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 'cache_chain_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/kmemtrace.h>
106 #include <linux/rcupdate.h>
107 #include <linux/string.h>
108 #include <linux/uaccess.h>
109 #include <linux/nodemask.h>
110 #include <linux/kmemleak.h>
111 #include <linux/mempolicy.h>
112 #include <linux/mutex.h>
113 #include <linux/fault-inject.h>
114 #include <linux/rtmutex.h>
115 #include <linux/reciprocal_div.h>
116 #include <linux/debugobjects.h>
117 #include <linux/kmemcheck.h>
119 #include <asm/cacheflush.h>
120 #include <asm/tlbflush.h>
121 #include <asm/page.h>
124 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
125 * 0 for faster, smaller code (especially in the critical paths).
127 * STATS - 1 to collect stats for /proc/slabinfo.
128 * 0 for faster, smaller code (especially in the critical paths).
130 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
133 #ifdef CONFIG_DEBUG_SLAB
136 #define FORCED_DEBUG 1
140 #define FORCED_DEBUG 0
143 /* Shouldn't this be in a header file somewhere? */
144 #define BYTES_PER_WORD sizeof(void *)
145 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
147 #ifndef ARCH_KMALLOC_FLAGS
148 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
151 /* Legal flag mask for kmem_cache_create(). */
153 # define CREATE_MASK (SLAB_RED_ZONE | \
154 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
157 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
158 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
159 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
161 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
163 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
164 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
165 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
171 * Bufctl's are used for linking objs within a slab
174 * This implementation relies on "struct page" for locating the cache &
175 * slab an object belongs to.
176 * This allows the bufctl structure to be small (one int), but limits
177 * the number of objects a slab (not a cache) can contain when off-slab
178 * bufctls are used. The limit is the size of the largest general cache
179 * that does not use off-slab slabs.
180 * For 32bit archs with 4 kB pages, is this 56.
181 * This is not serious, as it is only for large objects, when it is unwise
182 * to have too many per slab.
183 * Note: This limit can be raised by introducing a general cache whose size
184 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
187 typedef unsigned int kmem_bufctl_t;
188 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
189 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
190 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
191 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
196 * Manages the objs in a slab. Placed either at the beginning of mem allocated
197 * for a slab, or allocated from an general cache.
198 * Slabs are chained into three list: fully used, partial, fully free slabs.
201 struct list_head list;
202 unsigned long colouroff;
203 void *s_mem; /* including colour offset */
204 unsigned int inuse; /* num of objs active in slab */
206 unsigned short nodeid;
212 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
213 * arrange for kmem_freepages to be called via RCU. This is useful if
214 * we need to approach a kernel structure obliquely, from its address
215 * obtained without the usual locking. We can lock the structure to
216 * stabilize it and check it's still at the given address, only if we
217 * can be sure that the memory has not been meanwhile reused for some
218 * other kind of object (which our subsystem's lock might corrupt).
220 * rcu_read_lock before reading the address, then rcu_read_unlock after
221 * taking the spinlock within the structure expected at that address.
223 * We assume struct slab_rcu can overlay struct slab when destroying.
226 struct rcu_head head;
227 struct kmem_cache *cachep;
235 * - LIFO ordering, to hand out cache-warm objects from _alloc
236 * - reduce the number of linked list operations
237 * - reduce spinlock operations
239 * The limit is stored in the per-cpu structure to reduce the data cache
246 unsigned int batchcount;
247 unsigned int touched;
250 * Must have this definition in here for the proper
251 * alignment of array_cache. Also simplifies accessing
257 * bootstrap: The caches do not work without cpuarrays anymore, but the
258 * cpuarrays are allocated from the generic caches...
260 #define BOOT_CPUCACHE_ENTRIES 1
261 struct arraycache_init {
262 struct array_cache cache;
263 void *entries[BOOT_CPUCACHE_ENTRIES];
267 * The slab lists for all objects.
270 struct list_head slabs_partial; /* partial list first, better asm code */
271 struct list_head slabs_full;
272 struct list_head slabs_free;
273 unsigned long free_objects;
274 unsigned int free_limit;
275 unsigned int colour_next; /* Per-node cache coloring */
276 spinlock_t list_lock;
277 struct array_cache *shared; /* shared per node */
278 struct array_cache **alien; /* on other nodes */
279 unsigned long next_reap; /* updated without locking */
280 int free_touched; /* updated without locking */
284 * Need this for bootstrapping a per node allocator.
286 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
287 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
288 #define CACHE_CACHE 0
289 #define SIZE_AC MAX_NUMNODES
290 #define SIZE_L3 (2 * MAX_NUMNODES)
292 static int drain_freelist(struct kmem_cache *cache,
293 struct kmem_list3 *l3, int tofree);
294 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
296 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
297 static void cache_reap(struct work_struct *unused);
300 * This function must be completely optimized away if a constant is passed to
301 * it. Mostly the same as what is in linux/slab.h except it returns an index.
303 static __always_inline int index_of(const size_t size)
305 extern void __bad_size(void);
307 if (__builtin_constant_p(size)) {
315 #include <linux/kmalloc_sizes.h>
323 static int slab_early_init = 1;
325 #define INDEX_AC index_of(sizeof(struct arraycache_init))
326 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
328 static void kmem_list3_init(struct kmem_list3 *parent)
330 INIT_LIST_HEAD(&parent->slabs_full);
331 INIT_LIST_HEAD(&parent->slabs_partial);
332 INIT_LIST_HEAD(&parent->slabs_free);
333 parent->shared = NULL;
334 parent->alien = NULL;
335 parent->colour_next = 0;
336 spin_lock_init(&parent->list_lock);
337 parent->free_objects = 0;
338 parent->free_touched = 0;
341 #define MAKE_LIST(cachep, listp, slab, nodeid) \
343 INIT_LIST_HEAD(listp); \
344 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
347 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
349 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
350 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
351 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
354 #define CFLGS_OFF_SLAB (0x80000000UL)
355 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
357 #define BATCHREFILL_LIMIT 16
359 * Optimization question: fewer reaps means less probability for unnessary
360 * cpucache drain/refill cycles.
362 * OTOH the cpuarrays can contain lots of objects,
363 * which could lock up otherwise freeable slabs.
365 #define REAPTIMEOUT_CPUC (2*HZ)
366 #define REAPTIMEOUT_LIST3 (4*HZ)
369 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
370 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
371 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
372 #define STATS_INC_GROWN(x) ((x)->grown++)
373 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
374 #define STATS_SET_HIGH(x) \
376 if ((x)->num_active > (x)->high_mark) \
377 (x)->high_mark = (x)->num_active; \
379 #define STATS_INC_ERR(x) ((x)->errors++)
380 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
381 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
382 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
383 #define STATS_SET_FREEABLE(x, i) \
385 if ((x)->max_freeable < i) \
386 (x)->max_freeable = i; \
388 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
389 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
390 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
391 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
393 #define STATS_INC_ACTIVE(x) do { } while (0)
394 #define STATS_DEC_ACTIVE(x) do { } while (0)
395 #define STATS_INC_ALLOCED(x) do { } while (0)
396 #define STATS_INC_GROWN(x) do { } while (0)
397 #define STATS_ADD_REAPED(x,y) do { } while (0)
398 #define STATS_SET_HIGH(x) do { } while (0)
399 #define STATS_INC_ERR(x) do { } while (0)
400 #define STATS_INC_NODEALLOCS(x) do { } while (0)
401 #define STATS_INC_NODEFREES(x) do { } while (0)
402 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
403 #define STATS_SET_FREEABLE(x, i) do { } while (0)
404 #define STATS_INC_ALLOCHIT(x) do { } while (0)
405 #define STATS_INC_ALLOCMISS(x) do { } while (0)
406 #define STATS_INC_FREEHIT(x) do { } while (0)
407 #define STATS_INC_FREEMISS(x) do { } while (0)
413 * memory layout of objects:
415 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
416 * the end of an object is aligned with the end of the real
417 * allocation. Catches writes behind the end of the allocation.
418 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
420 * cachep->obj_offset: The real object.
421 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
422 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
423 * [BYTES_PER_WORD long]
425 static int obj_offset(struct kmem_cache *cachep)
427 return cachep->obj_offset;
430 static int obj_size(struct kmem_cache *cachep)
432 return cachep->obj_size;
435 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
437 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
438 return (unsigned long long*) (objp + obj_offset(cachep) -
439 sizeof(unsigned long long));
442 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
444 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
445 if (cachep->flags & SLAB_STORE_USER)
446 return (unsigned long long *)(objp + cachep->buffer_size -
447 sizeof(unsigned long long) -
449 return (unsigned long long *) (objp + cachep->buffer_size -
450 sizeof(unsigned long long));
453 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
455 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
456 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
461 #define obj_offset(x) 0
462 #define obj_size(cachep) (cachep->buffer_size)
463 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
464 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
465 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
469 #ifdef CONFIG_TRACING
470 size_t slab_buffer_size(struct kmem_cache *cachep)
472 return cachep->buffer_size;
474 EXPORT_SYMBOL(slab_buffer_size);
478 * Do not go above this order unless 0 objects fit into the slab.
480 #define BREAK_GFP_ORDER_HI 1
481 #define BREAK_GFP_ORDER_LO 0
482 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
485 * Functions for storing/retrieving the cachep and or slab from the page
486 * allocator. These are used to find the slab an obj belongs to. With kfree(),
487 * these are used to find the cache which an obj belongs to.
489 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
491 page->lru.next = (struct list_head *)cache;
494 static inline struct kmem_cache *page_get_cache(struct page *page)
496 page = compound_head(page);
497 BUG_ON(!PageSlab(page));
498 return (struct kmem_cache *)page->lru.next;
501 static inline void page_set_slab(struct page *page, struct slab *slab)
503 page->lru.prev = (struct list_head *)slab;
506 static inline struct slab *page_get_slab(struct page *page)
508 BUG_ON(!PageSlab(page));
509 return (struct slab *)page->lru.prev;
512 static inline struct kmem_cache *virt_to_cache(const void *obj)
514 struct page *page = virt_to_head_page(obj);
515 return page_get_cache(page);
518 static inline struct slab *virt_to_slab(const void *obj)
520 struct page *page = virt_to_head_page(obj);
521 return page_get_slab(page);
524 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
527 return slab->s_mem + cache->buffer_size * idx;
531 * We want to avoid an expensive divide : (offset / cache->buffer_size)
532 * Using the fact that buffer_size is a constant for a particular cache,
533 * we can replace (offset / cache->buffer_size) by
534 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
536 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
537 const struct slab *slab, void *obj)
539 u32 offset = (obj - slab->s_mem);
540 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
544 * These are the default caches for kmalloc. Custom caches can have other sizes.
546 struct cache_sizes malloc_sizes[] = {
547 #define CACHE(x) { .cs_size = (x) },
548 #include <linux/kmalloc_sizes.h>
552 EXPORT_SYMBOL(malloc_sizes);
554 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
560 static struct cache_names __initdata cache_names[] = {
561 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
562 #include <linux/kmalloc_sizes.h>
567 static struct arraycache_init initarray_cache __initdata =
568 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
569 static struct arraycache_init initarray_generic =
570 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
572 /* internal cache of cache description objs */
573 static struct kmem_cache cache_cache = {
575 .limit = BOOT_CPUCACHE_ENTRIES,
577 .buffer_size = sizeof(struct kmem_cache),
578 .name = "kmem_cache",
581 #define BAD_ALIEN_MAGIC 0x01020304ul
584 * chicken and egg problem: delay the per-cpu array allocation
585 * until the general caches are up.
596 * used by boot code to determine if it can use slab based allocator
598 int slab_is_available(void)
600 return g_cpucache_up >= EARLY;
603 #ifdef CONFIG_LOCKDEP
606 * Slab sometimes uses the kmalloc slabs to store the slab headers
607 * for other slabs "off slab".
608 * The locking for this is tricky in that it nests within the locks
609 * of all other slabs in a few places; to deal with this special
610 * locking we put on-slab caches into a separate lock-class.
612 * We set lock class for alien array caches which are up during init.
613 * The lock annotation will be lost if all cpus of a node goes down and
614 * then comes back up during hotplug
616 static struct lock_class_key on_slab_l3_key;
617 static struct lock_class_key on_slab_alc_key;
619 static void init_node_lock_keys(int q)
621 struct cache_sizes *s = malloc_sizes;
623 if (g_cpucache_up != FULL)
626 for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
627 struct array_cache **alc;
628 struct kmem_list3 *l3;
631 l3 = s->cs_cachep->nodelists[q];
632 if (!l3 || OFF_SLAB(s->cs_cachep))
634 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
637 * FIXME: This check for BAD_ALIEN_MAGIC
638 * should go away when common slab code is taught to
639 * work even without alien caches.
640 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
641 * for alloc_alien_cache,
643 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
647 lockdep_set_class(&alc[r]->lock,
653 static inline void init_lock_keys(void)
658 init_node_lock_keys(node);
661 static void init_node_lock_keys(int q)
665 static inline void init_lock_keys(void)
671 * Guard access to the cache-chain.
673 static DEFINE_MUTEX(cache_chain_mutex);
674 static struct list_head cache_chain;
676 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
678 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
680 return cachep->array[smp_processor_id()];
683 static inline struct kmem_cache *__find_general_cachep(size_t size,
686 struct cache_sizes *csizep = malloc_sizes;
689 /* This happens if someone tries to call
690 * kmem_cache_create(), or __kmalloc(), before
691 * the generic caches are initialized.
693 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
696 return ZERO_SIZE_PTR;
698 while (size > csizep->cs_size)
702 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
703 * has cs_{dma,}cachep==NULL. Thus no special case
704 * for large kmalloc calls required.
706 #ifdef CONFIG_ZONE_DMA
707 if (unlikely(gfpflags & GFP_DMA))
708 return csizep->cs_dmacachep;
710 return csizep->cs_cachep;
713 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
715 return __find_general_cachep(size, gfpflags);
718 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
720 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
724 * Calculate the number of objects and left-over bytes for a given buffer size.
726 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
727 size_t align, int flags, size_t *left_over,
732 size_t slab_size = PAGE_SIZE << gfporder;
735 * The slab management structure can be either off the slab or
736 * on it. For the latter case, the memory allocated for a
740 * - One kmem_bufctl_t for each object
741 * - Padding to respect alignment of @align
742 * - @buffer_size bytes for each object
744 * If the slab management structure is off the slab, then the
745 * alignment will already be calculated into the size. Because
746 * the slabs are all pages aligned, the objects will be at the
747 * correct alignment when allocated.
749 if (flags & CFLGS_OFF_SLAB) {
751 nr_objs = slab_size / buffer_size;
753 if (nr_objs > SLAB_LIMIT)
754 nr_objs = SLAB_LIMIT;
757 * Ignore padding for the initial guess. The padding
758 * is at most @align-1 bytes, and @buffer_size is at
759 * least @align. In the worst case, this result will
760 * be one greater than the number of objects that fit
761 * into the memory allocation when taking the padding
764 nr_objs = (slab_size - sizeof(struct slab)) /
765 (buffer_size + sizeof(kmem_bufctl_t));
768 * This calculated number will be either the right
769 * amount, or one greater than what we want.
771 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
775 if (nr_objs > SLAB_LIMIT)
776 nr_objs = SLAB_LIMIT;
778 mgmt_size = slab_mgmt_size(nr_objs, align);
781 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
784 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
786 static void __slab_error(const char *function, struct kmem_cache *cachep,
789 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
790 function, cachep->name, msg);
795 * By default on NUMA we use alien caches to stage the freeing of
796 * objects allocated from other nodes. This causes massive memory
797 * inefficiencies when using fake NUMA setup to split memory into a
798 * large number of small nodes, so it can be disabled on the command
802 static int use_alien_caches __read_mostly = 1;
803 static int __init noaliencache_setup(char *s)
805 use_alien_caches = 0;
808 __setup("noaliencache", noaliencache_setup);
812 * Special reaping functions for NUMA systems called from cache_reap().
813 * These take care of doing round robin flushing of alien caches (containing
814 * objects freed on different nodes from which they were allocated) and the
815 * flushing of remote pcps by calling drain_node_pages.
817 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
819 static void init_reap_node(int cpu)
823 node = next_node(cpu_to_node(cpu), node_online_map);
824 if (node == MAX_NUMNODES)
825 node = first_node(node_online_map);
827 per_cpu(slab_reap_node, cpu) = node;
830 static void next_reap_node(void)
832 int node = __get_cpu_var(slab_reap_node);
834 node = next_node(node, node_online_map);
835 if (unlikely(node >= MAX_NUMNODES))
836 node = first_node(node_online_map);
837 __get_cpu_var(slab_reap_node) = node;
841 #define init_reap_node(cpu) do { } while (0)
842 #define next_reap_node(void) do { } while (0)
846 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
847 * via the workqueue/eventd.
848 * Add the CPU number into the expiration time to minimize the possibility of
849 * the CPUs getting into lockstep and contending for the global cache chain
852 static void __cpuinit start_cpu_timer(int cpu)
854 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
857 * When this gets called from do_initcalls via cpucache_init(),
858 * init_workqueues() has already run, so keventd will be setup
861 if (keventd_up() && reap_work->work.func == NULL) {
863 INIT_DELAYED_WORK(reap_work, cache_reap);
864 schedule_delayed_work_on(cpu, reap_work,
865 __round_jiffies_relative(HZ, cpu));
869 static struct array_cache *alloc_arraycache(int node, int entries,
870 int batchcount, gfp_t gfp)
872 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
873 struct array_cache *nc = NULL;
875 nc = kmalloc_node(memsize, gfp, node);
877 * The array_cache structures contain pointers to free object.
878 * However, when such objects are allocated or transfered to another
879 * cache the pointers are not cleared and they could be counted as
880 * valid references during a kmemleak scan. Therefore, kmemleak must
881 * not scan such objects.
883 kmemleak_no_scan(nc);
887 nc->batchcount = batchcount;
889 spin_lock_init(&nc->lock);
895 * Transfer objects in one arraycache to another.
896 * Locking must be handled by the caller.
898 * Return the number of entries transferred.
900 static int transfer_objects(struct array_cache *to,
901 struct array_cache *from, unsigned int max)
903 /* Figure out how many entries to transfer */
904 int nr = min(min(from->avail, max), to->limit - to->avail);
909 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
919 #define drain_alien_cache(cachep, alien) do { } while (0)
920 #define reap_alien(cachep, l3) do { } while (0)
922 static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
924 return (struct array_cache **)BAD_ALIEN_MAGIC;
927 static inline void free_alien_cache(struct array_cache **ac_ptr)
931 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
936 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
942 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
943 gfp_t flags, int nodeid)
948 #else /* CONFIG_NUMA */
950 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
951 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
953 static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
955 struct array_cache **ac_ptr;
956 int memsize = sizeof(void *) * nr_node_ids;
961 ac_ptr = kzalloc_node(memsize, gfp, node);
964 if (i == node || !node_online(i))
966 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
968 for (i--; i >= 0; i--)
978 static void free_alien_cache(struct array_cache **ac_ptr)
989 static void __drain_alien_cache(struct kmem_cache *cachep,
990 struct array_cache *ac, int node)
992 struct kmem_list3 *rl3 = cachep->nodelists[node];
995 spin_lock(&rl3->list_lock);
997 * Stuff objects into the remote nodes shared array first.
998 * That way we could avoid the overhead of putting the objects
999 * into the free lists and getting them back later.
1002 transfer_objects(rl3->shared, ac, ac->limit);
1004 free_block(cachep, ac->entry, ac->avail, node);
1006 spin_unlock(&rl3->list_lock);
1011 * Called from cache_reap() to regularly drain alien caches round robin.
1013 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1015 int node = __get_cpu_var(slab_reap_node);
1018 struct array_cache *ac = l3->alien[node];
1020 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1021 __drain_alien_cache(cachep, ac, node);
1022 spin_unlock_irq(&ac->lock);
1027 static void drain_alien_cache(struct kmem_cache *cachep,
1028 struct array_cache **alien)
1031 struct array_cache *ac;
1032 unsigned long flags;
1034 for_each_online_node(i) {
1037 spin_lock_irqsave(&ac->lock, flags);
1038 __drain_alien_cache(cachep, ac, i);
1039 spin_unlock_irqrestore(&ac->lock, flags);
1044 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1046 struct slab *slabp = virt_to_slab(objp);
1047 int nodeid = slabp->nodeid;
1048 struct kmem_list3 *l3;
1049 struct array_cache *alien = NULL;
1052 node = numa_node_id();
1055 * Make sure we are not freeing a object from another node to the array
1056 * cache on this cpu.
1058 if (likely(slabp->nodeid == node))
1061 l3 = cachep->nodelists[node];
1062 STATS_INC_NODEFREES(cachep);
1063 if (l3->alien && l3->alien[nodeid]) {
1064 alien = l3->alien[nodeid];
1065 spin_lock(&alien->lock);
1066 if (unlikely(alien->avail == alien->limit)) {
1067 STATS_INC_ACOVERFLOW(cachep);
1068 __drain_alien_cache(cachep, alien, nodeid);
1070 alien->entry[alien->avail++] = objp;
1071 spin_unlock(&alien->lock);
1073 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1074 free_block(cachep, &objp, 1, nodeid);
1075 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1081 static void __cpuinit cpuup_canceled(long cpu)
1083 struct kmem_cache *cachep;
1084 struct kmem_list3 *l3 = NULL;
1085 int node = cpu_to_node(cpu);
1086 const struct cpumask *mask = cpumask_of_node(node);
1088 list_for_each_entry(cachep, &cache_chain, next) {
1089 struct array_cache *nc;
1090 struct array_cache *shared;
1091 struct array_cache **alien;
1093 /* cpu is dead; no one can alloc from it. */
1094 nc = cachep->array[cpu];
1095 cachep->array[cpu] = NULL;
1096 l3 = cachep->nodelists[node];
1099 goto free_array_cache;
1101 spin_lock_irq(&l3->list_lock);
1103 /* Free limit for this kmem_list3 */
1104 l3->free_limit -= cachep->batchcount;
1106 free_block(cachep, nc->entry, nc->avail, node);
1108 if (!cpumask_empty(mask)) {
1109 spin_unlock_irq(&l3->list_lock);
1110 goto free_array_cache;
1113 shared = l3->shared;
1115 free_block(cachep, shared->entry,
1116 shared->avail, node);
1123 spin_unlock_irq(&l3->list_lock);
1127 drain_alien_cache(cachep, alien);
1128 free_alien_cache(alien);
1134 * In the previous loop, all the objects were freed to
1135 * the respective cache's slabs, now we can go ahead and
1136 * shrink each nodelist to its limit.
1138 list_for_each_entry(cachep, &cache_chain, next) {
1139 l3 = cachep->nodelists[node];
1142 drain_freelist(cachep, l3, l3->free_objects);
1146 static int __cpuinit cpuup_prepare(long cpu)
1148 struct kmem_cache *cachep;
1149 struct kmem_list3 *l3 = NULL;
1150 int node = cpu_to_node(cpu);
1151 const int memsize = sizeof(struct kmem_list3);
1154 * We need to do this right in the beginning since
1155 * alloc_arraycache's are going to use this list.
1156 * kmalloc_node allows us to add the slab to the right
1157 * kmem_list3 and not this cpu's kmem_list3
1160 list_for_each_entry(cachep, &cache_chain, next) {
1162 * Set up the size64 kmemlist for cpu before we can
1163 * begin anything. Make sure some other cpu on this
1164 * node has not already allocated this
1166 if (!cachep->nodelists[node]) {
1167 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1170 kmem_list3_init(l3);
1171 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1172 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1175 * The l3s don't come and go as CPUs come and
1176 * go. cache_chain_mutex is sufficient
1179 cachep->nodelists[node] = l3;
1182 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1183 cachep->nodelists[node]->free_limit =
1184 (1 + nr_cpus_node(node)) *
1185 cachep->batchcount + cachep->num;
1186 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1190 * Now we can go ahead with allocating the shared arrays and
1193 list_for_each_entry(cachep, &cache_chain, next) {
1194 struct array_cache *nc;
1195 struct array_cache *shared = NULL;
1196 struct array_cache **alien = NULL;
1198 nc = alloc_arraycache(node, cachep->limit,
1199 cachep->batchcount, GFP_KERNEL);
1202 if (cachep->shared) {
1203 shared = alloc_arraycache(node,
1204 cachep->shared * cachep->batchcount,
1205 0xbaadf00d, GFP_KERNEL);
1211 if (use_alien_caches) {
1212 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1219 cachep->array[cpu] = nc;
1220 l3 = cachep->nodelists[node];
1223 spin_lock_irq(&l3->list_lock);
1226 * We are serialised from CPU_DEAD or
1227 * CPU_UP_CANCELLED by the cpucontrol lock
1229 l3->shared = shared;
1238 spin_unlock_irq(&l3->list_lock);
1240 free_alien_cache(alien);
1242 init_node_lock_keys(node);
1246 cpuup_canceled(cpu);
1250 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1251 unsigned long action, void *hcpu)
1253 long cpu = (long)hcpu;
1257 case CPU_UP_PREPARE:
1258 case CPU_UP_PREPARE_FROZEN:
1259 mutex_lock(&cache_chain_mutex);
1260 err = cpuup_prepare(cpu);
1261 mutex_unlock(&cache_chain_mutex);
1264 case CPU_ONLINE_FROZEN:
1265 start_cpu_timer(cpu);
1267 #ifdef CONFIG_HOTPLUG_CPU
1268 case CPU_DOWN_PREPARE:
1269 case CPU_DOWN_PREPARE_FROZEN:
1271 * Shutdown cache reaper. Note that the cache_chain_mutex is
1272 * held so that if cache_reap() is invoked it cannot do
1273 * anything expensive but will only modify reap_work
1274 * and reschedule the timer.
1276 cancel_rearming_delayed_work(&per_cpu(slab_reap_work, cpu));
1277 /* Now the cache_reaper is guaranteed to be not running. */
1278 per_cpu(slab_reap_work, cpu).work.func = NULL;
1280 case CPU_DOWN_FAILED:
1281 case CPU_DOWN_FAILED_FROZEN:
1282 start_cpu_timer(cpu);
1285 case CPU_DEAD_FROZEN:
1287 * Even if all the cpus of a node are down, we don't free the
1288 * kmem_list3 of any cache. This to avoid a race between
1289 * cpu_down, and a kmalloc allocation from another cpu for
1290 * memory from the node of the cpu going down. The list3
1291 * structure is usually allocated from kmem_cache_create() and
1292 * gets destroyed at kmem_cache_destroy().
1296 case CPU_UP_CANCELED:
1297 case CPU_UP_CANCELED_FROZEN:
1298 mutex_lock(&cache_chain_mutex);
1299 cpuup_canceled(cpu);
1300 mutex_unlock(&cache_chain_mutex);
1303 return err ? NOTIFY_BAD : NOTIFY_OK;
1306 static struct notifier_block __cpuinitdata cpucache_notifier = {
1307 &cpuup_callback, NULL, 0
1311 * swap the static kmem_list3 with kmalloced memory
1313 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1316 struct kmem_list3 *ptr;
1318 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
1321 memcpy(ptr, list, sizeof(struct kmem_list3));
1323 * Do not assume that spinlocks can be initialized via memcpy:
1325 spin_lock_init(&ptr->list_lock);
1327 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1328 cachep->nodelists[nodeid] = ptr;
1332 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1333 * size of kmem_list3.
1335 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1339 for_each_online_node(node) {
1340 cachep->nodelists[node] = &initkmem_list3[index + node];
1341 cachep->nodelists[node]->next_reap = jiffies +
1343 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1348 * Initialisation. Called after the page allocator have been initialised and
1349 * before smp_init().
1351 void __init kmem_cache_init(void)
1354 struct cache_sizes *sizes;
1355 struct cache_names *names;
1360 if (num_possible_nodes() == 1)
1361 use_alien_caches = 0;
1363 for (i = 0; i < NUM_INIT_LISTS; i++) {
1364 kmem_list3_init(&initkmem_list3[i]);
1365 if (i < MAX_NUMNODES)
1366 cache_cache.nodelists[i] = NULL;
1368 set_up_list3s(&cache_cache, CACHE_CACHE);
1371 * Fragmentation resistance on low memory - only use bigger
1372 * page orders on machines with more than 32MB of memory.
1374 if (totalram_pages > (32 << 20) >> PAGE_SHIFT)
1375 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1377 /* Bootstrap is tricky, because several objects are allocated
1378 * from caches that do not exist yet:
1379 * 1) initialize the cache_cache cache: it contains the struct
1380 * kmem_cache structures of all caches, except cache_cache itself:
1381 * cache_cache is statically allocated.
1382 * Initially an __init data area is used for the head array and the
1383 * kmem_list3 structures, it's replaced with a kmalloc allocated
1384 * array at the end of the bootstrap.
1385 * 2) Create the first kmalloc cache.
1386 * The struct kmem_cache for the new cache is allocated normally.
1387 * An __init data area is used for the head array.
1388 * 3) Create the remaining kmalloc caches, with minimally sized
1390 * 4) Replace the __init data head arrays for cache_cache and the first
1391 * kmalloc cache with kmalloc allocated arrays.
1392 * 5) Replace the __init data for kmem_list3 for cache_cache and
1393 * the other cache's with kmalloc allocated memory.
1394 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1397 node = numa_node_id();
1399 /* 1) create the cache_cache */
1400 INIT_LIST_HEAD(&cache_chain);
1401 list_add(&cache_cache.next, &cache_chain);
1402 cache_cache.colour_off = cache_line_size();
1403 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1404 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
1407 * struct kmem_cache size depends on nr_node_ids, which
1408 * can be less than MAX_NUMNODES.
1410 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1411 nr_node_ids * sizeof(struct kmem_list3 *);
1413 cache_cache.obj_size = cache_cache.buffer_size;
1415 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1417 cache_cache.reciprocal_buffer_size =
1418 reciprocal_value(cache_cache.buffer_size);
1420 for (order = 0; order < MAX_ORDER; order++) {
1421 cache_estimate(order, cache_cache.buffer_size,
1422 cache_line_size(), 0, &left_over, &cache_cache.num);
1423 if (cache_cache.num)
1426 BUG_ON(!cache_cache.num);
1427 cache_cache.gfporder = order;
1428 cache_cache.colour = left_over / cache_cache.colour_off;
1429 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1430 sizeof(struct slab), cache_line_size());
1432 /* 2+3) create the kmalloc caches */
1433 sizes = malloc_sizes;
1434 names = cache_names;
1437 * Initialize the caches that provide memory for the array cache and the
1438 * kmem_list3 structures first. Without this, further allocations will
1442 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1443 sizes[INDEX_AC].cs_size,
1444 ARCH_KMALLOC_MINALIGN,
1445 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1448 if (INDEX_AC != INDEX_L3) {
1449 sizes[INDEX_L3].cs_cachep =
1450 kmem_cache_create(names[INDEX_L3].name,
1451 sizes[INDEX_L3].cs_size,
1452 ARCH_KMALLOC_MINALIGN,
1453 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1457 slab_early_init = 0;
1459 while (sizes->cs_size != ULONG_MAX) {
1461 * For performance, all the general caches are L1 aligned.
1462 * This should be particularly beneficial on SMP boxes, as it
1463 * eliminates "false sharing".
1464 * Note for systems short on memory removing the alignment will
1465 * allow tighter packing of the smaller caches.
1467 if (!sizes->cs_cachep) {
1468 sizes->cs_cachep = kmem_cache_create(names->name,
1470 ARCH_KMALLOC_MINALIGN,
1471 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1474 #ifdef CONFIG_ZONE_DMA
1475 sizes->cs_dmacachep = kmem_cache_create(
1478 ARCH_KMALLOC_MINALIGN,
1479 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1486 /* 4) Replace the bootstrap head arrays */
1488 struct array_cache *ptr;
1490 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1492 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1493 memcpy(ptr, cpu_cache_get(&cache_cache),
1494 sizeof(struct arraycache_init));
1496 * Do not assume that spinlocks can be initialized via memcpy:
1498 spin_lock_init(&ptr->lock);
1500 cache_cache.array[smp_processor_id()] = ptr;
1502 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1504 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1505 != &initarray_generic.cache);
1506 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1507 sizeof(struct arraycache_init));
1509 * Do not assume that spinlocks can be initialized via memcpy:
1511 spin_lock_init(&ptr->lock);
1513 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1516 /* 5) Replace the bootstrap kmem_list3's */
1520 for_each_online_node(nid) {
1521 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
1523 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1524 &initkmem_list3[SIZE_AC + nid], nid);
1526 if (INDEX_AC != INDEX_L3) {
1527 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1528 &initkmem_list3[SIZE_L3 + nid], nid);
1533 g_cpucache_up = EARLY;
1536 void __init kmem_cache_init_late(void)
1538 struct kmem_cache *cachep;
1540 /* 6) resize the head arrays to their final sizes */
1541 mutex_lock(&cache_chain_mutex);
1542 list_for_each_entry(cachep, &cache_chain, next)
1543 if (enable_cpucache(cachep, GFP_NOWAIT))
1545 mutex_unlock(&cache_chain_mutex);
1548 g_cpucache_up = FULL;
1550 /* Annotate slab for lockdep -- annotate the malloc caches */
1554 * Register a cpu startup notifier callback that initializes
1555 * cpu_cache_get for all new cpus
1557 register_cpu_notifier(&cpucache_notifier);
1560 * The reap timers are started later, with a module init call: That part
1561 * of the kernel is not yet operational.
1565 static int __init cpucache_init(void)
1570 * Register the timers that return unneeded pages to the page allocator
1572 for_each_online_cpu(cpu)
1573 start_cpu_timer(cpu);
1576 __initcall(cpucache_init);
1579 * Interface to system's page allocator. No need to hold the cache-lock.
1581 * If we requested dmaable memory, we will get it. Even if we
1582 * did not request dmaable memory, we might get it, but that
1583 * would be relatively rare and ignorable.
1585 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1593 * Nommu uses slab's for process anonymous memory allocations, and thus
1594 * requires __GFP_COMP to properly refcount higher order allocations
1596 flags |= __GFP_COMP;
1599 flags |= cachep->gfpflags;
1600 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1601 flags |= __GFP_RECLAIMABLE;
1603 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1607 nr_pages = (1 << cachep->gfporder);
1608 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1609 add_zone_page_state(page_zone(page),
1610 NR_SLAB_RECLAIMABLE, nr_pages);
1612 add_zone_page_state(page_zone(page),
1613 NR_SLAB_UNRECLAIMABLE, nr_pages);
1614 for (i = 0; i < nr_pages; i++)
1615 __SetPageSlab(page + i);
1617 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1618 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1621 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1623 kmemcheck_mark_unallocated_pages(page, nr_pages);
1626 return page_address(page);
1630 * Interface to system's page release.
1632 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1634 unsigned long i = (1 << cachep->gfporder);
1635 struct page *page = virt_to_page(addr);
1636 const unsigned long nr_freed = i;
1638 kmemcheck_free_shadow(page, cachep->gfporder);
1640 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1641 sub_zone_page_state(page_zone(page),
1642 NR_SLAB_RECLAIMABLE, nr_freed);
1644 sub_zone_page_state(page_zone(page),
1645 NR_SLAB_UNRECLAIMABLE, nr_freed);
1647 BUG_ON(!PageSlab(page));
1648 __ClearPageSlab(page);
1651 if (current->reclaim_state)
1652 current->reclaim_state->reclaimed_slab += nr_freed;
1653 free_pages((unsigned long)addr, cachep->gfporder);
1656 static void kmem_rcu_free(struct rcu_head *head)
1658 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1659 struct kmem_cache *cachep = slab_rcu->cachep;
1661 kmem_freepages(cachep, slab_rcu->addr);
1662 if (OFF_SLAB(cachep))
1663 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1668 #ifdef CONFIG_DEBUG_PAGEALLOC
1669 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1670 unsigned long caller)
1672 int size = obj_size(cachep);
1674 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1676 if (size < 5 * sizeof(unsigned long))
1679 *addr++ = 0x12345678;
1681 *addr++ = smp_processor_id();
1682 size -= 3 * sizeof(unsigned long);
1684 unsigned long *sptr = &caller;
1685 unsigned long svalue;
1687 while (!kstack_end(sptr)) {
1689 if (kernel_text_address(svalue)) {
1691 size -= sizeof(unsigned long);
1692 if (size <= sizeof(unsigned long))
1698 *addr++ = 0x87654321;
1702 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1704 int size = obj_size(cachep);
1705 addr = &((char *)addr)[obj_offset(cachep)];
1707 memset(addr, val, size);
1708 *(unsigned char *)(addr + size - 1) = POISON_END;
1711 static void dump_line(char *data, int offset, int limit)
1714 unsigned char error = 0;
1717 printk(KERN_ERR "%03x:", offset);
1718 for (i = 0; i < limit; i++) {
1719 if (data[offset + i] != POISON_FREE) {
1720 error = data[offset + i];
1723 printk(" %02x", (unsigned char)data[offset + i]);
1727 if (bad_count == 1) {
1728 error ^= POISON_FREE;
1729 if (!(error & (error - 1))) {
1730 printk(KERN_ERR "Single bit error detected. Probably "
1733 printk(KERN_ERR "Run memtest86+ or a similar memory "
1736 printk(KERN_ERR "Run a memory test tool.\n");
1745 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1750 if (cachep->flags & SLAB_RED_ZONE) {
1751 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1752 *dbg_redzone1(cachep, objp),
1753 *dbg_redzone2(cachep, objp));
1756 if (cachep->flags & SLAB_STORE_USER) {
1757 printk(KERN_ERR "Last user: [<%p>]",
1758 *dbg_userword(cachep, objp));
1759 print_symbol("(%s)",
1760 (unsigned long)*dbg_userword(cachep, objp));
1763 realobj = (char *)objp + obj_offset(cachep);
1764 size = obj_size(cachep);
1765 for (i = 0; i < size && lines; i += 16, lines--) {
1768 if (i + limit > size)
1770 dump_line(realobj, i, limit);
1774 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1780 realobj = (char *)objp + obj_offset(cachep);
1781 size = obj_size(cachep);
1783 for (i = 0; i < size; i++) {
1784 char exp = POISON_FREE;
1787 if (realobj[i] != exp) {
1793 "Slab corruption: %s start=%p, len=%d\n",
1794 cachep->name, realobj, size);
1795 print_objinfo(cachep, objp, 0);
1797 /* Hexdump the affected line */
1800 if (i + limit > size)
1802 dump_line(realobj, i, limit);
1805 /* Limit to 5 lines */
1811 /* Print some data about the neighboring objects, if they
1814 struct slab *slabp = virt_to_slab(objp);
1817 objnr = obj_to_index(cachep, slabp, objp);
1819 objp = index_to_obj(cachep, slabp, objnr - 1);
1820 realobj = (char *)objp + obj_offset(cachep);
1821 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1823 print_objinfo(cachep, objp, 2);
1825 if (objnr + 1 < cachep->num) {
1826 objp = index_to_obj(cachep, slabp, objnr + 1);
1827 realobj = (char *)objp + obj_offset(cachep);
1828 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1830 print_objinfo(cachep, objp, 2);
1837 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1840 for (i = 0; i < cachep->num; i++) {
1841 void *objp = index_to_obj(cachep, slabp, i);
1843 if (cachep->flags & SLAB_POISON) {
1844 #ifdef CONFIG_DEBUG_PAGEALLOC
1845 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1847 kernel_map_pages(virt_to_page(objp),
1848 cachep->buffer_size / PAGE_SIZE, 1);
1850 check_poison_obj(cachep, objp);
1852 check_poison_obj(cachep, objp);
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, struct slab *slabp)
1872 * slab_destroy - destroy and release all objects in a slab
1873 * @cachep: cache pointer being destroyed
1874 * @slabp: slab pointer being destroyed
1876 * Destroy all the objs in a slab, and release the mem back to the system.
1877 * Before calling the slab must have been unlinked from the cache. The
1878 * cache-lock is not held/needed.
1880 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1882 void *addr = slabp->s_mem - slabp->colouroff;
1884 slab_destroy_debugcheck(cachep, slabp);
1885 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1886 struct slab_rcu *slab_rcu;
1888 slab_rcu = (struct slab_rcu *)slabp;
1889 slab_rcu->cachep = cachep;
1890 slab_rcu->addr = addr;
1891 call_rcu(&slab_rcu->head, kmem_rcu_free);
1893 kmem_freepages(cachep, addr);
1894 if (OFF_SLAB(cachep))
1895 kmem_cache_free(cachep->slabp_cache, slabp);
1899 static void __kmem_cache_destroy(struct kmem_cache *cachep)
1902 struct kmem_list3 *l3;
1904 for_each_online_cpu(i)
1905 kfree(cachep->array[i]);
1907 /* NUMA: free the list3 structures */
1908 for_each_online_node(i) {
1909 l3 = cachep->nodelists[i];
1912 free_alien_cache(l3->alien);
1916 kmem_cache_free(&cache_cache, cachep);
1921 * calculate_slab_order - calculate size (page order) of slabs
1922 * @cachep: pointer to the cache that is being created
1923 * @size: size of objects to be created in this cache.
1924 * @align: required alignment for the objects.
1925 * @flags: slab allocation flags
1927 * Also calculates the number of objects per slab.
1929 * This could be made much more intelligent. For now, try to avoid using
1930 * high order pages for slabs. When the gfp() functions are more friendly
1931 * towards high-order requests, this should be changed.
1933 static size_t calculate_slab_order(struct kmem_cache *cachep,
1934 size_t size, size_t align, unsigned long flags)
1936 unsigned long offslab_limit;
1937 size_t left_over = 0;
1940 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
1944 cache_estimate(gfporder, size, align, flags, &remainder, &num);
1948 if (flags & CFLGS_OFF_SLAB) {
1950 * Max number of objs-per-slab for caches which
1951 * use off-slab slabs. Needed to avoid a possible
1952 * looping condition in cache_grow().
1954 offslab_limit = size - sizeof(struct slab);
1955 offslab_limit /= sizeof(kmem_bufctl_t);
1957 if (num > offslab_limit)
1961 /* Found something acceptable - save it away */
1963 cachep->gfporder = gfporder;
1964 left_over = remainder;
1967 * A VFS-reclaimable slab tends to have most allocations
1968 * as GFP_NOFS and we really don't want to have to be allocating
1969 * higher-order pages when we are unable to shrink dcache.
1971 if (flags & SLAB_RECLAIM_ACCOUNT)
1975 * Large number of objects is good, but very large slabs are
1976 * currently bad for the gfp()s.
1978 if (gfporder >= slab_break_gfp_order)
1982 * Acceptable internal fragmentation?
1984 if (left_over * 8 <= (PAGE_SIZE << gfporder))
1990 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
1992 if (g_cpucache_up == FULL)
1993 return enable_cpucache(cachep, gfp);
1995 if (g_cpucache_up == NONE) {
1997 * Note: the first kmem_cache_create must create the cache
1998 * that's used by kmalloc(24), otherwise the creation of
1999 * further caches will BUG().
2001 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2004 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2005 * the first cache, then we need to set up all its list3s,
2006 * otherwise the creation of further caches will BUG().
2008 set_up_list3s(cachep, SIZE_AC);
2009 if (INDEX_AC == INDEX_L3)
2010 g_cpucache_up = PARTIAL_L3;
2012 g_cpucache_up = PARTIAL_AC;
2014 cachep->array[smp_processor_id()] =
2015 kmalloc(sizeof(struct arraycache_init), gfp);
2017 if (g_cpucache_up == PARTIAL_AC) {
2018 set_up_list3s(cachep, SIZE_L3);
2019 g_cpucache_up = PARTIAL_L3;
2022 for_each_online_node(node) {
2023 cachep->nodelists[node] =
2024 kmalloc_node(sizeof(struct kmem_list3),
2026 BUG_ON(!cachep->nodelists[node]);
2027 kmem_list3_init(cachep->nodelists[node]);
2031 cachep->nodelists[numa_node_id()]->next_reap =
2032 jiffies + REAPTIMEOUT_LIST3 +
2033 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
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;
2045 * kmem_cache_create - Create a cache.
2046 * @name: A string which is used in /proc/slabinfo to identify this cache.
2047 * @size: The size of objects to be created in this cache.
2048 * @align: The required alignment for the objects.
2049 * @flags: SLAB flags
2050 * @ctor: A constructor for the objects.
2052 * Returns a ptr to the cache on success, NULL on failure.
2053 * Cannot be called within a int, but can be interrupted.
2054 * The @ctor is run when new pages are allocated by the cache.
2056 * @name must be valid until the cache is destroyed. This implies that
2057 * the module calling this has to destroy the cache before getting unloaded.
2058 * Note that kmem_cache_name() is not guaranteed to return the same pointer,
2059 * therefore applications must manage it themselves.
2063 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2064 * to catch references to uninitialised memory.
2066 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2067 * for buffer overruns.
2069 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2070 * cacheline. This can be beneficial if you're counting cycles as closely
2074 kmem_cache_create (const char *name, size_t size, size_t align,
2075 unsigned long flags, void (*ctor)(void *))
2077 size_t left_over, slab_size, ralign;
2078 struct kmem_cache *cachep = NULL, *pc;
2082 * Sanity checks... these are all serious usage bugs.
2084 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2085 size > KMALLOC_MAX_SIZE) {
2086 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
2092 * We use cache_chain_mutex to ensure a consistent view of
2093 * cpu_online_mask as well. Please see cpuup_callback
2095 if (slab_is_available()) {
2097 mutex_lock(&cache_chain_mutex);
2100 list_for_each_entry(pc, &cache_chain, next) {
2105 * This happens when the module gets unloaded and doesn't
2106 * destroy its slab cache and no-one else reuses the vmalloc
2107 * area of the module. Print a warning.
2109 res = probe_kernel_address(pc->name, tmp);
2112 "SLAB: cache with size %d has lost its name\n",
2117 if (!strcmp(pc->name, name)) {
2119 "kmem_cache_create: duplicate cache %s\n", name);
2126 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2129 * Enable redzoning and last user accounting, except for caches with
2130 * large objects, if the increased size would increase the object size
2131 * above the next power of two: caches with object sizes just above a
2132 * power of two have a significant amount of internal fragmentation.
2134 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2135 2 * sizeof(unsigned long long)))
2136 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2137 if (!(flags & SLAB_DESTROY_BY_RCU))
2138 flags |= SLAB_POISON;
2140 if (flags & SLAB_DESTROY_BY_RCU)
2141 BUG_ON(flags & SLAB_POISON);
2144 * Always checks flags, a caller might be expecting debug support which
2147 BUG_ON(flags & ~CREATE_MASK);
2150 * Check that size is in terms of words. This is needed to avoid
2151 * unaligned accesses for some archs when redzoning is used, and makes
2152 * sure any on-slab bufctl's are also correctly aligned.
2154 if (size & (BYTES_PER_WORD - 1)) {
2155 size += (BYTES_PER_WORD - 1);
2156 size &= ~(BYTES_PER_WORD - 1);
2159 /* calculate the final buffer alignment: */
2161 /* 1) arch recommendation: can be overridden for debug */
2162 if (flags & SLAB_HWCACHE_ALIGN) {
2164 * Default alignment: as specified by the arch code. Except if
2165 * an object is really small, then squeeze multiple objects into
2168 ralign = cache_line_size();
2169 while (size <= ralign / 2)
2172 ralign = BYTES_PER_WORD;
2176 * Redzoning and user store require word alignment or possibly larger.
2177 * Note this will be overridden by architecture or caller mandated
2178 * alignment if either is greater than BYTES_PER_WORD.
2180 if (flags & SLAB_STORE_USER)
2181 ralign = BYTES_PER_WORD;
2183 if (flags & SLAB_RED_ZONE) {
2184 ralign = REDZONE_ALIGN;
2185 /* If redzoning, ensure that the second redzone is suitably
2186 * aligned, by adjusting the object size accordingly. */
2187 size += REDZONE_ALIGN - 1;
2188 size &= ~(REDZONE_ALIGN - 1);
2191 /* 2) arch mandated alignment */
2192 if (ralign < ARCH_SLAB_MINALIGN) {
2193 ralign = ARCH_SLAB_MINALIGN;
2195 /* 3) caller mandated alignment */
2196 if (ralign < align) {
2199 /* disable debug if necessary */
2200 if (ralign > __alignof__(unsigned long long))
2201 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2207 if (slab_is_available())
2212 /* Get cache's description obj. */
2213 cachep = kmem_cache_zalloc(&cache_cache, gfp);
2218 cachep->obj_size = size;
2221 * Both debugging options require word-alignment which is calculated
2224 if (flags & SLAB_RED_ZONE) {
2225 /* add space for red zone words */
2226 cachep->obj_offset += sizeof(unsigned long long);
2227 size += 2 * sizeof(unsigned long long);
2229 if (flags & SLAB_STORE_USER) {
2230 /* user store requires one word storage behind the end of
2231 * the real object. But if the second red zone needs to be
2232 * aligned to 64 bits, we must allow that much space.
2234 if (flags & SLAB_RED_ZONE)
2235 size += REDZONE_ALIGN;
2237 size += BYTES_PER_WORD;
2239 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2240 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2241 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2242 cachep->obj_offset += PAGE_SIZE - size;
2249 * Determine if the slab management is 'on' or 'off' slab.
2250 * (bootstrapping cannot cope with offslab caches so don't do
2251 * it too early on. Always use on-slab management when
2252 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2254 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2255 !(flags & SLAB_NOLEAKTRACE))
2257 * Size is large, assume best to place the slab management obj
2258 * off-slab (should allow better packing of objs).
2260 flags |= CFLGS_OFF_SLAB;
2262 size = ALIGN(size, align);
2264 left_over = calculate_slab_order(cachep, size, align, flags);
2268 "kmem_cache_create: couldn't create cache %s.\n", name);
2269 kmem_cache_free(&cache_cache, cachep);
2273 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2274 + sizeof(struct slab), align);
2277 * If the slab has been placed off-slab, and we have enough space then
2278 * move it on-slab. This is at the expense of any extra colouring.
2280 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2281 flags &= ~CFLGS_OFF_SLAB;
2282 left_over -= slab_size;
2285 if (flags & CFLGS_OFF_SLAB) {
2286 /* really off slab. No need for manual alignment */
2288 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2290 #ifdef CONFIG_PAGE_POISONING
2291 /* If we're going to use the generic kernel_map_pages()
2292 * poisoning, then it's going to smash the contents of
2293 * the redzone and userword anyhow, so switch them off.
2295 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2296 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2300 cachep->colour_off = cache_line_size();
2301 /* Offset must be a multiple of the alignment. */
2302 if (cachep->colour_off < align)
2303 cachep->colour_off = align;
2304 cachep->colour = left_over / cachep->colour_off;
2305 cachep->slab_size = slab_size;
2306 cachep->flags = flags;
2307 cachep->gfpflags = 0;
2308 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2309 cachep->gfpflags |= GFP_DMA;
2310 cachep->buffer_size = size;
2311 cachep->reciprocal_buffer_size = reciprocal_value(size);
2313 if (flags & CFLGS_OFF_SLAB) {
2314 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2316 * This is a possibility for one of the malloc_sizes caches.
2317 * But since we go off slab only for object size greater than
2318 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2319 * this should not happen at all.
2320 * But leave a BUG_ON for some lucky dude.
2322 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2324 cachep->ctor = ctor;
2325 cachep->name = name;
2327 if (setup_cpu_cache(cachep, gfp)) {
2328 __kmem_cache_destroy(cachep);
2333 /* cache setup completed, link it into the list */
2334 list_add(&cachep->next, &cache_chain);
2336 if (!cachep && (flags & SLAB_PANIC))
2337 panic("kmem_cache_create(): failed to create slab `%s'\n",
2339 if (slab_is_available()) {
2340 mutex_unlock(&cache_chain_mutex);
2345 EXPORT_SYMBOL(kmem_cache_create);
2348 static void check_irq_off(void)
2350 BUG_ON(!irqs_disabled());
2353 static void check_irq_on(void)
2355 BUG_ON(irqs_disabled());
2358 static void check_spinlock_acquired(struct kmem_cache *cachep)
2362 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2366 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2370 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2375 #define check_irq_off() do { } while(0)
2376 #define check_irq_on() do { } while(0)
2377 #define check_spinlock_acquired(x) do { } while(0)
2378 #define check_spinlock_acquired_node(x, y) do { } while(0)
2381 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2382 struct array_cache *ac,
2383 int force, int node);
2385 static void do_drain(void *arg)
2387 struct kmem_cache *cachep = arg;
2388 struct array_cache *ac;
2389 int node = numa_node_id();
2392 ac = cpu_cache_get(cachep);
2393 spin_lock(&cachep->nodelists[node]->list_lock);
2394 free_block(cachep, ac->entry, ac->avail, node);
2395 spin_unlock(&cachep->nodelists[node]->list_lock);
2399 static void drain_cpu_caches(struct kmem_cache *cachep)
2401 struct kmem_list3 *l3;
2404 on_each_cpu(do_drain, cachep, 1);
2406 for_each_online_node(node) {
2407 l3 = cachep->nodelists[node];
2408 if (l3 && l3->alien)
2409 drain_alien_cache(cachep, l3->alien);
2412 for_each_online_node(node) {
2413 l3 = cachep->nodelists[node];
2415 drain_array(cachep, l3, l3->shared, 1, node);
2420 * Remove slabs from the list of free slabs.
2421 * Specify the number of slabs to drain in tofree.
2423 * Returns the actual number of slabs released.
2425 static int drain_freelist(struct kmem_cache *cache,
2426 struct kmem_list3 *l3, int tofree)
2428 struct list_head *p;
2433 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2435 spin_lock_irq(&l3->list_lock);
2436 p = l3->slabs_free.prev;
2437 if (p == &l3->slabs_free) {
2438 spin_unlock_irq(&l3->list_lock);
2442 slabp = list_entry(p, struct slab, list);
2444 BUG_ON(slabp->inuse);
2446 list_del(&slabp->list);
2448 * Safe to drop the lock. The slab is no longer linked
2451 l3->free_objects -= cache->num;
2452 spin_unlock_irq(&l3->list_lock);
2453 slab_destroy(cache, slabp);
2460 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2461 static int __cache_shrink(struct kmem_cache *cachep)
2464 struct kmem_list3 *l3;
2466 drain_cpu_caches(cachep);
2469 for_each_online_node(i) {
2470 l3 = cachep->nodelists[i];
2474 drain_freelist(cachep, l3, l3->free_objects);
2476 ret += !list_empty(&l3->slabs_full) ||
2477 !list_empty(&l3->slabs_partial);
2479 return (ret ? 1 : 0);
2483 * kmem_cache_shrink - Shrink a cache.
2484 * @cachep: The cache to shrink.
2486 * Releases as many slabs as possible for a cache.
2487 * To help debugging, a zero exit status indicates all slabs were released.
2489 int kmem_cache_shrink(struct kmem_cache *cachep)
2492 BUG_ON(!cachep || in_interrupt());
2495 mutex_lock(&cache_chain_mutex);
2496 ret = __cache_shrink(cachep);
2497 mutex_unlock(&cache_chain_mutex);
2501 EXPORT_SYMBOL(kmem_cache_shrink);
2504 * kmem_cache_destroy - delete a cache
2505 * @cachep: the cache to destroy
2507 * Remove a &struct kmem_cache object from the slab cache.
2509 * It is expected this function will be called by a module when it is
2510 * unloaded. This will remove the cache completely, and avoid a duplicate
2511 * cache being allocated each time a module is loaded and unloaded, if the
2512 * module doesn't have persistent in-kernel storage across loads and unloads.
2514 * The cache must be empty before calling this function.
2516 * The caller must guarantee that noone will allocate memory from the cache
2517 * during the kmem_cache_destroy().
2519 void kmem_cache_destroy(struct kmem_cache *cachep)
2521 BUG_ON(!cachep || in_interrupt());
2523 /* Find the cache in the chain of caches. */
2525 mutex_lock(&cache_chain_mutex);
2527 * the chain is never empty, cache_cache is never destroyed
2529 list_del(&cachep->next);
2530 if (__cache_shrink(cachep)) {
2531 slab_error(cachep, "Can't free all objects");
2532 list_add(&cachep->next, &cache_chain);
2533 mutex_unlock(&cache_chain_mutex);
2538 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2541 __kmem_cache_destroy(cachep);
2542 mutex_unlock(&cache_chain_mutex);
2545 EXPORT_SYMBOL(kmem_cache_destroy);
2548 * Get the memory for a slab management obj.
2549 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2550 * always come from malloc_sizes caches. The slab descriptor cannot
2551 * come from the same cache which is getting created because,
2552 * when we are searching for an appropriate cache for these
2553 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2554 * If we are creating a malloc_sizes cache here it would not be visible to
2555 * kmem_find_general_cachep till the initialization is complete.
2556 * Hence we cannot have slabp_cache same as the original cache.
2558 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2559 int colour_off, gfp_t local_flags,
2564 if (OFF_SLAB(cachep)) {
2565 /* Slab management obj is off-slab. */
2566 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2567 local_flags, nodeid);
2569 * If the first object in the slab is leaked (it's allocated
2570 * but no one has a reference to it), we want to make sure
2571 * kmemleak does not treat the ->s_mem pointer as a reference
2572 * to the object. Otherwise we will not report the leak.
2574 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2579 slabp = objp + colour_off;
2580 colour_off += cachep->slab_size;
2583 slabp->colouroff = colour_off;
2584 slabp->s_mem = objp + colour_off;
2585 slabp->nodeid = nodeid;
2590 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2592 return (kmem_bufctl_t *) (slabp + 1);
2595 static void cache_init_objs(struct kmem_cache *cachep,
2600 for (i = 0; i < cachep->num; i++) {
2601 void *objp = index_to_obj(cachep, slabp, i);
2603 /* need to poison the objs? */
2604 if (cachep->flags & SLAB_POISON)
2605 poison_obj(cachep, objp, POISON_FREE);
2606 if (cachep->flags & SLAB_STORE_USER)
2607 *dbg_userword(cachep, objp) = NULL;
2609 if (cachep->flags & SLAB_RED_ZONE) {
2610 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2611 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2614 * Constructors are not allowed to allocate memory from the same
2615 * cache which they are a constructor for. Otherwise, deadlock.
2616 * They must also be threaded.
2618 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2619 cachep->ctor(objp + obj_offset(cachep));
2621 if (cachep->flags & SLAB_RED_ZONE) {
2622 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2623 slab_error(cachep, "constructor overwrote the"
2624 " end of an object");
2625 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2626 slab_error(cachep, "constructor overwrote the"
2627 " start of an object");
2629 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2630 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2631 kernel_map_pages(virt_to_page(objp),
2632 cachep->buffer_size / PAGE_SIZE, 0);
2637 slab_bufctl(slabp)[i] = i + 1;
2639 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2642 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2644 if (CONFIG_ZONE_DMA_FLAG) {
2645 if (flags & GFP_DMA)
2646 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2648 BUG_ON(cachep->gfpflags & GFP_DMA);
2652 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2655 void *objp = index_to_obj(cachep, slabp, slabp->free);
2659 next = slab_bufctl(slabp)[slabp->free];
2661 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2662 WARN_ON(slabp->nodeid != nodeid);
2669 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2670 void *objp, int nodeid)
2672 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2675 /* Verify that the slab belongs to the intended node */
2676 WARN_ON(slabp->nodeid != nodeid);
2678 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2679 printk(KERN_ERR "slab: double free detected in cache "
2680 "'%s', objp %p\n", cachep->name, objp);
2684 slab_bufctl(slabp)[objnr] = slabp->free;
2685 slabp->free = objnr;
2690 * Map pages beginning at addr to the given cache and slab. This is required
2691 * for the slab allocator to be able to lookup the cache and slab of a
2692 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2694 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2700 page = virt_to_page(addr);
2703 if (likely(!PageCompound(page)))
2704 nr_pages <<= cache->gfporder;
2707 page_set_cache(page, cache);
2708 page_set_slab(page, slab);
2710 } while (--nr_pages);
2714 * Grow (by 1) the number of slabs within a cache. This is called by
2715 * kmem_cache_alloc() when there are no active objs left in a cache.
2717 static int cache_grow(struct kmem_cache *cachep,
2718 gfp_t flags, int nodeid, void *objp)
2723 struct kmem_list3 *l3;
2726 * Be lazy and only check for valid flags here, keeping it out of the
2727 * critical path in kmem_cache_alloc().
2729 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2730 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2732 /* Take the l3 list lock to change the colour_next on this node */
2734 l3 = cachep->nodelists[nodeid];
2735 spin_lock(&l3->list_lock);
2737 /* Get colour for the slab, and cal the next value. */
2738 offset = l3->colour_next;
2740 if (l3->colour_next >= cachep->colour)
2741 l3->colour_next = 0;
2742 spin_unlock(&l3->list_lock);
2744 offset *= cachep->colour_off;
2746 if (local_flags & __GFP_WAIT)
2750 * The test for missing atomic flag is performed here, rather than
2751 * the more obvious place, simply to reduce the critical path length
2752 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2753 * will eventually be caught here (where it matters).
2755 kmem_flagcheck(cachep, flags);
2758 * Get mem for the objs. Attempt to allocate a physical page from
2762 objp = kmem_getpages(cachep, local_flags, nodeid);
2766 /* Get slab management. */
2767 slabp = alloc_slabmgmt(cachep, objp, offset,
2768 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2772 slab_map_pages(cachep, slabp, objp);
2774 cache_init_objs(cachep, slabp);
2776 if (local_flags & __GFP_WAIT)
2777 local_irq_disable();
2779 spin_lock(&l3->list_lock);
2781 /* Make slab active. */
2782 list_add_tail(&slabp->list, &(l3->slabs_free));
2783 STATS_INC_GROWN(cachep);
2784 l3->free_objects += cachep->num;
2785 spin_unlock(&l3->list_lock);
2788 kmem_freepages(cachep, objp);
2790 if (local_flags & __GFP_WAIT)
2791 local_irq_disable();
2798 * Perform extra freeing checks:
2799 * - detect bad pointers.
2800 * - POISON/RED_ZONE checking
2802 static void kfree_debugcheck(const void *objp)
2804 if (!virt_addr_valid(objp)) {
2805 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2806 (unsigned long)objp);
2811 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2813 unsigned long long redzone1, redzone2;
2815 redzone1 = *dbg_redzone1(cache, obj);
2816 redzone2 = *dbg_redzone2(cache, obj);
2821 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2824 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2825 slab_error(cache, "double free detected");
2827 slab_error(cache, "memory outside object was overwritten");
2829 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2830 obj, redzone1, redzone2);
2833 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2840 BUG_ON(virt_to_cache(objp) != cachep);
2842 objp -= obj_offset(cachep);
2843 kfree_debugcheck(objp);
2844 page = virt_to_head_page(objp);
2846 slabp = page_get_slab(page);
2848 if (cachep->flags & SLAB_RED_ZONE) {
2849 verify_redzone_free(cachep, objp);
2850 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2851 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2853 if (cachep->flags & SLAB_STORE_USER)
2854 *dbg_userword(cachep, objp) = caller;
2856 objnr = obj_to_index(cachep, slabp, objp);
2858 BUG_ON(objnr >= cachep->num);
2859 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2861 #ifdef CONFIG_DEBUG_SLAB_LEAK
2862 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2864 if (cachep->flags & SLAB_POISON) {
2865 #ifdef CONFIG_DEBUG_PAGEALLOC
2866 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2867 store_stackinfo(cachep, objp, (unsigned long)caller);
2868 kernel_map_pages(virt_to_page(objp),
2869 cachep->buffer_size / PAGE_SIZE, 0);
2871 poison_obj(cachep, objp, POISON_FREE);
2874 poison_obj(cachep, objp, POISON_FREE);
2880 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2885 /* Check slab's freelist to see if this obj is there. */
2886 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2888 if (entries > cachep->num || i >= cachep->num)
2891 if (entries != cachep->num - slabp->inuse) {
2893 printk(KERN_ERR "slab: Internal list corruption detected in "
2894 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2895 cachep->name, cachep->num, slabp, slabp->inuse);
2897 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2900 printk("\n%03x:", i);
2901 printk(" %02x", ((unsigned char *)slabp)[i]);
2908 #define kfree_debugcheck(x) do { } while(0)
2909 #define cache_free_debugcheck(x,objp,z) (objp)
2910 #define check_slabp(x,y) do { } while(0)
2913 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2916 struct kmem_list3 *l3;
2917 struct array_cache *ac;
2922 node = numa_node_id();
2923 ac = cpu_cache_get(cachep);
2924 batchcount = ac->batchcount;
2925 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2927 * If there was little recent activity on this cache, then
2928 * perform only a partial refill. Otherwise we could generate
2931 batchcount = BATCHREFILL_LIMIT;
2933 l3 = cachep->nodelists[node];
2935 BUG_ON(ac->avail > 0 || !l3);
2936 spin_lock(&l3->list_lock);
2938 /* See if we can refill from the shared array */
2939 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
2940 l3->shared->touched = 1;
2944 while (batchcount > 0) {
2945 struct list_head *entry;
2947 /* Get slab alloc is to come from. */
2948 entry = l3->slabs_partial.next;
2949 if (entry == &l3->slabs_partial) {
2950 l3->free_touched = 1;
2951 entry = l3->slabs_free.next;
2952 if (entry == &l3->slabs_free)
2956 slabp = list_entry(entry, struct slab, list);
2957 check_slabp(cachep, slabp);
2958 check_spinlock_acquired(cachep);
2961 * The slab was either on partial or free list so
2962 * there must be at least one object available for
2965 BUG_ON(slabp->inuse >= cachep->num);
2967 while (slabp->inuse < cachep->num && batchcount--) {
2968 STATS_INC_ALLOCED(cachep);
2969 STATS_INC_ACTIVE(cachep);
2970 STATS_SET_HIGH(cachep);
2972 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2975 check_slabp(cachep, slabp);
2977 /* move slabp to correct slabp list: */
2978 list_del(&slabp->list);
2979 if (slabp->free == BUFCTL_END)
2980 list_add(&slabp->list, &l3->slabs_full);
2982 list_add(&slabp->list, &l3->slabs_partial);
2986 l3->free_objects -= ac->avail;
2988 spin_unlock(&l3->list_lock);
2990 if (unlikely(!ac->avail)) {
2992 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
2994 /* cache_grow can reenable interrupts, then ac could change. */
2995 ac = cpu_cache_get(cachep);
2996 if (!x && ac->avail == 0) /* no objects in sight? abort */
2999 if (!ac->avail) /* objects refilled by interrupt? */
3003 return ac->entry[--ac->avail];
3006 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3009 might_sleep_if(flags & __GFP_WAIT);
3011 kmem_flagcheck(cachep, flags);
3016 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3017 gfp_t flags, void *objp, void *caller)
3021 if (cachep->flags & SLAB_POISON) {
3022 #ifdef CONFIG_DEBUG_PAGEALLOC
3023 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3024 kernel_map_pages(virt_to_page(objp),
3025 cachep->buffer_size / PAGE_SIZE, 1);
3027 check_poison_obj(cachep, objp);
3029 check_poison_obj(cachep, objp);
3031 poison_obj(cachep, objp, POISON_INUSE);
3033 if (cachep->flags & SLAB_STORE_USER)
3034 *dbg_userword(cachep, objp) = caller;
3036 if (cachep->flags & SLAB_RED_ZONE) {
3037 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3038 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3039 slab_error(cachep, "double free, or memory outside"
3040 " object was overwritten");
3042 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3043 objp, *dbg_redzone1(cachep, objp),
3044 *dbg_redzone2(cachep, objp));
3046 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3047 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3049 #ifdef CONFIG_DEBUG_SLAB_LEAK
3054 slabp = page_get_slab(virt_to_head_page(objp));
3055 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3056 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3059 objp += obj_offset(cachep);
3060 if (cachep->ctor && cachep->flags & SLAB_POISON)
3062 #if ARCH_SLAB_MINALIGN
3063 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3064 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3065 objp, ARCH_SLAB_MINALIGN);
3071 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3074 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3076 if (cachep == &cache_cache)
3079 return should_failslab(obj_size(cachep), flags, cachep->flags);
3082 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3085 struct array_cache *ac;
3089 ac = cpu_cache_get(cachep);
3090 if (likely(ac->avail)) {
3091 STATS_INC_ALLOCHIT(cachep);
3093 objp = ac->entry[--ac->avail];
3095 STATS_INC_ALLOCMISS(cachep);
3096 objp = cache_alloc_refill(cachep, flags);
3098 * the 'ac' may be updated by cache_alloc_refill(),
3099 * and kmemleak_erase() requires its correct value.
3101 ac = cpu_cache_get(cachep);
3104 * To avoid a false negative, if an object that is in one of the
3105 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3106 * treat the array pointers as a reference to the object.
3109 kmemleak_erase(&ac->entry[ac->avail]);
3115 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3117 * If we are in_interrupt, then process context, including cpusets and
3118 * mempolicy, may not apply and should not be used for allocation policy.
3120 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3122 int nid_alloc, nid_here;
3124 if (in_interrupt() || (flags & __GFP_THISNODE))
3126 nid_alloc = nid_here = numa_node_id();
3127 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3128 nid_alloc = cpuset_mem_spread_node();
3129 else if (current->mempolicy)
3130 nid_alloc = slab_node(current->mempolicy);
3131 if (nid_alloc != nid_here)
3132 return ____cache_alloc_node(cachep, flags, nid_alloc);
3137 * Fallback function if there was no memory available and no objects on a
3138 * certain node and fall back is permitted. First we scan all the
3139 * available nodelists for available objects. If that fails then we
3140 * perform an allocation without specifying a node. This allows the page
3141 * allocator to do its reclaim / fallback magic. We then insert the
3142 * slab into the proper nodelist and then allocate from it.
3144 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3146 struct zonelist *zonelist;
3150 enum zone_type high_zoneidx = gfp_zone(flags);
3154 if (flags & __GFP_THISNODE)
3157 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
3158 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3162 * Look through allowed nodes for objects available
3163 * from existing per node queues.
3165 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3166 nid = zone_to_nid(zone);
3168 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3169 cache->nodelists[nid] &&
3170 cache->nodelists[nid]->free_objects) {
3171 obj = ____cache_alloc_node(cache,
3172 flags | GFP_THISNODE, nid);
3180 * This allocation will be performed within the constraints
3181 * of the current cpuset / memory policy requirements.
3182 * We may trigger various forms of reclaim on the allowed
3183 * set and go into memory reserves if necessary.
3185 if (local_flags & __GFP_WAIT)
3187 kmem_flagcheck(cache, flags);
3188 obj = kmem_getpages(cache, local_flags, numa_node_id());
3189 if (local_flags & __GFP_WAIT)
3190 local_irq_disable();
3193 * Insert into the appropriate per node queues
3195 nid = page_to_nid(virt_to_page(obj));
3196 if (cache_grow(cache, flags, nid, obj)) {
3197 obj = ____cache_alloc_node(cache,
3198 flags | GFP_THISNODE, nid);
3201 * Another processor may allocate the
3202 * objects in the slab since we are
3203 * not holding any locks.
3207 /* cache_grow already freed obj */
3216 * A interface to enable slab creation on nodeid
3218 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3221 struct list_head *entry;
3223 struct kmem_list3 *l3;
3227 l3 = cachep->nodelists[nodeid];
3232 spin_lock(&l3->list_lock);
3233 entry = l3->slabs_partial.next;
3234 if (entry == &l3->slabs_partial) {
3235 l3->free_touched = 1;
3236 entry = l3->slabs_free.next;
3237 if (entry == &l3->slabs_free)
3241 slabp = list_entry(entry, struct slab, list);
3242 check_spinlock_acquired_node(cachep, nodeid);
3243 check_slabp(cachep, slabp);
3245 STATS_INC_NODEALLOCS(cachep);
3246 STATS_INC_ACTIVE(cachep);
3247 STATS_SET_HIGH(cachep);
3249 BUG_ON(slabp->inuse == cachep->num);
3251 obj = slab_get_obj(cachep, slabp, nodeid);
3252 check_slabp(cachep, slabp);
3254 /* move slabp to correct slabp list: */
3255 list_del(&slabp->list);
3257 if (slabp->free == BUFCTL_END)
3258 list_add(&slabp->list, &l3->slabs_full);
3260 list_add(&slabp->list, &l3->slabs_partial);
3262 spin_unlock(&l3->list_lock);
3266 spin_unlock(&l3->list_lock);
3267 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3271 return fallback_alloc(cachep, flags);
3278 * kmem_cache_alloc_node - Allocate an object on the specified node
3279 * @cachep: The cache to allocate from.
3280 * @flags: See kmalloc().
3281 * @nodeid: node number of the target node.
3282 * @caller: return address of caller, used for debug information
3284 * Identical to kmem_cache_alloc but it will allocate memory on the given
3285 * node, which can improve the performance for cpu bound structures.
3287 * Fallback to other node is possible if __GFP_THISNODE is not set.
3289 static __always_inline void *
3290 __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3293 unsigned long save_flags;
3296 flags &= gfp_allowed_mask;
3298 lockdep_trace_alloc(flags);
3300 if (slab_should_failslab(cachep, flags))
3303 cache_alloc_debugcheck_before(cachep, flags);
3304 local_irq_save(save_flags);
3307 nodeid = numa_node_id();
3309 if (unlikely(!cachep->nodelists[nodeid])) {
3310 /* Node not bootstrapped yet */
3311 ptr = fallback_alloc(cachep, flags);
3315 if (nodeid == numa_node_id()) {
3317 * Use the locally cached objects if possible.
3318 * However ____cache_alloc does not allow fallback
3319 * to other nodes. It may fail while we still have
3320 * objects on other nodes available.
3322 ptr = ____cache_alloc(cachep, flags);
3326 /* ___cache_alloc_node can fall back to other nodes */
3327 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3329 local_irq_restore(save_flags);
3330 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3331 kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags,
3335 kmemcheck_slab_alloc(cachep, flags, ptr, obj_size(cachep));
3337 if (unlikely((flags & __GFP_ZERO) && ptr))
3338 memset(ptr, 0, obj_size(cachep));
3343 static __always_inline void *
3344 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3348 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3349 objp = alternate_node_alloc(cache, flags);
3353 objp = ____cache_alloc(cache, flags);
3356 * We may just have run out of memory on the local node.
3357 * ____cache_alloc_node() knows how to locate memory on other nodes
3360 objp = ____cache_alloc_node(cache, flags, numa_node_id());
3367 static __always_inline void *
3368 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3370 return ____cache_alloc(cachep, flags);
3373 #endif /* CONFIG_NUMA */
3375 static __always_inline void *
3376 __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3378 unsigned long save_flags;
3381 flags &= gfp_allowed_mask;
3383 lockdep_trace_alloc(flags);
3385 if (slab_should_failslab(cachep, flags))
3388 cache_alloc_debugcheck_before(cachep, flags);
3389 local_irq_save(save_flags);
3390 objp = __do_cache_alloc(cachep, flags);
3391 local_irq_restore(save_flags);
3392 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3393 kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags,
3398 kmemcheck_slab_alloc(cachep, flags, objp, obj_size(cachep));
3400 if (unlikely((flags & __GFP_ZERO) && objp))
3401 memset(objp, 0, obj_size(cachep));
3407 * Caller needs to acquire correct kmem_list's list_lock
3409 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3413 struct kmem_list3 *l3;
3415 for (i = 0; i < nr_objects; i++) {
3416 void *objp = objpp[i];
3419 slabp = virt_to_slab(objp);
3420 l3 = cachep->nodelists[node];
3421 list_del(&slabp->list);
3422 check_spinlock_acquired_node(cachep, node);
3423 check_slabp(cachep, slabp);
3424 slab_put_obj(cachep, slabp, objp, node);
3425 STATS_DEC_ACTIVE(cachep);
3427 check_slabp(cachep, slabp);
3429 /* fixup slab chains */
3430 if (slabp->inuse == 0) {
3431 if (l3->free_objects > l3->free_limit) {
3432 l3->free_objects -= cachep->num;
3433 /* No need to drop any previously held
3434 * lock here, even if we have a off-slab slab
3435 * descriptor it is guaranteed to come from
3436 * a different cache, refer to comments before
3439 slab_destroy(cachep, slabp);
3441 list_add(&slabp->list, &l3->slabs_free);
3444 /* Unconditionally move a slab to the end of the
3445 * partial list on free - maximum time for the
3446 * other objects to be freed, too.
3448 list_add_tail(&slabp->list, &l3->slabs_partial);
3453 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3456 struct kmem_list3 *l3;
3457 int node = numa_node_id();
3459 batchcount = ac->batchcount;
3461 BUG_ON(!batchcount || batchcount > ac->avail);
3464 l3 = cachep->nodelists[node];
3465 spin_lock(&l3->list_lock);
3467 struct array_cache *shared_array = l3->shared;
3468 int max = shared_array->limit - shared_array->avail;
3470 if (batchcount > max)
3472 memcpy(&(shared_array->entry[shared_array->avail]),
3473 ac->entry, sizeof(void *) * batchcount);
3474 shared_array->avail += batchcount;
3479 free_block(cachep, ac->entry, batchcount, node);
3484 struct list_head *p;
3486 p = l3->slabs_free.next;
3487 while (p != &(l3->slabs_free)) {
3490 slabp = list_entry(p, struct slab, list);
3491 BUG_ON(slabp->inuse);
3496 STATS_SET_FREEABLE(cachep, i);
3499 spin_unlock(&l3->list_lock);
3500 ac->avail -= batchcount;
3501 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3505 * Release an obj back to its cache. If the obj has a constructed state, it must
3506 * be in this state _before_ it is released. Called with disabled ints.
3508 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3510 struct array_cache *ac = cpu_cache_get(cachep);
3513 kmemleak_free_recursive(objp, cachep->flags);
3514 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3516 kmemcheck_slab_free(cachep, objp, obj_size(cachep));
3519 * Skip calling cache_free_alien() when the platform is not numa.
3520 * This will avoid cache misses that happen while accessing slabp (which
3521 * is per page memory reference) to get nodeid. Instead use a global
3522 * variable to skip the call, which is mostly likely to be present in
3525 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3528 if (likely(ac->avail < ac->limit)) {
3529 STATS_INC_FREEHIT(cachep);
3530 ac->entry[ac->avail++] = objp;
3533 STATS_INC_FREEMISS(cachep);
3534 cache_flusharray(cachep, ac);
3535 ac->entry[ac->avail++] = objp;
3540 * kmem_cache_alloc - Allocate an object
3541 * @cachep: The cache to allocate from.
3542 * @flags: See kmalloc().
3544 * Allocate an object from this cache. The flags are only relevant
3545 * if the cache has no available objects.
3547 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3549 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3551 trace_kmem_cache_alloc(_RET_IP_, ret,
3552 obj_size(cachep), cachep->buffer_size, flags);
3556 EXPORT_SYMBOL(kmem_cache_alloc);
3558 #ifdef CONFIG_TRACING
3559 void *kmem_cache_alloc_notrace(struct kmem_cache *cachep, gfp_t flags)
3561 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3563 EXPORT_SYMBOL(kmem_cache_alloc_notrace);
3567 * kmem_ptr_validate - check if an untrusted pointer might be a slab entry.
3568 * @cachep: the cache we're checking against
3569 * @ptr: pointer to validate
3571 * This verifies that the untrusted pointer looks sane;
3572 * it is _not_ a guarantee that the pointer is actually
3573 * part of the slab cache in question, but it at least
3574 * validates that the pointer can be dereferenced and
3575 * looks half-way sane.
3577 * Currently only used for dentry validation.
3579 int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
3581 unsigned long size = cachep->buffer_size;
3584 if (unlikely(!kern_ptr_validate(ptr, size)))
3586 page = virt_to_page(ptr);
3587 if (unlikely(!PageSlab(page)))
3589 if (unlikely(page_get_cache(page) != cachep))
3597 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3599 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3600 __builtin_return_address(0));
3602 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3603 obj_size(cachep), cachep->buffer_size,
3608 EXPORT_SYMBOL(kmem_cache_alloc_node);
3610 #ifdef CONFIG_TRACING
3611 void *kmem_cache_alloc_node_notrace(struct kmem_cache *cachep,
3615 return __cache_alloc_node(cachep, flags, nodeid,
3616 __builtin_return_address(0));
3618 EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
3621 static __always_inline void *
3622 __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3624 struct kmem_cache *cachep;
3627 cachep = kmem_find_general_cachep(size, flags);
3628 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3630 ret = kmem_cache_alloc_node_notrace(cachep, flags, node);
3632 trace_kmalloc_node((unsigned long) caller, ret,
3633 size, cachep->buffer_size, flags, node);
3638 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3639 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3641 return __do_kmalloc_node(size, flags, node,
3642 __builtin_return_address(0));
3644 EXPORT_SYMBOL(__kmalloc_node);
3646 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3647 int node, unsigned long caller)
3649 return __do_kmalloc_node(size, flags, node, (void *)caller);
3651 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3653 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3655 return __do_kmalloc_node(size, flags, node, NULL);
3657 EXPORT_SYMBOL(__kmalloc_node);
3658 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3659 #endif /* CONFIG_NUMA */
3662 * __do_kmalloc - allocate memory
3663 * @size: how many bytes of memory are required.
3664 * @flags: the type of memory to allocate (see kmalloc).
3665 * @caller: function caller for debug tracking of the caller
3667 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3670 struct kmem_cache *cachep;
3673 /* If you want to save a few bytes .text space: replace
3675 * Then kmalloc uses the uninlined functions instead of the inline
3678 cachep = __find_general_cachep(size, flags);
3679 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3681 ret = __cache_alloc(cachep, flags, caller);
3683 trace_kmalloc((unsigned long) caller, ret,
3684 size, cachep->buffer_size, flags);
3690 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3691 void *__kmalloc(size_t size, gfp_t flags)
3693 return __do_kmalloc(size, flags, __builtin_return_address(0));
3695 EXPORT_SYMBOL(__kmalloc);
3697 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3699 return __do_kmalloc(size, flags, (void *)caller);
3701 EXPORT_SYMBOL(__kmalloc_track_caller);
3704 void *__kmalloc(size_t size, gfp_t flags)
3706 return __do_kmalloc(size, flags, NULL);
3708 EXPORT_SYMBOL(__kmalloc);
3712 * kmem_cache_free - Deallocate an object
3713 * @cachep: The cache the allocation was from.
3714 * @objp: The previously allocated object.
3716 * Free an object which was previously allocated from this
3719 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3721 unsigned long flags;
3723 local_irq_save(flags);
3724 debug_check_no_locks_freed(objp, obj_size(cachep));
3725 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3726 debug_check_no_obj_freed(objp, obj_size(cachep));
3727 __cache_free(cachep, objp);
3728 local_irq_restore(flags);
3730 trace_kmem_cache_free(_RET_IP_, objp);
3732 EXPORT_SYMBOL(kmem_cache_free);
3735 * kfree - free previously allocated memory
3736 * @objp: pointer returned by kmalloc.
3738 * If @objp is NULL, no operation is performed.
3740 * Don't free memory not originally allocated by kmalloc()
3741 * or you will run into trouble.
3743 void kfree(const void *objp)
3745 struct kmem_cache *c;
3746 unsigned long flags;
3748 trace_kfree(_RET_IP_, objp);
3750 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3752 local_irq_save(flags);
3753 kfree_debugcheck(objp);
3754 c = virt_to_cache(objp);
3755 debug_check_no_locks_freed(objp, obj_size(c));
3756 debug_check_no_obj_freed(objp, obj_size(c));
3757 __cache_free(c, (void *)objp);
3758 local_irq_restore(flags);
3760 EXPORT_SYMBOL(kfree);
3762 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3764 return obj_size(cachep);
3766 EXPORT_SYMBOL(kmem_cache_size);
3768 const char *kmem_cache_name(struct kmem_cache *cachep)
3770 return cachep->name;
3772 EXPORT_SYMBOL_GPL(kmem_cache_name);
3775 * This initializes kmem_list3 or resizes various caches for all nodes.
3777 static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3780 struct kmem_list3 *l3;
3781 struct array_cache *new_shared;
3782 struct array_cache **new_alien = NULL;
3784 for_each_online_node(node) {
3786 if (use_alien_caches) {
3787 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3793 if (cachep->shared) {
3794 new_shared = alloc_arraycache(node,
3795 cachep->shared*cachep->batchcount,
3798 free_alien_cache(new_alien);
3803 l3 = cachep->nodelists[node];
3805 struct array_cache *shared = l3->shared;
3807 spin_lock_irq(&l3->list_lock);
3810 free_block(cachep, shared->entry,
3811 shared->avail, node);
3813 l3->shared = new_shared;
3815 l3->alien = new_alien;
3818 l3->free_limit = (1 + nr_cpus_node(node)) *
3819 cachep->batchcount + cachep->num;
3820 spin_unlock_irq(&l3->list_lock);
3822 free_alien_cache(new_alien);
3825 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
3827 free_alien_cache(new_alien);
3832 kmem_list3_init(l3);
3833 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3834 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3835 l3->shared = new_shared;
3836 l3->alien = new_alien;
3837 l3->free_limit = (1 + nr_cpus_node(node)) *
3838 cachep->batchcount + cachep->num;
3839 cachep->nodelists[node] = l3;
3844 if (!cachep->next.next) {
3845 /* Cache is not active yet. Roll back what we did */
3848 if (cachep->nodelists[node]) {
3849 l3 = cachep->nodelists[node];
3852 free_alien_cache(l3->alien);
3854 cachep->nodelists[node] = NULL;
3862 struct ccupdate_struct {
3863 struct kmem_cache *cachep;
3864 struct array_cache *new[NR_CPUS];
3867 static void do_ccupdate_local(void *info)
3869 struct ccupdate_struct *new = info;
3870 struct array_cache *old;
3873 old = cpu_cache_get(new->cachep);
3875 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3876 new->new[smp_processor_id()] = old;
3879 /* Always called with the cache_chain_mutex held */
3880 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3881 int batchcount, int shared, gfp_t gfp)
3883 struct ccupdate_struct *new;
3886 new = kzalloc(sizeof(*new), gfp);
3890 for_each_online_cpu(i) {
3891 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
3894 for (i--; i >= 0; i--)
3900 new->cachep = cachep;
3902 on_each_cpu(do_ccupdate_local, (void *)new, 1);
3905 cachep->batchcount = batchcount;
3906 cachep->limit = limit;
3907 cachep->shared = shared;
3909 for_each_online_cpu(i) {
3910 struct array_cache *ccold = new->new[i];
3913 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3914 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3915 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3919 return alloc_kmemlist(cachep, gfp);
3922 /* Called with cache_chain_mutex held always */
3923 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3929 * The head array serves three purposes:
3930 * - create a LIFO ordering, i.e. return objects that are cache-warm
3931 * - reduce the number of spinlock operations.
3932 * - reduce the number of linked list operations on the slab and
3933 * bufctl chains: array operations are cheaper.
3934 * The numbers are guessed, we should auto-tune as described by
3937 if (cachep->buffer_size > 131072)
3939 else if (cachep->buffer_size > PAGE_SIZE)
3941 else if (cachep->buffer_size > 1024)
3943 else if (cachep->buffer_size > 256)
3949 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3950 * allocation behaviour: Most allocs on one cpu, most free operations
3951 * on another cpu. For these cases, an efficient object passing between
3952 * cpus is necessary. This is provided by a shared array. The array
3953 * replaces Bonwick's magazine layer.
3954 * On uniprocessor, it's functionally equivalent (but less efficient)
3955 * to a larger limit. Thus disabled by default.
3958 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
3963 * With debugging enabled, large batchcount lead to excessively long
3964 * periods with disabled local interrupts. Limit the batchcount
3969 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
3971 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3972 cachep->name, -err);
3977 * Drain an array if it contains any elements taking the l3 lock only if
3978 * necessary. Note that the l3 listlock also protects the array_cache
3979 * if drain_array() is used on the shared array.
3981 void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
3982 struct array_cache *ac, int force, int node)
3986 if (!ac || !ac->avail)
3988 if (ac->touched && !force) {
3991 spin_lock_irq(&l3->list_lock);
3993 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3994 if (tofree > ac->avail)
3995 tofree = (ac->avail + 1) / 2;
3996 free_block(cachep, ac->entry, tofree, node);
3997 ac->avail -= tofree;
3998 memmove(ac->entry, &(ac->entry[tofree]),
3999 sizeof(void *) * ac->avail);
4001 spin_unlock_irq(&l3->list_lock);
4006 * cache_reap - Reclaim memory from caches.
4007 * @w: work descriptor
4009 * Called from workqueue/eventd every few seconds.
4011 * - clear the per-cpu caches for this CPU.
4012 * - return freeable pages to the main free memory pool.
4014 * If we cannot acquire the cache chain mutex then just give up - we'll try
4015 * again on the next iteration.
4017 static void cache_reap(struct work_struct *w)
4019 struct kmem_cache *searchp;
4020 struct kmem_list3 *l3;
4021 int node = numa_node_id();
4022 struct delayed_work *work = to_delayed_work(w);
4024 if (!mutex_trylock(&cache_chain_mutex))
4025 /* Give up. Setup the next iteration. */
4028 list_for_each_entry(searchp, &cache_chain, next) {
4032 * We only take the l3 lock if absolutely necessary and we
4033 * have established with reasonable certainty that
4034 * we can do some work if the lock was obtained.
4036 l3 = searchp->nodelists[node];
4038 reap_alien(searchp, l3);
4040 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4043 * These are racy checks but it does not matter
4044 * if we skip one check or scan twice.
4046 if (time_after(l3->next_reap, jiffies))
4049 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4051 drain_array(searchp, l3, l3->shared, 0, node);
4053 if (l3->free_touched)
4054 l3->free_touched = 0;
4058 freed = drain_freelist(searchp, l3, (l3->free_limit +
4059 5 * searchp->num - 1) / (5 * searchp->num));
4060 STATS_ADD_REAPED(searchp, freed);
4066 mutex_unlock(&cache_chain_mutex);
4069 /* Set up the next iteration */
4070 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4073 #ifdef CONFIG_SLABINFO
4075 static void print_slabinfo_header(struct seq_file *m)
4078 * Output format version, so at least we can change it
4079 * without _too_ many complaints.
4082 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4084 seq_puts(m, "slabinfo - version: 2.1\n");
4086 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4087 "<objperslab> <pagesperslab>");
4088 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4089 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4091 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4092 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4093 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4098 static void *s_start(struct seq_file *m, loff_t *pos)
4102 mutex_lock(&cache_chain_mutex);
4104 print_slabinfo_header(m);
4106 return seq_list_start(&cache_chain, *pos);
4109 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4111 return seq_list_next(p, &cache_chain, pos);
4114 static void s_stop(struct seq_file *m, void *p)
4116 mutex_unlock(&cache_chain_mutex);
4119 static int s_show(struct seq_file *m, void *p)
4121 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4123 unsigned long active_objs;
4124 unsigned long num_objs;
4125 unsigned long active_slabs = 0;
4126 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4130 struct kmem_list3 *l3;
4134 for_each_online_node(node) {
4135 l3 = cachep->nodelists[node];
4140 spin_lock_irq(&l3->list_lock);
4142 list_for_each_entry(slabp, &l3->slabs_full, list) {
4143 if (slabp->inuse != cachep->num && !error)
4144 error = "slabs_full accounting error";
4145 active_objs += cachep->num;
4148 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4149 if (slabp->inuse == cachep->num && !error)
4150 error = "slabs_partial inuse accounting error";
4151 if (!slabp->inuse && !error)
4152 error = "slabs_partial/inuse accounting error";
4153 active_objs += slabp->inuse;
4156 list_for_each_entry(slabp, &l3->slabs_free, list) {
4157 if (slabp->inuse && !error)
4158 error = "slabs_free/inuse accounting error";
4161 free_objects += l3->free_objects;
4163 shared_avail += l3->shared->avail;
4165 spin_unlock_irq(&l3->list_lock);
4167 num_slabs += active_slabs;
4168 num_objs = num_slabs * cachep->num;
4169 if (num_objs - active_objs != free_objects && !error)
4170 error = "free_objects accounting error";
4172 name = cachep->name;
4174 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4176 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4177 name, active_objs, num_objs, cachep->buffer_size,
4178 cachep->num, (1 << cachep->gfporder));
4179 seq_printf(m, " : tunables %4u %4u %4u",
4180 cachep->limit, cachep->batchcount, cachep->shared);
4181 seq_printf(m, " : slabdata %6lu %6lu %6lu",
4182 active_slabs, num_slabs, shared_avail);
4185 unsigned long high = cachep->high_mark;
4186 unsigned long allocs = cachep->num_allocations;
4187 unsigned long grown = cachep->grown;
4188 unsigned long reaped = cachep->reaped;
4189 unsigned long errors = cachep->errors;
4190 unsigned long max_freeable = cachep->max_freeable;
4191 unsigned long node_allocs = cachep->node_allocs;
4192 unsigned long node_frees = cachep->node_frees;
4193 unsigned long overflows = cachep->node_overflow;
4195 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
4196 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
4197 reaped, errors, max_freeable, node_allocs,
4198 node_frees, overflows);
4202 unsigned long allochit = atomic_read(&cachep->allochit);
4203 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4204 unsigned long freehit = atomic_read(&cachep->freehit);
4205 unsigned long freemiss = atomic_read(&cachep->freemiss);
4207 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4208 allochit, allocmiss, freehit, freemiss);
4216 * slabinfo_op - iterator that generates /proc/slabinfo
4225 * num-pages-per-slab
4226 * + further values on SMP and with statistics enabled
4229 static const struct seq_operations slabinfo_op = {
4236 #define MAX_SLABINFO_WRITE 128
4238 * slabinfo_write - Tuning for the slab allocator
4240 * @buffer: user buffer
4241 * @count: data length
4244 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4245 size_t count, loff_t *ppos)
4247 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4248 int limit, batchcount, shared, res;
4249 struct kmem_cache *cachep;
4251 if (count > MAX_SLABINFO_WRITE)
4253 if (copy_from_user(&kbuf, buffer, count))
4255 kbuf[MAX_SLABINFO_WRITE] = '\0';
4257 tmp = strchr(kbuf, ' ');
4262 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4265 /* Find the cache in the chain of caches. */
4266 mutex_lock(&cache_chain_mutex);
4268 list_for_each_entry(cachep, &cache_chain, next) {
4269 if (!strcmp(cachep->name, kbuf)) {
4270 if (limit < 1 || batchcount < 1 ||
4271 batchcount > limit || shared < 0) {
4274 res = do_tune_cpucache(cachep, limit,
4281 mutex_unlock(&cache_chain_mutex);
4287 static int slabinfo_open(struct inode *inode, struct file *file)
4289 return seq_open(file, &slabinfo_op);
4292 static const struct file_operations proc_slabinfo_operations = {
4293 .open = slabinfo_open,
4295 .write = slabinfo_write,
4296 .llseek = seq_lseek,
4297 .release = seq_release,
4300 #ifdef CONFIG_DEBUG_SLAB_LEAK
4302 static void *leaks_start(struct seq_file *m, loff_t *pos)
4304 mutex_lock(&cache_chain_mutex);
4305 return seq_list_start(&cache_chain, *pos);
4308 static inline int add_caller(unsigned long *n, unsigned long v)
4318 unsigned long *q = p + 2 * i;
4332 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4338 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4344 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4345 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4347 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4352 static void show_symbol(struct seq_file *m, unsigned long address)
4354 #ifdef CONFIG_KALLSYMS
4355 unsigned long offset, size;
4356 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4358 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4359 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4361 seq_printf(m, " [%s]", modname);
4365 seq_printf(m, "%p", (void *)address);
4368 static int leaks_show(struct seq_file *m, void *p)
4370 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4372 struct kmem_list3 *l3;
4374 unsigned long *n = m->private;
4378 if (!(cachep->flags & SLAB_STORE_USER))
4380 if (!(cachep->flags & SLAB_RED_ZONE))
4383 /* OK, we can do it */
4387 for_each_online_node(node) {
4388 l3 = cachep->nodelists[node];
4393 spin_lock_irq(&l3->list_lock);
4395 list_for_each_entry(slabp, &l3->slabs_full, list)
4396 handle_slab(n, cachep, slabp);
4397 list_for_each_entry(slabp, &l3->slabs_partial, list)
4398 handle_slab(n, cachep, slabp);
4399 spin_unlock_irq(&l3->list_lock);
4401 name = cachep->name;
4403 /* Increase the buffer size */
4404 mutex_unlock(&cache_chain_mutex);
4405 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4407 /* Too bad, we are really out */
4409 mutex_lock(&cache_chain_mutex);
4412 *(unsigned long *)m->private = n[0] * 2;
4414 mutex_lock(&cache_chain_mutex);
4415 /* Now make sure this entry will be retried */
4419 for (i = 0; i < n[1]; i++) {
4420 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4421 show_symbol(m, n[2*i+2]);
4428 static const struct seq_operations slabstats_op = {
4429 .start = leaks_start,
4435 static int slabstats_open(struct inode *inode, struct file *file)
4437 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4440 ret = seq_open(file, &slabstats_op);
4442 struct seq_file *m = file->private_data;
4443 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4452 static const struct file_operations proc_slabstats_operations = {
4453 .open = slabstats_open,
4455 .llseek = seq_lseek,
4456 .release = seq_release_private,
4460 static int __init slab_proc_init(void)
4462 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
4463 #ifdef CONFIG_DEBUG_SLAB_LEAK
4464 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4468 module_init(slab_proc_init);
4472 * ksize - get the actual amount of memory allocated for a given object
4473 * @objp: Pointer to the object
4475 * kmalloc may internally round up allocations and return more memory
4476 * than requested. ksize() can be used to determine the actual amount of
4477 * memory allocated. The caller may use this additional memory, even though
4478 * a smaller amount of memory was initially specified with the kmalloc call.
4479 * The caller must guarantee that objp points to a valid object previously
4480 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4481 * must not be freed during the duration of the call.
4483 size_t ksize(const void *objp)
4486 if (unlikely(objp == ZERO_SIZE_PTR))
4489 return obj_size(virt_to_cache(objp));
4491 EXPORT_SYMBOL(ksize);