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/seq_file.h>
98 #include <linux/notifier.h>
99 #include <linux/kallsyms.h>
100 #include <linux/cpu.h>
101 #include <linux/sysctl.h>
102 #include <linux/module.h>
103 #include <linux/rcupdate.h>
104 #include <linux/string.h>
105 #include <linux/nodemask.h>
106 #include <linux/mempolicy.h>
107 #include <linux/mutex.h>
109 #include <asm/uaccess.h>
110 #include <asm/cacheflush.h>
111 #include <asm/tlbflush.h>
112 #include <asm/page.h>
115 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
116 * SLAB_RED_ZONE & SLAB_POISON.
117 * 0 for faster, smaller code (especially in the critical paths).
119 * STATS - 1 to collect stats for /proc/slabinfo.
120 * 0 for faster, smaller code (especially in the critical paths).
122 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
125 #ifdef CONFIG_DEBUG_SLAB
128 #define FORCED_DEBUG 1
132 #define FORCED_DEBUG 0
135 /* Shouldn't this be in a header file somewhere? */
136 #define BYTES_PER_WORD sizeof(void *)
138 #ifndef cache_line_size
139 #define cache_line_size() L1_CACHE_BYTES
142 #ifndef ARCH_KMALLOC_MINALIGN
144 * Enforce a minimum alignment for the kmalloc caches.
145 * Usually, the kmalloc caches are cache_line_size() aligned, except when
146 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
147 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
148 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
149 * Note that this flag disables some debug features.
151 #define ARCH_KMALLOC_MINALIGN 0
154 #ifndef ARCH_SLAB_MINALIGN
156 * Enforce a minimum alignment for all caches.
157 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
158 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
159 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
160 * some debug features.
162 #define ARCH_SLAB_MINALIGN 0
165 #ifndef ARCH_KMALLOC_FLAGS
166 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
169 /* Legal flag mask for kmem_cache_create(). */
171 # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
172 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
173 SLAB_NO_REAP | SLAB_CACHE_DMA | \
174 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
175 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
178 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | SLAB_NO_REAP | \
179 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
180 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
187 * Bufctl's are used for linking objs within a slab
190 * This implementation relies on "struct page" for locating the cache &
191 * slab an object belongs to.
192 * This allows the bufctl structure to be small (one int), but limits
193 * the number of objects a slab (not a cache) can contain when off-slab
194 * bufctls are used. The limit is the size of the largest general cache
195 * that does not use off-slab slabs.
196 * For 32bit archs with 4 kB pages, is this 56.
197 * This is not serious, as it is only for large objects, when it is unwise
198 * to have too many per slab.
199 * Note: This limit can be raised by introducing a general cache whose size
200 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
203 typedef unsigned int kmem_bufctl_t
;
204 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
205 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
206 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-2)
208 /* Max number of objs-per-slab for caches which use off-slab slabs.
209 * Needed to avoid a possible looping condition in cache_grow().
211 static unsigned long offslab_limit
;
216 * Manages the objs in a slab. Placed either at the beginning of mem allocated
217 * for a slab, or allocated from an general cache.
218 * Slabs are chained into three list: fully used, partial, fully free slabs.
221 struct list_head list
;
222 unsigned long colouroff
;
223 void *s_mem
; /* including colour offset */
224 unsigned int inuse
; /* num of objs active in slab */
226 unsigned short nodeid
;
232 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
233 * arrange for kmem_freepages to be called via RCU. This is useful if
234 * we need to approach a kernel structure obliquely, from its address
235 * obtained without the usual locking. We can lock the structure to
236 * stabilize it and check it's still at the given address, only if we
237 * can be sure that the memory has not been meanwhile reused for some
238 * other kind of object (which our subsystem's lock might corrupt).
240 * rcu_read_lock before reading the address, then rcu_read_unlock after
241 * taking the spinlock within the structure expected at that address.
243 * We assume struct slab_rcu can overlay struct slab when destroying.
246 struct rcu_head head
;
247 struct kmem_cache
*cachep
;
255 * - LIFO ordering, to hand out cache-warm objects from _alloc
256 * - reduce the number of linked list operations
257 * - reduce spinlock operations
259 * The limit is stored in the per-cpu structure to reduce the data cache
266 unsigned int batchcount
;
267 unsigned int touched
;
270 * Must have this definition in here for the proper
271 * alignment of array_cache. Also simplifies accessing
273 * [0] is for gcc 2.95. It should really be [].
277 /* bootstrap: The caches do not work without cpuarrays anymore,
278 * but the cpuarrays are allocated from the generic caches...
280 #define BOOT_CPUCACHE_ENTRIES 1
281 struct arraycache_init
{
282 struct array_cache cache
;
283 void *entries
[BOOT_CPUCACHE_ENTRIES
];
287 * The slab lists for all objects.
290 struct list_head slabs_partial
; /* partial list first, better asm code */
291 struct list_head slabs_full
;
292 struct list_head slabs_free
;
293 unsigned long free_objects
;
294 unsigned long next_reap
;
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 */
304 * Need this for bootstrapping a per node allocator.
306 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
307 struct kmem_list3 __initdata initkmem_list3
[NUM_INIT_LISTS
];
308 #define CACHE_CACHE 0
310 #define SIZE_L3 (1 + MAX_NUMNODES)
313 * This function must be completely optimized away if
314 * a constant is passed to it. Mostly the same as
315 * what is in linux/slab.h except it returns an
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 unsigned int batchcount
;
379 unsigned int buffer_size
;
380 /* 2) touched by every alloc & free from the backend */
381 struct kmem_list3
*nodelists
[MAX_NUMNODES
];
382 unsigned int flags
; /* constant flags */
383 unsigned int num
; /* # of objs per slab */
386 /* 3) cache_grow/shrink */
387 /* order of pgs per slab (2^n) */
388 unsigned int gfporder
;
390 /* force GFP flags, e.g. GFP_DMA */
393 size_t colour
; /* cache colouring range */
394 unsigned int colour_off
; /* colour offset */
395 struct kmem_cache
*slabp_cache
;
396 unsigned int slab_size
;
397 unsigned int dflags
; /* dynamic flags */
399 /* constructor func */
400 void (*ctor
) (void *, struct kmem_cache
*, unsigned long);
402 /* de-constructor func */
403 void (*dtor
) (void *, struct kmem_cache
*, unsigned long);
405 /* 4) cache creation/removal */
407 struct list_head next
;
411 unsigned long num_active
;
412 unsigned long num_allocations
;
413 unsigned long high_mark
;
415 unsigned long reaped
;
416 unsigned long errors
;
417 unsigned long max_freeable
;
418 unsigned long node_allocs
;
419 unsigned long node_frees
;
427 * If debugging is enabled, then the allocator can add additional
428 * fields and/or padding to every object. buffer_size contains the total
429 * object size including these internal fields, the following two
430 * variables contain the offset to the user object and its size.
437 #define CFLGS_OFF_SLAB (0x80000000UL)
438 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
440 #define BATCHREFILL_LIMIT 16
441 /* Optimization question: fewer reaps means less
442 * probability for unnessary cpucache drain/refill cycles.
444 * OTOH the cpuarrays can contain lots of objects,
445 * which could lock up otherwise freeable slabs.
447 #define REAPTIMEOUT_CPUC (2*HZ)
448 #define REAPTIMEOUT_LIST3 (4*HZ)
451 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
452 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
453 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
454 #define STATS_INC_GROWN(x) ((x)->grown++)
455 #define STATS_INC_REAPED(x) ((x)->reaped++)
456 #define STATS_SET_HIGH(x) do { if ((x)->num_active > (x)->high_mark) \
457 (x)->high_mark = (x)->num_active; \
459 #define STATS_INC_ERR(x) ((x)->errors++)
460 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
461 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
462 #define STATS_SET_FREEABLE(x, i) \
463 do { if ((x)->max_freeable < i) \
464 (x)->max_freeable = i; \
467 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
468 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
469 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
470 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
472 #define STATS_INC_ACTIVE(x) do { } while (0)
473 #define STATS_DEC_ACTIVE(x) do { } while (0)
474 #define STATS_INC_ALLOCED(x) do { } while (0)
475 #define STATS_INC_GROWN(x) do { } while (0)
476 #define STATS_INC_REAPED(x) do { } while (0)
477 #define STATS_SET_HIGH(x) do { } while (0)
478 #define STATS_INC_ERR(x) do { } while (0)
479 #define STATS_INC_NODEALLOCS(x) do { } while (0)
480 #define STATS_INC_NODEFREES(x) do { } while (0)
481 #define STATS_SET_FREEABLE(x, i) \
484 #define STATS_INC_ALLOCHIT(x) do { } while (0)
485 #define STATS_INC_ALLOCMISS(x) do { } while (0)
486 #define STATS_INC_FREEHIT(x) do { } while (0)
487 #define STATS_INC_FREEMISS(x) do { } while (0)
491 /* Magic nums for obj red zoning.
492 * Placed in the first word before and the first word after an obj.
494 #define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
495 #define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
497 /* ...and for poisoning */
498 #define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
499 #define POISON_FREE 0x6b /* for use-after-free poisoning */
500 #define POISON_END 0xa5 /* end-byte of poisoning */
502 /* memory layout of objects:
504 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
505 * the end of an object is aligned with the end of the real
506 * allocation. Catches writes behind the end of the allocation.
507 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
509 * cachep->obj_offset: The real object.
510 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
511 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address [BYTES_PER_WORD long]
513 static int obj_offset(struct kmem_cache
*cachep
)
515 return cachep
->obj_offset
;
518 static int obj_size(struct kmem_cache
*cachep
)
520 return cachep
->obj_size
;
523 static unsigned long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
525 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
526 return (unsigned long*) (objp
+obj_offset(cachep
)-BYTES_PER_WORD
);
529 static unsigned long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
531 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
532 if (cachep
->flags
& SLAB_STORE_USER
)
533 return (unsigned long *)(objp
+ cachep
->buffer_size
-
535 return (unsigned long *)(objp
+ cachep
->buffer_size
- BYTES_PER_WORD
);
538 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
540 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
541 return (void **)(objp
+ cachep
->buffer_size
- BYTES_PER_WORD
);
546 #define obj_offset(x) 0
547 #define obj_size(cachep) (cachep->buffer_size)
548 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
549 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
550 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
555 * Maximum size of an obj (in 2^order pages)
556 * and absolute limit for the gfp order.
558 #if defined(CONFIG_LARGE_ALLOCS)
559 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
560 #define MAX_GFP_ORDER 13 /* up to 32Mb */
561 #elif defined(CONFIG_MMU)
562 #define MAX_OBJ_ORDER 5 /* 32 pages */
563 #define MAX_GFP_ORDER 5 /* 32 pages */
565 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
566 #define MAX_GFP_ORDER 8 /* up to 1Mb */
570 * Do not go above this order unless 0 objects fit into the slab.
572 #define BREAK_GFP_ORDER_HI 1
573 #define BREAK_GFP_ORDER_LO 0
574 static int slab_break_gfp_order
= BREAK_GFP_ORDER_LO
;
576 /* Functions for storing/retrieving the cachep and or slab from the
577 * global 'mem_map'. These are used to find the slab an obj belongs to.
578 * With kfree(), these are used to find the cache which an obj belongs to.
580 static inline void page_set_cache(struct page
*page
, struct kmem_cache
*cache
)
582 page
->lru
.next
= (struct list_head
*)cache
;
585 static inline struct kmem_cache
*page_get_cache(struct page
*page
)
587 return (struct kmem_cache
*)page
->lru
.next
;
590 static inline void page_set_slab(struct page
*page
, struct slab
*slab
)
592 page
->lru
.prev
= (struct list_head
*)slab
;
595 static inline struct slab
*page_get_slab(struct page
*page
)
597 return (struct slab
*)page
->lru
.prev
;
600 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
602 struct page
*page
= virt_to_page(obj
);
603 return page_get_cache(page
);
606 static inline struct slab
*virt_to_slab(const void *obj
)
608 struct page
*page
= virt_to_page(obj
);
609 return page_get_slab(page
);
612 /* These are the default caches for kmalloc. Custom caches can have other sizes. */
613 struct cache_sizes malloc_sizes
[] = {
614 #define CACHE(x) { .cs_size = (x) },
615 #include <linux/kmalloc_sizes.h>
619 EXPORT_SYMBOL(malloc_sizes
);
621 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
627 static struct cache_names __initdata cache_names
[] = {
628 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
629 #include <linux/kmalloc_sizes.h>
634 static struct arraycache_init initarray_cache __initdata
=
635 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
636 static struct arraycache_init initarray_generic
=
637 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
639 /* internal cache of cache description objs */
640 static struct kmem_cache cache_cache
= {
642 .limit
= BOOT_CPUCACHE_ENTRIES
,
644 .buffer_size
= sizeof(struct kmem_cache
),
645 .flags
= SLAB_NO_REAP
,
646 .spinlock
= SPIN_LOCK_UNLOCKED
,
647 .name
= "kmem_cache",
649 .obj_size
= sizeof(struct kmem_cache
),
653 /* Guard access to the cache-chain. */
654 static DEFINE_MUTEX(cache_chain_mutex
);
655 static struct list_head cache_chain
;
658 * vm_enough_memory() looks at this to determine how many
659 * slab-allocated pages are possibly freeable under pressure
661 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
663 atomic_t slab_reclaim_pages
;
666 * chicken and egg problem: delay the per-cpu array allocation
667 * until the general caches are up.
676 static DEFINE_PER_CPU(struct work_struct
, reap_work
);
678 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
, int node
);
679 static void enable_cpucache(struct kmem_cache
*cachep
);
680 static void cache_reap(void *unused
);
681 static int __node_shrink(struct kmem_cache
*cachep
, int node
);
683 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
685 return cachep
->array
[smp_processor_id()];
688 static inline struct kmem_cache
*__find_general_cachep(size_t size
, gfp_t gfpflags
)
690 struct cache_sizes
*csizep
= malloc_sizes
;
693 /* This happens if someone tries to call
694 * kmem_cache_create(), or __kmalloc(), before
695 * the generic caches are initialized.
697 BUG_ON(malloc_sizes
[INDEX_AC
].cs_cachep
== NULL
);
699 while (size
> csizep
->cs_size
)
703 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
704 * has cs_{dma,}cachep==NULL. Thus no special case
705 * for large kmalloc calls required.
707 if (unlikely(gfpflags
& GFP_DMA
))
708 return csizep
->cs_dmacachep
;
709 return csizep
->cs_cachep
;
712 struct kmem_cache
*kmem_find_general_cachep(size_t size
, gfp_t gfpflags
)
714 return __find_general_cachep(size
, gfpflags
);
716 EXPORT_SYMBOL(kmem_find_general_cachep
);
718 static size_t slab_mgmt_size(size_t nr_objs
, size_t align
)
720 return ALIGN(sizeof(struct slab
)+nr_objs
*sizeof(kmem_bufctl_t
), align
);
723 /* Calculate the number of objects and left-over bytes for a given
725 static void cache_estimate(unsigned long gfporder
, size_t buffer_size
,
726 size_t align
, int flags
, size_t *left_over
,
731 size_t slab_size
= PAGE_SIZE
<< gfporder
;
734 * The slab management structure can be either off the slab or
735 * on it. For the latter case, the memory allocated for a
739 * - One kmem_bufctl_t for each object
740 * - Padding to respect alignment of @align
741 * - @buffer_size bytes for each object
743 * If the slab management structure is off the slab, then the
744 * alignment will already be calculated into the size. Because
745 * the slabs are all pages aligned, the objects will be at the
746 * correct alignment when allocated.
748 if (flags
& CFLGS_OFF_SLAB
) {
750 nr_objs
= slab_size
/ buffer_size
;
752 if (nr_objs
> SLAB_LIMIT
)
753 nr_objs
= SLAB_LIMIT
;
756 * Ignore padding for the initial guess. The padding
757 * is at most @align-1 bytes, and @buffer_size is at
758 * least @align. In the worst case, this result will
759 * be one greater than the number of objects that fit
760 * into the memory allocation when taking the padding
763 nr_objs
= (slab_size
- sizeof(struct slab
)) /
764 (buffer_size
+ sizeof(kmem_bufctl_t
));
767 * This calculated number will be either the right
768 * amount, or one greater than what we want.
770 if (slab_mgmt_size(nr_objs
, align
) + nr_objs
*buffer_size
774 if (nr_objs
> SLAB_LIMIT
)
775 nr_objs
= SLAB_LIMIT
;
777 mgmt_size
= slab_mgmt_size(nr_objs
, align
);
780 *left_over
= slab_size
- nr_objs
*buffer_size
- mgmt_size
;
783 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
785 static void __slab_error(const char *function
, struct kmem_cache
*cachep
, char *msg
)
787 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
788 function
, cachep
->name
, msg
);
793 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
794 * via the workqueue/eventd.
795 * Add the CPU number into the expiration time to minimize the possibility of
796 * the CPUs getting into lockstep and contending for the global cache chain
799 static void __devinit
start_cpu_timer(int cpu
)
801 struct work_struct
*reap_work
= &per_cpu(reap_work
, cpu
);
804 * When this gets called from do_initcalls via cpucache_init(),
805 * init_workqueues() has already run, so keventd will be setup
808 if (keventd_up() && reap_work
->func
== NULL
) {
809 INIT_WORK(reap_work
, cache_reap
, NULL
);
810 schedule_delayed_work_on(cpu
, reap_work
, HZ
+ 3 * cpu
);
814 static struct array_cache
*alloc_arraycache(int node
, int entries
,
817 int memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
818 struct array_cache
*nc
= NULL
;
820 nc
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
824 nc
->batchcount
= batchcount
;
826 spin_lock_init(&nc
->lock
);
832 static void *__cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
834 static struct array_cache
**alloc_alien_cache(int node
, int limit
)
836 struct array_cache
**ac_ptr
;
837 int memsize
= sizeof(void *) * MAX_NUMNODES
;
842 ac_ptr
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
845 if (i
== node
|| !node_online(i
)) {
849 ac_ptr
[i
] = alloc_arraycache(node
, limit
, 0xbaadf00d);
851 for (i
--; i
<= 0; i
--)
861 static void free_alien_cache(struct array_cache
**ac_ptr
)
874 static void __drain_alien_cache(struct kmem_cache
*cachep
,
875 struct array_cache
*ac
, int node
)
877 struct kmem_list3
*rl3
= cachep
->nodelists
[node
];
880 spin_lock(&rl3
->list_lock
);
881 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
883 spin_unlock(&rl3
->list_lock
);
887 static void drain_alien_cache(struct kmem_cache
*cachep
, struct array_cache
**alien
)
890 struct array_cache
*ac
;
893 for_each_online_node(i
) {
896 spin_lock_irqsave(&ac
->lock
, flags
);
897 __drain_alien_cache(cachep
, ac
, i
);
898 spin_unlock_irqrestore(&ac
->lock
, flags
);
904 #define drain_alien_cache(cachep, alien) do { } while (0)
906 static inline struct array_cache
**alloc_alien_cache(int node
, int limit
)
908 return (struct array_cache
**) 0x01020304ul
;
911 static inline void free_alien_cache(struct array_cache
**ac_ptr
)
917 static int __devinit
cpuup_callback(struct notifier_block
*nfb
,
918 unsigned long action
, void *hcpu
)
920 long cpu
= (long)hcpu
;
921 struct kmem_cache
*cachep
;
922 struct kmem_list3
*l3
= NULL
;
923 int node
= cpu_to_node(cpu
);
924 int memsize
= sizeof(struct kmem_list3
);
928 mutex_lock(&cache_chain_mutex
);
929 /* we need to do this right in the beginning since
930 * alloc_arraycache's are going to use this list.
931 * kmalloc_node allows us to add the slab to the right
932 * kmem_list3 and not this cpu's kmem_list3
935 list_for_each_entry(cachep
, &cache_chain
, next
) {
936 /* setup the size64 kmemlist for cpu before we can
937 * begin anything. Make sure some other cpu on this
938 * node has not already allocated this
940 if (!cachep
->nodelists
[node
]) {
941 if (!(l3
= kmalloc_node(memsize
,
945 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
946 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
949 * The l3s don't come and go as CPUs come and
950 * go. cache_chain_mutex is sufficient
953 cachep
->nodelists
[node
] = l3
;
956 spin_lock_irq(&cachep
->nodelists
[node
]->list_lock
);
957 cachep
->nodelists
[node
]->free_limit
=
958 (1 + nr_cpus_node(node
)) *
959 cachep
->batchcount
+ cachep
->num
;
960 spin_unlock_irq(&cachep
->nodelists
[node
]->list_lock
);
963 /* Now we can go ahead with allocating the shared array's
965 list_for_each_entry(cachep
, &cache_chain
, next
) {
966 struct array_cache
*nc
;
967 struct array_cache
*shared
;
968 struct array_cache
**alien
;
970 nc
= alloc_arraycache(node
, cachep
->limit
,
974 shared
= alloc_arraycache(node
,
975 cachep
->shared
* cachep
->batchcount
,
980 alien
= alloc_alien_cache(node
, cachep
->limit
);
983 cachep
->array
[cpu
] = nc
;
985 l3
= cachep
->nodelists
[node
];
988 spin_lock_irq(&l3
->list_lock
);
991 * We are serialised from CPU_DEAD or
992 * CPU_UP_CANCELLED by the cpucontrol lock
1003 spin_unlock_irq(&l3
->list_lock
);
1006 free_alien_cache(alien
);
1008 mutex_unlock(&cache_chain_mutex
);
1011 start_cpu_timer(cpu
);
1013 #ifdef CONFIG_HOTPLUG_CPU
1016 * Even if all the cpus of a node are down, we don't free the
1017 * kmem_list3 of any cache. This to avoid a race between
1018 * cpu_down, and a kmalloc allocation from another cpu for
1019 * memory from the node of the cpu going down. The list3
1020 * structure is usually allocated from kmem_cache_create() and
1021 * gets destroyed at kmem_cache_destroy().
1024 case CPU_UP_CANCELED
:
1025 mutex_lock(&cache_chain_mutex
);
1027 list_for_each_entry(cachep
, &cache_chain
, next
) {
1028 struct array_cache
*nc
;
1029 struct array_cache
*shared
;
1030 struct array_cache
**alien
;
1033 mask
= node_to_cpumask(node
);
1034 /* cpu is dead; no one can alloc from it. */
1035 nc
= cachep
->array
[cpu
];
1036 cachep
->array
[cpu
] = NULL
;
1037 l3
= cachep
->nodelists
[node
];
1040 goto free_array_cache
;
1042 spin_lock_irq(&l3
->list_lock
);
1044 /* Free limit for this kmem_list3 */
1045 l3
->free_limit
-= cachep
->batchcount
;
1047 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
1049 if (!cpus_empty(mask
)) {
1050 spin_unlock_irq(&l3
->list_lock
);
1051 goto free_array_cache
;
1054 shared
= l3
->shared
;
1056 free_block(cachep
, l3
->shared
->entry
,
1057 l3
->shared
->avail
, node
);
1064 spin_unlock_irq(&l3
->list_lock
);
1068 drain_alien_cache(cachep
, alien
);
1069 free_alien_cache(alien
);
1075 * In the previous loop, all the objects were freed to
1076 * the respective cache's slabs, now we can go ahead and
1077 * shrink each nodelist to its limit.
1079 list_for_each_entry(cachep
, &cache_chain
, next
) {
1080 l3
= cachep
->nodelists
[node
];
1083 spin_lock_irq(&l3
->list_lock
);
1084 /* free slabs belonging to this node */
1085 __node_shrink(cachep
, node
);
1086 spin_unlock_irq(&l3
->list_lock
);
1088 mutex_unlock(&cache_chain_mutex
);
1094 mutex_unlock(&cache_chain_mutex
);
1098 static struct notifier_block cpucache_notifier
= { &cpuup_callback
, NULL
, 0 };
1101 * swap the static kmem_list3 with kmalloced memory
1103 static void init_list(struct kmem_cache
*cachep
, struct kmem_list3
*list
, int nodeid
)
1105 struct kmem_list3
*ptr
;
1107 BUG_ON(cachep
->nodelists
[nodeid
] != list
);
1108 ptr
= kmalloc_node(sizeof(struct kmem_list3
), GFP_KERNEL
, nodeid
);
1111 local_irq_disable();
1112 memcpy(ptr
, list
, sizeof(struct kmem_list3
));
1113 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1114 cachep
->nodelists
[nodeid
] = ptr
;
1119 * Called after the gfp() functions have been enabled, and before smp_init().
1121 void __init
kmem_cache_init(void)
1124 struct cache_sizes
*sizes
;
1125 struct cache_names
*names
;
1128 for (i
= 0; i
< NUM_INIT_LISTS
; i
++) {
1129 kmem_list3_init(&initkmem_list3
[i
]);
1130 if (i
< MAX_NUMNODES
)
1131 cache_cache
.nodelists
[i
] = NULL
;
1135 * Fragmentation resistance on low memory - only use bigger
1136 * page orders on machines with more than 32MB of memory.
1138 if (num_physpages
> (32 << 20) >> PAGE_SHIFT
)
1139 slab_break_gfp_order
= BREAK_GFP_ORDER_HI
;
1141 /* Bootstrap is tricky, because several objects are allocated
1142 * from caches that do not exist yet:
1143 * 1) initialize the cache_cache cache: it contains the struct kmem_cache
1144 * structures of all caches, except cache_cache itself: cache_cache
1145 * is statically allocated.
1146 * Initially an __init data area is used for the head array and the
1147 * kmem_list3 structures, it's replaced with a kmalloc allocated
1148 * array at the end of the bootstrap.
1149 * 2) Create the first kmalloc cache.
1150 * The struct kmem_cache for the new cache is allocated normally.
1151 * An __init data area is used for the head array.
1152 * 3) Create the remaining kmalloc caches, with minimally sized
1154 * 4) Replace the __init data head arrays for cache_cache and the first
1155 * kmalloc cache with kmalloc allocated arrays.
1156 * 5) Replace the __init data for kmem_list3 for cache_cache and
1157 * the other cache's with kmalloc allocated memory.
1158 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1161 /* 1) create the cache_cache */
1162 INIT_LIST_HEAD(&cache_chain
);
1163 list_add(&cache_cache
.next
, &cache_chain
);
1164 cache_cache
.colour_off
= cache_line_size();
1165 cache_cache
.array
[smp_processor_id()] = &initarray_cache
.cache
;
1166 cache_cache
.nodelists
[numa_node_id()] = &initkmem_list3
[CACHE_CACHE
];
1168 cache_cache
.buffer_size
= ALIGN(cache_cache
.buffer_size
, cache_line_size());
1170 cache_estimate(0, cache_cache
.buffer_size
, cache_line_size(), 0,
1171 &left_over
, &cache_cache
.num
);
1172 if (!cache_cache
.num
)
1175 cache_cache
.colour
= left_over
/ cache_cache
.colour_off
;
1176 cache_cache
.slab_size
= ALIGN(cache_cache
.num
* sizeof(kmem_bufctl_t
) +
1177 sizeof(struct slab
), cache_line_size());
1179 /* 2+3) create the kmalloc caches */
1180 sizes
= malloc_sizes
;
1181 names
= cache_names
;
1183 /* Initialize the caches that provide memory for the array cache
1184 * and the kmem_list3 structures first.
1185 * Without this, further allocations will bug
1188 sizes
[INDEX_AC
].cs_cachep
= kmem_cache_create(names
[INDEX_AC
].name
,
1189 sizes
[INDEX_AC
].cs_size
,
1190 ARCH_KMALLOC_MINALIGN
,
1191 (ARCH_KMALLOC_FLAGS
|
1192 SLAB_PANIC
), NULL
, NULL
);
1194 if (INDEX_AC
!= INDEX_L3
)
1195 sizes
[INDEX_L3
].cs_cachep
=
1196 kmem_cache_create(names
[INDEX_L3
].name
,
1197 sizes
[INDEX_L3
].cs_size
,
1198 ARCH_KMALLOC_MINALIGN
,
1199 (ARCH_KMALLOC_FLAGS
| SLAB_PANIC
), NULL
,
1202 while (sizes
->cs_size
!= ULONG_MAX
) {
1204 * For performance, all the general caches are L1 aligned.
1205 * This should be particularly beneficial on SMP boxes, as it
1206 * eliminates "false sharing".
1207 * Note for systems short on memory removing the alignment will
1208 * allow tighter packing of the smaller caches.
1210 if (!sizes
->cs_cachep
)
1211 sizes
->cs_cachep
= kmem_cache_create(names
->name
,
1213 ARCH_KMALLOC_MINALIGN
,
1218 /* Inc off-slab bufctl limit until the ceiling is hit. */
1219 if (!(OFF_SLAB(sizes
->cs_cachep
))) {
1220 offslab_limit
= sizes
->cs_size
- sizeof(struct slab
);
1221 offslab_limit
/= sizeof(kmem_bufctl_t
);
1224 sizes
->cs_dmacachep
= kmem_cache_create(names
->name_dma
,
1226 ARCH_KMALLOC_MINALIGN
,
1227 (ARCH_KMALLOC_FLAGS
|
1235 /* 4) Replace the bootstrap head arrays */
1239 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1241 local_irq_disable();
1242 BUG_ON(cpu_cache_get(&cache_cache
) != &initarray_cache
.cache
);
1243 memcpy(ptr
, cpu_cache_get(&cache_cache
),
1244 sizeof(struct arraycache_init
));
1245 cache_cache
.array
[smp_processor_id()] = ptr
;
1248 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1250 local_irq_disable();
1251 BUG_ON(cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
)
1252 != &initarray_generic
.cache
);
1253 memcpy(ptr
, cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
),
1254 sizeof(struct arraycache_init
));
1255 malloc_sizes
[INDEX_AC
].cs_cachep
->array
[smp_processor_id()] =
1259 /* 5) Replace the bootstrap kmem_list3's */
1262 /* Replace the static kmem_list3 structures for the boot cpu */
1263 init_list(&cache_cache
, &initkmem_list3
[CACHE_CACHE
],
1266 for_each_online_node(node
) {
1267 init_list(malloc_sizes
[INDEX_AC
].cs_cachep
,
1268 &initkmem_list3
[SIZE_AC
+ node
], node
);
1270 if (INDEX_AC
!= INDEX_L3
) {
1271 init_list(malloc_sizes
[INDEX_L3
].cs_cachep
,
1272 &initkmem_list3
[SIZE_L3
+ node
],
1278 /* 6) resize the head arrays to their final sizes */
1280 struct kmem_cache
*cachep
;
1281 mutex_lock(&cache_chain_mutex
);
1282 list_for_each_entry(cachep
, &cache_chain
, next
)
1283 enable_cpucache(cachep
);
1284 mutex_unlock(&cache_chain_mutex
);
1288 g_cpucache_up
= FULL
;
1290 /* Register a cpu startup notifier callback
1291 * that initializes cpu_cache_get for all new cpus
1293 register_cpu_notifier(&cpucache_notifier
);
1295 /* The reap timers are started later, with a module init call:
1296 * That part of the kernel is not yet operational.
1300 static int __init
cpucache_init(void)
1305 * Register the timers that return unneeded
1308 for_each_online_cpu(cpu
)
1309 start_cpu_timer(cpu
);
1314 __initcall(cpucache_init
);
1317 * Interface to system's page allocator. No need to hold the cache-lock.
1319 * If we requested dmaable memory, we will get it. Even if we
1320 * did not request dmaable memory, we might get it, but that
1321 * would be relatively rare and ignorable.
1323 static void *kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
1329 flags
|= cachep
->gfpflags
;
1330 page
= alloc_pages_node(nodeid
, flags
, cachep
->gfporder
);
1333 addr
= page_address(page
);
1335 i
= (1 << cachep
->gfporder
);
1336 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1337 atomic_add(i
, &slab_reclaim_pages
);
1338 add_page_state(nr_slab
, i
);
1347 * Interface to system's page release.
1349 static void kmem_freepages(struct kmem_cache
*cachep
, void *addr
)
1351 unsigned long i
= (1 << cachep
->gfporder
);
1352 struct page
*page
= virt_to_page(addr
);
1353 const unsigned long nr_freed
= i
;
1356 if (!TestClearPageSlab(page
))
1360 sub_page_state(nr_slab
, nr_freed
);
1361 if (current
->reclaim_state
)
1362 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1363 free_pages((unsigned long)addr
, cachep
->gfporder
);
1364 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1365 atomic_sub(1 << cachep
->gfporder
, &slab_reclaim_pages
);
1368 static void kmem_rcu_free(struct rcu_head
*head
)
1370 struct slab_rcu
*slab_rcu
= (struct slab_rcu
*)head
;
1371 struct kmem_cache
*cachep
= slab_rcu
->cachep
;
1373 kmem_freepages(cachep
, slab_rcu
->addr
);
1374 if (OFF_SLAB(cachep
))
1375 kmem_cache_free(cachep
->slabp_cache
, slab_rcu
);
1380 #ifdef CONFIG_DEBUG_PAGEALLOC
1381 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1382 unsigned long caller
)
1384 int size
= obj_size(cachep
);
1386 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1388 if (size
< 5 * sizeof(unsigned long))
1391 *addr
++ = 0x12345678;
1393 *addr
++ = smp_processor_id();
1394 size
-= 3 * sizeof(unsigned long);
1396 unsigned long *sptr
= &caller
;
1397 unsigned long svalue
;
1399 while (!kstack_end(sptr
)) {
1401 if (kernel_text_address(svalue
)) {
1403 size
-= sizeof(unsigned long);
1404 if (size
<= sizeof(unsigned long))
1410 *addr
++ = 0x87654321;
1414 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1416 int size
= obj_size(cachep
);
1417 addr
= &((char *)addr
)[obj_offset(cachep
)];
1419 memset(addr
, val
, size
);
1420 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1423 static void dump_line(char *data
, int offset
, int limit
)
1426 printk(KERN_ERR
"%03x:", offset
);
1427 for (i
= 0; i
< limit
; i
++) {
1428 printk(" %02x", (unsigned char)data
[offset
+ i
]);
1436 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1441 if (cachep
->flags
& SLAB_RED_ZONE
) {
1442 printk(KERN_ERR
"Redzone: 0x%lx/0x%lx.\n",
1443 *dbg_redzone1(cachep
, objp
),
1444 *dbg_redzone2(cachep
, objp
));
1447 if (cachep
->flags
& SLAB_STORE_USER
) {
1448 printk(KERN_ERR
"Last user: [<%p>]",
1449 *dbg_userword(cachep
, objp
));
1450 print_symbol("(%s)",
1451 (unsigned long)*dbg_userword(cachep
, objp
));
1454 realobj
= (char *)objp
+ obj_offset(cachep
);
1455 size
= obj_size(cachep
);
1456 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1459 if (i
+ limit
> size
)
1461 dump_line(realobj
, i
, limit
);
1465 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1471 realobj
= (char *)objp
+ obj_offset(cachep
);
1472 size
= obj_size(cachep
);
1474 for (i
= 0; i
< size
; i
++) {
1475 char exp
= POISON_FREE
;
1478 if (realobj
[i
] != exp
) {
1484 "Slab corruption: start=%p, len=%d\n",
1486 print_objinfo(cachep
, objp
, 0);
1488 /* Hexdump the affected line */
1491 if (i
+ limit
> size
)
1493 dump_line(realobj
, i
, limit
);
1496 /* Limit to 5 lines */
1502 /* Print some data about the neighboring objects, if they
1505 struct slab
*slabp
= virt_to_slab(objp
);
1508 objnr
= (unsigned)(objp
- slabp
->s_mem
) / cachep
->buffer_size
;
1510 objp
= slabp
->s_mem
+ (objnr
- 1) * cachep
->buffer_size
;
1511 realobj
= (char *)objp
+ obj_offset(cachep
);
1512 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
1514 print_objinfo(cachep
, objp
, 2);
1516 if (objnr
+ 1 < cachep
->num
) {
1517 objp
= slabp
->s_mem
+ (objnr
+ 1) * cachep
->buffer_size
;
1518 realobj
= (char *)objp
+ obj_offset(cachep
);
1519 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
1521 print_objinfo(cachep
, objp
, 2);
1529 * slab_destroy_objs - call the registered destructor for each object in
1530 * a slab that is to be destroyed.
1532 static void slab_destroy_objs(struct kmem_cache
*cachep
, struct slab
*slabp
)
1535 for (i
= 0; i
< cachep
->num
; i
++) {
1536 void *objp
= slabp
->s_mem
+ cachep
->buffer_size
* i
;
1538 if (cachep
->flags
& SLAB_POISON
) {
1539 #ifdef CONFIG_DEBUG_PAGEALLOC
1540 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0
1541 && OFF_SLAB(cachep
))
1542 kernel_map_pages(virt_to_page(objp
),
1543 cachep
->buffer_size
/ PAGE_SIZE
,
1546 check_poison_obj(cachep
, objp
);
1548 check_poison_obj(cachep
, objp
);
1551 if (cachep
->flags
& SLAB_RED_ZONE
) {
1552 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1553 slab_error(cachep
, "start of a freed object "
1555 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1556 slab_error(cachep
, "end of a freed object "
1559 if (cachep
->dtor
&& !(cachep
->flags
& SLAB_POISON
))
1560 (cachep
->dtor
) (objp
+ obj_offset(cachep
), cachep
, 0);
1564 static void slab_destroy_objs(struct kmem_cache
*cachep
, struct slab
*slabp
)
1568 for (i
= 0; i
< cachep
->num
; i
++) {
1569 void *objp
= slabp
->s_mem
+ cachep
->buffer_size
* i
;
1570 (cachep
->dtor
) (objp
, cachep
, 0);
1577 * Destroy all the objs in a slab, and release the mem back to the system.
1578 * Before calling the slab must have been unlinked from the cache.
1579 * The cache-lock is not held/needed.
1581 static void slab_destroy(struct kmem_cache
*cachep
, struct slab
*slabp
)
1583 void *addr
= slabp
->s_mem
- slabp
->colouroff
;
1585 slab_destroy_objs(cachep
, slabp
);
1586 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
1587 struct slab_rcu
*slab_rcu
;
1589 slab_rcu
= (struct slab_rcu
*)slabp
;
1590 slab_rcu
->cachep
= cachep
;
1591 slab_rcu
->addr
= addr
;
1592 call_rcu(&slab_rcu
->head
, kmem_rcu_free
);
1594 kmem_freepages(cachep
, addr
);
1595 if (OFF_SLAB(cachep
))
1596 kmem_cache_free(cachep
->slabp_cache
, slabp
);
1600 /* For setting up all the kmem_list3s for cache whose buffer_size is same
1601 as size of kmem_list3. */
1602 static void set_up_list3s(struct kmem_cache
*cachep
, int index
)
1606 for_each_online_node(node
) {
1607 cachep
->nodelists
[node
] = &initkmem_list3
[index
+ node
];
1608 cachep
->nodelists
[node
]->next_reap
= jiffies
+
1610 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1615 * calculate_slab_order - calculate size (page order) of slabs
1616 * @cachep: pointer to the cache that is being created
1617 * @size: size of objects to be created in this cache.
1618 * @align: required alignment for the objects.
1619 * @flags: slab allocation flags
1621 * Also calculates the number of objects per slab.
1623 * This could be made much more intelligent. For now, try to avoid using
1624 * high order pages for slabs. When the gfp() functions are more friendly
1625 * towards high-order requests, this should be changed.
1627 static inline size_t calculate_slab_order(struct kmem_cache
*cachep
,
1628 size_t size
, size_t align
, unsigned long flags
)
1630 size_t left_over
= 0;
1632 for (;; cachep
->gfporder
++) {
1636 if (cachep
->gfporder
> MAX_GFP_ORDER
) {
1641 cache_estimate(cachep
->gfporder
, size
, align
, flags
,
1645 /* More than offslab_limit objects will cause problems */
1646 if (flags
& CFLGS_OFF_SLAB
&& cachep
->num
> offslab_limit
)
1650 left_over
= remainder
;
1653 * Large number of objects is good, but very large slabs are
1654 * currently bad for the gfp()s.
1656 if (cachep
->gfporder
>= slab_break_gfp_order
)
1659 if ((left_over
* 8) <= (PAGE_SIZE
<< cachep
->gfporder
))
1660 /* Acceptable internal fragmentation */
1667 * kmem_cache_create - Create a cache.
1668 * @name: A string which is used in /proc/slabinfo to identify this cache.
1669 * @size: The size of objects to be created in this cache.
1670 * @align: The required alignment for the objects.
1671 * @flags: SLAB flags
1672 * @ctor: A constructor for the objects.
1673 * @dtor: A destructor for the objects.
1675 * Returns a ptr to the cache on success, NULL on failure.
1676 * Cannot be called within a int, but can be interrupted.
1677 * The @ctor is run when new pages are allocated by the cache
1678 * and the @dtor is run before the pages are handed back.
1680 * @name must be valid until the cache is destroyed. This implies that
1681 * the module calling this has to destroy the cache before getting
1686 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1687 * to catch references to uninitialised memory.
1689 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1690 * for buffer overruns.
1692 * %SLAB_NO_REAP - Don't automatically reap this cache when we're under
1695 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1696 * cacheline. This can be beneficial if you're counting cycles as closely
1700 kmem_cache_create (const char *name
, size_t size
, size_t align
,
1701 unsigned long flags
, void (*ctor
)(void*, struct kmem_cache
*, unsigned long),
1702 void (*dtor
)(void*, struct kmem_cache
*, unsigned long))
1704 size_t left_over
, slab_size
, ralign
;
1705 struct kmem_cache
*cachep
= NULL
;
1706 struct list_head
*p
;
1709 * Sanity checks... these are all serious usage bugs.
1713 (size
< BYTES_PER_WORD
) ||
1714 (size
> (1 << MAX_OBJ_ORDER
) * PAGE_SIZE
) || (dtor
&& !ctor
)) {
1715 printk(KERN_ERR
"%s: Early error in slab %s\n",
1716 __FUNCTION__
, name
);
1721 * Prevent CPUs from coming and going.
1722 * lock_cpu_hotplug() nests outside cache_chain_mutex
1726 mutex_lock(&cache_chain_mutex
);
1728 list_for_each(p
, &cache_chain
) {
1729 struct kmem_cache
*pc
= list_entry(p
, struct kmem_cache
, next
);
1730 mm_segment_t old_fs
= get_fs();
1735 * This happens when the module gets unloaded and doesn't
1736 * destroy its slab cache and no-one else reuses the vmalloc
1737 * area of the module. Print a warning.
1740 res
= __get_user(tmp
, pc
->name
);
1743 printk("SLAB: cache with size %d has lost its name\n",
1748 if (!strcmp(pc
->name
, name
)) {
1749 printk("kmem_cache_create: duplicate cache %s\n", name
);
1756 WARN_ON(strchr(name
, ' ')); /* It confuses parsers */
1757 if ((flags
& SLAB_DEBUG_INITIAL
) && !ctor
) {
1758 /* No constructor, but inital state check requested */
1759 printk(KERN_ERR
"%s: No con, but init state check "
1760 "requested - %s\n", __FUNCTION__
, name
);
1761 flags
&= ~SLAB_DEBUG_INITIAL
;
1765 * Enable redzoning and last user accounting, except for caches with
1766 * large objects, if the increased size would increase the object size
1767 * above the next power of two: caches with object sizes just above a
1768 * power of two have a significant amount of internal fragmentation.
1771 || fls(size
- 1) == fls(size
- 1 + 3 * BYTES_PER_WORD
)))
1772 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
1773 if (!(flags
& SLAB_DESTROY_BY_RCU
))
1774 flags
|= SLAB_POISON
;
1776 if (flags
& SLAB_DESTROY_BY_RCU
)
1777 BUG_ON(flags
& SLAB_POISON
);
1779 if (flags
& SLAB_DESTROY_BY_RCU
)
1783 * Always checks flags, a caller might be expecting debug
1784 * support which isn't available.
1786 if (flags
& ~CREATE_MASK
)
1789 /* Check that size is in terms of words. This is needed to avoid
1790 * unaligned accesses for some archs when redzoning is used, and makes
1791 * sure any on-slab bufctl's are also correctly aligned.
1793 if (size
& (BYTES_PER_WORD
- 1)) {
1794 size
+= (BYTES_PER_WORD
- 1);
1795 size
&= ~(BYTES_PER_WORD
- 1);
1798 /* calculate out the final buffer alignment: */
1799 /* 1) arch recommendation: can be overridden for debug */
1800 if (flags
& SLAB_HWCACHE_ALIGN
) {
1801 /* Default alignment: as specified by the arch code.
1802 * Except if an object is really small, then squeeze multiple
1803 * objects into one cacheline.
1805 ralign
= cache_line_size();
1806 while (size
<= ralign
/ 2)
1809 ralign
= BYTES_PER_WORD
;
1811 /* 2) arch mandated alignment: disables debug if necessary */
1812 if (ralign
< ARCH_SLAB_MINALIGN
) {
1813 ralign
= ARCH_SLAB_MINALIGN
;
1814 if (ralign
> BYTES_PER_WORD
)
1815 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
1817 /* 3) caller mandated alignment: disables debug if necessary */
1818 if (ralign
< align
) {
1820 if (ralign
> BYTES_PER_WORD
)
1821 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
1823 /* 4) Store it. Note that the debug code below can reduce
1824 * the alignment to BYTES_PER_WORD.
1828 /* Get cache's description obj. */
1829 cachep
= kmem_cache_alloc(&cache_cache
, SLAB_KERNEL
);
1832 memset(cachep
, 0, sizeof(struct kmem_cache
));
1835 cachep
->obj_size
= size
;
1837 if (flags
& SLAB_RED_ZONE
) {
1838 /* redzoning only works with word aligned caches */
1839 align
= BYTES_PER_WORD
;
1841 /* add space for red zone words */
1842 cachep
->obj_offset
+= BYTES_PER_WORD
;
1843 size
+= 2 * BYTES_PER_WORD
;
1845 if (flags
& SLAB_STORE_USER
) {
1846 /* user store requires word alignment and
1847 * one word storage behind the end of the real
1850 align
= BYTES_PER_WORD
;
1851 size
+= BYTES_PER_WORD
;
1853 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
1854 if (size
>= malloc_sizes
[INDEX_L3
+ 1].cs_size
1855 && cachep
->obj_size
> cache_line_size() && size
< PAGE_SIZE
) {
1856 cachep
->obj_offset
+= PAGE_SIZE
- size
;
1862 /* Determine if the slab management is 'on' or 'off' slab. */
1863 if (size
>= (PAGE_SIZE
>> 3))
1865 * Size is large, assume best to place the slab management obj
1866 * off-slab (should allow better packing of objs).
1868 flags
|= CFLGS_OFF_SLAB
;
1870 size
= ALIGN(size
, align
);
1872 if ((flags
& SLAB_RECLAIM_ACCOUNT
) && size
<= PAGE_SIZE
) {
1874 * A VFS-reclaimable slab tends to have most allocations
1875 * as GFP_NOFS and we really don't want to have to be allocating
1876 * higher-order pages when we are unable to shrink dcache.
1878 cachep
->gfporder
= 0;
1879 cache_estimate(cachep
->gfporder
, size
, align
, flags
,
1880 &left_over
, &cachep
->num
);
1882 left_over
= calculate_slab_order(cachep
, size
, align
, flags
);
1885 printk("kmem_cache_create: couldn't create cache %s.\n", name
);
1886 kmem_cache_free(&cache_cache
, cachep
);
1890 slab_size
= ALIGN(cachep
->num
* sizeof(kmem_bufctl_t
)
1891 + sizeof(struct slab
), align
);
1894 * If the slab has been placed off-slab, and we have enough space then
1895 * move it on-slab. This is at the expense of any extra colouring.
1897 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= slab_size
) {
1898 flags
&= ~CFLGS_OFF_SLAB
;
1899 left_over
-= slab_size
;
1902 if (flags
& CFLGS_OFF_SLAB
) {
1903 /* really off slab. No need for manual alignment */
1905 cachep
->num
* sizeof(kmem_bufctl_t
) + sizeof(struct slab
);
1908 cachep
->colour_off
= cache_line_size();
1909 /* Offset must be a multiple of the alignment. */
1910 if (cachep
->colour_off
< align
)
1911 cachep
->colour_off
= align
;
1912 cachep
->colour
= left_over
/ cachep
->colour_off
;
1913 cachep
->slab_size
= slab_size
;
1914 cachep
->flags
= flags
;
1915 cachep
->gfpflags
= 0;
1916 if (flags
& SLAB_CACHE_DMA
)
1917 cachep
->gfpflags
|= GFP_DMA
;
1918 spin_lock_init(&cachep
->spinlock
);
1919 cachep
->buffer_size
= size
;
1921 if (flags
& CFLGS_OFF_SLAB
)
1922 cachep
->slabp_cache
= kmem_find_general_cachep(slab_size
, 0u);
1923 cachep
->ctor
= ctor
;
1924 cachep
->dtor
= dtor
;
1925 cachep
->name
= name
;
1928 if (g_cpucache_up
== FULL
) {
1929 enable_cpucache(cachep
);
1931 if (g_cpucache_up
== NONE
) {
1932 /* Note: the first kmem_cache_create must create
1933 * the cache that's used by kmalloc(24), otherwise
1934 * the creation of further caches will BUG().
1936 cachep
->array
[smp_processor_id()] =
1937 &initarray_generic
.cache
;
1939 /* If the cache that's used by
1940 * kmalloc(sizeof(kmem_list3)) is the first cache,
1941 * then we need to set up all its list3s, otherwise
1942 * the creation of further caches will BUG().
1944 set_up_list3s(cachep
, SIZE_AC
);
1945 if (INDEX_AC
== INDEX_L3
)
1946 g_cpucache_up
= PARTIAL_L3
;
1948 g_cpucache_up
= PARTIAL_AC
;
1950 cachep
->array
[smp_processor_id()] =
1951 kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1953 if (g_cpucache_up
== PARTIAL_AC
) {
1954 set_up_list3s(cachep
, SIZE_L3
);
1955 g_cpucache_up
= PARTIAL_L3
;
1958 for_each_online_node(node
) {
1960 cachep
->nodelists
[node
] =
1962 (struct kmem_list3
),
1964 BUG_ON(!cachep
->nodelists
[node
]);
1965 kmem_list3_init(cachep
->
1970 cachep
->nodelists
[numa_node_id()]->next_reap
=
1971 jiffies
+ REAPTIMEOUT_LIST3
+
1972 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1974 BUG_ON(!cpu_cache_get(cachep
));
1975 cpu_cache_get(cachep
)->avail
= 0;
1976 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
1977 cpu_cache_get(cachep
)->batchcount
= 1;
1978 cpu_cache_get(cachep
)->touched
= 0;
1979 cachep
->batchcount
= 1;
1980 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
1983 /* cache setup completed, link it into the list */
1984 list_add(&cachep
->next
, &cache_chain
);
1986 if (!cachep
&& (flags
& SLAB_PANIC
))
1987 panic("kmem_cache_create(): failed to create slab `%s'\n",
1989 mutex_unlock(&cache_chain_mutex
);
1990 unlock_cpu_hotplug();
1993 EXPORT_SYMBOL(kmem_cache_create
);
1996 static void check_irq_off(void)
1998 BUG_ON(!irqs_disabled());
2001 static void check_irq_on(void)
2003 BUG_ON(irqs_disabled());
2006 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2010 assert_spin_locked(&cachep
->nodelists
[numa_node_id()]->list_lock
);
2014 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2018 assert_spin_locked(&cachep
->nodelists
[node
]->list_lock
);
2023 #define check_irq_off() do { } while(0)
2024 #define check_irq_on() do { } while(0)
2025 #define check_spinlock_acquired(x) do { } while(0)
2026 #define check_spinlock_acquired_node(x, y) do { } while(0)
2030 * Waits for all CPUs to execute func().
2032 static void smp_call_function_all_cpus(void (*func
)(void *arg
), void *arg
)
2037 local_irq_disable();
2041 if (smp_call_function(func
, arg
, 1, 1))
2047 static void drain_array_locked(struct kmem_cache
*cachep
, struct array_cache
*ac
,
2048 int force
, int node
);
2050 static void do_drain(void *arg
)
2052 struct kmem_cache
*cachep
= (struct kmem_cache
*) arg
;
2053 struct array_cache
*ac
;
2054 int node
= numa_node_id();
2057 ac
= cpu_cache_get(cachep
);
2058 spin_lock(&cachep
->nodelists
[node
]->list_lock
);
2059 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
2060 spin_unlock(&cachep
->nodelists
[node
]->list_lock
);
2064 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2066 struct kmem_list3
*l3
;
2069 smp_call_function_all_cpus(do_drain
, cachep
);
2071 for_each_online_node(node
) {
2072 l3
= cachep
->nodelists
[node
];
2074 spin_lock_irq(&l3
->list_lock
);
2075 drain_array_locked(cachep
, l3
->shared
, 1, node
);
2076 spin_unlock_irq(&l3
->list_lock
);
2078 drain_alien_cache(cachep
, l3
->alien
);
2083 static int __node_shrink(struct kmem_cache
*cachep
, int node
)
2086 struct kmem_list3
*l3
= cachep
->nodelists
[node
];
2090 struct list_head
*p
;
2092 p
= l3
->slabs_free
.prev
;
2093 if (p
== &l3
->slabs_free
)
2096 slabp
= list_entry(l3
->slabs_free
.prev
, struct slab
, list
);
2101 list_del(&slabp
->list
);
2103 l3
->free_objects
-= cachep
->num
;
2104 spin_unlock_irq(&l3
->list_lock
);
2105 slab_destroy(cachep
, slabp
);
2106 spin_lock_irq(&l3
->list_lock
);
2108 ret
= !list_empty(&l3
->slabs_full
) || !list_empty(&l3
->slabs_partial
);
2112 static int __cache_shrink(struct kmem_cache
*cachep
)
2115 struct kmem_list3
*l3
;
2117 drain_cpu_caches(cachep
);
2120 for_each_online_node(i
) {
2121 l3
= cachep
->nodelists
[i
];
2123 spin_lock_irq(&l3
->list_lock
);
2124 ret
+= __node_shrink(cachep
, i
);
2125 spin_unlock_irq(&l3
->list_lock
);
2128 return (ret
? 1 : 0);
2132 * kmem_cache_shrink - Shrink a cache.
2133 * @cachep: The cache to shrink.
2135 * Releases as many slabs as possible for a cache.
2136 * To help debugging, a zero exit status indicates all slabs were released.
2138 int kmem_cache_shrink(struct kmem_cache
*cachep
)
2140 if (!cachep
|| in_interrupt())
2143 return __cache_shrink(cachep
);
2145 EXPORT_SYMBOL(kmem_cache_shrink
);
2148 * kmem_cache_destroy - delete a cache
2149 * @cachep: the cache to destroy
2151 * Remove a struct kmem_cache object from the slab cache.
2152 * Returns 0 on success.
2154 * It is expected this function will be called by a module when it is
2155 * unloaded. This will remove the cache completely, and avoid a duplicate
2156 * cache being allocated each time a module is loaded and unloaded, if the
2157 * module doesn't have persistent in-kernel storage across loads and unloads.
2159 * The cache must be empty before calling this function.
2161 * The caller must guarantee that noone will allocate memory from the cache
2162 * during the kmem_cache_destroy().
2164 int kmem_cache_destroy(struct kmem_cache
*cachep
)
2167 struct kmem_list3
*l3
;
2169 if (!cachep
|| in_interrupt())
2172 /* Don't let CPUs to come and go */
2175 /* Find the cache in the chain of caches. */
2176 mutex_lock(&cache_chain_mutex
);
2178 * the chain is never empty, cache_cache is never destroyed
2180 list_del(&cachep
->next
);
2181 mutex_unlock(&cache_chain_mutex
);
2183 if (__cache_shrink(cachep
)) {
2184 slab_error(cachep
, "Can't free all objects");
2185 mutex_lock(&cache_chain_mutex
);
2186 list_add(&cachep
->next
, &cache_chain
);
2187 mutex_unlock(&cache_chain_mutex
);
2188 unlock_cpu_hotplug();
2192 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
2195 for_each_online_cpu(i
)
2196 kfree(cachep
->array
[i
]);
2198 /* NUMA: free the list3 structures */
2199 for_each_online_node(i
) {
2200 if ((l3
= cachep
->nodelists
[i
])) {
2202 free_alien_cache(l3
->alien
);
2206 kmem_cache_free(&cache_cache
, cachep
);
2208 unlock_cpu_hotplug();
2212 EXPORT_SYMBOL(kmem_cache_destroy
);
2214 /* Get the memory for a slab management obj. */
2215 static struct slab
*alloc_slabmgmt(struct kmem_cache
*cachep
, void *objp
,
2216 int colour_off
, gfp_t local_flags
)
2220 if (OFF_SLAB(cachep
)) {
2221 /* Slab management obj is off-slab. */
2222 slabp
= kmem_cache_alloc(cachep
->slabp_cache
, local_flags
);
2226 slabp
= objp
+ colour_off
;
2227 colour_off
+= cachep
->slab_size
;
2230 slabp
->colouroff
= colour_off
;
2231 slabp
->s_mem
= objp
+ colour_off
;
2236 static inline kmem_bufctl_t
*slab_bufctl(struct slab
*slabp
)
2238 return (kmem_bufctl_t
*) (slabp
+ 1);
2241 static void cache_init_objs(struct kmem_cache
*cachep
,
2242 struct slab
*slabp
, unsigned long ctor_flags
)
2246 for (i
= 0; i
< cachep
->num
; i
++) {
2247 void *objp
= slabp
->s_mem
+ cachep
->buffer_size
* i
;
2249 /* need to poison the objs? */
2250 if (cachep
->flags
& SLAB_POISON
)
2251 poison_obj(cachep
, objp
, POISON_FREE
);
2252 if (cachep
->flags
& SLAB_STORE_USER
)
2253 *dbg_userword(cachep
, objp
) = NULL
;
2255 if (cachep
->flags
& SLAB_RED_ZONE
) {
2256 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2257 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2260 * Constructors are not allowed to allocate memory from
2261 * the same cache which they are a constructor for.
2262 * Otherwise, deadlock. They must also be threaded.
2264 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2265 cachep
->ctor(objp
+ obj_offset(cachep
), cachep
,
2268 if (cachep
->flags
& SLAB_RED_ZONE
) {
2269 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2270 slab_error(cachep
, "constructor overwrote the"
2271 " end of an object");
2272 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2273 slab_error(cachep
, "constructor overwrote the"
2274 " start of an object");
2276 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
)
2277 && cachep
->flags
& SLAB_POISON
)
2278 kernel_map_pages(virt_to_page(objp
),
2279 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2282 cachep
->ctor(objp
, cachep
, ctor_flags
);
2284 slab_bufctl(slabp
)[i
] = i
+ 1;
2286 slab_bufctl(slabp
)[i
- 1] = BUFCTL_END
;
2290 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2292 if (flags
& SLAB_DMA
) {
2293 if (!(cachep
->gfpflags
& GFP_DMA
))
2296 if (cachep
->gfpflags
& GFP_DMA
)
2301 static void *slab_get_obj(struct kmem_cache
*cachep
, struct slab
*slabp
, int nodeid
)
2303 void *objp
= slabp
->s_mem
+ (slabp
->free
* cachep
->buffer_size
);
2307 next
= slab_bufctl(slabp
)[slabp
->free
];
2309 slab_bufctl(slabp
)[slabp
->free
] = BUFCTL_FREE
;
2310 WARN_ON(slabp
->nodeid
!= nodeid
);
2317 static void slab_put_obj(struct kmem_cache
*cachep
, struct slab
*slabp
, void *objp
,
2320 unsigned int objnr
= (unsigned)(objp
-slabp
->s_mem
) / cachep
->buffer_size
;
2323 /* Verify that the slab belongs to the intended node */
2324 WARN_ON(slabp
->nodeid
!= nodeid
);
2326 if (slab_bufctl(slabp
)[objnr
] != BUFCTL_FREE
) {
2327 printk(KERN_ERR
"slab: double free detected in cache "
2328 "'%s', objp %p\n", cachep
->name
, objp
);
2332 slab_bufctl(slabp
)[objnr
] = slabp
->free
;
2333 slabp
->free
= objnr
;
2337 static void set_slab_attr(struct kmem_cache
*cachep
, struct slab
*slabp
, void *objp
)
2342 /* Nasty!!!!!! I hope this is OK. */
2343 i
= 1 << cachep
->gfporder
;
2344 page
= virt_to_page(objp
);
2346 page_set_cache(page
, cachep
);
2347 page_set_slab(page
, slabp
);
2353 * Grow (by 1) the number of slabs within a cache. This is called by
2354 * kmem_cache_alloc() when there are no active objs left in a cache.
2356 static int cache_grow(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
2362 unsigned long ctor_flags
;
2363 struct kmem_list3
*l3
;
2365 /* Be lazy and only check for valid flags here,
2366 * keeping it out of the critical path in kmem_cache_alloc().
2368 if (flags
& ~(SLAB_DMA
| SLAB_LEVEL_MASK
| SLAB_NO_GROW
))
2370 if (flags
& SLAB_NO_GROW
)
2373 ctor_flags
= SLAB_CTOR_CONSTRUCTOR
;
2374 local_flags
= (flags
& SLAB_LEVEL_MASK
);
2375 if (!(local_flags
& __GFP_WAIT
))
2377 * Not allowed to sleep. Need to tell a constructor about
2378 * this - it might need to know...
2380 ctor_flags
|= SLAB_CTOR_ATOMIC
;
2382 /* Take the l3 list lock to change the colour_next on this node */
2384 l3
= cachep
->nodelists
[nodeid
];
2385 spin_lock(&l3
->list_lock
);
2387 /* Get colour for the slab, and cal the next value. */
2388 offset
= l3
->colour_next
;
2390 if (l3
->colour_next
>= cachep
->colour
)
2391 l3
->colour_next
= 0;
2392 spin_unlock(&l3
->list_lock
);
2394 offset
*= cachep
->colour_off
;
2396 if (local_flags
& __GFP_WAIT
)
2400 * The test for missing atomic flag is performed here, rather than
2401 * the more obvious place, simply to reduce the critical path length
2402 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2403 * will eventually be caught here (where it matters).
2405 kmem_flagcheck(cachep
, flags
);
2407 /* Get mem for the objs.
2408 * Attempt to allocate a physical page from 'nodeid',
2410 if (!(objp
= kmem_getpages(cachep
, flags
, nodeid
)))
2413 /* Get slab management. */
2414 if (!(slabp
= alloc_slabmgmt(cachep
, objp
, offset
, local_flags
)))
2417 slabp
->nodeid
= nodeid
;
2418 set_slab_attr(cachep
, slabp
, objp
);
2420 cache_init_objs(cachep
, slabp
, ctor_flags
);
2422 if (local_flags
& __GFP_WAIT
)
2423 local_irq_disable();
2425 spin_lock(&l3
->list_lock
);
2427 /* Make slab active. */
2428 list_add_tail(&slabp
->list
, &(l3
->slabs_free
));
2429 STATS_INC_GROWN(cachep
);
2430 l3
->free_objects
+= cachep
->num
;
2431 spin_unlock(&l3
->list_lock
);
2434 kmem_freepages(cachep
, objp
);
2436 if (local_flags
& __GFP_WAIT
)
2437 local_irq_disable();
2444 * Perform extra freeing checks:
2445 * - detect bad pointers.
2446 * - POISON/RED_ZONE checking
2447 * - destructor calls, for caches with POISON+dtor
2449 static void kfree_debugcheck(const void *objp
)
2453 if (!virt_addr_valid(objp
)) {
2454 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
2455 (unsigned long)objp
);
2458 page
= virt_to_page(objp
);
2459 if (!PageSlab(page
)) {
2460 printk(KERN_ERR
"kfree_debugcheck: bad ptr %lxh.\n",
2461 (unsigned long)objp
);
2466 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2473 objp
-= obj_offset(cachep
);
2474 kfree_debugcheck(objp
);
2475 page
= virt_to_page(objp
);
2477 if (page_get_cache(page
) != cachep
) {
2479 "mismatch in kmem_cache_free: expected cache %p, got %p\n",
2480 page_get_cache(page
), cachep
);
2481 printk(KERN_ERR
"%p is %s.\n", cachep
, cachep
->name
);
2482 printk(KERN_ERR
"%p is %s.\n", page_get_cache(page
),
2483 page_get_cache(page
)->name
);
2486 slabp
= page_get_slab(page
);
2488 if (cachep
->flags
& SLAB_RED_ZONE
) {
2489 if (*dbg_redzone1(cachep
, objp
) != RED_ACTIVE
2490 || *dbg_redzone2(cachep
, objp
) != RED_ACTIVE
) {
2492 "double free, or memory outside"
2493 " object was overwritten");
2495 "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2496 objp
, *dbg_redzone1(cachep
, objp
),
2497 *dbg_redzone2(cachep
, objp
));
2499 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2500 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2502 if (cachep
->flags
& SLAB_STORE_USER
)
2503 *dbg_userword(cachep
, objp
) = caller
;
2505 objnr
= (unsigned)(objp
- slabp
->s_mem
) / cachep
->buffer_size
;
2507 BUG_ON(objnr
>= cachep
->num
);
2508 BUG_ON(objp
!= slabp
->s_mem
+ objnr
* cachep
->buffer_size
);
2510 if (cachep
->flags
& SLAB_DEBUG_INITIAL
) {
2511 /* Need to call the slab's constructor so the
2512 * caller can perform a verify of its state (debugging).
2513 * Called without the cache-lock held.
2515 cachep
->ctor(objp
+ obj_offset(cachep
),
2516 cachep
, SLAB_CTOR_CONSTRUCTOR
| SLAB_CTOR_VERIFY
);
2518 if (cachep
->flags
& SLAB_POISON
&& cachep
->dtor
) {
2519 /* we want to cache poison the object,
2520 * call the destruction callback
2522 cachep
->dtor(objp
+ obj_offset(cachep
), cachep
, 0);
2524 if (cachep
->flags
& SLAB_POISON
) {
2525 #ifdef CONFIG_DEBUG_PAGEALLOC
2526 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
)) {
2527 store_stackinfo(cachep
, objp
, (unsigned long)caller
);
2528 kernel_map_pages(virt_to_page(objp
),
2529 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2531 poison_obj(cachep
, objp
, POISON_FREE
);
2534 poison_obj(cachep
, objp
, POISON_FREE
);
2540 static void check_slabp(struct kmem_cache
*cachep
, struct slab
*slabp
)
2545 /* Check slab's freelist to see if this obj is there. */
2546 for (i
= slabp
->free
; i
!= BUFCTL_END
; i
= slab_bufctl(slabp
)[i
]) {
2548 if (entries
> cachep
->num
|| i
>= cachep
->num
)
2551 if (entries
!= cachep
->num
- slabp
->inuse
) {
2554 "slab: Internal list corruption detected in cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2555 cachep
->name
, cachep
->num
, slabp
, slabp
->inuse
);
2557 i
< sizeof(slabp
) + cachep
->num
* sizeof(kmem_bufctl_t
);
2560 printk("\n%03x:", i
);
2561 printk(" %02x", ((unsigned char *)slabp
)[i
]);
2568 #define kfree_debugcheck(x) do { } while(0)
2569 #define cache_free_debugcheck(x,objp,z) (objp)
2570 #define check_slabp(x,y) do { } while(0)
2573 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
2576 struct kmem_list3
*l3
;
2577 struct array_cache
*ac
;
2580 ac
= cpu_cache_get(cachep
);
2582 batchcount
= ac
->batchcount
;
2583 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2584 /* if there was little recent activity on this
2585 * cache, then perform only a partial refill.
2586 * Otherwise we could generate refill bouncing.
2588 batchcount
= BATCHREFILL_LIMIT
;
2590 l3
= cachep
->nodelists
[numa_node_id()];
2592 BUG_ON(ac
->avail
> 0 || !l3
);
2593 spin_lock(&l3
->list_lock
);
2596 struct array_cache
*shared_array
= l3
->shared
;
2597 if (shared_array
->avail
) {
2598 if (batchcount
> shared_array
->avail
)
2599 batchcount
= shared_array
->avail
;
2600 shared_array
->avail
-= batchcount
;
2601 ac
->avail
= batchcount
;
2603 &(shared_array
->entry
[shared_array
->avail
]),
2604 sizeof(void *) * batchcount
);
2605 shared_array
->touched
= 1;
2609 while (batchcount
> 0) {
2610 struct list_head
*entry
;
2612 /* Get slab alloc is to come from. */
2613 entry
= l3
->slabs_partial
.next
;
2614 if (entry
== &l3
->slabs_partial
) {
2615 l3
->free_touched
= 1;
2616 entry
= l3
->slabs_free
.next
;
2617 if (entry
== &l3
->slabs_free
)
2621 slabp
= list_entry(entry
, struct slab
, list
);
2622 check_slabp(cachep
, slabp
);
2623 check_spinlock_acquired(cachep
);
2624 while (slabp
->inuse
< cachep
->num
&& batchcount
--) {
2625 STATS_INC_ALLOCED(cachep
);
2626 STATS_INC_ACTIVE(cachep
);
2627 STATS_SET_HIGH(cachep
);
2629 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, slabp
,
2632 check_slabp(cachep
, slabp
);
2634 /* move slabp to correct slabp list: */
2635 list_del(&slabp
->list
);
2636 if (slabp
->free
== BUFCTL_END
)
2637 list_add(&slabp
->list
, &l3
->slabs_full
);
2639 list_add(&slabp
->list
, &l3
->slabs_partial
);
2643 l3
->free_objects
-= ac
->avail
;
2645 spin_unlock(&l3
->list_lock
);
2647 if (unlikely(!ac
->avail
)) {
2649 x
= cache_grow(cachep
, flags
, numa_node_id());
2651 // cache_grow can reenable interrupts, then ac could change.
2652 ac
= cpu_cache_get(cachep
);
2653 if (!x
&& ac
->avail
== 0) // no objects in sight? abort
2656 if (!ac
->avail
) // objects refilled by interrupt?
2660 return ac
->entry
[--ac
->avail
];
2664 cache_alloc_debugcheck_before(struct kmem_cache
*cachep
, gfp_t flags
)
2666 might_sleep_if(flags
& __GFP_WAIT
);
2668 kmem_flagcheck(cachep
, flags
);
2673 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
, gfp_t flags
,
2674 void *objp
, void *caller
)
2678 if (cachep
->flags
& SLAB_POISON
) {
2679 #ifdef CONFIG_DEBUG_PAGEALLOC
2680 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
2681 kernel_map_pages(virt_to_page(objp
),
2682 cachep
->buffer_size
/ PAGE_SIZE
, 1);
2684 check_poison_obj(cachep
, objp
);
2686 check_poison_obj(cachep
, objp
);
2688 poison_obj(cachep
, objp
, POISON_INUSE
);
2690 if (cachep
->flags
& SLAB_STORE_USER
)
2691 *dbg_userword(cachep
, objp
) = caller
;
2693 if (cachep
->flags
& SLAB_RED_ZONE
) {
2694 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
2695 || *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
2697 "double free, or memory outside"
2698 " object was overwritten");
2700 "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2701 objp
, *dbg_redzone1(cachep
, objp
),
2702 *dbg_redzone2(cachep
, objp
));
2704 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
2705 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
2707 objp
+= obj_offset(cachep
);
2708 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
) {
2709 unsigned long ctor_flags
= SLAB_CTOR_CONSTRUCTOR
;
2711 if (!(flags
& __GFP_WAIT
))
2712 ctor_flags
|= SLAB_CTOR_ATOMIC
;
2714 cachep
->ctor(objp
, cachep
, ctor_flags
);
2719 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2722 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
2725 struct array_cache
*ac
;
2728 if (unlikely(current
->mempolicy
&& !in_interrupt())) {
2729 int nid
= slab_node(current
->mempolicy
);
2731 if (nid
!= numa_node_id())
2732 return __cache_alloc_node(cachep
, flags
, nid
);
2737 ac
= cpu_cache_get(cachep
);
2738 if (likely(ac
->avail
)) {
2739 STATS_INC_ALLOCHIT(cachep
);
2741 objp
= ac
->entry
[--ac
->avail
];
2743 STATS_INC_ALLOCMISS(cachep
);
2744 objp
= cache_alloc_refill(cachep
, flags
);
2749 static __always_inline
void *
2750 __cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
, void *caller
)
2752 unsigned long save_flags
;
2755 cache_alloc_debugcheck_before(cachep
, flags
);
2757 local_irq_save(save_flags
);
2758 objp
= ____cache_alloc(cachep
, flags
);
2759 local_irq_restore(save_flags
);
2760 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
,
2768 * A interface to enable slab creation on nodeid
2770 static void *__cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
2772 struct list_head
*entry
;
2774 struct kmem_list3
*l3
;
2778 l3
= cachep
->nodelists
[nodeid
];
2783 spin_lock(&l3
->list_lock
);
2784 entry
= l3
->slabs_partial
.next
;
2785 if (entry
== &l3
->slabs_partial
) {
2786 l3
->free_touched
= 1;
2787 entry
= l3
->slabs_free
.next
;
2788 if (entry
== &l3
->slabs_free
)
2792 slabp
= list_entry(entry
, struct slab
, list
);
2793 check_spinlock_acquired_node(cachep
, nodeid
);
2794 check_slabp(cachep
, slabp
);
2796 STATS_INC_NODEALLOCS(cachep
);
2797 STATS_INC_ACTIVE(cachep
);
2798 STATS_SET_HIGH(cachep
);
2800 BUG_ON(slabp
->inuse
== cachep
->num
);
2802 obj
= slab_get_obj(cachep
, slabp
, nodeid
);
2803 check_slabp(cachep
, slabp
);
2805 /* move slabp to correct slabp list: */
2806 list_del(&slabp
->list
);
2808 if (slabp
->free
== BUFCTL_END
) {
2809 list_add(&slabp
->list
, &l3
->slabs_full
);
2811 list_add(&slabp
->list
, &l3
->slabs_partial
);
2814 spin_unlock(&l3
->list_lock
);
2818 spin_unlock(&l3
->list_lock
);
2819 x
= cache_grow(cachep
, flags
, nodeid
);
2831 * Caller needs to acquire correct kmem_list's list_lock
2833 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int nr_objects
,
2837 struct kmem_list3
*l3
;
2839 for (i
= 0; i
< nr_objects
; i
++) {
2840 void *objp
= objpp
[i
];
2843 slabp
= virt_to_slab(objp
);
2844 l3
= cachep
->nodelists
[node
];
2845 list_del(&slabp
->list
);
2846 check_spinlock_acquired_node(cachep
, node
);
2847 check_slabp(cachep
, slabp
);
2848 slab_put_obj(cachep
, slabp
, objp
, node
);
2849 STATS_DEC_ACTIVE(cachep
);
2851 check_slabp(cachep
, slabp
);
2853 /* fixup slab chains */
2854 if (slabp
->inuse
== 0) {
2855 if (l3
->free_objects
> l3
->free_limit
) {
2856 l3
->free_objects
-= cachep
->num
;
2857 slab_destroy(cachep
, slabp
);
2859 list_add(&slabp
->list
, &l3
->slabs_free
);
2862 /* Unconditionally move a slab to the end of the
2863 * partial list on free - maximum time for the
2864 * other objects to be freed, too.
2866 list_add_tail(&slabp
->list
, &l3
->slabs_partial
);
2871 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
2874 struct kmem_list3
*l3
;
2875 int node
= numa_node_id();
2877 batchcount
= ac
->batchcount
;
2879 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
2882 l3
= cachep
->nodelists
[node
];
2883 spin_lock(&l3
->list_lock
);
2885 struct array_cache
*shared_array
= l3
->shared
;
2886 int max
= shared_array
->limit
- shared_array
->avail
;
2888 if (batchcount
> max
)
2890 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
2891 ac
->entry
, sizeof(void *) * batchcount
);
2892 shared_array
->avail
+= batchcount
;
2897 free_block(cachep
, ac
->entry
, batchcount
, node
);
2902 struct list_head
*p
;
2904 p
= l3
->slabs_free
.next
;
2905 while (p
!= &(l3
->slabs_free
)) {
2908 slabp
= list_entry(p
, struct slab
, list
);
2909 BUG_ON(slabp
->inuse
);
2914 STATS_SET_FREEABLE(cachep
, i
);
2917 spin_unlock(&l3
->list_lock
);
2918 ac
->avail
-= batchcount
;
2919 memmove(ac
->entry
, &(ac
->entry
[batchcount
]),
2920 sizeof(void *) * ac
->avail
);
2925 * Release an obj back to its cache. If the obj has a constructed
2926 * state, it must be in this state _before_ it is released.
2928 * Called with disabled ints.
2930 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
)
2932 struct array_cache
*ac
= cpu_cache_get(cachep
);
2935 objp
= cache_free_debugcheck(cachep
, objp
, __builtin_return_address(0));
2937 /* Make sure we are not freeing a object from another
2938 * node to the array cache on this cpu.
2943 slabp
= virt_to_slab(objp
);
2944 if (unlikely(slabp
->nodeid
!= numa_node_id())) {
2945 struct array_cache
*alien
= NULL
;
2946 int nodeid
= slabp
->nodeid
;
2947 struct kmem_list3
*l3
=
2948 cachep
->nodelists
[numa_node_id()];
2950 STATS_INC_NODEFREES(cachep
);
2951 if (l3
->alien
&& l3
->alien
[nodeid
]) {
2952 alien
= l3
->alien
[nodeid
];
2953 spin_lock(&alien
->lock
);
2954 if (unlikely(alien
->avail
== alien
->limit
))
2955 __drain_alien_cache(cachep
,
2957 alien
->entry
[alien
->avail
++] = objp
;
2958 spin_unlock(&alien
->lock
);
2960 spin_lock(&(cachep
->nodelists
[nodeid
])->
2962 free_block(cachep
, &objp
, 1, nodeid
);
2963 spin_unlock(&(cachep
->nodelists
[nodeid
])->
2970 if (likely(ac
->avail
< ac
->limit
)) {
2971 STATS_INC_FREEHIT(cachep
);
2972 ac
->entry
[ac
->avail
++] = objp
;
2975 STATS_INC_FREEMISS(cachep
);
2976 cache_flusharray(cachep
, ac
);
2977 ac
->entry
[ac
->avail
++] = objp
;
2982 * kmem_cache_alloc - Allocate an object
2983 * @cachep: The cache to allocate from.
2984 * @flags: See kmalloc().
2986 * Allocate an object from this cache. The flags are only relevant
2987 * if the cache has no available objects.
2989 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
2991 return __cache_alloc(cachep
, flags
, __builtin_return_address(0));
2993 EXPORT_SYMBOL(kmem_cache_alloc
);
2996 * kmem_ptr_validate - check if an untrusted pointer might
2998 * @cachep: the cache we're checking against
2999 * @ptr: pointer to validate
3001 * This verifies that the untrusted pointer looks sane:
3002 * it is _not_ a guarantee that the pointer is actually
3003 * part of the slab cache in question, but it at least
3004 * validates that the pointer can be dereferenced and
3005 * looks half-way sane.
3007 * Currently only used for dentry validation.
3009 int fastcall
kmem_ptr_validate(struct kmem_cache
*cachep
, void *ptr
)
3011 unsigned long addr
= (unsigned long)ptr
;
3012 unsigned long min_addr
= PAGE_OFFSET
;
3013 unsigned long align_mask
= BYTES_PER_WORD
- 1;
3014 unsigned long size
= cachep
->buffer_size
;
3017 if (unlikely(addr
< min_addr
))
3019 if (unlikely(addr
> (unsigned long)high_memory
- size
))
3021 if (unlikely(addr
& align_mask
))
3023 if (unlikely(!kern_addr_valid(addr
)))
3025 if (unlikely(!kern_addr_valid(addr
+ size
- 1)))
3027 page
= virt_to_page(ptr
);
3028 if (unlikely(!PageSlab(page
)))
3030 if (unlikely(page_get_cache(page
) != cachep
))
3039 * kmem_cache_alloc_node - Allocate an object on the specified node
3040 * @cachep: The cache to allocate from.
3041 * @flags: See kmalloc().
3042 * @nodeid: node number of the target node.
3044 * Identical to kmem_cache_alloc, except that this function is slow
3045 * and can sleep. And it will allocate memory on the given node, which
3046 * can improve the performance for cpu bound structures.
3047 * New and improved: it will now make sure that the object gets
3048 * put on the correct node list so that there is no false sharing.
3050 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3052 unsigned long save_flags
;
3055 cache_alloc_debugcheck_before(cachep
, flags
);
3056 local_irq_save(save_flags
);
3058 if (nodeid
== -1 || nodeid
== numa_node_id() ||
3059 !cachep
->nodelists
[nodeid
])
3060 ptr
= ____cache_alloc(cachep
, flags
);
3062 ptr
= __cache_alloc_node(cachep
, flags
, nodeid
);
3063 local_irq_restore(save_flags
);
3065 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
,
3066 __builtin_return_address(0));
3070 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3072 void *kmalloc_node(size_t size
, gfp_t flags
, int node
)
3074 struct kmem_cache
*cachep
;
3076 cachep
= kmem_find_general_cachep(size
, flags
);
3077 if (unlikely(cachep
== NULL
))
3079 return kmem_cache_alloc_node(cachep
, flags
, node
);
3081 EXPORT_SYMBOL(kmalloc_node
);
3085 * kmalloc - allocate memory
3086 * @size: how many bytes of memory are required.
3087 * @flags: the type of memory to allocate.
3089 * kmalloc is the normal method of allocating memory
3092 * The @flags argument may be one of:
3094 * %GFP_USER - Allocate memory on behalf of user. May sleep.
3096 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
3098 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
3100 * Additionally, the %GFP_DMA flag may be set to indicate the memory
3101 * must be suitable for DMA. This can mean different things on different
3102 * platforms. For example, on i386, it means that the memory must come
3103 * from the first 16MB.
3105 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3108 struct kmem_cache
*cachep
;
3110 /* If you want to save a few bytes .text space: replace
3112 * Then kmalloc uses the uninlined functions instead of the inline
3115 cachep
= __find_general_cachep(size
, flags
);
3116 if (unlikely(cachep
== NULL
))
3118 return __cache_alloc(cachep
, flags
, caller
);
3121 #ifndef CONFIG_DEBUG_SLAB
3123 void *__kmalloc(size_t size
, gfp_t flags
)
3125 return __do_kmalloc(size
, flags
, NULL
);
3127 EXPORT_SYMBOL(__kmalloc
);
3131 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, void *caller
)
3133 return __do_kmalloc(size
, flags
, caller
);
3135 EXPORT_SYMBOL(__kmalloc_track_caller
);
3141 * __alloc_percpu - allocate one copy of the object for every present
3142 * cpu in the system, zeroing them.
3143 * Objects should be dereferenced using the per_cpu_ptr macro only.
3145 * @size: how many bytes of memory are required.
3147 void *__alloc_percpu(size_t size
)
3150 struct percpu_data
*pdata
= kmalloc(sizeof(*pdata
), GFP_KERNEL
);
3156 * Cannot use for_each_online_cpu since a cpu may come online
3157 * and we have no way of figuring out how to fix the array
3158 * that we have allocated then....
3161 int node
= cpu_to_node(i
);
3163 if (node_online(node
))
3164 pdata
->ptrs
[i
] = kmalloc_node(size
, GFP_KERNEL
, node
);
3166 pdata
->ptrs
[i
] = kmalloc(size
, GFP_KERNEL
);
3168 if (!pdata
->ptrs
[i
])
3170 memset(pdata
->ptrs
[i
], 0, size
);
3173 /* Catch derefs w/o wrappers */
3174 return (void *)(~(unsigned long)pdata
);
3178 if (!cpu_possible(i
))
3180 kfree(pdata
->ptrs
[i
]);
3185 EXPORT_SYMBOL(__alloc_percpu
);
3189 * kmem_cache_free - Deallocate an object
3190 * @cachep: The cache the allocation was from.
3191 * @objp: The previously allocated object.
3193 * Free an object which was previously allocated from this
3196 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3198 unsigned long flags
;
3200 local_irq_save(flags
);
3201 __cache_free(cachep
, objp
);
3202 local_irq_restore(flags
);
3204 EXPORT_SYMBOL(kmem_cache_free
);
3207 * kfree - free previously allocated memory
3208 * @objp: pointer returned by kmalloc.
3210 * If @objp is NULL, no operation is performed.
3212 * Don't free memory not originally allocated by kmalloc()
3213 * or you will run into trouble.
3215 void kfree(const void *objp
)
3217 struct kmem_cache
*c
;
3218 unsigned long flags
;
3220 if (unlikely(!objp
))
3222 local_irq_save(flags
);
3223 kfree_debugcheck(objp
);
3224 c
= virt_to_cache(objp
);
3225 mutex_debug_check_no_locks_freed(objp
, obj_size(c
));
3226 __cache_free(c
, (void *)objp
);
3227 local_irq_restore(flags
);
3229 EXPORT_SYMBOL(kfree
);
3233 * free_percpu - free previously allocated percpu memory
3234 * @objp: pointer returned by alloc_percpu.
3236 * Don't free memory not originally allocated by alloc_percpu()
3237 * The complemented objp is to check for that.
3239 void free_percpu(const void *objp
)
3242 struct percpu_data
*p
= (struct percpu_data
*)(~(unsigned long)objp
);
3245 * We allocate for all cpus so we cannot use for online cpu here.
3251 EXPORT_SYMBOL(free_percpu
);
3254 unsigned int kmem_cache_size(struct kmem_cache
*cachep
)
3256 return obj_size(cachep
);
3258 EXPORT_SYMBOL(kmem_cache_size
);
3260 const char *kmem_cache_name(struct kmem_cache
*cachep
)
3262 return cachep
->name
;
3264 EXPORT_SYMBOL_GPL(kmem_cache_name
);
3267 * This initializes kmem_list3 for all nodes.
3269 static int alloc_kmemlist(struct kmem_cache
*cachep
)
3272 struct kmem_list3
*l3
;
3275 for_each_online_node(node
) {
3276 struct array_cache
*nc
= NULL
, *new;
3277 struct array_cache
**new_alien
= NULL
;
3279 if (!(new_alien
= alloc_alien_cache(node
, cachep
->limit
)))
3282 if (!(new = alloc_arraycache(node
, (cachep
->shared
*
3283 cachep
->batchcount
),
3286 if ((l3
= cachep
->nodelists
[node
])) {
3288 spin_lock_irq(&l3
->list_lock
);
3290 if ((nc
= cachep
->nodelists
[node
]->shared
))
3291 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
3294 if (!cachep
->nodelists
[node
]->alien
) {
3295 l3
->alien
= new_alien
;
3298 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3299 cachep
->batchcount
+ cachep
->num
;
3300 spin_unlock_irq(&l3
->list_lock
);
3302 free_alien_cache(new_alien
);
3305 if (!(l3
= kmalloc_node(sizeof(struct kmem_list3
),
3309 kmem_list3_init(l3
);
3310 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
3311 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
3313 l3
->alien
= new_alien
;
3314 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3315 cachep
->batchcount
+ cachep
->num
;
3316 cachep
->nodelists
[node
] = l3
;
3324 struct ccupdate_struct
{
3325 struct kmem_cache
*cachep
;
3326 struct array_cache
*new[NR_CPUS
];
3329 static void do_ccupdate_local(void *info
)
3331 struct ccupdate_struct
*new = (struct ccupdate_struct
*)info
;
3332 struct array_cache
*old
;
3335 old
= cpu_cache_get(new->cachep
);
3337 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
3338 new->new[smp_processor_id()] = old
;
3341 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
, int batchcount
,
3344 struct ccupdate_struct
new;
3347 memset(&new.new, 0, sizeof(new.new));
3348 for_each_online_cpu(i
) {
3350 alloc_arraycache(cpu_to_node(i
), limit
, batchcount
);
3352 for (i
--; i
>= 0; i
--)
3357 new.cachep
= cachep
;
3359 smp_call_function_all_cpus(do_ccupdate_local
, (void *)&new);
3362 spin_lock(&cachep
->spinlock
);
3363 cachep
->batchcount
= batchcount
;
3364 cachep
->limit
= limit
;
3365 cachep
->shared
= shared
;
3366 spin_unlock(&cachep
->spinlock
);
3368 for_each_online_cpu(i
) {
3369 struct array_cache
*ccold
= new.new[i
];
3372 spin_lock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3373 free_block(cachep
, ccold
->entry
, ccold
->avail
, cpu_to_node(i
));
3374 spin_unlock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3378 err
= alloc_kmemlist(cachep
);
3380 printk(KERN_ERR
"alloc_kmemlist failed for %s, error %d.\n",
3381 cachep
->name
, -err
);
3387 static void enable_cpucache(struct kmem_cache
*cachep
)
3392 /* The head array serves three purposes:
3393 * - create a LIFO ordering, i.e. return objects that are cache-warm
3394 * - reduce the number of spinlock operations.
3395 * - reduce the number of linked list operations on the slab and
3396 * bufctl chains: array operations are cheaper.
3397 * The numbers are guessed, we should auto-tune as described by
3400 if (cachep
->buffer_size
> 131072)
3402 else if (cachep
->buffer_size
> PAGE_SIZE
)
3404 else if (cachep
->buffer_size
> 1024)
3406 else if (cachep
->buffer_size
> 256)
3411 /* Cpu bound tasks (e.g. network routing) can exhibit cpu bound
3412 * allocation behaviour: Most allocs on one cpu, most free operations
3413 * on another cpu. For these cases, an efficient object passing between
3414 * cpus is necessary. This is provided by a shared array. The array
3415 * replaces Bonwick's magazine layer.
3416 * On uniprocessor, it's functionally equivalent (but less efficient)
3417 * to a larger limit. Thus disabled by default.
3421 if (cachep
->buffer_size
<= PAGE_SIZE
)
3426 /* With debugging enabled, large batchcount lead to excessively
3427 * long periods with disabled local interrupts. Limit the
3433 err
= do_tune_cpucache(cachep
, limit
, (limit
+ 1) / 2, shared
);
3435 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
3436 cachep
->name
, -err
);
3439 static void drain_array_locked(struct kmem_cache
*cachep
, struct array_cache
*ac
,
3440 int force
, int node
)
3444 check_spinlock_acquired_node(cachep
, node
);
3445 if (ac
->touched
&& !force
) {
3447 } else if (ac
->avail
) {
3448 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
3449 if (tofree
> ac
->avail
) {
3450 tofree
= (ac
->avail
+ 1) / 2;
3452 free_block(cachep
, ac
->entry
, tofree
, node
);
3453 ac
->avail
-= tofree
;
3454 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
3455 sizeof(void *) * ac
->avail
);
3460 * cache_reap - Reclaim memory from caches.
3461 * @unused: unused parameter
3463 * Called from workqueue/eventd every few seconds.
3465 * - clear the per-cpu caches for this CPU.
3466 * - return freeable pages to the main free memory pool.
3468 * If we cannot acquire the cache chain mutex then just give up - we'll
3469 * try again on the next iteration.
3471 static void cache_reap(void *unused
)
3473 struct list_head
*walk
;
3474 struct kmem_list3
*l3
;
3476 if (!mutex_trylock(&cache_chain_mutex
)) {
3477 /* Give up. Setup the next iteration. */
3478 schedule_delayed_work(&__get_cpu_var(reap_work
),
3483 list_for_each(walk
, &cache_chain
) {
3484 struct kmem_cache
*searchp
;
3485 struct list_head
*p
;
3489 searchp
= list_entry(walk
, struct kmem_cache
, next
);
3491 if (searchp
->flags
& SLAB_NO_REAP
)
3496 l3
= searchp
->nodelists
[numa_node_id()];
3498 drain_alien_cache(searchp
, l3
->alien
);
3499 spin_lock_irq(&l3
->list_lock
);
3501 drain_array_locked(searchp
, cpu_cache_get(searchp
), 0,
3504 if (time_after(l3
->next_reap
, jiffies
))
3507 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
;
3510 drain_array_locked(searchp
, l3
->shared
, 0,
3513 if (l3
->free_touched
) {
3514 l3
->free_touched
= 0;
3519 (l3
->free_limit
+ 5 * searchp
->num
-
3520 1) / (5 * searchp
->num
);
3522 p
= l3
->slabs_free
.next
;
3523 if (p
== &(l3
->slabs_free
))
3526 slabp
= list_entry(p
, struct slab
, list
);
3527 BUG_ON(slabp
->inuse
);
3528 list_del(&slabp
->list
);
3529 STATS_INC_REAPED(searchp
);
3531 /* Safe to drop the lock. The slab is no longer
3532 * linked to the cache.
3533 * searchp cannot disappear, we hold
3536 l3
->free_objects
-= searchp
->num
;
3537 spin_unlock_irq(&l3
->list_lock
);
3538 slab_destroy(searchp
, slabp
);
3539 spin_lock_irq(&l3
->list_lock
);
3540 } while (--tofree
> 0);
3542 spin_unlock_irq(&l3
->list_lock
);
3547 mutex_unlock(&cache_chain_mutex
);
3548 drain_remote_pages();
3549 /* Setup the next iteration */
3550 schedule_delayed_work(&__get_cpu_var(reap_work
), REAPTIMEOUT_CPUC
);
3553 #ifdef CONFIG_PROC_FS
3555 static void print_slabinfo_header(struct seq_file
*m
)
3558 * Output format version, so at least we can change it
3559 * without _too_ many complaints.
3562 seq_puts(m
, "slabinfo - version: 2.1 (statistics)\n");
3564 seq_puts(m
, "slabinfo - version: 2.1\n");
3566 seq_puts(m
, "# name <active_objs> <num_objs> <objsize> "
3567 "<objperslab> <pagesperslab>");
3568 seq_puts(m
, " : tunables <limit> <batchcount> <sharedfactor>");
3569 seq_puts(m
, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3571 seq_puts(m
, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3572 "<error> <maxfreeable> <nodeallocs> <remotefrees>");
3573 seq_puts(m
, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3578 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
3581 struct list_head
*p
;
3583 mutex_lock(&cache_chain_mutex
);
3585 print_slabinfo_header(m
);
3586 p
= cache_chain
.next
;
3589 if (p
== &cache_chain
)
3592 return list_entry(p
, struct kmem_cache
, next
);
3595 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
3597 struct kmem_cache
*cachep
= p
;
3599 return cachep
->next
.next
== &cache_chain
? NULL
3600 : list_entry(cachep
->next
.next
, struct kmem_cache
, next
);
3603 static void s_stop(struct seq_file
*m
, void *p
)
3605 mutex_unlock(&cache_chain_mutex
);
3608 static int s_show(struct seq_file
*m
, void *p
)
3610 struct kmem_cache
*cachep
= p
;
3611 struct list_head
*q
;
3613 unsigned long active_objs
;
3614 unsigned long num_objs
;
3615 unsigned long active_slabs
= 0;
3616 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
3620 struct kmem_list3
*l3
;
3622 spin_lock(&cachep
->spinlock
);
3625 for_each_online_node(node
) {
3626 l3
= cachep
->nodelists
[node
];
3631 spin_lock_irq(&l3
->list_lock
);
3633 list_for_each(q
, &l3
->slabs_full
) {
3634 slabp
= list_entry(q
, struct slab
, list
);
3635 if (slabp
->inuse
!= cachep
->num
&& !error
)
3636 error
= "slabs_full accounting error";
3637 active_objs
+= cachep
->num
;
3640 list_for_each(q
, &l3
->slabs_partial
) {
3641 slabp
= list_entry(q
, struct slab
, list
);
3642 if (slabp
->inuse
== cachep
->num
&& !error
)
3643 error
= "slabs_partial inuse accounting error";
3644 if (!slabp
->inuse
&& !error
)
3645 error
= "slabs_partial/inuse accounting error";
3646 active_objs
+= slabp
->inuse
;
3649 list_for_each(q
, &l3
->slabs_free
) {
3650 slabp
= list_entry(q
, struct slab
, list
);
3651 if (slabp
->inuse
&& !error
)
3652 error
= "slabs_free/inuse accounting error";
3655 free_objects
+= l3
->free_objects
;
3657 shared_avail
+= l3
->shared
->avail
;
3659 spin_unlock_irq(&l3
->list_lock
);
3661 num_slabs
+= active_slabs
;
3662 num_objs
= num_slabs
* cachep
->num
;
3663 if (num_objs
- active_objs
!= free_objects
&& !error
)
3664 error
= "free_objects accounting error";
3666 name
= cachep
->name
;
3668 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
3670 seq_printf(m
, "%-17s %6lu %6lu %6u %4u %4d",
3671 name
, active_objs
, num_objs
, cachep
->buffer_size
,
3672 cachep
->num
, (1 << cachep
->gfporder
));
3673 seq_printf(m
, " : tunables %4u %4u %4u",
3674 cachep
->limit
, cachep
->batchcount
, cachep
->shared
);
3675 seq_printf(m
, " : slabdata %6lu %6lu %6lu",
3676 active_slabs
, num_slabs
, shared_avail
);
3679 unsigned long high
= cachep
->high_mark
;
3680 unsigned long allocs
= cachep
->num_allocations
;
3681 unsigned long grown
= cachep
->grown
;
3682 unsigned long reaped
= cachep
->reaped
;
3683 unsigned long errors
= cachep
->errors
;
3684 unsigned long max_freeable
= cachep
->max_freeable
;
3685 unsigned long node_allocs
= cachep
->node_allocs
;
3686 unsigned long node_frees
= cachep
->node_frees
;
3688 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu \
3689 %4lu %4lu %4lu %4lu", allocs
, high
, grown
, reaped
, errors
, max_freeable
, node_allocs
, node_frees
);
3693 unsigned long allochit
= atomic_read(&cachep
->allochit
);
3694 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
3695 unsigned long freehit
= atomic_read(&cachep
->freehit
);
3696 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
3698 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
3699 allochit
, allocmiss
, freehit
, freemiss
);
3703 spin_unlock(&cachep
->spinlock
);
3708 * slabinfo_op - iterator that generates /proc/slabinfo
3717 * num-pages-per-slab
3718 * + further values on SMP and with statistics enabled
3721 struct seq_operations slabinfo_op
= {
3728 #define MAX_SLABINFO_WRITE 128
3730 * slabinfo_write - Tuning for the slab allocator
3732 * @buffer: user buffer
3733 * @count: data length
3736 ssize_t
slabinfo_write(struct file
*file
, const char __user
* buffer
,
3737 size_t count
, loff_t
*ppos
)
3739 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
3740 int limit
, batchcount
, shared
, res
;
3741 struct list_head
*p
;
3743 if (count
> MAX_SLABINFO_WRITE
)
3745 if (copy_from_user(&kbuf
, buffer
, count
))
3747 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
3749 tmp
= strchr(kbuf
, ' ');
3754 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
3757 /* Find the cache in the chain of caches. */
3758 mutex_lock(&cache_chain_mutex
);
3760 list_for_each(p
, &cache_chain
) {
3761 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
,
3764 if (!strcmp(cachep
->name
, kbuf
)) {
3767 batchcount
> limit
|| shared
< 0) {
3770 res
= do_tune_cpucache(cachep
, limit
,
3771 batchcount
, shared
);
3776 mutex_unlock(&cache_chain_mutex
);
3784 * ksize - get the actual amount of memory allocated for a given object
3785 * @objp: Pointer to the object
3787 * kmalloc may internally round up allocations and return more memory
3788 * than requested. ksize() can be used to determine the actual amount of
3789 * memory allocated. The caller may use this additional memory, even though
3790 * a smaller amount of memory was initially specified with the kmalloc call.
3791 * The caller must guarantee that objp points to a valid object previously
3792 * allocated with either kmalloc() or kmem_cache_alloc(). The object
3793 * must not be freed during the duration of the call.
3795 unsigned int ksize(const void *objp
)
3797 if (unlikely(objp
== NULL
))
3800 return obj_size(virt_to_cache(objp
));