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)
210 /* Max number of objs-per-slab for caches which use off-slab slabs.
211 * Needed to avoid a possible looping condition in cache_grow().
213 static unsigned long offslab_limit;
218 * Manages the objs in a slab. Placed either at the beginning of mem allocated
219 * for a slab, or allocated from an general cache.
220 * Slabs are chained into three list: fully used, partial, fully free slabs.
223 struct list_head list;
224 unsigned long colouroff;
225 void *s_mem; /* including colour offset */
226 unsigned int inuse; /* num of objs active in slab */
228 unsigned short nodeid;
234 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
235 * arrange for kmem_freepages to be called via RCU. This is useful if
236 * we need to approach a kernel structure obliquely, from its address
237 * obtained without the usual locking. We can lock the structure to
238 * stabilize it and check it's still at the given address, only if we
239 * can be sure that the memory has not been meanwhile reused for some
240 * other kind of object (which our subsystem's lock might corrupt).
242 * rcu_read_lock before reading the address, then rcu_read_unlock after
243 * taking the spinlock within the structure expected at that address.
245 * We assume struct slab_rcu can overlay struct slab when destroying.
248 struct rcu_head head;
249 struct kmem_cache *cachep;
257 * - LIFO ordering, to hand out cache-warm objects from _alloc
258 * - reduce the number of linked list operations
259 * - reduce spinlock operations
261 * The limit is stored in the per-cpu structure to reduce the data cache
268 unsigned int batchcount;
269 unsigned int touched;
272 * Must have this definition in here for the proper
273 * alignment of array_cache. Also simplifies accessing
275 * [0] is for gcc 2.95. It should really be [].
280 * bootstrap: The caches do not work without cpuarrays anymore, but the
281 * cpuarrays are allocated from the generic caches...
283 #define BOOT_CPUCACHE_ENTRIES 1
284 struct arraycache_init {
285 struct array_cache cache;
286 void *entries[BOOT_CPUCACHE_ENTRIES];
290 * The slab lists for all objects.
293 struct list_head slabs_partial; /* partial list first, better asm code */
294 struct list_head slabs_full;
295 struct list_head slabs_free;
296 unsigned long free_objects;
297 unsigned int free_limit;
298 unsigned int colour_next; /* Per-node cache coloring */
299 spinlock_t list_lock;
300 struct array_cache *shared; /* shared per node */
301 struct array_cache **alien; /* on other nodes */
302 unsigned long next_reap; /* updated without locking */
303 int free_touched; /* updated without locking */
307 * Need this for bootstrapping a per node allocator.
309 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
310 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
311 #define CACHE_CACHE 0
313 #define SIZE_L3 (1 + MAX_NUMNODES)
316 * This function must be completely optimized away if a constant is passed to
317 * it. Mostly the same as what is in linux/slab.h except it returns an index.
319 static __always_inline int index_of(const size_t size)
321 extern void __bad_size(void);
323 if (__builtin_constant_p(size)) {
331 #include "linux/kmalloc_sizes.h"
339 #define INDEX_AC index_of(sizeof(struct arraycache_init))
340 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
342 static void kmem_list3_init(struct kmem_list3 *parent)
344 INIT_LIST_HEAD(&parent->slabs_full);
345 INIT_LIST_HEAD(&parent->slabs_partial);
346 INIT_LIST_HEAD(&parent->slabs_free);
347 parent->shared = NULL;
348 parent->alien = NULL;
349 parent->colour_next = 0;
350 spin_lock_init(&parent->list_lock);
351 parent->free_objects = 0;
352 parent->free_touched = 0;
355 #define MAKE_LIST(cachep, listp, slab, nodeid) \
357 INIT_LIST_HEAD(listp); \
358 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
361 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
363 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
364 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
365 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
375 /* 1) per-cpu data, touched during every alloc/free */
376 struct array_cache *array[NR_CPUS];
377 /* 2) Cache tunables. Protected by cache_chain_mutex */
378 unsigned int batchcount;
382 unsigned int buffer_size;
383 /* 3) touched by every alloc & free from the backend */
384 struct kmem_list3 *nodelists[MAX_NUMNODES];
386 unsigned int flags; /* constant flags */
387 unsigned int num; /* # of objs per slab */
389 /* 4) cache_grow/shrink */
390 /* order of pgs per slab (2^n) */
391 unsigned int gfporder;
393 /* force GFP flags, e.g. GFP_DMA */
396 size_t colour; /* cache colouring range */
397 unsigned int colour_off; /* colour offset */
398 struct kmem_cache *slabp_cache;
399 unsigned int slab_size;
400 unsigned int dflags; /* dynamic flags */
402 /* constructor func */
403 void (*ctor) (void *, struct kmem_cache *, unsigned long);
405 /* de-constructor func */
406 void (*dtor) (void *, struct kmem_cache *, unsigned long);
408 /* 5) cache creation/removal */
410 struct list_head next;
414 unsigned long num_active;
415 unsigned long num_allocations;
416 unsigned long high_mark;
418 unsigned long reaped;
419 unsigned long errors;
420 unsigned long max_freeable;
421 unsigned long node_allocs;
422 unsigned long node_frees;
430 * If debugging is enabled, then the allocator can add additional
431 * fields and/or padding to every object. buffer_size contains the total
432 * object size including these internal fields, the following two
433 * variables contain the offset to the user object and its size.
440 #define CFLGS_OFF_SLAB (0x80000000UL)
441 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
443 #define BATCHREFILL_LIMIT 16
445 * Optimization question: fewer reaps means less probability for unnessary
446 * cpucache drain/refill cycles.
448 * OTOH the cpuarrays can contain lots of objects,
449 * which could lock up otherwise freeable slabs.
451 #define REAPTIMEOUT_CPUC (2*HZ)
452 #define REAPTIMEOUT_LIST3 (4*HZ)
455 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
456 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
457 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
458 #define STATS_INC_GROWN(x) ((x)->grown++)
459 #define STATS_INC_REAPED(x) ((x)->reaped++)
460 #define STATS_SET_HIGH(x) \
462 if ((x)->num_active > (x)->high_mark) \
463 (x)->high_mark = (x)->num_active; \
465 #define STATS_INC_ERR(x) ((x)->errors++)
466 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
467 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
468 #define STATS_SET_FREEABLE(x, i) \
470 if ((x)->max_freeable < i) \
471 (x)->max_freeable = i; \
473 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
474 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
475 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
476 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
478 #define STATS_INC_ACTIVE(x) do { } while (0)
479 #define STATS_DEC_ACTIVE(x) do { } while (0)
480 #define STATS_INC_ALLOCED(x) do { } while (0)
481 #define STATS_INC_GROWN(x) do { } while (0)
482 #define STATS_INC_REAPED(x) do { } while (0)
483 #define STATS_SET_HIGH(x) do { } while (0)
484 #define STATS_INC_ERR(x) do { } while (0)
485 #define STATS_INC_NODEALLOCS(x) do { } while (0)
486 #define STATS_INC_NODEFREES(x) do { } while (0)
487 #define STATS_SET_FREEABLE(x, i) do { } while (0)
488 #define STATS_INC_ALLOCHIT(x) do { } while (0)
489 #define STATS_INC_ALLOCMISS(x) do { } while (0)
490 #define STATS_INC_FREEHIT(x) do { } while (0)
491 #define STATS_INC_FREEMISS(x) do { } while (0)
496 * Magic nums for obj red zoning.
497 * Placed in the first word before and the first word after an obj.
499 #define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
500 #define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
502 /* ...and for poisoning */
503 #define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
504 #define POISON_FREE 0x6b /* for use-after-free poisoning */
505 #define POISON_END 0xa5 /* end-byte of poisoning */
508 * memory layout of objects:
510 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
511 * the end of an object is aligned with the end of the real
512 * allocation. Catches writes behind the end of the allocation.
513 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
515 * cachep->obj_offset: The real object.
516 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
517 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
518 * [BYTES_PER_WORD long]
520 static int obj_offset(struct kmem_cache *cachep)
522 return cachep->obj_offset;
525 static int obj_size(struct kmem_cache *cachep)
527 return cachep->obj_size;
530 static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
532 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
533 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD);
536 static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
538 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
539 if (cachep->flags & SLAB_STORE_USER)
540 return (unsigned long *)(objp + cachep->buffer_size -
542 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD);
545 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
547 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
548 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
553 #define obj_offset(x) 0
554 #define obj_size(cachep) (cachep->buffer_size)
555 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
556 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
557 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
562 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
565 #if defined(CONFIG_LARGE_ALLOCS)
566 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
567 #define MAX_GFP_ORDER 13 /* up to 32Mb */
568 #elif defined(CONFIG_MMU)
569 #define MAX_OBJ_ORDER 5 /* 32 pages */
570 #define MAX_GFP_ORDER 5 /* 32 pages */
572 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
573 #define MAX_GFP_ORDER 8 /* up to 1Mb */
577 * Do not go above this order unless 0 objects fit into the slab.
579 #define BREAK_GFP_ORDER_HI 1
580 #define BREAK_GFP_ORDER_LO 0
581 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
584 * Functions for storing/retrieving the cachep and or slab from the page
585 * allocator. These are used to find the slab an obj belongs to. With kfree(),
586 * these are used to find the cache which an obj belongs to.
588 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
590 page->lru.next = (struct list_head *)cache;
593 static inline struct kmem_cache *page_get_cache(struct page *page)
595 if (unlikely(PageCompound(page)))
596 page = (struct page *)page_private(page);
597 return (struct kmem_cache *)page->lru.next;
600 static inline void page_set_slab(struct page *page, struct slab *slab)
602 page->lru.prev = (struct list_head *)slab;
605 static inline struct slab *page_get_slab(struct page *page)
607 if (unlikely(PageCompound(page)))
608 page = (struct page *)page_private(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.
700 static DEFINE_PER_CPU(struct work_struct, reap_work);
702 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
704 static void enable_cpucache(struct kmem_cache *cachep);
705 static void cache_reap(void *unused);
706 static int __node_shrink(struct kmem_cache *cachep, int node);
708 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
710 return cachep->array[smp_processor_id()];
713 static inline struct kmem_cache *__find_general_cachep(size_t size,
716 struct cache_sizes *csizep = malloc_sizes;
719 /* This happens if someone tries to call
720 * kmem_cache_create(), or __kmalloc(), before
721 * the generic caches are initialized.
723 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
725 while (size > csizep->cs_size)
729 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
730 * has cs_{dma,}cachep==NULL. Thus no special case
731 * for large kmalloc calls required.
733 if (unlikely(gfpflags & GFP_DMA))
734 return csizep->cs_dmacachep;
735 return csizep->cs_cachep;
738 struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
740 return __find_general_cachep(size, gfpflags);
742 EXPORT_SYMBOL(kmem_find_general_cachep);
744 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
746 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
750 * Calculate the number of objects and left-over bytes for a given buffer size.
752 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
753 size_t align, int flags, size_t *left_over,
758 size_t slab_size = PAGE_SIZE << gfporder;
761 * The slab management structure can be either off the slab or
762 * on it. For the latter case, the memory allocated for a
766 * - One kmem_bufctl_t for each object
767 * - Padding to respect alignment of @align
768 * - @buffer_size bytes for each object
770 * If the slab management structure is off the slab, then the
771 * alignment will already be calculated into the size. Because
772 * the slabs are all pages aligned, the objects will be at the
773 * correct alignment when allocated.
775 if (flags & CFLGS_OFF_SLAB) {
777 nr_objs = slab_size / buffer_size;
779 if (nr_objs > SLAB_LIMIT)
780 nr_objs = SLAB_LIMIT;
783 * Ignore padding for the initial guess. The padding
784 * is at most @align-1 bytes, and @buffer_size is at
785 * least @align. In the worst case, this result will
786 * be one greater than the number of objects that fit
787 * into the memory allocation when taking the padding
790 nr_objs = (slab_size - sizeof(struct slab)) /
791 (buffer_size + sizeof(kmem_bufctl_t));
794 * This calculated number will be either the right
795 * amount, or one greater than what we want.
797 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
801 if (nr_objs > SLAB_LIMIT)
802 nr_objs = SLAB_LIMIT;
804 mgmt_size = slab_mgmt_size(nr_objs, align);
807 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
810 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
812 static void __slab_error(const char *function, struct kmem_cache *cachep,
815 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
816 function, cachep->name, msg);
822 * Special reaping functions for NUMA systems called from cache_reap().
823 * These take care of doing round robin flushing of alien caches (containing
824 * objects freed on different nodes from which they were allocated) and the
825 * flushing of remote pcps by calling drain_node_pages.
827 static DEFINE_PER_CPU(unsigned long, reap_node);
829 static void init_reap_node(int cpu)
833 node = next_node(cpu_to_node(cpu), node_online_map);
834 if (node == MAX_NUMNODES)
835 node = first_node(node_online_map);
837 __get_cpu_var(reap_node) = node;
840 static void next_reap_node(void)
842 int node = __get_cpu_var(reap_node);
845 * Also drain per cpu pages on remote zones
847 if (node != numa_node_id())
848 drain_node_pages(node);
850 node = next_node(node, node_online_map);
851 if (unlikely(node >= MAX_NUMNODES))
852 node = first_node(node_online_map);
853 __get_cpu_var(reap_node) = node;
857 #define init_reap_node(cpu) do { } while (0)
858 #define next_reap_node(void) do { } while (0)
862 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
863 * via the workqueue/eventd.
864 * Add the CPU number into the expiration time to minimize the possibility of
865 * the CPUs getting into lockstep and contending for the global cache chain
868 static void __devinit start_cpu_timer(int cpu)
870 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
873 * When this gets called from do_initcalls via cpucache_init(),
874 * init_workqueues() has already run, so keventd will be setup
877 if (keventd_up() && reap_work->func == NULL) {
879 INIT_WORK(reap_work, cache_reap, NULL);
880 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
884 static struct array_cache *alloc_arraycache(int node, int entries,
887 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
888 struct array_cache *nc = NULL;
890 nc = kmalloc_node(memsize, GFP_KERNEL, node);
894 nc->batchcount = batchcount;
896 spin_lock_init(&nc->lock);
902 * Transfer objects in one arraycache to another.
903 * Locking must be handled by the caller.
905 * Return the number of entries transferred.
907 static int transfer_objects(struct array_cache *to,
908 struct array_cache *from, unsigned int max)
910 /* Figure out how many entries to transfer */
911 int nr = min(min(from->avail, max), to->limit - to->avail);
916 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
926 static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int);
927 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
929 static struct array_cache **alloc_alien_cache(int node, int limit)
931 struct array_cache **ac_ptr;
932 int memsize = sizeof(void *) * MAX_NUMNODES;
937 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
940 if (i == node || !node_online(i)) {
944 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
946 for (i--; i <= 0; i--)
956 static void free_alien_cache(struct array_cache **ac_ptr)
967 static void __drain_alien_cache(struct kmem_cache *cachep,
968 struct array_cache *ac, int node)
970 struct kmem_list3 *rl3 = cachep->nodelists[node];
973 spin_lock(&rl3->list_lock);
975 * Stuff objects into the remote nodes shared array first.
976 * That way we could avoid the overhead of putting the objects
977 * into the free lists and getting them back later.
979 transfer_objects(rl3->shared, ac, ac->limit);
981 free_block(cachep, ac->entry, ac->avail, node);
983 spin_unlock(&rl3->list_lock);
988 * Called from cache_reap() to regularly drain alien caches round robin.
990 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
992 int node = __get_cpu_var(reap_node);
995 struct array_cache *ac = l3->alien[node];
997 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
998 __drain_alien_cache(cachep, ac, node);
999 spin_unlock_irq(&ac->lock);
1004 static void drain_alien_cache(struct kmem_cache *cachep,
1005 struct array_cache **alien)
1008 struct array_cache *ac;
1009 unsigned long flags;
1011 for_each_online_node(i) {
1014 spin_lock_irqsave(&ac->lock, flags);
1015 __drain_alien_cache(cachep, ac, i);
1016 spin_unlock_irqrestore(&ac->lock, flags);
1022 #define drain_alien_cache(cachep, alien) do { } while (0)
1023 #define reap_alien(cachep, l3) do { } while (0)
1025 static inline struct array_cache **alloc_alien_cache(int node, int limit)
1027 return (struct array_cache **) 0x01020304ul;
1030 static inline void free_alien_cache(struct array_cache **ac_ptr)
1036 static int __devinit cpuup_callback(struct notifier_block *nfb,
1037 unsigned long action, void *hcpu)
1039 long cpu = (long)hcpu;
1040 struct kmem_cache *cachep;
1041 struct kmem_list3 *l3 = NULL;
1042 int node = cpu_to_node(cpu);
1043 int memsize = sizeof(struct kmem_list3);
1046 case CPU_UP_PREPARE:
1047 mutex_lock(&cache_chain_mutex);
1049 * We need to do this right in the beginning since
1050 * alloc_arraycache's are going to use this list.
1051 * kmalloc_node allows us to add the slab to the right
1052 * kmem_list3 and not this cpu's kmem_list3
1055 list_for_each_entry(cachep, &cache_chain, next) {
1057 * Set up the size64 kmemlist for cpu before we can
1058 * begin anything. Make sure some other cpu on this
1059 * node has not already allocated this
1061 if (!cachep->nodelists[node]) {
1062 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1065 kmem_list3_init(l3);
1066 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1067 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1070 * The l3s don't come and go as CPUs come and
1071 * go. cache_chain_mutex is sufficient
1074 cachep->nodelists[node] = l3;
1077 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1078 cachep->nodelists[node]->free_limit =
1079 (1 + nr_cpus_node(node)) *
1080 cachep->batchcount + cachep->num;
1081 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1085 * Now we can go ahead with allocating the shared arrays and
1088 list_for_each_entry(cachep, &cache_chain, next) {
1089 struct array_cache *nc;
1090 struct array_cache *shared;
1091 struct array_cache **alien;
1093 nc = alloc_arraycache(node, cachep->limit,
1094 cachep->batchcount);
1097 shared = alloc_arraycache(node,
1098 cachep->shared * cachep->batchcount,
1103 alien = alloc_alien_cache(node, cachep->limit);
1106 cachep->array[cpu] = nc;
1107 l3 = cachep->nodelists[node];
1110 spin_lock_irq(&l3->list_lock);
1113 * We are serialised from CPU_DEAD or
1114 * CPU_UP_CANCELLED by the cpucontrol lock
1116 l3->shared = shared;
1125 spin_unlock_irq(&l3->list_lock);
1127 free_alien_cache(alien);
1129 mutex_unlock(&cache_chain_mutex);
1132 start_cpu_timer(cpu);
1134 #ifdef CONFIG_HOTPLUG_CPU
1137 * Even if all the cpus of a node are down, we don't free the
1138 * kmem_list3 of any cache. This to avoid a race between
1139 * cpu_down, and a kmalloc allocation from another cpu for
1140 * memory from the node of the cpu going down. The list3
1141 * structure is usually allocated from kmem_cache_create() and
1142 * gets destroyed at kmem_cache_destroy().
1145 case CPU_UP_CANCELED:
1146 mutex_lock(&cache_chain_mutex);
1147 list_for_each_entry(cachep, &cache_chain, next) {
1148 struct array_cache *nc;
1149 struct array_cache *shared;
1150 struct array_cache **alien;
1153 mask = node_to_cpumask(node);
1154 /* cpu is dead; no one can alloc from it. */
1155 nc = cachep->array[cpu];
1156 cachep->array[cpu] = NULL;
1157 l3 = cachep->nodelists[node];
1160 goto free_array_cache;
1162 spin_lock_irq(&l3->list_lock);
1164 /* Free limit for this kmem_list3 */
1165 l3->free_limit -= cachep->batchcount;
1167 free_block(cachep, nc->entry, nc->avail, node);
1169 if (!cpus_empty(mask)) {
1170 spin_unlock_irq(&l3->list_lock);
1171 goto free_array_cache;
1174 shared = l3->shared;
1176 free_block(cachep, l3->shared->entry,
1177 l3->shared->avail, node);
1184 spin_unlock_irq(&l3->list_lock);
1188 drain_alien_cache(cachep, alien);
1189 free_alien_cache(alien);
1195 * In the previous loop, all the objects were freed to
1196 * the respective cache's slabs, now we can go ahead and
1197 * shrink each nodelist to its limit.
1199 list_for_each_entry(cachep, &cache_chain, next) {
1200 l3 = cachep->nodelists[node];
1203 spin_lock_irq(&l3->list_lock);
1204 /* free slabs belonging to this node */
1205 __node_shrink(cachep, node);
1206 spin_unlock_irq(&l3->list_lock);
1208 mutex_unlock(&cache_chain_mutex);
1214 mutex_unlock(&cache_chain_mutex);
1218 static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
1221 * swap the static kmem_list3 with kmalloced memory
1223 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1226 struct kmem_list3 *ptr;
1228 BUG_ON(cachep->nodelists[nodeid] != list);
1229 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1232 local_irq_disable();
1233 memcpy(ptr, list, sizeof(struct kmem_list3));
1234 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1235 cachep->nodelists[nodeid] = ptr;
1240 * Initialisation. Called after the page allocator have been initialised and
1241 * before smp_init().
1243 void __init kmem_cache_init(void)
1246 struct cache_sizes *sizes;
1247 struct cache_names *names;
1251 for (i = 0; i < NUM_INIT_LISTS; i++) {
1252 kmem_list3_init(&initkmem_list3[i]);
1253 if (i < MAX_NUMNODES)
1254 cache_cache.nodelists[i] = NULL;
1258 * Fragmentation resistance on low memory - only use bigger
1259 * page orders on machines with more than 32MB of memory.
1261 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1262 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1264 /* Bootstrap is tricky, because several objects are allocated
1265 * from caches that do not exist yet:
1266 * 1) initialize the cache_cache cache: it contains the struct
1267 * kmem_cache structures of all caches, except cache_cache itself:
1268 * cache_cache is statically allocated.
1269 * Initially an __init data area is used for the head array and the
1270 * kmem_list3 structures, it's replaced with a kmalloc allocated
1271 * array at the end of the bootstrap.
1272 * 2) Create the first kmalloc cache.
1273 * The struct kmem_cache for the new cache is allocated normally.
1274 * An __init data area is used for the head array.
1275 * 3) Create the remaining kmalloc caches, with minimally sized
1277 * 4) Replace the __init data head arrays for cache_cache and the first
1278 * kmalloc cache with kmalloc allocated arrays.
1279 * 5) Replace the __init data for kmem_list3 for cache_cache and
1280 * the other cache's with kmalloc allocated memory.
1281 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1284 /* 1) create the cache_cache */
1285 INIT_LIST_HEAD(&cache_chain);
1286 list_add(&cache_cache.next, &cache_chain);
1287 cache_cache.colour_off = cache_line_size();
1288 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1289 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
1291 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1294 for (order = 0; order < MAX_ORDER; order++) {
1295 cache_estimate(order, cache_cache.buffer_size,
1296 cache_line_size(), 0, &left_over, &cache_cache.num);
1297 if (cache_cache.num)
1300 BUG_ON(!cache_cache.num);
1301 cache_cache.gfporder = order;
1302 cache_cache.colour = left_over / cache_cache.colour_off;
1303 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1304 sizeof(struct slab), cache_line_size());
1306 /* 2+3) create the kmalloc caches */
1307 sizes = malloc_sizes;
1308 names = cache_names;
1311 * Initialize the caches that provide memory for the array cache and the
1312 * kmem_list3 structures first. Without this, further allocations will
1316 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1317 sizes[INDEX_AC].cs_size,
1318 ARCH_KMALLOC_MINALIGN,
1319 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1322 if (INDEX_AC != INDEX_L3) {
1323 sizes[INDEX_L3].cs_cachep =
1324 kmem_cache_create(names[INDEX_L3].name,
1325 sizes[INDEX_L3].cs_size,
1326 ARCH_KMALLOC_MINALIGN,
1327 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1331 while (sizes->cs_size != ULONG_MAX) {
1333 * For performance, all the general caches are L1 aligned.
1334 * This should be particularly beneficial on SMP boxes, as it
1335 * eliminates "false sharing".
1336 * Note for systems short on memory removing the alignment will
1337 * allow tighter packing of the smaller caches.
1339 if (!sizes->cs_cachep) {
1340 sizes->cs_cachep = kmem_cache_create(names->name,
1342 ARCH_KMALLOC_MINALIGN,
1343 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1347 /* Inc off-slab bufctl limit until the ceiling is hit. */
1348 if (!(OFF_SLAB(sizes->cs_cachep))) {
1349 offslab_limit = sizes->cs_size - sizeof(struct slab);
1350 offslab_limit /= sizeof(kmem_bufctl_t);
1353 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
1355 ARCH_KMALLOC_MINALIGN,
1356 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1362 /* 4) Replace the bootstrap head arrays */
1366 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1368 local_irq_disable();
1369 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1370 memcpy(ptr, cpu_cache_get(&cache_cache),
1371 sizeof(struct arraycache_init));
1372 cache_cache.array[smp_processor_id()] = ptr;
1375 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1377 local_irq_disable();
1378 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1379 != &initarray_generic.cache);
1380 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1381 sizeof(struct arraycache_init));
1382 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1386 /* 5) Replace the bootstrap kmem_list3's */
1389 /* Replace the static kmem_list3 structures for the boot cpu */
1390 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
1393 for_each_online_node(node) {
1394 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1395 &initkmem_list3[SIZE_AC + node], node);
1397 if (INDEX_AC != INDEX_L3) {
1398 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1399 &initkmem_list3[SIZE_L3 + node],
1405 /* 6) resize the head arrays to their final sizes */
1407 struct kmem_cache *cachep;
1408 mutex_lock(&cache_chain_mutex);
1409 list_for_each_entry(cachep, &cache_chain, next)
1410 enable_cpucache(cachep);
1411 mutex_unlock(&cache_chain_mutex);
1415 g_cpucache_up = FULL;
1418 * Register a cpu startup notifier callback that initializes
1419 * cpu_cache_get for all new cpus
1421 register_cpu_notifier(&cpucache_notifier);
1424 * The reap timers are started later, with a module init call: That part
1425 * of the kernel is not yet operational.
1429 static int __init cpucache_init(void)
1434 * Register the timers that return unneeded pages to the page allocator
1436 for_each_online_cpu(cpu)
1437 start_cpu_timer(cpu);
1440 __initcall(cpucache_init);
1443 * Interface to system's page allocator. No need to hold the cache-lock.
1445 * If we requested dmaable memory, we will get it. Even if we
1446 * did not request dmaable memory, we might get it, but that
1447 * would be relatively rare and ignorable.
1449 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1455 flags |= cachep->gfpflags;
1456 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1459 addr = page_address(page);
1461 i = (1 << cachep->gfporder);
1462 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1463 atomic_add(i, &slab_reclaim_pages);
1464 add_page_state(nr_slab, i);
1466 __SetPageSlab(page);
1473 * Interface to system's page release.
1475 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1477 unsigned long i = (1 << cachep->gfporder);
1478 struct page *page = virt_to_page(addr);
1479 const unsigned long nr_freed = i;
1482 BUG_ON(!PageSlab(page));
1483 __ClearPageSlab(page);
1486 sub_page_state(nr_slab, nr_freed);
1487 if (current->reclaim_state)
1488 current->reclaim_state->reclaimed_slab += nr_freed;
1489 free_pages((unsigned long)addr, cachep->gfporder);
1490 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1491 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages);
1494 static void kmem_rcu_free(struct rcu_head *head)
1496 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1497 struct kmem_cache *cachep = slab_rcu->cachep;
1499 kmem_freepages(cachep, slab_rcu->addr);
1500 if (OFF_SLAB(cachep))
1501 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1506 #ifdef CONFIG_DEBUG_PAGEALLOC
1507 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1508 unsigned long caller)
1510 int size = obj_size(cachep);
1512 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1514 if (size < 5 * sizeof(unsigned long))
1517 *addr++ = 0x12345678;
1519 *addr++ = smp_processor_id();
1520 size -= 3 * sizeof(unsigned long);
1522 unsigned long *sptr = &caller;
1523 unsigned long svalue;
1525 while (!kstack_end(sptr)) {
1527 if (kernel_text_address(svalue)) {
1529 size -= sizeof(unsigned long);
1530 if (size <= sizeof(unsigned long))
1536 *addr++ = 0x87654321;
1540 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1542 int size = obj_size(cachep);
1543 addr = &((char *)addr)[obj_offset(cachep)];
1545 memset(addr, val, size);
1546 *(unsigned char *)(addr + size - 1) = POISON_END;
1549 static void dump_line(char *data, int offset, int limit)
1552 printk(KERN_ERR "%03x:", offset);
1553 for (i = 0; i < limit; i++)
1554 printk(" %02x", (unsigned char)data[offset + i]);
1561 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1566 if (cachep->flags & SLAB_RED_ZONE) {
1567 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
1568 *dbg_redzone1(cachep, objp),
1569 *dbg_redzone2(cachep, objp));
1572 if (cachep->flags & SLAB_STORE_USER) {
1573 printk(KERN_ERR "Last user: [<%p>]",
1574 *dbg_userword(cachep, objp));
1575 print_symbol("(%s)",
1576 (unsigned long)*dbg_userword(cachep, objp));
1579 realobj = (char *)objp + obj_offset(cachep);
1580 size = obj_size(cachep);
1581 for (i = 0; i < size && lines; i += 16, lines--) {
1584 if (i + limit > size)
1586 dump_line(realobj, i, limit);
1590 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1596 realobj = (char *)objp + obj_offset(cachep);
1597 size = obj_size(cachep);
1599 for (i = 0; i < size; i++) {
1600 char exp = POISON_FREE;
1603 if (realobj[i] != exp) {
1609 "Slab corruption: start=%p, len=%d\n",
1611 print_objinfo(cachep, objp, 0);
1613 /* Hexdump the affected line */
1616 if (i + limit > size)
1618 dump_line(realobj, i, limit);
1621 /* Limit to 5 lines */
1627 /* Print some data about the neighboring objects, if they
1630 struct slab *slabp = virt_to_slab(objp);
1633 objnr = obj_to_index(cachep, slabp, objp);
1635 objp = index_to_obj(cachep, slabp, objnr - 1);
1636 realobj = (char *)objp + obj_offset(cachep);
1637 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1639 print_objinfo(cachep, objp, 2);
1641 if (objnr + 1 < cachep->num) {
1642 objp = index_to_obj(cachep, slabp, objnr + 1);
1643 realobj = (char *)objp + obj_offset(cachep);
1644 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1646 print_objinfo(cachep, objp, 2);
1654 * slab_destroy_objs - destroy a slab and its objects
1655 * @cachep: cache pointer being destroyed
1656 * @slabp: slab pointer being destroyed
1658 * Call the registered destructor for each object in a slab that is being
1661 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1664 for (i = 0; i < cachep->num; i++) {
1665 void *objp = index_to_obj(cachep, slabp, i);
1667 if (cachep->flags & SLAB_POISON) {
1668 #ifdef CONFIG_DEBUG_PAGEALLOC
1669 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1671 kernel_map_pages(virt_to_page(objp),
1672 cachep->buffer_size / PAGE_SIZE, 1);
1674 check_poison_obj(cachep, objp);
1676 check_poison_obj(cachep, objp);
1679 if (cachep->flags & SLAB_RED_ZONE) {
1680 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1681 slab_error(cachep, "start of a freed object "
1683 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1684 slab_error(cachep, "end of a freed object "
1687 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1688 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
1692 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1696 for (i = 0; i < cachep->num; i++) {
1697 void *objp = index_to_obj(cachep, slabp, i);
1698 (cachep->dtor) (objp, cachep, 0);
1705 * slab_destroy - destroy and release all objects in a slab
1706 * @cachep: cache pointer being destroyed
1707 * @slabp: slab pointer being destroyed
1709 * Destroy all the objs in a slab, and release the mem back to the system.
1710 * Before calling the slab must have been unlinked from the cache. The
1711 * cache-lock is not held/needed.
1713 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1715 void *addr = slabp->s_mem - slabp->colouroff;
1717 slab_destroy_objs(cachep, slabp);
1718 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1719 struct slab_rcu *slab_rcu;
1721 slab_rcu = (struct slab_rcu *)slabp;
1722 slab_rcu->cachep = cachep;
1723 slab_rcu->addr = addr;
1724 call_rcu(&slab_rcu->head, kmem_rcu_free);
1726 kmem_freepages(cachep, addr);
1727 if (OFF_SLAB(cachep))
1728 kmem_cache_free(cachep->slabp_cache, slabp);
1733 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1734 * size of kmem_list3.
1736 static void set_up_list3s(struct kmem_cache *cachep, int index)
1740 for_each_online_node(node) {
1741 cachep->nodelists[node] = &initkmem_list3[index + node];
1742 cachep->nodelists[node]->next_reap = jiffies +
1744 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1749 * calculate_slab_order - calculate size (page order) of slabs
1750 * @cachep: pointer to the cache that is being created
1751 * @size: size of objects to be created in this cache.
1752 * @align: required alignment for the objects.
1753 * @flags: slab allocation flags
1755 * Also calculates the number of objects per slab.
1757 * This could be made much more intelligent. For now, try to avoid using
1758 * high order pages for slabs. When the gfp() functions are more friendly
1759 * towards high-order requests, this should be changed.
1761 static size_t calculate_slab_order(struct kmem_cache *cachep,
1762 size_t size, size_t align, unsigned long flags)
1764 size_t left_over = 0;
1767 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
1771 cache_estimate(gfporder, size, align, flags, &remainder, &num);
1775 /* More than offslab_limit objects will cause problems */
1776 if ((flags & CFLGS_OFF_SLAB) && num > offslab_limit)
1779 /* Found something acceptable - save it away */
1781 cachep->gfporder = gfporder;
1782 left_over = remainder;
1785 * A VFS-reclaimable slab tends to have most allocations
1786 * as GFP_NOFS and we really don't want to have to be allocating
1787 * higher-order pages when we are unable to shrink dcache.
1789 if (flags & SLAB_RECLAIM_ACCOUNT)
1793 * Large number of objects is good, but very large slabs are
1794 * currently bad for the gfp()s.
1796 if (gfporder >= slab_break_gfp_order)
1800 * Acceptable internal fragmentation?
1802 if (left_over * 8 <= (PAGE_SIZE << gfporder))
1808 static void setup_cpu_cache(struct kmem_cache *cachep)
1810 if (g_cpucache_up == FULL) {
1811 enable_cpucache(cachep);
1814 if (g_cpucache_up == NONE) {
1816 * Note: the first kmem_cache_create must create the cache
1817 * that's used by kmalloc(24), otherwise the creation of
1818 * further caches will BUG().
1820 cachep->array[smp_processor_id()] = &initarray_generic.cache;
1823 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
1824 * the first cache, then we need to set up all its list3s,
1825 * otherwise the creation of further caches will BUG().
1827 set_up_list3s(cachep, SIZE_AC);
1828 if (INDEX_AC == INDEX_L3)
1829 g_cpucache_up = PARTIAL_L3;
1831 g_cpucache_up = PARTIAL_AC;
1833 cachep->array[smp_processor_id()] =
1834 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1836 if (g_cpucache_up == PARTIAL_AC) {
1837 set_up_list3s(cachep, SIZE_L3);
1838 g_cpucache_up = PARTIAL_L3;
1841 for_each_online_node(node) {
1842 cachep->nodelists[node] =
1843 kmalloc_node(sizeof(struct kmem_list3),
1845 BUG_ON(!cachep->nodelists[node]);
1846 kmem_list3_init(cachep->nodelists[node]);
1850 cachep->nodelists[numa_node_id()]->next_reap =
1851 jiffies + REAPTIMEOUT_LIST3 +
1852 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1854 cpu_cache_get(cachep)->avail = 0;
1855 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1856 cpu_cache_get(cachep)->batchcount = 1;
1857 cpu_cache_get(cachep)->touched = 0;
1858 cachep->batchcount = 1;
1859 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1863 * kmem_cache_create - Create a cache.
1864 * @name: A string which is used in /proc/slabinfo to identify this cache.
1865 * @size: The size of objects to be created in this cache.
1866 * @align: The required alignment for the objects.
1867 * @flags: SLAB flags
1868 * @ctor: A constructor for the objects.
1869 * @dtor: A destructor for the objects.
1871 * Returns a ptr to the cache on success, NULL on failure.
1872 * Cannot be called within a int, but can be interrupted.
1873 * The @ctor is run when new pages are allocated by the cache
1874 * and the @dtor is run before the pages are handed back.
1876 * @name must be valid until the cache is destroyed. This implies that
1877 * the module calling this has to destroy the cache before getting unloaded.
1881 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1882 * to catch references to uninitialised memory.
1884 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1885 * for buffer overruns.
1887 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1888 * cacheline. This can be beneficial if you're counting cycles as closely
1892 kmem_cache_create (const char *name, size_t size, size_t align,
1893 unsigned long flags,
1894 void (*ctor)(void*, struct kmem_cache *, unsigned long),
1895 void (*dtor)(void*, struct kmem_cache *, unsigned long))
1897 size_t left_over, slab_size, ralign;
1898 struct kmem_cache *cachep = NULL;
1899 struct list_head *p;
1902 * Sanity checks... these are all serious usage bugs.
1904 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
1905 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
1906 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
1912 * Prevent CPUs from coming and going.
1913 * lock_cpu_hotplug() nests outside cache_chain_mutex
1917 mutex_lock(&cache_chain_mutex);
1919 list_for_each(p, &cache_chain) {
1920 struct kmem_cache *pc = list_entry(p, struct kmem_cache, next);
1921 mm_segment_t old_fs = get_fs();
1926 * This happens when the module gets unloaded and doesn't
1927 * destroy its slab cache and no-one else reuses the vmalloc
1928 * area of the module. Print a warning.
1931 res = __get_user(tmp, pc->name);
1934 printk("SLAB: cache with size %d has lost its name\n",
1939 if (!strcmp(pc->name, name)) {
1940 printk("kmem_cache_create: duplicate cache %s\n", name);
1947 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1948 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
1949 /* No constructor, but inital state check requested */
1950 printk(KERN_ERR "%s: No con, but init state check "
1951 "requested - %s\n", __FUNCTION__, name);
1952 flags &= ~SLAB_DEBUG_INITIAL;
1956 * Enable redzoning and last user accounting, except for caches with
1957 * large objects, if the increased size would increase the object size
1958 * above the next power of two: caches with object sizes just above a
1959 * power of two have a significant amount of internal fragmentation.
1961 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
1962 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
1963 if (!(flags & SLAB_DESTROY_BY_RCU))
1964 flags |= SLAB_POISON;
1966 if (flags & SLAB_DESTROY_BY_RCU)
1967 BUG_ON(flags & SLAB_POISON);
1969 if (flags & SLAB_DESTROY_BY_RCU)
1973 * Always checks flags, a caller might be expecting debug support which
1976 BUG_ON(flags & ~CREATE_MASK);
1979 * Check that size is in terms of words. This is needed to avoid
1980 * unaligned accesses for some archs when redzoning is used, and makes
1981 * sure any on-slab bufctl's are also correctly aligned.
1983 if (size & (BYTES_PER_WORD - 1)) {
1984 size += (BYTES_PER_WORD - 1);
1985 size &= ~(BYTES_PER_WORD - 1);
1988 /* calculate the final buffer alignment: */
1990 /* 1) arch recommendation: can be overridden for debug */
1991 if (flags & SLAB_HWCACHE_ALIGN) {
1993 * Default alignment: as specified by the arch code. Except if
1994 * an object is really small, then squeeze multiple objects into
1997 ralign = cache_line_size();
1998 while (size <= ralign / 2)
2001 ralign = BYTES_PER_WORD;
2003 /* 2) arch mandated alignment: disables debug if necessary */
2004 if (ralign < ARCH_SLAB_MINALIGN) {
2005 ralign = ARCH_SLAB_MINALIGN;
2006 if (ralign > BYTES_PER_WORD)
2007 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2009 /* 3) caller mandated alignment: disables debug if necessary */
2010 if (ralign < align) {
2012 if (ralign > BYTES_PER_WORD)
2013 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2016 * 4) Store it. Note that the debug code below can reduce
2017 * the alignment to BYTES_PER_WORD.
2021 /* Get cache's description obj. */
2022 cachep = kmem_cache_zalloc(&cache_cache, SLAB_KERNEL);
2027 cachep->obj_size = size;
2029 if (flags & SLAB_RED_ZONE) {
2030 /* redzoning only works with word aligned caches */
2031 align = BYTES_PER_WORD;
2033 /* add space for red zone words */
2034 cachep->obj_offset += BYTES_PER_WORD;
2035 size += 2 * BYTES_PER_WORD;
2037 if (flags & SLAB_STORE_USER) {
2038 /* user store requires word alignment and
2039 * one word storage behind the end of the real
2042 align = BYTES_PER_WORD;
2043 size += BYTES_PER_WORD;
2045 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2046 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2047 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2048 cachep->obj_offset += PAGE_SIZE - size;
2054 /* Determine if the slab management is 'on' or 'off' slab. */
2055 if (size >= (PAGE_SIZE >> 3))
2057 * Size is large, assume best to place the slab management obj
2058 * off-slab (should allow better packing of objs).
2060 flags |= CFLGS_OFF_SLAB;
2062 size = ALIGN(size, align);
2064 left_over = calculate_slab_order(cachep, size, align, flags);
2067 printk("kmem_cache_create: couldn't create cache %s.\n", name);
2068 kmem_cache_free(&cache_cache, cachep);
2072 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2073 + sizeof(struct slab), align);
2076 * If the slab has been placed off-slab, and we have enough space then
2077 * move it on-slab. This is at the expense of any extra colouring.
2079 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2080 flags &= ~CFLGS_OFF_SLAB;
2081 left_over -= slab_size;
2084 if (flags & CFLGS_OFF_SLAB) {
2085 /* really off slab. No need for manual alignment */
2087 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2090 cachep->colour_off = cache_line_size();
2091 /* Offset must be a multiple of the alignment. */
2092 if (cachep->colour_off < align)
2093 cachep->colour_off = align;
2094 cachep->colour = left_over / cachep->colour_off;
2095 cachep->slab_size = slab_size;
2096 cachep->flags = flags;
2097 cachep->gfpflags = 0;
2098 if (flags & SLAB_CACHE_DMA)
2099 cachep->gfpflags |= GFP_DMA;
2100 cachep->buffer_size = size;
2102 if (flags & CFLGS_OFF_SLAB)
2103 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2104 cachep->ctor = ctor;
2105 cachep->dtor = dtor;
2106 cachep->name = name;
2109 setup_cpu_cache(cachep);
2111 /* cache setup completed, link it into the list */
2112 list_add(&cachep->next, &cache_chain);
2114 if (!cachep && (flags & SLAB_PANIC))
2115 panic("kmem_cache_create(): failed to create slab `%s'\n",
2117 mutex_unlock(&cache_chain_mutex);
2118 unlock_cpu_hotplug();
2121 EXPORT_SYMBOL(kmem_cache_create);
2124 static void check_irq_off(void)
2126 BUG_ON(!irqs_disabled());
2129 static void check_irq_on(void)
2131 BUG_ON(irqs_disabled());
2134 static void check_spinlock_acquired(struct kmem_cache *cachep)
2138 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2142 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2146 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2151 #define check_irq_off() do { } while(0)
2152 #define check_irq_on() do { } while(0)
2153 #define check_spinlock_acquired(x) do { } while(0)
2154 #define check_spinlock_acquired_node(x, y) do { } while(0)
2157 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2158 struct array_cache *ac,
2159 int force, int node);
2161 static void do_drain(void *arg)
2163 struct kmem_cache *cachep = arg;
2164 struct array_cache *ac;
2165 int node = numa_node_id();
2168 ac = cpu_cache_get(cachep);
2169 spin_lock(&cachep->nodelists[node]->list_lock);
2170 free_block(cachep, ac->entry, ac->avail, node);
2171 spin_unlock(&cachep->nodelists[node]->list_lock);
2175 static void drain_cpu_caches(struct kmem_cache *cachep)
2177 struct kmem_list3 *l3;
2180 on_each_cpu(do_drain, cachep, 1, 1);
2182 for_each_online_node(node) {
2183 l3 = cachep->nodelists[node];
2185 drain_array(cachep, l3, l3->shared, 1, node);
2187 drain_alien_cache(cachep, l3->alien);
2192 static int __node_shrink(struct kmem_cache *cachep, int node)
2195 struct kmem_list3 *l3 = cachep->nodelists[node];
2199 struct list_head *p;
2201 p = l3->slabs_free.prev;
2202 if (p == &l3->slabs_free)
2205 slabp = list_entry(l3->slabs_free.prev, struct slab, list);
2207 BUG_ON(slabp->inuse);
2209 list_del(&slabp->list);
2211 l3->free_objects -= cachep->num;
2212 spin_unlock_irq(&l3->list_lock);
2213 slab_destroy(cachep, slabp);
2214 spin_lock_irq(&l3->list_lock);
2216 ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial);
2220 static int __cache_shrink(struct kmem_cache *cachep)
2223 struct kmem_list3 *l3;
2225 drain_cpu_caches(cachep);
2228 for_each_online_node(i) {
2229 l3 = cachep->nodelists[i];
2231 spin_lock_irq(&l3->list_lock);
2232 ret += __node_shrink(cachep, i);
2233 spin_unlock_irq(&l3->list_lock);
2236 return (ret ? 1 : 0);
2240 * kmem_cache_shrink - Shrink a cache.
2241 * @cachep: The cache to shrink.
2243 * Releases as many slabs as possible for a cache.
2244 * To help debugging, a zero exit status indicates all slabs were released.
2246 int kmem_cache_shrink(struct kmem_cache *cachep)
2248 BUG_ON(!cachep || in_interrupt());
2250 return __cache_shrink(cachep);
2252 EXPORT_SYMBOL(kmem_cache_shrink);
2255 * kmem_cache_destroy - delete a cache
2256 * @cachep: the cache to destroy
2258 * Remove a struct kmem_cache object from the slab cache.
2259 * Returns 0 on success.
2261 * It is expected this function will be called by a module when it is
2262 * unloaded. This will remove the cache completely, and avoid a duplicate
2263 * cache being allocated each time a module is loaded and unloaded, if the
2264 * module doesn't have persistent in-kernel storage across loads and unloads.
2266 * The cache must be empty before calling this function.
2268 * The caller must guarantee that noone will allocate memory from the cache
2269 * during the kmem_cache_destroy().
2271 int kmem_cache_destroy(struct kmem_cache *cachep)
2274 struct kmem_list3 *l3;
2276 BUG_ON(!cachep || in_interrupt());
2278 /* Don't let CPUs to come and go */
2281 /* Find the cache in the chain of caches. */
2282 mutex_lock(&cache_chain_mutex);
2284 * the chain is never empty, cache_cache is never destroyed
2286 list_del(&cachep->next);
2287 mutex_unlock(&cache_chain_mutex);
2289 if (__cache_shrink(cachep)) {
2290 slab_error(cachep, "Can't free all objects");
2291 mutex_lock(&cache_chain_mutex);
2292 list_add(&cachep->next, &cache_chain);
2293 mutex_unlock(&cache_chain_mutex);
2294 unlock_cpu_hotplug();
2298 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2301 for_each_online_cpu(i)
2302 kfree(cachep->array[i]);
2304 /* NUMA: free the list3 structures */
2305 for_each_online_node(i) {
2306 l3 = cachep->nodelists[i];
2309 free_alien_cache(l3->alien);
2313 kmem_cache_free(&cache_cache, cachep);
2314 unlock_cpu_hotplug();
2317 EXPORT_SYMBOL(kmem_cache_destroy);
2319 /* Get the memory for a slab management obj. */
2320 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2321 int colour_off, gfp_t local_flags)
2325 if (OFF_SLAB(cachep)) {
2326 /* Slab management obj is off-slab. */
2327 slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags);
2331 slabp = objp + colour_off;
2332 colour_off += cachep->slab_size;
2335 slabp->colouroff = colour_off;
2336 slabp->s_mem = objp + colour_off;
2340 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2342 return (kmem_bufctl_t *) (slabp + 1);
2345 static void cache_init_objs(struct kmem_cache *cachep,
2346 struct slab *slabp, unsigned long ctor_flags)
2350 for (i = 0; i < cachep->num; i++) {
2351 void *objp = index_to_obj(cachep, slabp, i);
2353 /* need to poison the objs? */
2354 if (cachep->flags & SLAB_POISON)
2355 poison_obj(cachep, objp, POISON_FREE);
2356 if (cachep->flags & SLAB_STORE_USER)
2357 *dbg_userword(cachep, objp) = NULL;
2359 if (cachep->flags & SLAB_RED_ZONE) {
2360 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2361 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2364 * Constructors are not allowed to allocate memory from the same
2365 * cache which they are a constructor for. Otherwise, deadlock.
2366 * They must also be threaded.
2368 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2369 cachep->ctor(objp + obj_offset(cachep), cachep,
2372 if (cachep->flags & SLAB_RED_ZONE) {
2373 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2374 slab_error(cachep, "constructor overwrote the"
2375 " end of an object");
2376 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2377 slab_error(cachep, "constructor overwrote the"
2378 " start of an object");
2380 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2381 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2382 kernel_map_pages(virt_to_page(objp),
2383 cachep->buffer_size / PAGE_SIZE, 0);
2386 cachep->ctor(objp, cachep, ctor_flags);
2388 slab_bufctl(slabp)[i] = i + 1;
2390 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2394 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2396 if (flags & SLAB_DMA)
2397 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2399 BUG_ON(cachep->gfpflags & GFP_DMA);
2402 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2405 void *objp = index_to_obj(cachep, slabp, slabp->free);
2409 next = slab_bufctl(slabp)[slabp->free];
2411 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2412 WARN_ON(slabp->nodeid != nodeid);
2419 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2420 void *objp, int nodeid)
2422 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2425 /* Verify that the slab belongs to the intended node */
2426 WARN_ON(slabp->nodeid != nodeid);
2428 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2429 printk(KERN_ERR "slab: double free detected in cache "
2430 "'%s', objp %p\n", cachep->name, objp);
2434 slab_bufctl(slabp)[objnr] = slabp->free;
2435 slabp->free = objnr;
2439 static void set_slab_attr(struct kmem_cache *cachep, struct slab *slabp,
2445 /* Nasty!!!!!! I hope this is OK. */
2446 page = virt_to_page(objp);
2449 if (likely(!PageCompound(page)))
2450 i <<= cachep->gfporder;
2452 page_set_cache(page, cachep);
2453 page_set_slab(page, slabp);
2459 * Grow (by 1) the number of slabs within a cache. This is called by
2460 * kmem_cache_alloc() when there are no active objs left in a cache.
2462 static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid)
2468 unsigned long ctor_flags;
2469 struct kmem_list3 *l3;
2472 * Be lazy and only check for valid flags here, keeping it out of the
2473 * critical path in kmem_cache_alloc().
2475 BUG_ON(flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW));
2476 if (flags & SLAB_NO_GROW)
2479 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2480 local_flags = (flags & SLAB_LEVEL_MASK);
2481 if (!(local_flags & __GFP_WAIT))
2483 * Not allowed to sleep. Need to tell a constructor about
2484 * this - it might need to know...
2486 ctor_flags |= SLAB_CTOR_ATOMIC;
2488 /* Take the l3 list lock to change the colour_next on this node */
2490 l3 = cachep->nodelists[nodeid];
2491 spin_lock(&l3->list_lock);
2493 /* Get colour for the slab, and cal the next value. */
2494 offset = l3->colour_next;
2496 if (l3->colour_next >= cachep->colour)
2497 l3->colour_next = 0;
2498 spin_unlock(&l3->list_lock);
2500 offset *= cachep->colour_off;
2502 if (local_flags & __GFP_WAIT)
2506 * The test for missing atomic flag is performed here, rather than
2507 * the more obvious place, simply to reduce the critical path length
2508 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2509 * will eventually be caught here (where it matters).
2511 kmem_flagcheck(cachep, flags);
2514 * Get mem for the objs. Attempt to allocate a physical page from
2517 objp = kmem_getpages(cachep, flags, nodeid);
2521 /* Get slab management. */
2522 slabp = alloc_slabmgmt(cachep, objp, offset, local_flags);
2526 slabp->nodeid = nodeid;
2527 set_slab_attr(cachep, slabp, objp);
2529 cache_init_objs(cachep, slabp, ctor_flags);
2531 if (local_flags & __GFP_WAIT)
2532 local_irq_disable();
2534 spin_lock(&l3->list_lock);
2536 /* Make slab active. */
2537 list_add_tail(&slabp->list, &(l3->slabs_free));
2538 STATS_INC_GROWN(cachep);
2539 l3->free_objects += cachep->num;
2540 spin_unlock(&l3->list_lock);
2543 kmem_freepages(cachep, objp);
2545 if (local_flags & __GFP_WAIT)
2546 local_irq_disable();
2553 * Perform extra freeing checks:
2554 * - detect bad pointers.
2555 * - POISON/RED_ZONE checking
2556 * - destructor calls, for caches with POISON+dtor
2558 static void kfree_debugcheck(const void *objp)
2562 if (!virt_addr_valid(objp)) {
2563 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2564 (unsigned long)objp);
2567 page = virt_to_page(objp);
2568 if (!PageSlab(page)) {
2569 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2570 (unsigned long)objp);
2575 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2582 objp -= obj_offset(cachep);
2583 kfree_debugcheck(objp);
2584 page = virt_to_page(objp);
2586 if (page_get_cache(page) != cachep) {
2587 printk(KERN_ERR "mismatch in kmem_cache_free: expected "
2588 "cache %p, got %p\n",
2589 page_get_cache(page), cachep);
2590 printk(KERN_ERR "%p is %s.\n", cachep, cachep->name);
2591 printk(KERN_ERR "%p is %s.\n", page_get_cache(page),
2592 page_get_cache(page)->name);
2595 slabp = page_get_slab(page);
2597 if (cachep->flags & SLAB_RED_ZONE) {
2598 if (*dbg_redzone1(cachep, objp) != RED_ACTIVE ||
2599 *dbg_redzone2(cachep, objp) != RED_ACTIVE) {
2600 slab_error(cachep, "double free, or memory outside"
2601 " object was overwritten");
2602 printk(KERN_ERR "%p: redzone 1:0x%lx, "
2603 "redzone 2:0x%lx.\n",
2604 objp, *dbg_redzone1(cachep, objp),
2605 *dbg_redzone2(cachep, objp));
2607 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2608 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2610 if (cachep->flags & SLAB_STORE_USER)
2611 *dbg_userword(cachep, objp) = caller;
2613 objnr = obj_to_index(cachep, slabp, objp);
2615 BUG_ON(objnr >= cachep->num);
2616 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2618 if (cachep->flags & SLAB_DEBUG_INITIAL) {
2620 * Need to call the slab's constructor so the caller can
2621 * perform a verify of its state (debugging). Called without
2622 * the cache-lock held.
2624 cachep->ctor(objp + obj_offset(cachep),
2625 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
2627 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2628 /* we want to cache poison the object,
2629 * call the destruction callback
2631 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
2633 #ifdef CONFIG_DEBUG_SLAB_LEAK
2634 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2636 if (cachep->flags & SLAB_POISON) {
2637 #ifdef CONFIG_DEBUG_PAGEALLOC
2638 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2639 store_stackinfo(cachep, objp, (unsigned long)caller);
2640 kernel_map_pages(virt_to_page(objp),
2641 cachep->buffer_size / PAGE_SIZE, 0);
2643 poison_obj(cachep, objp, POISON_FREE);
2646 poison_obj(cachep, objp, POISON_FREE);
2652 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2657 /* Check slab's freelist to see if this obj is there. */
2658 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2660 if (entries > cachep->num || i >= cachep->num)
2663 if (entries != cachep->num - slabp->inuse) {
2665 printk(KERN_ERR "slab: Internal list corruption detected in "
2666 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2667 cachep->name, cachep->num, slabp, slabp->inuse);
2669 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2672 printk("\n%03x:", i);
2673 printk(" %02x", ((unsigned char *)slabp)[i]);
2680 #define kfree_debugcheck(x) do { } while(0)
2681 #define cache_free_debugcheck(x,objp,z) (objp)
2682 #define check_slabp(x,y) do { } while(0)
2685 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2688 struct kmem_list3 *l3;
2689 struct array_cache *ac;
2692 ac = cpu_cache_get(cachep);
2694 batchcount = ac->batchcount;
2695 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2697 * If there was little recent activity on this cache, then
2698 * perform only a partial refill. Otherwise we could generate
2701 batchcount = BATCHREFILL_LIMIT;
2703 l3 = cachep->nodelists[numa_node_id()];
2705 BUG_ON(ac->avail > 0 || !l3);
2706 spin_lock(&l3->list_lock);
2708 /* See if we can refill from the shared array */
2709 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2712 while (batchcount > 0) {
2713 struct list_head *entry;
2715 /* Get slab alloc is to come from. */
2716 entry = l3->slabs_partial.next;
2717 if (entry == &l3->slabs_partial) {
2718 l3->free_touched = 1;
2719 entry = l3->slabs_free.next;
2720 if (entry == &l3->slabs_free)
2724 slabp = list_entry(entry, struct slab, list);
2725 check_slabp(cachep, slabp);
2726 check_spinlock_acquired(cachep);
2727 while (slabp->inuse < cachep->num && batchcount--) {
2728 STATS_INC_ALLOCED(cachep);
2729 STATS_INC_ACTIVE(cachep);
2730 STATS_SET_HIGH(cachep);
2732 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2735 check_slabp(cachep, slabp);
2737 /* move slabp to correct slabp list: */
2738 list_del(&slabp->list);
2739 if (slabp->free == BUFCTL_END)
2740 list_add(&slabp->list, &l3->slabs_full);
2742 list_add(&slabp->list, &l3->slabs_partial);
2746 l3->free_objects -= ac->avail;
2748 spin_unlock(&l3->list_lock);
2750 if (unlikely(!ac->avail)) {
2752 x = cache_grow(cachep, flags, numa_node_id());
2754 /* cache_grow can reenable interrupts, then ac could change. */
2755 ac = cpu_cache_get(cachep);
2756 if (!x && ac->avail == 0) /* no objects in sight? abort */
2759 if (!ac->avail) /* objects refilled by interrupt? */
2763 return ac->entry[--ac->avail];
2766 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2769 might_sleep_if(flags & __GFP_WAIT);
2771 kmem_flagcheck(cachep, flags);
2776 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2777 gfp_t flags, void *objp, void *caller)
2781 if (cachep->flags & SLAB_POISON) {
2782 #ifdef CONFIG_DEBUG_PAGEALLOC
2783 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
2784 kernel_map_pages(virt_to_page(objp),
2785 cachep->buffer_size / PAGE_SIZE, 1);
2787 check_poison_obj(cachep, objp);
2789 check_poison_obj(cachep, objp);
2791 poison_obj(cachep, objp, POISON_INUSE);
2793 if (cachep->flags & SLAB_STORE_USER)
2794 *dbg_userword(cachep, objp) = caller;
2796 if (cachep->flags & SLAB_RED_ZONE) {
2797 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2798 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2799 slab_error(cachep, "double free, or memory outside"
2800 " object was overwritten");
2802 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
2803 objp, *dbg_redzone1(cachep, objp),
2804 *dbg_redzone2(cachep, objp));
2806 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2807 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2809 #ifdef CONFIG_DEBUG_SLAB_LEAK
2814 slabp = page_get_slab(virt_to_page(objp));
2815 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
2816 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
2819 objp += obj_offset(cachep);
2820 if (cachep->ctor && cachep->flags & SLAB_POISON) {
2821 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2823 if (!(flags & __GFP_WAIT))
2824 ctor_flags |= SLAB_CTOR_ATOMIC;
2826 cachep->ctor(objp, cachep, ctor_flags);
2831 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2834 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
2837 struct array_cache *ac;
2840 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
2841 objp = alternate_node_alloc(cachep, flags);
2848 ac = cpu_cache_get(cachep);
2849 if (likely(ac->avail)) {
2850 STATS_INC_ALLOCHIT(cachep);
2852 objp = ac->entry[--ac->avail];
2854 STATS_INC_ALLOCMISS(cachep);
2855 objp = cache_alloc_refill(cachep, flags);
2860 static __always_inline void *__cache_alloc(struct kmem_cache *cachep,
2861 gfp_t flags, void *caller)
2863 unsigned long save_flags;
2866 cache_alloc_debugcheck_before(cachep, flags);
2868 local_irq_save(save_flags);
2869 objp = ____cache_alloc(cachep, flags);
2870 local_irq_restore(save_flags);
2871 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
2879 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
2881 * If we are in_interrupt, then process context, including cpusets and
2882 * mempolicy, may not apply and should not be used for allocation policy.
2884 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
2886 int nid_alloc, nid_here;
2890 nid_alloc = nid_here = numa_node_id();
2891 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
2892 nid_alloc = cpuset_mem_spread_node();
2893 else if (current->mempolicy)
2894 nid_alloc = slab_node(current->mempolicy);
2895 if (nid_alloc != nid_here)
2896 return __cache_alloc_node(cachep, flags, nid_alloc);
2901 * A interface to enable slab creation on nodeid
2903 static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
2906 struct list_head *entry;
2908 struct kmem_list3 *l3;
2912 l3 = cachep->nodelists[nodeid];
2917 spin_lock(&l3->list_lock);
2918 entry = l3->slabs_partial.next;
2919 if (entry == &l3->slabs_partial) {
2920 l3->free_touched = 1;
2921 entry = l3->slabs_free.next;
2922 if (entry == &l3->slabs_free)
2926 slabp = list_entry(entry, struct slab, list);
2927 check_spinlock_acquired_node(cachep, nodeid);
2928 check_slabp(cachep, slabp);
2930 STATS_INC_NODEALLOCS(cachep);
2931 STATS_INC_ACTIVE(cachep);
2932 STATS_SET_HIGH(cachep);
2934 BUG_ON(slabp->inuse == cachep->num);
2936 obj = slab_get_obj(cachep, slabp, nodeid);
2937 check_slabp(cachep, slabp);
2939 /* move slabp to correct slabp list: */
2940 list_del(&slabp->list);
2942 if (slabp->free == BUFCTL_END)
2943 list_add(&slabp->list, &l3->slabs_full);
2945 list_add(&slabp->list, &l3->slabs_partial);
2947 spin_unlock(&l3->list_lock);
2951 spin_unlock(&l3->list_lock);
2952 x = cache_grow(cachep, flags, nodeid);
2964 * Caller needs to acquire correct kmem_list's list_lock
2966 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
2970 struct kmem_list3 *l3;
2972 for (i = 0; i < nr_objects; i++) {
2973 void *objp = objpp[i];
2976 slabp = virt_to_slab(objp);
2977 l3 = cachep->nodelists[node];
2978 list_del(&slabp->list);
2979 check_spinlock_acquired_node(cachep, node);
2980 check_slabp(cachep, slabp);
2981 slab_put_obj(cachep, slabp, objp, node);
2982 STATS_DEC_ACTIVE(cachep);
2984 check_slabp(cachep, slabp);
2986 /* fixup slab chains */
2987 if (slabp->inuse == 0) {
2988 if (l3->free_objects > l3->free_limit) {
2989 l3->free_objects -= cachep->num;
2990 slab_destroy(cachep, slabp);
2992 list_add(&slabp->list, &l3->slabs_free);
2995 /* Unconditionally move a slab to the end of the
2996 * partial list on free - maximum time for the
2997 * other objects to be freed, too.
2999 list_add_tail(&slabp->list, &l3->slabs_partial);
3004 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3007 struct kmem_list3 *l3;
3008 int node = numa_node_id();
3010 batchcount = ac->batchcount;
3012 BUG_ON(!batchcount || batchcount > ac->avail);
3015 l3 = cachep->nodelists[node];
3016 spin_lock(&l3->list_lock);
3018 struct array_cache *shared_array = l3->shared;
3019 int max = shared_array->limit - shared_array->avail;
3021 if (batchcount > max)
3023 memcpy(&(shared_array->entry[shared_array->avail]),
3024 ac->entry, sizeof(void *) * batchcount);
3025 shared_array->avail += batchcount;
3030 free_block(cachep, ac->entry, batchcount, node);
3035 struct list_head *p;
3037 p = l3->slabs_free.next;
3038 while (p != &(l3->slabs_free)) {
3041 slabp = list_entry(p, struct slab, list);
3042 BUG_ON(slabp->inuse);
3047 STATS_SET_FREEABLE(cachep, i);
3050 spin_unlock(&l3->list_lock);
3051 ac->avail -= batchcount;
3052 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3056 * Release an obj back to its cache. If the obj has a constructed state, it must
3057 * be in this state _before_ it is released. Called with disabled ints.
3059 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3061 struct array_cache *ac = cpu_cache_get(cachep);
3064 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3066 /* Make sure we are not freeing a object from another
3067 * node to the array cache on this cpu.
3072 slabp = virt_to_slab(objp);
3073 if (unlikely(slabp->nodeid != numa_node_id())) {
3074 struct array_cache *alien = NULL;
3075 int nodeid = slabp->nodeid;
3076 struct kmem_list3 *l3;
3078 l3 = cachep->nodelists[numa_node_id()];
3079 STATS_INC_NODEFREES(cachep);
3080 if (l3->alien && l3->alien[nodeid]) {
3081 alien = l3->alien[nodeid];
3082 spin_lock(&alien->lock);
3083 if (unlikely(alien->avail == alien->limit))
3084 __drain_alien_cache(cachep,
3086 alien->entry[alien->avail++] = objp;
3087 spin_unlock(&alien->lock);
3089 spin_lock(&(cachep->nodelists[nodeid])->
3091 free_block(cachep, &objp, 1, nodeid);
3092 spin_unlock(&(cachep->nodelists[nodeid])->
3099 if (likely(ac->avail < ac->limit)) {
3100 STATS_INC_FREEHIT(cachep);
3101 ac->entry[ac->avail++] = objp;
3104 STATS_INC_FREEMISS(cachep);
3105 cache_flusharray(cachep, ac);
3106 ac->entry[ac->avail++] = objp;
3111 * kmem_cache_alloc - Allocate an object
3112 * @cachep: The cache to allocate from.
3113 * @flags: See kmalloc().
3115 * Allocate an object from this cache. The flags are only relevant
3116 * if the cache has no available objects.
3118 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3120 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3122 EXPORT_SYMBOL(kmem_cache_alloc);
3125 * kmem_cache_alloc - Allocate an object. The memory is set to zero.
3126 * @cache: The cache to allocate from.
3127 * @flags: See kmalloc().
3129 * Allocate an object from this cache and set the allocated memory to zero.
3130 * The flags are only relevant if the cache has no available objects.
3132 void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
3134 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
3136 memset(ret, 0, obj_size(cache));
3139 EXPORT_SYMBOL(kmem_cache_zalloc);
3142 * kmem_ptr_validate - check if an untrusted pointer might
3144 * @cachep: the cache we're checking against
3145 * @ptr: pointer to validate
3147 * This verifies that the untrusted pointer looks sane:
3148 * it is _not_ a guarantee that the pointer is actually
3149 * part of the slab cache in question, but it at least
3150 * validates that the pointer can be dereferenced and
3151 * looks half-way sane.
3153 * Currently only used for dentry validation.
3155 int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr)
3157 unsigned long addr = (unsigned long)ptr;
3158 unsigned long min_addr = PAGE_OFFSET;
3159 unsigned long align_mask = BYTES_PER_WORD - 1;
3160 unsigned long size = cachep->buffer_size;
3163 if (unlikely(addr < min_addr))
3165 if (unlikely(addr > (unsigned long)high_memory - size))
3167 if (unlikely(addr & align_mask))
3169 if (unlikely(!kern_addr_valid(addr)))
3171 if (unlikely(!kern_addr_valid(addr + size - 1)))
3173 page = virt_to_page(ptr);
3174 if (unlikely(!PageSlab(page)))
3176 if (unlikely(page_get_cache(page) != cachep))
3185 * kmem_cache_alloc_node - Allocate an object on the specified node
3186 * @cachep: The cache to allocate from.
3187 * @flags: See kmalloc().
3188 * @nodeid: node number of the target node.
3190 * Identical to kmem_cache_alloc, except that this function is slow
3191 * and can sleep. And it will allocate memory on the given node, which
3192 * can improve the performance for cpu bound structures.
3193 * New and improved: it will now make sure that the object gets
3194 * put on the correct node list so that there is no false sharing.
3196 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3198 unsigned long save_flags;
3201 cache_alloc_debugcheck_before(cachep, flags);
3202 local_irq_save(save_flags);
3204 if (nodeid == -1 || nodeid == numa_node_id() ||
3205 !cachep->nodelists[nodeid])
3206 ptr = ____cache_alloc(cachep, flags);
3208 ptr = __cache_alloc_node(cachep, flags, nodeid);
3209 local_irq_restore(save_flags);
3211 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr,
3212 __builtin_return_address(0));
3216 EXPORT_SYMBOL(kmem_cache_alloc_node);
3218 void *kmalloc_node(size_t size, gfp_t flags, int node)
3220 struct kmem_cache *cachep;
3222 cachep = kmem_find_general_cachep(size, flags);
3223 if (unlikely(cachep == NULL))
3225 return kmem_cache_alloc_node(cachep, flags, node);
3227 EXPORT_SYMBOL(kmalloc_node);
3231 * kmalloc - allocate memory
3232 * @size: how many bytes of memory are required.
3233 * @flags: the type of memory to allocate.
3234 * @caller: function caller for debug tracking of the caller
3236 * kmalloc is the normal method of allocating memory
3239 * The @flags argument may be one of:
3241 * %GFP_USER - Allocate memory on behalf of user. May sleep.
3243 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
3245 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
3247 * Additionally, the %GFP_DMA flag may be set to indicate the memory
3248 * must be suitable for DMA. This can mean different things on different
3249 * platforms. For example, on i386, it means that the memory must come
3250 * from the first 16MB.
3252 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3255 struct kmem_cache *cachep;
3257 /* If you want to save a few bytes .text space: replace
3259 * Then kmalloc uses the uninlined functions instead of the inline
3262 cachep = __find_general_cachep(size, flags);
3263 if (unlikely(cachep == NULL))
3265 return __cache_alloc(cachep, flags, caller);
3269 void *__kmalloc(size_t size, gfp_t flags)
3271 #ifndef CONFIG_DEBUG_SLAB
3272 return __do_kmalloc(size, flags, NULL);
3274 return __do_kmalloc(size, flags, __builtin_return_address(0));
3277 EXPORT_SYMBOL(__kmalloc);
3279 #ifdef CONFIG_DEBUG_SLAB
3280 void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3282 return __do_kmalloc(size, flags, caller);
3284 EXPORT_SYMBOL(__kmalloc_track_caller);
3289 * __alloc_percpu - allocate one copy of the object for every present
3290 * cpu in the system, zeroing them.
3291 * Objects should be dereferenced using the per_cpu_ptr macro only.
3293 * @size: how many bytes of memory are required.
3295 void *__alloc_percpu(size_t size)
3298 struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL);
3304 * Cannot use for_each_online_cpu since a cpu may come online
3305 * and we have no way of figuring out how to fix the array
3306 * that we have allocated then....
3308 for_each_possible_cpu(i) {
3309 int node = cpu_to_node(i);
3311 if (node_online(node))
3312 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
3314 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
3316 if (!pdata->ptrs[i])
3318 memset(pdata->ptrs[i], 0, size);
3321 /* Catch derefs w/o wrappers */
3322 return (void *)(~(unsigned long)pdata);
3326 if (!cpu_possible(i))
3328 kfree(pdata->ptrs[i]);
3333 EXPORT_SYMBOL(__alloc_percpu);
3337 * kmem_cache_free - Deallocate an object
3338 * @cachep: The cache the allocation was from.
3339 * @objp: The previously allocated object.
3341 * Free an object which was previously allocated from this
3344 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3346 unsigned long flags;
3348 local_irq_save(flags);
3349 __cache_free(cachep, objp);
3350 local_irq_restore(flags);
3352 EXPORT_SYMBOL(kmem_cache_free);
3355 * kfree - free previously allocated memory
3356 * @objp: pointer returned by kmalloc.
3358 * If @objp is NULL, no operation is performed.
3360 * Don't free memory not originally allocated by kmalloc()
3361 * or you will run into trouble.
3363 void kfree(const void *objp)
3365 struct kmem_cache *c;
3366 unsigned long flags;
3368 if (unlikely(!objp))
3370 local_irq_save(flags);
3371 kfree_debugcheck(objp);
3372 c = virt_to_cache(objp);
3373 mutex_debug_check_no_locks_freed(objp, obj_size(c));
3374 __cache_free(c, (void *)objp);
3375 local_irq_restore(flags);
3377 EXPORT_SYMBOL(kfree);
3381 * free_percpu - free previously allocated percpu memory
3382 * @objp: pointer returned by alloc_percpu.
3384 * Don't free memory not originally allocated by alloc_percpu()
3385 * The complemented objp is to check for that.
3387 void free_percpu(const void *objp)
3390 struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp);
3393 * We allocate for all cpus so we cannot use for online cpu here.
3395 for_each_possible_cpu(i)
3399 EXPORT_SYMBOL(free_percpu);
3402 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3404 return obj_size(cachep);
3406 EXPORT_SYMBOL(kmem_cache_size);
3408 const char *kmem_cache_name(struct kmem_cache *cachep)
3410 return cachep->name;
3412 EXPORT_SYMBOL_GPL(kmem_cache_name);
3415 * This initializes kmem_list3 or resizes varioius caches for all nodes.
3417 static int alloc_kmemlist(struct kmem_cache *cachep)
3420 struct kmem_list3 *l3;
3421 struct array_cache *new_shared;
3422 struct array_cache **new_alien;
3424 for_each_online_node(node) {
3426 new_alien = alloc_alien_cache(node, cachep->limit);
3430 new_shared = alloc_arraycache(node,
3431 cachep->shared*cachep->batchcount,
3434 free_alien_cache(new_alien);
3438 l3 = cachep->nodelists[node];
3440 struct array_cache *shared = l3->shared;
3442 spin_lock_irq(&l3->list_lock);
3445 free_block(cachep, shared->entry,
3446 shared->avail, node);
3448 l3->shared = new_shared;
3450 l3->alien = new_alien;
3453 l3->free_limit = (1 + nr_cpus_node(node)) *
3454 cachep->batchcount + cachep->num;
3455 spin_unlock_irq(&l3->list_lock);
3457 free_alien_cache(new_alien);
3460 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
3462 free_alien_cache(new_alien);
3467 kmem_list3_init(l3);
3468 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3469 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3470 l3->shared = new_shared;
3471 l3->alien = new_alien;
3472 l3->free_limit = (1 + nr_cpus_node(node)) *
3473 cachep->batchcount + cachep->num;
3474 cachep->nodelists[node] = l3;
3479 if (!cachep->next.next) {
3480 /* Cache is not active yet. Roll back what we did */
3483 if (cachep->nodelists[node]) {
3484 l3 = cachep->nodelists[node];
3487 free_alien_cache(l3->alien);
3489 cachep->nodelists[node] = NULL;
3497 struct ccupdate_struct {
3498 struct kmem_cache *cachep;
3499 struct array_cache *new[NR_CPUS];
3502 static void do_ccupdate_local(void *info)
3504 struct ccupdate_struct *new = info;
3505 struct array_cache *old;
3508 old = cpu_cache_get(new->cachep);
3510 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3511 new->new[smp_processor_id()] = old;
3514 /* Always called with the cache_chain_mutex held */
3515 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3516 int batchcount, int shared)
3518 struct ccupdate_struct new;
3521 memset(&new.new, 0, sizeof(new.new));
3522 for_each_online_cpu(i) {
3523 new.new[i] = alloc_arraycache(cpu_to_node(i), limit,
3526 for (i--; i >= 0; i--)
3531 new.cachep = cachep;
3533 on_each_cpu(do_ccupdate_local, (void *)&new, 1, 1);
3536 cachep->batchcount = batchcount;
3537 cachep->limit = limit;
3538 cachep->shared = shared;
3540 for_each_online_cpu(i) {
3541 struct array_cache *ccold = new.new[i];
3544 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3545 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3546 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3550 err = alloc_kmemlist(cachep);
3552 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
3553 cachep->name, -err);
3559 /* Called with cache_chain_mutex held always */
3560 static void enable_cpucache(struct kmem_cache *cachep)
3566 * The head array serves three purposes:
3567 * - create a LIFO ordering, i.e. return objects that are cache-warm
3568 * - reduce the number of spinlock operations.
3569 * - reduce the number of linked list operations on the slab and
3570 * bufctl chains: array operations are cheaper.
3571 * The numbers are guessed, we should auto-tune as described by
3574 if (cachep->buffer_size > 131072)
3576 else if (cachep->buffer_size > PAGE_SIZE)
3578 else if (cachep->buffer_size > 1024)
3580 else if (cachep->buffer_size > 256)
3586 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3587 * allocation behaviour: Most allocs on one cpu, most free operations
3588 * on another cpu. For these cases, an efficient object passing between
3589 * cpus is necessary. This is provided by a shared array. The array
3590 * replaces Bonwick's magazine layer.
3591 * On uniprocessor, it's functionally equivalent (but less efficient)
3592 * to a larger limit. Thus disabled by default.
3596 if (cachep->buffer_size <= PAGE_SIZE)
3602 * With debugging enabled, large batchcount lead to excessively long
3603 * periods with disabled local interrupts. Limit the batchcount
3608 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
3610 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3611 cachep->name, -err);
3615 * Drain an array if it contains any elements taking the l3 lock only if
3616 * necessary. Note that the l3 listlock also protects the array_cache
3617 * if drain_array() is used on the shared array.
3619 void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
3620 struct array_cache *ac, int force, int node)
3624 if (!ac || !ac->avail)
3626 if (ac->touched && !force) {
3629 spin_lock_irq(&l3->list_lock);
3631 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3632 if (tofree > ac->avail)
3633 tofree = (ac->avail + 1) / 2;
3634 free_block(cachep, ac->entry, tofree, node);
3635 ac->avail -= tofree;
3636 memmove(ac->entry, &(ac->entry[tofree]),
3637 sizeof(void *) * ac->avail);
3639 spin_unlock_irq(&l3->list_lock);
3644 * cache_reap - Reclaim memory from caches.
3645 * @unused: unused parameter
3647 * Called from workqueue/eventd every few seconds.
3649 * - clear the per-cpu caches for this CPU.
3650 * - return freeable pages to the main free memory pool.
3652 * If we cannot acquire the cache chain mutex then just give up - we'll try
3653 * again on the next iteration.
3655 static void cache_reap(void *unused)
3657 struct list_head *walk;
3658 struct kmem_list3 *l3;
3659 int node = numa_node_id();
3661 if (!mutex_trylock(&cache_chain_mutex)) {
3662 /* Give up. Setup the next iteration. */
3663 schedule_delayed_work(&__get_cpu_var(reap_work),
3668 list_for_each(walk, &cache_chain) {
3669 struct kmem_cache *searchp;
3670 struct list_head *p;
3674 searchp = list_entry(walk, struct kmem_cache, next);
3678 * We only take the l3 lock if absolutely necessary and we
3679 * have established with reasonable certainty that
3680 * we can do some work if the lock was obtained.
3682 l3 = searchp->nodelists[node];
3684 reap_alien(searchp, l3);
3686 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
3689 * These are racy checks but it does not matter
3690 * if we skip one check or scan twice.
3692 if (time_after(l3->next_reap, jiffies))
3695 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
3697 drain_array(searchp, l3, l3->shared, 0, node);
3699 if (l3->free_touched) {
3700 l3->free_touched = 0;
3704 tofree = (l3->free_limit + 5 * searchp->num - 1) /
3708 * Do not lock if there are no free blocks.
3710 if (list_empty(&l3->slabs_free))
3713 spin_lock_irq(&l3->list_lock);
3714 p = l3->slabs_free.next;
3715 if (p == &(l3->slabs_free)) {
3716 spin_unlock_irq(&l3->list_lock);
3720 slabp = list_entry(p, struct slab, list);
3721 BUG_ON(slabp->inuse);
3722 list_del(&slabp->list);
3723 STATS_INC_REAPED(searchp);
3726 * Safe to drop the lock. The slab is no longer linked
3727 * to the cache. searchp cannot disappear, we hold
3730 l3->free_objects -= searchp->num;
3731 spin_unlock_irq(&l3->list_lock);
3732 slab_destroy(searchp, slabp);
3733 } while (--tofree > 0);
3738 mutex_unlock(&cache_chain_mutex);
3740 /* Set up the next iteration */
3741 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
3744 #ifdef CONFIG_PROC_FS
3746 static void print_slabinfo_header(struct seq_file *m)
3749 * Output format version, so at least we can change it
3750 * without _too_ many complaints.
3753 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3755 seq_puts(m, "slabinfo - version: 2.1\n");
3757 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3758 "<objperslab> <pagesperslab>");
3759 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3760 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3762 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3763 "<error> <maxfreeable> <nodeallocs> <remotefrees>");
3764 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3769 static void *s_start(struct seq_file *m, loff_t *pos)
3772 struct list_head *p;
3774 mutex_lock(&cache_chain_mutex);
3776 print_slabinfo_header(m);
3777 p = cache_chain.next;
3780 if (p == &cache_chain)
3783 return list_entry(p, struct kmem_cache, next);
3786 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3788 struct kmem_cache *cachep = p;
3790 return cachep->next.next == &cache_chain ?
3791 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
3794 static void s_stop(struct seq_file *m, void *p)
3796 mutex_unlock(&cache_chain_mutex);
3799 static int s_show(struct seq_file *m, void *p)
3801 struct kmem_cache *cachep = p;
3802 struct list_head *q;
3804 unsigned long active_objs;
3805 unsigned long num_objs;
3806 unsigned long active_slabs = 0;
3807 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3811 struct kmem_list3 *l3;
3815 for_each_online_node(node) {
3816 l3 = cachep->nodelists[node];
3821 spin_lock_irq(&l3->list_lock);
3823 list_for_each(q, &l3->slabs_full) {
3824 slabp = list_entry(q, struct slab, list);
3825 if (slabp->inuse != cachep->num && !error)
3826 error = "slabs_full accounting error";
3827 active_objs += cachep->num;
3830 list_for_each(q, &l3->slabs_partial) {
3831 slabp = list_entry(q, struct slab, list);
3832 if (slabp->inuse == cachep->num && !error)
3833 error = "slabs_partial inuse accounting error";
3834 if (!slabp->inuse && !error)
3835 error = "slabs_partial/inuse accounting error";
3836 active_objs += slabp->inuse;
3839 list_for_each(q, &l3->slabs_free) {
3840 slabp = list_entry(q, struct slab, list);
3841 if (slabp->inuse && !error)
3842 error = "slabs_free/inuse accounting error";
3845 free_objects += l3->free_objects;
3847 shared_avail += l3->shared->avail;
3849 spin_unlock_irq(&l3->list_lock);
3851 num_slabs += active_slabs;
3852 num_objs = num_slabs * cachep->num;
3853 if (num_objs - active_objs != free_objects && !error)
3854 error = "free_objects accounting error";
3856 name = cachep->name;
3858 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3860 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
3861 name, active_objs, num_objs, cachep->buffer_size,
3862 cachep->num, (1 << cachep->gfporder));
3863 seq_printf(m, " : tunables %4u %4u %4u",
3864 cachep->limit, cachep->batchcount, cachep->shared);
3865 seq_printf(m, " : slabdata %6lu %6lu %6lu",
3866 active_slabs, num_slabs, shared_avail);
3869 unsigned long high = cachep->high_mark;
3870 unsigned long allocs = cachep->num_allocations;
3871 unsigned long grown = cachep->grown;
3872 unsigned long reaped = cachep->reaped;
3873 unsigned long errors = cachep->errors;
3874 unsigned long max_freeable = cachep->max_freeable;
3875 unsigned long node_allocs = cachep->node_allocs;
3876 unsigned long node_frees = cachep->node_frees;
3878 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
3879 %4lu %4lu %4lu %4lu", allocs, high, grown,
3880 reaped, errors, max_freeable, node_allocs,
3885 unsigned long allochit = atomic_read(&cachep->allochit);
3886 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3887 unsigned long freehit = atomic_read(&cachep->freehit);
3888 unsigned long freemiss = atomic_read(&cachep->freemiss);
3890 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
3891 allochit, allocmiss, freehit, freemiss);
3899 * slabinfo_op - iterator that generates /proc/slabinfo
3908 * num-pages-per-slab
3909 * + further values on SMP and with statistics enabled
3912 struct seq_operations slabinfo_op = {
3919 #define MAX_SLABINFO_WRITE 128
3921 * slabinfo_write - Tuning for the slab allocator
3923 * @buffer: user buffer
3924 * @count: data length
3927 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
3928 size_t count, loff_t *ppos)
3930 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
3931 int limit, batchcount, shared, res;
3932 struct list_head *p;
3934 if (count > MAX_SLABINFO_WRITE)
3936 if (copy_from_user(&kbuf, buffer, count))
3938 kbuf[MAX_SLABINFO_WRITE] = '\0';
3940 tmp = strchr(kbuf, ' ');
3945 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3948 /* Find the cache in the chain of caches. */
3949 mutex_lock(&cache_chain_mutex);
3951 list_for_each(p, &cache_chain) {
3952 struct kmem_cache *cachep;
3954 cachep = list_entry(p, struct kmem_cache, next);
3955 if (!strcmp(cachep->name, kbuf)) {
3956 if (limit < 1 || batchcount < 1 ||
3957 batchcount > limit || shared < 0) {
3960 res = do_tune_cpucache(cachep, limit,
3961 batchcount, shared);
3966 mutex_unlock(&cache_chain_mutex);
3972 #ifdef CONFIG_DEBUG_SLAB_LEAK
3974 static void *leaks_start(struct seq_file *m, loff_t *pos)
3977 struct list_head *p;
3979 mutex_lock(&cache_chain_mutex);
3980 p = cache_chain.next;
3983 if (p == &cache_chain)
3986 return list_entry(p, struct kmem_cache, next);
3989 static inline int add_caller(unsigned long *n, unsigned long v)
3999 unsigned long *q = p + 2 * i;
4013 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4019 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4025 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4026 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4028 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4033 static void show_symbol(struct seq_file *m, unsigned long address)
4035 #ifdef CONFIG_KALLSYMS
4038 unsigned long offset, size;
4039 char namebuf[KSYM_NAME_LEN+1];
4041 name = kallsyms_lookup(address, &size, &offset, &modname, namebuf);
4044 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4046 seq_printf(m, " [%s]", modname);
4050 seq_printf(m, "%p", (void *)address);
4053 static int leaks_show(struct seq_file *m, void *p)
4055 struct kmem_cache *cachep = p;
4056 struct list_head *q;
4058 struct kmem_list3 *l3;
4060 unsigned long *n = m->private;
4064 if (!(cachep->flags & SLAB_STORE_USER))
4066 if (!(cachep->flags & SLAB_RED_ZONE))
4069 /* OK, we can do it */
4073 for_each_online_node(node) {
4074 l3 = cachep->nodelists[node];
4079 spin_lock_irq(&l3->list_lock);
4081 list_for_each(q, &l3->slabs_full) {
4082 slabp = list_entry(q, struct slab, list);
4083 handle_slab(n, cachep, slabp);
4085 list_for_each(q, &l3->slabs_partial) {
4086 slabp = list_entry(q, struct slab, list);
4087 handle_slab(n, cachep, slabp);
4089 spin_unlock_irq(&l3->list_lock);
4091 name = cachep->name;
4093 /* Increase the buffer size */
4094 mutex_unlock(&cache_chain_mutex);
4095 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4097 /* Too bad, we are really out */
4099 mutex_lock(&cache_chain_mutex);
4102 *(unsigned long *)m->private = n[0] * 2;
4104 mutex_lock(&cache_chain_mutex);
4105 /* Now make sure this entry will be retried */
4109 for (i = 0; i < n[1]; i++) {
4110 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4111 show_symbol(m, n[2*i+2]);
4117 struct seq_operations slabstats_op = {
4118 .start = leaks_start,
4127 * ksize - get the actual amount of memory allocated for a given object
4128 * @objp: Pointer to the object
4130 * kmalloc may internally round up allocations and return more memory
4131 * than requested. ksize() can be used to determine the actual amount of
4132 * memory allocated. The caller may use this additional memory, even though
4133 * a smaller amount of memory was initially specified with the kmalloc call.
4134 * The caller must guarantee that objp points to a valid object previously
4135 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4136 * must not be freed during the duration of the call.
4138 unsigned int ksize(const void *objp)
4140 if (unlikely(objp == NULL))
4143 return obj_size(virt_to_cache(objp));