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/poison.h>
93 #include <linux/swap.h>
94 #include <linux/cache.h>
95 #include <linux/interrupt.h>
96 #include <linux/init.h>
97 #include <linux/compiler.h>
98 #include <linux/cpuset.h>
99 #include <linux/seq_file.h>
100 #include <linux/notifier.h>
101 #include <linux/kallsyms.h>
102 #include <linux/cpu.h>
103 #include <linux/sysctl.h>
104 #include <linux/module.h>
105 #include <linux/rcupdate.h>
106 #include <linux/string.h>
107 #include <linux/nodemask.h>
108 #include <linux/mempolicy.h>
109 #include <linux/mutex.h>
110 #include <linux/rtmutex.h>
112 #include <asm/uaccess.h>
113 #include <asm/cacheflush.h>
114 #include <asm/tlbflush.h>
115 #include <asm/page.h>
118 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
119 * SLAB_RED_ZONE & SLAB_POISON.
120 * 0 for faster, smaller code (especially in the critical paths).
122 * STATS - 1 to collect stats for /proc/slabinfo.
123 * 0 for faster, smaller code (especially in the critical paths).
125 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
128 #ifdef CONFIG_DEBUG_SLAB
131 #define FORCED_DEBUG 1
135 #define FORCED_DEBUG 0
138 /* Shouldn't this be in a header file somewhere? */
139 #define BYTES_PER_WORD sizeof(void *)
141 #ifndef cache_line_size
142 #define cache_line_size() L1_CACHE_BYTES
145 #ifndef ARCH_KMALLOC_MINALIGN
147 * Enforce a minimum alignment for the kmalloc caches.
148 * Usually, the kmalloc caches are cache_line_size() aligned, except when
149 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
150 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
151 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
152 * Note that this flag disables some debug features.
154 #define ARCH_KMALLOC_MINALIGN 0
157 #ifndef ARCH_SLAB_MINALIGN
159 * Enforce a minimum alignment for all caches.
160 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
161 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
162 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
163 * some debug features.
165 #define ARCH_SLAB_MINALIGN 0
168 #ifndef ARCH_KMALLOC_FLAGS
169 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
172 /* Legal flag mask for kmem_cache_create(). */
174 # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
175 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
177 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
178 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
179 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
181 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
182 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
183 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
184 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
190 * Bufctl's are used for linking objs within a slab
193 * This implementation relies on "struct page" for locating the cache &
194 * slab an object belongs to.
195 * This allows the bufctl structure to be small (one int), but limits
196 * the number of objects a slab (not a cache) can contain when off-slab
197 * bufctls are used. The limit is the size of the largest general cache
198 * that does not use off-slab slabs.
199 * For 32bit archs with 4 kB pages, is this 56.
200 * This is not serious, as it is only for large objects, when it is unwise
201 * to have too many per slab.
202 * Note: This limit can be raised by introducing a general cache whose size
203 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
206 typedef unsigned int kmem_bufctl_t;
207 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
208 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
209 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
210 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
215 * Manages the objs in a slab. Placed either at the beginning of mem allocated
216 * for a slab, or allocated from an general cache.
217 * Slabs are chained into three list: fully used, partial, fully free slabs.
220 struct list_head list;
221 unsigned long colouroff;
222 void *s_mem; /* including colour offset */
223 unsigned int inuse; /* num of objs active in slab */
225 unsigned short nodeid;
231 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
232 * arrange for kmem_freepages to be called via RCU. This is useful if
233 * we need to approach a kernel structure obliquely, from its address
234 * obtained without the usual locking. We can lock the structure to
235 * stabilize it and check it's still at the given address, only if we
236 * can be sure that the memory has not been meanwhile reused for some
237 * other kind of object (which our subsystem's lock might corrupt).
239 * rcu_read_lock before reading the address, then rcu_read_unlock after
240 * taking the spinlock within the structure expected at that address.
242 * We assume struct slab_rcu can overlay struct slab when destroying.
245 struct rcu_head head;
246 struct kmem_cache *cachep;
254 * - LIFO ordering, to hand out cache-warm objects from _alloc
255 * - reduce the number of linked list operations
256 * - reduce spinlock operations
258 * The limit is stored in the per-cpu structure to reduce the data cache
265 unsigned int batchcount;
266 unsigned int touched;
269 * Must have this definition in here for the proper
270 * alignment of array_cache. Also simplifies accessing
272 * [0] is for gcc 2.95. It should really be [].
277 * bootstrap: The caches do not work without cpuarrays anymore, but the
278 * cpuarrays are allocated from the generic caches...
280 #define BOOT_CPUCACHE_ENTRIES 1
281 struct arraycache_init {
282 struct array_cache cache;
283 void *entries[BOOT_CPUCACHE_ENTRIES];
287 * The slab lists for all objects.
290 struct list_head slabs_partial; /* partial list first, better asm code */
291 struct list_head slabs_full;
292 struct list_head slabs_free;
293 unsigned long free_objects;
294 unsigned int free_limit;
295 unsigned int colour_next; /* Per-node cache coloring */
296 spinlock_t list_lock;
297 struct array_cache *shared; /* shared per node */
298 struct array_cache **alien; /* on other nodes */
299 unsigned long next_reap; /* updated without locking */
300 int free_touched; /* updated without locking */
304 * Need this for bootstrapping a per node allocator.
306 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
307 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
308 #define CACHE_CACHE 0
310 #define SIZE_L3 (1 + MAX_NUMNODES)
312 static int drain_freelist(struct kmem_cache *cache,
313 struct kmem_list3 *l3, int tofree);
314 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
316 static void enable_cpucache(struct kmem_cache *cachep);
317 static void cache_reap(void *unused);
320 * This function must be completely optimized away if a constant is passed to
321 * it. Mostly the same as what is in linux/slab.h except it returns an index.
323 static __always_inline int index_of(const size_t size)
325 extern void __bad_size(void);
327 if (__builtin_constant_p(size)) {
335 #include "linux/kmalloc_sizes.h"
343 static int slab_early_init = 1;
345 #define INDEX_AC index_of(sizeof(struct arraycache_init))
346 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
348 static void kmem_list3_init(struct kmem_list3 *parent)
350 INIT_LIST_HEAD(&parent->slabs_full);
351 INIT_LIST_HEAD(&parent->slabs_partial);
352 INIT_LIST_HEAD(&parent->slabs_free);
353 parent->shared = NULL;
354 parent->alien = NULL;
355 parent->colour_next = 0;
356 spin_lock_init(&parent->list_lock);
357 parent->free_objects = 0;
358 parent->free_touched = 0;
361 #define MAKE_LIST(cachep, listp, slab, nodeid) \
363 INIT_LIST_HEAD(listp); \
364 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
367 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
369 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
370 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
371 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
381 /* 1) per-cpu data, touched during every alloc/free */
382 struct array_cache *array[NR_CPUS];
383 /* 2) Cache tunables. Protected by cache_chain_mutex */
384 unsigned int batchcount;
388 unsigned int buffer_size;
389 /* 3) touched by every alloc & free from the backend */
390 struct kmem_list3 *nodelists[MAX_NUMNODES];
392 unsigned int flags; /* constant flags */
393 unsigned int num; /* # of objs per slab */
395 /* 4) cache_grow/shrink */
396 /* order of pgs per slab (2^n) */
397 unsigned int gfporder;
399 /* force GFP flags, e.g. GFP_DMA */
402 size_t colour; /* cache colouring range */
403 unsigned int colour_off; /* colour offset */
404 struct kmem_cache *slabp_cache;
405 unsigned int slab_size;
406 unsigned int dflags; /* dynamic flags */
408 /* constructor func */
409 void (*ctor) (void *, struct kmem_cache *, unsigned long);
411 /* de-constructor func */
412 void (*dtor) (void *, struct kmem_cache *, unsigned long);
414 /* 5) cache creation/removal */
416 struct list_head next;
420 unsigned long num_active;
421 unsigned long num_allocations;
422 unsigned long high_mark;
424 unsigned long reaped;
425 unsigned long errors;
426 unsigned long max_freeable;
427 unsigned long node_allocs;
428 unsigned long node_frees;
429 unsigned long node_overflow;
437 * If debugging is enabled, then the allocator can add additional
438 * fields and/or padding to every object. buffer_size contains the total
439 * object size including these internal fields, the following two
440 * variables contain the offset to the user object and its size.
447 #define CFLGS_OFF_SLAB (0x80000000UL)
448 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
450 #define BATCHREFILL_LIMIT 16
452 * Optimization question: fewer reaps means less probability for unnessary
453 * cpucache drain/refill cycles.
455 * OTOH the cpuarrays can contain lots of objects,
456 * which could lock up otherwise freeable slabs.
458 #define REAPTIMEOUT_CPUC (2*HZ)
459 #define REAPTIMEOUT_LIST3 (4*HZ)
462 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
463 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
464 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
465 #define STATS_INC_GROWN(x) ((x)->grown++)
466 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
467 #define STATS_SET_HIGH(x) \
469 if ((x)->num_active > (x)->high_mark) \
470 (x)->high_mark = (x)->num_active; \
472 #define STATS_INC_ERR(x) ((x)->errors++)
473 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
474 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
475 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
476 #define STATS_SET_FREEABLE(x, i) \
478 if ((x)->max_freeable < i) \
479 (x)->max_freeable = i; \
481 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
482 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
483 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
484 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
486 #define STATS_INC_ACTIVE(x) do { } while (0)
487 #define STATS_DEC_ACTIVE(x) do { } while (0)
488 #define STATS_INC_ALLOCED(x) do { } while (0)
489 #define STATS_INC_GROWN(x) do { } while (0)
490 #define STATS_ADD_REAPED(x,y) do { } while (0)
491 #define STATS_SET_HIGH(x) do { } while (0)
492 #define STATS_INC_ERR(x) do { } while (0)
493 #define STATS_INC_NODEALLOCS(x) do { } while (0)
494 #define STATS_INC_NODEFREES(x) do { } while (0)
495 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
496 #define STATS_SET_FREEABLE(x, i) do { } while (0)
497 #define STATS_INC_ALLOCHIT(x) do { } while (0)
498 #define STATS_INC_ALLOCMISS(x) do { } while (0)
499 #define STATS_INC_FREEHIT(x) do { } while (0)
500 #define STATS_INC_FREEMISS(x) do { } while (0)
506 * memory layout of objects:
508 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
509 * the end of an object is aligned with the end of the real
510 * allocation. Catches writes behind the end of the allocation.
511 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
513 * cachep->obj_offset: The real object.
514 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
515 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
516 * [BYTES_PER_WORD long]
518 static int obj_offset(struct kmem_cache *cachep)
520 return cachep->obj_offset;
523 static int obj_size(struct kmem_cache *cachep)
525 return cachep->obj_size;
528 static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
530 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
531 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD);
534 static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
536 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
537 if (cachep->flags & SLAB_STORE_USER)
538 return (unsigned long *)(objp + cachep->buffer_size -
540 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD);
543 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
545 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
546 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
551 #define obj_offset(x) 0
552 #define obj_size(cachep) (cachep->buffer_size)
553 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
554 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
555 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
560 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
563 #if defined(CONFIG_LARGE_ALLOCS)
564 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
565 #define MAX_GFP_ORDER 13 /* up to 32Mb */
566 #elif defined(CONFIG_MMU)
567 #define MAX_OBJ_ORDER 5 /* 32 pages */
568 #define MAX_GFP_ORDER 5 /* 32 pages */
570 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
571 #define MAX_GFP_ORDER 8 /* up to 1Mb */
575 * Do not go above this order unless 0 objects fit into the slab.
577 #define BREAK_GFP_ORDER_HI 1
578 #define BREAK_GFP_ORDER_LO 0
579 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
582 * Functions for storing/retrieving the cachep and or slab from the page
583 * allocator. These are used to find the slab an obj belongs to. With kfree(),
584 * these are used to find the cache which an obj belongs to.
586 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
588 page->lru.next = (struct list_head *)cache;
591 static inline struct kmem_cache *page_get_cache(struct page *page)
593 if (unlikely(PageCompound(page)))
594 page = (struct page *)page_private(page);
595 BUG_ON(!PageSlab(page));
596 return (struct kmem_cache *)page->lru.next;
599 static inline void page_set_slab(struct page *page, struct slab *slab)
601 page->lru.prev = (struct list_head *)slab;
604 static inline struct slab *page_get_slab(struct page *page)
606 if (unlikely(PageCompound(page)))
607 page = (struct page *)page_private(page);
608 BUG_ON(!PageSlab(page));
609 return (struct slab *)page->lru.prev;
612 static inline struct kmem_cache *virt_to_cache(const void *obj)
614 struct page *page = virt_to_page(obj);
615 return page_get_cache(page);
618 static inline struct slab *virt_to_slab(const void *obj)
620 struct page *page = virt_to_page(obj);
621 return page_get_slab(page);
624 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
627 return slab->s_mem + cache->buffer_size * idx;
630 static inline unsigned int obj_to_index(struct kmem_cache *cache,
631 struct slab *slab, void *obj)
633 return (unsigned)(obj - slab->s_mem) / cache->buffer_size;
637 * These are the default caches for kmalloc. Custom caches can have other sizes.
639 struct cache_sizes malloc_sizes[] = {
640 #define CACHE(x) { .cs_size = (x) },
641 #include <linux/kmalloc_sizes.h>
645 EXPORT_SYMBOL(malloc_sizes);
647 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
653 static struct cache_names __initdata cache_names[] = {
654 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
655 #include <linux/kmalloc_sizes.h>
660 static struct arraycache_init initarray_cache __initdata =
661 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
662 static struct arraycache_init initarray_generic =
663 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
665 /* internal cache of cache description objs */
666 static struct kmem_cache cache_cache = {
668 .limit = BOOT_CPUCACHE_ENTRIES,
670 .buffer_size = sizeof(struct kmem_cache),
671 .name = "kmem_cache",
673 .obj_size = sizeof(struct kmem_cache),
677 /* Guard access to the cache-chain. */
678 static DEFINE_MUTEX(cache_chain_mutex);
679 static struct list_head cache_chain;
682 * vm_enough_memory() looks at this to determine how many slab-allocated pages
683 * are possibly freeable under pressure
685 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
687 atomic_t slab_reclaim_pages;
690 * chicken and egg problem: delay the per-cpu array allocation
691 * until the general caches are up.
701 * used by boot code to determine if it can use slab based allocator
703 int slab_is_available(void)
705 return g_cpucache_up == FULL;
708 static DEFINE_PER_CPU(struct work_struct, reap_work);
710 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
712 return cachep->array[smp_processor_id()];
715 static inline struct kmem_cache *__find_general_cachep(size_t size,
718 struct cache_sizes *csizep = malloc_sizes;
721 /* This happens if someone tries to call
722 * kmem_cache_create(), or __kmalloc(), before
723 * the generic caches are initialized.
725 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
727 while (size > csizep->cs_size)
731 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
732 * has cs_{dma,}cachep==NULL. Thus no special case
733 * for large kmalloc calls required.
735 if (unlikely(gfpflags & GFP_DMA))
736 return csizep->cs_dmacachep;
737 return csizep->cs_cachep;
740 struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
742 return __find_general_cachep(size, gfpflags);
744 EXPORT_SYMBOL(kmem_find_general_cachep);
746 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
748 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
752 * Calculate the number of objects and left-over bytes for a given buffer size.
754 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
755 size_t align, int flags, size_t *left_over,
760 size_t slab_size = PAGE_SIZE << gfporder;
763 * The slab management structure can be either off the slab or
764 * on it. For the latter case, the memory allocated for a
768 * - One kmem_bufctl_t for each object
769 * - Padding to respect alignment of @align
770 * - @buffer_size bytes for each object
772 * If the slab management structure is off the slab, then the
773 * alignment will already be calculated into the size. Because
774 * the slabs are all pages aligned, the objects will be at the
775 * correct alignment when allocated.
777 if (flags & CFLGS_OFF_SLAB) {
779 nr_objs = slab_size / buffer_size;
781 if (nr_objs > SLAB_LIMIT)
782 nr_objs = SLAB_LIMIT;
785 * Ignore padding for the initial guess. The padding
786 * is at most @align-1 bytes, and @buffer_size is at
787 * least @align. In the worst case, this result will
788 * be one greater than the number of objects that fit
789 * into the memory allocation when taking the padding
792 nr_objs = (slab_size - sizeof(struct slab)) /
793 (buffer_size + sizeof(kmem_bufctl_t));
796 * This calculated number will be either the right
797 * amount, or one greater than what we want.
799 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
803 if (nr_objs > SLAB_LIMIT)
804 nr_objs = SLAB_LIMIT;
806 mgmt_size = slab_mgmt_size(nr_objs, align);
809 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
812 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
814 static void __slab_error(const char *function, struct kmem_cache *cachep,
817 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
818 function, cachep->name, msg);
824 * Special reaping functions for NUMA systems called from cache_reap().
825 * These take care of doing round robin flushing of alien caches (containing
826 * objects freed on different nodes from which they were allocated) and the
827 * flushing of remote pcps by calling drain_node_pages.
829 static DEFINE_PER_CPU(unsigned long, reap_node);
831 static void init_reap_node(int cpu)
835 node = next_node(cpu_to_node(cpu), node_online_map);
836 if (node == MAX_NUMNODES)
837 node = first_node(node_online_map);
839 __get_cpu_var(reap_node) = node;
842 static void next_reap_node(void)
844 int node = __get_cpu_var(reap_node);
847 * Also drain per cpu pages on remote zones
849 if (node != numa_node_id())
850 drain_node_pages(node);
852 node = next_node(node, node_online_map);
853 if (unlikely(node >= MAX_NUMNODES))
854 node = first_node(node_online_map);
855 __get_cpu_var(reap_node) = node;
859 #define init_reap_node(cpu) do { } while (0)
860 #define next_reap_node(void) do { } while (0)
864 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
865 * via the workqueue/eventd.
866 * Add the CPU number into the expiration time to minimize the possibility of
867 * the CPUs getting into lockstep and contending for the global cache chain
870 static void __devinit start_cpu_timer(int cpu)
872 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
875 * When this gets called from do_initcalls via cpucache_init(),
876 * init_workqueues() has already run, so keventd will be setup
879 if (keventd_up() && reap_work->func == NULL) {
881 INIT_WORK(reap_work, cache_reap, NULL);
882 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
886 static struct array_cache *alloc_arraycache(int node, int entries,
889 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
890 struct array_cache *nc = NULL;
892 nc = kmalloc_node(memsize, GFP_KERNEL, node);
896 nc->batchcount = batchcount;
898 spin_lock_init(&nc->lock);
904 * Transfer objects in one arraycache to another.
905 * Locking must be handled by the caller.
907 * Return the number of entries transferred.
909 static int transfer_objects(struct array_cache *to,
910 struct array_cache *from, unsigned int max)
912 /* Figure out how many entries to transfer */
913 int nr = min(min(from->avail, max), to->limit - to->avail);
918 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
928 static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int);
929 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
931 static struct array_cache **alloc_alien_cache(int node, int limit)
933 struct array_cache **ac_ptr;
934 int memsize = sizeof(void *) * MAX_NUMNODES;
939 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
942 if (i == node || !node_online(i)) {
946 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
948 for (i--; i <= 0; i--)
958 static void free_alien_cache(struct array_cache **ac_ptr)
969 static void __drain_alien_cache(struct kmem_cache *cachep,
970 struct array_cache *ac, int node)
972 struct kmem_list3 *rl3 = cachep->nodelists[node];
975 spin_lock(&rl3->list_lock);
977 * Stuff objects into the remote nodes shared array first.
978 * That way we could avoid the overhead of putting the objects
979 * into the free lists and getting them back later.
982 transfer_objects(rl3->shared, ac, ac->limit);
984 free_block(cachep, ac->entry, ac->avail, node);
986 spin_unlock(&rl3->list_lock);
991 * Called from cache_reap() to regularly drain alien caches round robin.
993 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
995 int node = __get_cpu_var(reap_node);
998 struct array_cache *ac = l3->alien[node];
1000 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1001 __drain_alien_cache(cachep, ac, node);
1002 spin_unlock_irq(&ac->lock);
1007 static void drain_alien_cache(struct kmem_cache *cachep,
1008 struct array_cache **alien)
1011 struct array_cache *ac;
1012 unsigned long flags;
1014 for_each_online_node(i) {
1017 spin_lock_irqsave(&ac->lock, flags);
1018 __drain_alien_cache(cachep, ac, i);
1019 spin_unlock_irqrestore(&ac->lock, flags);
1024 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1026 struct slab *slabp = virt_to_slab(objp);
1027 int nodeid = slabp->nodeid;
1028 struct kmem_list3 *l3;
1029 struct array_cache *alien = NULL;
1032 * Make sure we are not freeing a object from another node to the array
1033 * cache on this cpu.
1035 if (likely(slabp->nodeid == numa_node_id()))
1038 l3 = cachep->nodelists[numa_node_id()];
1039 STATS_INC_NODEFREES(cachep);
1040 if (l3->alien && l3->alien[nodeid]) {
1041 alien = l3->alien[nodeid];
1042 spin_lock(&alien->lock);
1043 if (unlikely(alien->avail == alien->limit)) {
1044 STATS_INC_ACOVERFLOW(cachep);
1045 __drain_alien_cache(cachep, alien, nodeid);
1047 alien->entry[alien->avail++] = objp;
1048 spin_unlock(&alien->lock);
1050 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1051 free_block(cachep, &objp, 1, nodeid);
1052 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1059 #define drain_alien_cache(cachep, alien) do { } while (0)
1060 #define reap_alien(cachep, l3) do { } while (0)
1062 static inline struct array_cache **alloc_alien_cache(int node, int limit)
1064 return (struct array_cache **) 0x01020304ul;
1067 static inline void free_alien_cache(struct array_cache **ac_ptr)
1071 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1078 static int __devinit cpuup_callback(struct notifier_block *nfb,
1079 unsigned long action, void *hcpu)
1081 long cpu = (long)hcpu;
1082 struct kmem_cache *cachep;
1083 struct kmem_list3 *l3 = NULL;
1084 int node = cpu_to_node(cpu);
1085 int memsize = sizeof(struct kmem_list3);
1088 case CPU_UP_PREPARE:
1089 mutex_lock(&cache_chain_mutex);
1091 * We need to do this right in the beginning since
1092 * alloc_arraycache's are going to use this list.
1093 * kmalloc_node allows us to add the slab to the right
1094 * kmem_list3 and not this cpu's kmem_list3
1097 list_for_each_entry(cachep, &cache_chain, next) {
1099 * Set up the size64 kmemlist for cpu before we can
1100 * begin anything. Make sure some other cpu on this
1101 * node has not already allocated this
1103 if (!cachep->nodelists[node]) {
1104 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1107 kmem_list3_init(l3);
1108 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1109 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1112 * The l3s don't come and go as CPUs come and
1113 * go. cache_chain_mutex is sufficient
1116 cachep->nodelists[node] = l3;
1119 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1120 cachep->nodelists[node]->free_limit =
1121 (1 + nr_cpus_node(node)) *
1122 cachep->batchcount + cachep->num;
1123 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1127 * Now we can go ahead with allocating the shared arrays and
1130 list_for_each_entry(cachep, &cache_chain, next) {
1131 struct array_cache *nc;
1132 struct array_cache *shared;
1133 struct array_cache **alien;
1135 nc = alloc_arraycache(node, cachep->limit,
1136 cachep->batchcount);
1139 shared = alloc_arraycache(node,
1140 cachep->shared * cachep->batchcount,
1145 alien = alloc_alien_cache(node, cachep->limit);
1148 cachep->array[cpu] = nc;
1149 l3 = cachep->nodelists[node];
1152 spin_lock_irq(&l3->list_lock);
1155 * We are serialised from CPU_DEAD or
1156 * CPU_UP_CANCELLED by the cpucontrol lock
1158 l3->shared = shared;
1167 spin_unlock_irq(&l3->list_lock);
1169 free_alien_cache(alien);
1171 mutex_unlock(&cache_chain_mutex);
1174 start_cpu_timer(cpu);
1176 #ifdef CONFIG_HOTPLUG_CPU
1179 * Even if all the cpus of a node are down, we don't free the
1180 * kmem_list3 of any cache. This to avoid a race between
1181 * cpu_down, and a kmalloc allocation from another cpu for
1182 * memory from the node of the cpu going down. The list3
1183 * structure is usually allocated from kmem_cache_create() and
1184 * gets destroyed at kmem_cache_destroy().
1187 case CPU_UP_CANCELED:
1188 mutex_lock(&cache_chain_mutex);
1189 list_for_each_entry(cachep, &cache_chain, next) {
1190 struct array_cache *nc;
1191 struct array_cache *shared;
1192 struct array_cache **alien;
1195 mask = node_to_cpumask(node);
1196 /* cpu is dead; no one can alloc from it. */
1197 nc = cachep->array[cpu];
1198 cachep->array[cpu] = NULL;
1199 l3 = cachep->nodelists[node];
1202 goto free_array_cache;
1204 spin_lock_irq(&l3->list_lock);
1206 /* Free limit for this kmem_list3 */
1207 l3->free_limit -= cachep->batchcount;
1209 free_block(cachep, nc->entry, nc->avail, node);
1211 if (!cpus_empty(mask)) {
1212 spin_unlock_irq(&l3->list_lock);
1213 goto free_array_cache;
1216 shared = l3->shared;
1218 free_block(cachep, l3->shared->entry,
1219 l3->shared->avail, node);
1226 spin_unlock_irq(&l3->list_lock);
1230 drain_alien_cache(cachep, alien);
1231 free_alien_cache(alien);
1237 * In the previous loop, all the objects were freed to
1238 * the respective cache's slabs, now we can go ahead and
1239 * shrink each nodelist to its limit.
1241 list_for_each_entry(cachep, &cache_chain, next) {
1242 l3 = cachep->nodelists[node];
1245 drain_freelist(cachep, l3, l3->free_objects);
1247 mutex_unlock(&cache_chain_mutex);
1253 mutex_unlock(&cache_chain_mutex);
1257 static struct notifier_block __cpuinitdata cpucache_notifier = {
1258 &cpuup_callback, NULL, 0
1262 * swap the static kmem_list3 with kmalloced memory
1264 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1267 struct kmem_list3 *ptr;
1269 BUG_ON(cachep->nodelists[nodeid] != list);
1270 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1273 local_irq_disable();
1274 memcpy(ptr, list, sizeof(struct kmem_list3));
1276 * Do not assume that spinlocks can be initialized via memcpy:
1278 spin_lock_init(&ptr->list_lock);
1280 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1281 cachep->nodelists[nodeid] = ptr;
1286 * Initialisation. Called after the page allocator have been initialised and
1287 * before smp_init().
1289 void __init kmem_cache_init(void)
1292 struct cache_sizes *sizes;
1293 struct cache_names *names;
1297 for (i = 0; i < NUM_INIT_LISTS; i++) {
1298 kmem_list3_init(&initkmem_list3[i]);
1299 if (i < MAX_NUMNODES)
1300 cache_cache.nodelists[i] = NULL;
1304 * Fragmentation resistance on low memory - only use bigger
1305 * page orders on machines with more than 32MB of memory.
1307 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1308 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1310 /* Bootstrap is tricky, because several objects are allocated
1311 * from caches that do not exist yet:
1312 * 1) initialize the cache_cache cache: it contains the struct
1313 * kmem_cache structures of all caches, except cache_cache itself:
1314 * cache_cache is statically allocated.
1315 * Initially an __init data area is used for the head array and the
1316 * kmem_list3 structures, it's replaced with a kmalloc allocated
1317 * array at the end of the bootstrap.
1318 * 2) Create the first kmalloc cache.
1319 * The struct kmem_cache for the new cache is allocated normally.
1320 * An __init data area is used for the head array.
1321 * 3) Create the remaining kmalloc caches, with minimally sized
1323 * 4) Replace the __init data head arrays for cache_cache and the first
1324 * kmalloc cache with kmalloc allocated arrays.
1325 * 5) Replace the __init data for kmem_list3 for cache_cache and
1326 * the other cache's with kmalloc allocated memory.
1327 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1330 /* 1) create the cache_cache */
1331 INIT_LIST_HEAD(&cache_chain);
1332 list_add(&cache_cache.next, &cache_chain);
1333 cache_cache.colour_off = cache_line_size();
1334 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1335 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
1337 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1340 for (order = 0; order < MAX_ORDER; order++) {
1341 cache_estimate(order, cache_cache.buffer_size,
1342 cache_line_size(), 0, &left_over, &cache_cache.num);
1343 if (cache_cache.num)
1346 BUG_ON(!cache_cache.num);
1347 cache_cache.gfporder = order;
1348 cache_cache.colour = left_over / cache_cache.colour_off;
1349 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1350 sizeof(struct slab), cache_line_size());
1352 /* 2+3) create the kmalloc caches */
1353 sizes = malloc_sizes;
1354 names = cache_names;
1357 * Initialize the caches that provide memory for the array cache and the
1358 * kmem_list3 structures first. Without this, further allocations will
1362 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1363 sizes[INDEX_AC].cs_size,
1364 ARCH_KMALLOC_MINALIGN,
1365 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1368 if (INDEX_AC != INDEX_L3) {
1369 sizes[INDEX_L3].cs_cachep =
1370 kmem_cache_create(names[INDEX_L3].name,
1371 sizes[INDEX_L3].cs_size,
1372 ARCH_KMALLOC_MINALIGN,
1373 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1377 slab_early_init = 0;
1379 while (sizes->cs_size != ULONG_MAX) {
1381 * For performance, all the general caches are L1 aligned.
1382 * This should be particularly beneficial on SMP boxes, as it
1383 * eliminates "false sharing".
1384 * Note for systems short on memory removing the alignment will
1385 * allow tighter packing of the smaller caches.
1387 if (!sizes->cs_cachep) {
1388 sizes->cs_cachep = kmem_cache_create(names->name,
1390 ARCH_KMALLOC_MINALIGN,
1391 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1395 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
1397 ARCH_KMALLOC_MINALIGN,
1398 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1404 /* 4) Replace the bootstrap head arrays */
1406 struct array_cache *ptr;
1408 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1410 local_irq_disable();
1411 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1412 memcpy(ptr, cpu_cache_get(&cache_cache),
1413 sizeof(struct arraycache_init));
1415 * Do not assume that spinlocks can be initialized via memcpy:
1417 spin_lock_init(&ptr->lock);
1419 cache_cache.array[smp_processor_id()] = ptr;
1422 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1424 local_irq_disable();
1425 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1426 != &initarray_generic.cache);
1427 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1428 sizeof(struct arraycache_init));
1430 * Do not assume that spinlocks can be initialized via memcpy:
1432 spin_lock_init(&ptr->lock);
1434 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1438 /* 5) Replace the bootstrap kmem_list3's */
1441 /* Replace the static kmem_list3 structures for the boot cpu */
1442 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
1445 for_each_online_node(node) {
1446 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1447 &initkmem_list3[SIZE_AC + node], node);
1449 if (INDEX_AC != INDEX_L3) {
1450 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1451 &initkmem_list3[SIZE_L3 + node],
1457 /* 6) resize the head arrays to their final sizes */
1459 struct kmem_cache *cachep;
1460 mutex_lock(&cache_chain_mutex);
1461 list_for_each_entry(cachep, &cache_chain, next)
1462 enable_cpucache(cachep);
1463 mutex_unlock(&cache_chain_mutex);
1467 g_cpucache_up = FULL;
1470 * Register a cpu startup notifier callback that initializes
1471 * cpu_cache_get for all new cpus
1473 register_cpu_notifier(&cpucache_notifier);
1476 * The reap timers are started later, with a module init call: That part
1477 * of the kernel is not yet operational.
1481 static int __init cpucache_init(void)
1486 * Register the timers that return unneeded pages to the page allocator
1488 for_each_online_cpu(cpu)
1489 start_cpu_timer(cpu);
1492 __initcall(cpucache_init);
1495 * Interface to system's page allocator. No need to hold the cache-lock.
1497 * If we requested dmaable memory, we will get it. Even if we
1498 * did not request dmaable memory, we might get it, but that
1499 * would be relatively rare and ignorable.
1501 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1509 * Nommu uses slab's for process anonymous memory allocations, and thus
1510 * requires __GFP_COMP to properly refcount higher order allocations
1512 flags |= __GFP_COMP;
1514 flags |= cachep->gfpflags;
1516 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1520 nr_pages = (1 << cachep->gfporder);
1521 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1522 atomic_add(nr_pages, &slab_reclaim_pages);
1523 add_zone_page_state(page_zone(page), NR_SLAB, nr_pages);
1524 for (i = 0; i < nr_pages; i++)
1525 __SetPageSlab(page + i);
1526 return page_address(page);
1530 * Interface to system's page release.
1532 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1534 unsigned long i = (1 << cachep->gfporder);
1535 struct page *page = virt_to_page(addr);
1536 const unsigned long nr_freed = i;
1538 sub_zone_page_state(page_zone(page), NR_SLAB, nr_freed);
1540 BUG_ON(!PageSlab(page));
1541 __ClearPageSlab(page);
1544 if (current->reclaim_state)
1545 current->reclaim_state->reclaimed_slab += nr_freed;
1546 free_pages((unsigned long)addr, cachep->gfporder);
1547 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1548 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages);
1551 static void kmem_rcu_free(struct rcu_head *head)
1553 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1554 struct kmem_cache *cachep = slab_rcu->cachep;
1556 kmem_freepages(cachep, slab_rcu->addr);
1557 if (OFF_SLAB(cachep))
1558 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1563 #ifdef CONFIG_DEBUG_PAGEALLOC
1564 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1565 unsigned long caller)
1567 int size = obj_size(cachep);
1569 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1571 if (size < 5 * sizeof(unsigned long))
1574 *addr++ = 0x12345678;
1576 *addr++ = smp_processor_id();
1577 size -= 3 * sizeof(unsigned long);
1579 unsigned long *sptr = &caller;
1580 unsigned long svalue;
1582 while (!kstack_end(sptr)) {
1584 if (kernel_text_address(svalue)) {
1586 size -= sizeof(unsigned long);
1587 if (size <= sizeof(unsigned long))
1593 *addr++ = 0x87654321;
1597 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1599 int size = obj_size(cachep);
1600 addr = &((char *)addr)[obj_offset(cachep)];
1602 memset(addr, val, size);
1603 *(unsigned char *)(addr + size - 1) = POISON_END;
1606 static void dump_line(char *data, int offset, int limit)
1609 printk(KERN_ERR "%03x:", offset);
1610 for (i = 0; i < limit; i++)
1611 printk(" %02x", (unsigned char)data[offset + i]);
1618 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1623 if (cachep->flags & SLAB_RED_ZONE) {
1624 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
1625 *dbg_redzone1(cachep, objp),
1626 *dbg_redzone2(cachep, objp));
1629 if (cachep->flags & SLAB_STORE_USER) {
1630 printk(KERN_ERR "Last user: [<%p>]",
1631 *dbg_userword(cachep, objp));
1632 print_symbol("(%s)",
1633 (unsigned long)*dbg_userword(cachep, objp));
1636 realobj = (char *)objp + obj_offset(cachep);
1637 size = obj_size(cachep);
1638 for (i = 0; i < size && lines; i += 16, lines--) {
1641 if (i + limit > size)
1643 dump_line(realobj, i, limit);
1647 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1653 realobj = (char *)objp + obj_offset(cachep);
1654 size = obj_size(cachep);
1656 for (i = 0; i < size; i++) {
1657 char exp = POISON_FREE;
1660 if (realobj[i] != exp) {
1666 "Slab corruption: start=%p, len=%d\n",
1668 print_objinfo(cachep, objp, 0);
1670 /* Hexdump the affected line */
1673 if (i + limit > size)
1675 dump_line(realobj, i, limit);
1678 /* Limit to 5 lines */
1684 /* Print some data about the neighboring objects, if they
1687 struct slab *slabp = virt_to_slab(objp);
1690 objnr = obj_to_index(cachep, slabp, objp);
1692 objp = index_to_obj(cachep, slabp, objnr - 1);
1693 realobj = (char *)objp + obj_offset(cachep);
1694 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1696 print_objinfo(cachep, objp, 2);
1698 if (objnr + 1 < cachep->num) {
1699 objp = index_to_obj(cachep, slabp, objnr + 1);
1700 realobj = (char *)objp + obj_offset(cachep);
1701 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1703 print_objinfo(cachep, objp, 2);
1711 * slab_destroy_objs - destroy a slab and its objects
1712 * @cachep: cache pointer being destroyed
1713 * @slabp: slab pointer being destroyed
1715 * Call the registered destructor for each object in a slab that is being
1718 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1721 for (i = 0; i < cachep->num; i++) {
1722 void *objp = index_to_obj(cachep, slabp, i);
1724 if (cachep->flags & SLAB_POISON) {
1725 #ifdef CONFIG_DEBUG_PAGEALLOC
1726 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1728 kernel_map_pages(virt_to_page(objp),
1729 cachep->buffer_size / PAGE_SIZE, 1);
1731 check_poison_obj(cachep, objp);
1733 check_poison_obj(cachep, objp);
1736 if (cachep->flags & SLAB_RED_ZONE) {
1737 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1738 slab_error(cachep, "start of a freed object "
1740 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1741 slab_error(cachep, "end of a freed object "
1744 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1745 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
1749 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1753 for (i = 0; i < cachep->num; i++) {
1754 void *objp = index_to_obj(cachep, slabp, i);
1755 (cachep->dtor) (objp, cachep, 0);
1762 * slab_destroy - destroy and release all objects in a slab
1763 * @cachep: cache pointer being destroyed
1764 * @slabp: slab pointer being destroyed
1766 * Destroy all the objs in a slab, and release the mem back to the system.
1767 * Before calling the slab must have been unlinked from the cache. The
1768 * cache-lock is not held/needed.
1770 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1772 void *addr = slabp->s_mem - slabp->colouroff;
1774 slab_destroy_objs(cachep, slabp);
1775 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1776 struct slab_rcu *slab_rcu;
1778 slab_rcu = (struct slab_rcu *)slabp;
1779 slab_rcu->cachep = cachep;
1780 slab_rcu->addr = addr;
1781 call_rcu(&slab_rcu->head, kmem_rcu_free);
1783 kmem_freepages(cachep, addr);
1784 if (OFF_SLAB(cachep))
1785 kmem_cache_free(cachep->slabp_cache, slabp);
1790 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1791 * size of kmem_list3.
1793 static void set_up_list3s(struct kmem_cache *cachep, int index)
1797 for_each_online_node(node) {
1798 cachep->nodelists[node] = &initkmem_list3[index + node];
1799 cachep->nodelists[node]->next_reap = jiffies +
1801 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1806 * calculate_slab_order - calculate size (page order) of slabs
1807 * @cachep: pointer to the cache that is being created
1808 * @size: size of objects to be created in this cache.
1809 * @align: required alignment for the objects.
1810 * @flags: slab allocation flags
1812 * Also calculates the number of objects per slab.
1814 * This could be made much more intelligent. For now, try to avoid using
1815 * high order pages for slabs. When the gfp() functions are more friendly
1816 * towards high-order requests, this should be changed.
1818 static size_t calculate_slab_order(struct kmem_cache *cachep,
1819 size_t size, size_t align, unsigned long flags)
1821 unsigned long offslab_limit;
1822 size_t left_over = 0;
1825 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
1829 cache_estimate(gfporder, size, align, flags, &remainder, &num);
1833 if (flags & CFLGS_OFF_SLAB) {
1835 * Max number of objs-per-slab for caches which
1836 * use off-slab slabs. Needed to avoid a possible
1837 * looping condition in cache_grow().
1839 offslab_limit = size - sizeof(struct slab);
1840 offslab_limit /= sizeof(kmem_bufctl_t);
1842 if (num > offslab_limit)
1846 /* Found something acceptable - save it away */
1848 cachep->gfporder = gfporder;
1849 left_over = remainder;
1852 * A VFS-reclaimable slab tends to have most allocations
1853 * as GFP_NOFS and we really don't want to have to be allocating
1854 * higher-order pages when we are unable to shrink dcache.
1856 if (flags & SLAB_RECLAIM_ACCOUNT)
1860 * Large number of objects is good, but very large slabs are
1861 * currently bad for the gfp()s.
1863 if (gfporder >= slab_break_gfp_order)
1867 * Acceptable internal fragmentation?
1869 if (left_over * 8 <= (PAGE_SIZE << gfporder))
1875 static void setup_cpu_cache(struct kmem_cache *cachep)
1877 if (g_cpucache_up == FULL) {
1878 enable_cpucache(cachep);
1881 if (g_cpucache_up == NONE) {
1883 * Note: the first kmem_cache_create must create the cache
1884 * that's used by kmalloc(24), otherwise the creation of
1885 * further caches will BUG().
1887 cachep->array[smp_processor_id()] = &initarray_generic.cache;
1890 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
1891 * the first cache, then we need to set up all its list3s,
1892 * otherwise the creation of further caches will BUG().
1894 set_up_list3s(cachep, SIZE_AC);
1895 if (INDEX_AC == INDEX_L3)
1896 g_cpucache_up = PARTIAL_L3;
1898 g_cpucache_up = PARTIAL_AC;
1900 cachep->array[smp_processor_id()] =
1901 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1903 if (g_cpucache_up == PARTIAL_AC) {
1904 set_up_list3s(cachep, SIZE_L3);
1905 g_cpucache_up = PARTIAL_L3;
1908 for_each_online_node(node) {
1909 cachep->nodelists[node] =
1910 kmalloc_node(sizeof(struct kmem_list3),
1912 BUG_ON(!cachep->nodelists[node]);
1913 kmem_list3_init(cachep->nodelists[node]);
1917 cachep->nodelists[numa_node_id()]->next_reap =
1918 jiffies + REAPTIMEOUT_LIST3 +
1919 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1921 cpu_cache_get(cachep)->avail = 0;
1922 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1923 cpu_cache_get(cachep)->batchcount = 1;
1924 cpu_cache_get(cachep)->touched = 0;
1925 cachep->batchcount = 1;
1926 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1930 * kmem_cache_create - Create a cache.
1931 * @name: A string which is used in /proc/slabinfo to identify this cache.
1932 * @size: The size of objects to be created in this cache.
1933 * @align: The required alignment for the objects.
1934 * @flags: SLAB flags
1935 * @ctor: A constructor for the objects.
1936 * @dtor: A destructor for the objects.
1938 * Returns a ptr to the cache on success, NULL on failure.
1939 * Cannot be called within a int, but can be interrupted.
1940 * The @ctor is run when new pages are allocated by the cache
1941 * and the @dtor is run before the pages are handed back.
1943 * @name must be valid until the cache is destroyed. This implies that
1944 * the module calling this has to destroy the cache before getting unloaded.
1948 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1949 * to catch references to uninitialised memory.
1951 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1952 * for buffer overruns.
1954 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1955 * cacheline. This can be beneficial if you're counting cycles as closely
1959 kmem_cache_create (const char *name, size_t size, size_t align,
1960 unsigned long flags,
1961 void (*ctor)(void*, struct kmem_cache *, unsigned long),
1962 void (*dtor)(void*, struct kmem_cache *, unsigned long))
1964 size_t left_over, slab_size, ralign;
1965 struct kmem_cache *cachep = NULL, *pc;
1968 * Sanity checks... these are all serious usage bugs.
1970 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
1971 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
1972 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
1978 * Prevent CPUs from coming and going.
1979 * lock_cpu_hotplug() nests outside cache_chain_mutex
1983 mutex_lock(&cache_chain_mutex);
1985 list_for_each_entry(pc, &cache_chain, next) {
1986 mm_segment_t old_fs = get_fs();
1991 * This happens when the module gets unloaded and doesn't
1992 * destroy its slab cache and no-one else reuses the vmalloc
1993 * area of the module. Print a warning.
1996 res = __get_user(tmp, pc->name);
1999 printk("SLAB: cache with size %d has lost its name\n",
2004 if (!strcmp(pc->name, name)) {
2005 printk("kmem_cache_create: duplicate cache %s\n", name);
2012 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2013 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
2014 /* No constructor, but inital state check requested */
2015 printk(KERN_ERR "%s: No con, but init state check "
2016 "requested - %s\n", __FUNCTION__, name);
2017 flags &= ~SLAB_DEBUG_INITIAL;
2021 * Enable redzoning and last user accounting, except for caches with
2022 * large objects, if the increased size would increase the object size
2023 * above the next power of two: caches with object sizes just above a
2024 * power of two have a significant amount of internal fragmentation.
2026 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
2027 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2028 if (!(flags & SLAB_DESTROY_BY_RCU))
2029 flags |= SLAB_POISON;
2031 if (flags & SLAB_DESTROY_BY_RCU)
2032 BUG_ON(flags & SLAB_POISON);
2034 if (flags & SLAB_DESTROY_BY_RCU)
2038 * Always checks flags, a caller might be expecting debug support which
2041 BUG_ON(flags & ~CREATE_MASK);
2044 * Check that size is in terms of words. This is needed to avoid
2045 * unaligned accesses for some archs when redzoning is used, and makes
2046 * sure any on-slab bufctl's are also correctly aligned.
2048 if (size & (BYTES_PER_WORD - 1)) {
2049 size += (BYTES_PER_WORD - 1);
2050 size &= ~(BYTES_PER_WORD - 1);
2053 /* calculate the final buffer alignment: */
2055 /* 1) arch recommendation: can be overridden for debug */
2056 if (flags & SLAB_HWCACHE_ALIGN) {
2058 * Default alignment: as specified by the arch code. Except if
2059 * an object is really small, then squeeze multiple objects into
2062 ralign = cache_line_size();
2063 while (size <= ralign / 2)
2066 ralign = BYTES_PER_WORD;
2068 /* 2) arch mandated alignment: disables debug if necessary */
2069 if (ralign < ARCH_SLAB_MINALIGN) {
2070 ralign = ARCH_SLAB_MINALIGN;
2071 if (ralign > BYTES_PER_WORD)
2072 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2074 /* 3) caller mandated alignment: disables debug if necessary */
2075 if (ralign < align) {
2077 if (ralign > BYTES_PER_WORD)
2078 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2081 * 4) Store it. Note that the debug code below can reduce
2082 * the alignment to BYTES_PER_WORD.
2086 /* Get cache's description obj. */
2087 cachep = kmem_cache_zalloc(&cache_cache, SLAB_KERNEL);
2092 cachep->obj_size = size;
2094 if (flags & SLAB_RED_ZONE) {
2095 /* redzoning only works with word aligned caches */
2096 align = BYTES_PER_WORD;
2098 /* add space for red zone words */
2099 cachep->obj_offset += BYTES_PER_WORD;
2100 size += 2 * BYTES_PER_WORD;
2102 if (flags & SLAB_STORE_USER) {
2103 /* user store requires word alignment and
2104 * one word storage behind the end of the real
2107 align = BYTES_PER_WORD;
2108 size += BYTES_PER_WORD;
2110 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2111 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2112 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2113 cachep->obj_offset += PAGE_SIZE - size;
2120 * Determine if the slab management is 'on' or 'off' slab.
2121 * (bootstrapping cannot cope with offslab caches so don't do
2124 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
2126 * Size is large, assume best to place the slab management obj
2127 * off-slab (should allow better packing of objs).
2129 flags |= CFLGS_OFF_SLAB;
2131 size = ALIGN(size, align);
2133 left_over = calculate_slab_order(cachep, size, align, flags);
2136 printk("kmem_cache_create: couldn't create cache %s.\n", name);
2137 kmem_cache_free(&cache_cache, cachep);
2141 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2142 + sizeof(struct slab), align);
2145 * If the slab has been placed off-slab, and we have enough space then
2146 * move it on-slab. This is at the expense of any extra colouring.
2148 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2149 flags &= ~CFLGS_OFF_SLAB;
2150 left_over -= slab_size;
2153 if (flags & CFLGS_OFF_SLAB) {
2154 /* really off slab. No need for manual alignment */
2156 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2159 cachep->colour_off = cache_line_size();
2160 /* Offset must be a multiple of the alignment. */
2161 if (cachep->colour_off < align)
2162 cachep->colour_off = align;
2163 cachep->colour = left_over / cachep->colour_off;
2164 cachep->slab_size = slab_size;
2165 cachep->flags = flags;
2166 cachep->gfpflags = 0;
2167 if (flags & SLAB_CACHE_DMA)
2168 cachep->gfpflags |= GFP_DMA;
2169 cachep->buffer_size = size;
2171 if (flags & CFLGS_OFF_SLAB)
2172 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2173 cachep->ctor = ctor;
2174 cachep->dtor = dtor;
2175 cachep->name = name;
2178 setup_cpu_cache(cachep);
2180 /* cache setup completed, link it into the list */
2181 list_add(&cachep->next, &cache_chain);
2183 if (!cachep && (flags & SLAB_PANIC))
2184 panic("kmem_cache_create(): failed to create slab `%s'\n",
2186 mutex_unlock(&cache_chain_mutex);
2187 unlock_cpu_hotplug();
2190 EXPORT_SYMBOL(kmem_cache_create);
2193 static void check_irq_off(void)
2195 BUG_ON(!irqs_disabled());
2198 static void check_irq_on(void)
2200 BUG_ON(irqs_disabled());
2203 static void check_spinlock_acquired(struct kmem_cache *cachep)
2207 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2211 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2215 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2220 #define check_irq_off() do { } while(0)
2221 #define check_irq_on() do { } while(0)
2222 #define check_spinlock_acquired(x) do { } while(0)
2223 #define check_spinlock_acquired_node(x, y) do { } while(0)
2226 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2227 struct array_cache *ac,
2228 int force, int node);
2230 static void do_drain(void *arg)
2232 struct kmem_cache *cachep = arg;
2233 struct array_cache *ac;
2234 int node = numa_node_id();
2237 ac = cpu_cache_get(cachep);
2238 spin_lock(&cachep->nodelists[node]->list_lock);
2239 free_block(cachep, ac->entry, ac->avail, node);
2240 spin_unlock(&cachep->nodelists[node]->list_lock);
2244 static void drain_cpu_caches(struct kmem_cache *cachep)
2246 struct kmem_list3 *l3;
2249 on_each_cpu(do_drain, cachep, 1, 1);
2251 for_each_online_node(node) {
2252 l3 = cachep->nodelists[node];
2253 if (l3 && l3->alien)
2254 drain_alien_cache(cachep, l3->alien);
2257 for_each_online_node(node) {
2258 l3 = cachep->nodelists[node];
2260 drain_array(cachep, l3, l3->shared, 1, node);
2265 * Remove slabs from the list of free slabs.
2266 * Specify the number of slabs to drain in tofree.
2268 * Returns the actual number of slabs released.
2270 static int drain_freelist(struct kmem_cache *cache,
2271 struct kmem_list3 *l3, int tofree)
2273 struct list_head *p;
2278 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2280 spin_lock_irq(&l3->list_lock);
2281 p = l3->slabs_free.prev;
2282 if (p == &l3->slabs_free) {
2283 spin_unlock_irq(&l3->list_lock);
2287 slabp = list_entry(p, struct slab, list);
2289 BUG_ON(slabp->inuse);
2291 list_del(&slabp->list);
2293 * Safe to drop the lock. The slab is no longer linked
2296 l3->free_objects -= cache->num;
2297 spin_unlock_irq(&l3->list_lock);
2298 slab_destroy(cache, slabp);
2305 static int __cache_shrink(struct kmem_cache *cachep)
2308 struct kmem_list3 *l3;
2310 drain_cpu_caches(cachep);
2313 for_each_online_node(i) {
2314 l3 = cachep->nodelists[i];
2318 drain_freelist(cachep, l3, l3->free_objects);
2320 ret += !list_empty(&l3->slabs_full) ||
2321 !list_empty(&l3->slabs_partial);
2323 return (ret ? 1 : 0);
2327 * kmem_cache_shrink - Shrink a cache.
2328 * @cachep: The cache to shrink.
2330 * Releases as many slabs as possible for a cache.
2331 * To help debugging, a zero exit status indicates all slabs were released.
2333 int kmem_cache_shrink(struct kmem_cache *cachep)
2335 BUG_ON(!cachep || in_interrupt());
2337 return __cache_shrink(cachep);
2339 EXPORT_SYMBOL(kmem_cache_shrink);
2342 * kmem_cache_destroy - delete a cache
2343 * @cachep: the cache to destroy
2345 * Remove a struct kmem_cache object from the slab cache.
2346 * Returns 0 on success.
2348 * It is expected this function will be called by a module when it is
2349 * unloaded. This will remove the cache completely, and avoid a duplicate
2350 * cache being allocated each time a module is loaded and unloaded, if the
2351 * module doesn't have persistent in-kernel storage across loads and unloads.
2353 * The cache must be empty before calling this function.
2355 * The caller must guarantee that noone will allocate memory from the cache
2356 * during the kmem_cache_destroy().
2358 int kmem_cache_destroy(struct kmem_cache *cachep)
2361 struct kmem_list3 *l3;
2363 BUG_ON(!cachep || in_interrupt());
2365 /* Don't let CPUs to come and go */
2368 /* Find the cache in the chain of caches. */
2369 mutex_lock(&cache_chain_mutex);
2371 * the chain is never empty, cache_cache is never destroyed
2373 list_del(&cachep->next);
2374 mutex_unlock(&cache_chain_mutex);
2376 if (__cache_shrink(cachep)) {
2377 slab_error(cachep, "Can't free all objects");
2378 mutex_lock(&cache_chain_mutex);
2379 list_add(&cachep->next, &cache_chain);
2380 mutex_unlock(&cache_chain_mutex);
2381 unlock_cpu_hotplug();
2385 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2388 for_each_online_cpu(i)
2389 kfree(cachep->array[i]);
2391 /* NUMA: free the list3 structures */
2392 for_each_online_node(i) {
2393 l3 = cachep->nodelists[i];
2396 free_alien_cache(l3->alien);
2400 kmem_cache_free(&cache_cache, cachep);
2401 unlock_cpu_hotplug();
2404 EXPORT_SYMBOL(kmem_cache_destroy);
2406 /* Get the memory for a slab management obj. */
2407 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2408 int colour_off, gfp_t local_flags,
2413 if (OFF_SLAB(cachep)) {
2414 /* Slab management obj is off-slab. */
2415 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2416 local_flags, nodeid);
2420 slabp = objp + colour_off;
2421 colour_off += cachep->slab_size;
2424 slabp->colouroff = colour_off;
2425 slabp->s_mem = objp + colour_off;
2426 slabp->nodeid = nodeid;
2430 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2432 return (kmem_bufctl_t *) (slabp + 1);
2435 static void cache_init_objs(struct kmem_cache *cachep,
2436 struct slab *slabp, unsigned long ctor_flags)
2440 for (i = 0; i < cachep->num; i++) {
2441 void *objp = index_to_obj(cachep, slabp, i);
2443 /* need to poison the objs? */
2444 if (cachep->flags & SLAB_POISON)
2445 poison_obj(cachep, objp, POISON_FREE);
2446 if (cachep->flags & SLAB_STORE_USER)
2447 *dbg_userword(cachep, objp) = NULL;
2449 if (cachep->flags & SLAB_RED_ZONE) {
2450 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2451 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2454 * Constructors are not allowed to allocate memory from the same
2455 * cache which they are a constructor for. Otherwise, deadlock.
2456 * They must also be threaded.
2458 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2459 cachep->ctor(objp + obj_offset(cachep), cachep,
2462 if (cachep->flags & SLAB_RED_ZONE) {
2463 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2464 slab_error(cachep, "constructor overwrote the"
2465 " end of an object");
2466 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2467 slab_error(cachep, "constructor overwrote the"
2468 " start of an object");
2470 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2471 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2472 kernel_map_pages(virt_to_page(objp),
2473 cachep->buffer_size / PAGE_SIZE, 0);
2476 cachep->ctor(objp, cachep, ctor_flags);
2478 slab_bufctl(slabp)[i] = i + 1;
2480 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2484 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2486 if (flags & SLAB_DMA)
2487 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2489 BUG_ON(cachep->gfpflags & GFP_DMA);
2492 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2495 void *objp = index_to_obj(cachep, slabp, slabp->free);
2499 next = slab_bufctl(slabp)[slabp->free];
2501 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2502 WARN_ON(slabp->nodeid != nodeid);
2509 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2510 void *objp, int nodeid)
2512 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2515 /* Verify that the slab belongs to the intended node */
2516 WARN_ON(slabp->nodeid != nodeid);
2518 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2519 printk(KERN_ERR "slab: double free detected in cache "
2520 "'%s', objp %p\n", cachep->name, objp);
2524 slab_bufctl(slabp)[objnr] = slabp->free;
2525 slabp->free = objnr;
2530 * Map pages beginning at addr to the given cache and slab. This is required
2531 * for the slab allocator to be able to lookup the cache and slab of a
2532 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2534 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2540 page = virt_to_page(addr);
2543 if (likely(!PageCompound(page)))
2544 nr_pages <<= cache->gfporder;
2547 page_set_cache(page, cache);
2548 page_set_slab(page, slab);
2550 } while (--nr_pages);
2554 * Grow (by 1) the number of slabs within a cache. This is called by
2555 * kmem_cache_alloc() when there are no active objs left in a cache.
2557 static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid)
2563 unsigned long ctor_flags;
2564 struct kmem_list3 *l3;
2567 * Be lazy and only check for valid flags here, keeping it out of the
2568 * critical path in kmem_cache_alloc().
2570 BUG_ON(flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW));
2571 if (flags & SLAB_NO_GROW)
2574 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2575 local_flags = (flags & SLAB_LEVEL_MASK);
2576 if (!(local_flags & __GFP_WAIT))
2578 * Not allowed to sleep. Need to tell a constructor about
2579 * this - it might need to know...
2581 ctor_flags |= SLAB_CTOR_ATOMIC;
2583 /* Take the l3 list lock to change the colour_next on this node */
2585 l3 = cachep->nodelists[nodeid];
2586 spin_lock(&l3->list_lock);
2588 /* Get colour for the slab, and cal the next value. */
2589 offset = l3->colour_next;
2591 if (l3->colour_next >= cachep->colour)
2592 l3->colour_next = 0;
2593 spin_unlock(&l3->list_lock);
2595 offset *= cachep->colour_off;
2597 if (local_flags & __GFP_WAIT)
2601 * The test for missing atomic flag is performed here, rather than
2602 * the more obvious place, simply to reduce the critical path length
2603 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2604 * will eventually be caught here (where it matters).
2606 kmem_flagcheck(cachep, flags);
2609 * Get mem for the objs. Attempt to allocate a physical page from
2612 objp = kmem_getpages(cachep, flags, nodeid);
2616 /* Get slab management. */
2617 slabp = alloc_slabmgmt(cachep, objp, offset, local_flags, nodeid);
2621 slabp->nodeid = nodeid;
2622 slab_map_pages(cachep, slabp, objp);
2624 cache_init_objs(cachep, slabp, ctor_flags);
2626 if (local_flags & __GFP_WAIT)
2627 local_irq_disable();
2629 spin_lock(&l3->list_lock);
2631 /* Make slab active. */
2632 list_add_tail(&slabp->list, &(l3->slabs_free));
2633 STATS_INC_GROWN(cachep);
2634 l3->free_objects += cachep->num;
2635 spin_unlock(&l3->list_lock);
2638 kmem_freepages(cachep, objp);
2640 if (local_flags & __GFP_WAIT)
2641 local_irq_disable();
2648 * Perform extra freeing checks:
2649 * - detect bad pointers.
2650 * - POISON/RED_ZONE checking
2651 * - destructor calls, for caches with POISON+dtor
2653 static void kfree_debugcheck(const void *objp)
2657 if (!virt_addr_valid(objp)) {
2658 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2659 (unsigned long)objp);
2662 page = virt_to_page(objp);
2663 if (!PageSlab(page)) {
2664 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2665 (unsigned long)objp);
2670 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2672 unsigned long redzone1, redzone2;
2674 redzone1 = *dbg_redzone1(cache, obj);
2675 redzone2 = *dbg_redzone2(cache, obj);
2680 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2683 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2684 slab_error(cache, "double free detected");
2686 slab_error(cache, "memory outside object was overwritten");
2688 printk(KERN_ERR "%p: redzone 1:0x%lx, redzone 2:0x%lx.\n",
2689 obj, redzone1, redzone2);
2692 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2699 objp -= obj_offset(cachep);
2700 kfree_debugcheck(objp);
2701 page = virt_to_page(objp);
2703 slabp = page_get_slab(page);
2705 if (cachep->flags & SLAB_RED_ZONE) {
2706 verify_redzone_free(cachep, objp);
2707 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2708 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2710 if (cachep->flags & SLAB_STORE_USER)
2711 *dbg_userword(cachep, objp) = caller;
2713 objnr = obj_to_index(cachep, slabp, objp);
2715 BUG_ON(objnr >= cachep->num);
2716 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2718 if (cachep->flags & SLAB_DEBUG_INITIAL) {
2720 * Need to call the slab's constructor so the caller can
2721 * perform a verify of its state (debugging). Called without
2722 * the cache-lock held.
2724 cachep->ctor(objp + obj_offset(cachep),
2725 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
2727 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2728 /* we want to cache poison the object,
2729 * call the destruction callback
2731 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
2733 #ifdef CONFIG_DEBUG_SLAB_LEAK
2734 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2736 if (cachep->flags & SLAB_POISON) {
2737 #ifdef CONFIG_DEBUG_PAGEALLOC
2738 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2739 store_stackinfo(cachep, objp, (unsigned long)caller);
2740 kernel_map_pages(virt_to_page(objp),
2741 cachep->buffer_size / PAGE_SIZE, 0);
2743 poison_obj(cachep, objp, POISON_FREE);
2746 poison_obj(cachep, objp, POISON_FREE);
2752 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2757 /* Check slab's freelist to see if this obj is there. */
2758 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2760 if (entries > cachep->num || i >= cachep->num)
2763 if (entries != cachep->num - slabp->inuse) {
2765 printk(KERN_ERR "slab: Internal list corruption detected in "
2766 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2767 cachep->name, cachep->num, slabp, slabp->inuse);
2769 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2772 printk("\n%03x:", i);
2773 printk(" %02x", ((unsigned char *)slabp)[i]);
2780 #define kfree_debugcheck(x) do { } while(0)
2781 #define cache_free_debugcheck(x,objp,z) (objp)
2782 #define check_slabp(x,y) do { } while(0)
2785 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2788 struct kmem_list3 *l3;
2789 struct array_cache *ac;
2792 ac = cpu_cache_get(cachep);
2794 batchcount = ac->batchcount;
2795 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2797 * If there was little recent activity on this cache, then
2798 * perform only a partial refill. Otherwise we could generate
2801 batchcount = BATCHREFILL_LIMIT;
2803 l3 = cachep->nodelists[numa_node_id()];
2805 BUG_ON(ac->avail > 0 || !l3);
2806 spin_lock(&l3->list_lock);
2808 /* See if we can refill from the shared array */
2809 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2812 while (batchcount > 0) {
2813 struct list_head *entry;
2815 /* Get slab alloc is to come from. */
2816 entry = l3->slabs_partial.next;
2817 if (entry == &l3->slabs_partial) {
2818 l3->free_touched = 1;
2819 entry = l3->slabs_free.next;
2820 if (entry == &l3->slabs_free)
2824 slabp = list_entry(entry, struct slab, list);
2825 check_slabp(cachep, slabp);
2826 check_spinlock_acquired(cachep);
2827 while (slabp->inuse < cachep->num && batchcount--) {
2828 STATS_INC_ALLOCED(cachep);
2829 STATS_INC_ACTIVE(cachep);
2830 STATS_SET_HIGH(cachep);
2832 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2835 check_slabp(cachep, slabp);
2837 /* move slabp to correct slabp list: */
2838 list_del(&slabp->list);
2839 if (slabp->free == BUFCTL_END)
2840 list_add(&slabp->list, &l3->slabs_full);
2842 list_add(&slabp->list, &l3->slabs_partial);
2846 l3->free_objects -= ac->avail;
2848 spin_unlock(&l3->list_lock);
2850 if (unlikely(!ac->avail)) {
2852 x = cache_grow(cachep, flags, numa_node_id());
2854 /* cache_grow can reenable interrupts, then ac could change. */
2855 ac = cpu_cache_get(cachep);
2856 if (!x && ac->avail == 0) /* no objects in sight? abort */
2859 if (!ac->avail) /* objects refilled by interrupt? */
2863 return ac->entry[--ac->avail];
2866 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2869 might_sleep_if(flags & __GFP_WAIT);
2871 kmem_flagcheck(cachep, flags);
2876 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2877 gfp_t flags, void *objp, void *caller)
2881 if (cachep->flags & SLAB_POISON) {
2882 #ifdef CONFIG_DEBUG_PAGEALLOC
2883 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
2884 kernel_map_pages(virt_to_page(objp),
2885 cachep->buffer_size / PAGE_SIZE, 1);
2887 check_poison_obj(cachep, objp);
2889 check_poison_obj(cachep, objp);
2891 poison_obj(cachep, objp, POISON_INUSE);
2893 if (cachep->flags & SLAB_STORE_USER)
2894 *dbg_userword(cachep, objp) = caller;
2896 if (cachep->flags & SLAB_RED_ZONE) {
2897 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2898 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2899 slab_error(cachep, "double free, or memory outside"
2900 " object was overwritten");
2902 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
2903 objp, *dbg_redzone1(cachep, objp),
2904 *dbg_redzone2(cachep, objp));
2906 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2907 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2909 #ifdef CONFIG_DEBUG_SLAB_LEAK
2914 slabp = page_get_slab(virt_to_page(objp));
2915 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
2916 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
2919 objp += obj_offset(cachep);
2920 if (cachep->ctor && cachep->flags & SLAB_POISON) {
2921 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2923 if (!(flags & __GFP_WAIT))
2924 ctor_flags |= SLAB_CTOR_ATOMIC;
2926 cachep->ctor(objp, cachep, ctor_flags);
2931 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2934 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
2937 struct array_cache *ac;
2940 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
2941 objp = alternate_node_alloc(cachep, flags);
2948 ac = cpu_cache_get(cachep);
2949 if (likely(ac->avail)) {
2950 STATS_INC_ALLOCHIT(cachep);
2952 objp = ac->entry[--ac->avail];
2954 STATS_INC_ALLOCMISS(cachep);
2955 objp = cache_alloc_refill(cachep, flags);
2960 static __always_inline void *__cache_alloc(struct kmem_cache *cachep,
2961 gfp_t flags, void *caller)
2963 unsigned long save_flags;
2966 cache_alloc_debugcheck_before(cachep, flags);
2968 local_irq_save(save_flags);
2969 objp = ____cache_alloc(cachep, flags);
2970 local_irq_restore(save_flags);
2971 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
2979 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
2981 * If we are in_interrupt, then process context, including cpusets and
2982 * mempolicy, may not apply and should not be used for allocation policy.
2984 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
2986 int nid_alloc, nid_here;
2990 nid_alloc = nid_here = numa_node_id();
2991 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
2992 nid_alloc = cpuset_mem_spread_node();
2993 else if (current->mempolicy)
2994 nid_alloc = slab_node(current->mempolicy);
2995 if (nid_alloc != nid_here)
2996 return __cache_alloc_node(cachep, flags, nid_alloc);
3001 * A interface to enable slab creation on nodeid
3003 static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3006 struct list_head *entry;
3008 struct kmem_list3 *l3;
3012 l3 = cachep->nodelists[nodeid];
3017 spin_lock(&l3->list_lock);
3018 entry = l3->slabs_partial.next;
3019 if (entry == &l3->slabs_partial) {
3020 l3->free_touched = 1;
3021 entry = l3->slabs_free.next;
3022 if (entry == &l3->slabs_free)
3026 slabp = list_entry(entry, struct slab, list);
3027 check_spinlock_acquired_node(cachep, nodeid);
3028 check_slabp(cachep, slabp);
3030 STATS_INC_NODEALLOCS(cachep);
3031 STATS_INC_ACTIVE(cachep);
3032 STATS_SET_HIGH(cachep);
3034 BUG_ON(slabp->inuse == cachep->num);
3036 obj = slab_get_obj(cachep, slabp, nodeid);
3037 check_slabp(cachep, slabp);
3039 /* move slabp to correct slabp list: */
3040 list_del(&slabp->list);
3042 if (slabp->free == BUFCTL_END)
3043 list_add(&slabp->list, &l3->slabs_full);
3045 list_add(&slabp->list, &l3->slabs_partial);
3047 spin_unlock(&l3->list_lock);
3051 spin_unlock(&l3->list_lock);
3052 x = cache_grow(cachep, flags, nodeid);
3064 * Caller needs to acquire correct kmem_list's list_lock
3066 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3070 struct kmem_list3 *l3;
3072 for (i = 0; i < nr_objects; i++) {
3073 void *objp = objpp[i];
3076 slabp = virt_to_slab(objp);
3077 l3 = cachep->nodelists[node];
3078 list_del(&slabp->list);
3079 check_spinlock_acquired_node(cachep, node);
3080 check_slabp(cachep, slabp);
3081 slab_put_obj(cachep, slabp, objp, node);
3082 STATS_DEC_ACTIVE(cachep);
3084 check_slabp(cachep, slabp);
3086 /* fixup slab chains */
3087 if (slabp->inuse == 0) {
3088 if (l3->free_objects > l3->free_limit) {
3089 l3->free_objects -= cachep->num;
3091 * It is safe to drop the lock. The slab is
3092 * no longer linked to the cache. cachep
3093 * cannot disappear - we are using it and
3094 * all destruction of caches must be
3095 * serialized properly by the user.
3097 spin_unlock(&l3->list_lock);
3098 slab_destroy(cachep, slabp);
3099 spin_lock(&l3->list_lock);
3101 list_add(&slabp->list, &l3->slabs_free);
3104 /* Unconditionally move a slab to the end of the
3105 * partial list on free - maximum time for the
3106 * other objects to be freed, too.
3108 list_add_tail(&slabp->list, &l3->slabs_partial);
3113 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3116 struct kmem_list3 *l3;
3117 int node = numa_node_id();
3119 batchcount = ac->batchcount;
3121 BUG_ON(!batchcount || batchcount > ac->avail);
3124 l3 = cachep->nodelists[node];
3125 spin_lock(&l3->list_lock);
3127 struct array_cache *shared_array = l3->shared;
3128 int max = shared_array->limit - shared_array->avail;
3130 if (batchcount > max)
3132 memcpy(&(shared_array->entry[shared_array->avail]),
3133 ac->entry, sizeof(void *) * batchcount);
3134 shared_array->avail += batchcount;
3139 free_block(cachep, ac->entry, batchcount, node);
3144 struct list_head *p;
3146 p = l3->slabs_free.next;
3147 while (p != &(l3->slabs_free)) {
3150 slabp = list_entry(p, struct slab, list);
3151 BUG_ON(slabp->inuse);
3156 STATS_SET_FREEABLE(cachep, i);
3159 spin_unlock(&l3->list_lock);
3160 ac->avail -= batchcount;
3161 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3165 * Release an obj back to its cache. If the obj has a constructed state, it must
3166 * be in this state _before_ it is released. Called with disabled ints.
3168 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3170 struct array_cache *ac = cpu_cache_get(cachep);
3173 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3175 if (cache_free_alien(cachep, objp))
3178 if (likely(ac->avail < ac->limit)) {
3179 STATS_INC_FREEHIT(cachep);
3180 ac->entry[ac->avail++] = objp;
3183 STATS_INC_FREEMISS(cachep);
3184 cache_flusharray(cachep, ac);
3185 ac->entry[ac->avail++] = objp;
3190 * kmem_cache_alloc - Allocate an object
3191 * @cachep: The cache to allocate from.
3192 * @flags: See kmalloc().
3194 * Allocate an object from this cache. The flags are only relevant
3195 * if the cache has no available objects.
3197 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3199 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3201 EXPORT_SYMBOL(kmem_cache_alloc);
3204 * kmem_cache_alloc - Allocate an object. The memory is set to zero.
3205 * @cache: The cache to allocate from.
3206 * @flags: See kmalloc().
3208 * Allocate an object from this cache and set the allocated memory to zero.
3209 * The flags are only relevant if the cache has no available objects.
3211 void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
3213 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
3215 memset(ret, 0, obj_size(cache));
3218 EXPORT_SYMBOL(kmem_cache_zalloc);
3221 * kmem_ptr_validate - check if an untrusted pointer might
3223 * @cachep: the cache we're checking against
3224 * @ptr: pointer to validate
3226 * This verifies that the untrusted pointer looks sane:
3227 * it is _not_ a guarantee that the pointer is actually
3228 * part of the slab cache in question, but it at least
3229 * validates that the pointer can be dereferenced and
3230 * looks half-way sane.
3232 * Currently only used for dentry validation.
3234 int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr)
3236 unsigned long addr = (unsigned long)ptr;
3237 unsigned long min_addr = PAGE_OFFSET;
3238 unsigned long align_mask = BYTES_PER_WORD - 1;
3239 unsigned long size = cachep->buffer_size;
3242 if (unlikely(addr < min_addr))
3244 if (unlikely(addr > (unsigned long)high_memory - size))
3246 if (unlikely(addr & align_mask))
3248 if (unlikely(!kern_addr_valid(addr)))
3250 if (unlikely(!kern_addr_valid(addr + size - 1)))
3252 page = virt_to_page(ptr);
3253 if (unlikely(!PageSlab(page)))
3255 if (unlikely(page_get_cache(page) != cachep))
3264 * kmem_cache_alloc_node - Allocate an object on the specified node
3265 * @cachep: The cache to allocate from.
3266 * @flags: See kmalloc().
3267 * @nodeid: node number of the target node.
3269 * Identical to kmem_cache_alloc, except that this function is slow
3270 * and can sleep. And it will allocate memory on the given node, which
3271 * can improve the performance for cpu bound structures.
3272 * New and improved: it will now make sure that the object gets
3273 * put on the correct node list so that there is no false sharing.
3275 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3277 unsigned long save_flags;
3280 cache_alloc_debugcheck_before(cachep, flags);
3281 local_irq_save(save_flags);
3283 if (nodeid == -1 || nodeid == numa_node_id() ||
3284 !cachep->nodelists[nodeid])
3285 ptr = ____cache_alloc(cachep, flags);
3287 ptr = __cache_alloc_node(cachep, flags, nodeid);
3288 local_irq_restore(save_flags);
3290 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr,
3291 __builtin_return_address(0));
3295 EXPORT_SYMBOL(kmem_cache_alloc_node);
3297 void *kmalloc_node(size_t size, gfp_t flags, int node)
3299 struct kmem_cache *cachep;
3301 cachep = kmem_find_general_cachep(size, flags);
3302 if (unlikely(cachep == NULL))
3304 return kmem_cache_alloc_node(cachep, flags, node);
3306 EXPORT_SYMBOL(kmalloc_node);
3310 * __do_kmalloc - allocate memory
3311 * @size: how many bytes of memory are required.
3312 * @flags: the type of memory to allocate (see kmalloc).
3313 * @caller: function caller for debug tracking of the caller
3315 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3318 struct kmem_cache *cachep;
3320 /* If you want to save a few bytes .text space: replace
3322 * Then kmalloc uses the uninlined functions instead of the inline
3325 cachep = __find_general_cachep(size, flags);
3326 if (unlikely(cachep == NULL))
3328 return __cache_alloc(cachep, flags, caller);
3332 void *__kmalloc(size_t size, gfp_t flags)
3334 #ifndef CONFIG_DEBUG_SLAB
3335 return __do_kmalloc(size, flags, NULL);
3337 return __do_kmalloc(size, flags, __builtin_return_address(0));
3340 EXPORT_SYMBOL(__kmalloc);
3342 #ifdef CONFIG_DEBUG_SLAB
3343 void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3345 return __do_kmalloc(size, flags, caller);
3347 EXPORT_SYMBOL(__kmalloc_track_caller);
3352 * __alloc_percpu - allocate one copy of the object for every present
3353 * cpu in the system, zeroing them.
3354 * Objects should be dereferenced using the per_cpu_ptr macro only.
3356 * @size: how many bytes of memory are required.
3358 void *__alloc_percpu(size_t size)
3361 struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL);
3367 * Cannot use for_each_online_cpu since a cpu may come online
3368 * and we have no way of figuring out how to fix the array
3369 * that we have allocated then....
3371 for_each_possible_cpu(i) {
3372 int node = cpu_to_node(i);
3374 if (node_online(node))
3375 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
3377 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
3379 if (!pdata->ptrs[i])
3381 memset(pdata->ptrs[i], 0, size);
3384 /* Catch derefs w/o wrappers */
3385 return (void *)(~(unsigned long)pdata);
3389 if (!cpu_possible(i))
3391 kfree(pdata->ptrs[i]);
3396 EXPORT_SYMBOL(__alloc_percpu);
3400 * kmem_cache_free - Deallocate an object
3401 * @cachep: The cache the allocation was from.
3402 * @objp: The previously allocated object.
3404 * Free an object which was previously allocated from this
3407 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3409 unsigned long flags;
3411 BUG_ON(virt_to_cache(objp) != cachep);
3413 local_irq_save(flags);
3414 __cache_free(cachep, objp);
3415 local_irq_restore(flags);
3417 EXPORT_SYMBOL(kmem_cache_free);
3420 * kfree - free previously allocated memory
3421 * @objp: pointer returned by kmalloc.
3423 * If @objp is NULL, no operation is performed.
3425 * Don't free memory not originally allocated by kmalloc()
3426 * or you will run into trouble.
3428 void kfree(const void *objp)
3430 struct kmem_cache *c;
3431 unsigned long flags;
3433 if (unlikely(!objp))
3435 local_irq_save(flags);
3436 kfree_debugcheck(objp);
3437 c = virt_to_cache(objp);
3438 debug_check_no_locks_freed(objp, obj_size(c));
3439 __cache_free(c, (void *)objp);
3440 local_irq_restore(flags);
3442 EXPORT_SYMBOL(kfree);
3446 * free_percpu - free previously allocated percpu memory
3447 * @objp: pointer returned by alloc_percpu.
3449 * Don't free memory not originally allocated by alloc_percpu()
3450 * The complemented objp is to check for that.
3452 void free_percpu(const void *objp)
3455 struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp);
3458 * We allocate for all cpus so we cannot use for online cpu here.
3460 for_each_possible_cpu(i)
3464 EXPORT_SYMBOL(free_percpu);
3467 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3469 return obj_size(cachep);
3471 EXPORT_SYMBOL(kmem_cache_size);
3473 const char *kmem_cache_name(struct kmem_cache *cachep)
3475 return cachep->name;
3477 EXPORT_SYMBOL_GPL(kmem_cache_name);
3480 * This initializes kmem_list3 or resizes varioius caches for all nodes.
3482 static int alloc_kmemlist(struct kmem_cache *cachep)
3485 struct kmem_list3 *l3;
3486 struct array_cache *new_shared;
3487 struct array_cache **new_alien;
3489 for_each_online_node(node) {
3491 new_alien = alloc_alien_cache(node, cachep->limit);
3495 new_shared = alloc_arraycache(node,
3496 cachep->shared*cachep->batchcount,
3499 free_alien_cache(new_alien);
3503 l3 = cachep->nodelists[node];
3505 struct array_cache *shared = l3->shared;
3507 spin_lock_irq(&l3->list_lock);
3510 free_block(cachep, shared->entry,
3511 shared->avail, node);
3513 l3->shared = new_shared;
3515 l3->alien = new_alien;
3518 l3->free_limit = (1 + nr_cpus_node(node)) *
3519 cachep->batchcount + cachep->num;
3520 spin_unlock_irq(&l3->list_lock);
3522 free_alien_cache(new_alien);
3525 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
3527 free_alien_cache(new_alien);
3532 kmem_list3_init(l3);
3533 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3534 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3535 l3->shared = new_shared;
3536 l3->alien = new_alien;
3537 l3->free_limit = (1 + nr_cpus_node(node)) *
3538 cachep->batchcount + cachep->num;
3539 cachep->nodelists[node] = l3;
3544 if (!cachep->next.next) {
3545 /* Cache is not active yet. Roll back what we did */
3548 if (cachep->nodelists[node]) {
3549 l3 = cachep->nodelists[node];
3552 free_alien_cache(l3->alien);
3554 cachep->nodelists[node] = NULL;
3562 struct ccupdate_struct {
3563 struct kmem_cache *cachep;
3564 struct array_cache *new[NR_CPUS];
3567 static void do_ccupdate_local(void *info)
3569 struct ccupdate_struct *new = info;
3570 struct array_cache *old;
3573 old = cpu_cache_get(new->cachep);
3575 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3576 new->new[smp_processor_id()] = old;
3579 /* Always called with the cache_chain_mutex held */
3580 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3581 int batchcount, int shared)
3583 struct ccupdate_struct new;
3586 memset(&new.new, 0, sizeof(new.new));
3587 for_each_online_cpu(i) {
3588 new.new[i] = alloc_arraycache(cpu_to_node(i), limit,
3591 for (i--; i >= 0; i--)
3596 new.cachep = cachep;
3598 on_each_cpu(do_ccupdate_local, (void *)&new, 1, 1);
3601 cachep->batchcount = batchcount;
3602 cachep->limit = limit;
3603 cachep->shared = shared;
3605 for_each_online_cpu(i) {
3606 struct array_cache *ccold = new.new[i];
3609 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3610 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3611 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3615 err = alloc_kmemlist(cachep);
3617 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
3618 cachep->name, -err);
3624 /* Called with cache_chain_mutex held always */
3625 static void enable_cpucache(struct kmem_cache *cachep)
3631 * The head array serves three purposes:
3632 * - create a LIFO ordering, i.e. return objects that are cache-warm
3633 * - reduce the number of spinlock operations.
3634 * - reduce the number of linked list operations on the slab and
3635 * bufctl chains: array operations are cheaper.
3636 * The numbers are guessed, we should auto-tune as described by
3639 if (cachep->buffer_size > 131072)
3641 else if (cachep->buffer_size > PAGE_SIZE)
3643 else if (cachep->buffer_size > 1024)
3645 else if (cachep->buffer_size > 256)
3651 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3652 * allocation behaviour: Most allocs on one cpu, most free operations
3653 * on another cpu. For these cases, an efficient object passing between
3654 * cpus is necessary. This is provided by a shared array. The array
3655 * replaces Bonwick's magazine layer.
3656 * On uniprocessor, it's functionally equivalent (but less efficient)
3657 * to a larger limit. Thus disabled by default.
3661 if (cachep->buffer_size <= PAGE_SIZE)
3667 * With debugging enabled, large batchcount lead to excessively long
3668 * periods with disabled local interrupts. Limit the batchcount
3673 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
3675 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3676 cachep->name, -err);
3680 * Drain an array if it contains any elements taking the l3 lock only if
3681 * necessary. Note that the l3 listlock also protects the array_cache
3682 * if drain_array() is used on the shared array.
3684 void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
3685 struct array_cache *ac, int force, int node)
3689 if (!ac || !ac->avail)
3691 if (ac->touched && !force) {
3694 spin_lock_irq(&l3->list_lock);
3696 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3697 if (tofree > ac->avail)
3698 tofree = (ac->avail + 1) / 2;
3699 free_block(cachep, ac->entry, tofree, node);
3700 ac->avail -= tofree;
3701 memmove(ac->entry, &(ac->entry[tofree]),
3702 sizeof(void *) * ac->avail);
3704 spin_unlock_irq(&l3->list_lock);
3709 * cache_reap - Reclaim memory from caches.
3710 * @unused: unused parameter
3712 * Called from workqueue/eventd every few seconds.
3714 * - clear the per-cpu caches for this CPU.
3715 * - return freeable pages to the main free memory pool.
3717 * If we cannot acquire the cache chain mutex then just give up - we'll try
3718 * again on the next iteration.
3720 static void cache_reap(void *unused)
3722 struct kmem_cache *searchp;
3723 struct kmem_list3 *l3;
3724 int node = numa_node_id();
3726 if (!mutex_trylock(&cache_chain_mutex)) {
3727 /* Give up. Setup the next iteration. */
3728 schedule_delayed_work(&__get_cpu_var(reap_work),
3733 list_for_each_entry(searchp, &cache_chain, next) {
3737 * We only take the l3 lock if absolutely necessary and we
3738 * have established with reasonable certainty that
3739 * we can do some work if the lock was obtained.
3741 l3 = searchp->nodelists[node];
3743 reap_alien(searchp, l3);
3745 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
3748 * These are racy checks but it does not matter
3749 * if we skip one check or scan twice.
3751 if (time_after(l3->next_reap, jiffies))
3754 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
3756 drain_array(searchp, l3, l3->shared, 0, node);
3758 if (l3->free_touched)
3759 l3->free_touched = 0;
3763 freed = drain_freelist(searchp, l3, (l3->free_limit +
3764 5 * searchp->num - 1) / (5 * searchp->num));
3765 STATS_ADD_REAPED(searchp, freed);
3771 mutex_unlock(&cache_chain_mutex);
3773 refresh_cpu_vm_stats(smp_processor_id());
3774 /* Set up the next iteration */
3775 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
3778 #ifdef CONFIG_PROC_FS
3780 static void print_slabinfo_header(struct seq_file *m)
3783 * Output format version, so at least we can change it
3784 * without _too_ many complaints.
3787 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3789 seq_puts(m, "slabinfo - version: 2.1\n");
3791 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3792 "<objperslab> <pagesperslab>");
3793 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3794 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3796 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3797 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
3798 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3803 static void *s_start(struct seq_file *m, loff_t *pos)
3806 struct list_head *p;
3808 mutex_lock(&cache_chain_mutex);
3810 print_slabinfo_header(m);
3811 p = cache_chain.next;
3814 if (p == &cache_chain)
3817 return list_entry(p, struct kmem_cache, next);
3820 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3822 struct kmem_cache *cachep = p;
3824 return cachep->next.next == &cache_chain ?
3825 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
3828 static void s_stop(struct seq_file *m, void *p)
3830 mutex_unlock(&cache_chain_mutex);
3833 static int s_show(struct seq_file *m, void *p)
3835 struct kmem_cache *cachep = p;
3837 unsigned long active_objs;
3838 unsigned long num_objs;
3839 unsigned long active_slabs = 0;
3840 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3844 struct kmem_list3 *l3;
3848 for_each_online_node(node) {
3849 l3 = cachep->nodelists[node];
3854 spin_lock_irq(&l3->list_lock);
3856 list_for_each_entry(slabp, &l3->slabs_full, list) {
3857 if (slabp->inuse != cachep->num && !error)
3858 error = "slabs_full accounting error";
3859 active_objs += cachep->num;
3862 list_for_each_entry(slabp, &l3->slabs_partial, list) {
3863 if (slabp->inuse == cachep->num && !error)
3864 error = "slabs_partial inuse accounting error";
3865 if (!slabp->inuse && !error)
3866 error = "slabs_partial/inuse accounting error";
3867 active_objs += slabp->inuse;
3870 list_for_each_entry(slabp, &l3->slabs_free, list) {
3871 if (slabp->inuse && !error)
3872 error = "slabs_free/inuse accounting error";
3875 free_objects += l3->free_objects;
3877 shared_avail += l3->shared->avail;
3879 spin_unlock_irq(&l3->list_lock);
3881 num_slabs += active_slabs;
3882 num_objs = num_slabs * cachep->num;
3883 if (num_objs - active_objs != free_objects && !error)
3884 error = "free_objects accounting error";
3886 name = cachep->name;
3888 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3890 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
3891 name, active_objs, num_objs, cachep->buffer_size,
3892 cachep->num, (1 << cachep->gfporder));
3893 seq_printf(m, " : tunables %4u %4u %4u",
3894 cachep->limit, cachep->batchcount, cachep->shared);
3895 seq_printf(m, " : slabdata %6lu %6lu %6lu",
3896 active_slabs, num_slabs, shared_avail);
3899 unsigned long high = cachep->high_mark;
3900 unsigned long allocs = cachep->num_allocations;
3901 unsigned long grown = cachep->grown;
3902 unsigned long reaped = cachep->reaped;
3903 unsigned long errors = cachep->errors;
3904 unsigned long max_freeable = cachep->max_freeable;
3905 unsigned long node_allocs = cachep->node_allocs;
3906 unsigned long node_frees = cachep->node_frees;
3907 unsigned long overflows = cachep->node_overflow;
3909 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
3910 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
3911 reaped, errors, max_freeable, node_allocs,
3912 node_frees, overflows);
3916 unsigned long allochit = atomic_read(&cachep->allochit);
3917 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3918 unsigned long freehit = atomic_read(&cachep->freehit);
3919 unsigned long freemiss = atomic_read(&cachep->freemiss);
3921 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
3922 allochit, allocmiss, freehit, freemiss);
3930 * slabinfo_op - iterator that generates /proc/slabinfo
3939 * num-pages-per-slab
3940 * + further values on SMP and with statistics enabled
3943 struct seq_operations slabinfo_op = {
3950 #define MAX_SLABINFO_WRITE 128
3952 * slabinfo_write - Tuning for the slab allocator
3954 * @buffer: user buffer
3955 * @count: data length
3958 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
3959 size_t count, loff_t *ppos)
3961 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
3962 int limit, batchcount, shared, res;
3963 struct kmem_cache *cachep;
3965 if (count > MAX_SLABINFO_WRITE)
3967 if (copy_from_user(&kbuf, buffer, count))
3969 kbuf[MAX_SLABINFO_WRITE] = '\0';
3971 tmp = strchr(kbuf, ' ');
3976 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3979 /* Find the cache in the chain of caches. */
3980 mutex_lock(&cache_chain_mutex);
3982 list_for_each_entry(cachep, &cache_chain, next) {
3983 if (!strcmp(cachep->name, kbuf)) {
3984 if (limit < 1 || batchcount < 1 ||
3985 batchcount > limit || shared < 0) {
3988 res = do_tune_cpucache(cachep, limit,
3989 batchcount, shared);
3994 mutex_unlock(&cache_chain_mutex);
4000 #ifdef CONFIG_DEBUG_SLAB_LEAK
4002 static void *leaks_start(struct seq_file *m, loff_t *pos)
4005 struct list_head *p;
4007 mutex_lock(&cache_chain_mutex);
4008 p = cache_chain.next;
4011 if (p == &cache_chain)
4014 return list_entry(p, struct kmem_cache, next);
4017 static inline int add_caller(unsigned long *n, unsigned long v)
4027 unsigned long *q = p + 2 * i;
4041 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4047 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4053 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4054 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4056 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4061 static void show_symbol(struct seq_file *m, unsigned long address)
4063 #ifdef CONFIG_KALLSYMS
4066 unsigned long offset, size;
4067 char namebuf[KSYM_NAME_LEN+1];
4069 name = kallsyms_lookup(address, &size, &offset, &modname, namebuf);
4072 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4074 seq_printf(m, " [%s]", modname);
4078 seq_printf(m, "%p", (void *)address);
4081 static int leaks_show(struct seq_file *m, void *p)
4083 struct kmem_cache *cachep = p;
4085 struct kmem_list3 *l3;
4087 unsigned long *n = m->private;
4091 if (!(cachep->flags & SLAB_STORE_USER))
4093 if (!(cachep->flags & SLAB_RED_ZONE))
4096 /* OK, we can do it */
4100 for_each_online_node(node) {
4101 l3 = cachep->nodelists[node];
4106 spin_lock_irq(&l3->list_lock);
4108 list_for_each_entry(slabp, &l3->slabs_full, list)
4109 handle_slab(n, cachep, slabp);
4110 list_for_each_entry(slabp, &l3->slabs_partial, list)
4111 handle_slab(n, cachep, slabp);
4112 spin_unlock_irq(&l3->list_lock);
4114 name = cachep->name;
4116 /* Increase the buffer size */
4117 mutex_unlock(&cache_chain_mutex);
4118 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4120 /* Too bad, we are really out */
4122 mutex_lock(&cache_chain_mutex);
4125 *(unsigned long *)m->private = n[0] * 2;
4127 mutex_lock(&cache_chain_mutex);
4128 /* Now make sure this entry will be retried */
4132 for (i = 0; i < n[1]; i++) {
4133 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4134 show_symbol(m, n[2*i+2]);
4140 struct seq_operations slabstats_op = {
4141 .start = leaks_start,
4150 * ksize - get the actual amount of memory allocated for a given object
4151 * @objp: Pointer to the object
4153 * kmalloc may internally round up allocations and return more memory
4154 * than requested. ksize() can be used to determine the actual amount of
4155 * memory allocated. The caller may use this additional memory, even though
4156 * a smaller amount of memory was initially specified with the kmalloc call.
4157 * The caller must guarantee that objp points to a valid object previously
4158 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4159 * must not be freed during the duration of the call.
4161 unsigned int ksize(const void *objp)
4163 if (unlikely(objp == NULL))
4166 return obj_size(virt_to_cache(objp));