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
;
1129 for (i
= 0; i
< NUM_INIT_LISTS
; i
++) {
1130 kmem_list3_init(&initkmem_list3
[i
]);
1131 if (i
< MAX_NUMNODES
)
1132 cache_cache
.nodelists
[i
] = NULL
;
1136 * Fragmentation resistance on low memory - only use bigger
1137 * page orders on machines with more than 32MB of memory.
1139 if (num_physpages
> (32 << 20) >> PAGE_SHIFT
)
1140 slab_break_gfp_order
= BREAK_GFP_ORDER_HI
;
1142 /* Bootstrap is tricky, because several objects are allocated
1143 * from caches that do not exist yet:
1144 * 1) initialize the cache_cache cache: it contains the struct kmem_cache
1145 * structures of all caches, except cache_cache itself: cache_cache
1146 * is statically allocated.
1147 * Initially an __init data area is used for the head array and the
1148 * kmem_list3 structures, it's replaced with a kmalloc allocated
1149 * array at the end of the bootstrap.
1150 * 2) Create the first kmalloc cache.
1151 * The struct kmem_cache for the new cache is allocated normally.
1152 * An __init data area is used for the head array.
1153 * 3) Create the remaining kmalloc caches, with minimally sized
1155 * 4) Replace the __init data head arrays for cache_cache and the first
1156 * kmalloc cache with kmalloc allocated arrays.
1157 * 5) Replace the __init data for kmem_list3 for cache_cache and
1158 * the other cache's with kmalloc allocated memory.
1159 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1162 /* 1) create the cache_cache */
1163 INIT_LIST_HEAD(&cache_chain
);
1164 list_add(&cache_cache
.next
, &cache_chain
);
1165 cache_cache
.colour_off
= cache_line_size();
1166 cache_cache
.array
[smp_processor_id()] = &initarray_cache
.cache
;
1167 cache_cache
.nodelists
[numa_node_id()] = &initkmem_list3
[CACHE_CACHE
];
1169 cache_cache
.buffer_size
= ALIGN(cache_cache
.buffer_size
, cache_line_size());
1171 for (order
= 0; order
< MAX_ORDER
; order
++) {
1172 cache_estimate(order
, cache_cache
.buffer_size
,
1173 cache_line_size(), 0, &left_over
, &cache_cache
.num
);
1174 if (cache_cache
.num
)
1177 if (!cache_cache
.num
)
1179 cache_cache
.gfporder
= order
;
1180 cache_cache
.colour
= left_over
/ cache_cache
.colour_off
;
1181 cache_cache
.slab_size
= ALIGN(cache_cache
.num
* sizeof(kmem_bufctl_t
) +
1182 sizeof(struct slab
), cache_line_size());
1184 /* 2+3) create the kmalloc caches */
1185 sizes
= malloc_sizes
;
1186 names
= cache_names
;
1188 /* Initialize the caches that provide memory for the array cache
1189 * and the kmem_list3 structures first.
1190 * Without this, further allocations will bug
1193 sizes
[INDEX_AC
].cs_cachep
= kmem_cache_create(names
[INDEX_AC
].name
,
1194 sizes
[INDEX_AC
].cs_size
,
1195 ARCH_KMALLOC_MINALIGN
,
1196 (ARCH_KMALLOC_FLAGS
|
1197 SLAB_PANIC
), NULL
, NULL
);
1199 if (INDEX_AC
!= INDEX_L3
)
1200 sizes
[INDEX_L3
].cs_cachep
=
1201 kmem_cache_create(names
[INDEX_L3
].name
,
1202 sizes
[INDEX_L3
].cs_size
,
1203 ARCH_KMALLOC_MINALIGN
,
1204 (ARCH_KMALLOC_FLAGS
| SLAB_PANIC
), NULL
,
1207 while (sizes
->cs_size
!= ULONG_MAX
) {
1209 * For performance, all the general caches are L1 aligned.
1210 * This should be particularly beneficial on SMP boxes, as it
1211 * eliminates "false sharing".
1212 * Note for systems short on memory removing the alignment will
1213 * allow tighter packing of the smaller caches.
1215 if (!sizes
->cs_cachep
)
1216 sizes
->cs_cachep
= kmem_cache_create(names
->name
,
1218 ARCH_KMALLOC_MINALIGN
,
1223 /* Inc off-slab bufctl limit until the ceiling is hit. */
1224 if (!(OFF_SLAB(sizes
->cs_cachep
))) {
1225 offslab_limit
= sizes
->cs_size
- sizeof(struct slab
);
1226 offslab_limit
/= sizeof(kmem_bufctl_t
);
1229 sizes
->cs_dmacachep
= kmem_cache_create(names
->name_dma
,
1231 ARCH_KMALLOC_MINALIGN
,
1232 (ARCH_KMALLOC_FLAGS
|
1240 /* 4) Replace the bootstrap head arrays */
1244 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1246 local_irq_disable();
1247 BUG_ON(cpu_cache_get(&cache_cache
) != &initarray_cache
.cache
);
1248 memcpy(ptr
, cpu_cache_get(&cache_cache
),
1249 sizeof(struct arraycache_init
));
1250 cache_cache
.array
[smp_processor_id()] = ptr
;
1253 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1255 local_irq_disable();
1256 BUG_ON(cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
)
1257 != &initarray_generic
.cache
);
1258 memcpy(ptr
, cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
),
1259 sizeof(struct arraycache_init
));
1260 malloc_sizes
[INDEX_AC
].cs_cachep
->array
[smp_processor_id()] =
1264 /* 5) Replace the bootstrap kmem_list3's */
1267 /* Replace the static kmem_list3 structures for the boot cpu */
1268 init_list(&cache_cache
, &initkmem_list3
[CACHE_CACHE
],
1271 for_each_online_node(node
) {
1272 init_list(malloc_sizes
[INDEX_AC
].cs_cachep
,
1273 &initkmem_list3
[SIZE_AC
+ node
], node
);
1275 if (INDEX_AC
!= INDEX_L3
) {
1276 init_list(malloc_sizes
[INDEX_L3
].cs_cachep
,
1277 &initkmem_list3
[SIZE_L3
+ node
],
1283 /* 6) resize the head arrays to their final sizes */
1285 struct kmem_cache
*cachep
;
1286 mutex_lock(&cache_chain_mutex
);
1287 list_for_each_entry(cachep
, &cache_chain
, next
)
1288 enable_cpucache(cachep
);
1289 mutex_unlock(&cache_chain_mutex
);
1293 g_cpucache_up
= FULL
;
1295 /* Register a cpu startup notifier callback
1296 * that initializes cpu_cache_get for all new cpus
1298 register_cpu_notifier(&cpucache_notifier
);
1300 /* The reap timers are started later, with a module init call:
1301 * That part of the kernel is not yet operational.
1305 static int __init
cpucache_init(void)
1310 * Register the timers that return unneeded
1313 for_each_online_cpu(cpu
)
1314 start_cpu_timer(cpu
);
1319 __initcall(cpucache_init
);
1322 * Interface to system's page allocator. No need to hold the cache-lock.
1324 * If we requested dmaable memory, we will get it. Even if we
1325 * did not request dmaable memory, we might get it, but that
1326 * would be relatively rare and ignorable.
1328 static void *kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
1334 flags
|= cachep
->gfpflags
;
1335 page
= alloc_pages_node(nodeid
, flags
, cachep
->gfporder
);
1338 addr
= page_address(page
);
1340 i
= (1 << cachep
->gfporder
);
1341 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1342 atomic_add(i
, &slab_reclaim_pages
);
1343 add_page_state(nr_slab
, i
);
1352 * Interface to system's page release.
1354 static void kmem_freepages(struct kmem_cache
*cachep
, void *addr
)
1356 unsigned long i
= (1 << cachep
->gfporder
);
1357 struct page
*page
= virt_to_page(addr
);
1358 const unsigned long nr_freed
= i
;
1361 if (!TestClearPageSlab(page
))
1365 sub_page_state(nr_slab
, nr_freed
);
1366 if (current
->reclaim_state
)
1367 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1368 free_pages((unsigned long)addr
, cachep
->gfporder
);
1369 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1370 atomic_sub(1 << cachep
->gfporder
, &slab_reclaim_pages
);
1373 static void kmem_rcu_free(struct rcu_head
*head
)
1375 struct slab_rcu
*slab_rcu
= (struct slab_rcu
*)head
;
1376 struct kmem_cache
*cachep
= slab_rcu
->cachep
;
1378 kmem_freepages(cachep
, slab_rcu
->addr
);
1379 if (OFF_SLAB(cachep
))
1380 kmem_cache_free(cachep
->slabp_cache
, slab_rcu
);
1385 #ifdef CONFIG_DEBUG_PAGEALLOC
1386 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1387 unsigned long caller
)
1389 int size
= obj_size(cachep
);
1391 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1393 if (size
< 5 * sizeof(unsigned long))
1396 *addr
++ = 0x12345678;
1398 *addr
++ = smp_processor_id();
1399 size
-= 3 * sizeof(unsigned long);
1401 unsigned long *sptr
= &caller
;
1402 unsigned long svalue
;
1404 while (!kstack_end(sptr
)) {
1406 if (kernel_text_address(svalue
)) {
1408 size
-= sizeof(unsigned long);
1409 if (size
<= sizeof(unsigned long))
1415 *addr
++ = 0x87654321;
1419 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1421 int size
= obj_size(cachep
);
1422 addr
= &((char *)addr
)[obj_offset(cachep
)];
1424 memset(addr
, val
, size
);
1425 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1428 static void dump_line(char *data
, int offset
, int limit
)
1431 printk(KERN_ERR
"%03x:", offset
);
1432 for (i
= 0; i
< limit
; i
++) {
1433 printk(" %02x", (unsigned char)data
[offset
+ i
]);
1441 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1446 if (cachep
->flags
& SLAB_RED_ZONE
) {
1447 printk(KERN_ERR
"Redzone: 0x%lx/0x%lx.\n",
1448 *dbg_redzone1(cachep
, objp
),
1449 *dbg_redzone2(cachep
, objp
));
1452 if (cachep
->flags
& SLAB_STORE_USER
) {
1453 printk(KERN_ERR
"Last user: [<%p>]",
1454 *dbg_userword(cachep
, objp
));
1455 print_symbol("(%s)",
1456 (unsigned long)*dbg_userword(cachep
, objp
));
1459 realobj
= (char *)objp
+ obj_offset(cachep
);
1460 size
= obj_size(cachep
);
1461 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1464 if (i
+ limit
> size
)
1466 dump_line(realobj
, i
, limit
);
1470 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1476 realobj
= (char *)objp
+ obj_offset(cachep
);
1477 size
= obj_size(cachep
);
1479 for (i
= 0; i
< size
; i
++) {
1480 char exp
= POISON_FREE
;
1483 if (realobj
[i
] != exp
) {
1489 "Slab corruption: start=%p, len=%d\n",
1491 print_objinfo(cachep
, objp
, 0);
1493 /* Hexdump the affected line */
1496 if (i
+ limit
> size
)
1498 dump_line(realobj
, i
, limit
);
1501 /* Limit to 5 lines */
1507 /* Print some data about the neighboring objects, if they
1510 struct slab
*slabp
= virt_to_slab(objp
);
1513 objnr
= (unsigned)(objp
- slabp
->s_mem
) / cachep
->buffer_size
;
1515 objp
= slabp
->s_mem
+ (objnr
- 1) * cachep
->buffer_size
;
1516 realobj
= (char *)objp
+ obj_offset(cachep
);
1517 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
1519 print_objinfo(cachep
, objp
, 2);
1521 if (objnr
+ 1 < cachep
->num
) {
1522 objp
= slabp
->s_mem
+ (objnr
+ 1) * cachep
->buffer_size
;
1523 realobj
= (char *)objp
+ obj_offset(cachep
);
1524 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
1526 print_objinfo(cachep
, objp
, 2);
1534 * slab_destroy_objs - call the registered destructor for each object in
1535 * a slab that is to be destroyed.
1537 static void slab_destroy_objs(struct kmem_cache
*cachep
, struct slab
*slabp
)
1540 for (i
= 0; i
< cachep
->num
; i
++) {
1541 void *objp
= slabp
->s_mem
+ cachep
->buffer_size
* i
;
1543 if (cachep
->flags
& SLAB_POISON
) {
1544 #ifdef CONFIG_DEBUG_PAGEALLOC
1545 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0
1546 && OFF_SLAB(cachep
))
1547 kernel_map_pages(virt_to_page(objp
),
1548 cachep
->buffer_size
/ PAGE_SIZE
,
1551 check_poison_obj(cachep
, objp
);
1553 check_poison_obj(cachep
, objp
);
1556 if (cachep
->flags
& SLAB_RED_ZONE
) {
1557 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1558 slab_error(cachep
, "start of a freed object "
1560 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1561 slab_error(cachep
, "end of a freed object "
1564 if (cachep
->dtor
&& !(cachep
->flags
& SLAB_POISON
))
1565 (cachep
->dtor
) (objp
+ obj_offset(cachep
), cachep
, 0);
1569 static void slab_destroy_objs(struct kmem_cache
*cachep
, struct slab
*slabp
)
1573 for (i
= 0; i
< cachep
->num
; i
++) {
1574 void *objp
= slabp
->s_mem
+ cachep
->buffer_size
* i
;
1575 (cachep
->dtor
) (objp
, cachep
, 0);
1582 * Destroy all the objs in a slab, and release the mem back to the system.
1583 * Before calling the slab must have been unlinked from the cache.
1584 * The cache-lock is not held/needed.
1586 static void slab_destroy(struct kmem_cache
*cachep
, struct slab
*slabp
)
1588 void *addr
= slabp
->s_mem
- slabp
->colouroff
;
1590 slab_destroy_objs(cachep
, slabp
);
1591 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
1592 struct slab_rcu
*slab_rcu
;
1594 slab_rcu
= (struct slab_rcu
*)slabp
;
1595 slab_rcu
->cachep
= cachep
;
1596 slab_rcu
->addr
= addr
;
1597 call_rcu(&slab_rcu
->head
, kmem_rcu_free
);
1599 kmem_freepages(cachep
, addr
);
1600 if (OFF_SLAB(cachep
))
1601 kmem_cache_free(cachep
->slabp_cache
, slabp
);
1605 /* For setting up all the kmem_list3s for cache whose buffer_size is same
1606 as size of kmem_list3. */
1607 static void set_up_list3s(struct kmem_cache
*cachep
, int index
)
1611 for_each_online_node(node
) {
1612 cachep
->nodelists
[node
] = &initkmem_list3
[index
+ node
];
1613 cachep
->nodelists
[node
]->next_reap
= jiffies
+
1615 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1620 * calculate_slab_order - calculate size (page order) of slabs
1621 * @cachep: pointer to the cache that is being created
1622 * @size: size of objects to be created in this cache.
1623 * @align: required alignment for the objects.
1624 * @flags: slab allocation flags
1626 * Also calculates the number of objects per slab.
1628 * This could be made much more intelligent. For now, try to avoid using
1629 * high order pages for slabs. When the gfp() functions are more friendly
1630 * towards high-order requests, this should be changed.
1632 static inline size_t calculate_slab_order(struct kmem_cache
*cachep
,
1633 size_t size
, size_t align
, unsigned long flags
)
1635 size_t left_over
= 0;
1638 for (gfporder
= 0 ; gfporder
<= MAX_GFP_ORDER
; gfporder
++) {
1642 cache_estimate(gfporder
, size
, align
, flags
, &remainder
, &num
);
1646 /* More than offslab_limit objects will cause problems */
1647 if ((flags
& CFLGS_OFF_SLAB
) && num
> offslab_limit
)
1650 /* Found something acceptable - save it away */
1652 cachep
->gfporder
= gfporder
;
1653 left_over
= remainder
;
1656 * A VFS-reclaimable slab tends to have most allocations
1657 * as GFP_NOFS and we really don't want to have to be allocating
1658 * higher-order pages when we are unable to shrink dcache.
1660 if (flags
& SLAB_RECLAIM_ACCOUNT
)
1664 * Large number of objects is good, but very large slabs are
1665 * currently bad for the gfp()s.
1667 if (gfporder
>= slab_break_gfp_order
)
1671 * Acceptable internal fragmentation?
1673 if ((left_over
* 8) <= (PAGE_SIZE
<< gfporder
))
1680 * kmem_cache_create - Create a cache.
1681 * @name: A string which is used in /proc/slabinfo to identify this cache.
1682 * @size: The size of objects to be created in this cache.
1683 * @align: The required alignment for the objects.
1684 * @flags: SLAB flags
1685 * @ctor: A constructor for the objects.
1686 * @dtor: A destructor for the objects.
1688 * Returns a ptr to the cache on success, NULL on failure.
1689 * Cannot be called within a int, but can be interrupted.
1690 * The @ctor is run when new pages are allocated by the cache
1691 * and the @dtor is run before the pages are handed back.
1693 * @name must be valid until the cache is destroyed. This implies that
1694 * the module calling this has to destroy the cache before getting
1699 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1700 * to catch references to uninitialised memory.
1702 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1703 * for buffer overruns.
1705 * %SLAB_NO_REAP - Don't automatically reap this cache when we're under
1708 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1709 * cacheline. This can be beneficial if you're counting cycles as closely
1713 kmem_cache_create (const char *name
, size_t size
, size_t align
,
1714 unsigned long flags
, void (*ctor
)(void*, struct kmem_cache
*, unsigned long),
1715 void (*dtor
)(void*, struct kmem_cache
*, unsigned long))
1717 size_t left_over
, slab_size
, ralign
;
1718 struct kmem_cache
*cachep
= NULL
;
1719 struct list_head
*p
;
1722 * Sanity checks... these are all serious usage bugs.
1726 (size
< BYTES_PER_WORD
) ||
1727 (size
> (1 << MAX_OBJ_ORDER
) * PAGE_SIZE
) || (dtor
&& !ctor
)) {
1728 printk(KERN_ERR
"%s: Early error in slab %s\n",
1729 __FUNCTION__
, name
);
1734 * Prevent CPUs from coming and going.
1735 * lock_cpu_hotplug() nests outside cache_chain_mutex
1739 mutex_lock(&cache_chain_mutex
);
1741 list_for_each(p
, &cache_chain
) {
1742 struct kmem_cache
*pc
= list_entry(p
, struct kmem_cache
, next
);
1743 mm_segment_t old_fs
= get_fs();
1748 * This happens when the module gets unloaded and doesn't
1749 * destroy its slab cache and no-one else reuses the vmalloc
1750 * area of the module. Print a warning.
1753 res
= __get_user(tmp
, pc
->name
);
1756 printk("SLAB: cache with size %d has lost its name\n",
1761 if (!strcmp(pc
->name
, name
)) {
1762 printk("kmem_cache_create: duplicate cache %s\n", name
);
1769 WARN_ON(strchr(name
, ' ')); /* It confuses parsers */
1770 if ((flags
& SLAB_DEBUG_INITIAL
) && !ctor
) {
1771 /* No constructor, but inital state check requested */
1772 printk(KERN_ERR
"%s: No con, but init state check "
1773 "requested - %s\n", __FUNCTION__
, name
);
1774 flags
&= ~SLAB_DEBUG_INITIAL
;
1778 * Enable redzoning and last user accounting, except for caches with
1779 * large objects, if the increased size would increase the object size
1780 * above the next power of two: caches with object sizes just above a
1781 * power of two have a significant amount of internal fragmentation.
1784 || fls(size
- 1) == fls(size
- 1 + 3 * BYTES_PER_WORD
)))
1785 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
1786 if (!(flags
& SLAB_DESTROY_BY_RCU
))
1787 flags
|= SLAB_POISON
;
1789 if (flags
& SLAB_DESTROY_BY_RCU
)
1790 BUG_ON(flags
& SLAB_POISON
);
1792 if (flags
& SLAB_DESTROY_BY_RCU
)
1796 * Always checks flags, a caller might be expecting debug
1797 * support which isn't available.
1799 if (flags
& ~CREATE_MASK
)
1802 /* Check that size is in terms of words. This is needed to avoid
1803 * unaligned accesses for some archs when redzoning is used, and makes
1804 * sure any on-slab bufctl's are also correctly aligned.
1806 if (size
& (BYTES_PER_WORD
- 1)) {
1807 size
+= (BYTES_PER_WORD
- 1);
1808 size
&= ~(BYTES_PER_WORD
- 1);
1811 /* calculate out the final buffer alignment: */
1812 /* 1) arch recommendation: can be overridden for debug */
1813 if (flags
& SLAB_HWCACHE_ALIGN
) {
1814 /* Default alignment: as specified by the arch code.
1815 * Except if an object is really small, then squeeze multiple
1816 * objects into one cacheline.
1818 ralign
= cache_line_size();
1819 while (size
<= ralign
/ 2)
1822 ralign
= BYTES_PER_WORD
;
1824 /* 2) arch mandated alignment: disables debug if necessary */
1825 if (ralign
< ARCH_SLAB_MINALIGN
) {
1826 ralign
= ARCH_SLAB_MINALIGN
;
1827 if (ralign
> BYTES_PER_WORD
)
1828 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
1830 /* 3) caller mandated alignment: disables debug if necessary */
1831 if (ralign
< align
) {
1833 if (ralign
> BYTES_PER_WORD
)
1834 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
1836 /* 4) Store it. Note that the debug code below can reduce
1837 * the alignment to BYTES_PER_WORD.
1841 /* Get cache's description obj. */
1842 cachep
= kmem_cache_alloc(&cache_cache
, SLAB_KERNEL
);
1845 memset(cachep
, 0, sizeof(struct kmem_cache
));
1848 cachep
->obj_size
= size
;
1850 if (flags
& SLAB_RED_ZONE
) {
1851 /* redzoning only works with word aligned caches */
1852 align
= BYTES_PER_WORD
;
1854 /* add space for red zone words */
1855 cachep
->obj_offset
+= BYTES_PER_WORD
;
1856 size
+= 2 * BYTES_PER_WORD
;
1858 if (flags
& SLAB_STORE_USER
) {
1859 /* user store requires word alignment and
1860 * one word storage behind the end of the real
1863 align
= BYTES_PER_WORD
;
1864 size
+= BYTES_PER_WORD
;
1866 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
1867 if (size
>= malloc_sizes
[INDEX_L3
+ 1].cs_size
1868 && cachep
->obj_size
> cache_line_size() && size
< PAGE_SIZE
) {
1869 cachep
->obj_offset
+= PAGE_SIZE
- size
;
1875 /* Determine if the slab management is 'on' or 'off' slab. */
1876 if (size
>= (PAGE_SIZE
>> 3))
1878 * Size is large, assume best to place the slab management obj
1879 * off-slab (should allow better packing of objs).
1881 flags
|= CFLGS_OFF_SLAB
;
1883 size
= ALIGN(size
, align
);
1885 left_over
= calculate_slab_order(cachep
, size
, align
, flags
);
1888 printk("kmem_cache_create: couldn't create cache %s.\n", name
);
1889 kmem_cache_free(&cache_cache
, cachep
);
1893 slab_size
= ALIGN(cachep
->num
* sizeof(kmem_bufctl_t
)
1894 + sizeof(struct slab
), align
);
1897 * If the slab has been placed off-slab, and we have enough space then
1898 * move it on-slab. This is at the expense of any extra colouring.
1900 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= slab_size
) {
1901 flags
&= ~CFLGS_OFF_SLAB
;
1902 left_over
-= slab_size
;
1905 if (flags
& CFLGS_OFF_SLAB
) {
1906 /* really off slab. No need for manual alignment */
1908 cachep
->num
* sizeof(kmem_bufctl_t
) + sizeof(struct slab
);
1911 cachep
->colour_off
= cache_line_size();
1912 /* Offset must be a multiple of the alignment. */
1913 if (cachep
->colour_off
< align
)
1914 cachep
->colour_off
= align
;
1915 cachep
->colour
= left_over
/ cachep
->colour_off
;
1916 cachep
->slab_size
= slab_size
;
1917 cachep
->flags
= flags
;
1918 cachep
->gfpflags
= 0;
1919 if (flags
& SLAB_CACHE_DMA
)
1920 cachep
->gfpflags
|= GFP_DMA
;
1921 spin_lock_init(&cachep
->spinlock
);
1922 cachep
->buffer_size
= size
;
1924 if (flags
& CFLGS_OFF_SLAB
)
1925 cachep
->slabp_cache
= kmem_find_general_cachep(slab_size
, 0u);
1926 cachep
->ctor
= ctor
;
1927 cachep
->dtor
= dtor
;
1928 cachep
->name
= name
;
1931 if (g_cpucache_up
== FULL
) {
1932 enable_cpucache(cachep
);
1934 if (g_cpucache_up
== NONE
) {
1935 /* Note: the first kmem_cache_create must create
1936 * the cache that's used by kmalloc(24), otherwise
1937 * the creation of further caches will BUG().
1939 cachep
->array
[smp_processor_id()] =
1940 &initarray_generic
.cache
;
1942 /* If the cache that's used by
1943 * kmalloc(sizeof(kmem_list3)) is the first cache,
1944 * then we need to set up all its list3s, otherwise
1945 * the creation of further caches will BUG().
1947 set_up_list3s(cachep
, SIZE_AC
);
1948 if (INDEX_AC
== INDEX_L3
)
1949 g_cpucache_up
= PARTIAL_L3
;
1951 g_cpucache_up
= PARTIAL_AC
;
1953 cachep
->array
[smp_processor_id()] =
1954 kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1956 if (g_cpucache_up
== PARTIAL_AC
) {
1957 set_up_list3s(cachep
, SIZE_L3
);
1958 g_cpucache_up
= PARTIAL_L3
;
1961 for_each_online_node(node
) {
1963 cachep
->nodelists
[node
] =
1965 (struct kmem_list3
),
1967 BUG_ON(!cachep
->nodelists
[node
]);
1968 kmem_list3_init(cachep
->
1973 cachep
->nodelists
[numa_node_id()]->next_reap
=
1974 jiffies
+ REAPTIMEOUT_LIST3
+
1975 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1977 BUG_ON(!cpu_cache_get(cachep
));
1978 cpu_cache_get(cachep
)->avail
= 0;
1979 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
1980 cpu_cache_get(cachep
)->batchcount
= 1;
1981 cpu_cache_get(cachep
)->touched
= 0;
1982 cachep
->batchcount
= 1;
1983 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
1986 /* cache setup completed, link it into the list */
1987 list_add(&cachep
->next
, &cache_chain
);
1989 if (!cachep
&& (flags
& SLAB_PANIC
))
1990 panic("kmem_cache_create(): failed to create slab `%s'\n",
1992 mutex_unlock(&cache_chain_mutex
);
1993 unlock_cpu_hotplug();
1996 EXPORT_SYMBOL(kmem_cache_create
);
1999 static void check_irq_off(void)
2001 BUG_ON(!irqs_disabled());
2004 static void check_irq_on(void)
2006 BUG_ON(irqs_disabled());
2009 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2013 assert_spin_locked(&cachep
->nodelists
[numa_node_id()]->list_lock
);
2017 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2021 assert_spin_locked(&cachep
->nodelists
[node
]->list_lock
);
2026 #define check_irq_off() do { } while(0)
2027 #define check_irq_on() do { } while(0)
2028 #define check_spinlock_acquired(x) do { } while(0)
2029 #define check_spinlock_acquired_node(x, y) do { } while(0)
2033 * Waits for all CPUs to execute func().
2035 static void smp_call_function_all_cpus(void (*func
)(void *arg
), void *arg
)
2040 local_irq_disable();
2044 if (smp_call_function(func
, arg
, 1, 1))
2050 static void drain_array_locked(struct kmem_cache
*cachep
, struct array_cache
*ac
,
2051 int force
, int node
);
2053 static void do_drain(void *arg
)
2055 struct kmem_cache
*cachep
= (struct kmem_cache
*) arg
;
2056 struct array_cache
*ac
;
2057 int node
= numa_node_id();
2060 ac
= cpu_cache_get(cachep
);
2061 spin_lock(&cachep
->nodelists
[node
]->list_lock
);
2062 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
2063 spin_unlock(&cachep
->nodelists
[node
]->list_lock
);
2067 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2069 struct kmem_list3
*l3
;
2072 smp_call_function_all_cpus(do_drain
, cachep
);
2074 for_each_online_node(node
) {
2075 l3
= cachep
->nodelists
[node
];
2077 spin_lock_irq(&l3
->list_lock
);
2078 drain_array_locked(cachep
, l3
->shared
, 1, node
);
2079 spin_unlock_irq(&l3
->list_lock
);
2081 drain_alien_cache(cachep
, l3
->alien
);
2086 static int __node_shrink(struct kmem_cache
*cachep
, int node
)
2089 struct kmem_list3
*l3
= cachep
->nodelists
[node
];
2093 struct list_head
*p
;
2095 p
= l3
->slabs_free
.prev
;
2096 if (p
== &l3
->slabs_free
)
2099 slabp
= list_entry(l3
->slabs_free
.prev
, struct slab
, list
);
2104 list_del(&slabp
->list
);
2106 l3
->free_objects
-= cachep
->num
;
2107 spin_unlock_irq(&l3
->list_lock
);
2108 slab_destroy(cachep
, slabp
);
2109 spin_lock_irq(&l3
->list_lock
);
2111 ret
= !list_empty(&l3
->slabs_full
) || !list_empty(&l3
->slabs_partial
);
2115 static int __cache_shrink(struct kmem_cache
*cachep
)
2118 struct kmem_list3
*l3
;
2120 drain_cpu_caches(cachep
);
2123 for_each_online_node(i
) {
2124 l3
= cachep
->nodelists
[i
];
2126 spin_lock_irq(&l3
->list_lock
);
2127 ret
+= __node_shrink(cachep
, i
);
2128 spin_unlock_irq(&l3
->list_lock
);
2131 return (ret
? 1 : 0);
2135 * kmem_cache_shrink - Shrink a cache.
2136 * @cachep: The cache to shrink.
2138 * Releases as many slabs as possible for a cache.
2139 * To help debugging, a zero exit status indicates all slabs were released.
2141 int kmem_cache_shrink(struct kmem_cache
*cachep
)
2143 if (!cachep
|| in_interrupt())
2146 return __cache_shrink(cachep
);
2148 EXPORT_SYMBOL(kmem_cache_shrink
);
2151 * kmem_cache_destroy - delete a cache
2152 * @cachep: the cache to destroy
2154 * Remove a struct kmem_cache object from the slab cache.
2155 * Returns 0 on success.
2157 * It is expected this function will be called by a module when it is
2158 * unloaded. This will remove the cache completely, and avoid a duplicate
2159 * cache being allocated each time a module is loaded and unloaded, if the
2160 * module doesn't have persistent in-kernel storage across loads and unloads.
2162 * The cache must be empty before calling this function.
2164 * The caller must guarantee that noone will allocate memory from the cache
2165 * during the kmem_cache_destroy().
2167 int kmem_cache_destroy(struct kmem_cache
*cachep
)
2170 struct kmem_list3
*l3
;
2172 if (!cachep
|| in_interrupt())
2175 /* Don't let CPUs to come and go */
2178 /* Find the cache in the chain of caches. */
2179 mutex_lock(&cache_chain_mutex
);
2181 * the chain is never empty, cache_cache is never destroyed
2183 list_del(&cachep
->next
);
2184 mutex_unlock(&cache_chain_mutex
);
2186 if (__cache_shrink(cachep
)) {
2187 slab_error(cachep
, "Can't free all objects");
2188 mutex_lock(&cache_chain_mutex
);
2189 list_add(&cachep
->next
, &cache_chain
);
2190 mutex_unlock(&cache_chain_mutex
);
2191 unlock_cpu_hotplug();
2195 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
2198 for_each_online_cpu(i
)
2199 kfree(cachep
->array
[i
]);
2201 /* NUMA: free the list3 structures */
2202 for_each_online_node(i
) {
2203 if ((l3
= cachep
->nodelists
[i
])) {
2205 free_alien_cache(l3
->alien
);
2209 kmem_cache_free(&cache_cache
, cachep
);
2211 unlock_cpu_hotplug();
2215 EXPORT_SYMBOL(kmem_cache_destroy
);
2217 /* Get the memory for a slab management obj. */
2218 static struct slab
*alloc_slabmgmt(struct kmem_cache
*cachep
, void *objp
,
2219 int colour_off
, gfp_t local_flags
)
2223 if (OFF_SLAB(cachep
)) {
2224 /* Slab management obj is off-slab. */
2225 slabp
= kmem_cache_alloc(cachep
->slabp_cache
, local_flags
);
2229 slabp
= objp
+ colour_off
;
2230 colour_off
+= cachep
->slab_size
;
2233 slabp
->colouroff
= colour_off
;
2234 slabp
->s_mem
= objp
+ colour_off
;
2239 static inline kmem_bufctl_t
*slab_bufctl(struct slab
*slabp
)
2241 return (kmem_bufctl_t
*) (slabp
+ 1);
2244 static void cache_init_objs(struct kmem_cache
*cachep
,
2245 struct slab
*slabp
, unsigned long ctor_flags
)
2249 for (i
= 0; i
< cachep
->num
; i
++) {
2250 void *objp
= slabp
->s_mem
+ cachep
->buffer_size
* i
;
2252 /* need to poison the objs? */
2253 if (cachep
->flags
& SLAB_POISON
)
2254 poison_obj(cachep
, objp
, POISON_FREE
);
2255 if (cachep
->flags
& SLAB_STORE_USER
)
2256 *dbg_userword(cachep
, objp
) = NULL
;
2258 if (cachep
->flags
& SLAB_RED_ZONE
) {
2259 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2260 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2263 * Constructors are not allowed to allocate memory from
2264 * the same cache which they are a constructor for.
2265 * Otherwise, deadlock. They must also be threaded.
2267 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2268 cachep
->ctor(objp
+ obj_offset(cachep
), cachep
,
2271 if (cachep
->flags
& SLAB_RED_ZONE
) {
2272 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2273 slab_error(cachep
, "constructor overwrote the"
2274 " end of an object");
2275 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2276 slab_error(cachep
, "constructor overwrote the"
2277 " start of an object");
2279 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
)
2280 && cachep
->flags
& SLAB_POISON
)
2281 kernel_map_pages(virt_to_page(objp
),
2282 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2285 cachep
->ctor(objp
, cachep
, ctor_flags
);
2287 slab_bufctl(slabp
)[i
] = i
+ 1;
2289 slab_bufctl(slabp
)[i
- 1] = BUFCTL_END
;
2293 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2295 if (flags
& SLAB_DMA
) {
2296 if (!(cachep
->gfpflags
& GFP_DMA
))
2299 if (cachep
->gfpflags
& GFP_DMA
)
2304 static void *slab_get_obj(struct kmem_cache
*cachep
, struct slab
*slabp
, int nodeid
)
2306 void *objp
= slabp
->s_mem
+ (slabp
->free
* cachep
->buffer_size
);
2310 next
= slab_bufctl(slabp
)[slabp
->free
];
2312 slab_bufctl(slabp
)[slabp
->free
] = BUFCTL_FREE
;
2313 WARN_ON(slabp
->nodeid
!= nodeid
);
2320 static void slab_put_obj(struct kmem_cache
*cachep
, struct slab
*slabp
, void *objp
,
2323 unsigned int objnr
= (unsigned)(objp
-slabp
->s_mem
) / cachep
->buffer_size
;
2326 /* Verify that the slab belongs to the intended node */
2327 WARN_ON(slabp
->nodeid
!= nodeid
);
2329 if (slab_bufctl(slabp
)[objnr
] != BUFCTL_FREE
) {
2330 printk(KERN_ERR
"slab: double free detected in cache "
2331 "'%s', objp %p\n", cachep
->name
, objp
);
2335 slab_bufctl(slabp
)[objnr
] = slabp
->free
;
2336 slabp
->free
= objnr
;
2340 static void set_slab_attr(struct kmem_cache
*cachep
, struct slab
*slabp
, void *objp
)
2345 /* Nasty!!!!!! I hope this is OK. */
2346 i
= 1 << cachep
->gfporder
;
2347 page
= virt_to_page(objp
);
2349 page_set_cache(page
, cachep
);
2350 page_set_slab(page
, slabp
);
2356 * Grow (by 1) the number of slabs within a cache. This is called by
2357 * kmem_cache_alloc() when there are no active objs left in a cache.
2359 static int cache_grow(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
2365 unsigned long ctor_flags
;
2366 struct kmem_list3
*l3
;
2368 /* Be lazy and only check for valid flags here,
2369 * keeping it out of the critical path in kmem_cache_alloc().
2371 if (flags
& ~(SLAB_DMA
| SLAB_LEVEL_MASK
| SLAB_NO_GROW
))
2373 if (flags
& SLAB_NO_GROW
)
2376 ctor_flags
= SLAB_CTOR_CONSTRUCTOR
;
2377 local_flags
= (flags
& SLAB_LEVEL_MASK
);
2378 if (!(local_flags
& __GFP_WAIT
))
2380 * Not allowed to sleep. Need to tell a constructor about
2381 * this - it might need to know...
2383 ctor_flags
|= SLAB_CTOR_ATOMIC
;
2385 /* Take the l3 list lock to change the colour_next on this node */
2387 l3
= cachep
->nodelists
[nodeid
];
2388 spin_lock(&l3
->list_lock
);
2390 /* Get colour for the slab, and cal the next value. */
2391 offset
= l3
->colour_next
;
2393 if (l3
->colour_next
>= cachep
->colour
)
2394 l3
->colour_next
= 0;
2395 spin_unlock(&l3
->list_lock
);
2397 offset
*= cachep
->colour_off
;
2399 if (local_flags
& __GFP_WAIT
)
2403 * The test for missing atomic flag is performed here, rather than
2404 * the more obvious place, simply to reduce the critical path length
2405 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2406 * will eventually be caught here (where it matters).
2408 kmem_flagcheck(cachep
, flags
);
2410 /* Get mem for the objs.
2411 * Attempt to allocate a physical page from 'nodeid',
2413 if (!(objp
= kmem_getpages(cachep
, flags
, nodeid
)))
2416 /* Get slab management. */
2417 if (!(slabp
= alloc_slabmgmt(cachep
, objp
, offset
, local_flags
)))
2420 slabp
->nodeid
= nodeid
;
2421 set_slab_attr(cachep
, slabp
, objp
);
2423 cache_init_objs(cachep
, slabp
, ctor_flags
);
2425 if (local_flags
& __GFP_WAIT
)
2426 local_irq_disable();
2428 spin_lock(&l3
->list_lock
);
2430 /* Make slab active. */
2431 list_add_tail(&slabp
->list
, &(l3
->slabs_free
));
2432 STATS_INC_GROWN(cachep
);
2433 l3
->free_objects
+= cachep
->num
;
2434 spin_unlock(&l3
->list_lock
);
2437 kmem_freepages(cachep
, objp
);
2439 if (local_flags
& __GFP_WAIT
)
2440 local_irq_disable();
2447 * Perform extra freeing checks:
2448 * - detect bad pointers.
2449 * - POISON/RED_ZONE checking
2450 * - destructor calls, for caches with POISON+dtor
2452 static void kfree_debugcheck(const void *objp
)
2456 if (!virt_addr_valid(objp
)) {
2457 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
2458 (unsigned long)objp
);
2461 page
= virt_to_page(objp
);
2462 if (!PageSlab(page
)) {
2463 printk(KERN_ERR
"kfree_debugcheck: bad ptr %lxh.\n",
2464 (unsigned long)objp
);
2469 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2476 objp
-= obj_offset(cachep
);
2477 kfree_debugcheck(objp
);
2478 page
= virt_to_page(objp
);
2480 if (page_get_cache(page
) != cachep
) {
2482 "mismatch in kmem_cache_free: expected cache %p, got %p\n",
2483 page_get_cache(page
), cachep
);
2484 printk(KERN_ERR
"%p is %s.\n", cachep
, cachep
->name
);
2485 printk(KERN_ERR
"%p is %s.\n", page_get_cache(page
),
2486 page_get_cache(page
)->name
);
2489 slabp
= page_get_slab(page
);
2491 if (cachep
->flags
& SLAB_RED_ZONE
) {
2492 if (*dbg_redzone1(cachep
, objp
) != RED_ACTIVE
2493 || *dbg_redzone2(cachep
, objp
) != RED_ACTIVE
) {
2495 "double free, or memory outside"
2496 " object was overwritten");
2498 "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2499 objp
, *dbg_redzone1(cachep
, objp
),
2500 *dbg_redzone2(cachep
, objp
));
2502 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2503 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2505 if (cachep
->flags
& SLAB_STORE_USER
)
2506 *dbg_userword(cachep
, objp
) = caller
;
2508 objnr
= (unsigned)(objp
- slabp
->s_mem
) / cachep
->buffer_size
;
2510 BUG_ON(objnr
>= cachep
->num
);
2511 BUG_ON(objp
!= slabp
->s_mem
+ objnr
* cachep
->buffer_size
);
2513 if (cachep
->flags
& SLAB_DEBUG_INITIAL
) {
2514 /* Need to call the slab's constructor so the
2515 * caller can perform a verify of its state (debugging).
2516 * Called without the cache-lock held.
2518 cachep
->ctor(objp
+ obj_offset(cachep
),
2519 cachep
, SLAB_CTOR_CONSTRUCTOR
| SLAB_CTOR_VERIFY
);
2521 if (cachep
->flags
& SLAB_POISON
&& cachep
->dtor
) {
2522 /* we want to cache poison the object,
2523 * call the destruction callback
2525 cachep
->dtor(objp
+ obj_offset(cachep
), cachep
, 0);
2527 if (cachep
->flags
& SLAB_POISON
) {
2528 #ifdef CONFIG_DEBUG_PAGEALLOC
2529 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
)) {
2530 store_stackinfo(cachep
, objp
, (unsigned long)caller
);
2531 kernel_map_pages(virt_to_page(objp
),
2532 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2534 poison_obj(cachep
, objp
, POISON_FREE
);
2537 poison_obj(cachep
, objp
, POISON_FREE
);
2543 static void check_slabp(struct kmem_cache
*cachep
, struct slab
*slabp
)
2548 /* Check slab's freelist to see if this obj is there. */
2549 for (i
= slabp
->free
; i
!= BUFCTL_END
; i
= slab_bufctl(slabp
)[i
]) {
2551 if (entries
> cachep
->num
|| i
>= cachep
->num
)
2554 if (entries
!= cachep
->num
- slabp
->inuse
) {
2557 "slab: Internal list corruption detected in cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2558 cachep
->name
, cachep
->num
, slabp
, slabp
->inuse
);
2560 i
< sizeof(*slabp
) + cachep
->num
* sizeof(kmem_bufctl_t
);
2563 printk("\n%03x:", i
);
2564 printk(" %02x", ((unsigned char *)slabp
)[i
]);
2571 #define kfree_debugcheck(x) do { } while(0)
2572 #define cache_free_debugcheck(x,objp,z) (objp)
2573 #define check_slabp(x,y) do { } while(0)
2576 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
2579 struct kmem_list3
*l3
;
2580 struct array_cache
*ac
;
2583 ac
= cpu_cache_get(cachep
);
2585 batchcount
= ac
->batchcount
;
2586 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2587 /* if there was little recent activity on this
2588 * cache, then perform only a partial refill.
2589 * Otherwise we could generate refill bouncing.
2591 batchcount
= BATCHREFILL_LIMIT
;
2593 l3
= cachep
->nodelists
[numa_node_id()];
2595 BUG_ON(ac
->avail
> 0 || !l3
);
2596 spin_lock(&l3
->list_lock
);
2599 struct array_cache
*shared_array
= l3
->shared
;
2600 if (shared_array
->avail
) {
2601 if (batchcount
> shared_array
->avail
)
2602 batchcount
= shared_array
->avail
;
2603 shared_array
->avail
-= batchcount
;
2604 ac
->avail
= batchcount
;
2606 &(shared_array
->entry
[shared_array
->avail
]),
2607 sizeof(void *) * batchcount
);
2608 shared_array
->touched
= 1;
2612 while (batchcount
> 0) {
2613 struct list_head
*entry
;
2615 /* Get slab alloc is to come from. */
2616 entry
= l3
->slabs_partial
.next
;
2617 if (entry
== &l3
->slabs_partial
) {
2618 l3
->free_touched
= 1;
2619 entry
= l3
->slabs_free
.next
;
2620 if (entry
== &l3
->slabs_free
)
2624 slabp
= list_entry(entry
, struct slab
, list
);
2625 check_slabp(cachep
, slabp
);
2626 check_spinlock_acquired(cachep
);
2627 while (slabp
->inuse
< cachep
->num
&& batchcount
--) {
2628 STATS_INC_ALLOCED(cachep
);
2629 STATS_INC_ACTIVE(cachep
);
2630 STATS_SET_HIGH(cachep
);
2632 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, slabp
,
2635 check_slabp(cachep
, slabp
);
2637 /* move slabp to correct slabp list: */
2638 list_del(&slabp
->list
);
2639 if (slabp
->free
== BUFCTL_END
)
2640 list_add(&slabp
->list
, &l3
->slabs_full
);
2642 list_add(&slabp
->list
, &l3
->slabs_partial
);
2646 l3
->free_objects
-= ac
->avail
;
2648 spin_unlock(&l3
->list_lock
);
2650 if (unlikely(!ac
->avail
)) {
2652 x
= cache_grow(cachep
, flags
, numa_node_id());
2654 // cache_grow can reenable interrupts, then ac could change.
2655 ac
= cpu_cache_get(cachep
);
2656 if (!x
&& ac
->avail
== 0) // no objects in sight? abort
2659 if (!ac
->avail
) // objects refilled by interrupt?
2663 return ac
->entry
[--ac
->avail
];
2667 cache_alloc_debugcheck_before(struct kmem_cache
*cachep
, gfp_t flags
)
2669 might_sleep_if(flags
& __GFP_WAIT
);
2671 kmem_flagcheck(cachep
, flags
);
2676 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
, gfp_t flags
,
2677 void *objp
, void *caller
)
2681 if (cachep
->flags
& SLAB_POISON
) {
2682 #ifdef CONFIG_DEBUG_PAGEALLOC
2683 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
2684 kernel_map_pages(virt_to_page(objp
),
2685 cachep
->buffer_size
/ PAGE_SIZE
, 1);
2687 check_poison_obj(cachep
, objp
);
2689 check_poison_obj(cachep
, objp
);
2691 poison_obj(cachep
, objp
, POISON_INUSE
);
2693 if (cachep
->flags
& SLAB_STORE_USER
)
2694 *dbg_userword(cachep
, objp
) = caller
;
2696 if (cachep
->flags
& SLAB_RED_ZONE
) {
2697 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
2698 || *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
2700 "double free, or memory outside"
2701 " object was overwritten");
2703 "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2704 objp
, *dbg_redzone1(cachep
, objp
),
2705 *dbg_redzone2(cachep
, objp
));
2707 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
2708 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
2710 objp
+= obj_offset(cachep
);
2711 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
) {
2712 unsigned long ctor_flags
= SLAB_CTOR_CONSTRUCTOR
;
2714 if (!(flags
& __GFP_WAIT
))
2715 ctor_flags
|= SLAB_CTOR_ATOMIC
;
2717 cachep
->ctor(objp
, cachep
, ctor_flags
);
2722 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2725 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
2728 struct array_cache
*ac
;
2731 if (unlikely(current
->mempolicy
&& !in_interrupt())) {
2732 int nid
= slab_node(current
->mempolicy
);
2734 if (nid
!= numa_node_id())
2735 return __cache_alloc_node(cachep
, flags
, nid
);
2740 ac
= cpu_cache_get(cachep
);
2741 if (likely(ac
->avail
)) {
2742 STATS_INC_ALLOCHIT(cachep
);
2744 objp
= ac
->entry
[--ac
->avail
];
2746 STATS_INC_ALLOCMISS(cachep
);
2747 objp
= cache_alloc_refill(cachep
, flags
);
2752 static __always_inline
void *
2753 __cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
, void *caller
)
2755 unsigned long save_flags
;
2758 cache_alloc_debugcheck_before(cachep
, flags
);
2760 local_irq_save(save_flags
);
2761 objp
= ____cache_alloc(cachep
, flags
);
2762 local_irq_restore(save_flags
);
2763 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
,
2771 * A interface to enable slab creation on nodeid
2773 static void *__cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
2775 struct list_head
*entry
;
2777 struct kmem_list3
*l3
;
2781 l3
= cachep
->nodelists
[nodeid
];
2786 spin_lock(&l3
->list_lock
);
2787 entry
= l3
->slabs_partial
.next
;
2788 if (entry
== &l3
->slabs_partial
) {
2789 l3
->free_touched
= 1;
2790 entry
= l3
->slabs_free
.next
;
2791 if (entry
== &l3
->slabs_free
)
2795 slabp
= list_entry(entry
, struct slab
, list
);
2796 check_spinlock_acquired_node(cachep
, nodeid
);
2797 check_slabp(cachep
, slabp
);
2799 STATS_INC_NODEALLOCS(cachep
);
2800 STATS_INC_ACTIVE(cachep
);
2801 STATS_SET_HIGH(cachep
);
2803 BUG_ON(slabp
->inuse
== cachep
->num
);
2805 obj
= slab_get_obj(cachep
, slabp
, nodeid
);
2806 check_slabp(cachep
, slabp
);
2808 /* move slabp to correct slabp list: */
2809 list_del(&slabp
->list
);
2811 if (slabp
->free
== BUFCTL_END
) {
2812 list_add(&slabp
->list
, &l3
->slabs_full
);
2814 list_add(&slabp
->list
, &l3
->slabs_partial
);
2817 spin_unlock(&l3
->list_lock
);
2821 spin_unlock(&l3
->list_lock
);
2822 x
= cache_grow(cachep
, flags
, nodeid
);
2834 * Caller needs to acquire correct kmem_list's list_lock
2836 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int nr_objects
,
2840 struct kmem_list3
*l3
;
2842 for (i
= 0; i
< nr_objects
; i
++) {
2843 void *objp
= objpp
[i
];
2846 slabp
= virt_to_slab(objp
);
2847 l3
= cachep
->nodelists
[node
];
2848 list_del(&slabp
->list
);
2849 check_spinlock_acquired_node(cachep
, node
);
2850 check_slabp(cachep
, slabp
);
2851 slab_put_obj(cachep
, slabp
, objp
, node
);
2852 STATS_DEC_ACTIVE(cachep
);
2854 check_slabp(cachep
, slabp
);
2856 /* fixup slab chains */
2857 if (slabp
->inuse
== 0) {
2858 if (l3
->free_objects
> l3
->free_limit
) {
2859 l3
->free_objects
-= cachep
->num
;
2860 slab_destroy(cachep
, slabp
);
2862 list_add(&slabp
->list
, &l3
->slabs_free
);
2865 /* Unconditionally move a slab to the end of the
2866 * partial list on free - maximum time for the
2867 * other objects to be freed, too.
2869 list_add_tail(&slabp
->list
, &l3
->slabs_partial
);
2874 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
2877 struct kmem_list3
*l3
;
2878 int node
= numa_node_id();
2880 batchcount
= ac
->batchcount
;
2882 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
2885 l3
= cachep
->nodelists
[node
];
2886 spin_lock(&l3
->list_lock
);
2888 struct array_cache
*shared_array
= l3
->shared
;
2889 int max
= shared_array
->limit
- shared_array
->avail
;
2891 if (batchcount
> max
)
2893 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
2894 ac
->entry
, sizeof(void *) * batchcount
);
2895 shared_array
->avail
+= batchcount
;
2900 free_block(cachep
, ac
->entry
, batchcount
, node
);
2905 struct list_head
*p
;
2907 p
= l3
->slabs_free
.next
;
2908 while (p
!= &(l3
->slabs_free
)) {
2911 slabp
= list_entry(p
, struct slab
, list
);
2912 BUG_ON(slabp
->inuse
);
2917 STATS_SET_FREEABLE(cachep
, i
);
2920 spin_unlock(&l3
->list_lock
);
2921 ac
->avail
-= batchcount
;
2922 memmove(ac
->entry
, &(ac
->entry
[batchcount
]),
2923 sizeof(void *) * ac
->avail
);
2928 * Release an obj back to its cache. If the obj has a constructed
2929 * state, it must be in this state _before_ it is released.
2931 * Called with disabled ints.
2933 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
)
2935 struct array_cache
*ac
= cpu_cache_get(cachep
);
2938 objp
= cache_free_debugcheck(cachep
, objp
, __builtin_return_address(0));
2940 /* Make sure we are not freeing a object from another
2941 * node to the array cache on this cpu.
2946 slabp
= virt_to_slab(objp
);
2947 if (unlikely(slabp
->nodeid
!= numa_node_id())) {
2948 struct array_cache
*alien
= NULL
;
2949 int nodeid
= slabp
->nodeid
;
2950 struct kmem_list3
*l3
=
2951 cachep
->nodelists
[numa_node_id()];
2953 STATS_INC_NODEFREES(cachep
);
2954 if (l3
->alien
&& l3
->alien
[nodeid
]) {
2955 alien
= l3
->alien
[nodeid
];
2956 spin_lock(&alien
->lock
);
2957 if (unlikely(alien
->avail
== alien
->limit
))
2958 __drain_alien_cache(cachep
,
2960 alien
->entry
[alien
->avail
++] = objp
;
2961 spin_unlock(&alien
->lock
);
2963 spin_lock(&(cachep
->nodelists
[nodeid
])->
2965 free_block(cachep
, &objp
, 1, nodeid
);
2966 spin_unlock(&(cachep
->nodelists
[nodeid
])->
2973 if (likely(ac
->avail
< ac
->limit
)) {
2974 STATS_INC_FREEHIT(cachep
);
2975 ac
->entry
[ac
->avail
++] = objp
;
2978 STATS_INC_FREEMISS(cachep
);
2979 cache_flusharray(cachep
, ac
);
2980 ac
->entry
[ac
->avail
++] = objp
;
2985 * kmem_cache_alloc - Allocate an object
2986 * @cachep: The cache to allocate from.
2987 * @flags: See kmalloc().
2989 * Allocate an object from this cache. The flags are only relevant
2990 * if the cache has no available objects.
2992 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
2994 return __cache_alloc(cachep
, flags
, __builtin_return_address(0));
2996 EXPORT_SYMBOL(kmem_cache_alloc
);
2999 * kmem_ptr_validate - check if an untrusted pointer might
3001 * @cachep: the cache we're checking against
3002 * @ptr: pointer to validate
3004 * This verifies that the untrusted pointer looks sane:
3005 * it is _not_ a guarantee that the pointer is actually
3006 * part of the slab cache in question, but it at least
3007 * validates that the pointer can be dereferenced and
3008 * looks half-way sane.
3010 * Currently only used for dentry validation.
3012 int fastcall
kmem_ptr_validate(struct kmem_cache
*cachep
, void *ptr
)
3014 unsigned long addr
= (unsigned long)ptr
;
3015 unsigned long min_addr
= PAGE_OFFSET
;
3016 unsigned long align_mask
= BYTES_PER_WORD
- 1;
3017 unsigned long size
= cachep
->buffer_size
;
3020 if (unlikely(addr
< min_addr
))
3022 if (unlikely(addr
> (unsigned long)high_memory
- size
))
3024 if (unlikely(addr
& align_mask
))
3026 if (unlikely(!kern_addr_valid(addr
)))
3028 if (unlikely(!kern_addr_valid(addr
+ size
- 1)))
3030 page
= virt_to_page(ptr
);
3031 if (unlikely(!PageSlab(page
)))
3033 if (unlikely(page_get_cache(page
) != cachep
))
3042 * kmem_cache_alloc_node - Allocate an object on the specified node
3043 * @cachep: The cache to allocate from.
3044 * @flags: See kmalloc().
3045 * @nodeid: node number of the target node.
3047 * Identical to kmem_cache_alloc, except that this function is slow
3048 * and can sleep. And it will allocate memory on the given node, which
3049 * can improve the performance for cpu bound structures.
3050 * New and improved: it will now make sure that the object gets
3051 * put on the correct node list so that there is no false sharing.
3053 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3055 unsigned long save_flags
;
3058 cache_alloc_debugcheck_before(cachep
, flags
);
3059 local_irq_save(save_flags
);
3061 if (nodeid
== -1 || nodeid
== numa_node_id() ||
3062 !cachep
->nodelists
[nodeid
])
3063 ptr
= ____cache_alloc(cachep
, flags
);
3065 ptr
= __cache_alloc_node(cachep
, flags
, nodeid
);
3066 local_irq_restore(save_flags
);
3068 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
,
3069 __builtin_return_address(0));
3073 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3075 void *kmalloc_node(size_t size
, gfp_t flags
, int node
)
3077 struct kmem_cache
*cachep
;
3079 cachep
= kmem_find_general_cachep(size
, flags
);
3080 if (unlikely(cachep
== NULL
))
3082 return kmem_cache_alloc_node(cachep
, flags
, node
);
3084 EXPORT_SYMBOL(kmalloc_node
);
3088 * kmalloc - allocate memory
3089 * @size: how many bytes of memory are required.
3090 * @flags: the type of memory to allocate.
3092 * kmalloc is the normal method of allocating memory
3095 * The @flags argument may be one of:
3097 * %GFP_USER - Allocate memory on behalf of user. May sleep.
3099 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
3101 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
3103 * Additionally, the %GFP_DMA flag may be set to indicate the memory
3104 * must be suitable for DMA. This can mean different things on different
3105 * platforms. For example, on i386, it means that the memory must come
3106 * from the first 16MB.
3108 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3111 struct kmem_cache
*cachep
;
3113 /* If you want to save a few bytes .text space: replace
3115 * Then kmalloc uses the uninlined functions instead of the inline
3118 cachep
= __find_general_cachep(size
, flags
);
3119 if (unlikely(cachep
== NULL
))
3121 return __cache_alloc(cachep
, flags
, caller
);
3124 #ifndef CONFIG_DEBUG_SLAB
3126 void *__kmalloc(size_t size
, gfp_t flags
)
3128 return __do_kmalloc(size
, flags
, NULL
);
3130 EXPORT_SYMBOL(__kmalloc
);
3134 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, void *caller
)
3136 return __do_kmalloc(size
, flags
, caller
);
3138 EXPORT_SYMBOL(__kmalloc_track_caller
);
3144 * __alloc_percpu - allocate one copy of the object for every present
3145 * cpu in the system, zeroing them.
3146 * Objects should be dereferenced using the per_cpu_ptr macro only.
3148 * @size: how many bytes of memory are required.
3150 void *__alloc_percpu(size_t size
)
3153 struct percpu_data
*pdata
= kmalloc(sizeof(*pdata
), GFP_KERNEL
);
3159 * Cannot use for_each_online_cpu since a cpu may come online
3160 * and we have no way of figuring out how to fix the array
3161 * that we have allocated then....
3164 int node
= cpu_to_node(i
);
3166 if (node_online(node
))
3167 pdata
->ptrs
[i
] = kmalloc_node(size
, GFP_KERNEL
, node
);
3169 pdata
->ptrs
[i
] = kmalloc(size
, GFP_KERNEL
);
3171 if (!pdata
->ptrs
[i
])
3173 memset(pdata
->ptrs
[i
], 0, size
);
3176 /* Catch derefs w/o wrappers */
3177 return (void *)(~(unsigned long)pdata
);
3181 if (!cpu_possible(i
))
3183 kfree(pdata
->ptrs
[i
]);
3188 EXPORT_SYMBOL(__alloc_percpu
);
3192 * kmem_cache_free - Deallocate an object
3193 * @cachep: The cache the allocation was from.
3194 * @objp: The previously allocated object.
3196 * Free an object which was previously allocated from this
3199 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3201 unsigned long flags
;
3203 local_irq_save(flags
);
3204 __cache_free(cachep
, objp
);
3205 local_irq_restore(flags
);
3207 EXPORT_SYMBOL(kmem_cache_free
);
3210 * kfree - free previously allocated memory
3211 * @objp: pointer returned by kmalloc.
3213 * If @objp is NULL, no operation is performed.
3215 * Don't free memory not originally allocated by kmalloc()
3216 * or you will run into trouble.
3218 void kfree(const void *objp
)
3220 struct kmem_cache
*c
;
3221 unsigned long flags
;
3223 if (unlikely(!objp
))
3225 local_irq_save(flags
);
3226 kfree_debugcheck(objp
);
3227 c
= virt_to_cache(objp
);
3228 mutex_debug_check_no_locks_freed(objp
, obj_size(c
));
3229 __cache_free(c
, (void *)objp
);
3230 local_irq_restore(flags
);
3232 EXPORT_SYMBOL(kfree
);
3236 * free_percpu - free previously allocated percpu memory
3237 * @objp: pointer returned by alloc_percpu.
3239 * Don't free memory not originally allocated by alloc_percpu()
3240 * The complemented objp is to check for that.
3242 void free_percpu(const void *objp
)
3245 struct percpu_data
*p
= (struct percpu_data
*)(~(unsigned long)objp
);
3248 * We allocate for all cpus so we cannot use for online cpu here.
3254 EXPORT_SYMBOL(free_percpu
);
3257 unsigned int kmem_cache_size(struct kmem_cache
*cachep
)
3259 return obj_size(cachep
);
3261 EXPORT_SYMBOL(kmem_cache_size
);
3263 const char *kmem_cache_name(struct kmem_cache
*cachep
)
3265 return cachep
->name
;
3267 EXPORT_SYMBOL_GPL(kmem_cache_name
);
3270 * This initializes kmem_list3 for all nodes.
3272 static int alloc_kmemlist(struct kmem_cache
*cachep
)
3275 struct kmem_list3
*l3
;
3278 for_each_online_node(node
) {
3279 struct array_cache
*nc
= NULL
, *new;
3280 struct array_cache
**new_alien
= NULL
;
3282 if (!(new_alien
= alloc_alien_cache(node
, cachep
->limit
)))
3285 if (!(new = alloc_arraycache(node
, (cachep
->shared
*
3286 cachep
->batchcount
),
3289 if ((l3
= cachep
->nodelists
[node
])) {
3291 spin_lock_irq(&l3
->list_lock
);
3293 if ((nc
= cachep
->nodelists
[node
]->shared
))
3294 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
3297 if (!cachep
->nodelists
[node
]->alien
) {
3298 l3
->alien
= new_alien
;
3301 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3302 cachep
->batchcount
+ cachep
->num
;
3303 spin_unlock_irq(&l3
->list_lock
);
3305 free_alien_cache(new_alien
);
3308 if (!(l3
= kmalloc_node(sizeof(struct kmem_list3
),
3312 kmem_list3_init(l3
);
3313 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
3314 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
3316 l3
->alien
= new_alien
;
3317 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3318 cachep
->batchcount
+ cachep
->num
;
3319 cachep
->nodelists
[node
] = l3
;
3327 struct ccupdate_struct
{
3328 struct kmem_cache
*cachep
;
3329 struct array_cache
*new[NR_CPUS
];
3332 static void do_ccupdate_local(void *info
)
3334 struct ccupdate_struct
*new = (struct ccupdate_struct
*)info
;
3335 struct array_cache
*old
;
3338 old
= cpu_cache_get(new->cachep
);
3340 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
3341 new->new[smp_processor_id()] = old
;
3344 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
, int batchcount
,
3347 struct ccupdate_struct
new;
3350 memset(&new.new, 0, sizeof(new.new));
3351 for_each_online_cpu(i
) {
3353 alloc_arraycache(cpu_to_node(i
), limit
, batchcount
);
3355 for (i
--; i
>= 0; i
--)
3360 new.cachep
= cachep
;
3362 smp_call_function_all_cpus(do_ccupdate_local
, (void *)&new);
3365 spin_lock(&cachep
->spinlock
);
3366 cachep
->batchcount
= batchcount
;
3367 cachep
->limit
= limit
;
3368 cachep
->shared
= shared
;
3369 spin_unlock(&cachep
->spinlock
);
3371 for_each_online_cpu(i
) {
3372 struct array_cache
*ccold
= new.new[i
];
3375 spin_lock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3376 free_block(cachep
, ccold
->entry
, ccold
->avail
, cpu_to_node(i
));
3377 spin_unlock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3381 err
= alloc_kmemlist(cachep
);
3383 printk(KERN_ERR
"alloc_kmemlist failed for %s, error %d.\n",
3384 cachep
->name
, -err
);
3390 static void enable_cpucache(struct kmem_cache
*cachep
)
3395 /* The head array serves three purposes:
3396 * - create a LIFO ordering, i.e. return objects that are cache-warm
3397 * - reduce the number of spinlock operations.
3398 * - reduce the number of linked list operations on the slab and
3399 * bufctl chains: array operations are cheaper.
3400 * The numbers are guessed, we should auto-tune as described by
3403 if (cachep
->buffer_size
> 131072)
3405 else if (cachep
->buffer_size
> PAGE_SIZE
)
3407 else if (cachep
->buffer_size
> 1024)
3409 else if (cachep
->buffer_size
> 256)
3414 /* Cpu bound tasks (e.g. network routing) can exhibit cpu bound
3415 * allocation behaviour: Most allocs on one cpu, most free operations
3416 * on another cpu. For these cases, an efficient object passing between
3417 * cpus is necessary. This is provided by a shared array. The array
3418 * replaces Bonwick's magazine layer.
3419 * On uniprocessor, it's functionally equivalent (but less efficient)
3420 * to a larger limit. Thus disabled by default.
3424 if (cachep
->buffer_size
<= PAGE_SIZE
)
3429 /* With debugging enabled, large batchcount lead to excessively
3430 * long periods with disabled local interrupts. Limit the
3436 err
= do_tune_cpucache(cachep
, limit
, (limit
+ 1) / 2, shared
);
3438 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
3439 cachep
->name
, -err
);
3442 static void drain_array_locked(struct kmem_cache
*cachep
, struct array_cache
*ac
,
3443 int force
, int node
)
3447 check_spinlock_acquired_node(cachep
, node
);
3448 if (ac
->touched
&& !force
) {
3450 } else if (ac
->avail
) {
3451 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
3452 if (tofree
> ac
->avail
) {
3453 tofree
= (ac
->avail
+ 1) / 2;
3455 free_block(cachep
, ac
->entry
, tofree
, node
);
3456 ac
->avail
-= tofree
;
3457 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
3458 sizeof(void *) * ac
->avail
);
3463 * cache_reap - Reclaim memory from caches.
3464 * @unused: unused parameter
3466 * Called from workqueue/eventd every few seconds.
3468 * - clear the per-cpu caches for this CPU.
3469 * - return freeable pages to the main free memory pool.
3471 * If we cannot acquire the cache chain mutex then just give up - we'll
3472 * try again on the next iteration.
3474 static void cache_reap(void *unused
)
3476 struct list_head
*walk
;
3477 struct kmem_list3
*l3
;
3479 if (!mutex_trylock(&cache_chain_mutex
)) {
3480 /* Give up. Setup the next iteration. */
3481 schedule_delayed_work(&__get_cpu_var(reap_work
),
3486 list_for_each(walk
, &cache_chain
) {
3487 struct kmem_cache
*searchp
;
3488 struct list_head
*p
;
3492 searchp
= list_entry(walk
, struct kmem_cache
, next
);
3494 if (searchp
->flags
& SLAB_NO_REAP
)
3499 l3
= searchp
->nodelists
[numa_node_id()];
3501 drain_alien_cache(searchp
, l3
->alien
);
3502 spin_lock_irq(&l3
->list_lock
);
3504 drain_array_locked(searchp
, cpu_cache_get(searchp
), 0,
3507 if (time_after(l3
->next_reap
, jiffies
))
3510 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
;
3513 drain_array_locked(searchp
, l3
->shared
, 0,
3516 if (l3
->free_touched
) {
3517 l3
->free_touched
= 0;
3522 (l3
->free_limit
+ 5 * searchp
->num
-
3523 1) / (5 * searchp
->num
);
3525 p
= l3
->slabs_free
.next
;
3526 if (p
== &(l3
->slabs_free
))
3529 slabp
= list_entry(p
, struct slab
, list
);
3530 BUG_ON(slabp
->inuse
);
3531 list_del(&slabp
->list
);
3532 STATS_INC_REAPED(searchp
);
3534 /* Safe to drop the lock. The slab is no longer
3535 * linked to the cache.
3536 * searchp cannot disappear, we hold
3539 l3
->free_objects
-= searchp
->num
;
3540 spin_unlock_irq(&l3
->list_lock
);
3541 slab_destroy(searchp
, slabp
);
3542 spin_lock_irq(&l3
->list_lock
);
3543 } while (--tofree
> 0);
3545 spin_unlock_irq(&l3
->list_lock
);
3550 mutex_unlock(&cache_chain_mutex
);
3551 drain_remote_pages();
3552 /* Setup the next iteration */
3553 schedule_delayed_work(&__get_cpu_var(reap_work
), REAPTIMEOUT_CPUC
);
3556 #ifdef CONFIG_PROC_FS
3558 static void print_slabinfo_header(struct seq_file
*m
)
3561 * Output format version, so at least we can change it
3562 * without _too_ many complaints.
3565 seq_puts(m
, "slabinfo - version: 2.1 (statistics)\n");
3567 seq_puts(m
, "slabinfo - version: 2.1\n");
3569 seq_puts(m
, "# name <active_objs> <num_objs> <objsize> "
3570 "<objperslab> <pagesperslab>");
3571 seq_puts(m
, " : tunables <limit> <batchcount> <sharedfactor>");
3572 seq_puts(m
, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3574 seq_puts(m
, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3575 "<error> <maxfreeable> <nodeallocs> <remotefrees>");
3576 seq_puts(m
, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3581 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
3584 struct list_head
*p
;
3586 mutex_lock(&cache_chain_mutex
);
3588 print_slabinfo_header(m
);
3589 p
= cache_chain
.next
;
3592 if (p
== &cache_chain
)
3595 return list_entry(p
, struct kmem_cache
, next
);
3598 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
3600 struct kmem_cache
*cachep
= p
;
3602 return cachep
->next
.next
== &cache_chain
? NULL
3603 : list_entry(cachep
->next
.next
, struct kmem_cache
, next
);
3606 static void s_stop(struct seq_file
*m
, void *p
)
3608 mutex_unlock(&cache_chain_mutex
);
3611 static int s_show(struct seq_file
*m
, void *p
)
3613 struct kmem_cache
*cachep
= p
;
3614 struct list_head
*q
;
3616 unsigned long active_objs
;
3617 unsigned long num_objs
;
3618 unsigned long active_slabs
= 0;
3619 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
3623 struct kmem_list3
*l3
;
3625 spin_lock(&cachep
->spinlock
);
3628 for_each_online_node(node
) {
3629 l3
= cachep
->nodelists
[node
];
3634 spin_lock_irq(&l3
->list_lock
);
3636 list_for_each(q
, &l3
->slabs_full
) {
3637 slabp
= list_entry(q
, struct slab
, list
);
3638 if (slabp
->inuse
!= cachep
->num
&& !error
)
3639 error
= "slabs_full accounting error";
3640 active_objs
+= cachep
->num
;
3643 list_for_each(q
, &l3
->slabs_partial
) {
3644 slabp
= list_entry(q
, struct slab
, list
);
3645 if (slabp
->inuse
== cachep
->num
&& !error
)
3646 error
= "slabs_partial inuse accounting error";
3647 if (!slabp
->inuse
&& !error
)
3648 error
= "slabs_partial/inuse accounting error";
3649 active_objs
+= slabp
->inuse
;
3652 list_for_each(q
, &l3
->slabs_free
) {
3653 slabp
= list_entry(q
, struct slab
, list
);
3654 if (slabp
->inuse
&& !error
)
3655 error
= "slabs_free/inuse accounting error";
3658 free_objects
+= l3
->free_objects
;
3660 shared_avail
+= l3
->shared
->avail
;
3662 spin_unlock_irq(&l3
->list_lock
);
3664 num_slabs
+= active_slabs
;
3665 num_objs
= num_slabs
* cachep
->num
;
3666 if (num_objs
- active_objs
!= free_objects
&& !error
)
3667 error
= "free_objects accounting error";
3669 name
= cachep
->name
;
3671 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
3673 seq_printf(m
, "%-17s %6lu %6lu %6u %4u %4d",
3674 name
, active_objs
, num_objs
, cachep
->buffer_size
,
3675 cachep
->num
, (1 << cachep
->gfporder
));
3676 seq_printf(m
, " : tunables %4u %4u %4u",
3677 cachep
->limit
, cachep
->batchcount
, cachep
->shared
);
3678 seq_printf(m
, " : slabdata %6lu %6lu %6lu",
3679 active_slabs
, num_slabs
, shared_avail
);
3682 unsigned long high
= cachep
->high_mark
;
3683 unsigned long allocs
= cachep
->num_allocations
;
3684 unsigned long grown
= cachep
->grown
;
3685 unsigned long reaped
= cachep
->reaped
;
3686 unsigned long errors
= cachep
->errors
;
3687 unsigned long max_freeable
= cachep
->max_freeable
;
3688 unsigned long node_allocs
= cachep
->node_allocs
;
3689 unsigned long node_frees
= cachep
->node_frees
;
3691 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu \
3692 %4lu %4lu %4lu %4lu", allocs
, high
, grown
, reaped
, errors
, max_freeable
, node_allocs
, node_frees
);
3696 unsigned long allochit
= atomic_read(&cachep
->allochit
);
3697 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
3698 unsigned long freehit
= atomic_read(&cachep
->freehit
);
3699 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
3701 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
3702 allochit
, allocmiss
, freehit
, freemiss
);
3706 spin_unlock(&cachep
->spinlock
);
3711 * slabinfo_op - iterator that generates /proc/slabinfo
3720 * num-pages-per-slab
3721 * + further values on SMP and with statistics enabled
3724 struct seq_operations slabinfo_op
= {
3731 #define MAX_SLABINFO_WRITE 128
3733 * slabinfo_write - Tuning for the slab allocator
3735 * @buffer: user buffer
3736 * @count: data length
3739 ssize_t
slabinfo_write(struct file
*file
, const char __user
* buffer
,
3740 size_t count
, loff_t
*ppos
)
3742 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
3743 int limit
, batchcount
, shared
, res
;
3744 struct list_head
*p
;
3746 if (count
> MAX_SLABINFO_WRITE
)
3748 if (copy_from_user(&kbuf
, buffer
, count
))
3750 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
3752 tmp
= strchr(kbuf
, ' ');
3757 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
3760 /* Find the cache in the chain of caches. */
3761 mutex_lock(&cache_chain_mutex
);
3763 list_for_each(p
, &cache_chain
) {
3764 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
,
3767 if (!strcmp(cachep
->name
, kbuf
)) {
3770 batchcount
> limit
|| shared
< 0) {
3773 res
= do_tune_cpucache(cachep
, limit
,
3774 batchcount
, shared
);
3779 mutex_unlock(&cache_chain_mutex
);
3787 * ksize - get the actual amount of memory allocated for a given object
3788 * @objp: Pointer to the object
3790 * kmalloc may internally round up allocations and return more memory
3791 * than requested. ksize() can be used to determine the actual amount of
3792 * memory allocated. The caller may use this additional memory, even though
3793 * a smaller amount of memory was initially specified with the kmalloc call.
3794 * The caller must guarantee that objp points to a valid object previously
3795 * allocated with either kmalloc() or kmem_cache_alloc(). The object
3796 * must not be freed during the duration of the call.
3798 unsigned int ksize(const void *objp
)
3800 if (unlikely(objp
== NULL
))
3803 return obj_size(virt_to_cache(objp
));