3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same intializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/config.h>
90 #include <linux/slab.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/cpuset.h>
98 #include <linux/seq_file.h>
99 #include <linux/notifier.h>
100 #include <linux/kallsyms.h>
101 #include <linux/cpu.h>
102 #include <linux/sysctl.h>
103 #include <linux/module.h>
104 #include <linux/rcupdate.h>
105 #include <linux/string.h>
106 #include <linux/nodemask.h>
107 #include <linux/mempolicy.h>
108 #include <linux/mutex.h>
110 #include <asm/uaccess.h>
111 #include <asm/cacheflush.h>
112 #include <asm/tlbflush.h>
113 #include <asm/page.h>
116 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
117 * SLAB_RED_ZONE & SLAB_POISON.
118 * 0 for faster, smaller code (especially in the critical paths).
120 * STATS - 1 to collect stats for /proc/slabinfo.
121 * 0 for faster, smaller code (especially in the critical paths).
123 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
126 #ifdef CONFIG_DEBUG_SLAB
129 #define FORCED_DEBUG 1
133 #define FORCED_DEBUG 0
136 /* Shouldn't this be in a header file somewhere? */
137 #define BYTES_PER_WORD sizeof(void *)
139 #ifndef cache_line_size
140 #define cache_line_size() L1_CACHE_BYTES
143 #ifndef ARCH_KMALLOC_MINALIGN
145 * Enforce a minimum alignment for the kmalloc caches.
146 * Usually, the kmalloc caches are cache_line_size() aligned, except when
147 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
148 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
149 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
150 * Note that this flag disables some debug features.
152 #define ARCH_KMALLOC_MINALIGN 0
155 #ifndef ARCH_SLAB_MINALIGN
157 * Enforce a minimum alignment for all caches.
158 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
159 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
160 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
161 * some debug features.
163 #define ARCH_SLAB_MINALIGN 0
166 #ifndef ARCH_KMALLOC_FLAGS
167 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
170 /* Legal flag mask for kmem_cache_create(). */
172 # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
173 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
175 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
176 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
177 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
179 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
180 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
181 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
182 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
188 * Bufctl's are used for linking objs within a slab
191 * This implementation relies on "struct page" for locating the cache &
192 * slab an object belongs to.
193 * This allows the bufctl structure to be small (one int), but limits
194 * the number of objects a slab (not a cache) can contain when off-slab
195 * bufctls are used. The limit is the size of the largest general cache
196 * that does not use off-slab slabs.
197 * For 32bit archs with 4 kB pages, is this 56.
198 * This is not serious, as it is only for large objects, when it is unwise
199 * to have too many per slab.
200 * Note: This limit can be raised by introducing a general cache whose size
201 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
204 typedef unsigned int kmem_bufctl_t
;
205 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
206 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
207 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-2)
209 /* Max number of objs-per-slab for caches which use off-slab slabs.
210 * Needed to avoid a possible looping condition in cache_grow().
212 static unsigned long offslab_limit
;
217 * Manages the objs in a slab. Placed either at the beginning of mem allocated
218 * for a slab, or allocated from an general cache.
219 * Slabs are chained into three list: fully used, partial, fully free slabs.
222 struct list_head list
;
223 unsigned long colouroff
;
224 void *s_mem
; /* including colour offset */
225 unsigned int inuse
; /* num of objs active in slab */
227 unsigned short nodeid
;
233 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
234 * arrange for kmem_freepages to be called via RCU. This is useful if
235 * we need to approach a kernel structure obliquely, from its address
236 * obtained without the usual locking. We can lock the structure to
237 * stabilize it and check it's still at the given address, only if we
238 * can be sure that the memory has not been meanwhile reused for some
239 * other kind of object (which our subsystem's lock might corrupt).
241 * rcu_read_lock before reading the address, then rcu_read_unlock after
242 * taking the spinlock within the structure expected at that address.
244 * We assume struct slab_rcu can overlay struct slab when destroying.
247 struct rcu_head head
;
248 struct kmem_cache
*cachep
;
256 * - LIFO ordering, to hand out cache-warm objects from _alloc
257 * - reduce the number of linked list operations
258 * - reduce spinlock operations
260 * The limit is stored in the per-cpu structure to reduce the data cache
267 unsigned int batchcount
;
268 unsigned int touched
;
271 * Must have this definition in here for the proper
272 * alignment of array_cache. Also simplifies accessing
274 * [0] is for gcc 2.95. It should really be [].
279 * bootstrap: The caches do not work without cpuarrays anymore, but the
280 * cpuarrays are allocated from the generic caches...
282 #define BOOT_CPUCACHE_ENTRIES 1
283 struct arraycache_init
{
284 struct array_cache cache
;
285 void *entries
[BOOT_CPUCACHE_ENTRIES
];
289 * The slab lists for all objects.
292 struct list_head slabs_partial
; /* partial list first, better asm code */
293 struct list_head slabs_full
;
294 struct list_head slabs_free
;
295 unsigned long free_objects
;
296 unsigned int free_limit
;
297 unsigned int colour_next
; /* Per-node cache coloring */
298 spinlock_t list_lock
;
299 struct array_cache
*shared
; /* shared per node */
300 struct array_cache
**alien
; /* on other nodes */
301 unsigned long next_reap
; /* updated without locking */
302 int free_touched
; /* updated without locking */
306 * Need this for bootstrapping a per node allocator.
308 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
309 struct kmem_list3 __initdata initkmem_list3
[NUM_INIT_LISTS
];
310 #define CACHE_CACHE 0
312 #define SIZE_L3 (1 + MAX_NUMNODES)
315 * This function must be completely optimized away if a constant is passed to
316 * it. Mostly the same as what is in linux/slab.h except it returns an index.
318 static __always_inline
int index_of(const size_t size
)
320 extern void __bad_size(void);
322 if (__builtin_constant_p(size
)) {
330 #include "linux/kmalloc_sizes.h"
338 #define INDEX_AC index_of(sizeof(struct arraycache_init))
339 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
341 static void kmem_list3_init(struct kmem_list3
*parent
)
343 INIT_LIST_HEAD(&parent
->slabs_full
);
344 INIT_LIST_HEAD(&parent
->slabs_partial
);
345 INIT_LIST_HEAD(&parent
->slabs_free
);
346 parent
->shared
= NULL
;
347 parent
->alien
= NULL
;
348 parent
->colour_next
= 0;
349 spin_lock_init(&parent
->list_lock
);
350 parent
->free_objects
= 0;
351 parent
->free_touched
= 0;
354 #define MAKE_LIST(cachep, listp, slab, nodeid) \
356 INIT_LIST_HEAD(listp); \
357 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
360 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
362 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
363 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
364 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
374 /* 1) per-cpu data, touched during every alloc/free */
375 struct array_cache
*array
[NR_CPUS
];
376 /* 2) Cache tunables. Protected by cache_chain_mutex */
377 unsigned int batchcount
;
381 unsigned int buffer_size
;
382 /* 3) touched by every alloc & free from the backend */
383 struct kmem_list3
*nodelists
[MAX_NUMNODES
];
385 unsigned int flags
; /* constant flags */
386 unsigned int num
; /* # of objs per slab */
388 /* 4) cache_grow/shrink */
389 /* order of pgs per slab (2^n) */
390 unsigned int gfporder
;
392 /* force GFP flags, e.g. GFP_DMA */
395 size_t colour
; /* cache colouring range */
396 unsigned int colour_off
; /* colour offset */
397 struct kmem_cache
*slabp_cache
;
398 unsigned int slab_size
;
399 unsigned int dflags
; /* dynamic flags */
401 /* constructor func */
402 void (*ctor
) (void *, struct kmem_cache
*, unsigned long);
404 /* de-constructor func */
405 void (*dtor
) (void *, struct kmem_cache
*, unsigned long);
407 /* 5) cache creation/removal */
409 struct list_head next
;
413 unsigned long num_active
;
414 unsigned long num_allocations
;
415 unsigned long high_mark
;
417 unsigned long reaped
;
418 unsigned long errors
;
419 unsigned long max_freeable
;
420 unsigned long node_allocs
;
421 unsigned long node_frees
;
429 * If debugging is enabled, then the allocator can add additional
430 * fields and/or padding to every object. buffer_size contains the total
431 * object size including these internal fields, the following two
432 * variables contain the offset to the user object and its size.
439 #define CFLGS_OFF_SLAB (0x80000000UL)
440 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
442 #define BATCHREFILL_LIMIT 16
444 * Optimization question: fewer reaps means less probability for unnessary
445 * cpucache drain/refill cycles.
447 * OTOH the cpuarrays can contain lots of objects,
448 * which could lock up otherwise freeable slabs.
450 #define REAPTIMEOUT_CPUC (2*HZ)
451 #define REAPTIMEOUT_LIST3 (4*HZ)
454 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
455 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
456 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
457 #define STATS_INC_GROWN(x) ((x)->grown++)
458 #define STATS_INC_REAPED(x) ((x)->reaped++)
459 #define STATS_SET_HIGH(x) \
461 if ((x)->num_active > (x)->high_mark) \
462 (x)->high_mark = (x)->num_active; \
464 #define STATS_INC_ERR(x) ((x)->errors++)
465 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
466 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
467 #define STATS_SET_FREEABLE(x, i) \
469 if ((x)->max_freeable < i) \
470 (x)->max_freeable = i; \
472 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
473 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
474 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
475 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
477 #define STATS_INC_ACTIVE(x) do { } while (0)
478 #define STATS_DEC_ACTIVE(x) do { } while (0)
479 #define STATS_INC_ALLOCED(x) do { } while (0)
480 #define STATS_INC_GROWN(x) do { } while (0)
481 #define STATS_INC_REAPED(x) do { } while (0)
482 #define STATS_SET_HIGH(x) do { } while (0)
483 #define STATS_INC_ERR(x) do { } while (0)
484 #define STATS_INC_NODEALLOCS(x) do { } while (0)
485 #define STATS_INC_NODEFREES(x) do { } while (0)
486 #define STATS_SET_FREEABLE(x, i) do { } while (0)
487 #define STATS_INC_ALLOCHIT(x) do { } while (0)
488 #define STATS_INC_ALLOCMISS(x) do { } while (0)
489 #define STATS_INC_FREEHIT(x) do { } while (0)
490 #define STATS_INC_FREEMISS(x) do { } while (0)
495 * Magic nums for obj red zoning.
496 * Placed in the first word before and the first word after an obj.
498 #define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
499 #define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
501 /* ...and for poisoning */
502 #define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
503 #define POISON_FREE 0x6b /* for use-after-free poisoning */
504 #define POISON_END 0xa5 /* end-byte of poisoning */
507 * memory layout of objects:
509 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
510 * the end of an object is aligned with the end of the real
511 * allocation. Catches writes behind the end of the allocation.
512 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
514 * cachep->obj_offset: The real object.
515 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
516 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
517 * [BYTES_PER_WORD long]
519 static int obj_offset(struct kmem_cache
*cachep
)
521 return cachep
->obj_offset
;
524 static int obj_size(struct kmem_cache
*cachep
)
526 return cachep
->obj_size
;
529 static unsigned long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
531 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
532 return (unsigned long*) (objp
+obj_offset(cachep
)-BYTES_PER_WORD
);
535 static unsigned long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
537 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
538 if (cachep
->flags
& SLAB_STORE_USER
)
539 return (unsigned long *)(objp
+ cachep
->buffer_size
-
541 return (unsigned long *)(objp
+ cachep
->buffer_size
- BYTES_PER_WORD
);
544 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
546 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
547 return (void **)(objp
+ cachep
->buffer_size
- BYTES_PER_WORD
);
552 #define obj_offset(x) 0
553 #define obj_size(cachep) (cachep->buffer_size)
554 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
555 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
556 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
561 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
564 #if defined(CONFIG_LARGE_ALLOCS)
565 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
566 #define MAX_GFP_ORDER 13 /* up to 32Mb */
567 #elif defined(CONFIG_MMU)
568 #define MAX_OBJ_ORDER 5 /* 32 pages */
569 #define MAX_GFP_ORDER 5 /* 32 pages */
571 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
572 #define MAX_GFP_ORDER 8 /* up to 1Mb */
576 * Do not go above this order unless 0 objects fit into the slab.
578 #define BREAK_GFP_ORDER_HI 1
579 #define BREAK_GFP_ORDER_LO 0
580 static int slab_break_gfp_order
= BREAK_GFP_ORDER_LO
;
583 * Functions for storing/retrieving the cachep and or slab from the page
584 * allocator. These are used to find the slab an obj belongs to. With kfree(),
585 * these are used to find the cache which an obj belongs to.
587 static inline void page_set_cache(struct page
*page
, struct kmem_cache
*cache
)
589 page
->lru
.next
= (struct list_head
*)cache
;
592 static inline struct kmem_cache
*page_get_cache(struct page
*page
)
594 if (unlikely(PageCompound(page
)))
595 page
= (struct page
*)page_private(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 return (struct slab
*)page
->lru
.prev
;
611 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
613 struct page
*page
= virt_to_page(obj
);
614 return page_get_cache(page
);
617 static inline struct slab
*virt_to_slab(const void *obj
)
619 struct page
*page
= virt_to_page(obj
);
620 return page_get_slab(page
);
623 static inline void *index_to_obj(struct kmem_cache
*cache
, struct slab
*slab
,
626 return slab
->s_mem
+ cache
->buffer_size
* idx
;
629 static inline unsigned int obj_to_index(struct kmem_cache
*cache
,
630 struct slab
*slab
, void *obj
)
632 return (unsigned)(obj
- slab
->s_mem
) / cache
->buffer_size
;
636 * These are the default caches for kmalloc. Custom caches can have other sizes.
638 struct cache_sizes malloc_sizes
[] = {
639 #define CACHE(x) { .cs_size = (x) },
640 #include <linux/kmalloc_sizes.h>
644 EXPORT_SYMBOL(malloc_sizes
);
646 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
652 static struct cache_names __initdata cache_names
[] = {
653 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
654 #include <linux/kmalloc_sizes.h>
659 static struct arraycache_init initarray_cache __initdata
=
660 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
661 static struct arraycache_init initarray_generic
=
662 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
664 /* internal cache of cache description objs */
665 static struct kmem_cache cache_cache
= {
667 .limit
= BOOT_CPUCACHE_ENTRIES
,
669 .buffer_size
= sizeof(struct kmem_cache
),
670 .name
= "kmem_cache",
672 .obj_size
= sizeof(struct kmem_cache
),
676 /* Guard access to the cache-chain. */
677 static DEFINE_MUTEX(cache_chain_mutex
);
678 static struct list_head cache_chain
;
681 * vm_enough_memory() looks at this to determine how many slab-allocated pages
682 * are possibly freeable under pressure
684 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
686 atomic_t slab_reclaim_pages
;
689 * chicken and egg problem: delay the per-cpu array allocation
690 * until the general caches are up.
699 static DEFINE_PER_CPU(struct work_struct
, reap_work
);
701 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
703 static void enable_cpucache(struct kmem_cache
*cachep
);
704 static void cache_reap(void *unused
);
705 static int __node_shrink(struct kmem_cache
*cachep
, int node
);
707 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
709 return cachep
->array
[smp_processor_id()];
712 static inline struct kmem_cache
*__find_general_cachep(size_t size
,
715 struct cache_sizes
*csizep
= malloc_sizes
;
718 /* This happens if someone tries to call
719 * kmem_cache_create(), or __kmalloc(), before
720 * the generic caches are initialized.
722 BUG_ON(malloc_sizes
[INDEX_AC
].cs_cachep
== NULL
);
724 while (size
> csizep
->cs_size
)
728 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
729 * has cs_{dma,}cachep==NULL. Thus no special case
730 * for large kmalloc calls required.
732 if (unlikely(gfpflags
& GFP_DMA
))
733 return csizep
->cs_dmacachep
;
734 return csizep
->cs_cachep
;
737 struct kmem_cache
*kmem_find_general_cachep(size_t size
, gfp_t gfpflags
)
739 return __find_general_cachep(size
, gfpflags
);
741 EXPORT_SYMBOL(kmem_find_general_cachep
);
743 static size_t slab_mgmt_size(size_t nr_objs
, size_t align
)
745 return ALIGN(sizeof(struct slab
)+nr_objs
*sizeof(kmem_bufctl_t
), align
);
749 * Calculate the number of objects and left-over bytes for a given buffer size.
751 static void cache_estimate(unsigned long gfporder
, size_t buffer_size
,
752 size_t align
, int flags
, size_t *left_over
,
757 size_t slab_size
= PAGE_SIZE
<< gfporder
;
760 * The slab management structure can be either off the slab or
761 * on it. For the latter case, the memory allocated for a
765 * - One kmem_bufctl_t for each object
766 * - Padding to respect alignment of @align
767 * - @buffer_size bytes for each object
769 * If the slab management structure is off the slab, then the
770 * alignment will already be calculated into the size. Because
771 * the slabs are all pages aligned, the objects will be at the
772 * correct alignment when allocated.
774 if (flags
& CFLGS_OFF_SLAB
) {
776 nr_objs
= slab_size
/ buffer_size
;
778 if (nr_objs
> SLAB_LIMIT
)
779 nr_objs
= SLAB_LIMIT
;
782 * Ignore padding for the initial guess. The padding
783 * is at most @align-1 bytes, and @buffer_size is at
784 * least @align. In the worst case, this result will
785 * be one greater than the number of objects that fit
786 * into the memory allocation when taking the padding
789 nr_objs
= (slab_size
- sizeof(struct slab
)) /
790 (buffer_size
+ sizeof(kmem_bufctl_t
));
793 * This calculated number will be either the right
794 * amount, or one greater than what we want.
796 if (slab_mgmt_size(nr_objs
, align
) + nr_objs
*buffer_size
800 if (nr_objs
> SLAB_LIMIT
)
801 nr_objs
= SLAB_LIMIT
;
803 mgmt_size
= slab_mgmt_size(nr_objs
, align
);
806 *left_over
= slab_size
- nr_objs
*buffer_size
- mgmt_size
;
809 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
811 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
814 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
815 function
, cachep
->name
, msg
);
821 * Special reaping functions for NUMA systems called from cache_reap().
822 * These take care of doing round robin flushing of alien caches (containing
823 * objects freed on different nodes from which they were allocated) and the
824 * flushing of remote pcps by calling drain_node_pages.
826 static DEFINE_PER_CPU(unsigned long, reap_node
);
828 static void init_reap_node(int cpu
)
832 node
= next_node(cpu_to_node(cpu
), node_online_map
);
833 if (node
== MAX_NUMNODES
)
834 node
= first_node(node_online_map
);
836 __get_cpu_var(reap_node
) = node
;
839 static void next_reap_node(void)
841 int node
= __get_cpu_var(reap_node
);
844 * Also drain per cpu pages on remote zones
846 if (node
!= numa_node_id())
847 drain_node_pages(node
);
849 node
= next_node(node
, node_online_map
);
850 if (unlikely(node
>= MAX_NUMNODES
))
851 node
= first_node(node_online_map
);
852 __get_cpu_var(reap_node
) = node
;
856 #define init_reap_node(cpu) do { } while (0)
857 #define next_reap_node(void) do { } while (0)
861 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
862 * via the workqueue/eventd.
863 * Add the CPU number into the expiration time to minimize the possibility of
864 * the CPUs getting into lockstep and contending for the global cache chain
867 static void __devinit
start_cpu_timer(int cpu
)
869 struct work_struct
*reap_work
= &per_cpu(reap_work
, cpu
);
872 * When this gets called from do_initcalls via cpucache_init(),
873 * init_workqueues() has already run, so keventd will be setup
876 if (keventd_up() && reap_work
->func
== NULL
) {
878 INIT_WORK(reap_work
, cache_reap
, NULL
);
879 schedule_delayed_work_on(cpu
, reap_work
, HZ
+ 3 * cpu
);
883 static struct array_cache
*alloc_arraycache(int node
, int entries
,
886 int memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
887 struct array_cache
*nc
= NULL
;
889 nc
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
893 nc
->batchcount
= batchcount
;
895 spin_lock_init(&nc
->lock
);
901 static void *__cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
902 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
904 static struct array_cache
**alloc_alien_cache(int node
, int limit
)
906 struct array_cache
**ac_ptr
;
907 int memsize
= sizeof(void *) * MAX_NUMNODES
;
912 ac_ptr
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
915 if (i
== node
|| !node_online(i
)) {
919 ac_ptr
[i
] = alloc_arraycache(node
, limit
, 0xbaadf00d);
921 for (i
--; i
<= 0; i
--)
931 static void free_alien_cache(struct array_cache
**ac_ptr
)
942 static void __drain_alien_cache(struct kmem_cache
*cachep
,
943 struct array_cache
*ac
, int node
)
945 struct kmem_list3
*rl3
= cachep
->nodelists
[node
];
948 spin_lock(&rl3
->list_lock
);
949 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
951 spin_unlock(&rl3
->list_lock
);
956 * Called from cache_reap() to regularly drain alien caches round robin.
958 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_list3
*l3
)
960 int node
= __get_cpu_var(reap_node
);
963 struct array_cache
*ac
= l3
->alien
[node
];
964 if (ac
&& ac
->avail
) {
965 spin_lock_irq(&ac
->lock
);
966 __drain_alien_cache(cachep
, ac
, node
);
967 spin_unlock_irq(&ac
->lock
);
972 static void drain_alien_cache(struct kmem_cache
*cachep
,
973 struct array_cache
**alien
)
976 struct array_cache
*ac
;
979 for_each_online_node(i
) {
982 spin_lock_irqsave(&ac
->lock
, flags
);
983 __drain_alien_cache(cachep
, ac
, i
);
984 spin_unlock_irqrestore(&ac
->lock
, flags
);
990 #define drain_alien_cache(cachep, alien) do { } while (0)
991 #define reap_alien(cachep, l3) do { } while (0)
993 static inline struct array_cache
**alloc_alien_cache(int node
, int limit
)
995 return (struct array_cache
**) 0x01020304ul
;
998 static inline void free_alien_cache(struct array_cache
**ac_ptr
)
1004 static int __devinit
cpuup_callback(struct notifier_block
*nfb
,
1005 unsigned long action
, void *hcpu
)
1007 long cpu
= (long)hcpu
;
1008 struct kmem_cache
*cachep
;
1009 struct kmem_list3
*l3
= NULL
;
1010 int node
= cpu_to_node(cpu
);
1011 int memsize
= sizeof(struct kmem_list3
);
1014 case CPU_UP_PREPARE
:
1015 mutex_lock(&cache_chain_mutex
);
1017 * We need to do this right in the beginning since
1018 * alloc_arraycache's are going to use this list.
1019 * kmalloc_node allows us to add the slab to the right
1020 * kmem_list3 and not this cpu's kmem_list3
1023 list_for_each_entry(cachep
, &cache_chain
, next
) {
1025 * Set up the size64 kmemlist for cpu before we can
1026 * begin anything. Make sure some other cpu on this
1027 * node has not already allocated this
1029 if (!cachep
->nodelists
[node
]) {
1030 l3
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1033 kmem_list3_init(l3
);
1034 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
1035 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1038 * The l3s don't come and go as CPUs come and
1039 * go. cache_chain_mutex is sufficient
1042 cachep
->nodelists
[node
] = l3
;
1045 spin_lock_irq(&cachep
->nodelists
[node
]->list_lock
);
1046 cachep
->nodelists
[node
]->free_limit
=
1047 (1 + nr_cpus_node(node
)) *
1048 cachep
->batchcount
+ cachep
->num
;
1049 spin_unlock_irq(&cachep
->nodelists
[node
]->list_lock
);
1053 * Now we can go ahead with allocating the shared arrays and
1056 list_for_each_entry(cachep
, &cache_chain
, next
) {
1057 struct array_cache
*nc
;
1058 struct array_cache
*shared
;
1059 struct array_cache
**alien
;
1061 nc
= alloc_arraycache(node
, cachep
->limit
,
1062 cachep
->batchcount
);
1065 shared
= alloc_arraycache(node
,
1066 cachep
->shared
* cachep
->batchcount
,
1071 alien
= alloc_alien_cache(node
, cachep
->limit
);
1074 cachep
->array
[cpu
] = nc
;
1075 l3
= cachep
->nodelists
[node
];
1078 spin_lock_irq(&l3
->list_lock
);
1081 * We are serialised from CPU_DEAD or
1082 * CPU_UP_CANCELLED by the cpucontrol lock
1084 l3
->shared
= shared
;
1093 spin_unlock_irq(&l3
->list_lock
);
1095 free_alien_cache(alien
);
1097 mutex_unlock(&cache_chain_mutex
);
1100 start_cpu_timer(cpu
);
1102 #ifdef CONFIG_HOTPLUG_CPU
1105 * Even if all the cpus of a node are down, we don't free the
1106 * kmem_list3 of any cache. This to avoid a race between
1107 * cpu_down, and a kmalloc allocation from another cpu for
1108 * memory from the node of the cpu going down. The list3
1109 * structure is usually allocated from kmem_cache_create() and
1110 * gets destroyed at kmem_cache_destroy().
1113 case CPU_UP_CANCELED
:
1114 mutex_lock(&cache_chain_mutex
);
1115 list_for_each_entry(cachep
, &cache_chain
, next
) {
1116 struct array_cache
*nc
;
1117 struct array_cache
*shared
;
1118 struct array_cache
**alien
;
1121 mask
= node_to_cpumask(node
);
1122 /* cpu is dead; no one can alloc from it. */
1123 nc
= cachep
->array
[cpu
];
1124 cachep
->array
[cpu
] = NULL
;
1125 l3
= cachep
->nodelists
[node
];
1128 goto free_array_cache
;
1130 spin_lock_irq(&l3
->list_lock
);
1132 /* Free limit for this kmem_list3 */
1133 l3
->free_limit
-= cachep
->batchcount
;
1135 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
1137 if (!cpus_empty(mask
)) {
1138 spin_unlock_irq(&l3
->list_lock
);
1139 goto free_array_cache
;
1142 shared
= l3
->shared
;
1144 free_block(cachep
, l3
->shared
->entry
,
1145 l3
->shared
->avail
, node
);
1152 spin_unlock_irq(&l3
->list_lock
);
1156 drain_alien_cache(cachep
, alien
);
1157 free_alien_cache(alien
);
1163 * In the previous loop, all the objects were freed to
1164 * the respective cache's slabs, now we can go ahead and
1165 * shrink each nodelist to its limit.
1167 list_for_each_entry(cachep
, &cache_chain
, next
) {
1168 l3
= cachep
->nodelists
[node
];
1171 spin_lock_irq(&l3
->list_lock
);
1172 /* free slabs belonging to this node */
1173 __node_shrink(cachep
, node
);
1174 spin_unlock_irq(&l3
->list_lock
);
1176 mutex_unlock(&cache_chain_mutex
);
1182 mutex_unlock(&cache_chain_mutex
);
1186 static struct notifier_block cpucache_notifier
= { &cpuup_callback
, NULL
, 0 };
1189 * swap the static kmem_list3 with kmalloced memory
1191 static void init_list(struct kmem_cache
*cachep
, struct kmem_list3
*list
,
1194 struct kmem_list3
*ptr
;
1196 BUG_ON(cachep
->nodelists
[nodeid
] != list
);
1197 ptr
= kmalloc_node(sizeof(struct kmem_list3
), GFP_KERNEL
, nodeid
);
1200 local_irq_disable();
1201 memcpy(ptr
, list
, sizeof(struct kmem_list3
));
1202 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1203 cachep
->nodelists
[nodeid
] = ptr
;
1208 * Initialisation. Called after the page allocator have been initialised and
1209 * before smp_init().
1211 void __init
kmem_cache_init(void)
1214 struct cache_sizes
*sizes
;
1215 struct cache_names
*names
;
1219 for (i
= 0; i
< NUM_INIT_LISTS
; i
++) {
1220 kmem_list3_init(&initkmem_list3
[i
]);
1221 if (i
< MAX_NUMNODES
)
1222 cache_cache
.nodelists
[i
] = NULL
;
1226 * Fragmentation resistance on low memory - only use bigger
1227 * page orders on machines with more than 32MB of memory.
1229 if (num_physpages
> (32 << 20) >> PAGE_SHIFT
)
1230 slab_break_gfp_order
= BREAK_GFP_ORDER_HI
;
1232 /* Bootstrap is tricky, because several objects are allocated
1233 * from caches that do not exist yet:
1234 * 1) initialize the cache_cache cache: it contains the struct
1235 * kmem_cache structures of all caches, except cache_cache itself:
1236 * cache_cache is statically allocated.
1237 * Initially an __init data area is used for the head array and the
1238 * kmem_list3 structures, it's replaced with a kmalloc allocated
1239 * array at the end of the bootstrap.
1240 * 2) Create the first kmalloc cache.
1241 * The struct kmem_cache for the new cache is allocated normally.
1242 * An __init data area is used for the head array.
1243 * 3) Create the remaining kmalloc caches, with minimally sized
1245 * 4) Replace the __init data head arrays for cache_cache and the first
1246 * kmalloc cache with kmalloc allocated arrays.
1247 * 5) Replace the __init data for kmem_list3 for cache_cache and
1248 * the other cache's with kmalloc allocated memory.
1249 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1252 /* 1) create the cache_cache */
1253 INIT_LIST_HEAD(&cache_chain
);
1254 list_add(&cache_cache
.next
, &cache_chain
);
1255 cache_cache
.colour_off
= cache_line_size();
1256 cache_cache
.array
[smp_processor_id()] = &initarray_cache
.cache
;
1257 cache_cache
.nodelists
[numa_node_id()] = &initkmem_list3
[CACHE_CACHE
];
1259 cache_cache
.buffer_size
= ALIGN(cache_cache
.buffer_size
,
1262 for (order
= 0; order
< MAX_ORDER
; order
++) {
1263 cache_estimate(order
, cache_cache
.buffer_size
,
1264 cache_line_size(), 0, &left_over
, &cache_cache
.num
);
1265 if (cache_cache
.num
)
1268 if (!cache_cache
.num
)
1270 cache_cache
.gfporder
= order
;
1271 cache_cache
.colour
= left_over
/ cache_cache
.colour_off
;
1272 cache_cache
.slab_size
= ALIGN(cache_cache
.num
* sizeof(kmem_bufctl_t
) +
1273 sizeof(struct slab
), cache_line_size());
1275 /* 2+3) create the kmalloc caches */
1276 sizes
= malloc_sizes
;
1277 names
= cache_names
;
1280 * Initialize the caches that provide memory for the array cache and the
1281 * kmem_list3 structures first. Without this, further allocations will
1285 sizes
[INDEX_AC
].cs_cachep
= kmem_cache_create(names
[INDEX_AC
].name
,
1286 sizes
[INDEX_AC
].cs_size
,
1287 ARCH_KMALLOC_MINALIGN
,
1288 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1291 if (INDEX_AC
!= INDEX_L3
) {
1292 sizes
[INDEX_L3
].cs_cachep
=
1293 kmem_cache_create(names
[INDEX_L3
].name
,
1294 sizes
[INDEX_L3
].cs_size
,
1295 ARCH_KMALLOC_MINALIGN
,
1296 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1300 while (sizes
->cs_size
!= ULONG_MAX
) {
1302 * For performance, all the general caches are L1 aligned.
1303 * This should be particularly beneficial on SMP boxes, as it
1304 * eliminates "false sharing".
1305 * Note for systems short on memory removing the alignment will
1306 * allow tighter packing of the smaller caches.
1308 if (!sizes
->cs_cachep
) {
1309 sizes
->cs_cachep
= kmem_cache_create(names
->name
,
1311 ARCH_KMALLOC_MINALIGN
,
1312 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1316 /* Inc off-slab bufctl limit until the ceiling is hit. */
1317 if (!(OFF_SLAB(sizes
->cs_cachep
))) {
1318 offslab_limit
= sizes
->cs_size
- sizeof(struct slab
);
1319 offslab_limit
/= sizeof(kmem_bufctl_t
);
1322 sizes
->cs_dmacachep
= kmem_cache_create(names
->name_dma
,
1324 ARCH_KMALLOC_MINALIGN
,
1325 ARCH_KMALLOC_FLAGS
|SLAB_CACHE_DMA
|
1331 /* 4) Replace the bootstrap head arrays */
1335 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1337 local_irq_disable();
1338 BUG_ON(cpu_cache_get(&cache_cache
) != &initarray_cache
.cache
);
1339 memcpy(ptr
, cpu_cache_get(&cache_cache
),
1340 sizeof(struct arraycache_init
));
1341 cache_cache
.array
[smp_processor_id()] = ptr
;
1344 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1346 local_irq_disable();
1347 BUG_ON(cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
)
1348 != &initarray_generic
.cache
);
1349 memcpy(ptr
, cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
),
1350 sizeof(struct arraycache_init
));
1351 malloc_sizes
[INDEX_AC
].cs_cachep
->array
[smp_processor_id()] =
1355 /* 5) Replace the bootstrap kmem_list3's */
1358 /* Replace the static kmem_list3 structures for the boot cpu */
1359 init_list(&cache_cache
, &initkmem_list3
[CACHE_CACHE
],
1362 for_each_online_node(node
) {
1363 init_list(malloc_sizes
[INDEX_AC
].cs_cachep
,
1364 &initkmem_list3
[SIZE_AC
+ node
], node
);
1366 if (INDEX_AC
!= INDEX_L3
) {
1367 init_list(malloc_sizes
[INDEX_L3
].cs_cachep
,
1368 &initkmem_list3
[SIZE_L3
+ node
],
1374 /* 6) resize the head arrays to their final sizes */
1376 struct kmem_cache
*cachep
;
1377 mutex_lock(&cache_chain_mutex
);
1378 list_for_each_entry(cachep
, &cache_chain
, next
)
1379 enable_cpucache(cachep
);
1380 mutex_unlock(&cache_chain_mutex
);
1384 g_cpucache_up
= FULL
;
1387 * Register a cpu startup notifier callback that initializes
1388 * cpu_cache_get for all new cpus
1390 register_cpu_notifier(&cpucache_notifier
);
1393 * The reap timers are started later, with a module init call: That part
1394 * of the kernel is not yet operational.
1398 static int __init
cpucache_init(void)
1403 * Register the timers that return unneeded pages to the page allocator
1405 for_each_online_cpu(cpu
)
1406 start_cpu_timer(cpu
);
1409 __initcall(cpucache_init
);
1412 * Interface to system's page allocator. No need to hold the cache-lock.
1414 * If we requested dmaable memory, we will get it. Even if we
1415 * did not request dmaable memory, we might get it, but that
1416 * would be relatively rare and ignorable.
1418 static void *kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
1424 flags
|= cachep
->gfpflags
;
1425 page
= alloc_pages_node(nodeid
, flags
, cachep
->gfporder
);
1428 addr
= page_address(page
);
1430 i
= (1 << cachep
->gfporder
);
1431 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1432 atomic_add(i
, &slab_reclaim_pages
);
1433 add_page_state(nr_slab
, i
);
1435 __SetPageSlab(page
);
1442 * Interface to system's page release.
1444 static void kmem_freepages(struct kmem_cache
*cachep
, void *addr
)
1446 unsigned long i
= (1 << cachep
->gfporder
);
1447 struct page
*page
= virt_to_page(addr
);
1448 const unsigned long nr_freed
= i
;
1451 BUG_ON(!PageSlab(page
));
1452 __ClearPageSlab(page
);
1455 sub_page_state(nr_slab
, nr_freed
);
1456 if (current
->reclaim_state
)
1457 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1458 free_pages((unsigned long)addr
, cachep
->gfporder
);
1459 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1460 atomic_sub(1 << cachep
->gfporder
, &slab_reclaim_pages
);
1463 static void kmem_rcu_free(struct rcu_head
*head
)
1465 struct slab_rcu
*slab_rcu
= (struct slab_rcu
*)head
;
1466 struct kmem_cache
*cachep
= slab_rcu
->cachep
;
1468 kmem_freepages(cachep
, slab_rcu
->addr
);
1469 if (OFF_SLAB(cachep
))
1470 kmem_cache_free(cachep
->slabp_cache
, slab_rcu
);
1475 #ifdef CONFIG_DEBUG_PAGEALLOC
1476 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1477 unsigned long caller
)
1479 int size
= obj_size(cachep
);
1481 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1483 if (size
< 5 * sizeof(unsigned long))
1486 *addr
++ = 0x12345678;
1488 *addr
++ = smp_processor_id();
1489 size
-= 3 * sizeof(unsigned long);
1491 unsigned long *sptr
= &caller
;
1492 unsigned long svalue
;
1494 while (!kstack_end(sptr
)) {
1496 if (kernel_text_address(svalue
)) {
1498 size
-= sizeof(unsigned long);
1499 if (size
<= sizeof(unsigned long))
1505 *addr
++ = 0x87654321;
1509 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1511 int size
= obj_size(cachep
);
1512 addr
= &((char *)addr
)[obj_offset(cachep
)];
1514 memset(addr
, val
, size
);
1515 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1518 static void dump_line(char *data
, int offset
, int limit
)
1521 printk(KERN_ERR
"%03x:", offset
);
1522 for (i
= 0; i
< limit
; i
++)
1523 printk(" %02x", (unsigned char)data
[offset
+ i
]);
1530 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1535 if (cachep
->flags
& SLAB_RED_ZONE
) {
1536 printk(KERN_ERR
"Redzone: 0x%lx/0x%lx.\n",
1537 *dbg_redzone1(cachep
, objp
),
1538 *dbg_redzone2(cachep
, objp
));
1541 if (cachep
->flags
& SLAB_STORE_USER
) {
1542 printk(KERN_ERR
"Last user: [<%p>]",
1543 *dbg_userword(cachep
, objp
));
1544 print_symbol("(%s)",
1545 (unsigned long)*dbg_userword(cachep
, objp
));
1548 realobj
= (char *)objp
+ obj_offset(cachep
);
1549 size
= obj_size(cachep
);
1550 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1553 if (i
+ limit
> size
)
1555 dump_line(realobj
, i
, limit
);
1559 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1565 realobj
= (char *)objp
+ obj_offset(cachep
);
1566 size
= obj_size(cachep
);
1568 for (i
= 0; i
< size
; i
++) {
1569 char exp
= POISON_FREE
;
1572 if (realobj
[i
] != exp
) {
1578 "Slab corruption: start=%p, len=%d\n",
1580 print_objinfo(cachep
, objp
, 0);
1582 /* Hexdump the affected line */
1585 if (i
+ limit
> size
)
1587 dump_line(realobj
, i
, limit
);
1590 /* Limit to 5 lines */
1596 /* Print some data about the neighboring objects, if they
1599 struct slab
*slabp
= virt_to_slab(objp
);
1602 objnr
= obj_to_index(cachep
, slabp
, objp
);
1604 objp
= index_to_obj(cachep
, slabp
, objnr
- 1);
1605 realobj
= (char *)objp
+ obj_offset(cachep
);
1606 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
1608 print_objinfo(cachep
, objp
, 2);
1610 if (objnr
+ 1 < cachep
->num
) {
1611 objp
= index_to_obj(cachep
, slabp
, objnr
+ 1);
1612 realobj
= (char *)objp
+ obj_offset(cachep
);
1613 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
1615 print_objinfo(cachep
, objp
, 2);
1623 * slab_destroy_objs - destroy a slab and its objects
1624 * @cachep: cache pointer being destroyed
1625 * @slabp: slab pointer being destroyed
1627 * Call the registered destructor for each object in a slab that is being
1630 static void slab_destroy_objs(struct kmem_cache
*cachep
, struct slab
*slabp
)
1633 for (i
= 0; i
< cachep
->num
; i
++) {
1634 void *objp
= index_to_obj(cachep
, slabp
, i
);
1636 if (cachep
->flags
& SLAB_POISON
) {
1637 #ifdef CONFIG_DEBUG_PAGEALLOC
1638 if (cachep
->buffer_size
% PAGE_SIZE
== 0 &&
1640 kernel_map_pages(virt_to_page(objp
),
1641 cachep
->buffer_size
/ PAGE_SIZE
, 1);
1643 check_poison_obj(cachep
, objp
);
1645 check_poison_obj(cachep
, objp
);
1648 if (cachep
->flags
& SLAB_RED_ZONE
) {
1649 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1650 slab_error(cachep
, "start of a freed object "
1652 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1653 slab_error(cachep
, "end of a freed object "
1656 if (cachep
->dtor
&& !(cachep
->flags
& SLAB_POISON
))
1657 (cachep
->dtor
) (objp
+ obj_offset(cachep
), cachep
, 0);
1661 static void slab_destroy_objs(struct kmem_cache
*cachep
, struct slab
*slabp
)
1665 for (i
= 0; i
< cachep
->num
; i
++) {
1666 void *objp
= index_to_obj(cachep
, slabp
, i
);
1667 (cachep
->dtor
) (objp
, cachep
, 0);
1674 * slab_destroy - destroy and release all objects in a slab
1675 * @cachep: cache pointer being destroyed
1676 * @slabp: slab pointer being destroyed
1678 * Destroy all the objs in a slab, and release the mem back to the system.
1679 * Before calling the slab must have been unlinked from the cache. The
1680 * cache-lock is not held/needed.
1682 static void slab_destroy(struct kmem_cache
*cachep
, struct slab
*slabp
)
1684 void *addr
= slabp
->s_mem
- slabp
->colouroff
;
1686 slab_destroy_objs(cachep
, slabp
);
1687 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
1688 struct slab_rcu
*slab_rcu
;
1690 slab_rcu
= (struct slab_rcu
*)slabp
;
1691 slab_rcu
->cachep
= cachep
;
1692 slab_rcu
->addr
= addr
;
1693 call_rcu(&slab_rcu
->head
, kmem_rcu_free
);
1695 kmem_freepages(cachep
, addr
);
1696 if (OFF_SLAB(cachep
))
1697 kmem_cache_free(cachep
->slabp_cache
, slabp
);
1702 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1703 * size of kmem_list3.
1705 static void set_up_list3s(struct kmem_cache
*cachep
, int index
)
1709 for_each_online_node(node
) {
1710 cachep
->nodelists
[node
] = &initkmem_list3
[index
+ node
];
1711 cachep
->nodelists
[node
]->next_reap
= jiffies
+
1713 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1718 * calculate_slab_order - calculate size (page order) of slabs
1719 * @cachep: pointer to the cache that is being created
1720 * @size: size of objects to be created in this cache.
1721 * @align: required alignment for the objects.
1722 * @flags: slab allocation flags
1724 * Also calculates the number of objects per slab.
1726 * This could be made much more intelligent. For now, try to avoid using
1727 * high order pages for slabs. When the gfp() functions are more friendly
1728 * towards high-order requests, this should be changed.
1730 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
1731 size_t size
, size_t align
, unsigned long flags
)
1733 size_t left_over
= 0;
1736 for (gfporder
= 0; gfporder
<= MAX_GFP_ORDER
; gfporder
++) {
1740 cache_estimate(gfporder
, size
, align
, flags
, &remainder
, &num
);
1744 /* More than offslab_limit objects will cause problems */
1745 if ((flags
& CFLGS_OFF_SLAB
) && num
> offslab_limit
)
1748 /* Found something acceptable - save it away */
1750 cachep
->gfporder
= gfporder
;
1751 left_over
= remainder
;
1754 * A VFS-reclaimable slab tends to have most allocations
1755 * as GFP_NOFS and we really don't want to have to be allocating
1756 * higher-order pages when we are unable to shrink dcache.
1758 if (flags
& SLAB_RECLAIM_ACCOUNT
)
1762 * Large number of objects is good, but very large slabs are
1763 * currently bad for the gfp()s.
1765 if (gfporder
>= slab_break_gfp_order
)
1769 * Acceptable internal fragmentation?
1771 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
1777 static void setup_cpu_cache(struct kmem_cache
*cachep
)
1779 if (g_cpucache_up
== FULL
) {
1780 enable_cpucache(cachep
);
1783 if (g_cpucache_up
== NONE
) {
1785 * Note: the first kmem_cache_create must create the cache
1786 * that's used by kmalloc(24), otherwise the creation of
1787 * further caches will BUG().
1789 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
1792 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
1793 * the first cache, then we need to set up all its list3s,
1794 * otherwise the creation of further caches will BUG().
1796 set_up_list3s(cachep
, SIZE_AC
);
1797 if (INDEX_AC
== INDEX_L3
)
1798 g_cpucache_up
= PARTIAL_L3
;
1800 g_cpucache_up
= PARTIAL_AC
;
1802 cachep
->array
[smp_processor_id()] =
1803 kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1805 if (g_cpucache_up
== PARTIAL_AC
) {
1806 set_up_list3s(cachep
, SIZE_L3
);
1807 g_cpucache_up
= PARTIAL_L3
;
1810 for_each_online_node(node
) {
1811 cachep
->nodelists
[node
] =
1812 kmalloc_node(sizeof(struct kmem_list3
),
1814 BUG_ON(!cachep
->nodelists
[node
]);
1815 kmem_list3_init(cachep
->nodelists
[node
]);
1819 cachep
->nodelists
[numa_node_id()]->next_reap
=
1820 jiffies
+ REAPTIMEOUT_LIST3
+
1821 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1823 cpu_cache_get(cachep
)->avail
= 0;
1824 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
1825 cpu_cache_get(cachep
)->batchcount
= 1;
1826 cpu_cache_get(cachep
)->touched
= 0;
1827 cachep
->batchcount
= 1;
1828 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
1832 * kmem_cache_create - Create a cache.
1833 * @name: A string which is used in /proc/slabinfo to identify this cache.
1834 * @size: The size of objects to be created in this cache.
1835 * @align: The required alignment for the objects.
1836 * @flags: SLAB flags
1837 * @ctor: A constructor for the objects.
1838 * @dtor: A destructor for the objects.
1840 * Returns a ptr to the cache on success, NULL on failure.
1841 * Cannot be called within a int, but can be interrupted.
1842 * The @ctor is run when new pages are allocated by the cache
1843 * and the @dtor is run before the pages are handed back.
1845 * @name must be valid until the cache is destroyed. This implies that
1846 * the module calling this has to destroy the cache before getting unloaded.
1850 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1851 * to catch references to uninitialised memory.
1853 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1854 * for buffer overruns.
1856 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1857 * cacheline. This can be beneficial if you're counting cycles as closely
1861 kmem_cache_create (const char *name
, size_t size
, size_t align
,
1862 unsigned long flags
,
1863 void (*ctor
)(void*, struct kmem_cache
*, unsigned long),
1864 void (*dtor
)(void*, struct kmem_cache
*, unsigned long))
1866 size_t left_over
, slab_size
, ralign
;
1867 struct kmem_cache
*cachep
= NULL
;
1868 struct list_head
*p
;
1871 * Sanity checks... these are all serious usage bugs.
1873 if (!name
|| in_interrupt() || (size
< BYTES_PER_WORD
) ||
1874 (size
> (1 << MAX_OBJ_ORDER
) * PAGE_SIZE
) || (dtor
&& !ctor
)) {
1875 printk(KERN_ERR
"%s: Early error in slab %s\n", __FUNCTION__
,
1881 * Prevent CPUs from coming and going.
1882 * lock_cpu_hotplug() nests outside cache_chain_mutex
1886 mutex_lock(&cache_chain_mutex
);
1888 list_for_each(p
, &cache_chain
) {
1889 struct kmem_cache
*pc
= list_entry(p
, struct kmem_cache
, next
);
1890 mm_segment_t old_fs
= get_fs();
1895 * This happens when the module gets unloaded and doesn't
1896 * destroy its slab cache and no-one else reuses the vmalloc
1897 * area of the module. Print a warning.
1900 res
= __get_user(tmp
, pc
->name
);
1903 printk("SLAB: cache with size %d has lost its name\n",
1908 if (!strcmp(pc
->name
, name
)) {
1909 printk("kmem_cache_create: duplicate cache %s\n", name
);
1916 WARN_ON(strchr(name
, ' ')); /* It confuses parsers */
1917 if ((flags
& SLAB_DEBUG_INITIAL
) && !ctor
) {
1918 /* No constructor, but inital state check requested */
1919 printk(KERN_ERR
"%s: No con, but init state check "
1920 "requested - %s\n", __FUNCTION__
, name
);
1921 flags
&= ~SLAB_DEBUG_INITIAL
;
1925 * Enable redzoning and last user accounting, except for caches with
1926 * large objects, if the increased size would increase the object size
1927 * above the next power of two: caches with object sizes just above a
1928 * power of two have a significant amount of internal fragmentation.
1930 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + 3 * BYTES_PER_WORD
))
1931 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
1932 if (!(flags
& SLAB_DESTROY_BY_RCU
))
1933 flags
|= SLAB_POISON
;
1935 if (flags
& SLAB_DESTROY_BY_RCU
)
1936 BUG_ON(flags
& SLAB_POISON
);
1938 if (flags
& SLAB_DESTROY_BY_RCU
)
1942 * Always checks flags, a caller might be expecting debug support which
1945 if (flags
& ~CREATE_MASK
)
1949 * Check that size is in terms of words. This is needed to avoid
1950 * unaligned accesses for some archs when redzoning is used, and makes
1951 * sure any on-slab bufctl's are also correctly aligned.
1953 if (size
& (BYTES_PER_WORD
- 1)) {
1954 size
+= (BYTES_PER_WORD
- 1);
1955 size
&= ~(BYTES_PER_WORD
- 1);
1958 /* calculate the final buffer alignment: */
1960 /* 1) arch recommendation: can be overridden for debug */
1961 if (flags
& SLAB_HWCACHE_ALIGN
) {
1963 * Default alignment: as specified by the arch code. Except if
1964 * an object is really small, then squeeze multiple objects into
1967 ralign
= cache_line_size();
1968 while (size
<= ralign
/ 2)
1971 ralign
= BYTES_PER_WORD
;
1973 /* 2) arch mandated alignment: disables debug if necessary */
1974 if (ralign
< ARCH_SLAB_MINALIGN
) {
1975 ralign
= ARCH_SLAB_MINALIGN
;
1976 if (ralign
> BYTES_PER_WORD
)
1977 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
1979 /* 3) caller mandated alignment: disables debug if necessary */
1980 if (ralign
< align
) {
1982 if (ralign
> BYTES_PER_WORD
)
1983 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
1986 * 4) Store it. Note that the debug code below can reduce
1987 * the alignment to BYTES_PER_WORD.
1991 /* Get cache's description obj. */
1992 cachep
= kmem_cache_alloc(&cache_cache
, SLAB_KERNEL
);
1995 memset(cachep
, 0, sizeof(struct kmem_cache
));
1998 cachep
->obj_size
= size
;
2000 if (flags
& SLAB_RED_ZONE
) {
2001 /* redzoning only works with word aligned caches */
2002 align
= BYTES_PER_WORD
;
2004 /* add space for red zone words */
2005 cachep
->obj_offset
+= BYTES_PER_WORD
;
2006 size
+= 2 * BYTES_PER_WORD
;
2008 if (flags
& SLAB_STORE_USER
) {
2009 /* user store requires word alignment and
2010 * one word storage behind the end of the real
2013 align
= BYTES_PER_WORD
;
2014 size
+= BYTES_PER_WORD
;
2016 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2017 if (size
>= malloc_sizes
[INDEX_L3
+ 1].cs_size
2018 && cachep
->obj_size
> cache_line_size() && size
< PAGE_SIZE
) {
2019 cachep
->obj_offset
+= PAGE_SIZE
- size
;
2025 /* Determine if the slab management is 'on' or 'off' slab. */
2026 if (size
>= (PAGE_SIZE
>> 3))
2028 * Size is large, assume best to place the slab management obj
2029 * off-slab (should allow better packing of objs).
2031 flags
|= CFLGS_OFF_SLAB
;
2033 size
= ALIGN(size
, align
);
2035 left_over
= calculate_slab_order(cachep
, size
, align
, flags
);
2038 printk("kmem_cache_create: couldn't create cache %s.\n", name
);
2039 kmem_cache_free(&cache_cache
, cachep
);
2043 slab_size
= ALIGN(cachep
->num
* sizeof(kmem_bufctl_t
)
2044 + sizeof(struct slab
), align
);
2047 * If the slab has been placed off-slab, and we have enough space then
2048 * move it on-slab. This is at the expense of any extra colouring.
2050 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= slab_size
) {
2051 flags
&= ~CFLGS_OFF_SLAB
;
2052 left_over
-= slab_size
;
2055 if (flags
& CFLGS_OFF_SLAB
) {
2056 /* really off slab. No need for manual alignment */
2058 cachep
->num
* sizeof(kmem_bufctl_t
) + sizeof(struct slab
);
2061 cachep
->colour_off
= cache_line_size();
2062 /* Offset must be a multiple of the alignment. */
2063 if (cachep
->colour_off
< align
)
2064 cachep
->colour_off
= align
;
2065 cachep
->colour
= left_over
/ cachep
->colour_off
;
2066 cachep
->slab_size
= slab_size
;
2067 cachep
->flags
= flags
;
2068 cachep
->gfpflags
= 0;
2069 if (flags
& SLAB_CACHE_DMA
)
2070 cachep
->gfpflags
|= GFP_DMA
;
2071 cachep
->buffer_size
= size
;
2073 if (flags
& CFLGS_OFF_SLAB
)
2074 cachep
->slabp_cache
= kmem_find_general_cachep(slab_size
, 0u);
2075 cachep
->ctor
= ctor
;
2076 cachep
->dtor
= dtor
;
2077 cachep
->name
= name
;
2080 setup_cpu_cache(cachep
);
2082 /* cache setup completed, link it into the list */
2083 list_add(&cachep
->next
, &cache_chain
);
2085 if (!cachep
&& (flags
& SLAB_PANIC
))
2086 panic("kmem_cache_create(): failed to create slab `%s'\n",
2088 mutex_unlock(&cache_chain_mutex
);
2089 unlock_cpu_hotplug();
2092 EXPORT_SYMBOL(kmem_cache_create
);
2095 static void check_irq_off(void)
2097 BUG_ON(!irqs_disabled());
2100 static void check_irq_on(void)
2102 BUG_ON(irqs_disabled());
2105 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2109 assert_spin_locked(&cachep
->nodelists
[numa_node_id()]->list_lock
);
2113 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2117 assert_spin_locked(&cachep
->nodelists
[node
]->list_lock
);
2122 #define check_irq_off() do { } while(0)
2123 #define check_irq_on() do { } while(0)
2124 #define check_spinlock_acquired(x) do { } while(0)
2125 #define check_spinlock_acquired_node(x, y) do { } while(0)
2128 static void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
2129 struct array_cache
*ac
,
2130 int force
, int node
);
2132 static void do_drain(void *arg
)
2134 struct kmem_cache
*cachep
= arg
;
2135 struct array_cache
*ac
;
2136 int node
= numa_node_id();
2139 ac
= cpu_cache_get(cachep
);
2140 spin_lock(&cachep
->nodelists
[node
]->list_lock
);
2141 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
2142 spin_unlock(&cachep
->nodelists
[node
]->list_lock
);
2146 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2148 struct kmem_list3
*l3
;
2151 on_each_cpu(do_drain
, cachep
, 1, 1);
2153 for_each_online_node(node
) {
2154 l3
= cachep
->nodelists
[node
];
2156 drain_array(cachep
, l3
, l3
->shared
, 1, node
);
2158 drain_alien_cache(cachep
, l3
->alien
);
2163 static int __node_shrink(struct kmem_cache
*cachep
, int node
)
2166 struct kmem_list3
*l3
= cachep
->nodelists
[node
];
2170 struct list_head
*p
;
2172 p
= l3
->slabs_free
.prev
;
2173 if (p
== &l3
->slabs_free
)
2176 slabp
= list_entry(l3
->slabs_free
.prev
, struct slab
, list
);
2181 list_del(&slabp
->list
);
2183 l3
->free_objects
-= cachep
->num
;
2184 spin_unlock_irq(&l3
->list_lock
);
2185 slab_destroy(cachep
, slabp
);
2186 spin_lock_irq(&l3
->list_lock
);
2188 ret
= !list_empty(&l3
->slabs_full
) || !list_empty(&l3
->slabs_partial
);
2192 static int __cache_shrink(struct kmem_cache
*cachep
)
2195 struct kmem_list3
*l3
;
2197 drain_cpu_caches(cachep
);
2200 for_each_online_node(i
) {
2201 l3
= cachep
->nodelists
[i
];
2203 spin_lock_irq(&l3
->list_lock
);
2204 ret
+= __node_shrink(cachep
, i
);
2205 spin_unlock_irq(&l3
->list_lock
);
2208 return (ret
? 1 : 0);
2212 * kmem_cache_shrink - Shrink a cache.
2213 * @cachep: The cache to shrink.
2215 * Releases as many slabs as possible for a cache.
2216 * To help debugging, a zero exit status indicates all slabs were released.
2218 int kmem_cache_shrink(struct kmem_cache
*cachep
)
2220 if (!cachep
|| in_interrupt())
2223 return __cache_shrink(cachep
);
2225 EXPORT_SYMBOL(kmem_cache_shrink
);
2228 * kmem_cache_destroy - delete a cache
2229 * @cachep: the cache to destroy
2231 * Remove a struct kmem_cache object from the slab cache.
2232 * Returns 0 on success.
2234 * It is expected this function will be called by a module when it is
2235 * unloaded. This will remove the cache completely, and avoid a duplicate
2236 * cache being allocated each time a module is loaded and unloaded, if the
2237 * module doesn't have persistent in-kernel storage across loads and unloads.
2239 * The cache must be empty before calling this function.
2241 * The caller must guarantee that noone will allocate memory from the cache
2242 * during the kmem_cache_destroy().
2244 int kmem_cache_destroy(struct kmem_cache
*cachep
)
2247 struct kmem_list3
*l3
;
2249 if (!cachep
|| in_interrupt())
2252 /* Don't let CPUs to come and go */
2255 /* Find the cache in the chain of caches. */
2256 mutex_lock(&cache_chain_mutex
);
2258 * the chain is never empty, cache_cache is never destroyed
2260 list_del(&cachep
->next
);
2261 mutex_unlock(&cache_chain_mutex
);
2263 if (__cache_shrink(cachep
)) {
2264 slab_error(cachep
, "Can't free all objects");
2265 mutex_lock(&cache_chain_mutex
);
2266 list_add(&cachep
->next
, &cache_chain
);
2267 mutex_unlock(&cache_chain_mutex
);
2268 unlock_cpu_hotplug();
2272 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
2275 for_each_online_cpu(i
)
2276 kfree(cachep
->array
[i
]);
2278 /* NUMA: free the list3 structures */
2279 for_each_online_node(i
) {
2280 l3
= cachep
->nodelists
[i
];
2283 free_alien_cache(l3
->alien
);
2287 kmem_cache_free(&cache_cache
, cachep
);
2288 unlock_cpu_hotplug();
2291 EXPORT_SYMBOL(kmem_cache_destroy
);
2293 /* Get the memory for a slab management obj. */
2294 static struct slab
*alloc_slabmgmt(struct kmem_cache
*cachep
, void *objp
,
2295 int colour_off
, gfp_t local_flags
)
2299 if (OFF_SLAB(cachep
)) {
2300 /* Slab management obj is off-slab. */
2301 slabp
= kmem_cache_alloc(cachep
->slabp_cache
, local_flags
);
2305 slabp
= objp
+ colour_off
;
2306 colour_off
+= cachep
->slab_size
;
2309 slabp
->colouroff
= colour_off
;
2310 slabp
->s_mem
= objp
+ colour_off
;
2314 static inline kmem_bufctl_t
*slab_bufctl(struct slab
*slabp
)
2316 return (kmem_bufctl_t
*) (slabp
+ 1);
2319 static void cache_init_objs(struct kmem_cache
*cachep
,
2320 struct slab
*slabp
, unsigned long ctor_flags
)
2324 for (i
= 0; i
< cachep
->num
; i
++) {
2325 void *objp
= index_to_obj(cachep
, slabp
, i
);
2327 /* need to poison the objs? */
2328 if (cachep
->flags
& SLAB_POISON
)
2329 poison_obj(cachep
, objp
, POISON_FREE
);
2330 if (cachep
->flags
& SLAB_STORE_USER
)
2331 *dbg_userword(cachep
, objp
) = NULL
;
2333 if (cachep
->flags
& SLAB_RED_ZONE
) {
2334 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2335 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2338 * Constructors are not allowed to allocate memory from the same
2339 * cache which they are a constructor for. Otherwise, deadlock.
2340 * They must also be threaded.
2342 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2343 cachep
->ctor(objp
+ obj_offset(cachep
), cachep
,
2346 if (cachep
->flags
& SLAB_RED_ZONE
) {
2347 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2348 slab_error(cachep
, "constructor overwrote the"
2349 " end of an object");
2350 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2351 slab_error(cachep
, "constructor overwrote the"
2352 " start of an object");
2354 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 &&
2355 OFF_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
)
2356 kernel_map_pages(virt_to_page(objp
),
2357 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2360 cachep
->ctor(objp
, cachep
, ctor_flags
);
2362 slab_bufctl(slabp
)[i
] = i
+ 1;
2364 slab_bufctl(slabp
)[i
- 1] = BUFCTL_END
;
2368 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2370 if (flags
& SLAB_DMA
)
2371 BUG_ON(!(cachep
->gfpflags
& GFP_DMA
));
2373 BUG_ON(cachep
->gfpflags
& GFP_DMA
);
2376 static void *slab_get_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2379 void *objp
= index_to_obj(cachep
, slabp
, slabp
->free
);
2383 next
= slab_bufctl(slabp
)[slabp
->free
];
2385 slab_bufctl(slabp
)[slabp
->free
] = BUFCTL_FREE
;
2386 WARN_ON(slabp
->nodeid
!= nodeid
);
2393 static void slab_put_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2394 void *objp
, int nodeid
)
2396 unsigned int objnr
= obj_to_index(cachep
, slabp
, objp
);
2399 /* Verify that the slab belongs to the intended node */
2400 WARN_ON(slabp
->nodeid
!= nodeid
);
2402 if (slab_bufctl(slabp
)[objnr
] != BUFCTL_FREE
) {
2403 printk(KERN_ERR
"slab: double free detected in cache "
2404 "'%s', objp %p\n", cachep
->name
, objp
);
2408 slab_bufctl(slabp
)[objnr
] = slabp
->free
;
2409 slabp
->free
= objnr
;
2413 static void set_slab_attr(struct kmem_cache
*cachep
, struct slab
*slabp
,
2419 /* Nasty!!!!!! I hope this is OK. */
2420 page
= virt_to_page(objp
);
2423 if (likely(!PageCompound(page
)))
2424 i
<<= cachep
->gfporder
;
2426 page_set_cache(page
, cachep
);
2427 page_set_slab(page
, slabp
);
2433 * Grow (by 1) the number of slabs within a cache. This is called by
2434 * kmem_cache_alloc() when there are no active objs left in a cache.
2436 static int cache_grow(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
2442 unsigned long ctor_flags
;
2443 struct kmem_list3
*l3
;
2446 * Be lazy and only check for valid flags here, keeping it out of the
2447 * critical path in kmem_cache_alloc().
2449 if (flags
& ~(SLAB_DMA
| SLAB_LEVEL_MASK
| SLAB_NO_GROW
))
2451 if (flags
& SLAB_NO_GROW
)
2454 ctor_flags
= SLAB_CTOR_CONSTRUCTOR
;
2455 local_flags
= (flags
& SLAB_LEVEL_MASK
);
2456 if (!(local_flags
& __GFP_WAIT
))
2458 * Not allowed to sleep. Need to tell a constructor about
2459 * this - it might need to know...
2461 ctor_flags
|= SLAB_CTOR_ATOMIC
;
2463 /* Take the l3 list lock to change the colour_next on this node */
2465 l3
= cachep
->nodelists
[nodeid
];
2466 spin_lock(&l3
->list_lock
);
2468 /* Get colour for the slab, and cal the next value. */
2469 offset
= l3
->colour_next
;
2471 if (l3
->colour_next
>= cachep
->colour
)
2472 l3
->colour_next
= 0;
2473 spin_unlock(&l3
->list_lock
);
2475 offset
*= cachep
->colour_off
;
2477 if (local_flags
& __GFP_WAIT
)
2481 * The test for missing atomic flag is performed here, rather than
2482 * the more obvious place, simply to reduce the critical path length
2483 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2484 * will eventually be caught here (where it matters).
2486 kmem_flagcheck(cachep
, flags
);
2489 * Get mem for the objs. Attempt to allocate a physical page from
2492 objp
= kmem_getpages(cachep
, flags
, nodeid
);
2496 /* Get slab management. */
2497 slabp
= alloc_slabmgmt(cachep
, objp
, offset
, local_flags
);
2501 slabp
->nodeid
= nodeid
;
2502 set_slab_attr(cachep
, slabp
, objp
);
2504 cache_init_objs(cachep
, slabp
, ctor_flags
);
2506 if (local_flags
& __GFP_WAIT
)
2507 local_irq_disable();
2509 spin_lock(&l3
->list_lock
);
2511 /* Make slab active. */
2512 list_add_tail(&slabp
->list
, &(l3
->slabs_free
));
2513 STATS_INC_GROWN(cachep
);
2514 l3
->free_objects
+= cachep
->num
;
2515 spin_unlock(&l3
->list_lock
);
2518 kmem_freepages(cachep
, objp
);
2520 if (local_flags
& __GFP_WAIT
)
2521 local_irq_disable();
2528 * Perform extra freeing checks:
2529 * - detect bad pointers.
2530 * - POISON/RED_ZONE checking
2531 * - destructor calls, for caches with POISON+dtor
2533 static void kfree_debugcheck(const void *objp
)
2537 if (!virt_addr_valid(objp
)) {
2538 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
2539 (unsigned long)objp
);
2542 page
= virt_to_page(objp
);
2543 if (!PageSlab(page
)) {
2544 printk(KERN_ERR
"kfree_debugcheck: bad ptr %lxh.\n",
2545 (unsigned long)objp
);
2550 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2557 objp
-= obj_offset(cachep
);
2558 kfree_debugcheck(objp
);
2559 page
= virt_to_page(objp
);
2561 if (page_get_cache(page
) != cachep
) {
2562 printk(KERN_ERR
"mismatch in kmem_cache_free: expected "
2563 "cache %p, got %p\n",
2564 page_get_cache(page
), cachep
);
2565 printk(KERN_ERR
"%p is %s.\n", cachep
, cachep
->name
);
2566 printk(KERN_ERR
"%p is %s.\n", page_get_cache(page
),
2567 page_get_cache(page
)->name
);
2570 slabp
= page_get_slab(page
);
2572 if (cachep
->flags
& SLAB_RED_ZONE
) {
2573 if (*dbg_redzone1(cachep
, objp
) != RED_ACTIVE
||
2574 *dbg_redzone2(cachep
, objp
) != RED_ACTIVE
) {
2575 slab_error(cachep
, "double free, or memory outside"
2576 " object was overwritten");
2577 printk(KERN_ERR
"%p: redzone 1:0x%lx, "
2578 "redzone 2:0x%lx.\n",
2579 objp
, *dbg_redzone1(cachep
, objp
),
2580 *dbg_redzone2(cachep
, objp
));
2582 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2583 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2585 if (cachep
->flags
& SLAB_STORE_USER
)
2586 *dbg_userword(cachep
, objp
) = caller
;
2588 objnr
= obj_to_index(cachep
, slabp
, objp
);
2590 BUG_ON(objnr
>= cachep
->num
);
2591 BUG_ON(objp
!= index_to_obj(cachep
, slabp
, objnr
));
2593 if (cachep
->flags
& SLAB_DEBUG_INITIAL
) {
2595 * Need to call the slab's constructor so the caller can
2596 * perform a verify of its state (debugging). Called without
2597 * the cache-lock held.
2599 cachep
->ctor(objp
+ obj_offset(cachep
),
2600 cachep
, SLAB_CTOR_CONSTRUCTOR
| SLAB_CTOR_VERIFY
);
2602 if (cachep
->flags
& SLAB_POISON
&& cachep
->dtor
) {
2603 /* we want to cache poison the object,
2604 * call the destruction callback
2606 cachep
->dtor(objp
+ obj_offset(cachep
), cachep
, 0);
2608 if (cachep
->flags
& SLAB_POISON
) {
2609 #ifdef CONFIG_DEBUG_PAGEALLOC
2610 if ((cachep
->buffer_size
% PAGE_SIZE
)==0 && OFF_SLAB(cachep
)) {
2611 store_stackinfo(cachep
, objp
, (unsigned long)caller
);
2612 kernel_map_pages(virt_to_page(objp
),
2613 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2615 poison_obj(cachep
, objp
, POISON_FREE
);
2618 poison_obj(cachep
, objp
, POISON_FREE
);
2624 static void check_slabp(struct kmem_cache
*cachep
, struct slab
*slabp
)
2629 /* Check slab's freelist to see if this obj is there. */
2630 for (i
= slabp
->free
; i
!= BUFCTL_END
; i
= slab_bufctl(slabp
)[i
]) {
2632 if (entries
> cachep
->num
|| i
>= cachep
->num
)
2635 if (entries
!= cachep
->num
- slabp
->inuse
) {
2637 printk(KERN_ERR
"slab: Internal list corruption detected in "
2638 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2639 cachep
->name
, cachep
->num
, slabp
, slabp
->inuse
);
2641 i
< sizeof(*slabp
) + cachep
->num
* sizeof(kmem_bufctl_t
);
2644 printk("\n%03x:", i
);
2645 printk(" %02x", ((unsigned char *)slabp
)[i
]);
2652 #define kfree_debugcheck(x) do { } while(0)
2653 #define cache_free_debugcheck(x,objp,z) (objp)
2654 #define check_slabp(x,y) do { } while(0)
2657 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
2660 struct kmem_list3
*l3
;
2661 struct array_cache
*ac
;
2664 ac
= cpu_cache_get(cachep
);
2666 batchcount
= ac
->batchcount
;
2667 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2669 * If there was little recent activity on this cache, then
2670 * perform only a partial refill. Otherwise we could generate
2673 batchcount
= BATCHREFILL_LIMIT
;
2675 l3
= cachep
->nodelists
[numa_node_id()];
2677 BUG_ON(ac
->avail
> 0 || !l3
);
2678 spin_lock(&l3
->list_lock
);
2681 struct array_cache
*shared_array
= l3
->shared
;
2682 if (shared_array
->avail
) {
2683 if (batchcount
> shared_array
->avail
)
2684 batchcount
= shared_array
->avail
;
2685 shared_array
->avail
-= batchcount
;
2686 ac
->avail
= batchcount
;
2688 &(shared_array
->entry
[shared_array
->avail
]),
2689 sizeof(void *) * batchcount
);
2690 shared_array
->touched
= 1;
2694 while (batchcount
> 0) {
2695 struct list_head
*entry
;
2697 /* Get slab alloc is to come from. */
2698 entry
= l3
->slabs_partial
.next
;
2699 if (entry
== &l3
->slabs_partial
) {
2700 l3
->free_touched
= 1;
2701 entry
= l3
->slabs_free
.next
;
2702 if (entry
== &l3
->slabs_free
)
2706 slabp
= list_entry(entry
, struct slab
, list
);
2707 check_slabp(cachep
, slabp
);
2708 check_spinlock_acquired(cachep
);
2709 while (slabp
->inuse
< cachep
->num
&& batchcount
--) {
2710 STATS_INC_ALLOCED(cachep
);
2711 STATS_INC_ACTIVE(cachep
);
2712 STATS_SET_HIGH(cachep
);
2714 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, slabp
,
2717 check_slabp(cachep
, slabp
);
2719 /* move slabp to correct slabp list: */
2720 list_del(&slabp
->list
);
2721 if (slabp
->free
== BUFCTL_END
)
2722 list_add(&slabp
->list
, &l3
->slabs_full
);
2724 list_add(&slabp
->list
, &l3
->slabs_partial
);
2728 l3
->free_objects
-= ac
->avail
;
2730 spin_unlock(&l3
->list_lock
);
2732 if (unlikely(!ac
->avail
)) {
2734 x
= cache_grow(cachep
, flags
, numa_node_id());
2736 /* cache_grow can reenable interrupts, then ac could change. */
2737 ac
= cpu_cache_get(cachep
);
2738 if (!x
&& ac
->avail
== 0) /* no objects in sight? abort */
2741 if (!ac
->avail
) /* objects refilled by interrupt? */
2745 return ac
->entry
[--ac
->avail
];
2748 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
2751 might_sleep_if(flags
& __GFP_WAIT
);
2753 kmem_flagcheck(cachep
, flags
);
2758 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
2759 gfp_t flags
, void *objp
, void *caller
)
2763 if (cachep
->flags
& SLAB_POISON
) {
2764 #ifdef CONFIG_DEBUG_PAGEALLOC
2765 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
2766 kernel_map_pages(virt_to_page(objp
),
2767 cachep
->buffer_size
/ PAGE_SIZE
, 1);
2769 check_poison_obj(cachep
, objp
);
2771 check_poison_obj(cachep
, objp
);
2773 poison_obj(cachep
, objp
, POISON_INUSE
);
2775 if (cachep
->flags
& SLAB_STORE_USER
)
2776 *dbg_userword(cachep
, objp
) = caller
;
2778 if (cachep
->flags
& SLAB_RED_ZONE
) {
2779 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
2780 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
2781 slab_error(cachep
, "double free, or memory outside"
2782 " object was overwritten");
2784 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
2785 objp
, *dbg_redzone1(cachep
, objp
),
2786 *dbg_redzone2(cachep
, objp
));
2788 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
2789 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
2791 objp
+= obj_offset(cachep
);
2792 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
) {
2793 unsigned long ctor_flags
= SLAB_CTOR_CONSTRUCTOR
;
2795 if (!(flags
& __GFP_WAIT
))
2796 ctor_flags
|= SLAB_CTOR_ATOMIC
;
2798 cachep
->ctor(objp
, cachep
, ctor_flags
);
2803 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2806 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
2809 struct array_cache
*ac
;
2812 if (unlikely(current
->flags
& (PF_SPREAD_SLAB
| PF_MEMPOLICY
))) {
2813 objp
= alternate_node_alloc(cachep
, flags
);
2820 ac
= cpu_cache_get(cachep
);
2821 if (likely(ac
->avail
)) {
2822 STATS_INC_ALLOCHIT(cachep
);
2824 objp
= ac
->entry
[--ac
->avail
];
2826 STATS_INC_ALLOCMISS(cachep
);
2827 objp
= cache_alloc_refill(cachep
, flags
);
2832 static __always_inline
void *__cache_alloc(struct kmem_cache
*cachep
,
2833 gfp_t flags
, void *caller
)
2835 unsigned long save_flags
;
2838 cache_alloc_debugcheck_before(cachep
, flags
);
2840 local_irq_save(save_flags
);
2841 objp
= ____cache_alloc(cachep
, flags
);
2842 local_irq_restore(save_flags
);
2843 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
,
2851 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
2853 * If we are in_interrupt, then process context, including cpusets and
2854 * mempolicy, may not apply and should not be used for allocation policy.
2856 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
2858 int nid_alloc
, nid_here
;
2862 nid_alloc
= nid_here
= numa_node_id();
2863 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
2864 nid_alloc
= cpuset_mem_spread_node();
2865 else if (current
->mempolicy
)
2866 nid_alloc
= slab_node(current
->mempolicy
);
2867 if (nid_alloc
!= nid_here
)
2868 return __cache_alloc_node(cachep
, flags
, nid_alloc
);
2873 * A interface to enable slab creation on nodeid
2875 static void *__cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
2878 struct list_head
*entry
;
2880 struct kmem_list3
*l3
;
2884 l3
= cachep
->nodelists
[nodeid
];
2889 spin_lock(&l3
->list_lock
);
2890 entry
= l3
->slabs_partial
.next
;
2891 if (entry
== &l3
->slabs_partial
) {
2892 l3
->free_touched
= 1;
2893 entry
= l3
->slabs_free
.next
;
2894 if (entry
== &l3
->slabs_free
)
2898 slabp
= list_entry(entry
, struct slab
, list
);
2899 check_spinlock_acquired_node(cachep
, nodeid
);
2900 check_slabp(cachep
, slabp
);
2902 STATS_INC_NODEALLOCS(cachep
);
2903 STATS_INC_ACTIVE(cachep
);
2904 STATS_SET_HIGH(cachep
);
2906 BUG_ON(slabp
->inuse
== cachep
->num
);
2908 obj
= slab_get_obj(cachep
, slabp
, nodeid
);
2909 check_slabp(cachep
, slabp
);
2911 /* move slabp to correct slabp list: */
2912 list_del(&slabp
->list
);
2914 if (slabp
->free
== BUFCTL_END
)
2915 list_add(&slabp
->list
, &l3
->slabs_full
);
2917 list_add(&slabp
->list
, &l3
->slabs_partial
);
2919 spin_unlock(&l3
->list_lock
);
2923 spin_unlock(&l3
->list_lock
);
2924 x
= cache_grow(cachep
, flags
, nodeid
);
2936 * Caller needs to acquire correct kmem_list's list_lock
2938 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int nr_objects
,
2942 struct kmem_list3
*l3
;
2944 for (i
= 0; i
< nr_objects
; i
++) {
2945 void *objp
= objpp
[i
];
2948 slabp
= virt_to_slab(objp
);
2949 l3
= cachep
->nodelists
[node
];
2950 list_del(&slabp
->list
);
2951 check_spinlock_acquired_node(cachep
, node
);
2952 check_slabp(cachep
, slabp
);
2953 slab_put_obj(cachep
, slabp
, objp
, node
);
2954 STATS_DEC_ACTIVE(cachep
);
2956 check_slabp(cachep
, slabp
);
2958 /* fixup slab chains */
2959 if (slabp
->inuse
== 0) {
2960 if (l3
->free_objects
> l3
->free_limit
) {
2961 l3
->free_objects
-= cachep
->num
;
2962 slab_destroy(cachep
, slabp
);
2964 list_add(&slabp
->list
, &l3
->slabs_free
);
2967 /* Unconditionally move a slab to the end of the
2968 * partial list on free - maximum time for the
2969 * other objects to be freed, too.
2971 list_add_tail(&slabp
->list
, &l3
->slabs_partial
);
2976 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
2979 struct kmem_list3
*l3
;
2980 int node
= numa_node_id();
2982 batchcount
= ac
->batchcount
;
2984 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
2987 l3
= cachep
->nodelists
[node
];
2988 spin_lock(&l3
->list_lock
);
2990 struct array_cache
*shared_array
= l3
->shared
;
2991 int max
= shared_array
->limit
- shared_array
->avail
;
2993 if (batchcount
> max
)
2995 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
2996 ac
->entry
, sizeof(void *) * batchcount
);
2997 shared_array
->avail
+= batchcount
;
3002 free_block(cachep
, ac
->entry
, batchcount
, node
);
3007 struct list_head
*p
;
3009 p
= l3
->slabs_free
.next
;
3010 while (p
!= &(l3
->slabs_free
)) {
3013 slabp
= list_entry(p
, struct slab
, list
);
3014 BUG_ON(slabp
->inuse
);
3019 STATS_SET_FREEABLE(cachep
, i
);
3022 spin_unlock(&l3
->list_lock
);
3023 ac
->avail
-= batchcount
;
3024 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3028 * Release an obj back to its cache. If the obj has a constructed state, it must
3029 * be in this state _before_ it is released. Called with disabled ints.
3031 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
)
3033 struct array_cache
*ac
= cpu_cache_get(cachep
);
3036 objp
= cache_free_debugcheck(cachep
, objp
, __builtin_return_address(0));
3038 /* Make sure we are not freeing a object from another
3039 * node to the array cache on this cpu.
3044 slabp
= virt_to_slab(objp
);
3045 if (unlikely(slabp
->nodeid
!= numa_node_id())) {
3046 struct array_cache
*alien
= NULL
;
3047 int nodeid
= slabp
->nodeid
;
3048 struct kmem_list3
*l3
;
3050 l3
= cachep
->nodelists
[numa_node_id()];
3051 STATS_INC_NODEFREES(cachep
);
3052 if (l3
->alien
&& l3
->alien
[nodeid
]) {
3053 alien
= l3
->alien
[nodeid
];
3054 spin_lock(&alien
->lock
);
3055 if (unlikely(alien
->avail
== alien
->limit
))
3056 __drain_alien_cache(cachep
,
3058 alien
->entry
[alien
->avail
++] = objp
;
3059 spin_unlock(&alien
->lock
);
3061 spin_lock(&(cachep
->nodelists
[nodeid
])->
3063 free_block(cachep
, &objp
, 1, nodeid
);
3064 spin_unlock(&(cachep
->nodelists
[nodeid
])->
3071 if (likely(ac
->avail
< ac
->limit
)) {
3072 STATS_INC_FREEHIT(cachep
);
3073 ac
->entry
[ac
->avail
++] = objp
;
3076 STATS_INC_FREEMISS(cachep
);
3077 cache_flusharray(cachep
, ac
);
3078 ac
->entry
[ac
->avail
++] = objp
;
3083 * kmem_cache_alloc - Allocate an object
3084 * @cachep: The cache to allocate from.
3085 * @flags: See kmalloc().
3087 * Allocate an object from this cache. The flags are only relevant
3088 * if the cache has no available objects.
3090 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3092 return __cache_alloc(cachep
, flags
, __builtin_return_address(0));
3094 EXPORT_SYMBOL(kmem_cache_alloc
);
3097 * kmem_ptr_validate - check if an untrusted pointer might
3099 * @cachep: the cache we're checking against
3100 * @ptr: pointer to validate
3102 * This verifies that the untrusted pointer looks sane:
3103 * it is _not_ a guarantee that the pointer is actually
3104 * part of the slab cache in question, but it at least
3105 * validates that the pointer can be dereferenced and
3106 * looks half-way sane.
3108 * Currently only used for dentry validation.
3110 int fastcall
kmem_ptr_validate(struct kmem_cache
*cachep
, void *ptr
)
3112 unsigned long addr
= (unsigned long)ptr
;
3113 unsigned long min_addr
= PAGE_OFFSET
;
3114 unsigned long align_mask
= BYTES_PER_WORD
- 1;
3115 unsigned long size
= cachep
->buffer_size
;
3118 if (unlikely(addr
< min_addr
))
3120 if (unlikely(addr
> (unsigned long)high_memory
- size
))
3122 if (unlikely(addr
& align_mask
))
3124 if (unlikely(!kern_addr_valid(addr
)))
3126 if (unlikely(!kern_addr_valid(addr
+ size
- 1)))
3128 page
= virt_to_page(ptr
);
3129 if (unlikely(!PageSlab(page
)))
3131 if (unlikely(page_get_cache(page
) != cachep
))
3140 * kmem_cache_alloc_node - Allocate an object on the specified node
3141 * @cachep: The cache to allocate from.
3142 * @flags: See kmalloc().
3143 * @nodeid: node number of the target node.
3145 * Identical to kmem_cache_alloc, except that this function is slow
3146 * and can sleep. And it will allocate memory on the given node, which
3147 * can improve the performance for cpu bound structures.
3148 * New and improved: it will now make sure that the object gets
3149 * put on the correct node list so that there is no false sharing.
3151 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3153 unsigned long save_flags
;
3156 cache_alloc_debugcheck_before(cachep
, flags
);
3157 local_irq_save(save_flags
);
3159 if (nodeid
== -1 || nodeid
== numa_node_id() ||
3160 !cachep
->nodelists
[nodeid
])
3161 ptr
= ____cache_alloc(cachep
, flags
);
3163 ptr
= __cache_alloc_node(cachep
, flags
, nodeid
);
3164 local_irq_restore(save_flags
);
3166 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
,
3167 __builtin_return_address(0));
3171 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3173 void *kmalloc_node(size_t size
, gfp_t flags
, int node
)
3175 struct kmem_cache
*cachep
;
3177 cachep
= kmem_find_general_cachep(size
, flags
);
3178 if (unlikely(cachep
== NULL
))
3180 return kmem_cache_alloc_node(cachep
, flags
, node
);
3182 EXPORT_SYMBOL(kmalloc_node
);
3186 * kmalloc - allocate memory
3187 * @size: how many bytes of memory are required.
3188 * @flags: the type of memory to allocate.
3189 * @caller: function caller for debug tracking of the caller
3191 * kmalloc is the normal method of allocating memory
3194 * The @flags argument may be one of:
3196 * %GFP_USER - Allocate memory on behalf of user. May sleep.
3198 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
3200 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
3202 * Additionally, the %GFP_DMA flag may be set to indicate the memory
3203 * must be suitable for DMA. This can mean different things on different
3204 * platforms. For example, on i386, it means that the memory must come
3205 * from the first 16MB.
3207 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3210 struct kmem_cache
*cachep
;
3212 /* If you want to save a few bytes .text space: replace
3214 * Then kmalloc uses the uninlined functions instead of the inline
3217 cachep
= __find_general_cachep(size
, flags
);
3218 if (unlikely(cachep
== NULL
))
3220 return __cache_alloc(cachep
, flags
, caller
);
3223 #ifndef CONFIG_DEBUG_SLAB
3225 void *__kmalloc(size_t size
, gfp_t flags
)
3227 return __do_kmalloc(size
, flags
, NULL
);
3229 EXPORT_SYMBOL(__kmalloc
);
3233 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, void *caller
)
3235 return __do_kmalloc(size
, flags
, caller
);
3237 EXPORT_SYMBOL(__kmalloc_track_caller
);
3243 * __alloc_percpu - allocate one copy of the object for every present
3244 * cpu in the system, zeroing them.
3245 * Objects should be dereferenced using the per_cpu_ptr macro only.
3247 * @size: how many bytes of memory are required.
3249 void *__alloc_percpu(size_t size
)
3252 struct percpu_data
*pdata
= kmalloc(sizeof(*pdata
), GFP_KERNEL
);
3258 * Cannot use for_each_online_cpu since a cpu may come online
3259 * and we have no way of figuring out how to fix the array
3260 * that we have allocated then....
3263 int node
= cpu_to_node(i
);
3265 if (node_online(node
))
3266 pdata
->ptrs
[i
] = kmalloc_node(size
, GFP_KERNEL
, node
);
3268 pdata
->ptrs
[i
] = kmalloc(size
, GFP_KERNEL
);
3270 if (!pdata
->ptrs
[i
])
3272 memset(pdata
->ptrs
[i
], 0, size
);
3275 /* Catch derefs w/o wrappers */
3276 return (void *)(~(unsigned long)pdata
);
3280 if (!cpu_possible(i
))
3282 kfree(pdata
->ptrs
[i
]);
3287 EXPORT_SYMBOL(__alloc_percpu
);
3291 * kmem_cache_free - Deallocate an object
3292 * @cachep: The cache the allocation was from.
3293 * @objp: The previously allocated object.
3295 * Free an object which was previously allocated from this
3298 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3300 unsigned long flags
;
3302 local_irq_save(flags
);
3303 __cache_free(cachep
, objp
);
3304 local_irq_restore(flags
);
3306 EXPORT_SYMBOL(kmem_cache_free
);
3309 * kfree - free previously allocated memory
3310 * @objp: pointer returned by kmalloc.
3312 * If @objp is NULL, no operation is performed.
3314 * Don't free memory not originally allocated by kmalloc()
3315 * or you will run into trouble.
3317 void kfree(const void *objp
)
3319 struct kmem_cache
*c
;
3320 unsigned long flags
;
3322 if (unlikely(!objp
))
3324 local_irq_save(flags
);
3325 kfree_debugcheck(objp
);
3326 c
= virt_to_cache(objp
);
3327 mutex_debug_check_no_locks_freed(objp
, obj_size(c
));
3328 __cache_free(c
, (void *)objp
);
3329 local_irq_restore(flags
);
3331 EXPORT_SYMBOL(kfree
);
3335 * free_percpu - free previously allocated percpu memory
3336 * @objp: pointer returned by alloc_percpu.
3338 * Don't free memory not originally allocated by alloc_percpu()
3339 * The complemented objp is to check for that.
3341 void free_percpu(const void *objp
)
3344 struct percpu_data
*p
= (struct percpu_data
*)(~(unsigned long)objp
);
3347 * We allocate for all cpus so we cannot use for online cpu here.
3353 EXPORT_SYMBOL(free_percpu
);
3356 unsigned int kmem_cache_size(struct kmem_cache
*cachep
)
3358 return obj_size(cachep
);
3360 EXPORT_SYMBOL(kmem_cache_size
);
3362 const char *kmem_cache_name(struct kmem_cache
*cachep
)
3364 return cachep
->name
;
3366 EXPORT_SYMBOL_GPL(kmem_cache_name
);
3369 * This initializes kmem_list3 for all nodes.
3371 static int alloc_kmemlist(struct kmem_cache
*cachep
)
3374 struct kmem_list3
*l3
;
3377 for_each_online_node(node
) {
3378 struct array_cache
*nc
= NULL
, *new;
3379 struct array_cache
**new_alien
= NULL
;
3381 new_alien
= alloc_alien_cache(node
, cachep
->limit
);
3385 new = alloc_arraycache(node
, cachep
->shared
*cachep
->batchcount
,
3389 l3
= cachep
->nodelists
[node
];
3391 spin_lock_irq(&l3
->list_lock
);
3393 nc
= cachep
->nodelists
[node
]->shared
;
3395 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
3398 if (!cachep
->nodelists
[node
]->alien
) {
3399 l3
->alien
= new_alien
;
3402 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3403 cachep
->batchcount
+ cachep
->num
;
3404 spin_unlock_irq(&l3
->list_lock
);
3406 free_alien_cache(new_alien
);
3409 l3
= kmalloc_node(sizeof(struct kmem_list3
), GFP_KERNEL
, node
);
3413 kmem_list3_init(l3
);
3414 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
3415 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
3417 l3
->alien
= new_alien
;
3418 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3419 cachep
->batchcount
+ cachep
->num
;
3420 cachep
->nodelists
[node
] = l3
;
3428 struct ccupdate_struct
{
3429 struct kmem_cache
*cachep
;
3430 struct array_cache
*new[NR_CPUS
];
3433 static void do_ccupdate_local(void *info
)
3435 struct ccupdate_struct
*new = info
;
3436 struct array_cache
*old
;
3439 old
= cpu_cache_get(new->cachep
);
3441 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
3442 new->new[smp_processor_id()] = old
;
3445 /* Always called with the cache_chain_mutex held */
3446 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3447 int batchcount
, int shared
)
3449 struct ccupdate_struct
new;
3452 memset(&new.new, 0, sizeof(new.new));
3453 for_each_online_cpu(i
) {
3454 new.new[i
] = alloc_arraycache(cpu_to_node(i
), limit
,
3457 for (i
--; i
>= 0; i
--)
3462 new.cachep
= cachep
;
3464 on_each_cpu(do_ccupdate_local
, (void *)&new, 1, 1);
3467 cachep
->batchcount
= batchcount
;
3468 cachep
->limit
= limit
;
3469 cachep
->shared
= shared
;
3471 for_each_online_cpu(i
) {
3472 struct array_cache
*ccold
= new.new[i
];
3475 spin_lock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3476 free_block(cachep
, ccold
->entry
, ccold
->avail
, cpu_to_node(i
));
3477 spin_unlock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3481 err
= alloc_kmemlist(cachep
);
3483 printk(KERN_ERR
"alloc_kmemlist failed for %s, error %d.\n",
3484 cachep
->name
, -err
);
3490 /* Called with cache_chain_mutex held always */
3491 static void enable_cpucache(struct kmem_cache
*cachep
)
3497 * The head array serves three purposes:
3498 * - create a LIFO ordering, i.e. return objects that are cache-warm
3499 * - reduce the number of spinlock operations.
3500 * - reduce the number of linked list operations on the slab and
3501 * bufctl chains: array operations are cheaper.
3502 * The numbers are guessed, we should auto-tune as described by
3505 if (cachep
->buffer_size
> 131072)
3507 else if (cachep
->buffer_size
> PAGE_SIZE
)
3509 else if (cachep
->buffer_size
> 1024)
3511 else if (cachep
->buffer_size
> 256)
3517 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3518 * allocation behaviour: Most allocs on one cpu, most free operations
3519 * on another cpu. For these cases, an efficient object passing between
3520 * cpus is necessary. This is provided by a shared array. The array
3521 * replaces Bonwick's magazine layer.
3522 * On uniprocessor, it's functionally equivalent (but less efficient)
3523 * to a larger limit. Thus disabled by default.
3527 if (cachep
->buffer_size
<= PAGE_SIZE
)
3533 * With debugging enabled, large batchcount lead to excessively long
3534 * periods with disabled local interrupts. Limit the batchcount
3539 err
= do_tune_cpucache(cachep
, limit
, (limit
+ 1) / 2, shared
);
3541 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
3542 cachep
->name
, -err
);
3546 * Drain an array if it contains any elements taking the l3 lock only if
3547 * necessary. Note that the l3 listlock also protects the array_cache
3548 * if drain_array() is used on the shared array.
3550 void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
3551 struct array_cache
*ac
, int force
, int node
)
3555 if (!ac
|| !ac
->avail
)
3557 if (ac
->touched
&& !force
) {
3560 spin_lock_irq(&l3
->list_lock
);
3562 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
3563 if (tofree
> ac
->avail
)
3564 tofree
= (ac
->avail
+ 1) / 2;
3565 free_block(cachep
, ac
->entry
, tofree
, node
);
3566 ac
->avail
-= tofree
;
3567 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
3568 sizeof(void *) * ac
->avail
);
3570 spin_unlock_irq(&l3
->list_lock
);
3575 * cache_reap - Reclaim memory from caches.
3576 * @unused: unused parameter
3578 * Called from workqueue/eventd every few seconds.
3580 * - clear the per-cpu caches for this CPU.
3581 * - return freeable pages to the main free memory pool.
3583 * If we cannot acquire the cache chain mutex then just give up - we'll try
3584 * again on the next iteration.
3586 static void cache_reap(void *unused
)
3588 struct list_head
*walk
;
3589 struct kmem_list3
*l3
;
3590 int node
= numa_node_id();
3592 if (!mutex_trylock(&cache_chain_mutex
)) {
3593 /* Give up. Setup the next iteration. */
3594 schedule_delayed_work(&__get_cpu_var(reap_work
),
3599 list_for_each(walk
, &cache_chain
) {
3600 struct kmem_cache
*searchp
;
3601 struct list_head
*p
;
3605 searchp
= list_entry(walk
, struct kmem_cache
, next
);
3609 * We only take the l3 lock if absolutely necessary and we
3610 * have established with reasonable certainty that
3611 * we can do some work if the lock was obtained.
3613 l3
= searchp
->nodelists
[node
];
3615 reap_alien(searchp
, l3
);
3617 drain_array(searchp
, l3
, cpu_cache_get(searchp
), 0, node
);
3620 * These are racy checks but it does not matter
3621 * if we skip one check or scan twice.
3623 if (time_after(l3
->next_reap
, jiffies
))
3626 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
;
3628 drain_array(searchp
, l3
, l3
->shared
, 0, node
);
3630 if (l3
->free_touched
) {
3631 l3
->free_touched
= 0;
3635 tofree
= (l3
->free_limit
+ 5 * searchp
->num
- 1) /
3639 * Do not lock if there are no free blocks.
3641 if (list_empty(&l3
->slabs_free
))
3644 spin_lock_irq(&l3
->list_lock
);
3645 p
= l3
->slabs_free
.next
;
3646 if (p
== &(l3
->slabs_free
)) {
3647 spin_unlock_irq(&l3
->list_lock
);
3651 slabp
= list_entry(p
, struct slab
, list
);
3652 BUG_ON(slabp
->inuse
);
3653 list_del(&slabp
->list
);
3654 STATS_INC_REAPED(searchp
);
3657 * Safe to drop the lock. The slab is no longer linked
3658 * to the cache. searchp cannot disappear, we hold
3661 l3
->free_objects
-= searchp
->num
;
3662 spin_unlock_irq(&l3
->list_lock
);
3663 slab_destroy(searchp
, slabp
);
3664 } while (--tofree
> 0);
3669 mutex_unlock(&cache_chain_mutex
);
3671 /* Set up the next iteration */
3672 schedule_delayed_work(&__get_cpu_var(reap_work
), REAPTIMEOUT_CPUC
);
3675 #ifdef CONFIG_PROC_FS
3677 static void print_slabinfo_header(struct seq_file
*m
)
3680 * Output format version, so at least we can change it
3681 * without _too_ many complaints.
3684 seq_puts(m
, "slabinfo - version: 2.1 (statistics)\n");
3686 seq_puts(m
, "slabinfo - version: 2.1\n");
3688 seq_puts(m
, "# name <active_objs> <num_objs> <objsize> "
3689 "<objperslab> <pagesperslab>");
3690 seq_puts(m
, " : tunables <limit> <batchcount> <sharedfactor>");
3691 seq_puts(m
, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3693 seq_puts(m
, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3694 "<error> <maxfreeable> <nodeallocs> <remotefrees>");
3695 seq_puts(m
, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3700 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
3703 struct list_head
*p
;
3705 mutex_lock(&cache_chain_mutex
);
3707 print_slabinfo_header(m
);
3708 p
= cache_chain
.next
;
3711 if (p
== &cache_chain
)
3714 return list_entry(p
, struct kmem_cache
, next
);
3717 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
3719 struct kmem_cache
*cachep
= p
;
3721 return cachep
->next
.next
== &cache_chain
?
3722 NULL
: list_entry(cachep
->next
.next
, struct kmem_cache
, next
);
3725 static void s_stop(struct seq_file
*m
, void *p
)
3727 mutex_unlock(&cache_chain_mutex
);
3730 static int s_show(struct seq_file
*m
, void *p
)
3732 struct kmem_cache
*cachep
= p
;
3733 struct list_head
*q
;
3735 unsigned long active_objs
;
3736 unsigned long num_objs
;
3737 unsigned long active_slabs
= 0;
3738 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
3742 struct kmem_list3
*l3
;
3746 for_each_online_node(node
) {
3747 l3
= cachep
->nodelists
[node
];
3752 spin_lock_irq(&l3
->list_lock
);
3754 list_for_each(q
, &l3
->slabs_full
) {
3755 slabp
= list_entry(q
, struct slab
, list
);
3756 if (slabp
->inuse
!= cachep
->num
&& !error
)
3757 error
= "slabs_full accounting error";
3758 active_objs
+= cachep
->num
;
3761 list_for_each(q
, &l3
->slabs_partial
) {
3762 slabp
= list_entry(q
, struct slab
, list
);
3763 if (slabp
->inuse
== cachep
->num
&& !error
)
3764 error
= "slabs_partial inuse accounting error";
3765 if (!slabp
->inuse
&& !error
)
3766 error
= "slabs_partial/inuse accounting error";
3767 active_objs
+= slabp
->inuse
;
3770 list_for_each(q
, &l3
->slabs_free
) {
3771 slabp
= list_entry(q
, struct slab
, list
);
3772 if (slabp
->inuse
&& !error
)
3773 error
= "slabs_free/inuse accounting error";
3776 free_objects
+= l3
->free_objects
;
3778 shared_avail
+= l3
->shared
->avail
;
3780 spin_unlock_irq(&l3
->list_lock
);
3782 num_slabs
+= active_slabs
;
3783 num_objs
= num_slabs
* cachep
->num
;
3784 if (num_objs
- active_objs
!= free_objects
&& !error
)
3785 error
= "free_objects accounting error";
3787 name
= cachep
->name
;
3789 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
3791 seq_printf(m
, "%-17s %6lu %6lu %6u %4u %4d",
3792 name
, active_objs
, num_objs
, cachep
->buffer_size
,
3793 cachep
->num
, (1 << cachep
->gfporder
));
3794 seq_printf(m
, " : tunables %4u %4u %4u",
3795 cachep
->limit
, cachep
->batchcount
, cachep
->shared
);
3796 seq_printf(m
, " : slabdata %6lu %6lu %6lu",
3797 active_slabs
, num_slabs
, shared_avail
);
3800 unsigned long high
= cachep
->high_mark
;
3801 unsigned long allocs
= cachep
->num_allocations
;
3802 unsigned long grown
= cachep
->grown
;
3803 unsigned long reaped
= cachep
->reaped
;
3804 unsigned long errors
= cachep
->errors
;
3805 unsigned long max_freeable
= cachep
->max_freeable
;
3806 unsigned long node_allocs
= cachep
->node_allocs
;
3807 unsigned long node_frees
= cachep
->node_frees
;
3809 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu \
3810 %4lu %4lu %4lu %4lu", allocs
, high
, grown
,
3811 reaped
, errors
, max_freeable
, node_allocs
,
3816 unsigned long allochit
= atomic_read(&cachep
->allochit
);
3817 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
3818 unsigned long freehit
= atomic_read(&cachep
->freehit
);
3819 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
3821 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
3822 allochit
, allocmiss
, freehit
, freemiss
);
3830 * slabinfo_op - iterator that generates /proc/slabinfo
3839 * num-pages-per-slab
3840 * + further values on SMP and with statistics enabled
3843 struct seq_operations slabinfo_op
= {
3850 #define MAX_SLABINFO_WRITE 128
3852 * slabinfo_write - Tuning for the slab allocator
3854 * @buffer: user buffer
3855 * @count: data length
3858 ssize_t
slabinfo_write(struct file
*file
, const char __user
* buffer
,
3859 size_t count
, loff_t
*ppos
)
3861 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
3862 int limit
, batchcount
, shared
, res
;
3863 struct list_head
*p
;
3865 if (count
> MAX_SLABINFO_WRITE
)
3867 if (copy_from_user(&kbuf
, buffer
, count
))
3869 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
3871 tmp
= strchr(kbuf
, ' ');
3876 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
3879 /* Find the cache in the chain of caches. */
3880 mutex_lock(&cache_chain_mutex
);
3882 list_for_each(p
, &cache_chain
) {
3883 struct kmem_cache
*cachep
;
3885 cachep
= list_entry(p
, struct kmem_cache
, next
);
3886 if (!strcmp(cachep
->name
, kbuf
)) {
3887 if (limit
< 1 || batchcount
< 1 ||
3888 batchcount
> limit
|| shared
< 0) {
3891 res
= do_tune_cpucache(cachep
, limit
,
3892 batchcount
, shared
);
3897 mutex_unlock(&cache_chain_mutex
);
3905 * ksize - get the actual amount of memory allocated for a given object
3906 * @objp: Pointer to the object
3908 * kmalloc may internally round up allocations and return more memory
3909 * than requested. ksize() can be used to determine the actual amount of
3910 * memory allocated. The caller may use this additional memory, even though
3911 * a smaller amount of memory was initially specified with the kmalloc call.
3912 * The caller must guarantee that objp points to a valid object previously
3913 * allocated with either kmalloc() or kmem_cache_alloc(). The object
3914 * must not be freed during the duration of the call.
3916 unsigned int ksize(const void *objp
)
3918 if (unlikely(objp
== NULL
))
3921 return obj_size(virt_to_cache(objp
));