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)
172 * true if a page was allocated from pfmemalloc reserves for network-based
175 static bool pfmemalloc_active __read_mostly;
181 * - LIFO ordering, to hand out cache-warm objects from _alloc
182 * - reduce the number of linked list operations
183 * - reduce spinlock operations
185 * The limit is stored in the per-cpu structure to reduce the data cache
192 unsigned int batchcount;
193 unsigned int touched;
195 * Must have this definition in here for the proper
196 * alignment of array_cache. Also simplifies accessing
199 * Entries should not be directly dereferenced as
200 * entries belonging to slabs marked pfmemalloc will
201 * have the lower bits set SLAB_OBJ_PFMEMALLOC
207 struct array_cache ac;
210 #define SLAB_OBJ_PFMEMALLOC 1
211 static inline bool is_obj_pfmemalloc(void *objp)
213 return (unsigned long)objp & SLAB_OBJ_PFMEMALLOC;
216 static inline void set_obj_pfmemalloc(void **objp)
218 *objp = (void *)((unsigned long)*objp | SLAB_OBJ_PFMEMALLOC);
222 static inline void clear_obj_pfmemalloc(void **objp)
224 *objp = (void *)((unsigned long)*objp & ~SLAB_OBJ_PFMEMALLOC);
228 * Need this for bootstrapping a per node allocator.
230 #define NUM_INIT_LISTS (2 * MAX_NUMNODES)
231 static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
232 #define CACHE_CACHE 0
233 #define SIZE_NODE (MAX_NUMNODES)
235 static int drain_freelist(struct kmem_cache *cache,
236 struct kmem_cache_node *n, int tofree);
237 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
238 int node, struct list_head *list);
239 static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list);
240 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
241 static void cache_reap(struct work_struct *unused);
243 static int slab_early_init = 1;
245 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
247 static void kmem_cache_node_init(struct kmem_cache_node *parent)
249 INIT_LIST_HEAD(&parent->slabs_full);
250 INIT_LIST_HEAD(&parent->slabs_partial);
251 INIT_LIST_HEAD(&parent->slabs_free);
252 parent->shared = NULL;
253 parent->alien = NULL;
254 parent->colour_next = 0;
255 spin_lock_init(&parent->list_lock);
256 parent->free_objects = 0;
257 parent->free_touched = 0;
260 #define MAKE_LIST(cachep, listp, slab, nodeid) \
262 INIT_LIST_HEAD(listp); \
263 list_splice(&get_node(cachep, nodeid)->slab, listp); \
266 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
268 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
269 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
270 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
273 #define CFLGS_OFF_SLAB (0x80000000UL)
274 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
275 #define OFF_SLAB_MIN_SIZE (max_t(size_t, PAGE_SIZE >> 5, KMALLOC_MIN_SIZE + 1))
277 #define BATCHREFILL_LIMIT 16
279 * Optimization question: fewer reaps means less probability for unnessary
280 * cpucache drain/refill cycles.
282 * OTOH the cpuarrays can contain lots of objects,
283 * which could lock up otherwise freeable slabs.
285 #define REAPTIMEOUT_AC (2*HZ)
286 #define REAPTIMEOUT_NODE (4*HZ)
289 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
290 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
291 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
292 #define STATS_INC_GROWN(x) ((x)->grown++)
293 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
294 #define STATS_SET_HIGH(x) \
296 if ((x)->num_active > (x)->high_mark) \
297 (x)->high_mark = (x)->num_active; \
299 #define STATS_INC_ERR(x) ((x)->errors++)
300 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
301 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
302 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
303 #define STATS_SET_FREEABLE(x, i) \
305 if ((x)->max_freeable < i) \
306 (x)->max_freeable = i; \
308 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
309 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
310 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
311 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
313 #define STATS_INC_ACTIVE(x) do { } while (0)
314 #define STATS_DEC_ACTIVE(x) do { } while (0)
315 #define STATS_INC_ALLOCED(x) do { } while (0)
316 #define STATS_INC_GROWN(x) do { } while (0)
317 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
318 #define STATS_SET_HIGH(x) do { } while (0)
319 #define STATS_INC_ERR(x) do { } while (0)
320 #define STATS_INC_NODEALLOCS(x) do { } while (0)
321 #define STATS_INC_NODEFREES(x) do { } while (0)
322 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
323 #define STATS_SET_FREEABLE(x, i) do { } while (0)
324 #define STATS_INC_ALLOCHIT(x) do { } while (0)
325 #define STATS_INC_ALLOCMISS(x) do { } while (0)
326 #define STATS_INC_FREEHIT(x) do { } while (0)
327 #define STATS_INC_FREEMISS(x) do { } while (0)
333 * memory layout of objects:
335 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
336 * the end of an object is aligned with the end of the real
337 * allocation. Catches writes behind the end of the allocation.
338 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
340 * cachep->obj_offset: The real object.
341 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
342 * cachep->size - 1* BYTES_PER_WORD: last caller address
343 * [BYTES_PER_WORD long]
345 static int obj_offset(struct kmem_cache *cachep)
347 return cachep->obj_offset;
350 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
352 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
353 return (unsigned long long*) (objp + obj_offset(cachep) -
354 sizeof(unsigned long long));
357 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
359 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
360 if (cachep->flags & SLAB_STORE_USER)
361 return (unsigned long long *)(objp + cachep->size -
362 sizeof(unsigned long long) -
364 return (unsigned long long *) (objp + cachep->size -
365 sizeof(unsigned long long));
368 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
370 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
371 return (void **)(objp + cachep->size - BYTES_PER_WORD);
376 #define obj_offset(x) 0
377 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
378 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
379 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
383 #define OBJECT_FREE (0)
384 #define OBJECT_ACTIVE (1)
386 #ifdef CONFIG_DEBUG_SLAB_LEAK
388 static void set_obj_status(struct page *page, int idx, int val)
392 struct kmem_cache *cachep = page->slab_cache;
394 freelist_size = cachep->num * sizeof(freelist_idx_t);
395 status = (char *)page->freelist + freelist_size;
399 static inline bool is_store_user_clean(struct kmem_cache *cachep)
401 return atomic_read(&cachep->store_user_clean) == 1;
404 static inline void set_store_user_clean(struct kmem_cache *cachep)
406 atomic_set(&cachep->store_user_clean, 1);
409 static inline void set_store_user_dirty(struct kmem_cache *cachep)
411 if (is_store_user_clean(cachep))
412 atomic_set(&cachep->store_user_clean, 0);
416 static inline void set_obj_status(struct page *page, int idx, int val) {}
417 static inline void set_store_user_dirty(struct kmem_cache *cachep) {}
422 * Do not go above this order unless 0 objects fit into the slab or
423 * overridden on the command line.
425 #define SLAB_MAX_ORDER_HI 1
426 #define SLAB_MAX_ORDER_LO 0
427 static int slab_max_order = SLAB_MAX_ORDER_LO;
428 static bool slab_max_order_set __initdata;
430 static inline struct kmem_cache *virt_to_cache(const void *obj)
432 struct page *page = virt_to_head_page(obj);
433 return page->slab_cache;
436 static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
439 return page->s_mem + cache->size * idx;
443 * We want to avoid an expensive divide : (offset / cache->size)
444 * Using the fact that size is a constant for a particular cache,
445 * we can replace (offset / cache->size) by
446 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
448 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
449 const struct page *page, void *obj)
451 u32 offset = (obj - page->s_mem);
452 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
455 #define BOOT_CPUCACHE_ENTRIES 1
456 /* internal cache of cache description objs */
457 static struct kmem_cache kmem_cache_boot = {
459 .limit = BOOT_CPUCACHE_ENTRIES,
461 .size = sizeof(struct kmem_cache),
462 .name = "kmem_cache",
465 #define BAD_ALIEN_MAGIC 0x01020304ul
467 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
469 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
471 return this_cpu_ptr(cachep->cpu_cache);
474 static size_t calculate_freelist_size(int nr_objs, size_t align)
476 size_t freelist_size;
478 freelist_size = nr_objs * sizeof(freelist_idx_t);
479 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK))
480 freelist_size += nr_objs * sizeof(char);
483 freelist_size = ALIGN(freelist_size, align);
485 return freelist_size;
488 static int calculate_nr_objs(size_t slab_size, size_t buffer_size,
489 size_t idx_size, size_t align)
492 size_t remained_size;
493 size_t freelist_size;
496 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK))
497 extra_space = sizeof(char);
499 * Ignore padding for the initial guess. The padding
500 * is at most @align-1 bytes, and @buffer_size is at
501 * least @align. In the worst case, this result will
502 * be one greater than the number of objects that fit
503 * into the memory allocation when taking the padding
506 nr_objs = slab_size / (buffer_size + idx_size + extra_space);
509 * This calculated number will be either the right
510 * amount, or one greater than what we want.
512 remained_size = slab_size - nr_objs * buffer_size;
513 freelist_size = calculate_freelist_size(nr_objs, align);
514 if (remained_size < freelist_size)
521 * Calculate the number of objects and left-over bytes for a given buffer size.
523 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
524 size_t align, int flags, size_t *left_over,
529 size_t slab_size = PAGE_SIZE << gfporder;
532 * The slab management structure can be either off the slab or
533 * on it. For the latter case, the memory allocated for a
536 * - One freelist_idx_t for each object
537 * - Padding to respect alignment of @align
538 * - @buffer_size bytes for each object
540 * If the slab management structure is off the slab, then the
541 * alignment will already be calculated into the size. Because
542 * the slabs are all pages aligned, the objects will be at the
543 * correct alignment when allocated.
545 if (flags & CFLGS_OFF_SLAB) {
547 nr_objs = slab_size / buffer_size;
550 nr_objs = calculate_nr_objs(slab_size, buffer_size,
551 sizeof(freelist_idx_t), align);
552 mgmt_size = calculate_freelist_size(nr_objs, align);
555 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
559 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
561 static void __slab_error(const char *function, struct kmem_cache *cachep,
564 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
565 function, cachep->name, msg);
567 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
572 * By default on NUMA we use alien caches to stage the freeing of
573 * objects allocated from other nodes. This causes massive memory
574 * inefficiencies when using fake NUMA setup to split memory into a
575 * large number of small nodes, so it can be disabled on the command
579 static int use_alien_caches __read_mostly = 1;
580 static int __init noaliencache_setup(char *s)
582 use_alien_caches = 0;
585 __setup("noaliencache", noaliencache_setup);
587 static int __init slab_max_order_setup(char *str)
589 get_option(&str, &slab_max_order);
590 slab_max_order = slab_max_order < 0 ? 0 :
591 min(slab_max_order, MAX_ORDER - 1);
592 slab_max_order_set = true;
596 __setup("slab_max_order=", slab_max_order_setup);
600 * Special reaping functions for NUMA systems called from cache_reap().
601 * These take care of doing round robin flushing of alien caches (containing
602 * objects freed on different nodes from which they were allocated) and the
603 * flushing of remote pcps by calling drain_node_pages.
605 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
607 static void init_reap_node(int cpu)
611 node = next_node(cpu_to_mem(cpu), node_online_map);
612 if (node == MAX_NUMNODES)
613 node = first_node(node_online_map);
615 per_cpu(slab_reap_node, cpu) = node;
618 static void next_reap_node(void)
620 int node = __this_cpu_read(slab_reap_node);
622 node = next_node(node, node_online_map);
623 if (unlikely(node >= MAX_NUMNODES))
624 node = first_node(node_online_map);
625 __this_cpu_write(slab_reap_node, node);
629 #define init_reap_node(cpu) do { } while (0)
630 #define next_reap_node(void) do { } while (0)
634 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
635 * via the workqueue/eventd.
636 * Add the CPU number into the expiration time to minimize the possibility of
637 * the CPUs getting into lockstep and contending for the global cache chain
640 static void start_cpu_timer(int cpu)
642 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
645 * When this gets called from do_initcalls via cpucache_init(),
646 * init_workqueues() has already run, so keventd will be setup
649 if (keventd_up() && reap_work->work.func == NULL) {
651 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
652 schedule_delayed_work_on(cpu, reap_work,
653 __round_jiffies_relative(HZ, cpu));
657 static void init_arraycache(struct array_cache *ac, int limit, int batch)
660 * The array_cache structures contain pointers to free object.
661 * However, when such objects are allocated or transferred to another
662 * cache the pointers are not cleared and they could be counted as
663 * valid references during a kmemleak scan. Therefore, kmemleak must
664 * not scan such objects.
666 kmemleak_no_scan(ac);
670 ac->batchcount = batch;
675 static struct array_cache *alloc_arraycache(int node, int entries,
676 int batchcount, gfp_t gfp)
678 size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache);
679 struct array_cache *ac = NULL;
681 ac = kmalloc_node(memsize, gfp, node);
682 init_arraycache(ac, entries, batchcount);
686 static inline bool is_slab_pfmemalloc(struct page *page)
688 return PageSlabPfmemalloc(page);
691 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
692 static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
693 struct array_cache *ac)
695 struct kmem_cache_node *n = get_node(cachep, numa_mem_id());
699 if (!pfmemalloc_active)
702 spin_lock_irqsave(&n->list_lock, flags);
703 list_for_each_entry(page, &n->slabs_full, lru)
704 if (is_slab_pfmemalloc(page))
707 list_for_each_entry(page, &n->slabs_partial, lru)
708 if (is_slab_pfmemalloc(page))
711 list_for_each_entry(page, &n->slabs_free, lru)
712 if (is_slab_pfmemalloc(page))
715 pfmemalloc_active = false;
717 spin_unlock_irqrestore(&n->list_lock, flags);
720 static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
721 gfp_t flags, bool force_refill)
724 void *objp = ac->entry[--ac->avail];
726 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
727 if (unlikely(is_obj_pfmemalloc(objp))) {
728 struct kmem_cache_node *n;
730 if (gfp_pfmemalloc_allowed(flags)) {
731 clear_obj_pfmemalloc(&objp);
735 /* The caller cannot use PFMEMALLOC objects, find another one */
736 for (i = 0; i < ac->avail; i++) {
737 /* If a !PFMEMALLOC object is found, swap them */
738 if (!is_obj_pfmemalloc(ac->entry[i])) {
740 ac->entry[i] = ac->entry[ac->avail];
741 ac->entry[ac->avail] = objp;
747 * If there are empty slabs on the slabs_free list and we are
748 * being forced to refill the cache, mark this one !pfmemalloc.
750 n = get_node(cachep, numa_mem_id());
751 if (!list_empty(&n->slabs_free) && force_refill) {
752 struct page *page = virt_to_head_page(objp);
753 ClearPageSlabPfmemalloc(page);
754 clear_obj_pfmemalloc(&objp);
755 recheck_pfmemalloc_active(cachep, ac);
759 /* No !PFMEMALLOC objects available */
767 static inline void *ac_get_obj(struct kmem_cache *cachep,
768 struct array_cache *ac, gfp_t flags, bool force_refill)
772 if (unlikely(sk_memalloc_socks()))
773 objp = __ac_get_obj(cachep, ac, flags, force_refill);
775 objp = ac->entry[--ac->avail];
780 static noinline void *__ac_put_obj(struct kmem_cache *cachep,
781 struct array_cache *ac, void *objp)
783 if (unlikely(pfmemalloc_active)) {
784 /* Some pfmemalloc slabs exist, check if this is one */
785 struct page *page = virt_to_head_page(objp);
786 if (PageSlabPfmemalloc(page))
787 set_obj_pfmemalloc(&objp);
793 static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
796 if (unlikely(sk_memalloc_socks()))
797 objp = __ac_put_obj(cachep, ac, objp);
799 ac->entry[ac->avail++] = objp;
803 * Transfer objects in one arraycache to another.
804 * Locking must be handled by the caller.
806 * Return the number of entries transferred.
808 static int transfer_objects(struct array_cache *to,
809 struct array_cache *from, unsigned int max)
811 /* Figure out how many entries to transfer */
812 int nr = min3(from->avail, max, to->limit - to->avail);
817 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
827 #define drain_alien_cache(cachep, alien) do { } while (0)
828 #define reap_alien(cachep, n) do { } while (0)
830 static inline struct alien_cache **alloc_alien_cache(int node,
831 int limit, gfp_t gfp)
833 return (struct alien_cache **)BAD_ALIEN_MAGIC;
836 static inline void free_alien_cache(struct alien_cache **ac_ptr)
840 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
845 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
851 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
852 gfp_t flags, int nodeid)
857 static inline gfp_t gfp_exact_node(gfp_t flags)
862 #else /* CONFIG_NUMA */
864 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
865 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
867 static struct alien_cache *__alloc_alien_cache(int node, int entries,
868 int batch, gfp_t gfp)
870 size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
871 struct alien_cache *alc = NULL;
873 alc = kmalloc_node(memsize, gfp, node);
874 init_arraycache(&alc->ac, entries, batch);
875 spin_lock_init(&alc->lock);
879 static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
881 struct alien_cache **alc_ptr;
882 size_t memsize = sizeof(void *) * nr_node_ids;
887 alc_ptr = kzalloc_node(memsize, gfp, node);
892 if (i == node || !node_online(i))
894 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
896 for (i--; i >= 0; i--)
905 static void free_alien_cache(struct alien_cache **alc_ptr)
916 static void __drain_alien_cache(struct kmem_cache *cachep,
917 struct array_cache *ac, int node,
918 struct list_head *list)
920 struct kmem_cache_node *n = get_node(cachep, node);
923 spin_lock(&n->list_lock);
925 * Stuff objects into the remote nodes shared array first.
926 * That way we could avoid the overhead of putting the objects
927 * into the free lists and getting them back later.
930 transfer_objects(n->shared, ac, ac->limit);
932 free_block(cachep, ac->entry, ac->avail, node, list);
934 spin_unlock(&n->list_lock);
939 * Called from cache_reap() to regularly drain alien caches round robin.
941 static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
943 int node = __this_cpu_read(slab_reap_node);
946 struct alien_cache *alc = n->alien[node];
947 struct array_cache *ac;
951 if (ac->avail && spin_trylock_irq(&alc->lock)) {
954 __drain_alien_cache(cachep, ac, node, &list);
955 spin_unlock_irq(&alc->lock);
956 slabs_destroy(cachep, &list);
962 static void drain_alien_cache(struct kmem_cache *cachep,
963 struct alien_cache **alien)
966 struct alien_cache *alc;
967 struct array_cache *ac;
970 for_each_online_node(i) {
976 spin_lock_irqsave(&alc->lock, flags);
977 __drain_alien_cache(cachep, ac, i, &list);
978 spin_unlock_irqrestore(&alc->lock, flags);
979 slabs_destroy(cachep, &list);
984 static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
985 int node, int page_node)
987 struct kmem_cache_node *n;
988 struct alien_cache *alien = NULL;
989 struct array_cache *ac;
992 n = get_node(cachep, node);
993 STATS_INC_NODEFREES(cachep);
994 if (n->alien && n->alien[page_node]) {
995 alien = n->alien[page_node];
997 spin_lock(&alien->lock);
998 if (unlikely(ac->avail == ac->limit)) {
999 STATS_INC_ACOVERFLOW(cachep);
1000 __drain_alien_cache(cachep, ac, page_node, &list);
1002 ac_put_obj(cachep, ac, objp);
1003 spin_unlock(&alien->lock);
1004 slabs_destroy(cachep, &list);
1006 n = get_node(cachep, page_node);
1007 spin_lock(&n->list_lock);
1008 free_block(cachep, &objp, 1, page_node, &list);
1009 spin_unlock(&n->list_lock);
1010 slabs_destroy(cachep, &list);
1015 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1017 int page_node = page_to_nid(virt_to_page(objp));
1018 int node = numa_mem_id();
1020 * Make sure we are not freeing a object from another node to the array
1021 * cache on this cpu.
1023 if (likely(node == page_node))
1026 return __cache_free_alien(cachep, objp, node, page_node);
1030 * Construct gfp mask to allocate from a specific node but do not direct reclaim
1031 * or warn about failures. kswapd may still wake to reclaim in the background.
1033 static inline gfp_t gfp_exact_node(gfp_t flags)
1035 return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~__GFP_DIRECT_RECLAIM;
1040 * Allocates and initializes node for a node on each slab cache, used for
1041 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
1042 * will be allocated off-node since memory is not yet online for the new node.
1043 * When hotplugging memory or a cpu, existing node are not replaced if
1046 * Must hold slab_mutex.
1048 static int init_cache_node_node(int node)
1050 struct kmem_cache *cachep;
1051 struct kmem_cache_node *n;
1052 const size_t memsize = sizeof(struct kmem_cache_node);
1054 list_for_each_entry(cachep, &slab_caches, list) {
1056 * Set up the kmem_cache_node for cpu before we can
1057 * begin anything. Make sure some other cpu on this
1058 * node has not already allocated this
1060 n = get_node(cachep, node);
1062 n = kmalloc_node(memsize, GFP_KERNEL, node);
1065 kmem_cache_node_init(n);
1066 n->next_reap = jiffies + REAPTIMEOUT_NODE +
1067 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1070 * The kmem_cache_nodes don't come and go as CPUs
1071 * come and go. slab_mutex is sufficient
1074 cachep->node[node] = n;
1077 spin_lock_irq(&n->list_lock);
1079 (1 + nr_cpus_node(node)) *
1080 cachep->batchcount + cachep->num;
1081 spin_unlock_irq(&n->list_lock);
1086 static inline int slabs_tofree(struct kmem_cache *cachep,
1087 struct kmem_cache_node *n)
1089 return (n->free_objects + cachep->num - 1) / cachep->num;
1092 static void cpuup_canceled(long cpu)
1094 struct kmem_cache *cachep;
1095 struct kmem_cache_node *n = NULL;
1096 int node = cpu_to_mem(cpu);
1097 const struct cpumask *mask = cpumask_of_node(node);
1099 list_for_each_entry(cachep, &slab_caches, list) {
1100 struct array_cache *nc;
1101 struct array_cache *shared;
1102 struct alien_cache **alien;
1105 n = get_node(cachep, node);
1109 spin_lock_irq(&n->list_lock);
1111 /* Free limit for this kmem_cache_node */
1112 n->free_limit -= cachep->batchcount;
1114 /* cpu is dead; no one can alloc from it. */
1115 nc = per_cpu_ptr(cachep->cpu_cache, cpu);
1117 free_block(cachep, nc->entry, nc->avail, node, &list);
1121 if (!cpumask_empty(mask)) {
1122 spin_unlock_irq(&n->list_lock);
1128 free_block(cachep, shared->entry,
1129 shared->avail, node, &list);
1136 spin_unlock_irq(&n->list_lock);
1140 drain_alien_cache(cachep, alien);
1141 free_alien_cache(alien);
1145 slabs_destroy(cachep, &list);
1148 * In the previous loop, all the objects were freed to
1149 * the respective cache's slabs, now we can go ahead and
1150 * shrink each nodelist to its limit.
1152 list_for_each_entry(cachep, &slab_caches, list) {
1153 n = get_node(cachep, node);
1156 drain_freelist(cachep, n, slabs_tofree(cachep, n));
1160 static int cpuup_prepare(long cpu)
1162 struct kmem_cache *cachep;
1163 struct kmem_cache_node *n = NULL;
1164 int node = cpu_to_mem(cpu);
1168 * We need to do this right in the beginning since
1169 * alloc_arraycache's are going to use this list.
1170 * kmalloc_node allows us to add the slab to the right
1171 * kmem_cache_node and not this cpu's kmem_cache_node
1173 err = init_cache_node_node(node);
1178 * Now we can go ahead with allocating the shared arrays and
1181 list_for_each_entry(cachep, &slab_caches, list) {
1182 struct array_cache *shared = NULL;
1183 struct alien_cache **alien = NULL;
1185 if (cachep->shared) {
1186 shared = alloc_arraycache(node,
1187 cachep->shared * cachep->batchcount,
1188 0xbaadf00d, GFP_KERNEL);
1192 if (use_alien_caches) {
1193 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1199 n = get_node(cachep, node);
1202 spin_lock_irq(&n->list_lock);
1205 * We are serialised from CPU_DEAD or
1206 * CPU_UP_CANCELLED by the cpucontrol lock
1217 spin_unlock_irq(&n->list_lock);
1219 free_alien_cache(alien);
1224 cpuup_canceled(cpu);
1228 static int cpuup_callback(struct notifier_block *nfb,
1229 unsigned long action, void *hcpu)
1231 long cpu = (long)hcpu;
1235 case CPU_UP_PREPARE:
1236 case CPU_UP_PREPARE_FROZEN:
1237 mutex_lock(&slab_mutex);
1238 err = cpuup_prepare(cpu);
1239 mutex_unlock(&slab_mutex);
1242 case CPU_ONLINE_FROZEN:
1243 start_cpu_timer(cpu);
1245 #ifdef CONFIG_HOTPLUG_CPU
1246 case CPU_DOWN_PREPARE:
1247 case CPU_DOWN_PREPARE_FROZEN:
1249 * Shutdown cache reaper. Note that the slab_mutex is
1250 * held so that if cache_reap() is invoked it cannot do
1251 * anything expensive but will only modify reap_work
1252 * and reschedule the timer.
1254 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1255 /* Now the cache_reaper is guaranteed to be not running. */
1256 per_cpu(slab_reap_work, cpu).work.func = NULL;
1258 case CPU_DOWN_FAILED:
1259 case CPU_DOWN_FAILED_FROZEN:
1260 start_cpu_timer(cpu);
1263 case CPU_DEAD_FROZEN:
1265 * Even if all the cpus of a node are down, we don't free the
1266 * kmem_cache_node of any cache. This to avoid a race between
1267 * cpu_down, and a kmalloc allocation from another cpu for
1268 * memory from the node of the cpu going down. The node
1269 * structure is usually allocated from kmem_cache_create() and
1270 * gets destroyed at kmem_cache_destroy().
1274 case CPU_UP_CANCELED:
1275 case CPU_UP_CANCELED_FROZEN:
1276 mutex_lock(&slab_mutex);
1277 cpuup_canceled(cpu);
1278 mutex_unlock(&slab_mutex);
1281 return notifier_from_errno(err);
1284 static struct notifier_block cpucache_notifier = {
1285 &cpuup_callback, NULL, 0
1288 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1290 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1291 * Returns -EBUSY if all objects cannot be drained so that the node is not
1294 * Must hold slab_mutex.
1296 static int __meminit drain_cache_node_node(int node)
1298 struct kmem_cache *cachep;
1301 list_for_each_entry(cachep, &slab_caches, list) {
1302 struct kmem_cache_node *n;
1304 n = get_node(cachep, node);
1308 drain_freelist(cachep, n, slabs_tofree(cachep, n));
1310 if (!list_empty(&n->slabs_full) ||
1311 !list_empty(&n->slabs_partial)) {
1319 static int __meminit slab_memory_callback(struct notifier_block *self,
1320 unsigned long action, void *arg)
1322 struct memory_notify *mnb = arg;
1326 nid = mnb->status_change_nid;
1331 case MEM_GOING_ONLINE:
1332 mutex_lock(&slab_mutex);
1333 ret = init_cache_node_node(nid);
1334 mutex_unlock(&slab_mutex);
1336 case MEM_GOING_OFFLINE:
1337 mutex_lock(&slab_mutex);
1338 ret = drain_cache_node_node(nid);
1339 mutex_unlock(&slab_mutex);
1343 case MEM_CANCEL_ONLINE:
1344 case MEM_CANCEL_OFFLINE:
1348 return notifier_from_errno(ret);
1350 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1353 * swap the static kmem_cache_node with kmalloced memory
1355 static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1358 struct kmem_cache_node *ptr;
1360 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1363 memcpy(ptr, list, sizeof(struct kmem_cache_node));
1365 * Do not assume that spinlocks can be initialized via memcpy:
1367 spin_lock_init(&ptr->list_lock);
1369 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1370 cachep->node[nodeid] = ptr;
1374 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1375 * size of kmem_cache_node.
1377 static void __init set_up_node(struct kmem_cache *cachep, int index)
1381 for_each_online_node(node) {
1382 cachep->node[node] = &init_kmem_cache_node[index + node];
1383 cachep->node[node]->next_reap = jiffies +
1385 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1390 * Initialisation. Called after the page allocator have been initialised and
1391 * before smp_init().
1393 void __init kmem_cache_init(void)
1397 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1398 sizeof(struct rcu_head));
1399 kmem_cache = &kmem_cache_boot;
1401 if (num_possible_nodes() == 1)
1402 use_alien_caches = 0;
1404 for (i = 0; i < NUM_INIT_LISTS; i++)
1405 kmem_cache_node_init(&init_kmem_cache_node[i]);
1408 * Fragmentation resistance on low memory - only use bigger
1409 * page orders on machines with more than 32MB of memory if
1410 * not overridden on the command line.
1412 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1413 slab_max_order = SLAB_MAX_ORDER_HI;
1415 /* Bootstrap is tricky, because several objects are allocated
1416 * from caches that do not exist yet:
1417 * 1) initialize the kmem_cache cache: it contains the struct
1418 * kmem_cache structures of all caches, except kmem_cache itself:
1419 * kmem_cache is statically allocated.
1420 * Initially an __init data area is used for the head array and the
1421 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1422 * array at the end of the bootstrap.
1423 * 2) Create the first kmalloc cache.
1424 * The struct kmem_cache for the new cache is allocated normally.
1425 * An __init data area is used for the head array.
1426 * 3) Create the remaining kmalloc caches, with minimally sized
1428 * 4) Replace the __init data head arrays for kmem_cache and the first
1429 * kmalloc cache with kmalloc allocated arrays.
1430 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1431 * the other cache's with kmalloc allocated memory.
1432 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1435 /* 1) create the kmem_cache */
1438 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1440 create_boot_cache(kmem_cache, "kmem_cache",
1441 offsetof(struct kmem_cache, node) +
1442 nr_node_ids * sizeof(struct kmem_cache_node *),
1443 SLAB_HWCACHE_ALIGN);
1444 list_add(&kmem_cache->list, &slab_caches);
1445 slab_state = PARTIAL;
1448 * Initialize the caches that provide memory for the kmem_cache_node
1449 * structures first. Without this, further allocations will bug.
1451 kmalloc_caches[INDEX_NODE] = create_kmalloc_cache("kmalloc-node",
1452 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
1453 slab_state = PARTIAL_NODE;
1454 setup_kmalloc_cache_index_table();
1456 slab_early_init = 0;
1458 /* 5) Replace the bootstrap kmem_cache_node */
1462 for_each_online_node(nid) {
1463 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
1465 init_list(kmalloc_caches[INDEX_NODE],
1466 &init_kmem_cache_node[SIZE_NODE + nid], nid);
1470 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1473 void __init kmem_cache_init_late(void)
1475 struct kmem_cache *cachep;
1479 /* 6) resize the head arrays to their final sizes */
1480 mutex_lock(&slab_mutex);
1481 list_for_each_entry(cachep, &slab_caches, list)
1482 if (enable_cpucache(cachep, GFP_NOWAIT))
1484 mutex_unlock(&slab_mutex);
1490 * Register a cpu startup notifier callback that initializes
1491 * cpu_cache_get for all new cpus
1493 register_cpu_notifier(&cpucache_notifier);
1497 * Register a memory hotplug callback that initializes and frees
1500 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1504 * The reap timers are started later, with a module init call: That part
1505 * of the kernel is not yet operational.
1509 static int __init cpucache_init(void)
1514 * Register the timers that return unneeded pages to the page allocator
1516 for_each_online_cpu(cpu)
1517 start_cpu_timer(cpu);
1523 __initcall(cpucache_init);
1525 static noinline void
1526 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1529 struct kmem_cache_node *n;
1531 unsigned long flags;
1533 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1534 DEFAULT_RATELIMIT_BURST);
1536 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1540 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1542 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
1543 cachep->name, cachep->size, cachep->gfporder);
1545 for_each_kmem_cache_node(cachep, node, n) {
1546 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1547 unsigned long active_slabs = 0, num_slabs = 0;
1549 spin_lock_irqsave(&n->list_lock, flags);
1550 list_for_each_entry(page, &n->slabs_full, lru) {
1551 active_objs += cachep->num;
1554 list_for_each_entry(page, &n->slabs_partial, lru) {
1555 active_objs += page->active;
1558 list_for_each_entry(page, &n->slabs_free, lru)
1561 free_objects += n->free_objects;
1562 spin_unlock_irqrestore(&n->list_lock, flags);
1564 num_slabs += active_slabs;
1565 num_objs = num_slabs * cachep->num;
1567 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1568 node, active_slabs, num_slabs, active_objs, num_objs,
1575 * Interface to system's page allocator. No need to hold the
1576 * kmem_cache_node ->list_lock.
1578 * If we requested dmaable memory, we will get it. Even if we
1579 * did not request dmaable memory, we might get it, but that
1580 * would be relatively rare and ignorable.
1582 static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1588 flags |= cachep->allocflags;
1589 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1590 flags |= __GFP_RECLAIMABLE;
1592 page = __alloc_pages_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1594 slab_out_of_memory(cachep, flags, nodeid);
1598 if (memcg_charge_slab(page, flags, cachep->gfporder, cachep)) {
1599 __free_pages(page, cachep->gfporder);
1603 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1604 if (page_is_pfmemalloc(page))
1605 pfmemalloc_active = true;
1607 nr_pages = (1 << cachep->gfporder);
1608 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1609 add_zone_page_state(page_zone(page),
1610 NR_SLAB_RECLAIMABLE, nr_pages);
1612 add_zone_page_state(page_zone(page),
1613 NR_SLAB_UNRECLAIMABLE, nr_pages);
1614 __SetPageSlab(page);
1615 if (page_is_pfmemalloc(page))
1616 SetPageSlabPfmemalloc(page);
1618 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1619 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1622 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1624 kmemcheck_mark_unallocated_pages(page, nr_pages);
1631 * Interface to system's page release.
1633 static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
1635 const unsigned long nr_freed = (1 << cachep->gfporder);
1637 kmemcheck_free_shadow(page, cachep->gfporder);
1639 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1640 sub_zone_page_state(page_zone(page),
1641 NR_SLAB_RECLAIMABLE, nr_freed);
1643 sub_zone_page_state(page_zone(page),
1644 NR_SLAB_UNRECLAIMABLE, nr_freed);
1646 BUG_ON(!PageSlab(page));
1647 __ClearPageSlabPfmemalloc(page);
1648 __ClearPageSlab(page);
1649 page_mapcount_reset(page);
1650 page->mapping = NULL;
1652 if (current->reclaim_state)
1653 current->reclaim_state->reclaimed_slab += nr_freed;
1654 __free_kmem_pages(page, cachep->gfporder);
1657 static void kmem_rcu_free(struct rcu_head *head)
1659 struct kmem_cache *cachep;
1662 page = container_of(head, struct page, rcu_head);
1663 cachep = page->slab_cache;
1665 kmem_freepages(cachep, page);
1669 static bool is_debug_pagealloc_cache(struct kmem_cache *cachep)
1671 if (debug_pagealloc_enabled() && OFF_SLAB(cachep) &&
1672 (cachep->size % PAGE_SIZE) == 0)
1678 #ifdef CONFIG_DEBUG_PAGEALLOC
1679 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1680 unsigned long caller)
1682 int size = cachep->object_size;
1684 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1686 if (size < 5 * sizeof(unsigned long))
1689 *addr++ = 0x12345678;
1691 *addr++ = smp_processor_id();
1692 size -= 3 * sizeof(unsigned long);
1694 unsigned long *sptr = &caller;
1695 unsigned long svalue;
1697 while (!kstack_end(sptr)) {
1699 if (kernel_text_address(svalue)) {
1701 size -= sizeof(unsigned long);
1702 if (size <= sizeof(unsigned long))
1708 *addr++ = 0x87654321;
1711 static void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1712 int map, unsigned long caller)
1714 if (!is_debug_pagealloc_cache(cachep))
1718 store_stackinfo(cachep, objp, caller);
1720 kernel_map_pages(virt_to_page(objp), cachep->size / PAGE_SIZE, map);
1724 static inline void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1725 int map, unsigned long caller) {}
1729 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1731 int size = cachep->object_size;
1732 addr = &((char *)addr)[obj_offset(cachep)];
1734 memset(addr, val, size);
1735 *(unsigned char *)(addr + size - 1) = POISON_END;
1738 static void dump_line(char *data, int offset, int limit)
1741 unsigned char error = 0;
1744 printk(KERN_ERR "%03x: ", offset);
1745 for (i = 0; i < limit; i++) {
1746 if (data[offset + i] != POISON_FREE) {
1747 error = data[offset + i];
1751 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1752 &data[offset], limit, 1);
1754 if (bad_count == 1) {
1755 error ^= POISON_FREE;
1756 if (!(error & (error - 1))) {
1757 printk(KERN_ERR "Single bit error detected. Probably "
1760 printk(KERN_ERR "Run memtest86+ or a similar memory "
1763 printk(KERN_ERR "Run a memory test tool.\n");
1772 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1777 if (cachep->flags & SLAB_RED_ZONE) {
1778 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1779 *dbg_redzone1(cachep, objp),
1780 *dbg_redzone2(cachep, objp));
1783 if (cachep->flags & SLAB_STORE_USER) {
1784 printk(KERN_ERR "Last user: [<%p>](%pSR)\n",
1785 *dbg_userword(cachep, objp),
1786 *dbg_userword(cachep, objp));
1788 realobj = (char *)objp + obj_offset(cachep);
1789 size = cachep->object_size;
1790 for (i = 0; i < size && lines; i += 16, lines--) {
1793 if (i + limit > size)
1795 dump_line(realobj, i, limit);
1799 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1805 if (is_debug_pagealloc_cache(cachep))
1808 realobj = (char *)objp + obj_offset(cachep);
1809 size = cachep->object_size;
1811 for (i = 0; i < size; i++) {
1812 char exp = POISON_FREE;
1815 if (realobj[i] != exp) {
1821 "Slab corruption (%s): %s start=%p, len=%d\n",
1822 print_tainted(), cachep->name, realobj, size);
1823 print_objinfo(cachep, objp, 0);
1825 /* Hexdump the affected line */
1828 if (i + limit > size)
1830 dump_line(realobj, i, limit);
1833 /* Limit to 5 lines */
1839 /* Print some data about the neighboring objects, if they
1842 struct page *page = virt_to_head_page(objp);
1845 objnr = obj_to_index(cachep, page, objp);
1847 objp = index_to_obj(cachep, page, objnr - 1);
1848 realobj = (char *)objp + obj_offset(cachep);
1849 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1851 print_objinfo(cachep, objp, 2);
1853 if (objnr + 1 < cachep->num) {
1854 objp = index_to_obj(cachep, page, objnr + 1);
1855 realobj = (char *)objp + obj_offset(cachep);
1856 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1858 print_objinfo(cachep, objp, 2);
1865 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1869 for (i = 0; i < cachep->num; i++) {
1870 void *objp = index_to_obj(cachep, page, i);
1872 if (cachep->flags & SLAB_POISON) {
1873 check_poison_obj(cachep, objp);
1874 slab_kernel_map(cachep, objp, 1, 0);
1876 if (cachep->flags & SLAB_RED_ZONE) {
1877 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1878 slab_error(cachep, "start of a freed object "
1880 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1881 slab_error(cachep, "end of a freed object "
1887 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1894 * slab_destroy - destroy and release all objects in a slab
1895 * @cachep: cache pointer being destroyed
1896 * @page: page pointer being destroyed
1898 * Destroy all the objs in a slab page, and release the mem back to the system.
1899 * Before calling the slab page must have been unlinked from the cache. The
1900 * kmem_cache_node ->list_lock is not held/needed.
1902 static void slab_destroy(struct kmem_cache *cachep, struct page *page)
1906 freelist = page->freelist;
1907 slab_destroy_debugcheck(cachep, page);
1908 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
1909 call_rcu(&page->rcu_head, kmem_rcu_free);
1911 kmem_freepages(cachep, page);
1914 * From now on, we don't use freelist
1915 * although actual page can be freed in rcu context
1917 if (OFF_SLAB(cachep))
1918 kmem_cache_free(cachep->freelist_cache, freelist);
1921 static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1923 struct page *page, *n;
1925 list_for_each_entry_safe(page, n, list, lru) {
1926 list_del(&page->lru);
1927 slab_destroy(cachep, page);
1932 * calculate_slab_order - calculate size (page order) of slabs
1933 * @cachep: pointer to the cache that is being created
1934 * @size: size of objects to be created in this cache.
1935 * @align: required alignment for the objects.
1936 * @flags: slab allocation flags
1938 * Also calculates the number of objects per slab.
1940 * This could be made much more intelligent. For now, try to avoid using
1941 * high order pages for slabs. When the gfp() functions are more friendly
1942 * towards high-order requests, this should be changed.
1944 static size_t calculate_slab_order(struct kmem_cache *cachep,
1945 size_t size, size_t align, unsigned long flags)
1947 unsigned long offslab_limit;
1948 size_t left_over = 0;
1951 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
1955 cache_estimate(gfporder, size, align, flags, &remainder, &num);
1959 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1960 if (num > SLAB_OBJ_MAX_NUM)
1963 if (flags & CFLGS_OFF_SLAB) {
1964 size_t freelist_size_per_obj = sizeof(freelist_idx_t);
1966 * Max number of objs-per-slab for caches which
1967 * use off-slab slabs. Needed to avoid a possible
1968 * looping condition in cache_grow().
1970 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK))
1971 freelist_size_per_obj += sizeof(char);
1972 offslab_limit = size;
1973 offslab_limit /= freelist_size_per_obj;
1975 if (num > offslab_limit)
1979 /* Found something acceptable - save it away */
1981 cachep->gfporder = gfporder;
1982 left_over = remainder;
1985 * A VFS-reclaimable slab tends to have most allocations
1986 * as GFP_NOFS and we really don't want to have to be allocating
1987 * higher-order pages when we are unable to shrink dcache.
1989 if (flags & SLAB_RECLAIM_ACCOUNT)
1993 * Large number of objects is good, but very large slabs are
1994 * currently bad for the gfp()s.
1996 if (gfporder >= slab_max_order)
2000 * Acceptable internal fragmentation?
2002 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2008 static struct array_cache __percpu *alloc_kmem_cache_cpus(
2009 struct kmem_cache *cachep, int entries, int batchcount)
2013 struct array_cache __percpu *cpu_cache;
2015 size = sizeof(void *) * entries + sizeof(struct array_cache);
2016 cpu_cache = __alloc_percpu(size, sizeof(void *));
2021 for_each_possible_cpu(cpu) {
2022 init_arraycache(per_cpu_ptr(cpu_cache, cpu),
2023 entries, batchcount);
2029 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2031 if (slab_state >= FULL)
2032 return enable_cpucache(cachep, gfp);
2034 cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
2035 if (!cachep->cpu_cache)
2038 if (slab_state == DOWN) {
2039 /* Creation of first cache (kmem_cache). */
2040 set_up_node(kmem_cache, CACHE_CACHE);
2041 } else if (slab_state == PARTIAL) {
2042 /* For kmem_cache_node */
2043 set_up_node(cachep, SIZE_NODE);
2047 for_each_online_node(node) {
2048 cachep->node[node] = kmalloc_node(
2049 sizeof(struct kmem_cache_node), gfp, node);
2050 BUG_ON(!cachep->node[node]);
2051 kmem_cache_node_init(cachep->node[node]);
2055 cachep->node[numa_mem_id()]->next_reap =
2056 jiffies + REAPTIMEOUT_NODE +
2057 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
2059 cpu_cache_get(cachep)->avail = 0;
2060 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2061 cpu_cache_get(cachep)->batchcount = 1;
2062 cpu_cache_get(cachep)->touched = 0;
2063 cachep->batchcount = 1;
2064 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2068 unsigned long kmem_cache_flags(unsigned long object_size,
2069 unsigned long flags, const char *name,
2070 void (*ctor)(void *))
2076 __kmem_cache_alias(const char *name, size_t size, size_t align,
2077 unsigned long flags, void (*ctor)(void *))
2079 struct kmem_cache *cachep;
2081 cachep = find_mergeable(size, align, flags, name, ctor);
2086 * Adjust the object sizes so that we clear
2087 * the complete object on kzalloc.
2089 cachep->object_size = max_t(int, cachep->object_size, size);
2095 * __kmem_cache_create - Create a cache.
2096 * @cachep: cache management descriptor
2097 * @flags: SLAB flags
2099 * Returns a ptr to the cache on success, NULL on failure.
2100 * Cannot be called within a int, but can be interrupted.
2101 * The @ctor is run when new pages are allocated by the cache.
2105 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2106 * to catch references to uninitialised memory.
2108 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2109 * for buffer overruns.
2111 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2112 * cacheline. This can be beneficial if you're counting cycles as closely
2116 __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2118 size_t left_over, freelist_size;
2119 size_t ralign = BYTES_PER_WORD;
2122 size_t size = cachep->size;
2127 * Enable redzoning and last user accounting, except for caches with
2128 * large objects, if the increased size would increase the object size
2129 * above the next power of two: caches with object sizes just above a
2130 * power of two have a significant amount of internal fragmentation.
2132 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2133 2 * sizeof(unsigned long long)))
2134 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2135 if (!(flags & SLAB_DESTROY_BY_RCU))
2136 flags |= SLAB_POISON;
2141 * Check that size is in terms of words. This is needed to avoid
2142 * unaligned accesses for some archs when redzoning is used, and makes
2143 * sure any on-slab bufctl's are also correctly aligned.
2145 if (size & (BYTES_PER_WORD - 1)) {
2146 size += (BYTES_PER_WORD - 1);
2147 size &= ~(BYTES_PER_WORD - 1);
2150 if (flags & SLAB_RED_ZONE) {
2151 ralign = REDZONE_ALIGN;
2152 /* If redzoning, ensure that the second redzone is suitably
2153 * aligned, by adjusting the object size accordingly. */
2154 size += REDZONE_ALIGN - 1;
2155 size &= ~(REDZONE_ALIGN - 1);
2158 /* 3) caller mandated alignment */
2159 if (ralign < cachep->align) {
2160 ralign = cachep->align;
2162 /* disable debug if necessary */
2163 if (ralign > __alignof__(unsigned long long))
2164 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2168 cachep->align = ralign;
2170 if (slab_is_available())
2178 * Both debugging options require word-alignment which is calculated
2181 if (flags & SLAB_RED_ZONE) {
2182 /* add space for red zone words */
2183 cachep->obj_offset += sizeof(unsigned long long);
2184 size += 2 * sizeof(unsigned long long);
2186 if (flags & SLAB_STORE_USER) {
2187 /* user store requires one word storage behind the end of
2188 * the real object. But if the second red zone needs to be
2189 * aligned to 64 bits, we must allow that much space.
2191 if (flags & SLAB_RED_ZONE)
2192 size += REDZONE_ALIGN;
2194 size += BYTES_PER_WORD;
2197 * To activate debug pagealloc, off-slab management is necessary
2198 * requirement. In early phase of initialization, small sized slab
2199 * doesn't get initialized so it would not be possible. So, we need
2200 * to check size >= 256. It guarantees that all necessary small
2201 * sized slab is initialized in current slab initialization sequence.
2203 if (debug_pagealloc_enabled() && (flags & SLAB_POISON) &&
2204 !slab_early_init && size >= kmalloc_size(INDEX_NODE) &&
2205 size >= 256 && cachep->object_size > cache_line_size() &&
2206 ALIGN(size, cachep->align) < PAGE_SIZE) {
2207 cachep->obj_offset += PAGE_SIZE - ALIGN(size, cachep->align);
2213 * Determine if the slab management is 'on' or 'off' slab.
2214 * (bootstrapping cannot cope with offslab caches so don't do
2215 * it too early on. Always use on-slab management when
2216 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2218 if (size >= OFF_SLAB_MIN_SIZE && !slab_early_init &&
2219 !(flags & SLAB_NOLEAKTRACE))
2221 * Size is large, assume best to place the slab management obj
2222 * off-slab (should allow better packing of objs).
2224 flags |= CFLGS_OFF_SLAB;
2226 size = ALIGN(size, cachep->align);
2228 * We should restrict the number of objects in a slab to implement
2229 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2231 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2232 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2234 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
2239 freelist_size = calculate_freelist_size(cachep->num, cachep->align);
2242 * If the slab has been placed off-slab, and we have enough space then
2243 * move it on-slab. This is at the expense of any extra colouring.
2245 if (flags & CFLGS_OFF_SLAB && left_over >= freelist_size) {
2246 flags &= ~CFLGS_OFF_SLAB;
2247 left_over -= freelist_size;
2250 if (flags & CFLGS_OFF_SLAB) {
2251 /* really off slab. No need for manual alignment */
2252 freelist_size = calculate_freelist_size(cachep->num, 0);
2255 cachep->colour_off = cache_line_size();
2256 /* Offset must be a multiple of the alignment. */
2257 if (cachep->colour_off < cachep->align)
2258 cachep->colour_off = cachep->align;
2259 cachep->colour = left_over / cachep->colour_off;
2260 cachep->freelist_size = freelist_size;
2261 cachep->flags = flags;
2262 cachep->allocflags = __GFP_COMP;
2263 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2264 cachep->allocflags |= GFP_DMA;
2265 cachep->size = size;
2266 cachep->reciprocal_buffer_size = reciprocal_value(size);
2270 * If we're going to use the generic kernel_map_pages()
2271 * poisoning, then it's going to smash the contents of
2272 * the redzone and userword anyhow, so switch them off.
2274 if (IS_ENABLED(CONFIG_PAGE_POISONING) &&
2275 (cachep->flags & SLAB_POISON) &&
2276 is_debug_pagealloc_cache(cachep))
2277 cachep->flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2280 if (OFF_SLAB(cachep)) {
2281 cachep->freelist_cache = kmalloc_slab(freelist_size, 0u);
2283 * This is a possibility for one of the kmalloc_{dma,}_caches.
2284 * But since we go off slab only for object size greater than
2285 * OFF_SLAB_MIN_SIZE, and kmalloc_{dma,}_caches get created
2286 * in ascending order,this should not happen at all.
2287 * But leave a BUG_ON for some lucky dude.
2289 BUG_ON(ZERO_OR_NULL_PTR(cachep->freelist_cache));
2292 err = setup_cpu_cache(cachep, gfp);
2294 __kmem_cache_release(cachep);
2302 static void check_irq_off(void)
2304 BUG_ON(!irqs_disabled());
2307 static void check_irq_on(void)
2309 BUG_ON(irqs_disabled());
2312 static void check_spinlock_acquired(struct kmem_cache *cachep)
2316 assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
2320 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2324 assert_spin_locked(&get_node(cachep, node)->list_lock);
2329 #define check_irq_off() do { } while(0)
2330 #define check_irq_on() do { } while(0)
2331 #define check_spinlock_acquired(x) do { } while(0)
2332 #define check_spinlock_acquired_node(x, y) do { } while(0)
2335 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
2336 struct array_cache *ac,
2337 int force, int node);
2339 static void do_drain(void *arg)
2341 struct kmem_cache *cachep = arg;
2342 struct array_cache *ac;
2343 int node = numa_mem_id();
2344 struct kmem_cache_node *n;
2348 ac = cpu_cache_get(cachep);
2349 n = get_node(cachep, node);
2350 spin_lock(&n->list_lock);
2351 free_block(cachep, ac->entry, ac->avail, node, &list);
2352 spin_unlock(&n->list_lock);
2353 slabs_destroy(cachep, &list);
2357 static void drain_cpu_caches(struct kmem_cache *cachep)
2359 struct kmem_cache_node *n;
2362 on_each_cpu(do_drain, cachep, 1);
2364 for_each_kmem_cache_node(cachep, node, n)
2366 drain_alien_cache(cachep, n->alien);
2368 for_each_kmem_cache_node(cachep, node, n)
2369 drain_array(cachep, n, n->shared, 1, node);
2373 * Remove slabs from the list of free slabs.
2374 * Specify the number of slabs to drain in tofree.
2376 * Returns the actual number of slabs released.
2378 static int drain_freelist(struct kmem_cache *cache,
2379 struct kmem_cache_node *n, int tofree)
2381 struct list_head *p;
2386 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2388 spin_lock_irq(&n->list_lock);
2389 p = n->slabs_free.prev;
2390 if (p == &n->slabs_free) {
2391 spin_unlock_irq(&n->list_lock);
2395 page = list_entry(p, struct page, lru);
2396 list_del(&page->lru);
2398 * Safe to drop the lock. The slab is no longer linked
2401 n->free_objects -= cache->num;
2402 spin_unlock_irq(&n->list_lock);
2403 slab_destroy(cache, page);
2410 int __kmem_cache_shrink(struct kmem_cache *cachep, bool deactivate)
2414 struct kmem_cache_node *n;
2416 drain_cpu_caches(cachep);
2419 for_each_kmem_cache_node(cachep, node, n) {
2420 drain_freelist(cachep, n, slabs_tofree(cachep, n));
2422 ret += !list_empty(&n->slabs_full) ||
2423 !list_empty(&n->slabs_partial);
2425 return (ret ? 1 : 0);
2428 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2430 return __kmem_cache_shrink(cachep, false);
2433 void __kmem_cache_release(struct kmem_cache *cachep)
2436 struct kmem_cache_node *n;
2438 free_percpu(cachep->cpu_cache);
2440 /* NUMA: free the node structures */
2441 for_each_kmem_cache_node(cachep, i, n) {
2443 free_alien_cache(n->alien);
2445 cachep->node[i] = NULL;
2450 * Get the memory for a slab management obj.
2452 * For a slab cache when the slab descriptor is off-slab, the
2453 * slab descriptor can't come from the same cache which is being created,
2454 * Because if it is the case, that means we defer the creation of
2455 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2456 * And we eventually call down to __kmem_cache_create(), which
2457 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2458 * This is a "chicken-and-egg" problem.
2460 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2461 * which are all initialized during kmem_cache_init().
2463 static void *alloc_slabmgmt(struct kmem_cache *cachep,
2464 struct page *page, int colour_off,
2465 gfp_t local_flags, int nodeid)
2468 void *addr = page_address(page);
2470 if (OFF_SLAB(cachep)) {
2471 /* Slab management obj is off-slab. */
2472 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
2473 local_flags, nodeid);
2477 freelist = addr + colour_off;
2478 colour_off += cachep->freelist_size;
2481 page->s_mem = addr + colour_off;
2485 static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
2487 return ((freelist_idx_t *)page->freelist)[idx];
2490 static inline void set_free_obj(struct page *page,
2491 unsigned int idx, freelist_idx_t val)
2493 ((freelist_idx_t *)(page->freelist))[idx] = val;
2496 static void cache_init_objs(struct kmem_cache *cachep,
2501 for (i = 0; i < cachep->num; i++) {
2502 void *objp = index_to_obj(cachep, page, i);
2504 if (cachep->flags & SLAB_STORE_USER)
2505 *dbg_userword(cachep, objp) = NULL;
2507 if (cachep->flags & SLAB_RED_ZONE) {
2508 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2509 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2512 * Constructors are not allowed to allocate memory from the same
2513 * cache which they are a constructor for. Otherwise, deadlock.
2514 * They must also be threaded.
2516 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2517 cachep->ctor(objp + obj_offset(cachep));
2519 if (cachep->flags & SLAB_RED_ZONE) {
2520 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2521 slab_error(cachep, "constructor overwrote the"
2522 " end of an object");
2523 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2524 slab_error(cachep, "constructor overwrote the"
2525 " start of an object");
2527 /* need to poison the objs? */
2528 if (cachep->flags & SLAB_POISON) {
2529 poison_obj(cachep, objp, POISON_FREE);
2530 slab_kernel_map(cachep, objp, 0, 0);
2536 set_obj_status(page, i, OBJECT_FREE);
2537 set_free_obj(page, i, i);
2541 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2543 if (CONFIG_ZONE_DMA_FLAG) {
2544 if (flags & GFP_DMA)
2545 BUG_ON(!(cachep->allocflags & GFP_DMA));
2547 BUG_ON(cachep->allocflags & GFP_DMA);
2551 static void *slab_get_obj(struct kmem_cache *cachep, struct page *page)
2555 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
2559 if (cachep->flags & SLAB_STORE_USER)
2560 set_store_user_dirty(cachep);
2566 static void slab_put_obj(struct kmem_cache *cachep,
2567 struct page *page, void *objp)
2569 unsigned int objnr = obj_to_index(cachep, page, objp);
2573 /* Verify double free bug */
2574 for (i = page->active; i < cachep->num; i++) {
2575 if (get_free_obj(page, i) == objnr) {
2576 printk(KERN_ERR "slab: double free detected in cache "
2577 "'%s', objp %p\n", cachep->name, objp);
2583 set_free_obj(page, page->active, objnr);
2587 * Map pages beginning at addr to the given cache and slab. This is required
2588 * for the slab allocator to be able to lookup the cache and slab of a
2589 * virtual address for kfree, ksize, and slab debugging.
2591 static void slab_map_pages(struct kmem_cache *cache, struct page *page,
2594 page->slab_cache = cache;
2595 page->freelist = freelist;
2599 * Grow (by 1) the number of slabs within a cache. This is called by
2600 * kmem_cache_alloc() when there are no active objs left in a cache.
2602 static int cache_grow(struct kmem_cache *cachep,
2603 gfp_t flags, int nodeid, struct page *page)
2608 struct kmem_cache_node *n;
2611 * Be lazy and only check for valid flags here, keeping it out of the
2612 * critical path in kmem_cache_alloc().
2614 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
2615 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
2618 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2620 /* Take the node list lock to change the colour_next on this node */
2622 n = get_node(cachep, nodeid);
2623 spin_lock(&n->list_lock);
2625 /* Get colour for the slab, and cal the next value. */
2626 offset = n->colour_next;
2628 if (n->colour_next >= cachep->colour)
2630 spin_unlock(&n->list_lock);
2632 offset *= cachep->colour_off;
2634 if (gfpflags_allow_blocking(local_flags))
2638 * The test for missing atomic flag is performed here, rather than
2639 * the more obvious place, simply to reduce the critical path length
2640 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2641 * will eventually be caught here (where it matters).
2643 kmem_flagcheck(cachep, flags);
2646 * Get mem for the objs. Attempt to allocate a physical page from
2650 page = kmem_getpages(cachep, local_flags, nodeid);
2654 /* Get slab management. */
2655 freelist = alloc_slabmgmt(cachep, page, offset,
2656 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2660 slab_map_pages(cachep, page, freelist);
2662 cache_init_objs(cachep, page);
2664 if (gfpflags_allow_blocking(local_flags))
2665 local_irq_disable();
2667 spin_lock(&n->list_lock);
2669 /* Make slab active. */
2670 list_add_tail(&page->lru, &(n->slabs_free));
2671 STATS_INC_GROWN(cachep);
2672 n->free_objects += cachep->num;
2673 spin_unlock(&n->list_lock);
2676 kmem_freepages(cachep, page);
2678 if (gfpflags_allow_blocking(local_flags))
2679 local_irq_disable();
2686 * Perform extra freeing checks:
2687 * - detect bad pointers.
2688 * - POISON/RED_ZONE checking
2690 static void kfree_debugcheck(const void *objp)
2692 if (!virt_addr_valid(objp)) {
2693 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2694 (unsigned long)objp);
2699 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2701 unsigned long long redzone1, redzone2;
2703 redzone1 = *dbg_redzone1(cache, obj);
2704 redzone2 = *dbg_redzone2(cache, obj);
2709 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2712 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2713 slab_error(cache, "double free detected");
2715 slab_error(cache, "memory outside object was overwritten");
2717 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2718 obj, redzone1, redzone2);
2721 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2722 unsigned long caller)
2727 BUG_ON(virt_to_cache(objp) != cachep);
2729 objp -= obj_offset(cachep);
2730 kfree_debugcheck(objp);
2731 page = virt_to_head_page(objp);
2733 if (cachep->flags & SLAB_RED_ZONE) {
2734 verify_redzone_free(cachep, objp);
2735 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2736 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2738 if (cachep->flags & SLAB_STORE_USER) {
2739 set_store_user_dirty(cachep);
2740 *dbg_userword(cachep, objp) = (void *)caller;
2743 objnr = obj_to_index(cachep, page, objp);
2745 BUG_ON(objnr >= cachep->num);
2746 BUG_ON(objp != index_to_obj(cachep, page, objnr));
2748 set_obj_status(page, objnr, OBJECT_FREE);
2749 if (cachep->flags & SLAB_POISON) {
2750 poison_obj(cachep, objp, POISON_FREE);
2751 slab_kernel_map(cachep, objp, 0, caller);
2757 #define kfree_debugcheck(x) do { } while(0)
2758 #define cache_free_debugcheck(x,objp,z) (objp)
2761 static struct page *get_first_slab(struct kmem_cache_node *n)
2765 page = list_first_entry_or_null(&n->slabs_partial,
2768 n->free_touched = 1;
2769 page = list_first_entry_or_null(&n->slabs_free,
2776 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
2780 struct kmem_cache_node *n;
2781 struct array_cache *ac;
2785 node = numa_mem_id();
2786 if (unlikely(force_refill))
2789 ac = cpu_cache_get(cachep);
2790 batchcount = ac->batchcount;
2791 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2793 * If there was little recent activity on this cache, then
2794 * perform only a partial refill. Otherwise we could generate
2797 batchcount = BATCHREFILL_LIMIT;
2799 n = get_node(cachep, node);
2801 BUG_ON(ac->avail > 0 || !n);
2802 spin_lock(&n->list_lock);
2804 /* See if we can refill from the shared array */
2805 if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2806 n->shared->touched = 1;
2810 while (batchcount > 0) {
2812 /* Get slab alloc is to come from. */
2813 page = get_first_slab(n);
2817 check_spinlock_acquired(cachep);
2820 * The slab was either on partial or free list so
2821 * there must be at least one object available for
2824 BUG_ON(page->active >= cachep->num);
2826 while (page->active < cachep->num && batchcount--) {
2827 STATS_INC_ALLOCED(cachep);
2828 STATS_INC_ACTIVE(cachep);
2829 STATS_SET_HIGH(cachep);
2831 ac_put_obj(cachep, ac, slab_get_obj(cachep, page));
2834 /* move slabp to correct slabp list: */
2835 list_del(&page->lru);
2836 if (page->active == cachep->num)
2837 list_add(&page->lru, &n->slabs_full);
2839 list_add(&page->lru, &n->slabs_partial);
2843 n->free_objects -= ac->avail;
2845 spin_unlock(&n->list_lock);
2847 if (unlikely(!ac->avail)) {
2850 x = cache_grow(cachep, gfp_exact_node(flags), node, NULL);
2852 /* cache_grow can reenable interrupts, then ac could change. */
2853 ac = cpu_cache_get(cachep);
2854 node = numa_mem_id();
2856 /* no objects in sight? abort */
2857 if (!x && (ac->avail == 0 || force_refill))
2860 if (!ac->avail) /* objects refilled by interrupt? */
2865 return ac_get_obj(cachep, ac, flags, force_refill);
2868 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2871 might_sleep_if(gfpflags_allow_blocking(flags));
2873 kmem_flagcheck(cachep, flags);
2878 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2879 gfp_t flags, void *objp, unsigned long caller)
2885 if (cachep->flags & SLAB_POISON) {
2886 check_poison_obj(cachep, objp);
2887 slab_kernel_map(cachep, objp, 1, 0);
2888 poison_obj(cachep, objp, POISON_INUSE);
2890 if (cachep->flags & SLAB_STORE_USER)
2891 *dbg_userword(cachep, objp) = (void *)caller;
2893 if (cachep->flags & SLAB_RED_ZONE) {
2894 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2895 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2896 slab_error(cachep, "double free, or memory outside"
2897 " object was overwritten");
2899 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2900 objp, *dbg_redzone1(cachep, objp),
2901 *dbg_redzone2(cachep, objp));
2903 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2904 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2907 page = virt_to_head_page(objp);
2908 set_obj_status(page, obj_to_index(cachep, page, objp), OBJECT_ACTIVE);
2909 objp += obj_offset(cachep);
2910 if (cachep->ctor && cachep->flags & SLAB_POISON)
2912 if (ARCH_SLAB_MINALIGN &&
2913 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
2914 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
2915 objp, (int)ARCH_SLAB_MINALIGN);
2920 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2923 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
2926 struct array_cache *ac;
2927 bool force_refill = false;
2931 ac = cpu_cache_get(cachep);
2932 if (likely(ac->avail)) {
2934 objp = ac_get_obj(cachep, ac, flags, false);
2937 * Allow for the possibility all avail objects are not allowed
2938 * by the current flags
2941 STATS_INC_ALLOCHIT(cachep);
2944 force_refill = true;
2947 STATS_INC_ALLOCMISS(cachep);
2948 objp = cache_alloc_refill(cachep, flags, force_refill);
2950 * the 'ac' may be updated by cache_alloc_refill(),
2951 * and kmemleak_erase() requires its correct value.
2953 ac = cpu_cache_get(cachep);
2957 * To avoid a false negative, if an object that is in one of the
2958 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
2959 * treat the array pointers as a reference to the object.
2962 kmemleak_erase(&ac->entry[ac->avail]);
2968 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
2970 * If we are in_interrupt, then process context, including cpusets and
2971 * mempolicy, may not apply and should not be used for allocation policy.
2973 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
2975 int nid_alloc, nid_here;
2977 if (in_interrupt() || (flags & __GFP_THISNODE))
2979 nid_alloc = nid_here = numa_mem_id();
2980 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
2981 nid_alloc = cpuset_slab_spread_node();
2982 else if (current->mempolicy)
2983 nid_alloc = mempolicy_slab_node();
2984 if (nid_alloc != nid_here)
2985 return ____cache_alloc_node(cachep, flags, nid_alloc);
2990 * Fallback function if there was no memory available and no objects on a
2991 * certain node and fall back is permitted. First we scan all the
2992 * available node for available objects. If that fails then we
2993 * perform an allocation without specifying a node. This allows the page
2994 * allocator to do its reclaim / fallback magic. We then insert the
2995 * slab into the proper nodelist and then allocate from it.
2997 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
2999 struct zonelist *zonelist;
3003 enum zone_type high_zoneidx = gfp_zone(flags);
3006 unsigned int cpuset_mems_cookie;
3008 if (flags & __GFP_THISNODE)
3011 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3014 cpuset_mems_cookie = read_mems_allowed_begin();
3015 zonelist = node_zonelist(mempolicy_slab_node(), flags);
3019 * Look through allowed nodes for objects available
3020 * from existing per node queues.
3022 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3023 nid = zone_to_nid(zone);
3025 if (cpuset_zone_allowed(zone, flags) &&
3026 get_node(cache, nid) &&
3027 get_node(cache, nid)->free_objects) {
3028 obj = ____cache_alloc_node(cache,
3029 gfp_exact_node(flags), nid);
3037 * This allocation will be performed within the constraints
3038 * of the current cpuset / memory policy requirements.
3039 * We may trigger various forms of reclaim on the allowed
3040 * set and go into memory reserves if necessary.
3044 if (gfpflags_allow_blocking(local_flags))
3046 kmem_flagcheck(cache, flags);
3047 page = kmem_getpages(cache, local_flags, numa_mem_id());
3048 if (gfpflags_allow_blocking(local_flags))
3049 local_irq_disable();
3052 * Insert into the appropriate per node queues
3054 nid = page_to_nid(page);
3055 if (cache_grow(cache, flags, nid, page)) {
3056 obj = ____cache_alloc_node(cache,
3057 gfp_exact_node(flags), nid);
3060 * Another processor may allocate the
3061 * objects in the slab since we are
3062 * not holding any locks.
3066 /* cache_grow already freed obj */
3072 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
3078 * A interface to enable slab creation on nodeid
3080 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3084 struct kmem_cache_node *n;
3088 VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
3089 n = get_node(cachep, nodeid);
3094 spin_lock(&n->list_lock);
3095 page = get_first_slab(n);
3099 check_spinlock_acquired_node(cachep, nodeid);
3101 STATS_INC_NODEALLOCS(cachep);
3102 STATS_INC_ACTIVE(cachep);
3103 STATS_SET_HIGH(cachep);
3105 BUG_ON(page->active == cachep->num);
3107 obj = slab_get_obj(cachep, page);
3109 /* move slabp to correct slabp list: */
3110 list_del(&page->lru);
3112 if (page->active == cachep->num)
3113 list_add(&page->lru, &n->slabs_full);
3115 list_add(&page->lru, &n->slabs_partial);
3117 spin_unlock(&n->list_lock);
3121 spin_unlock(&n->list_lock);
3122 x = cache_grow(cachep, gfp_exact_node(flags), nodeid, NULL);
3126 return fallback_alloc(cachep, flags);
3132 static __always_inline void *
3133 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3134 unsigned long caller)
3136 unsigned long save_flags;
3138 int slab_node = numa_mem_id();
3140 flags &= gfp_allowed_mask;
3141 cachep = slab_pre_alloc_hook(cachep, flags);
3142 if (unlikely(!cachep))
3145 cache_alloc_debugcheck_before(cachep, flags);
3146 local_irq_save(save_flags);
3148 if (nodeid == NUMA_NO_NODE)
3151 if (unlikely(!get_node(cachep, nodeid))) {
3152 /* Node not bootstrapped yet */
3153 ptr = fallback_alloc(cachep, flags);
3157 if (nodeid == slab_node) {
3159 * Use the locally cached objects if possible.
3160 * However ____cache_alloc does not allow fallback
3161 * to other nodes. It may fail while we still have
3162 * objects on other nodes available.
3164 ptr = ____cache_alloc(cachep, flags);
3168 /* ___cache_alloc_node can fall back to other nodes */
3169 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3171 local_irq_restore(save_flags);
3172 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3174 if (unlikely(flags & __GFP_ZERO) && ptr)
3175 memset(ptr, 0, cachep->object_size);
3177 slab_post_alloc_hook(cachep, flags, 1, &ptr);
3181 static __always_inline void *
3182 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3186 if (current->mempolicy || cpuset_do_slab_mem_spread()) {
3187 objp = alternate_node_alloc(cache, flags);
3191 objp = ____cache_alloc(cache, flags);
3194 * We may just have run out of memory on the local node.
3195 * ____cache_alloc_node() knows how to locate memory on other nodes
3198 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3205 static __always_inline void *
3206 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3208 return ____cache_alloc(cachep, flags);
3211 #endif /* CONFIG_NUMA */
3213 static __always_inline void *
3214 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3216 unsigned long save_flags;
3219 flags &= gfp_allowed_mask;
3220 cachep = slab_pre_alloc_hook(cachep, flags);
3221 if (unlikely(!cachep))
3224 cache_alloc_debugcheck_before(cachep, flags);
3225 local_irq_save(save_flags);
3226 objp = __do_cache_alloc(cachep, flags);
3227 local_irq_restore(save_flags);
3228 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3231 if (unlikely(flags & __GFP_ZERO) && objp)
3232 memset(objp, 0, cachep->object_size);
3234 slab_post_alloc_hook(cachep, flags, 1, &objp);
3239 * Caller needs to acquire correct kmem_cache_node's list_lock
3240 * @list: List of detached free slabs should be freed by caller
3242 static void free_block(struct kmem_cache *cachep, void **objpp,
3243 int nr_objects, int node, struct list_head *list)
3246 struct kmem_cache_node *n = get_node(cachep, node);
3248 for (i = 0; i < nr_objects; i++) {
3252 clear_obj_pfmemalloc(&objpp[i]);
3255 page = virt_to_head_page(objp);
3256 list_del(&page->lru);
3257 check_spinlock_acquired_node(cachep, node);
3258 slab_put_obj(cachep, page, objp);
3259 STATS_DEC_ACTIVE(cachep);
3262 /* fixup slab chains */
3263 if (page->active == 0) {
3264 if (n->free_objects > n->free_limit) {
3265 n->free_objects -= cachep->num;
3266 list_add_tail(&page->lru, list);
3268 list_add(&page->lru, &n->slabs_free);
3271 /* Unconditionally move a slab to the end of the
3272 * partial list on free - maximum time for the
3273 * other objects to be freed, too.
3275 list_add_tail(&page->lru, &n->slabs_partial);
3280 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3283 struct kmem_cache_node *n;
3284 int node = numa_mem_id();
3287 batchcount = ac->batchcount;
3290 n = get_node(cachep, node);
3291 spin_lock(&n->list_lock);
3293 struct array_cache *shared_array = n->shared;
3294 int max = shared_array->limit - shared_array->avail;
3296 if (batchcount > max)
3298 memcpy(&(shared_array->entry[shared_array->avail]),
3299 ac->entry, sizeof(void *) * batchcount);
3300 shared_array->avail += batchcount;
3305 free_block(cachep, ac->entry, batchcount, node, &list);
3312 list_for_each_entry(page, &n->slabs_free, lru) {
3313 BUG_ON(page->active);
3317 STATS_SET_FREEABLE(cachep, i);
3320 spin_unlock(&n->list_lock);
3321 slabs_destroy(cachep, &list);
3322 ac->avail -= batchcount;
3323 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3327 * Release an obj back to its cache. If the obj has a constructed state, it must
3328 * be in this state _before_ it is released. Called with disabled ints.
3330 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3331 unsigned long caller)
3333 struct array_cache *ac = cpu_cache_get(cachep);
3336 kmemleak_free_recursive(objp, cachep->flags);
3337 objp = cache_free_debugcheck(cachep, objp, caller);
3339 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3342 * Skip calling cache_free_alien() when the platform is not numa.
3343 * This will avoid cache misses that happen while accessing slabp (which
3344 * is per page memory reference) to get nodeid. Instead use a global
3345 * variable to skip the call, which is mostly likely to be present in
3348 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3351 if (ac->avail < ac->limit) {
3352 STATS_INC_FREEHIT(cachep);
3354 STATS_INC_FREEMISS(cachep);
3355 cache_flusharray(cachep, ac);
3358 ac_put_obj(cachep, ac, objp);
3362 * kmem_cache_alloc - Allocate an object
3363 * @cachep: The cache to allocate from.
3364 * @flags: See kmalloc().
3366 * Allocate an object from this cache. The flags are only relevant
3367 * if the cache has no available objects.
3369 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3371 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3373 trace_kmem_cache_alloc(_RET_IP_, ret,
3374 cachep->object_size, cachep->size, flags);
3378 EXPORT_SYMBOL(kmem_cache_alloc);
3380 static __always_inline void
3381 cache_alloc_debugcheck_after_bulk(struct kmem_cache *s, gfp_t flags,
3382 size_t size, void **p, unsigned long caller)
3386 for (i = 0; i < size; i++)
3387 p[i] = cache_alloc_debugcheck_after(s, flags, p[i], caller);
3390 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3395 s = slab_pre_alloc_hook(s, flags);
3399 cache_alloc_debugcheck_before(s, flags);
3401 local_irq_disable();
3402 for (i = 0; i < size; i++) {
3403 void *objp = __do_cache_alloc(s, flags);
3405 if (unlikely(!objp))
3411 cache_alloc_debugcheck_after_bulk(s, flags, size, p, _RET_IP_);
3413 /* Clear memory outside IRQ disabled section */
3414 if (unlikely(flags & __GFP_ZERO))
3415 for (i = 0; i < size; i++)
3416 memset(p[i], 0, s->object_size);
3418 slab_post_alloc_hook(s, flags, size, p);
3419 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3423 cache_alloc_debugcheck_after_bulk(s, flags, i, p, _RET_IP_);
3424 slab_post_alloc_hook(s, flags, i, p);
3425 __kmem_cache_free_bulk(s, i, p);
3428 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3430 #ifdef CONFIG_TRACING
3432 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3436 ret = slab_alloc(cachep, flags, _RET_IP_);
3438 trace_kmalloc(_RET_IP_, ret,
3439 size, cachep->size, flags);
3442 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3447 * kmem_cache_alloc_node - Allocate an object on the specified node
3448 * @cachep: The cache to allocate from.
3449 * @flags: See kmalloc().
3450 * @nodeid: node number of the target node.
3452 * Identical to kmem_cache_alloc but it will allocate memory on the given
3453 * node, which can improve the performance for cpu bound structures.
3455 * Fallback to other node is possible if __GFP_THISNODE is not set.
3457 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3459 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3461 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3462 cachep->object_size, cachep->size,
3467 EXPORT_SYMBOL(kmem_cache_alloc_node);
3469 #ifdef CONFIG_TRACING
3470 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3477 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3479 trace_kmalloc_node(_RET_IP_, ret,
3484 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3487 static __always_inline void *
3488 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3490 struct kmem_cache *cachep;
3492 cachep = kmalloc_slab(size, flags);
3493 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3495 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
3498 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3500 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3502 EXPORT_SYMBOL(__kmalloc_node);
3504 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3505 int node, unsigned long caller)
3507 return __do_kmalloc_node(size, flags, node, caller);
3509 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3510 #endif /* CONFIG_NUMA */
3513 * __do_kmalloc - allocate memory
3514 * @size: how many bytes of memory are required.
3515 * @flags: the type of memory to allocate (see kmalloc).
3516 * @caller: function caller for debug tracking of the caller
3518 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3519 unsigned long caller)
3521 struct kmem_cache *cachep;
3524 cachep = kmalloc_slab(size, flags);
3525 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3527 ret = slab_alloc(cachep, flags, caller);
3529 trace_kmalloc(caller, ret,
3530 size, cachep->size, flags);
3535 void *__kmalloc(size_t size, gfp_t flags)
3537 return __do_kmalloc(size, flags, _RET_IP_);
3539 EXPORT_SYMBOL(__kmalloc);
3541 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3543 return __do_kmalloc(size, flags, caller);
3545 EXPORT_SYMBOL(__kmalloc_track_caller);
3548 * kmem_cache_free - Deallocate an object
3549 * @cachep: The cache the allocation was from.
3550 * @objp: The previously allocated object.
3552 * Free an object which was previously allocated from this
3555 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3557 unsigned long flags;
3558 cachep = cache_from_obj(cachep, objp);
3562 local_irq_save(flags);
3563 debug_check_no_locks_freed(objp, cachep->object_size);
3564 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3565 debug_check_no_obj_freed(objp, cachep->object_size);
3566 __cache_free(cachep, objp, _RET_IP_);
3567 local_irq_restore(flags);
3569 trace_kmem_cache_free(_RET_IP_, objp);
3571 EXPORT_SYMBOL(kmem_cache_free);
3573 void kmem_cache_free_bulk(struct kmem_cache *orig_s, size_t size, void **p)
3575 struct kmem_cache *s;
3578 local_irq_disable();
3579 for (i = 0; i < size; i++) {
3582 if (!orig_s) /* called via kfree_bulk */
3583 s = virt_to_cache(objp);
3585 s = cache_from_obj(orig_s, objp);
3587 debug_check_no_locks_freed(objp, s->object_size);
3588 if (!(s->flags & SLAB_DEBUG_OBJECTS))
3589 debug_check_no_obj_freed(objp, s->object_size);
3591 __cache_free(s, objp, _RET_IP_);
3595 /* FIXME: add tracing */
3597 EXPORT_SYMBOL(kmem_cache_free_bulk);
3600 * kfree - free previously allocated memory
3601 * @objp: pointer returned by kmalloc.
3603 * If @objp is NULL, no operation is performed.
3605 * Don't free memory not originally allocated by kmalloc()
3606 * or you will run into trouble.
3608 void kfree(const void *objp)
3610 struct kmem_cache *c;
3611 unsigned long flags;
3613 trace_kfree(_RET_IP_, objp);
3615 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3617 local_irq_save(flags);
3618 kfree_debugcheck(objp);
3619 c = virt_to_cache(objp);
3620 debug_check_no_locks_freed(objp, c->object_size);
3622 debug_check_no_obj_freed(objp, c->object_size);
3623 __cache_free(c, (void *)objp, _RET_IP_);
3624 local_irq_restore(flags);
3626 EXPORT_SYMBOL(kfree);
3629 * This initializes kmem_cache_node or resizes various caches for all nodes.
3631 static int alloc_kmem_cache_node(struct kmem_cache *cachep, gfp_t gfp)
3634 struct kmem_cache_node *n;
3635 struct array_cache *new_shared;
3636 struct alien_cache **new_alien = NULL;
3638 for_each_online_node(node) {
3640 if (use_alien_caches) {
3641 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3647 if (cachep->shared) {
3648 new_shared = alloc_arraycache(node,
3649 cachep->shared*cachep->batchcount,
3652 free_alien_cache(new_alien);
3657 n = get_node(cachep, node);
3659 struct array_cache *shared = n->shared;
3662 spin_lock_irq(&n->list_lock);
3665 free_block(cachep, shared->entry,
3666 shared->avail, node, &list);
3668 n->shared = new_shared;
3670 n->alien = new_alien;
3673 n->free_limit = (1 + nr_cpus_node(node)) *
3674 cachep->batchcount + cachep->num;
3675 spin_unlock_irq(&n->list_lock);
3676 slabs_destroy(cachep, &list);
3678 free_alien_cache(new_alien);
3681 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3683 free_alien_cache(new_alien);
3688 kmem_cache_node_init(n);
3689 n->next_reap = jiffies + REAPTIMEOUT_NODE +
3690 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
3691 n->shared = new_shared;
3692 n->alien = new_alien;
3693 n->free_limit = (1 + nr_cpus_node(node)) *
3694 cachep->batchcount + cachep->num;
3695 cachep->node[node] = n;
3700 if (!cachep->list.next) {
3701 /* Cache is not active yet. Roll back what we did */
3704 n = get_node(cachep, node);
3707 free_alien_cache(n->alien);
3709 cachep->node[node] = NULL;
3717 /* Always called with the slab_mutex held */
3718 static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
3719 int batchcount, int shared, gfp_t gfp)
3721 struct array_cache __percpu *cpu_cache, *prev;
3724 cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3728 prev = cachep->cpu_cache;
3729 cachep->cpu_cache = cpu_cache;
3730 kick_all_cpus_sync();
3733 cachep->batchcount = batchcount;
3734 cachep->limit = limit;
3735 cachep->shared = shared;
3740 for_each_online_cpu(cpu) {
3743 struct kmem_cache_node *n;
3744 struct array_cache *ac = per_cpu_ptr(prev, cpu);
3746 node = cpu_to_mem(cpu);
3747 n = get_node(cachep, node);
3748 spin_lock_irq(&n->list_lock);
3749 free_block(cachep, ac->entry, ac->avail, node, &list);
3750 spin_unlock_irq(&n->list_lock);
3751 slabs_destroy(cachep, &list);
3756 return alloc_kmem_cache_node(cachep, gfp);
3759 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3760 int batchcount, int shared, gfp_t gfp)
3763 struct kmem_cache *c;
3765 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3767 if (slab_state < FULL)
3770 if ((ret < 0) || !is_root_cache(cachep))
3773 lockdep_assert_held(&slab_mutex);
3774 for_each_memcg_cache(c, cachep) {
3775 /* return value determined by the root cache only */
3776 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3782 /* Called with slab_mutex held always */
3783 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3790 if (!is_root_cache(cachep)) {
3791 struct kmem_cache *root = memcg_root_cache(cachep);
3792 limit = root->limit;
3793 shared = root->shared;
3794 batchcount = root->batchcount;
3797 if (limit && shared && batchcount)
3800 * The head array serves three purposes:
3801 * - create a LIFO ordering, i.e. return objects that are cache-warm
3802 * - reduce the number of spinlock operations.
3803 * - reduce the number of linked list operations on the slab and
3804 * bufctl chains: array operations are cheaper.
3805 * The numbers are guessed, we should auto-tune as described by
3808 if (cachep->size > 131072)
3810 else if (cachep->size > PAGE_SIZE)
3812 else if (cachep->size > 1024)
3814 else if (cachep->size > 256)
3820 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3821 * allocation behaviour: Most allocs on one cpu, most free operations
3822 * on another cpu. For these cases, an efficient object passing between
3823 * cpus is necessary. This is provided by a shared array. The array
3824 * replaces Bonwick's magazine layer.
3825 * On uniprocessor, it's functionally equivalent (but less efficient)
3826 * to a larger limit. Thus disabled by default.
3829 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
3834 * With debugging enabled, large batchcount lead to excessively long
3835 * periods with disabled local interrupts. Limit the batchcount
3840 batchcount = (limit + 1) / 2;
3842 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3844 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3845 cachep->name, -err);
3850 * Drain an array if it contains any elements taking the node lock only if
3851 * necessary. Note that the node listlock also protects the array_cache
3852 * if drain_array() is used on the shared array.
3854 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
3855 struct array_cache *ac, int force, int node)
3860 if (!ac || !ac->avail)
3862 if (ac->touched && !force) {
3865 spin_lock_irq(&n->list_lock);
3867 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3868 if (tofree > ac->avail)
3869 tofree = (ac->avail + 1) / 2;
3870 free_block(cachep, ac->entry, tofree, node, &list);
3871 ac->avail -= tofree;
3872 memmove(ac->entry, &(ac->entry[tofree]),
3873 sizeof(void *) * ac->avail);
3875 spin_unlock_irq(&n->list_lock);
3876 slabs_destroy(cachep, &list);
3881 * cache_reap - Reclaim memory from caches.
3882 * @w: work descriptor
3884 * Called from workqueue/eventd every few seconds.
3886 * - clear the per-cpu caches for this CPU.
3887 * - return freeable pages to the main free memory pool.
3889 * If we cannot acquire the cache chain mutex then just give up - we'll try
3890 * again on the next iteration.
3892 static void cache_reap(struct work_struct *w)
3894 struct kmem_cache *searchp;
3895 struct kmem_cache_node *n;
3896 int node = numa_mem_id();
3897 struct delayed_work *work = to_delayed_work(w);
3899 if (!mutex_trylock(&slab_mutex))
3900 /* Give up. Setup the next iteration. */
3903 list_for_each_entry(searchp, &slab_caches, list) {
3907 * We only take the node lock if absolutely necessary and we
3908 * have established with reasonable certainty that
3909 * we can do some work if the lock was obtained.
3911 n = get_node(searchp, node);
3913 reap_alien(searchp, n);
3915 drain_array(searchp, n, cpu_cache_get(searchp), 0, node);
3918 * These are racy checks but it does not matter
3919 * if we skip one check or scan twice.
3921 if (time_after(n->next_reap, jiffies))
3924 n->next_reap = jiffies + REAPTIMEOUT_NODE;
3926 drain_array(searchp, n, n->shared, 0, node);
3928 if (n->free_touched)
3929 n->free_touched = 0;
3933 freed = drain_freelist(searchp, n, (n->free_limit +
3934 5 * searchp->num - 1) / (5 * searchp->num));
3935 STATS_ADD_REAPED(searchp, freed);
3941 mutex_unlock(&slab_mutex);
3944 /* Set up the next iteration */
3945 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC));
3948 #ifdef CONFIG_SLABINFO
3949 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
3952 unsigned long active_objs;
3953 unsigned long num_objs;
3954 unsigned long active_slabs = 0;
3955 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3959 struct kmem_cache_node *n;
3963 for_each_kmem_cache_node(cachep, node, n) {
3966 spin_lock_irq(&n->list_lock);
3968 list_for_each_entry(page, &n->slabs_full, lru) {
3969 if (page->active != cachep->num && !error)
3970 error = "slabs_full accounting error";
3971 active_objs += cachep->num;
3974 list_for_each_entry(page, &n->slabs_partial, lru) {
3975 if (page->active == cachep->num && !error)
3976 error = "slabs_partial accounting error";
3977 if (!page->active && !error)
3978 error = "slabs_partial accounting error";
3979 active_objs += page->active;
3982 list_for_each_entry(page, &n->slabs_free, lru) {
3983 if (page->active && !error)
3984 error = "slabs_free accounting error";
3987 free_objects += n->free_objects;
3989 shared_avail += n->shared->avail;
3991 spin_unlock_irq(&n->list_lock);
3993 num_slabs += active_slabs;
3994 num_objs = num_slabs * cachep->num;
3995 if (num_objs - active_objs != free_objects && !error)
3996 error = "free_objects accounting error";
3998 name = cachep->name;
4000 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4002 sinfo->active_objs = active_objs;
4003 sinfo->num_objs = num_objs;
4004 sinfo->active_slabs = active_slabs;
4005 sinfo->num_slabs = num_slabs;
4006 sinfo->shared_avail = shared_avail;
4007 sinfo->limit = cachep->limit;
4008 sinfo->batchcount = cachep->batchcount;
4009 sinfo->shared = cachep->shared;
4010 sinfo->objects_per_slab = cachep->num;
4011 sinfo->cache_order = cachep->gfporder;
4014 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4018 unsigned long high = cachep->high_mark;
4019 unsigned long allocs = cachep->num_allocations;
4020 unsigned long grown = cachep->grown;
4021 unsigned long reaped = cachep->reaped;
4022 unsigned long errors = cachep->errors;
4023 unsigned long max_freeable = cachep->max_freeable;
4024 unsigned long node_allocs = cachep->node_allocs;
4025 unsigned long node_frees = cachep->node_frees;
4026 unsigned long overflows = cachep->node_overflow;
4028 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4029 "%4lu %4lu %4lu %4lu %4lu",
4030 allocs, high, grown,
4031 reaped, errors, max_freeable, node_allocs,
4032 node_frees, overflows);
4036 unsigned long allochit = atomic_read(&cachep->allochit);
4037 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4038 unsigned long freehit = atomic_read(&cachep->freehit);
4039 unsigned long freemiss = atomic_read(&cachep->freemiss);
4041 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4042 allochit, allocmiss, freehit, freemiss);
4047 #define MAX_SLABINFO_WRITE 128
4049 * slabinfo_write - Tuning for the slab allocator
4051 * @buffer: user buffer
4052 * @count: data length
4055 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4056 size_t count, loff_t *ppos)
4058 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4059 int limit, batchcount, shared, res;
4060 struct kmem_cache *cachep;
4062 if (count > MAX_SLABINFO_WRITE)
4064 if (copy_from_user(&kbuf, buffer, count))
4066 kbuf[MAX_SLABINFO_WRITE] = '\0';
4068 tmp = strchr(kbuf, ' ');
4073 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4076 /* Find the cache in the chain of caches. */
4077 mutex_lock(&slab_mutex);
4079 list_for_each_entry(cachep, &slab_caches, list) {
4080 if (!strcmp(cachep->name, kbuf)) {
4081 if (limit < 1 || batchcount < 1 ||
4082 batchcount > limit || shared < 0) {
4085 res = do_tune_cpucache(cachep, limit,
4092 mutex_unlock(&slab_mutex);
4098 #ifdef CONFIG_DEBUG_SLAB_LEAK
4100 static inline int add_caller(unsigned long *n, unsigned long v)
4110 unsigned long *q = p + 2 * i;
4124 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4130 static void handle_slab(unsigned long *n, struct kmem_cache *c,
4139 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
4142 for (j = page->active; j < c->num; j++) {
4143 if (get_free_obj(page, j) == i) {
4153 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4154 * mapping is established when actual object allocation and
4155 * we could mistakenly access the unmapped object in the cpu
4158 if (probe_kernel_read(&v, dbg_userword(c, p), sizeof(v)))
4161 if (!add_caller(n, v))
4166 static void show_symbol(struct seq_file *m, unsigned long address)
4168 #ifdef CONFIG_KALLSYMS
4169 unsigned long offset, size;
4170 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4172 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4173 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4175 seq_printf(m, " [%s]", modname);
4179 seq_printf(m, "%p", (void *)address);
4182 static int leaks_show(struct seq_file *m, void *p)
4184 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4186 struct kmem_cache_node *n;
4188 unsigned long *x = m->private;
4192 if (!(cachep->flags & SLAB_STORE_USER))
4194 if (!(cachep->flags & SLAB_RED_ZONE))
4198 * Set store_user_clean and start to grab stored user information
4199 * for all objects on this cache. If some alloc/free requests comes
4200 * during the processing, information would be wrong so restart
4204 set_store_user_clean(cachep);
4205 drain_cpu_caches(cachep);
4209 for_each_kmem_cache_node(cachep, node, n) {
4212 spin_lock_irq(&n->list_lock);
4214 list_for_each_entry(page, &n->slabs_full, lru)
4215 handle_slab(x, cachep, page);
4216 list_for_each_entry(page, &n->slabs_partial, lru)
4217 handle_slab(x, cachep, page);
4218 spin_unlock_irq(&n->list_lock);
4220 } while (!is_store_user_clean(cachep));
4222 name = cachep->name;
4224 /* Increase the buffer size */
4225 mutex_unlock(&slab_mutex);
4226 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4228 /* Too bad, we are really out */
4230 mutex_lock(&slab_mutex);
4233 *(unsigned long *)m->private = x[0] * 2;
4235 mutex_lock(&slab_mutex);
4236 /* Now make sure this entry will be retried */
4240 for (i = 0; i < x[1]; i++) {
4241 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4242 show_symbol(m, x[2*i+2]);
4249 static const struct seq_operations slabstats_op = {
4250 .start = slab_start,
4256 static int slabstats_open(struct inode *inode, struct file *file)
4260 n = __seq_open_private(file, &slabstats_op, PAGE_SIZE);
4264 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4269 static const struct file_operations proc_slabstats_operations = {
4270 .open = slabstats_open,
4272 .llseek = seq_lseek,
4273 .release = seq_release_private,
4277 static int __init slab_proc_init(void)
4279 #ifdef CONFIG_DEBUG_SLAB_LEAK
4280 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4284 module_init(slab_proc_init);
4288 * ksize - get the actual amount of memory allocated for a given object
4289 * @objp: Pointer to the object
4291 * kmalloc may internally round up allocations and return more memory
4292 * than requested. ksize() can be used to determine the actual amount of
4293 * memory allocated. The caller may use this additional memory, even though
4294 * a smaller amount of memory was initially specified with the kmalloc call.
4295 * The caller must guarantee that objp points to a valid object previously
4296 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4297 * must not be freed during the duration of the call.
4299 size_t ksize(const void *objp)
4302 if (unlikely(objp == ZERO_SIZE_PTR))
4305 return virt_to_cache(objp)->object_size;
4307 EXPORT_SYMBOL(ksize);