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 int slab_early_init = 1;
218 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
220 static void kmem_cache_node_init(struct kmem_cache_node *parent)
222 INIT_LIST_HEAD(&parent->slabs_full);
223 INIT_LIST_HEAD(&parent->slabs_partial);
224 INIT_LIST_HEAD(&parent->slabs_free);
225 parent->shared = NULL;
226 parent->alien = NULL;
227 parent->colour_next = 0;
228 spin_lock_init(&parent->list_lock);
229 parent->free_objects = 0;
230 parent->free_touched = 0;
233 #define MAKE_LIST(cachep, listp, slab, nodeid) \
235 INIT_LIST_HEAD(listp); \
236 list_splice(&get_node(cachep, nodeid)->slab, listp); \
239 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
241 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
242 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
243 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
246 #define CFLGS_OBJFREELIST_SLAB (0x40000000UL)
247 #define CFLGS_OFF_SLAB (0x80000000UL)
248 #define OBJFREELIST_SLAB(x) ((x)->flags & CFLGS_OBJFREELIST_SLAB)
249 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
251 #define BATCHREFILL_LIMIT 16
253 * Optimization question: fewer reaps means less probability for unnessary
254 * cpucache drain/refill cycles.
256 * OTOH the cpuarrays can contain lots of objects,
257 * which could lock up otherwise freeable slabs.
259 #define REAPTIMEOUT_AC (2*HZ)
260 #define REAPTIMEOUT_NODE (4*HZ)
263 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
264 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
265 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
266 #define STATS_INC_GROWN(x) ((x)->grown++)
267 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
268 #define STATS_SET_HIGH(x) \
270 if ((x)->num_active > (x)->high_mark) \
271 (x)->high_mark = (x)->num_active; \
273 #define STATS_INC_ERR(x) ((x)->errors++)
274 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
275 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
276 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
277 #define STATS_SET_FREEABLE(x, i) \
279 if ((x)->max_freeable < i) \
280 (x)->max_freeable = i; \
282 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
283 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
284 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
285 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
287 #define STATS_INC_ACTIVE(x) do { } while (0)
288 #define STATS_DEC_ACTIVE(x) do { } while (0)
289 #define STATS_INC_ALLOCED(x) do { } while (0)
290 #define STATS_INC_GROWN(x) do { } while (0)
291 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
292 #define STATS_SET_HIGH(x) do { } while (0)
293 #define STATS_INC_ERR(x) do { } while (0)
294 #define STATS_INC_NODEALLOCS(x) do { } while (0)
295 #define STATS_INC_NODEFREES(x) do { } while (0)
296 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
297 #define STATS_SET_FREEABLE(x, i) do { } while (0)
298 #define STATS_INC_ALLOCHIT(x) do { } while (0)
299 #define STATS_INC_ALLOCMISS(x) do { } while (0)
300 #define STATS_INC_FREEHIT(x) do { } while (0)
301 #define STATS_INC_FREEMISS(x) do { } while (0)
307 * memory layout of objects:
309 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
310 * the end of an object is aligned with the end of the real
311 * allocation. Catches writes behind the end of the allocation.
312 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
314 * cachep->obj_offset: The real object.
315 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
316 * cachep->size - 1* BYTES_PER_WORD: last caller address
317 * [BYTES_PER_WORD long]
319 static int obj_offset(struct kmem_cache *cachep)
321 return cachep->obj_offset;
324 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
326 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
327 return (unsigned long long*) (objp + obj_offset(cachep) -
328 sizeof(unsigned long long));
331 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
333 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
334 if (cachep->flags & SLAB_STORE_USER)
335 return (unsigned long long *)(objp + cachep->size -
336 sizeof(unsigned long long) -
338 return (unsigned long long *) (objp + cachep->size -
339 sizeof(unsigned long long));
342 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
344 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
345 return (void **)(objp + cachep->size - BYTES_PER_WORD);
350 #define obj_offset(x) 0
351 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
352 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
353 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
357 #ifdef CONFIG_DEBUG_SLAB_LEAK
359 static inline bool is_store_user_clean(struct kmem_cache *cachep)
361 return atomic_read(&cachep->store_user_clean) == 1;
364 static inline void set_store_user_clean(struct kmem_cache *cachep)
366 atomic_set(&cachep->store_user_clean, 1);
369 static inline void set_store_user_dirty(struct kmem_cache *cachep)
371 if (is_store_user_clean(cachep))
372 atomic_set(&cachep->store_user_clean, 0);
376 static inline void set_store_user_dirty(struct kmem_cache *cachep) {}
381 * Do not go above this order unless 0 objects fit into the slab or
382 * overridden on the command line.
384 #define SLAB_MAX_ORDER_HI 1
385 #define SLAB_MAX_ORDER_LO 0
386 static int slab_max_order = SLAB_MAX_ORDER_LO;
387 static bool slab_max_order_set __initdata;
389 static inline struct kmem_cache *virt_to_cache(const void *obj)
391 struct page *page = virt_to_head_page(obj);
392 return page->slab_cache;
395 static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
398 return page->s_mem + cache->size * idx;
402 * We want to avoid an expensive divide : (offset / cache->size)
403 * Using the fact that size is a constant for a particular cache,
404 * we can replace (offset / cache->size) by
405 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
407 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
408 const struct page *page, void *obj)
410 u32 offset = (obj - page->s_mem);
411 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
414 #define BOOT_CPUCACHE_ENTRIES 1
415 /* internal cache of cache description objs */
416 static struct kmem_cache kmem_cache_boot = {
418 .limit = BOOT_CPUCACHE_ENTRIES,
420 .size = sizeof(struct kmem_cache),
421 .name = "kmem_cache",
424 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
426 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
428 return this_cpu_ptr(cachep->cpu_cache);
432 * Calculate the number of objects and left-over bytes for a given buffer size.
434 static unsigned int cache_estimate(unsigned long gfporder, size_t buffer_size,
435 unsigned long flags, size_t *left_over)
438 size_t slab_size = PAGE_SIZE << gfporder;
441 * The slab management structure can be either off the slab or
442 * on it. For the latter case, the memory allocated for a
445 * - @buffer_size bytes for each object
446 * - One freelist_idx_t for each object
448 * We don't need to consider alignment of freelist because
449 * freelist will be at the end of slab page. The objects will be
450 * at the correct alignment.
452 * If the slab management structure is off the slab, then the
453 * alignment will already be calculated into the size. Because
454 * the slabs are all pages aligned, the objects will be at the
455 * correct alignment when allocated.
457 if (flags & (CFLGS_OBJFREELIST_SLAB | CFLGS_OFF_SLAB)) {
458 num = slab_size / buffer_size;
459 *left_over = slab_size % buffer_size;
461 num = slab_size / (buffer_size + sizeof(freelist_idx_t));
462 *left_over = slab_size %
463 (buffer_size + sizeof(freelist_idx_t));
470 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
472 static void __slab_error(const char *function, struct kmem_cache *cachep,
475 pr_err("slab error in %s(): cache `%s': %s\n",
476 function, cachep->name, msg);
478 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
483 * By default on NUMA we use alien caches to stage the freeing of
484 * objects allocated from other nodes. This causes massive memory
485 * inefficiencies when using fake NUMA setup to split memory into a
486 * large number of small nodes, so it can be disabled on the command
490 static int use_alien_caches __read_mostly = 1;
491 static int __init noaliencache_setup(char *s)
493 use_alien_caches = 0;
496 __setup("noaliencache", noaliencache_setup);
498 static int __init slab_max_order_setup(char *str)
500 get_option(&str, &slab_max_order);
501 slab_max_order = slab_max_order < 0 ? 0 :
502 min(slab_max_order, MAX_ORDER - 1);
503 slab_max_order_set = true;
507 __setup("slab_max_order=", slab_max_order_setup);
511 * Special reaping functions for NUMA systems called from cache_reap().
512 * These take care of doing round robin flushing of alien caches (containing
513 * objects freed on different nodes from which they were allocated) and the
514 * flushing of remote pcps by calling drain_node_pages.
516 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
518 static void init_reap_node(int cpu)
522 node = next_node(cpu_to_mem(cpu), node_online_map);
523 if (node == MAX_NUMNODES)
524 node = first_node(node_online_map);
526 per_cpu(slab_reap_node, cpu) = node;
529 static void next_reap_node(void)
531 int node = __this_cpu_read(slab_reap_node);
533 node = next_node(node, node_online_map);
534 if (unlikely(node >= MAX_NUMNODES))
535 node = first_node(node_online_map);
536 __this_cpu_write(slab_reap_node, node);
540 #define init_reap_node(cpu) do { } while (0)
541 #define next_reap_node(void) do { } while (0)
545 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
546 * via the workqueue/eventd.
547 * Add the CPU number into the expiration time to minimize the possibility of
548 * the CPUs getting into lockstep and contending for the global cache chain
551 static void start_cpu_timer(int cpu)
553 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
556 * When this gets called from do_initcalls via cpucache_init(),
557 * init_workqueues() has already run, so keventd will be setup
560 if (keventd_up() && reap_work->work.func == NULL) {
562 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
563 schedule_delayed_work_on(cpu, reap_work,
564 __round_jiffies_relative(HZ, cpu));
568 static void init_arraycache(struct array_cache *ac, int limit, int batch)
571 * The array_cache structures contain pointers to free object.
572 * However, when such objects are allocated or transferred to another
573 * cache the pointers are not cleared and they could be counted as
574 * valid references during a kmemleak scan. Therefore, kmemleak must
575 * not scan such objects.
577 kmemleak_no_scan(ac);
581 ac->batchcount = batch;
586 static struct array_cache *alloc_arraycache(int node, int entries,
587 int batchcount, gfp_t gfp)
589 size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache);
590 struct array_cache *ac = NULL;
592 ac = kmalloc_node(memsize, gfp, node);
593 init_arraycache(ac, entries, batchcount);
597 static noinline void cache_free_pfmemalloc(struct kmem_cache *cachep,
598 struct page *page, void *objp)
600 struct kmem_cache_node *n;
604 page_node = page_to_nid(page);
605 n = get_node(cachep, page_node);
607 spin_lock(&n->list_lock);
608 free_block(cachep, &objp, 1, page_node, &list);
609 spin_unlock(&n->list_lock);
611 slabs_destroy(cachep, &list);
615 * Transfer objects in one arraycache to another.
616 * Locking must be handled by the caller.
618 * Return the number of entries transferred.
620 static int transfer_objects(struct array_cache *to,
621 struct array_cache *from, unsigned int max)
623 /* Figure out how many entries to transfer */
624 int nr = min3(from->avail, max, to->limit - to->avail);
629 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
639 #define drain_alien_cache(cachep, alien) do { } while (0)
640 #define reap_alien(cachep, n) do { } while (0)
642 static inline struct alien_cache **alloc_alien_cache(int node,
643 int limit, gfp_t gfp)
648 static inline void free_alien_cache(struct alien_cache **ac_ptr)
652 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
657 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
663 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
664 gfp_t flags, int nodeid)
669 static inline gfp_t gfp_exact_node(gfp_t flags)
671 return flags & ~__GFP_NOFAIL;
674 #else /* CONFIG_NUMA */
676 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
677 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
679 static struct alien_cache *__alloc_alien_cache(int node, int entries,
680 int batch, gfp_t gfp)
682 size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
683 struct alien_cache *alc = NULL;
685 alc = kmalloc_node(memsize, gfp, node);
686 init_arraycache(&alc->ac, entries, batch);
687 spin_lock_init(&alc->lock);
691 static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
693 struct alien_cache **alc_ptr;
694 size_t memsize = sizeof(void *) * nr_node_ids;
699 alc_ptr = kzalloc_node(memsize, gfp, node);
704 if (i == node || !node_online(i))
706 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
708 for (i--; i >= 0; i--)
717 static void free_alien_cache(struct alien_cache **alc_ptr)
728 static void __drain_alien_cache(struct kmem_cache *cachep,
729 struct array_cache *ac, int node,
730 struct list_head *list)
732 struct kmem_cache_node *n = get_node(cachep, node);
735 spin_lock(&n->list_lock);
737 * Stuff objects into the remote nodes shared array first.
738 * That way we could avoid the overhead of putting the objects
739 * into the free lists and getting them back later.
742 transfer_objects(n->shared, ac, ac->limit);
744 free_block(cachep, ac->entry, ac->avail, node, list);
746 spin_unlock(&n->list_lock);
751 * Called from cache_reap() to regularly drain alien caches round robin.
753 static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
755 int node = __this_cpu_read(slab_reap_node);
758 struct alien_cache *alc = n->alien[node];
759 struct array_cache *ac;
763 if (ac->avail && spin_trylock_irq(&alc->lock)) {
766 __drain_alien_cache(cachep, ac, node, &list);
767 spin_unlock_irq(&alc->lock);
768 slabs_destroy(cachep, &list);
774 static void drain_alien_cache(struct kmem_cache *cachep,
775 struct alien_cache **alien)
778 struct alien_cache *alc;
779 struct array_cache *ac;
782 for_each_online_node(i) {
788 spin_lock_irqsave(&alc->lock, flags);
789 __drain_alien_cache(cachep, ac, i, &list);
790 spin_unlock_irqrestore(&alc->lock, flags);
791 slabs_destroy(cachep, &list);
796 static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
797 int node, int page_node)
799 struct kmem_cache_node *n;
800 struct alien_cache *alien = NULL;
801 struct array_cache *ac;
804 n = get_node(cachep, node);
805 STATS_INC_NODEFREES(cachep);
806 if (n->alien && n->alien[page_node]) {
807 alien = n->alien[page_node];
809 spin_lock(&alien->lock);
810 if (unlikely(ac->avail == ac->limit)) {
811 STATS_INC_ACOVERFLOW(cachep);
812 __drain_alien_cache(cachep, ac, page_node, &list);
814 ac->entry[ac->avail++] = objp;
815 spin_unlock(&alien->lock);
816 slabs_destroy(cachep, &list);
818 n = get_node(cachep, page_node);
819 spin_lock(&n->list_lock);
820 free_block(cachep, &objp, 1, page_node, &list);
821 spin_unlock(&n->list_lock);
822 slabs_destroy(cachep, &list);
827 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
829 int page_node = page_to_nid(virt_to_page(objp));
830 int node = numa_mem_id();
832 * Make sure we are not freeing a object from another node to the array
835 if (likely(node == page_node))
838 return __cache_free_alien(cachep, objp, node, page_node);
842 * Construct gfp mask to allocate from a specific node but do not reclaim or
843 * warn about failures.
845 static inline gfp_t gfp_exact_node(gfp_t flags)
847 return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
852 * Allocates and initializes node for a node on each slab cache, used for
853 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
854 * will be allocated off-node since memory is not yet online for the new node.
855 * When hotplugging memory or a cpu, existing node are not replaced if
858 * Must hold slab_mutex.
860 static int init_cache_node_node(int node)
862 struct kmem_cache *cachep;
863 struct kmem_cache_node *n;
864 const size_t memsize = sizeof(struct kmem_cache_node);
866 list_for_each_entry(cachep, &slab_caches, list) {
868 * Set up the kmem_cache_node for cpu before we can
869 * begin anything. Make sure some other cpu on this
870 * node has not already allocated this
872 n = get_node(cachep, node);
874 n = kmalloc_node(memsize, GFP_KERNEL, node);
877 kmem_cache_node_init(n);
878 n->next_reap = jiffies + REAPTIMEOUT_NODE +
879 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
882 * The kmem_cache_nodes don't come and go as CPUs
883 * come and go. slab_mutex is sufficient
886 cachep->node[node] = n;
889 spin_lock_irq(&n->list_lock);
891 (1 + nr_cpus_node(node)) *
892 cachep->batchcount + cachep->num;
893 spin_unlock_irq(&n->list_lock);
898 static inline int slabs_tofree(struct kmem_cache *cachep,
899 struct kmem_cache_node *n)
901 return (n->free_objects + cachep->num - 1) / cachep->num;
904 static void cpuup_canceled(long cpu)
906 struct kmem_cache *cachep;
907 struct kmem_cache_node *n = NULL;
908 int node = cpu_to_mem(cpu);
909 const struct cpumask *mask = cpumask_of_node(node);
911 list_for_each_entry(cachep, &slab_caches, list) {
912 struct array_cache *nc;
913 struct array_cache *shared;
914 struct alien_cache **alien;
917 n = get_node(cachep, node);
921 spin_lock_irq(&n->list_lock);
923 /* Free limit for this kmem_cache_node */
924 n->free_limit -= cachep->batchcount;
926 /* cpu is dead; no one can alloc from it. */
927 nc = per_cpu_ptr(cachep->cpu_cache, cpu);
929 free_block(cachep, nc->entry, nc->avail, node, &list);
933 if (!cpumask_empty(mask)) {
934 spin_unlock_irq(&n->list_lock);
940 free_block(cachep, shared->entry,
941 shared->avail, node, &list);
948 spin_unlock_irq(&n->list_lock);
952 drain_alien_cache(cachep, alien);
953 free_alien_cache(alien);
957 slabs_destroy(cachep, &list);
960 * In the previous loop, all the objects were freed to
961 * the respective cache's slabs, now we can go ahead and
962 * shrink each nodelist to its limit.
964 list_for_each_entry(cachep, &slab_caches, list) {
965 n = get_node(cachep, node);
968 drain_freelist(cachep, n, slabs_tofree(cachep, n));
972 static int cpuup_prepare(long cpu)
974 struct kmem_cache *cachep;
975 struct kmem_cache_node *n = NULL;
976 int node = cpu_to_mem(cpu);
980 * We need to do this right in the beginning since
981 * alloc_arraycache's are going to use this list.
982 * kmalloc_node allows us to add the slab to the right
983 * kmem_cache_node and not this cpu's kmem_cache_node
985 err = init_cache_node_node(node);
990 * Now we can go ahead with allocating the shared arrays and
993 list_for_each_entry(cachep, &slab_caches, list) {
994 struct array_cache *shared = NULL;
995 struct alien_cache **alien = NULL;
997 if (cachep->shared) {
998 shared = alloc_arraycache(node,
999 cachep->shared * cachep->batchcount,
1000 0xbaadf00d, GFP_KERNEL);
1004 if (use_alien_caches) {
1005 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1011 n = get_node(cachep, node);
1014 spin_lock_irq(&n->list_lock);
1017 * We are serialised from CPU_DEAD or
1018 * CPU_UP_CANCELLED by the cpucontrol lock
1029 spin_unlock_irq(&n->list_lock);
1031 free_alien_cache(alien);
1036 cpuup_canceled(cpu);
1040 static int cpuup_callback(struct notifier_block *nfb,
1041 unsigned long action, void *hcpu)
1043 long cpu = (long)hcpu;
1047 case CPU_UP_PREPARE:
1048 case CPU_UP_PREPARE_FROZEN:
1049 mutex_lock(&slab_mutex);
1050 err = cpuup_prepare(cpu);
1051 mutex_unlock(&slab_mutex);
1054 case CPU_ONLINE_FROZEN:
1055 start_cpu_timer(cpu);
1057 #ifdef CONFIG_HOTPLUG_CPU
1058 case CPU_DOWN_PREPARE:
1059 case CPU_DOWN_PREPARE_FROZEN:
1061 * Shutdown cache reaper. Note that the slab_mutex is
1062 * held so that if cache_reap() is invoked it cannot do
1063 * anything expensive but will only modify reap_work
1064 * and reschedule the timer.
1066 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1067 /* Now the cache_reaper is guaranteed to be not running. */
1068 per_cpu(slab_reap_work, cpu).work.func = NULL;
1070 case CPU_DOWN_FAILED:
1071 case CPU_DOWN_FAILED_FROZEN:
1072 start_cpu_timer(cpu);
1075 case CPU_DEAD_FROZEN:
1077 * Even if all the cpus of a node are down, we don't free the
1078 * kmem_cache_node of any cache. This to avoid a race between
1079 * cpu_down, and a kmalloc allocation from another cpu for
1080 * memory from the node of the cpu going down. The node
1081 * structure is usually allocated from kmem_cache_create() and
1082 * gets destroyed at kmem_cache_destroy().
1086 case CPU_UP_CANCELED:
1087 case CPU_UP_CANCELED_FROZEN:
1088 mutex_lock(&slab_mutex);
1089 cpuup_canceled(cpu);
1090 mutex_unlock(&slab_mutex);
1093 return notifier_from_errno(err);
1096 static struct notifier_block cpucache_notifier = {
1097 &cpuup_callback, NULL, 0
1100 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1102 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1103 * Returns -EBUSY if all objects cannot be drained so that the node is not
1106 * Must hold slab_mutex.
1108 static int __meminit drain_cache_node_node(int node)
1110 struct kmem_cache *cachep;
1113 list_for_each_entry(cachep, &slab_caches, list) {
1114 struct kmem_cache_node *n;
1116 n = get_node(cachep, node);
1120 drain_freelist(cachep, n, slabs_tofree(cachep, n));
1122 if (!list_empty(&n->slabs_full) ||
1123 !list_empty(&n->slabs_partial)) {
1131 static int __meminit slab_memory_callback(struct notifier_block *self,
1132 unsigned long action, void *arg)
1134 struct memory_notify *mnb = arg;
1138 nid = mnb->status_change_nid;
1143 case MEM_GOING_ONLINE:
1144 mutex_lock(&slab_mutex);
1145 ret = init_cache_node_node(nid);
1146 mutex_unlock(&slab_mutex);
1148 case MEM_GOING_OFFLINE:
1149 mutex_lock(&slab_mutex);
1150 ret = drain_cache_node_node(nid);
1151 mutex_unlock(&slab_mutex);
1155 case MEM_CANCEL_ONLINE:
1156 case MEM_CANCEL_OFFLINE:
1160 return notifier_from_errno(ret);
1162 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1165 * swap the static kmem_cache_node with kmalloced memory
1167 static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1170 struct kmem_cache_node *ptr;
1172 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1175 memcpy(ptr, list, sizeof(struct kmem_cache_node));
1177 * Do not assume that spinlocks can be initialized via memcpy:
1179 spin_lock_init(&ptr->list_lock);
1181 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1182 cachep->node[nodeid] = ptr;
1186 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1187 * size of kmem_cache_node.
1189 static void __init set_up_node(struct kmem_cache *cachep, int index)
1193 for_each_online_node(node) {
1194 cachep->node[node] = &init_kmem_cache_node[index + node];
1195 cachep->node[node]->next_reap = jiffies +
1197 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1202 * Initialisation. Called after the page allocator have been initialised and
1203 * before smp_init().
1205 void __init kmem_cache_init(void)
1209 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1210 sizeof(struct rcu_head));
1211 kmem_cache = &kmem_cache_boot;
1213 if (!IS_ENABLED(CONFIG_NUMA) || num_possible_nodes() == 1)
1214 use_alien_caches = 0;
1216 for (i = 0; i < NUM_INIT_LISTS; i++)
1217 kmem_cache_node_init(&init_kmem_cache_node[i]);
1220 * Fragmentation resistance on low memory - only use bigger
1221 * page orders on machines with more than 32MB of memory if
1222 * not overridden on the command line.
1224 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1225 slab_max_order = SLAB_MAX_ORDER_HI;
1227 /* Bootstrap is tricky, because several objects are allocated
1228 * from caches that do not exist yet:
1229 * 1) initialize the kmem_cache cache: it contains the struct
1230 * kmem_cache structures of all caches, except kmem_cache itself:
1231 * kmem_cache is statically allocated.
1232 * Initially an __init data area is used for the head array and the
1233 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1234 * array at the end of the bootstrap.
1235 * 2) Create the first kmalloc cache.
1236 * The struct kmem_cache for the new cache is allocated normally.
1237 * An __init data area is used for the head array.
1238 * 3) Create the remaining kmalloc caches, with minimally sized
1240 * 4) Replace the __init data head arrays for kmem_cache and the first
1241 * kmalloc cache with kmalloc allocated arrays.
1242 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1243 * the other cache's with kmalloc allocated memory.
1244 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1247 /* 1) create the kmem_cache */
1250 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1252 create_boot_cache(kmem_cache, "kmem_cache",
1253 offsetof(struct kmem_cache, node) +
1254 nr_node_ids * sizeof(struct kmem_cache_node *),
1255 SLAB_HWCACHE_ALIGN);
1256 list_add(&kmem_cache->list, &slab_caches);
1257 slab_state = PARTIAL;
1260 * Initialize the caches that provide memory for the kmem_cache_node
1261 * structures first. Without this, further allocations will bug.
1263 kmalloc_caches[INDEX_NODE] = create_kmalloc_cache("kmalloc-node",
1264 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
1265 slab_state = PARTIAL_NODE;
1266 setup_kmalloc_cache_index_table();
1268 slab_early_init = 0;
1270 /* 5) Replace the bootstrap kmem_cache_node */
1274 for_each_online_node(nid) {
1275 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
1277 init_list(kmalloc_caches[INDEX_NODE],
1278 &init_kmem_cache_node[SIZE_NODE + nid], nid);
1282 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1285 void __init kmem_cache_init_late(void)
1287 struct kmem_cache *cachep;
1291 /* 6) resize the head arrays to their final sizes */
1292 mutex_lock(&slab_mutex);
1293 list_for_each_entry(cachep, &slab_caches, list)
1294 if (enable_cpucache(cachep, GFP_NOWAIT))
1296 mutex_unlock(&slab_mutex);
1302 * Register a cpu startup notifier callback that initializes
1303 * cpu_cache_get for all new cpus
1305 register_cpu_notifier(&cpucache_notifier);
1309 * Register a memory hotplug callback that initializes and frees
1312 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1316 * The reap timers are started later, with a module init call: That part
1317 * of the kernel is not yet operational.
1321 static int __init cpucache_init(void)
1326 * Register the timers that return unneeded pages to the page allocator
1328 for_each_online_cpu(cpu)
1329 start_cpu_timer(cpu);
1335 __initcall(cpucache_init);
1337 static noinline void
1338 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1341 struct kmem_cache_node *n;
1343 unsigned long flags;
1345 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1346 DEFAULT_RATELIMIT_BURST);
1348 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1351 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1352 nodeid, gfpflags, &gfpflags);
1353 pr_warn(" cache: %s, object size: %d, order: %d\n",
1354 cachep->name, cachep->size, cachep->gfporder);
1356 for_each_kmem_cache_node(cachep, node, n) {
1357 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1358 unsigned long active_slabs = 0, num_slabs = 0;
1360 spin_lock_irqsave(&n->list_lock, flags);
1361 list_for_each_entry(page, &n->slabs_full, lru) {
1362 active_objs += cachep->num;
1365 list_for_each_entry(page, &n->slabs_partial, lru) {
1366 active_objs += page->active;
1369 list_for_each_entry(page, &n->slabs_free, lru)
1372 free_objects += n->free_objects;
1373 spin_unlock_irqrestore(&n->list_lock, flags);
1375 num_slabs += active_slabs;
1376 num_objs = num_slabs * cachep->num;
1377 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1378 node, active_slabs, num_slabs, active_objs, num_objs,
1385 * Interface to system's page allocator. No need to hold the
1386 * kmem_cache_node ->list_lock.
1388 * If we requested dmaable memory, we will get it. Even if we
1389 * did not request dmaable memory, we might get it, but that
1390 * would be relatively rare and ignorable.
1392 static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1398 flags |= cachep->allocflags;
1399 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1400 flags |= __GFP_RECLAIMABLE;
1402 page = __alloc_pages_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1404 slab_out_of_memory(cachep, flags, nodeid);
1408 if (memcg_charge_slab(page, flags, cachep->gfporder, cachep)) {
1409 __free_pages(page, cachep->gfporder);
1413 nr_pages = (1 << cachep->gfporder);
1414 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1415 add_zone_page_state(page_zone(page),
1416 NR_SLAB_RECLAIMABLE, nr_pages);
1418 add_zone_page_state(page_zone(page),
1419 NR_SLAB_UNRECLAIMABLE, nr_pages);
1421 __SetPageSlab(page);
1422 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1423 if (sk_memalloc_socks() && page_is_pfmemalloc(page))
1424 SetPageSlabPfmemalloc(page);
1426 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1427 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1430 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1432 kmemcheck_mark_unallocated_pages(page, nr_pages);
1439 * Interface to system's page release.
1441 static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
1443 int order = cachep->gfporder;
1444 unsigned long nr_freed = (1 << order);
1446 kmemcheck_free_shadow(page, order);
1448 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1449 sub_zone_page_state(page_zone(page),
1450 NR_SLAB_RECLAIMABLE, nr_freed);
1452 sub_zone_page_state(page_zone(page),
1453 NR_SLAB_UNRECLAIMABLE, nr_freed);
1455 BUG_ON(!PageSlab(page));
1456 __ClearPageSlabPfmemalloc(page);
1457 __ClearPageSlab(page);
1458 page_mapcount_reset(page);
1459 page->mapping = NULL;
1461 if (current->reclaim_state)
1462 current->reclaim_state->reclaimed_slab += nr_freed;
1463 memcg_uncharge_slab(page, order, cachep);
1464 __free_pages(page, order);
1467 static void kmem_rcu_free(struct rcu_head *head)
1469 struct kmem_cache *cachep;
1472 page = container_of(head, struct page, rcu_head);
1473 cachep = page->slab_cache;
1475 kmem_freepages(cachep, page);
1479 static bool is_debug_pagealloc_cache(struct kmem_cache *cachep)
1481 if (debug_pagealloc_enabled() && OFF_SLAB(cachep) &&
1482 (cachep->size % PAGE_SIZE) == 0)
1488 #ifdef CONFIG_DEBUG_PAGEALLOC
1489 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1490 unsigned long caller)
1492 int size = cachep->object_size;
1494 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1496 if (size < 5 * sizeof(unsigned long))
1499 *addr++ = 0x12345678;
1501 *addr++ = smp_processor_id();
1502 size -= 3 * sizeof(unsigned long);
1504 unsigned long *sptr = &caller;
1505 unsigned long svalue;
1507 while (!kstack_end(sptr)) {
1509 if (kernel_text_address(svalue)) {
1511 size -= sizeof(unsigned long);
1512 if (size <= sizeof(unsigned long))
1518 *addr++ = 0x87654321;
1521 static void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1522 int map, unsigned long caller)
1524 if (!is_debug_pagealloc_cache(cachep))
1528 store_stackinfo(cachep, objp, caller);
1530 kernel_map_pages(virt_to_page(objp), cachep->size / PAGE_SIZE, map);
1534 static inline void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1535 int map, unsigned long caller) {}
1539 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1541 int size = cachep->object_size;
1542 addr = &((char *)addr)[obj_offset(cachep)];
1544 memset(addr, val, size);
1545 *(unsigned char *)(addr + size - 1) = POISON_END;
1548 static void dump_line(char *data, int offset, int limit)
1551 unsigned char error = 0;
1554 pr_err("%03x: ", offset);
1555 for (i = 0; i < limit; i++) {
1556 if (data[offset + i] != POISON_FREE) {
1557 error = data[offset + i];
1561 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1562 &data[offset], limit, 1);
1564 if (bad_count == 1) {
1565 error ^= POISON_FREE;
1566 if (!(error & (error - 1))) {
1567 pr_err("Single bit error detected. Probably bad RAM.\n");
1569 pr_err("Run memtest86+ or a similar memory test tool.\n");
1571 pr_err("Run a memory test tool.\n");
1580 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1585 if (cachep->flags & SLAB_RED_ZONE) {
1586 pr_err("Redzone: 0x%llx/0x%llx\n",
1587 *dbg_redzone1(cachep, objp),
1588 *dbg_redzone2(cachep, objp));
1591 if (cachep->flags & SLAB_STORE_USER) {
1592 pr_err("Last user: [<%p>](%pSR)\n",
1593 *dbg_userword(cachep, objp),
1594 *dbg_userword(cachep, objp));
1596 realobj = (char *)objp + obj_offset(cachep);
1597 size = cachep->object_size;
1598 for (i = 0; i < size && lines; i += 16, lines--) {
1601 if (i + limit > size)
1603 dump_line(realobj, i, limit);
1607 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1613 if (is_debug_pagealloc_cache(cachep))
1616 realobj = (char *)objp + obj_offset(cachep);
1617 size = cachep->object_size;
1619 for (i = 0; i < size; i++) {
1620 char exp = POISON_FREE;
1623 if (realobj[i] != exp) {
1628 pr_err("Slab corruption (%s): %s start=%p, len=%d\n",
1629 print_tainted(), cachep->name,
1631 print_objinfo(cachep, objp, 0);
1633 /* Hexdump the affected line */
1636 if (i + limit > size)
1638 dump_line(realobj, i, limit);
1641 /* Limit to 5 lines */
1647 /* Print some data about the neighboring objects, if they
1650 struct page *page = virt_to_head_page(objp);
1653 objnr = obj_to_index(cachep, page, objp);
1655 objp = index_to_obj(cachep, page, objnr - 1);
1656 realobj = (char *)objp + obj_offset(cachep);
1657 pr_err("Prev obj: start=%p, len=%d\n", realobj, size);
1658 print_objinfo(cachep, objp, 2);
1660 if (objnr + 1 < cachep->num) {
1661 objp = index_to_obj(cachep, page, objnr + 1);
1662 realobj = (char *)objp + obj_offset(cachep);
1663 pr_err("Next obj: start=%p, len=%d\n", realobj, size);
1664 print_objinfo(cachep, objp, 2);
1671 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1676 if (OBJFREELIST_SLAB(cachep) && cachep->flags & SLAB_POISON) {
1677 poison_obj(cachep, page->freelist - obj_offset(cachep),
1681 for (i = 0; i < cachep->num; i++) {
1682 void *objp = index_to_obj(cachep, page, i);
1684 if (cachep->flags & SLAB_POISON) {
1685 check_poison_obj(cachep, objp);
1686 slab_kernel_map(cachep, objp, 1, 0);
1688 if (cachep->flags & SLAB_RED_ZONE) {
1689 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1690 slab_error(cachep, "start of a freed object was overwritten");
1691 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1692 slab_error(cachep, "end of a freed object was overwritten");
1697 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1704 * slab_destroy - destroy and release all objects in a slab
1705 * @cachep: cache pointer being destroyed
1706 * @page: page pointer being destroyed
1708 * Destroy all the objs in a slab page, and release the mem back to the system.
1709 * Before calling the slab page must have been unlinked from the cache. The
1710 * kmem_cache_node ->list_lock is not held/needed.
1712 static void slab_destroy(struct kmem_cache *cachep, struct page *page)
1716 freelist = page->freelist;
1717 slab_destroy_debugcheck(cachep, page);
1718 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
1719 call_rcu(&page->rcu_head, kmem_rcu_free);
1721 kmem_freepages(cachep, page);
1724 * From now on, we don't use freelist
1725 * although actual page can be freed in rcu context
1727 if (OFF_SLAB(cachep))
1728 kmem_cache_free(cachep->freelist_cache, freelist);
1731 static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1733 struct page *page, *n;
1735 list_for_each_entry_safe(page, n, list, lru) {
1736 list_del(&page->lru);
1737 slab_destroy(cachep, page);
1742 * calculate_slab_order - calculate size (page order) of slabs
1743 * @cachep: pointer to the cache that is being created
1744 * @size: size of objects to be created in this cache.
1745 * @flags: slab allocation flags
1747 * Also calculates the number of objects per slab.
1749 * This could be made much more intelligent. For now, try to avoid using
1750 * high order pages for slabs. When the gfp() functions are more friendly
1751 * towards high-order requests, this should be changed.
1753 static size_t calculate_slab_order(struct kmem_cache *cachep,
1754 size_t size, unsigned long flags)
1756 size_t left_over = 0;
1759 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
1763 num = cache_estimate(gfporder, size, flags, &remainder);
1767 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1768 if (num > SLAB_OBJ_MAX_NUM)
1771 if (flags & CFLGS_OFF_SLAB) {
1772 struct kmem_cache *freelist_cache;
1773 size_t freelist_size;
1775 freelist_size = num * sizeof(freelist_idx_t);
1776 freelist_cache = kmalloc_slab(freelist_size, 0u);
1777 if (!freelist_cache)
1781 * Needed to avoid possible looping condition
1784 if (OFF_SLAB(freelist_cache))
1787 /* check if off slab has enough benefit */
1788 if (freelist_cache->size > cachep->size / 2)
1792 /* Found something acceptable - save it away */
1794 cachep->gfporder = gfporder;
1795 left_over = remainder;
1798 * A VFS-reclaimable slab tends to have most allocations
1799 * as GFP_NOFS and we really don't want to have to be allocating
1800 * higher-order pages when we are unable to shrink dcache.
1802 if (flags & SLAB_RECLAIM_ACCOUNT)
1806 * Large number of objects is good, but very large slabs are
1807 * currently bad for the gfp()s.
1809 if (gfporder >= slab_max_order)
1813 * Acceptable internal fragmentation?
1815 if (left_over * 8 <= (PAGE_SIZE << gfporder))
1821 static struct array_cache __percpu *alloc_kmem_cache_cpus(
1822 struct kmem_cache *cachep, int entries, int batchcount)
1826 struct array_cache __percpu *cpu_cache;
1828 size = sizeof(void *) * entries + sizeof(struct array_cache);
1829 cpu_cache = __alloc_percpu(size, sizeof(void *));
1834 for_each_possible_cpu(cpu) {
1835 init_arraycache(per_cpu_ptr(cpu_cache, cpu),
1836 entries, batchcount);
1842 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
1844 if (slab_state >= FULL)
1845 return enable_cpucache(cachep, gfp);
1847 cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
1848 if (!cachep->cpu_cache)
1851 if (slab_state == DOWN) {
1852 /* Creation of first cache (kmem_cache). */
1853 set_up_node(kmem_cache, CACHE_CACHE);
1854 } else if (slab_state == PARTIAL) {
1855 /* For kmem_cache_node */
1856 set_up_node(cachep, SIZE_NODE);
1860 for_each_online_node(node) {
1861 cachep->node[node] = kmalloc_node(
1862 sizeof(struct kmem_cache_node), gfp, node);
1863 BUG_ON(!cachep->node[node]);
1864 kmem_cache_node_init(cachep->node[node]);
1868 cachep->node[numa_mem_id()]->next_reap =
1869 jiffies + REAPTIMEOUT_NODE +
1870 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1872 cpu_cache_get(cachep)->avail = 0;
1873 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1874 cpu_cache_get(cachep)->batchcount = 1;
1875 cpu_cache_get(cachep)->touched = 0;
1876 cachep->batchcount = 1;
1877 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1881 unsigned long kmem_cache_flags(unsigned long object_size,
1882 unsigned long flags, const char *name,
1883 void (*ctor)(void *))
1889 __kmem_cache_alias(const char *name, size_t size, size_t align,
1890 unsigned long flags, void (*ctor)(void *))
1892 struct kmem_cache *cachep;
1894 cachep = find_mergeable(size, align, flags, name, ctor);
1899 * Adjust the object sizes so that we clear
1900 * the complete object on kzalloc.
1902 cachep->object_size = max_t(int, cachep->object_size, size);
1907 static bool set_objfreelist_slab_cache(struct kmem_cache *cachep,
1908 size_t size, unsigned long flags)
1914 if (cachep->ctor || flags & SLAB_DESTROY_BY_RCU)
1917 left = calculate_slab_order(cachep, size,
1918 flags | CFLGS_OBJFREELIST_SLAB);
1922 if (cachep->num * sizeof(freelist_idx_t) > cachep->object_size)
1925 cachep->colour = left / cachep->colour_off;
1930 static bool set_off_slab_cache(struct kmem_cache *cachep,
1931 size_t size, unsigned long flags)
1938 * Always use on-slab management when SLAB_NOLEAKTRACE
1939 * to avoid recursive calls into kmemleak.
1941 if (flags & SLAB_NOLEAKTRACE)
1945 * Size is large, assume best to place the slab management obj
1946 * off-slab (should allow better packing of objs).
1948 left = calculate_slab_order(cachep, size, flags | CFLGS_OFF_SLAB);
1953 * If the slab has been placed off-slab, and we have enough space then
1954 * move it on-slab. This is at the expense of any extra colouring.
1956 if (left >= cachep->num * sizeof(freelist_idx_t))
1959 cachep->colour = left / cachep->colour_off;
1964 static bool set_on_slab_cache(struct kmem_cache *cachep,
1965 size_t size, unsigned long flags)
1971 left = calculate_slab_order(cachep, size, flags);
1975 cachep->colour = left / cachep->colour_off;
1981 * __kmem_cache_create - Create a cache.
1982 * @cachep: cache management descriptor
1983 * @flags: SLAB flags
1985 * Returns a ptr to the cache on success, NULL on failure.
1986 * Cannot be called within a int, but can be interrupted.
1987 * The @ctor is run when new pages are allocated by the cache.
1991 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1992 * to catch references to uninitialised memory.
1994 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1995 * for buffer overruns.
1997 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1998 * cacheline. This can be beneficial if you're counting cycles as closely
2002 __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2004 size_t ralign = BYTES_PER_WORD;
2007 size_t size = cachep->size;
2012 * Enable redzoning and last user accounting, except for caches with
2013 * large objects, if the increased size would increase the object size
2014 * above the next power of two: caches with object sizes just above a
2015 * power of two have a significant amount of internal fragmentation.
2017 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2018 2 * sizeof(unsigned long long)))
2019 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2020 if (!(flags & SLAB_DESTROY_BY_RCU))
2021 flags |= SLAB_POISON;
2026 * Check that size is in terms of words. This is needed to avoid
2027 * unaligned accesses for some archs when redzoning is used, and makes
2028 * sure any on-slab bufctl's are also correctly aligned.
2030 if (size & (BYTES_PER_WORD - 1)) {
2031 size += (BYTES_PER_WORD - 1);
2032 size &= ~(BYTES_PER_WORD - 1);
2035 if (flags & SLAB_RED_ZONE) {
2036 ralign = REDZONE_ALIGN;
2037 /* If redzoning, ensure that the second redzone is suitably
2038 * aligned, by adjusting the object size accordingly. */
2039 size += REDZONE_ALIGN - 1;
2040 size &= ~(REDZONE_ALIGN - 1);
2043 /* 3) caller mandated alignment */
2044 if (ralign < cachep->align) {
2045 ralign = cachep->align;
2047 /* disable debug if necessary */
2048 if (ralign > __alignof__(unsigned long long))
2049 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2053 cachep->align = ralign;
2054 cachep->colour_off = cache_line_size();
2055 /* Offset must be a multiple of the alignment. */
2056 if (cachep->colour_off < cachep->align)
2057 cachep->colour_off = cachep->align;
2059 if (slab_is_available())
2067 * Both debugging options require word-alignment which is calculated
2070 if (flags & SLAB_RED_ZONE) {
2071 /* add space for red zone words */
2072 cachep->obj_offset += sizeof(unsigned long long);
2073 size += 2 * sizeof(unsigned long long);
2075 if (flags & SLAB_STORE_USER) {
2076 /* user store requires one word storage behind the end of
2077 * the real object. But if the second red zone needs to be
2078 * aligned to 64 bits, we must allow that much space.
2080 if (flags & SLAB_RED_ZONE)
2081 size += REDZONE_ALIGN;
2083 size += BYTES_PER_WORD;
2087 kasan_cache_create(cachep, &size, &flags);
2089 size = ALIGN(size, cachep->align);
2091 * We should restrict the number of objects in a slab to implement
2092 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2094 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2095 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2099 * To activate debug pagealloc, off-slab management is necessary
2100 * requirement. In early phase of initialization, small sized slab
2101 * doesn't get initialized so it would not be possible. So, we need
2102 * to check size >= 256. It guarantees that all necessary small
2103 * sized slab is initialized in current slab initialization sequence.
2105 if (debug_pagealloc_enabled() && (flags & SLAB_POISON) &&
2106 size >= 256 && cachep->object_size > cache_line_size()) {
2107 if (size < PAGE_SIZE || size % PAGE_SIZE == 0) {
2108 size_t tmp_size = ALIGN(size, PAGE_SIZE);
2110 if (set_off_slab_cache(cachep, tmp_size, flags)) {
2111 flags |= CFLGS_OFF_SLAB;
2112 cachep->obj_offset += tmp_size - size;
2120 if (set_objfreelist_slab_cache(cachep, size, flags)) {
2121 flags |= CFLGS_OBJFREELIST_SLAB;
2125 if (set_off_slab_cache(cachep, size, flags)) {
2126 flags |= CFLGS_OFF_SLAB;
2130 if (set_on_slab_cache(cachep, size, flags))
2136 cachep->freelist_size = cachep->num * sizeof(freelist_idx_t);
2137 cachep->flags = flags;
2138 cachep->allocflags = __GFP_COMP;
2139 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2140 cachep->allocflags |= GFP_DMA;
2141 cachep->size = size;
2142 cachep->reciprocal_buffer_size = reciprocal_value(size);
2146 * If we're going to use the generic kernel_map_pages()
2147 * poisoning, then it's going to smash the contents of
2148 * the redzone and userword anyhow, so switch them off.
2150 if (IS_ENABLED(CONFIG_PAGE_POISONING) &&
2151 (cachep->flags & SLAB_POISON) &&
2152 is_debug_pagealloc_cache(cachep))
2153 cachep->flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2156 if (OFF_SLAB(cachep)) {
2157 cachep->freelist_cache =
2158 kmalloc_slab(cachep->freelist_size, 0u);
2161 err = setup_cpu_cache(cachep, gfp);
2163 __kmem_cache_release(cachep);
2171 static void check_irq_off(void)
2173 BUG_ON(!irqs_disabled());
2176 static void check_irq_on(void)
2178 BUG_ON(irqs_disabled());
2181 static void check_mutex_acquired(void)
2183 BUG_ON(!mutex_is_locked(&slab_mutex));
2186 static void check_spinlock_acquired(struct kmem_cache *cachep)
2190 assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
2194 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2198 assert_spin_locked(&get_node(cachep, node)->list_lock);
2203 #define check_irq_off() do { } while(0)
2204 #define check_irq_on() do { } while(0)
2205 #define check_mutex_acquired() do { } while(0)
2206 #define check_spinlock_acquired(x) do { } while(0)
2207 #define check_spinlock_acquired_node(x, y) do { } while(0)
2210 static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac,
2211 int node, bool free_all, struct list_head *list)
2215 if (!ac || !ac->avail)
2218 tofree = free_all ? ac->avail : (ac->limit + 4) / 5;
2219 if (tofree > ac->avail)
2220 tofree = (ac->avail + 1) / 2;
2222 free_block(cachep, ac->entry, tofree, node, list);
2223 ac->avail -= tofree;
2224 memmove(ac->entry, &(ac->entry[tofree]), sizeof(void *) * ac->avail);
2227 static void do_drain(void *arg)
2229 struct kmem_cache *cachep = arg;
2230 struct array_cache *ac;
2231 int node = numa_mem_id();
2232 struct kmem_cache_node *n;
2236 ac = cpu_cache_get(cachep);
2237 n = get_node(cachep, node);
2238 spin_lock(&n->list_lock);
2239 free_block(cachep, ac->entry, ac->avail, node, &list);
2240 spin_unlock(&n->list_lock);
2241 slabs_destroy(cachep, &list);
2245 static void drain_cpu_caches(struct kmem_cache *cachep)
2247 struct kmem_cache_node *n;
2251 on_each_cpu(do_drain, cachep, 1);
2253 for_each_kmem_cache_node(cachep, node, n)
2255 drain_alien_cache(cachep, n->alien);
2257 for_each_kmem_cache_node(cachep, node, n) {
2258 spin_lock_irq(&n->list_lock);
2259 drain_array_locked(cachep, n->shared, node, true, &list);
2260 spin_unlock_irq(&n->list_lock);
2262 slabs_destroy(cachep, &list);
2267 * Remove slabs from the list of free slabs.
2268 * Specify the number of slabs to drain in tofree.
2270 * Returns the actual number of slabs released.
2272 static int drain_freelist(struct kmem_cache *cache,
2273 struct kmem_cache_node *n, int tofree)
2275 struct list_head *p;
2280 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2282 spin_lock_irq(&n->list_lock);
2283 p = n->slabs_free.prev;
2284 if (p == &n->slabs_free) {
2285 spin_unlock_irq(&n->list_lock);
2289 page = list_entry(p, struct page, lru);
2290 list_del(&page->lru);
2292 * Safe to drop the lock. The slab is no longer linked
2295 n->free_objects -= cache->num;
2296 spin_unlock_irq(&n->list_lock);
2297 slab_destroy(cache, page);
2304 int __kmem_cache_shrink(struct kmem_cache *cachep, bool deactivate)
2308 struct kmem_cache_node *n;
2310 drain_cpu_caches(cachep);
2313 for_each_kmem_cache_node(cachep, node, n) {
2314 drain_freelist(cachep, n, slabs_tofree(cachep, n));
2316 ret += !list_empty(&n->slabs_full) ||
2317 !list_empty(&n->slabs_partial);
2319 return (ret ? 1 : 0);
2322 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2324 return __kmem_cache_shrink(cachep, false);
2327 void __kmem_cache_release(struct kmem_cache *cachep)
2330 struct kmem_cache_node *n;
2332 free_percpu(cachep->cpu_cache);
2334 /* NUMA: free the node structures */
2335 for_each_kmem_cache_node(cachep, i, n) {
2337 free_alien_cache(n->alien);
2339 cachep->node[i] = NULL;
2344 * Get the memory for a slab management obj.
2346 * For a slab cache when the slab descriptor is off-slab, the
2347 * slab descriptor can't come from the same cache which is being created,
2348 * Because if it is the case, that means we defer the creation of
2349 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2350 * And we eventually call down to __kmem_cache_create(), which
2351 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2352 * This is a "chicken-and-egg" problem.
2354 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2355 * which are all initialized during kmem_cache_init().
2357 static void *alloc_slabmgmt(struct kmem_cache *cachep,
2358 struct page *page, int colour_off,
2359 gfp_t local_flags, int nodeid)
2362 void *addr = page_address(page);
2364 page->s_mem = addr + colour_off;
2367 if (OBJFREELIST_SLAB(cachep))
2369 else if (OFF_SLAB(cachep)) {
2370 /* Slab management obj is off-slab. */
2371 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
2372 local_flags, nodeid);
2376 /* We will use last bytes at the slab for freelist */
2377 freelist = addr + (PAGE_SIZE << cachep->gfporder) -
2378 cachep->freelist_size;
2384 static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
2386 return ((freelist_idx_t *)page->freelist)[idx];
2389 static inline void set_free_obj(struct page *page,
2390 unsigned int idx, freelist_idx_t val)
2392 ((freelist_idx_t *)(page->freelist))[idx] = val;
2395 static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page)
2400 for (i = 0; i < cachep->num; i++) {
2401 void *objp = index_to_obj(cachep, page, i);
2403 if (cachep->flags & SLAB_STORE_USER)
2404 *dbg_userword(cachep, objp) = NULL;
2406 if (cachep->flags & SLAB_RED_ZONE) {
2407 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2408 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2411 * Constructors are not allowed to allocate memory from the same
2412 * cache which they are a constructor for. Otherwise, deadlock.
2413 * They must also be threaded.
2415 if (cachep->ctor && !(cachep->flags & SLAB_POISON)) {
2416 kasan_unpoison_object_data(cachep,
2417 objp + obj_offset(cachep));
2418 cachep->ctor(objp + obj_offset(cachep));
2419 kasan_poison_object_data(
2420 cachep, objp + obj_offset(cachep));
2423 if (cachep->flags & SLAB_RED_ZONE) {
2424 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2425 slab_error(cachep, "constructor overwrote the end of an object");
2426 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2427 slab_error(cachep, "constructor overwrote the start of an object");
2429 /* need to poison the objs? */
2430 if (cachep->flags & SLAB_POISON) {
2431 poison_obj(cachep, objp, POISON_FREE);
2432 slab_kernel_map(cachep, objp, 0, 0);
2438 static void cache_init_objs(struct kmem_cache *cachep,
2444 cache_init_objs_debug(cachep, page);
2446 if (OBJFREELIST_SLAB(cachep)) {
2447 page->freelist = index_to_obj(cachep, page, cachep->num - 1) +
2451 for (i = 0; i < cachep->num; i++) {
2452 /* constructor could break poison info */
2453 if (DEBUG == 0 && cachep->ctor) {
2454 objp = index_to_obj(cachep, page, i);
2455 kasan_unpoison_object_data(cachep, objp);
2457 kasan_poison_object_data(cachep, objp);
2460 set_free_obj(page, i, i);
2464 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2466 if (CONFIG_ZONE_DMA_FLAG) {
2467 if (flags & GFP_DMA)
2468 BUG_ON(!(cachep->allocflags & GFP_DMA));
2470 BUG_ON(cachep->allocflags & GFP_DMA);
2474 static void *slab_get_obj(struct kmem_cache *cachep, struct page *page)
2478 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
2482 if (cachep->flags & SLAB_STORE_USER)
2483 set_store_user_dirty(cachep);
2489 static void slab_put_obj(struct kmem_cache *cachep,
2490 struct page *page, void *objp)
2492 unsigned int objnr = obj_to_index(cachep, page, objp);
2496 /* Verify double free bug */
2497 for (i = page->active; i < cachep->num; i++) {
2498 if (get_free_obj(page, i) == objnr) {
2499 pr_err("slab: double free detected in cache '%s', objp %p\n",
2500 cachep->name, objp);
2506 if (!page->freelist)
2507 page->freelist = objp + obj_offset(cachep);
2509 set_free_obj(page, page->active, objnr);
2513 * Map pages beginning at addr to the given cache and slab. This is required
2514 * for the slab allocator to be able to lookup the cache and slab of a
2515 * virtual address for kfree, ksize, and slab debugging.
2517 static void slab_map_pages(struct kmem_cache *cache, struct page *page,
2520 page->slab_cache = cache;
2521 page->freelist = freelist;
2525 * Grow (by 1) the number of slabs within a cache. This is called by
2526 * kmem_cache_alloc() when there are no active objs left in a cache.
2528 static int cache_grow(struct kmem_cache *cachep,
2529 gfp_t flags, int nodeid, struct page *page)
2534 struct kmem_cache_node *n;
2537 * Be lazy and only check for valid flags here, keeping it out of the
2538 * critical path in kmem_cache_alloc().
2540 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
2541 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
2544 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2546 /* Take the node list lock to change the colour_next on this node */
2548 n = get_node(cachep, nodeid);
2549 spin_lock(&n->list_lock);
2551 /* Get colour for the slab, and cal the next value. */
2552 offset = n->colour_next;
2554 if (n->colour_next >= cachep->colour)
2556 spin_unlock(&n->list_lock);
2558 offset *= cachep->colour_off;
2560 if (gfpflags_allow_blocking(local_flags))
2564 * The test for missing atomic flag is performed here, rather than
2565 * the more obvious place, simply to reduce the critical path length
2566 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2567 * will eventually be caught here (where it matters).
2569 kmem_flagcheck(cachep, flags);
2572 * Get mem for the objs. Attempt to allocate a physical page from
2576 page = kmem_getpages(cachep, local_flags, nodeid);
2580 /* Get slab management. */
2581 freelist = alloc_slabmgmt(cachep, page, offset,
2582 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2583 if (OFF_SLAB(cachep) && !freelist)
2586 slab_map_pages(cachep, page, freelist);
2588 kasan_poison_slab(page);
2589 cache_init_objs(cachep, page);
2591 if (gfpflags_allow_blocking(local_flags))
2592 local_irq_disable();
2594 spin_lock(&n->list_lock);
2596 /* Make slab active. */
2597 list_add_tail(&page->lru, &(n->slabs_free));
2598 STATS_INC_GROWN(cachep);
2599 n->free_objects += cachep->num;
2600 spin_unlock(&n->list_lock);
2603 kmem_freepages(cachep, page);
2605 if (gfpflags_allow_blocking(local_flags))
2606 local_irq_disable();
2613 * Perform extra freeing checks:
2614 * - detect bad pointers.
2615 * - POISON/RED_ZONE checking
2617 static void kfree_debugcheck(const void *objp)
2619 if (!virt_addr_valid(objp)) {
2620 pr_err("kfree_debugcheck: out of range ptr %lxh\n",
2621 (unsigned long)objp);
2626 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2628 unsigned long long redzone1, redzone2;
2630 redzone1 = *dbg_redzone1(cache, obj);
2631 redzone2 = *dbg_redzone2(cache, obj);
2636 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2639 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2640 slab_error(cache, "double free detected");
2642 slab_error(cache, "memory outside object was overwritten");
2644 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2645 obj, redzone1, redzone2);
2648 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2649 unsigned long caller)
2654 BUG_ON(virt_to_cache(objp) != cachep);
2656 objp -= obj_offset(cachep);
2657 kfree_debugcheck(objp);
2658 page = virt_to_head_page(objp);
2660 if (cachep->flags & SLAB_RED_ZONE) {
2661 verify_redzone_free(cachep, objp);
2662 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2663 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2665 if (cachep->flags & SLAB_STORE_USER) {
2666 set_store_user_dirty(cachep);
2667 *dbg_userword(cachep, objp) = (void *)caller;
2670 objnr = obj_to_index(cachep, page, objp);
2672 BUG_ON(objnr >= cachep->num);
2673 BUG_ON(objp != index_to_obj(cachep, page, objnr));
2675 if (cachep->flags & SLAB_POISON) {
2676 poison_obj(cachep, objp, POISON_FREE);
2677 slab_kernel_map(cachep, objp, 0, caller);
2683 #define kfree_debugcheck(x) do { } while(0)
2684 #define cache_free_debugcheck(x,objp,z) (objp)
2687 static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
2695 objp = next - obj_offset(cachep);
2696 next = *(void **)next;
2697 poison_obj(cachep, objp, POISON_FREE);
2702 static inline void fixup_slab_list(struct kmem_cache *cachep,
2703 struct kmem_cache_node *n, struct page *page,
2706 /* move slabp to correct slabp list: */
2707 list_del(&page->lru);
2708 if (page->active == cachep->num) {
2709 list_add(&page->lru, &n->slabs_full);
2710 if (OBJFREELIST_SLAB(cachep)) {
2712 /* Poisoning will be done without holding the lock */
2713 if (cachep->flags & SLAB_POISON) {
2714 void **objp = page->freelist;
2720 page->freelist = NULL;
2723 list_add(&page->lru, &n->slabs_partial);
2726 /* Try to find non-pfmemalloc slab if needed */
2727 static noinline struct page *get_valid_first_slab(struct kmem_cache_node *n,
2728 struct page *page, bool pfmemalloc)
2736 if (!PageSlabPfmemalloc(page))
2739 /* No need to keep pfmemalloc slab if we have enough free objects */
2740 if (n->free_objects > n->free_limit) {
2741 ClearPageSlabPfmemalloc(page);
2745 /* Move pfmemalloc slab to the end of list to speed up next search */
2746 list_del(&page->lru);
2748 list_add_tail(&page->lru, &n->slabs_free);
2750 list_add_tail(&page->lru, &n->slabs_partial);
2752 list_for_each_entry(page, &n->slabs_partial, lru) {
2753 if (!PageSlabPfmemalloc(page))
2757 list_for_each_entry(page, &n->slabs_free, lru) {
2758 if (!PageSlabPfmemalloc(page))
2765 static struct page *get_first_slab(struct kmem_cache_node *n, bool pfmemalloc)
2769 page = list_first_entry_or_null(&n->slabs_partial,
2772 n->free_touched = 1;
2773 page = list_first_entry_or_null(&n->slabs_free,
2777 if (sk_memalloc_socks())
2778 return get_valid_first_slab(n, page, pfmemalloc);
2783 static noinline void *cache_alloc_pfmemalloc(struct kmem_cache *cachep,
2784 struct kmem_cache_node *n, gfp_t flags)
2790 if (!gfp_pfmemalloc_allowed(flags))
2793 spin_lock(&n->list_lock);
2794 page = get_first_slab(n, true);
2796 spin_unlock(&n->list_lock);
2800 obj = slab_get_obj(cachep, page);
2803 fixup_slab_list(cachep, n, page, &list);
2805 spin_unlock(&n->list_lock);
2806 fixup_objfreelist_debug(cachep, &list);
2811 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2814 struct kmem_cache_node *n;
2815 struct array_cache *ac;
2820 node = numa_mem_id();
2823 ac = cpu_cache_get(cachep);
2824 batchcount = ac->batchcount;
2825 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2827 * If there was little recent activity on this cache, then
2828 * perform only a partial refill. Otherwise we could generate
2831 batchcount = BATCHREFILL_LIMIT;
2833 n = get_node(cachep, node);
2835 BUG_ON(ac->avail > 0 || !n);
2836 spin_lock(&n->list_lock);
2838 /* See if we can refill from the shared array */
2839 if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2840 n->shared->touched = 1;
2844 while (batchcount > 0) {
2846 /* Get slab alloc is to come from. */
2847 page = get_first_slab(n, false);
2851 check_spinlock_acquired(cachep);
2854 * The slab was either on partial or free list so
2855 * there must be at least one object available for
2858 BUG_ON(page->active >= cachep->num);
2860 while (page->active < cachep->num && batchcount--) {
2861 STATS_INC_ALLOCED(cachep);
2862 STATS_INC_ACTIVE(cachep);
2863 STATS_SET_HIGH(cachep);
2865 ac->entry[ac->avail++] = slab_get_obj(cachep, page);
2868 fixup_slab_list(cachep, n, page, &list);
2872 n->free_objects -= ac->avail;
2874 spin_unlock(&n->list_lock);
2875 fixup_objfreelist_debug(cachep, &list);
2877 if (unlikely(!ac->avail)) {
2880 /* Check if we can use obj in pfmemalloc slab */
2881 if (sk_memalloc_socks()) {
2882 void *obj = cache_alloc_pfmemalloc(cachep, n, flags);
2888 x = cache_grow(cachep, gfp_exact_node(flags), node, NULL);
2890 /* cache_grow can reenable interrupts, then ac could change. */
2891 ac = cpu_cache_get(cachep);
2892 node = numa_mem_id();
2894 /* no objects in sight? abort */
2895 if (!x && ac->avail == 0)
2898 if (!ac->avail) /* objects refilled by interrupt? */
2903 return ac->entry[--ac->avail];
2906 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2909 might_sleep_if(gfpflags_allow_blocking(flags));
2911 kmem_flagcheck(cachep, flags);
2916 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2917 gfp_t flags, void *objp, unsigned long caller)
2921 if (cachep->flags & SLAB_POISON) {
2922 check_poison_obj(cachep, objp);
2923 slab_kernel_map(cachep, objp, 1, 0);
2924 poison_obj(cachep, objp, POISON_INUSE);
2926 if (cachep->flags & SLAB_STORE_USER)
2927 *dbg_userword(cachep, objp) = (void *)caller;
2929 if (cachep->flags & SLAB_RED_ZONE) {
2930 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2931 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2932 slab_error(cachep, "double free, or memory outside object was overwritten");
2933 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2934 objp, *dbg_redzone1(cachep, objp),
2935 *dbg_redzone2(cachep, objp));
2937 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2938 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2941 objp += obj_offset(cachep);
2942 if (cachep->ctor && cachep->flags & SLAB_POISON)
2944 if (ARCH_SLAB_MINALIGN &&
2945 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
2946 pr_err("0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
2947 objp, (int)ARCH_SLAB_MINALIGN);
2952 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2955 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
2958 struct array_cache *ac;
2962 ac = cpu_cache_get(cachep);
2963 if (likely(ac->avail)) {
2965 objp = ac->entry[--ac->avail];
2967 STATS_INC_ALLOCHIT(cachep);
2971 STATS_INC_ALLOCMISS(cachep);
2972 objp = cache_alloc_refill(cachep, flags);
2974 * the 'ac' may be updated by cache_alloc_refill(),
2975 * and kmemleak_erase() requires its correct value.
2977 ac = cpu_cache_get(cachep);
2981 * To avoid a false negative, if an object that is in one of the
2982 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
2983 * treat the array pointers as a reference to the object.
2986 kmemleak_erase(&ac->entry[ac->avail]);
2992 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
2994 * If we are in_interrupt, then process context, including cpusets and
2995 * mempolicy, may not apply and should not be used for allocation policy.
2997 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
2999 int nid_alloc, nid_here;
3001 if (in_interrupt() || (flags & __GFP_THISNODE))
3003 nid_alloc = nid_here = numa_mem_id();
3004 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3005 nid_alloc = cpuset_slab_spread_node();
3006 else if (current->mempolicy)
3007 nid_alloc = mempolicy_slab_node();
3008 if (nid_alloc != nid_here)
3009 return ____cache_alloc_node(cachep, flags, nid_alloc);
3014 * Fallback function if there was no memory available and no objects on a
3015 * certain node and fall back is permitted. First we scan all the
3016 * available node for available objects. If that fails then we
3017 * perform an allocation without specifying a node. This allows the page
3018 * allocator to do its reclaim / fallback magic. We then insert the
3019 * slab into the proper nodelist and then allocate from it.
3021 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3023 struct zonelist *zonelist;
3027 enum zone_type high_zoneidx = gfp_zone(flags);
3030 unsigned int cpuset_mems_cookie;
3032 if (flags & __GFP_THISNODE)
3035 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3038 cpuset_mems_cookie = read_mems_allowed_begin();
3039 zonelist = node_zonelist(mempolicy_slab_node(), flags);
3043 * Look through allowed nodes for objects available
3044 * from existing per node queues.
3046 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3047 nid = zone_to_nid(zone);
3049 if (cpuset_zone_allowed(zone, flags) &&
3050 get_node(cache, nid) &&
3051 get_node(cache, nid)->free_objects) {
3052 obj = ____cache_alloc_node(cache,
3053 gfp_exact_node(flags), nid);
3061 * This allocation will be performed within the constraints
3062 * of the current cpuset / memory policy requirements.
3063 * We may trigger various forms of reclaim on the allowed
3064 * set and go into memory reserves if necessary.
3068 if (gfpflags_allow_blocking(local_flags))
3070 kmem_flagcheck(cache, flags);
3071 page = kmem_getpages(cache, local_flags, numa_mem_id());
3072 if (gfpflags_allow_blocking(local_flags))
3073 local_irq_disable();
3076 * Insert into the appropriate per node queues
3078 nid = page_to_nid(page);
3079 if (cache_grow(cache, flags, nid, page)) {
3080 obj = ____cache_alloc_node(cache,
3081 gfp_exact_node(flags), nid);
3084 * Another processor may allocate the
3085 * objects in the slab since we are
3086 * not holding any locks.
3090 /* cache_grow already freed obj */
3096 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
3102 * A interface to enable slab creation on nodeid
3104 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3108 struct kmem_cache_node *n;
3113 VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
3114 n = get_node(cachep, nodeid);
3119 spin_lock(&n->list_lock);
3120 page = get_first_slab(n, false);
3124 check_spinlock_acquired_node(cachep, nodeid);
3126 STATS_INC_NODEALLOCS(cachep);
3127 STATS_INC_ACTIVE(cachep);
3128 STATS_SET_HIGH(cachep);
3130 BUG_ON(page->active == cachep->num);
3132 obj = slab_get_obj(cachep, page);
3135 fixup_slab_list(cachep, n, page, &list);
3137 spin_unlock(&n->list_lock);
3138 fixup_objfreelist_debug(cachep, &list);
3142 spin_unlock(&n->list_lock);
3143 x = cache_grow(cachep, gfp_exact_node(flags), nodeid, NULL);
3147 return fallback_alloc(cachep, flags);
3153 static __always_inline void *
3154 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3155 unsigned long caller)
3157 unsigned long save_flags;
3159 int slab_node = numa_mem_id();
3161 flags &= gfp_allowed_mask;
3162 cachep = slab_pre_alloc_hook(cachep, flags);
3163 if (unlikely(!cachep))
3166 cache_alloc_debugcheck_before(cachep, flags);
3167 local_irq_save(save_flags);
3169 if (nodeid == NUMA_NO_NODE)
3172 if (unlikely(!get_node(cachep, nodeid))) {
3173 /* Node not bootstrapped yet */
3174 ptr = fallback_alloc(cachep, flags);
3178 if (nodeid == slab_node) {
3180 * Use the locally cached objects if possible.
3181 * However ____cache_alloc does not allow fallback
3182 * to other nodes. It may fail while we still have
3183 * objects on other nodes available.
3185 ptr = ____cache_alloc(cachep, flags);
3189 /* ___cache_alloc_node can fall back to other nodes */
3190 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3192 local_irq_restore(save_flags);
3193 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3195 if (unlikely(flags & __GFP_ZERO) && ptr)
3196 memset(ptr, 0, cachep->object_size);
3198 slab_post_alloc_hook(cachep, flags, 1, &ptr);
3202 static __always_inline void *
3203 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3207 if (current->mempolicy || cpuset_do_slab_mem_spread()) {
3208 objp = alternate_node_alloc(cache, flags);
3212 objp = ____cache_alloc(cache, flags);
3215 * We may just have run out of memory on the local node.
3216 * ____cache_alloc_node() knows how to locate memory on other nodes
3219 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3226 static __always_inline void *
3227 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3229 return ____cache_alloc(cachep, flags);
3232 #endif /* CONFIG_NUMA */
3234 static __always_inline void *
3235 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3237 unsigned long save_flags;
3240 flags &= gfp_allowed_mask;
3241 cachep = slab_pre_alloc_hook(cachep, flags);
3242 if (unlikely(!cachep))
3245 cache_alloc_debugcheck_before(cachep, flags);
3246 local_irq_save(save_flags);
3247 objp = __do_cache_alloc(cachep, flags);
3248 local_irq_restore(save_flags);
3249 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3252 if (unlikely(flags & __GFP_ZERO) && objp)
3253 memset(objp, 0, cachep->object_size);
3255 slab_post_alloc_hook(cachep, flags, 1, &objp);
3260 * Caller needs to acquire correct kmem_cache_node's list_lock
3261 * @list: List of detached free slabs should be freed by caller
3263 static void free_block(struct kmem_cache *cachep, void **objpp,
3264 int nr_objects, int node, struct list_head *list)
3267 struct kmem_cache_node *n = get_node(cachep, node);
3269 for (i = 0; i < nr_objects; i++) {
3275 page = virt_to_head_page(objp);
3276 list_del(&page->lru);
3277 check_spinlock_acquired_node(cachep, node);
3278 slab_put_obj(cachep, page, objp);
3279 STATS_DEC_ACTIVE(cachep);
3282 /* fixup slab chains */
3283 if (page->active == 0) {
3284 if (n->free_objects > n->free_limit) {
3285 n->free_objects -= cachep->num;
3286 list_add_tail(&page->lru, list);
3288 list_add(&page->lru, &n->slabs_free);
3291 /* Unconditionally move a slab to the end of the
3292 * partial list on free - maximum time for the
3293 * other objects to be freed, too.
3295 list_add_tail(&page->lru, &n->slabs_partial);
3300 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3303 struct kmem_cache_node *n;
3304 int node = numa_mem_id();
3307 batchcount = ac->batchcount;
3310 n = get_node(cachep, node);
3311 spin_lock(&n->list_lock);
3313 struct array_cache *shared_array = n->shared;
3314 int max = shared_array->limit - shared_array->avail;
3316 if (batchcount > max)
3318 memcpy(&(shared_array->entry[shared_array->avail]),
3319 ac->entry, sizeof(void *) * batchcount);
3320 shared_array->avail += batchcount;
3325 free_block(cachep, ac->entry, batchcount, node, &list);
3332 list_for_each_entry(page, &n->slabs_free, lru) {
3333 BUG_ON(page->active);
3337 STATS_SET_FREEABLE(cachep, i);
3340 spin_unlock(&n->list_lock);
3341 slabs_destroy(cachep, &list);
3342 ac->avail -= batchcount;
3343 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3347 * Release an obj back to its cache. If the obj has a constructed state, it must
3348 * be in this state _before_ it is released. Called with disabled ints.
3350 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3351 unsigned long caller)
3353 struct array_cache *ac = cpu_cache_get(cachep);
3355 kasan_slab_free(cachep, objp);
3358 kmemleak_free_recursive(objp, cachep->flags);
3359 objp = cache_free_debugcheck(cachep, objp, caller);
3361 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3364 * Skip calling cache_free_alien() when the platform is not numa.
3365 * This will avoid cache misses that happen while accessing slabp (which
3366 * is per page memory reference) to get nodeid. Instead use a global
3367 * variable to skip the call, which is mostly likely to be present in
3370 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3373 if (ac->avail < ac->limit) {
3374 STATS_INC_FREEHIT(cachep);
3376 STATS_INC_FREEMISS(cachep);
3377 cache_flusharray(cachep, ac);
3380 if (sk_memalloc_socks()) {
3381 struct page *page = virt_to_head_page(objp);
3383 if (unlikely(PageSlabPfmemalloc(page))) {
3384 cache_free_pfmemalloc(cachep, page, objp);
3389 ac->entry[ac->avail++] = objp;
3393 * kmem_cache_alloc - Allocate an object
3394 * @cachep: The cache to allocate from.
3395 * @flags: See kmalloc().
3397 * Allocate an object from this cache. The flags are only relevant
3398 * if the cache has no available objects.
3400 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3402 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3404 kasan_slab_alloc(cachep, ret, flags);
3405 trace_kmem_cache_alloc(_RET_IP_, ret,
3406 cachep->object_size, cachep->size, flags);
3410 EXPORT_SYMBOL(kmem_cache_alloc);
3412 static __always_inline void
3413 cache_alloc_debugcheck_after_bulk(struct kmem_cache *s, gfp_t flags,
3414 size_t size, void **p, unsigned long caller)
3418 for (i = 0; i < size; i++)
3419 p[i] = cache_alloc_debugcheck_after(s, flags, p[i], caller);
3422 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3427 s = slab_pre_alloc_hook(s, flags);
3431 cache_alloc_debugcheck_before(s, flags);
3433 local_irq_disable();
3434 for (i = 0; i < size; i++) {
3435 void *objp = __do_cache_alloc(s, flags);
3437 if (unlikely(!objp))
3443 cache_alloc_debugcheck_after_bulk(s, flags, size, p, _RET_IP_);
3445 /* Clear memory outside IRQ disabled section */
3446 if (unlikely(flags & __GFP_ZERO))
3447 for (i = 0; i < size; i++)
3448 memset(p[i], 0, s->object_size);
3450 slab_post_alloc_hook(s, flags, size, p);
3451 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3455 cache_alloc_debugcheck_after_bulk(s, flags, i, p, _RET_IP_);
3456 slab_post_alloc_hook(s, flags, i, p);
3457 __kmem_cache_free_bulk(s, i, p);
3460 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3462 #ifdef CONFIG_TRACING
3464 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3468 ret = slab_alloc(cachep, flags, _RET_IP_);
3470 kasan_kmalloc(cachep, ret, size, flags);
3471 trace_kmalloc(_RET_IP_, ret,
3472 size, cachep->size, flags);
3475 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3480 * kmem_cache_alloc_node - Allocate an object on the specified node
3481 * @cachep: The cache to allocate from.
3482 * @flags: See kmalloc().
3483 * @nodeid: node number of the target node.
3485 * Identical to kmem_cache_alloc but it will allocate memory on the given
3486 * node, which can improve the performance for cpu bound structures.
3488 * Fallback to other node is possible if __GFP_THISNODE is not set.
3490 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3492 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3494 kasan_slab_alloc(cachep, ret, flags);
3495 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3496 cachep->object_size, cachep->size,
3501 EXPORT_SYMBOL(kmem_cache_alloc_node);
3503 #ifdef CONFIG_TRACING
3504 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3511 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3513 kasan_kmalloc(cachep, ret, size, flags);
3514 trace_kmalloc_node(_RET_IP_, ret,
3519 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3522 static __always_inline void *
3523 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3525 struct kmem_cache *cachep;
3528 cachep = kmalloc_slab(size, flags);
3529 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3531 ret = kmem_cache_alloc_node_trace(cachep, flags, node, size);
3532 kasan_kmalloc(cachep, ret, size, flags);
3537 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3539 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3541 EXPORT_SYMBOL(__kmalloc_node);
3543 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3544 int node, unsigned long caller)
3546 return __do_kmalloc_node(size, flags, node, caller);
3548 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3549 #endif /* CONFIG_NUMA */
3552 * __do_kmalloc - allocate memory
3553 * @size: how many bytes of memory are required.
3554 * @flags: the type of memory to allocate (see kmalloc).
3555 * @caller: function caller for debug tracking of the caller
3557 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3558 unsigned long caller)
3560 struct kmem_cache *cachep;
3563 cachep = kmalloc_slab(size, flags);
3564 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3566 ret = slab_alloc(cachep, flags, caller);
3568 kasan_kmalloc(cachep, ret, size, flags);
3569 trace_kmalloc(caller, ret,
3570 size, cachep->size, flags);
3575 void *__kmalloc(size_t size, gfp_t flags)
3577 return __do_kmalloc(size, flags, _RET_IP_);
3579 EXPORT_SYMBOL(__kmalloc);
3581 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3583 return __do_kmalloc(size, flags, caller);
3585 EXPORT_SYMBOL(__kmalloc_track_caller);
3588 * kmem_cache_free - Deallocate an object
3589 * @cachep: The cache the allocation was from.
3590 * @objp: The previously allocated object.
3592 * Free an object which was previously allocated from this
3595 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3597 unsigned long flags;
3598 cachep = cache_from_obj(cachep, objp);
3602 local_irq_save(flags);
3603 debug_check_no_locks_freed(objp, cachep->object_size);
3604 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3605 debug_check_no_obj_freed(objp, cachep->object_size);
3606 __cache_free(cachep, objp, _RET_IP_);
3607 local_irq_restore(flags);
3609 trace_kmem_cache_free(_RET_IP_, objp);
3611 EXPORT_SYMBOL(kmem_cache_free);
3613 void kmem_cache_free_bulk(struct kmem_cache *orig_s, size_t size, void **p)
3615 struct kmem_cache *s;
3618 local_irq_disable();
3619 for (i = 0; i < size; i++) {
3622 if (!orig_s) /* called via kfree_bulk */
3623 s = virt_to_cache(objp);
3625 s = cache_from_obj(orig_s, objp);
3627 debug_check_no_locks_freed(objp, s->object_size);
3628 if (!(s->flags & SLAB_DEBUG_OBJECTS))
3629 debug_check_no_obj_freed(objp, s->object_size);
3631 __cache_free(s, objp, _RET_IP_);
3635 /* FIXME: add tracing */
3637 EXPORT_SYMBOL(kmem_cache_free_bulk);
3640 * kfree - free previously allocated memory
3641 * @objp: pointer returned by kmalloc.
3643 * If @objp is NULL, no operation is performed.
3645 * Don't free memory not originally allocated by kmalloc()
3646 * or you will run into trouble.
3648 void kfree(const void *objp)
3650 struct kmem_cache *c;
3651 unsigned long flags;
3653 trace_kfree(_RET_IP_, objp);
3655 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3657 local_irq_save(flags);
3658 kfree_debugcheck(objp);
3659 c = virt_to_cache(objp);
3660 debug_check_no_locks_freed(objp, c->object_size);
3662 debug_check_no_obj_freed(objp, c->object_size);
3663 __cache_free(c, (void *)objp, _RET_IP_);
3664 local_irq_restore(flags);
3666 EXPORT_SYMBOL(kfree);
3669 * This initializes kmem_cache_node or resizes various caches for all nodes.
3671 static int alloc_kmem_cache_node(struct kmem_cache *cachep, gfp_t gfp)
3674 struct kmem_cache_node *n;
3675 struct array_cache *new_shared;
3676 struct alien_cache **new_alien = NULL;
3678 for_each_online_node(node) {
3680 if (use_alien_caches) {
3681 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3687 if (cachep->shared) {
3688 new_shared = alloc_arraycache(node,
3689 cachep->shared*cachep->batchcount,
3692 free_alien_cache(new_alien);
3697 n = get_node(cachep, node);
3699 struct array_cache *shared = n->shared;
3702 spin_lock_irq(&n->list_lock);
3705 free_block(cachep, shared->entry,
3706 shared->avail, node, &list);
3708 n->shared = new_shared;
3710 n->alien = new_alien;
3713 n->free_limit = (1 + nr_cpus_node(node)) *
3714 cachep->batchcount + cachep->num;
3715 spin_unlock_irq(&n->list_lock);
3716 slabs_destroy(cachep, &list);
3718 free_alien_cache(new_alien);
3721 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3723 free_alien_cache(new_alien);
3728 kmem_cache_node_init(n);
3729 n->next_reap = jiffies + REAPTIMEOUT_NODE +
3730 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
3731 n->shared = new_shared;
3732 n->alien = new_alien;
3733 n->free_limit = (1 + nr_cpus_node(node)) *
3734 cachep->batchcount + cachep->num;
3735 cachep->node[node] = n;
3740 if (!cachep->list.next) {
3741 /* Cache is not active yet. Roll back what we did */
3744 n = get_node(cachep, node);
3747 free_alien_cache(n->alien);
3749 cachep->node[node] = NULL;
3757 /* Always called with the slab_mutex held */
3758 static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
3759 int batchcount, int shared, gfp_t gfp)
3761 struct array_cache __percpu *cpu_cache, *prev;
3764 cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3768 prev = cachep->cpu_cache;
3769 cachep->cpu_cache = cpu_cache;
3770 kick_all_cpus_sync();
3773 cachep->batchcount = batchcount;
3774 cachep->limit = limit;
3775 cachep->shared = shared;
3780 for_each_online_cpu(cpu) {
3783 struct kmem_cache_node *n;
3784 struct array_cache *ac = per_cpu_ptr(prev, cpu);
3786 node = cpu_to_mem(cpu);
3787 n = get_node(cachep, node);
3788 spin_lock_irq(&n->list_lock);
3789 free_block(cachep, ac->entry, ac->avail, node, &list);
3790 spin_unlock_irq(&n->list_lock);
3791 slabs_destroy(cachep, &list);
3796 return alloc_kmem_cache_node(cachep, gfp);
3799 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3800 int batchcount, int shared, gfp_t gfp)
3803 struct kmem_cache *c;
3805 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3807 if (slab_state < FULL)
3810 if ((ret < 0) || !is_root_cache(cachep))
3813 lockdep_assert_held(&slab_mutex);
3814 for_each_memcg_cache(c, cachep) {
3815 /* return value determined by the root cache only */
3816 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3822 /* Called with slab_mutex held always */
3823 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3830 if (!is_root_cache(cachep)) {
3831 struct kmem_cache *root = memcg_root_cache(cachep);
3832 limit = root->limit;
3833 shared = root->shared;
3834 batchcount = root->batchcount;
3837 if (limit && shared && batchcount)
3840 * The head array serves three purposes:
3841 * - create a LIFO ordering, i.e. return objects that are cache-warm
3842 * - reduce the number of spinlock operations.
3843 * - reduce the number of linked list operations on the slab and
3844 * bufctl chains: array operations are cheaper.
3845 * The numbers are guessed, we should auto-tune as described by
3848 if (cachep->size > 131072)
3850 else if (cachep->size > PAGE_SIZE)
3852 else if (cachep->size > 1024)
3854 else if (cachep->size > 256)
3860 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3861 * allocation behaviour: Most allocs on one cpu, most free operations
3862 * on another cpu. For these cases, an efficient object passing between
3863 * cpus is necessary. This is provided by a shared array. The array
3864 * replaces Bonwick's magazine layer.
3865 * On uniprocessor, it's functionally equivalent (but less efficient)
3866 * to a larger limit. Thus disabled by default.
3869 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
3874 * With debugging enabled, large batchcount lead to excessively long
3875 * periods with disabled local interrupts. Limit the batchcount
3880 batchcount = (limit + 1) / 2;
3882 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3884 pr_err("enable_cpucache failed for %s, error %d\n",
3885 cachep->name, -err);
3890 * Drain an array if it contains any elements taking the node lock only if
3891 * necessary. Note that the node listlock also protects the array_cache
3892 * if drain_array() is used on the shared array.
3894 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
3895 struct array_cache *ac, int node)
3899 /* ac from n->shared can be freed if we don't hold the slab_mutex. */
3900 check_mutex_acquired();
3902 if (!ac || !ac->avail)
3910 spin_lock_irq(&n->list_lock);
3911 drain_array_locked(cachep, ac, node, false, &list);
3912 spin_unlock_irq(&n->list_lock);
3914 slabs_destroy(cachep, &list);
3918 * cache_reap - Reclaim memory from caches.
3919 * @w: work descriptor
3921 * Called from workqueue/eventd every few seconds.
3923 * - clear the per-cpu caches for this CPU.
3924 * - return freeable pages to the main free memory pool.
3926 * If we cannot acquire the cache chain mutex then just give up - we'll try
3927 * again on the next iteration.
3929 static void cache_reap(struct work_struct *w)
3931 struct kmem_cache *searchp;
3932 struct kmem_cache_node *n;
3933 int node = numa_mem_id();
3934 struct delayed_work *work = to_delayed_work(w);
3936 if (!mutex_trylock(&slab_mutex))
3937 /* Give up. Setup the next iteration. */
3940 list_for_each_entry(searchp, &slab_caches, list) {
3944 * We only take the node lock if absolutely necessary and we
3945 * have established with reasonable certainty that
3946 * we can do some work if the lock was obtained.
3948 n = get_node(searchp, node);
3950 reap_alien(searchp, n);
3952 drain_array(searchp, n, cpu_cache_get(searchp), node);
3955 * These are racy checks but it does not matter
3956 * if we skip one check or scan twice.
3958 if (time_after(n->next_reap, jiffies))
3961 n->next_reap = jiffies + REAPTIMEOUT_NODE;
3963 drain_array(searchp, n, n->shared, node);
3965 if (n->free_touched)
3966 n->free_touched = 0;
3970 freed = drain_freelist(searchp, n, (n->free_limit +
3971 5 * searchp->num - 1) / (5 * searchp->num));
3972 STATS_ADD_REAPED(searchp, freed);
3978 mutex_unlock(&slab_mutex);
3981 /* Set up the next iteration */
3982 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC));
3985 #ifdef CONFIG_SLABINFO
3986 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
3989 unsigned long active_objs;
3990 unsigned long num_objs;
3991 unsigned long active_slabs = 0;
3992 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3996 struct kmem_cache_node *n;
4000 for_each_kmem_cache_node(cachep, node, n) {
4003 spin_lock_irq(&n->list_lock);
4005 list_for_each_entry(page, &n->slabs_full, lru) {
4006 if (page->active != cachep->num && !error)
4007 error = "slabs_full accounting error";
4008 active_objs += cachep->num;
4011 list_for_each_entry(page, &n->slabs_partial, lru) {
4012 if (page->active == cachep->num && !error)
4013 error = "slabs_partial accounting error";
4014 if (!page->active && !error)
4015 error = "slabs_partial accounting error";
4016 active_objs += page->active;
4019 list_for_each_entry(page, &n->slabs_free, lru) {
4020 if (page->active && !error)
4021 error = "slabs_free accounting error";
4024 free_objects += n->free_objects;
4026 shared_avail += n->shared->avail;
4028 spin_unlock_irq(&n->list_lock);
4030 num_slabs += active_slabs;
4031 num_objs = num_slabs * cachep->num;
4032 if (num_objs - active_objs != free_objects && !error)
4033 error = "free_objects accounting error";
4035 name = cachep->name;
4037 pr_err("slab: cache %s error: %s\n", name, error);
4039 sinfo->active_objs = active_objs;
4040 sinfo->num_objs = num_objs;
4041 sinfo->active_slabs = active_slabs;
4042 sinfo->num_slabs = num_slabs;
4043 sinfo->shared_avail = shared_avail;
4044 sinfo->limit = cachep->limit;
4045 sinfo->batchcount = cachep->batchcount;
4046 sinfo->shared = cachep->shared;
4047 sinfo->objects_per_slab = cachep->num;
4048 sinfo->cache_order = cachep->gfporder;
4051 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4055 unsigned long high = cachep->high_mark;
4056 unsigned long allocs = cachep->num_allocations;
4057 unsigned long grown = cachep->grown;
4058 unsigned long reaped = cachep->reaped;
4059 unsigned long errors = cachep->errors;
4060 unsigned long max_freeable = cachep->max_freeable;
4061 unsigned long node_allocs = cachep->node_allocs;
4062 unsigned long node_frees = cachep->node_frees;
4063 unsigned long overflows = cachep->node_overflow;
4065 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
4066 allocs, high, grown,
4067 reaped, errors, max_freeable, node_allocs,
4068 node_frees, overflows);
4072 unsigned long allochit = atomic_read(&cachep->allochit);
4073 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4074 unsigned long freehit = atomic_read(&cachep->freehit);
4075 unsigned long freemiss = atomic_read(&cachep->freemiss);
4077 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4078 allochit, allocmiss, freehit, freemiss);
4083 #define MAX_SLABINFO_WRITE 128
4085 * slabinfo_write - Tuning for the slab allocator
4087 * @buffer: user buffer
4088 * @count: data length
4091 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4092 size_t count, loff_t *ppos)
4094 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4095 int limit, batchcount, shared, res;
4096 struct kmem_cache *cachep;
4098 if (count > MAX_SLABINFO_WRITE)
4100 if (copy_from_user(&kbuf, buffer, count))
4102 kbuf[MAX_SLABINFO_WRITE] = '\0';
4104 tmp = strchr(kbuf, ' ');
4109 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4112 /* Find the cache in the chain of caches. */
4113 mutex_lock(&slab_mutex);
4115 list_for_each_entry(cachep, &slab_caches, list) {
4116 if (!strcmp(cachep->name, kbuf)) {
4117 if (limit < 1 || batchcount < 1 ||
4118 batchcount > limit || shared < 0) {
4121 res = do_tune_cpucache(cachep, limit,
4128 mutex_unlock(&slab_mutex);
4134 #ifdef CONFIG_DEBUG_SLAB_LEAK
4136 static inline int add_caller(unsigned long *n, unsigned long v)
4146 unsigned long *q = p + 2 * i;
4160 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4166 static void handle_slab(unsigned long *n, struct kmem_cache *c,
4175 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
4178 for (j = page->active; j < c->num; j++) {
4179 if (get_free_obj(page, j) == i) {
4189 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4190 * mapping is established when actual object allocation and
4191 * we could mistakenly access the unmapped object in the cpu
4194 if (probe_kernel_read(&v, dbg_userword(c, p), sizeof(v)))
4197 if (!add_caller(n, v))
4202 static void show_symbol(struct seq_file *m, unsigned long address)
4204 #ifdef CONFIG_KALLSYMS
4205 unsigned long offset, size;
4206 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4208 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4209 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4211 seq_printf(m, " [%s]", modname);
4215 seq_printf(m, "%p", (void *)address);
4218 static int leaks_show(struct seq_file *m, void *p)
4220 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4222 struct kmem_cache_node *n;
4224 unsigned long *x = m->private;
4228 if (!(cachep->flags & SLAB_STORE_USER))
4230 if (!(cachep->flags & SLAB_RED_ZONE))
4234 * Set store_user_clean and start to grab stored user information
4235 * for all objects on this cache. If some alloc/free requests comes
4236 * during the processing, information would be wrong so restart
4240 set_store_user_clean(cachep);
4241 drain_cpu_caches(cachep);
4245 for_each_kmem_cache_node(cachep, node, n) {
4248 spin_lock_irq(&n->list_lock);
4250 list_for_each_entry(page, &n->slabs_full, lru)
4251 handle_slab(x, cachep, page);
4252 list_for_each_entry(page, &n->slabs_partial, lru)
4253 handle_slab(x, cachep, page);
4254 spin_unlock_irq(&n->list_lock);
4256 } while (!is_store_user_clean(cachep));
4258 name = cachep->name;
4260 /* Increase the buffer size */
4261 mutex_unlock(&slab_mutex);
4262 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4264 /* Too bad, we are really out */
4266 mutex_lock(&slab_mutex);
4269 *(unsigned long *)m->private = x[0] * 2;
4271 mutex_lock(&slab_mutex);
4272 /* Now make sure this entry will be retried */
4276 for (i = 0; i < x[1]; i++) {
4277 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4278 show_symbol(m, x[2*i+2]);
4285 static const struct seq_operations slabstats_op = {
4286 .start = slab_start,
4292 static int slabstats_open(struct inode *inode, struct file *file)
4296 n = __seq_open_private(file, &slabstats_op, PAGE_SIZE);
4300 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4305 static const struct file_operations proc_slabstats_operations = {
4306 .open = slabstats_open,
4308 .llseek = seq_lseek,
4309 .release = seq_release_private,
4313 static int __init slab_proc_init(void)
4315 #ifdef CONFIG_DEBUG_SLAB_LEAK
4316 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4320 module_init(slab_proc_init);
4324 * ksize - get the actual amount of memory allocated for a given object
4325 * @objp: Pointer to the object
4327 * kmalloc may internally round up allocations and return more memory
4328 * than requested. ksize() can be used to determine the actual amount of
4329 * memory allocated. The caller may use this additional memory, even though
4330 * a smaller amount of memory was initially specified with the kmalloc call.
4331 * The caller must guarantee that objp points to a valid object previously
4332 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4333 * must not be freed during the duration of the call.
4335 size_t ksize(const void *objp)
4340 if (unlikely(objp == ZERO_SIZE_PTR))
4343 size = virt_to_cache(objp)->object_size;
4344 /* We assume that ksize callers could use the whole allocated area,
4345 * so we need to unpoison this area.
4347 kasan_krealloc(objp, size, GFP_NOWAIT);
4351 EXPORT_SYMBOL(ksize);