558 page_mapcount_reset(page);
1559 page->mapping = NULL;
1561 if (current->reclaim_state)
1562 current->reclaim_state->reclaimed_slab += nr_freed;
1563 memcg_uncharge_slab(page, order, cachep);
1564 __free_pages(page, order);
1567 static void kmem_rcu_free(struct rcu_head *head)
1569 struct kmem_cache *cachep;
1572 page = container_of(head, struct page, rcu_head);
1573 cachep = page->slab_cache;
1575 kmem_freepages(cachep, page);
1579 static bool is_debug_pagealloc_cache(struct kmem_cache *cachep)
1581 if (debug_pagealloc_enabled() && OFF_SLAB(cachep) &&
1582 (cachep->size % PAGE_SIZE) == 0)
1588 #ifdef CONFIG_DEBUG_PAGEALLOC
1589 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1590 unsigned long caller)
1592 int size = cachep->object_size;
1594 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1596 if (size < 5 * sizeof(unsigned long))
1599 *addr++ = 0x12345678;
1601 *addr++ = smp_processor_id();
1602 size -= 3 * sizeof(unsigned long);
1604 unsigned long *sptr = &caller;
1605 unsigned long svalue;
1607 while (!kstack_end(sptr)) {
1609 if (kernel_text_address(svalue)) {
1611 size -= sizeof(unsigned long);
1612 if (size <= sizeof(unsigned long))
1618 *addr++ = 0x87654321;
1621 static void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1622 int map, unsigned long caller)
1624 if (!is_debug_pagealloc_cache(cachep))
1628 store_stackinfo(cachep, objp, caller);
1630 kernel_map_pages(virt_to_page(objp), cachep->size / PAGE_SIZE, map);
1634 static inline void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1635 int map, unsigned long caller) {}
1639 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1641 int size = cachep->object_size;
1642 addr = &((char *)addr)[obj_offset(cachep)];
1644 memset(addr, val, size);
1645 *(unsigned char *)(addr + size - 1) = POISON_END;
1648 static void dump_line(char *data, int offset, int limit)
1651 unsigned char error = 0;
1654 pr_err("%03x: ", offset);
1655 for (i = 0; i < limit; i++) {
1656 if (data[offset + i] != POISON_FREE) {
1657 error = data[offset + i];
1661 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1662 &data[offset], limit, 1);
1664 if (bad_count == 1) {
1665 error ^= POISON_FREE;
1666 if (!(error & (error - 1))) {
1667 pr_err("Single bit error detected. Probably bad RAM.\n");
1669 pr_err("Run memtest86+ or a similar memory test tool.\n");
1671 pr_err("Run a memory test tool.\n");
1680 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1685 if (cachep->flags & SLAB_RED_ZONE) {
1686 pr_err("Redzone: 0x%llx/0x%llx\n",
1687 *dbg_redzone1(cachep, objp),
1688 *dbg_redzone2(cachep, objp));
1691 if (cachep->flags & SLAB_STORE_USER) {
1692 pr_err("Last user: [<%p>](%pSR)\n",
1693 *dbg_userword(cachep, objp),
1694 *dbg_userword(cachep, objp));
1696 realobj = (char *)objp + obj_offset(cachep);
1697 size = cachep->object_size;
1698 for (i = 0; i < size && lines; i += 16, lines--) {
1701 if (i + limit > size)
1703 dump_line(realobj, i, limit);
1707 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1713 if (is_debug_pagealloc_cache(cachep))
1716 realobj = (char *)objp + obj_offset(cachep);
1717 size = cachep->object_size;
1719 for (i = 0; i < size; i++) {
1720 char exp = POISON_FREE;
1723 if (realobj[i] != exp) {
1728 pr_err("Slab corruption (%s): %s start=%p, len=%d\n",
1729 print_tainted(), cachep->name,
1731 print_objinfo(cachep, objp, 0);
1733 /* Hexdump the affected line */
1736 if (i + limit > size)
1738 dump_line(realobj, i, limit);
1741 /* Limit to 5 lines */
1747 /* Print some data about the neighboring objects, if they
1750 struct page *page = virt_to_head_page(objp);
1753 objnr = obj_to_index(cachep, page, objp);
1755 objp = index_to_obj(cachep, page, objnr - 1);
1756 realobj = (char *)objp + obj_offset(cachep);
1757 pr_err("Prev obj: start=%p, len=%d\n", realobj, size);
1758 print_objinfo(cachep, objp, 2);
1760 if (objnr + 1 < cachep->num) {
1761 objp = index_to_obj(cachep, page, objnr + 1);
1762 realobj = (char *)objp + obj_offset(cachep);
1763 pr_err("Next obj: start=%p, len=%d\n", realobj, size);
1764 print_objinfo(cachep, objp, 2);
1771 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1776 if (OBJFREELIST_SLAB(cachep) && cachep->flags & SLAB_POISON) {
1777 poison_obj(cachep, page->freelist - obj_offset(cachep),
1781 for (i = 0; i < cachep->num; i++) {
1782 void *objp = index_to_obj(cachep, page, i);
1784 if (cachep->flags & SLAB_POISON) {
1785 check_poison_obj(cachep, objp);
1786 slab_kernel_map(cachep, objp, 1, 0);
1788 if (cachep->flags & SLAB_RED_ZONE) {
1789 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1790 slab_error(cachep, "start of a freed object was overwritten");
1791 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1792 slab_error(cachep, "end of a freed object was overwritten");
1797 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1804 * slab_destroy - destroy and release all objects in a slab
1805 * @cachep: cache pointer being destroyed
1806 * @page: page pointer being destroyed
1808 * Destroy all the objs in a slab page, and release the mem back to the system.
1809 * Before calling the slab page must have been unlinked from the cache. The
1810 * kmem_cache_node ->list_lock is not held/needed.
1812 static void slab_destroy(struct kmem_cache *cachep, struct page *page)
1816 freelist = page->freelist;
1817 slab_destroy_debugcheck(cachep, page);
1818 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
1819 call_rcu(&page->rcu_head, kmem_rcu_free);
1821 kmem_freepages(cachep, page);
1824 * From now on, we don't use freelist
1825 * although actual page can be freed in rcu context
1827 if (OFF_SLAB(cachep))
1828 kmem_cache_free(cachep->freelist_cache, freelist);
1831 static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1833 struct page *page, *n;
1835 list_for_each_entry_safe(page, n, list, lru) {
1836 list_del(&page->lru);
1837 slab_destroy(cachep, page);
1842 * calculate_slab_order - calculate size (page order) of slabs
1843 * @cachep: pointer to the cache that is being created
1844 * @size: size of objects to be created in this cache.
1845 * @flags: slab allocation flags
1847 * Also calculates the number of objects per slab.
1849 * This could be made much more intelligent. For now, try to avoid using
1850 * high order pages for slabs. When the gfp() functions are more friendly
1851 * towards high-order requests, this should be changed.
1853 static size_t calculate_slab_order(struct kmem_cache *cachep,
1854 size_t size, unsigned long flags)
1856 size_t left_over = 0;
1859 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
1863 num = cache_estimate(gfporder, size, flags, &remainder);
1867 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1868 if (num > SLAB_OBJ_MAX_NUM)
1871 if (flags & CFLGS_OFF_SLAB) {
1872 struct kmem_cache *freelist_cache;
1873 size_t freelist_size;
1875 freelist_size = num * sizeof(freelist_idx_t);
1876 freelist_cache = kmalloc_slab(freelist_size, 0u);
1877 if (!freelist_cache)
1881 * Needed to avoid possible looping condition
1882 * in cache_grow_begin()
1884 if (OFF_SLAB(freelist_cache))
1887 /* check if off slab has enough benefit */
1888 if (freelist_cache->size > cachep->size / 2)
1892 /* Found something acceptable - save it away */
1894 cachep->gfporder = gfporder;
1895 left_over = remainder;
1898 * A VFS-reclaimable slab tends to have most allocations
1899 * as GFP_NOFS and we really don't want to have to be allocating
1900 * higher-order pages when we are unable to shrink dcache.
1902 if (flags & SLAB_RECLAIM_ACCOUNT)
1906 * Large number of objects is good, but very large slabs are
1907 * currently bad for the gfp()s.
1909 if (gfporder >= slab_max_order)
1913 * Acceptable internal fragmentation?
1915 if (left_over * 8 <= (PAGE_SIZE << gfporder))
1921 static struct array_cache __percpu *alloc_kmem_cache_cpus(
1922 struct kmem_cache *cachep, int entries, int batchcount)
1926 struct array_cache __percpu *cpu_cache;
1928 size = sizeof(void *) * entries + sizeof(struct array_cache);
1929 cpu_cache = __alloc_percpu(size, sizeof(void *));
1934 for_each_possible_cpu(cpu) {
1935 init_arraycache(per_cpu_ptr(cpu_cache, cpu),
1936 entries, batchcount);
1942 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
1944 if (slab_state >= FULL)
1945 return enable_cpucache(cachep, gfp);
1947 cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
1948 if (!cachep->cpu_cache)
1951 if (slab_state == DOWN) {
1952 /* Creation of first cache (kmem_cache). */
1953 set_up_node(kmem_cache, CACHE_CACHE);
1954 } else if (slab_state == PARTIAL) {
1955 /* For kmem_cache_node */
1956 set_up_node(cachep, SIZE_NODE);
1960 for_each_online_node(node) {
1961 cachep->node[node] = kmalloc_node(
1962 sizeof(struct kmem_cache_node), gfp, node);
1963 BUG_ON(!cachep->node[node]);
1964 kmem_cache_node_init(cachep->node[node]);
1968 cachep->node[numa_mem_id()]->next_reap =
1969 jiffies + REAPTIMEOUT_NODE +
1970 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1972 cpu_cache_get(cachep)->avail = 0;
1973 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1974 cpu_cache_get(cachep)->batchcount = 1;
1975 cpu_cache_get(cachep)->touched = 0;
1976 cachep->batchcount = 1;
1977 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1981 unsigned long kmem_cache_flags(unsigned long object_size,
1982 unsigned long flags, const char *name,
1983 void (*ctor)(void *))
1989 __kmem_cache_alias(const char *name, size_t size, size_t align,
1990 unsigned long flags, void (*ctor)(void *))
1992 struct kmem_cache *cachep;
1994 cachep = find_mergeable(size, align, flags, name, ctor);
1999 * Adjust the object sizes so that we clear
2000 * the complete object on kzalloc.
2002 cachep->object_size = max_t(int, cachep->object_size, size);
2007 static bool set_objfreelist_slab_cache(struct kmem_cache *cachep,
2008 size_t size, unsigned long flags)
2014 if (cachep->ctor || flags & SLAB_DESTROY_BY_RCU)
2017 left = calculate_slab_order(cachep, size,
2018 flags | CFLGS_OBJFREELIST_SLAB);
2022 if (cachep->num * sizeof(freelist_idx_t) > cachep->object_size)
2025 cachep->colour = left / cachep->colour_off;
2030 static bool set_off_slab_cache(struct kmem_cache *cachep,
2031 size_t size, unsigned long flags)
2038 * Always use on-slab management when SLAB_NOLEAKTRACE
2039 * to avoid recursive calls into kmemleak.
2041 if (flags & SLAB_NOLEAKTRACE)
2045 * Size is large, assume best to place the slab management obj
2046 * off-slab (should allow better packing of objs).
2048 left = calculate_slab_order(cachep, size, flags | CFLGS_OFF_SLAB);
2053 * If the slab has been placed off-slab, and we have enough space then
2054 * move it on-slab. This is at the expense of any extra colouring.
2056 if (left >= cachep->num * sizeof(freelist_idx_t))
2059 cachep->colour = left / cachep->colour_off;
2064 static bool set_on_slab_cache(struct kmem_cache *cachep,
2065 size_t size, unsigned long flags)
2071 left = calculate_slab_order(cachep, size, flags);
2075 cachep->colour = left / cachep->colour_off;
2081 * __kmem_cache_create - Create a cache.
2082 * @cachep: cache management descriptor
2083 * @flags: SLAB flags
2085 * Returns a ptr to the cache on success, NULL on failure.
2086 * Cannot be called within a int, but can be interrupted.
2087 * The @ctor is run when new pages are allocated by the cache.
2091 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2092 * to catch references to uninitialised memory.
2094 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2095 * for buffer overruns.
2097 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2098 * cacheline. This can be beneficial if you're counting cycles as closely
2102 __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2104 size_t ralign = BYTES_PER_WORD;
2107 size_t size = cachep->size;
2112 * Enable redzoning and last user accounting, except for caches with
2113 * large objects, if the increased size would increase the object size
2114 * above the next power of two: caches with object sizes just above a
2115 * power of two have a significant amount of internal fragmentation.
2117 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2118 2 * sizeof(unsigned long long)))
2119 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2120 if (!(flags & SLAB_DESTROY_BY_RCU))
2121 flags |= SLAB_POISON;
2126 * Check that size is in terms of words. This is needed to avoid
2127 * unaligned accesses for some archs when redzoning is used, and makes
2128 * sure any on-slab bufctl's are also correctly aligned.
2130 if (size & (BYTES_PER_WORD - 1)) {
2131 size += (BYTES_PER_WORD - 1);
2132 size &= ~(BYTES_PER_WORD - 1);
2135 if (flags & SLAB_RED_ZONE) {
2136 ralign = REDZONE_ALIGN;
2137 /* If redzoning, ensure that the second redzone is suitably
2138 * aligned, by adjusting the object size accordingly. */
2139 size += REDZONE_ALIGN - 1;
2140 size &= ~(REDZONE_ALIGN - 1);
2143 /* 3) caller mandated alignment */
2144 if (ralign < cachep->align) {
2145 ralign = cachep->align;
2147 /* disable debug if necessary */
2148 if (ralign > __alignof__(unsigned long long))
2149 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2153 cachep->align = ralign;
2154 cachep->colour_off = cache_line_size();
2155 /* Offset must be a multiple of the alignment. */
2156 if (cachep->colour_off < cachep->align)
2157 cachep->colour_off = cachep->align;
2159 if (slab_is_available())
2167 * Both debugging options require word-alignment which is calculated
2170 if (flags & SLAB_RED_ZONE) {
2171 /* add space for red zone words */
2172 cachep->obj_offset += sizeof(unsigned long long);
2173 size += 2 * sizeof(unsigned long long);
2175 if (flags & SLAB_STORE_USER) {
2176 /* user store requires one word storage behind the end of
2177 * the real object. But if the second red zone needs to be
2178 * aligned to 64 bits, we must allow that much space.
2180 if (flags & SLAB_RED_ZONE)
2181 size += REDZONE_ALIGN;
2183 size += BYTES_PER_WORD;
2187 kasan_cache_create(cachep, &size, &flags);
2189 size = ALIGN(size, cachep->align);
2191 * We should restrict the number of objects in a slab to implement
2192 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2194 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2195 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2199 * To activate debug pagealloc, off-slab management is necessary
2200 * requirement. In early phase of initialization, small sized slab
2201 * doesn't get initialized so it would not be possible. So, we need
2202 * to check size >= 256. It guarantees that all necessary small
2203 * sized slab is initialized in current slab initialization sequence.
2205 if (debug_pagealloc_enabled() && (flags & SLAB_POISON) &&
2206 size >= 256 && cachep->object_size > cache_line_size()) {
2207 if (size < PAGE_SIZE || size % PAGE_SIZE == 0) {
2208 size_t tmp_size = ALIGN(size, PAGE_SIZE);
2210 if (set_off_slab_cache(cachep, tmp_size, flags)) {
2211 flags |= CFLGS_OFF_SLAB;
2212 cachep->obj_offset += tmp_size - size;
2220 if (set_objfreelist_slab_cache(cachep, size, flags)) {
2221 flags |= CFLGS_OBJFREELIST_SLAB;
2225 if (set_off_slab_cache(cachep, size, flags)) {
2226 flags |= CFLGS_OFF_SLAB;
2230 if (set_on_slab_cache(cachep, size, flags))
2236 cachep->freelist_size = cachep->num * sizeof(freelist_idx_t);
2237 cachep->flags = flags;
2238 cachep->allocflags = __GFP_COMP;
2239 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2240 cachep->allocflags |= GFP_DMA;
2241 cachep->size = size;
2242 cachep->reciprocal_buffer_size = reciprocal_value(size);
2246 * If we're going to use the generic kernel_map_pages()
2247 * poisoning, then it's going to smash the contents of
2248 * the redzone and userword anyhow, so switch them off.
2250 if (IS_ENABLED(CONFIG_PAGE_POISONING) &&
2251 (cachep->flags & SLAB_POISON) &&
2252 is_debug_pagealloc_cache(cachep))
2253 cachep->flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2256 if (OFF_SLAB(cachep)) {
2257 cachep->freelist_cache =
2258 kmalloc_slab(cachep->freelist_size, 0u);
2261 err = setup_cpu_cache(cachep, gfp);
2263 __kmem_cache_release(cachep);
2271 static void check_irq_off(void)
2273 BUG_ON(!irqs_disabled());
2276 static void check_irq_on(void)
2278 BUG_ON(irqs_disabled());
2281 static void check_mutex_acquired(void)
2283 BUG_ON(!mutex_is_locked(&slab_mutex));
2286 static void check_spinlock_acquired(struct kmem_cache *cachep)
2290 assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
2294 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2298 assert_spin_locked(&get_node(cachep, node)->list_lock);
2303 #define check_irq_off() do { } while(0)
2304 #define check_irq_on() do { } while(0)
2305 #define check_mutex_acquired() do { } while(0)
2306 #define check_spinlock_acquired(x) do { } while(0)
2307 #define check_spinlock_acquired_node(x, y) do { } while(0)
2310 static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac,
2311 int node, bool free_all, struct list_head *list)
2315 if (!ac || !ac->avail)
2318 tofree = free_all ? ac->avail : (ac->limit + 4) / 5;
2319 if (tofree > ac->avail)
2320 tofree = (ac->avail + 1) / 2;
2322 free_block(cachep, ac->entry, tofree, node, list);
2323 ac->avail -= tofree;
2324 memmove(ac->entry, &(ac->entry[tofree]), sizeof(void *) * ac->avail);
2327 static void do_drain(void *arg)
2329 struct kmem_cache *cachep = arg;
2330 struct array_cache *ac;
2331 int node = numa_mem_id();
2332 struct kmem_cache_node *n;
2336 ac = cpu_cache_get(cachep);
2337 n = get_node(cachep, node);
2338 spin_lock(&n->list_lock);
2339 free_block(cachep, ac->entry, ac->avail, node, &list);
2340 spin_unlock(&n->list_lock);
2341 slabs_destroy(cachep, &list);
2345 static void drain_cpu_caches(struct kmem_cache *cachep)
2347 struct kmem_cache_node *n;
2351 on_each_cpu(do_drain, cachep, 1);
2353 for_each_kmem_cache_node(cachep, node, n)
2355 drain_alien_cache(cachep, n->alien);
2357 for_each_kmem_cache_node(cachep, node, n) {
2358 spin_lock_irq(&n->list_lock);
2359 drain_array_locked(cachep, n->shared, node, true, &list);
2360 spin_unlock_irq(&n->list_lock);
2362 slabs_destroy(cachep, &list);
2367 * Remove slabs from the list of free slabs.
2368 * Specify the number of slabs to drain in tofree.
2370 * Returns the actual number of slabs released.
2372 static int drain_freelist(struct kmem_cache *cache,
2373 struct kmem_cache_node *n, int tofree)
2375 struct list_head *p;
2380 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2382 spin_lock_irq(&n->list_lock);
2383 p = n->slabs_free.prev;
2384 if (p == &n->slabs_free) {
2385 spin_unlock_irq(&n->list_lock);
2389 page = list_entry(p, struct page, lru);
2390 list_del(&page->lru);
2392 * Safe to drop the lock. The slab is no longer linked
2395 n->free_objects -= cache->num;
2396 spin_unlock_irq(&n->list_lock);
2397 slab_destroy(cache, page);
2404 int __kmem_cache_shrink(struct kmem_cache *cachep, bool deactivate)
2408 struct kmem_cache_node *n;
2410 drain_cpu_caches(cachep);
2413 for_each_kmem_cache_node(cachep, node, n) {
2414 drain_freelist(cachep, n, INT_MAX);
2416 ret += !list_empty(&n->slabs_full) ||
2417 !list_empty(&n->slabs_partial);
2419 return (ret ? 1 : 0);
2422 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2424 return __kmem_cache_shrink(cachep, false);
2427 void __kmem_cache_release(struct kmem_cache *cachep)
2430 struct kmem_cache_node *n;
2432 cache_random_seq_destroy(cachep);
2434 free_percpu(cachep->cpu_cache);
2436 /* NUMA: free the node structures */
2437 for_each_kmem_cache_node(cachep, i, n) {
2439 free_alien_cache(n->alien);
2441 cachep->node[i] = NULL;
2446 * Get the memory for a slab management obj.
2448 * For a slab cache when the slab descriptor is off-slab, the
2449 * slab descriptor can't come from the same cache which is being created,
2450 * Because if it is the case, that means we defer the creation of
2451 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2452 * And we eventually call down to __kmem_cache_create(), which
2453 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2454 * This is a "chicken-and-egg" problem.
2456 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2457 * which are all initialized during kmem_cache_init().
2459 static void *alloc_slabmgmt(struct kmem_cache *cachep,
2460 struct page *page, int colour_off,
2461 gfp_t local_flags, int nodeid)
2464 void *addr = page_address(page);
2466 page->s_mem = addr + colour_off;
2469 if (OBJFREELIST_SLAB(cachep))
2471 else if (OFF_SLAB(cachep)) {
2472 /* Slab management obj is off-slab. */
2473 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
2474 local_flags, nodeid);
2478 /* We will use last bytes at the slab for freelist */
2479 freelist = addr + (PAGE_SIZE << cachep->gfporder) -
2480 cachep->freelist_size;
2486 static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
2488 return ((freelist_idx_t *)page->freelist)[idx];
2491 static inline void set_free_obj(struct page *page,
2492 unsigned int idx, freelist_idx_t val)
2494 ((freelist_idx_t *)(page->freelist))[idx] = val;
2497 static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page)
2502 for (i = 0; i < cachep->num; i++) {
2503 void *objp = index_to_obj(cachep, page, i);
2505 if (cachep->flags & SLAB_STORE_USER)
2506 *dbg_userword(cachep, objp) = NULL;
2508 if (cachep->flags & SLAB_RED_ZONE) {
2509 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2510 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2513 * Constructors are not allowed to allocate memory from the same
2514 * cache which they are a constructor for. Otherwise, deadlock.
2515 * They must also be threaded.
2517 if (cachep->ctor && !(cachep->flags & SLAB_POISON)) {
2518 kasan_unpoison_object_data(cachep,
2519 objp + obj_offset(cachep));
2520 cachep->ctor(objp + obj_offset(cachep));
2521 kasan_poison_object_data(
2522 cachep, objp + obj_offset(cachep));
2525 if (cachep->flags & SLAB_RED_ZONE) {
2526 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2527 slab_error(cachep, "constructor overwrote the end of an object");
2528 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2529 slab_error(cachep, "constructor overwrote the start of an object");
2531 /* need to poison the objs? */
2532 if (cachep->flags & SLAB_POISON) {
2533 poison_obj(cachep, objp, POISON_FREE);
2534 slab_kernel_map(cachep, objp, 0, 0);
2540 #ifdef CONFIG_SLAB_FREELIST_RANDOM
2541 /* Hold information during a freelist initialization */
2542 union freelist_init_state {
2545 freelist_idx_t *list;
2549 struct rnd_state rnd_state;
2553 * Initialize the state based on the randomization methode available.
2554 * return true if the pre-computed list is available, false otherwize.
2556 static bool freelist_state_initialize(union freelist_init_state *state,
2557 struct kmem_cache *cachep,
2563 /* Use best entropy available to define a random shift */
2564 get_random_bytes_arch(&rand, sizeof(rand));
2566 /* Use a random state if the pre-computed list is not available */
2567 if (!cachep->random_seq) {
2568 prandom_seed_state(&state->rnd_state, rand);
2571 state->list = cachep->random_seq;
2572 state->count = count;
2580 /* Get the next entry on the list and randomize it using a random shift */
2581 static freelist_idx_t next_random_slot(union freelist_init_state *state)
2583 return (state->list[state->pos++] + state->rand) % state->count;
2587 * Shuffle the freelist initialization state based on pre-computed lists.
2588 * return true if the list was successfully shuffled, false otherwise.
2590 static bool shuffle_freelist(struct kmem_cache *cachep, struct page *page)
2592 unsigned int objfreelist = 0, i, count = cachep->num;
2593 union freelist_init_state state;
2599 precomputed = freelist_state_initialize(&state, cachep, count);
2601 /* Take a random entry as the objfreelist */
2602 if (OBJFREELIST_SLAB(cachep)) {
2604 objfreelist = count - 1;
2606 objfreelist = next_random_slot(&state);
2607 page->freelist = index_to_obj(cachep, page, objfreelist) +
2613 * On early boot, generate the list dynamically.
2614 * Later use a pre-computed list for speed.
2617 freelist_randomize(&state.rnd_state, page->freelist, count);
2619 for (i = 0; i < count; i++)
2620 set_free_obj(page, i, next_random_slot(&state));
2623 if (OBJFREELIST_SLAB(cachep))
2624 set_free_obj(page, cachep->num - 1, objfreelist);
2629 static inline bool shuffle_freelist(struct kmem_cache *cachep,
2634 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
2636 static void cache_init_objs(struct kmem_cache *cachep,
2643 cache_init_objs_debug(cachep, page);
2645 /* Try to randomize the freelist if enabled */
2646 shuffled = shuffle_freelist(cachep, page);
2648 if (!shuffled && OBJFREELIST_SLAB(cachep)) {
2649 page->freelist = index_to_obj(cachep, page, cachep->num - 1) +
2653 for (i = 0; i < cachep->num; i++) {
2654 /* constructor could break poison info */
2655 if (DEBUG == 0 && cachep->ctor) {
2656 objp = index_to_obj(cachep, page, i);
2657 kasan_unpoison_object_data(cachep, objp);
2659 kasan_poison_object_data(cachep, objp);
2663 set_free_obj(page, i, i);
2667 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2669 if (CONFIG_ZONE_DMA_FLAG) {
2670 if (flags & GFP_DMA)
2671 BUG_ON(!(cachep->allocflags & GFP_DMA));
2673 BUG_ON(cachep->allocflags & GFP_DMA);
2677 static void *slab_get_obj(struct kmem_cache *cachep, struct page *page)
2681 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
2685 if (cachep->flags & SLAB_STORE_USER)
2686 set_store_user_dirty(cachep);
2692 static void slab_put_obj(struct kmem_cache *cachep,
2693 struct page *page, void *objp)
2695 unsigned int objnr = obj_to_index(cachep, page, objp);
2699 /* Verify double free bug */
2700 for (i = page->active; i < cachep->num; i++) {
2701 if (get_free_obj(page, i) == objnr) {
2702 pr_err("slab: double free detected in cache '%s', objp %p\n",
2703 cachep->name, objp);
2709 if (!page->freelist)
2710 page->freelist = objp + obj_offset(cachep);
2712 set_free_obj(page, page->active, objnr);
2716 * Map pages beginning at addr to the given cache and slab. This is required
2717 * for the slab allocator to be able to lookup the cache and slab of a
2718 * virtual address for kfree, ksize, and slab debugging.
2720 static void slab_map_pages(struct kmem_cache *cache, struct page *page,
2723 page->slab_cache = cache;
2724 page->freelist = freelist;
2728 * Grow (by 1) the number of slabs within a cache. This is called by
2729 * kmem_cache_alloc() when there are no active objs left in a cache.
2731 static struct page *cache_grow_begin(struct kmem_cache *cachep,
2732 gfp_t flags, int nodeid)
2738 struct kmem_cache_node *n;
2742 * Be lazy and only check for valid flags here, keeping it out of the
2743 * critical path in kmem_cache_alloc().
2745 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
2746 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
2749 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2752 if (gfpflags_allow_blocking(local_flags))
2756 * The test for missing atomic flag is performed here, rather than
2757 * the more obvious place, simply to reduce the critical path length
2758 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2759 * will eventually be caught here (where it matters).
2761 kmem_flagcheck(cachep, flags);
2764 * Get mem for the objs. Attempt to allocate a physical page from
2767 page = kmem_getpages(cachep, local_flags, nodeid);
2771 page_node = page_to_nid(page);
2772 n = get_node(cachep, page_node);
2774 /* Get colour for the slab, and cal the next value. */
2776 if (n->colour_next >= cachep->colour)
2779 offset = n->colour_next;
2780 if (offset >= cachep->colour)
2783 offset *= cachep->colour_off;
2785 /* Get slab management. */
2786 freelist = alloc_slabmgmt(cachep, page, offset,
2787 local_flags & ~GFP_CONSTRAINT_MASK, page_node);
2788 if (OFF_SLAB(cachep) && !freelist)
2791 slab_map_pages(cachep, page, freelist);
2793 kasan_poison_slab(page);
2794 cache_init_objs(cachep, page);
2796 if (gfpflags_allow_blocking(local_flags))
2797 local_irq_disable();
2802 kmem_freepages(cachep, page);
2804 if (gfpflags_allow_blocking(local_flags))
2805 local_irq_disable();
2809 static void cache_grow_end(struct kmem_cache *cachep, struct page *page)
2811 struct kmem_cache_node *n;
2819 INIT_LIST_HEAD(&page->lru);
2820 n = get_node(cachep, page_to_nid(page));
2822 spin_lock(&n->list_lock);
2824 list_add_tail(&page->lru, &(n->slabs_free));
2826 fixup_slab_list(cachep, n, page, &list);
2827 STATS_INC_GROWN(cachep);
2828 n->free_objects += cachep->num - page->active;
2829 spin_unlock(&n->list_lock);
2831 fixup_objfreelist_debug(cachep, &list);
2837 * Perform extra freeing checks:
2838 * - detect bad pointers.
2839 * - POISON/RED_ZONE checking
2841 static void kfree_debugcheck(const void *objp)
2843 if (!virt_addr_valid(objp)) {
2844 pr_err("kfree_debugcheck: out of range ptr %lxh\n",
2845 (unsigned long)objp);
2850 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2852 unsigned long long redzone1, redzone2;
2854 redzone1 = *dbg_redzone1(cache, obj);
2855 redzone2 = *dbg_redzone2(cache, obj);
2860 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2863 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2864 slab_error(cache, "double free detected");
2866 slab_error(cache, "memory outside object was overwritten");
2868 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2869 obj, redzone1, redzone2);
2872 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2873 unsigned long caller)
2878 BUG_ON(virt_to_cache(objp) != cachep);
2880 objp -= obj_offset(cachep);
2881 kfree_debugcheck(objp);
2882 page = virt_to_head_page(objp);
2884 if (cachep->flags & SLAB_RED_ZONE) {
2885 verify_redzone_free(cachep, objp);
2886 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2887 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2889 if (cachep->flags & SLAB_STORE_USER) {
2890 set_store_user_dirty(cachep);
2891 *dbg_userword(cachep, objp) = (void *)caller;
2894 objnr = obj_to_index(cachep, page, objp);
2896 BUG_ON(objnr >= cachep->num);
2897 BUG_ON(objp != index_to_obj(cachep, page, objnr));
2899 if (cachep->flags & SLAB_POISON) {
2900 poison_obj(cachep, objp, POISON_FREE);
2901 slab_kernel_map(cachep, objp, 0, caller);
2907 #define kfree_debugcheck(x) do { } while(0)
2908 #define cache_free_debugcheck(x,objp,z) (objp)
2911 static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
2919 objp = next - obj_offset(cachep);
2920 next = *(void **)next;
2921 poison_obj(cachep, objp, POISON_FREE);
2926 static inline void fixup_slab_list(struct kmem_cache *cachep,
2927 struct kmem_cache_node *n, struct page *page,
2930 /* move slabp to correct slabp list: */
2931 list_del(&page->lru);
2932 if (page->active == cachep->num) {
2933 list_add(&page->lru, &n->slabs_full);
2934 if (OBJFREELIST_SLAB(cachep)) {
2936 /* Poisoning will be done without holding the lock */
2937 if (cachep->flags & SLAB_POISON) {
2938 void **objp = page->freelist;
2944 page->freelist = NULL;
2947 list_add(&page->lru, &n->slabs_partial);
2950 /* Try to find non-pfmemalloc slab if needed */
2951 static noinline struct page *get_valid_first_slab(struct kmem_cache_node *n,
2952 struct page *page, bool pfmemalloc)
2960 if (!PageSlabPfmemalloc(page))
2963 /* No need to keep pfmemalloc slab if we have enough free objects */
2964 if (n->free_objects > n->free_limit) {
2965 ClearPageSlabPfmemalloc(page);
2969 /* Move pfmemalloc slab to the end of list to speed up next search */
2970 list_del(&page->lru);
2972 list_add_tail(&page->lru, &n->slabs_free);
2974 list_add_tail(&page->lru, &n->slabs_partial);
2976 list_for_each_entry(page, &n->slabs_partial, lru) {
2977 if (!PageSlabPfmemalloc(page))
2981 list_for_each_entry(page, &n->slabs_free, lru) {
2982 if (!PageSlabPfmemalloc(page))
2989 static struct page *get_first_slab(struct kmem_cache_node *n, bool pfmemalloc)
2993 page = list_first_entry_or_null(&n->slabs_partial,
2996 n->free_touched = 1;
2997 page = list_first_entry_or_null(&n->slabs_free,
3001 if (sk_memalloc_socks())
3002 return get_valid_first_slab(n, page, pfmemalloc);
3007 static noinline void *cache_alloc_pfmemalloc(struct kmem_cache *cachep,
3008 struct kmem_cache_node *n, gfp_t flags)
3014 if (!gfp_pfmemalloc_allowed(flags))
3017 spin_lock(&n->list_lock);
3018 page = get_first_slab(n, true);
3020 spin_unlock(&n->list_lock);
3024 obj = slab_get_obj(cachep, page);
3027 fixup_slab_list(cachep, n, page, &list);
3029 spin_unlock(&n->list_lock);
3030 fixup_objfreelist_debug(cachep, &list);
3036 * Slab list should be fixed up by fixup_slab_list() for existing slab
3037 * or cache_grow_end() for new slab
3039 static __always_inline int alloc_block(struct kmem_cache *cachep,
3040 struct array_cache *ac, struct page *page, int batchcount)
3043 * There must be at least one object available for
3046 BUG_ON(page->active >= cachep->num);
3048 while (page->active < cachep->num && batchcount--) {
3049 STATS_INC_ALLOCED(cachep);
3050 STATS_INC_ACTIVE(cachep);
3051 STATS_SET_HIGH(cachep);
3053 ac->entry[ac->avail++] = slab_get_obj(cachep, page);
3059 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
3062 struct kmem_cache_node *n;
3063 struct array_cache *ac, *shared;
3069 node = numa_mem_id();
3071 ac = cpu_cache_get(cachep);
3072 batchcount = ac->batchcount;
3073 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
3075 * If there was little recent activity on this cache, then
3076 * perform only a partial refill. Otherwise we could generate
3079 batchcount = BATCHREFILL_LIMIT;
3081 n = get_node(cachep, node);
3083 BUG_ON(ac->avail > 0 || !n);
3084 shared = READ_ONCE(n->shared);
3085 if (!n->free_objects && (!shared || !shared->avail))
3088 spin_lock(&n->list_lock);
3089 shared = READ_ONCE(n->shared);
3091 /* See if we can refill from the shared array */
3092 if (shared && transfer_objects(ac, shared, batchcount)) {
3093 shared->touched = 1;
3097 while (batchcount > 0) {
3098 /* Get slab alloc is to come from. */
3099 page = get_first_slab(n, false);
3103 check_spinlock_acquired(cachep);
3105 batchcount = alloc_block(cachep, ac, page, batchcount);
3106 fixup_slab_list(cachep, n, page, &list);
3110 n->free_objects -= ac->avail;
3112 spin_unlock(&n->list_lock);
3113 fixup_objfreelist_debug(cachep, &list);
3116 if (unlikely(!ac->avail)) {
3117 /* Check if we can use obj in pfmemalloc slab */
3118 if (sk_memalloc_socks()) {
3119 void *obj = cache_alloc_pfmemalloc(cachep, n, flags);
3125 page = cache_grow_begin(cachep, gfp_exact_node(flags), node);
3128 * cache_grow_begin() can reenable interrupts,
3129 * then ac could change.
3131 ac = cpu_cache_get(cachep);
3132 if (!ac->avail && page)
3133 alloc_block(cachep, ac, page, batchcount);
3134 cache_grow_end(cachep, page);
3141 return ac->entry[--ac->avail];
3144 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3147 might_sleep_if(gfpflags_allow_blocking(flags));
3149 kmem_flagcheck(cachep, flags);
3154 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3155 gfp_t flags, void *objp, unsigned long caller)
3159 if (cachep->flags & SLAB_POISON) {
3160 check_poison_obj(cachep, objp);
3161 slab_kernel_map(cachep, objp, 1, 0);
3162 poison_obj(cachep, objp, POISON_INUSE);
3164 if (cachep->flags & SLAB_STORE_USER)
3165 *dbg_userword(cachep, objp) = (void *)caller;
3167 if (cachep->flags & SLAB_RED_ZONE) {
3168 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3169 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3170 slab_error(cachep, "double free, or memory outside object was overwritten");
3171 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3172 objp, *dbg_redzone1(cachep, objp),
3173 *dbg_redzone2(cachep, objp));
3175 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3176 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3179 objp += obj_offset(cachep);
3180 if (cachep->ctor && cachep->flags & SLAB_POISON)
3182 if (ARCH_SLAB_MINALIGN &&
3183 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3184 pr_err("0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3185 objp, (int)ARCH_SLAB_MINALIGN);
3190 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3193 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3196 struct array_cache *ac;
3200 ac = cpu_cache_get(cachep);
3201 if (likely(ac->avail)) {
3203 objp = ac->entry[--ac->avail];
3205 STATS_INC_ALLOCHIT(cachep);
3209 STATS_INC_ALLOCMISS(cachep);
3210 objp = cache_alloc_refill(cachep, flags);
3212 * the 'ac' may be updated by cache_alloc_refill(),
3213 * and kmemleak_erase() requires its correct value.
3215 ac = cpu_cache_get(cachep);
3219 * To avoid a false negative, if an object that is in one of the
3220 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3221 * treat the array pointers as a reference to the object.
3224 kmemleak_erase(&ac->entry[ac->avail]);
3230 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
3232 * If we are in_interrupt, then process context, including cpusets and
3233 * mempolicy, may not apply and should not be used for allocation policy.
3235 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3237 int nid_alloc, nid_here;
3239 if (in_interrupt() || (flags & __GFP_THISNODE))
3241 nid_alloc = nid_here = numa_mem_id();
3242 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3243 nid_alloc = cpuset_slab_spread_node();
3244 else if (current->mempolicy)
3245 nid_alloc = mempolicy_slab_node();
3246 if (nid_alloc != nid_here)
3247 return ____cache_alloc_node(cachep, flags, nid_alloc);
3252 * Fallback function if there was no memory available and no objects on a
3253 * certain node and fall back is permitted. First we scan all the
3254 * available node for available objects. If that fails then we
3255 * perform an allocation without specifying a node. This allows the page
3256 * allocator to do its reclaim / fallback magic. We then insert the
3257 * slab into the proper nodelist and then allocate from it.
3259 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3261 struct zonelist *zonelist;
3264 enum zone_type high_zoneidx = gfp_zone(flags);
3268 unsigned int cpuset_mems_cookie;
3270 if (flags & __GFP_THISNODE)
3274 cpuset_mems_cookie = read_mems_allowed_begin();
3275 zonelist = node_zonelist(mempolicy_slab_node(), flags);
3279 * Look through allowed nodes for objects available
3280 * from existing per node queues.
3282 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3283 nid = zone_to_nid(zone);
3285 if (cpuset_zone_allowed(zone, flags) &&
3286 get_node(cache, nid) &&
3287 get_node(cache, nid)->free_objects) {
3288 obj = ____cache_alloc_node(cache,
3289 gfp_exact_node(flags), nid);
3297 * This allocation will be performed within the constraints
3298 * of the current cpuset / memory policy requirements.
3299 * We may trigger various forms of reclaim on the allowed
3300 * set and go into memory reserves if necessary.
3302 page = cache_grow_begin(cache, flags, numa_mem_id());
3303 cache_grow_end(cache, page);
3305 nid = page_to_nid(page);
3306 obj = ____cache_alloc_node(cache,
3307 gfp_exact_node(flags), nid);
3310 * Another processor may allocate the objects in
3311 * the slab since we are not holding any locks.
3318 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
3324 * A interface to enable slab creation on nodeid
3326 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3330 struct kmem_cache_node *n;
3334 VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
3335 n = get_node(cachep, nodeid);
3339 spin_lock(&n->list_lock);
3340 page = get_first_slab(n, false);
3344 check_spinlock_acquired_node(cachep, nodeid);
3346 STATS_INC_NODEALLOCS(cachep);
3347 STATS_INC_ACTIVE(cachep);
3348 STATS_SET_HIGH(cachep);
3350 BUG_ON(page->active == cachep->num);
3352 obj = slab_get_obj(cachep, page);
3355 fixup_slab_list(cachep, n, page, &list);
3357 spin_unlock(&n->list_lock);
3358 fixup_objfreelist_debug(cachep, &list);
3362 spin_unlock(&n->list_lock);
3363 page = cache_grow_begin(cachep, gfp_exact_node(flags), nodeid);
3365 /* This slab isn't counted yet so don't update free_objects */
3366 obj = slab_get_obj(cachep, page);
3368 cache_grow_end(cachep, page);
3370 return obj ? obj : fallback_alloc(cachep, flags);
3373 static __always_inline void *
3374 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3375 unsigned long caller)
3377 unsigned long save_flags;
3379 int slab_node = numa_mem_id();
3381 flags &= gfp_allowed_mask;
3382 cachep = slab_pre_alloc_hook(cachep, flags);
3383 if (unlikely(!cachep))
3386 cache_alloc_debugcheck_before(cachep, flags);
3387 local_irq_save(save_flags);
3389 if (nodeid == NUMA_NO_NODE)
3392 if (unlikely(!get_node(cachep, nodeid))) {
3393 /* Node not bootstrapped yet */
3394 ptr = fallback_alloc(cachep, flags);
3398 if (nodeid == slab_node) {
3400 * Use the locally cached objects if possible.
3401 * However ____cache_alloc does not allow fallback
3402 * to other nodes. It may fail while we still have
3403 * objects on other nodes available.
3405 ptr = ____cache_alloc(cachep, flags);
3409 /* ___cache_alloc_node can fall back to other nodes */
3410 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3412 local_irq_restore(save_flags);
3413 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3415 if (unlikely(flags & __GFP_ZERO) && ptr)
3416 memset(ptr, 0, cachep->object_size);
3418 slab_post_alloc_hook(cachep, flags, 1, &ptr);
3422 static __always_inline void *
3423 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3427 if (current->mempolicy || cpuset_do_slab_mem_spread()) {
3428 objp = alternate_node_alloc(cache, flags);
3432 objp = ____cache_alloc(cache, flags);
3435 * We may just have run out of memory on the local node.
3436 * ____cache_alloc_node() knows how to locate memory on other nodes
3439 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3446 static __always_inline void *
3447 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3449 return ____cache_alloc(cachep, flags);
3452 #endif /* CONFIG_NUMA */
3454 static __always_inline void *
3455 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3457 unsigned long save_flags;
3460 flags &= gfp_allowed_mask;
3461 cachep = slab_pre_alloc_hook(cachep, flags);
3462 if (unlikely(!cachep))
3465 cache_alloc_debugcheck_before(cachep, flags);
3466 local_irq_save(save_flags);
3467 objp = __do_cache_alloc(cachep, flags);
3468 local_irq_restore(save_flags);
3469 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3472 if (unlikely(flags & __GFP_ZERO) && objp)
3473 memset(objp, 0, cachep->object_size);
3475 slab_post_alloc_hook(cachep, flags, 1, &objp);
3480 * Caller needs to acquire correct kmem_cache_node's list_lock
3481 * @list: List of detached free slabs should be freed by caller
3483 static void free_block(struct kmem_cache *cachep, void **objpp,
3484 int nr_objects, int node, struct list_head *list)
3487 struct kmem_cache_node *n = get_node(cachep, node);
3490 n->free_objects += nr_objects;
3492 for (i = 0; i < nr_objects; i++) {
3498 page = virt_to_head_page(objp);
3499 list_del(&page->lru);
3500 check_spinlock_acquired_node(cachep, node);
3501 slab_put_obj(cachep, page, objp);
3502 STATS_DEC_ACTIVE(cachep);
3504 /* fixup slab chains */
3505 if (page->active == 0)
3506 list_add(&page->lru, &n->slabs_free);
3508 /* Unconditionally move a slab to the end of the
3509 * partial list on free - maximum time for the
3510 * other objects to be freed, too.
3512 list_add_tail(&page->lru, &n->slabs_partial);
3516 while (n->free_objects > n->free_limit && !list_empty(&n->slabs_free)) {
3517 n->free_objects -= cachep->num;
3519 page = list_last_entry(&n->slabs_free, struct page, lru);
3520 list_del(&page->lru);
3521 list_add(&page->lru, list);
3525 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3528 struct kmem_cache_node *n;
3529 int node = numa_mem_id();
3532 batchcount = ac->batchcount;
3535 n = get_node(cachep, node);
3536 spin_lock(&n->list_lock);
3538 struct array_cache *shared_array = n->shared;
3539 int max = shared_array->limit - shared_array->avail;
3541 if (batchcount > max)
3543 memcpy(&(shared_array->entry[shared_array->avail]),
3544 ac->entry, sizeof(void *) * batchcount);
3545 shared_array->avail += batchcount;
3550 free_block(cachep, ac->entry, batchcount, node, &list);
3557 list_for_each_entry(page, &n->slabs_free, lru) {
3558 BUG_ON(page->active);
3562 STATS_SET_FREEABLE(cachep, i);
3565 spin_unlock(&n->list_lock);
3566 slabs_destroy(cachep, &list);
3567 ac->avail -= batchcount;
3568 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3572 * Release an obj back to its cache. If the obj has a constructed state, it must
3573 * be in this state _before_ it is released. Called with disabled ints.
3575 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3576 unsigned long caller)
3578 struct array_cache *ac = cpu_cache_get(cachep);
3580 kasan_slab_free(cachep, objp);
3583 kmemleak_free_recursive(objp, cachep->flags);
3584 objp = cache_free_debugcheck(cachep, objp, caller);
3586 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3589 * Skip calling cache_free_alien() when the platform is not numa.
3590 * This will avoid cache misses that happen while accessing slabp (which
3591 * is per page memory reference) to get nodeid. Instead use a global
3592 * variable to skip the call, which is mostly likely to be present in
3595 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3598 if (ac->avail < ac->limit) {
3599 STATS_INC_FREEHIT(cachep);
3601 STATS_INC_FREEMISS(cachep);
3602 cache_flusharray(cachep, ac);
3605 if (sk_memalloc_socks()) {
3606 struct page *page = virt_to_head_page(objp);
3608 if (unlikely(PageSlabPfmemalloc(page))) {
3609 cache_free_pfmemalloc(cachep, page, objp);
3614 ac->entry[ac->avail++] = objp;
3618 * kmem_cache_alloc - Allocate an object
3619 * @cachep: The cache to allocate from.
3620 * @flags: See kmalloc().
3622 * Allocate an object from this cache. The flags are only relevant
3623 * if the cache has no available objects.
3625 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3627 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3629 kasan_slab_alloc(cachep, ret, flags);
3630 trace_kmem_cache_alloc(_RET_IP_, ret,
3631 cachep->object_size, cachep->size, flags);
3635 EXPORT_SYMBOL(kmem_cache_alloc);
3637 static __always_inline void
3638 cache_alloc_debugcheck_after_bulk(struct kmem_cache *s, gfp_t flags,
3639 size_t size, void **p, unsigned long caller)
3643 for (i = 0; i < size; i++)
3644 p[i] = cache_alloc_debugcheck_after(s, flags, p[i], caller);
3647 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3652 s = slab_pre_alloc_hook(s, flags);
3656 cache_alloc_debugcheck_before(s, flags);
3658 local_irq_disable();
3659 for (i = 0; i < size; i++) {
3660 void *objp = __do_cache_alloc(s, flags);
3662 if (unlikely(!objp))
3668 cache_alloc_debugcheck_after_bulk(s, flags, size, p, _RET_IP_);
3670 /* Clear memory outside IRQ disabled section */
3671 if (unlikely(flags & __GFP_ZERO))
3672 for (i = 0; i < size; i++)
3673 memset(p[i], 0, s->object_size);
3675 slab_post_alloc_hook(s, flags, size, p);
3676 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3680 cache_alloc_debugcheck_after_bulk(s, flags, i, p, _RET_IP_);
3681 slab_post_alloc_hook(s, flags, i, p);
3682 __kmem_cache_free_bulk(s, i, p);
3685 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3687 #ifdef CONFIG_TRACING
3689 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3693 ret = slab_alloc(cachep, flags, _RET_IP_);
3695 kasan_kmalloc(cachep, ret, size, flags);
3696 trace_kmalloc(_RET_IP_, ret,
3697 size, cachep->size, flags);
3700 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3705 * kmem_cache_alloc_node - Allocate an object on the specified node
3706 * @cachep: The cache to allocate from.
3707 * @flags: See kmalloc().
3708 * @nodeid: node number of the target node.
3710 * Identical to kmem_cache_alloc but it will allocate memory on the given
3711 * node, which can improve the performance for cpu bound structures.
3713 * Fallback to other node is possible if __GFP_THISNODE is not set.
3715 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3717 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3719 kasan_slab_alloc(cachep, ret, flags);
3720 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3721 cachep->object_size, cachep->size,
3726 EXPORT_SYMBOL(kmem_cache_alloc_node);
3728 #ifdef CONFIG_TRACING
3729 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3736 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3738 kasan_kmalloc(cachep, ret, size, flags);
3739 trace_kmalloc_node(_RET_IP_, ret,
3744 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3747 static __always_inline void *
3748 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3750 struct kmem_cache *cachep;
3753 cachep = kmalloc_slab(size, flags);
3754 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3756 ret = kmem_cache_alloc_node_trace(cachep, flags, node, size);
3757 kasan_kmalloc(cachep, ret, size, flags);
3762 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3764 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3766 EXPORT_SYMBOL(__kmalloc_node);
3768 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3769 int node, unsigned long caller)
3771 return __do_kmalloc_node(size, flags, node, caller);
3773 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3774 #endif /* CONFIG_NUMA */
3777 * __do_kmalloc - allocate memory
3778 * @size: how many bytes of memory are required.
3779 * @flags: the type of memory to allocate (see kmalloc).
3780 * @caller: function caller for debug tracking of the caller
3782 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3783 unsigned long caller)
3785 struct kmem_cache *cachep;
3788 cachep = kmalloc_slab(size, flags);
3789 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3791 ret = slab_alloc(cachep, flags, caller);
3793 kasan_kmalloc(cachep, ret, size, flags);
3794 trace_kmalloc(caller, ret,
3795 size, cachep->size, flags);
3800 void *__kmalloc(size_t size, gfp_t flags)
3802 return __do_kmalloc(size, flags, _RET_IP_);
3804 EXPORT_SYMBOL(__kmalloc);
3806 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3808 return __do_kmalloc(size, flags, caller);
3810 EXPORT_SYMBOL(__kmalloc_track_caller);
3813 * kmem_cache_free - Deallocate an object
3814 * @cachep: The cache the allocation was from.
3815 * @objp: The previously allocated object.
3817 * Free an object which was previously allocated from this
3820 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3822 unsigned long flags;
3823 cachep = cache_from_obj(cachep, objp);
3827 local_irq_save(flags);
3828 debug_check_no_locks_freed(objp, cachep->object_size);
3829 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3830 debug_check_no_obj_freed(objp, cachep->object_size);
3831 __cache_free(cachep, objp, _RET_IP_);
3832 local_irq_restore(flags);
3834 trace_kmem_cache_free(_RET_IP_, objp);
3836 EXPORT_SYMBOL(kmem_cache_free);
3838 void kmem_cache_free_bulk(struct kmem_cache *orig_s, size_t size, void **p)
3840 struct kmem_cache *s;
3843 local_irq_disable();
3844 for (i = 0; i < size; i++) {
3847 if (!orig_s) /* called via kfree_bulk */
3848 s = virt_to_cache(objp);
3850 s = cache_from_obj(orig_s, objp);
3852 debug_check_no_locks_freed(objp, s->object_size);
3853 if (!(s->flags & SLAB_DEBUG_OBJECTS))
3854 debug_check_no_obj_freed(objp, s->object_size);
3856 __cache_free(s, objp, _RET_IP_);
3860 /* FIXME: add tracing */
3862 EXPORT_SYMBOL(kmem_cache_free_bulk);
3865 * kfree - free previously allocated memory
3866 * @objp: pointer returned by kmalloc.
3868 * If @objp is NULL, no operation is performed.
3870 * Don't free memory not originally allocated by kmalloc()
3871 * or you will run into trouble.
3873 void kfree(const void *objp)
3875 struct kmem_cache *c;
3876 unsigned long flags;
3878 trace_kfree(_RET_IP_, objp);
3880 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3882 local_irq_save(flags);
3883 kfree_debugcheck(objp);
3884 c = virt_to_cache(objp);
3885 debug_check_no_locks_freed(objp, c->object_size);
3887 debug_check_no_obj_freed(objp, c->object_size);
3888 __cache_free(c, (void *)objp, _RET_IP_);
3889 local_irq_restore(flags);
3891 EXPORT_SYMBOL(kfree);
3894 * This initializes kmem_cache_node or resizes various caches for all nodes.
3896 static int setup_kmem_cache_nodes(struct kmem_cache *cachep, gfp_t gfp)
3900 struct kmem_cache_node *n;
3902 for_each_online_node(node) {
3903 ret = setup_kmem_cache_node(cachep, node, gfp, true);
3912 if (!cachep->list.next) {
3913 /* Cache is not active yet. Roll back what we did */
3916 n = get_node(cachep, node);
3919 free_alien_cache(n->alien);
3921 cachep->node[node] = NULL;
3929 /* Always called with the slab_mutex held */
3930 static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
3931 int batchcount, int shared, gfp_t gfp)
3933 struct array_cache __percpu *cpu_cache, *prev;
3936 cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3940 prev = cachep->cpu_cache;
3941 cachep->cpu_cache = cpu_cache;
3942 kick_all_cpus_sync();
3945 cachep->batchcount = batchcount;
3946 cachep->limit = limit;
3947 cachep->shared = shared;
3952 for_each_online_cpu(cpu) {
3955 struct kmem_cache_node *n;
3956 struct array_cache *ac = per_cpu_ptr(prev, cpu);
3958 node = cpu_to_mem(cpu);
3959 n = get_node(cachep, node);
3960 spin_lock_irq(&n->list_lock);
3961 free_block(cachep, ac->entry, ac->avail, node, &list);
3962 spin_unlock_irq(&n->list_lock);
3963 slabs_destroy(cachep, &list);
3968 return setup_kmem_cache_nodes(cachep, gfp);
3971 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3972 int batchcount, int shared, gfp_t gfp)
3975 struct kmem_cache *c;
3977 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3979 if (slab_state < FULL)
3982 if ((ret < 0) || !is_root_cache(cachep))
3985 lockdep_assert_held(&slab_mutex);
3986 for_each_memcg_cache(c, cachep) {
3987 /* return value determined by the root cache only */
3988 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3994 /* Called with slab_mutex held always */
3995 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
4002 err = cache_random_seq_create(cachep, gfp);
4006 if (!is_root_cache(cachep)) {
4007 struct kmem_cache *root = memcg_root_cache(cachep);
4008 limit = root->limit;
4009 shared = root->shared;
4010 batchcount = root->batchcount;
4013 if (limit && shared && batchcount)
4016 * The head array serves three purposes:
4017 * - create a LIFO ordering, i.e. return objects that are cache-warm
4018 * - reduce the number of spinlock operations.
4019 * - reduce the number of linked list operations on the slab and
4020 * bufctl chains: array operations are cheaper.
4021 * The numbers are guessed, we should auto-tune as described by
4024 if (cachep->size > 131072)
4026 else if (cachep->size > PAGE_SIZE)
4028 else if (cachep->size > 1024)
4030 else if (cachep->size > 256)
4036 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4037 * allocation behaviour: Most allocs on one cpu, most free operations
4038 * on another cpu. For these cases, an efficient object passing between
4039 * cpus is necessary. This is provided by a shared array. The array
4040 * replaces Bonwick's magazine layer.
4041 * On uniprocessor, it's functionally equivalent (but less efficient)
4042 * to a larger limit. Thus disabled by default.
4045 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
4050 * With debugging enabled, large batchcount lead to excessively long
4051 * periods with disabled local interrupts. Limit the batchcount
4056 batchcount = (limit + 1) / 2;
4058 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
4061 pr_err("enable_cpucache failed for %s, error %d\n",
4062 cachep->name, -err);
4067 * Drain an array if it contains any elements taking the node lock only if
4068 * necessary. Note that the node listlock also protects the array_cache
4069 * if drain_array() is used on the shared array.
4071 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
4072 struct array_cache *ac, int node)
4076 /* ac from n->shared can be freed if we don't hold the slab_mutex. */
4077 check_mutex_acquired();
4079 if (!ac || !ac->avail)
4087 spin_lock_irq(&n->list_lock);
4088 drain_array_locked(cachep, ac, node, false, &list);
4089 spin_unlock_irq(&n->list_lock);
4091 slabs_destroy(cachep, &list);
4095 * cache_reap - Reclaim memory from caches.
4096 * @w: work descriptor
4098 * Called from workqueue/eventd every few seconds.
4100 * - clear the per-cpu caches for this CPU.
4101 * - return freeable pages to the main free memory pool.
4103 * If we cannot acquire the cache chain mutex then just give up - we'll try
4104 * again on the next iteration.
4106 static void cache_reap(struct work_struct *w)
4108 struct kmem_cache *searchp;
4109 struct kmem_cache_node *n;
4110 int node = numa_mem_id();
4111 struct delayed_work *work = to_delayed_work(w);
4113 if (!mutex_trylock(&slab_mutex))
4114 /* Give up. Setup the next iteration. */
4117 list_for_each_entry(searchp, &slab_caches, list) {
4121 * We only take the node lock if absolutely necessary and we
4122 * have established with reasonable certainty that
4123 * we can do some work if the lock was obtained.
4125 n = get_node(searchp, node);
4127 reap_alien(searchp, n);
4129 drain_array(searchp, n, cpu_cache_get(searchp), node);
4132 * These are racy checks but it does not matter
4133 * if we skip one check or scan twice.
4135 if (time_after(n->next_reap, jiffies))
4138 n->next_reap = jiffies + REAPTIMEOUT_NODE;
4140 drain_array(searchp, n, n->shared, node);
4142 if (n->free_touched)
4143 n->free_touched = 0;
4147 freed = drain_freelist(searchp, n, (n->free_limit +
4148 5 * searchp->num - 1) / (5 * searchp->num));
4149 STATS_ADD_REAPED(searchp, freed);
4155 mutex_unlock(&slab_mutex);
4158 /* Set up the next iteration */
4159 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC));
4162 #ifdef CONFIG_SLABINFO
4163 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
4166 unsigned long active_objs;
4167 unsigned long num_objs;
4168 unsigned long active_slabs = 0;
4169 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4173 struct kmem_cache_node *n;
4177 for_each_kmem_cache_node(cachep, node, n) {
4180 spin_lock_irq(&n->list_lock);
4182 list_for_each_entry(page, &n->slabs_full, lru) {
4183 if (page->active != cachep->num && !error)
4184 error = "slabs_full accounting error";
4185 active_objs += cachep->num;
4188 list_for_each_entry(page, &n->slabs_partial, lru) {
4189 if (page->active == cachep->num && !error)
4190 error = "slabs_partial accounting error";
4191 if (!page->active && !error)
4192 error = "slabs_partial accounting error";
4193 active_objs += page->active;
4196 list_for_each_entry(page, &n->slabs_free, lru) {
4197 if (page->active && !error)
4198 error = "slabs_free accounting error";
4201 free_objects += n->free_objects;
4203 shared_avail += n->shared->avail;
4205 spin_unlock_irq(&n->list_lock);
4207 num_slabs += active_slabs;
4208 num_objs = num_slabs * cachep->num;
4209 if (num_objs - active_objs != free_objects && !error)
4210 error = "free_objects accounting error";
4212 name = cachep->name;
4214 pr_err("slab: cache %s error: %s\n", name, error);
4216 sinfo->active_objs = active_objs;
4217 sinfo->num_objs = num_objs;
4218 sinfo->active_slabs = active_slabs;
4219 sinfo->num_slabs = num_slabs;
4220 sinfo->shared_avail = shared_avail;
4221 sinfo->limit = cachep->limit;
4222 sinfo->batchcount = cachep->batchcount;
4223 sinfo->shared = cachep->shared;
4224 sinfo->objects_per_slab = cachep->num;
4225 sinfo->cache_order = cachep->gfporder;
4228 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4232 unsigned long high = cachep->high_mark;
4233 unsigned long allocs = cachep->num_allocations;
4234 unsigned long grown = cachep->grown;
4235 unsigned long reaped = cachep->reaped;
4236 unsigned long errors = cachep->errors;
4237 unsigned long max_freeable = cachep->max_freeable;
4238 unsigned long node_allocs = cachep->node_allocs;
4239 unsigned long node_frees = cachep->node_frees;
4240 unsigned long overflows = cachep->node_overflow;
4242 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
4243 allocs, high, grown,
4244 reaped, errors, max_freeable, node_allocs,
4245 node_frees, overflows);
4249 unsigned long allochit = atomic_read(&cachep->allochit);
4250 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4251 unsigned long freehit = atomic_read(&cachep->freehit);
4252 unsigned long freemiss = atomic_read(&cachep->freemiss);
4254 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4255 allochit, allocmiss, freehit, freemiss);
4260 #define MAX_SLABINFO_WRITE 128
4262 * slabinfo_write - Tuning for the slab allocator
4264 * @buffer: user buffer
4265 * @count: data length
4268 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4269 size_t count, loff_t *ppos)
4271 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4272 int limit, batchcount, shared, res;
4273 struct kmem_cache *cachep;
4275 if (count > MAX_SLABINFO_WRITE)
4277 if (copy_from_user(&kbuf, buffer, count))
4279 kbuf[MAX_SLABINFO_WRITE] = '\0';
4281 tmp = strchr(kbuf, ' ');
4286 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4289 /* Find the cache in the chain of caches. */
4290 mutex_lock(&slab_mutex);
4292 list_for_each_entry(cachep, &slab_caches, list) {
4293 if (!strcmp(cachep->name, kbuf)) {
4294 if (limit < 1 || batchcount < 1 ||
4295 batchcount > limit || shared < 0) {
4298 res = do_tune_cpucache(cachep, limit,
4305 mutex_unlock(&slab_mutex);
4311 #ifdef CONFIG_DEBUG_SLAB_LEAK
4313 static inline int add_caller(unsigned long *n, unsigned long v)
4323 unsigned long *q = p + 2 * i;
4337 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4343 static void handle_slab(unsigned long *n, struct kmem_cache *c,
4352 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
4355 for (j = page->active; j < c->num; j++) {
4356 if (get_free_obj(page, j) == i) {
4366 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4367 * mapping is established when actual object allocation and
4368 * we could mistakenly access the unmapped object in the cpu
4371 if (probe_kernel_read(&v, dbg_userword(c, p), sizeof(v)))
4374 if (!add_caller(n, v))
4379 static void show_symbol(struct seq_file *m, unsigned long address)
4381 #ifdef CONFIG_KALLSYMS
4382 unsigned long offset, size;
4383 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4385 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4386 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4388 seq_printf(m, " [%s]", modname);
4392 seq_printf(m, "%p", (void *)address);
4395 static int leaks_show(struct seq_file *m, void *p)
4397 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4399 struct kmem_cache_node *n;
4401 unsigned long *x = m->private;
4405 if (!(cachep->flags & SLAB_STORE_USER))
4407 if (!(cachep->flags & SLAB_RED_ZONE))
4411 * Set store_user_clean and start to grab stored user information
4412 * for all objects on this cache. If some alloc/free requests comes
4413 * during the processing, information would be wrong so restart
4417 set_store_user_clean(cachep);
4418 drain_cpu_caches(cachep);
4422 for_each_kmem_cache_node(cachep, node, n) {
4425 spin_lock_irq(&n->list_lock);
4427 list_for_each_entry(page, &n->slabs_full, lru)
4428 handle_slab(x, cachep, page);
4429 list_for_each_entry(page, &n->slabs_partial, lru)
4430 handle_slab(x, cachep, page);
4431 spin_unlock_irq(&n->list_lock);
4433 } while (!is_store_user_clean(cachep));
4435 name = cachep->name;
4437 /* Increase the buffer size */
4438 mutex_unlock(&slab_mutex);
4439 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4441 /* Too bad, we are really out */
4443 mutex_lock(&slab_mutex);
4446 *(unsigned long *)m->private = x[0] * 2;
4448 mutex_lock(&slab_mutex);
4449 /* Now make sure this entry will be retried */
4453 for (i = 0; i < x[1]; i++) {
4454 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4455 show_symbol(m, x[2*i+2]);
4462 static const struct seq_operations slabstats_op = {
4463 .start = slab_start,
4469 static int slabstats_open(struct inode *inode, struct file *file)
4473 n = __seq_open_private(file, &slabstats_op, PAGE_SIZE);
4477 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4482 static const struct file_operations proc_slabstats_operations = {
4483 .open = slabstats_open,
4485 .llseek = seq_lseek,
4486 .release = seq_release_private,
4490 static int __init slab_proc_init(void)
4492 #ifdef CONFIG_DEBUG_SLAB_LEAK
4493 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4497 module_init(slab_proc_init);
4501 * ksize - get the actual amount of memory allocated for a given object
4502 * @objp: Pointer to the object
4504 * kmalloc may internally round up allocations and return more memory
4505 * than requested. ksize() can be used to determine the actual amount of
4506 * memory allocated. The caller may use this additional memory, even though
4507 * a smaller amount of memory was initially specified with the kmalloc call.
4508 * The caller must guarantee that objp points to a valid object previously
4509 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4510 * must not be freed during the duration of the call.
4512 size_t ksize(const void *objp)
4517 if (unlikely(objp == ZERO_SIZE_PTR))
4520 size = virt_to_cache(objp)->object_size;
4521 /* We assume that ksize callers could use the whole allocated area,
4522 * so we need to unpoison this area.
4524 kasan_krealloc(objp, size, GFP_NOWAIT);
4528 EXPORT_SYMBOL(ksize);