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 BUG_ON(!PageSlab(page));
596 return (struct kmem_cache *)page->lru.next;
599 static inline void page_set_slab(struct page *page, struct slab *slab)
601 page->lru.prev = (struct list_head *)slab;
604 static inline struct slab *page_get_slab(struct page *page)
606 if (unlikely(PageCompound(page)))
607 page = (struct page *)page_private(page);
608 BUG_ON(!PageSlab(page));
609 return (struct slab *)page->lru.prev;
612 static inline struct kmem_cache *virt_to_cache(const void *obj)
614 struct page *page = virt_to_page(obj);
615 return page_get_cache(page);
618 static inline struct slab *virt_to_slab(const void *obj)
620 struct page *page = virt_to_page(obj);
621 return page_get_slab(page);
624 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
627 return slab->s_mem + cache->buffer_size * idx;
630 static inline unsigned int obj_to_index(struct kmem_cache *cache,
631 struct slab *slab, void *obj)
633 return (unsigned)(obj - slab->s_mem) / cache->buffer_size;
637 * These are the default caches for kmalloc. Custom caches can have other sizes.
639 struct cache_sizes malloc_sizes[] = {
640 #define CACHE(x) { .cs_size = (x) },
641 #include <linux/kmalloc_sizes.h>
645 EXPORT_SYMBOL(malloc_sizes);
647 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
653 static struct cache_names __initdata cache_names[] = {
654 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
655 #include <linux/kmalloc_sizes.h>
660 static struct arraycache_init initarray_cache __initdata =
661 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
662 static struct arraycache_init initarray_generic =
663 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
665 /* internal cache of cache description objs */
666 static struct kmem_cache cache_cache = {
668 .limit = BOOT_CPUCACHE_ENTRIES,
670 .buffer_size = sizeof(struct kmem_cache),
671 .name = "kmem_cache",
673 .obj_size = sizeof(struct kmem_cache),
677 /* Guard access to the cache-chain. */
678 static DEFINE_MUTEX(cache_chain_mutex);
679 static struct list_head cache_chain;
682 * vm_enough_memory() looks at this to determine how many slab-allocated pages
683 * are possibly freeable under pressure
685 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
687 atomic_t slab_reclaim_pages;
690 * chicken and egg problem: delay the per-cpu array allocation
691 * until the general caches are up.
701 * used by boot code to determine if it can use slab based allocator
703 int slab_is_available(void)
705 return g_cpucache_up == FULL;
708 static DEFINE_PER_CPU(struct work_struct, reap_work);
710 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
712 static void enable_cpucache(struct kmem_cache *cachep);
713 static void cache_reap(void *unused);
714 static int __node_shrink(struct kmem_cache *cachep, int node);
716 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
718 return cachep->array[smp_processor_id()];
721 static inline struct kmem_cache *__find_general_cachep(size_t size,
724 struct cache_sizes *csizep = malloc_sizes;
727 /* This happens if someone tries to call
728 * kmem_cache_create(), or __kmalloc(), before
729 * the generic caches are initialized.
731 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
733 while (size > csizep->cs_size)
737 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
738 * has cs_{dma,}cachep==NULL. Thus no special case
739 * for large kmalloc calls required.
741 if (unlikely(gfpflags & GFP_DMA))
742 return csizep->cs_dmacachep;
743 return csizep->cs_cachep;
746 struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
748 return __find_general_cachep(size, gfpflags);
750 EXPORT_SYMBOL(kmem_find_general_cachep);
752 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
754 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
758 * Calculate the number of objects and left-over bytes for a given buffer size.
760 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
761 size_t align, int flags, size_t *left_over,
766 size_t slab_size = PAGE_SIZE << gfporder;
769 * The slab management structure can be either off the slab or
770 * on it. For the latter case, the memory allocated for a
774 * - One kmem_bufctl_t for each object
775 * - Padding to respect alignment of @align
776 * - @buffer_size bytes for each object
778 * If the slab management structure is off the slab, then the
779 * alignment will already be calculated into the size. Because
780 * the slabs are all pages aligned, the objects will be at the
781 * correct alignment when allocated.
783 if (flags & CFLGS_OFF_SLAB) {
785 nr_objs = slab_size / buffer_size;
787 if (nr_objs > SLAB_LIMIT)
788 nr_objs = SLAB_LIMIT;
791 * Ignore padding for the initial guess. The padding
792 * is at most @align-1 bytes, and @buffer_size is at
793 * least @align. In the worst case, this result will
794 * be one greater than the number of objects that fit
795 * into the memory allocation when taking the padding
798 nr_objs = (slab_size - sizeof(struct slab)) /
799 (buffer_size + sizeof(kmem_bufctl_t));
802 * This calculated number will be either the right
803 * amount, or one greater than what we want.
805 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
809 if (nr_objs > SLAB_LIMIT)
810 nr_objs = SLAB_LIMIT;
812 mgmt_size = slab_mgmt_size(nr_objs, align);
815 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
818 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
820 static void __slab_error(const char *function, struct kmem_cache *cachep,
823 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
824 function, cachep->name, msg);
830 * Special reaping functions for NUMA systems called from cache_reap().
831 * These take care of doing round robin flushing of alien caches (containing
832 * objects freed on different nodes from which they were allocated) and the
833 * flushing of remote pcps by calling drain_node_pages.
835 static DEFINE_PER_CPU(unsigned long, reap_node);
837 static void init_reap_node(int cpu)
841 node = next_node(cpu_to_node(cpu), node_online_map);
842 if (node == MAX_NUMNODES)
843 node = first_node(node_online_map);
845 __get_cpu_var(reap_node) = node;
848 static void next_reap_node(void)
850 int node = __get_cpu_var(reap_node);
853 * Also drain per cpu pages on remote zones
855 if (node != numa_node_id())
856 drain_node_pages(node);
858 node = next_node(node, node_online_map);
859 if (unlikely(node >= MAX_NUMNODES))
860 node = first_node(node_online_map);
861 __get_cpu_var(reap_node) = node;
865 #define init_reap_node(cpu) do { } while (0)
866 #define next_reap_node(void) do { } while (0)
870 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
871 * via the workqueue/eventd.
872 * Add the CPU number into the expiration time to minimize the possibility of
873 * the CPUs getting into lockstep and contending for the global cache chain
876 static void __devinit start_cpu_timer(int cpu)
878 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
881 * When this gets called from do_initcalls via cpucache_init(),
882 * init_workqueues() has already run, so keventd will be setup
885 if (keventd_up() && reap_work->func == NULL) {
887 INIT_WORK(reap_work, cache_reap, NULL);
888 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
892 static struct array_cache *alloc_arraycache(int node, int entries,
895 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
896 struct array_cache *nc = NULL;
898 nc = kmalloc_node(memsize, GFP_KERNEL, node);
902 nc->batchcount = batchcount;
904 spin_lock_init(&nc->lock);
910 * Transfer objects in one arraycache to another.
911 * Locking must be handled by the caller.
913 * Return the number of entries transferred.
915 static int transfer_objects(struct array_cache *to,
916 struct array_cache *from, unsigned int max)
918 /* Figure out how many entries to transfer */
919 int nr = min(min(from->avail, max), to->limit - to->avail);
924 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
934 static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int);
935 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
937 static struct array_cache **alloc_alien_cache(int node, int limit)
939 struct array_cache **ac_ptr;
940 int memsize = sizeof(void *) * MAX_NUMNODES;
945 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
948 if (i == node || !node_online(i)) {
952 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
954 for (i--; i <= 0; i--)
964 static void free_alien_cache(struct array_cache **ac_ptr)
975 static void __drain_alien_cache(struct kmem_cache *cachep,
976 struct array_cache *ac, int node)
978 struct kmem_list3 *rl3 = cachep->nodelists[node];
981 spin_lock(&rl3->list_lock);
983 * Stuff objects into the remote nodes shared array first.
984 * That way we could avoid the overhead of putting the objects
985 * into the free lists and getting them back later.
988 transfer_objects(rl3->shared, ac, ac->limit);
990 free_block(cachep, ac->entry, ac->avail, node);
992 spin_unlock(&rl3->list_lock);
997 * Called from cache_reap() to regularly drain alien caches round robin.
999 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1001 int node = __get_cpu_var(reap_node);
1004 struct array_cache *ac = l3->alien[node];
1006 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1007 __drain_alien_cache(cachep, ac, node);
1008 spin_unlock_irq(&ac->lock);
1013 static void drain_alien_cache(struct kmem_cache *cachep,
1014 struct array_cache **alien)
1017 struct array_cache *ac;
1018 unsigned long flags;
1020 for_each_online_node(i) {
1023 spin_lock_irqsave(&ac->lock, flags);
1024 __drain_alien_cache(cachep, ac, i);
1025 spin_unlock_irqrestore(&ac->lock, flags);
1030 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1032 struct slab *slabp = virt_to_slab(objp);
1033 int nodeid = slabp->nodeid;
1034 struct kmem_list3 *l3;
1035 struct array_cache *alien = NULL;
1038 * Make sure we are not freeing a object from another node to the array
1039 * cache on this cpu.
1041 if (likely(slabp->nodeid == numa_node_id()))
1044 l3 = cachep->nodelists[numa_node_id()];
1045 STATS_INC_NODEFREES(cachep);
1046 if (l3->alien && l3->alien[nodeid]) {
1047 alien = l3->alien[nodeid];
1048 spin_lock(&alien->lock);
1049 if (unlikely(alien->avail == alien->limit)) {
1050 STATS_INC_ACOVERFLOW(cachep);
1051 __drain_alien_cache(cachep, alien, nodeid);
1053 alien->entry[alien->avail++] = objp;
1054 spin_unlock(&alien->lock);
1056 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1057 free_block(cachep, &objp, 1, nodeid);
1058 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1065 #define drain_alien_cache(cachep, alien) do { } while (0)
1066 #define reap_alien(cachep, l3) do { } while (0)
1068 static inline struct array_cache **alloc_alien_cache(int node, int limit)
1070 return (struct array_cache **) 0x01020304ul;
1073 static inline void free_alien_cache(struct array_cache **ac_ptr)
1077 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1084 static int cpuup_callback(struct notifier_block *nfb,
1085 unsigned long action, void *hcpu)
1087 long cpu = (long)hcpu;
1088 struct kmem_cache *cachep;
1089 struct kmem_list3 *l3 = NULL;
1090 int node = cpu_to_node(cpu);
1091 int memsize = sizeof(struct kmem_list3);
1094 case CPU_UP_PREPARE:
1095 mutex_lock(&cache_chain_mutex);
1097 * We need to do this right in the beginning since
1098 * alloc_arraycache's are going to use this list.
1099 * kmalloc_node allows us to add the slab to the right
1100 * kmem_list3 and not this cpu's kmem_list3
1103 list_for_each_entry(cachep, &cache_chain, next) {
1105 * Set up the size64 kmemlist for cpu before we can
1106 * begin anything. Make sure some other cpu on this
1107 * node has not already allocated this
1109 if (!cachep->nodelists[node]) {
1110 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1113 kmem_list3_init(l3);
1114 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1115 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1118 * The l3s don't come and go as CPUs come and
1119 * go. cache_chain_mutex is sufficient
1122 cachep->nodelists[node] = l3;
1125 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1126 cachep->nodelists[node]->free_limit =
1127 (1 + nr_cpus_node(node)) *
1128 cachep->batchcount + cachep->num;
1129 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1133 * Now we can go ahead with allocating the shared arrays and
1136 list_for_each_entry(cachep, &cache_chain, next) {
1137 struct array_cache *nc;
1138 struct array_cache *shared;
1139 struct array_cache **alien;
1141 nc = alloc_arraycache(node, cachep->limit,
1142 cachep->batchcount);
1145 shared = alloc_arraycache(node,
1146 cachep->shared * cachep->batchcount,
1151 alien = alloc_alien_cache(node, cachep->limit);
1154 cachep->array[cpu] = nc;
1155 l3 = cachep->nodelists[node];
1158 spin_lock_irq(&l3->list_lock);
1161 * We are serialised from CPU_DEAD or
1162 * CPU_UP_CANCELLED by the cpucontrol lock
1164 l3->shared = shared;
1173 spin_unlock_irq(&l3->list_lock);
1175 free_alien_cache(alien);
1177 mutex_unlock(&cache_chain_mutex);
1180 start_cpu_timer(cpu);
1182 #ifdef CONFIG_HOTPLUG_CPU
1185 * Even if all the cpus of a node are down, we don't free the
1186 * kmem_list3 of any cache. This to avoid a race between
1187 * cpu_down, and a kmalloc allocation from another cpu for
1188 * memory from the node of the cpu going down. The list3
1189 * structure is usually allocated from kmem_cache_create() and
1190 * gets destroyed at kmem_cache_destroy().
1193 case CPU_UP_CANCELED:
1194 mutex_lock(&cache_chain_mutex);
1195 list_for_each_entry(cachep, &cache_chain, next) {
1196 struct array_cache *nc;
1197 struct array_cache *shared;
1198 struct array_cache **alien;
1201 mask = node_to_cpumask(node);
1202 /* cpu is dead; no one can alloc from it. */
1203 nc = cachep->array[cpu];
1204 cachep->array[cpu] = NULL;
1205 l3 = cachep->nodelists[node];
1208 goto free_array_cache;
1210 spin_lock_irq(&l3->list_lock);
1212 /* Free limit for this kmem_list3 */
1213 l3->free_limit -= cachep->batchcount;
1215 free_block(cachep, nc->entry, nc->avail, node);
1217 if (!cpus_empty(mask)) {
1218 spin_unlock_irq(&l3->list_lock);
1219 goto free_array_cache;
1222 shared = l3->shared;
1224 free_block(cachep, l3->shared->entry,
1225 l3->shared->avail, node);
1232 spin_unlock_irq(&l3->list_lock);
1236 drain_alien_cache(cachep, alien);
1237 free_alien_cache(alien);
1243 * In the previous loop, all the objects were freed to
1244 * the respective cache's slabs, now we can go ahead and
1245 * shrink each nodelist to its limit.
1247 list_for_each_entry(cachep, &cache_chain, next) {
1248 l3 = cachep->nodelists[node];
1251 spin_lock_irq(&l3->list_lock);
1252 /* free slabs belonging to this node */
1253 __node_shrink(cachep, node);
1254 spin_unlock_irq(&l3->list_lock);
1256 mutex_unlock(&cache_chain_mutex);
1262 mutex_unlock(&cache_chain_mutex);
1266 static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
1269 * swap the static kmem_list3 with kmalloced memory
1271 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1274 struct kmem_list3 *ptr;
1276 BUG_ON(cachep->nodelists[nodeid] != list);
1277 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1280 local_irq_disable();
1281 memcpy(ptr, list, sizeof(struct kmem_list3));
1282 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1283 cachep->nodelists[nodeid] = ptr;
1288 * Initialisation. Called after the page allocator have been initialised and
1289 * before smp_init().
1291 void __init kmem_cache_init(void)
1294 struct cache_sizes *sizes;
1295 struct cache_names *names;
1299 for (i = 0; i < NUM_INIT_LISTS; i++) {
1300 kmem_list3_init(&initkmem_list3[i]);
1301 if (i < MAX_NUMNODES)
1302 cache_cache.nodelists[i] = NULL;
1306 * Fragmentation resistance on low memory - only use bigger
1307 * page orders on machines with more than 32MB of memory.
1309 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1310 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1312 /* Bootstrap is tricky, because several objects are allocated
1313 * from caches that do not exist yet:
1314 * 1) initialize the cache_cache cache: it contains the struct
1315 * kmem_cache structures of all caches, except cache_cache itself:
1316 * cache_cache is statically allocated.
1317 * Initially an __init data area is used for the head array and the
1318 * kmem_list3 structures, it's replaced with a kmalloc allocated
1319 * array at the end of the bootstrap.
1320 * 2) Create the first kmalloc cache.
1321 * The struct kmem_cache for the new cache is allocated normally.
1322 * An __init data area is used for the head array.
1323 * 3) Create the remaining kmalloc caches, with minimally sized
1325 * 4) Replace the __init data head arrays for cache_cache and the first
1326 * kmalloc cache with kmalloc allocated arrays.
1327 * 5) Replace the __init data for kmem_list3 for cache_cache and
1328 * the other cache's with kmalloc allocated memory.
1329 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1332 /* 1) create the cache_cache */
1333 INIT_LIST_HEAD(&cache_chain);
1334 list_add(&cache_cache.next, &cache_chain);
1335 cache_cache.colour_off = cache_line_size();
1336 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1337 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
1339 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1342 for (order = 0; order < MAX_ORDER; order++) {
1343 cache_estimate(order, cache_cache.buffer_size,
1344 cache_line_size(), 0, &left_over, &cache_cache.num);
1345 if (cache_cache.num)
1348 BUG_ON(!cache_cache.num);
1349 cache_cache.gfporder = order;
1350 cache_cache.colour = left_over / cache_cache.colour_off;
1351 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1352 sizeof(struct slab), cache_line_size());
1354 /* 2+3) create the kmalloc caches */
1355 sizes = malloc_sizes;
1356 names = cache_names;
1359 * Initialize the caches that provide memory for the array cache and the
1360 * kmem_list3 structures first. Without this, further allocations will
1364 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1365 sizes[INDEX_AC].cs_size,
1366 ARCH_KMALLOC_MINALIGN,
1367 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1370 if (INDEX_AC != INDEX_L3) {
1371 sizes[INDEX_L3].cs_cachep =
1372 kmem_cache_create(names[INDEX_L3].name,
1373 sizes[INDEX_L3].cs_size,
1374 ARCH_KMALLOC_MINALIGN,
1375 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1379 while (sizes->cs_size != ULONG_MAX) {
1381 * For performance, all the general caches are L1 aligned.
1382 * This should be particularly beneficial on SMP boxes, as it
1383 * eliminates "false sharing".
1384 * Note for systems short on memory removing the alignment will
1385 * allow tighter packing of the smaller caches.
1387 if (!sizes->cs_cachep) {
1388 sizes->cs_cachep = kmem_cache_create(names->name,
1390 ARCH_KMALLOC_MINALIGN,
1391 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1395 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
1397 ARCH_KMALLOC_MINALIGN,
1398 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1404 /* 4) Replace the bootstrap head arrays */
1408 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1410 local_irq_disable();
1411 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1412 memcpy(ptr, cpu_cache_get(&cache_cache),
1413 sizeof(struct arraycache_init));
1414 cache_cache.array[smp_processor_id()] = ptr;
1417 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1419 local_irq_disable();
1420 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1421 != &initarray_generic.cache);
1422 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1423 sizeof(struct arraycache_init));
1424 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1428 /* 5) Replace the bootstrap kmem_list3's */
1431 /* Replace the static kmem_list3 structures for the boot cpu */
1432 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
1435 for_each_online_node(node) {
1436 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1437 &initkmem_list3[SIZE_AC + node], node);
1439 if (INDEX_AC != INDEX_L3) {
1440 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1441 &initkmem_list3[SIZE_L3 + node],
1447 /* 6) resize the head arrays to their final sizes */
1449 struct kmem_cache *cachep;
1450 mutex_lock(&cache_chain_mutex);
1451 list_for_each_entry(cachep, &cache_chain, next)
1452 enable_cpucache(cachep);
1453 mutex_unlock(&cache_chain_mutex);
1457 g_cpucache_up = FULL;
1460 * Register a cpu startup notifier callback that initializes
1461 * cpu_cache_get for all new cpus
1463 register_cpu_notifier(&cpucache_notifier);
1466 * The reap timers are started later, with a module init call: That part
1467 * of the kernel is not yet operational.
1471 static int __init cpucache_init(void)
1476 * Register the timers that return unneeded pages to the page allocator
1478 for_each_online_cpu(cpu)
1479 start_cpu_timer(cpu);
1482 __initcall(cpucache_init);
1485 * Interface to system's page allocator. No need to hold the cache-lock.
1487 * If we requested dmaable memory, we will get it. Even if we
1488 * did not request dmaable memory, we might get it, but that
1489 * would be relatively rare and ignorable.
1491 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1499 * Nommu uses slab's for process anonymous memory allocations, and thus
1500 * requires __GFP_COMP to properly refcount higher order allocations
1502 flags |= __GFP_COMP;
1504 flags |= cachep->gfpflags;
1506 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1510 nr_pages = (1 << cachep->gfporder);
1511 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1512 atomic_add(nr_pages, &slab_reclaim_pages);
1513 add_page_state(nr_slab, nr_pages);
1514 for (i = 0; i < nr_pages; i++)
1515 __SetPageSlab(page + i);
1516 return page_address(page);
1520 * Interface to system's page release.
1522 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1524 unsigned long i = (1 << cachep->gfporder);
1525 struct page *page = virt_to_page(addr);
1526 const unsigned long nr_freed = i;
1529 BUG_ON(!PageSlab(page));
1530 __ClearPageSlab(page);
1533 sub_page_state(nr_slab, nr_freed);
1534 if (current->reclaim_state)
1535 current->reclaim_state->reclaimed_slab += nr_freed;
1536 free_pages((unsigned long)addr, cachep->gfporder);
1537 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1538 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages);
1541 static void kmem_rcu_free(struct rcu_head *head)
1543 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1544 struct kmem_cache *cachep = slab_rcu->cachep;
1546 kmem_freepages(cachep, slab_rcu->addr);
1547 if (OFF_SLAB(cachep))
1548 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1553 #ifdef CONFIG_DEBUG_PAGEALLOC
1554 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1555 unsigned long caller)
1557 int size = obj_size(cachep);
1559 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1561 if (size < 5 * sizeof(unsigned long))
1564 *addr++ = 0x12345678;
1566 *addr++ = smp_processor_id();
1567 size -= 3 * sizeof(unsigned long);
1569 unsigned long *sptr = &caller;
1570 unsigned long svalue;
1572 while (!kstack_end(sptr)) {
1574 if (kernel_text_address(svalue)) {
1576 size -= sizeof(unsigned long);
1577 if (size <= sizeof(unsigned long))
1583 *addr++ = 0x87654321;
1587 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1589 int size = obj_size(cachep);
1590 addr = &((char *)addr)[obj_offset(cachep)];
1592 memset(addr, val, size);
1593 *(unsigned char *)(addr + size - 1) = POISON_END;
1596 static void dump_line(char *data, int offset, int limit)
1599 printk(KERN_ERR "%03x:", offset);
1600 for (i = 0; i < limit; i++)
1601 printk(" %02x", (unsigned char)data[offset + i]);
1608 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1613 if (cachep->flags & SLAB_RED_ZONE) {
1614 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
1615 *dbg_redzone1(cachep, objp),
1616 *dbg_redzone2(cachep, objp));
1619 if (cachep->flags & SLAB_STORE_USER) {
1620 printk(KERN_ERR "Last user: [<%p>]",
1621 *dbg_userword(cachep, objp));
1622 print_symbol("(%s)",
1623 (unsigned long)*dbg_userword(cachep, objp));
1626 realobj = (char *)objp + obj_offset(cachep);
1627 size = obj_size(cachep);
1628 for (i = 0; i < size && lines; i += 16, lines--) {
1631 if (i + limit > size)
1633 dump_line(realobj, i, limit);
1637 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1643 realobj = (char *)objp + obj_offset(cachep);
1644 size = obj_size(cachep);
1646 for (i = 0; i < size; i++) {
1647 char exp = POISON_FREE;
1650 if (realobj[i] != exp) {
1656 "Slab corruption: start=%p, len=%d\n",
1658 print_objinfo(cachep, objp, 0);
1660 /* Hexdump the affected line */
1663 if (i + limit > size)
1665 dump_line(realobj, i, limit);
1668 /* Limit to 5 lines */
1674 /* Print some data about the neighboring objects, if they
1677 struct slab *slabp = virt_to_slab(objp);
1680 objnr = obj_to_index(cachep, slabp, objp);
1682 objp = index_to_obj(cachep, slabp, objnr - 1);
1683 realobj = (char *)objp + obj_offset(cachep);
1684 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1686 print_objinfo(cachep, objp, 2);
1688 if (objnr + 1 < cachep->num) {
1689 objp = index_to_obj(cachep, slabp, objnr + 1);
1690 realobj = (char *)objp + obj_offset(cachep);
1691 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1693 print_objinfo(cachep, objp, 2);
1701 * slab_destroy_objs - destroy a slab and its objects
1702 * @cachep: cache pointer being destroyed
1703 * @slabp: slab pointer being destroyed
1705 * Call the registered destructor for each object in a slab that is being
1708 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1711 for (i = 0; i < cachep->num; i++) {
1712 void *objp = index_to_obj(cachep, slabp, i);
1714 if (cachep->flags & SLAB_POISON) {
1715 #ifdef CONFIG_DEBUG_PAGEALLOC
1716 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1718 kernel_map_pages(virt_to_page(objp),
1719 cachep->buffer_size / PAGE_SIZE, 1);
1721 check_poison_obj(cachep, objp);
1723 check_poison_obj(cachep, objp);
1726 if (cachep->flags & SLAB_RED_ZONE) {
1727 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1728 slab_error(cachep, "start of a freed object "
1730 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1731 slab_error(cachep, "end of a freed object "
1734 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1735 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
1739 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1743 for (i = 0; i < cachep->num; i++) {
1744 void *objp = index_to_obj(cachep, slabp, i);
1745 (cachep->dtor) (objp, cachep, 0);
1752 * slab_destroy - destroy and release all objects in a slab
1753 * @cachep: cache pointer being destroyed
1754 * @slabp: slab pointer being destroyed
1756 * Destroy all the objs in a slab, and release the mem back to the system.
1757 * Before calling the slab must have been unlinked from the cache. The
1758 * cache-lock is not held/needed.
1760 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1762 void *addr = slabp->s_mem - slabp->colouroff;
1764 slab_destroy_objs(cachep, slabp);
1765 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1766 struct slab_rcu *slab_rcu;
1768 slab_rcu = (struct slab_rcu *)slabp;
1769 slab_rcu->cachep = cachep;
1770 slab_rcu->addr = addr;
1771 call_rcu(&slab_rcu->head, kmem_rcu_free);
1773 kmem_freepages(cachep, addr);
1774 if (OFF_SLAB(cachep))
1775 kmem_cache_free(cachep->slabp_cache, slabp);
1780 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1781 * size of kmem_list3.
1783 static void set_up_list3s(struct kmem_cache *cachep, int index)
1787 for_each_online_node(node) {
1788 cachep->nodelists[node] = &initkmem_list3[index + node];
1789 cachep->nodelists[node]->next_reap = jiffies +
1791 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1796 * calculate_slab_order - calculate size (page order) of slabs
1797 * @cachep: pointer to the cache that is being created
1798 * @size: size of objects to be created in this cache.
1799 * @align: required alignment for the objects.
1800 * @flags: slab allocation flags
1802 * Also calculates the number of objects per slab.
1804 * This could be made much more intelligent. For now, try to avoid using
1805 * high order pages for slabs. When the gfp() functions are more friendly
1806 * towards high-order requests, this should be changed.
1808 static size_t calculate_slab_order(struct kmem_cache *cachep,
1809 size_t size, size_t align, unsigned long flags)
1811 unsigned long offslab_limit;
1812 size_t left_over = 0;
1815 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
1819 cache_estimate(gfporder, size, align, flags, &remainder, &num);
1823 if (flags & CFLGS_OFF_SLAB) {
1825 * Max number of objs-per-slab for caches which
1826 * use off-slab slabs. Needed to avoid a possible
1827 * looping condition in cache_grow().
1829 offslab_limit = size - sizeof(struct slab);
1830 offslab_limit /= sizeof(kmem_bufctl_t);
1832 if (num > offslab_limit)
1836 /* Found something acceptable - save it away */
1838 cachep->gfporder = gfporder;
1839 left_over = remainder;
1842 * A VFS-reclaimable slab tends to have most allocations
1843 * as GFP_NOFS and we really don't want to have to be allocating
1844 * higher-order pages when we are unable to shrink dcache.
1846 if (flags & SLAB_RECLAIM_ACCOUNT)
1850 * Large number of objects is good, but very large slabs are
1851 * currently bad for the gfp()s.
1853 if (gfporder >= slab_break_gfp_order)
1857 * Acceptable internal fragmentation?
1859 if (left_over * 8 <= (PAGE_SIZE << gfporder))
1865 static void setup_cpu_cache(struct kmem_cache *cachep)
1867 if (g_cpucache_up == FULL) {
1868 enable_cpucache(cachep);
1871 if (g_cpucache_up == NONE) {
1873 * Note: the first kmem_cache_create must create the cache
1874 * that's used by kmalloc(24), otherwise the creation of
1875 * further caches will BUG().
1877 cachep->array[smp_processor_id()] = &initarray_generic.cache;
1880 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
1881 * the first cache, then we need to set up all its list3s,
1882 * otherwise the creation of further caches will BUG().
1884 set_up_list3s(cachep, SIZE_AC);
1885 if (INDEX_AC == INDEX_L3)
1886 g_cpucache_up = PARTIAL_L3;
1888 g_cpucache_up = PARTIAL_AC;
1890 cachep->array[smp_processor_id()] =
1891 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1893 if (g_cpucache_up == PARTIAL_AC) {
1894 set_up_list3s(cachep, SIZE_L3);
1895 g_cpucache_up = PARTIAL_L3;
1898 for_each_online_node(node) {
1899 cachep->nodelists[node] =
1900 kmalloc_node(sizeof(struct kmem_list3),
1902 BUG_ON(!cachep->nodelists[node]);
1903 kmem_list3_init(cachep->nodelists[node]);
1907 cachep->nodelists[numa_node_id()]->next_reap =
1908 jiffies + REAPTIMEOUT_LIST3 +
1909 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1911 cpu_cache_get(cachep)->avail = 0;
1912 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1913 cpu_cache_get(cachep)->batchcount = 1;
1914 cpu_cache_get(cachep)->touched = 0;
1915 cachep->batchcount = 1;
1916 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1920 * kmem_cache_create - Create a cache.
1921 * @name: A string which is used in /proc/slabinfo to identify this cache.
1922 * @size: The size of objects to be created in this cache.
1923 * @align: The required alignment for the objects.
1924 * @flags: SLAB flags
1925 * @ctor: A constructor for the objects.
1926 * @dtor: A destructor for the objects.
1928 * Returns a ptr to the cache on success, NULL on failure.
1929 * Cannot be called within a int, but can be interrupted.
1930 * The @ctor is run when new pages are allocated by the cache
1931 * and the @dtor is run before the pages are handed back.
1933 * @name must be valid until the cache is destroyed. This implies that
1934 * the module calling this has to destroy the cache before getting unloaded.
1938 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1939 * to catch references to uninitialised memory.
1941 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1942 * for buffer overruns.
1944 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1945 * cacheline. This can be beneficial if you're counting cycles as closely
1949 kmem_cache_create (const char *name, size_t size, size_t align,
1950 unsigned long flags,
1951 void (*ctor)(void*, struct kmem_cache *, unsigned long),
1952 void (*dtor)(void*, struct kmem_cache *, unsigned long))
1954 size_t left_over, slab_size, ralign;
1955 struct kmem_cache *cachep = NULL, *pc;
1958 * Sanity checks... these are all serious usage bugs.
1960 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
1961 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
1962 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
1968 * Prevent CPUs from coming and going.
1969 * lock_cpu_hotplug() nests outside cache_chain_mutex
1973 mutex_lock(&cache_chain_mutex);
1975 list_for_each_entry(pc, &cache_chain, next) {
1976 mm_segment_t old_fs = get_fs();
1981 * This happens when the module gets unloaded and doesn't
1982 * destroy its slab cache and no-one else reuses the vmalloc
1983 * area of the module. Print a warning.
1986 res = __get_user(tmp, pc->name);
1989 printk("SLAB: cache with size %d has lost its name\n",
1994 if (!strcmp(pc->name, name)) {
1995 printk("kmem_cache_create: duplicate cache %s\n", name);
2002 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2003 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
2004 /* No constructor, but inital state check requested */
2005 printk(KERN_ERR "%s: No con, but init state check "
2006 "requested - %s\n", __FUNCTION__, name);
2007 flags &= ~SLAB_DEBUG_INITIAL;
2011 * Enable redzoning and last user accounting, except for caches with
2012 * large objects, if the increased size would increase the object size
2013 * above the next power of two: caches with object sizes just above a
2014 * power of two have a significant amount of internal fragmentation.
2016 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
2017 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2018 if (!(flags & SLAB_DESTROY_BY_RCU))
2019 flags |= SLAB_POISON;
2021 if (flags & SLAB_DESTROY_BY_RCU)
2022 BUG_ON(flags & SLAB_POISON);
2024 if (flags & SLAB_DESTROY_BY_RCU)
2028 * Always checks flags, a caller might be expecting debug support which
2031 BUG_ON(flags & ~CREATE_MASK);
2034 * Check that size is in terms of words. This is needed to avoid
2035 * unaligned accesses for some archs when redzoning is used, and makes
2036 * sure any on-slab bufctl's are also correctly aligned.
2038 if (size & (BYTES_PER_WORD - 1)) {
2039 size += (BYTES_PER_WORD - 1);
2040 size &= ~(BYTES_PER_WORD - 1);
2043 /* calculate the final buffer alignment: */
2045 /* 1) arch recommendation: can be overridden for debug */
2046 if (flags & SLAB_HWCACHE_ALIGN) {
2048 * Default alignment: as specified by the arch code. Except if
2049 * an object is really small, then squeeze multiple objects into
2052 ralign = cache_line_size();
2053 while (size <= ralign / 2)
2056 ralign = BYTES_PER_WORD;
2058 /* 2) arch mandated alignment: disables debug if necessary */
2059 if (ralign < ARCH_SLAB_MINALIGN) {
2060 ralign = ARCH_SLAB_MINALIGN;
2061 if (ralign > BYTES_PER_WORD)
2062 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2064 /* 3) caller mandated alignment: disables debug if necessary */
2065 if (ralign < align) {
2067 if (ralign > BYTES_PER_WORD)
2068 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2071 * 4) Store it. Note that the debug code below can reduce
2072 * the alignment to BYTES_PER_WORD.
2076 /* Get cache's description obj. */
2077 cachep = kmem_cache_zalloc(&cache_cache, SLAB_KERNEL);
2082 cachep->obj_size = size;
2084 if (flags & SLAB_RED_ZONE) {
2085 /* redzoning only works with word aligned caches */
2086 align = BYTES_PER_WORD;
2088 /* add space for red zone words */
2089 cachep->obj_offset += BYTES_PER_WORD;
2090 size += 2 * BYTES_PER_WORD;
2092 if (flags & SLAB_STORE_USER) {
2093 /* user store requires word alignment and
2094 * one word storage behind the end of the real
2097 align = BYTES_PER_WORD;
2098 size += BYTES_PER_WORD;
2100 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2101 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2102 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2103 cachep->obj_offset += PAGE_SIZE - size;
2109 /* Determine if the slab management is 'on' or 'off' slab. */
2110 if (size >= (PAGE_SIZE >> 3))
2112 * Size is large, assume best to place the slab management obj
2113 * off-slab (should allow better packing of objs).
2115 flags |= CFLGS_OFF_SLAB;
2117 size = ALIGN(size, align);
2119 left_over = calculate_slab_order(cachep, size, align, flags);
2122 printk("kmem_cache_create: couldn't create cache %s.\n", name);
2123 kmem_cache_free(&cache_cache, cachep);
2127 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2128 + sizeof(struct slab), align);
2131 * If the slab has been placed off-slab, and we have enough space then
2132 * move it on-slab. This is at the expense of any extra colouring.
2134 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2135 flags &= ~CFLGS_OFF_SLAB;
2136 left_over -= slab_size;
2139 if (flags & CFLGS_OFF_SLAB) {
2140 /* really off slab. No need for manual alignment */
2142 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2145 cachep->colour_off = cache_line_size();
2146 /* Offset must be a multiple of the alignment. */
2147 if (cachep->colour_off < align)
2148 cachep->colour_off = align;
2149 cachep->colour = left_over / cachep->colour_off;
2150 cachep->slab_size = slab_size;
2151 cachep->flags = flags;
2152 cachep->gfpflags = 0;
2153 if (flags & SLAB_CACHE_DMA)
2154 cachep->gfpflags |= GFP_DMA;
2155 cachep->buffer_size = size;
2157 if (flags & CFLGS_OFF_SLAB)
2158 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2159 cachep->ctor = ctor;
2160 cachep->dtor = dtor;
2161 cachep->name = name;
2164 setup_cpu_cache(cachep);
2166 /* cache setup completed, link it into the list */
2167 list_add(&cachep->next, &cache_chain);
2169 if (!cachep && (flags & SLAB_PANIC))
2170 panic("kmem_cache_create(): failed to create slab `%s'\n",
2172 mutex_unlock(&cache_chain_mutex);
2173 unlock_cpu_hotplug();
2176 EXPORT_SYMBOL(kmem_cache_create);
2179 static void check_irq_off(void)
2181 BUG_ON(!irqs_disabled());
2184 static void check_irq_on(void)
2186 BUG_ON(irqs_disabled());
2189 static void check_spinlock_acquired(struct kmem_cache *cachep)
2193 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2197 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2201 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2206 #define check_irq_off() do { } while(0)
2207 #define check_irq_on() do { } while(0)
2208 #define check_spinlock_acquired(x) do { } while(0)
2209 #define check_spinlock_acquired_node(x, y) do { } while(0)
2212 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2213 struct array_cache *ac,
2214 int force, int node);
2216 static void do_drain(void *arg)
2218 struct kmem_cache *cachep = arg;
2219 struct array_cache *ac;
2220 int node = numa_node_id();
2223 ac = cpu_cache_get(cachep);
2224 spin_lock(&cachep->nodelists[node]->list_lock);
2225 free_block(cachep, ac->entry, ac->avail, node);
2226 spin_unlock(&cachep->nodelists[node]->list_lock);
2230 static void drain_cpu_caches(struct kmem_cache *cachep)
2232 struct kmem_list3 *l3;
2235 on_each_cpu(do_drain, cachep, 1, 1);
2237 for_each_online_node(node) {
2238 l3 = cachep->nodelists[node];
2239 if (l3 && l3->alien)
2240 drain_alien_cache(cachep, l3->alien);
2243 for_each_online_node(node) {
2244 l3 = cachep->nodelists[node];
2246 drain_array(cachep, l3, l3->shared, 1, node);
2250 static int __node_shrink(struct kmem_cache *cachep, int node)
2253 struct kmem_list3 *l3 = cachep->nodelists[node];
2257 struct list_head *p;
2259 p = l3->slabs_free.prev;
2260 if (p == &l3->slabs_free)
2263 slabp = list_entry(l3->slabs_free.prev, struct slab, list);
2265 BUG_ON(slabp->inuse);
2267 list_del(&slabp->list);
2269 l3->free_objects -= cachep->num;
2270 spin_unlock_irq(&l3->list_lock);
2271 slab_destroy(cachep, slabp);
2272 spin_lock_irq(&l3->list_lock);
2274 ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial);
2278 static int __cache_shrink(struct kmem_cache *cachep)
2281 struct kmem_list3 *l3;
2283 drain_cpu_caches(cachep);
2286 for_each_online_node(i) {
2287 l3 = cachep->nodelists[i];
2289 spin_lock_irq(&l3->list_lock);
2290 ret += __node_shrink(cachep, i);
2291 spin_unlock_irq(&l3->list_lock);
2294 return (ret ? 1 : 0);
2298 * kmem_cache_shrink - Shrink a cache.
2299 * @cachep: The cache to shrink.
2301 * Releases as many slabs as possible for a cache.
2302 * To help debugging, a zero exit status indicates all slabs were released.
2304 int kmem_cache_shrink(struct kmem_cache *cachep)
2306 BUG_ON(!cachep || in_interrupt());
2308 return __cache_shrink(cachep);
2310 EXPORT_SYMBOL(kmem_cache_shrink);
2313 * kmem_cache_destroy - delete a cache
2314 * @cachep: the cache to destroy
2316 * Remove a struct kmem_cache object from the slab cache.
2317 * Returns 0 on success.
2319 * It is expected this function will be called by a module when it is
2320 * unloaded. This will remove the cache completely, and avoid a duplicate
2321 * cache being allocated each time a module is loaded and unloaded, if the
2322 * module doesn't have persistent in-kernel storage across loads and unloads.
2324 * The cache must be empty before calling this function.
2326 * The caller must guarantee that noone will allocate memory from the cache
2327 * during the kmem_cache_destroy().
2329 int kmem_cache_destroy(struct kmem_cache *cachep)
2332 struct kmem_list3 *l3;
2334 BUG_ON(!cachep || in_interrupt());
2336 /* Don't let CPUs to come and go */
2339 /* Find the cache in the chain of caches. */
2340 mutex_lock(&cache_chain_mutex);
2342 * the chain is never empty, cache_cache is never destroyed
2344 list_del(&cachep->next);
2345 mutex_unlock(&cache_chain_mutex);
2347 if (__cache_shrink(cachep)) {
2348 slab_error(cachep, "Can't free all objects");
2349 mutex_lock(&cache_chain_mutex);
2350 list_add(&cachep->next, &cache_chain);
2351 mutex_unlock(&cache_chain_mutex);
2352 unlock_cpu_hotplug();
2356 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2359 for_each_online_cpu(i)
2360 kfree(cachep->array[i]);
2362 /* NUMA: free the list3 structures */
2363 for_each_online_node(i) {
2364 l3 = cachep->nodelists[i];
2367 free_alien_cache(l3->alien);
2371 kmem_cache_free(&cache_cache, cachep);
2372 unlock_cpu_hotplug();
2375 EXPORT_SYMBOL(kmem_cache_destroy);
2377 /* Get the memory for a slab management obj. */
2378 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2379 int colour_off, gfp_t local_flags,
2384 if (OFF_SLAB(cachep)) {
2385 /* Slab management obj is off-slab. */
2386 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2387 local_flags, nodeid);
2391 slabp = objp + colour_off;
2392 colour_off += cachep->slab_size;
2395 slabp->colouroff = colour_off;
2396 slabp->s_mem = objp + colour_off;
2397 slabp->nodeid = nodeid;
2401 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2403 return (kmem_bufctl_t *) (slabp + 1);
2406 static void cache_init_objs(struct kmem_cache *cachep,
2407 struct slab *slabp, unsigned long ctor_flags)
2411 for (i = 0; i < cachep->num; i++) {
2412 void *objp = index_to_obj(cachep, slabp, i);
2414 /* need to poison the objs? */
2415 if (cachep->flags & SLAB_POISON)
2416 poison_obj(cachep, objp, POISON_FREE);
2417 if (cachep->flags & SLAB_STORE_USER)
2418 *dbg_userword(cachep, objp) = NULL;
2420 if (cachep->flags & SLAB_RED_ZONE) {
2421 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2422 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2425 * Constructors are not allowed to allocate memory from the same
2426 * cache which they are a constructor for. Otherwise, deadlock.
2427 * They must also be threaded.
2429 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2430 cachep->ctor(objp + obj_offset(cachep), cachep,
2433 if (cachep->flags & SLAB_RED_ZONE) {
2434 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2435 slab_error(cachep, "constructor overwrote the"
2436 " end of an object");
2437 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2438 slab_error(cachep, "constructor overwrote the"
2439 " start of an object");
2441 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2442 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2443 kernel_map_pages(virt_to_page(objp),
2444 cachep->buffer_size / PAGE_SIZE, 0);
2447 cachep->ctor(objp, cachep, ctor_flags);
2449 slab_bufctl(slabp)[i] = i + 1;
2451 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2455 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2457 if (flags & SLAB_DMA)
2458 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2460 BUG_ON(cachep->gfpflags & GFP_DMA);
2463 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2466 void *objp = index_to_obj(cachep, slabp, slabp->free);
2470 next = slab_bufctl(slabp)[slabp->free];
2472 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2473 WARN_ON(slabp->nodeid != nodeid);
2480 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2481 void *objp, int nodeid)
2483 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2486 /* Verify that the slab belongs to the intended node */
2487 WARN_ON(slabp->nodeid != nodeid);
2489 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2490 printk(KERN_ERR "slab: double free detected in cache "
2491 "'%s', objp %p\n", cachep->name, objp);
2495 slab_bufctl(slabp)[objnr] = slabp->free;
2496 slabp->free = objnr;
2501 * Map pages beginning at addr to the given cache and slab. This is required
2502 * for the slab allocator to be able to lookup the cache and slab of a
2503 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2505 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2511 page = virt_to_page(addr);
2514 if (likely(!PageCompound(page)))
2515 nr_pages <<= cache->gfporder;
2518 page_set_cache(page, cache);
2519 page_set_slab(page, slab);
2521 } while (--nr_pages);
2525 * Grow (by 1) the number of slabs within a cache. This is called by
2526 * kmem_cache_alloc() when there are no active objs left in a cache.
2528 static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid)
2534 unsigned long ctor_flags;
2535 struct kmem_list3 *l3;
2538 * Be lazy and only check for valid flags here, keeping it out of the
2539 * critical path in kmem_cache_alloc().
2541 BUG_ON(flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW));
2542 if (flags & SLAB_NO_GROW)
2545 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2546 local_flags = (flags & SLAB_LEVEL_MASK);
2547 if (!(local_flags & __GFP_WAIT))
2549 * Not allowed to sleep. Need to tell a constructor about
2550 * this - it might need to know...
2552 ctor_flags |= SLAB_CTOR_ATOMIC;
2554 /* Take the l3 list lock to change the colour_next on this node */
2556 l3 = cachep->nodelists[nodeid];
2557 spin_lock(&l3->list_lock);
2559 /* Get colour for the slab, and cal the next value. */
2560 offset = l3->colour_next;
2562 if (l3->colour_next >= cachep->colour)
2563 l3->colour_next = 0;
2564 spin_unlock(&l3->list_lock);
2566 offset *= cachep->colour_off;
2568 if (local_flags & __GFP_WAIT)
2572 * The test for missing atomic flag is performed here, rather than
2573 * the more obvious place, simply to reduce the critical path length
2574 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2575 * will eventually be caught here (where it matters).
2577 kmem_flagcheck(cachep, flags);
2580 * Get mem for the objs. Attempt to allocate a physical page from
2583 objp = kmem_getpages(cachep, flags, nodeid);
2587 /* Get slab management. */
2588 slabp = alloc_slabmgmt(cachep, objp, offset, local_flags, nodeid);
2592 slabp->nodeid = nodeid;
2593 slab_map_pages(cachep, slabp, objp);
2595 cache_init_objs(cachep, slabp, ctor_flags);
2597 if (local_flags & __GFP_WAIT)
2598 local_irq_disable();
2600 spin_lock(&l3->list_lock);
2602 /* Make slab active. */
2603 list_add_tail(&slabp->list, &(l3->slabs_free));
2604 STATS_INC_GROWN(cachep);
2605 l3->free_objects += cachep->num;
2606 spin_unlock(&l3->list_lock);
2609 kmem_freepages(cachep, objp);
2611 if (local_flags & __GFP_WAIT)
2612 local_irq_disable();
2619 * Perform extra freeing checks:
2620 * - detect bad pointers.
2621 * - POISON/RED_ZONE checking
2622 * - destructor calls, for caches with POISON+dtor
2624 static void kfree_debugcheck(const void *objp)
2628 if (!virt_addr_valid(objp)) {
2629 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2630 (unsigned long)objp);
2633 page = virt_to_page(objp);
2634 if (!PageSlab(page)) {
2635 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2636 (unsigned long)objp);
2641 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2643 unsigned long redzone1, redzone2;
2645 redzone1 = *dbg_redzone1(cache, obj);
2646 redzone2 = *dbg_redzone2(cache, obj);
2651 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2654 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2655 slab_error(cache, "double free detected");
2657 slab_error(cache, "memory outside object was overwritten");
2659 printk(KERN_ERR "%p: redzone 1:0x%lx, redzone 2:0x%lx.\n",
2660 obj, redzone1, redzone2);
2663 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2670 objp -= obj_offset(cachep);
2671 kfree_debugcheck(objp);
2672 page = virt_to_page(objp);
2674 slabp = page_get_slab(page);
2676 if (cachep->flags & SLAB_RED_ZONE) {
2677 verify_redzone_free(cachep, objp);
2678 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2679 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2681 if (cachep->flags & SLAB_STORE_USER)
2682 *dbg_userword(cachep, objp) = caller;
2684 objnr = obj_to_index(cachep, slabp, objp);
2686 BUG_ON(objnr >= cachep->num);
2687 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2689 if (cachep->flags & SLAB_DEBUG_INITIAL) {
2691 * Need to call the slab's constructor so the caller can
2692 * perform a verify of its state (debugging). Called without
2693 * the cache-lock held.
2695 cachep->ctor(objp + obj_offset(cachep),
2696 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
2698 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2699 /* we want to cache poison the object,
2700 * call the destruction callback
2702 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
2704 #ifdef CONFIG_DEBUG_SLAB_LEAK
2705 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2707 if (cachep->flags & SLAB_POISON) {
2708 #ifdef CONFIG_DEBUG_PAGEALLOC
2709 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2710 store_stackinfo(cachep, objp, (unsigned long)caller);
2711 kernel_map_pages(virt_to_page(objp),
2712 cachep->buffer_size / PAGE_SIZE, 0);
2714 poison_obj(cachep, objp, POISON_FREE);
2717 poison_obj(cachep, objp, POISON_FREE);
2723 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2728 /* Check slab's freelist to see if this obj is there. */
2729 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2731 if (entries > cachep->num || i >= cachep->num)
2734 if (entries != cachep->num - slabp->inuse) {
2736 printk(KERN_ERR "slab: Internal list corruption detected in "
2737 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2738 cachep->name, cachep->num, slabp, slabp->inuse);
2740 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2743 printk("\n%03x:", i);
2744 printk(" %02x", ((unsigned char *)slabp)[i]);
2751 #define kfree_debugcheck(x) do { } while(0)
2752 #define cache_free_debugcheck(x,objp,z) (objp)
2753 #define check_slabp(x,y) do { } while(0)
2756 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2759 struct kmem_list3 *l3;
2760 struct array_cache *ac;
2763 ac = cpu_cache_get(cachep);
2765 batchcount = ac->batchcount;
2766 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2768 * If there was little recent activity on this cache, then
2769 * perform only a partial refill. Otherwise we could generate
2772 batchcount = BATCHREFILL_LIMIT;
2774 l3 = cachep->nodelists[numa_node_id()];
2776 BUG_ON(ac->avail > 0 || !l3);
2777 spin_lock(&l3->list_lock);
2779 /* See if we can refill from the shared array */
2780 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2783 while (batchcount > 0) {
2784 struct list_head *entry;
2786 /* Get slab alloc is to come from. */
2787 entry = l3->slabs_partial.next;
2788 if (entry == &l3->slabs_partial) {
2789 l3->free_touched = 1;
2790 entry = l3->slabs_free.next;
2791 if (entry == &l3->slabs_free)
2795 slabp = list_entry(entry, struct slab, list);
2796 check_slabp(cachep, slabp);
2797 check_spinlock_acquired(cachep);
2798 while (slabp->inuse < cachep->num && batchcount--) {
2799 STATS_INC_ALLOCED(cachep);
2800 STATS_INC_ACTIVE(cachep);
2801 STATS_SET_HIGH(cachep);
2803 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2806 check_slabp(cachep, slabp);
2808 /* move slabp to correct slabp list: */
2809 list_del(&slabp->list);
2810 if (slabp->free == BUFCTL_END)
2811 list_add(&slabp->list, &l3->slabs_full);
2813 list_add(&slabp->list, &l3->slabs_partial);
2817 l3->free_objects -= ac->avail;
2819 spin_unlock(&l3->list_lock);
2821 if (unlikely(!ac->avail)) {
2823 x = cache_grow(cachep, flags, numa_node_id());
2825 /* cache_grow can reenable interrupts, then ac could change. */
2826 ac = cpu_cache_get(cachep);
2827 if (!x && ac->avail == 0) /* no objects in sight? abort */
2830 if (!ac->avail) /* objects refilled by interrupt? */
2834 return ac->entry[--ac->avail];
2837 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2840 might_sleep_if(flags & __GFP_WAIT);
2842 kmem_flagcheck(cachep, flags);
2847 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2848 gfp_t flags, void *objp, void *caller)
2852 if (cachep->flags & SLAB_POISON) {
2853 #ifdef CONFIG_DEBUG_PAGEALLOC
2854 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
2855 kernel_map_pages(virt_to_page(objp),
2856 cachep->buffer_size / PAGE_SIZE, 1);
2858 check_poison_obj(cachep, objp);
2860 check_poison_obj(cachep, objp);
2862 poison_obj(cachep, objp, POISON_INUSE);
2864 if (cachep->flags & SLAB_STORE_USER)
2865 *dbg_userword(cachep, objp) = caller;
2867 if (cachep->flags & SLAB_RED_ZONE) {
2868 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2869 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2870 slab_error(cachep, "double free, or memory outside"
2871 " object was overwritten");
2873 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
2874 objp, *dbg_redzone1(cachep, objp),
2875 *dbg_redzone2(cachep, objp));
2877 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2878 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2880 #ifdef CONFIG_DEBUG_SLAB_LEAK
2885 slabp = page_get_slab(virt_to_page(objp));
2886 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
2887 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
2890 objp += obj_offset(cachep);
2891 if (cachep->ctor && cachep->flags & SLAB_POISON) {
2892 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2894 if (!(flags & __GFP_WAIT))
2895 ctor_flags |= SLAB_CTOR_ATOMIC;
2897 cachep->ctor(objp, cachep, ctor_flags);
2902 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2905 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
2908 struct array_cache *ac;
2911 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
2912 objp = alternate_node_alloc(cachep, flags);
2919 ac = cpu_cache_get(cachep);
2920 if (likely(ac->avail)) {
2921 STATS_INC_ALLOCHIT(cachep);
2923 objp = ac->entry[--ac->avail];
2925 STATS_INC_ALLOCMISS(cachep);
2926 objp = cache_alloc_refill(cachep, flags);
2931 static __always_inline void *__cache_alloc(struct kmem_cache *cachep,
2932 gfp_t flags, void *caller)
2934 unsigned long save_flags;
2937 cache_alloc_debugcheck_before(cachep, flags);
2939 local_irq_save(save_flags);
2940 objp = ____cache_alloc(cachep, flags);
2941 local_irq_restore(save_flags);
2942 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
2950 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
2952 * If we are in_interrupt, then process context, including cpusets and
2953 * mempolicy, may not apply and should not be used for allocation policy.
2955 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
2957 int nid_alloc, nid_here;
2961 nid_alloc = nid_here = numa_node_id();
2962 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
2963 nid_alloc = cpuset_mem_spread_node();
2964 else if (current->mempolicy)
2965 nid_alloc = slab_node(current->mempolicy);
2966 if (nid_alloc != nid_here)
2967 return __cache_alloc_node(cachep, flags, nid_alloc);
2972 * A interface to enable slab creation on nodeid
2974 static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
2977 struct list_head *entry;
2979 struct kmem_list3 *l3;
2983 l3 = cachep->nodelists[nodeid];
2988 spin_lock(&l3->list_lock);
2989 entry = l3->slabs_partial.next;
2990 if (entry == &l3->slabs_partial) {
2991 l3->free_touched = 1;
2992 entry = l3->slabs_free.next;
2993 if (entry == &l3->slabs_free)
2997 slabp = list_entry(entry, struct slab, list);
2998 check_spinlock_acquired_node(cachep, nodeid);
2999 check_slabp(cachep, slabp);
3001 STATS_INC_NODEALLOCS(cachep);
3002 STATS_INC_ACTIVE(cachep);
3003 STATS_SET_HIGH(cachep);
3005 BUG_ON(slabp->inuse == cachep->num);
3007 obj = slab_get_obj(cachep, slabp, nodeid);
3008 check_slabp(cachep, slabp);
3010 /* move slabp to correct slabp list: */
3011 list_del(&slabp->list);
3013 if (slabp->free == BUFCTL_END)
3014 list_add(&slabp->list, &l3->slabs_full);
3016 list_add(&slabp->list, &l3->slabs_partial);
3018 spin_unlock(&l3->list_lock);
3022 spin_unlock(&l3->list_lock);
3023 x = cache_grow(cachep, flags, nodeid);
3035 * Caller needs to acquire correct kmem_list's list_lock
3037 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3041 struct kmem_list3 *l3;
3043 for (i = 0; i < nr_objects; i++) {
3044 void *objp = objpp[i];
3047 slabp = virt_to_slab(objp);
3048 l3 = cachep->nodelists[node];
3049 list_del(&slabp->list);
3050 check_spinlock_acquired_node(cachep, node);
3051 check_slabp(cachep, slabp);
3052 slab_put_obj(cachep, slabp, objp, node);
3053 STATS_DEC_ACTIVE(cachep);
3055 check_slabp(cachep, slabp);
3057 /* fixup slab chains */
3058 if (slabp->inuse == 0) {
3059 if (l3->free_objects > l3->free_limit) {
3060 l3->free_objects -= cachep->num;
3061 slab_destroy(cachep, slabp);
3063 list_add(&slabp->list, &l3->slabs_free);
3066 /* Unconditionally move a slab to the end of the
3067 * partial list on free - maximum time for the
3068 * other objects to be freed, too.
3070 list_add_tail(&slabp->list, &l3->slabs_partial);
3075 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3078 struct kmem_list3 *l3;
3079 int node = numa_node_id();
3081 batchcount = ac->batchcount;
3083 BUG_ON(!batchcount || batchcount > ac->avail);
3086 l3 = cachep->nodelists[node];
3087 spin_lock(&l3->list_lock);
3089 struct array_cache *shared_array = l3->shared;
3090 int max = shared_array->limit - shared_array->avail;
3092 if (batchcount > max)
3094 memcpy(&(shared_array->entry[shared_array->avail]),
3095 ac->entry, sizeof(void *) * batchcount);
3096 shared_array->avail += batchcount;
3101 free_block(cachep, ac->entry, batchcount, node);
3106 struct list_head *p;
3108 p = l3->slabs_free.next;
3109 while (p != &(l3->slabs_free)) {
3112 slabp = list_entry(p, struct slab, list);
3113 BUG_ON(slabp->inuse);
3118 STATS_SET_FREEABLE(cachep, i);
3121 spin_unlock(&l3->list_lock);
3122 ac->avail -= batchcount;
3123 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3127 * Release an obj back to its cache. If the obj has a constructed state, it must
3128 * be in this state _before_ it is released. Called with disabled ints.
3130 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3132 struct array_cache *ac = cpu_cache_get(cachep);
3135 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3137 if (cache_free_alien(cachep, objp))
3140 if (likely(ac->avail < ac->limit)) {
3141 STATS_INC_FREEHIT(cachep);
3142 ac->entry[ac->avail++] = objp;
3145 STATS_INC_FREEMISS(cachep);
3146 cache_flusharray(cachep, ac);
3147 ac->entry[ac->avail++] = objp;
3152 * kmem_cache_alloc - Allocate an object
3153 * @cachep: The cache to allocate from.
3154 * @flags: See kmalloc().
3156 * Allocate an object from this cache. The flags are only relevant
3157 * if the cache has no available objects.
3159 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3161 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3163 EXPORT_SYMBOL(kmem_cache_alloc);
3166 * kmem_cache_alloc - Allocate an object. The memory is set to zero.
3167 * @cache: The cache to allocate from.
3168 * @flags: See kmalloc().
3170 * Allocate an object from this cache and set the allocated memory to zero.
3171 * The flags are only relevant if the cache has no available objects.
3173 void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
3175 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
3177 memset(ret, 0, obj_size(cache));
3180 EXPORT_SYMBOL(kmem_cache_zalloc);
3183 * kmem_ptr_validate - check if an untrusted pointer might
3185 * @cachep: the cache we're checking against
3186 * @ptr: pointer to validate
3188 * This verifies that the untrusted pointer looks sane:
3189 * it is _not_ a guarantee that the pointer is actually
3190 * part of the slab cache in question, but it at least
3191 * validates that the pointer can be dereferenced and
3192 * looks half-way sane.
3194 * Currently only used for dentry validation.
3196 int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr)
3198 unsigned long addr = (unsigned long)ptr;
3199 unsigned long min_addr = PAGE_OFFSET;
3200 unsigned long align_mask = BYTES_PER_WORD - 1;
3201 unsigned long size = cachep->buffer_size;
3204 if (unlikely(addr < min_addr))
3206 if (unlikely(addr > (unsigned long)high_memory - size))
3208 if (unlikely(addr & align_mask))
3210 if (unlikely(!kern_addr_valid(addr)))
3212 if (unlikely(!kern_addr_valid(addr + size - 1)))
3214 page = virt_to_page(ptr);
3215 if (unlikely(!PageSlab(page)))
3217 if (unlikely(page_get_cache(page) != cachep))
3226 * kmem_cache_alloc_node - Allocate an object on the specified node
3227 * @cachep: The cache to allocate from.
3228 * @flags: See kmalloc().
3229 * @nodeid: node number of the target node.
3231 * Identical to kmem_cache_alloc, except that this function is slow
3232 * and can sleep. And it will allocate memory on the given node, which
3233 * can improve the performance for cpu bound structures.
3234 * New and improved: it will now make sure that the object gets
3235 * put on the correct node list so that there is no false sharing.
3237 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3239 unsigned long save_flags;
3242 cache_alloc_debugcheck_before(cachep, flags);
3243 local_irq_save(save_flags);
3245 if (nodeid == -1 || nodeid == numa_node_id() ||
3246 !cachep->nodelists[nodeid])
3247 ptr = ____cache_alloc(cachep, flags);
3249 ptr = __cache_alloc_node(cachep, flags, nodeid);
3250 local_irq_restore(save_flags);
3252 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr,
3253 __builtin_return_address(0));
3257 EXPORT_SYMBOL(kmem_cache_alloc_node);
3259 void *kmalloc_node(size_t size, gfp_t flags, int node)
3261 struct kmem_cache *cachep;
3263 cachep = kmem_find_general_cachep(size, flags);
3264 if (unlikely(cachep == NULL))
3266 return kmem_cache_alloc_node(cachep, flags, node);
3268 EXPORT_SYMBOL(kmalloc_node);
3272 * kmalloc - allocate memory
3273 * @size: how many bytes of memory are required.
3274 * @flags: the type of memory to allocate.
3275 * @caller: function caller for debug tracking of the caller
3277 * kmalloc is the normal method of allocating memory
3280 * The @flags argument may be one of:
3282 * %GFP_USER - Allocate memory on behalf of user. May sleep.
3284 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
3286 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
3288 * Additionally, the %GFP_DMA flag may be set to indicate the memory
3289 * must be suitable for DMA. This can mean different things on different
3290 * platforms. For example, on i386, it means that the memory must come
3291 * from the first 16MB.
3293 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3296 struct kmem_cache *cachep;
3298 /* If you want to save a few bytes .text space: replace
3300 * Then kmalloc uses the uninlined functions instead of the inline
3303 cachep = __find_general_cachep(size, flags);
3304 if (unlikely(cachep == NULL))
3306 return __cache_alloc(cachep, flags, caller);
3310 void *__kmalloc(size_t size, gfp_t flags)
3312 #ifndef CONFIG_DEBUG_SLAB
3313 return __do_kmalloc(size, flags, NULL);
3315 return __do_kmalloc(size, flags, __builtin_return_address(0));
3318 EXPORT_SYMBOL(__kmalloc);
3320 #ifdef CONFIG_DEBUG_SLAB
3321 void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3323 return __do_kmalloc(size, flags, caller);
3325 EXPORT_SYMBOL(__kmalloc_track_caller);
3330 * __alloc_percpu - allocate one copy of the object for every present
3331 * cpu in the system, zeroing them.
3332 * Objects should be dereferenced using the per_cpu_ptr macro only.
3334 * @size: how many bytes of memory are required.
3336 void *__alloc_percpu(size_t size)
3339 struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL);
3345 * Cannot use for_each_online_cpu since a cpu may come online
3346 * and we have no way of figuring out how to fix the array
3347 * that we have allocated then....
3349 for_each_possible_cpu(i) {
3350 int node = cpu_to_node(i);
3352 if (node_online(node))
3353 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
3355 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
3357 if (!pdata->ptrs[i])
3359 memset(pdata->ptrs[i], 0, size);
3362 /* Catch derefs w/o wrappers */
3363 return (void *)(~(unsigned long)pdata);
3367 if (!cpu_possible(i))
3369 kfree(pdata->ptrs[i]);
3374 EXPORT_SYMBOL(__alloc_percpu);
3378 * kmem_cache_free - Deallocate an object
3379 * @cachep: The cache the allocation was from.
3380 * @objp: The previously allocated object.
3382 * Free an object which was previously allocated from this
3385 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3387 unsigned long flags;
3389 BUG_ON(virt_to_cache(objp) != cachep);
3391 local_irq_save(flags);
3392 __cache_free(cachep, objp);
3393 local_irq_restore(flags);
3395 EXPORT_SYMBOL(kmem_cache_free);
3398 * kfree - free previously allocated memory
3399 * @objp: pointer returned by kmalloc.
3401 * If @objp is NULL, no operation is performed.
3403 * Don't free memory not originally allocated by kmalloc()
3404 * or you will run into trouble.
3406 void kfree(const void *objp)
3408 struct kmem_cache *c;
3409 unsigned long flags;
3411 if (unlikely(!objp))
3413 local_irq_save(flags);
3414 kfree_debugcheck(objp);
3415 c = virt_to_cache(objp);
3416 mutex_debug_check_no_locks_freed(objp, obj_size(c));
3417 __cache_free(c, (void *)objp);
3418 local_irq_restore(flags);
3420 EXPORT_SYMBOL(kfree);
3424 * free_percpu - free previously allocated percpu memory
3425 * @objp: pointer returned by alloc_percpu.
3427 * Don't free memory not originally allocated by alloc_percpu()
3428 * The complemented objp is to check for that.
3430 void free_percpu(const void *objp)
3433 struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp);
3436 * We allocate for all cpus so we cannot use for online cpu here.
3438 for_each_possible_cpu(i)
3442 EXPORT_SYMBOL(free_percpu);
3445 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3447 return obj_size(cachep);
3449 EXPORT_SYMBOL(kmem_cache_size);
3451 const char *kmem_cache_name(struct kmem_cache *cachep)
3453 return cachep->name;
3455 EXPORT_SYMBOL_GPL(kmem_cache_name);
3458 * This initializes kmem_list3 or resizes varioius caches for all nodes.
3460 static int alloc_kmemlist(struct kmem_cache *cachep)
3463 struct kmem_list3 *l3;
3464 struct array_cache *new_shared;
3465 struct array_cache **new_alien;
3467 for_each_online_node(node) {
3469 new_alien = alloc_alien_cache(node, cachep->limit);
3473 new_shared = alloc_arraycache(node,
3474 cachep->shared*cachep->batchcount,
3477 free_alien_cache(new_alien);
3481 l3 = cachep->nodelists[node];
3483 struct array_cache *shared = l3->shared;
3485 spin_lock_irq(&l3->list_lock);
3488 free_block(cachep, shared->entry,
3489 shared->avail, node);
3491 l3->shared = new_shared;
3493 l3->alien = new_alien;
3496 l3->free_limit = (1 + nr_cpus_node(node)) *
3497 cachep->batchcount + cachep->num;
3498 spin_unlock_irq(&l3->list_lock);
3500 free_alien_cache(new_alien);
3503 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
3505 free_alien_cache(new_alien);
3510 kmem_list3_init(l3);
3511 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3512 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3513 l3->shared = new_shared;
3514 l3->alien = new_alien;
3515 l3->free_limit = (1 + nr_cpus_node(node)) *
3516 cachep->batchcount + cachep->num;
3517 cachep->nodelists[node] = l3;
3522 if (!cachep->next.next) {
3523 /* Cache is not active yet. Roll back what we did */
3526 if (cachep->nodelists[node]) {
3527 l3 = cachep->nodelists[node];
3530 free_alien_cache(l3->alien);
3532 cachep->nodelists[node] = NULL;
3540 struct ccupdate_struct {
3541 struct kmem_cache *cachep;
3542 struct array_cache *new[NR_CPUS];
3545 static void do_ccupdate_local(void *info)
3547 struct ccupdate_struct *new = info;
3548 struct array_cache *old;
3551 old = cpu_cache_get(new->cachep);
3553 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3554 new->new[smp_processor_id()] = old;
3557 /* Always called with the cache_chain_mutex held */
3558 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3559 int batchcount, int shared)
3561 struct ccupdate_struct new;
3564 memset(&new.new, 0, sizeof(new.new));
3565 for_each_online_cpu(i) {
3566 new.new[i] = alloc_arraycache(cpu_to_node(i), limit,
3569 for (i--; i >= 0; i--)
3574 new.cachep = cachep;
3576 on_each_cpu(do_ccupdate_local, (void *)&new, 1, 1);
3579 cachep->batchcount = batchcount;
3580 cachep->limit = limit;
3581 cachep->shared = shared;
3583 for_each_online_cpu(i) {
3584 struct array_cache *ccold = new.new[i];
3587 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3588 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3589 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3593 err = alloc_kmemlist(cachep);
3595 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
3596 cachep->name, -err);
3602 /* Called with cache_chain_mutex held always */
3603 static void enable_cpucache(struct kmem_cache *cachep)
3609 * The head array serves three purposes:
3610 * - create a LIFO ordering, i.e. return objects that are cache-warm
3611 * - reduce the number of spinlock operations.
3612 * - reduce the number of linked list operations on the slab and
3613 * bufctl chains: array operations are cheaper.
3614 * The numbers are guessed, we should auto-tune as described by
3617 if (cachep->buffer_size > 131072)
3619 else if (cachep->buffer_size > PAGE_SIZE)
3621 else if (cachep->buffer_size > 1024)
3623 else if (cachep->buffer_size > 256)
3629 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3630 * allocation behaviour: Most allocs on one cpu, most free operations
3631 * on another cpu. For these cases, an efficient object passing between
3632 * cpus is necessary. This is provided by a shared array. The array
3633 * replaces Bonwick's magazine layer.
3634 * On uniprocessor, it's functionally equivalent (but less efficient)
3635 * to a larger limit. Thus disabled by default.
3639 if (cachep->buffer_size <= PAGE_SIZE)
3645 * With debugging enabled, large batchcount lead to excessively long
3646 * periods with disabled local interrupts. Limit the batchcount
3651 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
3653 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3654 cachep->name, -err);
3658 * Drain an array if it contains any elements taking the l3 lock only if
3659 * necessary. Note that the l3 listlock also protects the array_cache
3660 * if drain_array() is used on the shared array.
3662 void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
3663 struct array_cache *ac, int force, int node)
3667 if (!ac || !ac->avail)
3669 if (ac->touched && !force) {
3672 spin_lock_irq(&l3->list_lock);
3674 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3675 if (tofree > ac->avail)
3676 tofree = (ac->avail + 1) / 2;
3677 free_block(cachep, ac->entry, tofree, node);
3678 ac->avail -= tofree;
3679 memmove(ac->entry, &(ac->entry[tofree]),
3680 sizeof(void *) * ac->avail);
3682 spin_unlock_irq(&l3->list_lock);
3687 * cache_reap - Reclaim memory from caches.
3688 * @unused: unused parameter
3690 * Called from workqueue/eventd every few seconds.
3692 * - clear the per-cpu caches for this CPU.
3693 * - return freeable pages to the main free memory pool.
3695 * If we cannot acquire the cache chain mutex then just give up - we'll try
3696 * again on the next iteration.
3698 static void cache_reap(void *unused)
3700 struct kmem_cache *searchp;
3701 struct kmem_list3 *l3;
3702 int node = numa_node_id();
3704 if (!mutex_trylock(&cache_chain_mutex)) {
3705 /* Give up. Setup the next iteration. */
3706 schedule_delayed_work(&__get_cpu_var(reap_work),
3711 list_for_each_entry(searchp, &cache_chain, next) {
3712 struct list_head *p;
3719 * We only take the l3 lock if absolutely necessary and we
3720 * have established with reasonable certainty that
3721 * we can do some work if the lock was obtained.
3723 l3 = searchp->nodelists[node];
3725 reap_alien(searchp, l3);
3727 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
3730 * These are racy checks but it does not matter
3731 * if we skip one check or scan twice.
3733 if (time_after(l3->next_reap, jiffies))
3736 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
3738 drain_array(searchp, l3, l3->shared, 0, node);
3740 if (l3->free_touched) {
3741 l3->free_touched = 0;
3745 tofree = (l3->free_limit + 5 * searchp->num - 1) /
3749 * Do not lock if there are no free blocks.
3751 if (list_empty(&l3->slabs_free))
3754 spin_lock_irq(&l3->list_lock);
3755 p = l3->slabs_free.next;
3756 if (p == &(l3->slabs_free)) {
3757 spin_unlock_irq(&l3->list_lock);
3761 slabp = list_entry(p, struct slab, list);
3762 BUG_ON(slabp->inuse);
3763 list_del(&slabp->list);
3764 STATS_INC_REAPED(searchp);
3767 * Safe to drop the lock. The slab is no longer linked
3768 * to the cache. searchp cannot disappear, we hold
3771 l3->free_objects -= searchp->num;
3772 spin_unlock_irq(&l3->list_lock);
3773 slab_destroy(searchp, slabp);
3774 } while (--tofree > 0);
3779 mutex_unlock(&cache_chain_mutex);
3781 /* Set up the next iteration */
3782 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
3785 #ifdef CONFIG_PROC_FS
3787 static void print_slabinfo_header(struct seq_file *m)
3790 * Output format version, so at least we can change it
3791 * without _too_ many complaints.
3794 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3796 seq_puts(m, "slabinfo - version: 2.1\n");
3798 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3799 "<objperslab> <pagesperslab>");
3800 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3801 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3803 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3804 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
3805 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3810 static void *s_start(struct seq_file *m, loff_t *pos)
3813 struct list_head *p;
3815 mutex_lock(&cache_chain_mutex);
3817 print_slabinfo_header(m);
3818 p = cache_chain.next;
3821 if (p == &cache_chain)
3824 return list_entry(p, struct kmem_cache, next);
3827 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3829 struct kmem_cache *cachep = p;
3831 return cachep->next.next == &cache_chain ?
3832 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
3835 static void s_stop(struct seq_file *m, void *p)
3837 mutex_unlock(&cache_chain_mutex);
3840 static int s_show(struct seq_file *m, void *p)
3842 struct kmem_cache *cachep = p;
3844 unsigned long active_objs;
3845 unsigned long num_objs;
3846 unsigned long active_slabs = 0;
3847 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3851 struct kmem_list3 *l3;
3855 for_each_online_node(node) {
3856 l3 = cachep->nodelists[node];
3861 spin_lock_irq(&l3->list_lock);
3863 list_for_each_entry(slabp, &l3->slabs_full, list) {
3864 if (slabp->inuse != cachep->num && !error)
3865 error = "slabs_full accounting error";
3866 active_objs += cachep->num;
3869 list_for_each_entry(slabp, &l3->slabs_partial, list) {
3870 if (slabp->inuse == cachep->num && !error)
3871 error = "slabs_partial inuse accounting error";
3872 if (!slabp->inuse && !error)
3873 error = "slabs_partial/inuse accounting error";
3874 active_objs += slabp->inuse;
3877 list_for_each_entry(slabp, &l3->slabs_free, list) {
3878 if (slabp->inuse && !error)
3879 error = "slabs_free/inuse accounting error";
3882 free_objects += l3->free_objects;
3884 shared_avail += l3->shared->avail;
3886 spin_unlock_irq(&l3->list_lock);
3888 num_slabs += active_slabs;
3889 num_objs = num_slabs * cachep->num;
3890 if (num_objs - active_objs != free_objects && !error)
3891 error = "free_objects accounting error";
3893 name = cachep->name;
3895 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3897 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
3898 name, active_objs, num_objs, cachep->buffer_size,
3899 cachep->num, (1 << cachep->gfporder));
3900 seq_printf(m, " : tunables %4u %4u %4u",
3901 cachep->limit, cachep->batchcount, cachep->shared);
3902 seq_printf(m, " : slabdata %6lu %6lu %6lu",
3903 active_slabs, num_slabs, shared_avail);
3906 unsigned long high = cachep->high_mark;
3907 unsigned long allocs = cachep->num_allocations;
3908 unsigned long grown = cachep->grown;
3909 unsigned long reaped = cachep->reaped;
3910 unsigned long errors = cachep->errors;
3911 unsigned long max_freeable = cachep->max_freeable;
3912 unsigned long node_allocs = cachep->node_allocs;
3913 unsigned long node_frees = cachep->node_frees;
3914 unsigned long overflows = cachep->node_overflow;
3916 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
3917 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
3918 reaped, errors, max_freeable, node_allocs,
3919 node_frees, overflows);
3923 unsigned long allochit = atomic_read(&cachep->allochit);
3924 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3925 unsigned long freehit = atomic_read(&cachep->freehit);
3926 unsigned long freemiss = atomic_read(&cachep->freemiss);
3928 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
3929 allochit, allocmiss, freehit, freemiss);
3937 * slabinfo_op - iterator that generates /proc/slabinfo
3946 * num-pages-per-slab
3947 * + further values on SMP and with statistics enabled
3950 struct seq_operations slabinfo_op = {
3957 #define MAX_SLABINFO_WRITE 128
3959 * slabinfo_write - Tuning for the slab allocator
3961 * @buffer: user buffer
3962 * @count: data length
3965 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
3966 size_t count, loff_t *ppos)
3968 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
3969 int limit, batchcount, shared, res;
3970 struct kmem_cache *cachep;
3972 if (count > MAX_SLABINFO_WRITE)
3974 if (copy_from_user(&kbuf, buffer, count))
3976 kbuf[MAX_SLABINFO_WRITE] = '\0';
3978 tmp = strchr(kbuf, ' ');
3983 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3986 /* Find the cache in the chain of caches. */
3987 mutex_lock(&cache_chain_mutex);
3989 list_for_each_entry(cachep, &cache_chain, next) {
3990 if (!strcmp(cachep->name, kbuf)) {
3991 if (limit < 1 || batchcount < 1 ||
3992 batchcount > limit || shared < 0) {
3995 res = do_tune_cpucache(cachep, limit,
3996 batchcount, shared);
4001 mutex_unlock(&cache_chain_mutex);
4007 #ifdef CONFIG_DEBUG_SLAB_LEAK
4009 static void *leaks_start(struct seq_file *m, loff_t *pos)
4012 struct list_head *p;
4014 mutex_lock(&cache_chain_mutex);
4015 p = cache_chain.next;
4018 if (p == &cache_chain)
4021 return list_entry(p, struct kmem_cache, next);
4024 static inline int add_caller(unsigned long *n, unsigned long v)
4034 unsigned long *q = p + 2 * i;
4048 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4054 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4060 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4061 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4063 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4068 static void show_symbol(struct seq_file *m, unsigned long address)
4070 #ifdef CONFIG_KALLSYMS
4073 unsigned long offset, size;
4074 char namebuf[KSYM_NAME_LEN+1];
4076 name = kallsyms_lookup(address, &size, &offset, &modname, namebuf);
4079 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4081 seq_printf(m, " [%s]", modname);
4085 seq_printf(m, "%p", (void *)address);
4088 static int leaks_show(struct seq_file *m, void *p)
4090 struct kmem_cache *cachep = p;
4092 struct kmem_list3 *l3;
4094 unsigned long *n = m->private;
4098 if (!(cachep->flags & SLAB_STORE_USER))
4100 if (!(cachep->flags & SLAB_RED_ZONE))
4103 /* OK, we can do it */
4107 for_each_online_node(node) {
4108 l3 = cachep->nodelists[node];
4113 spin_lock_irq(&l3->list_lock);
4115 list_for_each_entry(slabp, &l3->slabs_full, list)
4116 handle_slab(n, cachep, slabp);
4117 list_for_each_entry(slabp, &l3->slabs_partial, list)
4118 handle_slab(n, cachep, slabp);
4119 spin_unlock_irq(&l3->list_lock);
4121 name = cachep->name;
4123 /* Increase the buffer size */
4124 mutex_unlock(&cache_chain_mutex);
4125 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4127 /* Too bad, we are really out */
4129 mutex_lock(&cache_chain_mutex);
4132 *(unsigned long *)m->private = n[0] * 2;
4134 mutex_lock(&cache_chain_mutex);
4135 /* Now make sure this entry will be retried */
4139 for (i = 0; i < n[1]; i++) {
4140 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4141 show_symbol(m, n[2*i+2]);
4147 struct seq_operations slabstats_op = {
4148 .start = leaks_start,
4157 * ksize - get the actual amount of memory allocated for a given object
4158 * @objp: Pointer to the object
4160 * kmalloc may internally round up allocations and return more memory
4161 * than requested. ksize() can be used to determine the actual amount of
4162 * memory allocated. The caller may use this additional memory, even though
4163 * a smaller amount of memory was initially specified with the kmalloc call.
4164 * The caller must guarantee that objp points to a valid object previously
4165 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4166 * must not be freed during the duration of the call.
4168 unsigned int ksize(const void *objp)
4170 if (unlikely(objp == NULL))
4173 return obj_size(virt_to_cache(objp));