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 intializations 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/config.h>
90 #include <linux/slab.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/seq_file.h>
98 #include <linux/notifier.h>
99 #include <linux/kallsyms.h>
100 #include <linux/cpu.h>
101 #include <linux/sysctl.h>
102 #include <linux/module.h>
103 #include <linux/rcupdate.h>
104 #include <linux/string.h>
105 #include <linux/nodemask.h>
106 #include <linux/mempolicy.h>
107 #include <linux/mutex.h>
109 #include <asm/uaccess.h>
110 #include <asm/cacheflush.h>
111 #include <asm/tlbflush.h>
112 #include <asm/page.h>
115 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
116 * SLAB_RED_ZONE & SLAB_POISON.
117 * 0 for faster, smaller code (especially in the critical paths).
119 * STATS - 1 to collect stats for /proc/slabinfo.
120 * 0 for faster, smaller code (especially in the critical paths).
122 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
125 #ifdef CONFIG_DEBUG_SLAB
128 #define FORCED_DEBUG 1
132 #define FORCED_DEBUG 0
135 /* Shouldn't this be in a header file somewhere? */
136 #define BYTES_PER_WORD sizeof(void *)
138 #ifndef cache_line_size
139 #define cache_line_size() L1_CACHE_BYTES
142 #ifndef ARCH_KMALLOC_MINALIGN
144 * Enforce a minimum alignment for the kmalloc caches.
145 * Usually, the kmalloc caches are cache_line_size() aligned, except when
146 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
147 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
148 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
149 * Note that this flag disables some debug features.
151 #define ARCH_KMALLOC_MINALIGN 0
154 #ifndef ARCH_SLAB_MINALIGN
156 * Enforce a minimum alignment for all caches.
157 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
158 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
159 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
160 * some debug features.
162 #define ARCH_SLAB_MINALIGN 0
165 #ifndef ARCH_KMALLOC_FLAGS
166 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
169 /* Legal flag mask for kmem_cache_create(). */
171 # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
172 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
174 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
175 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
178 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
179 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
180 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
187 * Bufctl's are used for linking objs within a slab
190 * This implementation relies on "struct page" for locating the cache &
191 * slab an object belongs to.
192 * This allows the bufctl structure to be small (one int), but limits
193 * the number of objects a slab (not a cache) can contain when off-slab
194 * bufctls are used. The limit is the size of the largest general cache
195 * that does not use off-slab slabs.
196 * For 32bit archs with 4 kB pages, is this 56.
197 * This is not serious, as it is only for large objects, when it is unwise
198 * to have too many per slab.
199 * Note: This limit can be raised by introducing a general cache whose size
200 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
203 typedef unsigned int kmem_bufctl_t;
204 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
205 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
206 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-2)
208 /* Max number of objs-per-slab for caches which use off-slab slabs.
209 * Needed to avoid a possible looping condition in cache_grow().
211 static unsigned long offslab_limit;
216 * Manages the objs in a slab. Placed either at the beginning of mem allocated
217 * for a slab, or allocated from an general cache.
218 * Slabs are chained into three list: fully used, partial, fully free slabs.
221 struct list_head list;
222 unsigned long colouroff;
223 void *s_mem; /* including colour offset */
224 unsigned int inuse; /* num of objs active in slab */
226 unsigned short nodeid;
232 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
233 * arrange for kmem_freepages to be called via RCU. This is useful if
234 * we need to approach a kernel structure obliquely, from its address
235 * obtained without the usual locking. We can lock the structure to
236 * stabilize it and check it's still at the given address, only if we
237 * can be sure that the memory has not been meanwhile reused for some
238 * other kind of object (which our subsystem's lock might corrupt).
240 * rcu_read_lock before reading the address, then rcu_read_unlock after
241 * taking the spinlock within the structure expected at that address.
243 * We assume struct slab_rcu can overlay struct slab when destroying.
246 struct rcu_head head;
247 struct kmem_cache *cachep;
255 * - LIFO ordering, to hand out cache-warm objects from _alloc
256 * - reduce the number of linked list operations
257 * - reduce spinlock operations
259 * The limit is stored in the per-cpu structure to reduce the data cache
266 unsigned int batchcount;
267 unsigned int touched;
270 * Must have this definition in here for the proper
271 * alignment of array_cache. Also simplifies accessing
273 * [0] is for gcc 2.95. It should really be [].
278 * bootstrap: The caches do not work without cpuarrays anymore, but the
279 * cpuarrays are allocated from the generic caches...
281 #define BOOT_CPUCACHE_ENTRIES 1
282 struct arraycache_init {
283 struct array_cache cache;
284 void *entries[BOOT_CPUCACHE_ENTRIES];
288 * The slab lists for all objects.
291 struct list_head slabs_partial; /* partial list first, better asm code */
292 struct list_head slabs_full;
293 struct list_head slabs_free;
294 unsigned long free_objects;
295 unsigned long next_reap;
297 unsigned int free_limit;
298 unsigned int colour_next; /* Per-node cache coloring */
299 spinlock_t list_lock;
300 struct array_cache *shared; /* shared per node */
301 struct array_cache **alien; /* on other nodes */
305 * Need this for bootstrapping a per node allocator.
307 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
308 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
309 #define CACHE_CACHE 0
311 #define SIZE_L3 (1 + MAX_NUMNODES)
314 * This function must be completely optimized away if a constant is passed to
315 * it. Mostly the same as what is in linux/slab.h except it returns an index.
317 static __always_inline int index_of(const size_t size)
319 extern void __bad_size(void);
321 if (__builtin_constant_p(size)) {
329 #include "linux/kmalloc_sizes.h"
337 #define INDEX_AC index_of(sizeof(struct arraycache_init))
338 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
340 static void kmem_list3_init(struct kmem_list3 *parent)
342 INIT_LIST_HEAD(&parent->slabs_full);
343 INIT_LIST_HEAD(&parent->slabs_partial);
344 INIT_LIST_HEAD(&parent->slabs_free);
345 parent->shared = NULL;
346 parent->alien = NULL;
347 parent->colour_next = 0;
348 spin_lock_init(&parent->list_lock);
349 parent->free_objects = 0;
350 parent->free_touched = 0;
353 #define MAKE_LIST(cachep, listp, slab, nodeid) \
355 INIT_LIST_HEAD(listp); \
356 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
359 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
361 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
362 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
363 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
373 /* 1) per-cpu data, touched during every alloc/free */
374 struct array_cache *array[NR_CPUS];
375 /* 2) Cache tunables. Protected by cache_chain_mutex */
376 unsigned int batchcount;
380 unsigned int buffer_size;
381 /* 3) touched by every alloc & free from the backend */
382 struct kmem_list3 *nodelists[MAX_NUMNODES];
384 unsigned int flags; /* constant flags */
385 unsigned int num; /* # of objs per slab */
387 /* 4) cache_grow/shrink */
388 /* order of pgs per slab (2^n) */
389 unsigned int gfporder;
391 /* force GFP flags, e.g. GFP_DMA */
394 size_t colour; /* cache colouring range */
395 unsigned int colour_off; /* colour offset */
396 struct kmem_cache *slabp_cache;
397 unsigned int slab_size;
398 unsigned int dflags; /* dynamic flags */
400 /* constructor func */
401 void (*ctor) (void *, struct kmem_cache *, unsigned long);
403 /* de-constructor func */
404 void (*dtor) (void *, struct kmem_cache *, unsigned long);
406 /* 5) cache creation/removal */
408 struct list_head next;
412 unsigned long num_active;
413 unsigned long num_allocations;
414 unsigned long high_mark;
416 unsigned long reaped;
417 unsigned long errors;
418 unsigned long max_freeable;
419 unsigned long node_allocs;
420 unsigned long node_frees;
428 * If debugging is enabled, then the allocator can add additional
429 * fields and/or padding to every object. buffer_size contains the total
430 * object size including these internal fields, the following two
431 * variables contain the offset to the user object and its size.
438 #define CFLGS_OFF_SLAB (0x80000000UL)
439 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
441 #define BATCHREFILL_LIMIT 16
443 * Optimization question: fewer reaps means less probability for unnessary
444 * cpucache drain/refill cycles.
446 * OTOH the cpuarrays can contain lots of objects,
447 * which could lock up otherwise freeable slabs.
449 #define REAPTIMEOUT_CPUC (2*HZ)
450 #define REAPTIMEOUT_LIST3 (4*HZ)
453 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
454 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
455 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
456 #define STATS_INC_GROWN(x) ((x)->grown++)
457 #define STATS_INC_REAPED(x) ((x)->reaped++)
458 #define STATS_SET_HIGH(x) \
460 if ((x)->num_active > (x)->high_mark) \
461 (x)->high_mark = (x)->num_active; \
463 #define STATS_INC_ERR(x) ((x)->errors++)
464 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
465 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
466 #define STATS_SET_FREEABLE(x, i) \
468 if ((x)->max_freeable < i) \
469 (x)->max_freeable = i; \
471 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
472 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
473 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
474 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
476 #define STATS_INC_ACTIVE(x) do { } while (0)
477 #define STATS_DEC_ACTIVE(x) do { } while (0)
478 #define STATS_INC_ALLOCED(x) do { } while (0)
479 #define STATS_INC_GROWN(x) do { } while (0)
480 #define STATS_INC_REAPED(x) do { } while (0)
481 #define STATS_SET_HIGH(x) do { } while (0)
482 #define STATS_INC_ERR(x) do { } while (0)
483 #define STATS_INC_NODEALLOCS(x) do { } while (0)
484 #define STATS_INC_NODEFREES(x) do { } while (0)
485 #define STATS_SET_FREEABLE(x, i) do { } while (0)
486 #define STATS_INC_ALLOCHIT(x) do { } while (0)
487 #define STATS_INC_ALLOCMISS(x) do { } while (0)
488 #define STATS_INC_FREEHIT(x) do { } while (0)
489 #define STATS_INC_FREEMISS(x) do { } while (0)
494 * Magic nums for obj red zoning.
495 * Placed in the first word before and the first word after an obj.
497 #define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
498 #define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
500 /* ...and for poisoning */
501 #define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
502 #define POISON_FREE 0x6b /* for use-after-free poisoning */
503 #define POISON_END 0xa5 /* end-byte of poisoning */
506 * memory layout of objects:
508 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
509 * the end of an object is aligned with the end of the real
510 * allocation. Catches writes behind the end of the allocation.
511 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
513 * cachep->obj_offset: The real object.
514 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
515 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
516 * [BYTES_PER_WORD long]
518 static int obj_offset(struct kmem_cache *cachep)
520 return cachep->obj_offset;
523 static int obj_size(struct kmem_cache *cachep)
525 return cachep->obj_size;
528 static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
530 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
531 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD);
534 static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
536 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
537 if (cachep->flags & SLAB_STORE_USER)
538 return (unsigned long *)(objp + cachep->buffer_size -
540 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD);
543 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
545 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
546 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
551 #define obj_offset(x) 0
552 #define obj_size(cachep) (cachep->buffer_size)
553 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
554 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
555 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
560 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
563 #if defined(CONFIG_LARGE_ALLOCS)
564 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
565 #define MAX_GFP_ORDER 13 /* up to 32Mb */
566 #elif defined(CONFIG_MMU)
567 #define MAX_OBJ_ORDER 5 /* 32 pages */
568 #define MAX_GFP_ORDER 5 /* 32 pages */
570 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
571 #define MAX_GFP_ORDER 8 /* up to 1Mb */
575 * Do not go above this order unless 0 objects fit into the slab.
577 #define BREAK_GFP_ORDER_HI 1
578 #define BREAK_GFP_ORDER_LO 0
579 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
582 * Functions for storing/retrieving the cachep and or slab from the page
583 * allocator. These are used to find the slab an obj belongs to. With kfree(),
584 * these are used to find the cache which an obj belongs to.
586 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
588 page->lru.next = (struct list_head *)cache;
591 static inline struct kmem_cache *page_get_cache(struct page *page)
593 if (unlikely(PageCompound(page)))
594 page = (struct page *)page_private(page);
595 return (struct kmem_cache *)page->lru.next;
598 static inline void page_set_slab(struct page *page, struct slab *slab)
600 page->lru.prev = (struct list_head *)slab;
603 static inline struct slab *page_get_slab(struct page *page)
605 if (unlikely(PageCompound(page)))
606 page = (struct page *)page_private(page);
607 return (struct slab *)page->lru.prev;
610 static inline struct kmem_cache *virt_to_cache(const void *obj)
612 struct page *page = virt_to_page(obj);
613 return page_get_cache(page);
616 static inline struct slab *virt_to_slab(const void *obj)
618 struct page *page = virt_to_page(obj);
619 return page_get_slab(page);
622 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
625 return slab->s_mem + cache->buffer_size * idx;
628 static inline unsigned int obj_to_index(struct kmem_cache *cache,
629 struct slab *slab, void *obj)
631 return (unsigned)(obj - slab->s_mem) / cache->buffer_size;
635 * These are the default caches for kmalloc. Custom caches can have other sizes.
637 struct cache_sizes malloc_sizes[] = {
638 #define CACHE(x) { .cs_size = (x) },
639 #include <linux/kmalloc_sizes.h>
643 EXPORT_SYMBOL(malloc_sizes);
645 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
651 static struct cache_names __initdata cache_names[] = {
652 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
653 #include <linux/kmalloc_sizes.h>
658 static struct arraycache_init initarray_cache __initdata =
659 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
660 static struct arraycache_init initarray_generic =
661 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
663 /* internal cache of cache description objs */
664 static struct kmem_cache cache_cache = {
666 .limit = BOOT_CPUCACHE_ENTRIES,
668 .buffer_size = sizeof(struct kmem_cache),
669 .name = "kmem_cache",
671 .obj_size = sizeof(struct kmem_cache),
675 /* Guard access to the cache-chain. */
676 static DEFINE_MUTEX(cache_chain_mutex);
677 static struct list_head cache_chain;
680 * vm_enough_memory() looks at this to determine how many slab-allocated pages
681 * are possibly freeable under pressure
683 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
685 atomic_t slab_reclaim_pages;
688 * chicken and egg problem: delay the per-cpu array allocation
689 * until the general caches are up.
698 static DEFINE_PER_CPU(struct work_struct, reap_work);
700 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
702 static void enable_cpucache(struct kmem_cache *cachep);
703 static void cache_reap(void *unused);
704 static int __node_shrink(struct kmem_cache *cachep, int node);
706 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
708 return cachep->array[smp_processor_id()];
711 static inline struct kmem_cache *__find_general_cachep(size_t size,
714 struct cache_sizes *csizep = malloc_sizes;
717 /* This happens if someone tries to call
718 * kmem_cache_create(), or __kmalloc(), before
719 * the generic caches are initialized.
721 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
723 while (size > csizep->cs_size)
727 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
728 * has cs_{dma,}cachep==NULL. Thus no special case
729 * for large kmalloc calls required.
731 if (unlikely(gfpflags & GFP_DMA))
732 return csizep->cs_dmacachep;
733 return csizep->cs_cachep;
736 struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
738 return __find_general_cachep(size, gfpflags);
740 EXPORT_SYMBOL(kmem_find_general_cachep);
742 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
744 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
748 * Calculate the number of objects and left-over bytes for a given buffer size.
750 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
751 size_t align, int flags, size_t *left_over,
756 size_t slab_size = PAGE_SIZE << gfporder;
759 * The slab management structure can be either off the slab or
760 * on it. For the latter case, the memory allocated for a
764 * - One kmem_bufctl_t for each object
765 * - Padding to respect alignment of @align
766 * - @buffer_size bytes for each object
768 * If the slab management structure is off the slab, then the
769 * alignment will already be calculated into the size. Because
770 * the slabs are all pages aligned, the objects will be at the
771 * correct alignment when allocated.
773 if (flags & CFLGS_OFF_SLAB) {
775 nr_objs = slab_size / buffer_size;
777 if (nr_objs > SLAB_LIMIT)
778 nr_objs = SLAB_LIMIT;
781 * Ignore padding for the initial guess. The padding
782 * is at most @align-1 bytes, and @buffer_size is at
783 * least @align. In the worst case, this result will
784 * be one greater than the number of objects that fit
785 * into the memory allocation when taking the padding
788 nr_objs = (slab_size - sizeof(struct slab)) /
789 (buffer_size + sizeof(kmem_bufctl_t));
792 * This calculated number will be either the right
793 * amount, or one greater than what we want.
795 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
799 if (nr_objs > SLAB_LIMIT)
800 nr_objs = SLAB_LIMIT;
802 mgmt_size = slab_mgmt_size(nr_objs, align);
805 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
808 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
810 static void __slab_error(const char *function, struct kmem_cache *cachep,
813 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
814 function, cachep->name, msg);
820 * Special reaping functions for NUMA systems called from cache_reap().
821 * These take care of doing round robin flushing of alien caches (containing
822 * objects freed on different nodes from which they were allocated) and the
823 * flushing of remote pcps by calling drain_node_pages.
825 static DEFINE_PER_CPU(unsigned long, reap_node);
827 static void init_reap_node(int cpu)
831 node = next_node(cpu_to_node(cpu), node_online_map);
832 if (node == MAX_NUMNODES)
835 __get_cpu_var(reap_node) = node;
838 static void next_reap_node(void)
840 int node = __get_cpu_var(reap_node);
843 * Also drain per cpu pages on remote zones
845 if (node != numa_node_id())
846 drain_node_pages(node);
848 node = next_node(node, node_online_map);
849 if (unlikely(node >= MAX_NUMNODES))
850 node = first_node(node_online_map);
851 __get_cpu_var(reap_node) = node;
855 #define init_reap_node(cpu) do { } while (0)
856 #define next_reap_node(void) do { } while (0)
860 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
861 * via the workqueue/eventd.
862 * Add the CPU number into the expiration time to minimize the possibility of
863 * the CPUs getting into lockstep and contending for the global cache chain
866 static void __devinit start_cpu_timer(int cpu)
868 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
871 * When this gets called from do_initcalls via cpucache_init(),
872 * init_workqueues() has already run, so keventd will be setup
875 if (keventd_up() && reap_work->func == NULL) {
877 INIT_WORK(reap_work, cache_reap, NULL);
878 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
882 static struct array_cache *alloc_arraycache(int node, int entries,
885 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
886 struct array_cache *nc = NULL;
888 nc = kmalloc_node(memsize, GFP_KERNEL, node);
892 nc->batchcount = batchcount;
894 spin_lock_init(&nc->lock);
900 static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int);
902 static struct array_cache **alloc_alien_cache(int node, int limit)
904 struct array_cache **ac_ptr;
905 int memsize = sizeof(void *) * MAX_NUMNODES;
910 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
913 if (i == node || !node_online(i)) {
917 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
919 for (i--; i <= 0; i--)
929 static void free_alien_cache(struct array_cache **ac_ptr)
940 static void __drain_alien_cache(struct kmem_cache *cachep,
941 struct array_cache *ac, int node)
943 struct kmem_list3 *rl3 = cachep->nodelists[node];
946 spin_lock(&rl3->list_lock);
947 free_block(cachep, ac->entry, ac->avail, node);
949 spin_unlock(&rl3->list_lock);
954 * Called from cache_reap() to regularly drain alien caches round robin.
956 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
958 int node = __get_cpu_var(reap_node);
961 struct array_cache *ac = l3->alien[node];
962 if (ac && ac->avail) {
963 spin_lock_irq(&ac->lock);
964 __drain_alien_cache(cachep, ac, node);
965 spin_unlock_irq(&ac->lock);
970 static void drain_alien_cache(struct kmem_cache *cachep,
971 struct array_cache **alien)
974 struct array_cache *ac;
977 for_each_online_node(i) {
980 spin_lock_irqsave(&ac->lock, flags);
981 __drain_alien_cache(cachep, ac, i);
982 spin_unlock_irqrestore(&ac->lock, flags);
988 #define drain_alien_cache(cachep, alien) do { } while (0)
989 #define reap_alien(cachep, l3) do { } while (0)
991 static inline struct array_cache **alloc_alien_cache(int node, int limit)
993 return (struct array_cache **) 0x01020304ul;
996 static inline void free_alien_cache(struct array_cache **ac_ptr)
1002 static int __devinit cpuup_callback(struct notifier_block *nfb,
1003 unsigned long action, void *hcpu)
1005 long cpu = (long)hcpu;
1006 struct kmem_cache *cachep;
1007 struct kmem_list3 *l3 = NULL;
1008 int node = cpu_to_node(cpu);
1009 int memsize = sizeof(struct kmem_list3);
1012 case CPU_UP_PREPARE:
1013 mutex_lock(&cache_chain_mutex);
1015 * We need to do this right in the beginning since
1016 * alloc_arraycache's are going to use this list.
1017 * kmalloc_node allows us to add the slab to the right
1018 * kmem_list3 and not this cpu's kmem_list3
1021 list_for_each_entry(cachep, &cache_chain, next) {
1023 * Set up the size64 kmemlist for cpu before we can
1024 * begin anything. Make sure some other cpu on this
1025 * node has not already allocated this
1027 if (!cachep->nodelists[node]) {
1028 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1031 kmem_list3_init(l3);
1032 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1033 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1036 * The l3s don't come and go as CPUs come and
1037 * go. cache_chain_mutex is sufficient
1040 cachep->nodelists[node] = l3;
1043 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1044 cachep->nodelists[node]->free_limit =
1045 (1 + nr_cpus_node(node)) *
1046 cachep->batchcount + cachep->num;
1047 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1051 * Now we can go ahead with allocating the shared arrays and
1054 list_for_each_entry(cachep, &cache_chain, next) {
1055 struct array_cache *nc;
1056 struct array_cache *shared;
1057 struct array_cache **alien;
1059 nc = alloc_arraycache(node, cachep->limit,
1060 cachep->batchcount);
1063 shared = alloc_arraycache(node,
1064 cachep->shared * cachep->batchcount,
1069 alien = alloc_alien_cache(node, cachep->limit);
1072 cachep->array[cpu] = nc;
1073 l3 = cachep->nodelists[node];
1076 spin_lock_irq(&l3->list_lock);
1079 * We are serialised from CPU_DEAD or
1080 * CPU_UP_CANCELLED by the cpucontrol lock
1082 l3->shared = shared;
1091 spin_unlock_irq(&l3->list_lock);
1093 free_alien_cache(alien);
1095 mutex_unlock(&cache_chain_mutex);
1098 start_cpu_timer(cpu);
1100 #ifdef CONFIG_HOTPLUG_CPU
1103 * Even if all the cpus of a node are down, we don't free the
1104 * kmem_list3 of any cache. This to avoid a race between
1105 * cpu_down, and a kmalloc allocation from another cpu for
1106 * memory from the node of the cpu going down. The list3
1107 * structure is usually allocated from kmem_cache_create() and
1108 * gets destroyed at kmem_cache_destroy().
1111 case CPU_UP_CANCELED:
1112 mutex_lock(&cache_chain_mutex);
1113 list_for_each_entry(cachep, &cache_chain, next) {
1114 struct array_cache *nc;
1115 struct array_cache *shared;
1116 struct array_cache **alien;
1119 mask = node_to_cpumask(node);
1120 /* cpu is dead; no one can alloc from it. */
1121 nc = cachep->array[cpu];
1122 cachep->array[cpu] = NULL;
1123 l3 = cachep->nodelists[node];
1126 goto free_array_cache;
1128 spin_lock_irq(&l3->list_lock);
1130 /* Free limit for this kmem_list3 */
1131 l3->free_limit -= cachep->batchcount;
1133 free_block(cachep, nc->entry, nc->avail, node);
1135 if (!cpus_empty(mask)) {
1136 spin_unlock_irq(&l3->list_lock);
1137 goto free_array_cache;
1140 shared = l3->shared;
1142 free_block(cachep, l3->shared->entry,
1143 l3->shared->avail, node);
1150 spin_unlock_irq(&l3->list_lock);
1154 drain_alien_cache(cachep, alien);
1155 free_alien_cache(alien);
1161 * In the previous loop, all the objects were freed to
1162 * the respective cache's slabs, now we can go ahead and
1163 * shrink each nodelist to its limit.
1165 list_for_each_entry(cachep, &cache_chain, next) {
1166 l3 = cachep->nodelists[node];
1169 spin_lock_irq(&l3->list_lock);
1170 /* free slabs belonging to this node */
1171 __node_shrink(cachep, node);
1172 spin_unlock_irq(&l3->list_lock);
1174 mutex_unlock(&cache_chain_mutex);
1180 mutex_unlock(&cache_chain_mutex);
1184 static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
1187 * swap the static kmem_list3 with kmalloced memory
1189 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1192 struct kmem_list3 *ptr;
1194 BUG_ON(cachep->nodelists[nodeid] != list);
1195 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1198 local_irq_disable();
1199 memcpy(ptr, list, sizeof(struct kmem_list3));
1200 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1201 cachep->nodelists[nodeid] = ptr;
1206 * Initialisation. Called after the page allocator have been initialised and
1207 * before smp_init().
1209 void __init kmem_cache_init(void)
1212 struct cache_sizes *sizes;
1213 struct cache_names *names;
1217 for (i = 0; i < NUM_INIT_LISTS; i++) {
1218 kmem_list3_init(&initkmem_list3[i]);
1219 if (i < MAX_NUMNODES)
1220 cache_cache.nodelists[i] = NULL;
1224 * Fragmentation resistance on low memory - only use bigger
1225 * page orders on machines with more than 32MB of memory.
1227 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1228 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1230 /* Bootstrap is tricky, because several objects are allocated
1231 * from caches that do not exist yet:
1232 * 1) initialize the cache_cache cache: it contains the struct
1233 * kmem_cache structures of all caches, except cache_cache itself:
1234 * cache_cache is statically allocated.
1235 * Initially an __init data area is used for the head array and the
1236 * kmem_list3 structures, it's replaced with a kmalloc allocated
1237 * array at the end of the bootstrap.
1238 * 2) Create the first kmalloc cache.
1239 * The struct kmem_cache for the new cache is allocated normally.
1240 * An __init data area is used for the head array.
1241 * 3) Create the remaining kmalloc caches, with minimally sized
1243 * 4) Replace the __init data head arrays for cache_cache and the first
1244 * kmalloc cache with kmalloc allocated arrays.
1245 * 5) Replace the __init data for kmem_list3 for cache_cache and
1246 * the other cache's with kmalloc allocated memory.
1247 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1250 /* 1) create the cache_cache */
1251 INIT_LIST_HEAD(&cache_chain);
1252 list_add(&cache_cache.next, &cache_chain);
1253 cache_cache.colour_off = cache_line_size();
1254 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1255 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
1257 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1260 for (order = 0; order < MAX_ORDER; order++) {
1261 cache_estimate(order, cache_cache.buffer_size,
1262 cache_line_size(), 0, &left_over, &cache_cache.num);
1263 if (cache_cache.num)
1266 if (!cache_cache.num)
1268 cache_cache.gfporder = order;
1269 cache_cache.colour = left_over / cache_cache.colour_off;
1270 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1271 sizeof(struct slab), cache_line_size());
1273 /* 2+3) create the kmalloc caches */
1274 sizes = malloc_sizes;
1275 names = cache_names;
1278 * Initialize the caches that provide memory for the array cache and the
1279 * kmem_list3 structures first. Without this, further allocations will
1283 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1284 sizes[INDEX_AC].cs_size,
1285 ARCH_KMALLOC_MINALIGN,
1286 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1289 if (INDEX_AC != INDEX_L3) {
1290 sizes[INDEX_L3].cs_cachep =
1291 kmem_cache_create(names[INDEX_L3].name,
1292 sizes[INDEX_L3].cs_size,
1293 ARCH_KMALLOC_MINALIGN,
1294 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1298 while (sizes->cs_size != ULONG_MAX) {
1300 * For performance, all the general caches are L1 aligned.
1301 * This should be particularly beneficial on SMP boxes, as it
1302 * eliminates "false sharing".
1303 * Note for systems short on memory removing the alignment will
1304 * allow tighter packing of the smaller caches.
1306 if (!sizes->cs_cachep) {
1307 sizes->cs_cachep = kmem_cache_create(names->name,
1309 ARCH_KMALLOC_MINALIGN,
1310 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1314 /* Inc off-slab bufctl limit until the ceiling is hit. */
1315 if (!(OFF_SLAB(sizes->cs_cachep))) {
1316 offslab_limit = sizes->cs_size - sizeof(struct slab);
1317 offslab_limit /= sizeof(kmem_bufctl_t);
1320 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
1322 ARCH_KMALLOC_MINALIGN,
1323 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1329 /* 4) Replace the bootstrap head arrays */
1333 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1335 local_irq_disable();
1336 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1337 memcpy(ptr, cpu_cache_get(&cache_cache),
1338 sizeof(struct arraycache_init));
1339 cache_cache.array[smp_processor_id()] = ptr;
1342 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1344 local_irq_disable();
1345 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1346 != &initarray_generic.cache);
1347 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1348 sizeof(struct arraycache_init));
1349 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1353 /* 5) Replace the bootstrap kmem_list3's */
1356 /* Replace the static kmem_list3 structures for the boot cpu */
1357 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
1360 for_each_online_node(node) {
1361 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1362 &initkmem_list3[SIZE_AC + node], node);
1364 if (INDEX_AC != INDEX_L3) {
1365 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1366 &initkmem_list3[SIZE_L3 + node],
1372 /* 6) resize the head arrays to their final sizes */
1374 struct kmem_cache *cachep;
1375 mutex_lock(&cache_chain_mutex);
1376 list_for_each_entry(cachep, &cache_chain, next)
1377 enable_cpucache(cachep);
1378 mutex_unlock(&cache_chain_mutex);
1382 g_cpucache_up = FULL;
1385 * Register a cpu startup notifier callback that initializes
1386 * cpu_cache_get for all new cpus
1388 register_cpu_notifier(&cpucache_notifier);
1391 * The reap timers are started later, with a module init call: That part
1392 * of the kernel is not yet operational.
1396 static int __init cpucache_init(void)
1401 * Register the timers that return unneeded pages to the page allocator
1403 for_each_online_cpu(cpu)
1404 start_cpu_timer(cpu);
1407 __initcall(cpucache_init);
1410 * Interface to system's page allocator. No need to hold the cache-lock.
1412 * If we requested dmaable memory, we will get it. Even if we
1413 * did not request dmaable memory, we might get it, but that
1414 * would be relatively rare and ignorable.
1416 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1422 flags |= cachep->gfpflags;
1423 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1426 addr = page_address(page);
1428 i = (1 << cachep->gfporder);
1429 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1430 atomic_add(i, &slab_reclaim_pages);
1431 add_page_state(nr_slab, i);
1433 __SetPageSlab(page);
1440 * Interface to system's page release.
1442 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1444 unsigned long i = (1 << cachep->gfporder);
1445 struct page *page = virt_to_page(addr);
1446 const unsigned long nr_freed = i;
1449 BUG_ON(!PageSlab(page));
1450 __ClearPageSlab(page);
1453 sub_page_state(nr_slab, nr_freed);
1454 if (current->reclaim_state)
1455 current->reclaim_state->reclaimed_slab += nr_freed;
1456 free_pages((unsigned long)addr, cachep->gfporder);
1457 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1458 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages);
1461 static void kmem_rcu_free(struct rcu_head *head)
1463 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1464 struct kmem_cache *cachep = slab_rcu->cachep;
1466 kmem_freepages(cachep, slab_rcu->addr);
1467 if (OFF_SLAB(cachep))
1468 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1473 #ifdef CONFIG_DEBUG_PAGEALLOC
1474 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1475 unsigned long caller)
1477 int size = obj_size(cachep);
1479 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1481 if (size < 5 * sizeof(unsigned long))
1484 *addr++ = 0x12345678;
1486 *addr++ = smp_processor_id();
1487 size -= 3 * sizeof(unsigned long);
1489 unsigned long *sptr = &caller;
1490 unsigned long svalue;
1492 while (!kstack_end(sptr)) {
1494 if (kernel_text_address(svalue)) {
1496 size -= sizeof(unsigned long);
1497 if (size <= sizeof(unsigned long))
1503 *addr++ = 0x87654321;
1507 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1509 int size = obj_size(cachep);
1510 addr = &((char *)addr)[obj_offset(cachep)];
1512 memset(addr, val, size);
1513 *(unsigned char *)(addr + size - 1) = POISON_END;
1516 static void dump_line(char *data, int offset, int limit)
1519 printk(KERN_ERR "%03x:", offset);
1520 for (i = 0; i < limit; i++)
1521 printk(" %02x", (unsigned char)data[offset + i]);
1528 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1533 if (cachep->flags & SLAB_RED_ZONE) {
1534 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
1535 *dbg_redzone1(cachep, objp),
1536 *dbg_redzone2(cachep, objp));
1539 if (cachep->flags & SLAB_STORE_USER) {
1540 printk(KERN_ERR "Last user: [<%p>]",
1541 *dbg_userword(cachep, objp));
1542 print_symbol("(%s)",
1543 (unsigned long)*dbg_userword(cachep, objp));
1546 realobj = (char *)objp + obj_offset(cachep);
1547 size = obj_size(cachep);
1548 for (i = 0; i < size && lines; i += 16, lines--) {
1551 if (i + limit > size)
1553 dump_line(realobj, i, limit);
1557 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1563 realobj = (char *)objp + obj_offset(cachep);
1564 size = obj_size(cachep);
1566 for (i = 0; i < size; i++) {
1567 char exp = POISON_FREE;
1570 if (realobj[i] != exp) {
1576 "Slab corruption: start=%p, len=%d\n",
1578 print_objinfo(cachep, objp, 0);
1580 /* Hexdump the affected line */
1583 if (i + limit > size)
1585 dump_line(realobj, i, limit);
1588 /* Limit to 5 lines */
1594 /* Print some data about the neighboring objects, if they
1597 struct slab *slabp = virt_to_slab(objp);
1600 objnr = obj_to_index(cachep, slabp, objp);
1602 objp = index_to_obj(cachep, slabp, objnr - 1);
1603 realobj = (char *)objp + obj_offset(cachep);
1604 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1606 print_objinfo(cachep, objp, 2);
1608 if (objnr + 1 < cachep->num) {
1609 objp = index_to_obj(cachep, slabp, objnr + 1);
1610 realobj = (char *)objp + obj_offset(cachep);
1611 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1613 print_objinfo(cachep, objp, 2);
1621 * slab_destroy_objs - destroy a slab and its objects
1622 * @cachep: cache pointer being destroyed
1623 * @slabp: slab pointer being destroyed
1625 * Call the registered destructor for each object in a slab that is being
1628 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1631 for (i = 0; i < cachep->num; i++) {
1632 void *objp = index_to_obj(cachep, slabp, i);
1634 if (cachep->flags & SLAB_POISON) {
1635 #ifdef CONFIG_DEBUG_PAGEALLOC
1636 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1638 kernel_map_pages(virt_to_page(objp),
1639 cachep->buffer_size / PAGE_SIZE, 1);
1641 check_poison_obj(cachep, objp);
1643 check_poison_obj(cachep, objp);
1646 if (cachep->flags & SLAB_RED_ZONE) {
1647 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1648 slab_error(cachep, "start of a freed object "
1650 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1651 slab_error(cachep, "end of a freed object "
1654 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1655 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
1659 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1663 for (i = 0; i < cachep->num; i++) {
1664 void *objp = index_to_obj(cachep, slabp, i);
1665 (cachep->dtor) (objp, cachep, 0);
1672 * slab_destroy - destroy and release all objects in a slab
1673 * @cachep: cache pointer being destroyed
1674 * @slabp: slab pointer being destroyed
1676 * Destroy all the objs in a slab, and release the mem back to the system.
1677 * Before calling the slab must have been unlinked from the cache. The
1678 * cache-lock is not held/needed.
1680 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1682 void *addr = slabp->s_mem - slabp->colouroff;
1684 slab_destroy_objs(cachep, slabp);
1685 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1686 struct slab_rcu *slab_rcu;
1688 slab_rcu = (struct slab_rcu *)slabp;
1689 slab_rcu->cachep = cachep;
1690 slab_rcu->addr = addr;
1691 call_rcu(&slab_rcu->head, kmem_rcu_free);
1693 kmem_freepages(cachep, addr);
1694 if (OFF_SLAB(cachep))
1695 kmem_cache_free(cachep->slabp_cache, slabp);
1700 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1701 * size of kmem_list3.
1703 static void set_up_list3s(struct kmem_cache *cachep, int index)
1707 for_each_online_node(node) {
1708 cachep->nodelists[node] = &initkmem_list3[index + node];
1709 cachep->nodelists[node]->next_reap = jiffies +
1711 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1716 * calculate_slab_order - calculate size (page order) of slabs
1717 * @cachep: pointer to the cache that is being created
1718 * @size: size of objects to be created in this cache.
1719 * @align: required alignment for the objects.
1720 * @flags: slab allocation flags
1722 * Also calculates the number of objects per slab.
1724 * This could be made much more intelligent. For now, try to avoid using
1725 * high order pages for slabs. When the gfp() functions are more friendly
1726 * towards high-order requests, this should be changed.
1728 static size_t calculate_slab_order(struct kmem_cache *cachep,
1729 size_t size, size_t align, unsigned long flags)
1731 size_t left_over = 0;
1734 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
1738 cache_estimate(gfporder, size, align, flags, &remainder, &num);
1742 /* More than offslab_limit objects will cause problems */
1743 if ((flags & CFLGS_OFF_SLAB) && num > offslab_limit)
1746 /* Found something acceptable - save it away */
1748 cachep->gfporder = gfporder;
1749 left_over = remainder;
1752 * A VFS-reclaimable slab tends to have most allocations
1753 * as GFP_NOFS and we really don't want to have to be allocating
1754 * higher-order pages when we are unable to shrink dcache.
1756 if (flags & SLAB_RECLAIM_ACCOUNT)
1760 * Large number of objects is good, but very large slabs are
1761 * currently bad for the gfp()s.
1763 if (gfporder >= slab_break_gfp_order)
1767 * Acceptable internal fragmentation?
1769 if (left_over * 8 <= (PAGE_SIZE << gfporder))
1775 static void setup_cpu_cache(struct kmem_cache *cachep)
1777 if (g_cpucache_up == FULL) {
1778 enable_cpucache(cachep);
1781 if (g_cpucache_up == NONE) {
1783 * Note: the first kmem_cache_create must create the cache
1784 * that's used by kmalloc(24), otherwise the creation of
1785 * further caches will BUG().
1787 cachep->array[smp_processor_id()] = &initarray_generic.cache;
1790 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
1791 * the first cache, then we need to set up all its list3s,
1792 * otherwise the creation of further caches will BUG().
1794 set_up_list3s(cachep, SIZE_AC);
1795 if (INDEX_AC == INDEX_L3)
1796 g_cpucache_up = PARTIAL_L3;
1798 g_cpucache_up = PARTIAL_AC;
1800 cachep->array[smp_processor_id()] =
1801 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1803 if (g_cpucache_up == PARTIAL_AC) {
1804 set_up_list3s(cachep, SIZE_L3);
1805 g_cpucache_up = PARTIAL_L3;
1808 for_each_online_node(node) {
1809 cachep->nodelists[node] =
1810 kmalloc_node(sizeof(struct kmem_list3),
1812 BUG_ON(!cachep->nodelists[node]);
1813 kmem_list3_init(cachep->nodelists[node]);
1817 cachep->nodelists[numa_node_id()]->next_reap =
1818 jiffies + REAPTIMEOUT_LIST3 +
1819 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1821 cpu_cache_get(cachep)->avail = 0;
1822 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1823 cpu_cache_get(cachep)->batchcount = 1;
1824 cpu_cache_get(cachep)->touched = 0;
1825 cachep->batchcount = 1;
1826 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1830 * kmem_cache_create - Create a cache.
1831 * @name: A string which is used in /proc/slabinfo to identify this cache.
1832 * @size: The size of objects to be created in this cache.
1833 * @align: The required alignment for the objects.
1834 * @flags: SLAB flags
1835 * @ctor: A constructor for the objects.
1836 * @dtor: A destructor for the objects.
1838 * Returns a ptr to the cache on success, NULL on failure.
1839 * Cannot be called within a int, but can be interrupted.
1840 * The @ctor is run when new pages are allocated by the cache
1841 * and the @dtor is run before the pages are handed back.
1843 * @name must be valid until the cache is destroyed. This implies that
1844 * the module calling this has to destroy the cache before getting unloaded.
1848 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1849 * to catch references to uninitialised memory.
1851 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1852 * for buffer overruns.
1854 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1855 * cacheline. This can be beneficial if you're counting cycles as closely
1859 kmem_cache_create (const char *name, size_t size, size_t align,
1860 unsigned long flags,
1861 void (*ctor)(void*, struct kmem_cache *, unsigned long),
1862 void (*dtor)(void*, struct kmem_cache *, unsigned long))
1864 size_t left_over, slab_size, ralign;
1865 struct kmem_cache *cachep = NULL;
1866 struct list_head *p;
1869 * Sanity checks... these are all serious usage bugs.
1871 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
1872 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
1873 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
1879 * Prevent CPUs from coming and going.
1880 * lock_cpu_hotplug() nests outside cache_chain_mutex
1884 mutex_lock(&cache_chain_mutex);
1886 list_for_each(p, &cache_chain) {
1887 struct kmem_cache *pc = list_entry(p, struct kmem_cache, next);
1888 mm_segment_t old_fs = get_fs();
1893 * This happens when the module gets unloaded and doesn't
1894 * destroy its slab cache and no-one else reuses the vmalloc
1895 * area of the module. Print a warning.
1898 res = __get_user(tmp, pc->name);
1901 printk("SLAB: cache with size %d has lost its name\n",
1906 if (!strcmp(pc->name, name)) {
1907 printk("kmem_cache_create: duplicate cache %s\n", name);
1914 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1915 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
1916 /* No constructor, but inital state check requested */
1917 printk(KERN_ERR "%s: No con, but init state check "
1918 "requested - %s\n", __FUNCTION__, name);
1919 flags &= ~SLAB_DEBUG_INITIAL;
1923 * Enable redzoning and last user accounting, except for caches with
1924 * large objects, if the increased size would increase the object size
1925 * above the next power of two: caches with object sizes just above a
1926 * power of two have a significant amount of internal fragmentation.
1928 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
1929 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
1930 if (!(flags & SLAB_DESTROY_BY_RCU))
1931 flags |= SLAB_POISON;
1933 if (flags & SLAB_DESTROY_BY_RCU)
1934 BUG_ON(flags & SLAB_POISON);
1936 if (flags & SLAB_DESTROY_BY_RCU)
1940 * Always checks flags, a caller might be expecting debug support which
1943 if (flags & ~CREATE_MASK)
1947 * Check that size is in terms of words. This is needed to avoid
1948 * unaligned accesses for some archs when redzoning is used, and makes
1949 * sure any on-slab bufctl's are also correctly aligned.
1951 if (size & (BYTES_PER_WORD - 1)) {
1952 size += (BYTES_PER_WORD - 1);
1953 size &= ~(BYTES_PER_WORD - 1);
1956 /* calculate the final buffer alignment: */
1958 /* 1) arch recommendation: can be overridden for debug */
1959 if (flags & SLAB_HWCACHE_ALIGN) {
1961 * Default alignment: as specified by the arch code. Except if
1962 * an object is really small, then squeeze multiple objects into
1965 ralign = cache_line_size();
1966 while (size <= ralign / 2)
1969 ralign = BYTES_PER_WORD;
1971 /* 2) arch mandated alignment: disables debug if necessary */
1972 if (ralign < ARCH_SLAB_MINALIGN) {
1973 ralign = ARCH_SLAB_MINALIGN;
1974 if (ralign > BYTES_PER_WORD)
1975 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
1977 /* 3) caller mandated alignment: disables debug if necessary */
1978 if (ralign < align) {
1980 if (ralign > BYTES_PER_WORD)
1981 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
1984 * 4) Store it. Note that the debug code below can reduce
1985 * the alignment to BYTES_PER_WORD.
1989 /* Get cache's description obj. */
1990 cachep = kmem_cache_alloc(&cache_cache, SLAB_KERNEL);
1993 memset(cachep, 0, sizeof(struct kmem_cache));
1996 cachep->obj_size = size;
1998 if (flags & SLAB_RED_ZONE) {
1999 /* redzoning only works with word aligned caches */
2000 align = BYTES_PER_WORD;
2002 /* add space for red zone words */
2003 cachep->obj_offset += BYTES_PER_WORD;
2004 size += 2 * BYTES_PER_WORD;
2006 if (flags & SLAB_STORE_USER) {
2007 /* user store requires word alignment and
2008 * one word storage behind the end of the real
2011 align = BYTES_PER_WORD;
2012 size += BYTES_PER_WORD;
2014 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2015 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2016 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2017 cachep->obj_offset += PAGE_SIZE - size;
2023 /* Determine if the slab management is 'on' or 'off' slab. */
2024 if (size >= (PAGE_SIZE >> 3))
2026 * Size is large, assume best to place the slab management obj
2027 * off-slab (should allow better packing of objs).
2029 flags |= CFLGS_OFF_SLAB;
2031 size = ALIGN(size, align);
2033 left_over = calculate_slab_order(cachep, size, align, flags);
2036 printk("kmem_cache_create: couldn't create cache %s.\n", name);
2037 kmem_cache_free(&cache_cache, cachep);
2041 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2042 + sizeof(struct slab), align);
2045 * If the slab has been placed off-slab, and we have enough space then
2046 * move it on-slab. This is at the expense of any extra colouring.
2048 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2049 flags &= ~CFLGS_OFF_SLAB;
2050 left_over -= slab_size;
2053 if (flags & CFLGS_OFF_SLAB) {
2054 /* really off slab. No need for manual alignment */
2056 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2059 cachep->colour_off = cache_line_size();
2060 /* Offset must be a multiple of the alignment. */
2061 if (cachep->colour_off < align)
2062 cachep->colour_off = align;
2063 cachep->colour = left_over / cachep->colour_off;
2064 cachep->slab_size = slab_size;
2065 cachep->flags = flags;
2066 cachep->gfpflags = 0;
2067 if (flags & SLAB_CACHE_DMA)
2068 cachep->gfpflags |= GFP_DMA;
2069 cachep->buffer_size = size;
2071 if (flags & CFLGS_OFF_SLAB)
2072 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2073 cachep->ctor = ctor;
2074 cachep->dtor = dtor;
2075 cachep->name = name;
2078 setup_cpu_cache(cachep);
2080 /* cache setup completed, link it into the list */
2081 list_add(&cachep->next, &cache_chain);
2083 if (!cachep && (flags & SLAB_PANIC))
2084 panic("kmem_cache_create(): failed to create slab `%s'\n",
2086 mutex_unlock(&cache_chain_mutex);
2087 unlock_cpu_hotplug();
2090 EXPORT_SYMBOL(kmem_cache_create);
2093 static void check_irq_off(void)
2095 BUG_ON(!irqs_disabled());
2098 static void check_irq_on(void)
2100 BUG_ON(irqs_disabled());
2103 static void check_spinlock_acquired(struct kmem_cache *cachep)
2107 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2111 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2115 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2120 #define check_irq_off() do { } while(0)
2121 #define check_irq_on() do { } while(0)
2122 #define check_spinlock_acquired(x) do { } while(0)
2123 #define check_spinlock_acquired_node(x, y) do { } while(0)
2126 static void drain_array_locked(struct kmem_cache *cachep,
2127 struct array_cache *ac, int force, int node);
2129 static void do_drain(void *arg)
2131 struct kmem_cache *cachep = arg;
2132 struct array_cache *ac;
2133 int node = numa_node_id();
2136 ac = cpu_cache_get(cachep);
2137 spin_lock(&cachep->nodelists[node]->list_lock);
2138 free_block(cachep, ac->entry, ac->avail, node);
2139 spin_unlock(&cachep->nodelists[node]->list_lock);
2143 static void drain_cpu_caches(struct kmem_cache *cachep)
2145 struct kmem_list3 *l3;
2148 on_each_cpu(do_drain, cachep, 1, 1);
2150 for_each_online_node(node) {
2151 l3 = cachep->nodelists[node];
2153 spin_lock_irq(&l3->list_lock);
2154 drain_array_locked(cachep, l3->shared, 1, node);
2155 spin_unlock_irq(&l3->list_lock);
2157 drain_alien_cache(cachep, l3->alien);
2162 static int __node_shrink(struct kmem_cache *cachep, int node)
2165 struct kmem_list3 *l3 = cachep->nodelists[node];
2169 struct list_head *p;
2171 p = l3->slabs_free.prev;
2172 if (p == &l3->slabs_free)
2175 slabp = list_entry(l3->slabs_free.prev, struct slab, list);
2180 list_del(&slabp->list);
2182 l3->free_objects -= cachep->num;
2183 spin_unlock_irq(&l3->list_lock);
2184 slab_destroy(cachep, slabp);
2185 spin_lock_irq(&l3->list_lock);
2187 ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial);
2191 static int __cache_shrink(struct kmem_cache *cachep)
2194 struct kmem_list3 *l3;
2196 drain_cpu_caches(cachep);
2199 for_each_online_node(i) {
2200 l3 = cachep->nodelists[i];
2202 spin_lock_irq(&l3->list_lock);
2203 ret += __node_shrink(cachep, i);
2204 spin_unlock_irq(&l3->list_lock);
2207 return (ret ? 1 : 0);
2211 * kmem_cache_shrink - Shrink a cache.
2212 * @cachep: The cache to shrink.
2214 * Releases as many slabs as possible for a cache.
2215 * To help debugging, a zero exit status indicates all slabs were released.
2217 int kmem_cache_shrink(struct kmem_cache *cachep)
2219 if (!cachep || in_interrupt())
2222 return __cache_shrink(cachep);
2224 EXPORT_SYMBOL(kmem_cache_shrink);
2227 * kmem_cache_destroy - delete a cache
2228 * @cachep: the cache to destroy
2230 * Remove a struct kmem_cache object from the slab cache.
2231 * Returns 0 on success.
2233 * It is expected this function will be called by a module when it is
2234 * unloaded. This will remove the cache completely, and avoid a duplicate
2235 * cache being allocated each time a module is loaded and unloaded, if the
2236 * module doesn't have persistent in-kernel storage across loads and unloads.
2238 * The cache must be empty before calling this function.
2240 * The caller must guarantee that noone will allocate memory from the cache
2241 * during the kmem_cache_destroy().
2243 int kmem_cache_destroy(struct kmem_cache *cachep)
2246 struct kmem_list3 *l3;
2248 if (!cachep || in_interrupt())
2251 /* Don't let CPUs to come and go */
2254 /* Find the cache in the chain of caches. */
2255 mutex_lock(&cache_chain_mutex);
2257 * the chain is never empty, cache_cache is never destroyed
2259 list_del(&cachep->next);
2260 mutex_unlock(&cache_chain_mutex);
2262 if (__cache_shrink(cachep)) {
2263 slab_error(cachep, "Can't free all objects");
2264 mutex_lock(&cache_chain_mutex);
2265 list_add(&cachep->next, &cache_chain);
2266 mutex_unlock(&cache_chain_mutex);
2267 unlock_cpu_hotplug();
2271 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2274 for_each_online_cpu(i)
2275 kfree(cachep->array[i]);
2277 /* NUMA: free the list3 structures */
2278 for_each_online_node(i) {
2279 l3 = cachep->nodelists[i];
2282 free_alien_cache(l3->alien);
2286 kmem_cache_free(&cache_cache, cachep);
2287 unlock_cpu_hotplug();
2290 EXPORT_SYMBOL(kmem_cache_destroy);
2292 /* Get the memory for a slab management obj. */
2293 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2294 int colour_off, gfp_t local_flags)
2298 if (OFF_SLAB(cachep)) {
2299 /* Slab management obj is off-slab. */
2300 slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags);
2304 slabp = objp + colour_off;
2305 colour_off += cachep->slab_size;
2308 slabp->colouroff = colour_off;
2309 slabp->s_mem = objp + colour_off;
2313 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2315 return (kmem_bufctl_t *) (slabp + 1);
2318 static void cache_init_objs(struct kmem_cache *cachep,
2319 struct slab *slabp, unsigned long ctor_flags)
2323 for (i = 0; i < cachep->num; i++) {
2324 void *objp = index_to_obj(cachep, slabp, i);
2326 /* need to poison the objs? */
2327 if (cachep->flags & SLAB_POISON)
2328 poison_obj(cachep, objp, POISON_FREE);
2329 if (cachep->flags & SLAB_STORE_USER)
2330 *dbg_userword(cachep, objp) = NULL;
2332 if (cachep->flags & SLAB_RED_ZONE) {
2333 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2334 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2337 * Constructors are not allowed to allocate memory from the same
2338 * cache which they are a constructor for. Otherwise, deadlock.
2339 * They must also be threaded.
2341 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2342 cachep->ctor(objp + obj_offset(cachep), cachep,
2345 if (cachep->flags & SLAB_RED_ZONE) {
2346 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2347 slab_error(cachep, "constructor overwrote the"
2348 " end of an object");
2349 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2350 slab_error(cachep, "constructor overwrote the"
2351 " start of an object");
2353 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2354 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2355 kernel_map_pages(virt_to_page(objp),
2356 cachep->buffer_size / PAGE_SIZE, 0);
2359 cachep->ctor(objp, cachep, ctor_flags);
2361 slab_bufctl(slabp)[i] = i + 1;
2363 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2367 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2369 if (flags & SLAB_DMA)
2370 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2372 BUG_ON(cachep->gfpflags & GFP_DMA);
2375 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2378 void *objp = index_to_obj(cachep, slabp, slabp->free);
2382 next = slab_bufctl(slabp)[slabp->free];
2384 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2385 WARN_ON(slabp->nodeid != nodeid);
2392 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2393 void *objp, int nodeid)
2395 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2398 /* Verify that the slab belongs to the intended node */
2399 WARN_ON(slabp->nodeid != nodeid);
2401 if (slab_bufctl(slabp)[objnr] != BUFCTL_FREE) {
2402 printk(KERN_ERR "slab: double free detected in cache "
2403 "'%s', objp %p\n", cachep->name, objp);
2407 slab_bufctl(slabp)[objnr] = slabp->free;
2408 slabp->free = objnr;
2412 static void set_slab_attr(struct kmem_cache *cachep, struct slab *slabp,
2418 /* Nasty!!!!!! I hope this is OK. */
2419 page = virt_to_page(objp);
2422 if (likely(!PageCompound(page)))
2423 i <<= cachep->gfporder;
2425 page_set_cache(page, cachep);
2426 page_set_slab(page, slabp);
2432 * Grow (by 1) the number of slabs within a cache. This is called by
2433 * kmem_cache_alloc() when there are no active objs left in a cache.
2435 static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid)
2441 unsigned long ctor_flags;
2442 struct kmem_list3 *l3;
2445 * Be lazy and only check for valid flags here, keeping it out of the
2446 * critical path in kmem_cache_alloc().
2448 if (flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW))
2450 if (flags & SLAB_NO_GROW)
2453 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2454 local_flags = (flags & SLAB_LEVEL_MASK);
2455 if (!(local_flags & __GFP_WAIT))
2457 * Not allowed to sleep. Need to tell a constructor about
2458 * this - it might need to know...
2460 ctor_flags |= SLAB_CTOR_ATOMIC;
2462 /* Take the l3 list lock to change the colour_next on this node */
2464 l3 = cachep->nodelists[nodeid];
2465 spin_lock(&l3->list_lock);
2467 /* Get colour for the slab, and cal the next value. */
2468 offset = l3->colour_next;
2470 if (l3->colour_next >= cachep->colour)
2471 l3->colour_next = 0;
2472 spin_unlock(&l3->list_lock);
2474 offset *= cachep->colour_off;
2476 if (local_flags & __GFP_WAIT)
2480 * The test for missing atomic flag is performed here, rather than
2481 * the more obvious place, simply to reduce the critical path length
2482 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2483 * will eventually be caught here (where it matters).
2485 kmem_flagcheck(cachep, flags);
2488 * Get mem for the objs. Attempt to allocate a physical page from
2491 objp = kmem_getpages(cachep, flags, nodeid);
2495 /* Get slab management. */
2496 slabp = alloc_slabmgmt(cachep, objp, offset, local_flags);
2500 slabp->nodeid = nodeid;
2501 set_slab_attr(cachep, slabp, objp);
2503 cache_init_objs(cachep, slabp, ctor_flags);
2505 if (local_flags & __GFP_WAIT)
2506 local_irq_disable();
2508 spin_lock(&l3->list_lock);
2510 /* Make slab active. */
2511 list_add_tail(&slabp->list, &(l3->slabs_free));
2512 STATS_INC_GROWN(cachep);
2513 l3->free_objects += cachep->num;
2514 spin_unlock(&l3->list_lock);
2517 kmem_freepages(cachep, objp);
2519 if (local_flags & __GFP_WAIT)
2520 local_irq_disable();
2527 * Perform extra freeing checks:
2528 * - detect bad pointers.
2529 * - POISON/RED_ZONE checking
2530 * - destructor calls, for caches with POISON+dtor
2532 static void kfree_debugcheck(const void *objp)
2536 if (!virt_addr_valid(objp)) {
2537 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2538 (unsigned long)objp);
2541 page = virt_to_page(objp);
2542 if (!PageSlab(page)) {
2543 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2544 (unsigned long)objp);
2549 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2556 objp -= obj_offset(cachep);
2557 kfree_debugcheck(objp);
2558 page = virt_to_page(objp);
2560 if (page_get_cache(page) != cachep) {
2561 printk(KERN_ERR "mismatch in kmem_cache_free: expected "
2562 "cache %p, got %p\n",
2563 page_get_cache(page), cachep);
2564 printk(KERN_ERR "%p is %s.\n", cachep, cachep->name);
2565 printk(KERN_ERR "%p is %s.\n", page_get_cache(page),
2566 page_get_cache(page)->name);
2569 slabp = page_get_slab(page);
2571 if (cachep->flags & SLAB_RED_ZONE) {
2572 if (*dbg_redzone1(cachep, objp) != RED_ACTIVE ||
2573 *dbg_redzone2(cachep, objp) != RED_ACTIVE) {
2574 slab_error(cachep, "double free, or memory outside"
2575 " object was overwritten");
2576 printk(KERN_ERR "%p: redzone 1:0x%lx, "
2577 "redzone 2:0x%lx.\n",
2578 objp, *dbg_redzone1(cachep, objp),
2579 *dbg_redzone2(cachep, objp));
2581 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2582 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2584 if (cachep->flags & SLAB_STORE_USER)
2585 *dbg_userword(cachep, objp) = caller;
2587 objnr = obj_to_index(cachep, slabp, objp);
2589 BUG_ON(objnr >= cachep->num);
2590 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2592 if (cachep->flags & SLAB_DEBUG_INITIAL) {
2594 * Need to call the slab's constructor so the caller can
2595 * perform a verify of its state (debugging). Called without
2596 * the cache-lock held.
2598 cachep->ctor(objp + obj_offset(cachep),
2599 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
2601 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2602 /* we want to cache poison the object,
2603 * call the destruction callback
2605 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
2607 if (cachep->flags & SLAB_POISON) {
2608 #ifdef CONFIG_DEBUG_PAGEALLOC
2609 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2610 store_stackinfo(cachep, objp, (unsigned long)caller);
2611 kernel_map_pages(virt_to_page(objp),
2612 cachep->buffer_size / PAGE_SIZE, 0);
2614 poison_obj(cachep, objp, POISON_FREE);
2617 poison_obj(cachep, objp, POISON_FREE);
2623 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2628 /* Check slab's freelist to see if this obj is there. */
2629 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2631 if (entries > cachep->num || i >= cachep->num)
2634 if (entries != cachep->num - slabp->inuse) {
2636 printk(KERN_ERR "slab: Internal list corruption detected in "
2637 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2638 cachep->name, cachep->num, slabp, slabp->inuse);
2640 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2643 printk("\n%03x:", i);
2644 printk(" %02x", ((unsigned char *)slabp)[i]);
2651 #define kfree_debugcheck(x) do { } while(0)
2652 #define cache_free_debugcheck(x,objp,z) (objp)
2653 #define check_slabp(x,y) do { } while(0)
2656 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2659 struct kmem_list3 *l3;
2660 struct array_cache *ac;
2663 ac = cpu_cache_get(cachep);
2665 batchcount = ac->batchcount;
2666 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2668 * If there was little recent activity on this cache, then
2669 * perform only a partial refill. Otherwise we could generate
2672 batchcount = BATCHREFILL_LIMIT;
2674 l3 = cachep->nodelists[numa_node_id()];
2676 BUG_ON(ac->avail > 0 || !l3);
2677 spin_lock(&l3->list_lock);
2680 struct array_cache *shared_array = l3->shared;
2681 if (shared_array->avail) {
2682 if (batchcount > shared_array->avail)
2683 batchcount = shared_array->avail;
2684 shared_array->avail -= batchcount;
2685 ac->avail = batchcount;
2687 &(shared_array->entry[shared_array->avail]),
2688 sizeof(void *) * batchcount);
2689 shared_array->touched = 1;
2693 while (batchcount > 0) {
2694 struct list_head *entry;
2696 /* Get slab alloc is to come from. */
2697 entry = l3->slabs_partial.next;
2698 if (entry == &l3->slabs_partial) {
2699 l3->free_touched = 1;
2700 entry = l3->slabs_free.next;
2701 if (entry == &l3->slabs_free)
2705 slabp = list_entry(entry, struct slab, list);
2706 check_slabp(cachep, slabp);
2707 check_spinlock_acquired(cachep);
2708 while (slabp->inuse < cachep->num && batchcount--) {
2709 STATS_INC_ALLOCED(cachep);
2710 STATS_INC_ACTIVE(cachep);
2711 STATS_SET_HIGH(cachep);
2713 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2716 check_slabp(cachep, slabp);
2718 /* move slabp to correct slabp list: */
2719 list_del(&slabp->list);
2720 if (slabp->free == BUFCTL_END)
2721 list_add(&slabp->list, &l3->slabs_full);
2723 list_add(&slabp->list, &l3->slabs_partial);
2727 l3->free_objects -= ac->avail;
2729 spin_unlock(&l3->list_lock);
2731 if (unlikely(!ac->avail)) {
2733 x = cache_grow(cachep, flags, numa_node_id());
2735 /* cache_grow can reenable interrupts, then ac could change. */
2736 ac = cpu_cache_get(cachep);
2737 if (!x && ac->avail == 0) /* no objects in sight? abort */
2740 if (!ac->avail) /* objects refilled by interrupt? */
2744 return ac->entry[--ac->avail];
2747 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2750 might_sleep_if(flags & __GFP_WAIT);
2752 kmem_flagcheck(cachep, flags);
2757 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2758 gfp_t flags, void *objp, void *caller)
2762 if (cachep->flags & SLAB_POISON) {
2763 #ifdef CONFIG_DEBUG_PAGEALLOC
2764 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
2765 kernel_map_pages(virt_to_page(objp),
2766 cachep->buffer_size / PAGE_SIZE, 1);
2768 check_poison_obj(cachep, objp);
2770 check_poison_obj(cachep, objp);
2772 poison_obj(cachep, objp, POISON_INUSE);
2774 if (cachep->flags & SLAB_STORE_USER)
2775 *dbg_userword(cachep, objp) = caller;
2777 if (cachep->flags & SLAB_RED_ZONE) {
2778 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2779 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2780 slab_error(cachep, "double free, or memory outside"
2781 " object was overwritten");
2783 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
2784 objp, *dbg_redzone1(cachep, objp),
2785 *dbg_redzone2(cachep, objp));
2787 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2788 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2790 objp += obj_offset(cachep);
2791 if (cachep->ctor && cachep->flags & SLAB_POISON) {
2792 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2794 if (!(flags & __GFP_WAIT))
2795 ctor_flags |= SLAB_CTOR_ATOMIC;
2797 cachep->ctor(objp, cachep, ctor_flags);
2802 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2805 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
2808 struct array_cache *ac;
2811 if (unlikely(current->mempolicy && !in_interrupt())) {
2812 int nid = slab_node(current->mempolicy);
2814 if (nid != numa_node_id())
2815 return __cache_alloc_node(cachep, flags, nid);
2820 ac = cpu_cache_get(cachep);
2821 if (likely(ac->avail)) {
2822 STATS_INC_ALLOCHIT(cachep);
2824 objp = ac->entry[--ac->avail];
2826 STATS_INC_ALLOCMISS(cachep);
2827 objp = cache_alloc_refill(cachep, flags);
2832 static __always_inline void *__cache_alloc(struct kmem_cache *cachep,
2833 gfp_t flags, void *caller)
2835 unsigned long save_flags;
2838 cache_alloc_debugcheck_before(cachep, flags);
2840 local_irq_save(save_flags);
2841 objp = ____cache_alloc(cachep, flags);
2842 local_irq_restore(save_flags);
2843 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
2851 * A interface to enable slab creation on nodeid
2853 static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
2856 struct list_head *entry;
2858 struct kmem_list3 *l3;
2862 l3 = cachep->nodelists[nodeid];
2867 spin_lock(&l3->list_lock);
2868 entry = l3->slabs_partial.next;
2869 if (entry == &l3->slabs_partial) {
2870 l3->free_touched = 1;
2871 entry = l3->slabs_free.next;
2872 if (entry == &l3->slabs_free)
2876 slabp = list_entry(entry, struct slab, list);
2877 check_spinlock_acquired_node(cachep, nodeid);
2878 check_slabp(cachep, slabp);
2880 STATS_INC_NODEALLOCS(cachep);
2881 STATS_INC_ACTIVE(cachep);
2882 STATS_SET_HIGH(cachep);
2884 BUG_ON(slabp->inuse == cachep->num);
2886 obj = slab_get_obj(cachep, slabp, nodeid);
2887 check_slabp(cachep, slabp);
2889 /* move slabp to correct slabp list: */
2890 list_del(&slabp->list);
2892 if (slabp->free == BUFCTL_END)
2893 list_add(&slabp->list, &l3->slabs_full);
2895 list_add(&slabp->list, &l3->slabs_partial);
2897 spin_unlock(&l3->list_lock);
2901 spin_unlock(&l3->list_lock);
2902 x = cache_grow(cachep, flags, nodeid);
2914 * Caller needs to acquire correct kmem_list's list_lock
2916 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
2920 struct kmem_list3 *l3;
2922 for (i = 0; i < nr_objects; i++) {
2923 void *objp = objpp[i];
2926 slabp = virt_to_slab(objp);
2927 l3 = cachep->nodelists[node];
2928 list_del(&slabp->list);
2929 check_spinlock_acquired_node(cachep, node);
2930 check_slabp(cachep, slabp);
2931 slab_put_obj(cachep, slabp, objp, node);
2932 STATS_DEC_ACTIVE(cachep);
2934 check_slabp(cachep, slabp);
2936 /* fixup slab chains */
2937 if (slabp->inuse == 0) {
2938 if (l3->free_objects > l3->free_limit) {
2939 l3->free_objects -= cachep->num;
2940 slab_destroy(cachep, slabp);
2942 list_add(&slabp->list, &l3->slabs_free);
2945 /* Unconditionally move a slab to the end of the
2946 * partial list on free - maximum time for the
2947 * other objects to be freed, too.
2949 list_add_tail(&slabp->list, &l3->slabs_partial);
2954 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
2957 struct kmem_list3 *l3;
2958 int node = numa_node_id();
2960 batchcount = ac->batchcount;
2962 BUG_ON(!batchcount || batchcount > ac->avail);
2965 l3 = cachep->nodelists[node];
2966 spin_lock(&l3->list_lock);
2968 struct array_cache *shared_array = l3->shared;
2969 int max = shared_array->limit - shared_array->avail;
2971 if (batchcount > max)
2973 memcpy(&(shared_array->entry[shared_array->avail]),
2974 ac->entry, sizeof(void *) * batchcount);
2975 shared_array->avail += batchcount;
2980 free_block(cachep, ac->entry, batchcount, node);
2985 struct list_head *p;
2987 p = l3->slabs_free.next;
2988 while (p != &(l3->slabs_free)) {
2991 slabp = list_entry(p, struct slab, list);
2992 BUG_ON(slabp->inuse);
2997 STATS_SET_FREEABLE(cachep, i);
3000 spin_unlock(&l3->list_lock);
3001 ac->avail -= batchcount;
3002 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3006 * Release an obj back to its cache. If the obj has a constructed state, it must
3007 * be in this state _before_ it is released. Called with disabled ints.
3009 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3011 struct array_cache *ac = cpu_cache_get(cachep);
3014 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3016 /* Make sure we are not freeing a object from another
3017 * node to the array cache on this cpu.
3022 slabp = virt_to_slab(objp);
3023 if (unlikely(slabp->nodeid != numa_node_id())) {
3024 struct array_cache *alien = NULL;
3025 int nodeid = slabp->nodeid;
3026 struct kmem_list3 *l3;
3028 l3 = cachep->nodelists[numa_node_id()];
3029 STATS_INC_NODEFREES(cachep);
3030 if (l3->alien && l3->alien[nodeid]) {
3031 alien = l3->alien[nodeid];
3032 spin_lock(&alien->lock);
3033 if (unlikely(alien->avail == alien->limit))
3034 __drain_alien_cache(cachep,
3036 alien->entry[alien->avail++] = objp;
3037 spin_unlock(&alien->lock);
3039 spin_lock(&(cachep->nodelists[nodeid])->
3041 free_block(cachep, &objp, 1, nodeid);
3042 spin_unlock(&(cachep->nodelists[nodeid])->
3049 if (likely(ac->avail < ac->limit)) {
3050 STATS_INC_FREEHIT(cachep);
3051 ac->entry[ac->avail++] = objp;
3054 STATS_INC_FREEMISS(cachep);
3055 cache_flusharray(cachep, ac);
3056 ac->entry[ac->avail++] = objp;
3061 * kmem_cache_alloc - Allocate an object
3062 * @cachep: The cache to allocate from.
3063 * @flags: See kmalloc().
3065 * Allocate an object from this cache. The flags are only relevant
3066 * if the cache has no available objects.
3068 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3070 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3072 EXPORT_SYMBOL(kmem_cache_alloc);
3075 * kmem_ptr_validate - check if an untrusted pointer might
3077 * @cachep: the cache we're checking against
3078 * @ptr: pointer to validate
3080 * This verifies that the untrusted pointer looks sane:
3081 * it is _not_ a guarantee that the pointer is actually
3082 * part of the slab cache in question, but it at least
3083 * validates that the pointer can be dereferenced and
3084 * looks half-way sane.
3086 * Currently only used for dentry validation.
3088 int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr)
3090 unsigned long addr = (unsigned long)ptr;
3091 unsigned long min_addr = PAGE_OFFSET;
3092 unsigned long align_mask = BYTES_PER_WORD - 1;
3093 unsigned long size = cachep->buffer_size;
3096 if (unlikely(addr < min_addr))
3098 if (unlikely(addr > (unsigned long)high_memory - size))
3100 if (unlikely(addr & align_mask))
3102 if (unlikely(!kern_addr_valid(addr)))
3104 if (unlikely(!kern_addr_valid(addr + size - 1)))
3106 page = virt_to_page(ptr);
3107 if (unlikely(!PageSlab(page)))
3109 if (unlikely(page_get_cache(page) != cachep))
3118 * kmem_cache_alloc_node - Allocate an object on the specified node
3119 * @cachep: The cache to allocate from.
3120 * @flags: See kmalloc().
3121 * @nodeid: node number of the target node.
3123 * Identical to kmem_cache_alloc, except that this function is slow
3124 * and can sleep. And it will allocate memory on the given node, which
3125 * can improve the performance for cpu bound structures.
3126 * New and improved: it will now make sure that the object gets
3127 * put on the correct node list so that there is no false sharing.
3129 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3131 unsigned long save_flags;
3134 cache_alloc_debugcheck_before(cachep, flags);
3135 local_irq_save(save_flags);
3137 if (nodeid == -1 || nodeid == numa_node_id() ||
3138 !cachep->nodelists[nodeid])
3139 ptr = ____cache_alloc(cachep, flags);
3141 ptr = __cache_alloc_node(cachep, flags, nodeid);
3142 local_irq_restore(save_flags);
3144 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr,
3145 __builtin_return_address(0));
3149 EXPORT_SYMBOL(kmem_cache_alloc_node);
3151 void *kmalloc_node(size_t size, gfp_t flags, int node)
3153 struct kmem_cache *cachep;
3155 cachep = kmem_find_general_cachep(size, flags);
3156 if (unlikely(cachep == NULL))
3158 return kmem_cache_alloc_node(cachep, flags, node);
3160 EXPORT_SYMBOL(kmalloc_node);
3164 * kmalloc - allocate memory
3165 * @size: how many bytes of memory are required.
3166 * @flags: the type of memory to allocate.
3167 * @caller: function caller for debug tracking of the caller
3169 * kmalloc is the normal method of allocating memory
3172 * The @flags argument may be one of:
3174 * %GFP_USER - Allocate memory on behalf of user. May sleep.
3176 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
3178 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
3180 * Additionally, the %GFP_DMA flag may be set to indicate the memory
3181 * must be suitable for DMA. This can mean different things on different
3182 * platforms. For example, on i386, it means that the memory must come
3183 * from the first 16MB.
3185 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3188 struct kmem_cache *cachep;
3190 /* If you want to save a few bytes .text space: replace
3192 * Then kmalloc uses the uninlined functions instead of the inline
3195 cachep = __find_general_cachep(size, flags);
3196 if (unlikely(cachep == NULL))
3198 return __cache_alloc(cachep, flags, caller);
3201 #ifndef CONFIG_DEBUG_SLAB
3203 void *__kmalloc(size_t size, gfp_t flags)
3205 return __do_kmalloc(size, flags, NULL);
3207 EXPORT_SYMBOL(__kmalloc);
3211 void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3213 return __do_kmalloc(size, flags, caller);
3215 EXPORT_SYMBOL(__kmalloc_track_caller);
3221 * __alloc_percpu - allocate one copy of the object for every present
3222 * cpu in the system, zeroing them.
3223 * Objects should be dereferenced using the per_cpu_ptr macro only.
3225 * @size: how many bytes of memory are required.
3227 void *__alloc_percpu(size_t size)
3230 struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL);
3236 * Cannot use for_each_online_cpu since a cpu may come online
3237 * and we have no way of figuring out how to fix the array
3238 * that we have allocated then....
3241 int node = cpu_to_node(i);
3243 if (node_online(node))
3244 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
3246 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
3248 if (!pdata->ptrs[i])
3250 memset(pdata->ptrs[i], 0, size);
3253 /* Catch derefs w/o wrappers */
3254 return (void *)(~(unsigned long)pdata);
3258 if (!cpu_possible(i))
3260 kfree(pdata->ptrs[i]);
3265 EXPORT_SYMBOL(__alloc_percpu);
3269 * kmem_cache_free - Deallocate an object
3270 * @cachep: The cache the allocation was from.
3271 * @objp: The previously allocated object.
3273 * Free an object which was previously allocated from this
3276 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3278 unsigned long flags;
3280 local_irq_save(flags);
3281 __cache_free(cachep, objp);
3282 local_irq_restore(flags);
3284 EXPORT_SYMBOL(kmem_cache_free);
3287 * kfree - free previously allocated memory
3288 * @objp: pointer returned by kmalloc.
3290 * If @objp is NULL, no operation is performed.
3292 * Don't free memory not originally allocated by kmalloc()
3293 * or you will run into trouble.
3295 void kfree(const void *objp)
3297 struct kmem_cache *c;
3298 unsigned long flags;
3300 if (unlikely(!objp))
3302 local_irq_save(flags);
3303 kfree_debugcheck(objp);
3304 c = virt_to_cache(objp);
3305 mutex_debug_check_no_locks_freed(objp, obj_size(c));
3306 __cache_free(c, (void *)objp);
3307 local_irq_restore(flags);
3309 EXPORT_SYMBOL(kfree);
3313 * free_percpu - free previously allocated percpu memory
3314 * @objp: pointer returned by alloc_percpu.
3316 * Don't free memory not originally allocated by alloc_percpu()
3317 * The complemented objp is to check for that.
3319 void free_percpu(const void *objp)
3322 struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp);
3325 * We allocate for all cpus so we cannot use for online cpu here.
3331 EXPORT_SYMBOL(free_percpu);
3334 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3336 return obj_size(cachep);
3338 EXPORT_SYMBOL(kmem_cache_size);
3340 const char *kmem_cache_name(struct kmem_cache *cachep)
3342 return cachep->name;
3344 EXPORT_SYMBOL_GPL(kmem_cache_name);
3347 * This initializes kmem_list3 for all nodes.
3349 static int alloc_kmemlist(struct kmem_cache *cachep)
3352 struct kmem_list3 *l3;
3355 for_each_online_node(node) {
3356 struct array_cache *nc = NULL, *new;
3357 struct array_cache **new_alien = NULL;
3359 new_alien = alloc_alien_cache(node, cachep->limit);
3363 new = alloc_arraycache(node, cachep->shared*cachep->batchcount,
3367 l3 = cachep->nodelists[node];
3369 spin_lock_irq(&l3->list_lock);
3371 nc = cachep->nodelists[node]->shared;
3373 free_block(cachep, nc->entry, nc->avail, node);
3376 if (!cachep->nodelists[node]->alien) {
3377 l3->alien = new_alien;
3380 l3->free_limit = (1 + nr_cpus_node(node)) *
3381 cachep->batchcount + cachep->num;
3382 spin_unlock_irq(&l3->list_lock);
3384 free_alien_cache(new_alien);
3387 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
3391 kmem_list3_init(l3);
3392 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3393 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3395 l3->alien = new_alien;
3396 l3->free_limit = (1 + nr_cpus_node(node)) *
3397 cachep->batchcount + cachep->num;
3398 cachep->nodelists[node] = l3;
3406 struct ccupdate_struct {
3407 struct kmem_cache *cachep;
3408 struct array_cache *new[NR_CPUS];
3411 static void do_ccupdate_local(void *info)
3413 struct ccupdate_struct *new = info;
3414 struct array_cache *old;
3417 old = cpu_cache_get(new->cachep);
3419 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3420 new->new[smp_processor_id()] = old;
3423 /* Always called with the cache_chain_mutex held */
3424 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3425 int batchcount, int shared)
3427 struct ccupdate_struct new;
3430 memset(&new.new, 0, sizeof(new.new));
3431 for_each_online_cpu(i) {
3432 new.new[i] = alloc_arraycache(cpu_to_node(i), limit,
3435 for (i--; i >= 0; i--)
3440 new.cachep = cachep;
3442 on_each_cpu(do_ccupdate_local, (void *)&new, 1, 1);
3445 cachep->batchcount = batchcount;
3446 cachep->limit = limit;
3447 cachep->shared = shared;
3449 for_each_online_cpu(i) {
3450 struct array_cache *ccold = new.new[i];
3453 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3454 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3455 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3459 err = alloc_kmemlist(cachep);
3461 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
3462 cachep->name, -err);
3468 /* Called with cache_chain_mutex held always */
3469 static void enable_cpucache(struct kmem_cache *cachep)
3475 * The head array serves three purposes:
3476 * - create a LIFO ordering, i.e. return objects that are cache-warm
3477 * - reduce the number of spinlock operations.
3478 * - reduce the number of linked list operations on the slab and
3479 * bufctl chains: array operations are cheaper.
3480 * The numbers are guessed, we should auto-tune as described by
3483 if (cachep->buffer_size > 131072)
3485 else if (cachep->buffer_size > PAGE_SIZE)
3487 else if (cachep->buffer_size > 1024)
3489 else if (cachep->buffer_size > 256)
3495 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3496 * allocation behaviour: Most allocs on one cpu, most free operations
3497 * on another cpu. For these cases, an efficient object passing between
3498 * cpus is necessary. This is provided by a shared array. The array
3499 * replaces Bonwick's magazine layer.
3500 * On uniprocessor, it's functionally equivalent (but less efficient)
3501 * to a larger limit. Thus disabled by default.
3505 if (cachep->buffer_size <= PAGE_SIZE)
3511 * With debugging enabled, large batchcount lead to excessively long
3512 * periods with disabled local interrupts. Limit the batchcount
3517 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
3519 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3520 cachep->name, -err);
3523 static void drain_array_locked(struct kmem_cache *cachep,
3524 struct array_cache *ac, int force, int node)
3528 check_spinlock_acquired_node(cachep, node);
3529 if (ac->touched && !force) {
3531 } else if (ac->avail) {
3532 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3533 if (tofree > ac->avail)
3534 tofree = (ac->avail + 1) / 2;
3535 free_block(cachep, ac->entry, tofree, node);
3536 ac->avail -= tofree;
3537 memmove(ac->entry, &(ac->entry[tofree]),
3538 sizeof(void *) * ac->avail);
3543 * cache_reap - Reclaim memory from caches.
3544 * @unused: unused parameter
3546 * Called from workqueue/eventd every few seconds.
3548 * - clear the per-cpu caches for this CPU.
3549 * - return freeable pages to the main free memory pool.
3551 * If we cannot acquire the cache chain mutex then just give up - we'll try
3552 * again on the next iteration.
3554 static void cache_reap(void *unused)
3556 struct list_head *walk;
3557 struct kmem_list3 *l3;
3559 if (!mutex_trylock(&cache_chain_mutex)) {
3560 /* Give up. Setup the next iteration. */
3561 schedule_delayed_work(&__get_cpu_var(reap_work),
3566 list_for_each(walk, &cache_chain) {
3567 struct kmem_cache *searchp;
3568 struct list_head *p;
3572 searchp = list_entry(walk, struct kmem_cache, next);
3575 l3 = searchp->nodelists[numa_node_id()];
3576 reap_alien(searchp, l3);
3577 spin_lock_irq(&l3->list_lock);
3579 drain_array_locked(searchp, cpu_cache_get(searchp), 0,
3582 if (time_after(l3->next_reap, jiffies))
3585 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
3588 drain_array_locked(searchp, l3->shared, 0,
3591 if (l3->free_touched) {
3592 l3->free_touched = 0;
3596 tofree = (l3->free_limit + 5 * searchp->num - 1) /
3599 p = l3->slabs_free.next;
3600 if (p == &(l3->slabs_free))
3603 slabp = list_entry(p, struct slab, list);
3604 BUG_ON(slabp->inuse);
3605 list_del(&slabp->list);
3606 STATS_INC_REAPED(searchp);
3609 * Safe to drop the lock. The slab is no longer linked
3610 * to the cache. searchp cannot disappear, we hold
3613 l3->free_objects -= searchp->num;
3614 spin_unlock_irq(&l3->list_lock);
3615 slab_destroy(searchp, slabp);
3616 spin_lock_irq(&l3->list_lock);
3617 } while (--tofree > 0);
3619 spin_unlock_irq(&l3->list_lock);
3623 mutex_unlock(&cache_chain_mutex);
3625 /* Set up the next iteration */
3626 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
3629 #ifdef CONFIG_PROC_FS
3631 static void print_slabinfo_header(struct seq_file *m)
3634 * Output format version, so at least we can change it
3635 * without _too_ many complaints.
3638 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3640 seq_puts(m, "slabinfo - version: 2.1\n");
3642 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3643 "<objperslab> <pagesperslab>");
3644 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3645 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3647 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3648 "<error> <maxfreeable> <nodeallocs> <remotefrees>");
3649 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3654 static void *s_start(struct seq_file *m, loff_t *pos)
3657 struct list_head *p;
3659 mutex_lock(&cache_chain_mutex);
3661 print_slabinfo_header(m);
3662 p = cache_chain.next;
3665 if (p == &cache_chain)
3668 return list_entry(p, struct kmem_cache, next);
3671 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3673 struct kmem_cache *cachep = p;
3675 return cachep->next.next == &cache_chain ?
3676 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
3679 static void s_stop(struct seq_file *m, void *p)
3681 mutex_unlock(&cache_chain_mutex);
3684 static int s_show(struct seq_file *m, void *p)
3686 struct kmem_cache *cachep = p;
3687 struct list_head *q;
3689 unsigned long active_objs;
3690 unsigned long num_objs;
3691 unsigned long active_slabs = 0;
3692 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3696 struct kmem_list3 *l3;
3700 for_each_online_node(node) {
3701 l3 = cachep->nodelists[node];
3706 spin_lock_irq(&l3->list_lock);
3708 list_for_each(q, &l3->slabs_full) {
3709 slabp = list_entry(q, struct slab, list);
3710 if (slabp->inuse != cachep->num && !error)
3711 error = "slabs_full accounting error";
3712 active_objs += cachep->num;
3715 list_for_each(q, &l3->slabs_partial) {
3716 slabp = list_entry(q, struct slab, list);
3717 if (slabp->inuse == cachep->num && !error)
3718 error = "slabs_partial inuse accounting error";
3719 if (!slabp->inuse && !error)
3720 error = "slabs_partial/inuse accounting error";
3721 active_objs += slabp->inuse;
3724 list_for_each(q, &l3->slabs_free) {
3725 slabp = list_entry(q, struct slab, list);
3726 if (slabp->inuse && !error)
3727 error = "slabs_free/inuse accounting error";
3730 free_objects += l3->free_objects;
3732 shared_avail += l3->shared->avail;
3734 spin_unlock_irq(&l3->list_lock);
3736 num_slabs += active_slabs;
3737 num_objs = num_slabs * cachep->num;
3738 if (num_objs - active_objs != free_objects && !error)
3739 error = "free_objects accounting error";
3741 name = cachep->name;
3743 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3745 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
3746 name, active_objs, num_objs, cachep->buffer_size,
3747 cachep->num, (1 << cachep->gfporder));
3748 seq_printf(m, " : tunables %4u %4u %4u",
3749 cachep->limit, cachep->batchcount, cachep->shared);
3750 seq_printf(m, " : slabdata %6lu %6lu %6lu",
3751 active_slabs, num_slabs, shared_avail);
3754 unsigned long high = cachep->high_mark;
3755 unsigned long allocs = cachep->num_allocations;
3756 unsigned long grown = cachep->grown;
3757 unsigned long reaped = cachep->reaped;
3758 unsigned long errors = cachep->errors;
3759 unsigned long max_freeable = cachep->max_freeable;
3760 unsigned long node_allocs = cachep->node_allocs;
3761 unsigned long node_frees = cachep->node_frees;
3763 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
3764 %4lu %4lu %4lu %4lu", allocs, high, grown,
3765 reaped, errors, max_freeable, node_allocs,
3770 unsigned long allochit = atomic_read(&cachep->allochit);
3771 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3772 unsigned long freehit = atomic_read(&cachep->freehit);
3773 unsigned long freemiss = atomic_read(&cachep->freemiss);
3775 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
3776 allochit, allocmiss, freehit, freemiss);
3784 * slabinfo_op - iterator that generates /proc/slabinfo
3793 * num-pages-per-slab
3794 * + further values on SMP and with statistics enabled
3797 struct seq_operations slabinfo_op = {
3804 #define MAX_SLABINFO_WRITE 128
3806 * slabinfo_write - Tuning for the slab allocator
3808 * @buffer: user buffer
3809 * @count: data length
3812 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
3813 size_t count, loff_t *ppos)
3815 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
3816 int limit, batchcount, shared, res;
3817 struct list_head *p;
3819 if (count > MAX_SLABINFO_WRITE)
3821 if (copy_from_user(&kbuf, buffer, count))
3823 kbuf[MAX_SLABINFO_WRITE] = '\0';
3825 tmp = strchr(kbuf, ' ');
3830 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3833 /* Find the cache in the chain of caches. */
3834 mutex_lock(&cache_chain_mutex);
3836 list_for_each(p, &cache_chain) {
3837 struct kmem_cache *cachep;
3839 cachep = list_entry(p, struct kmem_cache, next);
3840 if (!strcmp(cachep->name, kbuf)) {
3841 if (limit < 1 || batchcount < 1 ||
3842 batchcount > limit || shared < 0) {
3845 res = do_tune_cpucache(cachep, limit,
3846 batchcount, shared);
3851 mutex_unlock(&cache_chain_mutex);
3859 * ksize - get the actual amount of memory allocated for a given object
3860 * @objp: Pointer to the object
3862 * kmalloc may internally round up allocations and return more memory
3863 * than requested. ksize() can be used to determine the actual amount of
3864 * memory allocated. The caller may use this additional memory, even though
3865 * a smaller amount of memory was initially specified with the kmalloc call.
3866 * The caller must guarantee that objp points to a valid object previously
3867 * allocated with either kmalloc() or kmem_cache_alloc(). The object
3868 * must not be freed during the duration of the call.
3870 unsigned int ksize(const void *objp)
3872 if (unlikely(objp == NULL))
3875 return obj_size(virt_to_cache(objp));