3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same intializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/config.h>
90 #include <linux/slab.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/seq_file.h>
98 #include <linux/notifier.h>
99 #include <linux/kallsyms.h>
100 #include <linux/cpu.h>
101 #include <linux/sysctl.h>
102 #include <linux/module.h>
103 #include <linux/rcupdate.h>
104 #include <linux/string.h>
105 #include <linux/nodemask.h>
106 #include <linux/mempolicy.h>
107 #include <linux/mutex.h>
109 #include <asm/uaccess.h>
110 #include <asm/cacheflush.h>
111 #include <asm/tlbflush.h>
112 #include <asm/page.h>
115 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
116 * SLAB_RED_ZONE & SLAB_POISON.
117 * 0 for faster, smaller code (especially in the critical paths).
119 * STATS - 1 to collect stats for /proc/slabinfo.
120 * 0 for faster, smaller code (especially in the critical paths).
122 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
125 #ifdef CONFIG_DEBUG_SLAB
128 #define FORCED_DEBUG 1
132 #define FORCED_DEBUG 0
135 /* Shouldn't this be in a header file somewhere? */
136 #define BYTES_PER_WORD sizeof(void *)
138 #ifndef cache_line_size
139 #define cache_line_size() L1_CACHE_BYTES
142 #ifndef ARCH_KMALLOC_MINALIGN
144 * Enforce a minimum alignment for the kmalloc caches.
145 * Usually, the kmalloc caches are cache_line_size() aligned, except when
146 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
147 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
148 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
149 * Note that this flag disables some debug features.
151 #define ARCH_KMALLOC_MINALIGN 0
154 #ifndef ARCH_SLAB_MINALIGN
156 * Enforce a minimum alignment for all caches.
157 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
158 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
159 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
160 * some debug features.
162 #define ARCH_SLAB_MINALIGN 0
165 #ifndef ARCH_KMALLOC_FLAGS
166 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
169 /* Legal flag mask for kmem_cache_create(). */
171 # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
172 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
173 SLAB_NO_REAP | SLAB_CACHE_DMA | \
174 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
175 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
178 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | SLAB_NO_REAP | \
179 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
180 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
187 * Bufctl's are used for linking objs within a slab
190 * This implementation relies on "struct page" for locating the cache &
191 * slab an object belongs to.
192 * This allows the bufctl structure to be small (one int), but limits
193 * the number of objects a slab (not a cache) can contain when off-slab
194 * bufctls are used. The limit is the size of the largest general cache
195 * that does not use off-slab slabs.
196 * For 32bit archs with 4 kB pages, is this 56.
197 * This is not serious, as it is only for large objects, when it is unwise
198 * to have too many per slab.
199 * Note: This limit can be raised by introducing a general cache whose size
200 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
203 typedef unsigned int kmem_bufctl_t;
204 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
205 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
206 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-2)
208 /* Max number of objs-per-slab for caches which use off-slab slabs.
209 * Needed to avoid a possible looping condition in cache_grow().
211 static unsigned long offslab_limit;
216 * Manages the objs in a slab. Placed either at the beginning of mem allocated
217 * for a slab, or allocated from an general cache.
218 * Slabs are chained into three list: fully used, partial, fully free slabs.
221 struct list_head list;
222 unsigned long colouroff;
223 void *s_mem; /* including colour offset */
224 unsigned int inuse; /* num of objs active in slab */
226 unsigned short nodeid;
232 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
233 * arrange for kmem_freepages to be called via RCU. This is useful if
234 * we need to approach a kernel structure obliquely, from its address
235 * obtained without the usual locking. We can lock the structure to
236 * stabilize it and check it's still at the given address, only if we
237 * can be sure that the memory has not been meanwhile reused for some
238 * other kind of object (which our subsystem's lock might corrupt).
240 * rcu_read_lock before reading the address, then rcu_read_unlock after
241 * taking the spinlock within the structure expected at that address.
243 * We assume struct slab_rcu can overlay struct slab when destroying.
246 struct rcu_head head;
247 struct kmem_cache *cachep;
255 * - LIFO ordering, to hand out cache-warm objects from _alloc
256 * - reduce the number of linked list operations
257 * - reduce spinlock operations
259 * The limit is stored in the per-cpu structure to reduce the data cache
266 unsigned int batchcount;
267 unsigned int touched;
270 * Must have this definition in here for the proper
271 * alignment of array_cache. Also simplifies accessing
273 * [0] is for gcc 2.95. It should really be [].
277 /* bootstrap: The caches do not work without cpuarrays anymore,
278 * but the cpuarrays are allocated from the generic caches...
280 #define BOOT_CPUCACHE_ENTRIES 1
281 struct arraycache_init {
282 struct array_cache cache;
283 void *entries[BOOT_CPUCACHE_ENTRIES];
287 * The slab lists for all objects.
290 struct list_head slabs_partial; /* partial list first, better asm code */
291 struct list_head slabs_full;
292 struct list_head slabs_free;
293 unsigned long free_objects;
294 unsigned long next_reap;
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 */
304 * Need this for bootstrapping a per node allocator.
306 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
307 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
308 #define CACHE_CACHE 0
310 #define SIZE_L3 (1 + MAX_NUMNODES)
313 * This function must be completely optimized away if
314 * a constant is passed to it. Mostly the same as
315 * what is in linux/slab.h except it returns an
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 unsigned int batchcount;
379 unsigned int buffer_size;
380 /* 2) touched by every alloc & free from the backend */
381 struct kmem_list3 *nodelists[MAX_NUMNODES];
382 unsigned int flags; /* constant flags */
383 unsigned int num; /* # of objs per slab */
386 /* 3) cache_grow/shrink */
387 /* order of pgs per slab (2^n) */
388 unsigned int gfporder;
390 /* force GFP flags, e.g. GFP_DMA */
393 size_t colour; /* cache colouring range */
394 unsigned int colour_off; /* colour offset */
395 struct kmem_cache *slabp_cache;
396 unsigned int slab_size;
397 unsigned int dflags; /* dynamic flags */
399 /* constructor func */
400 void (*ctor) (void *, struct kmem_cache *, unsigned long);
402 /* de-constructor func */
403 void (*dtor) (void *, struct kmem_cache *, unsigned long);
405 /* 4) cache creation/removal */
407 struct list_head next;
411 unsigned long num_active;
412 unsigned long num_allocations;
413 unsigned long high_mark;
415 unsigned long reaped;
416 unsigned long errors;
417 unsigned long max_freeable;
418 unsigned long node_allocs;
419 unsigned long node_frees;
427 * If debugging is enabled, then the allocator can add additional
428 * fields and/or padding to every object. buffer_size contains the total
429 * object size including these internal fields, the following two
430 * variables contain the offset to the user object and its size.
437 #define CFLGS_OFF_SLAB (0x80000000UL)
438 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
440 #define BATCHREFILL_LIMIT 16
441 /* Optimization question: fewer reaps means less
442 * probability for unnessary cpucache drain/refill cycles.
444 * OTOH the cpuarrays can contain lots of objects,
445 * which could lock up otherwise freeable slabs.
447 #define REAPTIMEOUT_CPUC (2*HZ)
448 #define REAPTIMEOUT_LIST3 (4*HZ)
451 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
452 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
453 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
454 #define STATS_INC_GROWN(x) ((x)->grown++)
455 #define STATS_INC_REAPED(x) ((x)->reaped++)
456 #define STATS_SET_HIGH(x) do { if ((x)->num_active > (x)->high_mark) \
457 (x)->high_mark = (x)->num_active; \
459 #define STATS_INC_ERR(x) ((x)->errors++)
460 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
461 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
462 #define STATS_SET_FREEABLE(x, i) \
463 do { if ((x)->max_freeable < i) \
464 (x)->max_freeable = i; \
467 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
468 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
469 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
470 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
472 #define STATS_INC_ACTIVE(x) do { } while (0)
473 #define STATS_DEC_ACTIVE(x) do { } while (0)
474 #define STATS_INC_ALLOCED(x) do { } while (0)
475 #define STATS_INC_GROWN(x) do { } while (0)
476 #define STATS_INC_REAPED(x) do { } while (0)
477 #define STATS_SET_HIGH(x) do { } while (0)
478 #define STATS_INC_ERR(x) do { } while (0)
479 #define STATS_INC_NODEALLOCS(x) do { } while (0)
480 #define STATS_INC_NODEFREES(x) do { } while (0)
481 #define STATS_SET_FREEABLE(x, i) \
484 #define STATS_INC_ALLOCHIT(x) do { } while (0)
485 #define STATS_INC_ALLOCMISS(x) do { } while (0)
486 #define STATS_INC_FREEHIT(x) do { } while (0)
487 #define STATS_INC_FREEMISS(x) do { } while (0)
491 /* Magic nums for obj red zoning.
492 * Placed in the first word before and the first word after an obj.
494 #define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
495 #define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
497 /* ...and for poisoning */
498 #define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
499 #define POISON_FREE 0x6b /* for use-after-free poisoning */
500 #define POISON_END 0xa5 /* end-byte of poisoning */
502 /* memory layout of objects:
504 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
505 * the end of an object is aligned with the end of the real
506 * allocation. Catches writes behind the end of the allocation.
507 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
509 * cachep->obj_offset: The real object.
510 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
511 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address [BYTES_PER_WORD long]
513 static int obj_offset(struct kmem_cache *cachep)
515 return cachep->obj_offset;
518 static int obj_size(struct kmem_cache *cachep)
520 return cachep->obj_size;
523 static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
525 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
526 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD);
529 static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
531 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
532 if (cachep->flags & SLAB_STORE_USER)
533 return (unsigned long *)(objp + cachep->buffer_size -
535 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD);
538 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
540 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
541 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
546 #define obj_offset(x) 0
547 #define obj_size(cachep) (cachep->buffer_size)
548 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
549 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
550 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
555 * Maximum size of an obj (in 2^order pages)
556 * and absolute limit for the gfp order.
558 #if defined(CONFIG_LARGE_ALLOCS)
559 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
560 #define MAX_GFP_ORDER 13 /* up to 32Mb */
561 #elif defined(CONFIG_MMU)
562 #define MAX_OBJ_ORDER 5 /* 32 pages */
563 #define MAX_GFP_ORDER 5 /* 32 pages */
565 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
566 #define MAX_GFP_ORDER 8 /* up to 1Mb */
570 * Do not go above this order unless 0 objects fit into the slab.
572 #define BREAK_GFP_ORDER_HI 1
573 #define BREAK_GFP_ORDER_LO 0
574 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
576 /* Functions for storing/retrieving the cachep and or slab from the
577 * global 'mem_map'. These are used to find the slab an obj belongs to.
578 * With kfree(), these are used to find the cache which an obj belongs to.
580 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
582 page->lru.next = (struct list_head *)cache;
585 static inline struct kmem_cache *page_get_cache(struct page *page)
587 return (struct kmem_cache *)page->lru.next;
590 static inline void page_set_slab(struct page *page, struct slab *slab)
592 page->lru.prev = (struct list_head *)slab;
595 static inline struct slab *page_get_slab(struct page *page)
597 return (struct slab *)page->lru.prev;
600 static inline struct kmem_cache *virt_to_cache(const void *obj)
602 struct page *page = virt_to_page(obj);
603 return page_get_cache(page);
606 static inline struct slab *virt_to_slab(const void *obj)
608 struct page *page = virt_to_page(obj);
609 return page_get_slab(page);
612 /* These are the default caches for kmalloc. Custom caches can have other sizes. */
613 struct cache_sizes malloc_sizes[] = {
614 #define CACHE(x) { .cs_size = (x) },
615 #include <linux/kmalloc_sizes.h>
619 EXPORT_SYMBOL(malloc_sizes);
621 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
627 static struct cache_names __initdata cache_names[] = {
628 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
629 #include <linux/kmalloc_sizes.h>
634 static struct arraycache_init initarray_cache __initdata =
635 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
636 static struct arraycache_init initarray_generic =
637 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
639 /* internal cache of cache description objs */
640 static struct kmem_cache cache_cache = {
642 .limit = BOOT_CPUCACHE_ENTRIES,
644 .buffer_size = sizeof(struct kmem_cache),
645 .flags = SLAB_NO_REAP,
646 .spinlock = SPIN_LOCK_UNLOCKED,
647 .name = "kmem_cache",
649 .obj_size = sizeof(struct kmem_cache),
653 /* Guard access to the cache-chain. */
654 static DEFINE_MUTEX(cache_chain_mutex);
655 static struct list_head cache_chain;
658 * vm_enough_memory() looks at this to determine how many
659 * slab-allocated pages are possibly freeable under pressure
661 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
663 atomic_t slab_reclaim_pages;
666 * chicken and egg problem: delay the per-cpu array allocation
667 * until the general caches are up.
676 static DEFINE_PER_CPU(struct work_struct, reap_work);
678 static void free_block(struct kmem_cache *cachep, void **objpp, int len, int node);
679 static void enable_cpucache(struct kmem_cache *cachep);
680 static void cache_reap(void *unused);
681 static int __node_shrink(struct kmem_cache *cachep, int node);
683 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
685 return cachep->array[smp_processor_id()];
688 static inline struct kmem_cache *__find_general_cachep(size_t size, gfp_t gfpflags)
690 struct cache_sizes *csizep = malloc_sizes;
693 /* This happens if someone tries to call
694 * kmem_cache_create(), or __kmalloc(), before
695 * the generic caches are initialized.
697 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
699 while (size > csizep->cs_size)
703 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
704 * has cs_{dma,}cachep==NULL. Thus no special case
705 * for large kmalloc calls required.
707 if (unlikely(gfpflags & GFP_DMA))
708 return csizep->cs_dmacachep;
709 return csizep->cs_cachep;
712 struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
714 return __find_general_cachep(size, gfpflags);
716 EXPORT_SYMBOL(kmem_find_general_cachep);
718 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
720 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
723 /* Calculate the number of objects and left-over bytes for a given
725 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
726 size_t align, int flags, size_t *left_over,
731 size_t slab_size = PAGE_SIZE << gfporder;
734 * The slab management structure can be either off the slab or
735 * on it. For the latter case, the memory allocated for a
739 * - One kmem_bufctl_t for each object
740 * - Padding to respect alignment of @align
741 * - @buffer_size bytes for each object
743 * If the slab management structure is off the slab, then the
744 * alignment will already be calculated into the size. Because
745 * the slabs are all pages aligned, the objects will be at the
746 * correct alignment when allocated.
748 if (flags & CFLGS_OFF_SLAB) {
750 nr_objs = slab_size / buffer_size;
752 if (nr_objs > SLAB_LIMIT)
753 nr_objs = SLAB_LIMIT;
756 * Ignore padding for the initial guess. The padding
757 * is at most @align-1 bytes, and @buffer_size is at
758 * least @align. In the worst case, this result will
759 * be one greater than the number of objects that fit
760 * into the memory allocation when taking the padding
763 nr_objs = (slab_size - sizeof(struct slab)) /
764 (buffer_size + sizeof(kmem_bufctl_t));
767 * This calculated number will be either the right
768 * amount, or one greater than what we want.
770 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
774 if (nr_objs > SLAB_LIMIT)
775 nr_objs = SLAB_LIMIT;
777 mgmt_size = slab_mgmt_size(nr_objs, align);
780 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
783 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
785 static void __slab_error(const char *function, struct kmem_cache *cachep, char *msg)
787 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
788 function, cachep->name, msg);
793 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
794 * via the workqueue/eventd.
795 * Add the CPU number into the expiration time to minimize the possibility of
796 * the CPUs getting into lockstep and contending for the global cache chain
799 static void __devinit start_cpu_timer(int cpu)
801 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
804 * When this gets called from do_initcalls via cpucache_init(),
805 * init_workqueues() has already run, so keventd will be setup
808 if (keventd_up() && reap_work->func == NULL) {
809 INIT_WORK(reap_work, cache_reap, NULL);
810 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
814 static struct array_cache *alloc_arraycache(int node, int entries,
817 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
818 struct array_cache *nc = NULL;
820 nc = kmalloc_node(memsize, GFP_KERNEL, node);
824 nc->batchcount = batchcount;
826 spin_lock_init(&nc->lock);
832 static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int);
834 static struct array_cache **alloc_alien_cache(int node, int limit)
836 struct array_cache **ac_ptr;
837 int memsize = sizeof(void *) * MAX_NUMNODES;
842 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
845 if (i == node || !node_online(i)) {
849 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
851 for (i--; i <= 0; i--)
861 static void free_alien_cache(struct array_cache **ac_ptr)
874 static void __drain_alien_cache(struct kmem_cache *cachep,
875 struct array_cache *ac, int node)
877 struct kmem_list3 *rl3 = cachep->nodelists[node];
880 spin_lock(&rl3->list_lock);
881 free_block(cachep, ac->entry, ac->avail, node);
883 spin_unlock(&rl3->list_lock);
887 static void drain_alien_cache(struct kmem_cache *cachep, struct array_cache **alien)
890 struct array_cache *ac;
893 for_each_online_node(i) {
896 spin_lock_irqsave(&ac->lock, flags);
897 __drain_alien_cache(cachep, ac, i);
898 spin_unlock_irqrestore(&ac->lock, flags);
903 #define alloc_alien_cache(node, limit) do { } while (0)
904 #define drain_alien_cache(cachep, alien) do { } while (0)
906 static inline void free_alien_cache(struct array_cache **ac_ptr)
911 static int __devinit cpuup_callback(struct notifier_block *nfb,
912 unsigned long action, void *hcpu)
914 long cpu = (long)hcpu;
915 struct kmem_cache *cachep;
916 struct kmem_list3 *l3 = NULL;
917 int node = cpu_to_node(cpu);
918 int memsize = sizeof(struct kmem_list3);
922 mutex_lock(&cache_chain_mutex);
923 /* we need to do this right in the beginning since
924 * alloc_arraycache's are going to use this list.
925 * kmalloc_node allows us to add the slab to the right
926 * kmem_list3 and not this cpu's kmem_list3
929 list_for_each_entry(cachep, &cache_chain, next) {
930 /* setup the size64 kmemlist for cpu before we can
931 * begin anything. Make sure some other cpu on this
932 * node has not already allocated this
934 if (!cachep->nodelists[node]) {
935 if (!(l3 = kmalloc_node(memsize,
939 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
940 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
943 * The l3s don't come and go as CPUs come and
944 * go. cache_chain_mutex is sufficient
947 cachep->nodelists[node] = l3;
950 spin_lock_irq(&cachep->nodelists[node]->list_lock);
951 cachep->nodelists[node]->free_limit =
952 (1 + nr_cpus_node(node)) *
953 cachep->batchcount + cachep->num;
954 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
957 /* Now we can go ahead with allocating the shared array's
959 list_for_each_entry(cachep, &cache_chain, next) {
960 struct array_cache *nc;
961 struct array_cache *shared;
962 struct array_cache **alien;
964 nc = alloc_arraycache(node, cachep->limit,
968 shared = alloc_arraycache(node,
969 cachep->shared * cachep->batchcount,
974 alien = alloc_alien_cache(node, cachep->limit);
978 cachep->array[cpu] = nc;
980 l3 = cachep->nodelists[node];
983 spin_lock_irq(&l3->list_lock);
986 * We are serialised from CPU_DEAD or
987 * CPU_UP_CANCELLED by the cpucontrol lock
998 spin_unlock_irq(&l3->list_lock);
1001 free_alien_cache(alien);
1003 mutex_unlock(&cache_chain_mutex);
1006 start_cpu_timer(cpu);
1008 #ifdef CONFIG_HOTPLUG_CPU
1011 * Even if all the cpus of a node are down, we don't free the
1012 * kmem_list3 of any cache. This to avoid a race between
1013 * cpu_down, and a kmalloc allocation from another cpu for
1014 * memory from the node of the cpu going down. The list3
1015 * structure is usually allocated from kmem_cache_create() and
1016 * gets destroyed at kmem_cache_destroy().
1019 case CPU_UP_CANCELED:
1020 mutex_lock(&cache_chain_mutex);
1022 list_for_each_entry(cachep, &cache_chain, next) {
1023 struct array_cache *nc;
1024 struct array_cache *shared;
1025 struct array_cache **alien;
1028 mask = node_to_cpumask(node);
1029 /* cpu is dead; no one can alloc from it. */
1030 nc = cachep->array[cpu];
1031 cachep->array[cpu] = NULL;
1032 l3 = cachep->nodelists[node];
1035 goto free_array_cache;
1037 spin_lock_irq(&l3->list_lock);
1039 /* Free limit for this kmem_list3 */
1040 l3->free_limit -= cachep->batchcount;
1042 free_block(cachep, nc->entry, nc->avail, node);
1044 if (!cpus_empty(mask)) {
1045 spin_unlock_irq(&l3->list_lock);
1046 goto free_array_cache;
1049 shared = l3->shared;
1051 free_block(cachep, l3->shared->entry,
1052 l3->shared->avail, node);
1059 spin_unlock_irq(&l3->list_lock);
1063 drain_alien_cache(cachep, alien);
1064 free_alien_cache(alien);
1070 * In the previous loop, all the objects were freed to
1071 * the respective cache's slabs, now we can go ahead and
1072 * shrink each nodelist to its limit.
1074 list_for_each_entry(cachep, &cache_chain, next) {
1075 l3 = cachep->nodelists[node];
1078 spin_lock_irq(&l3->list_lock);
1079 /* free slabs belonging to this node */
1080 __node_shrink(cachep, node);
1081 spin_unlock_irq(&l3->list_lock);
1083 mutex_unlock(&cache_chain_mutex);
1089 mutex_unlock(&cache_chain_mutex);
1093 static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
1096 * swap the static kmem_list3 with kmalloced memory
1098 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list, int nodeid)
1100 struct kmem_list3 *ptr;
1102 BUG_ON(cachep->nodelists[nodeid] != list);
1103 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1106 local_irq_disable();
1107 memcpy(ptr, list, sizeof(struct kmem_list3));
1108 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1109 cachep->nodelists[nodeid] = ptr;
1114 * Called after the gfp() functions have been enabled, and before smp_init().
1116 void __init kmem_cache_init(void)
1119 struct cache_sizes *sizes;
1120 struct cache_names *names;
1123 for (i = 0; i < NUM_INIT_LISTS; i++) {
1124 kmem_list3_init(&initkmem_list3[i]);
1125 if (i < MAX_NUMNODES)
1126 cache_cache.nodelists[i] = NULL;
1130 * Fragmentation resistance on low memory - only use bigger
1131 * page orders on machines with more than 32MB of memory.
1133 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1134 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1136 /* Bootstrap is tricky, because several objects are allocated
1137 * from caches that do not exist yet:
1138 * 1) initialize the cache_cache cache: it contains the struct kmem_cache
1139 * structures of all caches, except cache_cache itself: cache_cache
1140 * is statically allocated.
1141 * Initially an __init data area is used for the head array and the
1142 * kmem_list3 structures, it's replaced with a kmalloc allocated
1143 * array at the end of the bootstrap.
1144 * 2) Create the first kmalloc cache.
1145 * The struct kmem_cache for the new cache is allocated normally.
1146 * An __init data area is used for the head array.
1147 * 3) Create the remaining kmalloc caches, with minimally sized
1149 * 4) Replace the __init data head arrays for cache_cache and the first
1150 * kmalloc cache with kmalloc allocated arrays.
1151 * 5) Replace the __init data for kmem_list3 for cache_cache and
1152 * the other cache's with kmalloc allocated memory.
1153 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1156 /* 1) create the cache_cache */
1157 INIT_LIST_HEAD(&cache_chain);
1158 list_add(&cache_cache.next, &cache_chain);
1159 cache_cache.colour_off = cache_line_size();
1160 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1161 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
1163 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, cache_line_size());
1165 cache_estimate(0, cache_cache.buffer_size, cache_line_size(), 0,
1166 &left_over, &cache_cache.num);
1167 if (!cache_cache.num)
1170 cache_cache.colour = left_over / cache_cache.colour_off;
1171 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1172 sizeof(struct slab), cache_line_size());
1174 /* 2+3) create the kmalloc caches */
1175 sizes = malloc_sizes;
1176 names = cache_names;
1178 /* Initialize the caches that provide memory for the array cache
1179 * and the kmem_list3 structures first.
1180 * Without this, further allocations will bug
1183 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1184 sizes[INDEX_AC].cs_size,
1185 ARCH_KMALLOC_MINALIGN,
1186 (ARCH_KMALLOC_FLAGS |
1187 SLAB_PANIC), NULL, NULL);
1189 if (INDEX_AC != INDEX_L3)
1190 sizes[INDEX_L3].cs_cachep =
1191 kmem_cache_create(names[INDEX_L3].name,
1192 sizes[INDEX_L3].cs_size,
1193 ARCH_KMALLOC_MINALIGN,
1194 (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL,
1197 while (sizes->cs_size != ULONG_MAX) {
1199 * For performance, all the general caches are L1 aligned.
1200 * This should be particularly beneficial on SMP boxes, as it
1201 * eliminates "false sharing".
1202 * Note for systems short on memory removing the alignment will
1203 * allow tighter packing of the smaller caches.
1205 if (!sizes->cs_cachep)
1206 sizes->cs_cachep = kmem_cache_create(names->name,
1208 ARCH_KMALLOC_MINALIGN,
1213 /* Inc off-slab bufctl limit until the ceiling is hit. */
1214 if (!(OFF_SLAB(sizes->cs_cachep))) {
1215 offslab_limit = sizes->cs_size - sizeof(struct slab);
1216 offslab_limit /= sizeof(kmem_bufctl_t);
1219 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
1221 ARCH_KMALLOC_MINALIGN,
1222 (ARCH_KMALLOC_FLAGS |
1230 /* 4) Replace the bootstrap head arrays */
1234 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1236 local_irq_disable();
1237 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1238 memcpy(ptr, cpu_cache_get(&cache_cache),
1239 sizeof(struct arraycache_init));
1240 cache_cache.array[smp_processor_id()] = ptr;
1243 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1245 local_irq_disable();
1246 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1247 != &initarray_generic.cache);
1248 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1249 sizeof(struct arraycache_init));
1250 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1254 /* 5) Replace the bootstrap kmem_list3's */
1257 /* Replace the static kmem_list3 structures for the boot cpu */
1258 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
1261 for_each_online_node(node) {
1262 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1263 &initkmem_list3[SIZE_AC + node], node);
1265 if (INDEX_AC != INDEX_L3) {
1266 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1267 &initkmem_list3[SIZE_L3 + node],
1273 /* 6) resize the head arrays to their final sizes */
1275 struct kmem_cache *cachep;
1276 mutex_lock(&cache_chain_mutex);
1277 list_for_each_entry(cachep, &cache_chain, next)
1278 enable_cpucache(cachep);
1279 mutex_unlock(&cache_chain_mutex);
1283 g_cpucache_up = FULL;
1285 /* Register a cpu startup notifier callback
1286 * that initializes cpu_cache_get for all new cpus
1288 register_cpu_notifier(&cpucache_notifier);
1290 /* The reap timers are started later, with a module init call:
1291 * That part of the kernel is not yet operational.
1295 static int __init cpucache_init(void)
1300 * Register the timers that return unneeded
1303 for_each_online_cpu(cpu)
1304 start_cpu_timer(cpu);
1309 __initcall(cpucache_init);
1312 * Interface to system's page allocator. No need to hold the cache-lock.
1314 * If we requested dmaable memory, we will get it. Even if we
1315 * did not request dmaable memory, we might get it, but that
1316 * would be relatively rare and ignorable.
1318 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1324 flags |= cachep->gfpflags;
1325 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1328 addr = page_address(page);
1330 i = (1 << cachep->gfporder);
1331 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1332 atomic_add(i, &slab_reclaim_pages);
1333 add_page_state(nr_slab, i);
1342 * Interface to system's page release.
1344 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1346 unsigned long i = (1 << cachep->gfporder);
1347 struct page *page = virt_to_page(addr);
1348 const unsigned long nr_freed = i;
1351 if (!TestClearPageSlab(page))
1355 sub_page_state(nr_slab, nr_freed);
1356 if (current->reclaim_state)
1357 current->reclaim_state->reclaimed_slab += nr_freed;
1358 free_pages((unsigned long)addr, cachep->gfporder);
1359 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1360 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages);
1363 static void kmem_rcu_free(struct rcu_head *head)
1365 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1366 struct kmem_cache *cachep = slab_rcu->cachep;
1368 kmem_freepages(cachep, slab_rcu->addr);
1369 if (OFF_SLAB(cachep))
1370 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1375 #ifdef CONFIG_DEBUG_PAGEALLOC
1376 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1377 unsigned long caller)
1379 int size = obj_size(cachep);
1381 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1383 if (size < 5 * sizeof(unsigned long))
1386 *addr++ = 0x12345678;
1388 *addr++ = smp_processor_id();
1389 size -= 3 * sizeof(unsigned long);
1391 unsigned long *sptr = &caller;
1392 unsigned long svalue;
1394 while (!kstack_end(sptr)) {
1396 if (kernel_text_address(svalue)) {
1398 size -= sizeof(unsigned long);
1399 if (size <= sizeof(unsigned long))
1405 *addr++ = 0x87654321;
1409 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1411 int size = obj_size(cachep);
1412 addr = &((char *)addr)[obj_offset(cachep)];
1414 memset(addr, val, size);
1415 *(unsigned char *)(addr + size - 1) = POISON_END;
1418 static void dump_line(char *data, int offset, int limit)
1421 printk(KERN_ERR "%03x:", offset);
1422 for (i = 0; i < limit; i++) {
1423 printk(" %02x", (unsigned char)data[offset + i]);
1431 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1436 if (cachep->flags & SLAB_RED_ZONE) {
1437 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
1438 *dbg_redzone1(cachep, objp),
1439 *dbg_redzone2(cachep, objp));
1442 if (cachep->flags & SLAB_STORE_USER) {
1443 printk(KERN_ERR "Last user: [<%p>]",
1444 *dbg_userword(cachep, objp));
1445 print_symbol("(%s)",
1446 (unsigned long)*dbg_userword(cachep, objp));
1449 realobj = (char *)objp + obj_offset(cachep);
1450 size = obj_size(cachep);
1451 for (i = 0; i < size && lines; i += 16, lines--) {
1454 if (i + limit > size)
1456 dump_line(realobj, i, limit);
1460 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1466 realobj = (char *)objp + obj_offset(cachep);
1467 size = obj_size(cachep);
1469 for (i = 0; i < size; i++) {
1470 char exp = POISON_FREE;
1473 if (realobj[i] != exp) {
1479 "Slab corruption: start=%p, len=%d\n",
1481 print_objinfo(cachep, objp, 0);
1483 /* Hexdump the affected line */
1486 if (i + limit > size)
1488 dump_line(realobj, i, limit);
1491 /* Limit to 5 lines */
1497 /* Print some data about the neighboring objects, if they
1500 struct slab *slabp = virt_to_slab(objp);
1503 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
1505 objp = slabp->s_mem + (objnr - 1) * cachep->buffer_size;
1506 realobj = (char *)objp + obj_offset(cachep);
1507 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1509 print_objinfo(cachep, objp, 2);
1511 if (objnr + 1 < cachep->num) {
1512 objp = slabp->s_mem + (objnr + 1) * cachep->buffer_size;
1513 realobj = (char *)objp + obj_offset(cachep);
1514 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1516 print_objinfo(cachep, objp, 2);
1524 * slab_destroy_objs - call the registered destructor for each object in
1525 * a slab that is to be destroyed.
1527 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1530 for (i = 0; i < cachep->num; i++) {
1531 void *objp = slabp->s_mem + cachep->buffer_size * i;
1533 if (cachep->flags & SLAB_POISON) {
1534 #ifdef CONFIG_DEBUG_PAGEALLOC
1535 if ((cachep->buffer_size % PAGE_SIZE) == 0
1536 && OFF_SLAB(cachep))
1537 kernel_map_pages(virt_to_page(objp),
1538 cachep->buffer_size / PAGE_SIZE,
1541 check_poison_obj(cachep, objp);
1543 check_poison_obj(cachep, objp);
1546 if (cachep->flags & SLAB_RED_ZONE) {
1547 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1548 slab_error(cachep, "start of a freed object "
1550 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1551 slab_error(cachep, "end of a freed object "
1554 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1555 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
1559 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1563 for (i = 0; i < cachep->num; i++) {
1564 void *objp = slabp->s_mem + cachep->buffer_size * i;
1565 (cachep->dtor) (objp, cachep, 0);
1572 * Destroy all the objs in a slab, and release the mem back to the system.
1573 * Before calling the slab must have been unlinked from the cache.
1574 * The cache-lock is not held/needed.
1576 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1578 void *addr = slabp->s_mem - slabp->colouroff;
1580 slab_destroy_objs(cachep, slabp);
1581 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1582 struct slab_rcu *slab_rcu;
1584 slab_rcu = (struct slab_rcu *)slabp;
1585 slab_rcu->cachep = cachep;
1586 slab_rcu->addr = addr;
1587 call_rcu(&slab_rcu->head, kmem_rcu_free);
1589 kmem_freepages(cachep, addr);
1590 if (OFF_SLAB(cachep))
1591 kmem_cache_free(cachep->slabp_cache, slabp);
1595 /* For setting up all the kmem_list3s for cache whose buffer_size is same
1596 as size of kmem_list3. */
1597 static void set_up_list3s(struct kmem_cache *cachep, int index)
1601 for_each_online_node(node) {
1602 cachep->nodelists[node] = &initkmem_list3[index + node];
1603 cachep->nodelists[node]->next_reap = jiffies +
1605 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1610 * calculate_slab_order - calculate size (page order) of slabs
1611 * @cachep: pointer to the cache that is being created
1612 * @size: size of objects to be created in this cache.
1613 * @align: required alignment for the objects.
1614 * @flags: slab allocation flags
1616 * Also calculates the number of objects per slab.
1618 * This could be made much more intelligent. For now, try to avoid using
1619 * high order pages for slabs. When the gfp() functions are more friendly
1620 * towards high-order requests, this should be changed.
1622 static inline size_t calculate_slab_order(struct kmem_cache *cachep,
1623 size_t size, size_t align, unsigned long flags)
1625 size_t left_over = 0;
1627 for (;; cachep->gfporder++) {
1631 if (cachep->gfporder > MAX_GFP_ORDER) {
1636 cache_estimate(cachep->gfporder, size, align, flags,
1640 /* More than offslab_limit objects will cause problems */
1641 if (flags & CFLGS_OFF_SLAB && cachep->num > offslab_limit)
1645 left_over = remainder;
1648 * Large number of objects is good, but very large slabs are
1649 * currently bad for the gfp()s.
1651 if (cachep->gfporder >= slab_break_gfp_order)
1654 if ((left_over * 8) <= (PAGE_SIZE << cachep->gfporder))
1655 /* Acceptable internal fragmentation */
1662 * kmem_cache_create - Create a cache.
1663 * @name: A string which is used in /proc/slabinfo to identify this cache.
1664 * @size: The size of objects to be created in this cache.
1665 * @align: The required alignment for the objects.
1666 * @flags: SLAB flags
1667 * @ctor: A constructor for the objects.
1668 * @dtor: A destructor for the objects.
1670 * Returns a ptr to the cache on success, NULL on failure.
1671 * Cannot be called within a int, but can be interrupted.
1672 * The @ctor is run when new pages are allocated by the cache
1673 * and the @dtor is run before the pages are handed back.
1675 * @name must be valid until the cache is destroyed. This implies that
1676 * the module calling this has to destroy the cache before getting
1681 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1682 * to catch references to uninitialised memory.
1684 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1685 * for buffer overruns.
1687 * %SLAB_NO_REAP - Don't automatically reap this cache when we're under
1690 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1691 * cacheline. This can be beneficial if you're counting cycles as closely
1695 kmem_cache_create (const char *name, size_t size, size_t align,
1696 unsigned long flags, void (*ctor)(void*, struct kmem_cache *, unsigned long),
1697 void (*dtor)(void*, struct kmem_cache *, unsigned long))
1699 size_t left_over, slab_size, ralign;
1700 struct kmem_cache *cachep = NULL;
1701 struct list_head *p;
1704 * Sanity checks... these are all serious usage bugs.
1708 (size < BYTES_PER_WORD) ||
1709 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
1710 printk(KERN_ERR "%s: Early error in slab %s\n",
1711 __FUNCTION__, name);
1715 mutex_lock(&cache_chain_mutex);
1717 list_for_each(p, &cache_chain) {
1718 struct kmem_cache *pc = list_entry(p, struct kmem_cache, next);
1719 mm_segment_t old_fs = get_fs();
1724 * This happens when the module gets unloaded and doesn't
1725 * destroy its slab cache and no-one else reuses the vmalloc
1726 * area of the module. Print a warning.
1729 res = __get_user(tmp, pc->name);
1732 printk("SLAB: cache with size %d has lost its name\n",
1737 if (!strcmp(pc->name, name)) {
1738 printk("kmem_cache_create: duplicate cache %s\n", name);
1745 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1746 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
1747 /* No constructor, but inital state check requested */
1748 printk(KERN_ERR "%s: No con, but init state check "
1749 "requested - %s\n", __FUNCTION__, name);
1750 flags &= ~SLAB_DEBUG_INITIAL;
1754 * Enable redzoning and last user accounting, except for caches with
1755 * large objects, if the increased size would increase the object size
1756 * above the next power of two: caches with object sizes just above a
1757 * power of two have a significant amount of internal fragmentation.
1760 || fls(size - 1) == fls(size - 1 + 3 * BYTES_PER_WORD)))
1761 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
1762 if (!(flags & SLAB_DESTROY_BY_RCU))
1763 flags |= SLAB_POISON;
1765 if (flags & SLAB_DESTROY_BY_RCU)
1766 BUG_ON(flags & SLAB_POISON);
1768 if (flags & SLAB_DESTROY_BY_RCU)
1772 * Always checks flags, a caller might be expecting debug
1773 * support which isn't available.
1775 if (flags & ~CREATE_MASK)
1778 /* Check that size is in terms of words. This is needed to avoid
1779 * unaligned accesses for some archs when redzoning is used, and makes
1780 * sure any on-slab bufctl's are also correctly aligned.
1782 if (size & (BYTES_PER_WORD - 1)) {
1783 size += (BYTES_PER_WORD - 1);
1784 size &= ~(BYTES_PER_WORD - 1);
1787 /* calculate out the final buffer alignment: */
1788 /* 1) arch recommendation: can be overridden for debug */
1789 if (flags & SLAB_HWCACHE_ALIGN) {
1790 /* Default alignment: as specified by the arch code.
1791 * Except if an object is really small, then squeeze multiple
1792 * objects into one cacheline.
1794 ralign = cache_line_size();
1795 while (size <= ralign / 2)
1798 ralign = BYTES_PER_WORD;
1800 /* 2) arch mandated alignment: disables debug if necessary */
1801 if (ralign < ARCH_SLAB_MINALIGN) {
1802 ralign = ARCH_SLAB_MINALIGN;
1803 if (ralign > BYTES_PER_WORD)
1804 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
1806 /* 3) caller mandated alignment: disables debug if necessary */
1807 if (ralign < align) {
1809 if (ralign > BYTES_PER_WORD)
1810 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
1812 /* 4) Store it. Note that the debug code below can reduce
1813 * the alignment to BYTES_PER_WORD.
1817 /* Get cache's description obj. */
1818 cachep = kmem_cache_alloc(&cache_cache, SLAB_KERNEL);
1821 memset(cachep, 0, sizeof(struct kmem_cache));
1824 cachep->obj_size = size;
1826 if (flags & SLAB_RED_ZONE) {
1827 /* redzoning only works with word aligned caches */
1828 align = BYTES_PER_WORD;
1830 /* add space for red zone words */
1831 cachep->obj_offset += BYTES_PER_WORD;
1832 size += 2 * BYTES_PER_WORD;
1834 if (flags & SLAB_STORE_USER) {
1835 /* user store requires word alignment and
1836 * one word storage behind the end of the real
1839 align = BYTES_PER_WORD;
1840 size += BYTES_PER_WORD;
1842 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
1843 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
1844 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
1845 cachep->obj_offset += PAGE_SIZE - size;
1851 /* Determine if the slab management is 'on' or 'off' slab. */
1852 if (size >= (PAGE_SIZE >> 3))
1854 * Size is large, assume best to place the slab management obj
1855 * off-slab (should allow better packing of objs).
1857 flags |= CFLGS_OFF_SLAB;
1859 size = ALIGN(size, align);
1861 if ((flags & SLAB_RECLAIM_ACCOUNT) && size <= PAGE_SIZE) {
1863 * A VFS-reclaimable slab tends to have most allocations
1864 * as GFP_NOFS and we really don't want to have to be allocating
1865 * higher-order pages when we are unable to shrink dcache.
1867 cachep->gfporder = 0;
1868 cache_estimate(cachep->gfporder, size, align, flags,
1869 &left_over, &cachep->num);
1871 left_over = calculate_slab_order(cachep, size, align, flags);
1874 printk("kmem_cache_create: couldn't create cache %s.\n", name);
1875 kmem_cache_free(&cache_cache, cachep);
1879 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
1880 + sizeof(struct slab), align);
1883 * If the slab has been placed off-slab, and we have enough space then
1884 * move it on-slab. This is at the expense of any extra colouring.
1886 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
1887 flags &= ~CFLGS_OFF_SLAB;
1888 left_over -= slab_size;
1891 if (flags & CFLGS_OFF_SLAB) {
1892 /* really off slab. No need for manual alignment */
1894 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
1897 cachep->colour_off = cache_line_size();
1898 /* Offset must be a multiple of the alignment. */
1899 if (cachep->colour_off < align)
1900 cachep->colour_off = align;
1901 cachep->colour = left_over / cachep->colour_off;
1902 cachep->slab_size = slab_size;
1903 cachep->flags = flags;
1904 cachep->gfpflags = 0;
1905 if (flags & SLAB_CACHE_DMA)
1906 cachep->gfpflags |= GFP_DMA;
1907 spin_lock_init(&cachep->spinlock);
1908 cachep->buffer_size = size;
1910 if (flags & CFLGS_OFF_SLAB)
1911 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
1912 cachep->ctor = ctor;
1913 cachep->dtor = dtor;
1914 cachep->name = name;
1916 /* Don't let CPUs to come and go */
1919 if (g_cpucache_up == FULL) {
1920 enable_cpucache(cachep);
1922 if (g_cpucache_up == NONE) {
1923 /* Note: the first kmem_cache_create must create
1924 * the cache that's used by kmalloc(24), otherwise
1925 * the creation of further caches will BUG().
1927 cachep->array[smp_processor_id()] =
1928 &initarray_generic.cache;
1930 /* If the cache that's used by
1931 * kmalloc(sizeof(kmem_list3)) is the first cache,
1932 * then we need to set up all its list3s, otherwise
1933 * the creation of further caches will BUG().
1935 set_up_list3s(cachep, SIZE_AC);
1936 if (INDEX_AC == INDEX_L3)
1937 g_cpucache_up = PARTIAL_L3;
1939 g_cpucache_up = PARTIAL_AC;
1941 cachep->array[smp_processor_id()] =
1942 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1944 if (g_cpucache_up == PARTIAL_AC) {
1945 set_up_list3s(cachep, SIZE_L3);
1946 g_cpucache_up = PARTIAL_L3;
1949 for_each_online_node(node) {
1951 cachep->nodelists[node] =
1953 (struct kmem_list3),
1955 BUG_ON(!cachep->nodelists[node]);
1956 kmem_list3_init(cachep->
1961 cachep->nodelists[numa_node_id()]->next_reap =
1962 jiffies + REAPTIMEOUT_LIST3 +
1963 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1965 BUG_ON(!cpu_cache_get(cachep));
1966 cpu_cache_get(cachep)->avail = 0;
1967 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1968 cpu_cache_get(cachep)->batchcount = 1;
1969 cpu_cache_get(cachep)->touched = 0;
1970 cachep->batchcount = 1;
1971 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1974 /* cache setup completed, link it into the list */
1975 list_add(&cachep->next, &cache_chain);
1976 unlock_cpu_hotplug();
1978 if (!cachep && (flags & SLAB_PANIC))
1979 panic("kmem_cache_create(): failed to create slab `%s'\n",
1981 mutex_unlock(&cache_chain_mutex);
1984 EXPORT_SYMBOL(kmem_cache_create);
1987 static void check_irq_off(void)
1989 BUG_ON(!irqs_disabled());
1992 static void check_irq_on(void)
1994 BUG_ON(irqs_disabled());
1997 static void check_spinlock_acquired(struct kmem_cache *cachep)
2001 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2005 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2009 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2014 #define check_irq_off() do { } while(0)
2015 #define check_irq_on() do { } while(0)
2016 #define check_spinlock_acquired(x) do { } while(0)
2017 #define check_spinlock_acquired_node(x, y) do { } while(0)
2021 * Waits for all CPUs to execute func().
2023 static void smp_call_function_all_cpus(void (*func)(void *arg), void *arg)
2028 local_irq_disable();
2032 if (smp_call_function(func, arg, 1, 1))
2038 static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac,
2039 int force, int node);
2041 static void do_drain(void *arg)
2043 struct kmem_cache *cachep = (struct kmem_cache *) arg;
2044 struct array_cache *ac;
2045 int node = numa_node_id();
2048 ac = cpu_cache_get(cachep);
2049 spin_lock(&cachep->nodelists[node]->list_lock);
2050 free_block(cachep, ac->entry, ac->avail, node);
2051 spin_unlock(&cachep->nodelists[node]->list_lock);
2055 static void drain_cpu_caches(struct kmem_cache *cachep)
2057 struct kmem_list3 *l3;
2060 smp_call_function_all_cpus(do_drain, cachep);
2062 for_each_online_node(node) {
2063 l3 = cachep->nodelists[node];
2065 spin_lock_irq(&l3->list_lock);
2066 drain_array_locked(cachep, l3->shared, 1, node);
2067 spin_unlock_irq(&l3->list_lock);
2069 drain_alien_cache(cachep, l3->alien);
2074 static int __node_shrink(struct kmem_cache *cachep, int node)
2077 struct kmem_list3 *l3 = cachep->nodelists[node];
2081 struct list_head *p;
2083 p = l3->slabs_free.prev;
2084 if (p == &l3->slabs_free)
2087 slabp = list_entry(l3->slabs_free.prev, struct slab, list);
2092 list_del(&slabp->list);
2094 l3->free_objects -= cachep->num;
2095 spin_unlock_irq(&l3->list_lock);
2096 slab_destroy(cachep, slabp);
2097 spin_lock_irq(&l3->list_lock);
2099 ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial);
2103 static int __cache_shrink(struct kmem_cache *cachep)
2106 struct kmem_list3 *l3;
2108 drain_cpu_caches(cachep);
2111 for_each_online_node(i) {
2112 l3 = cachep->nodelists[i];
2114 spin_lock_irq(&l3->list_lock);
2115 ret += __node_shrink(cachep, i);
2116 spin_unlock_irq(&l3->list_lock);
2119 return (ret ? 1 : 0);
2123 * kmem_cache_shrink - Shrink a cache.
2124 * @cachep: The cache to shrink.
2126 * Releases as many slabs as possible for a cache.
2127 * To help debugging, a zero exit status indicates all slabs were released.
2129 int kmem_cache_shrink(struct kmem_cache *cachep)
2131 if (!cachep || in_interrupt())
2134 return __cache_shrink(cachep);
2136 EXPORT_SYMBOL(kmem_cache_shrink);
2139 * kmem_cache_destroy - delete a cache
2140 * @cachep: the cache to destroy
2142 * Remove a struct kmem_cache object from the slab cache.
2143 * Returns 0 on success.
2145 * It is expected this function will be called by a module when it is
2146 * unloaded. This will remove the cache completely, and avoid a duplicate
2147 * cache being allocated each time a module is loaded and unloaded, if the
2148 * module doesn't have persistent in-kernel storage across loads and unloads.
2150 * The cache must be empty before calling this function.
2152 * The caller must guarantee that noone will allocate memory from the cache
2153 * during the kmem_cache_destroy().
2155 int kmem_cache_destroy(struct kmem_cache *cachep)
2158 struct kmem_list3 *l3;
2160 if (!cachep || in_interrupt())
2163 /* Don't let CPUs to come and go */
2166 /* Find the cache in the chain of caches. */
2167 mutex_lock(&cache_chain_mutex);
2169 * the chain is never empty, cache_cache is never destroyed
2171 list_del(&cachep->next);
2172 mutex_unlock(&cache_chain_mutex);
2174 if (__cache_shrink(cachep)) {
2175 slab_error(cachep, "Can't free all objects");
2176 mutex_lock(&cache_chain_mutex);
2177 list_add(&cachep->next, &cache_chain);
2178 mutex_unlock(&cache_chain_mutex);
2179 unlock_cpu_hotplug();
2183 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2186 for_each_online_cpu(i)
2187 kfree(cachep->array[i]);
2189 /* NUMA: free the list3 structures */
2190 for_each_online_node(i) {
2191 if ((l3 = cachep->nodelists[i])) {
2193 free_alien_cache(l3->alien);
2197 kmem_cache_free(&cache_cache, cachep);
2199 unlock_cpu_hotplug();
2203 EXPORT_SYMBOL(kmem_cache_destroy);
2205 /* Get the memory for a slab management obj. */
2206 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2207 int colour_off, gfp_t local_flags)
2211 if (OFF_SLAB(cachep)) {
2212 /* Slab management obj is off-slab. */
2213 slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags);
2217 slabp = objp + colour_off;
2218 colour_off += cachep->slab_size;
2221 slabp->colouroff = colour_off;
2222 slabp->s_mem = objp + colour_off;
2227 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2229 return (kmem_bufctl_t *) (slabp + 1);
2232 static void cache_init_objs(struct kmem_cache *cachep,
2233 struct slab *slabp, unsigned long ctor_flags)
2237 for (i = 0; i < cachep->num; i++) {
2238 void *objp = slabp->s_mem + cachep->buffer_size * i;
2240 /* need to poison the objs? */
2241 if (cachep->flags & SLAB_POISON)
2242 poison_obj(cachep, objp, POISON_FREE);
2243 if (cachep->flags & SLAB_STORE_USER)
2244 *dbg_userword(cachep, objp) = NULL;
2246 if (cachep->flags & SLAB_RED_ZONE) {
2247 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2248 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2251 * Constructors are not allowed to allocate memory from
2252 * the same cache which they are a constructor for.
2253 * Otherwise, deadlock. They must also be threaded.
2255 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2256 cachep->ctor(objp + obj_offset(cachep), cachep,
2259 if (cachep->flags & SLAB_RED_ZONE) {
2260 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2261 slab_error(cachep, "constructor overwrote the"
2262 " end of an object");
2263 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2264 slab_error(cachep, "constructor overwrote the"
2265 " start of an object");
2267 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)
2268 && cachep->flags & SLAB_POISON)
2269 kernel_map_pages(virt_to_page(objp),
2270 cachep->buffer_size / PAGE_SIZE, 0);
2273 cachep->ctor(objp, cachep, ctor_flags);
2275 slab_bufctl(slabp)[i] = i + 1;
2277 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2281 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2283 if (flags & SLAB_DMA) {
2284 if (!(cachep->gfpflags & GFP_DMA))
2287 if (cachep->gfpflags & GFP_DMA)
2292 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, int nodeid)
2294 void *objp = slabp->s_mem + (slabp->free * cachep->buffer_size);
2298 next = slab_bufctl(slabp)[slabp->free];
2300 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2301 WARN_ON(slabp->nodeid != nodeid);
2308 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, void *objp,
2311 unsigned int objnr = (unsigned)(objp-slabp->s_mem) / cachep->buffer_size;
2314 /* Verify that the slab belongs to the intended node */
2315 WARN_ON(slabp->nodeid != nodeid);
2317 if (slab_bufctl(slabp)[objnr] != BUFCTL_FREE) {
2318 printk(KERN_ERR "slab: double free detected in cache "
2319 "'%s', objp %p\n", cachep->name, objp);
2323 slab_bufctl(slabp)[objnr] = slabp->free;
2324 slabp->free = objnr;
2328 static void set_slab_attr(struct kmem_cache *cachep, struct slab *slabp, void *objp)
2333 /* Nasty!!!!!! I hope this is OK. */
2334 i = 1 << cachep->gfporder;
2335 page = virt_to_page(objp);
2337 page_set_cache(page, cachep);
2338 page_set_slab(page, slabp);
2344 * Grow (by 1) the number of slabs within a cache. This is called by
2345 * kmem_cache_alloc() when there are no active objs left in a cache.
2347 static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid)
2353 unsigned long ctor_flags;
2354 struct kmem_list3 *l3;
2356 /* Be lazy and only check for valid flags here,
2357 * keeping it out of the critical path in kmem_cache_alloc().
2359 if (flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW))
2361 if (flags & SLAB_NO_GROW)
2364 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2365 local_flags = (flags & SLAB_LEVEL_MASK);
2366 if (!(local_flags & __GFP_WAIT))
2368 * Not allowed to sleep. Need to tell a constructor about
2369 * this - it might need to know...
2371 ctor_flags |= SLAB_CTOR_ATOMIC;
2373 /* Take the l3 list lock to change the colour_next on this node */
2375 l3 = cachep->nodelists[nodeid];
2376 spin_lock(&l3->list_lock);
2378 /* Get colour for the slab, and cal the next value. */
2379 offset = l3->colour_next;
2381 if (l3->colour_next >= cachep->colour)
2382 l3->colour_next = 0;
2383 spin_unlock(&l3->list_lock);
2385 offset *= cachep->colour_off;
2387 if (local_flags & __GFP_WAIT)
2391 * The test for missing atomic flag is performed here, rather than
2392 * the more obvious place, simply to reduce the critical path length
2393 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2394 * will eventually be caught here (where it matters).
2396 kmem_flagcheck(cachep, flags);
2398 /* Get mem for the objs.
2399 * Attempt to allocate a physical page from 'nodeid',
2401 if (!(objp = kmem_getpages(cachep, flags, nodeid)))
2404 /* Get slab management. */
2405 if (!(slabp = alloc_slabmgmt(cachep, objp, offset, local_flags)))
2408 slabp->nodeid = nodeid;
2409 set_slab_attr(cachep, slabp, objp);
2411 cache_init_objs(cachep, slabp, ctor_flags);
2413 if (local_flags & __GFP_WAIT)
2414 local_irq_disable();
2416 spin_lock(&l3->list_lock);
2418 /* Make slab active. */
2419 list_add_tail(&slabp->list, &(l3->slabs_free));
2420 STATS_INC_GROWN(cachep);
2421 l3->free_objects += cachep->num;
2422 spin_unlock(&l3->list_lock);
2425 kmem_freepages(cachep, objp);
2427 if (local_flags & __GFP_WAIT)
2428 local_irq_disable();
2435 * Perform extra freeing checks:
2436 * - detect bad pointers.
2437 * - POISON/RED_ZONE checking
2438 * - destructor calls, for caches with POISON+dtor
2440 static void kfree_debugcheck(const void *objp)
2444 if (!virt_addr_valid(objp)) {
2445 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2446 (unsigned long)objp);
2449 page = virt_to_page(objp);
2450 if (!PageSlab(page)) {
2451 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2452 (unsigned long)objp);
2457 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2464 objp -= obj_offset(cachep);
2465 kfree_debugcheck(objp);
2466 page = virt_to_page(objp);
2468 if (page_get_cache(page) != cachep) {
2470 "mismatch in kmem_cache_free: expected cache %p, got %p\n",
2471 page_get_cache(page), cachep);
2472 printk(KERN_ERR "%p is %s.\n", cachep, cachep->name);
2473 printk(KERN_ERR "%p is %s.\n", page_get_cache(page),
2474 page_get_cache(page)->name);
2477 slabp = page_get_slab(page);
2479 if (cachep->flags & SLAB_RED_ZONE) {
2480 if (*dbg_redzone1(cachep, objp) != RED_ACTIVE
2481 || *dbg_redzone2(cachep, objp) != RED_ACTIVE) {
2483 "double free, or memory outside"
2484 " object was overwritten");
2486 "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2487 objp, *dbg_redzone1(cachep, objp),
2488 *dbg_redzone2(cachep, objp));
2490 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2491 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2493 if (cachep->flags & SLAB_STORE_USER)
2494 *dbg_userword(cachep, objp) = caller;
2496 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
2498 BUG_ON(objnr >= cachep->num);
2499 BUG_ON(objp != slabp->s_mem + objnr * cachep->buffer_size);
2501 if (cachep->flags & SLAB_DEBUG_INITIAL) {
2502 /* Need to call the slab's constructor so the
2503 * caller can perform a verify of its state (debugging).
2504 * Called without the cache-lock held.
2506 cachep->ctor(objp + obj_offset(cachep),
2507 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
2509 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2510 /* we want to cache poison the object,
2511 * call the destruction callback
2513 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
2515 if (cachep->flags & SLAB_POISON) {
2516 #ifdef CONFIG_DEBUG_PAGEALLOC
2517 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) {
2518 store_stackinfo(cachep, objp, (unsigned long)caller);
2519 kernel_map_pages(virt_to_page(objp),
2520 cachep->buffer_size / PAGE_SIZE, 0);
2522 poison_obj(cachep, objp, POISON_FREE);
2525 poison_obj(cachep, objp, POISON_FREE);
2531 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2536 /* Check slab's freelist to see if this obj is there. */
2537 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2539 if (entries > cachep->num || i >= cachep->num)
2542 if (entries != cachep->num - slabp->inuse) {
2545 "slab: Internal list corruption detected in cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2546 cachep->name, cachep->num, slabp, slabp->inuse);
2548 i < sizeof(slabp) + cachep->num * sizeof(kmem_bufctl_t);
2551 printk("\n%03x:", i);
2552 printk(" %02x", ((unsigned char *)slabp)[i]);
2559 #define kfree_debugcheck(x) do { } while(0)
2560 #define cache_free_debugcheck(x,objp,z) (objp)
2561 #define check_slabp(x,y) do { } while(0)
2564 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2567 struct kmem_list3 *l3;
2568 struct array_cache *ac;
2571 ac = cpu_cache_get(cachep);
2573 batchcount = ac->batchcount;
2574 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2575 /* if there was little recent activity on this
2576 * cache, then perform only a partial refill.
2577 * Otherwise we could generate refill bouncing.
2579 batchcount = BATCHREFILL_LIMIT;
2581 l3 = cachep->nodelists[numa_node_id()];
2583 BUG_ON(ac->avail > 0 || !l3);
2584 spin_lock(&l3->list_lock);
2587 struct array_cache *shared_array = l3->shared;
2588 if (shared_array->avail) {
2589 if (batchcount > shared_array->avail)
2590 batchcount = shared_array->avail;
2591 shared_array->avail -= batchcount;
2592 ac->avail = batchcount;
2594 &(shared_array->entry[shared_array->avail]),
2595 sizeof(void *) * batchcount);
2596 shared_array->touched = 1;
2600 while (batchcount > 0) {
2601 struct list_head *entry;
2603 /* Get slab alloc is to come from. */
2604 entry = l3->slabs_partial.next;
2605 if (entry == &l3->slabs_partial) {
2606 l3->free_touched = 1;
2607 entry = l3->slabs_free.next;
2608 if (entry == &l3->slabs_free)
2612 slabp = list_entry(entry, struct slab, list);
2613 check_slabp(cachep, slabp);
2614 check_spinlock_acquired(cachep);
2615 while (slabp->inuse < cachep->num && batchcount--) {
2616 STATS_INC_ALLOCED(cachep);
2617 STATS_INC_ACTIVE(cachep);
2618 STATS_SET_HIGH(cachep);
2620 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2623 check_slabp(cachep, slabp);
2625 /* move slabp to correct slabp list: */
2626 list_del(&slabp->list);
2627 if (slabp->free == BUFCTL_END)
2628 list_add(&slabp->list, &l3->slabs_full);
2630 list_add(&slabp->list, &l3->slabs_partial);
2634 l3->free_objects -= ac->avail;
2636 spin_unlock(&l3->list_lock);
2638 if (unlikely(!ac->avail)) {
2640 x = cache_grow(cachep, flags, numa_node_id());
2642 // cache_grow can reenable interrupts, then ac could change.
2643 ac = cpu_cache_get(cachep);
2644 if (!x && ac->avail == 0) // no objects in sight? abort
2647 if (!ac->avail) // objects refilled by interrupt?
2651 return ac->entry[--ac->avail];
2655 cache_alloc_debugcheck_before(struct kmem_cache *cachep, gfp_t flags)
2657 might_sleep_if(flags & __GFP_WAIT);
2659 kmem_flagcheck(cachep, flags);
2664 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, gfp_t flags,
2665 void *objp, void *caller)
2669 if (cachep->flags & SLAB_POISON) {
2670 #ifdef CONFIG_DEBUG_PAGEALLOC
2671 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
2672 kernel_map_pages(virt_to_page(objp),
2673 cachep->buffer_size / PAGE_SIZE, 1);
2675 check_poison_obj(cachep, objp);
2677 check_poison_obj(cachep, objp);
2679 poison_obj(cachep, objp, POISON_INUSE);
2681 if (cachep->flags & SLAB_STORE_USER)
2682 *dbg_userword(cachep, objp) = caller;
2684 if (cachep->flags & SLAB_RED_ZONE) {
2685 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE
2686 || *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2688 "double free, or memory outside"
2689 " object was overwritten");
2691 "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2692 objp, *dbg_redzone1(cachep, objp),
2693 *dbg_redzone2(cachep, objp));
2695 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2696 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2698 objp += obj_offset(cachep);
2699 if (cachep->ctor && cachep->flags & SLAB_POISON) {
2700 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2702 if (!(flags & __GFP_WAIT))
2703 ctor_flags |= SLAB_CTOR_ATOMIC;
2705 cachep->ctor(objp, cachep, ctor_flags);
2710 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2713 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
2716 struct array_cache *ac;
2719 if (unlikely(current->mempolicy && !in_interrupt())) {
2720 int nid = slab_node(current->mempolicy);
2722 if (nid != numa_node_id())
2723 return __cache_alloc_node(cachep, flags, nid);
2728 ac = cpu_cache_get(cachep);
2729 if (likely(ac->avail)) {
2730 STATS_INC_ALLOCHIT(cachep);
2732 objp = ac->entry[--ac->avail];
2734 STATS_INC_ALLOCMISS(cachep);
2735 objp = cache_alloc_refill(cachep, flags);
2740 static __always_inline void *
2741 __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
2743 unsigned long save_flags;
2746 cache_alloc_debugcheck_before(cachep, flags);
2748 local_irq_save(save_flags);
2749 objp = ____cache_alloc(cachep, flags);
2750 local_irq_restore(save_flags);
2751 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
2759 * A interface to enable slab creation on nodeid
2761 static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
2763 struct list_head *entry;
2765 struct kmem_list3 *l3;
2769 l3 = cachep->nodelists[nodeid];
2774 spin_lock(&l3->list_lock);
2775 entry = l3->slabs_partial.next;
2776 if (entry == &l3->slabs_partial) {
2777 l3->free_touched = 1;
2778 entry = l3->slabs_free.next;
2779 if (entry == &l3->slabs_free)
2783 slabp = list_entry(entry, struct slab, list);
2784 check_spinlock_acquired_node(cachep, nodeid);
2785 check_slabp(cachep, slabp);
2787 STATS_INC_NODEALLOCS(cachep);
2788 STATS_INC_ACTIVE(cachep);
2789 STATS_SET_HIGH(cachep);
2791 BUG_ON(slabp->inuse == cachep->num);
2793 obj = slab_get_obj(cachep, slabp, nodeid);
2794 check_slabp(cachep, slabp);
2796 /* move slabp to correct slabp list: */
2797 list_del(&slabp->list);
2799 if (slabp->free == BUFCTL_END) {
2800 list_add(&slabp->list, &l3->slabs_full);
2802 list_add(&slabp->list, &l3->slabs_partial);
2805 spin_unlock(&l3->list_lock);
2809 spin_unlock(&l3->list_lock);
2810 x = cache_grow(cachep, flags, nodeid);
2822 * Caller needs to acquire correct kmem_list's list_lock
2824 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
2828 struct kmem_list3 *l3;
2830 for (i = 0; i < nr_objects; i++) {
2831 void *objp = objpp[i];
2834 slabp = virt_to_slab(objp);
2835 l3 = cachep->nodelists[node];
2836 list_del(&slabp->list);
2837 check_spinlock_acquired_node(cachep, node);
2838 check_slabp(cachep, slabp);
2839 slab_put_obj(cachep, slabp, objp, node);
2840 STATS_DEC_ACTIVE(cachep);
2842 check_slabp(cachep, slabp);
2844 /* fixup slab chains */
2845 if (slabp->inuse == 0) {
2846 if (l3->free_objects > l3->free_limit) {
2847 l3->free_objects -= cachep->num;
2848 slab_destroy(cachep, slabp);
2850 list_add(&slabp->list, &l3->slabs_free);
2853 /* Unconditionally move a slab to the end of the
2854 * partial list on free - maximum time for the
2855 * other objects to be freed, too.
2857 list_add_tail(&slabp->list, &l3->slabs_partial);
2862 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
2865 struct kmem_list3 *l3;
2866 int node = numa_node_id();
2868 batchcount = ac->batchcount;
2870 BUG_ON(!batchcount || batchcount > ac->avail);
2873 l3 = cachep->nodelists[node];
2874 spin_lock(&l3->list_lock);
2876 struct array_cache *shared_array = l3->shared;
2877 int max = shared_array->limit - shared_array->avail;
2879 if (batchcount > max)
2881 memcpy(&(shared_array->entry[shared_array->avail]),
2882 ac->entry, sizeof(void *) * batchcount);
2883 shared_array->avail += batchcount;
2888 free_block(cachep, ac->entry, batchcount, node);
2893 struct list_head *p;
2895 p = l3->slabs_free.next;
2896 while (p != &(l3->slabs_free)) {
2899 slabp = list_entry(p, struct slab, list);
2900 BUG_ON(slabp->inuse);
2905 STATS_SET_FREEABLE(cachep, i);
2908 spin_unlock(&l3->list_lock);
2909 ac->avail -= batchcount;
2910 memmove(ac->entry, &(ac->entry[batchcount]),
2911 sizeof(void *) * ac->avail);
2916 * Release an obj back to its cache. If the obj has a constructed
2917 * state, it must be in this state _before_ it is released.
2919 * Called with disabled ints.
2921 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
2923 struct array_cache *ac = cpu_cache_get(cachep);
2926 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
2928 /* Make sure we are not freeing a object from another
2929 * node to the array cache on this cpu.
2934 slabp = virt_to_slab(objp);
2935 if (unlikely(slabp->nodeid != numa_node_id())) {
2936 struct array_cache *alien = NULL;
2937 int nodeid = slabp->nodeid;
2938 struct kmem_list3 *l3 =
2939 cachep->nodelists[numa_node_id()];
2941 STATS_INC_NODEFREES(cachep);
2942 if (l3->alien && l3->alien[nodeid]) {
2943 alien = l3->alien[nodeid];
2944 spin_lock(&alien->lock);
2945 if (unlikely(alien->avail == alien->limit))
2946 __drain_alien_cache(cachep,
2948 alien->entry[alien->avail++] = objp;
2949 spin_unlock(&alien->lock);
2951 spin_lock(&(cachep->nodelists[nodeid])->
2953 free_block(cachep, &objp, 1, nodeid);
2954 spin_unlock(&(cachep->nodelists[nodeid])->
2961 if (likely(ac->avail < ac->limit)) {
2962 STATS_INC_FREEHIT(cachep);
2963 ac->entry[ac->avail++] = objp;
2966 STATS_INC_FREEMISS(cachep);
2967 cache_flusharray(cachep, ac);
2968 ac->entry[ac->avail++] = objp;
2973 * kmem_cache_alloc - Allocate an object
2974 * @cachep: The cache to allocate from.
2975 * @flags: See kmalloc().
2977 * Allocate an object from this cache. The flags are only relevant
2978 * if the cache has no available objects.
2980 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
2982 return __cache_alloc(cachep, flags, __builtin_return_address(0));
2984 EXPORT_SYMBOL(kmem_cache_alloc);
2987 * kmem_ptr_validate - check if an untrusted pointer might
2989 * @cachep: the cache we're checking against
2990 * @ptr: pointer to validate
2992 * This verifies that the untrusted pointer looks sane:
2993 * it is _not_ a guarantee that the pointer is actually
2994 * part of the slab cache in question, but it at least
2995 * validates that the pointer can be dereferenced and
2996 * looks half-way sane.
2998 * Currently only used for dentry validation.
3000 int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr)
3002 unsigned long addr = (unsigned long)ptr;
3003 unsigned long min_addr = PAGE_OFFSET;
3004 unsigned long align_mask = BYTES_PER_WORD - 1;
3005 unsigned long size = cachep->buffer_size;
3008 if (unlikely(addr < min_addr))
3010 if (unlikely(addr > (unsigned long)high_memory - size))
3012 if (unlikely(addr & align_mask))
3014 if (unlikely(!kern_addr_valid(addr)))
3016 if (unlikely(!kern_addr_valid(addr + size - 1)))
3018 page = virt_to_page(ptr);
3019 if (unlikely(!PageSlab(page)))
3021 if (unlikely(page_get_cache(page) != cachep))
3030 * kmem_cache_alloc_node - Allocate an object on the specified node
3031 * @cachep: The cache to allocate from.
3032 * @flags: See kmalloc().
3033 * @nodeid: node number of the target node.
3035 * Identical to kmem_cache_alloc, except that this function is slow
3036 * and can sleep. And it will allocate memory on the given node, which
3037 * can improve the performance for cpu bound structures.
3038 * New and improved: it will now make sure that the object gets
3039 * put on the correct node list so that there is no false sharing.
3041 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3043 unsigned long save_flags;
3046 cache_alloc_debugcheck_before(cachep, flags);
3047 local_irq_save(save_flags);
3049 if (nodeid == -1 || nodeid == numa_node_id() ||
3050 !cachep->nodelists[nodeid])
3051 ptr = ____cache_alloc(cachep, flags);
3053 ptr = __cache_alloc_node(cachep, flags, nodeid);
3054 local_irq_restore(save_flags);
3056 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr,
3057 __builtin_return_address(0));
3061 EXPORT_SYMBOL(kmem_cache_alloc_node);
3063 void *kmalloc_node(size_t size, gfp_t flags, int node)
3065 struct kmem_cache *cachep;
3067 cachep = kmem_find_general_cachep(size, flags);
3068 if (unlikely(cachep == NULL))
3070 return kmem_cache_alloc_node(cachep, flags, node);
3072 EXPORT_SYMBOL(kmalloc_node);
3076 * kmalloc - allocate memory
3077 * @size: how many bytes of memory are required.
3078 * @flags: the type of memory to allocate.
3080 * kmalloc is the normal method of allocating memory
3083 * The @flags argument may be one of:
3085 * %GFP_USER - Allocate memory on behalf of user. May sleep.
3087 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
3089 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
3091 * Additionally, the %GFP_DMA flag may be set to indicate the memory
3092 * must be suitable for DMA. This can mean different things on different
3093 * platforms. For example, on i386, it means that the memory must come
3094 * from the first 16MB.
3096 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3099 struct kmem_cache *cachep;
3101 /* If you want to save a few bytes .text space: replace
3103 * Then kmalloc uses the uninlined functions instead of the inline
3106 cachep = __find_general_cachep(size, flags);
3107 if (unlikely(cachep == NULL))
3109 return __cache_alloc(cachep, flags, caller);
3112 #ifndef CONFIG_DEBUG_SLAB
3114 void *__kmalloc(size_t size, gfp_t flags)
3116 return __do_kmalloc(size, flags, NULL);
3118 EXPORT_SYMBOL(__kmalloc);
3122 void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3124 return __do_kmalloc(size, flags, caller);
3126 EXPORT_SYMBOL(__kmalloc_track_caller);
3132 * __alloc_percpu - allocate one copy of the object for every present
3133 * cpu in the system, zeroing them.
3134 * Objects should be dereferenced using the per_cpu_ptr macro only.
3136 * @size: how many bytes of memory are required.
3138 void *__alloc_percpu(size_t size)
3141 struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL);
3147 * Cannot use for_each_online_cpu since a cpu may come online
3148 * and we have no way of figuring out how to fix the array
3149 * that we have allocated then....
3152 int node = cpu_to_node(i);
3154 if (node_online(node))
3155 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
3157 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
3159 if (!pdata->ptrs[i])
3161 memset(pdata->ptrs[i], 0, size);
3164 /* Catch derefs w/o wrappers */
3165 return (void *)(~(unsigned long)pdata);
3169 if (!cpu_possible(i))
3171 kfree(pdata->ptrs[i]);
3176 EXPORT_SYMBOL(__alloc_percpu);
3180 * kmem_cache_free - Deallocate an object
3181 * @cachep: The cache the allocation was from.
3182 * @objp: The previously allocated object.
3184 * Free an object which was previously allocated from this
3187 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3189 unsigned long flags;
3191 local_irq_save(flags);
3192 __cache_free(cachep, objp);
3193 local_irq_restore(flags);
3195 EXPORT_SYMBOL(kmem_cache_free);
3198 * kfree - free previously allocated memory
3199 * @objp: pointer returned by kmalloc.
3201 * If @objp is NULL, no operation is performed.
3203 * Don't free memory not originally allocated by kmalloc()
3204 * or you will run into trouble.
3206 void kfree(const void *objp)
3208 struct kmem_cache *c;
3209 unsigned long flags;
3211 if (unlikely(!objp))
3213 local_irq_save(flags);
3214 kfree_debugcheck(objp);
3215 c = virt_to_cache(objp);
3216 mutex_debug_check_no_locks_freed(objp, obj_size(c));
3217 __cache_free(c, (void *)objp);
3218 local_irq_restore(flags);
3220 EXPORT_SYMBOL(kfree);
3224 * free_percpu - free previously allocated percpu memory
3225 * @objp: pointer returned by alloc_percpu.
3227 * Don't free memory not originally allocated by alloc_percpu()
3228 * The complemented objp is to check for that.
3230 void free_percpu(const void *objp)
3233 struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp);
3236 * We allocate for all cpus so we cannot use for online cpu here.
3242 EXPORT_SYMBOL(free_percpu);
3245 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3247 return obj_size(cachep);
3249 EXPORT_SYMBOL(kmem_cache_size);
3251 const char *kmem_cache_name(struct kmem_cache *cachep)
3253 return cachep->name;
3255 EXPORT_SYMBOL_GPL(kmem_cache_name);
3258 * This initializes kmem_list3 for all nodes.
3260 static int alloc_kmemlist(struct kmem_cache *cachep)
3263 struct kmem_list3 *l3;
3266 for_each_online_node(node) {
3267 struct array_cache *nc = NULL, *new;
3268 struct array_cache **new_alien = NULL;
3270 if (!(new_alien = alloc_alien_cache(node, cachep->limit)))
3273 if (!(new = alloc_arraycache(node, (cachep->shared *
3274 cachep->batchcount),
3277 if ((l3 = cachep->nodelists[node])) {
3279 spin_lock_irq(&l3->list_lock);
3281 if ((nc = cachep->nodelists[node]->shared))
3282 free_block(cachep, nc->entry, nc->avail, node);
3285 if (!cachep->nodelists[node]->alien) {
3286 l3->alien = new_alien;
3289 l3->free_limit = (1 + nr_cpus_node(node)) *
3290 cachep->batchcount + cachep->num;
3291 spin_unlock_irq(&l3->list_lock);
3293 free_alien_cache(new_alien);
3296 if (!(l3 = kmalloc_node(sizeof(struct kmem_list3),
3300 kmem_list3_init(l3);
3301 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3302 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3304 l3->alien = new_alien;
3305 l3->free_limit = (1 + nr_cpus_node(node)) *
3306 cachep->batchcount + cachep->num;
3307 cachep->nodelists[node] = l3;
3315 struct ccupdate_struct {
3316 struct kmem_cache *cachep;
3317 struct array_cache *new[NR_CPUS];
3320 static void do_ccupdate_local(void *info)
3322 struct ccupdate_struct *new = (struct ccupdate_struct *)info;
3323 struct array_cache *old;
3326 old = cpu_cache_get(new->cachep);
3328 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3329 new->new[smp_processor_id()] = old;
3332 static int do_tune_cpucache(struct kmem_cache *cachep, int limit, int batchcount,
3335 struct ccupdate_struct new;
3338 memset(&new.new, 0, sizeof(new.new));
3339 for_each_online_cpu(i) {
3341 alloc_arraycache(cpu_to_node(i), limit, batchcount);
3343 for (i--; i >= 0; i--)
3348 new.cachep = cachep;
3350 smp_call_function_all_cpus(do_ccupdate_local, (void *)&new);
3353 spin_lock(&cachep->spinlock);
3354 cachep->batchcount = batchcount;
3355 cachep->limit = limit;
3356 cachep->shared = shared;
3357 spin_unlock(&cachep->spinlock);
3359 for_each_online_cpu(i) {
3360 struct array_cache *ccold = new.new[i];
3363 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3364 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3365 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3369 err = alloc_kmemlist(cachep);
3371 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
3372 cachep->name, -err);
3378 static void enable_cpucache(struct kmem_cache *cachep)
3383 /* The head array serves three purposes:
3384 * - create a LIFO ordering, i.e. return objects that are cache-warm
3385 * - reduce the number of spinlock operations.
3386 * - reduce the number of linked list operations on the slab and
3387 * bufctl chains: array operations are cheaper.
3388 * The numbers are guessed, we should auto-tune as described by
3391 if (cachep->buffer_size > 131072)
3393 else if (cachep->buffer_size > PAGE_SIZE)
3395 else if (cachep->buffer_size > 1024)
3397 else if (cachep->buffer_size > 256)
3402 /* Cpu bound tasks (e.g. network routing) can exhibit cpu bound
3403 * allocation behaviour: Most allocs on one cpu, most free operations
3404 * on another cpu. For these cases, an efficient object passing between
3405 * cpus is necessary. This is provided by a shared array. The array
3406 * replaces Bonwick's magazine layer.
3407 * On uniprocessor, it's functionally equivalent (but less efficient)
3408 * to a larger limit. Thus disabled by default.
3412 if (cachep->buffer_size <= PAGE_SIZE)
3417 /* With debugging enabled, large batchcount lead to excessively
3418 * long periods with disabled local interrupts. Limit the
3424 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
3426 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3427 cachep->name, -err);
3430 static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac,
3431 int force, int node)
3435 check_spinlock_acquired_node(cachep, node);
3436 if (ac->touched && !force) {
3438 } else if (ac->avail) {
3439 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3440 if (tofree > ac->avail) {
3441 tofree = (ac->avail + 1) / 2;
3443 free_block(cachep, ac->entry, tofree, node);
3444 ac->avail -= tofree;
3445 memmove(ac->entry, &(ac->entry[tofree]),
3446 sizeof(void *) * ac->avail);
3451 * cache_reap - Reclaim memory from caches.
3452 * @unused: unused parameter
3454 * Called from workqueue/eventd every few seconds.
3456 * - clear the per-cpu caches for this CPU.
3457 * - return freeable pages to the main free memory pool.
3459 * If we cannot acquire the cache chain mutex then just give up - we'll
3460 * try again on the next iteration.
3462 static void cache_reap(void *unused)
3464 struct list_head *walk;
3465 struct kmem_list3 *l3;
3467 if (!mutex_trylock(&cache_chain_mutex)) {
3468 /* Give up. Setup the next iteration. */
3469 schedule_delayed_work(&__get_cpu_var(reap_work),
3474 list_for_each(walk, &cache_chain) {
3475 struct kmem_cache *searchp;
3476 struct list_head *p;
3480 searchp = list_entry(walk, struct kmem_cache, next);
3482 if (searchp->flags & SLAB_NO_REAP)
3487 l3 = searchp->nodelists[numa_node_id()];
3489 drain_alien_cache(searchp, l3->alien);
3490 spin_lock_irq(&l3->list_lock);
3492 drain_array_locked(searchp, cpu_cache_get(searchp), 0,
3495 if (time_after(l3->next_reap, jiffies))
3498 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
3501 drain_array_locked(searchp, l3->shared, 0,
3504 if (l3->free_touched) {
3505 l3->free_touched = 0;
3510 (l3->free_limit + 5 * searchp->num -
3511 1) / (5 * searchp->num);
3513 p = l3->slabs_free.next;
3514 if (p == &(l3->slabs_free))
3517 slabp = list_entry(p, struct slab, list);
3518 BUG_ON(slabp->inuse);
3519 list_del(&slabp->list);
3520 STATS_INC_REAPED(searchp);
3522 /* Safe to drop the lock. The slab is no longer
3523 * linked to the cache.
3524 * searchp cannot disappear, we hold
3527 l3->free_objects -= searchp->num;
3528 spin_unlock_irq(&l3->list_lock);
3529 slab_destroy(searchp, slabp);
3530 spin_lock_irq(&l3->list_lock);
3531 } while (--tofree > 0);
3533 spin_unlock_irq(&l3->list_lock);
3538 mutex_unlock(&cache_chain_mutex);
3539 drain_remote_pages();
3540 /* Setup the next iteration */
3541 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
3544 #ifdef CONFIG_PROC_FS
3546 static void print_slabinfo_header(struct seq_file *m)
3549 * Output format version, so at least we can change it
3550 * without _too_ many complaints.
3553 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3555 seq_puts(m, "slabinfo - version: 2.1\n");
3557 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3558 "<objperslab> <pagesperslab>");
3559 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3560 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3562 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3563 "<error> <maxfreeable> <nodeallocs> <remotefrees>");
3564 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3569 static void *s_start(struct seq_file *m, loff_t *pos)
3572 struct list_head *p;
3574 mutex_lock(&cache_chain_mutex);
3576 print_slabinfo_header(m);
3577 p = cache_chain.next;
3580 if (p == &cache_chain)
3583 return list_entry(p, struct kmem_cache, next);
3586 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3588 struct kmem_cache *cachep = p;
3590 return cachep->next.next == &cache_chain ? NULL
3591 : list_entry(cachep->next.next, struct kmem_cache, next);
3594 static void s_stop(struct seq_file *m, void *p)
3596 mutex_unlock(&cache_chain_mutex);
3599 static int s_show(struct seq_file *m, void *p)
3601 struct kmem_cache *cachep = p;
3602 struct list_head *q;
3604 unsigned long active_objs;
3605 unsigned long num_objs;
3606 unsigned long active_slabs = 0;
3607 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3611 struct kmem_list3 *l3;
3613 spin_lock(&cachep->spinlock);
3616 for_each_online_node(node) {
3617 l3 = cachep->nodelists[node];
3622 spin_lock_irq(&l3->list_lock);
3624 list_for_each(q, &l3->slabs_full) {
3625 slabp = list_entry(q, struct slab, list);
3626 if (slabp->inuse != cachep->num && !error)
3627 error = "slabs_full accounting error";
3628 active_objs += cachep->num;
3631 list_for_each(q, &l3->slabs_partial) {
3632 slabp = list_entry(q, struct slab, list);
3633 if (slabp->inuse == cachep->num && !error)
3634 error = "slabs_partial inuse accounting error";
3635 if (!slabp->inuse && !error)
3636 error = "slabs_partial/inuse accounting error";
3637 active_objs += slabp->inuse;
3640 list_for_each(q, &l3->slabs_free) {
3641 slabp = list_entry(q, struct slab, list);
3642 if (slabp->inuse && !error)
3643 error = "slabs_free/inuse accounting error";
3646 free_objects += l3->free_objects;
3648 shared_avail += l3->shared->avail;
3650 spin_unlock_irq(&l3->list_lock);
3652 num_slabs += active_slabs;
3653 num_objs = num_slabs * cachep->num;
3654 if (num_objs - active_objs != free_objects && !error)
3655 error = "free_objects accounting error";
3657 name = cachep->name;
3659 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3661 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
3662 name, active_objs, num_objs, cachep->buffer_size,
3663 cachep->num, (1 << cachep->gfporder));
3664 seq_printf(m, " : tunables %4u %4u %4u",
3665 cachep->limit, cachep->batchcount, cachep->shared);
3666 seq_printf(m, " : slabdata %6lu %6lu %6lu",
3667 active_slabs, num_slabs, shared_avail);
3670 unsigned long high = cachep->high_mark;
3671 unsigned long allocs = cachep->num_allocations;
3672 unsigned long grown = cachep->grown;
3673 unsigned long reaped = cachep->reaped;
3674 unsigned long errors = cachep->errors;
3675 unsigned long max_freeable = cachep->max_freeable;
3676 unsigned long node_allocs = cachep->node_allocs;
3677 unsigned long node_frees = cachep->node_frees;
3679 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
3680 %4lu %4lu %4lu %4lu", allocs, high, grown, reaped, errors, max_freeable, node_allocs, node_frees);
3684 unsigned long allochit = atomic_read(&cachep->allochit);
3685 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3686 unsigned long freehit = atomic_read(&cachep->freehit);
3687 unsigned long freemiss = atomic_read(&cachep->freemiss);
3689 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
3690 allochit, allocmiss, freehit, freemiss);
3694 spin_unlock(&cachep->spinlock);
3699 * slabinfo_op - iterator that generates /proc/slabinfo
3708 * num-pages-per-slab
3709 * + further values on SMP and with statistics enabled
3712 struct seq_operations slabinfo_op = {
3719 #define MAX_SLABINFO_WRITE 128
3721 * slabinfo_write - Tuning for the slab allocator
3723 * @buffer: user buffer
3724 * @count: data length
3727 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
3728 size_t count, loff_t *ppos)
3730 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
3731 int limit, batchcount, shared, res;
3732 struct list_head *p;
3734 if (count > MAX_SLABINFO_WRITE)
3736 if (copy_from_user(&kbuf, buffer, count))
3738 kbuf[MAX_SLABINFO_WRITE] = '\0';
3740 tmp = strchr(kbuf, ' ');
3745 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3748 /* Find the cache in the chain of caches. */
3749 mutex_lock(&cache_chain_mutex);
3751 list_for_each(p, &cache_chain) {
3752 struct kmem_cache *cachep = list_entry(p, struct kmem_cache,
3755 if (!strcmp(cachep->name, kbuf)) {
3758 batchcount > limit || shared < 0) {
3761 res = do_tune_cpucache(cachep, limit,
3762 batchcount, shared);
3767 mutex_unlock(&cache_chain_mutex);
3775 * ksize - get the actual amount of memory allocated for a given object
3776 * @objp: Pointer to the object
3778 * kmalloc may internally round up allocations and return more memory
3779 * than requested. ksize() can be used to determine the actual amount of
3780 * memory allocated. The caller may use this additional memory, even though
3781 * a smaller amount of memory was initially specified with the kmalloc call.
3782 * The caller must guarantee that objp points to a valid object previously
3783 * allocated with either kmalloc() or kmem_cache_alloc(). The object
3784 * must not be freed during the duration of the call.
3786 unsigned int ksize(const void *objp)
3788 if (unlikely(objp == NULL))
3791 return obj_size(virt_to_cache(objp));