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 kmem_cache_t 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 kmem_cache_t *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 spinlock_t list_lock;
298 struct array_cache *shared; /* shared per node */
299 struct array_cache **alien; /* on other nodes */
303 * Need this for bootstrapping a per node allocator.
305 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
306 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
307 #define CACHE_CACHE 0
309 #define SIZE_L3 (1 + MAX_NUMNODES)
312 * This function must be completely optimized away if
313 * a constant is passed to it. Mostly the same as
314 * what is in linux/slab.h except it returns an
317 static __always_inline int index_of(const size_t size)
319 extern void __bad_size(void);
321 if (__builtin_constant_p(size)) {
329 #include "linux/kmalloc_sizes.h"
337 #define INDEX_AC index_of(sizeof(struct arraycache_init))
338 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
340 static void kmem_list3_init(struct kmem_list3 *parent)
342 INIT_LIST_HEAD(&parent->slabs_full);
343 INIT_LIST_HEAD(&parent->slabs_partial);
344 INIT_LIST_HEAD(&parent->slabs_free);
345 parent->shared = NULL;
346 parent->alien = NULL;
347 spin_lock_init(&parent->list_lock);
348 parent->free_objects = 0;
349 parent->free_touched = 0;
352 #define MAKE_LIST(cachep, listp, slab, nodeid) \
354 INIT_LIST_HEAD(listp); \
355 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
358 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
360 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
361 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
362 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
372 /* 1) per-cpu data, touched during every alloc/free */
373 struct array_cache *array[NR_CPUS];
374 unsigned int batchcount;
377 unsigned int buffer_size;
378 /* 2) touched by every alloc & free from the backend */
379 struct kmem_list3 *nodelists[MAX_NUMNODES];
380 unsigned int flags; /* constant flags */
381 unsigned int num; /* # of objs per slab */
384 /* 3) cache_grow/shrink */
385 /* order of pgs per slab (2^n) */
386 unsigned int gfporder;
388 /* force GFP flags, e.g. GFP_DMA */
391 size_t colour; /* cache colouring range */
392 unsigned int colour_off; /* colour offset */
393 unsigned int colour_next; /* cache colouring */
394 kmem_cache_t *slabp_cache;
395 unsigned int slab_size;
396 unsigned int dflags; /* dynamic flags */
398 /* constructor func */
399 void (*ctor) (void *, kmem_cache_t *, unsigned long);
401 /* de-constructor func */
402 void (*dtor) (void *, kmem_cache_t *, unsigned long);
404 /* 4) cache creation/removal */
406 struct list_head next;
410 unsigned long num_active;
411 unsigned long num_allocations;
412 unsigned long high_mark;
414 unsigned long reaped;
415 unsigned long errors;
416 unsigned long max_freeable;
417 unsigned long node_allocs;
418 unsigned long node_frees;
426 * If debugging is enabled, then the allocator can add additional
427 * fields and/or padding to every object. buffer_size contains the total
428 * object size including these internal fields, the following two
429 * variables contain the offset to the user object and its size.
436 #define CFLGS_OFF_SLAB (0x80000000UL)
437 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
439 #define BATCHREFILL_LIMIT 16
440 /* Optimization question: fewer reaps means less
441 * probability for unnessary cpucache drain/refill cycles.
443 * OTOH the cpuarrays can contain lots of objects,
444 * which could lock up otherwise freeable slabs.
446 #define REAPTIMEOUT_CPUC (2*HZ)
447 #define REAPTIMEOUT_LIST3 (4*HZ)
450 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
451 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
452 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
453 #define STATS_INC_GROWN(x) ((x)->grown++)
454 #define STATS_INC_REAPED(x) ((x)->reaped++)
455 #define STATS_SET_HIGH(x) do { if ((x)->num_active > (x)->high_mark) \
456 (x)->high_mark = (x)->num_active; \
458 #define STATS_INC_ERR(x) ((x)->errors++)
459 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
460 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
461 #define STATS_SET_FREEABLE(x, i) \
462 do { if ((x)->max_freeable < i) \
463 (x)->max_freeable = i; \
466 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
467 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
468 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
469 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
471 #define STATS_INC_ACTIVE(x) do { } while (0)
472 #define STATS_DEC_ACTIVE(x) do { } while (0)
473 #define STATS_INC_ALLOCED(x) do { } while (0)
474 #define STATS_INC_GROWN(x) do { } while (0)
475 #define STATS_INC_REAPED(x) do { } while (0)
476 #define STATS_SET_HIGH(x) do { } while (0)
477 #define STATS_INC_ERR(x) do { } while (0)
478 #define STATS_INC_NODEALLOCS(x) do { } while (0)
479 #define STATS_INC_NODEFREES(x) do { } while (0)
480 #define STATS_SET_FREEABLE(x, i) \
483 #define STATS_INC_ALLOCHIT(x) do { } while (0)
484 #define STATS_INC_ALLOCMISS(x) do { } while (0)
485 #define STATS_INC_FREEHIT(x) do { } while (0)
486 #define STATS_INC_FREEMISS(x) do { } while (0)
490 /* Magic nums for obj red zoning.
491 * Placed in the first word before and the first word after an obj.
493 #define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
494 #define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
496 /* ...and for poisoning */
497 #define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
498 #define POISON_FREE 0x6b /* for use-after-free poisoning */
499 #define POISON_END 0xa5 /* end-byte of poisoning */
501 /* memory layout of objects:
503 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
504 * the end of an object is aligned with the end of the real
505 * allocation. Catches writes behind the end of the allocation.
506 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
508 * cachep->obj_offset: The real object.
509 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
510 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address [BYTES_PER_WORD long]
512 static int obj_offset(kmem_cache_t *cachep)
514 return cachep->obj_offset;
517 static int obj_size(kmem_cache_t *cachep)
519 return cachep->obj_size;
522 static unsigned long *dbg_redzone1(kmem_cache_t *cachep, void *objp)
524 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
525 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD);
528 static unsigned long *dbg_redzone2(kmem_cache_t *cachep, void *objp)
530 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
531 if (cachep->flags & SLAB_STORE_USER)
532 return (unsigned long *)(objp + cachep->buffer_size -
534 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD);
537 static void **dbg_userword(kmem_cache_t *cachep, void *objp)
539 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
540 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
545 #define obj_offset(x) 0
546 #define obj_size(cachep) (cachep->buffer_size)
547 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
548 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
549 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
554 * Maximum size of an obj (in 2^order pages)
555 * and absolute limit for the gfp order.
557 #if defined(CONFIG_LARGE_ALLOCS)
558 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
559 #define MAX_GFP_ORDER 13 /* up to 32Mb */
560 #elif defined(CONFIG_MMU)
561 #define MAX_OBJ_ORDER 5 /* 32 pages */
562 #define MAX_GFP_ORDER 5 /* 32 pages */
564 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
565 #define MAX_GFP_ORDER 8 /* up to 1Mb */
569 * Do not go above this order unless 0 objects fit into the slab.
571 #define BREAK_GFP_ORDER_HI 1
572 #define BREAK_GFP_ORDER_LO 0
573 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
575 /* Functions for storing/retrieving the cachep and or slab from the
576 * global 'mem_map'. These are used to find the slab an obj belongs to.
577 * With kfree(), these are used to find the cache which an obj belongs to.
579 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
581 page->lru.next = (struct list_head *)cache;
584 static inline struct kmem_cache *page_get_cache(struct page *page)
586 return (struct kmem_cache *)page->lru.next;
589 static inline void page_set_slab(struct page *page, struct slab *slab)
591 page->lru.prev = (struct list_head *)slab;
594 static inline struct slab *page_get_slab(struct page *page)
596 return (struct slab *)page->lru.prev;
599 static inline struct kmem_cache *virt_to_cache(const void *obj)
601 struct page *page = virt_to_page(obj);
602 return page_get_cache(page);
605 static inline struct slab *virt_to_slab(const void *obj)
607 struct page *page = virt_to_page(obj);
608 return page_get_slab(page);
611 /* These are the default caches for kmalloc. Custom caches can have other sizes. */
612 struct cache_sizes malloc_sizes[] = {
613 #define CACHE(x) { .cs_size = (x) },
614 #include <linux/kmalloc_sizes.h>
618 EXPORT_SYMBOL(malloc_sizes);
620 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
626 static struct cache_names __initdata cache_names[] = {
627 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
628 #include <linux/kmalloc_sizes.h>
633 static struct arraycache_init initarray_cache __initdata =
634 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
635 static struct arraycache_init initarray_generic =
636 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
638 /* internal cache of cache description objs */
639 static kmem_cache_t cache_cache = {
641 .limit = BOOT_CPUCACHE_ENTRIES,
643 .buffer_size = sizeof(kmem_cache_t),
644 .flags = SLAB_NO_REAP,
645 .spinlock = SPIN_LOCK_UNLOCKED,
646 .name = "kmem_cache",
648 .obj_size = sizeof(kmem_cache_t),
652 /* Guard access to the cache-chain. */
653 static DEFINE_MUTEX(cache_chain_mutex);
654 static struct list_head cache_chain;
657 * vm_enough_memory() looks at this to determine how many
658 * slab-allocated pages are possibly freeable under pressure
660 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
662 atomic_t slab_reclaim_pages;
665 * chicken and egg problem: delay the per-cpu array allocation
666 * until the general caches are up.
675 static DEFINE_PER_CPU(struct work_struct, reap_work);
677 static void free_block(kmem_cache_t *cachep, void **objpp, int len, int node);
678 static void enable_cpucache(kmem_cache_t *cachep);
679 static void cache_reap(void *unused);
680 static int __node_shrink(kmem_cache_t *cachep, int node);
682 static inline struct array_cache *cpu_cache_get(kmem_cache_t *cachep)
684 return cachep->array[smp_processor_id()];
687 static inline kmem_cache_t *__find_general_cachep(size_t size, gfp_t gfpflags)
689 struct cache_sizes *csizep = malloc_sizes;
692 /* This happens if someone tries to call
693 * kmem_cache_create(), or __kmalloc(), before
694 * the generic caches are initialized.
696 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
698 while (size > csizep->cs_size)
702 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
703 * has cs_{dma,}cachep==NULL. Thus no special case
704 * for large kmalloc calls required.
706 if (unlikely(gfpflags & GFP_DMA))
707 return csizep->cs_dmacachep;
708 return csizep->cs_cachep;
711 kmem_cache_t *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
713 return __find_general_cachep(size, gfpflags);
715 EXPORT_SYMBOL(kmem_find_general_cachep);
717 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
719 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
722 /* Calculate the number of objects and left-over bytes for a given
724 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
725 size_t align, int flags, size_t *left_over,
730 size_t slab_size = PAGE_SIZE << gfporder;
733 * The slab management structure can be either off the slab or
734 * on it. For the latter case, the memory allocated for a
738 * - One kmem_bufctl_t for each object
739 * - Padding to respect alignment of @align
740 * - @buffer_size bytes for each object
742 * If the slab management structure is off the slab, then the
743 * alignment will already be calculated into the size. Because
744 * the slabs are all pages aligned, the objects will be at the
745 * correct alignment when allocated.
747 if (flags & CFLGS_OFF_SLAB) {
749 nr_objs = slab_size / buffer_size;
751 if (nr_objs > SLAB_LIMIT)
752 nr_objs = SLAB_LIMIT;
755 * Ignore padding for the initial guess. The padding
756 * is at most @align-1 bytes, and @buffer_size is at
757 * least @align. In the worst case, this result will
758 * be one greater than the number of objects that fit
759 * into the memory allocation when taking the padding
762 nr_objs = (slab_size - sizeof(struct slab)) /
763 (buffer_size + sizeof(kmem_bufctl_t));
766 * This calculated number will be either the right
767 * amount, or one greater than what we want.
769 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
773 if (nr_objs > SLAB_LIMIT)
774 nr_objs = SLAB_LIMIT;
776 mgmt_size = slab_mgmt_size(nr_objs, align);
779 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
782 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
784 static void __slab_error(const char *function, kmem_cache_t *cachep, char *msg)
786 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
787 function, cachep->name, msg);
792 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
793 * via the workqueue/eventd.
794 * Add the CPU number into the expiration time to minimize the possibility of
795 * the CPUs getting into lockstep and contending for the global cache chain
798 static void __devinit start_cpu_timer(int cpu)
800 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
803 * When this gets called from do_initcalls via cpucache_init(),
804 * init_workqueues() has already run, so keventd will be setup
807 if (keventd_up() && reap_work->func == NULL) {
808 INIT_WORK(reap_work, cache_reap, NULL);
809 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
813 static struct array_cache *alloc_arraycache(int node, int entries,
816 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
817 struct array_cache *nc = NULL;
819 nc = kmalloc_node(memsize, GFP_KERNEL, node);
823 nc->batchcount = batchcount;
825 spin_lock_init(&nc->lock);
831 static void *__cache_alloc_node(kmem_cache_t *, gfp_t, int);
833 static struct array_cache **alloc_alien_cache(int node, int limit)
835 struct array_cache **ac_ptr;
836 int memsize = sizeof(void *) * MAX_NUMNODES;
841 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
844 if (i == node || !node_online(i)) {
848 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
850 for (i--; i <= 0; i--)
860 static void free_alien_cache(struct array_cache **ac_ptr)
873 static void __drain_alien_cache(kmem_cache_t *cachep,
874 struct array_cache *ac, int node)
876 struct kmem_list3 *rl3 = cachep->nodelists[node];
879 spin_lock(&rl3->list_lock);
880 free_block(cachep, ac->entry, ac->avail, node);
882 spin_unlock(&rl3->list_lock);
886 static void drain_alien_cache(kmem_cache_t *cachep, struct kmem_list3 *l3)
889 struct array_cache *ac;
892 for_each_online_node(i) {
895 spin_lock_irqsave(&ac->lock, flags);
896 __drain_alien_cache(cachep, ac, i);
897 spin_unlock_irqrestore(&ac->lock, flags);
902 #define alloc_alien_cache(node, limit) do { } while (0)
903 #define free_alien_cache(ac_ptr) do { } while (0)
904 #define drain_alien_cache(cachep, l3) do { } while (0)
907 static int __devinit cpuup_callback(struct notifier_block *nfb,
908 unsigned long action, void *hcpu)
910 long cpu = (long)hcpu;
911 kmem_cache_t *cachep;
912 struct kmem_list3 *l3 = NULL;
913 int node = cpu_to_node(cpu);
914 int memsize = sizeof(struct kmem_list3);
918 mutex_lock(&cache_chain_mutex);
919 /* we need to do this right in the beginning since
920 * alloc_arraycache's are going to use this list.
921 * kmalloc_node allows us to add the slab to the right
922 * kmem_list3 and not this cpu's kmem_list3
925 list_for_each_entry(cachep, &cache_chain, next) {
926 /* setup the size64 kmemlist for cpu before we can
927 * begin anything. Make sure some other cpu on this
928 * node has not already allocated this
930 if (!cachep->nodelists[node]) {
931 if (!(l3 = kmalloc_node(memsize,
935 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
936 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
938 cachep->nodelists[node] = l3;
941 spin_lock_irq(&cachep->nodelists[node]->list_lock);
942 cachep->nodelists[node]->free_limit =
943 (1 + nr_cpus_node(node)) *
944 cachep->batchcount + cachep->num;
945 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
948 /* Now we can go ahead with allocating the shared array's
950 list_for_each_entry(cachep, &cache_chain, next) {
951 struct array_cache *nc;
953 nc = alloc_arraycache(node, cachep->limit,
957 cachep->array[cpu] = nc;
959 l3 = cachep->nodelists[node];
962 if (!(nc = alloc_arraycache(node,
968 /* we are serialised from CPU_DEAD or
969 CPU_UP_CANCELLED by the cpucontrol lock */
973 mutex_unlock(&cache_chain_mutex);
976 start_cpu_timer(cpu);
978 #ifdef CONFIG_HOTPLUG_CPU
981 case CPU_UP_CANCELED:
982 mutex_lock(&cache_chain_mutex);
984 list_for_each_entry(cachep, &cache_chain, next) {
985 struct array_cache *nc;
988 mask = node_to_cpumask(node);
989 spin_lock_irq(&cachep->spinlock);
990 /* cpu is dead; no one can alloc from it. */
991 nc = cachep->array[cpu];
992 cachep->array[cpu] = NULL;
993 l3 = cachep->nodelists[node];
998 spin_lock(&l3->list_lock);
1000 /* Free limit for this kmem_list3 */
1001 l3->free_limit -= cachep->batchcount;
1003 free_block(cachep, nc->entry, nc->avail, node);
1005 if (!cpus_empty(mask)) {
1006 spin_unlock(&l3->list_lock);
1011 free_block(cachep, l3->shared->entry,
1012 l3->shared->avail, node);
1017 drain_alien_cache(cachep, l3);
1018 free_alien_cache(l3->alien);
1022 /* free slabs belonging to this node */
1023 if (__node_shrink(cachep, node)) {
1024 cachep->nodelists[node] = NULL;
1025 spin_unlock(&l3->list_lock);
1028 spin_unlock(&l3->list_lock);
1031 spin_unlock_irq(&cachep->spinlock);
1034 mutex_unlock(&cache_chain_mutex);
1040 mutex_unlock(&cache_chain_mutex);
1044 static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
1047 * swap the static kmem_list3 with kmalloced memory
1049 static void init_list(kmem_cache_t *cachep, struct kmem_list3 *list, int nodeid)
1051 struct kmem_list3 *ptr;
1053 BUG_ON(cachep->nodelists[nodeid] != list);
1054 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1057 local_irq_disable();
1058 memcpy(ptr, list, sizeof(struct kmem_list3));
1059 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1060 cachep->nodelists[nodeid] = ptr;
1065 * Called after the gfp() functions have been enabled, and before smp_init().
1067 void __init kmem_cache_init(void)
1070 struct cache_sizes *sizes;
1071 struct cache_names *names;
1074 for (i = 0; i < NUM_INIT_LISTS; i++) {
1075 kmem_list3_init(&initkmem_list3[i]);
1076 if (i < MAX_NUMNODES)
1077 cache_cache.nodelists[i] = NULL;
1081 * Fragmentation resistance on low memory - only use bigger
1082 * page orders on machines with more than 32MB of memory.
1084 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1085 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1087 /* Bootstrap is tricky, because several objects are allocated
1088 * from caches that do not exist yet:
1089 * 1) initialize the cache_cache cache: it contains the kmem_cache_t
1090 * structures of all caches, except cache_cache itself: cache_cache
1091 * is statically allocated.
1092 * Initially an __init data area is used for the head array and the
1093 * kmem_list3 structures, it's replaced with a kmalloc allocated
1094 * array at the end of the bootstrap.
1095 * 2) Create the first kmalloc cache.
1096 * The kmem_cache_t for the new cache is allocated normally.
1097 * An __init data area is used for the head array.
1098 * 3) Create the remaining kmalloc caches, with minimally sized
1100 * 4) Replace the __init data head arrays for cache_cache and the first
1101 * kmalloc cache with kmalloc allocated arrays.
1102 * 5) Replace the __init data for kmem_list3 for cache_cache and
1103 * the other cache's with kmalloc allocated memory.
1104 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1107 /* 1) create the cache_cache */
1108 INIT_LIST_HEAD(&cache_chain);
1109 list_add(&cache_cache.next, &cache_chain);
1110 cache_cache.colour_off = cache_line_size();
1111 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1112 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
1114 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, cache_line_size());
1116 cache_estimate(0, cache_cache.buffer_size, cache_line_size(), 0,
1117 &left_over, &cache_cache.num);
1118 if (!cache_cache.num)
1121 cache_cache.colour = left_over / cache_cache.colour_off;
1122 cache_cache.colour_next = 0;
1123 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1124 sizeof(struct slab), cache_line_size());
1126 /* 2+3) create the kmalloc caches */
1127 sizes = malloc_sizes;
1128 names = cache_names;
1130 /* Initialize the caches that provide memory for the array cache
1131 * and the kmem_list3 structures first.
1132 * Without this, further allocations will bug
1135 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1136 sizes[INDEX_AC].cs_size,
1137 ARCH_KMALLOC_MINALIGN,
1138 (ARCH_KMALLOC_FLAGS |
1139 SLAB_PANIC), NULL, NULL);
1141 if (INDEX_AC != INDEX_L3)
1142 sizes[INDEX_L3].cs_cachep =
1143 kmem_cache_create(names[INDEX_L3].name,
1144 sizes[INDEX_L3].cs_size,
1145 ARCH_KMALLOC_MINALIGN,
1146 (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL,
1149 while (sizes->cs_size != ULONG_MAX) {
1151 * For performance, all the general caches are L1 aligned.
1152 * This should be particularly beneficial on SMP boxes, as it
1153 * eliminates "false sharing".
1154 * Note for systems short on memory removing the alignment will
1155 * allow tighter packing of the smaller caches.
1157 if (!sizes->cs_cachep)
1158 sizes->cs_cachep = kmem_cache_create(names->name,
1160 ARCH_KMALLOC_MINALIGN,
1165 /* Inc off-slab bufctl limit until the ceiling is hit. */
1166 if (!(OFF_SLAB(sizes->cs_cachep))) {
1167 offslab_limit = sizes->cs_size - sizeof(struct slab);
1168 offslab_limit /= sizeof(kmem_bufctl_t);
1171 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
1173 ARCH_KMALLOC_MINALIGN,
1174 (ARCH_KMALLOC_FLAGS |
1182 /* 4) Replace the bootstrap head arrays */
1186 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1188 local_irq_disable();
1189 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1190 memcpy(ptr, cpu_cache_get(&cache_cache),
1191 sizeof(struct arraycache_init));
1192 cache_cache.array[smp_processor_id()] = ptr;
1195 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1197 local_irq_disable();
1198 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1199 != &initarray_generic.cache);
1200 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1201 sizeof(struct arraycache_init));
1202 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1206 /* 5) Replace the bootstrap kmem_list3's */
1209 /* Replace the static kmem_list3 structures for the boot cpu */
1210 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
1213 for_each_online_node(node) {
1214 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1215 &initkmem_list3[SIZE_AC + node], node);
1217 if (INDEX_AC != INDEX_L3) {
1218 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1219 &initkmem_list3[SIZE_L3 + node],
1225 /* 6) resize the head arrays to their final sizes */
1227 kmem_cache_t *cachep;
1228 mutex_lock(&cache_chain_mutex);
1229 list_for_each_entry(cachep, &cache_chain, next)
1230 enable_cpucache(cachep);
1231 mutex_unlock(&cache_chain_mutex);
1235 g_cpucache_up = FULL;
1237 /* Register a cpu startup notifier callback
1238 * that initializes cpu_cache_get for all new cpus
1240 register_cpu_notifier(&cpucache_notifier);
1242 /* The reap timers are started later, with a module init call:
1243 * That part of the kernel is not yet operational.
1247 static int __init cpucache_init(void)
1252 * Register the timers that return unneeded
1255 for_each_online_cpu(cpu)
1256 start_cpu_timer(cpu);
1261 __initcall(cpucache_init);
1264 * Interface to system's page allocator. No need to hold the cache-lock.
1266 * If we requested dmaable memory, we will get it. Even if we
1267 * did not request dmaable memory, we might get it, but that
1268 * would be relatively rare and ignorable.
1270 static void *kmem_getpages(kmem_cache_t *cachep, gfp_t flags, int nodeid)
1276 flags |= cachep->gfpflags;
1277 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1280 addr = page_address(page);
1282 i = (1 << cachep->gfporder);
1283 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1284 atomic_add(i, &slab_reclaim_pages);
1285 add_page_state(nr_slab, i);
1294 * Interface to system's page release.
1296 static void kmem_freepages(kmem_cache_t *cachep, void *addr)
1298 unsigned long i = (1 << cachep->gfporder);
1299 struct page *page = virt_to_page(addr);
1300 const unsigned long nr_freed = i;
1303 if (!TestClearPageSlab(page))
1307 sub_page_state(nr_slab, nr_freed);
1308 if (current->reclaim_state)
1309 current->reclaim_state->reclaimed_slab += nr_freed;
1310 free_pages((unsigned long)addr, cachep->gfporder);
1311 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1312 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages);
1315 static void kmem_rcu_free(struct rcu_head *head)
1317 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1318 kmem_cache_t *cachep = slab_rcu->cachep;
1320 kmem_freepages(cachep, slab_rcu->addr);
1321 if (OFF_SLAB(cachep))
1322 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1327 #ifdef CONFIG_DEBUG_PAGEALLOC
1328 static void store_stackinfo(kmem_cache_t *cachep, unsigned long *addr,
1329 unsigned long caller)
1331 int size = obj_size(cachep);
1333 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1335 if (size < 5 * sizeof(unsigned long))
1338 *addr++ = 0x12345678;
1340 *addr++ = smp_processor_id();
1341 size -= 3 * sizeof(unsigned long);
1343 unsigned long *sptr = &caller;
1344 unsigned long svalue;
1346 while (!kstack_end(sptr)) {
1348 if (kernel_text_address(svalue)) {
1350 size -= sizeof(unsigned long);
1351 if (size <= sizeof(unsigned long))
1357 *addr++ = 0x87654321;
1361 static void poison_obj(kmem_cache_t *cachep, void *addr, unsigned char val)
1363 int size = obj_size(cachep);
1364 addr = &((char *)addr)[obj_offset(cachep)];
1366 memset(addr, val, size);
1367 *(unsigned char *)(addr + size - 1) = POISON_END;
1370 static void dump_line(char *data, int offset, int limit)
1373 printk(KERN_ERR "%03x:", offset);
1374 for (i = 0; i < limit; i++) {
1375 printk(" %02x", (unsigned char)data[offset + i]);
1383 static void print_objinfo(kmem_cache_t *cachep, void *objp, int lines)
1388 if (cachep->flags & SLAB_RED_ZONE) {
1389 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
1390 *dbg_redzone1(cachep, objp),
1391 *dbg_redzone2(cachep, objp));
1394 if (cachep->flags & SLAB_STORE_USER) {
1395 printk(KERN_ERR "Last user: [<%p>]",
1396 *dbg_userword(cachep, objp));
1397 print_symbol("(%s)",
1398 (unsigned long)*dbg_userword(cachep, objp));
1401 realobj = (char *)objp + obj_offset(cachep);
1402 size = obj_size(cachep);
1403 for (i = 0; i < size && lines; i += 16, lines--) {
1406 if (i + limit > size)
1408 dump_line(realobj, i, limit);
1412 static void check_poison_obj(kmem_cache_t *cachep, void *objp)
1418 realobj = (char *)objp + obj_offset(cachep);
1419 size = obj_size(cachep);
1421 for (i = 0; i < size; i++) {
1422 char exp = POISON_FREE;
1425 if (realobj[i] != exp) {
1431 "Slab corruption: start=%p, len=%d\n",
1433 print_objinfo(cachep, objp, 0);
1435 /* Hexdump the affected line */
1438 if (i + limit > size)
1440 dump_line(realobj, i, limit);
1443 /* Limit to 5 lines */
1449 /* Print some data about the neighboring objects, if they
1452 struct slab *slabp = virt_to_slab(objp);
1455 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
1457 objp = slabp->s_mem + (objnr - 1) * cachep->buffer_size;
1458 realobj = (char *)objp + obj_offset(cachep);
1459 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1461 print_objinfo(cachep, objp, 2);
1463 if (objnr + 1 < cachep->num) {
1464 objp = slabp->s_mem + (objnr + 1) * cachep->buffer_size;
1465 realobj = (char *)objp + obj_offset(cachep);
1466 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1468 print_objinfo(cachep, objp, 2);
1476 * slab_destroy_objs - call the registered destructor for each object in
1477 * a slab that is to be destroyed.
1479 static void slab_destroy_objs(kmem_cache_t *cachep, struct slab *slabp)
1482 for (i = 0; i < cachep->num; i++) {
1483 void *objp = slabp->s_mem + cachep->buffer_size * i;
1485 if (cachep->flags & SLAB_POISON) {
1486 #ifdef CONFIG_DEBUG_PAGEALLOC
1487 if ((cachep->buffer_size % PAGE_SIZE) == 0
1488 && OFF_SLAB(cachep))
1489 kernel_map_pages(virt_to_page(objp),
1490 cachep->buffer_size / PAGE_SIZE,
1493 check_poison_obj(cachep, objp);
1495 check_poison_obj(cachep, objp);
1498 if (cachep->flags & SLAB_RED_ZONE) {
1499 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1500 slab_error(cachep, "start of a freed object "
1502 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1503 slab_error(cachep, "end of a freed object "
1506 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1507 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
1511 static void slab_destroy_objs(kmem_cache_t *cachep, struct slab *slabp)
1515 for (i = 0; i < cachep->num; i++) {
1516 void *objp = slabp->s_mem + cachep->buffer_size * i;
1517 (cachep->dtor) (objp, cachep, 0);
1524 * Destroy all the objs in a slab, and release the mem back to the system.
1525 * Before calling the slab must have been unlinked from the cache.
1526 * The cache-lock is not held/needed.
1528 static void slab_destroy(kmem_cache_t *cachep, struct slab *slabp)
1530 void *addr = slabp->s_mem - slabp->colouroff;
1532 slab_destroy_objs(cachep, slabp);
1533 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1534 struct slab_rcu *slab_rcu;
1536 slab_rcu = (struct slab_rcu *)slabp;
1537 slab_rcu->cachep = cachep;
1538 slab_rcu->addr = addr;
1539 call_rcu(&slab_rcu->head, kmem_rcu_free);
1541 kmem_freepages(cachep, addr);
1542 if (OFF_SLAB(cachep))
1543 kmem_cache_free(cachep->slabp_cache, slabp);
1547 /* For setting up all the kmem_list3s for cache whose buffer_size is same
1548 as size of kmem_list3. */
1549 static void set_up_list3s(kmem_cache_t *cachep, int index)
1553 for_each_online_node(node) {
1554 cachep->nodelists[node] = &initkmem_list3[index + node];
1555 cachep->nodelists[node]->next_reap = jiffies +
1557 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1562 * calculate_slab_order - calculate size (page order) of slabs and the number
1563 * of objects per slab.
1565 * This could be made much more intelligent. For now, try to avoid using
1566 * high order pages for slabs. When the gfp() functions are more friendly
1567 * towards high-order requests, this should be changed.
1569 static inline size_t calculate_slab_order(kmem_cache_t *cachep, size_t size,
1570 size_t align, gfp_t flags)
1572 size_t left_over = 0;
1574 for (;; cachep->gfporder++) {
1578 if (cachep->gfporder > MAX_GFP_ORDER) {
1583 cache_estimate(cachep->gfporder, size, align, flags,
1587 /* More than offslab_limit objects will cause problems */
1588 if (flags & CFLGS_OFF_SLAB && cachep->num > offslab_limit)
1592 left_over = remainder;
1595 * Large number of objects is good, but very large slabs are
1596 * currently bad for the gfp()s.
1598 if (cachep->gfporder >= slab_break_gfp_order)
1601 if ((left_over * 8) <= (PAGE_SIZE << cachep->gfporder))
1602 /* Acceptable internal fragmentation */
1609 * kmem_cache_create - Create a cache.
1610 * @name: A string which is used in /proc/slabinfo to identify this cache.
1611 * @size: The size of objects to be created in this cache.
1612 * @align: The required alignment for the objects.
1613 * @flags: SLAB flags
1614 * @ctor: A constructor for the objects.
1615 * @dtor: A destructor for the objects.
1617 * Returns a ptr to the cache on success, NULL on failure.
1618 * Cannot be called within a int, but can be interrupted.
1619 * The @ctor is run when new pages are allocated by the cache
1620 * and the @dtor is run before the pages are handed back.
1622 * @name must be valid until the cache is destroyed. This implies that
1623 * the module calling this has to destroy the cache before getting
1628 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1629 * to catch references to uninitialised memory.
1631 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1632 * for buffer overruns.
1634 * %SLAB_NO_REAP - Don't automatically reap this cache when we're under
1637 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1638 * cacheline. This can be beneficial if you're counting cycles as closely
1642 kmem_cache_create (const char *name, size_t size, size_t align,
1643 unsigned long flags, void (*ctor)(void*, kmem_cache_t *, unsigned long),
1644 void (*dtor)(void*, kmem_cache_t *, unsigned long))
1646 size_t left_over, slab_size, ralign;
1647 kmem_cache_t *cachep = NULL;
1648 struct list_head *p;
1651 * Sanity checks... these are all serious usage bugs.
1655 (size < BYTES_PER_WORD) ||
1656 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
1657 printk(KERN_ERR "%s: Early error in slab %s\n",
1658 __FUNCTION__, name);
1662 mutex_lock(&cache_chain_mutex);
1664 list_for_each(p, &cache_chain) {
1665 kmem_cache_t *pc = list_entry(p, kmem_cache_t, next);
1666 mm_segment_t old_fs = get_fs();
1671 * This happens when the module gets unloaded and doesn't
1672 * destroy its slab cache and no-one else reuses the vmalloc
1673 * area of the module. Print a warning.
1676 res = __get_user(tmp, pc->name);
1679 printk("SLAB: cache with size %d has lost its name\n",
1684 if (!strcmp(pc->name, name)) {
1685 printk("kmem_cache_create: duplicate cache %s\n", name);
1692 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1693 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
1694 /* No constructor, but inital state check requested */
1695 printk(KERN_ERR "%s: No con, but init state check "
1696 "requested - %s\n", __FUNCTION__, name);
1697 flags &= ~SLAB_DEBUG_INITIAL;
1701 * Enable redzoning and last user accounting, except for caches with
1702 * large objects, if the increased size would increase the object size
1703 * above the next power of two: caches with object sizes just above a
1704 * power of two have a significant amount of internal fragmentation.
1707 || fls(size - 1) == fls(size - 1 + 3 * BYTES_PER_WORD)))
1708 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
1709 if (!(flags & SLAB_DESTROY_BY_RCU))
1710 flags |= SLAB_POISON;
1712 if (flags & SLAB_DESTROY_BY_RCU)
1713 BUG_ON(flags & SLAB_POISON);
1715 if (flags & SLAB_DESTROY_BY_RCU)
1719 * Always checks flags, a caller might be expecting debug
1720 * support which isn't available.
1722 if (flags & ~CREATE_MASK)
1725 /* Check that size is in terms of words. This is needed to avoid
1726 * unaligned accesses for some archs when redzoning is used, and makes
1727 * sure any on-slab bufctl's are also correctly aligned.
1729 if (size & (BYTES_PER_WORD - 1)) {
1730 size += (BYTES_PER_WORD - 1);
1731 size &= ~(BYTES_PER_WORD - 1);
1734 /* calculate out the final buffer alignment: */
1735 /* 1) arch recommendation: can be overridden for debug */
1736 if (flags & SLAB_HWCACHE_ALIGN) {
1737 /* Default alignment: as specified by the arch code.
1738 * Except if an object is really small, then squeeze multiple
1739 * objects into one cacheline.
1741 ralign = cache_line_size();
1742 while (size <= ralign / 2)
1745 ralign = BYTES_PER_WORD;
1747 /* 2) arch mandated alignment: disables debug if necessary */
1748 if (ralign < ARCH_SLAB_MINALIGN) {
1749 ralign = ARCH_SLAB_MINALIGN;
1750 if (ralign > BYTES_PER_WORD)
1751 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
1753 /* 3) caller mandated alignment: disables debug if necessary */
1754 if (ralign < align) {
1756 if (ralign > BYTES_PER_WORD)
1757 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
1759 /* 4) Store it. Note that the debug code below can reduce
1760 * the alignment to BYTES_PER_WORD.
1764 /* Get cache's description obj. */
1765 cachep = (kmem_cache_t *) kmem_cache_alloc(&cache_cache, SLAB_KERNEL);
1768 memset(cachep, 0, sizeof(kmem_cache_t));
1771 cachep->obj_size = size;
1773 if (flags & SLAB_RED_ZONE) {
1774 /* redzoning only works with word aligned caches */
1775 align = BYTES_PER_WORD;
1777 /* add space for red zone words */
1778 cachep->obj_offset += BYTES_PER_WORD;
1779 size += 2 * BYTES_PER_WORD;
1781 if (flags & SLAB_STORE_USER) {
1782 /* user store requires word alignment and
1783 * one word storage behind the end of the real
1786 align = BYTES_PER_WORD;
1787 size += BYTES_PER_WORD;
1789 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
1790 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
1791 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
1792 cachep->obj_offset += PAGE_SIZE - size;
1798 /* Determine if the slab management is 'on' or 'off' slab. */
1799 if (size >= (PAGE_SIZE >> 3))
1801 * Size is large, assume best to place the slab management obj
1802 * off-slab (should allow better packing of objs).
1804 flags |= CFLGS_OFF_SLAB;
1806 size = ALIGN(size, align);
1808 if ((flags & SLAB_RECLAIM_ACCOUNT) && size <= PAGE_SIZE) {
1810 * A VFS-reclaimable slab tends to have most allocations
1811 * as GFP_NOFS and we really don't want to have to be allocating
1812 * higher-order pages when we are unable to shrink dcache.
1814 cachep->gfporder = 0;
1815 cache_estimate(cachep->gfporder, size, align, flags,
1816 &left_over, &cachep->num);
1818 left_over = calculate_slab_order(cachep, size, align, flags);
1821 printk("kmem_cache_create: couldn't create cache %s.\n", name);
1822 kmem_cache_free(&cache_cache, cachep);
1826 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
1827 + sizeof(struct slab), align);
1830 * If the slab has been placed off-slab, and we have enough space then
1831 * move it on-slab. This is at the expense of any extra colouring.
1833 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
1834 flags &= ~CFLGS_OFF_SLAB;
1835 left_over -= slab_size;
1838 if (flags & CFLGS_OFF_SLAB) {
1839 /* really off slab. No need for manual alignment */
1841 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
1844 cachep->colour_off = cache_line_size();
1845 /* Offset must be a multiple of the alignment. */
1846 if (cachep->colour_off < align)
1847 cachep->colour_off = align;
1848 cachep->colour = left_over / cachep->colour_off;
1849 cachep->slab_size = slab_size;
1850 cachep->flags = flags;
1851 cachep->gfpflags = 0;
1852 if (flags & SLAB_CACHE_DMA)
1853 cachep->gfpflags |= GFP_DMA;
1854 spin_lock_init(&cachep->spinlock);
1855 cachep->buffer_size = size;
1857 if (flags & CFLGS_OFF_SLAB)
1858 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
1859 cachep->ctor = ctor;
1860 cachep->dtor = dtor;
1861 cachep->name = name;
1863 /* Don't let CPUs to come and go */
1866 if (g_cpucache_up == FULL) {
1867 enable_cpucache(cachep);
1869 if (g_cpucache_up == NONE) {
1870 /* Note: the first kmem_cache_create must create
1871 * the cache that's used by kmalloc(24), otherwise
1872 * the creation of further caches will BUG().
1874 cachep->array[smp_processor_id()] =
1875 &initarray_generic.cache;
1877 /* If the cache that's used by
1878 * kmalloc(sizeof(kmem_list3)) is the first cache,
1879 * then we need to set up all its list3s, otherwise
1880 * the creation of further caches will BUG().
1882 set_up_list3s(cachep, SIZE_AC);
1883 if (INDEX_AC == INDEX_L3)
1884 g_cpucache_up = PARTIAL_L3;
1886 g_cpucache_up = PARTIAL_AC;
1888 cachep->array[smp_processor_id()] =
1889 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1891 if (g_cpucache_up == PARTIAL_AC) {
1892 set_up_list3s(cachep, SIZE_L3);
1893 g_cpucache_up = PARTIAL_L3;
1896 for_each_online_node(node) {
1898 cachep->nodelists[node] =
1900 (struct kmem_list3),
1902 BUG_ON(!cachep->nodelists[node]);
1903 kmem_list3_init(cachep->
1908 cachep->nodelists[numa_node_id()]->next_reap =
1909 jiffies + REAPTIMEOUT_LIST3 +
1910 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1912 BUG_ON(!cpu_cache_get(cachep));
1913 cpu_cache_get(cachep)->avail = 0;
1914 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1915 cpu_cache_get(cachep)->batchcount = 1;
1916 cpu_cache_get(cachep)->touched = 0;
1917 cachep->batchcount = 1;
1918 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1921 /* cache setup completed, link it into the list */
1922 list_add(&cachep->next, &cache_chain);
1923 unlock_cpu_hotplug();
1925 if (!cachep && (flags & SLAB_PANIC))
1926 panic("kmem_cache_create(): failed to create slab `%s'\n",
1928 mutex_unlock(&cache_chain_mutex);
1931 EXPORT_SYMBOL(kmem_cache_create);
1934 static void check_irq_off(void)
1936 BUG_ON(!irqs_disabled());
1939 static void check_irq_on(void)
1941 BUG_ON(irqs_disabled());
1944 static void check_spinlock_acquired(kmem_cache_t *cachep)
1948 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
1952 static void check_spinlock_acquired_node(kmem_cache_t *cachep, int node)
1956 assert_spin_locked(&cachep->nodelists[node]->list_lock);
1961 #define check_irq_off() do { } while(0)
1962 #define check_irq_on() do { } while(0)
1963 #define check_spinlock_acquired(x) do { } while(0)
1964 #define check_spinlock_acquired_node(x, y) do { } while(0)
1968 * Waits for all CPUs to execute func().
1970 static void smp_call_function_all_cpus(void (*func)(void *arg), void *arg)
1975 local_irq_disable();
1979 if (smp_call_function(func, arg, 1, 1))
1985 static void drain_array_locked(kmem_cache_t *cachep, struct array_cache *ac,
1986 int force, int node);
1988 static void do_drain(void *arg)
1990 kmem_cache_t *cachep = (kmem_cache_t *) arg;
1991 struct array_cache *ac;
1992 int node = numa_node_id();
1995 ac = cpu_cache_get(cachep);
1996 spin_lock(&cachep->nodelists[node]->list_lock);
1997 free_block(cachep, ac->entry, ac->avail, node);
1998 spin_unlock(&cachep->nodelists[node]->list_lock);
2002 static void drain_cpu_caches(kmem_cache_t *cachep)
2004 struct kmem_list3 *l3;
2007 smp_call_function_all_cpus(do_drain, cachep);
2009 spin_lock_irq(&cachep->spinlock);
2010 for_each_online_node(node) {
2011 l3 = cachep->nodelists[node];
2013 spin_lock(&l3->list_lock);
2014 drain_array_locked(cachep, l3->shared, 1, node);
2015 spin_unlock(&l3->list_lock);
2017 drain_alien_cache(cachep, l3);
2020 spin_unlock_irq(&cachep->spinlock);
2023 static int __node_shrink(kmem_cache_t *cachep, int node)
2026 struct kmem_list3 *l3 = cachep->nodelists[node];
2030 struct list_head *p;
2032 p = l3->slabs_free.prev;
2033 if (p == &l3->slabs_free)
2036 slabp = list_entry(l3->slabs_free.prev, struct slab, list);
2041 list_del(&slabp->list);
2043 l3->free_objects -= cachep->num;
2044 spin_unlock_irq(&l3->list_lock);
2045 slab_destroy(cachep, slabp);
2046 spin_lock_irq(&l3->list_lock);
2048 ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial);
2052 static int __cache_shrink(kmem_cache_t *cachep)
2055 struct kmem_list3 *l3;
2057 drain_cpu_caches(cachep);
2060 for_each_online_node(i) {
2061 l3 = cachep->nodelists[i];
2063 spin_lock_irq(&l3->list_lock);
2064 ret += __node_shrink(cachep, i);
2065 spin_unlock_irq(&l3->list_lock);
2068 return (ret ? 1 : 0);
2072 * kmem_cache_shrink - Shrink a cache.
2073 * @cachep: The cache to shrink.
2075 * Releases as many slabs as possible for a cache.
2076 * To help debugging, a zero exit status indicates all slabs were released.
2078 int kmem_cache_shrink(kmem_cache_t *cachep)
2080 if (!cachep || in_interrupt())
2083 return __cache_shrink(cachep);
2085 EXPORT_SYMBOL(kmem_cache_shrink);
2088 * kmem_cache_destroy - delete a cache
2089 * @cachep: the cache to destroy
2091 * Remove a kmem_cache_t object from the slab cache.
2092 * Returns 0 on success.
2094 * It is expected this function will be called by a module when it is
2095 * unloaded. This will remove the cache completely, and avoid a duplicate
2096 * cache being allocated each time a module is loaded and unloaded, if the
2097 * module doesn't have persistent in-kernel storage across loads and unloads.
2099 * The cache must be empty before calling this function.
2101 * The caller must guarantee that noone will allocate memory from the cache
2102 * during the kmem_cache_destroy().
2104 int kmem_cache_destroy(kmem_cache_t *cachep)
2107 struct kmem_list3 *l3;
2109 if (!cachep || in_interrupt())
2112 /* Don't let CPUs to come and go */
2115 /* Find the cache in the chain of caches. */
2116 mutex_lock(&cache_chain_mutex);
2118 * the chain is never empty, cache_cache is never destroyed
2120 list_del(&cachep->next);
2121 mutex_unlock(&cache_chain_mutex);
2123 if (__cache_shrink(cachep)) {
2124 slab_error(cachep, "Can't free all objects");
2125 mutex_lock(&cache_chain_mutex);
2126 list_add(&cachep->next, &cache_chain);
2127 mutex_unlock(&cache_chain_mutex);
2128 unlock_cpu_hotplug();
2132 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2135 for_each_online_cpu(i)
2136 kfree(cachep->array[i]);
2138 /* NUMA: free the list3 structures */
2139 for_each_online_node(i) {
2140 if ((l3 = cachep->nodelists[i])) {
2142 free_alien_cache(l3->alien);
2146 kmem_cache_free(&cache_cache, cachep);
2148 unlock_cpu_hotplug();
2152 EXPORT_SYMBOL(kmem_cache_destroy);
2154 /* Get the memory for a slab management obj. */
2155 static struct slab *alloc_slabmgmt(kmem_cache_t *cachep, void *objp,
2156 int colour_off, gfp_t local_flags)
2160 if (OFF_SLAB(cachep)) {
2161 /* Slab management obj is off-slab. */
2162 slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags);
2166 slabp = objp + colour_off;
2167 colour_off += cachep->slab_size;
2170 slabp->colouroff = colour_off;
2171 slabp->s_mem = objp + colour_off;
2176 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2178 return (kmem_bufctl_t *) (slabp + 1);
2181 static void cache_init_objs(kmem_cache_t *cachep,
2182 struct slab *slabp, unsigned long ctor_flags)
2186 for (i = 0; i < cachep->num; i++) {
2187 void *objp = slabp->s_mem + cachep->buffer_size * i;
2189 /* need to poison the objs? */
2190 if (cachep->flags & SLAB_POISON)
2191 poison_obj(cachep, objp, POISON_FREE);
2192 if (cachep->flags & SLAB_STORE_USER)
2193 *dbg_userword(cachep, objp) = NULL;
2195 if (cachep->flags & SLAB_RED_ZONE) {
2196 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2197 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2200 * Constructors are not allowed to allocate memory from
2201 * the same cache which they are a constructor for.
2202 * Otherwise, deadlock. They must also be threaded.
2204 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2205 cachep->ctor(objp + obj_offset(cachep), cachep,
2208 if (cachep->flags & SLAB_RED_ZONE) {
2209 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2210 slab_error(cachep, "constructor overwrote the"
2211 " end of an object");
2212 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2213 slab_error(cachep, "constructor overwrote the"
2214 " start of an object");
2216 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)
2217 && cachep->flags & SLAB_POISON)
2218 kernel_map_pages(virt_to_page(objp),
2219 cachep->buffer_size / PAGE_SIZE, 0);
2222 cachep->ctor(objp, cachep, ctor_flags);
2224 slab_bufctl(slabp)[i] = i + 1;
2226 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2230 static void kmem_flagcheck(kmem_cache_t *cachep, gfp_t flags)
2232 if (flags & SLAB_DMA) {
2233 if (!(cachep->gfpflags & GFP_DMA))
2236 if (cachep->gfpflags & GFP_DMA)
2241 static void *slab_get_obj(kmem_cache_t *cachep, struct slab *slabp, int nodeid)
2243 void *objp = slabp->s_mem + (slabp->free * cachep->buffer_size);
2247 next = slab_bufctl(slabp)[slabp->free];
2249 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2250 WARN_ON(slabp->nodeid != nodeid);
2257 static void slab_put_obj(kmem_cache_t *cachep, struct slab *slabp, void *objp,
2260 unsigned int objnr = (unsigned)(objp-slabp->s_mem) / cachep->buffer_size;
2263 /* Verify that the slab belongs to the intended node */
2264 WARN_ON(slabp->nodeid != nodeid);
2266 if (slab_bufctl(slabp)[objnr] != BUFCTL_FREE) {
2267 printk(KERN_ERR "slab: double free detected in cache "
2268 "'%s', objp %p\n", cachep->name, objp);
2272 slab_bufctl(slabp)[objnr] = slabp->free;
2273 slabp->free = objnr;
2277 static void set_slab_attr(kmem_cache_t *cachep, struct slab *slabp, void *objp)
2282 /* Nasty!!!!!! I hope this is OK. */
2283 i = 1 << cachep->gfporder;
2284 page = virt_to_page(objp);
2286 page_set_cache(page, cachep);
2287 page_set_slab(page, slabp);
2293 * Grow (by 1) the number of slabs within a cache. This is called by
2294 * kmem_cache_alloc() when there are no active objs left in a cache.
2296 static int cache_grow(kmem_cache_t *cachep, gfp_t flags, int nodeid)
2302 unsigned long ctor_flags;
2303 struct kmem_list3 *l3;
2305 /* Be lazy and only check for valid flags here,
2306 * keeping it out of the critical path in kmem_cache_alloc().
2308 if (flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW))
2310 if (flags & SLAB_NO_GROW)
2313 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2314 local_flags = (flags & SLAB_LEVEL_MASK);
2315 if (!(local_flags & __GFP_WAIT))
2317 * Not allowed to sleep. Need to tell a constructor about
2318 * this - it might need to know...
2320 ctor_flags |= SLAB_CTOR_ATOMIC;
2322 /* About to mess with non-constant members - lock. */
2324 spin_lock(&cachep->spinlock);
2326 /* Get colour for the slab, and cal the next value. */
2327 offset = cachep->colour_next;
2328 cachep->colour_next++;
2329 if (cachep->colour_next >= cachep->colour)
2330 cachep->colour_next = 0;
2331 offset *= cachep->colour_off;
2333 spin_unlock(&cachep->spinlock);
2336 if (local_flags & __GFP_WAIT)
2340 * The test for missing atomic flag is performed here, rather than
2341 * the more obvious place, simply to reduce the critical path length
2342 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2343 * will eventually be caught here (where it matters).
2345 kmem_flagcheck(cachep, flags);
2347 /* Get mem for the objs.
2348 * Attempt to allocate a physical page from 'nodeid',
2350 if (!(objp = kmem_getpages(cachep, flags, nodeid)))
2353 /* Get slab management. */
2354 if (!(slabp = alloc_slabmgmt(cachep, objp, offset, local_flags)))
2357 slabp->nodeid = nodeid;
2358 set_slab_attr(cachep, slabp, objp);
2360 cache_init_objs(cachep, slabp, ctor_flags);
2362 if (local_flags & __GFP_WAIT)
2363 local_irq_disable();
2365 l3 = cachep->nodelists[nodeid];
2366 spin_lock(&l3->list_lock);
2368 /* Make slab active. */
2369 list_add_tail(&slabp->list, &(l3->slabs_free));
2370 STATS_INC_GROWN(cachep);
2371 l3->free_objects += cachep->num;
2372 spin_unlock(&l3->list_lock);
2375 kmem_freepages(cachep, objp);
2377 if (local_flags & __GFP_WAIT)
2378 local_irq_disable();
2385 * Perform extra freeing checks:
2386 * - detect bad pointers.
2387 * - POISON/RED_ZONE checking
2388 * - destructor calls, for caches with POISON+dtor
2390 static void kfree_debugcheck(const void *objp)
2394 if (!virt_addr_valid(objp)) {
2395 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2396 (unsigned long)objp);
2399 page = virt_to_page(objp);
2400 if (!PageSlab(page)) {
2401 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2402 (unsigned long)objp);
2407 static void *cache_free_debugcheck(kmem_cache_t *cachep, void *objp,
2414 objp -= obj_offset(cachep);
2415 kfree_debugcheck(objp);
2416 page = virt_to_page(objp);
2418 if (page_get_cache(page) != cachep) {
2420 "mismatch in kmem_cache_free: expected cache %p, got %p\n",
2421 page_get_cache(page), cachep);
2422 printk(KERN_ERR "%p is %s.\n", cachep, cachep->name);
2423 printk(KERN_ERR "%p is %s.\n", page_get_cache(page),
2424 page_get_cache(page)->name);
2427 slabp = page_get_slab(page);
2429 if (cachep->flags & SLAB_RED_ZONE) {
2430 if (*dbg_redzone1(cachep, objp) != RED_ACTIVE
2431 || *dbg_redzone2(cachep, objp) != RED_ACTIVE) {
2433 "double free, or memory outside"
2434 " object was overwritten");
2436 "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2437 objp, *dbg_redzone1(cachep, objp),
2438 *dbg_redzone2(cachep, objp));
2440 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2441 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2443 if (cachep->flags & SLAB_STORE_USER)
2444 *dbg_userword(cachep, objp) = caller;
2446 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
2448 BUG_ON(objnr >= cachep->num);
2449 BUG_ON(objp != slabp->s_mem + objnr * cachep->buffer_size);
2451 if (cachep->flags & SLAB_DEBUG_INITIAL) {
2452 /* Need to call the slab's constructor so the
2453 * caller can perform a verify of its state (debugging).
2454 * Called without the cache-lock held.
2456 cachep->ctor(objp + obj_offset(cachep),
2457 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
2459 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2460 /* we want to cache poison the object,
2461 * call the destruction callback
2463 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
2465 if (cachep->flags & SLAB_POISON) {
2466 #ifdef CONFIG_DEBUG_PAGEALLOC
2467 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) {
2468 store_stackinfo(cachep, objp, (unsigned long)caller);
2469 kernel_map_pages(virt_to_page(objp),
2470 cachep->buffer_size / PAGE_SIZE, 0);
2472 poison_obj(cachep, objp, POISON_FREE);
2475 poison_obj(cachep, objp, POISON_FREE);
2481 static void check_slabp(kmem_cache_t *cachep, struct slab *slabp)
2486 /* Check slab's freelist to see if this obj is there. */
2487 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2489 if (entries > cachep->num || i >= cachep->num)
2492 if (entries != cachep->num - slabp->inuse) {
2495 "slab: Internal list corruption detected in cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2496 cachep->name, cachep->num, slabp, slabp->inuse);
2498 i < sizeof(slabp) + cachep->num * sizeof(kmem_bufctl_t);
2501 printk("\n%03x:", i);
2502 printk(" %02x", ((unsigned char *)slabp)[i]);
2509 #define kfree_debugcheck(x) do { } while(0)
2510 #define cache_free_debugcheck(x,objp,z) (objp)
2511 #define check_slabp(x,y) do { } while(0)
2514 static void *cache_alloc_refill(kmem_cache_t *cachep, gfp_t flags)
2517 struct kmem_list3 *l3;
2518 struct array_cache *ac;
2521 ac = cpu_cache_get(cachep);
2523 batchcount = ac->batchcount;
2524 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2525 /* if there was little recent activity on this
2526 * cache, then perform only a partial refill.
2527 * Otherwise we could generate refill bouncing.
2529 batchcount = BATCHREFILL_LIMIT;
2531 l3 = cachep->nodelists[numa_node_id()];
2533 BUG_ON(ac->avail > 0 || !l3);
2534 spin_lock(&l3->list_lock);
2537 struct array_cache *shared_array = l3->shared;
2538 if (shared_array->avail) {
2539 if (batchcount > shared_array->avail)
2540 batchcount = shared_array->avail;
2541 shared_array->avail -= batchcount;
2542 ac->avail = batchcount;
2544 &(shared_array->entry[shared_array->avail]),
2545 sizeof(void *) * batchcount);
2546 shared_array->touched = 1;
2550 while (batchcount > 0) {
2551 struct list_head *entry;
2553 /* Get slab alloc is to come from. */
2554 entry = l3->slabs_partial.next;
2555 if (entry == &l3->slabs_partial) {
2556 l3->free_touched = 1;
2557 entry = l3->slabs_free.next;
2558 if (entry == &l3->slabs_free)
2562 slabp = list_entry(entry, struct slab, list);
2563 check_slabp(cachep, slabp);
2564 check_spinlock_acquired(cachep);
2565 while (slabp->inuse < cachep->num && batchcount--) {
2566 STATS_INC_ALLOCED(cachep);
2567 STATS_INC_ACTIVE(cachep);
2568 STATS_SET_HIGH(cachep);
2570 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2573 check_slabp(cachep, slabp);
2575 /* move slabp to correct slabp list: */
2576 list_del(&slabp->list);
2577 if (slabp->free == BUFCTL_END)
2578 list_add(&slabp->list, &l3->slabs_full);
2580 list_add(&slabp->list, &l3->slabs_partial);
2584 l3->free_objects -= ac->avail;
2586 spin_unlock(&l3->list_lock);
2588 if (unlikely(!ac->avail)) {
2590 x = cache_grow(cachep, flags, numa_node_id());
2592 // cache_grow can reenable interrupts, then ac could change.
2593 ac = cpu_cache_get(cachep);
2594 if (!x && ac->avail == 0) // no objects in sight? abort
2597 if (!ac->avail) // objects refilled by interrupt?
2601 return ac->entry[--ac->avail];
2605 cache_alloc_debugcheck_before(kmem_cache_t *cachep, gfp_t flags)
2607 might_sleep_if(flags & __GFP_WAIT);
2609 kmem_flagcheck(cachep, flags);
2614 static void *cache_alloc_debugcheck_after(kmem_cache_t *cachep, gfp_t flags,
2615 void *objp, void *caller)
2619 if (cachep->flags & SLAB_POISON) {
2620 #ifdef CONFIG_DEBUG_PAGEALLOC
2621 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
2622 kernel_map_pages(virt_to_page(objp),
2623 cachep->buffer_size / PAGE_SIZE, 1);
2625 check_poison_obj(cachep, objp);
2627 check_poison_obj(cachep, objp);
2629 poison_obj(cachep, objp, POISON_INUSE);
2631 if (cachep->flags & SLAB_STORE_USER)
2632 *dbg_userword(cachep, objp) = caller;
2634 if (cachep->flags & SLAB_RED_ZONE) {
2635 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE
2636 || *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2638 "double free, or memory outside"
2639 " object was overwritten");
2641 "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2642 objp, *dbg_redzone1(cachep, objp),
2643 *dbg_redzone2(cachep, objp));
2645 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2646 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2648 objp += obj_offset(cachep);
2649 if (cachep->ctor && cachep->flags & SLAB_POISON) {
2650 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2652 if (!(flags & __GFP_WAIT))
2653 ctor_flags |= SLAB_CTOR_ATOMIC;
2655 cachep->ctor(objp, cachep, ctor_flags);
2660 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2663 static inline void *____cache_alloc(kmem_cache_t *cachep, gfp_t flags)
2666 struct array_cache *ac;
2669 if (unlikely(current->mempolicy && !in_interrupt())) {
2670 int nid = slab_node(current->mempolicy);
2672 if (nid != numa_node_id())
2673 return __cache_alloc_node(cachep, flags, nid);
2678 ac = cpu_cache_get(cachep);
2679 if (likely(ac->avail)) {
2680 STATS_INC_ALLOCHIT(cachep);
2682 objp = ac->entry[--ac->avail];
2684 STATS_INC_ALLOCMISS(cachep);
2685 objp = cache_alloc_refill(cachep, flags);
2690 static inline void *__cache_alloc(kmem_cache_t *cachep, gfp_t flags)
2692 unsigned long save_flags;
2695 cache_alloc_debugcheck_before(cachep, flags);
2697 local_irq_save(save_flags);
2698 objp = ____cache_alloc(cachep, flags);
2699 local_irq_restore(save_flags);
2700 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
2701 __builtin_return_address(0));
2708 * A interface to enable slab creation on nodeid
2710 static void *__cache_alloc_node(kmem_cache_t *cachep, gfp_t flags, int nodeid)
2712 struct list_head *entry;
2714 struct kmem_list3 *l3;
2718 l3 = cachep->nodelists[nodeid];
2722 spin_lock(&l3->list_lock);
2723 entry = l3->slabs_partial.next;
2724 if (entry == &l3->slabs_partial) {
2725 l3->free_touched = 1;
2726 entry = l3->slabs_free.next;
2727 if (entry == &l3->slabs_free)
2731 slabp = list_entry(entry, struct slab, list);
2732 check_spinlock_acquired_node(cachep, nodeid);
2733 check_slabp(cachep, slabp);
2735 STATS_INC_NODEALLOCS(cachep);
2736 STATS_INC_ACTIVE(cachep);
2737 STATS_SET_HIGH(cachep);
2739 BUG_ON(slabp->inuse == cachep->num);
2741 obj = slab_get_obj(cachep, slabp, nodeid);
2742 check_slabp(cachep, slabp);
2744 /* move slabp to correct slabp list: */
2745 list_del(&slabp->list);
2747 if (slabp->free == BUFCTL_END) {
2748 list_add(&slabp->list, &l3->slabs_full);
2750 list_add(&slabp->list, &l3->slabs_partial);
2753 spin_unlock(&l3->list_lock);
2757 spin_unlock(&l3->list_lock);
2758 x = cache_grow(cachep, flags, nodeid);
2770 * Caller needs to acquire correct kmem_list's list_lock
2772 static void free_block(kmem_cache_t *cachep, void **objpp, int nr_objects,
2776 struct kmem_list3 *l3;
2778 for (i = 0; i < nr_objects; i++) {
2779 void *objp = objpp[i];
2782 slabp = virt_to_slab(objp);
2783 l3 = cachep->nodelists[node];
2784 list_del(&slabp->list);
2785 check_spinlock_acquired_node(cachep, node);
2786 check_slabp(cachep, slabp);
2787 slab_put_obj(cachep, slabp, objp, node);
2788 STATS_DEC_ACTIVE(cachep);
2790 check_slabp(cachep, slabp);
2792 /* fixup slab chains */
2793 if (slabp->inuse == 0) {
2794 if (l3->free_objects > l3->free_limit) {
2795 l3->free_objects -= cachep->num;
2796 slab_destroy(cachep, slabp);
2798 list_add(&slabp->list, &l3->slabs_free);
2801 /* Unconditionally move a slab to the end of the
2802 * partial list on free - maximum time for the
2803 * other objects to be freed, too.
2805 list_add_tail(&slabp->list, &l3->slabs_partial);
2810 static void cache_flusharray(kmem_cache_t *cachep, struct array_cache *ac)
2813 struct kmem_list3 *l3;
2814 int node = numa_node_id();
2816 batchcount = ac->batchcount;
2818 BUG_ON(!batchcount || batchcount > ac->avail);
2821 l3 = cachep->nodelists[node];
2822 spin_lock(&l3->list_lock);
2824 struct array_cache *shared_array = l3->shared;
2825 int max = shared_array->limit - shared_array->avail;
2827 if (batchcount > max)
2829 memcpy(&(shared_array->entry[shared_array->avail]),
2830 ac->entry, sizeof(void *) * batchcount);
2831 shared_array->avail += batchcount;
2836 free_block(cachep, ac->entry, batchcount, node);
2841 struct list_head *p;
2843 p = l3->slabs_free.next;
2844 while (p != &(l3->slabs_free)) {
2847 slabp = list_entry(p, struct slab, list);
2848 BUG_ON(slabp->inuse);
2853 STATS_SET_FREEABLE(cachep, i);
2856 spin_unlock(&l3->list_lock);
2857 ac->avail -= batchcount;
2858 memmove(ac->entry, &(ac->entry[batchcount]),
2859 sizeof(void *) * ac->avail);
2864 * Release an obj back to its cache. If the obj has a constructed
2865 * state, it must be in this state _before_ it is released.
2867 * Called with disabled ints.
2869 static inline void __cache_free(kmem_cache_t *cachep, void *objp)
2871 struct array_cache *ac = cpu_cache_get(cachep);
2874 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
2876 /* Make sure we are not freeing a object from another
2877 * node to the array cache on this cpu.
2882 slabp = virt_to_slab(objp);
2883 if (unlikely(slabp->nodeid != numa_node_id())) {
2884 struct array_cache *alien = NULL;
2885 int nodeid = slabp->nodeid;
2886 struct kmem_list3 *l3 =
2887 cachep->nodelists[numa_node_id()];
2889 STATS_INC_NODEFREES(cachep);
2890 if (l3->alien && l3->alien[nodeid]) {
2891 alien = l3->alien[nodeid];
2892 spin_lock(&alien->lock);
2893 if (unlikely(alien->avail == alien->limit))
2894 __drain_alien_cache(cachep,
2896 alien->entry[alien->avail++] = objp;
2897 spin_unlock(&alien->lock);
2899 spin_lock(&(cachep->nodelists[nodeid])->
2901 free_block(cachep, &objp, 1, nodeid);
2902 spin_unlock(&(cachep->nodelists[nodeid])->
2909 if (likely(ac->avail < ac->limit)) {
2910 STATS_INC_FREEHIT(cachep);
2911 ac->entry[ac->avail++] = objp;
2914 STATS_INC_FREEMISS(cachep);
2915 cache_flusharray(cachep, ac);
2916 ac->entry[ac->avail++] = objp;
2921 * kmem_cache_alloc - Allocate an object
2922 * @cachep: The cache to allocate from.
2923 * @flags: See kmalloc().
2925 * Allocate an object from this cache. The flags are only relevant
2926 * if the cache has no available objects.
2928 void *kmem_cache_alloc(kmem_cache_t *cachep, gfp_t flags)
2930 return __cache_alloc(cachep, flags);
2932 EXPORT_SYMBOL(kmem_cache_alloc);
2935 * kmem_ptr_validate - check if an untrusted pointer might
2937 * @cachep: the cache we're checking against
2938 * @ptr: pointer to validate
2940 * This verifies that the untrusted pointer looks sane:
2941 * it is _not_ a guarantee that the pointer is actually
2942 * part of the slab cache in question, but it at least
2943 * validates that the pointer can be dereferenced and
2944 * looks half-way sane.
2946 * Currently only used for dentry validation.
2948 int fastcall kmem_ptr_validate(kmem_cache_t *cachep, void *ptr)
2950 unsigned long addr = (unsigned long)ptr;
2951 unsigned long min_addr = PAGE_OFFSET;
2952 unsigned long align_mask = BYTES_PER_WORD - 1;
2953 unsigned long size = cachep->buffer_size;
2956 if (unlikely(addr < min_addr))
2958 if (unlikely(addr > (unsigned long)high_memory - size))
2960 if (unlikely(addr & align_mask))
2962 if (unlikely(!kern_addr_valid(addr)))
2964 if (unlikely(!kern_addr_valid(addr + size - 1)))
2966 page = virt_to_page(ptr);
2967 if (unlikely(!PageSlab(page)))
2969 if (unlikely(page_get_cache(page) != cachep))
2978 * kmem_cache_alloc_node - Allocate an object on the specified node
2979 * @cachep: The cache to allocate from.
2980 * @flags: See kmalloc().
2981 * @nodeid: node number of the target node.
2983 * Identical to kmem_cache_alloc, except that this function is slow
2984 * and can sleep. And it will allocate memory on the given node, which
2985 * can improve the performance for cpu bound structures.
2986 * New and improved: it will now make sure that the object gets
2987 * put on the correct node list so that there is no false sharing.
2989 void *kmem_cache_alloc_node(kmem_cache_t *cachep, gfp_t flags, int nodeid)
2991 unsigned long save_flags;
2994 cache_alloc_debugcheck_before(cachep, flags);
2995 local_irq_save(save_flags);
2997 if (nodeid == -1 || nodeid == numa_node_id() ||
2998 !cachep->nodelists[nodeid])
2999 ptr = ____cache_alloc(cachep, flags);
3001 ptr = __cache_alloc_node(cachep, flags, nodeid);
3002 local_irq_restore(save_flags);
3004 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr,
3005 __builtin_return_address(0));
3009 EXPORT_SYMBOL(kmem_cache_alloc_node);
3011 void *kmalloc_node(size_t size, gfp_t flags, int node)
3013 kmem_cache_t *cachep;
3015 cachep = kmem_find_general_cachep(size, flags);
3016 if (unlikely(cachep == NULL))
3018 return kmem_cache_alloc_node(cachep, flags, node);
3020 EXPORT_SYMBOL(kmalloc_node);
3024 * kmalloc - allocate memory
3025 * @size: how many bytes of memory are required.
3026 * @flags: the type of memory to allocate.
3028 * kmalloc is the normal method of allocating memory
3031 * The @flags argument may be one of:
3033 * %GFP_USER - Allocate memory on behalf of user. May sleep.
3035 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
3037 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
3039 * Additionally, the %GFP_DMA flag may be set to indicate the memory
3040 * must be suitable for DMA. This can mean different things on different
3041 * platforms. For example, on i386, it means that the memory must come
3042 * from the first 16MB.
3044 void *__kmalloc(size_t size, gfp_t flags)
3046 kmem_cache_t *cachep;
3048 /* If you want to save a few bytes .text space: replace
3050 * Then kmalloc uses the uninlined functions instead of the inline
3053 cachep = __find_general_cachep(size, flags);
3054 if (unlikely(cachep == NULL))
3056 return __cache_alloc(cachep, flags);
3058 EXPORT_SYMBOL(__kmalloc);
3062 * __alloc_percpu - allocate one copy of the object for every present
3063 * cpu in the system, zeroing them.
3064 * Objects should be dereferenced using the per_cpu_ptr macro only.
3066 * @size: how many bytes of memory are required.
3068 void *__alloc_percpu(size_t size)
3071 struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL);
3077 * Cannot use for_each_online_cpu since a cpu may come online
3078 * and we have no way of figuring out how to fix the array
3079 * that we have allocated then....
3082 int node = cpu_to_node(i);
3084 if (node_online(node))
3085 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
3087 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
3089 if (!pdata->ptrs[i])
3091 memset(pdata->ptrs[i], 0, size);
3094 /* Catch derefs w/o wrappers */
3095 return (void *)(~(unsigned long)pdata);
3099 if (!cpu_possible(i))
3101 kfree(pdata->ptrs[i]);
3106 EXPORT_SYMBOL(__alloc_percpu);
3110 * kmem_cache_free - Deallocate an object
3111 * @cachep: The cache the allocation was from.
3112 * @objp: The previously allocated object.
3114 * Free an object which was previously allocated from this
3117 void kmem_cache_free(kmem_cache_t *cachep, void *objp)
3119 unsigned long flags;
3121 local_irq_save(flags);
3122 __cache_free(cachep, objp);
3123 local_irq_restore(flags);
3125 EXPORT_SYMBOL(kmem_cache_free);
3128 * kfree - free previously allocated memory
3129 * @objp: pointer returned by kmalloc.
3131 * If @objp is NULL, no operation is performed.
3133 * Don't free memory not originally allocated by kmalloc()
3134 * or you will run into trouble.
3136 void kfree(const void *objp)
3139 unsigned long flags;
3141 if (unlikely(!objp))
3143 local_irq_save(flags);
3144 kfree_debugcheck(objp);
3145 c = virt_to_cache(objp);
3146 mutex_debug_check_no_locks_freed(objp, obj_size(c));
3147 __cache_free(c, (void *)objp);
3148 local_irq_restore(flags);
3150 EXPORT_SYMBOL(kfree);
3154 * free_percpu - free previously allocated percpu memory
3155 * @objp: pointer returned by alloc_percpu.
3157 * Don't free memory not originally allocated by alloc_percpu()
3158 * The complemented objp is to check for that.
3160 void free_percpu(const void *objp)
3163 struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp);
3166 * We allocate for all cpus so we cannot use for online cpu here.
3172 EXPORT_SYMBOL(free_percpu);
3175 unsigned int kmem_cache_size(kmem_cache_t *cachep)
3177 return obj_size(cachep);
3179 EXPORT_SYMBOL(kmem_cache_size);
3181 const char *kmem_cache_name(kmem_cache_t *cachep)
3183 return cachep->name;
3185 EXPORT_SYMBOL_GPL(kmem_cache_name);
3188 * This initializes kmem_list3 for all nodes.
3190 static int alloc_kmemlist(kmem_cache_t *cachep)
3193 struct kmem_list3 *l3;
3196 for_each_online_node(node) {
3197 struct array_cache *nc = NULL, *new;
3198 struct array_cache **new_alien = NULL;
3200 if (!(new_alien = alloc_alien_cache(node, cachep->limit)))
3203 if (!(new = alloc_arraycache(node, (cachep->shared *
3204 cachep->batchcount),
3207 if ((l3 = cachep->nodelists[node])) {
3209 spin_lock_irq(&l3->list_lock);
3211 if ((nc = cachep->nodelists[node]->shared))
3212 free_block(cachep, nc->entry, nc->avail, node);
3215 if (!cachep->nodelists[node]->alien) {
3216 l3->alien = new_alien;
3219 l3->free_limit = (1 + nr_cpus_node(node)) *
3220 cachep->batchcount + cachep->num;
3221 spin_unlock_irq(&l3->list_lock);
3223 free_alien_cache(new_alien);
3226 if (!(l3 = kmalloc_node(sizeof(struct kmem_list3),
3230 kmem_list3_init(l3);
3231 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3232 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3234 l3->alien = new_alien;
3235 l3->free_limit = (1 + nr_cpus_node(node)) *
3236 cachep->batchcount + cachep->num;
3237 cachep->nodelists[node] = l3;
3245 struct ccupdate_struct {
3246 kmem_cache_t *cachep;
3247 struct array_cache *new[NR_CPUS];
3250 static void do_ccupdate_local(void *info)
3252 struct ccupdate_struct *new = (struct ccupdate_struct *)info;
3253 struct array_cache *old;
3256 old = cpu_cache_get(new->cachep);
3258 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3259 new->new[smp_processor_id()] = old;
3262 static int do_tune_cpucache(kmem_cache_t *cachep, int limit, int batchcount,
3265 struct ccupdate_struct new;
3268 memset(&new.new, 0, sizeof(new.new));
3269 for_each_online_cpu(i) {
3271 alloc_arraycache(cpu_to_node(i), limit, batchcount);
3273 for (i--; i >= 0; i--)
3278 new.cachep = cachep;
3280 smp_call_function_all_cpus(do_ccupdate_local, (void *)&new);
3283 spin_lock_irq(&cachep->spinlock);
3284 cachep->batchcount = batchcount;
3285 cachep->limit = limit;
3286 cachep->shared = shared;
3287 spin_unlock_irq(&cachep->spinlock);
3289 for_each_online_cpu(i) {
3290 struct array_cache *ccold = new.new[i];
3293 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3294 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3295 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3299 err = alloc_kmemlist(cachep);
3301 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
3302 cachep->name, -err);
3308 static void enable_cpucache(kmem_cache_t *cachep)
3313 /* The head array serves three purposes:
3314 * - create a LIFO ordering, i.e. return objects that are cache-warm
3315 * - reduce the number of spinlock operations.
3316 * - reduce the number of linked list operations on the slab and
3317 * bufctl chains: array operations are cheaper.
3318 * The numbers are guessed, we should auto-tune as described by
3321 if (cachep->buffer_size > 131072)
3323 else if (cachep->buffer_size > PAGE_SIZE)
3325 else if (cachep->buffer_size > 1024)
3327 else if (cachep->buffer_size > 256)
3332 /* Cpu bound tasks (e.g. network routing) can exhibit cpu bound
3333 * allocation behaviour: Most allocs on one cpu, most free operations
3334 * on another cpu. For these cases, an efficient object passing between
3335 * cpus is necessary. This is provided by a shared array. The array
3336 * replaces Bonwick's magazine layer.
3337 * On uniprocessor, it's functionally equivalent (but less efficient)
3338 * to a larger limit. Thus disabled by default.
3342 if (cachep->buffer_size <= PAGE_SIZE)
3347 /* With debugging enabled, large batchcount lead to excessively
3348 * long periods with disabled local interrupts. Limit the
3354 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
3356 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3357 cachep->name, -err);
3360 static void drain_array_locked(kmem_cache_t *cachep, struct array_cache *ac,
3361 int force, int node)
3365 check_spinlock_acquired_node(cachep, node);
3366 if (ac->touched && !force) {
3368 } else if (ac->avail) {
3369 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3370 if (tofree > ac->avail) {
3371 tofree = (ac->avail + 1) / 2;
3373 free_block(cachep, ac->entry, tofree, node);
3374 ac->avail -= tofree;
3375 memmove(ac->entry, &(ac->entry[tofree]),
3376 sizeof(void *) * ac->avail);
3381 * cache_reap - Reclaim memory from caches.
3382 * @unused: unused parameter
3384 * Called from workqueue/eventd every few seconds.
3386 * - clear the per-cpu caches for this CPU.
3387 * - return freeable pages to the main free memory pool.
3389 * If we cannot acquire the cache chain mutex then just give up - we'll
3390 * try again on the next iteration.
3392 static void cache_reap(void *unused)
3394 struct list_head *walk;
3395 struct kmem_list3 *l3;
3397 if (!mutex_trylock(&cache_chain_mutex)) {
3398 /* Give up. Setup the next iteration. */
3399 schedule_delayed_work(&__get_cpu_var(reap_work),
3404 list_for_each(walk, &cache_chain) {
3405 kmem_cache_t *searchp;
3406 struct list_head *p;
3410 searchp = list_entry(walk, kmem_cache_t, next);
3412 if (searchp->flags & SLAB_NO_REAP)
3417 l3 = searchp->nodelists[numa_node_id()];
3419 drain_alien_cache(searchp, l3);
3420 spin_lock_irq(&l3->list_lock);
3422 drain_array_locked(searchp, cpu_cache_get(searchp), 0,
3425 if (time_after(l3->next_reap, jiffies))
3428 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
3431 drain_array_locked(searchp, l3->shared, 0,
3434 if (l3->free_touched) {
3435 l3->free_touched = 0;
3440 (l3->free_limit + 5 * searchp->num -
3441 1) / (5 * searchp->num);
3443 p = l3->slabs_free.next;
3444 if (p == &(l3->slabs_free))
3447 slabp = list_entry(p, struct slab, list);
3448 BUG_ON(slabp->inuse);
3449 list_del(&slabp->list);
3450 STATS_INC_REAPED(searchp);
3452 /* Safe to drop the lock. The slab is no longer
3453 * linked to the cache.
3454 * searchp cannot disappear, we hold
3457 l3->free_objects -= searchp->num;
3458 spin_unlock_irq(&l3->list_lock);
3459 slab_destroy(searchp, slabp);
3460 spin_lock_irq(&l3->list_lock);
3461 } while (--tofree > 0);
3463 spin_unlock_irq(&l3->list_lock);
3468 mutex_unlock(&cache_chain_mutex);
3469 drain_remote_pages();
3470 /* Setup the next iteration */
3471 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
3474 #ifdef CONFIG_PROC_FS
3476 static void print_slabinfo_header(struct seq_file *m)
3479 * Output format version, so at least we can change it
3480 * without _too_ many complaints.
3483 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3485 seq_puts(m, "slabinfo - version: 2.1\n");
3487 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3488 "<objperslab> <pagesperslab>");
3489 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3490 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3492 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3493 "<error> <maxfreeable> <nodeallocs> <remotefrees>");
3494 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3499 static void *s_start(struct seq_file *m, loff_t *pos)
3502 struct list_head *p;
3504 mutex_lock(&cache_chain_mutex);
3506 print_slabinfo_header(m);
3507 p = cache_chain.next;
3510 if (p == &cache_chain)
3513 return list_entry(p, kmem_cache_t, next);
3516 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3518 kmem_cache_t *cachep = p;
3520 return cachep->next.next == &cache_chain ? NULL
3521 : list_entry(cachep->next.next, kmem_cache_t, next);
3524 static void s_stop(struct seq_file *m, void *p)
3526 mutex_unlock(&cache_chain_mutex);
3529 static int s_show(struct seq_file *m, void *p)
3531 kmem_cache_t *cachep = p;
3532 struct list_head *q;
3534 unsigned long active_objs;
3535 unsigned long num_objs;
3536 unsigned long active_slabs = 0;
3537 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3541 struct kmem_list3 *l3;
3544 spin_lock_irq(&cachep->spinlock);
3547 for_each_online_node(node) {
3548 l3 = cachep->nodelists[node];
3552 spin_lock(&l3->list_lock);
3554 list_for_each(q, &l3->slabs_full) {
3555 slabp = list_entry(q, struct slab, list);
3556 if (slabp->inuse != cachep->num && !error)
3557 error = "slabs_full accounting error";
3558 active_objs += cachep->num;
3561 list_for_each(q, &l3->slabs_partial) {
3562 slabp = list_entry(q, struct slab, list);
3563 if (slabp->inuse == cachep->num && !error)
3564 error = "slabs_partial inuse accounting error";
3565 if (!slabp->inuse && !error)
3566 error = "slabs_partial/inuse accounting error";
3567 active_objs += slabp->inuse;
3570 list_for_each(q, &l3->slabs_free) {
3571 slabp = list_entry(q, struct slab, list);
3572 if (slabp->inuse && !error)
3573 error = "slabs_free/inuse accounting error";
3576 free_objects += l3->free_objects;
3577 shared_avail += l3->shared->avail;
3579 spin_unlock(&l3->list_lock);
3581 num_slabs += active_slabs;
3582 num_objs = num_slabs * cachep->num;
3583 if (num_objs - active_objs != free_objects && !error)
3584 error = "free_objects accounting error";
3586 name = cachep->name;
3588 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3590 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
3591 name, active_objs, num_objs, cachep->buffer_size,
3592 cachep->num, (1 << cachep->gfporder));
3593 seq_printf(m, " : tunables %4u %4u %4u",
3594 cachep->limit, cachep->batchcount, cachep->shared);
3595 seq_printf(m, " : slabdata %6lu %6lu %6lu",
3596 active_slabs, num_slabs, shared_avail);
3599 unsigned long high = cachep->high_mark;
3600 unsigned long allocs = cachep->num_allocations;
3601 unsigned long grown = cachep->grown;
3602 unsigned long reaped = cachep->reaped;
3603 unsigned long errors = cachep->errors;
3604 unsigned long max_freeable = cachep->max_freeable;
3605 unsigned long node_allocs = cachep->node_allocs;
3606 unsigned long node_frees = cachep->node_frees;
3608 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
3609 %4lu %4lu %4lu %4lu", allocs, high, grown, reaped, errors, max_freeable, node_allocs, node_frees);
3613 unsigned long allochit = atomic_read(&cachep->allochit);
3614 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3615 unsigned long freehit = atomic_read(&cachep->freehit);
3616 unsigned long freemiss = atomic_read(&cachep->freemiss);
3618 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
3619 allochit, allocmiss, freehit, freemiss);
3623 spin_unlock_irq(&cachep->spinlock);
3628 * slabinfo_op - iterator that generates /proc/slabinfo
3637 * num-pages-per-slab
3638 * + further values on SMP and with statistics enabled
3641 struct seq_operations slabinfo_op = {
3648 #define MAX_SLABINFO_WRITE 128
3650 * slabinfo_write - Tuning for the slab allocator
3652 * @buffer: user buffer
3653 * @count: data length
3656 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
3657 size_t count, loff_t *ppos)
3659 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
3660 int limit, batchcount, shared, res;
3661 struct list_head *p;
3663 if (count > MAX_SLABINFO_WRITE)
3665 if (copy_from_user(&kbuf, buffer, count))
3667 kbuf[MAX_SLABINFO_WRITE] = '\0';
3669 tmp = strchr(kbuf, ' ');
3674 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3677 /* Find the cache in the chain of caches. */
3678 mutex_lock(&cache_chain_mutex);
3680 list_for_each(p, &cache_chain) {
3681 kmem_cache_t *cachep = list_entry(p, kmem_cache_t, next);
3683 if (!strcmp(cachep->name, kbuf)) {
3686 batchcount > limit || shared < 0) {
3689 res = do_tune_cpucache(cachep, limit,
3690 batchcount, shared);
3695 mutex_unlock(&cache_chain_mutex);
3703 * ksize - get the actual amount of memory allocated for a given object
3704 * @objp: Pointer to the object
3706 * kmalloc may internally round up allocations and return more memory
3707 * than requested. ksize() can be used to determine the actual amount of
3708 * memory allocated. The caller may use this additional memory, even though
3709 * a smaller amount of memory was initially specified with the kmalloc call.
3710 * The caller must guarantee that objp points to a valid object previously
3711 * allocated with either kmalloc() or kmem_cache_alloc(). The object
3712 * must not be freed during the duration of the call.
3714 unsigned int ksize(const void *objp)
3716 if (unlikely(objp == NULL))
3719 return obj_size(virt_to_cache(objp));