3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same intializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/config.h>
90 #include <linux/slab.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/cpuset.h>
98 #include <linux/seq_file.h>
99 #include <linux/notifier.h>
100 #include <linux/kallsyms.h>
101 #include <linux/cpu.h>
102 #include <linux/sysctl.h>
103 #include <linux/module.h>
104 #include <linux/rcupdate.h>
105 #include <linux/string.h>
106 #include <linux/nodemask.h>
107 #include <linux/mempolicy.h>
108 #include <linux/mutex.h>
110 #include <asm/uaccess.h>
111 #include <asm/cacheflush.h>
112 #include <asm/tlbflush.h>
113 #include <asm/page.h>
116 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
117 * SLAB_RED_ZONE & SLAB_POISON.
118 * 0 for faster, smaller code (especially in the critical paths).
120 * STATS - 1 to collect stats for /proc/slabinfo.
121 * 0 for faster, smaller code (especially in the critical paths).
123 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
126 #ifdef CONFIG_DEBUG_SLAB
129 #define FORCED_DEBUG 1
133 #define FORCED_DEBUG 0
136 /* Shouldn't this be in a header file somewhere? */
137 #define BYTES_PER_WORD sizeof(void *)
139 #ifndef cache_line_size
140 #define cache_line_size() L1_CACHE_BYTES
143 #ifndef ARCH_KMALLOC_MINALIGN
145 * Enforce a minimum alignment for the kmalloc caches.
146 * Usually, the kmalloc caches are cache_line_size() aligned, except when
147 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
148 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
149 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
150 * Note that this flag disables some debug features.
152 #define ARCH_KMALLOC_MINALIGN 0
155 #ifndef ARCH_SLAB_MINALIGN
157 * Enforce a minimum alignment for all caches.
158 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
159 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
160 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
161 * some debug features.
163 #define ARCH_SLAB_MINALIGN 0
166 #ifndef ARCH_KMALLOC_FLAGS
167 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
170 /* Legal flag mask for kmem_cache_create(). */
172 # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
173 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
175 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
176 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
177 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
179 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
180 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
181 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
182 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
188 * Bufctl's are used for linking objs within a slab
191 * This implementation relies on "struct page" for locating the cache &
192 * slab an object belongs to.
193 * This allows the bufctl structure to be small (one int), but limits
194 * the number of objects a slab (not a cache) can contain when off-slab
195 * bufctls are used. The limit is the size of the largest general cache
196 * that does not use off-slab slabs.
197 * For 32bit archs with 4 kB pages, is this 56.
198 * This is not serious, as it is only for large objects, when it is unwise
199 * to have too many per slab.
200 * Note: This limit can be raised by introducing a general cache whose size
201 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
204 typedef unsigned int kmem_bufctl_t;
205 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
206 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
207 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
208 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
213 * Manages the objs in a slab. Placed either at the beginning of mem allocated
214 * for a slab, or allocated from an general cache.
215 * Slabs are chained into three list: fully used, partial, fully free slabs.
218 struct list_head list;
219 unsigned long colouroff;
220 void *s_mem; /* including colour offset */
221 unsigned int inuse; /* num of objs active in slab */
223 unsigned short nodeid;
229 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
230 * arrange for kmem_freepages to be called via RCU. This is useful if
231 * we need to approach a kernel structure obliquely, from its address
232 * obtained without the usual locking. We can lock the structure to
233 * stabilize it and check it's still at the given address, only if we
234 * can be sure that the memory has not been meanwhile reused for some
235 * other kind of object (which our subsystem's lock might corrupt).
237 * rcu_read_lock before reading the address, then rcu_read_unlock after
238 * taking the spinlock within the structure expected at that address.
240 * We assume struct slab_rcu can overlay struct slab when destroying.
243 struct rcu_head head;
244 struct kmem_cache *cachep;
252 * - LIFO ordering, to hand out cache-warm objects from _alloc
253 * - reduce the number of linked list operations
254 * - reduce spinlock operations
256 * The limit is stored in the per-cpu structure to reduce the data cache
263 unsigned int batchcount;
264 unsigned int touched;
267 * Must have this definition in here for the proper
268 * alignment of array_cache. Also simplifies accessing
270 * [0] is for gcc 2.95. It should really be [].
275 * bootstrap: The caches do not work without cpuarrays anymore, but the
276 * cpuarrays are allocated from the generic caches...
278 #define BOOT_CPUCACHE_ENTRIES 1
279 struct arraycache_init {
280 struct array_cache cache;
281 void *entries[BOOT_CPUCACHE_ENTRIES];
285 * The slab lists for all objects.
288 struct list_head slabs_partial; /* partial list first, better asm code */
289 struct list_head slabs_full;
290 struct list_head slabs_free;
291 unsigned long free_objects;
292 unsigned int free_limit;
293 unsigned int colour_next; /* Per-node cache coloring */
294 spinlock_t list_lock;
295 struct array_cache *shared; /* shared per node */
296 struct array_cache **alien; /* on other nodes */
297 unsigned long next_reap; /* updated without locking */
298 int free_touched; /* updated without locking */
302 * Need this for bootstrapping a per node allocator.
304 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
305 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
306 #define CACHE_CACHE 0
308 #define SIZE_L3 (1 + MAX_NUMNODES)
311 * This function must be completely optimized away if a constant is passed to
312 * it. Mostly the same as what is in linux/slab.h except it returns an index.
314 static __always_inline int index_of(const size_t size)
316 extern void __bad_size(void);
318 if (__builtin_constant_p(size)) {
326 #include "linux/kmalloc_sizes.h"
334 #define INDEX_AC index_of(sizeof(struct arraycache_init))
335 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
337 static void kmem_list3_init(struct kmem_list3 *parent)
339 INIT_LIST_HEAD(&parent->slabs_full);
340 INIT_LIST_HEAD(&parent->slabs_partial);
341 INIT_LIST_HEAD(&parent->slabs_free);
342 parent->shared = NULL;
343 parent->alien = NULL;
344 parent->colour_next = 0;
345 spin_lock_init(&parent->list_lock);
346 parent->free_objects = 0;
347 parent->free_touched = 0;
350 #define MAKE_LIST(cachep, listp, slab, nodeid) \
352 INIT_LIST_HEAD(listp); \
353 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
356 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
358 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
359 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
360 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
370 /* 1) per-cpu data, touched during every alloc/free */
371 struct array_cache *array[NR_CPUS];
372 /* 2) Cache tunables. Protected by cache_chain_mutex */
373 unsigned int batchcount;
377 unsigned int buffer_size;
378 /* 3) touched by every alloc & free from the backend */
379 struct kmem_list3 *nodelists[MAX_NUMNODES];
381 unsigned int flags; /* constant flags */
382 unsigned int num; /* # of objs per slab */
384 /* 4) 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 struct kmem_cache *slabp_cache;
394 unsigned int slab_size;
395 unsigned int dflags; /* dynamic flags */
397 /* constructor func */
398 void (*ctor) (void *, struct kmem_cache *, unsigned long);
400 /* de-constructor func */
401 void (*dtor) (void *, struct kmem_cache *, unsigned long);
403 /* 5) cache creation/removal */
405 struct list_head next;
409 unsigned long num_active;
410 unsigned long num_allocations;
411 unsigned long high_mark;
413 unsigned long reaped;
414 unsigned long errors;
415 unsigned long max_freeable;
416 unsigned long node_allocs;
417 unsigned long node_frees;
418 unsigned long node_overflow;
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
441 * Optimization question: fewer reaps means less probability for unnessary
442 * 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) \
458 if ((x)->num_active > (x)->high_mark) \
459 (x)->high_mark = (x)->num_active; \
461 #define STATS_INC_ERR(x) ((x)->errors++)
462 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
463 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
464 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
465 #define STATS_SET_FREEABLE(x, i) \
467 if ((x)->max_freeable < i) \
468 (x)->max_freeable = i; \
470 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
471 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
472 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
473 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
475 #define STATS_INC_ACTIVE(x) do { } while (0)
476 #define STATS_DEC_ACTIVE(x) do { } while (0)
477 #define STATS_INC_ALLOCED(x) do { } while (0)
478 #define STATS_INC_GROWN(x) do { } while (0)
479 #define STATS_INC_REAPED(x) do { } while (0)
480 #define STATS_SET_HIGH(x) do { } while (0)
481 #define STATS_INC_ERR(x) do { } while (0)
482 #define STATS_INC_NODEALLOCS(x) do { } while (0)
483 #define STATS_INC_NODEFREES(x) do { } while (0)
484 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
485 #define STATS_SET_FREEABLE(x, i) do { } while (0)
486 #define STATS_INC_ALLOCHIT(x) do { } while (0)
487 #define STATS_INC_ALLOCMISS(x) do { } while (0)
488 #define STATS_INC_FREEHIT(x) do { } while (0)
489 #define STATS_INC_FREEMISS(x) do { } while (0)
494 * Magic nums for obj red zoning.
495 * Placed in the first word before and the first word after an obj.
497 #define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
498 #define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
500 /* ...and for poisoning */
501 #define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
502 #define POISON_FREE 0x6b /* for use-after-free poisoning */
503 #define POISON_END 0xa5 /* end-byte of poisoning */
506 * memory layout of objects:
508 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
509 * the end of an object is aligned with the end of the real
510 * allocation. Catches writes behind the end of the allocation.
511 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
513 * cachep->obj_offset: The real object.
514 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
515 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
516 * [BYTES_PER_WORD long]
518 static int obj_offset(struct kmem_cache *cachep)
520 return cachep->obj_offset;
523 static int obj_size(struct kmem_cache *cachep)
525 return cachep->obj_size;
528 static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
530 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
531 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD);
534 static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
536 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
537 if (cachep->flags & SLAB_STORE_USER)
538 return (unsigned long *)(objp + cachep->buffer_size -
540 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD);
543 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
545 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
546 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
551 #define obj_offset(x) 0
552 #define obj_size(cachep) (cachep->buffer_size)
553 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
554 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
555 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
560 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
563 #if defined(CONFIG_LARGE_ALLOCS)
564 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
565 #define MAX_GFP_ORDER 13 /* up to 32Mb */
566 #elif defined(CONFIG_MMU)
567 #define MAX_OBJ_ORDER 5 /* 32 pages */
568 #define MAX_GFP_ORDER 5 /* 32 pages */
570 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
571 #define MAX_GFP_ORDER 8 /* up to 1Mb */
575 * Do not go above this order unless 0 objects fit into the slab.
577 #define BREAK_GFP_ORDER_HI 1
578 #define BREAK_GFP_ORDER_LO 0
579 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
582 * Functions for storing/retrieving the cachep and or slab from the page
583 * allocator. These are used to find the slab an obj belongs to. With kfree(),
584 * these are used to find the cache which an obj belongs to.
586 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
588 page->lru.next = (struct list_head *)cache;
591 static inline struct kmem_cache *page_get_cache(struct page *page)
593 if (unlikely(PageCompound(page)))
594 page = (struct page *)page_private(page);
595 return (struct kmem_cache *)page->lru.next;
598 static inline void page_set_slab(struct page *page, struct slab *slab)
600 page->lru.prev = (struct list_head *)slab;
603 static inline struct slab *page_get_slab(struct page *page)
605 if (unlikely(PageCompound(page)))
606 page = (struct page *)page_private(page);
607 return (struct slab *)page->lru.prev;
610 static inline struct kmem_cache *virt_to_cache(const void *obj)
612 struct page *page = virt_to_page(obj);
613 return page_get_cache(page);
616 static inline struct slab *virt_to_slab(const void *obj)
618 struct page *page = virt_to_page(obj);
619 return page_get_slab(page);
622 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
625 return slab->s_mem + cache->buffer_size * idx;
628 static inline unsigned int obj_to_index(struct kmem_cache *cache,
629 struct slab *slab, void *obj)
631 return (unsigned)(obj - slab->s_mem) / cache->buffer_size;
635 * These are the default caches for kmalloc. Custom caches can have other sizes.
637 struct cache_sizes malloc_sizes[] = {
638 #define CACHE(x) { .cs_size = (x) },
639 #include <linux/kmalloc_sizes.h>
643 EXPORT_SYMBOL(malloc_sizes);
645 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
651 static struct cache_names __initdata cache_names[] = {
652 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
653 #include <linux/kmalloc_sizes.h>
658 static struct arraycache_init initarray_cache __initdata =
659 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
660 static struct arraycache_init initarray_generic =
661 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
663 /* internal cache of cache description objs */
664 static struct kmem_cache cache_cache = {
666 .limit = BOOT_CPUCACHE_ENTRIES,
668 .buffer_size = sizeof(struct kmem_cache),
669 .name = "kmem_cache",
671 .obj_size = sizeof(struct kmem_cache),
675 /* Guard access to the cache-chain. */
676 static DEFINE_MUTEX(cache_chain_mutex);
677 static struct list_head cache_chain;
680 * vm_enough_memory() looks at this to determine how many slab-allocated pages
681 * are possibly freeable under pressure
683 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
685 atomic_t slab_reclaim_pages;
688 * chicken and egg problem: delay the per-cpu array allocation
689 * until the general caches are up.
699 * used by boot code to determine if it can use slab based allocator
701 int slab_is_available(void)
703 return g_cpucache_up == FULL;
706 static DEFINE_PER_CPU(struct work_struct, reap_work);
708 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
710 static void enable_cpucache(struct kmem_cache *cachep);
711 static void cache_reap(void *unused);
712 static int __node_shrink(struct kmem_cache *cachep, int node);
714 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
716 return cachep->array[smp_processor_id()];
719 static inline struct kmem_cache *__find_general_cachep(size_t size,
722 struct cache_sizes *csizep = malloc_sizes;
725 /* This happens if someone tries to call
726 * kmem_cache_create(), or __kmalloc(), before
727 * the generic caches are initialized.
729 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
731 while (size > csizep->cs_size)
735 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
736 * has cs_{dma,}cachep==NULL. Thus no special case
737 * for large kmalloc calls required.
739 if (unlikely(gfpflags & GFP_DMA))
740 return csizep->cs_dmacachep;
741 return csizep->cs_cachep;
744 struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
746 return __find_general_cachep(size, gfpflags);
748 EXPORT_SYMBOL(kmem_find_general_cachep);
750 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
752 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
756 * Calculate the number of objects and left-over bytes for a given buffer size.
758 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
759 size_t align, int flags, size_t *left_over,
764 size_t slab_size = PAGE_SIZE << gfporder;
767 * The slab management structure can be either off the slab or
768 * on it. For the latter case, the memory allocated for a
772 * - One kmem_bufctl_t for each object
773 * - Padding to respect alignment of @align
774 * - @buffer_size bytes for each object
776 * If the slab management structure is off the slab, then the
777 * alignment will already be calculated into the size. Because
778 * the slabs are all pages aligned, the objects will be at the
779 * correct alignment when allocated.
781 if (flags & CFLGS_OFF_SLAB) {
783 nr_objs = slab_size / buffer_size;
785 if (nr_objs > SLAB_LIMIT)
786 nr_objs = SLAB_LIMIT;
789 * Ignore padding for the initial guess. The padding
790 * is at most @align-1 bytes, and @buffer_size is at
791 * least @align. In the worst case, this result will
792 * be one greater than the number of objects that fit
793 * into the memory allocation when taking the padding
796 nr_objs = (slab_size - sizeof(struct slab)) /
797 (buffer_size + sizeof(kmem_bufctl_t));
800 * This calculated number will be either the right
801 * amount, or one greater than what we want.
803 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
807 if (nr_objs > SLAB_LIMIT)
808 nr_objs = SLAB_LIMIT;
810 mgmt_size = slab_mgmt_size(nr_objs, align);
813 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
816 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
818 static void __slab_error(const char *function, struct kmem_cache *cachep,
821 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
822 function, cachep->name, msg);
828 * Special reaping functions for NUMA systems called from cache_reap().
829 * These take care of doing round robin flushing of alien caches (containing
830 * objects freed on different nodes from which they were allocated) and the
831 * flushing of remote pcps by calling drain_node_pages.
833 static DEFINE_PER_CPU(unsigned long, reap_node);
835 static void init_reap_node(int cpu)
839 node = next_node(cpu_to_node(cpu), node_online_map);
840 if (node == MAX_NUMNODES)
841 node = first_node(node_online_map);
843 __get_cpu_var(reap_node) = node;
846 static void next_reap_node(void)
848 int node = __get_cpu_var(reap_node);
851 * Also drain per cpu pages on remote zones
853 if (node != numa_node_id())
854 drain_node_pages(node);
856 node = next_node(node, node_online_map);
857 if (unlikely(node >= MAX_NUMNODES))
858 node = first_node(node_online_map);
859 __get_cpu_var(reap_node) = node;
863 #define init_reap_node(cpu) do { } while (0)
864 #define next_reap_node(void) do { } while (0)
868 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
869 * via the workqueue/eventd.
870 * Add the CPU number into the expiration time to minimize the possibility of
871 * the CPUs getting into lockstep and contending for the global cache chain
874 static void __devinit start_cpu_timer(int cpu)
876 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
879 * When this gets called from do_initcalls via cpucache_init(),
880 * init_workqueues() has already run, so keventd will be setup
883 if (keventd_up() && reap_work->func == NULL) {
885 INIT_WORK(reap_work, cache_reap, NULL);
886 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
890 static struct array_cache *alloc_arraycache(int node, int entries,
893 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
894 struct array_cache *nc = NULL;
896 nc = kmalloc_node(memsize, GFP_KERNEL, node);
900 nc->batchcount = batchcount;
902 spin_lock_init(&nc->lock);
908 * Transfer objects in one arraycache to another.
909 * Locking must be handled by the caller.
911 * Return the number of entries transferred.
913 static int transfer_objects(struct array_cache *to,
914 struct array_cache *from, unsigned int max)
916 /* Figure out how many entries to transfer */
917 int nr = min(min(from->avail, max), to->limit - to->avail);
922 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
932 static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int);
933 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
935 static struct array_cache **alloc_alien_cache(int node, int limit)
937 struct array_cache **ac_ptr;
938 int memsize = sizeof(void *) * MAX_NUMNODES;
943 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
946 if (i == node || !node_online(i)) {
950 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
952 for (i--; i <= 0; i--)
962 static void free_alien_cache(struct array_cache **ac_ptr)
973 static void __drain_alien_cache(struct kmem_cache *cachep,
974 struct array_cache *ac, int node)
976 struct kmem_list3 *rl3 = cachep->nodelists[node];
979 spin_lock(&rl3->list_lock);
981 * Stuff objects into the remote nodes shared array first.
982 * That way we could avoid the overhead of putting the objects
983 * into the free lists and getting them back later.
986 transfer_objects(rl3->shared, ac, ac->limit);
988 free_block(cachep, ac->entry, ac->avail, node);
990 spin_unlock(&rl3->list_lock);
995 * Called from cache_reap() to regularly drain alien caches round robin.
997 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
999 int node = __get_cpu_var(reap_node);
1002 struct array_cache *ac = l3->alien[node];
1004 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1005 __drain_alien_cache(cachep, ac, node);
1006 spin_unlock_irq(&ac->lock);
1011 static void drain_alien_cache(struct kmem_cache *cachep,
1012 struct array_cache **alien)
1015 struct array_cache *ac;
1016 unsigned long flags;
1018 for_each_online_node(i) {
1021 spin_lock_irqsave(&ac->lock, flags);
1022 __drain_alien_cache(cachep, ac, i);
1023 spin_unlock_irqrestore(&ac->lock, flags);
1028 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1030 struct slab *slabp = virt_to_slab(objp);
1031 int nodeid = slabp->nodeid;
1032 struct kmem_list3 *l3;
1033 struct array_cache *alien = NULL;
1036 * Make sure we are not freeing a object from another node to the array
1037 * cache on this cpu.
1039 if (likely(slabp->nodeid == numa_node_id()))
1042 l3 = cachep->nodelists[numa_node_id()];
1043 STATS_INC_NODEFREES(cachep);
1044 if (l3->alien && l3->alien[nodeid]) {
1045 alien = l3->alien[nodeid];
1046 spin_lock(&alien->lock);
1047 if (unlikely(alien->avail == alien->limit)) {
1048 STATS_INC_ACOVERFLOW(cachep);
1049 __drain_alien_cache(cachep, alien, nodeid);
1051 alien->entry[alien->avail++] = objp;
1052 spin_unlock(&alien->lock);
1054 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1055 free_block(cachep, &objp, 1, nodeid);
1056 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1063 #define drain_alien_cache(cachep, alien) do { } while (0)
1064 #define reap_alien(cachep, l3) do { } while (0)
1066 static inline struct array_cache **alloc_alien_cache(int node, int limit)
1068 return (struct array_cache **) 0x01020304ul;
1071 static inline void free_alien_cache(struct array_cache **ac_ptr)
1075 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1082 static int cpuup_callback(struct notifier_block *nfb,
1083 unsigned long action, void *hcpu)
1085 long cpu = (long)hcpu;
1086 struct kmem_cache *cachep;
1087 struct kmem_list3 *l3 = NULL;
1088 int node = cpu_to_node(cpu);
1089 int memsize = sizeof(struct kmem_list3);
1092 case CPU_UP_PREPARE:
1093 mutex_lock(&cache_chain_mutex);
1095 * We need to do this right in the beginning since
1096 * alloc_arraycache's are going to use this list.
1097 * kmalloc_node allows us to add the slab to the right
1098 * kmem_list3 and not this cpu's kmem_list3
1101 list_for_each_entry(cachep, &cache_chain, next) {
1103 * Set up the size64 kmemlist for cpu before we can
1104 * begin anything. Make sure some other cpu on this
1105 * node has not already allocated this
1107 if (!cachep->nodelists[node]) {
1108 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1111 kmem_list3_init(l3);
1112 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1113 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1116 * The l3s don't come and go as CPUs come and
1117 * go. cache_chain_mutex is sufficient
1120 cachep->nodelists[node] = l3;
1123 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1124 cachep->nodelists[node]->free_limit =
1125 (1 + nr_cpus_node(node)) *
1126 cachep->batchcount + cachep->num;
1127 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1131 * Now we can go ahead with allocating the shared arrays and
1134 list_for_each_entry(cachep, &cache_chain, next) {
1135 struct array_cache *nc;
1136 struct array_cache *shared;
1137 struct array_cache **alien;
1139 nc = alloc_arraycache(node, cachep->limit,
1140 cachep->batchcount);
1143 shared = alloc_arraycache(node,
1144 cachep->shared * cachep->batchcount,
1149 alien = alloc_alien_cache(node, cachep->limit);
1152 cachep->array[cpu] = nc;
1153 l3 = cachep->nodelists[node];
1156 spin_lock_irq(&l3->list_lock);
1159 * We are serialised from CPU_DEAD or
1160 * CPU_UP_CANCELLED by the cpucontrol lock
1162 l3->shared = shared;
1171 spin_unlock_irq(&l3->list_lock);
1173 free_alien_cache(alien);
1175 mutex_unlock(&cache_chain_mutex);
1178 start_cpu_timer(cpu);
1180 #ifdef CONFIG_HOTPLUG_CPU
1183 * Even if all the cpus of a node are down, we don't free the
1184 * kmem_list3 of any cache. This to avoid a race between
1185 * cpu_down, and a kmalloc allocation from another cpu for
1186 * memory from the node of the cpu going down. The list3
1187 * structure is usually allocated from kmem_cache_create() and
1188 * gets destroyed at kmem_cache_destroy().
1191 case CPU_UP_CANCELED:
1192 mutex_lock(&cache_chain_mutex);
1193 list_for_each_entry(cachep, &cache_chain, next) {
1194 struct array_cache *nc;
1195 struct array_cache *shared;
1196 struct array_cache **alien;
1199 mask = node_to_cpumask(node);
1200 /* cpu is dead; no one can alloc from it. */
1201 nc = cachep->array[cpu];
1202 cachep->array[cpu] = NULL;
1203 l3 = cachep->nodelists[node];
1206 goto free_array_cache;
1208 spin_lock_irq(&l3->list_lock);
1210 /* Free limit for this kmem_list3 */
1211 l3->free_limit -= cachep->batchcount;
1213 free_block(cachep, nc->entry, nc->avail, node);
1215 if (!cpus_empty(mask)) {
1216 spin_unlock_irq(&l3->list_lock);
1217 goto free_array_cache;
1220 shared = l3->shared;
1222 free_block(cachep, l3->shared->entry,
1223 l3->shared->avail, node);
1230 spin_unlock_irq(&l3->list_lock);
1234 drain_alien_cache(cachep, alien);
1235 free_alien_cache(alien);
1241 * In the previous loop, all the objects were freed to
1242 * the respective cache's slabs, now we can go ahead and
1243 * shrink each nodelist to its limit.
1245 list_for_each_entry(cachep, &cache_chain, next) {
1246 l3 = cachep->nodelists[node];
1249 spin_lock_irq(&l3->list_lock);
1250 /* free slabs belonging to this node */
1251 __node_shrink(cachep, node);
1252 spin_unlock_irq(&l3->list_lock);
1254 mutex_unlock(&cache_chain_mutex);
1260 mutex_unlock(&cache_chain_mutex);
1264 static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
1267 * swap the static kmem_list3 with kmalloced memory
1269 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1272 struct kmem_list3 *ptr;
1274 BUG_ON(cachep->nodelists[nodeid] != list);
1275 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1278 local_irq_disable();
1279 memcpy(ptr, list, sizeof(struct kmem_list3));
1280 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1281 cachep->nodelists[nodeid] = ptr;
1286 * Initialisation. Called after the page allocator have been initialised and
1287 * before smp_init().
1289 void __init kmem_cache_init(void)
1292 struct cache_sizes *sizes;
1293 struct cache_names *names;
1297 for (i = 0; i < NUM_INIT_LISTS; i++) {
1298 kmem_list3_init(&initkmem_list3[i]);
1299 if (i < MAX_NUMNODES)
1300 cache_cache.nodelists[i] = NULL;
1304 * Fragmentation resistance on low memory - only use bigger
1305 * page orders on machines with more than 32MB of memory.
1307 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1308 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1310 /* Bootstrap is tricky, because several objects are allocated
1311 * from caches that do not exist yet:
1312 * 1) initialize the cache_cache cache: it contains the struct
1313 * kmem_cache structures of all caches, except cache_cache itself:
1314 * cache_cache is statically allocated.
1315 * Initially an __init data area is used for the head array and the
1316 * kmem_list3 structures, it's replaced with a kmalloc allocated
1317 * array at the end of the bootstrap.
1318 * 2) Create the first kmalloc cache.
1319 * The struct kmem_cache for the new cache is allocated normally.
1320 * An __init data area is used for the head array.
1321 * 3) Create the remaining kmalloc caches, with minimally sized
1323 * 4) Replace the __init data head arrays for cache_cache and the first
1324 * kmalloc cache with kmalloc allocated arrays.
1325 * 5) Replace the __init data for kmem_list3 for cache_cache and
1326 * the other cache's with kmalloc allocated memory.
1327 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1330 /* 1) create the cache_cache */
1331 INIT_LIST_HEAD(&cache_chain);
1332 list_add(&cache_cache.next, &cache_chain);
1333 cache_cache.colour_off = cache_line_size();
1334 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1335 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
1337 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1340 for (order = 0; order < MAX_ORDER; order++) {
1341 cache_estimate(order, cache_cache.buffer_size,
1342 cache_line_size(), 0, &left_over, &cache_cache.num);
1343 if (cache_cache.num)
1346 BUG_ON(!cache_cache.num);
1347 cache_cache.gfporder = order;
1348 cache_cache.colour = left_over / cache_cache.colour_off;
1349 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1350 sizeof(struct slab), cache_line_size());
1352 /* 2+3) create the kmalloc caches */
1353 sizes = malloc_sizes;
1354 names = cache_names;
1357 * Initialize the caches that provide memory for the array cache and the
1358 * kmem_list3 structures first. Without this, further allocations will
1362 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1363 sizes[INDEX_AC].cs_size,
1364 ARCH_KMALLOC_MINALIGN,
1365 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1368 if (INDEX_AC != INDEX_L3) {
1369 sizes[INDEX_L3].cs_cachep =
1370 kmem_cache_create(names[INDEX_L3].name,
1371 sizes[INDEX_L3].cs_size,
1372 ARCH_KMALLOC_MINALIGN,
1373 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1377 while (sizes->cs_size != ULONG_MAX) {
1379 * For performance, all the general caches are L1 aligned.
1380 * This should be particularly beneficial on SMP boxes, as it
1381 * eliminates "false sharing".
1382 * Note for systems short on memory removing the alignment will
1383 * allow tighter packing of the smaller caches.
1385 if (!sizes->cs_cachep) {
1386 sizes->cs_cachep = kmem_cache_create(names->name,
1388 ARCH_KMALLOC_MINALIGN,
1389 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1393 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
1395 ARCH_KMALLOC_MINALIGN,
1396 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1402 /* 4) Replace the bootstrap head arrays */
1406 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1408 local_irq_disable();
1409 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1410 memcpy(ptr, cpu_cache_get(&cache_cache),
1411 sizeof(struct arraycache_init));
1412 cache_cache.array[smp_processor_id()] = ptr;
1415 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1417 local_irq_disable();
1418 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1419 != &initarray_generic.cache);
1420 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1421 sizeof(struct arraycache_init));
1422 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1426 /* 5) Replace the bootstrap kmem_list3's */
1429 /* Replace the static kmem_list3 structures for the boot cpu */
1430 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
1433 for_each_online_node(node) {
1434 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1435 &initkmem_list3[SIZE_AC + node], node);
1437 if (INDEX_AC != INDEX_L3) {
1438 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1439 &initkmem_list3[SIZE_L3 + node],
1445 /* 6) resize the head arrays to their final sizes */
1447 struct kmem_cache *cachep;
1448 mutex_lock(&cache_chain_mutex);
1449 list_for_each_entry(cachep, &cache_chain, next)
1450 enable_cpucache(cachep);
1451 mutex_unlock(&cache_chain_mutex);
1455 g_cpucache_up = FULL;
1458 * Register a cpu startup notifier callback that initializes
1459 * cpu_cache_get for all new cpus
1461 register_cpu_notifier(&cpucache_notifier);
1464 * The reap timers are started later, with a module init call: That part
1465 * of the kernel is not yet operational.
1469 static int __init cpucache_init(void)
1474 * Register the timers that return unneeded pages to the page allocator
1476 for_each_online_cpu(cpu)
1477 start_cpu_timer(cpu);
1480 __initcall(cpucache_init);
1483 * Interface to system's page allocator. No need to hold the cache-lock.
1485 * If we requested dmaable memory, we will get it. Even if we
1486 * did not request dmaable memory, we might get it, but that
1487 * would be relatively rare and ignorable.
1489 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1497 * Nommu uses slab's for process anonymous memory allocations, and thus
1498 * requires __GFP_COMP to properly refcount higher order allocations
1500 flags |= __GFP_COMP;
1502 flags |= cachep->gfpflags;
1504 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1508 nr_pages = (1 << cachep->gfporder);
1509 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1510 atomic_add(nr_pages, &slab_reclaim_pages);
1511 add_page_state(nr_slab, nr_pages);
1512 for (i = 0; i < nr_pages; i++)
1513 __SetPageSlab(page + i);
1514 return page_address(page);
1518 * Interface to system's page release.
1520 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1522 unsigned long i = (1 << cachep->gfporder);
1523 struct page *page = virt_to_page(addr);
1524 const unsigned long nr_freed = i;
1527 BUG_ON(!PageSlab(page));
1528 __ClearPageSlab(page);
1531 sub_page_state(nr_slab, nr_freed);
1532 if (current->reclaim_state)
1533 current->reclaim_state->reclaimed_slab += nr_freed;
1534 free_pages((unsigned long)addr, cachep->gfporder);
1535 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1536 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages);
1539 static void kmem_rcu_free(struct rcu_head *head)
1541 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1542 struct kmem_cache *cachep = slab_rcu->cachep;
1544 kmem_freepages(cachep, slab_rcu->addr);
1545 if (OFF_SLAB(cachep))
1546 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1551 #ifdef CONFIG_DEBUG_PAGEALLOC
1552 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1553 unsigned long caller)
1555 int size = obj_size(cachep);
1557 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1559 if (size < 5 * sizeof(unsigned long))
1562 *addr++ = 0x12345678;
1564 *addr++ = smp_processor_id();
1565 size -= 3 * sizeof(unsigned long);
1567 unsigned long *sptr = &caller;
1568 unsigned long svalue;
1570 while (!kstack_end(sptr)) {
1572 if (kernel_text_address(svalue)) {
1574 size -= sizeof(unsigned long);
1575 if (size <= sizeof(unsigned long))
1581 *addr++ = 0x87654321;
1585 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1587 int size = obj_size(cachep);
1588 addr = &((char *)addr)[obj_offset(cachep)];
1590 memset(addr, val, size);
1591 *(unsigned char *)(addr + size - 1) = POISON_END;
1594 static void dump_line(char *data, int offset, int limit)
1597 printk(KERN_ERR "%03x:", offset);
1598 for (i = 0; i < limit; i++)
1599 printk(" %02x", (unsigned char)data[offset + i]);
1606 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1611 if (cachep->flags & SLAB_RED_ZONE) {
1612 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
1613 *dbg_redzone1(cachep, objp),
1614 *dbg_redzone2(cachep, objp));
1617 if (cachep->flags & SLAB_STORE_USER) {
1618 printk(KERN_ERR "Last user: [<%p>]",
1619 *dbg_userword(cachep, objp));
1620 print_symbol("(%s)",
1621 (unsigned long)*dbg_userword(cachep, objp));
1624 realobj = (char *)objp + obj_offset(cachep);
1625 size = obj_size(cachep);
1626 for (i = 0; i < size && lines; i += 16, lines--) {
1629 if (i + limit > size)
1631 dump_line(realobj, i, limit);
1635 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1641 realobj = (char *)objp + obj_offset(cachep);
1642 size = obj_size(cachep);
1644 for (i = 0; i < size; i++) {
1645 char exp = POISON_FREE;
1648 if (realobj[i] != exp) {
1654 "Slab corruption: start=%p, len=%d\n",
1656 print_objinfo(cachep, objp, 0);
1658 /* Hexdump the affected line */
1661 if (i + limit > size)
1663 dump_line(realobj, i, limit);
1666 /* Limit to 5 lines */
1672 /* Print some data about the neighboring objects, if they
1675 struct slab *slabp = virt_to_slab(objp);
1678 objnr = obj_to_index(cachep, slabp, objp);
1680 objp = index_to_obj(cachep, slabp, objnr - 1);
1681 realobj = (char *)objp + obj_offset(cachep);
1682 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1684 print_objinfo(cachep, objp, 2);
1686 if (objnr + 1 < cachep->num) {
1687 objp = index_to_obj(cachep, slabp, objnr + 1);
1688 realobj = (char *)objp + obj_offset(cachep);
1689 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1691 print_objinfo(cachep, objp, 2);
1699 * slab_destroy_objs - destroy a slab and its objects
1700 * @cachep: cache pointer being destroyed
1701 * @slabp: slab pointer being destroyed
1703 * Call the registered destructor for each object in a slab that is being
1706 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1709 for (i = 0; i < cachep->num; i++) {
1710 void *objp = index_to_obj(cachep, slabp, i);
1712 if (cachep->flags & SLAB_POISON) {
1713 #ifdef CONFIG_DEBUG_PAGEALLOC
1714 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1716 kernel_map_pages(virt_to_page(objp),
1717 cachep->buffer_size / PAGE_SIZE, 1);
1719 check_poison_obj(cachep, objp);
1721 check_poison_obj(cachep, objp);
1724 if (cachep->flags & SLAB_RED_ZONE) {
1725 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1726 slab_error(cachep, "start of a freed object "
1728 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1729 slab_error(cachep, "end of a freed object "
1732 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1733 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
1737 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1741 for (i = 0; i < cachep->num; i++) {
1742 void *objp = index_to_obj(cachep, slabp, i);
1743 (cachep->dtor) (objp, cachep, 0);
1750 * slab_destroy - destroy and release all objects in a slab
1751 * @cachep: cache pointer being destroyed
1752 * @slabp: slab pointer being destroyed
1754 * Destroy all the objs in a slab, and release the mem back to the system.
1755 * Before calling the slab must have been unlinked from the cache. The
1756 * cache-lock is not held/needed.
1758 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1760 void *addr = slabp->s_mem - slabp->colouroff;
1762 slab_destroy_objs(cachep, slabp);
1763 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1764 struct slab_rcu *slab_rcu;
1766 slab_rcu = (struct slab_rcu *)slabp;
1767 slab_rcu->cachep = cachep;
1768 slab_rcu->addr = addr;
1769 call_rcu(&slab_rcu->head, kmem_rcu_free);
1771 kmem_freepages(cachep, addr);
1772 if (OFF_SLAB(cachep))
1773 kmem_cache_free(cachep->slabp_cache, slabp);
1778 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1779 * size of kmem_list3.
1781 static void set_up_list3s(struct kmem_cache *cachep, int index)
1785 for_each_online_node(node) {
1786 cachep->nodelists[node] = &initkmem_list3[index + node];
1787 cachep->nodelists[node]->next_reap = jiffies +
1789 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1794 * calculate_slab_order - calculate size (page order) of slabs
1795 * @cachep: pointer to the cache that is being created
1796 * @size: size of objects to be created in this cache.
1797 * @align: required alignment for the objects.
1798 * @flags: slab allocation flags
1800 * Also calculates the number of objects per slab.
1802 * This could be made much more intelligent. For now, try to avoid using
1803 * high order pages for slabs. When the gfp() functions are more friendly
1804 * towards high-order requests, this should be changed.
1806 static size_t calculate_slab_order(struct kmem_cache *cachep,
1807 size_t size, size_t align, unsigned long flags)
1809 unsigned long offslab_limit;
1810 size_t left_over = 0;
1813 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
1817 cache_estimate(gfporder, size, align, flags, &remainder, &num);
1821 if (flags & CFLGS_OFF_SLAB) {
1823 * Max number of objs-per-slab for caches which
1824 * use off-slab slabs. Needed to avoid a possible
1825 * looping condition in cache_grow().
1827 offslab_limit = size - sizeof(struct slab);
1828 offslab_limit /= sizeof(kmem_bufctl_t);
1830 if (num > offslab_limit)
1834 /* Found something acceptable - save it away */
1836 cachep->gfporder = gfporder;
1837 left_over = remainder;
1840 * A VFS-reclaimable slab tends to have most allocations
1841 * as GFP_NOFS and we really don't want to have to be allocating
1842 * higher-order pages when we are unable to shrink dcache.
1844 if (flags & SLAB_RECLAIM_ACCOUNT)
1848 * Large number of objects is good, but very large slabs are
1849 * currently bad for the gfp()s.
1851 if (gfporder >= slab_break_gfp_order)
1855 * Acceptable internal fragmentation?
1857 if (left_over * 8 <= (PAGE_SIZE << gfporder))
1863 static void setup_cpu_cache(struct kmem_cache *cachep)
1865 if (g_cpucache_up == FULL) {
1866 enable_cpucache(cachep);
1869 if (g_cpucache_up == NONE) {
1871 * Note: the first kmem_cache_create must create the cache
1872 * that's used by kmalloc(24), otherwise the creation of
1873 * further caches will BUG().
1875 cachep->array[smp_processor_id()] = &initarray_generic.cache;
1878 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
1879 * the first cache, then we need to set up all its list3s,
1880 * otherwise 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) {
1897 cachep->nodelists[node] =
1898 kmalloc_node(sizeof(struct kmem_list3),
1900 BUG_ON(!cachep->nodelists[node]);
1901 kmem_list3_init(cachep->nodelists[node]);
1905 cachep->nodelists[numa_node_id()]->next_reap =
1906 jiffies + REAPTIMEOUT_LIST3 +
1907 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1909 cpu_cache_get(cachep)->avail = 0;
1910 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1911 cpu_cache_get(cachep)->batchcount = 1;
1912 cpu_cache_get(cachep)->touched = 0;
1913 cachep->batchcount = 1;
1914 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1918 * kmem_cache_create - Create a cache.
1919 * @name: A string which is used in /proc/slabinfo to identify this cache.
1920 * @size: The size of objects to be created in this cache.
1921 * @align: The required alignment for the objects.
1922 * @flags: SLAB flags
1923 * @ctor: A constructor for the objects.
1924 * @dtor: A destructor for the objects.
1926 * Returns a ptr to the cache on success, NULL on failure.
1927 * Cannot be called within a int, but can be interrupted.
1928 * The @ctor is run when new pages are allocated by the cache
1929 * and the @dtor is run before the pages are handed back.
1931 * @name must be valid until the cache is destroyed. This implies that
1932 * the module calling this has to destroy the cache before getting unloaded.
1936 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1937 * to catch references to uninitialised memory.
1939 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1940 * for buffer overruns.
1942 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1943 * cacheline. This can be beneficial if you're counting cycles as closely
1947 kmem_cache_create (const char *name, size_t size, size_t align,
1948 unsigned long flags,
1949 void (*ctor)(void*, struct kmem_cache *, unsigned long),
1950 void (*dtor)(void*, struct kmem_cache *, unsigned long))
1952 size_t left_over, slab_size, ralign;
1953 struct kmem_cache *cachep = NULL, *pc;
1956 * Sanity checks... these are all serious usage bugs.
1958 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
1959 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
1960 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
1966 * Prevent CPUs from coming and going.
1967 * lock_cpu_hotplug() nests outside cache_chain_mutex
1971 mutex_lock(&cache_chain_mutex);
1973 list_for_each_entry(pc, &cache_chain, next) {
1974 mm_segment_t old_fs = get_fs();
1979 * This happens when the module gets unloaded and doesn't
1980 * destroy its slab cache and no-one else reuses the vmalloc
1981 * area of the module. Print a warning.
1984 res = __get_user(tmp, pc->name);
1987 printk("SLAB: cache with size %d has lost its name\n",
1992 if (!strcmp(pc->name, name)) {
1993 printk("kmem_cache_create: duplicate cache %s\n", name);
2000 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2001 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
2002 /* No constructor, but inital state check requested */
2003 printk(KERN_ERR "%s: No con, but init state check "
2004 "requested - %s\n", __FUNCTION__, name);
2005 flags &= ~SLAB_DEBUG_INITIAL;
2009 * Enable redzoning and last user accounting, except for caches with
2010 * large objects, if the increased size would increase the object size
2011 * above the next power of two: caches with object sizes just above a
2012 * power of two have a significant amount of internal fragmentation.
2014 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
2015 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2016 if (!(flags & SLAB_DESTROY_BY_RCU))
2017 flags |= SLAB_POISON;
2019 if (flags & SLAB_DESTROY_BY_RCU)
2020 BUG_ON(flags & SLAB_POISON);
2022 if (flags & SLAB_DESTROY_BY_RCU)
2026 * Always checks flags, a caller might be expecting debug support which
2029 BUG_ON(flags & ~CREATE_MASK);
2032 * Check that size is in terms of words. This is needed to avoid
2033 * unaligned accesses for some archs when redzoning is used, and makes
2034 * sure any on-slab bufctl's are also correctly aligned.
2036 if (size & (BYTES_PER_WORD - 1)) {
2037 size += (BYTES_PER_WORD - 1);
2038 size &= ~(BYTES_PER_WORD - 1);
2041 /* calculate the final buffer alignment: */
2043 /* 1) arch recommendation: can be overridden for debug */
2044 if (flags & SLAB_HWCACHE_ALIGN) {
2046 * Default alignment: as specified by the arch code. Except if
2047 * an object is really small, then squeeze multiple objects into
2050 ralign = cache_line_size();
2051 while (size <= ralign / 2)
2054 ralign = BYTES_PER_WORD;
2056 /* 2) arch mandated alignment: disables debug if necessary */
2057 if (ralign < ARCH_SLAB_MINALIGN) {
2058 ralign = ARCH_SLAB_MINALIGN;
2059 if (ralign > BYTES_PER_WORD)
2060 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2062 /* 3) caller mandated alignment: disables debug if necessary */
2063 if (ralign < align) {
2065 if (ralign > BYTES_PER_WORD)
2066 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2069 * 4) Store it. Note that the debug code below can reduce
2070 * the alignment to BYTES_PER_WORD.
2074 /* Get cache's description obj. */
2075 cachep = kmem_cache_zalloc(&cache_cache, SLAB_KERNEL);
2080 cachep->obj_size = size;
2082 if (flags & SLAB_RED_ZONE) {
2083 /* redzoning only works with word aligned caches */
2084 align = BYTES_PER_WORD;
2086 /* add space for red zone words */
2087 cachep->obj_offset += BYTES_PER_WORD;
2088 size += 2 * BYTES_PER_WORD;
2090 if (flags & SLAB_STORE_USER) {
2091 /* user store requires word alignment and
2092 * one word storage behind the end of the real
2095 align = BYTES_PER_WORD;
2096 size += BYTES_PER_WORD;
2098 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2099 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2100 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2101 cachep->obj_offset += PAGE_SIZE - size;
2107 /* Determine if the slab management is 'on' or 'off' slab. */
2108 if (size >= (PAGE_SIZE >> 3))
2110 * Size is large, assume best to place the slab management obj
2111 * off-slab (should allow better packing of objs).
2113 flags |= CFLGS_OFF_SLAB;
2115 size = ALIGN(size, align);
2117 left_over = calculate_slab_order(cachep, size, align, flags);
2120 printk("kmem_cache_create: couldn't create cache %s.\n", name);
2121 kmem_cache_free(&cache_cache, cachep);
2125 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2126 + sizeof(struct slab), align);
2129 * If the slab has been placed off-slab, and we have enough space then
2130 * move it on-slab. This is at the expense of any extra colouring.
2132 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2133 flags &= ~CFLGS_OFF_SLAB;
2134 left_over -= slab_size;
2137 if (flags & CFLGS_OFF_SLAB) {
2138 /* really off slab. No need for manual alignment */
2140 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2143 cachep->colour_off = cache_line_size();
2144 /* Offset must be a multiple of the alignment. */
2145 if (cachep->colour_off < align)
2146 cachep->colour_off = align;
2147 cachep->colour = left_over / cachep->colour_off;
2148 cachep->slab_size = slab_size;
2149 cachep->flags = flags;
2150 cachep->gfpflags = 0;
2151 if (flags & SLAB_CACHE_DMA)
2152 cachep->gfpflags |= GFP_DMA;
2153 cachep->buffer_size = size;
2155 if (flags & CFLGS_OFF_SLAB)
2156 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2157 cachep->ctor = ctor;
2158 cachep->dtor = dtor;
2159 cachep->name = name;
2162 setup_cpu_cache(cachep);
2164 /* cache setup completed, link it into the list */
2165 list_add(&cachep->next, &cache_chain);
2167 if (!cachep && (flags & SLAB_PANIC))
2168 panic("kmem_cache_create(): failed to create slab `%s'\n",
2170 mutex_unlock(&cache_chain_mutex);
2171 unlock_cpu_hotplug();
2174 EXPORT_SYMBOL(kmem_cache_create);
2177 static void check_irq_off(void)
2179 BUG_ON(!irqs_disabled());
2182 static void check_irq_on(void)
2184 BUG_ON(irqs_disabled());
2187 static void check_spinlock_acquired(struct kmem_cache *cachep)
2191 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2195 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2199 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2204 #define check_irq_off() do { } while(0)
2205 #define check_irq_on() do { } while(0)
2206 #define check_spinlock_acquired(x) do { } while(0)
2207 #define check_spinlock_acquired_node(x, y) do { } while(0)
2210 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2211 struct array_cache *ac,
2212 int force, int node);
2214 static void do_drain(void *arg)
2216 struct kmem_cache *cachep = arg;
2217 struct array_cache *ac;
2218 int node = numa_node_id();
2221 ac = cpu_cache_get(cachep);
2222 spin_lock(&cachep->nodelists[node]->list_lock);
2223 free_block(cachep, ac->entry, ac->avail, node);
2224 spin_unlock(&cachep->nodelists[node]->list_lock);
2228 static void drain_cpu_caches(struct kmem_cache *cachep)
2230 struct kmem_list3 *l3;
2233 on_each_cpu(do_drain, cachep, 1, 1);
2235 for_each_online_node(node) {
2236 l3 = cachep->nodelists[node];
2237 if (l3 && l3->alien)
2238 drain_alien_cache(cachep, l3->alien);
2241 for_each_online_node(node) {
2242 l3 = cachep->nodelists[node];
2244 drain_array(cachep, l3, l3->shared, 1, node);
2248 static int __node_shrink(struct kmem_cache *cachep, int node)
2251 struct kmem_list3 *l3 = cachep->nodelists[node];
2255 struct list_head *p;
2257 p = l3->slabs_free.prev;
2258 if (p == &l3->slabs_free)
2261 slabp = list_entry(l3->slabs_free.prev, struct slab, list);
2263 BUG_ON(slabp->inuse);
2265 list_del(&slabp->list);
2267 l3->free_objects -= cachep->num;
2268 spin_unlock_irq(&l3->list_lock);
2269 slab_destroy(cachep, slabp);
2270 spin_lock_irq(&l3->list_lock);
2272 ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial);
2276 static int __cache_shrink(struct kmem_cache *cachep)
2279 struct kmem_list3 *l3;
2281 drain_cpu_caches(cachep);
2284 for_each_online_node(i) {
2285 l3 = cachep->nodelists[i];
2287 spin_lock_irq(&l3->list_lock);
2288 ret += __node_shrink(cachep, i);
2289 spin_unlock_irq(&l3->list_lock);
2292 return (ret ? 1 : 0);
2296 * kmem_cache_shrink - Shrink a cache.
2297 * @cachep: The cache to shrink.
2299 * Releases as many slabs as possible for a cache.
2300 * To help debugging, a zero exit status indicates all slabs were released.
2302 int kmem_cache_shrink(struct kmem_cache *cachep)
2304 BUG_ON(!cachep || in_interrupt());
2306 return __cache_shrink(cachep);
2308 EXPORT_SYMBOL(kmem_cache_shrink);
2311 * kmem_cache_destroy - delete a cache
2312 * @cachep: the cache to destroy
2314 * Remove a struct kmem_cache object from the slab cache.
2315 * Returns 0 on success.
2317 * It is expected this function will be called by a module when it is
2318 * unloaded. This will remove the cache completely, and avoid a duplicate
2319 * cache being allocated each time a module is loaded and unloaded, if the
2320 * module doesn't have persistent in-kernel storage across loads and unloads.
2322 * The cache must be empty before calling this function.
2324 * The caller must guarantee that noone will allocate memory from the cache
2325 * during the kmem_cache_destroy().
2327 int kmem_cache_destroy(struct kmem_cache *cachep)
2330 struct kmem_list3 *l3;
2332 BUG_ON(!cachep || in_interrupt());
2334 /* Don't let CPUs to come and go */
2337 /* Find the cache in the chain of caches. */
2338 mutex_lock(&cache_chain_mutex);
2340 * the chain is never empty, cache_cache is never destroyed
2342 list_del(&cachep->next);
2343 mutex_unlock(&cache_chain_mutex);
2345 if (__cache_shrink(cachep)) {
2346 slab_error(cachep, "Can't free all objects");
2347 mutex_lock(&cache_chain_mutex);
2348 list_add(&cachep->next, &cache_chain);
2349 mutex_unlock(&cache_chain_mutex);
2350 unlock_cpu_hotplug();
2354 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2357 for_each_online_cpu(i)
2358 kfree(cachep->array[i]);
2360 /* NUMA: free the list3 structures */
2361 for_each_online_node(i) {
2362 l3 = cachep->nodelists[i];
2365 free_alien_cache(l3->alien);
2369 kmem_cache_free(&cache_cache, cachep);
2370 unlock_cpu_hotplug();
2373 EXPORT_SYMBOL(kmem_cache_destroy);
2375 /* Get the memory for a slab management obj. */
2376 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2377 int colour_off, gfp_t local_flags,
2382 if (OFF_SLAB(cachep)) {
2383 /* Slab management obj is off-slab. */
2384 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2385 local_flags, nodeid);
2389 slabp = objp + colour_off;
2390 colour_off += cachep->slab_size;
2393 slabp->colouroff = colour_off;
2394 slabp->s_mem = objp + colour_off;
2395 slabp->nodeid = nodeid;
2399 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2401 return (kmem_bufctl_t *) (slabp + 1);
2404 static void cache_init_objs(struct kmem_cache *cachep,
2405 struct slab *slabp, unsigned long ctor_flags)
2409 for (i = 0; i < cachep->num; i++) {
2410 void *objp = index_to_obj(cachep, slabp, i);
2412 /* need to poison the objs? */
2413 if (cachep->flags & SLAB_POISON)
2414 poison_obj(cachep, objp, POISON_FREE);
2415 if (cachep->flags & SLAB_STORE_USER)
2416 *dbg_userword(cachep, objp) = NULL;
2418 if (cachep->flags & SLAB_RED_ZONE) {
2419 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2420 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2423 * Constructors are not allowed to allocate memory from the same
2424 * cache which they are a constructor for. Otherwise, deadlock.
2425 * They must also be threaded.
2427 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2428 cachep->ctor(objp + obj_offset(cachep), cachep,
2431 if (cachep->flags & SLAB_RED_ZONE) {
2432 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2433 slab_error(cachep, "constructor overwrote the"
2434 " end of an object");
2435 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2436 slab_error(cachep, "constructor overwrote the"
2437 " start of an object");
2439 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2440 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2441 kernel_map_pages(virt_to_page(objp),
2442 cachep->buffer_size / PAGE_SIZE, 0);
2445 cachep->ctor(objp, cachep, ctor_flags);
2447 slab_bufctl(slabp)[i] = i + 1;
2449 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2453 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2455 if (flags & SLAB_DMA)
2456 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2458 BUG_ON(cachep->gfpflags & GFP_DMA);
2461 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2464 void *objp = index_to_obj(cachep, slabp, slabp->free);
2468 next = slab_bufctl(slabp)[slabp->free];
2470 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2471 WARN_ON(slabp->nodeid != nodeid);
2478 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2479 void *objp, int nodeid)
2481 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2484 /* Verify that the slab belongs to the intended node */
2485 WARN_ON(slabp->nodeid != nodeid);
2487 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2488 printk(KERN_ERR "slab: double free detected in cache "
2489 "'%s', objp %p\n", cachep->name, objp);
2493 slab_bufctl(slabp)[objnr] = slabp->free;
2494 slabp->free = objnr;
2499 * Map pages beginning at addr to the given cache and slab. This is required
2500 * for the slab allocator to be able to lookup the cache and slab of a
2501 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2503 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2509 page = virt_to_page(addr);
2512 if (likely(!PageCompound(page)))
2513 nr_pages <<= cache->gfporder;
2516 page_set_cache(page, cache);
2517 page_set_slab(page, slab);
2519 } while (--nr_pages);
2523 * Grow (by 1) the number of slabs within a cache. This is called by
2524 * kmem_cache_alloc() when there are no active objs left in a cache.
2526 static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid)
2532 unsigned long ctor_flags;
2533 struct kmem_list3 *l3;
2536 * Be lazy and only check for valid flags here, keeping it out of the
2537 * critical path in kmem_cache_alloc().
2539 BUG_ON(flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW));
2540 if (flags & SLAB_NO_GROW)
2543 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2544 local_flags = (flags & SLAB_LEVEL_MASK);
2545 if (!(local_flags & __GFP_WAIT))
2547 * Not allowed to sleep. Need to tell a constructor about
2548 * this - it might need to know...
2550 ctor_flags |= SLAB_CTOR_ATOMIC;
2552 /* Take the l3 list lock to change the colour_next on this node */
2554 l3 = cachep->nodelists[nodeid];
2555 spin_lock(&l3->list_lock);
2557 /* Get colour for the slab, and cal the next value. */
2558 offset = l3->colour_next;
2560 if (l3->colour_next >= cachep->colour)
2561 l3->colour_next = 0;
2562 spin_unlock(&l3->list_lock);
2564 offset *= cachep->colour_off;
2566 if (local_flags & __GFP_WAIT)
2570 * The test for missing atomic flag is performed here, rather than
2571 * the more obvious place, simply to reduce the critical path length
2572 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2573 * will eventually be caught here (where it matters).
2575 kmem_flagcheck(cachep, flags);
2578 * Get mem for the objs. Attempt to allocate a physical page from
2581 objp = kmem_getpages(cachep, flags, nodeid);
2585 /* Get slab management. */
2586 slabp = alloc_slabmgmt(cachep, objp, offset, local_flags, nodeid);
2590 slabp->nodeid = nodeid;
2591 slab_map_pages(cachep, slabp, objp);
2593 cache_init_objs(cachep, slabp, ctor_flags);
2595 if (local_flags & __GFP_WAIT)
2596 local_irq_disable();
2598 spin_lock(&l3->list_lock);
2600 /* Make slab active. */
2601 list_add_tail(&slabp->list, &(l3->slabs_free));
2602 STATS_INC_GROWN(cachep);
2603 l3->free_objects += cachep->num;
2604 spin_unlock(&l3->list_lock);
2607 kmem_freepages(cachep, objp);
2609 if (local_flags & __GFP_WAIT)
2610 local_irq_disable();
2617 * Perform extra freeing checks:
2618 * - detect bad pointers.
2619 * - POISON/RED_ZONE checking
2620 * - destructor calls, for caches with POISON+dtor
2622 static void kfree_debugcheck(const void *objp)
2626 if (!virt_addr_valid(objp)) {
2627 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2628 (unsigned long)objp);
2631 page = virt_to_page(objp);
2632 if (!PageSlab(page)) {
2633 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2634 (unsigned long)objp);
2639 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2641 unsigned long redzone1, redzone2;
2643 redzone1 = *dbg_redzone1(cache, obj);
2644 redzone2 = *dbg_redzone2(cache, obj);
2649 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2652 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2653 slab_error(cache, "double free detected");
2655 slab_error(cache, "memory outside object was overwritten");
2657 printk(KERN_ERR "%p: redzone 1:0x%lx, redzone 2:0x%lx.\n",
2658 obj, redzone1, redzone2);
2661 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2668 objp -= obj_offset(cachep);
2669 kfree_debugcheck(objp);
2670 page = virt_to_page(objp);
2672 if (page_get_cache(page) != cachep) {
2673 printk(KERN_ERR "mismatch in kmem_cache_free: expected "
2674 "cache %p, got %p\n",
2675 page_get_cache(page), cachep);
2676 printk(KERN_ERR "%p is %s.\n", cachep, cachep->name);
2677 printk(KERN_ERR "%p is %s.\n", page_get_cache(page),
2678 page_get_cache(page)->name);
2681 slabp = page_get_slab(page);
2683 if (cachep->flags & SLAB_RED_ZONE) {
2684 verify_redzone_free(cachep, objp);
2685 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2686 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2688 if (cachep->flags & SLAB_STORE_USER)
2689 *dbg_userword(cachep, objp) = caller;
2691 objnr = obj_to_index(cachep, slabp, objp);
2693 BUG_ON(objnr >= cachep->num);
2694 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2696 if (cachep->flags & SLAB_DEBUG_INITIAL) {
2698 * Need to call the slab's constructor so the caller can
2699 * perform a verify of its state (debugging). Called without
2700 * the cache-lock held.
2702 cachep->ctor(objp + obj_offset(cachep),
2703 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
2705 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2706 /* we want to cache poison the object,
2707 * call the destruction callback
2709 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
2711 #ifdef CONFIG_DEBUG_SLAB_LEAK
2712 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2714 if (cachep->flags & SLAB_POISON) {
2715 #ifdef CONFIG_DEBUG_PAGEALLOC
2716 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2717 store_stackinfo(cachep, objp, (unsigned long)caller);
2718 kernel_map_pages(virt_to_page(objp),
2719 cachep->buffer_size / PAGE_SIZE, 0);
2721 poison_obj(cachep, objp, POISON_FREE);
2724 poison_obj(cachep, objp, POISON_FREE);
2730 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2735 /* Check slab's freelist to see if this obj is there. */
2736 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2738 if (entries > cachep->num || i >= cachep->num)
2741 if (entries != cachep->num - slabp->inuse) {
2743 printk(KERN_ERR "slab: Internal list corruption detected in "
2744 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2745 cachep->name, cachep->num, slabp, slabp->inuse);
2747 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2750 printk("\n%03x:", i);
2751 printk(" %02x", ((unsigned char *)slabp)[i]);
2758 #define kfree_debugcheck(x) do { } while(0)
2759 #define cache_free_debugcheck(x,objp,z) (objp)
2760 #define check_slabp(x,y) do { } while(0)
2763 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2766 struct kmem_list3 *l3;
2767 struct array_cache *ac;
2770 ac = cpu_cache_get(cachep);
2772 batchcount = ac->batchcount;
2773 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2775 * If there was little recent activity on this cache, then
2776 * perform only a partial refill. Otherwise we could generate
2779 batchcount = BATCHREFILL_LIMIT;
2781 l3 = cachep->nodelists[numa_node_id()];
2783 BUG_ON(ac->avail > 0 || !l3);
2784 spin_lock(&l3->list_lock);
2786 /* See if we can refill from the shared array */
2787 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2790 while (batchcount > 0) {
2791 struct list_head *entry;
2793 /* Get slab alloc is to come from. */
2794 entry = l3->slabs_partial.next;
2795 if (entry == &l3->slabs_partial) {
2796 l3->free_touched = 1;
2797 entry = l3->slabs_free.next;
2798 if (entry == &l3->slabs_free)
2802 slabp = list_entry(entry, struct slab, list);
2803 check_slabp(cachep, slabp);
2804 check_spinlock_acquired(cachep);
2805 while (slabp->inuse < cachep->num && batchcount--) {
2806 STATS_INC_ALLOCED(cachep);
2807 STATS_INC_ACTIVE(cachep);
2808 STATS_SET_HIGH(cachep);
2810 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2813 check_slabp(cachep, slabp);
2815 /* move slabp to correct slabp list: */
2816 list_del(&slabp->list);
2817 if (slabp->free == BUFCTL_END)
2818 list_add(&slabp->list, &l3->slabs_full);
2820 list_add(&slabp->list, &l3->slabs_partial);
2824 l3->free_objects -= ac->avail;
2826 spin_unlock(&l3->list_lock);
2828 if (unlikely(!ac->avail)) {
2830 x = cache_grow(cachep, flags, numa_node_id());
2832 /* cache_grow can reenable interrupts, then ac could change. */
2833 ac = cpu_cache_get(cachep);
2834 if (!x && ac->avail == 0) /* no objects in sight? abort */
2837 if (!ac->avail) /* objects refilled by interrupt? */
2841 return ac->entry[--ac->avail];
2844 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2847 might_sleep_if(flags & __GFP_WAIT);
2849 kmem_flagcheck(cachep, flags);
2854 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2855 gfp_t flags, void *objp, void *caller)
2859 if (cachep->flags & SLAB_POISON) {
2860 #ifdef CONFIG_DEBUG_PAGEALLOC
2861 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
2862 kernel_map_pages(virt_to_page(objp),
2863 cachep->buffer_size / PAGE_SIZE, 1);
2865 check_poison_obj(cachep, objp);
2867 check_poison_obj(cachep, objp);
2869 poison_obj(cachep, objp, POISON_INUSE);
2871 if (cachep->flags & SLAB_STORE_USER)
2872 *dbg_userword(cachep, objp) = caller;
2874 if (cachep->flags & SLAB_RED_ZONE) {
2875 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2876 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2877 slab_error(cachep, "double free, or memory outside"
2878 " object was overwritten");
2880 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
2881 objp, *dbg_redzone1(cachep, objp),
2882 *dbg_redzone2(cachep, objp));
2884 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2885 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2887 #ifdef CONFIG_DEBUG_SLAB_LEAK
2892 slabp = page_get_slab(virt_to_page(objp));
2893 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
2894 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
2897 objp += obj_offset(cachep);
2898 if (cachep->ctor && cachep->flags & SLAB_POISON) {
2899 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2901 if (!(flags & __GFP_WAIT))
2902 ctor_flags |= SLAB_CTOR_ATOMIC;
2904 cachep->ctor(objp, cachep, ctor_flags);
2909 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2912 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
2915 struct array_cache *ac;
2918 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
2919 objp = alternate_node_alloc(cachep, flags);
2926 ac = cpu_cache_get(cachep);
2927 if (likely(ac->avail)) {
2928 STATS_INC_ALLOCHIT(cachep);
2930 objp = ac->entry[--ac->avail];
2932 STATS_INC_ALLOCMISS(cachep);
2933 objp = cache_alloc_refill(cachep, flags);
2938 static __always_inline void *__cache_alloc(struct kmem_cache *cachep,
2939 gfp_t flags, void *caller)
2941 unsigned long save_flags;
2944 cache_alloc_debugcheck_before(cachep, flags);
2946 local_irq_save(save_flags);
2947 objp = ____cache_alloc(cachep, flags);
2948 local_irq_restore(save_flags);
2949 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
2957 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
2959 * If we are in_interrupt, then process context, including cpusets and
2960 * mempolicy, may not apply and should not be used for allocation policy.
2962 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
2964 int nid_alloc, nid_here;
2968 nid_alloc = nid_here = numa_node_id();
2969 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
2970 nid_alloc = cpuset_mem_spread_node();
2971 else if (current->mempolicy)
2972 nid_alloc = slab_node(current->mempolicy);
2973 if (nid_alloc != nid_here)
2974 return __cache_alloc_node(cachep, flags, nid_alloc);
2979 * A interface to enable slab creation on nodeid
2981 static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
2984 struct list_head *entry;
2986 struct kmem_list3 *l3;
2990 l3 = cachep->nodelists[nodeid];
2995 spin_lock(&l3->list_lock);
2996 entry = l3->slabs_partial.next;
2997 if (entry == &l3->slabs_partial) {
2998 l3->free_touched = 1;
2999 entry = l3->slabs_free.next;
3000 if (entry == &l3->slabs_free)
3004 slabp = list_entry(entry, struct slab, list);
3005 check_spinlock_acquired_node(cachep, nodeid);
3006 check_slabp(cachep, slabp);
3008 STATS_INC_NODEALLOCS(cachep);
3009 STATS_INC_ACTIVE(cachep);
3010 STATS_SET_HIGH(cachep);
3012 BUG_ON(slabp->inuse == cachep->num);
3014 obj = slab_get_obj(cachep, slabp, nodeid);
3015 check_slabp(cachep, slabp);
3017 /* move slabp to correct slabp list: */
3018 list_del(&slabp->list);
3020 if (slabp->free == BUFCTL_END)
3021 list_add(&slabp->list, &l3->slabs_full);
3023 list_add(&slabp->list, &l3->slabs_partial);
3025 spin_unlock(&l3->list_lock);
3029 spin_unlock(&l3->list_lock);
3030 x = cache_grow(cachep, flags, nodeid);
3042 * Caller needs to acquire correct kmem_list's list_lock
3044 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3048 struct kmem_list3 *l3;
3050 for (i = 0; i < nr_objects; i++) {
3051 void *objp = objpp[i];
3054 slabp = virt_to_slab(objp);
3055 l3 = cachep->nodelists[node];
3056 list_del(&slabp->list);
3057 check_spinlock_acquired_node(cachep, node);
3058 check_slabp(cachep, slabp);
3059 slab_put_obj(cachep, slabp, objp, node);
3060 STATS_DEC_ACTIVE(cachep);
3062 check_slabp(cachep, slabp);
3064 /* fixup slab chains */
3065 if (slabp->inuse == 0) {
3066 if (l3->free_objects > l3->free_limit) {
3067 l3->free_objects -= cachep->num;
3068 slab_destroy(cachep, slabp);
3070 list_add(&slabp->list, &l3->slabs_free);
3073 /* Unconditionally move a slab to the end of the
3074 * partial list on free - maximum time for the
3075 * other objects to be freed, too.
3077 list_add_tail(&slabp->list, &l3->slabs_partial);
3082 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3085 struct kmem_list3 *l3;
3086 int node = numa_node_id();
3088 batchcount = ac->batchcount;
3090 BUG_ON(!batchcount || batchcount > ac->avail);
3093 l3 = cachep->nodelists[node];
3094 spin_lock(&l3->list_lock);
3096 struct array_cache *shared_array = l3->shared;
3097 int max = shared_array->limit - shared_array->avail;
3099 if (batchcount > max)
3101 memcpy(&(shared_array->entry[shared_array->avail]),
3102 ac->entry, sizeof(void *) * batchcount);
3103 shared_array->avail += batchcount;
3108 free_block(cachep, ac->entry, batchcount, node);
3113 struct list_head *p;
3115 p = l3->slabs_free.next;
3116 while (p != &(l3->slabs_free)) {
3119 slabp = list_entry(p, struct slab, list);
3120 BUG_ON(slabp->inuse);
3125 STATS_SET_FREEABLE(cachep, i);
3128 spin_unlock(&l3->list_lock);
3129 ac->avail -= batchcount;
3130 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3134 * Release an obj back to its cache. If the obj has a constructed state, it must
3135 * be in this state _before_ it is released. Called with disabled ints.
3137 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3139 struct array_cache *ac = cpu_cache_get(cachep);
3142 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3144 if (cache_free_alien(cachep, objp))
3147 if (likely(ac->avail < ac->limit)) {
3148 STATS_INC_FREEHIT(cachep);
3149 ac->entry[ac->avail++] = objp;
3152 STATS_INC_FREEMISS(cachep);
3153 cache_flusharray(cachep, ac);
3154 ac->entry[ac->avail++] = objp;
3159 * kmem_cache_alloc - Allocate an object
3160 * @cachep: The cache to allocate from.
3161 * @flags: See kmalloc().
3163 * Allocate an object from this cache. The flags are only relevant
3164 * if the cache has no available objects.
3166 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3168 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3170 EXPORT_SYMBOL(kmem_cache_alloc);
3173 * kmem_cache_alloc - Allocate an object. The memory is set to zero.
3174 * @cache: The cache to allocate from.
3175 * @flags: See kmalloc().
3177 * Allocate an object from this cache and set the allocated memory to zero.
3178 * The flags are only relevant if the cache has no available objects.
3180 void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
3182 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
3184 memset(ret, 0, obj_size(cache));
3187 EXPORT_SYMBOL(kmem_cache_zalloc);
3190 * kmem_ptr_validate - check if an untrusted pointer might
3192 * @cachep: the cache we're checking against
3193 * @ptr: pointer to validate
3195 * This verifies that the untrusted pointer looks sane:
3196 * it is _not_ a guarantee that the pointer is actually
3197 * part of the slab cache in question, but it at least
3198 * validates that the pointer can be dereferenced and
3199 * looks half-way sane.
3201 * Currently only used for dentry validation.
3203 int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr)
3205 unsigned long addr = (unsigned long)ptr;
3206 unsigned long min_addr = PAGE_OFFSET;
3207 unsigned long align_mask = BYTES_PER_WORD - 1;
3208 unsigned long size = cachep->buffer_size;
3211 if (unlikely(addr < min_addr))
3213 if (unlikely(addr > (unsigned long)high_memory - size))
3215 if (unlikely(addr & align_mask))
3217 if (unlikely(!kern_addr_valid(addr)))
3219 if (unlikely(!kern_addr_valid(addr + size - 1)))
3221 page = virt_to_page(ptr);
3222 if (unlikely(!PageSlab(page)))
3224 if (unlikely(page_get_cache(page) != cachep))
3233 * kmem_cache_alloc_node - Allocate an object on the specified node
3234 * @cachep: The cache to allocate from.
3235 * @flags: See kmalloc().
3236 * @nodeid: node number of the target node.
3238 * Identical to kmem_cache_alloc, except that this function is slow
3239 * and can sleep. And it will allocate memory on the given node, which
3240 * can improve the performance for cpu bound structures.
3241 * New and improved: it will now make sure that the object gets
3242 * put on the correct node list so that there is no false sharing.
3244 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3246 unsigned long save_flags;
3249 cache_alloc_debugcheck_before(cachep, flags);
3250 local_irq_save(save_flags);
3252 if (nodeid == -1 || nodeid == numa_node_id() ||
3253 !cachep->nodelists[nodeid])
3254 ptr = ____cache_alloc(cachep, flags);
3256 ptr = __cache_alloc_node(cachep, flags, nodeid);
3257 local_irq_restore(save_flags);
3259 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr,
3260 __builtin_return_address(0));
3264 EXPORT_SYMBOL(kmem_cache_alloc_node);
3266 void *kmalloc_node(size_t size, gfp_t flags, int node)
3268 struct kmem_cache *cachep;
3270 cachep = kmem_find_general_cachep(size, flags);
3271 if (unlikely(cachep == NULL))
3273 return kmem_cache_alloc_node(cachep, flags, node);
3275 EXPORT_SYMBOL(kmalloc_node);
3279 * kmalloc - allocate memory
3280 * @size: how many bytes of memory are required.
3281 * @flags: the type of memory to allocate.
3282 * @caller: function caller for debug tracking of the caller
3284 * kmalloc is the normal method of allocating memory
3287 * The @flags argument may be one of:
3289 * %GFP_USER - Allocate memory on behalf of user. May sleep.
3291 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
3293 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
3295 * Additionally, the %GFP_DMA flag may be set to indicate the memory
3296 * must be suitable for DMA. This can mean different things on different
3297 * platforms. For example, on i386, it means that the memory must come
3298 * from the first 16MB.
3300 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3303 struct kmem_cache *cachep;
3305 /* If you want to save a few bytes .text space: replace
3307 * Then kmalloc uses the uninlined functions instead of the inline
3310 cachep = __find_general_cachep(size, flags);
3311 if (unlikely(cachep == NULL))
3313 return __cache_alloc(cachep, flags, caller);
3317 void *__kmalloc(size_t size, gfp_t flags)
3319 #ifndef CONFIG_DEBUG_SLAB
3320 return __do_kmalloc(size, flags, NULL);
3322 return __do_kmalloc(size, flags, __builtin_return_address(0));
3325 EXPORT_SYMBOL(__kmalloc);
3327 #ifdef CONFIG_DEBUG_SLAB
3328 void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3330 return __do_kmalloc(size, flags, caller);
3332 EXPORT_SYMBOL(__kmalloc_track_caller);
3337 * __alloc_percpu - allocate one copy of the object for every present
3338 * cpu in the system, zeroing them.
3339 * Objects should be dereferenced using the per_cpu_ptr macro only.
3341 * @size: how many bytes of memory are required.
3343 void *__alloc_percpu(size_t size)
3346 struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL);
3352 * Cannot use for_each_online_cpu since a cpu may come online
3353 * and we have no way of figuring out how to fix the array
3354 * that we have allocated then....
3356 for_each_possible_cpu(i) {
3357 int node = cpu_to_node(i);
3359 if (node_online(node))
3360 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
3362 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
3364 if (!pdata->ptrs[i])
3366 memset(pdata->ptrs[i], 0, size);
3369 /* Catch derefs w/o wrappers */
3370 return (void *)(~(unsigned long)pdata);
3374 if (!cpu_possible(i))
3376 kfree(pdata->ptrs[i]);
3381 EXPORT_SYMBOL(__alloc_percpu);
3385 * kmem_cache_free - Deallocate an object
3386 * @cachep: The cache the allocation was from.
3387 * @objp: The previously allocated object.
3389 * Free an object which was previously allocated from this
3392 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3394 unsigned long flags;
3396 local_irq_save(flags);
3397 __cache_free(cachep, objp);
3398 local_irq_restore(flags);
3400 EXPORT_SYMBOL(kmem_cache_free);
3403 * kfree - free previously allocated memory
3404 * @objp: pointer returned by kmalloc.
3406 * If @objp is NULL, no operation is performed.
3408 * Don't free memory not originally allocated by kmalloc()
3409 * or you will run into trouble.
3411 void kfree(const void *objp)
3413 struct kmem_cache *c;
3414 unsigned long flags;
3416 if (unlikely(!objp))
3418 local_irq_save(flags);
3419 kfree_debugcheck(objp);
3420 c = virt_to_cache(objp);
3421 mutex_debug_check_no_locks_freed(objp, obj_size(c));
3422 __cache_free(c, (void *)objp);
3423 local_irq_restore(flags);
3425 EXPORT_SYMBOL(kfree);
3429 * free_percpu - free previously allocated percpu memory
3430 * @objp: pointer returned by alloc_percpu.
3432 * Don't free memory not originally allocated by alloc_percpu()
3433 * The complemented objp is to check for that.
3435 void free_percpu(const void *objp)
3438 struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp);
3441 * We allocate for all cpus so we cannot use for online cpu here.
3443 for_each_possible_cpu(i)
3447 EXPORT_SYMBOL(free_percpu);
3450 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3452 return obj_size(cachep);
3454 EXPORT_SYMBOL(kmem_cache_size);
3456 const char *kmem_cache_name(struct kmem_cache *cachep)
3458 return cachep->name;
3460 EXPORT_SYMBOL_GPL(kmem_cache_name);
3463 * This initializes kmem_list3 or resizes varioius caches for all nodes.
3465 static int alloc_kmemlist(struct kmem_cache *cachep)
3468 struct kmem_list3 *l3;
3469 struct array_cache *new_shared;
3470 struct array_cache **new_alien;
3472 for_each_online_node(node) {
3474 new_alien = alloc_alien_cache(node, cachep->limit);
3478 new_shared = alloc_arraycache(node,
3479 cachep->shared*cachep->batchcount,
3482 free_alien_cache(new_alien);
3486 l3 = cachep->nodelists[node];
3488 struct array_cache *shared = l3->shared;
3490 spin_lock_irq(&l3->list_lock);
3493 free_block(cachep, shared->entry,
3494 shared->avail, node);
3496 l3->shared = new_shared;
3498 l3->alien = new_alien;
3501 l3->free_limit = (1 + nr_cpus_node(node)) *
3502 cachep->batchcount + cachep->num;
3503 spin_unlock_irq(&l3->list_lock);
3505 free_alien_cache(new_alien);
3508 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
3510 free_alien_cache(new_alien);
3515 kmem_list3_init(l3);
3516 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3517 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3518 l3->shared = new_shared;
3519 l3->alien = new_alien;
3520 l3->free_limit = (1 + nr_cpus_node(node)) *
3521 cachep->batchcount + cachep->num;
3522 cachep->nodelists[node] = l3;
3527 if (!cachep->next.next) {
3528 /* Cache is not active yet. Roll back what we did */
3531 if (cachep->nodelists[node]) {
3532 l3 = cachep->nodelists[node];
3535 free_alien_cache(l3->alien);
3537 cachep->nodelists[node] = NULL;
3545 struct ccupdate_struct {
3546 struct kmem_cache *cachep;
3547 struct array_cache *new[NR_CPUS];
3550 static void do_ccupdate_local(void *info)
3552 struct ccupdate_struct *new = info;
3553 struct array_cache *old;
3556 old = cpu_cache_get(new->cachep);
3558 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3559 new->new[smp_processor_id()] = old;
3562 /* Always called with the cache_chain_mutex held */
3563 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3564 int batchcount, int shared)
3566 struct ccupdate_struct new;
3569 memset(&new.new, 0, sizeof(new.new));
3570 for_each_online_cpu(i) {
3571 new.new[i] = alloc_arraycache(cpu_to_node(i), limit,
3574 for (i--; i >= 0; i--)
3579 new.cachep = cachep;
3581 on_each_cpu(do_ccupdate_local, (void *)&new, 1, 1);
3584 cachep->batchcount = batchcount;
3585 cachep->limit = limit;
3586 cachep->shared = shared;
3588 for_each_online_cpu(i) {
3589 struct array_cache *ccold = new.new[i];
3592 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3593 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3594 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3598 err = alloc_kmemlist(cachep);
3600 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
3601 cachep->name, -err);
3607 /* Called with cache_chain_mutex held always */
3608 static void enable_cpucache(struct kmem_cache *cachep)
3614 * The head array serves three purposes:
3615 * - create a LIFO ordering, i.e. return objects that are cache-warm
3616 * - reduce the number of spinlock operations.
3617 * - reduce the number of linked list operations on the slab and
3618 * bufctl chains: array operations are cheaper.
3619 * The numbers are guessed, we should auto-tune as described by
3622 if (cachep->buffer_size > 131072)
3624 else if (cachep->buffer_size > PAGE_SIZE)
3626 else if (cachep->buffer_size > 1024)
3628 else if (cachep->buffer_size > 256)
3634 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3635 * allocation behaviour: Most allocs on one cpu, most free operations
3636 * on another cpu. For these cases, an efficient object passing between
3637 * cpus is necessary. This is provided by a shared array. The array
3638 * replaces Bonwick's magazine layer.
3639 * On uniprocessor, it's functionally equivalent (but less efficient)
3640 * to a larger limit. Thus disabled by default.
3644 if (cachep->buffer_size <= PAGE_SIZE)
3650 * With debugging enabled, large batchcount lead to excessively long
3651 * periods with disabled local interrupts. Limit the batchcount
3656 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
3658 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3659 cachep->name, -err);
3663 * Drain an array if it contains any elements taking the l3 lock only if
3664 * necessary. Note that the l3 listlock also protects the array_cache
3665 * if drain_array() is used on the shared array.
3667 void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
3668 struct array_cache *ac, int force, int node)
3672 if (!ac || !ac->avail)
3674 if (ac->touched && !force) {
3677 spin_lock_irq(&l3->list_lock);
3679 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3680 if (tofree > ac->avail)
3681 tofree = (ac->avail + 1) / 2;
3682 free_block(cachep, ac->entry, tofree, node);
3683 ac->avail -= tofree;
3684 memmove(ac->entry, &(ac->entry[tofree]),
3685 sizeof(void *) * ac->avail);
3687 spin_unlock_irq(&l3->list_lock);
3692 * cache_reap - Reclaim memory from caches.
3693 * @unused: unused parameter
3695 * Called from workqueue/eventd every few seconds.
3697 * - clear the per-cpu caches for this CPU.
3698 * - return freeable pages to the main free memory pool.
3700 * If we cannot acquire the cache chain mutex then just give up - we'll try
3701 * again on the next iteration.
3703 static void cache_reap(void *unused)
3705 struct kmem_cache *searchp;
3706 struct kmem_list3 *l3;
3707 int node = numa_node_id();
3709 if (!mutex_trylock(&cache_chain_mutex)) {
3710 /* Give up. Setup the next iteration. */
3711 schedule_delayed_work(&__get_cpu_var(reap_work),
3716 list_for_each_entry(searchp, &cache_chain, next) {
3717 struct list_head *p;
3724 * We only take the l3 lock if absolutely necessary and we
3725 * have established with reasonable certainty that
3726 * we can do some work if the lock was obtained.
3728 l3 = searchp->nodelists[node];
3730 reap_alien(searchp, l3);
3732 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
3735 * These are racy checks but it does not matter
3736 * if we skip one check or scan twice.
3738 if (time_after(l3->next_reap, jiffies))
3741 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
3743 drain_array(searchp, l3, l3->shared, 0, node);
3745 if (l3->free_touched) {
3746 l3->free_touched = 0;
3750 tofree = (l3->free_limit + 5 * searchp->num - 1) /
3754 * Do not lock if there are no free blocks.
3756 if (list_empty(&l3->slabs_free))
3759 spin_lock_irq(&l3->list_lock);
3760 p = l3->slabs_free.next;
3761 if (p == &(l3->slabs_free)) {
3762 spin_unlock_irq(&l3->list_lock);
3766 slabp = list_entry(p, struct slab, list);
3767 BUG_ON(slabp->inuse);
3768 list_del(&slabp->list);
3769 STATS_INC_REAPED(searchp);
3772 * Safe to drop the lock. The slab is no longer linked
3773 * to the cache. searchp cannot disappear, we hold
3776 l3->free_objects -= searchp->num;
3777 spin_unlock_irq(&l3->list_lock);
3778 slab_destroy(searchp, slabp);
3779 } while (--tofree > 0);
3784 mutex_unlock(&cache_chain_mutex);
3786 /* Set up the next iteration */
3787 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
3790 #ifdef CONFIG_PROC_FS
3792 static void print_slabinfo_header(struct seq_file *m)
3795 * Output format version, so at least we can change it
3796 * without _too_ many complaints.
3799 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3801 seq_puts(m, "slabinfo - version: 2.1\n");
3803 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3804 "<objperslab> <pagesperslab>");
3805 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3806 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3808 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3809 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
3810 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3815 static void *s_start(struct seq_file *m, loff_t *pos)
3818 struct list_head *p;
3820 mutex_lock(&cache_chain_mutex);
3822 print_slabinfo_header(m);
3823 p = cache_chain.next;
3826 if (p == &cache_chain)
3829 return list_entry(p, struct kmem_cache, next);
3832 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3834 struct kmem_cache *cachep = p;
3836 return cachep->next.next == &cache_chain ?
3837 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
3840 static void s_stop(struct seq_file *m, void *p)
3842 mutex_unlock(&cache_chain_mutex);
3845 static int s_show(struct seq_file *m, void *p)
3847 struct kmem_cache *cachep = p;
3849 unsigned long active_objs;
3850 unsigned long num_objs;
3851 unsigned long active_slabs = 0;
3852 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3856 struct kmem_list3 *l3;
3860 for_each_online_node(node) {
3861 l3 = cachep->nodelists[node];
3866 spin_lock_irq(&l3->list_lock);
3868 list_for_each_entry(slabp, &l3->slabs_full, list) {
3869 if (slabp->inuse != cachep->num && !error)
3870 error = "slabs_full accounting error";
3871 active_objs += cachep->num;
3874 list_for_each_entry(slabp, &l3->slabs_partial, list) {
3875 if (slabp->inuse == cachep->num && !error)
3876 error = "slabs_partial inuse accounting error";
3877 if (!slabp->inuse && !error)
3878 error = "slabs_partial/inuse accounting error";
3879 active_objs += slabp->inuse;
3882 list_for_each_entry(slabp, &l3->slabs_free, list) {
3883 if (slabp->inuse && !error)
3884 error = "slabs_free/inuse accounting error";
3887 free_objects += l3->free_objects;
3889 shared_avail += l3->shared->avail;
3891 spin_unlock_irq(&l3->list_lock);
3893 num_slabs += active_slabs;
3894 num_objs = num_slabs * cachep->num;
3895 if (num_objs - active_objs != free_objects && !error)
3896 error = "free_objects accounting error";
3898 name = cachep->name;
3900 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3902 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
3903 name, active_objs, num_objs, cachep->buffer_size,
3904 cachep->num, (1 << cachep->gfporder));
3905 seq_printf(m, " : tunables %4u %4u %4u",
3906 cachep->limit, cachep->batchcount, cachep->shared);
3907 seq_printf(m, " : slabdata %6lu %6lu %6lu",
3908 active_slabs, num_slabs, shared_avail);
3911 unsigned long high = cachep->high_mark;
3912 unsigned long allocs = cachep->num_allocations;
3913 unsigned long grown = cachep->grown;
3914 unsigned long reaped = cachep->reaped;
3915 unsigned long errors = cachep->errors;
3916 unsigned long max_freeable = cachep->max_freeable;
3917 unsigned long node_allocs = cachep->node_allocs;
3918 unsigned long node_frees = cachep->node_frees;
3919 unsigned long overflows = cachep->node_overflow;
3921 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
3922 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
3923 reaped, errors, max_freeable, node_allocs,
3924 node_frees, overflows);
3928 unsigned long allochit = atomic_read(&cachep->allochit);
3929 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3930 unsigned long freehit = atomic_read(&cachep->freehit);
3931 unsigned long freemiss = atomic_read(&cachep->freemiss);
3933 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
3934 allochit, allocmiss, freehit, freemiss);
3942 * slabinfo_op - iterator that generates /proc/slabinfo
3951 * num-pages-per-slab
3952 * + further values on SMP and with statistics enabled
3955 struct seq_operations slabinfo_op = {
3962 #define MAX_SLABINFO_WRITE 128
3964 * slabinfo_write - Tuning for the slab allocator
3966 * @buffer: user buffer
3967 * @count: data length
3970 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
3971 size_t count, loff_t *ppos)
3973 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
3974 int limit, batchcount, shared, res;
3975 struct kmem_cache *cachep;
3977 if (count > MAX_SLABINFO_WRITE)
3979 if (copy_from_user(&kbuf, buffer, count))
3981 kbuf[MAX_SLABINFO_WRITE] = '\0';
3983 tmp = strchr(kbuf, ' ');
3988 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3991 /* Find the cache in the chain of caches. */
3992 mutex_lock(&cache_chain_mutex);
3994 list_for_each_entry(cachep, &cache_chain, next) {
3995 if (!strcmp(cachep->name, kbuf)) {
3996 if (limit < 1 || batchcount < 1 ||
3997 batchcount > limit || shared < 0) {
4000 res = do_tune_cpucache(cachep, limit,
4001 batchcount, shared);
4006 mutex_unlock(&cache_chain_mutex);
4012 #ifdef CONFIG_DEBUG_SLAB_LEAK
4014 static void *leaks_start(struct seq_file *m, loff_t *pos)
4017 struct list_head *p;
4019 mutex_lock(&cache_chain_mutex);
4020 p = cache_chain.next;
4023 if (p == &cache_chain)
4026 return list_entry(p, struct kmem_cache, next);
4029 static inline int add_caller(unsigned long *n, unsigned long v)
4039 unsigned long *q = p + 2 * i;
4053 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4059 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4065 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4066 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4068 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4073 static void show_symbol(struct seq_file *m, unsigned long address)
4075 #ifdef CONFIG_KALLSYMS
4078 unsigned long offset, size;
4079 char namebuf[KSYM_NAME_LEN+1];
4081 name = kallsyms_lookup(address, &size, &offset, &modname, namebuf);
4084 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4086 seq_printf(m, " [%s]", modname);
4090 seq_printf(m, "%p", (void *)address);
4093 static int leaks_show(struct seq_file *m, void *p)
4095 struct kmem_cache *cachep = p;
4097 struct kmem_list3 *l3;
4099 unsigned long *n = m->private;
4103 if (!(cachep->flags & SLAB_STORE_USER))
4105 if (!(cachep->flags & SLAB_RED_ZONE))
4108 /* OK, we can do it */
4112 for_each_online_node(node) {
4113 l3 = cachep->nodelists[node];
4118 spin_lock_irq(&l3->list_lock);
4120 list_for_each_entry(slabp, &l3->slabs_full, list)
4121 handle_slab(n, cachep, slabp);
4122 list_for_each_entry(slabp, &l3->slabs_partial, list)
4123 handle_slab(n, cachep, slabp);
4124 spin_unlock_irq(&l3->list_lock);
4126 name = cachep->name;
4128 /* Increase the buffer size */
4129 mutex_unlock(&cache_chain_mutex);
4130 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4132 /* Too bad, we are really out */
4134 mutex_lock(&cache_chain_mutex);
4137 *(unsigned long *)m->private = n[0] * 2;
4139 mutex_lock(&cache_chain_mutex);
4140 /* Now make sure this entry will be retried */
4144 for (i = 0; i < n[1]; i++) {
4145 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4146 show_symbol(m, n[2*i+2]);
4152 struct seq_operations slabstats_op = {
4153 .start = leaks_start,
4162 * ksize - get the actual amount of memory allocated for a given object
4163 * @objp: Pointer to the object
4165 * kmalloc may internally round up allocations and return more memory
4166 * than requested. ksize() can be used to determine the actual amount of
4167 * memory allocated. The caller may use this additional memory, even though
4168 * a smaller amount of memory was initially specified with the kmalloc call.
4169 * The caller must guarantee that objp points to a valid object previously
4170 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4171 * must not be freed during the duration of the call.
4173 unsigned int ksize(const void *objp)
4175 if (unlikely(objp == NULL))
4178 return obj_size(virt_to_cache(objp));