3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same intializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/config.h>
90 #include <linux/slab.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/seq_file.h>
98 #include <linux/notifier.h>
99 #include <linux/kallsyms.h>
100 #include <linux/cpu.h>
101 #include <linux/sysctl.h>
102 #include <linux/module.h>
103 #include <linux/rcupdate.h>
104 #include <linux/string.h>
105 #include <linux/nodemask.h>
106 #include <linux/mempolicy.h>
107 #include <linux/mutex.h>
109 #include <asm/uaccess.h>
110 #include <asm/cacheflush.h>
111 #include <asm/tlbflush.h>
112 #include <asm/page.h>
115 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
116 * SLAB_RED_ZONE & SLAB_POISON.
117 * 0 for faster, smaller code (especially in the critical paths).
119 * STATS - 1 to collect stats for /proc/slabinfo.
120 * 0 for faster, smaller code (especially in the critical paths).
122 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
125 #ifdef CONFIG_DEBUG_SLAB
128 #define FORCED_DEBUG 1
132 #define FORCED_DEBUG 0
135 /* Shouldn't this be in a header file somewhere? */
136 #define BYTES_PER_WORD sizeof(void *)
138 #ifndef cache_line_size
139 #define cache_line_size() L1_CACHE_BYTES
142 #ifndef ARCH_KMALLOC_MINALIGN
144 * Enforce a minimum alignment for the kmalloc caches.
145 * Usually, the kmalloc caches are cache_line_size() aligned, except when
146 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
147 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
148 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
149 * Note that this flag disables some debug features.
151 #define ARCH_KMALLOC_MINALIGN 0
154 #ifndef ARCH_SLAB_MINALIGN
156 * Enforce a minimum alignment for all caches.
157 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
158 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
159 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
160 * some debug features.
162 #define ARCH_SLAB_MINALIGN 0
165 #ifndef ARCH_KMALLOC_FLAGS
166 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
169 /* Legal flag mask for kmem_cache_create(). */
171 # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
172 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
173 SLAB_NO_REAP | SLAB_CACHE_DMA | \
174 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
175 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
178 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | SLAB_NO_REAP | \
179 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
180 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
187 * Bufctl's are used for linking objs within a slab
190 * This implementation relies on "struct page" for locating the cache &
191 * slab an object belongs to.
192 * This allows the bufctl structure to be small (one int), but limits
193 * the number of objects a slab (not a cache) can contain when off-slab
194 * bufctls are used. The limit is the size of the largest general cache
195 * that does not use off-slab slabs.
196 * For 32bit archs with 4 kB pages, is this 56.
197 * This is not serious, as it is only for large objects, when it is unwise
198 * to have too many per slab.
199 * Note: This limit can be raised by introducing a general cache whose size
200 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
203 typedef unsigned int kmem_bufctl_t;
204 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
205 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
206 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-2)
208 /* Max number of objs-per-slab for caches which use off-slab slabs.
209 * Needed to avoid a possible looping condition in cache_grow().
211 static unsigned long offslab_limit;
216 * Manages the objs in a slab. Placed either at the beginning of mem allocated
217 * for a slab, or allocated from an general cache.
218 * Slabs are chained into three list: fully used, partial, fully free slabs.
221 struct list_head list;
222 unsigned long colouroff;
223 void *s_mem; /* including colour offset */
224 unsigned int inuse; /* num of objs active in slab */
226 unsigned short nodeid;
232 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
233 * arrange for kmem_freepages to be called via RCU. This is useful if
234 * we need to approach a kernel structure obliquely, from its address
235 * obtained without the usual locking. We can lock the structure to
236 * stabilize it and check it's still at the given address, only if we
237 * can be sure that the memory has not been meanwhile reused for some
238 * other kind of object (which our subsystem's lock might corrupt).
240 * rcu_read_lock before reading the address, then rcu_read_unlock after
241 * taking the spinlock within the structure expected at that address.
243 * We assume struct slab_rcu can overlay struct slab when destroying.
246 struct rcu_head head;
247 struct kmem_cache *cachep;
255 * - LIFO ordering, to hand out cache-warm objects from _alloc
256 * - reduce the number of linked list operations
257 * - reduce spinlock operations
259 * The limit is stored in the per-cpu structure to reduce the data cache
266 unsigned int batchcount;
267 unsigned int touched;
270 * Must have this definition in here for the proper
271 * alignment of array_cache. Also simplifies accessing
273 * [0] is for gcc 2.95. It should really be [].
277 /* bootstrap: The caches do not work without cpuarrays anymore,
278 * but the cpuarrays are allocated from the generic caches...
280 #define BOOT_CPUCACHE_ENTRIES 1
281 struct arraycache_init {
282 struct array_cache cache;
283 void *entries[BOOT_CPUCACHE_ENTRIES];
287 * The slab lists for all objects.
290 struct list_head slabs_partial; /* partial list first, better asm code */
291 struct list_head slabs_full;
292 struct list_head slabs_free;
293 unsigned long free_objects;
294 unsigned long next_reap;
296 unsigned int free_limit;
297 unsigned int colour_next; /* Per-node cache coloring */
298 spinlock_t list_lock;
299 struct array_cache *shared; /* shared per node */
300 struct array_cache **alien; /* on other nodes */
304 * Need this for bootstrapping a per node allocator.
306 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
307 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
308 #define CACHE_CACHE 0
310 #define SIZE_L3 (1 + MAX_NUMNODES)
313 * This function must be completely optimized away if
314 * a constant is passed to it. Mostly the same as
315 * what is in linux/slab.h except it returns an
318 static __always_inline int index_of(const size_t size)
320 extern void __bad_size(void);
322 if (__builtin_constant_p(size)) {
330 #include "linux/kmalloc_sizes.h"
338 #define INDEX_AC index_of(sizeof(struct arraycache_init))
339 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
341 static void kmem_list3_init(struct kmem_list3 *parent)
343 INIT_LIST_HEAD(&parent->slabs_full);
344 INIT_LIST_HEAD(&parent->slabs_partial);
345 INIT_LIST_HEAD(&parent->slabs_free);
346 parent->shared = NULL;
347 parent->alien = NULL;
348 parent->colour_next = 0;
349 spin_lock_init(&parent->list_lock);
350 parent->free_objects = 0;
351 parent->free_touched = 0;
354 #define MAKE_LIST(cachep, listp, slab, nodeid) \
356 INIT_LIST_HEAD(listp); \
357 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
360 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
362 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
363 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
364 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
374 /* 1) per-cpu data, touched during every alloc/free */
375 struct array_cache *array[NR_CPUS];
376 unsigned int batchcount;
379 unsigned int buffer_size;
380 /* 2) touched by every alloc & free from the backend */
381 struct kmem_list3 *nodelists[MAX_NUMNODES];
382 unsigned int flags; /* constant flags */
383 unsigned int num; /* # of objs per slab */
386 /* 3) cache_grow/shrink */
387 /* order of pgs per slab (2^n) */
388 unsigned int gfporder;
390 /* force GFP flags, e.g. GFP_DMA */
393 size_t colour; /* cache colouring range */
394 unsigned int colour_off; /* colour offset */
395 struct kmem_cache *slabp_cache;
396 unsigned int slab_size;
397 unsigned int dflags; /* dynamic flags */
399 /* constructor func */
400 void (*ctor) (void *, struct kmem_cache *, unsigned long);
402 /* de-constructor func */
403 void (*dtor) (void *, struct kmem_cache *, unsigned long);
405 /* 4) cache creation/removal */
407 struct list_head next;
411 unsigned long num_active;
412 unsigned long num_allocations;
413 unsigned long high_mark;
415 unsigned long reaped;
416 unsigned long errors;
417 unsigned long max_freeable;
418 unsigned long node_allocs;
419 unsigned long node_frees;
427 * If debugging is enabled, then the allocator can add additional
428 * fields and/or padding to every object. buffer_size contains the total
429 * object size including these internal fields, the following two
430 * variables contain the offset to the user object and its size.
437 #define CFLGS_OFF_SLAB (0x80000000UL)
438 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
440 #define BATCHREFILL_LIMIT 16
441 /* Optimization question: fewer reaps means less
442 * probability for unnessary cpucache drain/refill cycles.
444 * OTOH the cpuarrays can contain lots of objects,
445 * which could lock up otherwise freeable slabs.
447 #define REAPTIMEOUT_CPUC (2*HZ)
448 #define REAPTIMEOUT_LIST3 (4*HZ)
451 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
452 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
453 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
454 #define STATS_INC_GROWN(x) ((x)->grown++)
455 #define STATS_INC_REAPED(x) ((x)->reaped++)
456 #define STATS_SET_HIGH(x) do { if ((x)->num_active > (x)->high_mark) \
457 (x)->high_mark = (x)->num_active; \
459 #define STATS_INC_ERR(x) ((x)->errors++)
460 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
461 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
462 #define STATS_SET_FREEABLE(x, i) \
463 do { if ((x)->max_freeable < i) \
464 (x)->max_freeable = i; \
467 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
468 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
469 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
470 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
472 #define STATS_INC_ACTIVE(x) do { } while (0)
473 #define STATS_DEC_ACTIVE(x) do { } while (0)
474 #define STATS_INC_ALLOCED(x) do { } while (0)
475 #define STATS_INC_GROWN(x) do { } while (0)
476 #define STATS_INC_REAPED(x) do { } while (0)
477 #define STATS_SET_HIGH(x) do { } while (0)
478 #define STATS_INC_ERR(x) do { } while (0)
479 #define STATS_INC_NODEALLOCS(x) do { } while (0)
480 #define STATS_INC_NODEFREES(x) do { } while (0)
481 #define STATS_SET_FREEABLE(x, i) \
484 #define STATS_INC_ALLOCHIT(x) do { } while (0)
485 #define STATS_INC_ALLOCMISS(x) do { } while (0)
486 #define STATS_INC_FREEHIT(x) do { } while (0)
487 #define STATS_INC_FREEMISS(x) do { } while (0)
491 /* Magic nums for obj red zoning.
492 * Placed in the first word before and the first word after an obj.
494 #define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
495 #define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
497 /* ...and for poisoning */
498 #define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
499 #define POISON_FREE 0x6b /* for use-after-free poisoning */
500 #define POISON_END 0xa5 /* end-byte of poisoning */
502 /* memory layout of objects:
504 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
505 * the end of an object is aligned with the end of the real
506 * allocation. Catches writes behind the end of the allocation.
507 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
509 * cachep->obj_offset: The real object.
510 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
511 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address [BYTES_PER_WORD long]
513 static int obj_offset(struct kmem_cache *cachep)
515 return cachep->obj_offset;
518 static int obj_size(struct kmem_cache *cachep)
520 return cachep->obj_size;
523 static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
525 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
526 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD);
529 static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
531 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
532 if (cachep->flags & SLAB_STORE_USER)
533 return (unsigned long *)(objp + cachep->buffer_size -
535 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD);
538 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
540 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
541 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
546 #define obj_offset(x) 0
547 #define obj_size(cachep) (cachep->buffer_size)
548 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
549 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
550 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
555 * Maximum size of an obj (in 2^order pages)
556 * and absolute limit for the gfp order.
558 #if defined(CONFIG_LARGE_ALLOCS)
559 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
560 #define MAX_GFP_ORDER 13 /* up to 32Mb */
561 #elif defined(CONFIG_MMU)
562 #define MAX_OBJ_ORDER 5 /* 32 pages */
563 #define MAX_GFP_ORDER 5 /* 32 pages */
565 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
566 #define MAX_GFP_ORDER 8 /* up to 1Mb */
570 * Do not go above this order unless 0 objects fit into the slab.
572 #define BREAK_GFP_ORDER_HI 1
573 #define BREAK_GFP_ORDER_LO 0
574 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
576 /* Functions for storing/retrieving the cachep and or slab from the
577 * global 'mem_map'. These are used to find the slab an obj belongs to.
578 * With kfree(), these are used to find the cache which an obj belongs to.
580 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
582 page->lru.next = (struct list_head *)cache;
585 static inline struct kmem_cache *page_get_cache(struct page *page)
587 return (struct kmem_cache *)page->lru.next;
590 static inline void page_set_slab(struct page *page, struct slab *slab)
592 page->lru.prev = (struct list_head *)slab;
595 static inline struct slab *page_get_slab(struct page *page)
597 return (struct slab *)page->lru.prev;
600 static inline struct kmem_cache *virt_to_cache(const void *obj)
602 struct page *page = virt_to_page(obj);
603 return page_get_cache(page);
606 static inline struct slab *virt_to_slab(const void *obj)
608 struct page *page = virt_to_page(obj);
609 return page_get_slab(page);
612 /* These are the default caches for kmalloc. Custom caches can have other sizes. */
613 struct cache_sizes malloc_sizes[] = {
614 #define CACHE(x) { .cs_size = (x) },
615 #include <linux/kmalloc_sizes.h>
619 EXPORT_SYMBOL(malloc_sizes);
621 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
627 static struct cache_names __initdata cache_names[] = {
628 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
629 #include <linux/kmalloc_sizes.h>
634 static struct arraycache_init initarray_cache __initdata =
635 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
636 static struct arraycache_init initarray_generic =
637 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
639 /* internal cache of cache description objs */
640 static struct kmem_cache cache_cache = {
642 .limit = BOOT_CPUCACHE_ENTRIES,
644 .buffer_size = sizeof(struct kmem_cache),
645 .flags = SLAB_NO_REAP,
646 .spinlock = SPIN_LOCK_UNLOCKED,
647 .name = "kmem_cache",
649 .obj_size = sizeof(struct kmem_cache),
653 /* Guard access to the cache-chain. */
654 static DEFINE_MUTEX(cache_chain_mutex);
655 static struct list_head cache_chain;
658 * vm_enough_memory() looks at this to determine how many
659 * slab-allocated pages are possibly freeable under pressure
661 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
663 atomic_t slab_reclaim_pages;
666 * chicken and egg problem: delay the per-cpu array allocation
667 * until the general caches are up.
676 static DEFINE_PER_CPU(struct work_struct, reap_work);
678 static void free_block(struct kmem_cache *cachep, void **objpp, int len, int node);
679 static void enable_cpucache(struct kmem_cache *cachep);
680 static void cache_reap(void *unused);
681 static int __node_shrink(struct kmem_cache *cachep, int node);
683 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
685 return cachep->array[smp_processor_id()];
688 static inline struct kmem_cache *__find_general_cachep(size_t size, gfp_t gfpflags)
690 struct cache_sizes *csizep = malloc_sizes;
693 /* This happens if someone tries to call
694 * kmem_cache_create(), or __kmalloc(), before
695 * the generic caches are initialized.
697 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
699 while (size > csizep->cs_size)
703 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
704 * has cs_{dma,}cachep==NULL. Thus no special case
705 * for large kmalloc calls required.
707 if (unlikely(gfpflags & GFP_DMA))
708 return csizep->cs_dmacachep;
709 return csizep->cs_cachep;
712 struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
714 return __find_general_cachep(size, gfpflags);
716 EXPORT_SYMBOL(kmem_find_general_cachep);
718 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
720 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
723 /* Calculate the number of objects and left-over bytes for a given
725 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
726 size_t align, int flags, size_t *left_over,
731 size_t slab_size = PAGE_SIZE << gfporder;
734 * The slab management structure can be either off the slab or
735 * on it. For the latter case, the memory allocated for a
739 * - One kmem_bufctl_t for each object
740 * - Padding to respect alignment of @align
741 * - @buffer_size bytes for each object
743 * If the slab management structure is off the slab, then the
744 * alignment will already be calculated into the size. Because
745 * the slabs are all pages aligned, the objects will be at the
746 * correct alignment when allocated.
748 if (flags & CFLGS_OFF_SLAB) {
750 nr_objs = slab_size / buffer_size;
752 if (nr_objs > SLAB_LIMIT)
753 nr_objs = SLAB_LIMIT;
756 * Ignore padding for the initial guess. The padding
757 * is at most @align-1 bytes, and @buffer_size is at
758 * least @align. In the worst case, this result will
759 * be one greater than the number of objects that fit
760 * into the memory allocation when taking the padding
763 nr_objs = (slab_size - sizeof(struct slab)) /
764 (buffer_size + sizeof(kmem_bufctl_t));
767 * This calculated number will be either the right
768 * amount, or one greater than what we want.
770 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
774 if (nr_objs > SLAB_LIMIT)
775 nr_objs = SLAB_LIMIT;
777 mgmt_size = slab_mgmt_size(nr_objs, align);
780 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
783 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
785 static void __slab_error(const char *function, struct kmem_cache *cachep, char *msg)
787 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
788 function, cachep->name, msg);
794 * Special reaping functions for NUMA systems called from cache_reap().
795 * These take care of doing round robin flushing of alien caches (containing
796 * objects freed on different nodes from which they were allocated) and the
797 * flushing of remote pcps by calling drain_node_pages.
799 static DEFINE_PER_CPU(unsigned long, reap_node);
801 static void init_reap_node(int cpu)
805 node = next_node(cpu_to_node(cpu), node_online_map);
806 if (node == MAX_NUMNODES)
809 __get_cpu_var(reap_node) = node;
812 static void next_reap_node(void)
814 int node = __get_cpu_var(reap_node);
817 * Also drain per cpu pages on remote zones
819 if (node != numa_node_id())
820 drain_node_pages(node);
822 node = next_node(node, node_online_map);
823 if (unlikely(node >= MAX_NUMNODES))
824 node = first_node(node_online_map);
825 __get_cpu_var(reap_node) = node;
829 #define init_reap_node(cpu) do { } while (0)
830 #define next_reap_node(void) do { } while (0)
834 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
835 * via the workqueue/eventd.
836 * Add the CPU number into the expiration time to minimize the possibility of
837 * the CPUs getting into lockstep and contending for the global cache chain
840 static void __devinit start_cpu_timer(int cpu)
842 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
845 * When this gets called from do_initcalls via cpucache_init(),
846 * init_workqueues() has already run, so keventd will be setup
849 if (keventd_up() && reap_work->func == NULL) {
851 INIT_WORK(reap_work, cache_reap, NULL);
852 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
856 static struct array_cache *alloc_arraycache(int node, int entries,
859 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
860 struct array_cache *nc = NULL;
862 nc = kmalloc_node(memsize, GFP_KERNEL, node);
866 nc->batchcount = batchcount;
868 spin_lock_init(&nc->lock);
874 static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int);
876 static struct array_cache **alloc_alien_cache(int node, int limit)
878 struct array_cache **ac_ptr;
879 int memsize = sizeof(void *) * MAX_NUMNODES;
884 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
887 if (i == node || !node_online(i)) {
891 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
893 for (i--; i <= 0; i--)
903 static void free_alien_cache(struct array_cache **ac_ptr)
916 static void __drain_alien_cache(struct kmem_cache *cachep,
917 struct array_cache *ac, int node)
919 struct kmem_list3 *rl3 = cachep->nodelists[node];
922 spin_lock(&rl3->list_lock);
923 free_block(cachep, ac->entry, ac->avail, node);
925 spin_unlock(&rl3->list_lock);
930 * Called from cache_reap() to regularly drain alien caches round robin.
932 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
934 int node = __get_cpu_var(reap_node);
937 struct array_cache *ac = l3->alien[node];
938 if (ac && ac->avail) {
939 spin_lock_irq(&ac->lock);
940 __drain_alien_cache(cachep, ac, node);
941 spin_unlock_irq(&ac->lock);
946 static void drain_alien_cache(struct kmem_cache *cachep, struct array_cache **alien)
949 struct array_cache *ac;
952 for_each_online_node(i) {
955 spin_lock_irqsave(&ac->lock, flags);
956 __drain_alien_cache(cachep, ac, i);
957 spin_unlock_irqrestore(&ac->lock, flags);
963 #define drain_alien_cache(cachep, alien) do { } while (0)
964 #define reap_alien(cachep, l3) do { } while (0)
966 static inline struct array_cache **alloc_alien_cache(int node, int limit)
968 return (struct array_cache **) 0x01020304ul;
971 static inline void free_alien_cache(struct array_cache **ac_ptr)
977 static int __devinit cpuup_callback(struct notifier_block *nfb,
978 unsigned long action, void *hcpu)
980 long cpu = (long)hcpu;
981 struct kmem_cache *cachep;
982 struct kmem_list3 *l3 = NULL;
983 int node = cpu_to_node(cpu);
984 int memsize = sizeof(struct kmem_list3);
988 mutex_lock(&cache_chain_mutex);
989 /* we need to do this right in the beginning since
990 * alloc_arraycache's are going to use this list.
991 * kmalloc_node allows us to add the slab to the right
992 * kmem_list3 and not this cpu's kmem_list3
995 list_for_each_entry(cachep, &cache_chain, next) {
996 /* setup the size64 kmemlist for cpu before we can
997 * begin anything. Make sure some other cpu on this
998 * node has not already allocated this
1000 if (!cachep->nodelists[node]) {
1001 if (!(l3 = kmalloc_node(memsize,
1004 kmem_list3_init(l3);
1005 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1006 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1009 * The l3s don't come and go as CPUs come and
1010 * go. cache_chain_mutex is sufficient
1013 cachep->nodelists[node] = l3;
1016 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1017 cachep->nodelists[node]->free_limit =
1018 (1 + nr_cpus_node(node)) *
1019 cachep->batchcount + cachep->num;
1020 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1023 /* Now we can go ahead with allocating the shared array's
1025 list_for_each_entry(cachep, &cache_chain, next) {
1026 struct array_cache *nc;
1027 struct array_cache *shared;
1028 struct array_cache **alien;
1030 nc = alloc_arraycache(node, cachep->limit,
1031 cachep->batchcount);
1034 shared = alloc_arraycache(node,
1035 cachep->shared * cachep->batchcount,
1040 alien = alloc_alien_cache(node, cachep->limit);
1043 cachep->array[cpu] = nc;
1045 l3 = cachep->nodelists[node];
1048 spin_lock_irq(&l3->list_lock);
1051 * We are serialised from CPU_DEAD or
1052 * CPU_UP_CANCELLED by the cpucontrol lock
1054 l3->shared = shared;
1063 spin_unlock_irq(&l3->list_lock);
1066 free_alien_cache(alien);
1068 mutex_unlock(&cache_chain_mutex);
1071 start_cpu_timer(cpu);
1073 #ifdef CONFIG_HOTPLUG_CPU
1076 * Even if all the cpus of a node are down, we don't free the
1077 * kmem_list3 of any cache. This to avoid a race between
1078 * cpu_down, and a kmalloc allocation from another cpu for
1079 * memory from the node of the cpu going down. The list3
1080 * structure is usually allocated from kmem_cache_create() and
1081 * gets destroyed at kmem_cache_destroy().
1084 case CPU_UP_CANCELED:
1085 mutex_lock(&cache_chain_mutex);
1087 list_for_each_entry(cachep, &cache_chain, next) {
1088 struct array_cache *nc;
1089 struct array_cache *shared;
1090 struct array_cache **alien;
1093 mask = node_to_cpumask(node);
1094 /* cpu is dead; no one can alloc from it. */
1095 nc = cachep->array[cpu];
1096 cachep->array[cpu] = NULL;
1097 l3 = cachep->nodelists[node];
1100 goto free_array_cache;
1102 spin_lock_irq(&l3->list_lock);
1104 /* Free limit for this kmem_list3 */
1105 l3->free_limit -= cachep->batchcount;
1107 free_block(cachep, nc->entry, nc->avail, node);
1109 if (!cpus_empty(mask)) {
1110 spin_unlock_irq(&l3->list_lock);
1111 goto free_array_cache;
1114 shared = l3->shared;
1116 free_block(cachep, l3->shared->entry,
1117 l3->shared->avail, node);
1124 spin_unlock_irq(&l3->list_lock);
1128 drain_alien_cache(cachep, alien);
1129 free_alien_cache(alien);
1135 * In the previous loop, all the objects were freed to
1136 * the respective cache's slabs, now we can go ahead and
1137 * shrink each nodelist to its limit.
1139 list_for_each_entry(cachep, &cache_chain, next) {
1140 l3 = cachep->nodelists[node];
1143 spin_lock_irq(&l3->list_lock);
1144 /* free slabs belonging to this node */
1145 __node_shrink(cachep, node);
1146 spin_unlock_irq(&l3->list_lock);
1148 mutex_unlock(&cache_chain_mutex);
1154 mutex_unlock(&cache_chain_mutex);
1158 static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
1161 * swap the static kmem_list3 with kmalloced memory
1163 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list, int nodeid)
1165 struct kmem_list3 *ptr;
1167 BUG_ON(cachep->nodelists[nodeid] != list);
1168 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1171 local_irq_disable();
1172 memcpy(ptr, list, sizeof(struct kmem_list3));
1173 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1174 cachep->nodelists[nodeid] = ptr;
1179 * Called after the gfp() functions have been enabled, and before smp_init().
1181 void __init kmem_cache_init(void)
1184 struct cache_sizes *sizes;
1185 struct cache_names *names;
1189 for (i = 0; i < NUM_INIT_LISTS; i++) {
1190 kmem_list3_init(&initkmem_list3[i]);
1191 if (i < MAX_NUMNODES)
1192 cache_cache.nodelists[i] = NULL;
1196 * Fragmentation resistance on low memory - only use bigger
1197 * page orders on machines with more than 32MB of memory.
1199 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1200 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1202 /* Bootstrap is tricky, because several objects are allocated
1203 * from caches that do not exist yet:
1204 * 1) initialize the cache_cache cache: it contains the struct kmem_cache
1205 * structures of all caches, except cache_cache itself: cache_cache
1206 * is statically allocated.
1207 * Initially an __init data area is used for the head array and the
1208 * kmem_list3 structures, it's replaced with a kmalloc allocated
1209 * array at the end of the bootstrap.
1210 * 2) Create the first kmalloc cache.
1211 * The struct kmem_cache for the new cache is allocated normally.
1212 * An __init data area is used for the head array.
1213 * 3) Create the remaining kmalloc caches, with minimally sized
1215 * 4) Replace the __init data head arrays for cache_cache and the first
1216 * kmalloc cache with kmalloc allocated arrays.
1217 * 5) Replace the __init data for kmem_list3 for cache_cache and
1218 * the other cache's with kmalloc allocated memory.
1219 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1222 /* 1) create the cache_cache */
1223 INIT_LIST_HEAD(&cache_chain);
1224 list_add(&cache_cache.next, &cache_chain);
1225 cache_cache.colour_off = cache_line_size();
1226 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1227 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
1229 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, cache_line_size());
1231 for (order = 0; order < MAX_ORDER; order++) {
1232 cache_estimate(order, cache_cache.buffer_size,
1233 cache_line_size(), 0, &left_over, &cache_cache.num);
1234 if (cache_cache.num)
1237 if (!cache_cache.num)
1239 cache_cache.gfporder = order;
1240 cache_cache.colour = left_over / cache_cache.colour_off;
1241 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1242 sizeof(struct slab), cache_line_size());
1244 /* 2+3) create the kmalloc caches */
1245 sizes = malloc_sizes;
1246 names = cache_names;
1248 /* Initialize the caches that provide memory for the array cache
1249 * and the kmem_list3 structures first.
1250 * Without this, further allocations will bug
1253 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1254 sizes[INDEX_AC].cs_size,
1255 ARCH_KMALLOC_MINALIGN,
1256 (ARCH_KMALLOC_FLAGS |
1257 SLAB_PANIC), NULL, NULL);
1259 if (INDEX_AC != INDEX_L3)
1260 sizes[INDEX_L3].cs_cachep =
1261 kmem_cache_create(names[INDEX_L3].name,
1262 sizes[INDEX_L3].cs_size,
1263 ARCH_KMALLOC_MINALIGN,
1264 (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL,
1267 while (sizes->cs_size != ULONG_MAX) {
1269 * For performance, all the general caches are L1 aligned.
1270 * This should be particularly beneficial on SMP boxes, as it
1271 * eliminates "false sharing".
1272 * Note for systems short on memory removing the alignment will
1273 * allow tighter packing of the smaller caches.
1275 if (!sizes->cs_cachep)
1276 sizes->cs_cachep = kmem_cache_create(names->name,
1278 ARCH_KMALLOC_MINALIGN,
1283 /* Inc off-slab bufctl limit until the ceiling is hit. */
1284 if (!(OFF_SLAB(sizes->cs_cachep))) {
1285 offslab_limit = sizes->cs_size - sizeof(struct slab);
1286 offslab_limit /= sizeof(kmem_bufctl_t);
1289 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
1291 ARCH_KMALLOC_MINALIGN,
1292 (ARCH_KMALLOC_FLAGS |
1300 /* 4) Replace the bootstrap head arrays */
1304 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1306 local_irq_disable();
1307 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1308 memcpy(ptr, cpu_cache_get(&cache_cache),
1309 sizeof(struct arraycache_init));
1310 cache_cache.array[smp_processor_id()] = ptr;
1313 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1315 local_irq_disable();
1316 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1317 != &initarray_generic.cache);
1318 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1319 sizeof(struct arraycache_init));
1320 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1324 /* 5) Replace the bootstrap kmem_list3's */
1327 /* Replace the static kmem_list3 structures for the boot cpu */
1328 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
1331 for_each_online_node(node) {
1332 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1333 &initkmem_list3[SIZE_AC + node], node);
1335 if (INDEX_AC != INDEX_L3) {
1336 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1337 &initkmem_list3[SIZE_L3 + node],
1343 /* 6) resize the head arrays to their final sizes */
1345 struct kmem_cache *cachep;
1346 mutex_lock(&cache_chain_mutex);
1347 list_for_each_entry(cachep, &cache_chain, next)
1348 enable_cpucache(cachep);
1349 mutex_unlock(&cache_chain_mutex);
1353 g_cpucache_up = FULL;
1355 /* Register a cpu startup notifier callback
1356 * that initializes cpu_cache_get for all new cpus
1358 register_cpu_notifier(&cpucache_notifier);
1360 /* The reap timers are started later, with a module init call:
1361 * That part of the kernel is not yet operational.
1365 static int __init cpucache_init(void)
1370 * Register the timers that return unneeded
1373 for_each_online_cpu(cpu)
1374 start_cpu_timer(cpu);
1379 __initcall(cpucache_init);
1382 * Interface to system's page allocator. No need to hold the cache-lock.
1384 * If we requested dmaable memory, we will get it. Even if we
1385 * did not request dmaable memory, we might get it, but that
1386 * would be relatively rare and ignorable.
1388 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1394 flags |= cachep->gfpflags;
1395 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1398 addr = page_address(page);
1400 i = (1 << cachep->gfporder);
1401 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1402 atomic_add(i, &slab_reclaim_pages);
1403 add_page_state(nr_slab, i);
1412 * Interface to system's page release.
1414 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1416 unsigned long i = (1 << cachep->gfporder);
1417 struct page *page = virt_to_page(addr);
1418 const unsigned long nr_freed = i;
1421 if (!TestClearPageSlab(page))
1425 sub_page_state(nr_slab, nr_freed);
1426 if (current->reclaim_state)
1427 current->reclaim_state->reclaimed_slab += nr_freed;
1428 free_pages((unsigned long)addr, cachep->gfporder);
1429 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1430 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages);
1433 static void kmem_rcu_free(struct rcu_head *head)
1435 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1436 struct kmem_cache *cachep = slab_rcu->cachep;
1438 kmem_freepages(cachep, slab_rcu->addr);
1439 if (OFF_SLAB(cachep))
1440 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1445 #ifdef CONFIG_DEBUG_PAGEALLOC
1446 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1447 unsigned long caller)
1449 int size = obj_size(cachep);
1451 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1453 if (size < 5 * sizeof(unsigned long))
1456 *addr++ = 0x12345678;
1458 *addr++ = smp_processor_id();
1459 size -= 3 * sizeof(unsigned long);
1461 unsigned long *sptr = &caller;
1462 unsigned long svalue;
1464 while (!kstack_end(sptr)) {
1466 if (kernel_text_address(svalue)) {
1468 size -= sizeof(unsigned long);
1469 if (size <= sizeof(unsigned long))
1475 *addr++ = 0x87654321;
1479 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1481 int size = obj_size(cachep);
1482 addr = &((char *)addr)[obj_offset(cachep)];
1484 memset(addr, val, size);
1485 *(unsigned char *)(addr + size - 1) = POISON_END;
1488 static void dump_line(char *data, int offset, int limit)
1491 printk(KERN_ERR "%03x:", offset);
1492 for (i = 0; i < limit; i++) {
1493 printk(" %02x", (unsigned char)data[offset + i]);
1501 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1506 if (cachep->flags & SLAB_RED_ZONE) {
1507 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
1508 *dbg_redzone1(cachep, objp),
1509 *dbg_redzone2(cachep, objp));
1512 if (cachep->flags & SLAB_STORE_USER) {
1513 printk(KERN_ERR "Last user: [<%p>]",
1514 *dbg_userword(cachep, objp));
1515 print_symbol("(%s)",
1516 (unsigned long)*dbg_userword(cachep, objp));
1519 realobj = (char *)objp + obj_offset(cachep);
1520 size = obj_size(cachep);
1521 for (i = 0; i < size && lines; i += 16, lines--) {
1524 if (i + limit > size)
1526 dump_line(realobj, i, limit);
1530 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1536 realobj = (char *)objp + obj_offset(cachep);
1537 size = obj_size(cachep);
1539 for (i = 0; i < size; i++) {
1540 char exp = POISON_FREE;
1543 if (realobj[i] != exp) {
1549 "Slab corruption: start=%p, len=%d\n",
1551 print_objinfo(cachep, objp, 0);
1553 /* Hexdump the affected line */
1556 if (i + limit > size)
1558 dump_line(realobj, i, limit);
1561 /* Limit to 5 lines */
1567 /* Print some data about the neighboring objects, if they
1570 struct slab *slabp = virt_to_slab(objp);
1573 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
1575 objp = slabp->s_mem + (objnr - 1) * cachep->buffer_size;
1576 realobj = (char *)objp + obj_offset(cachep);
1577 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1579 print_objinfo(cachep, objp, 2);
1581 if (objnr + 1 < cachep->num) {
1582 objp = slabp->s_mem + (objnr + 1) * cachep->buffer_size;
1583 realobj = (char *)objp + obj_offset(cachep);
1584 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1586 print_objinfo(cachep, objp, 2);
1594 * slab_destroy_objs - call the registered destructor for each object in
1595 * a slab that is to be destroyed.
1597 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1600 for (i = 0; i < cachep->num; i++) {
1601 void *objp = slabp->s_mem + cachep->buffer_size * i;
1603 if (cachep->flags & SLAB_POISON) {
1604 #ifdef CONFIG_DEBUG_PAGEALLOC
1605 if ((cachep->buffer_size % PAGE_SIZE) == 0
1606 && OFF_SLAB(cachep))
1607 kernel_map_pages(virt_to_page(objp),
1608 cachep->buffer_size / PAGE_SIZE,
1611 check_poison_obj(cachep, objp);
1613 check_poison_obj(cachep, objp);
1616 if (cachep->flags & SLAB_RED_ZONE) {
1617 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1618 slab_error(cachep, "start of a freed object "
1620 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1621 slab_error(cachep, "end of a freed object "
1624 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1625 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
1629 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1633 for (i = 0; i < cachep->num; i++) {
1634 void *objp = slabp->s_mem + cachep->buffer_size * i;
1635 (cachep->dtor) (objp, cachep, 0);
1642 * Destroy all the objs in a slab, and release the mem back to the system.
1643 * Before calling the slab must have been unlinked from the cache.
1644 * The cache-lock is not held/needed.
1646 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1648 void *addr = slabp->s_mem - slabp->colouroff;
1650 slab_destroy_objs(cachep, slabp);
1651 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1652 struct slab_rcu *slab_rcu;
1654 slab_rcu = (struct slab_rcu *)slabp;
1655 slab_rcu->cachep = cachep;
1656 slab_rcu->addr = addr;
1657 call_rcu(&slab_rcu->head, kmem_rcu_free);
1659 kmem_freepages(cachep, addr);
1660 if (OFF_SLAB(cachep))
1661 kmem_cache_free(cachep->slabp_cache, slabp);
1665 /* For setting up all the kmem_list3s for cache whose buffer_size is same
1666 as size of kmem_list3. */
1667 static void set_up_list3s(struct kmem_cache *cachep, int index)
1671 for_each_online_node(node) {
1672 cachep->nodelists[node] = &initkmem_list3[index + node];
1673 cachep->nodelists[node]->next_reap = jiffies +
1675 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1680 * calculate_slab_order - calculate size (page order) of slabs
1681 * @cachep: pointer to the cache that is being created
1682 * @size: size of objects to be created in this cache.
1683 * @align: required alignment for the objects.
1684 * @flags: slab allocation flags
1686 * Also calculates the number of objects per slab.
1688 * This could be made much more intelligent. For now, try to avoid using
1689 * high order pages for slabs. When the gfp() functions are more friendly
1690 * towards high-order requests, this should be changed.
1692 static inline size_t calculate_slab_order(struct kmem_cache *cachep,
1693 size_t size, size_t align, unsigned long flags)
1695 size_t left_over = 0;
1698 for (gfporder = 0 ; gfporder <= MAX_GFP_ORDER; gfporder++) {
1702 cache_estimate(gfporder, size, align, flags, &remainder, &num);
1706 /* More than offslab_limit objects will cause problems */
1707 if ((flags & CFLGS_OFF_SLAB) && num > offslab_limit)
1710 /* Found something acceptable - save it away */
1712 cachep->gfporder = gfporder;
1713 left_over = remainder;
1716 * A VFS-reclaimable slab tends to have most allocations
1717 * as GFP_NOFS and we really don't want to have to be allocating
1718 * higher-order pages when we are unable to shrink dcache.
1720 if (flags & SLAB_RECLAIM_ACCOUNT)
1724 * Large number of objects is good, but very large slabs are
1725 * currently bad for the gfp()s.
1727 if (gfporder >= slab_break_gfp_order)
1731 * Acceptable internal fragmentation?
1733 if ((left_over * 8) <= (PAGE_SIZE << gfporder))
1740 * kmem_cache_create - Create a cache.
1741 * @name: A string which is used in /proc/slabinfo to identify this cache.
1742 * @size: The size of objects to be created in this cache.
1743 * @align: The required alignment for the objects.
1744 * @flags: SLAB flags
1745 * @ctor: A constructor for the objects.
1746 * @dtor: A destructor for the objects.
1748 * Returns a ptr to the cache on success, NULL on failure.
1749 * Cannot be called within a int, but can be interrupted.
1750 * The @ctor is run when new pages are allocated by the cache
1751 * and the @dtor is run before the pages are handed back.
1753 * @name must be valid until the cache is destroyed. This implies that
1754 * the module calling this has to destroy the cache before getting
1759 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1760 * to catch references to uninitialised memory.
1762 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1763 * for buffer overruns.
1765 * %SLAB_NO_REAP - Don't automatically reap this cache when we're under
1768 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1769 * cacheline. This can be beneficial if you're counting cycles as closely
1773 kmem_cache_create (const char *name, size_t size, size_t align,
1774 unsigned long flags, void (*ctor)(void*, struct kmem_cache *, unsigned long),
1775 void (*dtor)(void*, struct kmem_cache *, unsigned long))
1777 size_t left_over, slab_size, ralign;
1778 struct kmem_cache *cachep = NULL;
1779 struct list_head *p;
1782 * Sanity checks... these are all serious usage bugs.
1786 (size < BYTES_PER_WORD) ||
1787 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
1788 printk(KERN_ERR "%s: Early error in slab %s\n",
1789 __FUNCTION__, name);
1794 * Prevent CPUs from coming and going.
1795 * lock_cpu_hotplug() nests outside cache_chain_mutex
1799 mutex_lock(&cache_chain_mutex);
1801 list_for_each(p, &cache_chain) {
1802 struct kmem_cache *pc = list_entry(p, struct kmem_cache, next);
1803 mm_segment_t old_fs = get_fs();
1808 * This happens when the module gets unloaded and doesn't
1809 * destroy its slab cache and no-one else reuses the vmalloc
1810 * area of the module. Print a warning.
1813 res = __get_user(tmp, pc->name);
1816 printk("SLAB: cache with size %d has lost its name\n",
1821 if (!strcmp(pc->name, name)) {
1822 printk("kmem_cache_create: duplicate cache %s\n", name);
1829 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1830 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
1831 /* No constructor, but inital state check requested */
1832 printk(KERN_ERR "%s: No con, but init state check "
1833 "requested - %s\n", __FUNCTION__, name);
1834 flags &= ~SLAB_DEBUG_INITIAL;
1838 * Enable redzoning and last user accounting, except for caches with
1839 * large objects, if the increased size would increase the object size
1840 * above the next power of two: caches with object sizes just above a
1841 * power of two have a significant amount of internal fragmentation.
1844 || fls(size - 1) == fls(size - 1 + 3 * BYTES_PER_WORD)))
1845 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
1846 if (!(flags & SLAB_DESTROY_BY_RCU))
1847 flags |= SLAB_POISON;
1849 if (flags & SLAB_DESTROY_BY_RCU)
1850 BUG_ON(flags & SLAB_POISON);
1852 if (flags & SLAB_DESTROY_BY_RCU)
1856 * Always checks flags, a caller might be expecting debug
1857 * support which isn't available.
1859 if (flags & ~CREATE_MASK)
1862 /* Check that size is in terms of words. This is needed to avoid
1863 * unaligned accesses for some archs when redzoning is used, and makes
1864 * sure any on-slab bufctl's are also correctly aligned.
1866 if (size & (BYTES_PER_WORD - 1)) {
1867 size += (BYTES_PER_WORD - 1);
1868 size &= ~(BYTES_PER_WORD - 1);
1871 /* calculate out the final buffer alignment: */
1872 /* 1) arch recommendation: can be overridden for debug */
1873 if (flags & SLAB_HWCACHE_ALIGN) {
1874 /* Default alignment: as specified by the arch code.
1875 * Except if an object is really small, then squeeze multiple
1876 * objects into one cacheline.
1878 ralign = cache_line_size();
1879 while (size <= ralign / 2)
1882 ralign = BYTES_PER_WORD;
1884 /* 2) arch mandated alignment: disables debug if necessary */
1885 if (ralign < ARCH_SLAB_MINALIGN) {
1886 ralign = ARCH_SLAB_MINALIGN;
1887 if (ralign > BYTES_PER_WORD)
1888 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
1890 /* 3) caller mandated alignment: disables debug if necessary */
1891 if (ralign < align) {
1893 if (ralign > BYTES_PER_WORD)
1894 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
1896 /* 4) Store it. Note that the debug code below can reduce
1897 * the alignment to BYTES_PER_WORD.
1901 /* Get cache's description obj. */
1902 cachep = kmem_cache_alloc(&cache_cache, SLAB_KERNEL);
1905 memset(cachep, 0, sizeof(struct kmem_cache));
1908 cachep->obj_size = size;
1910 if (flags & SLAB_RED_ZONE) {
1911 /* redzoning only works with word aligned caches */
1912 align = BYTES_PER_WORD;
1914 /* add space for red zone words */
1915 cachep->obj_offset += BYTES_PER_WORD;
1916 size += 2 * BYTES_PER_WORD;
1918 if (flags & SLAB_STORE_USER) {
1919 /* user store requires word alignment and
1920 * one word storage behind the end of the real
1923 align = BYTES_PER_WORD;
1924 size += BYTES_PER_WORD;
1926 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
1927 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
1928 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
1929 cachep->obj_offset += PAGE_SIZE - size;
1935 /* Determine if the slab management is 'on' or 'off' slab. */
1936 if (size >= (PAGE_SIZE >> 3))
1938 * Size is large, assume best to place the slab management obj
1939 * off-slab (should allow better packing of objs).
1941 flags |= CFLGS_OFF_SLAB;
1943 size = ALIGN(size, align);
1945 left_over = calculate_slab_order(cachep, size, align, flags);
1948 printk("kmem_cache_create: couldn't create cache %s.\n", name);
1949 kmem_cache_free(&cache_cache, cachep);
1953 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
1954 + sizeof(struct slab), align);
1957 * If the slab has been placed off-slab, and we have enough space then
1958 * move it on-slab. This is at the expense of any extra colouring.
1960 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
1961 flags &= ~CFLGS_OFF_SLAB;
1962 left_over -= slab_size;
1965 if (flags & CFLGS_OFF_SLAB) {
1966 /* really off slab. No need for manual alignment */
1968 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
1971 cachep->colour_off = cache_line_size();
1972 /* Offset must be a multiple of the alignment. */
1973 if (cachep->colour_off < align)
1974 cachep->colour_off = align;
1975 cachep->colour = left_over / cachep->colour_off;
1976 cachep->slab_size = slab_size;
1977 cachep->flags = flags;
1978 cachep->gfpflags = 0;
1979 if (flags & SLAB_CACHE_DMA)
1980 cachep->gfpflags |= GFP_DMA;
1981 spin_lock_init(&cachep->spinlock);
1982 cachep->buffer_size = size;
1984 if (flags & CFLGS_OFF_SLAB)
1985 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
1986 cachep->ctor = ctor;
1987 cachep->dtor = dtor;
1988 cachep->name = name;
1991 if (g_cpucache_up == FULL) {
1992 enable_cpucache(cachep);
1994 if (g_cpucache_up == NONE) {
1995 /* Note: the first kmem_cache_create must create
1996 * the cache that's used by kmalloc(24), otherwise
1997 * the creation of further caches will BUG().
1999 cachep->array[smp_processor_id()] =
2000 &initarray_generic.cache;
2002 /* If the cache that's used by
2003 * kmalloc(sizeof(kmem_list3)) is the first cache,
2004 * then we need to set up all its list3s, otherwise
2005 * the creation of further caches will BUG().
2007 set_up_list3s(cachep, SIZE_AC);
2008 if (INDEX_AC == INDEX_L3)
2009 g_cpucache_up = PARTIAL_L3;
2011 g_cpucache_up = PARTIAL_AC;
2013 cachep->array[smp_processor_id()] =
2014 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
2016 if (g_cpucache_up == PARTIAL_AC) {
2017 set_up_list3s(cachep, SIZE_L3);
2018 g_cpucache_up = PARTIAL_L3;
2021 for_each_online_node(node) {
2023 cachep->nodelists[node] =
2025 (struct kmem_list3),
2027 BUG_ON(!cachep->nodelists[node]);
2028 kmem_list3_init(cachep->
2033 cachep->nodelists[numa_node_id()]->next_reap =
2034 jiffies + REAPTIMEOUT_LIST3 +
2035 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2037 BUG_ON(!cpu_cache_get(cachep));
2038 cpu_cache_get(cachep)->avail = 0;
2039 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2040 cpu_cache_get(cachep)->batchcount = 1;
2041 cpu_cache_get(cachep)->touched = 0;
2042 cachep->batchcount = 1;
2043 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2046 /* cache setup completed, link it into the list */
2047 list_add(&cachep->next, &cache_chain);
2049 if (!cachep && (flags & SLAB_PANIC))
2050 panic("kmem_cache_create(): failed to create slab `%s'\n",
2052 mutex_unlock(&cache_chain_mutex);
2053 unlock_cpu_hotplug();
2056 EXPORT_SYMBOL(kmem_cache_create);
2059 static void check_irq_off(void)
2061 BUG_ON(!irqs_disabled());
2064 static void check_irq_on(void)
2066 BUG_ON(irqs_disabled());
2069 static void check_spinlock_acquired(struct kmem_cache *cachep)
2073 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2077 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2081 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2086 #define check_irq_off() do { } while(0)
2087 #define check_irq_on() do { } while(0)
2088 #define check_spinlock_acquired(x) do { } while(0)
2089 #define check_spinlock_acquired_node(x, y) do { } while(0)
2093 * Waits for all CPUs to execute func().
2095 static void smp_call_function_all_cpus(void (*func)(void *arg), void *arg)
2100 local_irq_disable();
2104 if (smp_call_function(func, arg, 1, 1))
2110 static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac,
2111 int force, int node);
2113 static void do_drain(void *arg)
2115 struct kmem_cache *cachep = (struct kmem_cache *) arg;
2116 struct array_cache *ac;
2117 int node = numa_node_id();
2120 ac = cpu_cache_get(cachep);
2121 spin_lock(&cachep->nodelists[node]->list_lock);
2122 free_block(cachep, ac->entry, ac->avail, node);
2123 spin_unlock(&cachep->nodelists[node]->list_lock);
2127 static void drain_cpu_caches(struct kmem_cache *cachep)
2129 struct kmem_list3 *l3;
2132 smp_call_function_all_cpus(do_drain, cachep);
2134 for_each_online_node(node) {
2135 l3 = cachep->nodelists[node];
2137 spin_lock_irq(&l3->list_lock);
2138 drain_array_locked(cachep, l3->shared, 1, node);
2139 spin_unlock_irq(&l3->list_lock);
2141 drain_alien_cache(cachep, l3->alien);
2146 static int __node_shrink(struct kmem_cache *cachep, int node)
2149 struct kmem_list3 *l3 = cachep->nodelists[node];
2153 struct list_head *p;
2155 p = l3->slabs_free.prev;
2156 if (p == &l3->slabs_free)
2159 slabp = list_entry(l3->slabs_free.prev, struct slab, list);
2164 list_del(&slabp->list);
2166 l3->free_objects -= cachep->num;
2167 spin_unlock_irq(&l3->list_lock);
2168 slab_destroy(cachep, slabp);
2169 spin_lock_irq(&l3->list_lock);
2171 ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial);
2175 static int __cache_shrink(struct kmem_cache *cachep)
2178 struct kmem_list3 *l3;
2180 drain_cpu_caches(cachep);
2183 for_each_online_node(i) {
2184 l3 = cachep->nodelists[i];
2186 spin_lock_irq(&l3->list_lock);
2187 ret += __node_shrink(cachep, i);
2188 spin_unlock_irq(&l3->list_lock);
2191 return (ret ? 1 : 0);
2195 * kmem_cache_shrink - Shrink a cache.
2196 * @cachep: The cache to shrink.
2198 * Releases as many slabs as possible for a cache.
2199 * To help debugging, a zero exit status indicates all slabs were released.
2201 int kmem_cache_shrink(struct kmem_cache *cachep)
2203 if (!cachep || in_interrupt())
2206 return __cache_shrink(cachep);
2208 EXPORT_SYMBOL(kmem_cache_shrink);
2211 * kmem_cache_destroy - delete a cache
2212 * @cachep: the cache to destroy
2214 * Remove a struct kmem_cache object from the slab cache.
2215 * Returns 0 on success.
2217 * It is expected this function will be called by a module when it is
2218 * unloaded. This will remove the cache completely, and avoid a duplicate
2219 * cache being allocated each time a module is loaded and unloaded, if the
2220 * module doesn't have persistent in-kernel storage across loads and unloads.
2222 * The cache must be empty before calling this function.
2224 * The caller must guarantee that noone will allocate memory from the cache
2225 * during the kmem_cache_destroy().
2227 int kmem_cache_destroy(struct kmem_cache *cachep)
2230 struct kmem_list3 *l3;
2232 if (!cachep || in_interrupt())
2235 /* Don't let CPUs to come and go */
2238 /* Find the cache in the chain of caches. */
2239 mutex_lock(&cache_chain_mutex);
2241 * the chain is never empty, cache_cache is never destroyed
2243 list_del(&cachep->next);
2244 mutex_unlock(&cache_chain_mutex);
2246 if (__cache_shrink(cachep)) {
2247 slab_error(cachep, "Can't free all objects");
2248 mutex_lock(&cache_chain_mutex);
2249 list_add(&cachep->next, &cache_chain);
2250 mutex_unlock(&cache_chain_mutex);
2251 unlock_cpu_hotplug();
2255 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2258 for_each_online_cpu(i)
2259 kfree(cachep->array[i]);
2261 /* NUMA: free the list3 structures */
2262 for_each_online_node(i) {
2263 if ((l3 = cachep->nodelists[i])) {
2265 free_alien_cache(l3->alien);
2269 kmem_cache_free(&cache_cache, cachep);
2271 unlock_cpu_hotplug();
2275 EXPORT_SYMBOL(kmem_cache_destroy);
2277 /* Get the memory for a slab management obj. */
2278 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2279 int colour_off, gfp_t local_flags)
2283 if (OFF_SLAB(cachep)) {
2284 /* Slab management obj is off-slab. */
2285 slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags);
2289 slabp = objp + colour_off;
2290 colour_off += cachep->slab_size;
2293 slabp->colouroff = colour_off;
2294 slabp->s_mem = objp + colour_off;
2299 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2301 return (kmem_bufctl_t *) (slabp + 1);
2304 static void cache_init_objs(struct kmem_cache *cachep,
2305 struct slab *slabp, unsigned long ctor_flags)
2309 for (i = 0; i < cachep->num; i++) {
2310 void *objp = slabp->s_mem + cachep->buffer_size * i;
2312 /* need to poison the objs? */
2313 if (cachep->flags & SLAB_POISON)
2314 poison_obj(cachep, objp, POISON_FREE);
2315 if (cachep->flags & SLAB_STORE_USER)
2316 *dbg_userword(cachep, objp) = NULL;
2318 if (cachep->flags & SLAB_RED_ZONE) {
2319 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2320 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2323 * Constructors are not allowed to allocate memory from
2324 * the same cache which they are a constructor for.
2325 * Otherwise, deadlock. They must also be threaded.
2327 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2328 cachep->ctor(objp + obj_offset(cachep), cachep,
2331 if (cachep->flags & SLAB_RED_ZONE) {
2332 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2333 slab_error(cachep, "constructor overwrote the"
2334 " end of an object");
2335 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2336 slab_error(cachep, "constructor overwrote the"
2337 " start of an object");
2339 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)
2340 && cachep->flags & SLAB_POISON)
2341 kernel_map_pages(virt_to_page(objp),
2342 cachep->buffer_size / PAGE_SIZE, 0);
2345 cachep->ctor(objp, cachep, ctor_flags);
2347 slab_bufctl(slabp)[i] = i + 1;
2349 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2353 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2355 if (flags & SLAB_DMA) {
2356 if (!(cachep->gfpflags & GFP_DMA))
2359 if (cachep->gfpflags & GFP_DMA)
2364 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, int nodeid)
2366 void *objp = slabp->s_mem + (slabp->free * cachep->buffer_size);
2370 next = slab_bufctl(slabp)[slabp->free];
2372 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2373 WARN_ON(slabp->nodeid != nodeid);
2380 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, void *objp,
2383 unsigned int objnr = (unsigned)(objp-slabp->s_mem) / cachep->buffer_size;
2386 /* Verify that the slab belongs to the intended node */
2387 WARN_ON(slabp->nodeid != nodeid);
2389 if (slab_bufctl(slabp)[objnr] != BUFCTL_FREE) {
2390 printk(KERN_ERR "slab: double free detected in cache "
2391 "'%s', objp %p\n", cachep->name, objp);
2395 slab_bufctl(slabp)[objnr] = slabp->free;
2396 slabp->free = objnr;
2400 static void set_slab_attr(struct kmem_cache *cachep, struct slab *slabp, void *objp)
2405 /* Nasty!!!!!! I hope this is OK. */
2406 i = 1 << cachep->gfporder;
2407 page = virt_to_page(objp);
2409 page_set_cache(page, cachep);
2410 page_set_slab(page, slabp);
2416 * Grow (by 1) the number of slabs within a cache. This is called by
2417 * kmem_cache_alloc() when there are no active objs left in a cache.
2419 static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid)
2425 unsigned long ctor_flags;
2426 struct kmem_list3 *l3;
2428 /* Be lazy and only check for valid flags here,
2429 * keeping it out of the critical path in kmem_cache_alloc().
2431 if (flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW))
2433 if (flags & SLAB_NO_GROW)
2436 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2437 local_flags = (flags & SLAB_LEVEL_MASK);
2438 if (!(local_flags & __GFP_WAIT))
2440 * Not allowed to sleep. Need to tell a constructor about
2441 * this - it might need to know...
2443 ctor_flags |= SLAB_CTOR_ATOMIC;
2445 /* Take the l3 list lock to change the colour_next on this node */
2447 l3 = cachep->nodelists[nodeid];
2448 spin_lock(&l3->list_lock);
2450 /* Get colour for the slab, and cal the next value. */
2451 offset = l3->colour_next;
2453 if (l3->colour_next >= cachep->colour)
2454 l3->colour_next = 0;
2455 spin_unlock(&l3->list_lock);
2457 offset *= cachep->colour_off;
2459 if (local_flags & __GFP_WAIT)
2463 * The test for missing atomic flag is performed here, rather than
2464 * the more obvious place, simply to reduce the critical path length
2465 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2466 * will eventually be caught here (where it matters).
2468 kmem_flagcheck(cachep, flags);
2470 /* Get mem for the objs.
2471 * Attempt to allocate a physical page from 'nodeid',
2473 if (!(objp = kmem_getpages(cachep, flags, nodeid)))
2476 /* Get slab management. */
2477 if (!(slabp = alloc_slabmgmt(cachep, objp, offset, local_flags)))
2480 slabp->nodeid = nodeid;
2481 set_slab_attr(cachep, slabp, objp);
2483 cache_init_objs(cachep, slabp, ctor_flags);
2485 if (local_flags & __GFP_WAIT)
2486 local_irq_disable();
2488 spin_lock(&l3->list_lock);
2490 /* Make slab active. */
2491 list_add_tail(&slabp->list, &(l3->slabs_free));
2492 STATS_INC_GROWN(cachep);
2493 l3->free_objects += cachep->num;
2494 spin_unlock(&l3->list_lock);
2497 kmem_freepages(cachep, objp);
2499 if (local_flags & __GFP_WAIT)
2500 local_irq_disable();
2507 * Perform extra freeing checks:
2508 * - detect bad pointers.
2509 * - POISON/RED_ZONE checking
2510 * - destructor calls, for caches with POISON+dtor
2512 static void kfree_debugcheck(const void *objp)
2516 if (!virt_addr_valid(objp)) {
2517 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2518 (unsigned long)objp);
2521 page = virt_to_page(objp);
2522 if (!PageSlab(page)) {
2523 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2524 (unsigned long)objp);
2529 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2536 objp -= obj_offset(cachep);
2537 kfree_debugcheck(objp);
2538 page = virt_to_page(objp);
2540 if (page_get_cache(page) != cachep) {
2542 "mismatch in kmem_cache_free: expected cache %p, got %p\n",
2543 page_get_cache(page), cachep);
2544 printk(KERN_ERR "%p is %s.\n", cachep, cachep->name);
2545 printk(KERN_ERR "%p is %s.\n", page_get_cache(page),
2546 page_get_cache(page)->name);
2549 slabp = page_get_slab(page);
2551 if (cachep->flags & SLAB_RED_ZONE) {
2552 if (*dbg_redzone1(cachep, objp) != RED_ACTIVE
2553 || *dbg_redzone2(cachep, objp) != RED_ACTIVE) {
2555 "double free, or memory outside"
2556 " object was overwritten");
2558 "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2559 objp, *dbg_redzone1(cachep, objp),
2560 *dbg_redzone2(cachep, objp));
2562 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2563 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2565 if (cachep->flags & SLAB_STORE_USER)
2566 *dbg_userword(cachep, objp) = caller;
2568 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
2570 BUG_ON(objnr >= cachep->num);
2571 BUG_ON(objp != slabp->s_mem + objnr * cachep->buffer_size);
2573 if (cachep->flags & SLAB_DEBUG_INITIAL) {
2574 /* Need to call the slab's constructor so the
2575 * caller can perform a verify of its state (debugging).
2576 * Called without the cache-lock held.
2578 cachep->ctor(objp + obj_offset(cachep),
2579 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
2581 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2582 /* we want to cache poison the object,
2583 * call the destruction callback
2585 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
2587 if (cachep->flags & SLAB_POISON) {
2588 #ifdef CONFIG_DEBUG_PAGEALLOC
2589 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) {
2590 store_stackinfo(cachep, objp, (unsigned long)caller);
2591 kernel_map_pages(virt_to_page(objp),
2592 cachep->buffer_size / PAGE_SIZE, 0);
2594 poison_obj(cachep, objp, POISON_FREE);
2597 poison_obj(cachep, objp, POISON_FREE);
2603 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2608 /* Check slab's freelist to see if this obj is there. */
2609 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2611 if (entries > cachep->num || i >= cachep->num)
2614 if (entries != cachep->num - slabp->inuse) {
2617 "slab: Internal list corruption detected in cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2618 cachep->name, cachep->num, slabp, slabp->inuse);
2620 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2623 printk("\n%03x:", i);
2624 printk(" %02x", ((unsigned char *)slabp)[i]);
2631 #define kfree_debugcheck(x) do { } while(0)
2632 #define cache_free_debugcheck(x,objp,z) (objp)
2633 #define check_slabp(x,y) do { } while(0)
2636 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2639 struct kmem_list3 *l3;
2640 struct array_cache *ac;
2643 ac = cpu_cache_get(cachep);
2645 batchcount = ac->batchcount;
2646 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2647 /* if there was little recent activity on this
2648 * cache, then perform only a partial refill.
2649 * Otherwise we could generate refill bouncing.
2651 batchcount = BATCHREFILL_LIMIT;
2653 l3 = cachep->nodelists[numa_node_id()];
2655 BUG_ON(ac->avail > 0 || !l3);
2656 spin_lock(&l3->list_lock);
2659 struct array_cache *shared_array = l3->shared;
2660 if (shared_array->avail) {
2661 if (batchcount > shared_array->avail)
2662 batchcount = shared_array->avail;
2663 shared_array->avail -= batchcount;
2664 ac->avail = batchcount;
2666 &(shared_array->entry[shared_array->avail]),
2667 sizeof(void *) * batchcount);
2668 shared_array->touched = 1;
2672 while (batchcount > 0) {
2673 struct list_head *entry;
2675 /* Get slab alloc is to come from. */
2676 entry = l3->slabs_partial.next;
2677 if (entry == &l3->slabs_partial) {
2678 l3->free_touched = 1;
2679 entry = l3->slabs_free.next;
2680 if (entry == &l3->slabs_free)
2684 slabp = list_entry(entry, struct slab, list);
2685 check_slabp(cachep, slabp);
2686 check_spinlock_acquired(cachep);
2687 while (slabp->inuse < cachep->num && batchcount--) {
2688 STATS_INC_ALLOCED(cachep);
2689 STATS_INC_ACTIVE(cachep);
2690 STATS_SET_HIGH(cachep);
2692 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2695 check_slabp(cachep, slabp);
2697 /* move slabp to correct slabp list: */
2698 list_del(&slabp->list);
2699 if (slabp->free == BUFCTL_END)
2700 list_add(&slabp->list, &l3->slabs_full);
2702 list_add(&slabp->list, &l3->slabs_partial);
2706 l3->free_objects -= ac->avail;
2708 spin_unlock(&l3->list_lock);
2710 if (unlikely(!ac->avail)) {
2712 x = cache_grow(cachep, flags, numa_node_id());
2714 // cache_grow can reenable interrupts, then ac could change.
2715 ac = cpu_cache_get(cachep);
2716 if (!x && ac->avail == 0) // no objects in sight? abort
2719 if (!ac->avail) // objects refilled by interrupt?
2723 return ac->entry[--ac->avail];
2727 cache_alloc_debugcheck_before(struct kmem_cache *cachep, gfp_t flags)
2729 might_sleep_if(flags & __GFP_WAIT);
2731 kmem_flagcheck(cachep, flags);
2736 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, gfp_t flags,
2737 void *objp, void *caller)
2741 if (cachep->flags & SLAB_POISON) {
2742 #ifdef CONFIG_DEBUG_PAGEALLOC
2743 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
2744 kernel_map_pages(virt_to_page(objp),
2745 cachep->buffer_size / PAGE_SIZE, 1);
2747 check_poison_obj(cachep, objp);
2749 check_poison_obj(cachep, objp);
2751 poison_obj(cachep, objp, POISON_INUSE);
2753 if (cachep->flags & SLAB_STORE_USER)
2754 *dbg_userword(cachep, objp) = caller;
2756 if (cachep->flags & SLAB_RED_ZONE) {
2757 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE
2758 || *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2760 "double free, or memory outside"
2761 " object was overwritten");
2763 "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2764 objp, *dbg_redzone1(cachep, objp),
2765 *dbg_redzone2(cachep, objp));
2767 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2768 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2770 objp += obj_offset(cachep);
2771 if (cachep->ctor && cachep->flags & SLAB_POISON) {
2772 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2774 if (!(flags & __GFP_WAIT))
2775 ctor_flags |= SLAB_CTOR_ATOMIC;
2777 cachep->ctor(objp, cachep, ctor_flags);
2782 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2785 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
2788 struct array_cache *ac;
2791 if (unlikely(current->mempolicy && !in_interrupt())) {
2792 int nid = slab_node(current->mempolicy);
2794 if (nid != numa_node_id())
2795 return __cache_alloc_node(cachep, flags, nid);
2800 ac = cpu_cache_get(cachep);
2801 if (likely(ac->avail)) {
2802 STATS_INC_ALLOCHIT(cachep);
2804 objp = ac->entry[--ac->avail];
2806 STATS_INC_ALLOCMISS(cachep);
2807 objp = cache_alloc_refill(cachep, flags);
2812 static __always_inline void *
2813 __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
2815 unsigned long save_flags;
2818 cache_alloc_debugcheck_before(cachep, flags);
2820 local_irq_save(save_flags);
2821 objp = ____cache_alloc(cachep, flags);
2822 local_irq_restore(save_flags);
2823 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
2831 * A interface to enable slab creation on nodeid
2833 static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
2835 struct list_head *entry;
2837 struct kmem_list3 *l3;
2841 l3 = cachep->nodelists[nodeid];
2846 spin_lock(&l3->list_lock);
2847 entry = l3->slabs_partial.next;
2848 if (entry == &l3->slabs_partial) {
2849 l3->free_touched = 1;
2850 entry = l3->slabs_free.next;
2851 if (entry == &l3->slabs_free)
2855 slabp = list_entry(entry, struct slab, list);
2856 check_spinlock_acquired_node(cachep, nodeid);
2857 check_slabp(cachep, slabp);
2859 STATS_INC_NODEALLOCS(cachep);
2860 STATS_INC_ACTIVE(cachep);
2861 STATS_SET_HIGH(cachep);
2863 BUG_ON(slabp->inuse == cachep->num);
2865 obj = slab_get_obj(cachep, slabp, nodeid);
2866 check_slabp(cachep, slabp);
2868 /* move slabp to correct slabp list: */
2869 list_del(&slabp->list);
2871 if (slabp->free == BUFCTL_END) {
2872 list_add(&slabp->list, &l3->slabs_full);
2874 list_add(&slabp->list, &l3->slabs_partial);
2877 spin_unlock(&l3->list_lock);
2881 spin_unlock(&l3->list_lock);
2882 x = cache_grow(cachep, flags, nodeid);
2894 * Caller needs to acquire correct kmem_list's list_lock
2896 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
2900 struct kmem_list3 *l3;
2902 for (i = 0; i < nr_objects; i++) {
2903 void *objp = objpp[i];
2906 slabp = virt_to_slab(objp);
2907 l3 = cachep->nodelists[node];
2908 list_del(&slabp->list);
2909 check_spinlock_acquired_node(cachep, node);
2910 check_slabp(cachep, slabp);
2911 slab_put_obj(cachep, slabp, objp, node);
2912 STATS_DEC_ACTIVE(cachep);
2914 check_slabp(cachep, slabp);
2916 /* fixup slab chains */
2917 if (slabp->inuse == 0) {
2918 if (l3->free_objects > l3->free_limit) {
2919 l3->free_objects -= cachep->num;
2920 slab_destroy(cachep, slabp);
2922 list_add(&slabp->list, &l3->slabs_free);
2925 /* Unconditionally move a slab to the end of the
2926 * partial list on free - maximum time for the
2927 * other objects to be freed, too.
2929 list_add_tail(&slabp->list, &l3->slabs_partial);
2934 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
2937 struct kmem_list3 *l3;
2938 int node = numa_node_id();
2940 batchcount = ac->batchcount;
2942 BUG_ON(!batchcount || batchcount > ac->avail);
2945 l3 = cachep->nodelists[node];
2946 spin_lock(&l3->list_lock);
2948 struct array_cache *shared_array = l3->shared;
2949 int max = shared_array->limit - shared_array->avail;
2951 if (batchcount > max)
2953 memcpy(&(shared_array->entry[shared_array->avail]),
2954 ac->entry, sizeof(void *) * batchcount);
2955 shared_array->avail += batchcount;
2960 free_block(cachep, ac->entry, batchcount, node);
2965 struct list_head *p;
2967 p = l3->slabs_free.next;
2968 while (p != &(l3->slabs_free)) {
2971 slabp = list_entry(p, struct slab, list);
2972 BUG_ON(slabp->inuse);
2977 STATS_SET_FREEABLE(cachep, i);
2980 spin_unlock(&l3->list_lock);
2981 ac->avail -= batchcount;
2982 memmove(ac->entry, &(ac->entry[batchcount]),
2983 sizeof(void *) * ac->avail);
2988 * Release an obj back to its cache. If the obj has a constructed
2989 * state, it must be in this state _before_ it is released.
2991 * Called with disabled ints.
2993 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
2995 struct array_cache *ac = cpu_cache_get(cachep);
2998 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3000 /* Make sure we are not freeing a object from another
3001 * node to the array cache on this cpu.
3006 slabp = virt_to_slab(objp);
3007 if (unlikely(slabp->nodeid != numa_node_id())) {
3008 struct array_cache *alien = NULL;
3009 int nodeid = slabp->nodeid;
3010 struct kmem_list3 *l3 =
3011 cachep->nodelists[numa_node_id()];
3013 STATS_INC_NODEFREES(cachep);
3014 if (l3->alien && l3->alien[nodeid]) {
3015 alien = l3->alien[nodeid];
3016 spin_lock(&alien->lock);
3017 if (unlikely(alien->avail == alien->limit))
3018 __drain_alien_cache(cachep,
3020 alien->entry[alien->avail++] = objp;
3021 spin_unlock(&alien->lock);
3023 spin_lock(&(cachep->nodelists[nodeid])->
3025 free_block(cachep, &objp, 1, nodeid);
3026 spin_unlock(&(cachep->nodelists[nodeid])->
3033 if (likely(ac->avail < ac->limit)) {
3034 STATS_INC_FREEHIT(cachep);
3035 ac->entry[ac->avail++] = objp;
3038 STATS_INC_FREEMISS(cachep);
3039 cache_flusharray(cachep, ac);
3040 ac->entry[ac->avail++] = objp;
3045 * kmem_cache_alloc - Allocate an object
3046 * @cachep: The cache to allocate from.
3047 * @flags: See kmalloc().
3049 * Allocate an object from this cache. The flags are only relevant
3050 * if the cache has no available objects.
3052 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3054 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3056 EXPORT_SYMBOL(kmem_cache_alloc);
3059 * kmem_ptr_validate - check if an untrusted pointer might
3061 * @cachep: the cache we're checking against
3062 * @ptr: pointer to validate
3064 * This verifies that the untrusted pointer looks sane:
3065 * it is _not_ a guarantee that the pointer is actually
3066 * part of the slab cache in question, but it at least
3067 * validates that the pointer can be dereferenced and
3068 * looks half-way sane.
3070 * Currently only used for dentry validation.
3072 int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr)
3074 unsigned long addr = (unsigned long)ptr;
3075 unsigned long min_addr = PAGE_OFFSET;
3076 unsigned long align_mask = BYTES_PER_WORD - 1;
3077 unsigned long size = cachep->buffer_size;
3080 if (unlikely(addr < min_addr))
3082 if (unlikely(addr > (unsigned long)high_memory - size))
3084 if (unlikely(addr & align_mask))
3086 if (unlikely(!kern_addr_valid(addr)))
3088 if (unlikely(!kern_addr_valid(addr + size - 1)))
3090 page = virt_to_page(ptr);
3091 if (unlikely(!PageSlab(page)))
3093 if (unlikely(page_get_cache(page) != cachep))
3102 * kmem_cache_alloc_node - Allocate an object on the specified node
3103 * @cachep: The cache to allocate from.
3104 * @flags: See kmalloc().
3105 * @nodeid: node number of the target node.
3107 * Identical to kmem_cache_alloc, except that this function is slow
3108 * and can sleep. And it will allocate memory on the given node, which
3109 * can improve the performance for cpu bound structures.
3110 * New and improved: it will now make sure that the object gets
3111 * put on the correct node list so that there is no false sharing.
3113 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3115 unsigned long save_flags;
3118 cache_alloc_debugcheck_before(cachep, flags);
3119 local_irq_save(save_flags);
3121 if (nodeid == -1 || nodeid == numa_node_id() ||
3122 !cachep->nodelists[nodeid])
3123 ptr = ____cache_alloc(cachep, flags);
3125 ptr = __cache_alloc_node(cachep, flags, nodeid);
3126 local_irq_restore(save_flags);
3128 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr,
3129 __builtin_return_address(0));
3133 EXPORT_SYMBOL(kmem_cache_alloc_node);
3135 void *kmalloc_node(size_t size, gfp_t flags, int node)
3137 struct kmem_cache *cachep;
3139 cachep = kmem_find_general_cachep(size, flags);
3140 if (unlikely(cachep == NULL))
3142 return kmem_cache_alloc_node(cachep, flags, node);
3144 EXPORT_SYMBOL(kmalloc_node);
3148 * kmalloc - allocate memory
3149 * @size: how many bytes of memory are required.
3150 * @flags: the type of memory to allocate.
3152 * kmalloc is the normal method of allocating memory
3155 * The @flags argument may be one of:
3157 * %GFP_USER - Allocate memory on behalf of user. May sleep.
3159 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
3161 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
3163 * Additionally, the %GFP_DMA flag may be set to indicate the memory
3164 * must be suitable for DMA. This can mean different things on different
3165 * platforms. For example, on i386, it means that the memory must come
3166 * from the first 16MB.
3168 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3171 struct kmem_cache *cachep;
3173 /* If you want to save a few bytes .text space: replace
3175 * Then kmalloc uses the uninlined functions instead of the inline
3178 cachep = __find_general_cachep(size, flags);
3179 if (unlikely(cachep == NULL))
3181 return __cache_alloc(cachep, flags, caller);
3184 #ifndef CONFIG_DEBUG_SLAB
3186 void *__kmalloc(size_t size, gfp_t flags)
3188 return __do_kmalloc(size, flags, NULL);
3190 EXPORT_SYMBOL(__kmalloc);
3194 void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3196 return __do_kmalloc(size, flags, caller);
3198 EXPORT_SYMBOL(__kmalloc_track_caller);
3204 * __alloc_percpu - allocate one copy of the object for every present
3205 * cpu in the system, zeroing them.
3206 * Objects should be dereferenced using the per_cpu_ptr macro only.
3208 * @size: how many bytes of memory are required.
3210 void *__alloc_percpu(size_t size)
3213 struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL);
3219 * Cannot use for_each_online_cpu since a cpu may come online
3220 * and we have no way of figuring out how to fix the array
3221 * that we have allocated then....
3224 int node = cpu_to_node(i);
3226 if (node_online(node))
3227 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
3229 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
3231 if (!pdata->ptrs[i])
3233 memset(pdata->ptrs[i], 0, size);
3236 /* Catch derefs w/o wrappers */
3237 return (void *)(~(unsigned long)pdata);
3241 if (!cpu_possible(i))
3243 kfree(pdata->ptrs[i]);
3248 EXPORT_SYMBOL(__alloc_percpu);
3252 * kmem_cache_free - Deallocate an object
3253 * @cachep: The cache the allocation was from.
3254 * @objp: The previously allocated object.
3256 * Free an object which was previously allocated from this
3259 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3261 unsigned long flags;
3263 local_irq_save(flags);
3264 __cache_free(cachep, objp);
3265 local_irq_restore(flags);
3267 EXPORT_SYMBOL(kmem_cache_free);
3270 * kfree - free previously allocated memory
3271 * @objp: pointer returned by kmalloc.
3273 * If @objp is NULL, no operation is performed.
3275 * Don't free memory not originally allocated by kmalloc()
3276 * or you will run into trouble.
3278 void kfree(const void *objp)
3280 struct kmem_cache *c;
3281 unsigned long flags;
3283 if (unlikely(!objp))
3285 local_irq_save(flags);
3286 kfree_debugcheck(objp);
3287 c = virt_to_cache(objp);
3288 mutex_debug_check_no_locks_freed(objp, obj_size(c));
3289 __cache_free(c, (void *)objp);
3290 local_irq_restore(flags);
3292 EXPORT_SYMBOL(kfree);
3296 * free_percpu - free previously allocated percpu memory
3297 * @objp: pointer returned by alloc_percpu.
3299 * Don't free memory not originally allocated by alloc_percpu()
3300 * The complemented objp is to check for that.
3302 void free_percpu(const void *objp)
3305 struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp);
3308 * We allocate for all cpus so we cannot use for online cpu here.
3314 EXPORT_SYMBOL(free_percpu);
3317 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3319 return obj_size(cachep);
3321 EXPORT_SYMBOL(kmem_cache_size);
3323 const char *kmem_cache_name(struct kmem_cache *cachep)
3325 return cachep->name;
3327 EXPORT_SYMBOL_GPL(kmem_cache_name);
3330 * This initializes kmem_list3 for all nodes.
3332 static int alloc_kmemlist(struct kmem_cache *cachep)
3335 struct kmem_list3 *l3;
3338 for_each_online_node(node) {
3339 struct array_cache *nc = NULL, *new;
3340 struct array_cache **new_alien = NULL;
3342 if (!(new_alien = alloc_alien_cache(node, cachep->limit)))
3345 if (!(new = alloc_arraycache(node, (cachep->shared *
3346 cachep->batchcount),
3349 if ((l3 = cachep->nodelists[node])) {
3351 spin_lock_irq(&l3->list_lock);
3353 if ((nc = cachep->nodelists[node]->shared))
3354 free_block(cachep, nc->entry, nc->avail, node);
3357 if (!cachep->nodelists[node]->alien) {
3358 l3->alien = new_alien;
3361 l3->free_limit = (1 + nr_cpus_node(node)) *
3362 cachep->batchcount + cachep->num;
3363 spin_unlock_irq(&l3->list_lock);
3365 free_alien_cache(new_alien);
3368 if (!(l3 = kmalloc_node(sizeof(struct kmem_list3),
3372 kmem_list3_init(l3);
3373 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3374 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3376 l3->alien = new_alien;
3377 l3->free_limit = (1 + nr_cpus_node(node)) *
3378 cachep->batchcount + cachep->num;
3379 cachep->nodelists[node] = l3;
3387 struct ccupdate_struct {
3388 struct kmem_cache *cachep;
3389 struct array_cache *new[NR_CPUS];
3392 static void do_ccupdate_local(void *info)
3394 struct ccupdate_struct *new = (struct ccupdate_struct *)info;
3395 struct array_cache *old;
3398 old = cpu_cache_get(new->cachep);
3400 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3401 new->new[smp_processor_id()] = old;
3404 static int do_tune_cpucache(struct kmem_cache *cachep, int limit, int batchcount,
3407 struct ccupdate_struct new;
3410 memset(&new.new, 0, sizeof(new.new));
3411 for_each_online_cpu(i) {
3413 alloc_arraycache(cpu_to_node(i), limit, batchcount);
3415 for (i--; i >= 0; i--)
3420 new.cachep = cachep;
3422 smp_call_function_all_cpus(do_ccupdate_local, (void *)&new);
3425 spin_lock(&cachep->spinlock);
3426 cachep->batchcount = batchcount;
3427 cachep->limit = limit;
3428 cachep->shared = shared;
3429 spin_unlock(&cachep->spinlock);
3431 for_each_online_cpu(i) {
3432 struct array_cache *ccold = new.new[i];
3435 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3436 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3437 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3441 err = alloc_kmemlist(cachep);
3443 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
3444 cachep->name, -err);
3450 static void enable_cpucache(struct kmem_cache *cachep)
3455 /* The head array serves three purposes:
3456 * - create a LIFO ordering, i.e. return objects that are cache-warm
3457 * - reduce the number of spinlock operations.
3458 * - reduce the number of linked list operations on the slab and
3459 * bufctl chains: array operations are cheaper.
3460 * The numbers are guessed, we should auto-tune as described by
3463 if (cachep->buffer_size > 131072)
3465 else if (cachep->buffer_size > PAGE_SIZE)
3467 else if (cachep->buffer_size > 1024)
3469 else if (cachep->buffer_size > 256)
3474 /* Cpu bound tasks (e.g. network routing) can exhibit cpu bound
3475 * allocation behaviour: Most allocs on one cpu, most free operations
3476 * on another cpu. For these cases, an efficient object passing between
3477 * cpus is necessary. This is provided by a shared array. The array
3478 * replaces Bonwick's magazine layer.
3479 * On uniprocessor, it's functionally equivalent (but less efficient)
3480 * to a larger limit. Thus disabled by default.
3484 if (cachep->buffer_size <= PAGE_SIZE)
3489 /* With debugging enabled, large batchcount lead to excessively
3490 * long periods with disabled local interrupts. Limit the
3496 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
3498 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3499 cachep->name, -err);
3502 static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac,
3503 int force, int node)
3507 check_spinlock_acquired_node(cachep, node);
3508 if (ac->touched && !force) {
3510 } else if (ac->avail) {
3511 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3512 if (tofree > ac->avail) {
3513 tofree = (ac->avail + 1) / 2;
3515 free_block(cachep, ac->entry, tofree, node);
3516 ac->avail -= tofree;
3517 memmove(ac->entry, &(ac->entry[tofree]),
3518 sizeof(void *) * ac->avail);
3523 * cache_reap - Reclaim memory from caches.
3524 * @unused: unused parameter
3526 * Called from workqueue/eventd every few seconds.
3528 * - clear the per-cpu caches for this CPU.
3529 * - return freeable pages to the main free memory pool.
3531 * If we cannot acquire the cache chain mutex then just give up - we'll
3532 * try again on the next iteration.
3534 static void cache_reap(void *unused)
3536 struct list_head *walk;
3537 struct kmem_list3 *l3;
3539 if (!mutex_trylock(&cache_chain_mutex)) {
3540 /* Give up. Setup the next iteration. */
3541 schedule_delayed_work(&__get_cpu_var(reap_work),
3546 list_for_each(walk, &cache_chain) {
3547 struct kmem_cache *searchp;
3548 struct list_head *p;
3552 searchp = list_entry(walk, struct kmem_cache, next);
3554 if (searchp->flags & SLAB_NO_REAP)
3559 l3 = searchp->nodelists[numa_node_id()];
3560 reap_alien(searchp, l3);
3561 spin_lock_irq(&l3->list_lock);
3563 drain_array_locked(searchp, cpu_cache_get(searchp), 0,
3566 if (time_after(l3->next_reap, jiffies))
3569 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
3572 drain_array_locked(searchp, l3->shared, 0,
3575 if (l3->free_touched) {
3576 l3->free_touched = 0;
3581 (l3->free_limit + 5 * searchp->num -
3582 1) / (5 * searchp->num);
3584 p = l3->slabs_free.next;
3585 if (p == &(l3->slabs_free))
3588 slabp = list_entry(p, struct slab, list);
3589 BUG_ON(slabp->inuse);
3590 list_del(&slabp->list);
3591 STATS_INC_REAPED(searchp);
3593 /* Safe to drop the lock. The slab is no longer
3594 * linked to the cache.
3595 * searchp cannot disappear, we hold
3598 l3->free_objects -= searchp->num;
3599 spin_unlock_irq(&l3->list_lock);
3600 slab_destroy(searchp, slabp);
3601 spin_lock_irq(&l3->list_lock);
3602 } while (--tofree > 0);
3604 spin_unlock_irq(&l3->list_lock);
3609 mutex_unlock(&cache_chain_mutex);
3611 /* Setup the next iteration */
3612 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
3615 #ifdef CONFIG_PROC_FS
3617 static void print_slabinfo_header(struct seq_file *m)
3620 * Output format version, so at least we can change it
3621 * without _too_ many complaints.
3624 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3626 seq_puts(m, "slabinfo - version: 2.1\n");
3628 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3629 "<objperslab> <pagesperslab>");
3630 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3631 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3633 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3634 "<error> <maxfreeable> <nodeallocs> <remotefrees>");
3635 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3640 static void *s_start(struct seq_file *m, loff_t *pos)
3643 struct list_head *p;
3645 mutex_lock(&cache_chain_mutex);
3647 print_slabinfo_header(m);
3648 p = cache_chain.next;
3651 if (p == &cache_chain)
3654 return list_entry(p, struct kmem_cache, next);
3657 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3659 struct kmem_cache *cachep = p;
3661 return cachep->next.next == &cache_chain ? NULL
3662 : list_entry(cachep->next.next, struct kmem_cache, next);
3665 static void s_stop(struct seq_file *m, void *p)
3667 mutex_unlock(&cache_chain_mutex);
3670 static int s_show(struct seq_file *m, void *p)
3672 struct kmem_cache *cachep = p;
3673 struct list_head *q;
3675 unsigned long active_objs;
3676 unsigned long num_objs;
3677 unsigned long active_slabs = 0;
3678 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3682 struct kmem_list3 *l3;
3684 spin_lock(&cachep->spinlock);
3687 for_each_online_node(node) {
3688 l3 = cachep->nodelists[node];
3693 spin_lock_irq(&l3->list_lock);
3695 list_for_each(q, &l3->slabs_full) {
3696 slabp = list_entry(q, struct slab, list);
3697 if (slabp->inuse != cachep->num && !error)
3698 error = "slabs_full accounting error";
3699 active_objs += cachep->num;
3702 list_for_each(q, &l3->slabs_partial) {
3703 slabp = list_entry(q, struct slab, list);
3704 if (slabp->inuse == cachep->num && !error)
3705 error = "slabs_partial inuse accounting error";
3706 if (!slabp->inuse && !error)
3707 error = "slabs_partial/inuse accounting error";
3708 active_objs += slabp->inuse;
3711 list_for_each(q, &l3->slabs_free) {
3712 slabp = list_entry(q, struct slab, list);
3713 if (slabp->inuse && !error)
3714 error = "slabs_free/inuse accounting error";
3717 free_objects += l3->free_objects;
3719 shared_avail += l3->shared->avail;
3721 spin_unlock_irq(&l3->list_lock);
3723 num_slabs += active_slabs;
3724 num_objs = num_slabs * cachep->num;
3725 if (num_objs - active_objs != free_objects && !error)
3726 error = "free_objects accounting error";
3728 name = cachep->name;
3730 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3732 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
3733 name, active_objs, num_objs, cachep->buffer_size,
3734 cachep->num, (1 << cachep->gfporder));
3735 seq_printf(m, " : tunables %4u %4u %4u",
3736 cachep->limit, cachep->batchcount, cachep->shared);
3737 seq_printf(m, " : slabdata %6lu %6lu %6lu",
3738 active_slabs, num_slabs, shared_avail);
3741 unsigned long high = cachep->high_mark;
3742 unsigned long allocs = cachep->num_allocations;
3743 unsigned long grown = cachep->grown;
3744 unsigned long reaped = cachep->reaped;
3745 unsigned long errors = cachep->errors;
3746 unsigned long max_freeable = cachep->max_freeable;
3747 unsigned long node_allocs = cachep->node_allocs;
3748 unsigned long node_frees = cachep->node_frees;
3750 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
3751 %4lu %4lu %4lu %4lu", allocs, high, grown, reaped, errors, max_freeable, node_allocs, node_frees);
3755 unsigned long allochit = atomic_read(&cachep->allochit);
3756 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3757 unsigned long freehit = atomic_read(&cachep->freehit);
3758 unsigned long freemiss = atomic_read(&cachep->freemiss);
3760 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
3761 allochit, allocmiss, freehit, freemiss);
3765 spin_unlock(&cachep->spinlock);
3770 * slabinfo_op - iterator that generates /proc/slabinfo
3779 * num-pages-per-slab
3780 * + further values on SMP and with statistics enabled
3783 struct seq_operations slabinfo_op = {
3790 #define MAX_SLABINFO_WRITE 128
3792 * slabinfo_write - Tuning for the slab allocator
3794 * @buffer: user buffer
3795 * @count: data length
3798 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
3799 size_t count, loff_t *ppos)
3801 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
3802 int limit, batchcount, shared, res;
3803 struct list_head *p;
3805 if (count > MAX_SLABINFO_WRITE)
3807 if (copy_from_user(&kbuf, buffer, count))
3809 kbuf[MAX_SLABINFO_WRITE] = '\0';
3811 tmp = strchr(kbuf, ' ');
3816 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3819 /* Find the cache in the chain of caches. */
3820 mutex_lock(&cache_chain_mutex);
3822 list_for_each(p, &cache_chain) {
3823 struct kmem_cache *cachep = list_entry(p, struct kmem_cache,
3826 if (!strcmp(cachep->name, kbuf)) {
3829 batchcount > limit || shared < 0) {
3832 res = do_tune_cpucache(cachep, limit,
3833 batchcount, shared);
3838 mutex_unlock(&cache_chain_mutex);
3846 * ksize - get the actual amount of memory allocated for a given object
3847 * @objp: Pointer to the object
3849 * kmalloc may internally round up allocations and return more memory
3850 * than requested. ksize() can be used to determine the actual amount of
3851 * memory allocated. The caller may use this additional memory, even though
3852 * a smaller amount of memory was initially specified with the kmalloc call.
3853 * The caller must guarantee that objp points to a valid object previously
3854 * allocated with either kmalloc() or kmem_cache_alloc(). The object
3855 * must not be freed during the duration of the call.
3857 unsigned int ksize(const void *objp)
3859 if (unlikely(objp == NULL))
3862 return obj_size(virt_to_cache(objp));