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 [].
278 * bootstrap: The caches do not work without cpuarrays anymore, but the
279 * cpuarrays are allocated from the generic caches...
281 #define BOOT_CPUCACHE_ENTRIES 1
282 struct arraycache_init
{
283 struct array_cache cache
;
284 void *entries
[BOOT_CPUCACHE_ENTRIES
];
288 * The slab lists for all objects.
291 struct list_head slabs_partial
; /* partial list first, better asm code */
292 struct list_head slabs_full
;
293 struct list_head slabs_free
;
294 unsigned long free_objects
;
295 unsigned long next_reap
;
297 unsigned int free_limit
;
298 unsigned int colour_next
; /* Per-node cache coloring */
299 spinlock_t list_lock
;
300 struct array_cache
*shared
; /* shared per node */
301 struct array_cache
**alien
; /* on other nodes */
305 * Need this for bootstrapping a per node allocator.
307 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
308 struct kmem_list3 __initdata initkmem_list3
[NUM_INIT_LISTS
];
309 #define CACHE_CACHE 0
311 #define SIZE_L3 (1 + MAX_NUMNODES)
314 * This function must be completely optimized away if a constant is passed to
315 * it. Mostly the same as what is in linux/slab.h except it returns an index.
317 static __always_inline
int index_of(const size_t size
)
319 extern void __bad_size(void);
321 if (__builtin_constant_p(size
)) {
329 #include "linux/kmalloc_sizes.h"
337 #define INDEX_AC index_of(sizeof(struct arraycache_init))
338 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
340 static void kmem_list3_init(struct kmem_list3
*parent
)
342 INIT_LIST_HEAD(&parent
->slabs_full
);
343 INIT_LIST_HEAD(&parent
->slabs_partial
);
344 INIT_LIST_HEAD(&parent
->slabs_free
);
345 parent
->shared
= NULL
;
346 parent
->alien
= NULL
;
347 parent
->colour_next
= 0;
348 spin_lock_init(&parent
->list_lock
);
349 parent
->free_objects
= 0;
350 parent
->free_touched
= 0;
353 #define MAKE_LIST(cachep, listp, slab, nodeid) \
355 INIT_LIST_HEAD(listp); \
356 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
359 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
361 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
362 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
363 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
373 /* 1) per-cpu data, touched during every alloc/free */
374 struct array_cache
*array
[NR_CPUS
];
375 /* 2) Cache tunables. Protected by cache_chain_mutex */
376 unsigned int batchcount
;
380 unsigned int buffer_size
;
381 /* 3) touched by every alloc & free from the backend */
382 struct kmem_list3
*nodelists
[MAX_NUMNODES
];
384 unsigned int flags
; /* constant flags */
385 unsigned int num
; /* # of objs per slab */
387 /* 4) cache_grow/shrink */
388 /* order of pgs per slab (2^n) */
389 unsigned int gfporder
;
391 /* force GFP flags, e.g. GFP_DMA */
394 size_t colour
; /* cache colouring range */
395 unsigned int colour_off
; /* colour offset */
396 struct kmem_cache
*slabp_cache
;
397 unsigned int slab_size
;
398 unsigned int dflags
; /* dynamic flags */
400 /* constructor func */
401 void (*ctor
) (void *, struct kmem_cache
*, unsigned long);
403 /* de-constructor func */
404 void (*dtor
) (void *, struct kmem_cache
*, unsigned long);
406 /* 5) cache creation/removal */
408 struct list_head next
;
412 unsigned long num_active
;
413 unsigned long num_allocations
;
414 unsigned long high_mark
;
416 unsigned long reaped
;
417 unsigned long errors
;
418 unsigned long max_freeable
;
419 unsigned long node_allocs
;
420 unsigned long node_frees
;
428 * If debugging is enabled, then the allocator can add additional
429 * fields and/or padding to every object. buffer_size contains the total
430 * object size including these internal fields, the following two
431 * variables contain the offset to the user object and its size.
438 #define CFLGS_OFF_SLAB (0x80000000UL)
439 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
441 #define BATCHREFILL_LIMIT 16
443 * Optimization question: fewer reaps means less probability for unnessary
444 * cpucache drain/refill cycles.
446 * OTOH the cpuarrays can contain lots of objects,
447 * which could lock up otherwise freeable slabs.
449 #define REAPTIMEOUT_CPUC (2*HZ)
450 #define REAPTIMEOUT_LIST3 (4*HZ)
453 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
454 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
455 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
456 #define STATS_INC_GROWN(x) ((x)->grown++)
457 #define STATS_INC_REAPED(x) ((x)->reaped++)
458 #define STATS_SET_HIGH(x) \
460 if ((x)->num_active > (x)->high_mark) \
461 (x)->high_mark = (x)->num_active; \
463 #define STATS_INC_ERR(x) ((x)->errors++)
464 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
465 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
466 #define STATS_SET_FREEABLE(x, i) \
468 if ((x)->max_freeable < i) \
469 (x)->max_freeable = i; \
471 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
472 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
473 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
474 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
476 #define STATS_INC_ACTIVE(x) do { } while (0)
477 #define STATS_DEC_ACTIVE(x) do { } while (0)
478 #define STATS_INC_ALLOCED(x) do { } while (0)
479 #define STATS_INC_GROWN(x) do { } while (0)
480 #define STATS_INC_REAPED(x) do { } while (0)
481 #define STATS_SET_HIGH(x) do { } while (0)
482 #define STATS_INC_ERR(x) do { } while (0)
483 #define STATS_INC_NODEALLOCS(x) do { } while (0)
484 #define STATS_INC_NODEFREES(x) do { } while (0)
485 #define STATS_SET_FREEABLE(x, i) do { } while (0)
486 #define STATS_INC_ALLOCHIT(x) do { } while (0)
487 #define STATS_INC_ALLOCMISS(x) do { } while (0)
488 #define STATS_INC_FREEHIT(x) do { } while (0)
489 #define STATS_INC_FREEMISS(x) do { } while (0)
494 * Magic nums for obj red zoning.
495 * Placed in the first word before and the first word after an obj.
497 #define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
498 #define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
500 /* ...and for poisoning */
501 #define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
502 #define POISON_FREE 0x6b /* for use-after-free poisoning */
503 #define POISON_END 0xa5 /* end-byte of poisoning */
506 * memory layout of objects:
508 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
509 * the end of an object is aligned with the end of the real
510 * allocation. Catches writes behind the end of the allocation.
511 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
513 * cachep->obj_offset: The real object.
514 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
515 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
516 * [BYTES_PER_WORD long]
518 static int obj_offset(struct kmem_cache
*cachep
)
520 return cachep
->obj_offset
;
523 static int obj_size(struct kmem_cache
*cachep
)
525 return cachep
->obj_size
;
528 static unsigned long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
530 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
531 return (unsigned long*) (objp
+obj_offset(cachep
)-BYTES_PER_WORD
);
534 static unsigned long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
536 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
537 if (cachep
->flags
& SLAB_STORE_USER
)
538 return (unsigned long *)(objp
+ cachep
->buffer_size
-
540 return (unsigned long *)(objp
+ cachep
->buffer_size
- BYTES_PER_WORD
);
543 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
545 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
546 return (void **)(objp
+ cachep
->buffer_size
- BYTES_PER_WORD
);
551 #define obj_offset(x) 0
552 #define obj_size(cachep) (cachep->buffer_size)
553 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
554 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
555 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
560 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
563 #if defined(CONFIG_LARGE_ALLOCS)
564 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
565 #define MAX_GFP_ORDER 13 /* up to 32Mb */
566 #elif defined(CONFIG_MMU)
567 #define MAX_OBJ_ORDER 5 /* 32 pages */
568 #define MAX_GFP_ORDER 5 /* 32 pages */
570 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
571 #define MAX_GFP_ORDER 8 /* up to 1Mb */
575 * Do not go above this order unless 0 objects fit into the slab.
577 #define BREAK_GFP_ORDER_HI 1
578 #define BREAK_GFP_ORDER_LO 0
579 static int slab_break_gfp_order
= BREAK_GFP_ORDER_LO
;
582 * Functions for storing/retrieving the cachep and or slab from the page
583 * allocator. These are used to find the slab an obj belongs to. With kfree(),
584 * these are used to find the cache which an obj belongs to.
586 static inline void page_set_cache(struct page
*page
, struct kmem_cache
*cache
)
588 page
->lru
.next
= (struct list_head
*)cache
;
591 static inline struct kmem_cache
*page_get_cache(struct page
*page
)
593 return (struct kmem_cache
*)page
->lru
.next
;
596 static inline void page_set_slab(struct page
*page
, struct slab
*slab
)
598 page
->lru
.prev
= (struct list_head
*)slab
;
601 static inline struct slab
*page_get_slab(struct page
*page
)
603 return (struct slab
*)page
->lru
.prev
;
606 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
608 struct page
*page
= virt_to_page(obj
);
609 return page_get_cache(page
);
612 static inline struct slab
*virt_to_slab(const void *obj
)
614 struct page
*page
= virt_to_page(obj
);
615 return page_get_slab(page
);
618 static inline void *index_to_obj(struct kmem_cache
*cache
, struct slab
*slab
,
621 return slab
->s_mem
+ cache
->buffer_size
* idx
;
624 static inline unsigned int obj_to_index(struct kmem_cache
*cache
,
625 struct slab
*slab
, void *obj
)
627 return (unsigned)(obj
- slab
->s_mem
) / cache
->buffer_size
;
631 * These are the default caches for kmalloc. Custom caches can have other sizes.
633 struct cache_sizes malloc_sizes
[] = {
634 #define CACHE(x) { .cs_size = (x) },
635 #include <linux/kmalloc_sizes.h>
639 EXPORT_SYMBOL(malloc_sizes
);
641 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
647 static struct cache_names __initdata cache_names
[] = {
648 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
649 #include <linux/kmalloc_sizes.h>
654 static struct arraycache_init initarray_cache __initdata
=
655 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
656 static struct arraycache_init initarray_generic
=
657 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
659 /* internal cache of cache description objs */
660 static struct kmem_cache cache_cache
= {
662 .limit
= BOOT_CPUCACHE_ENTRIES
,
664 .buffer_size
= sizeof(struct kmem_cache
),
665 .flags
= SLAB_NO_REAP
,
666 .name
= "kmem_cache",
668 .obj_size
= sizeof(struct kmem_cache
),
672 /* Guard access to the cache-chain. */
673 static DEFINE_MUTEX(cache_chain_mutex
);
674 static struct list_head cache_chain
;
677 * vm_enough_memory() looks at this to determine how many slab-allocated pages
678 * are possibly freeable under pressure
680 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
682 atomic_t slab_reclaim_pages
;
685 * chicken and egg problem: delay the per-cpu array allocation
686 * until the general caches are up.
695 static DEFINE_PER_CPU(struct work_struct
, reap_work
);
697 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
699 static void enable_cpucache(struct kmem_cache
*cachep
);
700 static void cache_reap(void *unused
);
701 static int __node_shrink(struct kmem_cache
*cachep
, int node
);
703 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
705 return cachep
->array
[smp_processor_id()];
708 static inline struct kmem_cache
*__find_general_cachep(size_t size
,
711 struct cache_sizes
*csizep
= malloc_sizes
;
714 /* This happens if someone tries to call
715 * kmem_cache_create(), or __kmalloc(), before
716 * the generic caches are initialized.
718 BUG_ON(malloc_sizes
[INDEX_AC
].cs_cachep
== NULL
);
720 while (size
> csizep
->cs_size
)
724 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
725 * has cs_{dma,}cachep==NULL. Thus no special case
726 * for large kmalloc calls required.
728 if (unlikely(gfpflags
& GFP_DMA
))
729 return csizep
->cs_dmacachep
;
730 return csizep
->cs_cachep
;
733 struct kmem_cache
*kmem_find_general_cachep(size_t size
, gfp_t gfpflags
)
735 return __find_general_cachep(size
, gfpflags
);
737 EXPORT_SYMBOL(kmem_find_general_cachep
);
739 static size_t slab_mgmt_size(size_t nr_objs
, size_t align
)
741 return ALIGN(sizeof(struct slab
)+nr_objs
*sizeof(kmem_bufctl_t
), align
);
745 * Calculate the number of objects and left-over bytes for a given buffer size.
747 static void cache_estimate(unsigned long gfporder
, size_t buffer_size
,
748 size_t align
, int flags
, size_t *left_over
,
753 size_t slab_size
= PAGE_SIZE
<< gfporder
;
756 * The slab management structure can be either off the slab or
757 * on it. For the latter case, the memory allocated for a
761 * - One kmem_bufctl_t for each object
762 * - Padding to respect alignment of @align
763 * - @buffer_size bytes for each object
765 * If the slab management structure is off the slab, then the
766 * alignment will already be calculated into the size. Because
767 * the slabs are all pages aligned, the objects will be at the
768 * correct alignment when allocated.
770 if (flags
& CFLGS_OFF_SLAB
) {
772 nr_objs
= slab_size
/ buffer_size
;
774 if (nr_objs
> SLAB_LIMIT
)
775 nr_objs
= SLAB_LIMIT
;
778 * Ignore padding for the initial guess. The padding
779 * is at most @align-1 bytes, and @buffer_size is at
780 * least @align. In the worst case, this result will
781 * be one greater than the number of objects that fit
782 * into the memory allocation when taking the padding
785 nr_objs
= (slab_size
- sizeof(struct slab
)) /
786 (buffer_size
+ sizeof(kmem_bufctl_t
));
789 * This calculated number will be either the right
790 * amount, or one greater than what we want.
792 if (slab_mgmt_size(nr_objs
, align
) + nr_objs
*buffer_size
796 if (nr_objs
> SLAB_LIMIT
)
797 nr_objs
= SLAB_LIMIT
;
799 mgmt_size
= slab_mgmt_size(nr_objs
, align
);
802 *left_over
= slab_size
- nr_objs
*buffer_size
- mgmt_size
;
805 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
807 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
810 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
811 function
, cachep
->name
, msg
);
817 * Special reaping functions for NUMA systems called from cache_reap().
818 * These take care of doing round robin flushing of alien caches (containing
819 * objects freed on different nodes from which they were allocated) and the
820 * flushing of remote pcps by calling drain_node_pages.
822 static DEFINE_PER_CPU(unsigned long, reap_node
);
824 static void init_reap_node(int cpu
)
828 node
= next_node(cpu_to_node(cpu
), node_online_map
);
829 if (node
== MAX_NUMNODES
)
832 __get_cpu_var(reap_node
) = node
;
835 static void next_reap_node(void)
837 int node
= __get_cpu_var(reap_node
);
840 * Also drain per cpu pages on remote zones
842 if (node
!= numa_node_id())
843 drain_node_pages(node
);
845 node
= next_node(node
, node_online_map
);
846 if (unlikely(node
>= MAX_NUMNODES
))
847 node
= first_node(node_online_map
);
848 __get_cpu_var(reap_node
) = node
;
852 #define init_reap_node(cpu) do { } while (0)
853 #define next_reap_node(void) do { } while (0)
857 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
858 * via the workqueue/eventd.
859 * Add the CPU number into the expiration time to minimize the possibility of
860 * the CPUs getting into lockstep and contending for the global cache chain
863 static void __devinit
start_cpu_timer(int cpu
)
865 struct work_struct
*reap_work
= &per_cpu(reap_work
, cpu
);
868 * When this gets called from do_initcalls via cpucache_init(),
869 * init_workqueues() has already run, so keventd will be setup
872 if (keventd_up() && reap_work
->func
== NULL
) {
874 INIT_WORK(reap_work
, cache_reap
, NULL
);
875 schedule_delayed_work_on(cpu
, reap_work
, HZ
+ 3 * cpu
);
879 static struct array_cache
*alloc_arraycache(int node
, int entries
,
882 int memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
883 struct array_cache
*nc
= NULL
;
885 nc
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
889 nc
->batchcount
= batchcount
;
891 spin_lock_init(&nc
->lock
);
897 static void *__cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
899 static struct array_cache
**alloc_alien_cache(int node
, int limit
)
901 struct array_cache
**ac_ptr
;
902 int memsize
= sizeof(void *) * MAX_NUMNODES
;
907 ac_ptr
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
910 if (i
== node
|| !node_online(i
)) {
914 ac_ptr
[i
] = alloc_arraycache(node
, limit
, 0xbaadf00d);
916 for (i
--; i
<= 0; i
--)
926 static void free_alien_cache(struct array_cache
**ac_ptr
)
937 static void __drain_alien_cache(struct kmem_cache
*cachep
,
938 struct array_cache
*ac
, int node
)
940 struct kmem_list3
*rl3
= cachep
->nodelists
[node
];
943 spin_lock(&rl3
->list_lock
);
944 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
946 spin_unlock(&rl3
->list_lock
);
951 * Called from cache_reap() to regularly drain alien caches round robin.
953 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_list3
*l3
)
955 int node
= __get_cpu_var(reap_node
);
958 struct array_cache
*ac
= l3
->alien
[node
];
959 if (ac
&& ac
->avail
) {
960 spin_lock_irq(&ac
->lock
);
961 __drain_alien_cache(cachep
, ac
, node
);
962 spin_unlock_irq(&ac
->lock
);
967 static void drain_alien_cache(struct kmem_cache
*cachep
,
968 struct array_cache
**alien
)
971 struct array_cache
*ac
;
974 for_each_online_node(i
) {
977 spin_lock_irqsave(&ac
->lock
, flags
);
978 __drain_alien_cache(cachep
, ac
, i
);
979 spin_unlock_irqrestore(&ac
->lock
, flags
);
985 #define drain_alien_cache(cachep, alien) do { } while (0)
986 #define reap_alien(cachep, l3) do { } while (0)
988 static inline struct array_cache
**alloc_alien_cache(int node
, int limit
)
990 return (struct array_cache
**) 0x01020304ul
;
993 static inline void free_alien_cache(struct array_cache
**ac_ptr
)
999 static int __devinit
cpuup_callback(struct notifier_block
*nfb
,
1000 unsigned long action
, void *hcpu
)
1002 long cpu
= (long)hcpu
;
1003 struct kmem_cache
*cachep
;
1004 struct kmem_list3
*l3
= NULL
;
1005 int node
= cpu_to_node(cpu
);
1006 int memsize
= sizeof(struct kmem_list3
);
1009 case CPU_UP_PREPARE
:
1010 mutex_lock(&cache_chain_mutex
);
1012 * We need to do this right in the beginning since
1013 * alloc_arraycache's are going to use this list.
1014 * kmalloc_node allows us to add the slab to the right
1015 * kmem_list3 and not this cpu's kmem_list3
1018 list_for_each_entry(cachep
, &cache_chain
, next
) {
1020 * Set up the size64 kmemlist for cpu before we can
1021 * begin anything. Make sure some other cpu on this
1022 * node has not already allocated this
1024 if (!cachep
->nodelists
[node
]) {
1025 l3
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1028 kmem_list3_init(l3
);
1029 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
1030 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1033 * The l3s don't come and go as CPUs come and
1034 * go. cache_chain_mutex is sufficient
1037 cachep
->nodelists
[node
] = l3
;
1040 spin_lock_irq(&cachep
->nodelists
[node
]->list_lock
);
1041 cachep
->nodelists
[node
]->free_limit
=
1042 (1 + nr_cpus_node(node
)) *
1043 cachep
->batchcount
+ cachep
->num
;
1044 spin_unlock_irq(&cachep
->nodelists
[node
]->list_lock
);
1048 * Now we can go ahead with allocating the shared arrays and
1051 list_for_each_entry(cachep
, &cache_chain
, next
) {
1052 struct array_cache
*nc
;
1053 struct array_cache
*shared
;
1054 struct array_cache
**alien
;
1056 nc
= alloc_arraycache(node
, cachep
->limit
,
1057 cachep
->batchcount
);
1060 shared
= alloc_arraycache(node
,
1061 cachep
->shared
* cachep
->batchcount
,
1066 alien
= alloc_alien_cache(node
, cachep
->limit
);
1069 cachep
->array
[cpu
] = nc
;
1070 l3
= cachep
->nodelists
[node
];
1073 spin_lock_irq(&l3
->list_lock
);
1076 * We are serialised from CPU_DEAD or
1077 * CPU_UP_CANCELLED by the cpucontrol lock
1079 l3
->shared
= shared
;
1088 spin_unlock_irq(&l3
->list_lock
);
1090 free_alien_cache(alien
);
1092 mutex_unlock(&cache_chain_mutex
);
1095 start_cpu_timer(cpu
);
1097 #ifdef CONFIG_HOTPLUG_CPU
1100 * Even if all the cpus of a node are down, we don't free the
1101 * kmem_list3 of any cache. This to avoid a race between
1102 * cpu_down, and a kmalloc allocation from another cpu for
1103 * memory from the node of the cpu going down. The list3
1104 * structure is usually allocated from kmem_cache_create() and
1105 * gets destroyed at kmem_cache_destroy().
1108 case CPU_UP_CANCELED
:
1109 mutex_lock(&cache_chain_mutex
);
1110 list_for_each_entry(cachep
, &cache_chain
, next
) {
1111 struct array_cache
*nc
;
1112 struct array_cache
*shared
;
1113 struct array_cache
**alien
;
1116 mask
= node_to_cpumask(node
);
1117 /* cpu is dead; no one can alloc from it. */
1118 nc
= cachep
->array
[cpu
];
1119 cachep
->array
[cpu
] = NULL
;
1120 l3
= cachep
->nodelists
[node
];
1123 goto free_array_cache
;
1125 spin_lock_irq(&l3
->list_lock
);
1127 /* Free limit for this kmem_list3 */
1128 l3
->free_limit
-= cachep
->batchcount
;
1130 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
1132 if (!cpus_empty(mask
)) {
1133 spin_unlock_irq(&l3
->list_lock
);
1134 goto free_array_cache
;
1137 shared
= l3
->shared
;
1139 free_block(cachep
, l3
->shared
->entry
,
1140 l3
->shared
->avail
, node
);
1147 spin_unlock_irq(&l3
->list_lock
);
1151 drain_alien_cache(cachep
, alien
);
1152 free_alien_cache(alien
);
1158 * In the previous loop, all the objects were freed to
1159 * the respective cache's slabs, now we can go ahead and
1160 * shrink each nodelist to its limit.
1162 list_for_each_entry(cachep
, &cache_chain
, next
) {
1163 l3
= cachep
->nodelists
[node
];
1166 spin_lock_irq(&l3
->list_lock
);
1167 /* free slabs belonging to this node */
1168 __node_shrink(cachep
, node
);
1169 spin_unlock_irq(&l3
->list_lock
);
1171 mutex_unlock(&cache_chain_mutex
);
1177 mutex_unlock(&cache_chain_mutex
);
1181 static struct notifier_block cpucache_notifier
= { &cpuup_callback
, NULL
, 0 };
1184 * swap the static kmem_list3 with kmalloced memory
1186 static void init_list(struct kmem_cache
*cachep
, struct kmem_list3
*list
,
1189 struct kmem_list3
*ptr
;
1191 BUG_ON(cachep
->nodelists
[nodeid
] != list
);
1192 ptr
= kmalloc_node(sizeof(struct kmem_list3
), GFP_KERNEL
, nodeid
);
1195 local_irq_disable();
1196 memcpy(ptr
, list
, sizeof(struct kmem_list3
));
1197 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1198 cachep
->nodelists
[nodeid
] = ptr
;
1203 * Initialisation. Called after the page allocator have been initialised and
1204 * before smp_init().
1206 void __init
kmem_cache_init(void)
1209 struct cache_sizes
*sizes
;
1210 struct cache_names
*names
;
1214 for (i
= 0; i
< NUM_INIT_LISTS
; i
++) {
1215 kmem_list3_init(&initkmem_list3
[i
]);
1216 if (i
< MAX_NUMNODES
)
1217 cache_cache
.nodelists
[i
] = NULL
;
1221 * Fragmentation resistance on low memory - only use bigger
1222 * page orders on machines with more than 32MB of memory.
1224 if (num_physpages
> (32 << 20) >> PAGE_SHIFT
)
1225 slab_break_gfp_order
= BREAK_GFP_ORDER_HI
;
1227 /* Bootstrap is tricky, because several objects are allocated
1228 * from caches that do not exist yet:
1229 * 1) initialize the cache_cache cache: it contains the struct
1230 * kmem_cache structures of all caches, except cache_cache itself:
1231 * cache_cache is statically allocated.
1232 * Initially an __init data area is used for the head array and the
1233 * kmem_list3 structures, it's replaced with a kmalloc allocated
1234 * array at the end of the bootstrap.
1235 * 2) Create the first kmalloc cache.
1236 * The struct kmem_cache for the new cache is allocated normally.
1237 * An __init data area is used for the head array.
1238 * 3) Create the remaining kmalloc caches, with minimally sized
1240 * 4) Replace the __init data head arrays for cache_cache and the first
1241 * kmalloc cache with kmalloc allocated arrays.
1242 * 5) Replace the __init data for kmem_list3 for cache_cache and
1243 * the other cache's with kmalloc allocated memory.
1244 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1247 /* 1) create the cache_cache */
1248 INIT_LIST_HEAD(&cache_chain
);
1249 list_add(&cache_cache
.next
, &cache_chain
);
1250 cache_cache
.colour_off
= cache_line_size();
1251 cache_cache
.array
[smp_processor_id()] = &initarray_cache
.cache
;
1252 cache_cache
.nodelists
[numa_node_id()] = &initkmem_list3
[CACHE_CACHE
];
1254 cache_cache
.buffer_size
= ALIGN(cache_cache
.buffer_size
,
1257 for (order
= 0; order
< MAX_ORDER
; order
++) {
1258 cache_estimate(order
, cache_cache
.buffer_size
,
1259 cache_line_size(), 0, &left_over
, &cache_cache
.num
);
1260 if (cache_cache
.num
)
1263 if (!cache_cache
.num
)
1265 cache_cache
.gfporder
= order
;
1266 cache_cache
.colour
= left_over
/ cache_cache
.colour_off
;
1267 cache_cache
.slab_size
= ALIGN(cache_cache
.num
* sizeof(kmem_bufctl_t
) +
1268 sizeof(struct slab
), cache_line_size());
1270 /* 2+3) create the kmalloc caches */
1271 sizes
= malloc_sizes
;
1272 names
= cache_names
;
1275 * Initialize the caches that provide memory for the array cache and the
1276 * kmem_list3 structures first. Without this, further allocations will
1280 sizes
[INDEX_AC
].cs_cachep
= kmem_cache_create(names
[INDEX_AC
].name
,
1281 sizes
[INDEX_AC
].cs_size
,
1282 ARCH_KMALLOC_MINALIGN
,
1283 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1286 if (INDEX_AC
!= INDEX_L3
) {
1287 sizes
[INDEX_L3
].cs_cachep
=
1288 kmem_cache_create(names
[INDEX_L3
].name
,
1289 sizes
[INDEX_L3
].cs_size
,
1290 ARCH_KMALLOC_MINALIGN
,
1291 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1295 while (sizes
->cs_size
!= ULONG_MAX
) {
1297 * For performance, all the general caches are L1 aligned.
1298 * This should be particularly beneficial on SMP boxes, as it
1299 * eliminates "false sharing".
1300 * Note for systems short on memory removing the alignment will
1301 * allow tighter packing of the smaller caches.
1303 if (!sizes
->cs_cachep
) {
1304 sizes
->cs_cachep
= kmem_cache_create(names
->name
,
1306 ARCH_KMALLOC_MINALIGN
,
1307 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1311 /* Inc off-slab bufctl limit until the ceiling is hit. */
1312 if (!(OFF_SLAB(sizes
->cs_cachep
))) {
1313 offslab_limit
= sizes
->cs_size
- sizeof(struct slab
);
1314 offslab_limit
/= sizeof(kmem_bufctl_t
);
1317 sizes
->cs_dmacachep
= kmem_cache_create(names
->name_dma
,
1319 ARCH_KMALLOC_MINALIGN
,
1320 ARCH_KMALLOC_FLAGS
|SLAB_CACHE_DMA
|
1326 /* 4) Replace the bootstrap head arrays */
1330 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1332 local_irq_disable();
1333 BUG_ON(cpu_cache_get(&cache_cache
) != &initarray_cache
.cache
);
1334 memcpy(ptr
, cpu_cache_get(&cache_cache
),
1335 sizeof(struct arraycache_init
));
1336 cache_cache
.array
[smp_processor_id()] = ptr
;
1339 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1341 local_irq_disable();
1342 BUG_ON(cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
)
1343 != &initarray_generic
.cache
);
1344 memcpy(ptr
, cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
),
1345 sizeof(struct arraycache_init
));
1346 malloc_sizes
[INDEX_AC
].cs_cachep
->array
[smp_processor_id()] =
1350 /* 5) Replace the bootstrap kmem_list3's */
1353 /* Replace the static kmem_list3 structures for the boot cpu */
1354 init_list(&cache_cache
, &initkmem_list3
[CACHE_CACHE
],
1357 for_each_online_node(node
) {
1358 init_list(malloc_sizes
[INDEX_AC
].cs_cachep
,
1359 &initkmem_list3
[SIZE_AC
+ node
], node
);
1361 if (INDEX_AC
!= INDEX_L3
) {
1362 init_list(malloc_sizes
[INDEX_L3
].cs_cachep
,
1363 &initkmem_list3
[SIZE_L3
+ node
],
1369 /* 6) resize the head arrays to their final sizes */
1371 struct kmem_cache
*cachep
;
1372 mutex_lock(&cache_chain_mutex
);
1373 list_for_each_entry(cachep
, &cache_chain
, next
)
1374 enable_cpucache(cachep
);
1375 mutex_unlock(&cache_chain_mutex
);
1379 g_cpucache_up
= FULL
;
1382 * Register a cpu startup notifier callback that initializes
1383 * cpu_cache_get for all new cpus
1385 register_cpu_notifier(&cpucache_notifier
);
1388 * The reap timers are started later, with a module init call: That part
1389 * of the kernel is not yet operational.
1393 static int __init
cpucache_init(void)
1398 * Register the timers that return unneeded pages to the page allocator
1400 for_each_online_cpu(cpu
)
1401 start_cpu_timer(cpu
);
1404 __initcall(cpucache_init
);
1407 * Interface to system's page allocator. No need to hold the cache-lock.
1409 * If we requested dmaable memory, we will get it. Even if we
1410 * did not request dmaable memory, we might get it, but that
1411 * would be relatively rare and ignorable.
1413 static void *kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
1419 flags
|= cachep
->gfpflags
;
1420 page
= alloc_pages_node(nodeid
, flags
, cachep
->gfporder
);
1423 addr
= page_address(page
);
1425 i
= (1 << cachep
->gfporder
);
1426 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1427 atomic_add(i
, &slab_reclaim_pages
);
1428 add_page_state(nr_slab
, i
);
1430 __SetPageSlab(page
);
1437 * Interface to system's page release.
1439 static void kmem_freepages(struct kmem_cache
*cachep
, void *addr
)
1441 unsigned long i
= (1 << cachep
->gfporder
);
1442 struct page
*page
= virt_to_page(addr
);
1443 const unsigned long nr_freed
= i
;
1446 BUG_ON(!PageSlab(page
));
1447 __ClearPageSlab(page
);
1450 sub_page_state(nr_slab
, nr_freed
);
1451 if (current
->reclaim_state
)
1452 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1453 free_pages((unsigned long)addr
, cachep
->gfporder
);
1454 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1455 atomic_sub(1 << cachep
->gfporder
, &slab_reclaim_pages
);
1458 static void kmem_rcu_free(struct rcu_head
*head
)
1460 struct slab_rcu
*slab_rcu
= (struct slab_rcu
*)head
;
1461 struct kmem_cache
*cachep
= slab_rcu
->cachep
;
1463 kmem_freepages(cachep
, slab_rcu
->addr
);
1464 if (OFF_SLAB(cachep
))
1465 kmem_cache_free(cachep
->slabp_cache
, slab_rcu
);
1470 #ifdef CONFIG_DEBUG_PAGEALLOC
1471 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1472 unsigned long caller
)
1474 int size
= obj_size(cachep
);
1476 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1478 if (size
< 5 * sizeof(unsigned long))
1481 *addr
++ = 0x12345678;
1483 *addr
++ = smp_processor_id();
1484 size
-= 3 * sizeof(unsigned long);
1486 unsigned long *sptr
= &caller
;
1487 unsigned long svalue
;
1489 while (!kstack_end(sptr
)) {
1491 if (kernel_text_address(svalue
)) {
1493 size
-= sizeof(unsigned long);
1494 if (size
<= sizeof(unsigned long))
1500 *addr
++ = 0x87654321;
1504 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1506 int size
= obj_size(cachep
);
1507 addr
= &((char *)addr
)[obj_offset(cachep
)];
1509 memset(addr
, val
, size
);
1510 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1513 static void dump_line(char *data
, int offset
, int limit
)
1516 printk(KERN_ERR
"%03x:", offset
);
1517 for (i
= 0; i
< limit
; i
++)
1518 printk(" %02x", (unsigned char)data
[offset
+ i
]);
1525 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1530 if (cachep
->flags
& SLAB_RED_ZONE
) {
1531 printk(KERN_ERR
"Redzone: 0x%lx/0x%lx.\n",
1532 *dbg_redzone1(cachep
, objp
),
1533 *dbg_redzone2(cachep
, objp
));
1536 if (cachep
->flags
& SLAB_STORE_USER
) {
1537 printk(KERN_ERR
"Last user: [<%p>]",
1538 *dbg_userword(cachep
, objp
));
1539 print_symbol("(%s)",
1540 (unsigned long)*dbg_userword(cachep
, objp
));
1543 realobj
= (char *)objp
+ obj_offset(cachep
);
1544 size
= obj_size(cachep
);
1545 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1548 if (i
+ limit
> size
)
1550 dump_line(realobj
, i
, limit
);
1554 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1560 realobj
= (char *)objp
+ obj_offset(cachep
);
1561 size
= obj_size(cachep
);
1563 for (i
= 0; i
< size
; i
++) {
1564 char exp
= POISON_FREE
;
1567 if (realobj
[i
] != exp
) {
1573 "Slab corruption: start=%p, len=%d\n",
1575 print_objinfo(cachep
, objp
, 0);
1577 /* Hexdump the affected line */
1580 if (i
+ limit
> size
)
1582 dump_line(realobj
, i
, limit
);
1585 /* Limit to 5 lines */
1591 /* Print some data about the neighboring objects, if they
1594 struct slab
*slabp
= virt_to_slab(objp
);
1597 objnr
= obj_to_index(cachep
, slabp
, objp
);
1599 objp
= index_to_obj(cachep
, slabp
, objnr
- 1);
1600 realobj
= (char *)objp
+ obj_offset(cachep
);
1601 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
1603 print_objinfo(cachep
, objp
, 2);
1605 if (objnr
+ 1 < cachep
->num
) {
1606 objp
= index_to_obj(cachep
, slabp
, objnr
+ 1);
1607 realobj
= (char *)objp
+ obj_offset(cachep
);
1608 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
1610 print_objinfo(cachep
, objp
, 2);
1618 * slab_destroy_objs - destroy a slab and its objects
1619 * @cachep: cache pointer being destroyed
1620 * @slabp: slab pointer being destroyed
1622 * Call the registered destructor for each object in a slab that is being
1625 static void slab_destroy_objs(struct kmem_cache
*cachep
, struct slab
*slabp
)
1628 for (i
= 0; i
< cachep
->num
; i
++) {
1629 void *objp
= index_to_obj(cachep
, slabp
, i
);
1631 if (cachep
->flags
& SLAB_POISON
) {
1632 #ifdef CONFIG_DEBUG_PAGEALLOC
1633 if (cachep
->buffer_size
% PAGE_SIZE
== 0 &&
1635 kernel_map_pages(virt_to_page(objp
),
1636 cachep
->buffer_size
/ PAGE_SIZE
, 1);
1638 check_poison_obj(cachep
, objp
);
1640 check_poison_obj(cachep
, objp
);
1643 if (cachep
->flags
& SLAB_RED_ZONE
) {
1644 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1645 slab_error(cachep
, "start of a freed object "
1647 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1648 slab_error(cachep
, "end of a freed object "
1651 if (cachep
->dtor
&& !(cachep
->flags
& SLAB_POISON
))
1652 (cachep
->dtor
) (objp
+ obj_offset(cachep
), cachep
, 0);
1656 static void slab_destroy_objs(struct kmem_cache
*cachep
, struct slab
*slabp
)
1660 for (i
= 0; i
< cachep
->num
; i
++) {
1661 void *objp
= index_to_obj(cachep
, slabp
, i
);
1662 (cachep
->dtor
) (objp
, cachep
, 0);
1669 * slab_destroy - destroy and release all objects in a slab
1670 * @cachep: cache pointer being destroyed
1671 * @slabp: slab pointer being destroyed
1673 * Destroy all the objs in a slab, and release the mem back to the system.
1674 * Before calling the slab must have been unlinked from the cache. The
1675 * cache-lock is not held/needed.
1677 static void slab_destroy(struct kmem_cache
*cachep
, struct slab
*slabp
)
1679 void *addr
= slabp
->s_mem
- slabp
->colouroff
;
1681 slab_destroy_objs(cachep
, slabp
);
1682 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
1683 struct slab_rcu
*slab_rcu
;
1685 slab_rcu
= (struct slab_rcu
*)slabp
;
1686 slab_rcu
->cachep
= cachep
;
1687 slab_rcu
->addr
= addr
;
1688 call_rcu(&slab_rcu
->head
, kmem_rcu_free
);
1690 kmem_freepages(cachep
, addr
);
1691 if (OFF_SLAB(cachep
))
1692 kmem_cache_free(cachep
->slabp_cache
, slabp
);
1697 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1698 * size of kmem_list3.
1700 static void set_up_list3s(struct kmem_cache
*cachep
, int index
)
1704 for_each_online_node(node
) {
1705 cachep
->nodelists
[node
] = &initkmem_list3
[index
+ node
];
1706 cachep
->nodelists
[node
]->next_reap
= jiffies
+
1708 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1713 * calculate_slab_order - calculate size (page order) of slabs
1714 * @cachep: pointer to the cache that is being created
1715 * @size: size of objects to be created in this cache.
1716 * @align: required alignment for the objects.
1717 * @flags: slab allocation flags
1719 * Also calculates the number of objects per slab.
1721 * This could be made much more intelligent. For now, try to avoid using
1722 * high order pages for slabs. When the gfp() functions are more friendly
1723 * towards high-order requests, this should be changed.
1725 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
1726 size_t size
, size_t align
, unsigned long flags
)
1728 size_t left_over
= 0;
1731 for (gfporder
= 0; gfporder
<= MAX_GFP_ORDER
; gfporder
++) {
1735 cache_estimate(gfporder
, size
, align
, flags
, &remainder
, &num
);
1739 /* More than offslab_limit objects will cause problems */
1740 if ((flags
& CFLGS_OFF_SLAB
) && num
> offslab_limit
)
1743 /* Found something acceptable - save it away */
1745 cachep
->gfporder
= gfporder
;
1746 left_over
= remainder
;
1749 * A VFS-reclaimable slab tends to have most allocations
1750 * as GFP_NOFS and we really don't want to have to be allocating
1751 * higher-order pages when we are unable to shrink dcache.
1753 if (flags
& SLAB_RECLAIM_ACCOUNT
)
1757 * Large number of objects is good, but very large slabs are
1758 * currently bad for the gfp()s.
1760 if (gfporder
>= slab_break_gfp_order
)
1764 * Acceptable internal fragmentation?
1766 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
1772 static void setup_cpu_cache(struct kmem_cache
*cachep
)
1774 if (g_cpucache_up
== FULL
) {
1775 enable_cpucache(cachep
);
1778 if (g_cpucache_up
== NONE
) {
1780 * Note: the first kmem_cache_create must create the cache
1781 * that's used by kmalloc(24), otherwise the creation of
1782 * further caches will BUG().
1784 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
1787 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
1788 * the first cache, then we need to set up all its list3s,
1789 * otherwise the creation of further caches will BUG().
1791 set_up_list3s(cachep
, SIZE_AC
);
1792 if (INDEX_AC
== INDEX_L3
)
1793 g_cpucache_up
= PARTIAL_L3
;
1795 g_cpucache_up
= PARTIAL_AC
;
1797 cachep
->array
[smp_processor_id()] =
1798 kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1800 if (g_cpucache_up
== PARTIAL_AC
) {
1801 set_up_list3s(cachep
, SIZE_L3
);
1802 g_cpucache_up
= PARTIAL_L3
;
1805 for_each_online_node(node
) {
1806 cachep
->nodelists
[node
] =
1807 kmalloc_node(sizeof(struct kmem_list3
),
1809 BUG_ON(!cachep
->nodelists
[node
]);
1810 kmem_list3_init(cachep
->nodelists
[node
]);
1814 cachep
->nodelists
[numa_node_id()]->next_reap
=
1815 jiffies
+ REAPTIMEOUT_LIST3
+
1816 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1818 cpu_cache_get(cachep
)->avail
= 0;
1819 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
1820 cpu_cache_get(cachep
)->batchcount
= 1;
1821 cpu_cache_get(cachep
)->touched
= 0;
1822 cachep
->batchcount
= 1;
1823 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
1827 * kmem_cache_create - Create a cache.
1828 * @name: A string which is used in /proc/slabinfo to identify this cache.
1829 * @size: The size of objects to be created in this cache.
1830 * @align: The required alignment for the objects.
1831 * @flags: SLAB flags
1832 * @ctor: A constructor for the objects.
1833 * @dtor: A destructor for the objects.
1835 * Returns a ptr to the cache on success, NULL on failure.
1836 * Cannot be called within a int, but can be interrupted.
1837 * The @ctor is run when new pages are allocated by the cache
1838 * and the @dtor is run before the pages are handed back.
1840 * @name must be valid until the cache is destroyed. This implies that
1841 * the module calling this has to destroy the cache before getting unloaded.
1845 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1846 * to catch references to uninitialised memory.
1848 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1849 * for buffer overruns.
1851 * %SLAB_NO_REAP - Don't automatically reap this cache when we're under
1854 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1855 * cacheline. This can be beneficial if you're counting cycles as closely
1859 kmem_cache_create (const char *name
, size_t size
, size_t align
,
1860 unsigned long flags
,
1861 void (*ctor
)(void*, struct kmem_cache
*, unsigned long),
1862 void (*dtor
)(void*, struct kmem_cache
*, unsigned long))
1864 size_t left_over
, slab_size
, ralign
;
1865 struct kmem_cache
*cachep
= NULL
;
1866 struct list_head
*p
;
1869 * Sanity checks... these are all serious usage bugs.
1871 if (!name
|| in_interrupt() || (size
< BYTES_PER_WORD
) ||
1872 (size
> (1 << MAX_OBJ_ORDER
) * PAGE_SIZE
) || (dtor
&& !ctor
)) {
1873 printk(KERN_ERR
"%s: Early error in slab %s\n", __FUNCTION__
,
1879 * Prevent CPUs from coming and going.
1880 * lock_cpu_hotplug() nests outside cache_chain_mutex
1884 mutex_lock(&cache_chain_mutex
);
1886 list_for_each(p
, &cache_chain
) {
1887 struct kmem_cache
*pc
= list_entry(p
, struct kmem_cache
, next
);
1888 mm_segment_t old_fs
= get_fs();
1893 * This happens when the module gets unloaded and doesn't
1894 * destroy its slab cache and no-one else reuses the vmalloc
1895 * area of the module. Print a warning.
1898 res
= __get_user(tmp
, pc
->name
);
1901 printk("SLAB: cache with size %d has lost its name\n",
1906 if (!strcmp(pc
->name
, name
)) {
1907 printk("kmem_cache_create: duplicate cache %s\n", name
);
1914 WARN_ON(strchr(name
, ' ')); /* It confuses parsers */
1915 if ((flags
& SLAB_DEBUG_INITIAL
) && !ctor
) {
1916 /* No constructor, but inital state check requested */
1917 printk(KERN_ERR
"%s: No con, but init state check "
1918 "requested - %s\n", __FUNCTION__
, name
);
1919 flags
&= ~SLAB_DEBUG_INITIAL
;
1923 * Enable redzoning and last user accounting, except for caches with
1924 * large objects, if the increased size would increase the object size
1925 * above the next power of two: caches with object sizes just above a
1926 * power of two have a significant amount of internal fragmentation.
1928 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + 3 * BYTES_PER_WORD
))
1929 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
1930 if (!(flags
& SLAB_DESTROY_BY_RCU
))
1931 flags
|= SLAB_POISON
;
1933 if (flags
& SLAB_DESTROY_BY_RCU
)
1934 BUG_ON(flags
& SLAB_POISON
);
1936 if (flags
& SLAB_DESTROY_BY_RCU
)
1940 * Always checks flags, a caller might be expecting debug support which
1943 if (flags
& ~CREATE_MASK
)
1947 * Check that size is in terms of words. This is needed to avoid
1948 * unaligned accesses for some archs when redzoning is used, and makes
1949 * sure any on-slab bufctl's are also correctly aligned.
1951 if (size
& (BYTES_PER_WORD
- 1)) {
1952 size
+= (BYTES_PER_WORD
- 1);
1953 size
&= ~(BYTES_PER_WORD
- 1);
1956 /* calculate the final buffer alignment: */
1958 /* 1) arch recommendation: can be overridden for debug */
1959 if (flags
& SLAB_HWCACHE_ALIGN
) {
1961 * Default alignment: as specified by the arch code. Except if
1962 * an object is really small, then squeeze multiple objects into
1965 ralign
= cache_line_size();
1966 while (size
<= ralign
/ 2)
1969 ralign
= BYTES_PER_WORD
;
1971 /* 2) arch mandated alignment: disables debug if necessary */
1972 if (ralign
< ARCH_SLAB_MINALIGN
) {
1973 ralign
= ARCH_SLAB_MINALIGN
;
1974 if (ralign
> BYTES_PER_WORD
)
1975 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
1977 /* 3) caller mandated alignment: disables debug if necessary */
1978 if (ralign
< align
) {
1980 if (ralign
> BYTES_PER_WORD
)
1981 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
1984 * 4) Store it. Note that the debug code below can reduce
1985 * the alignment to BYTES_PER_WORD.
1989 /* Get cache's description obj. */
1990 cachep
= kmem_cache_alloc(&cache_cache
, SLAB_KERNEL
);
1993 memset(cachep
, 0, sizeof(struct kmem_cache
));
1996 cachep
->obj_size
= size
;
1998 if (flags
& SLAB_RED_ZONE
) {
1999 /* redzoning only works with word aligned caches */
2000 align
= BYTES_PER_WORD
;
2002 /* add space for red zone words */
2003 cachep
->obj_offset
+= BYTES_PER_WORD
;
2004 size
+= 2 * BYTES_PER_WORD
;
2006 if (flags
& SLAB_STORE_USER
) {
2007 /* user store requires word alignment and
2008 * one word storage behind the end of the real
2011 align
= BYTES_PER_WORD
;
2012 size
+= BYTES_PER_WORD
;
2014 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2015 if (size
>= malloc_sizes
[INDEX_L3
+ 1].cs_size
2016 && cachep
->obj_size
> cache_line_size() && size
< PAGE_SIZE
) {
2017 cachep
->obj_offset
+= PAGE_SIZE
- size
;
2023 /* Determine if the slab management is 'on' or 'off' slab. */
2024 if (size
>= (PAGE_SIZE
>> 3))
2026 * Size is large, assume best to place the slab management obj
2027 * off-slab (should allow better packing of objs).
2029 flags
|= CFLGS_OFF_SLAB
;
2031 size
= ALIGN(size
, align
);
2033 left_over
= calculate_slab_order(cachep
, size
, align
, flags
);
2036 printk("kmem_cache_create: couldn't create cache %s.\n", name
);
2037 kmem_cache_free(&cache_cache
, cachep
);
2041 slab_size
= ALIGN(cachep
->num
* sizeof(kmem_bufctl_t
)
2042 + sizeof(struct slab
), align
);
2045 * If the slab has been placed off-slab, and we have enough space then
2046 * move it on-slab. This is at the expense of any extra colouring.
2048 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= slab_size
) {
2049 flags
&= ~CFLGS_OFF_SLAB
;
2050 left_over
-= slab_size
;
2053 if (flags
& CFLGS_OFF_SLAB
) {
2054 /* really off slab. No need for manual alignment */
2056 cachep
->num
* sizeof(kmem_bufctl_t
) + sizeof(struct slab
);
2059 cachep
->colour_off
= cache_line_size();
2060 /* Offset must be a multiple of the alignment. */
2061 if (cachep
->colour_off
< align
)
2062 cachep
->colour_off
= align
;
2063 cachep
->colour
= left_over
/ cachep
->colour_off
;
2064 cachep
->slab_size
= slab_size
;
2065 cachep
->flags
= flags
;
2066 cachep
->gfpflags
= 0;
2067 if (flags
& SLAB_CACHE_DMA
)
2068 cachep
->gfpflags
|= GFP_DMA
;
2069 cachep
->buffer_size
= size
;
2071 if (flags
& CFLGS_OFF_SLAB
)
2072 cachep
->slabp_cache
= kmem_find_general_cachep(slab_size
, 0u);
2073 cachep
->ctor
= ctor
;
2074 cachep
->dtor
= dtor
;
2075 cachep
->name
= name
;
2078 setup_cpu_cache(cachep
);
2080 /* cache setup completed, link it into the list */
2081 list_add(&cachep
->next
, &cache_chain
);
2083 if (!cachep
&& (flags
& SLAB_PANIC
))
2084 panic("kmem_cache_create(): failed to create slab `%s'\n",
2086 mutex_unlock(&cache_chain_mutex
);
2087 unlock_cpu_hotplug();
2090 EXPORT_SYMBOL(kmem_cache_create
);
2093 static void check_irq_off(void)
2095 BUG_ON(!irqs_disabled());
2098 static void check_irq_on(void)
2100 BUG_ON(irqs_disabled());
2103 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2107 assert_spin_locked(&cachep
->nodelists
[numa_node_id()]->list_lock
);
2111 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2115 assert_spin_locked(&cachep
->nodelists
[node
]->list_lock
);
2120 #define check_irq_off() do { } while(0)
2121 #define check_irq_on() do { } while(0)
2122 #define check_spinlock_acquired(x) do { } while(0)
2123 #define check_spinlock_acquired_node(x, y) do { } while(0)
2127 * Waits for all CPUs to execute func().
2129 static void smp_call_function_all_cpus(void (*func
)(void *arg
), void *arg
)
2133 local_irq_disable();
2137 if (smp_call_function(func
, arg
, 1, 1))
2143 static void drain_array_locked(struct kmem_cache
*cachep
,
2144 struct array_cache
*ac
, int force
, int node
);
2146 static void do_drain(void *arg
)
2148 struct kmem_cache
*cachep
= arg
;
2149 struct array_cache
*ac
;
2150 int node
= numa_node_id();
2153 ac
= cpu_cache_get(cachep
);
2154 spin_lock(&cachep
->nodelists
[node
]->list_lock
);
2155 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
2156 spin_unlock(&cachep
->nodelists
[node
]->list_lock
);
2160 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2162 struct kmem_list3
*l3
;
2165 smp_call_function_all_cpus(do_drain
, cachep
);
2167 for_each_online_node(node
) {
2168 l3
= cachep
->nodelists
[node
];
2170 spin_lock_irq(&l3
->list_lock
);
2171 drain_array_locked(cachep
, l3
->shared
, 1, node
);
2172 spin_unlock_irq(&l3
->list_lock
);
2174 drain_alien_cache(cachep
, l3
->alien
);
2179 static int __node_shrink(struct kmem_cache
*cachep
, int node
)
2182 struct kmem_list3
*l3
= cachep
->nodelists
[node
];
2186 struct list_head
*p
;
2188 p
= l3
->slabs_free
.prev
;
2189 if (p
== &l3
->slabs_free
)
2192 slabp
= list_entry(l3
->slabs_free
.prev
, struct slab
, list
);
2197 list_del(&slabp
->list
);
2199 l3
->free_objects
-= cachep
->num
;
2200 spin_unlock_irq(&l3
->list_lock
);
2201 slab_destroy(cachep
, slabp
);
2202 spin_lock_irq(&l3
->list_lock
);
2204 ret
= !list_empty(&l3
->slabs_full
) || !list_empty(&l3
->slabs_partial
);
2208 static int __cache_shrink(struct kmem_cache
*cachep
)
2211 struct kmem_list3
*l3
;
2213 drain_cpu_caches(cachep
);
2216 for_each_online_node(i
) {
2217 l3
= cachep
->nodelists
[i
];
2219 spin_lock_irq(&l3
->list_lock
);
2220 ret
+= __node_shrink(cachep
, i
);
2221 spin_unlock_irq(&l3
->list_lock
);
2224 return (ret
? 1 : 0);
2228 * kmem_cache_shrink - Shrink a cache.
2229 * @cachep: The cache to shrink.
2231 * Releases as many slabs as possible for a cache.
2232 * To help debugging, a zero exit status indicates all slabs were released.
2234 int kmem_cache_shrink(struct kmem_cache
*cachep
)
2236 if (!cachep
|| in_interrupt())
2239 return __cache_shrink(cachep
);
2241 EXPORT_SYMBOL(kmem_cache_shrink
);
2244 * kmem_cache_destroy - delete a cache
2245 * @cachep: the cache to destroy
2247 * Remove a struct kmem_cache object from the slab cache.
2248 * Returns 0 on success.
2250 * It is expected this function will be called by a module when it is
2251 * unloaded. This will remove the cache completely, and avoid a duplicate
2252 * cache being allocated each time a module is loaded and unloaded, if the
2253 * module doesn't have persistent in-kernel storage across loads and unloads.
2255 * The cache must be empty before calling this function.
2257 * The caller must guarantee that noone will allocate memory from the cache
2258 * during the kmem_cache_destroy().
2260 int kmem_cache_destroy(struct kmem_cache
*cachep
)
2263 struct kmem_list3
*l3
;
2265 if (!cachep
|| in_interrupt())
2268 /* Don't let CPUs to come and go */
2271 /* Find the cache in the chain of caches. */
2272 mutex_lock(&cache_chain_mutex
);
2274 * the chain is never empty, cache_cache is never destroyed
2276 list_del(&cachep
->next
);
2277 mutex_unlock(&cache_chain_mutex
);
2279 if (__cache_shrink(cachep
)) {
2280 slab_error(cachep
, "Can't free all objects");
2281 mutex_lock(&cache_chain_mutex
);
2282 list_add(&cachep
->next
, &cache_chain
);
2283 mutex_unlock(&cache_chain_mutex
);
2284 unlock_cpu_hotplug();
2288 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
2291 for_each_online_cpu(i
)
2292 kfree(cachep
->array
[i
]);
2294 /* NUMA: free the list3 structures */
2295 for_each_online_node(i
) {
2296 l3
= cachep
->nodelists
[i
];
2299 free_alien_cache(l3
->alien
);
2303 kmem_cache_free(&cache_cache
, cachep
);
2304 unlock_cpu_hotplug();
2307 EXPORT_SYMBOL(kmem_cache_destroy
);
2309 /* Get the memory for a slab management obj. */
2310 static struct slab
*alloc_slabmgmt(struct kmem_cache
*cachep
, void *objp
,
2311 int colour_off
, gfp_t local_flags
)
2315 if (OFF_SLAB(cachep
)) {
2316 /* Slab management obj is off-slab. */
2317 slabp
= kmem_cache_alloc(cachep
->slabp_cache
, local_flags
);
2321 slabp
= objp
+ colour_off
;
2322 colour_off
+= cachep
->slab_size
;
2325 slabp
->colouroff
= colour_off
;
2326 slabp
->s_mem
= objp
+ colour_off
;
2330 static inline kmem_bufctl_t
*slab_bufctl(struct slab
*slabp
)
2332 return (kmem_bufctl_t
*) (slabp
+ 1);
2335 static void cache_init_objs(struct kmem_cache
*cachep
,
2336 struct slab
*slabp
, unsigned long ctor_flags
)
2340 for (i
= 0; i
< cachep
->num
; i
++) {
2341 void *objp
= index_to_obj(cachep
, slabp
, i
);
2343 /* need to poison the objs? */
2344 if (cachep
->flags
& SLAB_POISON
)
2345 poison_obj(cachep
, objp
, POISON_FREE
);
2346 if (cachep
->flags
& SLAB_STORE_USER
)
2347 *dbg_userword(cachep
, objp
) = NULL
;
2349 if (cachep
->flags
& SLAB_RED_ZONE
) {
2350 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2351 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2354 * Constructors are not allowed to allocate memory from the same
2355 * cache which they are a constructor for. Otherwise, deadlock.
2356 * They must also be threaded.
2358 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2359 cachep
->ctor(objp
+ obj_offset(cachep
), cachep
,
2362 if (cachep
->flags
& SLAB_RED_ZONE
) {
2363 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2364 slab_error(cachep
, "constructor overwrote the"
2365 " end of an object");
2366 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2367 slab_error(cachep
, "constructor overwrote the"
2368 " start of an object");
2370 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 &&
2371 OFF_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
)
2372 kernel_map_pages(virt_to_page(objp
),
2373 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2376 cachep
->ctor(objp
, cachep
, ctor_flags
);
2378 slab_bufctl(slabp
)[i
] = i
+ 1;
2380 slab_bufctl(slabp
)[i
- 1] = BUFCTL_END
;
2384 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2386 if (flags
& SLAB_DMA
)
2387 BUG_ON(!(cachep
->gfpflags
& GFP_DMA
));
2389 BUG_ON(cachep
->gfpflags
& GFP_DMA
);
2392 static void *slab_get_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2395 void *objp
= index_to_obj(cachep
, slabp
, slabp
->free
);
2399 next
= slab_bufctl(slabp
)[slabp
->free
];
2401 slab_bufctl(slabp
)[slabp
->free
] = BUFCTL_FREE
;
2402 WARN_ON(slabp
->nodeid
!= nodeid
);
2409 static void slab_put_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2410 void *objp
, int nodeid
)
2412 unsigned int objnr
= obj_to_index(cachep
, slabp
, objp
);
2415 /* Verify that the slab belongs to the intended node */
2416 WARN_ON(slabp
->nodeid
!= nodeid
);
2418 if (slab_bufctl(slabp
)[objnr
] != BUFCTL_FREE
) {
2419 printk(KERN_ERR
"slab: double free detected in cache "
2420 "'%s', objp %p\n", cachep
->name
, objp
);
2424 slab_bufctl(slabp
)[objnr
] = slabp
->free
;
2425 slabp
->free
= objnr
;
2429 static void set_slab_attr(struct kmem_cache
*cachep
, struct slab
*slabp
,
2435 /* Nasty!!!!!! I hope this is OK. */
2436 i
= 1 << cachep
->gfporder
;
2437 page
= virt_to_page(objp
);
2439 page_set_cache(page
, cachep
);
2440 page_set_slab(page
, slabp
);
2446 * Grow (by 1) the number of slabs within a cache. This is called by
2447 * kmem_cache_alloc() when there are no active objs left in a cache.
2449 static int cache_grow(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
2455 unsigned long ctor_flags
;
2456 struct kmem_list3
*l3
;
2459 * Be lazy and only check for valid flags here, keeping it out of the
2460 * critical path in kmem_cache_alloc().
2462 if (flags
& ~(SLAB_DMA
| SLAB_LEVEL_MASK
| SLAB_NO_GROW
))
2464 if (flags
& SLAB_NO_GROW
)
2467 ctor_flags
= SLAB_CTOR_CONSTRUCTOR
;
2468 local_flags
= (flags
& SLAB_LEVEL_MASK
);
2469 if (!(local_flags
& __GFP_WAIT
))
2471 * Not allowed to sleep. Need to tell a constructor about
2472 * this - it might need to know...
2474 ctor_flags
|= SLAB_CTOR_ATOMIC
;
2476 /* Take the l3 list lock to change the colour_next on this node */
2478 l3
= cachep
->nodelists
[nodeid
];
2479 spin_lock(&l3
->list_lock
);
2481 /* Get colour for the slab, and cal the next value. */
2482 offset
= l3
->colour_next
;
2484 if (l3
->colour_next
>= cachep
->colour
)
2485 l3
->colour_next
= 0;
2486 spin_unlock(&l3
->list_lock
);
2488 offset
*= cachep
->colour_off
;
2490 if (local_flags
& __GFP_WAIT
)
2494 * The test for missing atomic flag is performed here, rather than
2495 * the more obvious place, simply to reduce the critical path length
2496 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2497 * will eventually be caught here (where it matters).
2499 kmem_flagcheck(cachep
, flags
);
2502 * Get mem for the objs. Attempt to allocate a physical page from
2505 objp
= kmem_getpages(cachep
, flags
, nodeid
);
2509 /* Get slab management. */
2510 slabp
= alloc_slabmgmt(cachep
, objp
, offset
, local_flags
);
2514 slabp
->nodeid
= nodeid
;
2515 set_slab_attr(cachep
, slabp
, objp
);
2517 cache_init_objs(cachep
, slabp
, ctor_flags
);
2519 if (local_flags
& __GFP_WAIT
)
2520 local_irq_disable();
2522 spin_lock(&l3
->list_lock
);
2524 /* Make slab active. */
2525 list_add_tail(&slabp
->list
, &(l3
->slabs_free
));
2526 STATS_INC_GROWN(cachep
);
2527 l3
->free_objects
+= cachep
->num
;
2528 spin_unlock(&l3
->list_lock
);
2531 kmem_freepages(cachep
, objp
);
2533 if (local_flags
& __GFP_WAIT
)
2534 local_irq_disable();
2541 * Perform extra freeing checks:
2542 * - detect bad pointers.
2543 * - POISON/RED_ZONE checking
2544 * - destructor calls, for caches with POISON+dtor
2546 static void kfree_debugcheck(const void *objp
)
2550 if (!virt_addr_valid(objp
)) {
2551 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
2552 (unsigned long)objp
);
2555 page
= virt_to_page(objp
);
2556 if (!PageSlab(page
)) {
2557 printk(KERN_ERR
"kfree_debugcheck: bad ptr %lxh.\n",
2558 (unsigned long)objp
);
2563 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2570 objp
-= obj_offset(cachep
);
2571 kfree_debugcheck(objp
);
2572 page
= virt_to_page(objp
);
2574 if (page_get_cache(page
) != cachep
) {
2575 printk(KERN_ERR
"mismatch in kmem_cache_free: expected "
2576 "cache %p, got %p\n",
2577 page_get_cache(page
), cachep
);
2578 printk(KERN_ERR
"%p is %s.\n", cachep
, cachep
->name
);
2579 printk(KERN_ERR
"%p is %s.\n", page_get_cache(page
),
2580 page_get_cache(page
)->name
);
2583 slabp
= page_get_slab(page
);
2585 if (cachep
->flags
& SLAB_RED_ZONE
) {
2586 if (*dbg_redzone1(cachep
, objp
) != RED_ACTIVE
||
2587 *dbg_redzone2(cachep
, objp
) != RED_ACTIVE
) {
2588 slab_error(cachep
, "double free, or memory outside"
2589 " object was overwritten");
2590 printk(KERN_ERR
"%p: redzone 1:0x%lx, "
2591 "redzone 2:0x%lx.\n",
2592 objp
, *dbg_redzone1(cachep
, objp
),
2593 *dbg_redzone2(cachep
, objp
));
2595 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2596 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2598 if (cachep
->flags
& SLAB_STORE_USER
)
2599 *dbg_userword(cachep
, objp
) = caller
;
2601 objnr
= obj_to_index(cachep
, slabp
, objp
);
2603 BUG_ON(objnr
>= cachep
->num
);
2604 BUG_ON(objp
!= index_to_obj(cachep
, slabp
, objnr
));
2606 if (cachep
->flags
& SLAB_DEBUG_INITIAL
) {
2608 * Need to call the slab's constructor so the caller can
2609 * perform a verify of its state (debugging). Called without
2610 * the cache-lock held.
2612 cachep
->ctor(objp
+ obj_offset(cachep
),
2613 cachep
, SLAB_CTOR_CONSTRUCTOR
| SLAB_CTOR_VERIFY
);
2615 if (cachep
->flags
& SLAB_POISON
&& cachep
->dtor
) {
2616 /* we want to cache poison the object,
2617 * call the destruction callback
2619 cachep
->dtor(objp
+ obj_offset(cachep
), cachep
, 0);
2621 if (cachep
->flags
& SLAB_POISON
) {
2622 #ifdef CONFIG_DEBUG_PAGEALLOC
2623 if ((cachep
->buffer_size
% PAGE_SIZE
)==0 && OFF_SLAB(cachep
)) {
2624 store_stackinfo(cachep
, objp
, (unsigned long)caller
);
2625 kernel_map_pages(virt_to_page(objp
),
2626 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2628 poison_obj(cachep
, objp
, POISON_FREE
);
2631 poison_obj(cachep
, objp
, POISON_FREE
);
2637 static void check_slabp(struct kmem_cache
*cachep
, struct slab
*slabp
)
2642 /* Check slab's freelist to see if this obj is there. */
2643 for (i
= slabp
->free
; i
!= BUFCTL_END
; i
= slab_bufctl(slabp
)[i
]) {
2645 if (entries
> cachep
->num
|| i
>= cachep
->num
)
2648 if (entries
!= cachep
->num
- slabp
->inuse
) {
2650 printk(KERN_ERR
"slab: Internal list corruption detected in "
2651 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2652 cachep
->name
, cachep
->num
, slabp
, slabp
->inuse
);
2654 i
< sizeof(*slabp
) + cachep
->num
* sizeof(kmem_bufctl_t
);
2657 printk("\n%03x:", i
);
2658 printk(" %02x", ((unsigned char *)slabp
)[i
]);
2665 #define kfree_debugcheck(x) do { } while(0)
2666 #define cache_free_debugcheck(x,objp,z) (objp)
2667 #define check_slabp(x,y) do { } while(0)
2670 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
2673 struct kmem_list3
*l3
;
2674 struct array_cache
*ac
;
2677 ac
= cpu_cache_get(cachep
);
2679 batchcount
= ac
->batchcount
;
2680 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2682 * If there was little recent activity on this cache, then
2683 * perform only a partial refill. Otherwise we could generate
2686 batchcount
= BATCHREFILL_LIMIT
;
2688 l3
= cachep
->nodelists
[numa_node_id()];
2690 BUG_ON(ac
->avail
> 0 || !l3
);
2691 spin_lock(&l3
->list_lock
);
2694 struct array_cache
*shared_array
= l3
->shared
;
2695 if (shared_array
->avail
) {
2696 if (batchcount
> shared_array
->avail
)
2697 batchcount
= shared_array
->avail
;
2698 shared_array
->avail
-= batchcount
;
2699 ac
->avail
= batchcount
;
2701 &(shared_array
->entry
[shared_array
->avail
]),
2702 sizeof(void *) * batchcount
);
2703 shared_array
->touched
= 1;
2707 while (batchcount
> 0) {
2708 struct list_head
*entry
;
2710 /* Get slab alloc is to come from. */
2711 entry
= l3
->slabs_partial
.next
;
2712 if (entry
== &l3
->slabs_partial
) {
2713 l3
->free_touched
= 1;
2714 entry
= l3
->slabs_free
.next
;
2715 if (entry
== &l3
->slabs_free
)
2719 slabp
= list_entry(entry
, struct slab
, list
);
2720 check_slabp(cachep
, slabp
);
2721 check_spinlock_acquired(cachep
);
2722 while (slabp
->inuse
< cachep
->num
&& batchcount
--) {
2723 STATS_INC_ALLOCED(cachep
);
2724 STATS_INC_ACTIVE(cachep
);
2725 STATS_SET_HIGH(cachep
);
2727 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, slabp
,
2730 check_slabp(cachep
, slabp
);
2732 /* move slabp to correct slabp list: */
2733 list_del(&slabp
->list
);
2734 if (slabp
->free
== BUFCTL_END
)
2735 list_add(&slabp
->list
, &l3
->slabs_full
);
2737 list_add(&slabp
->list
, &l3
->slabs_partial
);
2741 l3
->free_objects
-= ac
->avail
;
2743 spin_unlock(&l3
->list_lock
);
2745 if (unlikely(!ac
->avail
)) {
2747 x
= cache_grow(cachep
, flags
, numa_node_id());
2749 /* cache_grow can reenable interrupts, then ac could change. */
2750 ac
= cpu_cache_get(cachep
);
2751 if (!x
&& ac
->avail
== 0) /* no objects in sight? abort */
2754 if (!ac
->avail
) /* objects refilled by interrupt? */
2758 return ac
->entry
[--ac
->avail
];
2761 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
2764 might_sleep_if(flags
& __GFP_WAIT
);
2766 kmem_flagcheck(cachep
, flags
);
2771 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
2772 gfp_t flags
, void *objp
, void *caller
)
2776 if (cachep
->flags
& SLAB_POISON
) {
2777 #ifdef CONFIG_DEBUG_PAGEALLOC
2778 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
2779 kernel_map_pages(virt_to_page(objp
),
2780 cachep
->buffer_size
/ PAGE_SIZE
, 1);
2782 check_poison_obj(cachep
, objp
);
2784 check_poison_obj(cachep
, objp
);
2786 poison_obj(cachep
, objp
, POISON_INUSE
);
2788 if (cachep
->flags
& SLAB_STORE_USER
)
2789 *dbg_userword(cachep
, objp
) = caller
;
2791 if (cachep
->flags
& SLAB_RED_ZONE
) {
2792 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
2793 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
2794 slab_error(cachep
, "double free, or memory outside"
2795 " object was overwritten");
2797 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
2798 objp
, *dbg_redzone1(cachep
, objp
),
2799 *dbg_redzone2(cachep
, objp
));
2801 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
2802 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
2804 objp
+= obj_offset(cachep
);
2805 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
) {
2806 unsigned long ctor_flags
= SLAB_CTOR_CONSTRUCTOR
;
2808 if (!(flags
& __GFP_WAIT
))
2809 ctor_flags
|= SLAB_CTOR_ATOMIC
;
2811 cachep
->ctor(objp
, cachep
, ctor_flags
);
2816 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2819 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
2822 struct array_cache
*ac
;
2825 if (unlikely(current
->mempolicy
&& !in_interrupt())) {
2826 int nid
= slab_node(current
->mempolicy
);
2828 if (nid
!= numa_node_id())
2829 return __cache_alloc_node(cachep
, flags
, nid
);
2834 ac
= cpu_cache_get(cachep
);
2835 if (likely(ac
->avail
)) {
2836 STATS_INC_ALLOCHIT(cachep
);
2838 objp
= ac
->entry
[--ac
->avail
];
2840 STATS_INC_ALLOCMISS(cachep
);
2841 objp
= cache_alloc_refill(cachep
, flags
);
2846 static __always_inline
void *__cache_alloc(struct kmem_cache
*cachep
,
2847 gfp_t flags
, void *caller
)
2849 unsigned long save_flags
;
2852 cache_alloc_debugcheck_before(cachep
, flags
);
2854 local_irq_save(save_flags
);
2855 objp
= ____cache_alloc(cachep
, flags
);
2856 local_irq_restore(save_flags
);
2857 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
,
2865 * A interface to enable slab creation on nodeid
2867 static void *__cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
2870 struct list_head
*entry
;
2872 struct kmem_list3
*l3
;
2876 l3
= cachep
->nodelists
[nodeid
];
2881 spin_lock(&l3
->list_lock
);
2882 entry
= l3
->slabs_partial
.next
;
2883 if (entry
== &l3
->slabs_partial
) {
2884 l3
->free_touched
= 1;
2885 entry
= l3
->slabs_free
.next
;
2886 if (entry
== &l3
->slabs_free
)
2890 slabp
= list_entry(entry
, struct slab
, list
);
2891 check_spinlock_acquired_node(cachep
, nodeid
);
2892 check_slabp(cachep
, slabp
);
2894 STATS_INC_NODEALLOCS(cachep
);
2895 STATS_INC_ACTIVE(cachep
);
2896 STATS_SET_HIGH(cachep
);
2898 BUG_ON(slabp
->inuse
== cachep
->num
);
2900 obj
= slab_get_obj(cachep
, slabp
, nodeid
);
2901 check_slabp(cachep
, slabp
);
2903 /* move slabp to correct slabp list: */
2904 list_del(&slabp
->list
);
2906 if (slabp
->free
== BUFCTL_END
)
2907 list_add(&slabp
->list
, &l3
->slabs_full
);
2909 list_add(&slabp
->list
, &l3
->slabs_partial
);
2911 spin_unlock(&l3
->list_lock
);
2915 spin_unlock(&l3
->list_lock
);
2916 x
= cache_grow(cachep
, flags
, nodeid
);
2928 * Caller needs to acquire correct kmem_list's list_lock
2930 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int nr_objects
,
2934 struct kmem_list3
*l3
;
2936 for (i
= 0; i
< nr_objects
; i
++) {
2937 void *objp
= objpp
[i
];
2940 slabp
= virt_to_slab(objp
);
2941 l3
= cachep
->nodelists
[node
];
2942 list_del(&slabp
->list
);
2943 check_spinlock_acquired_node(cachep
, node
);
2944 check_slabp(cachep
, slabp
);
2945 slab_put_obj(cachep
, slabp
, objp
, node
);
2946 STATS_DEC_ACTIVE(cachep
);
2948 check_slabp(cachep
, slabp
);
2950 /* fixup slab chains */
2951 if (slabp
->inuse
== 0) {
2952 if (l3
->free_objects
> l3
->free_limit
) {
2953 l3
->free_objects
-= cachep
->num
;
2954 slab_destroy(cachep
, slabp
);
2956 list_add(&slabp
->list
, &l3
->slabs_free
);
2959 /* Unconditionally move a slab to the end of the
2960 * partial list on free - maximum time for the
2961 * other objects to be freed, too.
2963 list_add_tail(&slabp
->list
, &l3
->slabs_partial
);
2968 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
2971 struct kmem_list3
*l3
;
2972 int node
= numa_node_id();
2974 batchcount
= ac
->batchcount
;
2976 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
2979 l3
= cachep
->nodelists
[node
];
2980 spin_lock(&l3
->list_lock
);
2982 struct array_cache
*shared_array
= l3
->shared
;
2983 int max
= shared_array
->limit
- shared_array
->avail
;
2985 if (batchcount
> max
)
2987 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
2988 ac
->entry
, sizeof(void *) * batchcount
);
2989 shared_array
->avail
+= batchcount
;
2994 free_block(cachep
, ac
->entry
, batchcount
, node
);
2999 struct list_head
*p
;
3001 p
= l3
->slabs_free
.next
;
3002 while (p
!= &(l3
->slabs_free
)) {
3005 slabp
= list_entry(p
, struct slab
, list
);
3006 BUG_ON(slabp
->inuse
);
3011 STATS_SET_FREEABLE(cachep
, i
);
3014 spin_unlock(&l3
->list_lock
);
3015 ac
->avail
-= batchcount
;
3016 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3020 * Release an obj back to its cache. If the obj has a constructed state, it must
3021 * be in this state _before_ it is released. Called with disabled ints.
3023 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
)
3025 struct array_cache
*ac
= cpu_cache_get(cachep
);
3028 objp
= cache_free_debugcheck(cachep
, objp
, __builtin_return_address(0));
3030 /* Make sure we are not freeing a object from another
3031 * node to the array cache on this cpu.
3036 slabp
= virt_to_slab(objp
);
3037 if (unlikely(slabp
->nodeid
!= numa_node_id())) {
3038 struct array_cache
*alien
= NULL
;
3039 int nodeid
= slabp
->nodeid
;
3040 struct kmem_list3
*l3
;
3042 l3
= cachep
->nodelists
[numa_node_id()];
3043 STATS_INC_NODEFREES(cachep
);
3044 if (l3
->alien
&& l3
->alien
[nodeid
]) {
3045 alien
= l3
->alien
[nodeid
];
3046 spin_lock(&alien
->lock
);
3047 if (unlikely(alien
->avail
== alien
->limit
))
3048 __drain_alien_cache(cachep
,
3050 alien
->entry
[alien
->avail
++] = objp
;
3051 spin_unlock(&alien
->lock
);
3053 spin_lock(&(cachep
->nodelists
[nodeid
])->
3055 free_block(cachep
, &objp
, 1, nodeid
);
3056 spin_unlock(&(cachep
->nodelists
[nodeid
])->
3063 if (likely(ac
->avail
< ac
->limit
)) {
3064 STATS_INC_FREEHIT(cachep
);
3065 ac
->entry
[ac
->avail
++] = objp
;
3068 STATS_INC_FREEMISS(cachep
);
3069 cache_flusharray(cachep
, ac
);
3070 ac
->entry
[ac
->avail
++] = objp
;
3075 * kmem_cache_alloc - Allocate an object
3076 * @cachep: The cache to allocate from.
3077 * @flags: See kmalloc().
3079 * Allocate an object from this cache. The flags are only relevant
3080 * if the cache has no available objects.
3082 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3084 return __cache_alloc(cachep
, flags
, __builtin_return_address(0));
3086 EXPORT_SYMBOL(kmem_cache_alloc
);
3089 * kmem_ptr_validate - check if an untrusted pointer might
3091 * @cachep: the cache we're checking against
3092 * @ptr: pointer to validate
3094 * This verifies that the untrusted pointer looks sane:
3095 * it is _not_ a guarantee that the pointer is actually
3096 * part of the slab cache in question, but it at least
3097 * validates that the pointer can be dereferenced and
3098 * looks half-way sane.
3100 * Currently only used for dentry validation.
3102 int fastcall
kmem_ptr_validate(struct kmem_cache
*cachep
, void *ptr
)
3104 unsigned long addr
= (unsigned long)ptr
;
3105 unsigned long min_addr
= PAGE_OFFSET
;
3106 unsigned long align_mask
= BYTES_PER_WORD
- 1;
3107 unsigned long size
= cachep
->buffer_size
;
3110 if (unlikely(addr
< min_addr
))
3112 if (unlikely(addr
> (unsigned long)high_memory
- size
))
3114 if (unlikely(addr
& align_mask
))
3116 if (unlikely(!kern_addr_valid(addr
)))
3118 if (unlikely(!kern_addr_valid(addr
+ size
- 1)))
3120 page
= virt_to_page(ptr
);
3121 if (unlikely(!PageSlab(page
)))
3123 if (unlikely(page_get_cache(page
) != cachep
))
3132 * kmem_cache_alloc_node - Allocate an object on the specified node
3133 * @cachep: The cache to allocate from.
3134 * @flags: See kmalloc().
3135 * @nodeid: node number of the target node.
3137 * Identical to kmem_cache_alloc, except that this function is slow
3138 * and can sleep. And it will allocate memory on the given node, which
3139 * can improve the performance for cpu bound structures.
3140 * New and improved: it will now make sure that the object gets
3141 * put on the correct node list so that there is no false sharing.
3143 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3145 unsigned long save_flags
;
3148 cache_alloc_debugcheck_before(cachep
, flags
);
3149 local_irq_save(save_flags
);
3151 if (nodeid
== -1 || nodeid
== numa_node_id() ||
3152 !cachep
->nodelists
[nodeid
])
3153 ptr
= ____cache_alloc(cachep
, flags
);
3155 ptr
= __cache_alloc_node(cachep
, flags
, nodeid
);
3156 local_irq_restore(save_flags
);
3158 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
,
3159 __builtin_return_address(0));
3163 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3165 void *kmalloc_node(size_t size
, gfp_t flags
, int node
)
3167 struct kmem_cache
*cachep
;
3169 cachep
= kmem_find_general_cachep(size
, flags
);
3170 if (unlikely(cachep
== NULL
))
3172 return kmem_cache_alloc_node(cachep
, flags
, node
);
3174 EXPORT_SYMBOL(kmalloc_node
);
3178 * kmalloc - allocate memory
3179 * @size: how many bytes of memory are required.
3180 * @flags: the type of memory to allocate.
3181 * @caller: function caller for debug tracking of the caller
3183 * kmalloc is the normal method of allocating memory
3186 * The @flags argument may be one of:
3188 * %GFP_USER - Allocate memory on behalf of user. May sleep.
3190 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
3192 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
3194 * Additionally, the %GFP_DMA flag may be set to indicate the memory
3195 * must be suitable for DMA. This can mean different things on different
3196 * platforms. For example, on i386, it means that the memory must come
3197 * from the first 16MB.
3199 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3202 struct kmem_cache
*cachep
;
3204 /* If you want to save a few bytes .text space: replace
3206 * Then kmalloc uses the uninlined functions instead of the inline
3209 cachep
= __find_general_cachep(size
, flags
);
3210 if (unlikely(cachep
== NULL
))
3212 return __cache_alloc(cachep
, flags
, caller
);
3215 #ifndef CONFIG_DEBUG_SLAB
3217 void *__kmalloc(size_t size
, gfp_t flags
)
3219 return __do_kmalloc(size
, flags
, NULL
);
3221 EXPORT_SYMBOL(__kmalloc
);
3225 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, void *caller
)
3227 return __do_kmalloc(size
, flags
, caller
);
3229 EXPORT_SYMBOL(__kmalloc_track_caller
);
3235 * __alloc_percpu - allocate one copy of the object for every present
3236 * cpu in the system, zeroing them.
3237 * Objects should be dereferenced using the per_cpu_ptr macro only.
3239 * @size: how many bytes of memory are required.
3241 void *__alloc_percpu(size_t size
)
3244 struct percpu_data
*pdata
= kmalloc(sizeof(*pdata
), GFP_KERNEL
);
3250 * Cannot use for_each_online_cpu since a cpu may come online
3251 * and we have no way of figuring out how to fix the array
3252 * that we have allocated then....
3255 int node
= cpu_to_node(i
);
3257 if (node_online(node
))
3258 pdata
->ptrs
[i
] = kmalloc_node(size
, GFP_KERNEL
, node
);
3260 pdata
->ptrs
[i
] = kmalloc(size
, GFP_KERNEL
);
3262 if (!pdata
->ptrs
[i
])
3264 memset(pdata
->ptrs
[i
], 0, size
);
3267 /* Catch derefs w/o wrappers */
3268 return (void *)(~(unsigned long)pdata
);
3272 if (!cpu_possible(i
))
3274 kfree(pdata
->ptrs
[i
]);
3279 EXPORT_SYMBOL(__alloc_percpu
);
3283 * kmem_cache_free - Deallocate an object
3284 * @cachep: The cache the allocation was from.
3285 * @objp: The previously allocated object.
3287 * Free an object which was previously allocated from this
3290 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3292 unsigned long flags
;
3294 local_irq_save(flags
);
3295 __cache_free(cachep
, objp
);
3296 local_irq_restore(flags
);
3298 EXPORT_SYMBOL(kmem_cache_free
);
3301 * kfree - free previously allocated memory
3302 * @objp: pointer returned by kmalloc.
3304 * If @objp is NULL, no operation is performed.
3306 * Don't free memory not originally allocated by kmalloc()
3307 * or you will run into trouble.
3309 void kfree(const void *objp
)
3311 struct kmem_cache
*c
;
3312 unsigned long flags
;
3314 if (unlikely(!objp
))
3316 local_irq_save(flags
);
3317 kfree_debugcheck(objp
);
3318 c
= virt_to_cache(objp
);
3319 mutex_debug_check_no_locks_freed(objp
, obj_size(c
));
3320 __cache_free(c
, (void *)objp
);
3321 local_irq_restore(flags
);
3323 EXPORT_SYMBOL(kfree
);
3327 * free_percpu - free previously allocated percpu memory
3328 * @objp: pointer returned by alloc_percpu.
3330 * Don't free memory not originally allocated by alloc_percpu()
3331 * The complemented objp is to check for that.
3333 void free_percpu(const void *objp
)
3336 struct percpu_data
*p
= (struct percpu_data
*)(~(unsigned long)objp
);
3339 * We allocate for all cpus so we cannot use for online cpu here.
3345 EXPORT_SYMBOL(free_percpu
);
3348 unsigned int kmem_cache_size(struct kmem_cache
*cachep
)
3350 return obj_size(cachep
);
3352 EXPORT_SYMBOL(kmem_cache_size
);
3354 const char *kmem_cache_name(struct kmem_cache
*cachep
)
3356 return cachep
->name
;
3358 EXPORT_SYMBOL_GPL(kmem_cache_name
);
3361 * This initializes kmem_list3 for all nodes.
3363 static int alloc_kmemlist(struct kmem_cache
*cachep
)
3366 struct kmem_list3
*l3
;
3369 for_each_online_node(node
) {
3370 struct array_cache
*nc
= NULL
, *new;
3371 struct array_cache
**new_alien
= NULL
;
3373 new_alien
= alloc_alien_cache(node
, cachep
->limit
);
3377 new = alloc_arraycache(node
, cachep
->shared
*cachep
->batchcount
,
3381 l3
= cachep
->nodelists
[node
];
3383 spin_lock_irq(&l3
->list_lock
);
3385 nc
= cachep
->nodelists
[node
]->shared
;
3387 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
3390 if (!cachep
->nodelists
[node
]->alien
) {
3391 l3
->alien
= new_alien
;
3394 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3395 cachep
->batchcount
+ cachep
->num
;
3396 spin_unlock_irq(&l3
->list_lock
);
3398 free_alien_cache(new_alien
);
3401 l3
= kmalloc_node(sizeof(struct kmem_list3
), GFP_KERNEL
, node
);
3405 kmem_list3_init(l3
);
3406 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
3407 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
3409 l3
->alien
= new_alien
;
3410 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3411 cachep
->batchcount
+ cachep
->num
;
3412 cachep
->nodelists
[node
] = l3
;
3420 struct ccupdate_struct
{
3421 struct kmem_cache
*cachep
;
3422 struct array_cache
*new[NR_CPUS
];
3425 static void do_ccupdate_local(void *info
)
3427 struct ccupdate_struct
*new = info
;
3428 struct array_cache
*old
;
3431 old
= cpu_cache_get(new->cachep
);
3433 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
3434 new->new[smp_processor_id()] = old
;
3437 /* Always called with the cache_chain_mutex held */
3438 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3439 int batchcount
, int shared
)
3441 struct ccupdate_struct
new;
3444 memset(&new.new, 0, sizeof(new.new));
3445 for_each_online_cpu(i
) {
3446 new.new[i
] = alloc_arraycache(cpu_to_node(i
), limit
,
3449 for (i
--; i
>= 0; i
--)
3454 new.cachep
= cachep
;
3456 smp_call_function_all_cpus(do_ccupdate_local
, (void *)&new);
3459 cachep
->batchcount
= batchcount
;
3460 cachep
->limit
= limit
;
3461 cachep
->shared
= shared
;
3463 for_each_online_cpu(i
) {
3464 struct array_cache
*ccold
= new.new[i
];
3467 spin_lock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3468 free_block(cachep
, ccold
->entry
, ccold
->avail
, cpu_to_node(i
));
3469 spin_unlock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3473 err
= alloc_kmemlist(cachep
);
3475 printk(KERN_ERR
"alloc_kmemlist failed for %s, error %d.\n",
3476 cachep
->name
, -err
);
3482 /* Called with cache_chain_mutex held always */
3483 static void enable_cpucache(struct kmem_cache
*cachep
)
3489 * The head array serves three purposes:
3490 * - create a LIFO ordering, i.e. return objects that are cache-warm
3491 * - reduce the number of spinlock operations.
3492 * - reduce the number of linked list operations on the slab and
3493 * bufctl chains: array operations are cheaper.
3494 * The numbers are guessed, we should auto-tune as described by
3497 if (cachep
->buffer_size
> 131072)
3499 else if (cachep
->buffer_size
> PAGE_SIZE
)
3501 else if (cachep
->buffer_size
> 1024)
3503 else if (cachep
->buffer_size
> 256)
3509 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3510 * allocation behaviour: Most allocs on one cpu, most free operations
3511 * on another cpu. For these cases, an efficient object passing between
3512 * cpus is necessary. This is provided by a shared array. The array
3513 * replaces Bonwick's magazine layer.
3514 * On uniprocessor, it's functionally equivalent (but less efficient)
3515 * to a larger limit. Thus disabled by default.
3519 if (cachep
->buffer_size
<= PAGE_SIZE
)
3525 * With debugging enabled, large batchcount lead to excessively long
3526 * periods with disabled local interrupts. Limit the batchcount
3531 err
= do_tune_cpucache(cachep
, limit
, (limit
+ 1) / 2, shared
);
3533 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
3534 cachep
->name
, -err
);
3537 static void drain_array_locked(struct kmem_cache
*cachep
,
3538 struct array_cache
*ac
, int force
, int node
)
3542 check_spinlock_acquired_node(cachep
, node
);
3543 if (ac
->touched
&& !force
) {
3545 } else if (ac
->avail
) {
3546 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
3547 if (tofree
> ac
->avail
)
3548 tofree
= (ac
->avail
+ 1) / 2;
3549 free_block(cachep
, ac
->entry
, tofree
, node
);
3550 ac
->avail
-= tofree
;
3551 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
3552 sizeof(void *) * ac
->avail
);
3557 * cache_reap - Reclaim memory from caches.
3558 * @unused: unused parameter
3560 * Called from workqueue/eventd every few seconds.
3562 * - clear the per-cpu caches for this CPU.
3563 * - return freeable pages to the main free memory pool.
3565 * If we cannot acquire the cache chain mutex then just give up - we'll try
3566 * again on the next iteration.
3568 static void cache_reap(void *unused
)
3570 struct list_head
*walk
;
3571 struct kmem_list3
*l3
;
3573 if (!mutex_trylock(&cache_chain_mutex
)) {
3574 /* Give up. Setup the next iteration. */
3575 schedule_delayed_work(&__get_cpu_var(reap_work
),
3580 list_for_each(walk
, &cache_chain
) {
3581 struct kmem_cache
*searchp
;
3582 struct list_head
*p
;
3586 searchp
= list_entry(walk
, struct kmem_cache
, next
);
3588 if (searchp
->flags
& SLAB_NO_REAP
)
3593 l3
= searchp
->nodelists
[numa_node_id()];
3594 reap_alien(searchp
, l3
);
3595 spin_lock_irq(&l3
->list_lock
);
3597 drain_array_locked(searchp
, cpu_cache_get(searchp
), 0,
3600 if (time_after(l3
->next_reap
, jiffies
))
3603 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
;
3606 drain_array_locked(searchp
, l3
->shared
, 0,
3609 if (l3
->free_touched
) {
3610 l3
->free_touched
= 0;
3614 tofree
= (l3
->free_limit
+ 5 * searchp
->num
- 1) /
3617 p
= l3
->slabs_free
.next
;
3618 if (p
== &(l3
->slabs_free
))
3621 slabp
= list_entry(p
, struct slab
, list
);
3622 BUG_ON(slabp
->inuse
);
3623 list_del(&slabp
->list
);
3624 STATS_INC_REAPED(searchp
);
3627 * Safe to drop the lock. The slab is no longer linked
3628 * to the cache. searchp cannot disappear, we hold
3631 l3
->free_objects
-= searchp
->num
;
3632 spin_unlock_irq(&l3
->list_lock
);
3633 slab_destroy(searchp
, slabp
);
3634 spin_lock_irq(&l3
->list_lock
);
3635 } while (--tofree
> 0);
3637 spin_unlock_irq(&l3
->list_lock
);
3642 mutex_unlock(&cache_chain_mutex
);
3644 /* Set up the next iteration */
3645 schedule_delayed_work(&__get_cpu_var(reap_work
), REAPTIMEOUT_CPUC
);
3648 #ifdef CONFIG_PROC_FS
3650 static void print_slabinfo_header(struct seq_file
*m
)
3653 * Output format version, so at least we can change it
3654 * without _too_ many complaints.
3657 seq_puts(m
, "slabinfo - version: 2.1 (statistics)\n");
3659 seq_puts(m
, "slabinfo - version: 2.1\n");
3661 seq_puts(m
, "# name <active_objs> <num_objs> <objsize> "
3662 "<objperslab> <pagesperslab>");
3663 seq_puts(m
, " : tunables <limit> <batchcount> <sharedfactor>");
3664 seq_puts(m
, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3666 seq_puts(m
, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3667 "<error> <maxfreeable> <nodeallocs> <remotefrees>");
3668 seq_puts(m
, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3673 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
3676 struct list_head
*p
;
3678 mutex_lock(&cache_chain_mutex
);
3680 print_slabinfo_header(m
);
3681 p
= cache_chain
.next
;
3684 if (p
== &cache_chain
)
3687 return list_entry(p
, struct kmem_cache
, next
);
3690 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
3692 struct kmem_cache
*cachep
= p
;
3694 return cachep
->next
.next
== &cache_chain
?
3695 NULL
: list_entry(cachep
->next
.next
, struct kmem_cache
, next
);
3698 static void s_stop(struct seq_file
*m
, void *p
)
3700 mutex_unlock(&cache_chain_mutex
);
3703 static int s_show(struct seq_file
*m
, void *p
)
3705 struct kmem_cache
*cachep
= p
;
3706 struct list_head
*q
;
3708 unsigned long active_objs
;
3709 unsigned long num_objs
;
3710 unsigned long active_slabs
= 0;
3711 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
3715 struct kmem_list3
*l3
;
3719 for_each_online_node(node
) {
3720 l3
= cachep
->nodelists
[node
];
3725 spin_lock_irq(&l3
->list_lock
);
3727 list_for_each(q
, &l3
->slabs_full
) {
3728 slabp
= list_entry(q
, struct slab
, list
);
3729 if (slabp
->inuse
!= cachep
->num
&& !error
)
3730 error
= "slabs_full accounting error";
3731 active_objs
+= cachep
->num
;
3734 list_for_each(q
, &l3
->slabs_partial
) {
3735 slabp
= list_entry(q
, struct slab
, list
);
3736 if (slabp
->inuse
== cachep
->num
&& !error
)
3737 error
= "slabs_partial inuse accounting error";
3738 if (!slabp
->inuse
&& !error
)
3739 error
= "slabs_partial/inuse accounting error";
3740 active_objs
+= slabp
->inuse
;
3743 list_for_each(q
, &l3
->slabs_free
) {
3744 slabp
= list_entry(q
, struct slab
, list
);
3745 if (slabp
->inuse
&& !error
)
3746 error
= "slabs_free/inuse accounting error";
3749 free_objects
+= l3
->free_objects
;
3751 shared_avail
+= l3
->shared
->avail
;
3753 spin_unlock_irq(&l3
->list_lock
);
3755 num_slabs
+= active_slabs
;
3756 num_objs
= num_slabs
* cachep
->num
;
3757 if (num_objs
- active_objs
!= free_objects
&& !error
)
3758 error
= "free_objects accounting error";
3760 name
= cachep
->name
;
3762 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
3764 seq_printf(m
, "%-17s %6lu %6lu %6u %4u %4d",
3765 name
, active_objs
, num_objs
, cachep
->buffer_size
,
3766 cachep
->num
, (1 << cachep
->gfporder
));
3767 seq_printf(m
, " : tunables %4u %4u %4u",
3768 cachep
->limit
, cachep
->batchcount
, cachep
->shared
);
3769 seq_printf(m
, " : slabdata %6lu %6lu %6lu",
3770 active_slabs
, num_slabs
, shared_avail
);
3773 unsigned long high
= cachep
->high_mark
;
3774 unsigned long allocs
= cachep
->num_allocations
;
3775 unsigned long grown
= cachep
->grown
;
3776 unsigned long reaped
= cachep
->reaped
;
3777 unsigned long errors
= cachep
->errors
;
3778 unsigned long max_freeable
= cachep
->max_freeable
;
3779 unsigned long node_allocs
= cachep
->node_allocs
;
3780 unsigned long node_frees
= cachep
->node_frees
;
3782 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu \
3783 %4lu %4lu %4lu %4lu", allocs
, high
, grown
,
3784 reaped
, errors
, max_freeable
, node_allocs
,
3789 unsigned long allochit
= atomic_read(&cachep
->allochit
);
3790 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
3791 unsigned long freehit
= atomic_read(&cachep
->freehit
);
3792 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
3794 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
3795 allochit
, allocmiss
, freehit
, freemiss
);
3803 * slabinfo_op - iterator that generates /proc/slabinfo
3812 * num-pages-per-slab
3813 * + further values on SMP and with statistics enabled
3816 struct seq_operations slabinfo_op
= {
3823 #define MAX_SLABINFO_WRITE 128
3825 * slabinfo_write - Tuning for the slab allocator
3827 * @buffer: user buffer
3828 * @count: data length
3831 ssize_t
slabinfo_write(struct file
*file
, const char __user
* buffer
,
3832 size_t count
, loff_t
*ppos
)
3834 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
3835 int limit
, batchcount
, shared
, res
;
3836 struct list_head
*p
;
3838 if (count
> MAX_SLABINFO_WRITE
)
3840 if (copy_from_user(&kbuf
, buffer
, count
))
3842 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
3844 tmp
= strchr(kbuf
, ' ');
3849 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
3852 /* Find the cache in the chain of caches. */
3853 mutex_lock(&cache_chain_mutex
);
3855 list_for_each(p
, &cache_chain
) {
3856 struct kmem_cache
*cachep
;
3858 cachep
= list_entry(p
, struct kmem_cache
, next
);
3859 if (!strcmp(cachep
->name
, kbuf
)) {
3860 if (limit
< 1 || batchcount
< 1 ||
3861 batchcount
> limit
|| shared
< 0) {
3864 res
= do_tune_cpucache(cachep
, limit
,
3865 batchcount
, shared
);
3870 mutex_unlock(&cache_chain_mutex
);
3878 * ksize - get the actual amount of memory allocated for a given object
3879 * @objp: Pointer to the object
3881 * kmalloc may internally round up allocations and return more memory
3882 * than requested. ksize() can be used to determine the actual amount of
3883 * memory allocated. The caller may use this additional memory, even though
3884 * a smaller amount of memory was initially specified with the kmalloc call.
3885 * The caller must guarantee that objp points to a valid object previously
3886 * allocated with either kmalloc() or kmem_cache_alloc(). The object
3887 * must not be freed during the duration of the call.
3889 unsigned int ksize(const void *objp
)
3891 if (unlikely(objp
== NULL
))
3894 return obj_size(virt_to_cache(objp
));