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 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
615 return slab->s_mem + cache->buffer_size * idx;
618 static inline unsigned int obj_to_index(struct kmem_cache *cache,
619 struct slab *slab, void *obj)
621 return (unsigned)(obj - slab->s_mem) / cache->buffer_size;
624 /* These are the default caches for kmalloc. Custom caches can have other sizes. */
625 struct cache_sizes malloc_sizes[] = {
626 #define CACHE(x) { .cs_size = (x) },
627 #include <linux/kmalloc_sizes.h>
631 EXPORT_SYMBOL(malloc_sizes);
633 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
639 static struct cache_names __initdata cache_names[] = {
640 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
641 #include <linux/kmalloc_sizes.h>
646 static struct arraycache_init initarray_cache __initdata =
647 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
648 static struct arraycache_init initarray_generic =
649 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
651 /* internal cache of cache description objs */
652 static struct kmem_cache cache_cache = {
654 .limit = BOOT_CPUCACHE_ENTRIES,
656 .buffer_size = sizeof(struct kmem_cache),
657 .flags = SLAB_NO_REAP,
658 .spinlock = SPIN_LOCK_UNLOCKED,
659 .name = "kmem_cache",
661 .obj_size = sizeof(struct kmem_cache),
665 /* Guard access to the cache-chain. */
666 static DEFINE_MUTEX(cache_chain_mutex);
667 static struct list_head cache_chain;
670 * vm_enough_memory() looks at this to determine how many
671 * slab-allocated pages are possibly freeable under pressure
673 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
675 atomic_t slab_reclaim_pages;
678 * chicken and egg problem: delay the per-cpu array allocation
679 * until the general caches are up.
688 static DEFINE_PER_CPU(struct work_struct, reap_work);
690 static void free_block(struct kmem_cache *cachep, void **objpp, int len, int node);
691 static void enable_cpucache(struct kmem_cache *cachep);
692 static void cache_reap(void *unused);
693 static int __node_shrink(struct kmem_cache *cachep, int node);
695 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
697 return cachep->array[smp_processor_id()];
700 static inline struct kmem_cache *__find_general_cachep(size_t size, gfp_t gfpflags)
702 struct cache_sizes *csizep = malloc_sizes;
705 /* This happens if someone tries to call
706 * kmem_cache_create(), or __kmalloc(), before
707 * the generic caches are initialized.
709 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
711 while (size > csizep->cs_size)
715 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
716 * has cs_{dma,}cachep==NULL. Thus no special case
717 * for large kmalloc calls required.
719 if (unlikely(gfpflags & GFP_DMA))
720 return csizep->cs_dmacachep;
721 return csizep->cs_cachep;
724 struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
726 return __find_general_cachep(size, gfpflags);
728 EXPORT_SYMBOL(kmem_find_general_cachep);
730 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
732 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
735 /* Calculate the number of objects and left-over bytes for a given
737 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
738 size_t align, int flags, size_t *left_over,
743 size_t slab_size = PAGE_SIZE << gfporder;
746 * The slab management structure can be either off the slab or
747 * on it. For the latter case, the memory allocated for a
751 * - One kmem_bufctl_t for each object
752 * - Padding to respect alignment of @align
753 * - @buffer_size bytes for each object
755 * If the slab management structure is off the slab, then the
756 * alignment will already be calculated into the size. Because
757 * the slabs are all pages aligned, the objects will be at the
758 * correct alignment when allocated.
760 if (flags & CFLGS_OFF_SLAB) {
762 nr_objs = slab_size / buffer_size;
764 if (nr_objs > SLAB_LIMIT)
765 nr_objs = SLAB_LIMIT;
768 * Ignore padding for the initial guess. The padding
769 * is at most @align-1 bytes, and @buffer_size is at
770 * least @align. In the worst case, this result will
771 * be one greater than the number of objects that fit
772 * into the memory allocation when taking the padding
775 nr_objs = (slab_size - sizeof(struct slab)) /
776 (buffer_size + sizeof(kmem_bufctl_t));
779 * This calculated number will be either the right
780 * amount, or one greater than what we want.
782 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
786 if (nr_objs > SLAB_LIMIT)
787 nr_objs = SLAB_LIMIT;
789 mgmt_size = slab_mgmt_size(nr_objs, align);
792 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
795 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
797 static void __slab_error(const char *function, struct kmem_cache *cachep, char *msg)
799 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
800 function, cachep->name, msg);
806 * Special reaping functions for NUMA systems called from cache_reap().
807 * These take care of doing round robin flushing of alien caches (containing
808 * objects freed on different nodes from which they were allocated) and the
809 * flushing of remote pcps by calling drain_node_pages.
811 static DEFINE_PER_CPU(unsigned long, reap_node);
813 static void init_reap_node(int cpu)
817 node = next_node(cpu_to_node(cpu), node_online_map);
818 if (node == MAX_NUMNODES)
821 __get_cpu_var(reap_node) = node;
824 static void next_reap_node(void)
826 int node = __get_cpu_var(reap_node);
829 * Also drain per cpu pages on remote zones
831 if (node != numa_node_id())
832 drain_node_pages(node);
834 node = next_node(node, node_online_map);
835 if (unlikely(node >= MAX_NUMNODES))
836 node = first_node(node_online_map);
837 __get_cpu_var(reap_node) = node;
841 #define init_reap_node(cpu) do { } while (0)
842 #define next_reap_node(void) do { } while (0)
846 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
847 * via the workqueue/eventd.
848 * Add the CPU number into the expiration time to minimize the possibility of
849 * the CPUs getting into lockstep and contending for the global cache chain
852 static void __devinit start_cpu_timer(int cpu)
854 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
857 * When this gets called from do_initcalls via cpucache_init(),
858 * init_workqueues() has already run, so keventd will be setup
861 if (keventd_up() && reap_work->func == NULL) {
863 INIT_WORK(reap_work, cache_reap, NULL);
864 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
868 static struct array_cache *alloc_arraycache(int node, int entries,
871 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
872 struct array_cache *nc = NULL;
874 nc = kmalloc_node(memsize, GFP_KERNEL, node);
878 nc->batchcount = batchcount;
880 spin_lock_init(&nc->lock);
886 static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int);
888 static struct array_cache **alloc_alien_cache(int node, int limit)
890 struct array_cache **ac_ptr;
891 int memsize = sizeof(void *) * MAX_NUMNODES;
896 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
899 if (i == node || !node_online(i)) {
903 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
905 for (i--; i <= 0; i--)
915 static void free_alien_cache(struct array_cache **ac_ptr)
928 static void __drain_alien_cache(struct kmem_cache *cachep,
929 struct array_cache *ac, int node)
931 struct kmem_list3 *rl3 = cachep->nodelists[node];
934 spin_lock(&rl3->list_lock);
935 free_block(cachep, ac->entry, ac->avail, node);
937 spin_unlock(&rl3->list_lock);
942 * Called from cache_reap() to regularly drain alien caches round robin.
944 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
946 int node = __get_cpu_var(reap_node);
949 struct array_cache *ac = l3->alien[node];
950 if (ac && ac->avail) {
951 spin_lock_irq(&ac->lock);
952 __drain_alien_cache(cachep, ac, node);
953 spin_unlock_irq(&ac->lock);
958 static void drain_alien_cache(struct kmem_cache *cachep, struct array_cache **alien)
961 struct array_cache *ac;
964 for_each_online_node(i) {
967 spin_lock_irqsave(&ac->lock, flags);
968 __drain_alien_cache(cachep, ac, i);
969 spin_unlock_irqrestore(&ac->lock, flags);
975 #define drain_alien_cache(cachep, alien) do { } while (0)
976 #define reap_alien(cachep, l3) do { } while (0)
978 static inline struct array_cache **alloc_alien_cache(int node, int limit)
980 return (struct array_cache **) 0x01020304ul;
983 static inline void free_alien_cache(struct array_cache **ac_ptr)
989 static int __devinit cpuup_callback(struct notifier_block *nfb,
990 unsigned long action, void *hcpu)
992 long cpu = (long)hcpu;
993 struct kmem_cache *cachep;
994 struct kmem_list3 *l3 = NULL;
995 int node = cpu_to_node(cpu);
996 int memsize = sizeof(struct kmem_list3);
1000 mutex_lock(&cache_chain_mutex);
1001 /* we need to do this right in the beginning since
1002 * alloc_arraycache's are going to use this list.
1003 * kmalloc_node allows us to add the slab to the right
1004 * kmem_list3 and not this cpu's kmem_list3
1007 list_for_each_entry(cachep, &cache_chain, next) {
1008 /* setup the size64 kmemlist for cpu before we can
1009 * begin anything. Make sure some other cpu on this
1010 * node has not already allocated this
1012 if (!cachep->nodelists[node]) {
1013 if (!(l3 = kmalloc_node(memsize,
1016 kmem_list3_init(l3);
1017 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1018 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1021 * The l3s don't come and go as CPUs come and
1022 * go. cache_chain_mutex is sufficient
1025 cachep->nodelists[node] = l3;
1028 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1029 cachep->nodelists[node]->free_limit =
1030 (1 + nr_cpus_node(node)) *
1031 cachep->batchcount + cachep->num;
1032 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1035 /* Now we can go ahead with allocating the shared array's
1037 list_for_each_entry(cachep, &cache_chain, next) {
1038 struct array_cache *nc;
1039 struct array_cache *shared;
1040 struct array_cache **alien;
1042 nc = alloc_arraycache(node, cachep->limit,
1043 cachep->batchcount);
1046 shared = alloc_arraycache(node,
1047 cachep->shared * cachep->batchcount,
1052 alien = alloc_alien_cache(node, cachep->limit);
1055 cachep->array[cpu] = nc;
1057 l3 = cachep->nodelists[node];
1060 spin_lock_irq(&l3->list_lock);
1063 * We are serialised from CPU_DEAD or
1064 * CPU_UP_CANCELLED by the cpucontrol lock
1066 l3->shared = shared;
1075 spin_unlock_irq(&l3->list_lock);
1078 free_alien_cache(alien);
1080 mutex_unlock(&cache_chain_mutex);
1083 start_cpu_timer(cpu);
1085 #ifdef CONFIG_HOTPLUG_CPU
1088 * Even if all the cpus of a node are down, we don't free the
1089 * kmem_list3 of any cache. This to avoid a race between
1090 * cpu_down, and a kmalloc allocation from another cpu for
1091 * memory from the node of the cpu going down. The list3
1092 * structure is usually allocated from kmem_cache_create() and
1093 * gets destroyed at kmem_cache_destroy().
1096 case CPU_UP_CANCELED:
1097 mutex_lock(&cache_chain_mutex);
1099 list_for_each_entry(cachep, &cache_chain, next) {
1100 struct array_cache *nc;
1101 struct array_cache *shared;
1102 struct array_cache **alien;
1105 mask = node_to_cpumask(node);
1106 /* cpu is dead; no one can alloc from it. */
1107 nc = cachep->array[cpu];
1108 cachep->array[cpu] = NULL;
1109 l3 = cachep->nodelists[node];
1112 goto free_array_cache;
1114 spin_lock_irq(&l3->list_lock);
1116 /* Free limit for this kmem_list3 */
1117 l3->free_limit -= cachep->batchcount;
1119 free_block(cachep, nc->entry, nc->avail, node);
1121 if (!cpus_empty(mask)) {
1122 spin_unlock_irq(&l3->list_lock);
1123 goto free_array_cache;
1126 shared = l3->shared;
1128 free_block(cachep, l3->shared->entry,
1129 l3->shared->avail, node);
1136 spin_unlock_irq(&l3->list_lock);
1140 drain_alien_cache(cachep, alien);
1141 free_alien_cache(alien);
1147 * In the previous loop, all the objects were freed to
1148 * the respective cache's slabs, now we can go ahead and
1149 * shrink each nodelist to its limit.
1151 list_for_each_entry(cachep, &cache_chain, next) {
1152 l3 = cachep->nodelists[node];
1155 spin_lock_irq(&l3->list_lock);
1156 /* free slabs belonging to this node */
1157 __node_shrink(cachep, node);
1158 spin_unlock_irq(&l3->list_lock);
1160 mutex_unlock(&cache_chain_mutex);
1166 mutex_unlock(&cache_chain_mutex);
1170 static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
1173 * swap the static kmem_list3 with kmalloced memory
1175 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list, int nodeid)
1177 struct kmem_list3 *ptr;
1179 BUG_ON(cachep->nodelists[nodeid] != list);
1180 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1183 local_irq_disable();
1184 memcpy(ptr, list, sizeof(struct kmem_list3));
1185 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1186 cachep->nodelists[nodeid] = ptr;
1191 * Called after the gfp() functions have been enabled, and before smp_init().
1193 void __init kmem_cache_init(void)
1196 struct cache_sizes *sizes;
1197 struct cache_names *names;
1201 for (i = 0; i < NUM_INIT_LISTS; i++) {
1202 kmem_list3_init(&initkmem_list3[i]);
1203 if (i < MAX_NUMNODES)
1204 cache_cache.nodelists[i] = NULL;
1208 * Fragmentation resistance on low memory - only use bigger
1209 * page orders on machines with more than 32MB of memory.
1211 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1212 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1214 /* Bootstrap is tricky, because several objects are allocated
1215 * from caches that do not exist yet:
1216 * 1) initialize the cache_cache cache: it contains the struct kmem_cache
1217 * structures of all caches, except cache_cache itself: cache_cache
1218 * is statically allocated.
1219 * Initially an __init data area is used for the head array and the
1220 * kmem_list3 structures, it's replaced with a kmalloc allocated
1221 * array at the end of the bootstrap.
1222 * 2) Create the first kmalloc cache.
1223 * The struct kmem_cache for the new cache is allocated normally.
1224 * An __init data area is used for the head array.
1225 * 3) Create the remaining kmalloc caches, with minimally sized
1227 * 4) Replace the __init data head arrays for cache_cache and the first
1228 * kmalloc cache with kmalloc allocated arrays.
1229 * 5) Replace the __init data for kmem_list3 for cache_cache and
1230 * the other cache's with kmalloc allocated memory.
1231 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1234 /* 1) create the cache_cache */
1235 INIT_LIST_HEAD(&cache_chain);
1236 list_add(&cache_cache.next, &cache_chain);
1237 cache_cache.colour_off = cache_line_size();
1238 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1239 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
1241 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, cache_line_size());
1243 for (order = 0; order < MAX_ORDER; order++) {
1244 cache_estimate(order, cache_cache.buffer_size,
1245 cache_line_size(), 0, &left_over, &cache_cache.num);
1246 if (cache_cache.num)
1249 if (!cache_cache.num)
1251 cache_cache.gfporder = order;
1252 cache_cache.colour = left_over / cache_cache.colour_off;
1253 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1254 sizeof(struct slab), cache_line_size());
1256 /* 2+3) create the kmalloc caches */
1257 sizes = malloc_sizes;
1258 names = cache_names;
1260 /* Initialize the caches that provide memory for the array cache
1261 * and the kmem_list3 structures first.
1262 * Without this, further allocations will bug
1265 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1266 sizes[INDEX_AC].cs_size,
1267 ARCH_KMALLOC_MINALIGN,
1268 (ARCH_KMALLOC_FLAGS |
1269 SLAB_PANIC), NULL, NULL);
1271 if (INDEX_AC != INDEX_L3)
1272 sizes[INDEX_L3].cs_cachep =
1273 kmem_cache_create(names[INDEX_L3].name,
1274 sizes[INDEX_L3].cs_size,
1275 ARCH_KMALLOC_MINALIGN,
1276 (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL,
1279 while (sizes->cs_size != ULONG_MAX) {
1281 * For performance, all the general caches are L1 aligned.
1282 * This should be particularly beneficial on SMP boxes, as it
1283 * eliminates "false sharing".
1284 * Note for systems short on memory removing the alignment will
1285 * allow tighter packing of the smaller caches.
1287 if (!sizes->cs_cachep)
1288 sizes->cs_cachep = kmem_cache_create(names->name,
1290 ARCH_KMALLOC_MINALIGN,
1295 /* Inc off-slab bufctl limit until the ceiling is hit. */
1296 if (!(OFF_SLAB(sizes->cs_cachep))) {
1297 offslab_limit = sizes->cs_size - sizeof(struct slab);
1298 offslab_limit /= sizeof(kmem_bufctl_t);
1301 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
1303 ARCH_KMALLOC_MINALIGN,
1304 (ARCH_KMALLOC_FLAGS |
1312 /* 4) Replace the bootstrap head arrays */
1316 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1318 local_irq_disable();
1319 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1320 memcpy(ptr, cpu_cache_get(&cache_cache),
1321 sizeof(struct arraycache_init));
1322 cache_cache.array[smp_processor_id()] = ptr;
1325 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1327 local_irq_disable();
1328 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1329 != &initarray_generic.cache);
1330 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1331 sizeof(struct arraycache_init));
1332 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1336 /* 5) Replace the bootstrap kmem_list3's */
1339 /* Replace the static kmem_list3 structures for the boot cpu */
1340 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
1343 for_each_online_node(node) {
1344 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1345 &initkmem_list3[SIZE_AC + node], node);
1347 if (INDEX_AC != INDEX_L3) {
1348 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1349 &initkmem_list3[SIZE_L3 + node],
1355 /* 6) resize the head arrays to their final sizes */
1357 struct kmem_cache *cachep;
1358 mutex_lock(&cache_chain_mutex);
1359 list_for_each_entry(cachep, &cache_chain, next)
1360 enable_cpucache(cachep);
1361 mutex_unlock(&cache_chain_mutex);
1365 g_cpucache_up = FULL;
1367 /* Register a cpu startup notifier callback
1368 * that initializes cpu_cache_get for all new cpus
1370 register_cpu_notifier(&cpucache_notifier);
1372 /* The reap timers are started later, with a module init call:
1373 * That part of the kernel is not yet operational.
1377 static int __init cpucache_init(void)
1382 * Register the timers that return unneeded
1385 for_each_online_cpu(cpu)
1386 start_cpu_timer(cpu);
1391 __initcall(cpucache_init);
1394 * Interface to system's page allocator. No need to hold the cache-lock.
1396 * If we requested dmaable memory, we will get it. Even if we
1397 * did not request dmaable memory, we might get it, but that
1398 * would be relatively rare and ignorable.
1400 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1406 flags |= cachep->gfpflags;
1407 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1410 addr = page_address(page);
1412 i = (1 << cachep->gfporder);
1413 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1414 atomic_add(i, &slab_reclaim_pages);
1415 add_page_state(nr_slab, i);
1417 __SetPageSlab(page);
1424 * Interface to system's page release.
1426 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1428 unsigned long i = (1 << cachep->gfporder);
1429 struct page *page = virt_to_page(addr);
1430 const unsigned long nr_freed = i;
1433 BUG_ON(!PageSlab(page));
1434 __ClearPageSlab(page);
1437 sub_page_state(nr_slab, nr_freed);
1438 if (current->reclaim_state)
1439 current->reclaim_state->reclaimed_slab += nr_freed;
1440 free_pages((unsigned long)addr, cachep->gfporder);
1441 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1442 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages);
1445 static void kmem_rcu_free(struct rcu_head *head)
1447 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1448 struct kmem_cache *cachep = slab_rcu->cachep;
1450 kmem_freepages(cachep, slab_rcu->addr);
1451 if (OFF_SLAB(cachep))
1452 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1457 #ifdef CONFIG_DEBUG_PAGEALLOC
1458 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1459 unsigned long caller)
1461 int size = obj_size(cachep);
1463 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1465 if (size < 5 * sizeof(unsigned long))
1468 *addr++ = 0x12345678;
1470 *addr++ = smp_processor_id();
1471 size -= 3 * sizeof(unsigned long);
1473 unsigned long *sptr = &caller;
1474 unsigned long svalue;
1476 while (!kstack_end(sptr)) {
1478 if (kernel_text_address(svalue)) {
1480 size -= sizeof(unsigned long);
1481 if (size <= sizeof(unsigned long))
1487 *addr++ = 0x87654321;
1491 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1493 int size = obj_size(cachep);
1494 addr = &((char *)addr)[obj_offset(cachep)];
1496 memset(addr, val, size);
1497 *(unsigned char *)(addr + size - 1) = POISON_END;
1500 static void dump_line(char *data, int offset, int limit)
1503 printk(KERN_ERR "%03x:", offset);
1504 for (i = 0; i < limit; i++) {
1505 printk(" %02x", (unsigned char)data[offset + i]);
1513 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1518 if (cachep->flags & SLAB_RED_ZONE) {
1519 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
1520 *dbg_redzone1(cachep, objp),
1521 *dbg_redzone2(cachep, objp));
1524 if (cachep->flags & SLAB_STORE_USER) {
1525 printk(KERN_ERR "Last user: [<%p>]",
1526 *dbg_userword(cachep, objp));
1527 print_symbol("(%s)",
1528 (unsigned long)*dbg_userword(cachep, objp));
1531 realobj = (char *)objp + obj_offset(cachep);
1532 size = obj_size(cachep);
1533 for (i = 0; i < size && lines; i += 16, lines--) {
1536 if (i + limit > size)
1538 dump_line(realobj, i, limit);
1542 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1548 realobj = (char *)objp + obj_offset(cachep);
1549 size = obj_size(cachep);
1551 for (i = 0; i < size; i++) {
1552 char exp = POISON_FREE;
1555 if (realobj[i] != exp) {
1561 "Slab corruption: start=%p, len=%d\n",
1563 print_objinfo(cachep, objp, 0);
1565 /* Hexdump the affected line */
1568 if (i + limit > size)
1570 dump_line(realobj, i, limit);
1573 /* Limit to 5 lines */
1579 /* Print some data about the neighboring objects, if they
1582 struct slab *slabp = virt_to_slab(objp);
1585 objnr = obj_to_index(cachep, slabp, objp);
1587 objp = index_to_obj(cachep, slabp, objnr - 1);
1588 realobj = (char *)objp + obj_offset(cachep);
1589 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1591 print_objinfo(cachep, objp, 2);
1593 if (objnr + 1 < cachep->num) {
1594 objp = index_to_obj(cachep, slabp, objnr + 1);
1595 realobj = (char *)objp + obj_offset(cachep);
1596 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1598 print_objinfo(cachep, objp, 2);
1606 * slab_destroy_objs - call the registered destructor for each object in
1607 * a slab that is to be destroyed.
1609 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1612 for (i = 0; i < cachep->num; i++) {
1613 void *objp = index_to_obj(cachep, slabp, i);
1615 if (cachep->flags & SLAB_POISON) {
1616 #ifdef CONFIG_DEBUG_PAGEALLOC
1617 if ((cachep->buffer_size % PAGE_SIZE) == 0
1618 && OFF_SLAB(cachep))
1619 kernel_map_pages(virt_to_page(objp),
1620 cachep->buffer_size / PAGE_SIZE,
1623 check_poison_obj(cachep, objp);
1625 check_poison_obj(cachep, objp);
1628 if (cachep->flags & SLAB_RED_ZONE) {
1629 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1630 slab_error(cachep, "start of a freed object "
1632 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1633 slab_error(cachep, "end of a freed object "
1636 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1637 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
1641 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1645 for (i = 0; i < cachep->num; i++) {
1646 void *objp = index_to_obj(cachep, slabp, i);
1647 (cachep->dtor) (objp, cachep, 0);
1654 * Destroy all the objs in a slab, and release the mem back to the system.
1655 * Before calling the slab must have been unlinked from the cache.
1656 * The cache-lock is not held/needed.
1658 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1660 void *addr = slabp->s_mem - slabp->colouroff;
1662 slab_destroy_objs(cachep, slabp);
1663 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1664 struct slab_rcu *slab_rcu;
1666 slab_rcu = (struct slab_rcu *)slabp;
1667 slab_rcu->cachep = cachep;
1668 slab_rcu->addr = addr;
1669 call_rcu(&slab_rcu->head, kmem_rcu_free);
1671 kmem_freepages(cachep, addr);
1672 if (OFF_SLAB(cachep))
1673 kmem_cache_free(cachep->slabp_cache, slabp);
1677 /* For setting up all the kmem_list3s for cache whose buffer_size is same
1678 as size of kmem_list3. */
1679 static void set_up_list3s(struct kmem_cache *cachep, int index)
1683 for_each_online_node(node) {
1684 cachep->nodelists[node] = &initkmem_list3[index + node];
1685 cachep->nodelists[node]->next_reap = jiffies +
1687 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1692 * calculate_slab_order - calculate size (page order) of slabs
1693 * @cachep: pointer to the cache that is being created
1694 * @size: size of objects to be created in this cache.
1695 * @align: required alignment for the objects.
1696 * @flags: slab allocation flags
1698 * Also calculates the number of objects per slab.
1700 * This could be made much more intelligent. For now, try to avoid using
1701 * high order pages for slabs. When the gfp() functions are more friendly
1702 * towards high-order requests, this should be changed.
1704 static inline size_t calculate_slab_order(struct kmem_cache *cachep,
1705 size_t size, size_t align, unsigned long flags)
1707 size_t left_over = 0;
1710 for (gfporder = 0 ; gfporder <= MAX_GFP_ORDER; gfporder++) {
1714 cache_estimate(gfporder, size, align, flags, &remainder, &num);
1718 /* More than offslab_limit objects will cause problems */
1719 if ((flags & CFLGS_OFF_SLAB) && num > offslab_limit)
1722 /* Found something acceptable - save it away */
1724 cachep->gfporder = gfporder;
1725 left_over = remainder;
1728 * A VFS-reclaimable slab tends to have most allocations
1729 * as GFP_NOFS and we really don't want to have to be allocating
1730 * higher-order pages when we are unable to shrink dcache.
1732 if (flags & SLAB_RECLAIM_ACCOUNT)
1736 * Large number of objects is good, but very large slabs are
1737 * currently bad for the gfp()s.
1739 if (gfporder >= slab_break_gfp_order)
1743 * Acceptable internal fragmentation?
1745 if ((left_over * 8) <= (PAGE_SIZE << gfporder))
1751 static void setup_cpu_cache(struct kmem_cache *cachep)
1753 if (g_cpucache_up == FULL) {
1754 enable_cpucache(cachep);
1757 if (g_cpucache_up == NONE) {
1759 * Note: the first kmem_cache_create must create the cache
1760 * that's used by kmalloc(24), otherwise the creation of
1761 * further caches will BUG().
1763 cachep->array[smp_processor_id()] = &initarray_generic.cache;
1766 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
1767 * the first cache, then we need to set up all its list3s,
1768 * otherwise the creation of further caches will BUG().
1770 set_up_list3s(cachep, SIZE_AC);
1771 if (INDEX_AC == INDEX_L3)
1772 g_cpucache_up = PARTIAL_L3;
1774 g_cpucache_up = PARTIAL_AC;
1776 cachep->array[smp_processor_id()] =
1777 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1779 if (g_cpucache_up == PARTIAL_AC) {
1780 set_up_list3s(cachep, SIZE_L3);
1781 g_cpucache_up = PARTIAL_L3;
1784 for_each_online_node(node) {
1785 cachep->nodelists[node] =
1786 kmalloc_node(sizeof(struct kmem_list3),
1788 BUG_ON(!cachep->nodelists[node]);
1789 kmem_list3_init(cachep->nodelists[node]);
1793 cachep->nodelists[numa_node_id()]->next_reap =
1794 jiffies + REAPTIMEOUT_LIST3 +
1795 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1797 cpu_cache_get(cachep)->avail = 0;
1798 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1799 cpu_cache_get(cachep)->batchcount = 1;
1800 cpu_cache_get(cachep)->touched = 0;
1801 cachep->batchcount = 1;
1802 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1806 * kmem_cache_create - Create a cache.
1807 * @name: A string which is used in /proc/slabinfo to identify this cache.
1808 * @size: The size of objects to be created in this cache.
1809 * @align: The required alignment for the objects.
1810 * @flags: SLAB flags
1811 * @ctor: A constructor for the objects.
1812 * @dtor: A destructor for the objects.
1814 * Returns a ptr to the cache on success, NULL on failure.
1815 * Cannot be called within a int, but can be interrupted.
1816 * The @ctor is run when new pages are allocated by the cache
1817 * and the @dtor is run before the pages are handed back.
1819 * @name must be valid until the cache is destroyed. This implies that
1820 * the module calling this has to destroy the cache before getting
1825 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1826 * to catch references to uninitialised memory.
1828 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1829 * for buffer overruns.
1831 * %SLAB_NO_REAP - Don't automatically reap this cache when we're under
1834 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1835 * cacheline. This can be beneficial if you're counting cycles as closely
1839 kmem_cache_create (const char *name, size_t size, size_t align,
1840 unsigned long flags, void (*ctor)(void*, struct kmem_cache *, unsigned long),
1841 void (*dtor)(void*, struct kmem_cache *, unsigned long))
1843 size_t left_over, slab_size, ralign;
1844 struct kmem_cache *cachep = NULL;
1845 struct list_head *p;
1848 * Sanity checks... these are all serious usage bugs.
1852 (size < BYTES_PER_WORD) ||
1853 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
1854 printk(KERN_ERR "%s: Early error in slab %s\n",
1855 __FUNCTION__, name);
1860 * Prevent CPUs from coming and going.
1861 * lock_cpu_hotplug() nests outside cache_chain_mutex
1865 mutex_lock(&cache_chain_mutex);
1867 list_for_each(p, &cache_chain) {
1868 struct kmem_cache *pc = list_entry(p, struct kmem_cache, next);
1869 mm_segment_t old_fs = get_fs();
1874 * This happens when the module gets unloaded and doesn't
1875 * destroy its slab cache and no-one else reuses the vmalloc
1876 * area of the module. Print a warning.
1879 res = __get_user(tmp, pc->name);
1882 printk("SLAB: cache with size %d has lost its name\n",
1887 if (!strcmp(pc->name, name)) {
1888 printk("kmem_cache_create: duplicate cache %s\n", name);
1895 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1896 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
1897 /* No constructor, but inital state check requested */
1898 printk(KERN_ERR "%s: No con, but init state check "
1899 "requested - %s\n", __FUNCTION__, name);
1900 flags &= ~SLAB_DEBUG_INITIAL;
1904 * Enable redzoning and last user accounting, except for caches with
1905 * large objects, if the increased size would increase the object size
1906 * above the next power of two: caches with object sizes just above a
1907 * power of two have a significant amount of internal fragmentation.
1910 || fls(size - 1) == fls(size - 1 + 3 * BYTES_PER_WORD)))
1911 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
1912 if (!(flags & SLAB_DESTROY_BY_RCU))
1913 flags |= SLAB_POISON;
1915 if (flags & SLAB_DESTROY_BY_RCU)
1916 BUG_ON(flags & SLAB_POISON);
1918 if (flags & SLAB_DESTROY_BY_RCU)
1922 * Always checks flags, a caller might be expecting debug
1923 * support which isn't available.
1925 if (flags & ~CREATE_MASK)
1928 /* Check that size is in terms of words. This is needed to avoid
1929 * unaligned accesses for some archs when redzoning is used, and makes
1930 * sure any on-slab bufctl's are also correctly aligned.
1932 if (size & (BYTES_PER_WORD - 1)) {
1933 size += (BYTES_PER_WORD - 1);
1934 size &= ~(BYTES_PER_WORD - 1);
1937 /* calculate out the final buffer alignment: */
1938 /* 1) arch recommendation: can be overridden for debug */
1939 if (flags & SLAB_HWCACHE_ALIGN) {
1940 /* Default alignment: as specified by the arch code.
1941 * Except if an object is really small, then squeeze multiple
1942 * objects into one cacheline.
1944 ralign = cache_line_size();
1945 while (size <= ralign / 2)
1948 ralign = BYTES_PER_WORD;
1950 /* 2) arch mandated alignment: disables debug if necessary */
1951 if (ralign < ARCH_SLAB_MINALIGN) {
1952 ralign = ARCH_SLAB_MINALIGN;
1953 if (ralign > BYTES_PER_WORD)
1954 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
1956 /* 3) caller mandated alignment: disables debug if necessary */
1957 if (ralign < align) {
1959 if (ralign > BYTES_PER_WORD)
1960 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
1962 /* 4) Store it. Note that the debug code below can reduce
1963 * the alignment to BYTES_PER_WORD.
1967 /* Get cache's description obj. */
1968 cachep = kmem_cache_alloc(&cache_cache, SLAB_KERNEL);
1971 memset(cachep, 0, sizeof(struct kmem_cache));
1974 cachep->obj_size = size;
1976 if (flags & SLAB_RED_ZONE) {
1977 /* redzoning only works with word aligned caches */
1978 align = BYTES_PER_WORD;
1980 /* add space for red zone words */
1981 cachep->obj_offset += BYTES_PER_WORD;
1982 size += 2 * BYTES_PER_WORD;
1984 if (flags & SLAB_STORE_USER) {
1985 /* user store requires word alignment and
1986 * one word storage behind the end of the real
1989 align = BYTES_PER_WORD;
1990 size += BYTES_PER_WORD;
1992 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
1993 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
1994 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
1995 cachep->obj_offset += PAGE_SIZE - size;
2001 /* Determine if the slab management is 'on' or 'off' slab. */
2002 if (size >= (PAGE_SIZE >> 3))
2004 * Size is large, assume best to place the slab management obj
2005 * off-slab (should allow better packing of objs).
2007 flags |= CFLGS_OFF_SLAB;
2009 size = ALIGN(size, align);
2011 left_over = calculate_slab_order(cachep, size, align, flags);
2014 printk("kmem_cache_create: couldn't create cache %s.\n", name);
2015 kmem_cache_free(&cache_cache, cachep);
2019 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2020 + sizeof(struct slab), align);
2023 * If the slab has been placed off-slab, and we have enough space then
2024 * move it on-slab. This is at the expense of any extra colouring.
2026 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2027 flags &= ~CFLGS_OFF_SLAB;
2028 left_over -= slab_size;
2031 if (flags & CFLGS_OFF_SLAB) {
2032 /* really off slab. No need for manual alignment */
2034 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2037 cachep->colour_off = cache_line_size();
2038 /* Offset must be a multiple of the alignment. */
2039 if (cachep->colour_off < align)
2040 cachep->colour_off = align;
2041 cachep->colour = left_over / cachep->colour_off;
2042 cachep->slab_size = slab_size;
2043 cachep->flags = flags;
2044 cachep->gfpflags = 0;
2045 if (flags & SLAB_CACHE_DMA)
2046 cachep->gfpflags |= GFP_DMA;
2047 spin_lock_init(&cachep->spinlock);
2048 cachep->buffer_size = size;
2050 if (flags & CFLGS_OFF_SLAB)
2051 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2052 cachep->ctor = ctor;
2053 cachep->dtor = dtor;
2054 cachep->name = name;
2057 setup_cpu_cache(cachep);
2059 /* cache setup completed, link it into the list */
2060 list_add(&cachep->next, &cache_chain);
2062 if (!cachep && (flags & SLAB_PANIC))
2063 panic("kmem_cache_create(): failed to create slab `%s'\n",
2065 mutex_unlock(&cache_chain_mutex);
2066 unlock_cpu_hotplug();
2069 EXPORT_SYMBOL(kmem_cache_create);
2072 static void check_irq_off(void)
2074 BUG_ON(!irqs_disabled());
2077 static void check_irq_on(void)
2079 BUG_ON(irqs_disabled());
2082 static void check_spinlock_acquired(struct kmem_cache *cachep)
2086 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2090 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2094 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2099 #define check_irq_off() do { } while(0)
2100 #define check_irq_on() do { } while(0)
2101 #define check_spinlock_acquired(x) do { } while(0)
2102 #define check_spinlock_acquired_node(x, y) do { } while(0)
2106 * Waits for all CPUs to execute func().
2108 static void smp_call_function_all_cpus(void (*func)(void *arg), void *arg)
2113 local_irq_disable();
2117 if (smp_call_function(func, arg, 1, 1))
2123 static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac,
2124 int force, int node);
2126 static void do_drain(void *arg)
2128 struct kmem_cache *cachep = (struct kmem_cache *) arg;
2129 struct array_cache *ac;
2130 int node = numa_node_id();
2133 ac = cpu_cache_get(cachep);
2134 spin_lock(&cachep->nodelists[node]->list_lock);
2135 free_block(cachep, ac->entry, ac->avail, node);
2136 spin_unlock(&cachep->nodelists[node]->list_lock);
2140 static void drain_cpu_caches(struct kmem_cache *cachep)
2142 struct kmem_list3 *l3;
2145 smp_call_function_all_cpus(do_drain, cachep);
2147 for_each_online_node(node) {
2148 l3 = cachep->nodelists[node];
2150 spin_lock_irq(&l3->list_lock);
2151 drain_array_locked(cachep, l3->shared, 1, node);
2152 spin_unlock_irq(&l3->list_lock);
2154 drain_alien_cache(cachep, l3->alien);
2159 static int __node_shrink(struct kmem_cache *cachep, int node)
2162 struct kmem_list3 *l3 = cachep->nodelists[node];
2166 struct list_head *p;
2168 p = l3->slabs_free.prev;
2169 if (p == &l3->slabs_free)
2172 slabp = list_entry(l3->slabs_free.prev, struct slab, list);
2177 list_del(&slabp->list);
2179 l3->free_objects -= cachep->num;
2180 spin_unlock_irq(&l3->list_lock);
2181 slab_destroy(cachep, slabp);
2182 spin_lock_irq(&l3->list_lock);
2184 ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial);
2188 static int __cache_shrink(struct kmem_cache *cachep)
2191 struct kmem_list3 *l3;
2193 drain_cpu_caches(cachep);
2196 for_each_online_node(i) {
2197 l3 = cachep->nodelists[i];
2199 spin_lock_irq(&l3->list_lock);
2200 ret += __node_shrink(cachep, i);
2201 spin_unlock_irq(&l3->list_lock);
2204 return (ret ? 1 : 0);
2208 * kmem_cache_shrink - Shrink a cache.
2209 * @cachep: The cache to shrink.
2211 * Releases as many slabs as possible for a cache.
2212 * To help debugging, a zero exit status indicates all slabs were released.
2214 int kmem_cache_shrink(struct kmem_cache *cachep)
2216 if (!cachep || in_interrupt())
2219 return __cache_shrink(cachep);
2221 EXPORT_SYMBOL(kmem_cache_shrink);
2224 * kmem_cache_destroy - delete a cache
2225 * @cachep: the cache to destroy
2227 * Remove a struct kmem_cache object from the slab cache.
2228 * Returns 0 on success.
2230 * It is expected this function will be called by a module when it is
2231 * unloaded. This will remove the cache completely, and avoid a duplicate
2232 * cache being allocated each time a module is loaded and unloaded, if the
2233 * module doesn't have persistent in-kernel storage across loads and unloads.
2235 * The cache must be empty before calling this function.
2237 * The caller must guarantee that noone will allocate memory from the cache
2238 * during the kmem_cache_destroy().
2240 int kmem_cache_destroy(struct kmem_cache *cachep)
2243 struct kmem_list3 *l3;
2245 if (!cachep || in_interrupt())
2248 /* Don't let CPUs to come and go */
2251 /* Find the cache in the chain of caches. */
2252 mutex_lock(&cache_chain_mutex);
2254 * the chain is never empty, cache_cache is never destroyed
2256 list_del(&cachep->next);
2257 mutex_unlock(&cache_chain_mutex);
2259 if (__cache_shrink(cachep)) {
2260 slab_error(cachep, "Can't free all objects");
2261 mutex_lock(&cache_chain_mutex);
2262 list_add(&cachep->next, &cache_chain);
2263 mutex_unlock(&cache_chain_mutex);
2264 unlock_cpu_hotplug();
2268 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2271 for_each_online_cpu(i)
2272 kfree(cachep->array[i]);
2274 /* NUMA: free the list3 structures */
2275 for_each_online_node(i) {
2276 if ((l3 = cachep->nodelists[i])) {
2278 free_alien_cache(l3->alien);
2282 kmem_cache_free(&cache_cache, cachep);
2284 unlock_cpu_hotplug();
2288 EXPORT_SYMBOL(kmem_cache_destroy);
2290 /* Get the memory for a slab management obj. */
2291 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2292 int colour_off, gfp_t local_flags)
2296 if (OFF_SLAB(cachep)) {
2297 /* Slab management obj is off-slab. */
2298 slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags);
2302 slabp = objp + colour_off;
2303 colour_off += cachep->slab_size;
2306 slabp->colouroff = colour_off;
2307 slabp->s_mem = objp + colour_off;
2312 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2314 return (kmem_bufctl_t *) (slabp + 1);
2317 static void cache_init_objs(struct kmem_cache *cachep,
2318 struct slab *slabp, unsigned long ctor_flags)
2322 for (i = 0; i < cachep->num; i++) {
2323 void *objp = index_to_obj(cachep, slabp, i);
2325 /* need to poison the objs? */
2326 if (cachep->flags & SLAB_POISON)
2327 poison_obj(cachep, objp, POISON_FREE);
2328 if (cachep->flags & SLAB_STORE_USER)
2329 *dbg_userword(cachep, objp) = NULL;
2331 if (cachep->flags & SLAB_RED_ZONE) {
2332 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2333 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2336 * Constructors are not allowed to allocate memory from
2337 * the same cache which they are a constructor for.
2338 * Otherwise, deadlock. They must also be threaded.
2340 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2341 cachep->ctor(objp + obj_offset(cachep), cachep,
2344 if (cachep->flags & SLAB_RED_ZONE) {
2345 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2346 slab_error(cachep, "constructor overwrote the"
2347 " end of an object");
2348 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2349 slab_error(cachep, "constructor overwrote the"
2350 " start of an object");
2352 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)
2353 && cachep->flags & SLAB_POISON)
2354 kernel_map_pages(virt_to_page(objp),
2355 cachep->buffer_size / PAGE_SIZE, 0);
2358 cachep->ctor(objp, cachep, ctor_flags);
2360 slab_bufctl(slabp)[i] = i + 1;
2362 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2366 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2368 if (flags & SLAB_DMA) {
2369 if (!(cachep->gfpflags & GFP_DMA))
2372 if (cachep->gfpflags & GFP_DMA)
2377 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, int nodeid)
2379 void *objp = index_to_obj(cachep, slabp, slabp->free);
2383 next = slab_bufctl(slabp)[slabp->free];
2385 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2386 WARN_ON(slabp->nodeid != nodeid);
2393 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, void *objp,
2396 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2399 /* Verify that the slab belongs to the intended node */
2400 WARN_ON(slabp->nodeid != nodeid);
2402 if (slab_bufctl(slabp)[objnr] != BUFCTL_FREE) {
2403 printk(KERN_ERR "slab: double free detected in cache "
2404 "'%s', objp %p\n", cachep->name, objp);
2408 slab_bufctl(slabp)[objnr] = slabp->free;
2409 slabp->free = objnr;
2413 static void set_slab_attr(struct kmem_cache *cachep, struct slab *slabp, void *objp)
2418 /* Nasty!!!!!! I hope this is OK. */
2419 i = 1 << cachep->gfporder;
2420 page = virt_to_page(objp);
2422 page_set_cache(page, cachep);
2423 page_set_slab(page, slabp);
2429 * Grow (by 1) the number of slabs within a cache. This is called by
2430 * kmem_cache_alloc() when there are no active objs left in a cache.
2432 static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid)
2438 unsigned long ctor_flags;
2439 struct kmem_list3 *l3;
2441 /* Be lazy and only check for valid flags here,
2442 * keeping it out of the critical path in kmem_cache_alloc().
2444 if (flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW))
2446 if (flags & SLAB_NO_GROW)
2449 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2450 local_flags = (flags & SLAB_LEVEL_MASK);
2451 if (!(local_flags & __GFP_WAIT))
2453 * Not allowed to sleep. Need to tell a constructor about
2454 * this - it might need to know...
2456 ctor_flags |= SLAB_CTOR_ATOMIC;
2458 /* Take the l3 list lock to change the colour_next on this node */
2460 l3 = cachep->nodelists[nodeid];
2461 spin_lock(&l3->list_lock);
2463 /* Get colour for the slab, and cal the next value. */
2464 offset = l3->colour_next;
2466 if (l3->colour_next >= cachep->colour)
2467 l3->colour_next = 0;
2468 spin_unlock(&l3->list_lock);
2470 offset *= cachep->colour_off;
2472 if (local_flags & __GFP_WAIT)
2476 * The test for missing atomic flag is performed here, rather than
2477 * the more obvious place, simply to reduce the critical path length
2478 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2479 * will eventually be caught here (where it matters).
2481 kmem_flagcheck(cachep, flags);
2483 /* Get mem for the objs.
2484 * Attempt to allocate a physical page from 'nodeid',
2486 if (!(objp = kmem_getpages(cachep, flags, nodeid)))
2489 /* Get slab management. */
2490 if (!(slabp = alloc_slabmgmt(cachep, objp, offset, local_flags)))
2493 slabp->nodeid = nodeid;
2494 set_slab_attr(cachep, slabp, objp);
2496 cache_init_objs(cachep, slabp, ctor_flags);
2498 if (local_flags & __GFP_WAIT)
2499 local_irq_disable();
2501 spin_lock(&l3->list_lock);
2503 /* Make slab active. */
2504 list_add_tail(&slabp->list, &(l3->slabs_free));
2505 STATS_INC_GROWN(cachep);
2506 l3->free_objects += cachep->num;
2507 spin_unlock(&l3->list_lock);
2510 kmem_freepages(cachep, objp);
2512 if (local_flags & __GFP_WAIT)
2513 local_irq_disable();
2520 * Perform extra freeing checks:
2521 * - detect bad pointers.
2522 * - POISON/RED_ZONE checking
2523 * - destructor calls, for caches with POISON+dtor
2525 static void kfree_debugcheck(const void *objp)
2529 if (!virt_addr_valid(objp)) {
2530 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2531 (unsigned long)objp);
2534 page = virt_to_page(objp);
2535 if (!PageSlab(page)) {
2536 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2537 (unsigned long)objp);
2542 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2549 objp -= obj_offset(cachep);
2550 kfree_debugcheck(objp);
2551 page = virt_to_page(objp);
2553 if (page_get_cache(page) != cachep) {
2555 "mismatch in kmem_cache_free: expected cache %p, got %p\n",
2556 page_get_cache(page), cachep);
2557 printk(KERN_ERR "%p is %s.\n", cachep, cachep->name);
2558 printk(KERN_ERR "%p is %s.\n", page_get_cache(page),
2559 page_get_cache(page)->name);
2562 slabp = page_get_slab(page);
2564 if (cachep->flags & SLAB_RED_ZONE) {
2565 if (*dbg_redzone1(cachep, objp) != RED_ACTIVE
2566 || *dbg_redzone2(cachep, objp) != RED_ACTIVE) {
2568 "double free, or memory outside"
2569 " object was overwritten");
2571 "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2572 objp, *dbg_redzone1(cachep, objp),
2573 *dbg_redzone2(cachep, objp));
2575 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2576 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2578 if (cachep->flags & SLAB_STORE_USER)
2579 *dbg_userword(cachep, objp) = caller;
2581 objnr = obj_to_index(cachep, slabp, objp);
2583 BUG_ON(objnr >= cachep->num);
2584 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2586 if (cachep->flags & SLAB_DEBUG_INITIAL) {
2587 /* Need to call the slab's constructor so the
2588 * caller can perform a verify of its state (debugging).
2589 * Called without the cache-lock held.
2591 cachep->ctor(objp + obj_offset(cachep),
2592 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
2594 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2595 /* we want to cache poison the object,
2596 * call the destruction callback
2598 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
2600 if (cachep->flags & SLAB_POISON) {
2601 #ifdef CONFIG_DEBUG_PAGEALLOC
2602 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) {
2603 store_stackinfo(cachep, objp, (unsigned long)caller);
2604 kernel_map_pages(virt_to_page(objp),
2605 cachep->buffer_size / PAGE_SIZE, 0);
2607 poison_obj(cachep, objp, POISON_FREE);
2610 poison_obj(cachep, objp, POISON_FREE);
2616 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2621 /* Check slab's freelist to see if this obj is there. */
2622 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2624 if (entries > cachep->num || i >= cachep->num)
2627 if (entries != cachep->num - slabp->inuse) {
2630 "slab: Internal list corruption detected in cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2631 cachep->name, cachep->num, slabp, slabp->inuse);
2633 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2636 printk("\n%03x:", i);
2637 printk(" %02x", ((unsigned char *)slabp)[i]);
2644 #define kfree_debugcheck(x) do { } while(0)
2645 #define cache_free_debugcheck(x,objp,z) (objp)
2646 #define check_slabp(x,y) do { } while(0)
2649 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2652 struct kmem_list3 *l3;
2653 struct array_cache *ac;
2656 ac = cpu_cache_get(cachep);
2658 batchcount = ac->batchcount;
2659 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2660 /* if there was little recent activity on this
2661 * cache, then perform only a partial refill.
2662 * Otherwise we could generate refill bouncing.
2664 batchcount = BATCHREFILL_LIMIT;
2666 l3 = cachep->nodelists[numa_node_id()];
2668 BUG_ON(ac->avail > 0 || !l3);
2669 spin_lock(&l3->list_lock);
2672 struct array_cache *shared_array = l3->shared;
2673 if (shared_array->avail) {
2674 if (batchcount > shared_array->avail)
2675 batchcount = shared_array->avail;
2676 shared_array->avail -= batchcount;
2677 ac->avail = batchcount;
2679 &(shared_array->entry[shared_array->avail]),
2680 sizeof(void *) * batchcount);
2681 shared_array->touched = 1;
2685 while (batchcount > 0) {
2686 struct list_head *entry;
2688 /* Get slab alloc is to come from. */
2689 entry = l3->slabs_partial.next;
2690 if (entry == &l3->slabs_partial) {
2691 l3->free_touched = 1;
2692 entry = l3->slabs_free.next;
2693 if (entry == &l3->slabs_free)
2697 slabp = list_entry(entry, struct slab, list);
2698 check_slabp(cachep, slabp);
2699 check_spinlock_acquired(cachep);
2700 while (slabp->inuse < cachep->num && batchcount--) {
2701 STATS_INC_ALLOCED(cachep);
2702 STATS_INC_ACTIVE(cachep);
2703 STATS_SET_HIGH(cachep);
2705 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2708 check_slabp(cachep, slabp);
2710 /* move slabp to correct slabp list: */
2711 list_del(&slabp->list);
2712 if (slabp->free == BUFCTL_END)
2713 list_add(&slabp->list, &l3->slabs_full);
2715 list_add(&slabp->list, &l3->slabs_partial);
2719 l3->free_objects -= ac->avail;
2721 spin_unlock(&l3->list_lock);
2723 if (unlikely(!ac->avail)) {
2725 x = cache_grow(cachep, flags, numa_node_id());
2727 // cache_grow can reenable interrupts, then ac could change.
2728 ac = cpu_cache_get(cachep);
2729 if (!x && ac->avail == 0) // no objects in sight? abort
2732 if (!ac->avail) // objects refilled by interrupt?
2736 return ac->entry[--ac->avail];
2740 cache_alloc_debugcheck_before(struct kmem_cache *cachep, gfp_t flags)
2742 might_sleep_if(flags & __GFP_WAIT);
2744 kmem_flagcheck(cachep, flags);
2749 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, gfp_t flags,
2750 void *objp, void *caller)
2754 if (cachep->flags & SLAB_POISON) {
2755 #ifdef CONFIG_DEBUG_PAGEALLOC
2756 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
2757 kernel_map_pages(virt_to_page(objp),
2758 cachep->buffer_size / PAGE_SIZE, 1);
2760 check_poison_obj(cachep, objp);
2762 check_poison_obj(cachep, objp);
2764 poison_obj(cachep, objp, POISON_INUSE);
2766 if (cachep->flags & SLAB_STORE_USER)
2767 *dbg_userword(cachep, objp) = caller;
2769 if (cachep->flags & SLAB_RED_ZONE) {
2770 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE
2771 || *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2773 "double free, or memory outside"
2774 " object was overwritten");
2776 "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2777 objp, *dbg_redzone1(cachep, objp),
2778 *dbg_redzone2(cachep, objp));
2780 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2781 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2783 objp += obj_offset(cachep);
2784 if (cachep->ctor && cachep->flags & SLAB_POISON) {
2785 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2787 if (!(flags & __GFP_WAIT))
2788 ctor_flags |= SLAB_CTOR_ATOMIC;
2790 cachep->ctor(objp, cachep, ctor_flags);
2795 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2798 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
2801 struct array_cache *ac;
2804 if (unlikely(current->mempolicy && !in_interrupt())) {
2805 int nid = slab_node(current->mempolicy);
2807 if (nid != numa_node_id())
2808 return __cache_alloc_node(cachep, flags, nid);
2813 ac = cpu_cache_get(cachep);
2814 if (likely(ac->avail)) {
2815 STATS_INC_ALLOCHIT(cachep);
2817 objp = ac->entry[--ac->avail];
2819 STATS_INC_ALLOCMISS(cachep);
2820 objp = cache_alloc_refill(cachep, flags);
2825 static __always_inline void *
2826 __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
2828 unsigned long save_flags;
2831 cache_alloc_debugcheck_before(cachep, flags);
2833 local_irq_save(save_flags);
2834 objp = ____cache_alloc(cachep, flags);
2835 local_irq_restore(save_flags);
2836 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
2844 * A interface to enable slab creation on nodeid
2846 static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
2848 struct list_head *entry;
2850 struct kmem_list3 *l3;
2854 l3 = cachep->nodelists[nodeid];
2859 spin_lock(&l3->list_lock);
2860 entry = l3->slabs_partial.next;
2861 if (entry == &l3->slabs_partial) {
2862 l3->free_touched = 1;
2863 entry = l3->slabs_free.next;
2864 if (entry == &l3->slabs_free)
2868 slabp = list_entry(entry, struct slab, list);
2869 check_spinlock_acquired_node(cachep, nodeid);
2870 check_slabp(cachep, slabp);
2872 STATS_INC_NODEALLOCS(cachep);
2873 STATS_INC_ACTIVE(cachep);
2874 STATS_SET_HIGH(cachep);
2876 BUG_ON(slabp->inuse == cachep->num);
2878 obj = slab_get_obj(cachep, slabp, nodeid);
2879 check_slabp(cachep, slabp);
2881 /* move slabp to correct slabp list: */
2882 list_del(&slabp->list);
2884 if (slabp->free == BUFCTL_END) {
2885 list_add(&slabp->list, &l3->slabs_full);
2887 list_add(&slabp->list, &l3->slabs_partial);
2890 spin_unlock(&l3->list_lock);
2894 spin_unlock(&l3->list_lock);
2895 x = cache_grow(cachep, flags, nodeid);
2907 * Caller needs to acquire correct kmem_list's list_lock
2909 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
2913 struct kmem_list3 *l3;
2915 for (i = 0; i < nr_objects; i++) {
2916 void *objp = objpp[i];
2919 slabp = virt_to_slab(objp);
2920 l3 = cachep->nodelists[node];
2921 list_del(&slabp->list);
2922 check_spinlock_acquired_node(cachep, node);
2923 check_slabp(cachep, slabp);
2924 slab_put_obj(cachep, slabp, objp, node);
2925 STATS_DEC_ACTIVE(cachep);
2927 check_slabp(cachep, slabp);
2929 /* fixup slab chains */
2930 if (slabp->inuse == 0) {
2931 if (l3->free_objects > l3->free_limit) {
2932 l3->free_objects -= cachep->num;
2933 slab_destroy(cachep, slabp);
2935 list_add(&slabp->list, &l3->slabs_free);
2938 /* Unconditionally move a slab to the end of the
2939 * partial list on free - maximum time for the
2940 * other objects to be freed, too.
2942 list_add_tail(&slabp->list, &l3->slabs_partial);
2947 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
2950 struct kmem_list3 *l3;
2951 int node = numa_node_id();
2953 batchcount = ac->batchcount;
2955 BUG_ON(!batchcount || batchcount > ac->avail);
2958 l3 = cachep->nodelists[node];
2959 spin_lock(&l3->list_lock);
2961 struct array_cache *shared_array = l3->shared;
2962 int max = shared_array->limit - shared_array->avail;
2964 if (batchcount > max)
2966 memcpy(&(shared_array->entry[shared_array->avail]),
2967 ac->entry, sizeof(void *) * batchcount);
2968 shared_array->avail += batchcount;
2973 free_block(cachep, ac->entry, batchcount, node);
2978 struct list_head *p;
2980 p = l3->slabs_free.next;
2981 while (p != &(l3->slabs_free)) {
2984 slabp = list_entry(p, struct slab, list);
2985 BUG_ON(slabp->inuse);
2990 STATS_SET_FREEABLE(cachep, i);
2993 spin_unlock(&l3->list_lock);
2994 ac->avail -= batchcount;
2995 memmove(ac->entry, &(ac->entry[batchcount]),
2996 sizeof(void *) * ac->avail);
3001 * Release an obj back to its cache. If the obj has a constructed
3002 * state, it must be in this state _before_ it is released.
3004 * Called with disabled ints.
3006 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3008 struct array_cache *ac = cpu_cache_get(cachep);
3011 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3013 /* Make sure we are not freeing a object from another
3014 * node to the array cache on this cpu.
3019 slabp = virt_to_slab(objp);
3020 if (unlikely(slabp->nodeid != numa_node_id())) {
3021 struct array_cache *alien = NULL;
3022 int nodeid = slabp->nodeid;
3023 struct kmem_list3 *l3 =
3024 cachep->nodelists[numa_node_id()];
3026 STATS_INC_NODEFREES(cachep);
3027 if (l3->alien && l3->alien[nodeid]) {
3028 alien = l3->alien[nodeid];
3029 spin_lock(&alien->lock);
3030 if (unlikely(alien->avail == alien->limit))
3031 __drain_alien_cache(cachep,
3033 alien->entry[alien->avail++] = objp;
3034 spin_unlock(&alien->lock);
3036 spin_lock(&(cachep->nodelists[nodeid])->
3038 free_block(cachep, &objp, 1, nodeid);
3039 spin_unlock(&(cachep->nodelists[nodeid])->
3046 if (likely(ac->avail < ac->limit)) {
3047 STATS_INC_FREEHIT(cachep);
3048 ac->entry[ac->avail++] = objp;
3051 STATS_INC_FREEMISS(cachep);
3052 cache_flusharray(cachep, ac);
3053 ac->entry[ac->avail++] = objp;
3058 * kmem_cache_alloc - Allocate an object
3059 * @cachep: The cache to allocate from.
3060 * @flags: See kmalloc().
3062 * Allocate an object from this cache. The flags are only relevant
3063 * if the cache has no available objects.
3065 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3067 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3069 EXPORT_SYMBOL(kmem_cache_alloc);
3072 * kmem_ptr_validate - check if an untrusted pointer might
3074 * @cachep: the cache we're checking against
3075 * @ptr: pointer to validate
3077 * This verifies that the untrusted pointer looks sane:
3078 * it is _not_ a guarantee that the pointer is actually
3079 * part of the slab cache in question, but it at least
3080 * validates that the pointer can be dereferenced and
3081 * looks half-way sane.
3083 * Currently only used for dentry validation.
3085 int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr)
3087 unsigned long addr = (unsigned long)ptr;
3088 unsigned long min_addr = PAGE_OFFSET;
3089 unsigned long align_mask = BYTES_PER_WORD - 1;
3090 unsigned long size = cachep->buffer_size;
3093 if (unlikely(addr < min_addr))
3095 if (unlikely(addr > (unsigned long)high_memory - size))
3097 if (unlikely(addr & align_mask))
3099 if (unlikely(!kern_addr_valid(addr)))
3101 if (unlikely(!kern_addr_valid(addr + size - 1)))
3103 page = virt_to_page(ptr);
3104 if (unlikely(!PageSlab(page)))
3106 if (unlikely(page_get_cache(page) != cachep))
3115 * kmem_cache_alloc_node - Allocate an object on the specified node
3116 * @cachep: The cache to allocate from.
3117 * @flags: See kmalloc().
3118 * @nodeid: node number of the target node.
3120 * Identical to kmem_cache_alloc, except that this function is slow
3121 * and can sleep. And it will allocate memory on the given node, which
3122 * can improve the performance for cpu bound structures.
3123 * New and improved: it will now make sure that the object gets
3124 * put on the correct node list so that there is no false sharing.
3126 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3128 unsigned long save_flags;
3131 cache_alloc_debugcheck_before(cachep, flags);
3132 local_irq_save(save_flags);
3134 if (nodeid == -1 || nodeid == numa_node_id() ||
3135 !cachep->nodelists[nodeid])
3136 ptr = ____cache_alloc(cachep, flags);
3138 ptr = __cache_alloc_node(cachep, flags, nodeid);
3139 local_irq_restore(save_flags);
3141 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr,
3142 __builtin_return_address(0));
3146 EXPORT_SYMBOL(kmem_cache_alloc_node);
3148 void *kmalloc_node(size_t size, gfp_t flags, int node)
3150 struct kmem_cache *cachep;
3152 cachep = kmem_find_general_cachep(size, flags);
3153 if (unlikely(cachep == NULL))
3155 return kmem_cache_alloc_node(cachep, flags, node);
3157 EXPORT_SYMBOL(kmalloc_node);
3161 * kmalloc - allocate memory
3162 * @size: how many bytes of memory are required.
3163 * @flags: the type of memory to allocate.
3165 * kmalloc is the normal method of allocating memory
3168 * The @flags argument may be one of:
3170 * %GFP_USER - Allocate memory on behalf of user. May sleep.
3172 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
3174 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
3176 * Additionally, the %GFP_DMA flag may be set to indicate the memory
3177 * must be suitable for DMA. This can mean different things on different
3178 * platforms. For example, on i386, it means that the memory must come
3179 * from the first 16MB.
3181 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3184 struct kmem_cache *cachep;
3186 /* If you want to save a few bytes .text space: replace
3188 * Then kmalloc uses the uninlined functions instead of the inline
3191 cachep = __find_general_cachep(size, flags);
3192 if (unlikely(cachep == NULL))
3194 return __cache_alloc(cachep, flags, caller);
3197 #ifndef CONFIG_DEBUG_SLAB
3199 void *__kmalloc(size_t size, gfp_t flags)
3201 return __do_kmalloc(size, flags, NULL);
3203 EXPORT_SYMBOL(__kmalloc);
3207 void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3209 return __do_kmalloc(size, flags, caller);
3211 EXPORT_SYMBOL(__kmalloc_track_caller);
3217 * __alloc_percpu - allocate one copy of the object for every present
3218 * cpu in the system, zeroing them.
3219 * Objects should be dereferenced using the per_cpu_ptr macro only.
3221 * @size: how many bytes of memory are required.
3223 void *__alloc_percpu(size_t size)
3226 struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL);
3232 * Cannot use for_each_online_cpu since a cpu may come online
3233 * and we have no way of figuring out how to fix the array
3234 * that we have allocated then....
3237 int node = cpu_to_node(i);
3239 if (node_online(node))
3240 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
3242 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
3244 if (!pdata->ptrs[i])
3246 memset(pdata->ptrs[i], 0, size);
3249 /* Catch derefs w/o wrappers */
3250 return (void *)(~(unsigned long)pdata);
3254 if (!cpu_possible(i))
3256 kfree(pdata->ptrs[i]);
3261 EXPORT_SYMBOL(__alloc_percpu);
3265 * kmem_cache_free - Deallocate an object
3266 * @cachep: The cache the allocation was from.
3267 * @objp: The previously allocated object.
3269 * Free an object which was previously allocated from this
3272 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3274 unsigned long flags;
3276 local_irq_save(flags);
3277 __cache_free(cachep, objp);
3278 local_irq_restore(flags);
3280 EXPORT_SYMBOL(kmem_cache_free);
3283 * kfree - free previously allocated memory
3284 * @objp: pointer returned by kmalloc.
3286 * If @objp is NULL, no operation is performed.
3288 * Don't free memory not originally allocated by kmalloc()
3289 * or you will run into trouble.
3291 void kfree(const void *objp)
3293 struct kmem_cache *c;
3294 unsigned long flags;
3296 if (unlikely(!objp))
3298 local_irq_save(flags);
3299 kfree_debugcheck(objp);
3300 c = virt_to_cache(objp);
3301 mutex_debug_check_no_locks_freed(objp, obj_size(c));
3302 __cache_free(c, (void *)objp);
3303 local_irq_restore(flags);
3305 EXPORT_SYMBOL(kfree);
3309 * free_percpu - free previously allocated percpu memory
3310 * @objp: pointer returned by alloc_percpu.
3312 * Don't free memory not originally allocated by alloc_percpu()
3313 * The complemented objp is to check for that.
3315 void free_percpu(const void *objp)
3318 struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp);
3321 * We allocate for all cpus so we cannot use for online cpu here.
3327 EXPORT_SYMBOL(free_percpu);
3330 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3332 return obj_size(cachep);
3334 EXPORT_SYMBOL(kmem_cache_size);
3336 const char *kmem_cache_name(struct kmem_cache *cachep)
3338 return cachep->name;
3340 EXPORT_SYMBOL_GPL(kmem_cache_name);
3343 * This initializes kmem_list3 for all nodes.
3345 static int alloc_kmemlist(struct kmem_cache *cachep)
3348 struct kmem_list3 *l3;
3351 for_each_online_node(node) {
3352 struct array_cache *nc = NULL, *new;
3353 struct array_cache **new_alien = NULL;
3355 if (!(new_alien = alloc_alien_cache(node, cachep->limit)))
3358 if (!(new = alloc_arraycache(node, (cachep->shared *
3359 cachep->batchcount),
3362 if ((l3 = cachep->nodelists[node])) {
3364 spin_lock_irq(&l3->list_lock);
3366 if ((nc = cachep->nodelists[node]->shared))
3367 free_block(cachep, nc->entry, nc->avail, node);
3370 if (!cachep->nodelists[node]->alien) {
3371 l3->alien = new_alien;
3374 l3->free_limit = (1 + nr_cpus_node(node)) *
3375 cachep->batchcount + cachep->num;
3376 spin_unlock_irq(&l3->list_lock);
3378 free_alien_cache(new_alien);
3381 if (!(l3 = kmalloc_node(sizeof(struct kmem_list3),
3385 kmem_list3_init(l3);
3386 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3387 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3389 l3->alien = new_alien;
3390 l3->free_limit = (1 + nr_cpus_node(node)) *
3391 cachep->batchcount + cachep->num;
3392 cachep->nodelists[node] = l3;
3400 struct ccupdate_struct {
3401 struct kmem_cache *cachep;
3402 struct array_cache *new[NR_CPUS];
3405 static void do_ccupdate_local(void *info)
3407 struct ccupdate_struct *new = (struct ccupdate_struct *)info;
3408 struct array_cache *old;
3411 old = cpu_cache_get(new->cachep);
3413 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3414 new->new[smp_processor_id()] = old;
3417 static int do_tune_cpucache(struct kmem_cache *cachep, int limit, int batchcount,
3420 struct ccupdate_struct new;
3423 memset(&new.new, 0, sizeof(new.new));
3424 for_each_online_cpu(i) {
3426 alloc_arraycache(cpu_to_node(i), limit, batchcount);
3428 for (i--; i >= 0; i--)
3433 new.cachep = cachep;
3435 smp_call_function_all_cpus(do_ccupdate_local, (void *)&new);
3438 spin_lock(&cachep->spinlock);
3439 cachep->batchcount = batchcount;
3440 cachep->limit = limit;
3441 cachep->shared = shared;
3442 spin_unlock(&cachep->spinlock);
3444 for_each_online_cpu(i) {
3445 struct array_cache *ccold = new.new[i];
3448 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3449 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3450 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3454 err = alloc_kmemlist(cachep);
3456 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
3457 cachep->name, -err);
3463 static void enable_cpucache(struct kmem_cache *cachep)
3468 /* The head array serves three purposes:
3469 * - create a LIFO ordering, i.e. return objects that are cache-warm
3470 * - reduce the number of spinlock operations.
3471 * - reduce the number of linked list operations on the slab and
3472 * bufctl chains: array operations are cheaper.
3473 * The numbers are guessed, we should auto-tune as described by
3476 if (cachep->buffer_size > 131072)
3478 else if (cachep->buffer_size > PAGE_SIZE)
3480 else if (cachep->buffer_size > 1024)
3482 else if (cachep->buffer_size > 256)
3487 /* Cpu bound tasks (e.g. network routing) can exhibit cpu bound
3488 * allocation behaviour: Most allocs on one cpu, most free operations
3489 * on another cpu. For these cases, an efficient object passing between
3490 * cpus is necessary. This is provided by a shared array. The array
3491 * replaces Bonwick's magazine layer.
3492 * On uniprocessor, it's functionally equivalent (but less efficient)
3493 * to a larger limit. Thus disabled by default.
3497 if (cachep->buffer_size <= PAGE_SIZE)
3502 /* With debugging enabled, large batchcount lead to excessively
3503 * long periods with disabled local interrupts. Limit the
3509 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
3511 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3512 cachep->name, -err);
3515 static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac,
3516 int force, int node)
3520 check_spinlock_acquired_node(cachep, node);
3521 if (ac->touched && !force) {
3523 } else if (ac->avail) {
3524 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3525 if (tofree > ac->avail) {
3526 tofree = (ac->avail + 1) / 2;
3528 free_block(cachep, ac->entry, tofree, node);
3529 ac->avail -= tofree;
3530 memmove(ac->entry, &(ac->entry[tofree]),
3531 sizeof(void *) * ac->avail);
3536 * cache_reap - Reclaim memory from caches.
3537 * @unused: unused parameter
3539 * Called from workqueue/eventd every few seconds.
3541 * - clear the per-cpu caches for this CPU.
3542 * - return freeable pages to the main free memory pool.
3544 * If we cannot acquire the cache chain mutex then just give up - we'll
3545 * try again on the next iteration.
3547 static void cache_reap(void *unused)
3549 struct list_head *walk;
3550 struct kmem_list3 *l3;
3552 if (!mutex_trylock(&cache_chain_mutex)) {
3553 /* Give up. Setup the next iteration. */
3554 schedule_delayed_work(&__get_cpu_var(reap_work),
3559 list_for_each(walk, &cache_chain) {
3560 struct kmem_cache *searchp;
3561 struct list_head *p;
3565 searchp = list_entry(walk, struct kmem_cache, next);
3567 if (searchp->flags & SLAB_NO_REAP)
3572 l3 = searchp->nodelists[numa_node_id()];
3573 reap_alien(searchp, l3);
3574 spin_lock_irq(&l3->list_lock);
3576 drain_array_locked(searchp, cpu_cache_get(searchp), 0,
3579 if (time_after(l3->next_reap, jiffies))
3582 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
3585 drain_array_locked(searchp, l3->shared, 0,
3588 if (l3->free_touched) {
3589 l3->free_touched = 0;
3594 (l3->free_limit + 5 * searchp->num -
3595 1) / (5 * searchp->num);
3597 p = l3->slabs_free.next;
3598 if (p == &(l3->slabs_free))
3601 slabp = list_entry(p, struct slab, list);
3602 BUG_ON(slabp->inuse);
3603 list_del(&slabp->list);
3604 STATS_INC_REAPED(searchp);
3606 /* Safe to drop the lock. The slab is no longer
3607 * linked to the cache.
3608 * searchp cannot disappear, we hold
3611 l3->free_objects -= searchp->num;
3612 spin_unlock_irq(&l3->list_lock);
3613 slab_destroy(searchp, slabp);
3614 spin_lock_irq(&l3->list_lock);
3615 } while (--tofree > 0);
3617 spin_unlock_irq(&l3->list_lock);
3622 mutex_unlock(&cache_chain_mutex);
3624 /* Setup the next iteration */
3625 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
3628 #ifdef CONFIG_PROC_FS
3630 static void print_slabinfo_header(struct seq_file *m)
3633 * Output format version, so at least we can change it
3634 * without _too_ many complaints.
3637 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3639 seq_puts(m, "slabinfo - version: 2.1\n");
3641 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3642 "<objperslab> <pagesperslab>");
3643 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3644 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3646 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3647 "<error> <maxfreeable> <nodeallocs> <remotefrees>");
3648 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3653 static void *s_start(struct seq_file *m, loff_t *pos)
3656 struct list_head *p;
3658 mutex_lock(&cache_chain_mutex);
3660 print_slabinfo_header(m);
3661 p = cache_chain.next;
3664 if (p == &cache_chain)
3667 return list_entry(p, struct kmem_cache, next);
3670 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3672 struct kmem_cache *cachep = p;
3674 return cachep->next.next == &cache_chain ? NULL
3675 : list_entry(cachep->next.next, struct kmem_cache, next);
3678 static void s_stop(struct seq_file *m, void *p)
3680 mutex_unlock(&cache_chain_mutex);
3683 static int s_show(struct seq_file *m, void *p)
3685 struct kmem_cache *cachep = p;
3686 struct list_head *q;
3688 unsigned long active_objs;
3689 unsigned long num_objs;
3690 unsigned long active_slabs = 0;
3691 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3695 struct kmem_list3 *l3;
3697 spin_lock(&cachep->spinlock);
3700 for_each_online_node(node) {
3701 l3 = cachep->nodelists[node];
3706 spin_lock_irq(&l3->list_lock);
3708 list_for_each(q, &l3->slabs_full) {
3709 slabp = list_entry(q, struct slab, list);
3710 if (slabp->inuse != cachep->num && !error)
3711 error = "slabs_full accounting error";
3712 active_objs += cachep->num;
3715 list_for_each(q, &l3->slabs_partial) {
3716 slabp = list_entry(q, struct slab, list);
3717 if (slabp->inuse == cachep->num && !error)
3718 error = "slabs_partial inuse accounting error";
3719 if (!slabp->inuse && !error)
3720 error = "slabs_partial/inuse accounting error";
3721 active_objs += slabp->inuse;
3724 list_for_each(q, &l3->slabs_free) {
3725 slabp = list_entry(q, struct slab, list);
3726 if (slabp->inuse && !error)
3727 error = "slabs_free/inuse accounting error";
3730 free_objects += l3->free_objects;
3732 shared_avail += l3->shared->avail;
3734 spin_unlock_irq(&l3->list_lock);
3736 num_slabs += active_slabs;
3737 num_objs = num_slabs * cachep->num;
3738 if (num_objs - active_objs != free_objects && !error)
3739 error = "free_objects accounting error";
3741 name = cachep->name;
3743 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3745 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
3746 name, active_objs, num_objs, cachep->buffer_size,
3747 cachep->num, (1 << cachep->gfporder));
3748 seq_printf(m, " : tunables %4u %4u %4u",
3749 cachep->limit, cachep->batchcount, cachep->shared);
3750 seq_printf(m, " : slabdata %6lu %6lu %6lu",
3751 active_slabs, num_slabs, shared_avail);
3754 unsigned long high = cachep->high_mark;
3755 unsigned long allocs = cachep->num_allocations;
3756 unsigned long grown = cachep->grown;
3757 unsigned long reaped = cachep->reaped;
3758 unsigned long errors = cachep->errors;
3759 unsigned long max_freeable = cachep->max_freeable;
3760 unsigned long node_allocs = cachep->node_allocs;
3761 unsigned long node_frees = cachep->node_frees;
3763 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
3764 %4lu %4lu %4lu %4lu", allocs, high, grown, reaped, errors, max_freeable, node_allocs, node_frees);
3768 unsigned long allochit = atomic_read(&cachep->allochit);
3769 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3770 unsigned long freehit = atomic_read(&cachep->freehit);
3771 unsigned long freemiss = atomic_read(&cachep->freemiss);
3773 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
3774 allochit, allocmiss, freehit, freemiss);
3778 spin_unlock(&cachep->spinlock);
3783 * slabinfo_op - iterator that generates /proc/slabinfo
3792 * num-pages-per-slab
3793 * + further values on SMP and with statistics enabled
3796 struct seq_operations slabinfo_op = {
3803 #define MAX_SLABINFO_WRITE 128
3805 * slabinfo_write - Tuning for the slab allocator
3807 * @buffer: user buffer
3808 * @count: data length
3811 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
3812 size_t count, loff_t *ppos)
3814 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
3815 int limit, batchcount, shared, res;
3816 struct list_head *p;
3818 if (count > MAX_SLABINFO_WRITE)
3820 if (copy_from_user(&kbuf, buffer, count))
3822 kbuf[MAX_SLABINFO_WRITE] = '\0';
3824 tmp = strchr(kbuf, ' ');
3829 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3832 /* Find the cache in the chain of caches. */
3833 mutex_lock(&cache_chain_mutex);
3835 list_for_each(p, &cache_chain) {
3836 struct kmem_cache *cachep = list_entry(p, struct kmem_cache,
3839 if (!strcmp(cachep->name, kbuf)) {
3842 batchcount > limit || shared < 0) {
3845 res = do_tune_cpucache(cachep, limit,
3846 batchcount, shared);
3851 mutex_unlock(&cache_chain_mutex);
3859 * ksize - get the actual amount of memory allocated for a given object
3860 * @objp: Pointer to the object
3862 * kmalloc may internally round up allocations and return more memory
3863 * than requested. ksize() can be used to determine the actual amount of
3864 * memory allocated. The caller may use this additional memory, even though
3865 * a smaller amount of memory was initially specified with the kmalloc call.
3866 * The caller must guarantee that objp points to a valid object previously
3867 * allocated with either kmalloc() or kmem_cache_alloc(). The object
3868 * must not be freed during the duration of the call.
3870 unsigned int ksize(const void *objp)
3872 if (unlikely(objp == NULL))
3875 return obj_size(virt_to_cache(objp));