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 #ifdef CONFIG_DEBUG_SLAB_LEAK
385 static inline bool is_store_user_clean(struct kmem_cache *cachep)
387 return atomic_read(&cachep->store_user_clean) == 1;
390 static inline void set_store_user_clean(struct kmem_cache *cachep)
392 atomic_set(&cachep->store_user_clean, 1);
395 static inline void set_store_user_dirty(struct kmem_cache *cachep)
397 if (is_store_user_clean(cachep))
398 atomic_set(&cachep->store_user_clean, 0);
402 static inline void set_store_user_dirty(struct kmem_cache *cachep) {}
407 * Do not go above this order unless 0 objects fit into the slab or
408 * overridden on the command line.
410 #define SLAB_MAX_ORDER_HI 1
411 #define SLAB_MAX_ORDER_LO 0
412 static int slab_max_order = SLAB_MAX_ORDER_LO;
413 static bool slab_max_order_set __initdata;
415 static inline struct kmem_cache *virt_to_cache(const void *obj)
417 struct page *page = virt_to_head_page(obj);
418 return page->slab_cache;
421 static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
424 return page->s_mem + cache->size * idx;
428 * We want to avoid an expensive divide : (offset / cache->size)
429 * Using the fact that size is a constant for a particular cache,
430 * we can replace (offset / cache->size) by
431 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
433 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
434 const struct page *page, void *obj)
436 u32 offset = (obj - page->s_mem);
437 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
440 #define BOOT_CPUCACHE_ENTRIES 1
441 /* internal cache of cache description objs */
442 static struct kmem_cache kmem_cache_boot = {
444 .limit = BOOT_CPUCACHE_ENTRIES,
446 .size = sizeof(struct kmem_cache),
447 .name = "kmem_cache",
450 #define BAD_ALIEN_MAGIC 0x01020304ul
452 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
454 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
456 return this_cpu_ptr(cachep->cpu_cache);
459 static size_t calculate_freelist_size(int nr_objs, size_t align)
461 size_t freelist_size;
463 freelist_size = nr_objs * sizeof(freelist_idx_t);
465 freelist_size = ALIGN(freelist_size, align);
467 return freelist_size;
470 static int calculate_nr_objs(size_t slab_size, size_t buffer_size,
471 size_t idx_size, size_t align)
474 size_t remained_size;
475 size_t freelist_size;
478 * Ignore padding for the initial guess. The padding
479 * is at most @align-1 bytes, and @buffer_size is at
480 * least @align. In the worst case, this result will
481 * be one greater than the number of objects that fit
482 * into the memory allocation when taking the padding
485 nr_objs = slab_size / (buffer_size + idx_size);
488 * This calculated number will be either the right
489 * amount, or one greater than what we want.
491 remained_size = slab_size - nr_objs * buffer_size;
492 freelist_size = calculate_freelist_size(nr_objs, align);
493 if (remained_size < freelist_size)
500 * Calculate the number of objects and left-over bytes for a given buffer size.
502 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
503 size_t align, int flags, size_t *left_over,
508 size_t slab_size = PAGE_SIZE << gfporder;
511 * The slab management structure can be either off the slab or
512 * on it. For the latter case, the memory allocated for a
515 * - One freelist_idx_t for each object
516 * - Padding to respect alignment of @align
517 * - @buffer_size bytes for each object
519 * If the slab management structure is off the slab, then the
520 * alignment will already be calculated into the size. Because
521 * the slabs are all pages aligned, the objects will be at the
522 * correct alignment when allocated.
524 if (flags & CFLGS_OFF_SLAB) {
526 nr_objs = slab_size / buffer_size;
529 nr_objs = calculate_nr_objs(slab_size, buffer_size,
530 sizeof(freelist_idx_t), align);
531 mgmt_size = calculate_freelist_size(nr_objs, align);
534 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
538 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
540 static void __slab_error(const char *function, struct kmem_cache *cachep,
543 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
544 function, cachep->name, msg);
546 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
551 * By default on NUMA we use alien caches to stage the freeing of
552 * objects allocated from other nodes. This causes massive memory
553 * inefficiencies when using fake NUMA setup to split memory into a
554 * large number of small nodes, so it can be disabled on the command
558 static int use_alien_caches __read_mostly = 1;
559 static int __init noaliencache_setup(char *s)
561 use_alien_caches = 0;
564 __setup("noaliencache", noaliencache_setup);
566 static int __init slab_max_order_setup(char *str)
568 get_option(&str, &slab_max_order);
569 slab_max_order = slab_max_order < 0 ? 0 :
570 min(slab_max_order, MAX_ORDER - 1);
571 slab_max_order_set = true;
575 __setup("slab_max_order=", slab_max_order_setup);
579 * Special reaping functions for NUMA systems called from cache_reap().
580 * These take care of doing round robin flushing of alien caches (containing
581 * objects freed on different nodes from which they were allocated) and the
582 * flushing of remote pcps by calling drain_node_pages.
584 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
586 static void init_reap_node(int cpu)
590 node = next_node(cpu_to_mem(cpu), node_online_map);
591 if (node == MAX_NUMNODES)
592 node = first_node(node_online_map);
594 per_cpu(slab_reap_node, cpu) = node;
597 static void next_reap_node(void)
599 int node = __this_cpu_read(slab_reap_node);
601 node = next_node(node, node_online_map);
602 if (unlikely(node >= MAX_NUMNODES))
603 node = first_node(node_online_map);
604 __this_cpu_write(slab_reap_node, node);
608 #define init_reap_node(cpu) do { } while (0)
609 #define next_reap_node(void) do { } while (0)
613 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
614 * via the workqueue/eventd.
615 * Add the CPU number into the expiration time to minimize the possibility of
616 * the CPUs getting into lockstep and contending for the global cache chain
619 static void start_cpu_timer(int cpu)
621 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
624 * When this gets called from do_initcalls via cpucache_init(),
625 * init_workqueues() has already run, so keventd will be setup
628 if (keventd_up() && reap_work->work.func == NULL) {
630 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
631 schedule_delayed_work_on(cpu, reap_work,
632 __round_jiffies_relative(HZ, cpu));
636 static void init_arraycache(struct array_cache *ac, int limit, int batch)
639 * The array_cache structures contain pointers to free object.
640 * However, when such objects are allocated or transferred to another
641 * cache the pointers are not cleared and they could be counted as
642 * valid references during a kmemleak scan. Therefore, kmemleak must
643 * not scan such objects.
645 kmemleak_no_scan(ac);
649 ac->batchcount = batch;
654 static struct array_cache *alloc_arraycache(int node, int entries,
655 int batchcount, gfp_t gfp)
657 size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache);
658 struct array_cache *ac = NULL;
660 ac = kmalloc_node(memsize, gfp, node);
661 init_arraycache(ac, entries, batchcount);
665 static inline bool is_slab_pfmemalloc(struct page *page)
667 return PageSlabPfmemalloc(page);
670 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
671 static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
672 struct array_cache *ac)
674 struct kmem_cache_node *n = get_node(cachep, numa_mem_id());
678 if (!pfmemalloc_active)
681 spin_lock_irqsave(&n->list_lock, flags);
682 list_for_each_entry(page, &n->slabs_full, lru)
683 if (is_slab_pfmemalloc(page))
686 list_for_each_entry(page, &n->slabs_partial, lru)
687 if (is_slab_pfmemalloc(page))
690 list_for_each_entry(page, &n->slabs_free, lru)
691 if (is_slab_pfmemalloc(page))
694 pfmemalloc_active = false;
696 spin_unlock_irqrestore(&n->list_lock, flags);
699 static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
700 gfp_t flags, bool force_refill)
703 void *objp = ac->entry[--ac->avail];
705 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
706 if (unlikely(is_obj_pfmemalloc(objp))) {
707 struct kmem_cache_node *n;
709 if (gfp_pfmemalloc_allowed(flags)) {
710 clear_obj_pfmemalloc(&objp);
714 /* The caller cannot use PFMEMALLOC objects, find another one */
715 for (i = 0; i < ac->avail; i++) {
716 /* If a !PFMEMALLOC object is found, swap them */
717 if (!is_obj_pfmemalloc(ac->entry[i])) {
719 ac->entry[i] = ac->entry[ac->avail];
720 ac->entry[ac->avail] = objp;
726 * If there are empty slabs on the slabs_free list and we are
727 * being forced to refill the cache, mark this one !pfmemalloc.
729 n = get_node(cachep, numa_mem_id());
730 if (!list_empty(&n->slabs_free) && force_refill) {
731 struct page *page = virt_to_head_page(objp);
732 ClearPageSlabPfmemalloc(page);
733 clear_obj_pfmemalloc(&objp);
734 recheck_pfmemalloc_active(cachep, ac);
738 /* No !PFMEMALLOC objects available */
746 static inline void *ac_get_obj(struct kmem_cache *cachep,
747 struct array_cache *ac, gfp_t flags, bool force_refill)
751 if (unlikely(sk_memalloc_socks()))
752 objp = __ac_get_obj(cachep, ac, flags, force_refill);
754 objp = ac->entry[--ac->avail];
759 static noinline void *__ac_put_obj(struct kmem_cache *cachep,
760 struct array_cache *ac, void *objp)
762 if (unlikely(pfmemalloc_active)) {
763 /* Some pfmemalloc slabs exist, check if this is one */
764 struct page *page = virt_to_head_page(objp);
765 if (PageSlabPfmemalloc(page))
766 set_obj_pfmemalloc(&objp);
772 static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
775 if (unlikely(sk_memalloc_socks()))
776 objp = __ac_put_obj(cachep, ac, objp);
778 ac->entry[ac->avail++] = objp;
782 * Transfer objects in one arraycache to another.
783 * Locking must be handled by the caller.
785 * Return the number of entries transferred.
787 static int transfer_objects(struct array_cache *to,
788 struct array_cache *from, unsigned int max)
790 /* Figure out how many entries to transfer */
791 int nr = min3(from->avail, max, to->limit - to->avail);
796 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
806 #define drain_alien_cache(cachep, alien) do { } while (0)
807 #define reap_alien(cachep, n) do { } while (0)
809 static inline struct alien_cache **alloc_alien_cache(int node,
810 int limit, gfp_t gfp)
812 return (struct alien_cache **)BAD_ALIEN_MAGIC;
815 static inline void free_alien_cache(struct alien_cache **ac_ptr)
819 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
824 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
830 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
831 gfp_t flags, int nodeid)
836 static inline gfp_t gfp_exact_node(gfp_t flags)
841 #else /* CONFIG_NUMA */
843 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
844 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
846 static struct alien_cache *__alloc_alien_cache(int node, int entries,
847 int batch, gfp_t gfp)
849 size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
850 struct alien_cache *alc = NULL;
852 alc = kmalloc_node(memsize, gfp, node);
853 init_arraycache(&alc->ac, entries, batch);
854 spin_lock_init(&alc->lock);
858 static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
860 struct alien_cache **alc_ptr;
861 size_t memsize = sizeof(void *) * nr_node_ids;
866 alc_ptr = kzalloc_node(memsize, gfp, node);
871 if (i == node || !node_online(i))
873 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
875 for (i--; i >= 0; i--)
884 static void free_alien_cache(struct alien_cache **alc_ptr)
895 static void __drain_alien_cache(struct kmem_cache *cachep,
896 struct array_cache *ac, int node,
897 struct list_head *list)
899 struct kmem_cache_node *n = get_node(cachep, node);
902 spin_lock(&n->list_lock);
904 * Stuff objects into the remote nodes shared array first.
905 * That way we could avoid the overhead of putting the objects
906 * into the free lists and getting them back later.
909 transfer_objects(n->shared, ac, ac->limit);
911 free_block(cachep, ac->entry, ac->avail, node, list);
913 spin_unlock(&n->list_lock);
918 * Called from cache_reap() to regularly drain alien caches round robin.
920 static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
922 int node = __this_cpu_read(slab_reap_node);
925 struct alien_cache *alc = n->alien[node];
926 struct array_cache *ac;
930 if (ac->avail && spin_trylock_irq(&alc->lock)) {
933 __drain_alien_cache(cachep, ac, node, &list);
934 spin_unlock_irq(&alc->lock);
935 slabs_destroy(cachep, &list);
941 static void drain_alien_cache(struct kmem_cache *cachep,
942 struct alien_cache **alien)
945 struct alien_cache *alc;
946 struct array_cache *ac;
949 for_each_online_node(i) {
955 spin_lock_irqsave(&alc->lock, flags);
956 __drain_alien_cache(cachep, ac, i, &list);
957 spin_unlock_irqrestore(&alc->lock, flags);
958 slabs_destroy(cachep, &list);
963 static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
964 int node, int page_node)
966 struct kmem_cache_node *n;
967 struct alien_cache *alien = NULL;
968 struct array_cache *ac;
971 n = get_node(cachep, node);
972 STATS_INC_NODEFREES(cachep);
973 if (n->alien && n->alien[page_node]) {
974 alien = n->alien[page_node];
976 spin_lock(&alien->lock);
977 if (unlikely(ac->avail == ac->limit)) {
978 STATS_INC_ACOVERFLOW(cachep);
979 __drain_alien_cache(cachep, ac, page_node, &list);
981 ac_put_obj(cachep, ac, objp);
982 spin_unlock(&alien->lock);
983 slabs_destroy(cachep, &list);
985 n = get_node(cachep, page_node);
986 spin_lock(&n->list_lock);
987 free_block(cachep, &objp, 1, page_node, &list);
988 spin_unlock(&n->list_lock);
989 slabs_destroy(cachep, &list);
994 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
996 int page_node = page_to_nid(virt_to_page(objp));
997 int node = numa_mem_id();
999 * Make sure we are not freeing a object from another node to the array
1000 * cache on this cpu.
1002 if (likely(node == page_node))
1005 return __cache_free_alien(cachep, objp, node, page_node);
1009 * Construct gfp mask to allocate from a specific node but do not direct reclaim
1010 * or warn about failures. kswapd may still wake to reclaim in the background.
1012 static inline gfp_t gfp_exact_node(gfp_t flags)
1014 return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~__GFP_DIRECT_RECLAIM;
1019 * Allocates and initializes node for a node on each slab cache, used for
1020 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
1021 * will be allocated off-node since memory is not yet online for the new node.
1022 * When hotplugging memory or a cpu, existing node are not replaced if
1025 * Must hold slab_mutex.
1027 static int init_cache_node_node(int node)
1029 struct kmem_cache *cachep;
1030 struct kmem_cache_node *n;
1031 const size_t memsize = sizeof(struct kmem_cache_node);
1033 list_for_each_entry(cachep, &slab_caches, list) {
1035 * Set up the kmem_cache_node for cpu before we can
1036 * begin anything. Make sure some other cpu on this
1037 * node has not already allocated this
1039 n = get_node(cachep, node);
1041 n = kmalloc_node(memsize, GFP_KERNEL, node);
1044 kmem_cache_node_init(n);
1045 n->next_reap = jiffies + REAPTIMEOUT_NODE +
1046 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1049 * The kmem_cache_nodes don't come and go as CPUs
1050 * come and go. slab_mutex is sufficient
1053 cachep->node[node] = n;
1056 spin_lock_irq(&n->list_lock);
1058 (1 + nr_cpus_node(node)) *
1059 cachep->batchcount + cachep->num;
1060 spin_unlock_irq(&n->list_lock);
1065 static inline int slabs_tofree(struct kmem_cache *cachep,
1066 struct kmem_cache_node *n)
1068 return (n->free_objects + cachep->num - 1) / cachep->num;
1071 static void cpuup_canceled(long cpu)
1073 struct kmem_cache *cachep;
1074 struct kmem_cache_node *n = NULL;
1075 int node = cpu_to_mem(cpu);
1076 const struct cpumask *mask = cpumask_of_node(node);
1078 list_for_each_entry(cachep, &slab_caches, list) {
1079 struct array_cache *nc;
1080 struct array_cache *shared;
1081 struct alien_cache **alien;
1084 n = get_node(cachep, node);
1088 spin_lock_irq(&n->list_lock);
1090 /* Free limit for this kmem_cache_node */
1091 n->free_limit -= cachep->batchcount;
1093 /* cpu is dead; no one can alloc from it. */
1094 nc = per_cpu_ptr(cachep->cpu_cache, cpu);
1096 free_block(cachep, nc->entry, nc->avail, node, &list);
1100 if (!cpumask_empty(mask)) {
1101 spin_unlock_irq(&n->list_lock);
1107 free_block(cachep, shared->entry,
1108 shared->avail, node, &list);
1115 spin_unlock_irq(&n->list_lock);
1119 drain_alien_cache(cachep, alien);
1120 free_alien_cache(alien);
1124 slabs_destroy(cachep, &list);
1127 * In the previous loop, all the objects were freed to
1128 * the respective cache's slabs, now we can go ahead and
1129 * shrink each nodelist to its limit.
1131 list_for_each_entry(cachep, &slab_caches, list) {
1132 n = get_node(cachep, node);
1135 drain_freelist(cachep, n, slabs_tofree(cachep, n));
1139 static int cpuup_prepare(long cpu)
1141 struct kmem_cache *cachep;
1142 struct kmem_cache_node *n = NULL;
1143 int node = cpu_to_mem(cpu);
1147 * We need to do this right in the beginning since
1148 * alloc_arraycache's are going to use this list.
1149 * kmalloc_node allows us to add the slab to the right
1150 * kmem_cache_node and not this cpu's kmem_cache_node
1152 err = init_cache_node_node(node);
1157 * Now we can go ahead with allocating the shared arrays and
1160 list_for_each_entry(cachep, &slab_caches, list) {
1161 struct array_cache *shared = NULL;
1162 struct alien_cache **alien = NULL;
1164 if (cachep->shared) {
1165 shared = alloc_arraycache(node,
1166 cachep->shared * cachep->batchcount,
1167 0xbaadf00d, GFP_KERNEL);
1171 if (use_alien_caches) {
1172 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1178 n = get_node(cachep, node);
1181 spin_lock_irq(&n->list_lock);
1184 * We are serialised from CPU_DEAD or
1185 * CPU_UP_CANCELLED by the cpucontrol lock
1196 spin_unlock_irq(&n->list_lock);
1198 free_alien_cache(alien);
1203 cpuup_canceled(cpu);
1207 static int cpuup_callback(struct notifier_block *nfb,
1208 unsigned long action, void *hcpu)
1210 long cpu = (long)hcpu;
1214 case CPU_UP_PREPARE:
1215 case CPU_UP_PREPARE_FROZEN:
1216 mutex_lock(&slab_mutex);
1217 err = cpuup_prepare(cpu);
1218 mutex_unlock(&slab_mutex);
1221 case CPU_ONLINE_FROZEN:
1222 start_cpu_timer(cpu);
1224 #ifdef CONFIG_HOTPLUG_CPU
1225 case CPU_DOWN_PREPARE:
1226 case CPU_DOWN_PREPARE_FROZEN:
1228 * Shutdown cache reaper. Note that the slab_mutex is
1229 * held so that if cache_reap() is invoked it cannot do
1230 * anything expensive but will only modify reap_work
1231 * and reschedule the timer.
1233 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1234 /* Now the cache_reaper is guaranteed to be not running. */
1235 per_cpu(slab_reap_work, cpu).work.func = NULL;
1237 case CPU_DOWN_FAILED:
1238 case CPU_DOWN_FAILED_FROZEN:
1239 start_cpu_timer(cpu);
1242 case CPU_DEAD_FROZEN:
1244 * Even if all the cpus of a node are down, we don't free the
1245 * kmem_cache_node of any cache. This to avoid a race between
1246 * cpu_down, and a kmalloc allocation from another cpu for
1247 * memory from the node of the cpu going down. The node
1248 * structure is usually allocated from kmem_cache_create() and
1249 * gets destroyed at kmem_cache_destroy().
1253 case CPU_UP_CANCELED:
1254 case CPU_UP_CANCELED_FROZEN:
1255 mutex_lock(&slab_mutex);
1256 cpuup_canceled(cpu);
1257 mutex_unlock(&slab_mutex);
1260 return notifier_from_errno(err);
1263 static struct notifier_block cpucache_notifier = {
1264 &cpuup_callback, NULL, 0
1267 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1269 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1270 * Returns -EBUSY if all objects cannot be drained so that the node is not
1273 * Must hold slab_mutex.
1275 static int __meminit drain_cache_node_node(int node)
1277 struct kmem_cache *cachep;
1280 list_for_each_entry(cachep, &slab_caches, list) {
1281 struct kmem_cache_node *n;
1283 n = get_node(cachep, node);
1287 drain_freelist(cachep, n, slabs_tofree(cachep, n));
1289 if (!list_empty(&n->slabs_full) ||
1290 !list_empty(&n->slabs_partial)) {
1298 static int __meminit slab_memory_callback(struct notifier_block *self,
1299 unsigned long action, void *arg)
1301 struct memory_notify *mnb = arg;
1305 nid = mnb->status_change_nid;
1310 case MEM_GOING_ONLINE:
1311 mutex_lock(&slab_mutex);
1312 ret = init_cache_node_node(nid);
1313 mutex_unlock(&slab_mutex);
1315 case MEM_GOING_OFFLINE:
1316 mutex_lock(&slab_mutex);
1317 ret = drain_cache_node_node(nid);
1318 mutex_unlock(&slab_mutex);
1322 case MEM_CANCEL_ONLINE:
1323 case MEM_CANCEL_OFFLINE:
1327 return notifier_from_errno(ret);
1329 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1332 * swap the static kmem_cache_node with kmalloced memory
1334 static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1337 struct kmem_cache_node *ptr;
1339 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1342 memcpy(ptr, list, sizeof(struct kmem_cache_node));
1344 * Do not assume that spinlocks can be initialized via memcpy:
1346 spin_lock_init(&ptr->list_lock);
1348 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1349 cachep->node[nodeid] = ptr;
1353 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1354 * size of kmem_cache_node.
1356 static void __init set_up_node(struct kmem_cache *cachep, int index)
1360 for_each_online_node(node) {
1361 cachep->node[node] = &init_kmem_cache_node[index + node];
1362 cachep->node[node]->next_reap = jiffies +
1364 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1369 * Initialisation. Called after the page allocator have been initialised and
1370 * before smp_init().
1372 void __init kmem_cache_init(void)
1376 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1377 sizeof(struct rcu_head));
1378 kmem_cache = &kmem_cache_boot;
1380 if (num_possible_nodes() == 1)
1381 use_alien_caches = 0;
1383 for (i = 0; i < NUM_INIT_LISTS; i++)
1384 kmem_cache_node_init(&init_kmem_cache_node[i]);
1387 * Fragmentation resistance on low memory - only use bigger
1388 * page orders on machines with more than 32MB of memory if
1389 * not overridden on the command line.
1391 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1392 slab_max_order = SLAB_MAX_ORDER_HI;
1394 /* Bootstrap is tricky, because several objects are allocated
1395 * from caches that do not exist yet:
1396 * 1) initialize the kmem_cache cache: it contains the struct
1397 * kmem_cache structures of all caches, except kmem_cache itself:
1398 * kmem_cache is statically allocated.
1399 * Initially an __init data area is used for the head array and the
1400 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1401 * array at the end of the bootstrap.
1402 * 2) Create the first kmalloc cache.
1403 * The struct kmem_cache for the new cache is allocated normally.
1404 * An __init data area is used for the head array.
1405 * 3) Create the remaining kmalloc caches, with minimally sized
1407 * 4) Replace the __init data head arrays for kmem_cache and the first
1408 * kmalloc cache with kmalloc allocated arrays.
1409 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1410 * the other cache's with kmalloc allocated memory.
1411 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1414 /* 1) create the kmem_cache */
1417 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1419 create_boot_cache(kmem_cache, "kmem_cache",
1420 offsetof(struct kmem_cache, node) +
1421 nr_node_ids * sizeof(struct kmem_cache_node *),
1422 SLAB_HWCACHE_ALIGN);
1423 list_add(&kmem_cache->list, &slab_caches);
1424 slab_state = PARTIAL;
1427 * Initialize the caches that provide memory for the kmem_cache_node
1428 * structures first. Without this, further allocations will bug.
1430 kmalloc_caches[INDEX_NODE] = create_kmalloc_cache("kmalloc-node",
1431 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
1432 slab_state = PARTIAL_NODE;
1433 setup_kmalloc_cache_index_table();
1435 slab_early_init = 0;
1437 /* 5) Replace the bootstrap kmem_cache_node */
1441 for_each_online_node(nid) {
1442 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
1444 init_list(kmalloc_caches[INDEX_NODE],
1445 &init_kmem_cache_node[SIZE_NODE + nid], nid);
1449 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1452 void __init kmem_cache_init_late(void)
1454 struct kmem_cache *cachep;
1458 /* 6) resize the head arrays to their final sizes */
1459 mutex_lock(&slab_mutex);
1460 list_for_each_entry(cachep, &slab_caches, list)
1461 if (enable_cpucache(cachep, GFP_NOWAIT))
1463 mutex_unlock(&slab_mutex);
1469 * Register a cpu startup notifier callback that initializes
1470 * cpu_cache_get for all new cpus
1472 register_cpu_notifier(&cpucache_notifier);
1476 * Register a memory hotplug callback that initializes and frees
1479 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1483 * The reap timers are started later, with a module init call: That part
1484 * of the kernel is not yet operational.
1488 static int __init cpucache_init(void)
1493 * Register the timers that return unneeded pages to the page allocator
1495 for_each_online_cpu(cpu)
1496 start_cpu_timer(cpu);
1502 __initcall(cpucache_init);
1504 static noinline void
1505 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1508 struct kmem_cache_node *n;
1510 unsigned long flags;
1512 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1513 DEFAULT_RATELIMIT_BURST);
1515 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1519 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1521 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
1522 cachep->name, cachep->size, cachep->gfporder);
1524 for_each_kmem_cache_node(cachep, node, n) {
1525 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1526 unsigned long active_slabs = 0, num_slabs = 0;
1528 spin_lock_irqsave(&n->list_lock, flags);
1529 list_for_each_entry(page, &n->slabs_full, lru) {
1530 active_objs += cachep->num;
1533 list_for_each_entry(page, &n->slabs_partial, lru) {
1534 active_objs += page->active;
1537 list_for_each_entry(page, &n->slabs_free, lru)
1540 free_objects += n->free_objects;
1541 spin_unlock_irqrestore(&n->list_lock, flags);
1543 num_slabs += active_slabs;
1544 num_objs = num_slabs * cachep->num;
1546 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1547 node, active_slabs, num_slabs, active_objs, num_objs,
1554 * Interface to system's page allocator. No need to hold the
1555 * kmem_cache_node ->list_lock.
1557 * If we requested dmaable memory, we will get it. Even if we
1558 * did not request dmaable memory, we might get it, but that
1559 * would be relatively rare and ignorable.
1561 static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1567 flags |= cachep->allocflags;
1568 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1569 flags |= __GFP_RECLAIMABLE;
1571 page = __alloc_pages_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1573 slab_out_of_memory(cachep, flags, nodeid);
1577 if (memcg_charge_slab(page, flags, cachep->gfporder, cachep)) {
1578 __free_pages(page, cachep->gfporder);
1582 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1583 if (page_is_pfmemalloc(page))
1584 pfmemalloc_active = true;
1586 nr_pages = (1 << cachep->gfporder);
1587 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1588 add_zone_page_state(page_zone(page),
1589 NR_SLAB_RECLAIMABLE, nr_pages);
1591 add_zone_page_state(page_zone(page),
1592 NR_SLAB_UNRECLAIMABLE, nr_pages);
1593 __SetPageSlab(page);
1594 if (page_is_pfmemalloc(page))
1595 SetPageSlabPfmemalloc(page);
1597 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1598 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1601 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1603 kmemcheck_mark_unallocated_pages(page, nr_pages);
1610 * Interface to system's page release.
1612 static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
1614 const unsigned long nr_freed = (1 << cachep->gfporder);
1616 kmemcheck_free_shadow(page, cachep->gfporder);
1618 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1619 sub_zone_page_state(page_zone(page),
1620 NR_SLAB_RECLAIMABLE, nr_freed);
1622 sub_zone_page_state(page_zone(page),
1623 NR_SLAB_UNRECLAIMABLE, nr_freed);
1625 BUG_ON(!PageSlab(page));
1626 __ClearPageSlabPfmemalloc(page);
1627 __ClearPageSlab(page);
1628 page_mapcount_reset(page);
1629 page->mapping = NULL;
1631 if (current->reclaim_state)
1632 current->reclaim_state->reclaimed_slab += nr_freed;
1633 __free_kmem_pages(page, cachep->gfporder);
1636 static void kmem_rcu_free(struct rcu_head *head)
1638 struct kmem_cache *cachep;
1641 page = container_of(head, struct page, rcu_head);
1642 cachep = page->slab_cache;
1644 kmem_freepages(cachep, page);
1648 static bool is_debug_pagealloc_cache(struct kmem_cache *cachep)
1650 if (debug_pagealloc_enabled() && OFF_SLAB(cachep) &&
1651 (cachep->size % PAGE_SIZE) == 0)
1657 #ifdef CONFIG_DEBUG_PAGEALLOC
1658 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1659 unsigned long caller)
1661 int size = cachep->object_size;
1663 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1665 if (size < 5 * sizeof(unsigned long))
1668 *addr++ = 0x12345678;
1670 *addr++ = smp_processor_id();
1671 size -= 3 * sizeof(unsigned long);
1673 unsigned long *sptr = &caller;
1674 unsigned long svalue;
1676 while (!kstack_end(sptr)) {
1678 if (kernel_text_address(svalue)) {
1680 size -= sizeof(unsigned long);
1681 if (size <= sizeof(unsigned long))
1687 *addr++ = 0x87654321;
1690 static void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1691 int map, unsigned long caller)
1693 if (!is_debug_pagealloc_cache(cachep))
1697 store_stackinfo(cachep, objp, caller);
1699 kernel_map_pages(virt_to_page(objp), cachep->size / PAGE_SIZE, map);
1703 static inline void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1704 int map, unsigned long caller) {}
1708 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1710 int size = cachep->object_size;
1711 addr = &((char *)addr)[obj_offset(cachep)];
1713 memset(addr, val, size);
1714 *(unsigned char *)(addr + size - 1) = POISON_END;
1717 static void dump_line(char *data, int offset, int limit)
1720 unsigned char error = 0;
1723 printk(KERN_ERR "%03x: ", offset);
1724 for (i = 0; i < limit; i++) {
1725 if (data[offset + i] != POISON_FREE) {
1726 error = data[offset + i];
1730 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1731 &data[offset], limit, 1);
1733 if (bad_count == 1) {
1734 error ^= POISON_FREE;
1735 if (!(error & (error - 1))) {
1736 printk(KERN_ERR "Single bit error detected. Probably "
1739 printk(KERN_ERR "Run memtest86+ or a similar memory "
1742 printk(KERN_ERR "Run a memory test tool.\n");
1751 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1756 if (cachep->flags & SLAB_RED_ZONE) {
1757 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1758 *dbg_redzone1(cachep, objp),
1759 *dbg_redzone2(cachep, objp));
1762 if (cachep->flags & SLAB_STORE_USER) {
1763 printk(KERN_ERR "Last user: [<%p>](%pSR)\n",
1764 *dbg_userword(cachep, objp),
1765 *dbg_userword(cachep, objp));
1767 realobj = (char *)objp + obj_offset(cachep);
1768 size = cachep->object_size;
1769 for (i = 0; i < size && lines; i += 16, lines--) {
1772 if (i + limit > size)
1774 dump_line(realobj, i, limit);
1778 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1784 if (is_debug_pagealloc_cache(cachep))
1787 realobj = (char *)objp + obj_offset(cachep);
1788 size = cachep->object_size;
1790 for (i = 0; i < size; i++) {
1791 char exp = POISON_FREE;
1794 if (realobj[i] != exp) {
1800 "Slab corruption (%s): %s start=%p, len=%d\n",
1801 print_tainted(), cachep->name, realobj, size);
1802 print_objinfo(cachep, objp, 0);
1804 /* Hexdump the affected line */
1807 if (i + limit > size)
1809 dump_line(realobj, i, limit);
1812 /* Limit to 5 lines */
1818 /* Print some data about the neighboring objects, if they
1821 struct page *page = virt_to_head_page(objp);
1824 objnr = obj_to_index(cachep, page, objp);
1826 objp = index_to_obj(cachep, page, objnr - 1);
1827 realobj = (char *)objp + obj_offset(cachep);
1828 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1830 print_objinfo(cachep, objp, 2);
1832 if (objnr + 1 < cachep->num) {
1833 objp = index_to_obj(cachep, page, objnr + 1);
1834 realobj = (char *)objp + obj_offset(cachep);
1835 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1837 print_objinfo(cachep, objp, 2);
1844 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1848 for (i = 0; i < cachep->num; i++) {
1849 void *objp = index_to_obj(cachep, page, i);
1851 if (cachep->flags & SLAB_POISON) {
1852 check_poison_obj(cachep, objp);
1853 slab_kernel_map(cachep, objp, 1, 0);
1855 if (cachep->flags & SLAB_RED_ZONE) {
1856 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1857 slab_error(cachep, "start of a freed object "
1859 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1860 slab_error(cachep, "end of a freed object "
1866 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1873 * slab_destroy - destroy and release all objects in a slab
1874 * @cachep: cache pointer being destroyed
1875 * @page: page pointer being destroyed
1877 * Destroy all the objs in a slab page, and release the mem back to the system.
1878 * Before calling the slab page must have been unlinked from the cache. The
1879 * kmem_cache_node ->list_lock is not held/needed.
1881 static void slab_destroy(struct kmem_cache *cachep, struct page *page)
1885 freelist = page->freelist;
1886 slab_destroy_debugcheck(cachep, page);
1887 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
1888 call_rcu(&page->rcu_head, kmem_rcu_free);
1890 kmem_freepages(cachep, page);
1893 * From now on, we don't use freelist
1894 * although actual page can be freed in rcu context
1896 if (OFF_SLAB(cachep))
1897 kmem_cache_free(cachep->freelist_cache, freelist);
1900 static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1902 struct page *page, *n;
1904 list_for_each_entry_safe(page, n, list, lru) {
1905 list_del(&page->lru);
1906 slab_destroy(cachep, page);
1911 * calculate_slab_order - calculate size (page order) of slabs
1912 * @cachep: pointer to the cache that is being created
1913 * @size: size of objects to be created in this cache.
1914 * @align: required alignment for the objects.
1915 * @flags: slab allocation flags
1917 * Also calculates the number of objects per slab.
1919 * This could be made much more intelligent. For now, try to avoid using
1920 * high order pages for slabs. When the gfp() functions are more friendly
1921 * towards high-order requests, this should be changed.
1923 static size_t calculate_slab_order(struct kmem_cache *cachep,
1924 size_t size, size_t align, unsigned long flags)
1926 unsigned long offslab_limit;
1927 size_t left_over = 0;
1930 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
1934 cache_estimate(gfporder, size, align, flags, &remainder, &num);
1938 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1939 if (num > SLAB_OBJ_MAX_NUM)
1942 if (flags & CFLGS_OFF_SLAB) {
1944 * Max number of objs-per-slab for caches which
1945 * use off-slab slabs. Needed to avoid a possible
1946 * looping condition in cache_grow().
1948 offslab_limit = size;
1949 offslab_limit /= sizeof(freelist_idx_t);
1951 if (num > offslab_limit)
1955 /* Found something acceptable - save it away */
1957 cachep->gfporder = gfporder;
1958 left_over = remainder;
1961 * A VFS-reclaimable slab tends to have most allocations
1962 * as GFP_NOFS and we really don't want to have to be allocating
1963 * higher-order pages when we are unable to shrink dcache.
1965 if (flags & SLAB_RECLAIM_ACCOUNT)
1969 * Large number of objects is good, but very large slabs are
1970 * currently bad for the gfp()s.
1972 if (gfporder >= slab_max_order)
1976 * Acceptable internal fragmentation?
1978 if (left_over * 8 <= (PAGE_SIZE << gfporder))
1984 static struct array_cache __percpu *alloc_kmem_cache_cpus(
1985 struct kmem_cache *cachep, int entries, int batchcount)
1989 struct array_cache __percpu *cpu_cache;
1991 size = sizeof(void *) * entries + sizeof(struct array_cache);
1992 cpu_cache = __alloc_percpu(size, sizeof(void *));
1997 for_each_possible_cpu(cpu) {
1998 init_arraycache(per_cpu_ptr(cpu_cache, cpu),
1999 entries, batchcount);
2005 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2007 if (slab_state >= FULL)
2008 return enable_cpucache(cachep, gfp);
2010 cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
2011 if (!cachep->cpu_cache)
2014 if (slab_state == DOWN) {
2015 /* Creation of first cache (kmem_cache). */
2016 set_up_node(kmem_cache, CACHE_CACHE);
2017 } else if (slab_state == PARTIAL) {
2018 /* For kmem_cache_node */
2019 set_up_node(cachep, SIZE_NODE);
2023 for_each_online_node(node) {
2024 cachep->node[node] = kmalloc_node(
2025 sizeof(struct kmem_cache_node), gfp, node);
2026 BUG_ON(!cachep->node[node]);
2027 kmem_cache_node_init(cachep->node[node]);
2031 cachep->node[numa_mem_id()]->next_reap =
2032 jiffies + REAPTIMEOUT_NODE +
2033 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
2035 cpu_cache_get(cachep)->avail = 0;
2036 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2037 cpu_cache_get(cachep)->batchcount = 1;
2038 cpu_cache_get(cachep)->touched = 0;
2039 cachep->batchcount = 1;
2040 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2044 unsigned long kmem_cache_flags(unsigned long object_size,
2045 unsigned long flags, const char *name,
2046 void (*ctor)(void *))
2052 __kmem_cache_alias(const char *name, size_t size, size_t align,
2053 unsigned long flags, void (*ctor)(void *))
2055 struct kmem_cache *cachep;
2057 cachep = find_mergeable(size, align, flags, name, ctor);
2062 * Adjust the object sizes so that we clear
2063 * the complete object on kzalloc.
2065 cachep->object_size = max_t(int, cachep->object_size, size);
2071 * __kmem_cache_create - Create a cache.
2072 * @cachep: cache management descriptor
2073 * @flags: SLAB flags
2075 * Returns a ptr to the cache on success, NULL on failure.
2076 * Cannot be called within a int, but can be interrupted.
2077 * The @ctor is run when new pages are allocated by the cache.
2081 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2082 * to catch references to uninitialised memory.
2084 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2085 * for buffer overruns.
2087 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2088 * cacheline. This can be beneficial if you're counting cycles as closely
2092 __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2094 size_t left_over, freelist_size;
2095 size_t ralign = BYTES_PER_WORD;
2098 size_t size = cachep->size;
2103 * Enable redzoning and last user accounting, except for caches with
2104 * large objects, if the increased size would increase the object size
2105 * above the next power of two: caches with object sizes just above a
2106 * power of two have a significant amount of internal fragmentation.
2108 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2109 2 * sizeof(unsigned long long)))
2110 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2111 if (!(flags & SLAB_DESTROY_BY_RCU))
2112 flags |= SLAB_POISON;
2117 * Check that size is in terms of words. This is needed to avoid
2118 * unaligned accesses for some archs when redzoning is used, and makes
2119 * sure any on-slab bufctl's are also correctly aligned.
2121 if (size & (BYTES_PER_WORD - 1)) {
2122 size += (BYTES_PER_WORD - 1);
2123 size &= ~(BYTES_PER_WORD - 1);
2126 if (flags & SLAB_RED_ZONE) {
2127 ralign = REDZONE_ALIGN;
2128 /* If redzoning, ensure that the second redzone is suitably
2129 * aligned, by adjusting the object size accordingly. */
2130 size += REDZONE_ALIGN - 1;
2131 size &= ~(REDZONE_ALIGN - 1);
2134 /* 3) caller mandated alignment */
2135 if (ralign < cachep->align) {
2136 ralign = cachep->align;
2138 /* disable debug if necessary */
2139 if (ralign > __alignof__(unsigned long long))
2140 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2144 cachep->align = ralign;
2146 if (slab_is_available())
2154 * Both debugging options require word-alignment which is calculated
2157 if (flags & SLAB_RED_ZONE) {
2158 /* add space for red zone words */
2159 cachep->obj_offset += sizeof(unsigned long long);
2160 size += 2 * sizeof(unsigned long long);
2162 if (flags & SLAB_STORE_USER) {
2163 /* user store requires one word storage behind the end of
2164 * the real object. But if the second red zone needs to be
2165 * aligned to 64 bits, we must allow that much space.
2167 if (flags & SLAB_RED_ZONE)
2168 size += REDZONE_ALIGN;
2170 size += BYTES_PER_WORD;
2173 * To activate debug pagealloc, off-slab management is necessary
2174 * requirement. In early phase of initialization, small sized slab
2175 * doesn't get initialized so it would not be possible. So, we need
2176 * to check size >= 256. It guarantees that all necessary small
2177 * sized slab is initialized in current slab initialization sequence.
2179 if (debug_pagealloc_enabled() && (flags & SLAB_POISON) &&
2180 !slab_early_init && size >= kmalloc_size(INDEX_NODE) &&
2181 size >= 256 && cachep->object_size > cache_line_size() &&
2182 ALIGN(size, cachep->align) < PAGE_SIZE) {
2183 cachep->obj_offset += PAGE_SIZE - ALIGN(size, cachep->align);
2189 * Determine if the slab management is 'on' or 'off' slab.
2190 * (bootstrapping cannot cope with offslab caches so don't do
2191 * it too early on. Always use on-slab management when
2192 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2194 if (size >= OFF_SLAB_MIN_SIZE && !slab_early_init &&
2195 !(flags & SLAB_NOLEAKTRACE))
2197 * Size is large, assume best to place the slab management obj
2198 * off-slab (should allow better packing of objs).
2200 flags |= CFLGS_OFF_SLAB;
2202 size = ALIGN(size, cachep->align);
2204 * We should restrict the number of objects in a slab to implement
2205 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2207 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2208 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2210 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
2215 freelist_size = calculate_freelist_size(cachep->num, cachep->align);
2218 * If the slab has been placed off-slab, and we have enough space then
2219 * move it on-slab. This is at the expense of any extra colouring.
2221 if (flags & CFLGS_OFF_SLAB && left_over >= freelist_size) {
2222 flags &= ~CFLGS_OFF_SLAB;
2223 left_over -= freelist_size;
2226 if (flags & CFLGS_OFF_SLAB) {
2227 /* really off slab. No need for manual alignment */
2228 freelist_size = calculate_freelist_size(cachep->num, 0);
2231 cachep->colour_off = cache_line_size();
2232 /* Offset must be a multiple of the alignment. */
2233 if (cachep->colour_off < cachep->align)
2234 cachep->colour_off = cachep->align;
2235 cachep->colour = left_over / cachep->colour_off;
2236 cachep->freelist_size = freelist_size;
2237 cachep->flags = flags;
2238 cachep->allocflags = __GFP_COMP;
2239 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2240 cachep->allocflags |= GFP_DMA;
2241 cachep->size = size;
2242 cachep->reciprocal_buffer_size = reciprocal_value(size);
2246 * If we're going to use the generic kernel_map_pages()
2247 * poisoning, then it's going to smash the contents of
2248 * the redzone and userword anyhow, so switch them off.
2250 if (IS_ENABLED(CONFIG_PAGE_POISONING) &&
2251 (cachep->flags & SLAB_POISON) &&
2252 is_debug_pagealloc_cache(cachep))
2253 cachep->flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2256 if (OFF_SLAB(cachep)) {
2257 cachep->freelist_cache = kmalloc_slab(freelist_size, 0u);
2259 * This is a possibility for one of the kmalloc_{dma,}_caches.
2260 * But since we go off slab only for object size greater than
2261 * OFF_SLAB_MIN_SIZE, and kmalloc_{dma,}_caches get created
2262 * in ascending order,this should not happen at all.
2263 * But leave a BUG_ON for some lucky dude.
2265 BUG_ON(ZERO_OR_NULL_PTR(cachep->freelist_cache));
2268 err = setup_cpu_cache(cachep, gfp);
2270 __kmem_cache_release(cachep);
2278 static void check_irq_off(void)
2280 BUG_ON(!irqs_disabled());
2283 static void check_irq_on(void)
2285 BUG_ON(irqs_disabled());
2288 static void check_spinlock_acquired(struct kmem_cache *cachep)
2292 assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
2296 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2300 assert_spin_locked(&get_node(cachep, node)->list_lock);
2305 #define check_irq_off() do { } while(0)
2306 #define check_irq_on() do { } while(0)
2307 #define check_spinlock_acquired(x) do { } while(0)
2308 #define check_spinlock_acquired_node(x, y) do { } while(0)
2311 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
2312 struct array_cache *ac,
2313 int force, int node);
2315 static void do_drain(void *arg)
2317 struct kmem_cache *cachep = arg;
2318 struct array_cache *ac;
2319 int node = numa_mem_id();
2320 struct kmem_cache_node *n;
2324 ac = cpu_cache_get(cachep);
2325 n = get_node(cachep, node);
2326 spin_lock(&n->list_lock);
2327 free_block(cachep, ac->entry, ac->avail, node, &list);
2328 spin_unlock(&n->list_lock);
2329 slabs_destroy(cachep, &list);
2333 static void drain_cpu_caches(struct kmem_cache *cachep)
2335 struct kmem_cache_node *n;
2338 on_each_cpu(do_drain, cachep, 1);
2340 for_each_kmem_cache_node(cachep, node, n)
2342 drain_alien_cache(cachep, n->alien);
2344 for_each_kmem_cache_node(cachep, node, n)
2345 drain_array(cachep, n, n->shared, 1, node);
2349 * Remove slabs from the list of free slabs.
2350 * Specify the number of slabs to drain in tofree.
2352 * Returns the actual number of slabs released.
2354 static int drain_freelist(struct kmem_cache *cache,
2355 struct kmem_cache_node *n, int tofree)
2357 struct list_head *p;
2362 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2364 spin_lock_irq(&n->list_lock);
2365 p = n->slabs_free.prev;
2366 if (p == &n->slabs_free) {
2367 spin_unlock_irq(&n->list_lock);
2371 page = list_entry(p, struct page, lru);
2372 list_del(&page->lru);
2374 * Safe to drop the lock. The slab is no longer linked
2377 n->free_objects -= cache->num;
2378 spin_unlock_irq(&n->list_lock);
2379 slab_destroy(cache, page);
2386 int __kmem_cache_shrink(struct kmem_cache *cachep, bool deactivate)
2390 struct kmem_cache_node *n;
2392 drain_cpu_caches(cachep);
2395 for_each_kmem_cache_node(cachep, node, n) {
2396 drain_freelist(cachep, n, slabs_tofree(cachep, n));
2398 ret += !list_empty(&n->slabs_full) ||
2399 !list_empty(&n->slabs_partial);
2401 return (ret ? 1 : 0);
2404 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2406 return __kmem_cache_shrink(cachep, false);
2409 void __kmem_cache_release(struct kmem_cache *cachep)
2412 struct kmem_cache_node *n;
2414 free_percpu(cachep->cpu_cache);
2416 /* NUMA: free the node structures */
2417 for_each_kmem_cache_node(cachep, i, n) {
2419 free_alien_cache(n->alien);
2421 cachep->node[i] = NULL;
2426 * Get the memory for a slab management obj.
2428 * For a slab cache when the slab descriptor is off-slab, the
2429 * slab descriptor can't come from the same cache which is being created,
2430 * Because if it is the case, that means we defer the creation of
2431 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2432 * And we eventually call down to __kmem_cache_create(), which
2433 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2434 * This is a "chicken-and-egg" problem.
2436 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2437 * which are all initialized during kmem_cache_init().
2439 static void *alloc_slabmgmt(struct kmem_cache *cachep,
2440 struct page *page, int colour_off,
2441 gfp_t local_flags, int nodeid)
2444 void *addr = page_address(page);
2446 if (OFF_SLAB(cachep)) {
2447 /* Slab management obj is off-slab. */
2448 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
2449 local_flags, nodeid);
2453 freelist = addr + colour_off;
2454 colour_off += cachep->freelist_size;
2457 page->s_mem = addr + colour_off;
2461 static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
2463 return ((freelist_idx_t *)page->freelist)[idx];
2466 static inline void set_free_obj(struct page *page,
2467 unsigned int idx, freelist_idx_t val)
2469 ((freelist_idx_t *)(page->freelist))[idx] = val;
2472 static void cache_init_objs(struct kmem_cache *cachep,
2477 for (i = 0; i < cachep->num; i++) {
2478 void *objp = index_to_obj(cachep, page, i);
2480 if (cachep->flags & SLAB_STORE_USER)
2481 *dbg_userword(cachep, objp) = NULL;
2483 if (cachep->flags & SLAB_RED_ZONE) {
2484 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2485 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2488 * Constructors are not allowed to allocate memory from the same
2489 * cache which they are a constructor for. Otherwise, deadlock.
2490 * They must also be threaded.
2492 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2493 cachep->ctor(objp + obj_offset(cachep));
2495 if (cachep->flags & SLAB_RED_ZONE) {
2496 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2497 slab_error(cachep, "constructor overwrote the"
2498 " end of an object");
2499 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2500 slab_error(cachep, "constructor overwrote the"
2501 " start of an object");
2503 /* need to poison the objs? */
2504 if (cachep->flags & SLAB_POISON) {
2505 poison_obj(cachep, objp, POISON_FREE);
2506 slab_kernel_map(cachep, objp, 0, 0);
2512 set_free_obj(page, i, i);
2516 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2518 if (CONFIG_ZONE_DMA_FLAG) {
2519 if (flags & GFP_DMA)
2520 BUG_ON(!(cachep->allocflags & GFP_DMA));
2522 BUG_ON(cachep->allocflags & GFP_DMA);
2526 static void *slab_get_obj(struct kmem_cache *cachep, struct page *page)
2530 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
2534 if (cachep->flags & SLAB_STORE_USER)
2535 set_store_user_dirty(cachep);
2541 static void slab_put_obj(struct kmem_cache *cachep,
2542 struct page *page, void *objp)
2544 unsigned int objnr = obj_to_index(cachep, page, objp);
2548 /* Verify double free bug */
2549 for (i = page->active; i < cachep->num; i++) {
2550 if (get_free_obj(page, i) == objnr) {
2551 printk(KERN_ERR "slab: double free detected in cache "
2552 "'%s', objp %p\n", cachep->name, objp);
2558 set_free_obj(page, page->active, objnr);
2562 * Map pages beginning at addr to the given cache and slab. This is required
2563 * for the slab allocator to be able to lookup the cache and slab of a
2564 * virtual address for kfree, ksize, and slab debugging.
2566 static void slab_map_pages(struct kmem_cache *cache, struct page *page,
2569 page->slab_cache = cache;
2570 page->freelist = freelist;
2574 * Grow (by 1) the number of slabs within a cache. This is called by
2575 * kmem_cache_alloc() when there are no active objs left in a cache.
2577 static int cache_grow(struct kmem_cache *cachep,
2578 gfp_t flags, int nodeid, struct page *page)
2583 struct kmem_cache_node *n;
2586 * Be lazy and only check for valid flags here, keeping it out of the
2587 * critical path in kmem_cache_alloc().
2589 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
2590 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
2593 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2595 /* Take the node list lock to change the colour_next on this node */
2597 n = get_node(cachep, nodeid);
2598 spin_lock(&n->list_lock);
2600 /* Get colour for the slab, and cal the next value. */
2601 offset = n->colour_next;
2603 if (n->colour_next >= cachep->colour)
2605 spin_unlock(&n->list_lock);
2607 offset *= cachep->colour_off;
2609 if (gfpflags_allow_blocking(local_flags))
2613 * The test for missing atomic flag is performed here, rather than
2614 * the more obvious place, simply to reduce the critical path length
2615 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2616 * will eventually be caught here (where it matters).
2618 kmem_flagcheck(cachep, flags);
2621 * Get mem for the objs. Attempt to allocate a physical page from
2625 page = kmem_getpages(cachep, local_flags, nodeid);
2629 /* Get slab management. */
2630 freelist = alloc_slabmgmt(cachep, page, offset,
2631 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2635 slab_map_pages(cachep, page, freelist);
2637 cache_init_objs(cachep, page);
2639 if (gfpflags_allow_blocking(local_flags))
2640 local_irq_disable();
2642 spin_lock(&n->list_lock);
2644 /* Make slab active. */
2645 list_add_tail(&page->lru, &(n->slabs_free));
2646 STATS_INC_GROWN(cachep);
2647 n->free_objects += cachep->num;
2648 spin_unlock(&n->list_lock);
2651 kmem_freepages(cachep, page);
2653 if (gfpflags_allow_blocking(local_flags))
2654 local_irq_disable();
2661 * Perform extra freeing checks:
2662 * - detect bad pointers.
2663 * - POISON/RED_ZONE checking
2665 static void kfree_debugcheck(const void *objp)
2667 if (!virt_addr_valid(objp)) {
2668 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2669 (unsigned long)objp);
2674 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2676 unsigned long long redzone1, redzone2;
2678 redzone1 = *dbg_redzone1(cache, obj);
2679 redzone2 = *dbg_redzone2(cache, obj);
2684 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2687 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2688 slab_error(cache, "double free detected");
2690 slab_error(cache, "memory outside object was overwritten");
2692 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2693 obj, redzone1, redzone2);
2696 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2697 unsigned long caller)
2702 BUG_ON(virt_to_cache(objp) != cachep);
2704 objp -= obj_offset(cachep);
2705 kfree_debugcheck(objp);
2706 page = virt_to_head_page(objp);
2708 if (cachep->flags & SLAB_RED_ZONE) {
2709 verify_redzone_free(cachep, objp);
2710 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2711 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2713 if (cachep->flags & SLAB_STORE_USER) {
2714 set_store_user_dirty(cachep);
2715 *dbg_userword(cachep, objp) = (void *)caller;
2718 objnr = obj_to_index(cachep, page, objp);
2720 BUG_ON(objnr >= cachep->num);
2721 BUG_ON(objp != index_to_obj(cachep, page, objnr));
2723 if (cachep->flags & SLAB_POISON) {
2724 poison_obj(cachep, objp, POISON_FREE);
2725 slab_kernel_map(cachep, objp, 0, caller);
2731 #define kfree_debugcheck(x) do { } while(0)
2732 #define cache_free_debugcheck(x,objp,z) (objp)
2735 static struct page *get_first_slab(struct kmem_cache_node *n)
2739 page = list_first_entry_or_null(&n->slabs_partial,
2742 n->free_touched = 1;
2743 page = list_first_entry_or_null(&n->slabs_free,
2750 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
2754 struct kmem_cache_node *n;
2755 struct array_cache *ac;
2759 node = numa_mem_id();
2760 if (unlikely(force_refill))
2763 ac = cpu_cache_get(cachep);
2764 batchcount = ac->batchcount;
2765 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2767 * If there was little recent activity on this cache, then
2768 * perform only a partial refill. Otherwise we could generate
2771 batchcount = BATCHREFILL_LIMIT;
2773 n = get_node(cachep, node);
2775 BUG_ON(ac->avail > 0 || !n);
2776 spin_lock(&n->list_lock);
2778 /* See if we can refill from the shared array */
2779 if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2780 n->shared->touched = 1;
2784 while (batchcount > 0) {
2786 /* Get slab alloc is to come from. */
2787 page = get_first_slab(n);
2791 check_spinlock_acquired(cachep);
2794 * The slab was either on partial or free list so
2795 * there must be at least one object available for
2798 BUG_ON(page->active >= cachep->num);
2800 while (page->active < cachep->num && batchcount--) {
2801 STATS_INC_ALLOCED(cachep);
2802 STATS_INC_ACTIVE(cachep);
2803 STATS_SET_HIGH(cachep);
2805 ac_put_obj(cachep, ac, slab_get_obj(cachep, page));
2808 /* move slabp to correct slabp list: */
2809 list_del(&page->lru);
2810 if (page->active == cachep->num)
2811 list_add(&page->lru, &n->slabs_full);
2813 list_add(&page->lru, &n->slabs_partial);
2817 n->free_objects -= ac->avail;
2819 spin_unlock(&n->list_lock);
2821 if (unlikely(!ac->avail)) {
2824 x = cache_grow(cachep, gfp_exact_node(flags), node, NULL);
2826 /* cache_grow can reenable interrupts, then ac could change. */
2827 ac = cpu_cache_get(cachep);
2828 node = numa_mem_id();
2830 /* no objects in sight? abort */
2831 if (!x && (ac->avail == 0 || force_refill))
2834 if (!ac->avail) /* objects refilled by interrupt? */
2839 return ac_get_obj(cachep, ac, flags, force_refill);
2842 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2845 might_sleep_if(gfpflags_allow_blocking(flags));
2847 kmem_flagcheck(cachep, flags);
2852 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2853 gfp_t flags, void *objp, unsigned long caller)
2857 if (cachep->flags & SLAB_POISON) {
2858 check_poison_obj(cachep, objp);
2859 slab_kernel_map(cachep, objp, 1, 0);
2860 poison_obj(cachep, objp, POISON_INUSE);
2862 if (cachep->flags & SLAB_STORE_USER)
2863 *dbg_userword(cachep, objp) = (void *)caller;
2865 if (cachep->flags & SLAB_RED_ZONE) {
2866 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2867 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2868 slab_error(cachep, "double free, or memory outside"
2869 " object was overwritten");
2871 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2872 objp, *dbg_redzone1(cachep, objp),
2873 *dbg_redzone2(cachep, objp));
2875 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2876 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2879 objp += obj_offset(cachep);
2880 if (cachep->ctor && cachep->flags & SLAB_POISON)
2882 if (ARCH_SLAB_MINALIGN &&
2883 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
2884 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
2885 objp, (int)ARCH_SLAB_MINALIGN);
2890 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2893 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
2896 struct array_cache *ac;
2897 bool force_refill = false;
2901 ac = cpu_cache_get(cachep);
2902 if (likely(ac->avail)) {
2904 objp = ac_get_obj(cachep, ac, flags, false);
2907 * Allow for the possibility all avail objects are not allowed
2908 * by the current flags
2911 STATS_INC_ALLOCHIT(cachep);
2914 force_refill = true;
2917 STATS_INC_ALLOCMISS(cachep);
2918 objp = cache_alloc_refill(cachep, flags, force_refill);
2920 * the 'ac' may be updated by cache_alloc_refill(),
2921 * and kmemleak_erase() requires its correct value.
2923 ac = cpu_cache_get(cachep);
2927 * To avoid a false negative, if an object that is in one of the
2928 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
2929 * treat the array pointers as a reference to the object.
2932 kmemleak_erase(&ac->entry[ac->avail]);
2938 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
2940 * If we are in_interrupt, then process context, including cpusets and
2941 * mempolicy, may not apply and should not be used for allocation policy.
2943 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
2945 int nid_alloc, nid_here;
2947 if (in_interrupt() || (flags & __GFP_THISNODE))
2949 nid_alloc = nid_here = numa_mem_id();
2950 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
2951 nid_alloc = cpuset_slab_spread_node();
2952 else if (current->mempolicy)
2953 nid_alloc = mempolicy_slab_node();
2954 if (nid_alloc != nid_here)
2955 return ____cache_alloc_node(cachep, flags, nid_alloc);
2960 * Fallback function if there was no memory available and no objects on a
2961 * certain node and fall back is permitted. First we scan all the
2962 * available node for available objects. If that fails then we
2963 * perform an allocation without specifying a node. This allows the page
2964 * allocator to do its reclaim / fallback magic. We then insert the
2965 * slab into the proper nodelist and then allocate from it.
2967 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
2969 struct zonelist *zonelist;
2973 enum zone_type high_zoneidx = gfp_zone(flags);
2976 unsigned int cpuset_mems_cookie;
2978 if (flags & __GFP_THISNODE)
2981 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2984 cpuset_mems_cookie = read_mems_allowed_begin();
2985 zonelist = node_zonelist(mempolicy_slab_node(), flags);
2989 * Look through allowed nodes for objects available
2990 * from existing per node queues.
2992 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
2993 nid = zone_to_nid(zone);
2995 if (cpuset_zone_allowed(zone, flags) &&
2996 get_node(cache, nid) &&
2997 get_node(cache, nid)->free_objects) {
2998 obj = ____cache_alloc_node(cache,
2999 gfp_exact_node(flags), nid);
3007 * This allocation will be performed within the constraints
3008 * of the current cpuset / memory policy requirements.
3009 * We may trigger various forms of reclaim on the allowed
3010 * set and go into memory reserves if necessary.
3014 if (gfpflags_allow_blocking(local_flags))
3016 kmem_flagcheck(cache, flags);
3017 page = kmem_getpages(cache, local_flags, numa_mem_id());
3018 if (gfpflags_allow_blocking(local_flags))
3019 local_irq_disable();
3022 * Insert into the appropriate per node queues
3024 nid = page_to_nid(page);
3025 if (cache_grow(cache, flags, nid, page)) {
3026 obj = ____cache_alloc_node(cache,
3027 gfp_exact_node(flags), nid);
3030 * Another processor may allocate the
3031 * objects in the slab since we are
3032 * not holding any locks.
3036 /* cache_grow already freed obj */
3042 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
3048 * A interface to enable slab creation on nodeid
3050 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3054 struct kmem_cache_node *n;
3058 VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
3059 n = get_node(cachep, nodeid);
3064 spin_lock(&n->list_lock);
3065 page = get_first_slab(n);
3069 check_spinlock_acquired_node(cachep, nodeid);
3071 STATS_INC_NODEALLOCS(cachep);
3072 STATS_INC_ACTIVE(cachep);
3073 STATS_SET_HIGH(cachep);
3075 BUG_ON(page->active == cachep->num);
3077 obj = slab_get_obj(cachep, page);
3079 /* move slabp to correct slabp list: */
3080 list_del(&page->lru);
3082 if (page->active == cachep->num)
3083 list_add(&page->lru, &n->slabs_full);
3085 list_add(&page->lru, &n->slabs_partial);
3087 spin_unlock(&n->list_lock);
3091 spin_unlock(&n->list_lock);
3092 x = cache_grow(cachep, gfp_exact_node(flags), nodeid, NULL);
3096 return fallback_alloc(cachep, flags);
3102 static __always_inline void *
3103 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3104 unsigned long caller)
3106 unsigned long save_flags;
3108 int slab_node = numa_mem_id();
3110 flags &= gfp_allowed_mask;
3111 cachep = slab_pre_alloc_hook(cachep, flags);
3112 if (unlikely(!cachep))
3115 cache_alloc_debugcheck_before(cachep, flags);
3116 local_irq_save(save_flags);
3118 if (nodeid == NUMA_NO_NODE)
3121 if (unlikely(!get_node(cachep, nodeid))) {
3122 /* Node not bootstrapped yet */
3123 ptr = fallback_alloc(cachep, flags);
3127 if (nodeid == slab_node) {
3129 * Use the locally cached objects if possible.
3130 * However ____cache_alloc does not allow fallback
3131 * to other nodes. It may fail while we still have
3132 * objects on other nodes available.
3134 ptr = ____cache_alloc(cachep, flags);
3138 /* ___cache_alloc_node can fall back to other nodes */
3139 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3141 local_irq_restore(save_flags);
3142 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3144 if (unlikely(flags & __GFP_ZERO) && ptr)
3145 memset(ptr, 0, cachep->object_size);
3147 slab_post_alloc_hook(cachep, flags, 1, &ptr);
3151 static __always_inline void *
3152 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3156 if (current->mempolicy || cpuset_do_slab_mem_spread()) {
3157 objp = alternate_node_alloc(cache, flags);
3161 objp = ____cache_alloc(cache, flags);
3164 * We may just have run out of memory on the local node.
3165 * ____cache_alloc_node() knows how to locate memory on other nodes
3168 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3175 static __always_inline void *
3176 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3178 return ____cache_alloc(cachep, flags);
3181 #endif /* CONFIG_NUMA */
3183 static __always_inline void *
3184 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3186 unsigned long save_flags;
3189 flags &= gfp_allowed_mask;
3190 cachep = slab_pre_alloc_hook(cachep, flags);
3191 if (unlikely(!cachep))
3194 cache_alloc_debugcheck_before(cachep, flags);
3195 local_irq_save(save_flags);
3196 objp = __do_cache_alloc(cachep, flags);
3197 local_irq_restore(save_flags);
3198 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3201 if (unlikely(flags & __GFP_ZERO) && objp)
3202 memset(objp, 0, cachep->object_size);
3204 slab_post_alloc_hook(cachep, flags, 1, &objp);
3209 * Caller needs to acquire correct kmem_cache_node's list_lock
3210 * @list: List of detached free slabs should be freed by caller
3212 static void free_block(struct kmem_cache *cachep, void **objpp,
3213 int nr_objects, int node, struct list_head *list)
3216 struct kmem_cache_node *n = get_node(cachep, node);
3218 for (i = 0; i < nr_objects; i++) {
3222 clear_obj_pfmemalloc(&objpp[i]);
3225 page = virt_to_head_page(objp);
3226 list_del(&page->lru);
3227 check_spinlock_acquired_node(cachep, node);
3228 slab_put_obj(cachep, page, objp);
3229 STATS_DEC_ACTIVE(cachep);
3232 /* fixup slab chains */
3233 if (page->active == 0) {
3234 if (n->free_objects > n->free_limit) {
3235 n->free_objects -= cachep->num;
3236 list_add_tail(&page->lru, list);
3238 list_add(&page->lru, &n->slabs_free);
3241 /* Unconditionally move a slab to the end of the
3242 * partial list on free - maximum time for the
3243 * other objects to be freed, too.
3245 list_add_tail(&page->lru, &n->slabs_partial);
3250 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3253 struct kmem_cache_node *n;
3254 int node = numa_mem_id();
3257 batchcount = ac->batchcount;
3260 n = get_node(cachep, node);
3261 spin_lock(&n->list_lock);
3263 struct array_cache *shared_array = n->shared;
3264 int max = shared_array->limit - shared_array->avail;
3266 if (batchcount > max)
3268 memcpy(&(shared_array->entry[shared_array->avail]),
3269 ac->entry, sizeof(void *) * batchcount);
3270 shared_array->avail += batchcount;
3275 free_block(cachep, ac->entry, batchcount, node, &list);
3282 list_for_each_entry(page, &n->slabs_free, lru) {
3283 BUG_ON(page->active);
3287 STATS_SET_FREEABLE(cachep, i);
3290 spin_unlock(&n->list_lock);
3291 slabs_destroy(cachep, &list);
3292 ac->avail -= batchcount;
3293 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3297 * Release an obj back to its cache. If the obj has a constructed state, it must
3298 * be in this state _before_ it is released. Called with disabled ints.
3300 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3301 unsigned long caller)
3303 struct array_cache *ac = cpu_cache_get(cachep);
3306 kmemleak_free_recursive(objp, cachep->flags);
3307 objp = cache_free_debugcheck(cachep, objp, caller);
3309 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3312 * Skip calling cache_free_alien() when the platform is not numa.
3313 * This will avoid cache misses that happen while accessing slabp (which
3314 * is per page memory reference) to get nodeid. Instead use a global
3315 * variable to skip the call, which is mostly likely to be present in
3318 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3321 if (ac->avail < ac->limit) {
3322 STATS_INC_FREEHIT(cachep);
3324 STATS_INC_FREEMISS(cachep);
3325 cache_flusharray(cachep, ac);
3328 ac_put_obj(cachep, ac, objp);
3332 * kmem_cache_alloc - Allocate an object
3333 * @cachep: The cache to allocate from.
3334 * @flags: See kmalloc().
3336 * Allocate an object from this cache. The flags are only relevant
3337 * if the cache has no available objects.
3339 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3341 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3343 trace_kmem_cache_alloc(_RET_IP_, ret,
3344 cachep->object_size, cachep->size, flags);
3348 EXPORT_SYMBOL(kmem_cache_alloc);
3350 static __always_inline void
3351 cache_alloc_debugcheck_after_bulk(struct kmem_cache *s, gfp_t flags,
3352 size_t size, void **p, unsigned long caller)
3356 for (i = 0; i < size; i++)
3357 p[i] = cache_alloc_debugcheck_after(s, flags, p[i], caller);
3360 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3365 s = slab_pre_alloc_hook(s, flags);
3369 cache_alloc_debugcheck_before(s, flags);
3371 local_irq_disable();
3372 for (i = 0; i < size; i++) {
3373 void *objp = __do_cache_alloc(s, flags);
3375 if (unlikely(!objp))
3381 cache_alloc_debugcheck_after_bulk(s, flags, size, p, _RET_IP_);
3383 /* Clear memory outside IRQ disabled section */
3384 if (unlikely(flags & __GFP_ZERO))
3385 for (i = 0; i < size; i++)
3386 memset(p[i], 0, s->object_size);
3388 slab_post_alloc_hook(s, flags, size, p);
3389 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3393 cache_alloc_debugcheck_after_bulk(s, flags, i, p, _RET_IP_);
3394 slab_post_alloc_hook(s, flags, i, p);
3395 __kmem_cache_free_bulk(s, i, p);
3398 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3400 #ifdef CONFIG_TRACING
3402 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3406 ret = slab_alloc(cachep, flags, _RET_IP_);
3408 trace_kmalloc(_RET_IP_, ret,
3409 size, cachep->size, flags);
3412 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3417 * kmem_cache_alloc_node - Allocate an object on the specified node
3418 * @cachep: The cache to allocate from.
3419 * @flags: See kmalloc().
3420 * @nodeid: node number of the target node.
3422 * Identical to kmem_cache_alloc but it will allocate memory on the given
3423 * node, which can improve the performance for cpu bound structures.
3425 * Fallback to other node is possible if __GFP_THISNODE is not set.
3427 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3429 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3431 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3432 cachep->object_size, cachep->size,
3437 EXPORT_SYMBOL(kmem_cache_alloc_node);
3439 #ifdef CONFIG_TRACING
3440 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3447 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3449 trace_kmalloc_node(_RET_IP_, ret,
3454 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3457 static __always_inline void *
3458 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3460 struct kmem_cache *cachep;
3462 cachep = kmalloc_slab(size, flags);
3463 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3465 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
3468 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3470 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3472 EXPORT_SYMBOL(__kmalloc_node);
3474 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3475 int node, unsigned long caller)
3477 return __do_kmalloc_node(size, flags, node, caller);
3479 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3480 #endif /* CONFIG_NUMA */
3483 * __do_kmalloc - allocate memory
3484 * @size: how many bytes of memory are required.
3485 * @flags: the type of memory to allocate (see kmalloc).
3486 * @caller: function caller for debug tracking of the caller
3488 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3489 unsigned long caller)
3491 struct kmem_cache *cachep;
3494 cachep = kmalloc_slab(size, flags);
3495 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3497 ret = slab_alloc(cachep, flags, caller);
3499 trace_kmalloc(caller, ret,
3500 size, cachep->size, flags);
3505 void *__kmalloc(size_t size, gfp_t flags)
3507 return __do_kmalloc(size, flags, _RET_IP_);
3509 EXPORT_SYMBOL(__kmalloc);
3511 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3513 return __do_kmalloc(size, flags, caller);
3515 EXPORT_SYMBOL(__kmalloc_track_caller);
3518 * kmem_cache_free - Deallocate an object
3519 * @cachep: The cache the allocation was from.
3520 * @objp: The previously allocated object.
3522 * Free an object which was previously allocated from this
3525 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3527 unsigned long flags;
3528 cachep = cache_from_obj(cachep, objp);
3532 local_irq_save(flags);
3533 debug_check_no_locks_freed(objp, cachep->object_size);
3534 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3535 debug_check_no_obj_freed(objp, cachep->object_size);
3536 __cache_free(cachep, objp, _RET_IP_);
3537 local_irq_restore(flags);
3539 trace_kmem_cache_free(_RET_IP_, objp);
3541 EXPORT_SYMBOL(kmem_cache_free);
3543 void kmem_cache_free_bulk(struct kmem_cache *orig_s, size_t size, void **p)
3545 struct kmem_cache *s;
3548 local_irq_disable();
3549 for (i = 0; i < size; i++) {
3552 if (!orig_s) /* called via kfree_bulk */
3553 s = virt_to_cache(objp);
3555 s = cache_from_obj(orig_s, objp);
3557 debug_check_no_locks_freed(objp, s->object_size);
3558 if (!(s->flags & SLAB_DEBUG_OBJECTS))
3559 debug_check_no_obj_freed(objp, s->object_size);
3561 __cache_free(s, objp, _RET_IP_);
3565 /* FIXME: add tracing */
3567 EXPORT_SYMBOL(kmem_cache_free_bulk);
3570 * kfree - free previously allocated memory
3571 * @objp: pointer returned by kmalloc.
3573 * If @objp is NULL, no operation is performed.
3575 * Don't free memory not originally allocated by kmalloc()
3576 * or you will run into trouble.
3578 void kfree(const void *objp)
3580 struct kmem_cache *c;
3581 unsigned long flags;
3583 trace_kfree(_RET_IP_, objp);
3585 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3587 local_irq_save(flags);
3588 kfree_debugcheck(objp);
3589 c = virt_to_cache(objp);
3590 debug_check_no_locks_freed(objp, c->object_size);
3592 debug_check_no_obj_freed(objp, c->object_size);
3593 __cache_free(c, (void *)objp, _RET_IP_);
3594 local_irq_restore(flags);
3596 EXPORT_SYMBOL(kfree);
3599 * This initializes kmem_cache_node or resizes various caches for all nodes.
3601 static int alloc_kmem_cache_node(struct kmem_cache *cachep, gfp_t gfp)
3604 struct kmem_cache_node *n;
3605 struct array_cache *new_shared;
3606 struct alien_cache **new_alien = NULL;
3608 for_each_online_node(node) {
3610 if (use_alien_caches) {
3611 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3617 if (cachep->shared) {
3618 new_shared = alloc_arraycache(node,
3619 cachep->shared*cachep->batchcount,
3622 free_alien_cache(new_alien);
3627 n = get_node(cachep, node);
3629 struct array_cache *shared = n->shared;
3632 spin_lock_irq(&n->list_lock);
3635 free_block(cachep, shared->entry,
3636 shared->avail, node, &list);
3638 n->shared = new_shared;
3640 n->alien = new_alien;
3643 n->free_limit = (1 + nr_cpus_node(node)) *
3644 cachep->batchcount + cachep->num;
3645 spin_unlock_irq(&n->list_lock);
3646 slabs_destroy(cachep, &list);
3648 free_alien_cache(new_alien);
3651 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3653 free_alien_cache(new_alien);
3658 kmem_cache_node_init(n);
3659 n->next_reap = jiffies + REAPTIMEOUT_NODE +
3660 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
3661 n->shared = new_shared;
3662 n->alien = new_alien;
3663 n->free_limit = (1 + nr_cpus_node(node)) *
3664 cachep->batchcount + cachep->num;
3665 cachep->node[node] = n;
3670 if (!cachep->list.next) {
3671 /* Cache is not active yet. Roll back what we did */
3674 n = get_node(cachep, node);
3677 free_alien_cache(n->alien);
3679 cachep->node[node] = NULL;
3687 /* Always called with the slab_mutex held */
3688 static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
3689 int batchcount, int shared, gfp_t gfp)
3691 struct array_cache __percpu *cpu_cache, *prev;
3694 cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3698 prev = cachep->cpu_cache;
3699 cachep->cpu_cache = cpu_cache;
3700 kick_all_cpus_sync();
3703 cachep->batchcount = batchcount;
3704 cachep->limit = limit;
3705 cachep->shared = shared;
3710 for_each_online_cpu(cpu) {
3713 struct kmem_cache_node *n;
3714 struct array_cache *ac = per_cpu_ptr(prev, cpu);
3716 node = cpu_to_mem(cpu);
3717 n = get_node(cachep, node);
3718 spin_lock_irq(&n->list_lock);
3719 free_block(cachep, ac->entry, ac->avail, node, &list);
3720 spin_unlock_irq(&n->list_lock);
3721 slabs_destroy(cachep, &list);
3726 return alloc_kmem_cache_node(cachep, gfp);
3729 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3730 int batchcount, int shared, gfp_t gfp)
3733 struct kmem_cache *c;
3735 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3737 if (slab_state < FULL)
3740 if ((ret < 0) || !is_root_cache(cachep))
3743 lockdep_assert_held(&slab_mutex);
3744 for_each_memcg_cache(c, cachep) {
3745 /* return value determined by the root cache only */
3746 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3752 /* Called with slab_mutex held always */
3753 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3760 if (!is_root_cache(cachep)) {
3761 struct kmem_cache *root = memcg_root_cache(cachep);
3762 limit = root->limit;
3763 shared = root->shared;
3764 batchcount = root->batchcount;
3767 if (limit && shared && batchcount)
3770 * The head array serves three purposes:
3771 * - create a LIFO ordering, i.e. return objects that are cache-warm
3772 * - reduce the number of spinlock operations.
3773 * - reduce the number of linked list operations on the slab and
3774 * bufctl chains: array operations are cheaper.
3775 * The numbers are guessed, we should auto-tune as described by
3778 if (cachep->size > 131072)
3780 else if (cachep->size > PAGE_SIZE)
3782 else if (cachep->size > 1024)
3784 else if (cachep->size > 256)
3790 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3791 * allocation behaviour: Most allocs on one cpu, most free operations
3792 * on another cpu. For these cases, an efficient object passing between
3793 * cpus is necessary. This is provided by a shared array. The array
3794 * replaces Bonwick's magazine layer.
3795 * On uniprocessor, it's functionally equivalent (but less efficient)
3796 * to a larger limit. Thus disabled by default.
3799 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
3804 * With debugging enabled, large batchcount lead to excessively long
3805 * periods with disabled local interrupts. Limit the batchcount
3810 batchcount = (limit + 1) / 2;
3812 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3814 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3815 cachep->name, -err);
3820 * Drain an array if it contains any elements taking the node lock only if
3821 * necessary. Note that the node listlock also protects the array_cache
3822 * if drain_array() is used on the shared array.
3824 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
3825 struct array_cache *ac, int force, int node)
3830 if (!ac || !ac->avail)
3832 if (ac->touched && !force) {
3835 spin_lock_irq(&n->list_lock);
3837 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3838 if (tofree > ac->avail)
3839 tofree = (ac->avail + 1) / 2;
3840 free_block(cachep, ac->entry, tofree, node, &list);
3841 ac->avail -= tofree;
3842 memmove(ac->entry, &(ac->entry[tofree]),
3843 sizeof(void *) * ac->avail);
3845 spin_unlock_irq(&n->list_lock);
3846 slabs_destroy(cachep, &list);
3851 * cache_reap - Reclaim memory from caches.
3852 * @w: work descriptor
3854 * Called from workqueue/eventd every few seconds.
3856 * - clear the per-cpu caches for this CPU.
3857 * - return freeable pages to the main free memory pool.
3859 * If we cannot acquire the cache chain mutex then just give up - we'll try
3860 * again on the next iteration.
3862 static void cache_reap(struct work_struct *w)
3864 struct kmem_cache *searchp;
3865 struct kmem_cache_node *n;
3866 int node = numa_mem_id();
3867 struct delayed_work *work = to_delayed_work(w);
3869 if (!mutex_trylock(&slab_mutex))
3870 /* Give up. Setup the next iteration. */
3873 list_for_each_entry(searchp, &slab_caches, list) {
3877 * We only take the node lock if absolutely necessary and we
3878 * have established with reasonable certainty that
3879 * we can do some work if the lock was obtained.
3881 n = get_node(searchp, node);
3883 reap_alien(searchp, n);
3885 drain_array(searchp, n, cpu_cache_get(searchp), 0, node);
3888 * These are racy checks but it does not matter
3889 * if we skip one check or scan twice.
3891 if (time_after(n->next_reap, jiffies))
3894 n->next_reap = jiffies + REAPTIMEOUT_NODE;
3896 drain_array(searchp, n, n->shared, 0, node);
3898 if (n->free_touched)
3899 n->free_touched = 0;
3903 freed = drain_freelist(searchp, n, (n->free_limit +
3904 5 * searchp->num - 1) / (5 * searchp->num));
3905 STATS_ADD_REAPED(searchp, freed);
3911 mutex_unlock(&slab_mutex);
3914 /* Set up the next iteration */
3915 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC));
3918 #ifdef CONFIG_SLABINFO
3919 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
3922 unsigned long active_objs;
3923 unsigned long num_objs;
3924 unsigned long active_slabs = 0;
3925 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3929 struct kmem_cache_node *n;
3933 for_each_kmem_cache_node(cachep, node, n) {
3936 spin_lock_irq(&n->list_lock);
3938 list_for_each_entry(page, &n->slabs_full, lru) {
3939 if (page->active != cachep->num && !error)
3940 error = "slabs_full accounting error";
3941 active_objs += cachep->num;
3944 list_for_each_entry(page, &n->slabs_partial, lru) {
3945 if (page->active == cachep->num && !error)
3946 error = "slabs_partial accounting error";
3947 if (!page->active && !error)
3948 error = "slabs_partial accounting error";
3949 active_objs += page->active;
3952 list_for_each_entry(page, &n->slabs_free, lru) {
3953 if (page->active && !error)
3954 error = "slabs_free accounting error";
3957 free_objects += n->free_objects;
3959 shared_avail += n->shared->avail;
3961 spin_unlock_irq(&n->list_lock);
3963 num_slabs += active_slabs;
3964 num_objs = num_slabs * cachep->num;
3965 if (num_objs - active_objs != free_objects && !error)
3966 error = "free_objects accounting error";
3968 name = cachep->name;
3970 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3972 sinfo->active_objs = active_objs;
3973 sinfo->num_objs = num_objs;
3974 sinfo->active_slabs = active_slabs;
3975 sinfo->num_slabs = num_slabs;
3976 sinfo->shared_avail = shared_avail;
3977 sinfo->limit = cachep->limit;
3978 sinfo->batchcount = cachep->batchcount;
3979 sinfo->shared = cachep->shared;
3980 sinfo->objects_per_slab = cachep->num;
3981 sinfo->cache_order = cachep->gfporder;
3984 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
3988 unsigned long high = cachep->high_mark;
3989 unsigned long allocs = cachep->num_allocations;
3990 unsigned long grown = cachep->grown;
3991 unsigned long reaped = cachep->reaped;
3992 unsigned long errors = cachep->errors;
3993 unsigned long max_freeable = cachep->max_freeable;
3994 unsigned long node_allocs = cachep->node_allocs;
3995 unsigned long node_frees = cachep->node_frees;
3996 unsigned long overflows = cachep->node_overflow;
3998 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
3999 "%4lu %4lu %4lu %4lu %4lu",
4000 allocs, high, grown,
4001 reaped, errors, max_freeable, node_allocs,
4002 node_frees, overflows);
4006 unsigned long allochit = atomic_read(&cachep->allochit);
4007 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4008 unsigned long freehit = atomic_read(&cachep->freehit);
4009 unsigned long freemiss = atomic_read(&cachep->freemiss);
4011 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4012 allochit, allocmiss, freehit, freemiss);
4017 #define MAX_SLABINFO_WRITE 128
4019 * slabinfo_write - Tuning for the slab allocator
4021 * @buffer: user buffer
4022 * @count: data length
4025 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4026 size_t count, loff_t *ppos)
4028 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4029 int limit, batchcount, shared, res;
4030 struct kmem_cache *cachep;
4032 if (count > MAX_SLABINFO_WRITE)
4034 if (copy_from_user(&kbuf, buffer, count))
4036 kbuf[MAX_SLABINFO_WRITE] = '\0';
4038 tmp = strchr(kbuf, ' ');
4043 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4046 /* Find the cache in the chain of caches. */
4047 mutex_lock(&slab_mutex);
4049 list_for_each_entry(cachep, &slab_caches, list) {
4050 if (!strcmp(cachep->name, kbuf)) {
4051 if (limit < 1 || batchcount < 1 ||
4052 batchcount > limit || shared < 0) {
4055 res = do_tune_cpucache(cachep, limit,
4062 mutex_unlock(&slab_mutex);
4068 #ifdef CONFIG_DEBUG_SLAB_LEAK
4070 static inline int add_caller(unsigned long *n, unsigned long v)
4080 unsigned long *q = p + 2 * i;
4094 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4100 static void handle_slab(unsigned long *n, struct kmem_cache *c,
4109 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
4112 for (j = page->active; j < c->num; j++) {
4113 if (get_free_obj(page, j) == i) {
4123 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4124 * mapping is established when actual object allocation and
4125 * we could mistakenly access the unmapped object in the cpu
4128 if (probe_kernel_read(&v, dbg_userword(c, p), sizeof(v)))
4131 if (!add_caller(n, v))
4136 static void show_symbol(struct seq_file *m, unsigned long address)
4138 #ifdef CONFIG_KALLSYMS
4139 unsigned long offset, size;
4140 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4142 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4143 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4145 seq_printf(m, " [%s]", modname);
4149 seq_printf(m, "%p", (void *)address);
4152 static int leaks_show(struct seq_file *m, void *p)
4154 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4156 struct kmem_cache_node *n;
4158 unsigned long *x = m->private;
4162 if (!(cachep->flags & SLAB_STORE_USER))
4164 if (!(cachep->flags & SLAB_RED_ZONE))
4168 * Set store_user_clean and start to grab stored user information
4169 * for all objects on this cache. If some alloc/free requests comes
4170 * during the processing, information would be wrong so restart
4174 set_store_user_clean(cachep);
4175 drain_cpu_caches(cachep);
4179 for_each_kmem_cache_node(cachep, node, n) {
4182 spin_lock_irq(&n->list_lock);
4184 list_for_each_entry(page, &n->slabs_full, lru)
4185 handle_slab(x, cachep, page);
4186 list_for_each_entry(page, &n->slabs_partial, lru)
4187 handle_slab(x, cachep, page);
4188 spin_unlock_irq(&n->list_lock);
4190 } while (!is_store_user_clean(cachep));
4192 name = cachep->name;
4194 /* Increase the buffer size */
4195 mutex_unlock(&slab_mutex);
4196 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4198 /* Too bad, we are really out */
4200 mutex_lock(&slab_mutex);
4203 *(unsigned long *)m->private = x[0] * 2;
4205 mutex_lock(&slab_mutex);
4206 /* Now make sure this entry will be retried */
4210 for (i = 0; i < x[1]; i++) {
4211 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4212 show_symbol(m, x[2*i+2]);
4219 static const struct seq_operations slabstats_op = {
4220 .start = slab_start,
4226 static int slabstats_open(struct inode *inode, struct file *file)
4230 n = __seq_open_private(file, &slabstats_op, PAGE_SIZE);
4234 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4239 static const struct file_operations proc_slabstats_operations = {
4240 .open = slabstats_open,
4242 .llseek = seq_lseek,
4243 .release = seq_release_private,
4247 static int __init slab_proc_init(void)
4249 #ifdef CONFIG_DEBUG_SLAB_LEAK
4250 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4254 module_init(slab_proc_init);
4258 * ksize - get the actual amount of memory allocated for a given object
4259 * @objp: Pointer to the object
4261 * kmalloc may internally round up allocations and return more memory
4262 * than requested. ksize() can be used to determine the actual amount of
4263 * memory allocated. The caller may use this additional memory, even though
4264 * a smaller amount of memory was initially specified with the kmalloc call.
4265 * The caller must guarantee that objp points to a valid object previously
4266 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4267 * must not be freed during the duration of the call.
4269 size_t ksize(const void *objp)
4272 if (unlikely(objp == ZERO_SIZE_PTR))
4275 return virt_to_cache(objp)->object_size;
4277 EXPORT_SYMBOL(ksize);