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/cpuset.h>
98 #include <linux/seq_file.h>
99 #include <linux/notifier.h>
100 #include <linux/kallsyms.h>
101 #include <linux/cpu.h>
102 #include <linux/sysctl.h>
103 #include <linux/module.h>
104 #include <linux/rcupdate.h>
105 #include <linux/string.h>
106 #include <linux/nodemask.h>
107 #include <linux/mempolicy.h>
108 #include <linux/mutex.h>
110 #include <asm/uaccess.h>
111 #include <asm/cacheflush.h>
112 #include <asm/tlbflush.h>
113 #include <asm/page.h>
116 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
117 * SLAB_RED_ZONE & SLAB_POISON.
118 * 0 for faster, smaller code (especially in the critical paths).
120 * STATS - 1 to collect stats for /proc/slabinfo.
121 * 0 for faster, smaller code (especially in the critical paths).
123 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
126 #ifdef CONFIG_DEBUG_SLAB
129 #define FORCED_DEBUG 1
133 #define FORCED_DEBUG 0
136 /* Shouldn't this be in a header file somewhere? */
137 #define BYTES_PER_WORD sizeof(void *)
139 #ifndef cache_line_size
140 #define cache_line_size() L1_CACHE_BYTES
143 #ifndef ARCH_KMALLOC_MINALIGN
145 * Enforce a minimum alignment for the kmalloc caches.
146 * Usually, the kmalloc caches are cache_line_size() aligned, except when
147 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
148 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
149 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
150 * Note that this flag disables some debug features.
152 #define ARCH_KMALLOC_MINALIGN 0
155 #ifndef ARCH_SLAB_MINALIGN
157 * Enforce a minimum alignment for all caches.
158 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
159 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
160 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
161 * some debug features.
163 #define ARCH_SLAB_MINALIGN 0
166 #ifndef ARCH_KMALLOC_FLAGS
167 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
170 /* Legal flag mask for kmem_cache_create(). */
172 # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
173 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
175 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
176 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
177 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
179 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
180 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
181 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
182 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
188 * Bufctl's are used for linking objs within a slab
191 * This implementation relies on "struct page" for locating the cache &
192 * slab an object belongs to.
193 * This allows the bufctl structure to be small (one int), but limits
194 * the number of objects a slab (not a cache) can contain when off-slab
195 * bufctls are used. The limit is the size of the largest general cache
196 * that does not use off-slab slabs.
197 * For 32bit archs with 4 kB pages, is this 56.
198 * This is not serious, as it is only for large objects, when it is unwise
199 * to have too many per slab.
200 * Note: This limit can be raised by introducing a general cache whose size
201 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
204 typedef unsigned int kmem_bufctl_t;
205 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
206 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
207 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-2)
209 /* Max number of objs-per-slab for caches which use off-slab slabs.
210 * Needed to avoid a possible looping condition in cache_grow().
212 static unsigned long offslab_limit;
217 * Manages the objs in a slab. Placed either at the beginning of mem allocated
218 * for a slab, or allocated from an general cache.
219 * Slabs are chained into three list: fully used, partial, fully free slabs.
222 struct list_head list;
223 unsigned long colouroff;
224 void *s_mem; /* including colour offset */
225 unsigned int inuse; /* num of objs active in slab */
227 unsigned short nodeid;
233 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
234 * arrange for kmem_freepages to be called via RCU. This is useful if
235 * we need to approach a kernel structure obliquely, from its address
236 * obtained without the usual locking. We can lock the structure to
237 * stabilize it and check it's still at the given address, only if we
238 * can be sure that the memory has not been meanwhile reused for some
239 * other kind of object (which our subsystem's lock might corrupt).
241 * rcu_read_lock before reading the address, then rcu_read_unlock after
242 * taking the spinlock within the structure expected at that address.
244 * We assume struct slab_rcu can overlay struct slab when destroying.
247 struct rcu_head head;
248 struct kmem_cache *cachep;
256 * - LIFO ordering, to hand out cache-warm objects from _alloc
257 * - reduce the number of linked list operations
258 * - reduce spinlock operations
260 * The limit is stored in the per-cpu structure to reduce the data cache
267 unsigned int batchcount;
268 unsigned int touched;
271 * Must have this definition in here for the proper
272 * alignment of array_cache. Also simplifies accessing
274 * [0] is for gcc 2.95. It should really be [].
279 * bootstrap: The caches do not work without cpuarrays anymore, but the
280 * cpuarrays are allocated from the generic caches...
282 #define BOOT_CPUCACHE_ENTRIES 1
283 struct arraycache_init {
284 struct array_cache cache;
285 void *entries[BOOT_CPUCACHE_ENTRIES];
289 * The slab lists for all objects.
292 struct list_head slabs_partial; /* partial list first, better asm code */
293 struct list_head slabs_full;
294 struct list_head slabs_free;
295 unsigned long free_objects;
296 unsigned int free_limit;
297 unsigned int colour_next; /* Per-node cache coloring */
298 spinlock_t list_lock;
299 struct array_cache *shared; /* shared per node */
300 struct array_cache **alien; /* on other nodes */
301 unsigned long next_reap; /* updated without locking */
302 int free_touched; /* updated without locking */
306 * Need this for bootstrapping a per node allocator.
308 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
309 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
310 #define CACHE_CACHE 0
312 #define SIZE_L3 (1 + MAX_NUMNODES)
315 * This function must be completely optimized away if a constant is passed to
316 * it. Mostly the same as what is in linux/slab.h except it returns an index.
318 static __always_inline int index_of(const size_t size)
320 extern void __bad_size(void);
322 if (__builtin_constant_p(size)) {
330 #include "linux/kmalloc_sizes.h"
338 #define INDEX_AC index_of(sizeof(struct arraycache_init))
339 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
341 static void kmem_list3_init(struct kmem_list3 *parent)
343 INIT_LIST_HEAD(&parent->slabs_full);
344 INIT_LIST_HEAD(&parent->slabs_partial);
345 INIT_LIST_HEAD(&parent->slabs_free);
346 parent->shared = NULL;
347 parent->alien = NULL;
348 parent->colour_next = 0;
349 spin_lock_init(&parent->list_lock);
350 parent->free_objects = 0;
351 parent->free_touched = 0;
354 #define MAKE_LIST(cachep, listp, slab, nodeid) \
356 INIT_LIST_HEAD(listp); \
357 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
360 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
362 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
363 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
364 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
374 /* 1) per-cpu data, touched during every alloc/free */
375 struct array_cache *array[NR_CPUS];
376 /* 2) Cache tunables. Protected by cache_chain_mutex */
377 unsigned int batchcount;
381 unsigned int buffer_size;
382 /* 3) touched by every alloc & free from the backend */
383 struct kmem_list3 *nodelists[MAX_NUMNODES];
385 unsigned int flags; /* constant flags */
386 unsigned int num; /* # of objs per slab */
388 /* 4) cache_grow/shrink */
389 /* order of pgs per slab (2^n) */
390 unsigned int gfporder;
392 /* force GFP flags, e.g. GFP_DMA */
395 size_t colour; /* cache colouring range */
396 unsigned int colour_off; /* colour offset */
397 struct kmem_cache *slabp_cache;
398 unsigned int slab_size;
399 unsigned int dflags; /* dynamic flags */
401 /* constructor func */
402 void (*ctor) (void *, struct kmem_cache *, unsigned long);
404 /* de-constructor func */
405 void (*dtor) (void *, struct kmem_cache *, unsigned long);
407 /* 5) cache creation/removal */
409 struct list_head next;
413 unsigned long num_active;
414 unsigned long num_allocations;
415 unsigned long high_mark;
417 unsigned long reaped;
418 unsigned long errors;
419 unsigned long max_freeable;
420 unsigned long node_allocs;
421 unsigned long node_frees;
429 * If debugging is enabled, then the allocator can add additional
430 * fields and/or padding to every object. buffer_size contains the total
431 * object size including these internal fields, the following two
432 * variables contain the offset to the user object and its size.
439 #define CFLGS_OFF_SLAB (0x80000000UL)
440 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
442 #define BATCHREFILL_LIMIT 16
444 * Optimization question: fewer reaps means less probability for unnessary
445 * cpucache drain/refill cycles.
447 * OTOH the cpuarrays can contain lots of objects,
448 * which could lock up otherwise freeable slabs.
450 #define REAPTIMEOUT_CPUC (2*HZ)
451 #define REAPTIMEOUT_LIST3 (4*HZ)
454 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
455 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
456 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
457 #define STATS_INC_GROWN(x) ((x)->grown++)
458 #define STATS_INC_REAPED(x) ((x)->reaped++)
459 #define STATS_SET_HIGH(x) \
461 if ((x)->num_active > (x)->high_mark) \
462 (x)->high_mark = (x)->num_active; \
464 #define STATS_INC_ERR(x) ((x)->errors++)
465 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
466 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
467 #define STATS_SET_FREEABLE(x, i) \
469 if ((x)->max_freeable < i) \
470 (x)->max_freeable = i; \
472 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
473 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
474 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
475 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
477 #define STATS_INC_ACTIVE(x) do { } while (0)
478 #define STATS_DEC_ACTIVE(x) do { } while (0)
479 #define STATS_INC_ALLOCED(x) do { } while (0)
480 #define STATS_INC_GROWN(x) do { } while (0)
481 #define STATS_INC_REAPED(x) do { } while (0)
482 #define STATS_SET_HIGH(x) do { } while (0)
483 #define STATS_INC_ERR(x) do { } while (0)
484 #define STATS_INC_NODEALLOCS(x) do { } while (0)
485 #define STATS_INC_NODEFREES(x) do { } while (0)
486 #define STATS_SET_FREEABLE(x, i) do { } while (0)
487 #define STATS_INC_ALLOCHIT(x) do { } while (0)
488 #define STATS_INC_ALLOCMISS(x) do { } while (0)
489 #define STATS_INC_FREEHIT(x) do { } while (0)
490 #define STATS_INC_FREEMISS(x) do { } while (0)
495 * Magic nums for obj red zoning.
496 * Placed in the first word before and the first word after an obj.
498 #define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
499 #define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
501 /* ...and for poisoning */
502 #define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
503 #define POISON_FREE 0x6b /* for use-after-free poisoning */
504 #define POISON_END 0xa5 /* end-byte of poisoning */
507 * memory layout of objects:
509 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
510 * the end of an object is aligned with the end of the real
511 * allocation. Catches writes behind the end of the allocation.
512 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
514 * cachep->obj_offset: The real object.
515 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
516 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
517 * [BYTES_PER_WORD long]
519 static int obj_offset(struct kmem_cache *cachep)
521 return cachep->obj_offset;
524 static int obj_size(struct kmem_cache *cachep)
526 return cachep->obj_size;
529 static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
531 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
532 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD);
535 static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
537 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
538 if (cachep->flags & SLAB_STORE_USER)
539 return (unsigned long *)(objp + cachep->buffer_size -
541 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD);
544 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
546 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
547 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
552 #define obj_offset(x) 0
553 #define obj_size(cachep) (cachep->buffer_size)
554 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
555 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
556 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
561 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
564 #if defined(CONFIG_LARGE_ALLOCS)
565 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
566 #define MAX_GFP_ORDER 13 /* up to 32Mb */
567 #elif defined(CONFIG_MMU)
568 #define MAX_OBJ_ORDER 5 /* 32 pages */
569 #define MAX_GFP_ORDER 5 /* 32 pages */
571 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
572 #define MAX_GFP_ORDER 8 /* up to 1Mb */
576 * Do not go above this order unless 0 objects fit into the slab.
578 #define BREAK_GFP_ORDER_HI 1
579 #define BREAK_GFP_ORDER_LO 0
580 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
583 * Functions for storing/retrieving the cachep and or slab from the page
584 * allocator. These are used to find the slab an obj belongs to. With kfree(),
585 * these are used to find the cache which an obj belongs to.
587 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
589 page->lru.next = (struct list_head *)cache;
592 static inline struct kmem_cache *page_get_cache(struct page *page)
594 if (unlikely(PageCompound(page)))
595 page = (struct page *)page_private(page);
596 return (struct kmem_cache *)page->lru.next;
599 static inline void page_set_slab(struct page *page, struct slab *slab)
601 page->lru.prev = (struct list_head *)slab;
604 static inline struct slab *page_get_slab(struct page *page)
606 if (unlikely(PageCompound(page)))
607 page = (struct page *)page_private(page);
608 return (struct slab *)page->lru.prev;
611 static inline struct kmem_cache *virt_to_cache(const void *obj)
613 struct page *page = virt_to_page(obj);
614 return page_get_cache(page);
617 static inline struct slab *virt_to_slab(const void *obj)
619 struct page *page = virt_to_page(obj);
620 return page_get_slab(page);
623 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
626 return slab->s_mem + cache->buffer_size * idx;
629 static inline unsigned int obj_to_index(struct kmem_cache *cache,
630 struct slab *slab, void *obj)
632 return (unsigned)(obj - slab->s_mem) / cache->buffer_size;
636 * These are the default caches for kmalloc. Custom caches can have other sizes.
638 struct cache_sizes malloc_sizes[] = {
639 #define CACHE(x) { .cs_size = (x) },
640 #include <linux/kmalloc_sizes.h>
644 EXPORT_SYMBOL(malloc_sizes);
646 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
652 static struct cache_names __initdata cache_names[] = {
653 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
654 #include <linux/kmalloc_sizes.h>
659 static struct arraycache_init initarray_cache __initdata =
660 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
661 static struct arraycache_init initarray_generic =
662 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
664 /* internal cache of cache description objs */
665 static struct kmem_cache cache_cache = {
667 .limit = BOOT_CPUCACHE_ENTRIES,
669 .buffer_size = sizeof(struct kmem_cache),
670 .name = "kmem_cache",
672 .obj_size = sizeof(struct kmem_cache),
676 /* Guard access to the cache-chain. */
677 static DEFINE_MUTEX(cache_chain_mutex);
678 static struct list_head cache_chain;
681 * vm_enough_memory() looks at this to determine how many slab-allocated pages
682 * are possibly freeable under pressure
684 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
686 atomic_t slab_reclaim_pages;
689 * chicken and egg problem: delay the per-cpu array allocation
690 * until the general caches are up.
699 static DEFINE_PER_CPU(struct work_struct, reap_work);
701 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
703 static void enable_cpucache(struct kmem_cache *cachep);
704 static void cache_reap(void *unused);
705 static int __node_shrink(struct kmem_cache *cachep, int node);
707 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
709 return cachep->array[smp_processor_id()];
712 static inline struct kmem_cache *__find_general_cachep(size_t size,
715 struct cache_sizes *csizep = malloc_sizes;
718 /* This happens if someone tries to call
719 * kmem_cache_create(), or __kmalloc(), before
720 * the generic caches are initialized.
722 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
724 while (size > csizep->cs_size)
728 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
729 * has cs_{dma,}cachep==NULL. Thus no special case
730 * for large kmalloc calls required.
732 if (unlikely(gfpflags & GFP_DMA))
733 return csizep->cs_dmacachep;
734 return csizep->cs_cachep;
737 struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
739 return __find_general_cachep(size, gfpflags);
741 EXPORT_SYMBOL(kmem_find_general_cachep);
743 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
745 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
749 * Calculate the number of objects and left-over bytes for a given buffer size.
751 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
752 size_t align, int flags, size_t *left_over,
757 size_t slab_size = PAGE_SIZE << gfporder;
760 * The slab management structure can be either off the slab or
761 * on it. For the latter case, the memory allocated for a
765 * - One kmem_bufctl_t for each object
766 * - Padding to respect alignment of @align
767 * - @buffer_size bytes for each object
769 * If the slab management structure is off the slab, then the
770 * alignment will already be calculated into the size. Because
771 * the slabs are all pages aligned, the objects will be at the
772 * correct alignment when allocated.
774 if (flags & CFLGS_OFF_SLAB) {
776 nr_objs = slab_size / buffer_size;
778 if (nr_objs > SLAB_LIMIT)
779 nr_objs = SLAB_LIMIT;
782 * Ignore padding for the initial guess. The padding
783 * is at most @align-1 bytes, and @buffer_size is at
784 * least @align. In the worst case, this result will
785 * be one greater than the number of objects that fit
786 * into the memory allocation when taking the padding
789 nr_objs = (slab_size - sizeof(struct slab)) /
790 (buffer_size + sizeof(kmem_bufctl_t));
793 * This calculated number will be either the right
794 * amount, or one greater than what we want.
796 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
800 if (nr_objs > SLAB_LIMIT)
801 nr_objs = SLAB_LIMIT;
803 mgmt_size = slab_mgmt_size(nr_objs, align);
806 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
809 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
811 static void __slab_error(const char *function, struct kmem_cache *cachep,
814 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
815 function, cachep->name, msg);
821 * Special reaping functions for NUMA systems called from cache_reap().
822 * These take care of doing round robin flushing of alien caches (containing
823 * objects freed on different nodes from which they were allocated) and the
824 * flushing of remote pcps by calling drain_node_pages.
826 static DEFINE_PER_CPU(unsigned long, reap_node);
828 static void init_reap_node(int cpu)
832 node = next_node(cpu_to_node(cpu), node_online_map);
833 if (node == MAX_NUMNODES)
834 node = first_node(node_online_map);
836 __get_cpu_var(reap_node) = node;
839 static void next_reap_node(void)
841 int node = __get_cpu_var(reap_node);
844 * Also drain per cpu pages on remote zones
846 if (node != numa_node_id())
847 drain_node_pages(node);
849 node = next_node(node, node_online_map);
850 if (unlikely(node >= MAX_NUMNODES))
851 node = first_node(node_online_map);
852 __get_cpu_var(reap_node) = node;
856 #define init_reap_node(cpu) do { } while (0)
857 #define next_reap_node(void) do { } while (0)
861 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
862 * via the workqueue/eventd.
863 * Add the CPU number into the expiration time to minimize the possibility of
864 * the CPUs getting into lockstep and contending for the global cache chain
867 static void __devinit start_cpu_timer(int cpu)
869 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
872 * When this gets called from do_initcalls via cpucache_init(),
873 * init_workqueues() has already run, so keventd will be setup
876 if (keventd_up() && reap_work->func == NULL) {
878 INIT_WORK(reap_work, cache_reap, NULL);
879 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
883 static struct array_cache *alloc_arraycache(int node, int entries,
886 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
887 struct array_cache *nc = NULL;
889 nc = kmalloc_node(memsize, GFP_KERNEL, node);
893 nc->batchcount = batchcount;
895 spin_lock_init(&nc->lock);
901 static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int);
903 static struct array_cache **alloc_alien_cache(int node, int limit)
905 struct array_cache **ac_ptr;
906 int memsize = sizeof(void *) * MAX_NUMNODES;
911 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
914 if (i == node || !node_online(i)) {
918 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
920 for (i--; i <= 0; i--)
930 static void free_alien_cache(struct array_cache **ac_ptr)
941 static void __drain_alien_cache(struct kmem_cache *cachep,
942 struct array_cache *ac, int node)
944 struct kmem_list3 *rl3 = cachep->nodelists[node];
947 spin_lock(&rl3->list_lock);
948 free_block(cachep, ac->entry, ac->avail, node);
950 spin_unlock(&rl3->list_lock);
955 * Called from cache_reap() to regularly drain alien caches round robin.
957 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
959 int node = __get_cpu_var(reap_node);
962 struct array_cache *ac = l3->alien[node];
963 if (ac && ac->avail) {
964 spin_lock_irq(&ac->lock);
965 __drain_alien_cache(cachep, ac, node);
966 spin_unlock_irq(&ac->lock);
971 static void drain_alien_cache(struct kmem_cache *cachep,
972 struct array_cache **alien)
975 struct array_cache *ac;
978 for_each_online_node(i) {
981 spin_lock_irqsave(&ac->lock, flags);
982 __drain_alien_cache(cachep, ac, i);
983 spin_unlock_irqrestore(&ac->lock, flags);
989 #define drain_alien_cache(cachep, alien) do { } while (0)
990 #define reap_alien(cachep, l3) do { } while (0)
992 static inline struct array_cache **alloc_alien_cache(int node, int limit)
994 return (struct array_cache **) 0x01020304ul;
997 static inline void free_alien_cache(struct array_cache **ac_ptr)
1003 static int __devinit cpuup_callback(struct notifier_block *nfb,
1004 unsigned long action, void *hcpu)
1006 long cpu = (long)hcpu;
1007 struct kmem_cache *cachep;
1008 struct kmem_list3 *l3 = NULL;
1009 int node = cpu_to_node(cpu);
1010 int memsize = sizeof(struct kmem_list3);
1013 case CPU_UP_PREPARE:
1014 mutex_lock(&cache_chain_mutex);
1016 * We need to do this right in the beginning since
1017 * alloc_arraycache's are going to use this list.
1018 * kmalloc_node allows us to add the slab to the right
1019 * kmem_list3 and not this cpu's kmem_list3
1022 list_for_each_entry(cachep, &cache_chain, next) {
1024 * Set up the size64 kmemlist for cpu before we can
1025 * begin anything. Make sure some other cpu on this
1026 * node has not already allocated this
1028 if (!cachep->nodelists[node]) {
1029 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1032 kmem_list3_init(l3);
1033 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1034 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1037 * The l3s don't come and go as CPUs come and
1038 * go. cache_chain_mutex is sufficient
1041 cachep->nodelists[node] = l3;
1044 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1045 cachep->nodelists[node]->free_limit =
1046 (1 + nr_cpus_node(node)) *
1047 cachep->batchcount + cachep->num;
1048 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1052 * Now we can go ahead with allocating the shared arrays and
1055 list_for_each_entry(cachep, &cache_chain, next) {
1056 struct array_cache *nc;
1057 struct array_cache *shared;
1058 struct array_cache **alien;
1060 nc = alloc_arraycache(node, cachep->limit,
1061 cachep->batchcount);
1064 shared = alloc_arraycache(node,
1065 cachep->shared * cachep->batchcount,
1070 alien = alloc_alien_cache(node, cachep->limit);
1073 cachep->array[cpu] = nc;
1074 l3 = cachep->nodelists[node];
1077 spin_lock_irq(&l3->list_lock);
1080 * We are serialised from CPU_DEAD or
1081 * CPU_UP_CANCELLED by the cpucontrol lock
1083 l3->shared = shared;
1092 spin_unlock_irq(&l3->list_lock);
1094 free_alien_cache(alien);
1096 mutex_unlock(&cache_chain_mutex);
1099 start_cpu_timer(cpu);
1101 #ifdef CONFIG_HOTPLUG_CPU
1104 * Even if all the cpus of a node are down, we don't free the
1105 * kmem_list3 of any cache. This to avoid a race between
1106 * cpu_down, and a kmalloc allocation from another cpu for
1107 * memory from the node of the cpu going down. The list3
1108 * structure is usually allocated from kmem_cache_create() and
1109 * gets destroyed at kmem_cache_destroy().
1112 case CPU_UP_CANCELED:
1113 mutex_lock(&cache_chain_mutex);
1114 list_for_each_entry(cachep, &cache_chain, next) {
1115 struct array_cache *nc;
1116 struct array_cache *shared;
1117 struct array_cache **alien;
1120 mask = node_to_cpumask(node);
1121 /* cpu is dead; no one can alloc from it. */
1122 nc = cachep->array[cpu];
1123 cachep->array[cpu] = NULL;
1124 l3 = cachep->nodelists[node];
1127 goto free_array_cache;
1129 spin_lock_irq(&l3->list_lock);
1131 /* Free limit for this kmem_list3 */
1132 l3->free_limit -= cachep->batchcount;
1134 free_block(cachep, nc->entry, nc->avail, node);
1136 if (!cpus_empty(mask)) {
1137 spin_unlock_irq(&l3->list_lock);
1138 goto free_array_cache;
1141 shared = l3->shared;
1143 free_block(cachep, l3->shared->entry,
1144 l3->shared->avail, node);
1151 spin_unlock_irq(&l3->list_lock);
1155 drain_alien_cache(cachep, alien);
1156 free_alien_cache(alien);
1162 * In the previous loop, all the objects were freed to
1163 * the respective cache's slabs, now we can go ahead and
1164 * shrink each nodelist to its limit.
1166 list_for_each_entry(cachep, &cache_chain, next) {
1167 l3 = cachep->nodelists[node];
1170 spin_lock_irq(&l3->list_lock);
1171 /* free slabs belonging to this node */
1172 __node_shrink(cachep, node);
1173 spin_unlock_irq(&l3->list_lock);
1175 mutex_unlock(&cache_chain_mutex);
1181 mutex_unlock(&cache_chain_mutex);
1185 static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
1188 * swap the static kmem_list3 with kmalloced memory
1190 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1193 struct kmem_list3 *ptr;
1195 BUG_ON(cachep->nodelists[nodeid] != list);
1196 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1199 local_irq_disable();
1200 memcpy(ptr, list, sizeof(struct kmem_list3));
1201 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1202 cachep->nodelists[nodeid] = ptr;
1207 * Initialisation. Called after the page allocator have been initialised and
1208 * before smp_init().
1210 void __init kmem_cache_init(void)
1213 struct cache_sizes *sizes;
1214 struct cache_names *names;
1218 for (i = 0; i < NUM_INIT_LISTS; i++) {
1219 kmem_list3_init(&initkmem_list3[i]);
1220 if (i < MAX_NUMNODES)
1221 cache_cache.nodelists[i] = NULL;
1225 * Fragmentation resistance on low memory - only use bigger
1226 * page orders on machines with more than 32MB of memory.
1228 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1229 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1231 /* Bootstrap is tricky, because several objects are allocated
1232 * from caches that do not exist yet:
1233 * 1) initialize the cache_cache cache: it contains the struct
1234 * kmem_cache structures of all caches, except cache_cache itself:
1235 * cache_cache is statically allocated.
1236 * Initially an __init data area is used for the head array and the
1237 * kmem_list3 structures, it's replaced with a kmalloc allocated
1238 * array at the end of the bootstrap.
1239 * 2) Create the first kmalloc cache.
1240 * The struct kmem_cache for the new cache is allocated normally.
1241 * An __init data area is used for the head array.
1242 * 3) Create the remaining kmalloc caches, with minimally sized
1244 * 4) Replace the __init data head arrays for cache_cache and the first
1245 * kmalloc cache with kmalloc allocated arrays.
1246 * 5) Replace the __init data for kmem_list3 for cache_cache and
1247 * the other cache's with kmalloc allocated memory.
1248 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1251 /* 1) create the cache_cache */
1252 INIT_LIST_HEAD(&cache_chain);
1253 list_add(&cache_cache.next, &cache_chain);
1254 cache_cache.colour_off = cache_line_size();
1255 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1256 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
1258 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1261 for (order = 0; order < MAX_ORDER; order++) {
1262 cache_estimate(order, cache_cache.buffer_size,
1263 cache_line_size(), 0, &left_over, &cache_cache.num);
1264 if (cache_cache.num)
1267 if (!cache_cache.num)
1269 cache_cache.gfporder = order;
1270 cache_cache.colour = left_over / cache_cache.colour_off;
1271 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1272 sizeof(struct slab), cache_line_size());
1274 /* 2+3) create the kmalloc caches */
1275 sizes = malloc_sizes;
1276 names = cache_names;
1279 * Initialize the caches that provide memory for the array cache and the
1280 * kmem_list3 structures first. Without this, further allocations will
1284 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1285 sizes[INDEX_AC].cs_size,
1286 ARCH_KMALLOC_MINALIGN,
1287 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1290 if (INDEX_AC != INDEX_L3) {
1291 sizes[INDEX_L3].cs_cachep =
1292 kmem_cache_create(names[INDEX_L3].name,
1293 sizes[INDEX_L3].cs_size,
1294 ARCH_KMALLOC_MINALIGN,
1295 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1299 while (sizes->cs_size != ULONG_MAX) {
1301 * For performance, all the general caches are L1 aligned.
1302 * This should be particularly beneficial on SMP boxes, as it
1303 * eliminates "false sharing".
1304 * Note for systems short on memory removing the alignment will
1305 * allow tighter packing of the smaller caches.
1307 if (!sizes->cs_cachep) {
1308 sizes->cs_cachep = kmem_cache_create(names->name,
1310 ARCH_KMALLOC_MINALIGN,
1311 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1315 /* Inc off-slab bufctl limit until the ceiling is hit. */
1316 if (!(OFF_SLAB(sizes->cs_cachep))) {
1317 offslab_limit = sizes->cs_size - sizeof(struct slab);
1318 offslab_limit /= sizeof(kmem_bufctl_t);
1321 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
1323 ARCH_KMALLOC_MINALIGN,
1324 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1330 /* 4) Replace the bootstrap head arrays */
1334 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1336 local_irq_disable();
1337 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1338 memcpy(ptr, cpu_cache_get(&cache_cache),
1339 sizeof(struct arraycache_init));
1340 cache_cache.array[smp_processor_id()] = ptr;
1343 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1345 local_irq_disable();
1346 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1347 != &initarray_generic.cache);
1348 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1349 sizeof(struct arraycache_init));
1350 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1354 /* 5) Replace the bootstrap kmem_list3's */
1357 /* Replace the static kmem_list3 structures for the boot cpu */
1358 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
1361 for_each_online_node(node) {
1362 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1363 &initkmem_list3[SIZE_AC + node], node);
1365 if (INDEX_AC != INDEX_L3) {
1366 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1367 &initkmem_list3[SIZE_L3 + node],
1373 /* 6) resize the head arrays to their final sizes */
1375 struct kmem_cache *cachep;
1376 mutex_lock(&cache_chain_mutex);
1377 list_for_each_entry(cachep, &cache_chain, next)
1378 enable_cpucache(cachep);
1379 mutex_unlock(&cache_chain_mutex);
1383 g_cpucache_up = FULL;
1386 * Register a cpu startup notifier callback that initializes
1387 * cpu_cache_get for all new cpus
1389 register_cpu_notifier(&cpucache_notifier);
1392 * The reap timers are started later, with a module init call: That part
1393 * of the kernel is not yet operational.
1397 static int __init cpucache_init(void)
1402 * Register the timers that return unneeded pages to the page allocator
1404 for_each_online_cpu(cpu)
1405 start_cpu_timer(cpu);
1408 __initcall(cpucache_init);
1411 * Interface to system's page allocator. No need to hold the cache-lock.
1413 * If we requested dmaable memory, we will get it. Even if we
1414 * did not request dmaable memory, we might get it, but that
1415 * would be relatively rare and ignorable.
1417 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1423 flags |= cachep->gfpflags;
1424 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1427 addr = page_address(page);
1429 i = (1 << cachep->gfporder);
1430 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1431 atomic_add(i, &slab_reclaim_pages);
1432 add_page_state(nr_slab, i);
1434 __SetPageSlab(page);
1441 * Interface to system's page release.
1443 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1445 unsigned long i = (1 << cachep->gfporder);
1446 struct page *page = virt_to_page(addr);
1447 const unsigned long nr_freed = i;
1450 BUG_ON(!PageSlab(page));
1451 __ClearPageSlab(page);
1454 sub_page_state(nr_slab, nr_freed);
1455 if (current->reclaim_state)
1456 current->reclaim_state->reclaimed_slab += nr_freed;
1457 free_pages((unsigned long)addr, cachep->gfporder);
1458 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1459 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages);
1462 static void kmem_rcu_free(struct rcu_head *head)
1464 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1465 struct kmem_cache *cachep = slab_rcu->cachep;
1467 kmem_freepages(cachep, slab_rcu->addr);
1468 if (OFF_SLAB(cachep))
1469 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1474 #ifdef CONFIG_DEBUG_PAGEALLOC
1475 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1476 unsigned long caller)
1478 int size = obj_size(cachep);
1480 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1482 if (size < 5 * sizeof(unsigned long))
1485 *addr++ = 0x12345678;
1487 *addr++ = smp_processor_id();
1488 size -= 3 * sizeof(unsigned long);
1490 unsigned long *sptr = &caller;
1491 unsigned long svalue;
1493 while (!kstack_end(sptr)) {
1495 if (kernel_text_address(svalue)) {
1497 size -= sizeof(unsigned long);
1498 if (size <= sizeof(unsigned long))
1504 *addr++ = 0x87654321;
1508 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1510 int size = obj_size(cachep);
1511 addr = &((char *)addr)[obj_offset(cachep)];
1513 memset(addr, val, size);
1514 *(unsigned char *)(addr + size - 1) = POISON_END;
1517 static void dump_line(char *data, int offset, int limit)
1520 printk(KERN_ERR "%03x:", offset);
1521 for (i = 0; i < limit; i++)
1522 printk(" %02x", (unsigned char)data[offset + i]);
1529 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1534 if (cachep->flags & SLAB_RED_ZONE) {
1535 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
1536 *dbg_redzone1(cachep, objp),
1537 *dbg_redzone2(cachep, objp));
1540 if (cachep->flags & SLAB_STORE_USER) {
1541 printk(KERN_ERR "Last user: [<%p>]",
1542 *dbg_userword(cachep, objp));
1543 print_symbol("(%s)",
1544 (unsigned long)*dbg_userword(cachep, objp));
1547 realobj = (char *)objp + obj_offset(cachep);
1548 size = obj_size(cachep);
1549 for (i = 0; i < size && lines; i += 16, lines--) {
1552 if (i + limit > size)
1554 dump_line(realobj, i, limit);
1558 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1564 realobj = (char *)objp + obj_offset(cachep);
1565 size = obj_size(cachep);
1567 for (i = 0; i < size; i++) {
1568 char exp = POISON_FREE;
1571 if (realobj[i] != exp) {
1577 "Slab corruption: start=%p, len=%d\n",
1579 print_objinfo(cachep, objp, 0);
1581 /* Hexdump the affected line */
1584 if (i + limit > size)
1586 dump_line(realobj, i, limit);
1589 /* Limit to 5 lines */
1595 /* Print some data about the neighboring objects, if they
1598 struct slab *slabp = virt_to_slab(objp);
1601 objnr = obj_to_index(cachep, slabp, objp);
1603 objp = index_to_obj(cachep, slabp, objnr - 1);
1604 realobj = (char *)objp + obj_offset(cachep);
1605 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1607 print_objinfo(cachep, objp, 2);
1609 if (objnr + 1 < cachep->num) {
1610 objp = index_to_obj(cachep, slabp, objnr + 1);
1611 realobj = (char *)objp + obj_offset(cachep);
1612 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1614 print_objinfo(cachep, objp, 2);
1622 * slab_destroy_objs - destroy a slab and its objects
1623 * @cachep: cache pointer being destroyed
1624 * @slabp: slab pointer being destroyed
1626 * Call the registered destructor for each object in a slab that is being
1629 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1632 for (i = 0; i < cachep->num; i++) {
1633 void *objp = index_to_obj(cachep, slabp, i);
1635 if (cachep->flags & SLAB_POISON) {
1636 #ifdef CONFIG_DEBUG_PAGEALLOC
1637 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1639 kernel_map_pages(virt_to_page(objp),
1640 cachep->buffer_size / PAGE_SIZE, 1);
1642 check_poison_obj(cachep, objp);
1644 check_poison_obj(cachep, objp);
1647 if (cachep->flags & SLAB_RED_ZONE) {
1648 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1649 slab_error(cachep, "start of a freed object "
1651 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1652 slab_error(cachep, "end of a freed object "
1655 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1656 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
1660 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1664 for (i = 0; i < cachep->num; i++) {
1665 void *objp = index_to_obj(cachep, slabp, i);
1666 (cachep->dtor) (objp, cachep, 0);
1673 * slab_destroy - destroy and release all objects in a slab
1674 * @cachep: cache pointer being destroyed
1675 * @slabp: slab pointer being destroyed
1677 * Destroy all the objs in a slab, and release the mem back to the system.
1678 * Before calling the slab must have been unlinked from the cache. The
1679 * cache-lock is not held/needed.
1681 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1683 void *addr = slabp->s_mem - slabp->colouroff;
1685 slab_destroy_objs(cachep, slabp);
1686 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1687 struct slab_rcu *slab_rcu;
1689 slab_rcu = (struct slab_rcu *)slabp;
1690 slab_rcu->cachep = cachep;
1691 slab_rcu->addr = addr;
1692 call_rcu(&slab_rcu->head, kmem_rcu_free);
1694 kmem_freepages(cachep, addr);
1695 if (OFF_SLAB(cachep))
1696 kmem_cache_free(cachep->slabp_cache, slabp);
1701 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1702 * size of kmem_list3.
1704 static void set_up_list3s(struct kmem_cache *cachep, int index)
1708 for_each_online_node(node) {
1709 cachep->nodelists[node] = &initkmem_list3[index + node];
1710 cachep->nodelists[node]->next_reap = jiffies +
1712 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1717 * calculate_slab_order - calculate size (page order) of slabs
1718 * @cachep: pointer to the cache that is being created
1719 * @size: size of objects to be created in this cache.
1720 * @align: required alignment for the objects.
1721 * @flags: slab allocation flags
1723 * Also calculates the number of objects per slab.
1725 * This could be made much more intelligent. For now, try to avoid using
1726 * high order pages for slabs. When the gfp() functions are more friendly
1727 * towards high-order requests, this should be changed.
1729 static size_t calculate_slab_order(struct kmem_cache *cachep,
1730 size_t size, size_t align, unsigned long flags)
1732 size_t left_over = 0;
1735 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
1739 cache_estimate(gfporder, size, align, flags, &remainder, &num);
1743 /* More than offslab_limit objects will cause problems */
1744 if ((flags & CFLGS_OFF_SLAB) && num > offslab_limit)
1747 /* Found something acceptable - save it away */
1749 cachep->gfporder = gfporder;
1750 left_over = remainder;
1753 * A VFS-reclaimable slab tends to have most allocations
1754 * as GFP_NOFS and we really don't want to have to be allocating
1755 * higher-order pages when we are unable to shrink dcache.
1757 if (flags & SLAB_RECLAIM_ACCOUNT)
1761 * Large number of objects is good, but very large slabs are
1762 * currently bad for the gfp()s.
1764 if (gfporder >= slab_break_gfp_order)
1768 * Acceptable internal fragmentation?
1770 if (left_over * 8 <= (PAGE_SIZE << gfporder))
1776 static void setup_cpu_cache(struct kmem_cache *cachep)
1778 if (g_cpucache_up == FULL) {
1779 enable_cpucache(cachep);
1782 if (g_cpucache_up == NONE) {
1784 * Note: the first kmem_cache_create must create the cache
1785 * that's used by kmalloc(24), otherwise the creation of
1786 * further caches will BUG().
1788 cachep->array[smp_processor_id()] = &initarray_generic.cache;
1791 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
1792 * the first cache, then we need to set up all its list3s,
1793 * otherwise the creation of further caches will BUG().
1795 set_up_list3s(cachep, SIZE_AC);
1796 if (INDEX_AC == INDEX_L3)
1797 g_cpucache_up = PARTIAL_L3;
1799 g_cpucache_up = PARTIAL_AC;
1801 cachep->array[smp_processor_id()] =
1802 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1804 if (g_cpucache_up == PARTIAL_AC) {
1805 set_up_list3s(cachep, SIZE_L3);
1806 g_cpucache_up = PARTIAL_L3;
1809 for_each_online_node(node) {
1810 cachep->nodelists[node] =
1811 kmalloc_node(sizeof(struct kmem_list3),
1813 BUG_ON(!cachep->nodelists[node]);
1814 kmem_list3_init(cachep->nodelists[node]);
1818 cachep->nodelists[numa_node_id()]->next_reap =
1819 jiffies + REAPTIMEOUT_LIST3 +
1820 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1822 cpu_cache_get(cachep)->avail = 0;
1823 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1824 cpu_cache_get(cachep)->batchcount = 1;
1825 cpu_cache_get(cachep)->touched = 0;
1826 cachep->batchcount = 1;
1827 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1831 * kmem_cache_create - Create a cache.
1832 * @name: A string which is used in /proc/slabinfo to identify this cache.
1833 * @size: The size of objects to be created in this cache.
1834 * @align: The required alignment for the objects.
1835 * @flags: SLAB flags
1836 * @ctor: A constructor for the objects.
1837 * @dtor: A destructor for the objects.
1839 * Returns a ptr to the cache on success, NULL on failure.
1840 * Cannot be called within a int, but can be interrupted.
1841 * The @ctor is run when new pages are allocated by the cache
1842 * and the @dtor is run before the pages are handed back.
1844 * @name must be valid until the cache is destroyed. This implies that
1845 * the module calling this has to destroy the cache before getting unloaded.
1849 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1850 * to catch references to uninitialised memory.
1852 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1853 * for buffer overruns.
1855 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1856 * cacheline. This can be beneficial if you're counting cycles as closely
1860 kmem_cache_create (const char *name, size_t size, size_t align,
1861 unsigned long flags,
1862 void (*ctor)(void*, struct kmem_cache *, unsigned long),
1863 void (*dtor)(void*, struct kmem_cache *, unsigned long))
1865 size_t left_over, slab_size, ralign;
1866 struct kmem_cache *cachep = NULL;
1867 struct list_head *p;
1870 * Sanity checks... these are all serious usage bugs.
1872 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
1873 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
1874 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
1880 * Prevent CPUs from coming and going.
1881 * lock_cpu_hotplug() nests outside cache_chain_mutex
1885 mutex_lock(&cache_chain_mutex);
1887 list_for_each(p, &cache_chain) {
1888 struct kmem_cache *pc = list_entry(p, struct kmem_cache, next);
1889 mm_segment_t old_fs = get_fs();
1894 * This happens when the module gets unloaded and doesn't
1895 * destroy its slab cache and no-one else reuses the vmalloc
1896 * area of the module. Print a warning.
1899 res = __get_user(tmp, pc->name);
1902 printk("SLAB: cache with size %d has lost its name\n",
1907 if (!strcmp(pc->name, name)) {
1908 printk("kmem_cache_create: duplicate cache %s\n", name);
1915 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1916 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
1917 /* No constructor, but inital state check requested */
1918 printk(KERN_ERR "%s: No con, but init state check "
1919 "requested - %s\n", __FUNCTION__, name);
1920 flags &= ~SLAB_DEBUG_INITIAL;
1924 * Enable redzoning and last user accounting, except for caches with
1925 * large objects, if the increased size would increase the object size
1926 * above the next power of two: caches with object sizes just above a
1927 * power of two have a significant amount of internal fragmentation.
1929 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
1930 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
1931 if (!(flags & SLAB_DESTROY_BY_RCU))
1932 flags |= SLAB_POISON;
1934 if (flags & SLAB_DESTROY_BY_RCU)
1935 BUG_ON(flags & SLAB_POISON);
1937 if (flags & SLAB_DESTROY_BY_RCU)
1941 * Always checks flags, a caller might be expecting debug support which
1944 if (flags & ~CREATE_MASK)
1948 * Check that size is in terms of words. This is needed to avoid
1949 * unaligned accesses for some archs when redzoning is used, and makes
1950 * sure any on-slab bufctl's are also correctly aligned.
1952 if (size & (BYTES_PER_WORD - 1)) {
1953 size += (BYTES_PER_WORD - 1);
1954 size &= ~(BYTES_PER_WORD - 1);
1957 /* calculate the final buffer alignment: */
1959 /* 1) arch recommendation: can be overridden for debug */
1960 if (flags & SLAB_HWCACHE_ALIGN) {
1962 * Default alignment: as specified by the arch code. Except if
1963 * an object is really small, then squeeze multiple objects into
1966 ralign = cache_line_size();
1967 while (size <= ralign / 2)
1970 ralign = BYTES_PER_WORD;
1972 /* 2) arch mandated alignment: disables debug if necessary */
1973 if (ralign < ARCH_SLAB_MINALIGN) {
1974 ralign = ARCH_SLAB_MINALIGN;
1975 if (ralign > BYTES_PER_WORD)
1976 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
1978 /* 3) caller mandated alignment: disables debug if necessary */
1979 if (ralign < align) {
1981 if (ralign > BYTES_PER_WORD)
1982 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
1985 * 4) Store it. Note that the debug code below can reduce
1986 * the alignment to BYTES_PER_WORD.
1990 /* Get cache's description obj. */
1991 cachep = kmem_cache_alloc(&cache_cache, SLAB_KERNEL);
1994 memset(cachep, 0, sizeof(struct kmem_cache));
1997 cachep->obj_size = size;
1999 if (flags & SLAB_RED_ZONE) {
2000 /* redzoning only works with word aligned caches */
2001 align = BYTES_PER_WORD;
2003 /* add space for red zone words */
2004 cachep->obj_offset += BYTES_PER_WORD;
2005 size += 2 * BYTES_PER_WORD;
2007 if (flags & SLAB_STORE_USER) {
2008 /* user store requires word alignment and
2009 * one word storage behind the end of the real
2012 align = BYTES_PER_WORD;
2013 size += BYTES_PER_WORD;
2015 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2016 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2017 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2018 cachep->obj_offset += PAGE_SIZE - size;
2024 /* Determine if the slab management is 'on' or 'off' slab. */
2025 if (size >= (PAGE_SIZE >> 3))
2027 * Size is large, assume best to place the slab management obj
2028 * off-slab (should allow better packing of objs).
2030 flags |= CFLGS_OFF_SLAB;
2032 size = ALIGN(size, align);
2034 left_over = calculate_slab_order(cachep, size, align, flags);
2037 printk("kmem_cache_create: couldn't create cache %s.\n", name);
2038 kmem_cache_free(&cache_cache, cachep);
2042 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2043 + sizeof(struct slab), align);
2046 * If the slab has been placed off-slab, and we have enough space then
2047 * move it on-slab. This is at the expense of any extra colouring.
2049 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2050 flags &= ~CFLGS_OFF_SLAB;
2051 left_over -= slab_size;
2054 if (flags & CFLGS_OFF_SLAB) {
2055 /* really off slab. No need for manual alignment */
2057 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2060 cachep->colour_off = cache_line_size();
2061 /* Offset must be a multiple of the alignment. */
2062 if (cachep->colour_off < align)
2063 cachep->colour_off = align;
2064 cachep->colour = left_over / cachep->colour_off;
2065 cachep->slab_size = slab_size;
2066 cachep->flags = flags;
2067 cachep->gfpflags = 0;
2068 if (flags & SLAB_CACHE_DMA)
2069 cachep->gfpflags |= GFP_DMA;
2070 cachep->buffer_size = size;
2072 if (flags & CFLGS_OFF_SLAB)
2073 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2074 cachep->ctor = ctor;
2075 cachep->dtor = dtor;
2076 cachep->name = name;
2079 setup_cpu_cache(cachep);
2081 /* cache setup completed, link it into the list */
2082 list_add(&cachep->next, &cache_chain);
2084 if (!cachep && (flags & SLAB_PANIC))
2085 panic("kmem_cache_create(): failed to create slab `%s'\n",
2087 mutex_unlock(&cache_chain_mutex);
2088 unlock_cpu_hotplug();
2091 EXPORT_SYMBOL(kmem_cache_create);
2094 static void check_irq_off(void)
2096 BUG_ON(!irqs_disabled());
2099 static void check_irq_on(void)
2101 BUG_ON(irqs_disabled());
2104 static void check_spinlock_acquired(struct kmem_cache *cachep)
2108 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2112 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2116 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2121 #define check_irq_off() do { } while(0)
2122 #define check_irq_on() do { } while(0)
2123 #define check_spinlock_acquired(x) do { } while(0)
2124 #define check_spinlock_acquired_node(x, y) do { } while(0)
2127 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2128 struct array_cache *ac,
2129 int force, int node);
2131 static void do_drain(void *arg)
2133 struct kmem_cache *cachep = arg;
2134 struct array_cache *ac;
2135 int node = numa_node_id();
2138 ac = cpu_cache_get(cachep);
2139 spin_lock(&cachep->nodelists[node]->list_lock);
2140 free_block(cachep, ac->entry, ac->avail, node);
2141 spin_unlock(&cachep->nodelists[node]->list_lock);
2145 static void drain_cpu_caches(struct kmem_cache *cachep)
2147 struct kmem_list3 *l3;
2150 on_each_cpu(do_drain, cachep, 1, 1);
2152 for_each_online_node(node) {
2153 l3 = cachep->nodelists[node];
2155 drain_array(cachep, l3, l3->shared, 1, node);
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);
2817 if (unlikely(cpuset_do_slab_mem_spread() &&
2818 (cachep->flags & SLAB_MEM_SPREAD) &&
2820 int nid = cpuset_mem_spread_node();
2822 if (nid != numa_node_id())
2823 return __cache_alloc_node(cachep, flags, nid);
2828 ac = cpu_cache_get(cachep);
2829 if (likely(ac->avail)) {
2830 STATS_INC_ALLOCHIT(cachep);
2832 objp = ac->entry[--ac->avail];
2834 STATS_INC_ALLOCMISS(cachep);
2835 objp = cache_alloc_refill(cachep, flags);
2840 static __always_inline void *__cache_alloc(struct kmem_cache *cachep,
2841 gfp_t flags, void *caller)
2843 unsigned long save_flags;
2846 cache_alloc_debugcheck_before(cachep, flags);
2848 local_irq_save(save_flags);
2849 objp = ____cache_alloc(cachep, flags);
2850 local_irq_restore(save_flags);
2851 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
2859 * A interface to enable slab creation on nodeid
2861 static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
2864 struct list_head *entry;
2866 struct kmem_list3 *l3;
2870 l3 = cachep->nodelists[nodeid];
2875 spin_lock(&l3->list_lock);
2876 entry = l3->slabs_partial.next;
2877 if (entry == &l3->slabs_partial) {
2878 l3->free_touched = 1;
2879 entry = l3->slabs_free.next;
2880 if (entry == &l3->slabs_free)
2884 slabp = list_entry(entry, struct slab, list);
2885 check_spinlock_acquired_node(cachep, nodeid);
2886 check_slabp(cachep, slabp);
2888 STATS_INC_NODEALLOCS(cachep);
2889 STATS_INC_ACTIVE(cachep);
2890 STATS_SET_HIGH(cachep);
2892 BUG_ON(slabp->inuse == cachep->num);
2894 obj = slab_get_obj(cachep, slabp, nodeid);
2895 check_slabp(cachep, slabp);
2897 /* move slabp to correct slabp list: */
2898 list_del(&slabp->list);
2900 if (slabp->free == BUFCTL_END)
2901 list_add(&slabp->list, &l3->slabs_full);
2903 list_add(&slabp->list, &l3->slabs_partial);
2905 spin_unlock(&l3->list_lock);
2909 spin_unlock(&l3->list_lock);
2910 x = cache_grow(cachep, flags, nodeid);
2922 * Caller needs to acquire correct kmem_list's list_lock
2924 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
2928 struct kmem_list3 *l3;
2930 for (i = 0; i < nr_objects; i++) {
2931 void *objp = objpp[i];
2934 slabp = virt_to_slab(objp);
2935 l3 = cachep->nodelists[node];
2936 list_del(&slabp->list);
2937 check_spinlock_acquired_node(cachep, node);
2938 check_slabp(cachep, slabp);
2939 slab_put_obj(cachep, slabp, objp, node);
2940 STATS_DEC_ACTIVE(cachep);
2942 check_slabp(cachep, slabp);
2944 /* fixup slab chains */
2945 if (slabp->inuse == 0) {
2946 if (l3->free_objects > l3->free_limit) {
2947 l3->free_objects -= cachep->num;
2948 slab_destroy(cachep, slabp);
2950 list_add(&slabp->list, &l3->slabs_free);
2953 /* Unconditionally move a slab to the end of the
2954 * partial list on free - maximum time for the
2955 * other objects to be freed, too.
2957 list_add_tail(&slabp->list, &l3->slabs_partial);
2962 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
2965 struct kmem_list3 *l3;
2966 int node = numa_node_id();
2968 batchcount = ac->batchcount;
2970 BUG_ON(!batchcount || batchcount > ac->avail);
2973 l3 = cachep->nodelists[node];
2974 spin_lock(&l3->list_lock);
2976 struct array_cache *shared_array = l3->shared;
2977 int max = shared_array->limit - shared_array->avail;
2979 if (batchcount > max)
2981 memcpy(&(shared_array->entry[shared_array->avail]),
2982 ac->entry, sizeof(void *) * batchcount);
2983 shared_array->avail += batchcount;
2988 free_block(cachep, ac->entry, batchcount, node);
2993 struct list_head *p;
2995 p = l3->slabs_free.next;
2996 while (p != &(l3->slabs_free)) {
2999 slabp = list_entry(p, struct slab, list);
3000 BUG_ON(slabp->inuse);
3005 STATS_SET_FREEABLE(cachep, i);
3008 spin_unlock(&l3->list_lock);
3009 ac->avail -= batchcount;
3010 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3014 * Release an obj back to its cache. If the obj has a constructed state, it must
3015 * be in this state _before_ it is released. Called with disabled ints.
3017 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3019 struct array_cache *ac = cpu_cache_get(cachep);
3022 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3024 /* Make sure we are not freeing a object from another
3025 * node to the array cache on this cpu.
3030 slabp = virt_to_slab(objp);
3031 if (unlikely(slabp->nodeid != numa_node_id())) {
3032 struct array_cache *alien = NULL;
3033 int nodeid = slabp->nodeid;
3034 struct kmem_list3 *l3;
3036 l3 = cachep->nodelists[numa_node_id()];
3037 STATS_INC_NODEFREES(cachep);
3038 if (l3->alien && l3->alien[nodeid]) {
3039 alien = l3->alien[nodeid];
3040 spin_lock(&alien->lock);
3041 if (unlikely(alien->avail == alien->limit))
3042 __drain_alien_cache(cachep,
3044 alien->entry[alien->avail++] = objp;
3045 spin_unlock(&alien->lock);
3047 spin_lock(&(cachep->nodelists[nodeid])->
3049 free_block(cachep, &objp, 1, nodeid);
3050 spin_unlock(&(cachep->nodelists[nodeid])->
3057 if (likely(ac->avail < ac->limit)) {
3058 STATS_INC_FREEHIT(cachep);
3059 ac->entry[ac->avail++] = objp;
3062 STATS_INC_FREEMISS(cachep);
3063 cache_flusharray(cachep, ac);
3064 ac->entry[ac->avail++] = objp;
3069 * kmem_cache_alloc - Allocate an object
3070 * @cachep: The cache to allocate from.
3071 * @flags: See kmalloc().
3073 * Allocate an object from this cache. The flags are only relevant
3074 * if the cache has no available objects.
3076 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3078 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3080 EXPORT_SYMBOL(kmem_cache_alloc);
3083 * kmem_ptr_validate - check if an untrusted pointer might
3085 * @cachep: the cache we're checking against
3086 * @ptr: pointer to validate
3088 * This verifies that the untrusted pointer looks sane:
3089 * it is _not_ a guarantee that the pointer is actually
3090 * part of the slab cache in question, but it at least
3091 * validates that the pointer can be dereferenced and
3092 * looks half-way sane.
3094 * Currently only used for dentry validation.
3096 int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr)
3098 unsigned long addr = (unsigned long)ptr;
3099 unsigned long min_addr = PAGE_OFFSET;
3100 unsigned long align_mask = BYTES_PER_WORD - 1;
3101 unsigned long size = cachep->buffer_size;
3104 if (unlikely(addr < min_addr))
3106 if (unlikely(addr > (unsigned long)high_memory - size))
3108 if (unlikely(addr & align_mask))
3110 if (unlikely(!kern_addr_valid(addr)))
3112 if (unlikely(!kern_addr_valid(addr + size - 1)))
3114 page = virt_to_page(ptr);
3115 if (unlikely(!PageSlab(page)))
3117 if (unlikely(page_get_cache(page) != cachep))
3126 * kmem_cache_alloc_node - Allocate an object on the specified node
3127 * @cachep: The cache to allocate from.
3128 * @flags: See kmalloc().
3129 * @nodeid: node number of the target node.
3131 * Identical to kmem_cache_alloc, except that this function is slow
3132 * and can sleep. And it will allocate memory on the given node, which
3133 * can improve the performance for cpu bound structures.
3134 * New and improved: it will now make sure that the object gets
3135 * put on the correct node list so that there is no false sharing.
3137 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3139 unsigned long save_flags;
3142 cache_alloc_debugcheck_before(cachep, flags);
3143 local_irq_save(save_flags);
3145 if (nodeid == -1 || nodeid == numa_node_id() ||
3146 !cachep->nodelists[nodeid])
3147 ptr = ____cache_alloc(cachep, flags);
3149 ptr = __cache_alloc_node(cachep, flags, nodeid);
3150 local_irq_restore(save_flags);
3152 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr,
3153 __builtin_return_address(0));
3157 EXPORT_SYMBOL(kmem_cache_alloc_node);
3159 void *kmalloc_node(size_t size, gfp_t flags, int node)
3161 struct kmem_cache *cachep;
3163 cachep = kmem_find_general_cachep(size, flags);
3164 if (unlikely(cachep == NULL))
3166 return kmem_cache_alloc_node(cachep, flags, node);
3168 EXPORT_SYMBOL(kmalloc_node);
3172 * kmalloc - allocate memory
3173 * @size: how many bytes of memory are required.
3174 * @flags: the type of memory to allocate.
3175 * @caller: function caller for debug tracking of the caller
3177 * kmalloc is the normal method of allocating memory
3180 * The @flags argument may be one of:
3182 * %GFP_USER - Allocate memory on behalf of user. May sleep.
3184 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
3186 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
3188 * Additionally, the %GFP_DMA flag may be set to indicate the memory
3189 * must be suitable for DMA. This can mean different things on different
3190 * platforms. For example, on i386, it means that the memory must come
3191 * from the first 16MB.
3193 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3196 struct kmem_cache *cachep;
3198 /* If you want to save a few bytes .text space: replace
3200 * Then kmalloc uses the uninlined functions instead of the inline
3203 cachep = __find_general_cachep(size, flags);
3204 if (unlikely(cachep == NULL))
3206 return __cache_alloc(cachep, flags, caller);
3209 #ifndef CONFIG_DEBUG_SLAB
3211 void *__kmalloc(size_t size, gfp_t flags)
3213 return __do_kmalloc(size, flags, NULL);
3215 EXPORT_SYMBOL(__kmalloc);
3219 void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3221 return __do_kmalloc(size, flags, caller);
3223 EXPORT_SYMBOL(__kmalloc_track_caller);
3229 * __alloc_percpu - allocate one copy of the object for every present
3230 * cpu in the system, zeroing them.
3231 * Objects should be dereferenced using the per_cpu_ptr macro only.
3233 * @size: how many bytes of memory are required.
3235 void *__alloc_percpu(size_t size)
3238 struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL);
3244 * Cannot use for_each_online_cpu since a cpu may come online
3245 * and we have no way of figuring out how to fix the array
3246 * that we have allocated then....
3249 int node = cpu_to_node(i);
3251 if (node_online(node))
3252 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
3254 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
3256 if (!pdata->ptrs[i])
3258 memset(pdata->ptrs[i], 0, size);
3261 /* Catch derefs w/o wrappers */
3262 return (void *)(~(unsigned long)pdata);
3266 if (!cpu_possible(i))
3268 kfree(pdata->ptrs[i]);
3273 EXPORT_SYMBOL(__alloc_percpu);
3277 * kmem_cache_free - Deallocate an object
3278 * @cachep: The cache the allocation was from.
3279 * @objp: The previously allocated object.
3281 * Free an object which was previously allocated from this
3284 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3286 unsigned long flags;
3288 local_irq_save(flags);
3289 __cache_free(cachep, objp);
3290 local_irq_restore(flags);
3292 EXPORT_SYMBOL(kmem_cache_free);
3295 * kfree - free previously allocated memory
3296 * @objp: pointer returned by kmalloc.
3298 * If @objp is NULL, no operation is performed.
3300 * Don't free memory not originally allocated by kmalloc()
3301 * or you will run into trouble.
3303 void kfree(const void *objp)
3305 struct kmem_cache *c;
3306 unsigned long flags;
3308 if (unlikely(!objp))
3310 local_irq_save(flags);
3311 kfree_debugcheck(objp);
3312 c = virt_to_cache(objp);
3313 mutex_debug_check_no_locks_freed(objp, obj_size(c));
3314 __cache_free(c, (void *)objp);
3315 local_irq_restore(flags);
3317 EXPORT_SYMBOL(kfree);
3321 * free_percpu - free previously allocated percpu memory
3322 * @objp: pointer returned by alloc_percpu.
3324 * Don't free memory not originally allocated by alloc_percpu()
3325 * The complemented objp is to check for that.
3327 void free_percpu(const void *objp)
3330 struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp);
3333 * We allocate for all cpus so we cannot use for online cpu here.
3339 EXPORT_SYMBOL(free_percpu);
3342 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3344 return obj_size(cachep);
3346 EXPORT_SYMBOL(kmem_cache_size);
3348 const char *kmem_cache_name(struct kmem_cache *cachep)
3350 return cachep->name;
3352 EXPORT_SYMBOL_GPL(kmem_cache_name);
3355 * This initializes kmem_list3 for all nodes.
3357 static int alloc_kmemlist(struct kmem_cache *cachep)
3360 struct kmem_list3 *l3;
3363 for_each_online_node(node) {
3364 struct array_cache *nc = NULL, *new;
3365 struct array_cache **new_alien = NULL;
3367 new_alien = alloc_alien_cache(node, cachep->limit);
3371 new = alloc_arraycache(node, cachep->shared*cachep->batchcount,
3375 l3 = cachep->nodelists[node];
3377 spin_lock_irq(&l3->list_lock);
3379 nc = cachep->nodelists[node]->shared;
3381 free_block(cachep, nc->entry, nc->avail, node);
3384 if (!cachep->nodelists[node]->alien) {
3385 l3->alien = new_alien;
3388 l3->free_limit = (1 + nr_cpus_node(node)) *
3389 cachep->batchcount + cachep->num;
3390 spin_unlock_irq(&l3->list_lock);
3392 free_alien_cache(new_alien);
3395 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
3399 kmem_list3_init(l3);
3400 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3401 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3403 l3->alien = new_alien;
3404 l3->free_limit = (1 + nr_cpus_node(node)) *
3405 cachep->batchcount + cachep->num;
3406 cachep->nodelists[node] = l3;
3414 struct ccupdate_struct {
3415 struct kmem_cache *cachep;
3416 struct array_cache *new[NR_CPUS];
3419 static void do_ccupdate_local(void *info)
3421 struct ccupdate_struct *new = info;
3422 struct array_cache *old;
3425 old = cpu_cache_get(new->cachep);
3427 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3428 new->new[smp_processor_id()] = old;
3431 /* Always called with the cache_chain_mutex held */
3432 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3433 int batchcount, int shared)
3435 struct ccupdate_struct new;
3438 memset(&new.new, 0, sizeof(new.new));
3439 for_each_online_cpu(i) {
3440 new.new[i] = alloc_arraycache(cpu_to_node(i), limit,
3443 for (i--; i >= 0; i--)
3448 new.cachep = cachep;
3450 on_each_cpu(do_ccupdate_local, (void *)&new, 1, 1);
3453 cachep->batchcount = batchcount;
3454 cachep->limit = limit;
3455 cachep->shared = shared;
3457 for_each_online_cpu(i) {
3458 struct array_cache *ccold = new.new[i];
3461 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3462 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3463 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3467 err = alloc_kmemlist(cachep);
3469 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
3470 cachep->name, -err);
3476 /* Called with cache_chain_mutex held always */
3477 static void enable_cpucache(struct kmem_cache *cachep)
3483 * The head array serves three purposes:
3484 * - create a LIFO ordering, i.e. return objects that are cache-warm
3485 * - reduce the number of spinlock operations.
3486 * - reduce the number of linked list operations on the slab and
3487 * bufctl chains: array operations are cheaper.
3488 * The numbers are guessed, we should auto-tune as described by
3491 if (cachep->buffer_size > 131072)
3493 else if (cachep->buffer_size > PAGE_SIZE)
3495 else if (cachep->buffer_size > 1024)
3497 else if (cachep->buffer_size > 256)
3503 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3504 * allocation behaviour: Most allocs on one cpu, most free operations
3505 * on another cpu. For these cases, an efficient object passing between
3506 * cpus is necessary. This is provided by a shared array. The array
3507 * replaces Bonwick's magazine layer.
3508 * On uniprocessor, it's functionally equivalent (but less efficient)
3509 * to a larger limit. Thus disabled by default.
3513 if (cachep->buffer_size <= PAGE_SIZE)
3519 * With debugging enabled, large batchcount lead to excessively long
3520 * periods with disabled local interrupts. Limit the batchcount
3525 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
3527 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3528 cachep->name, -err);
3532 * Drain an array if it contains any elements taking the l3 lock only if
3533 * necessary. Note that the l3 listlock also protects the array_cache
3534 * if drain_array() is used on the shared array.
3536 void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
3537 struct array_cache *ac, int force, int node)
3541 if (!ac || !ac->avail)
3543 if (ac->touched && !force) {
3546 spin_lock_irq(&l3->list_lock);
3548 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3549 if (tofree > ac->avail)
3550 tofree = (ac->avail + 1) / 2;
3551 free_block(cachep, ac->entry, tofree, node);
3552 ac->avail -= tofree;
3553 memmove(ac->entry, &(ac->entry[tofree]),
3554 sizeof(void *) * ac->avail);
3556 spin_unlock_irq(&l3->list_lock);
3561 * cache_reap - Reclaim memory from caches.
3562 * @unused: unused parameter
3564 * Called from workqueue/eventd every few seconds.
3566 * - clear the per-cpu caches for this CPU.
3567 * - return freeable pages to the main free memory pool.
3569 * If we cannot acquire the cache chain mutex then just give up - we'll try
3570 * again on the next iteration.
3572 static void cache_reap(void *unused)
3574 struct list_head *walk;
3575 struct kmem_list3 *l3;
3576 int node = numa_node_id();
3578 if (!mutex_trylock(&cache_chain_mutex)) {
3579 /* Give up. Setup the next iteration. */
3580 schedule_delayed_work(&__get_cpu_var(reap_work),
3585 list_for_each(walk, &cache_chain) {
3586 struct kmem_cache *searchp;
3587 struct list_head *p;
3591 searchp = list_entry(walk, struct kmem_cache, next);
3595 * We only take the l3 lock if absolutely necessary and we
3596 * have established with reasonable certainty that
3597 * we can do some work if the lock was obtained.
3599 l3 = searchp->nodelists[node];
3601 reap_alien(searchp, l3);
3603 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
3606 * These are racy checks but it does not matter
3607 * if we skip one check or scan twice.
3609 if (time_after(l3->next_reap, jiffies))
3612 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
3614 drain_array(searchp, l3, l3->shared, 0, node);
3616 if (l3->free_touched) {
3617 l3->free_touched = 0;
3621 tofree = (l3->free_limit + 5 * searchp->num - 1) /
3625 * Do not lock if there are no free blocks.
3627 if (list_empty(&l3->slabs_free))
3630 spin_lock_irq(&l3->list_lock);
3631 p = l3->slabs_free.next;
3632 if (p == &(l3->slabs_free)) {
3633 spin_unlock_irq(&l3->list_lock);
3637 slabp = list_entry(p, struct slab, list);
3638 BUG_ON(slabp->inuse);
3639 list_del(&slabp->list);
3640 STATS_INC_REAPED(searchp);
3643 * Safe to drop the lock. The slab is no longer linked
3644 * to the cache. searchp cannot disappear, we hold
3647 l3->free_objects -= searchp->num;
3648 spin_unlock_irq(&l3->list_lock);
3649 slab_destroy(searchp, slabp);
3650 } while (--tofree > 0);
3655 mutex_unlock(&cache_chain_mutex);
3657 /* Set up the next iteration */
3658 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
3661 #ifdef CONFIG_PROC_FS
3663 static void print_slabinfo_header(struct seq_file *m)
3666 * Output format version, so at least we can change it
3667 * without _too_ many complaints.
3670 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3672 seq_puts(m, "slabinfo - version: 2.1\n");
3674 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3675 "<objperslab> <pagesperslab>");
3676 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3677 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3679 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3680 "<error> <maxfreeable> <nodeallocs> <remotefrees>");
3681 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3686 static void *s_start(struct seq_file *m, loff_t *pos)
3689 struct list_head *p;
3691 mutex_lock(&cache_chain_mutex);
3693 print_slabinfo_header(m);
3694 p = cache_chain.next;
3697 if (p == &cache_chain)
3700 return list_entry(p, struct kmem_cache, next);
3703 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3705 struct kmem_cache *cachep = p;
3707 return cachep->next.next == &cache_chain ?
3708 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
3711 static void s_stop(struct seq_file *m, void *p)
3713 mutex_unlock(&cache_chain_mutex);
3716 static int s_show(struct seq_file *m, void *p)
3718 struct kmem_cache *cachep = p;
3719 struct list_head *q;
3721 unsigned long active_objs;
3722 unsigned long num_objs;
3723 unsigned long active_slabs = 0;
3724 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3728 struct kmem_list3 *l3;
3732 for_each_online_node(node) {
3733 l3 = cachep->nodelists[node];
3738 spin_lock_irq(&l3->list_lock);
3740 list_for_each(q, &l3->slabs_full) {
3741 slabp = list_entry(q, struct slab, list);
3742 if (slabp->inuse != cachep->num && !error)
3743 error = "slabs_full accounting error";
3744 active_objs += cachep->num;
3747 list_for_each(q, &l3->slabs_partial) {
3748 slabp = list_entry(q, struct slab, list);
3749 if (slabp->inuse == cachep->num && !error)
3750 error = "slabs_partial inuse accounting error";
3751 if (!slabp->inuse && !error)
3752 error = "slabs_partial/inuse accounting error";
3753 active_objs += slabp->inuse;
3756 list_for_each(q, &l3->slabs_free) {
3757 slabp = list_entry(q, struct slab, list);
3758 if (slabp->inuse && !error)
3759 error = "slabs_free/inuse accounting error";
3762 free_objects += l3->free_objects;
3764 shared_avail += l3->shared->avail;
3766 spin_unlock_irq(&l3->list_lock);
3768 num_slabs += active_slabs;
3769 num_objs = num_slabs * cachep->num;
3770 if (num_objs - active_objs != free_objects && !error)
3771 error = "free_objects accounting error";
3773 name = cachep->name;
3775 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3777 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
3778 name, active_objs, num_objs, cachep->buffer_size,
3779 cachep->num, (1 << cachep->gfporder));
3780 seq_printf(m, " : tunables %4u %4u %4u",
3781 cachep->limit, cachep->batchcount, cachep->shared);
3782 seq_printf(m, " : slabdata %6lu %6lu %6lu",
3783 active_slabs, num_slabs, shared_avail);
3786 unsigned long high = cachep->high_mark;
3787 unsigned long allocs = cachep->num_allocations;
3788 unsigned long grown = cachep->grown;
3789 unsigned long reaped = cachep->reaped;
3790 unsigned long errors = cachep->errors;
3791 unsigned long max_freeable = cachep->max_freeable;
3792 unsigned long node_allocs = cachep->node_allocs;
3793 unsigned long node_frees = cachep->node_frees;
3795 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
3796 %4lu %4lu %4lu %4lu", allocs, high, grown,
3797 reaped, errors, max_freeable, node_allocs,
3802 unsigned long allochit = atomic_read(&cachep->allochit);
3803 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3804 unsigned long freehit = atomic_read(&cachep->freehit);
3805 unsigned long freemiss = atomic_read(&cachep->freemiss);
3807 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
3808 allochit, allocmiss, freehit, freemiss);
3816 * slabinfo_op - iterator that generates /proc/slabinfo
3825 * num-pages-per-slab
3826 * + further values on SMP and with statistics enabled
3829 struct seq_operations slabinfo_op = {
3836 #define MAX_SLABINFO_WRITE 128
3838 * slabinfo_write - Tuning for the slab allocator
3840 * @buffer: user buffer
3841 * @count: data length
3844 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
3845 size_t count, loff_t *ppos)
3847 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
3848 int limit, batchcount, shared, res;
3849 struct list_head *p;
3851 if (count > MAX_SLABINFO_WRITE)
3853 if (copy_from_user(&kbuf, buffer, count))
3855 kbuf[MAX_SLABINFO_WRITE] = '\0';
3857 tmp = strchr(kbuf, ' ');
3862 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3865 /* Find the cache in the chain of caches. */
3866 mutex_lock(&cache_chain_mutex);
3868 list_for_each(p, &cache_chain) {
3869 struct kmem_cache *cachep;
3871 cachep = list_entry(p, struct kmem_cache, next);
3872 if (!strcmp(cachep->name, kbuf)) {
3873 if (limit < 1 || batchcount < 1 ||
3874 batchcount > limit || shared < 0) {
3877 res = do_tune_cpucache(cachep, limit,
3878 batchcount, shared);
3883 mutex_unlock(&cache_chain_mutex);
3891 * ksize - get the actual amount of memory allocated for a given object
3892 * @objp: Pointer to the object
3894 * kmalloc may internally round up allocations and return more memory
3895 * than requested. ksize() can be used to determine the actual amount of
3896 * memory allocated. The caller may use this additional memory, even though
3897 * a smaller amount of memory was initially specified with the kmalloc call.
3898 * The caller must guarantee that objp points to a valid object previously
3899 * allocated with either kmalloc() or kmem_cache_alloc(). The object
3900 * must not be freed during the duration of the call.
3902 unsigned int ksize(const void *objp)
3904 if (unlikely(objp == NULL))
3907 return obj_size(virt_to_cache(objp));