3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same intializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/config.h>
90 #include <linux/slab.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/cpuset.h>
98 #include <linux/seq_file.h>
99 #include <linux/notifier.h>
100 #include <linux/kallsyms.h>
101 #include <linux/cpu.h>
102 #include <linux/sysctl.h>
103 #include <linux/module.h>
104 #include <linux/rcupdate.h>
105 #include <linux/string.h>
106 #include <linux/nodemask.h>
107 #include <linux/mempolicy.h>
108 #include <linux/mutex.h>
110 #include <asm/uaccess.h>
111 #include <asm/cacheflush.h>
112 #include <asm/tlbflush.h>
113 #include <asm/page.h>
116 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
117 * SLAB_RED_ZONE & SLAB_POISON.
118 * 0 for faster, smaller code (especially in the critical paths).
120 * STATS - 1 to collect stats for /proc/slabinfo.
121 * 0 for faster, smaller code (especially in the critical paths).
123 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
126 #ifdef CONFIG_DEBUG_SLAB
129 #define FORCED_DEBUG 1
133 #define FORCED_DEBUG 0
136 /* Shouldn't this be in a header file somewhere? */
137 #define BYTES_PER_WORD sizeof(void *)
139 #ifndef cache_line_size
140 #define cache_line_size() L1_CACHE_BYTES
143 #ifndef ARCH_KMALLOC_MINALIGN
145 * Enforce a minimum alignment for the kmalloc caches.
146 * Usually, the kmalloc caches are cache_line_size() aligned, except when
147 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
148 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
149 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
150 * Note that this flag disables some debug features.
152 #define ARCH_KMALLOC_MINALIGN 0
155 #ifndef ARCH_SLAB_MINALIGN
157 * Enforce a minimum alignment for all caches.
158 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
159 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
160 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
161 * some debug features.
163 #define ARCH_SLAB_MINALIGN 0
166 #ifndef ARCH_KMALLOC_FLAGS
167 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
170 /* Legal flag mask for kmem_cache_create(). */
172 # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
173 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
175 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
176 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
177 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
179 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
180 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
181 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
182 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
188 * Bufctl's are used for linking objs within a slab
191 * This implementation relies on "struct page" for locating the cache &
192 * slab an object belongs to.
193 * This allows the bufctl structure to be small (one int), but limits
194 * the number of objects a slab (not a cache) can contain when off-slab
195 * bufctls are used. The limit is the size of the largest general cache
196 * that does not use off-slab slabs.
197 * For 32bit archs with 4 kB pages, is this 56.
198 * This is not serious, as it is only for large objects, when it is unwise
199 * to have too many per slab.
200 * Note: This limit can be raised by introducing a general cache whose size
201 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
204 typedef unsigned int kmem_bufctl_t
;
205 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
206 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
207 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
208 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
210 /* Max number of objs-per-slab for caches which use off-slab slabs.
211 * Needed to avoid a possible looping condition in cache_grow().
213 static unsigned long offslab_limit
;
218 * Manages the objs in a slab. Placed either at the beginning of mem allocated
219 * for a slab, or allocated from an general cache.
220 * Slabs are chained into three list: fully used, partial, fully free slabs.
223 struct list_head list
;
224 unsigned long colouroff
;
225 void *s_mem
; /* including colour offset */
226 unsigned int inuse
; /* num of objs active in slab */
228 unsigned short nodeid
;
234 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
235 * arrange for kmem_freepages to be called via RCU. This is useful if
236 * we need to approach a kernel structure obliquely, from its address
237 * obtained without the usual locking. We can lock the structure to
238 * stabilize it and check it's still at the given address, only if we
239 * can be sure that the memory has not been meanwhile reused for some
240 * other kind of object (which our subsystem's lock might corrupt).
242 * rcu_read_lock before reading the address, then rcu_read_unlock after
243 * taking the spinlock within the structure expected at that address.
245 * We assume struct slab_rcu can overlay struct slab when destroying.
248 struct rcu_head head
;
249 struct kmem_cache
*cachep
;
257 * - LIFO ordering, to hand out cache-warm objects from _alloc
258 * - reduce the number of linked list operations
259 * - reduce spinlock operations
261 * The limit is stored in the per-cpu structure to reduce the data cache
268 unsigned int batchcount
;
269 unsigned int touched
;
272 * Must have this definition in here for the proper
273 * alignment of array_cache. Also simplifies accessing
275 * [0] is for gcc 2.95. It should really be [].
280 * bootstrap: The caches do not work without cpuarrays anymore, but the
281 * cpuarrays are allocated from the generic caches...
283 #define BOOT_CPUCACHE_ENTRIES 1
284 struct arraycache_init
{
285 struct array_cache cache
;
286 void *entries
[BOOT_CPUCACHE_ENTRIES
];
290 * The slab lists for all objects.
293 struct list_head slabs_partial
; /* partial list first, better asm code */
294 struct list_head slabs_full
;
295 struct list_head slabs_free
;
296 unsigned long free_objects
;
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 */
302 unsigned long next_reap
; /* updated without locking */
303 int free_touched
; /* updated without locking */
307 * Need this for bootstrapping a per node allocator.
309 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
310 struct kmem_list3 __initdata initkmem_list3
[NUM_INIT_LISTS
];
311 #define CACHE_CACHE 0
313 #define SIZE_L3 (1 + MAX_NUMNODES)
316 * This function must be completely optimized away if a constant is passed to
317 * it. Mostly the same as what is in linux/slab.h except it returns an index.
319 static __always_inline
int index_of(const size_t size
)
321 extern void __bad_size(void);
323 if (__builtin_constant_p(size
)) {
331 #include "linux/kmalloc_sizes.h"
339 #define INDEX_AC index_of(sizeof(struct arraycache_init))
340 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
342 static void kmem_list3_init(struct kmem_list3
*parent
)
344 INIT_LIST_HEAD(&parent
->slabs_full
);
345 INIT_LIST_HEAD(&parent
->slabs_partial
);
346 INIT_LIST_HEAD(&parent
->slabs_free
);
347 parent
->shared
= NULL
;
348 parent
->alien
= NULL
;
349 parent
->colour_next
= 0;
350 spin_lock_init(&parent
->list_lock
);
351 parent
->free_objects
= 0;
352 parent
->free_touched
= 0;
355 #define MAKE_LIST(cachep, listp, slab, nodeid) \
357 INIT_LIST_HEAD(listp); \
358 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
361 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
363 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
364 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
365 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
375 /* 1) per-cpu data, touched during every alloc/free */
376 struct array_cache
*array
[NR_CPUS
];
377 /* 2) Cache tunables. Protected by cache_chain_mutex */
378 unsigned int batchcount
;
382 unsigned int buffer_size
;
383 /* 3) touched by every alloc & free from the backend */
384 struct kmem_list3
*nodelists
[MAX_NUMNODES
];
386 unsigned int flags
; /* constant flags */
387 unsigned int num
; /* # of objs per slab */
389 /* 4) cache_grow/shrink */
390 /* order of pgs per slab (2^n) */
391 unsigned int gfporder
;
393 /* force GFP flags, e.g. GFP_DMA */
396 size_t colour
; /* cache colouring range */
397 unsigned int colour_off
; /* colour offset */
398 struct kmem_cache
*slabp_cache
;
399 unsigned int slab_size
;
400 unsigned int dflags
; /* dynamic flags */
402 /* constructor func */
403 void (*ctor
) (void *, struct kmem_cache
*, unsigned long);
405 /* de-constructor func */
406 void (*dtor
) (void *, struct kmem_cache
*, unsigned long);
408 /* 5) cache creation/removal */
410 struct list_head next
;
414 unsigned long num_active
;
415 unsigned long num_allocations
;
416 unsigned long high_mark
;
418 unsigned long reaped
;
419 unsigned long errors
;
420 unsigned long max_freeable
;
421 unsigned long node_allocs
;
422 unsigned long node_frees
;
423 unsigned long node_overflow
;
431 * If debugging is enabled, then the allocator can add additional
432 * fields and/or padding to every object. buffer_size contains the total
433 * object size including these internal fields, the following two
434 * variables contain the offset to the user object and its size.
441 #define CFLGS_OFF_SLAB (0x80000000UL)
442 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
444 #define BATCHREFILL_LIMIT 16
446 * Optimization question: fewer reaps means less probability for unnessary
447 * cpucache drain/refill cycles.
449 * OTOH the cpuarrays can contain lots of objects,
450 * which could lock up otherwise freeable slabs.
452 #define REAPTIMEOUT_CPUC (2*HZ)
453 #define REAPTIMEOUT_LIST3 (4*HZ)
456 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
457 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
458 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
459 #define STATS_INC_GROWN(x) ((x)->grown++)
460 #define STATS_INC_REAPED(x) ((x)->reaped++)
461 #define STATS_SET_HIGH(x) \
463 if ((x)->num_active > (x)->high_mark) \
464 (x)->high_mark = (x)->num_active; \
466 #define STATS_INC_ERR(x) ((x)->errors++)
467 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
468 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
469 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
470 #define STATS_SET_FREEABLE(x, i) \
472 if ((x)->max_freeable < i) \
473 (x)->max_freeable = i; \
475 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
476 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
477 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
478 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
480 #define STATS_INC_ACTIVE(x) do { } while (0)
481 #define STATS_DEC_ACTIVE(x) do { } while (0)
482 #define STATS_INC_ALLOCED(x) do { } while (0)
483 #define STATS_INC_GROWN(x) do { } while (0)
484 #define STATS_INC_REAPED(x) do { } while (0)
485 #define STATS_SET_HIGH(x) do { } while (0)
486 #define STATS_INC_ERR(x) do { } while (0)
487 #define STATS_INC_NODEALLOCS(x) do { } while (0)
488 #define STATS_INC_NODEFREES(x) do { } while (0)
489 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
490 #define STATS_SET_FREEABLE(x, i) do { } while (0)
491 #define STATS_INC_ALLOCHIT(x) do { } while (0)
492 #define STATS_INC_ALLOCMISS(x) do { } while (0)
493 #define STATS_INC_FREEHIT(x) do { } while (0)
494 #define STATS_INC_FREEMISS(x) do { } while (0)
499 * Magic nums for obj red zoning.
500 * Placed in the first word before and the first word after an obj.
502 #define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
503 #define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
505 /* ...and for poisoning */
506 #define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
507 #define POISON_FREE 0x6b /* for use-after-free poisoning */
508 #define POISON_END 0xa5 /* end-byte of poisoning */
511 * memory layout of objects:
513 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
514 * the end of an object is aligned with the end of the real
515 * allocation. Catches writes behind the end of the allocation.
516 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
518 * cachep->obj_offset: The real object.
519 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
520 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
521 * [BYTES_PER_WORD long]
523 static int obj_offset(struct kmem_cache
*cachep
)
525 return cachep
->obj_offset
;
528 static int obj_size(struct kmem_cache
*cachep
)
530 return cachep
->obj_size
;
533 static unsigned long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
535 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
536 return (unsigned long*) (objp
+obj_offset(cachep
)-BYTES_PER_WORD
);
539 static unsigned long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
541 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
542 if (cachep
->flags
& SLAB_STORE_USER
)
543 return (unsigned long *)(objp
+ cachep
->buffer_size
-
545 return (unsigned long *)(objp
+ cachep
->buffer_size
- BYTES_PER_WORD
);
548 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
550 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
551 return (void **)(objp
+ cachep
->buffer_size
- BYTES_PER_WORD
);
556 #define obj_offset(x) 0
557 #define obj_size(cachep) (cachep->buffer_size)
558 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
559 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
560 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
565 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
568 #if defined(CONFIG_LARGE_ALLOCS)
569 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
570 #define MAX_GFP_ORDER 13 /* up to 32Mb */
571 #elif defined(CONFIG_MMU)
572 #define MAX_OBJ_ORDER 5 /* 32 pages */
573 #define MAX_GFP_ORDER 5 /* 32 pages */
575 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
576 #define MAX_GFP_ORDER 8 /* up to 1Mb */
580 * Do not go above this order unless 0 objects fit into the slab.
582 #define BREAK_GFP_ORDER_HI 1
583 #define BREAK_GFP_ORDER_LO 0
584 static int slab_break_gfp_order
= BREAK_GFP_ORDER_LO
;
587 * Functions for storing/retrieving the cachep and or slab from the page
588 * allocator. These are used to find the slab an obj belongs to. With kfree(),
589 * these are used to find the cache which an obj belongs to.
591 static inline void page_set_cache(struct page
*page
, struct kmem_cache
*cache
)
593 page
->lru
.next
= (struct list_head
*)cache
;
596 static inline struct kmem_cache
*page_get_cache(struct page
*page
)
598 if (unlikely(PageCompound(page
)))
599 page
= (struct page
*)page_private(page
);
600 return (struct kmem_cache
*)page
->lru
.next
;
603 static inline void page_set_slab(struct page
*page
, struct slab
*slab
)
605 page
->lru
.prev
= (struct list_head
*)slab
;
608 static inline struct slab
*page_get_slab(struct page
*page
)
610 if (unlikely(PageCompound(page
)))
611 page
= (struct page
*)page_private(page
);
612 return (struct slab
*)page
->lru
.prev
;
615 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
617 struct page
*page
= virt_to_page(obj
);
618 return page_get_cache(page
);
621 static inline struct slab
*virt_to_slab(const void *obj
)
623 struct page
*page
= virt_to_page(obj
);
624 return page_get_slab(page
);
627 static inline void *index_to_obj(struct kmem_cache
*cache
, struct slab
*slab
,
630 return slab
->s_mem
+ cache
->buffer_size
* idx
;
633 static inline unsigned int obj_to_index(struct kmem_cache
*cache
,
634 struct slab
*slab
, void *obj
)
636 return (unsigned)(obj
- slab
->s_mem
) / cache
->buffer_size
;
640 * These are the default caches for kmalloc. Custom caches can have other sizes.
642 struct cache_sizes malloc_sizes
[] = {
643 #define CACHE(x) { .cs_size = (x) },
644 #include <linux/kmalloc_sizes.h>
648 EXPORT_SYMBOL(malloc_sizes
);
650 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
656 static struct cache_names __initdata cache_names
[] = {
657 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
658 #include <linux/kmalloc_sizes.h>
663 static struct arraycache_init initarray_cache __initdata
=
664 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
665 static struct arraycache_init initarray_generic
=
666 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
668 /* internal cache of cache description objs */
669 static struct kmem_cache cache_cache
= {
671 .limit
= BOOT_CPUCACHE_ENTRIES
,
673 .buffer_size
= sizeof(struct kmem_cache
),
674 .name
= "kmem_cache",
676 .obj_size
= sizeof(struct kmem_cache
),
680 /* Guard access to the cache-chain. */
681 static DEFINE_MUTEX(cache_chain_mutex
);
682 static struct list_head cache_chain
;
685 * vm_enough_memory() looks at this to determine how many slab-allocated pages
686 * are possibly freeable under pressure
688 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
690 atomic_t slab_reclaim_pages
;
693 * chicken and egg problem: delay the per-cpu array allocation
694 * until the general caches are up.
703 static DEFINE_PER_CPU(struct work_struct
, reap_work
);
705 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
707 static void enable_cpucache(struct kmem_cache
*cachep
);
708 static void cache_reap(void *unused
);
709 static int __node_shrink(struct kmem_cache
*cachep
, int node
);
711 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
713 return cachep
->array
[smp_processor_id()];
716 static inline struct kmem_cache
*__find_general_cachep(size_t size
,
719 struct cache_sizes
*csizep
= malloc_sizes
;
722 /* This happens if someone tries to call
723 * kmem_cache_create(), or __kmalloc(), before
724 * the generic caches are initialized.
726 BUG_ON(malloc_sizes
[INDEX_AC
].cs_cachep
== NULL
);
728 while (size
> csizep
->cs_size
)
732 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
733 * has cs_{dma,}cachep==NULL. Thus no special case
734 * for large kmalloc calls required.
736 if (unlikely(gfpflags
& GFP_DMA
))
737 return csizep
->cs_dmacachep
;
738 return csizep
->cs_cachep
;
741 struct kmem_cache
*kmem_find_general_cachep(size_t size
, gfp_t gfpflags
)
743 return __find_general_cachep(size
, gfpflags
);
745 EXPORT_SYMBOL(kmem_find_general_cachep
);
747 static size_t slab_mgmt_size(size_t nr_objs
, size_t align
)
749 return ALIGN(sizeof(struct slab
)+nr_objs
*sizeof(kmem_bufctl_t
), align
);
753 * Calculate the number of objects and left-over bytes for a given buffer size.
755 static void cache_estimate(unsigned long gfporder
, size_t buffer_size
,
756 size_t align
, int flags
, size_t *left_over
,
761 size_t slab_size
= PAGE_SIZE
<< gfporder
;
764 * The slab management structure can be either off the slab or
765 * on it. For the latter case, the memory allocated for a
769 * - One kmem_bufctl_t for each object
770 * - Padding to respect alignment of @align
771 * - @buffer_size bytes for each object
773 * If the slab management structure is off the slab, then the
774 * alignment will already be calculated into the size. Because
775 * the slabs are all pages aligned, the objects will be at the
776 * correct alignment when allocated.
778 if (flags
& CFLGS_OFF_SLAB
) {
780 nr_objs
= slab_size
/ buffer_size
;
782 if (nr_objs
> SLAB_LIMIT
)
783 nr_objs
= SLAB_LIMIT
;
786 * Ignore padding for the initial guess. The padding
787 * is at most @align-1 bytes, and @buffer_size is at
788 * least @align. In the worst case, this result will
789 * be one greater than the number of objects that fit
790 * into the memory allocation when taking the padding
793 nr_objs
= (slab_size
- sizeof(struct slab
)) /
794 (buffer_size
+ sizeof(kmem_bufctl_t
));
797 * This calculated number will be either the right
798 * amount, or one greater than what we want.
800 if (slab_mgmt_size(nr_objs
, align
) + nr_objs
*buffer_size
804 if (nr_objs
> SLAB_LIMIT
)
805 nr_objs
= SLAB_LIMIT
;
807 mgmt_size
= slab_mgmt_size(nr_objs
, align
);
810 *left_over
= slab_size
- nr_objs
*buffer_size
- mgmt_size
;
813 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
815 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
818 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
819 function
, cachep
->name
, msg
);
825 * Special reaping functions for NUMA systems called from cache_reap().
826 * These take care of doing round robin flushing of alien caches (containing
827 * objects freed on different nodes from which they were allocated) and the
828 * flushing of remote pcps by calling drain_node_pages.
830 static DEFINE_PER_CPU(unsigned long, reap_node
);
832 static void init_reap_node(int cpu
)
836 node
= next_node(cpu_to_node(cpu
), node_online_map
);
837 if (node
== MAX_NUMNODES
)
838 node
= first_node(node_online_map
);
840 __get_cpu_var(reap_node
) = node
;
843 static void next_reap_node(void)
845 int node
= __get_cpu_var(reap_node
);
848 * Also drain per cpu pages on remote zones
850 if (node
!= numa_node_id())
851 drain_node_pages(node
);
853 node
= next_node(node
, node_online_map
);
854 if (unlikely(node
>= MAX_NUMNODES
))
855 node
= first_node(node_online_map
);
856 __get_cpu_var(reap_node
) = node
;
860 #define init_reap_node(cpu) do { } while (0)
861 #define next_reap_node(void) do { } while (0)
865 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
866 * via the workqueue/eventd.
867 * Add the CPU number into the expiration time to minimize the possibility of
868 * the CPUs getting into lockstep and contending for the global cache chain
871 static void __devinit
start_cpu_timer(int cpu
)
873 struct work_struct
*reap_work
= &per_cpu(reap_work
, cpu
);
876 * When this gets called from do_initcalls via cpucache_init(),
877 * init_workqueues() has already run, so keventd will be setup
880 if (keventd_up() && reap_work
->func
== NULL
) {
882 INIT_WORK(reap_work
, cache_reap
, NULL
);
883 schedule_delayed_work_on(cpu
, reap_work
, HZ
+ 3 * cpu
);
887 static struct array_cache
*alloc_arraycache(int node
, int entries
,
890 int memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
891 struct array_cache
*nc
= NULL
;
893 nc
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
897 nc
->batchcount
= batchcount
;
899 spin_lock_init(&nc
->lock
);
905 * Transfer objects in one arraycache to another.
906 * Locking must be handled by the caller.
908 * Return the number of entries transferred.
910 static int transfer_objects(struct array_cache
*to
,
911 struct array_cache
*from
, unsigned int max
)
913 /* Figure out how many entries to transfer */
914 int nr
= min(min(from
->avail
, max
), to
->limit
- to
->avail
);
919 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
929 static void *__cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
930 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
932 static struct array_cache
**alloc_alien_cache(int node
, int limit
)
934 struct array_cache
**ac_ptr
;
935 int memsize
= sizeof(void *) * MAX_NUMNODES
;
940 ac_ptr
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
943 if (i
== node
|| !node_online(i
)) {
947 ac_ptr
[i
] = alloc_arraycache(node
, limit
, 0xbaadf00d);
949 for (i
--; i
<= 0; i
--)
959 static void free_alien_cache(struct array_cache
**ac_ptr
)
970 static void __drain_alien_cache(struct kmem_cache
*cachep
,
971 struct array_cache
*ac
, int node
)
973 struct kmem_list3
*rl3
= cachep
->nodelists
[node
];
976 spin_lock(&rl3
->list_lock
);
978 * Stuff objects into the remote nodes shared array first.
979 * That way we could avoid the overhead of putting the objects
980 * into the free lists and getting them back later.
983 transfer_objects(rl3
->shared
, ac
, ac
->limit
);
985 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
987 spin_unlock(&rl3
->list_lock
);
992 * Called from cache_reap() to regularly drain alien caches round robin.
994 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_list3
*l3
)
996 int node
= __get_cpu_var(reap_node
);
999 struct array_cache
*ac
= l3
->alien
[node
];
1001 if (ac
&& ac
->avail
&& spin_trylock_irq(&ac
->lock
)) {
1002 __drain_alien_cache(cachep
, ac
, node
);
1003 spin_unlock_irq(&ac
->lock
);
1008 static void drain_alien_cache(struct kmem_cache
*cachep
,
1009 struct array_cache
**alien
)
1012 struct array_cache
*ac
;
1013 unsigned long flags
;
1015 for_each_online_node(i
) {
1018 spin_lock_irqsave(&ac
->lock
, flags
);
1019 __drain_alien_cache(cachep
, ac
, i
);
1020 spin_unlock_irqrestore(&ac
->lock
, flags
);
1026 #define drain_alien_cache(cachep, alien) do { } while (0)
1027 #define reap_alien(cachep, l3) do { } while (0)
1029 static inline struct array_cache
**alloc_alien_cache(int node
, int limit
)
1031 return (struct array_cache
**) 0x01020304ul
;
1034 static inline void free_alien_cache(struct array_cache
**ac_ptr
)
1040 static int cpuup_callback(struct notifier_block
*nfb
,
1041 unsigned long action
, void *hcpu
)
1043 long cpu
= (long)hcpu
;
1044 struct kmem_cache
*cachep
;
1045 struct kmem_list3
*l3
= NULL
;
1046 int node
= cpu_to_node(cpu
);
1047 int memsize
= sizeof(struct kmem_list3
);
1050 case CPU_UP_PREPARE
:
1051 mutex_lock(&cache_chain_mutex
);
1053 * We need to do this right in the beginning since
1054 * alloc_arraycache's are going to use this list.
1055 * kmalloc_node allows us to add the slab to the right
1056 * kmem_list3 and not this cpu's kmem_list3
1059 list_for_each_entry(cachep
, &cache_chain
, next
) {
1061 * Set up the size64 kmemlist for cpu before we can
1062 * begin anything. Make sure some other cpu on this
1063 * node has not already allocated this
1065 if (!cachep
->nodelists
[node
]) {
1066 l3
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1069 kmem_list3_init(l3
);
1070 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
1071 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1074 * The l3s don't come and go as CPUs come and
1075 * go. cache_chain_mutex is sufficient
1078 cachep
->nodelists
[node
] = l3
;
1081 spin_lock_irq(&cachep
->nodelists
[node
]->list_lock
);
1082 cachep
->nodelists
[node
]->free_limit
=
1083 (1 + nr_cpus_node(node
)) *
1084 cachep
->batchcount
+ cachep
->num
;
1085 spin_unlock_irq(&cachep
->nodelists
[node
]->list_lock
);
1089 * Now we can go ahead with allocating the shared arrays and
1092 list_for_each_entry(cachep
, &cache_chain
, next
) {
1093 struct array_cache
*nc
;
1094 struct array_cache
*shared
;
1095 struct array_cache
**alien
;
1097 nc
= alloc_arraycache(node
, cachep
->limit
,
1098 cachep
->batchcount
);
1101 shared
= alloc_arraycache(node
,
1102 cachep
->shared
* cachep
->batchcount
,
1107 alien
= alloc_alien_cache(node
, cachep
->limit
);
1110 cachep
->array
[cpu
] = nc
;
1111 l3
= cachep
->nodelists
[node
];
1114 spin_lock_irq(&l3
->list_lock
);
1117 * We are serialised from CPU_DEAD or
1118 * CPU_UP_CANCELLED by the cpucontrol lock
1120 l3
->shared
= shared
;
1129 spin_unlock_irq(&l3
->list_lock
);
1131 free_alien_cache(alien
);
1133 mutex_unlock(&cache_chain_mutex
);
1136 start_cpu_timer(cpu
);
1138 #ifdef CONFIG_HOTPLUG_CPU
1141 * Even if all the cpus of a node are down, we don't free the
1142 * kmem_list3 of any cache. This to avoid a race between
1143 * cpu_down, and a kmalloc allocation from another cpu for
1144 * memory from the node of the cpu going down. The list3
1145 * structure is usually allocated from kmem_cache_create() and
1146 * gets destroyed at kmem_cache_destroy().
1149 case CPU_UP_CANCELED
:
1150 mutex_lock(&cache_chain_mutex
);
1151 list_for_each_entry(cachep
, &cache_chain
, next
) {
1152 struct array_cache
*nc
;
1153 struct array_cache
*shared
;
1154 struct array_cache
**alien
;
1157 mask
= node_to_cpumask(node
);
1158 /* cpu is dead; no one can alloc from it. */
1159 nc
= cachep
->array
[cpu
];
1160 cachep
->array
[cpu
] = NULL
;
1161 l3
= cachep
->nodelists
[node
];
1164 goto free_array_cache
;
1166 spin_lock_irq(&l3
->list_lock
);
1168 /* Free limit for this kmem_list3 */
1169 l3
->free_limit
-= cachep
->batchcount
;
1171 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
1173 if (!cpus_empty(mask
)) {
1174 spin_unlock_irq(&l3
->list_lock
);
1175 goto free_array_cache
;
1178 shared
= l3
->shared
;
1180 free_block(cachep
, l3
->shared
->entry
,
1181 l3
->shared
->avail
, node
);
1188 spin_unlock_irq(&l3
->list_lock
);
1192 drain_alien_cache(cachep
, alien
);
1193 free_alien_cache(alien
);
1199 * In the previous loop, all the objects were freed to
1200 * the respective cache's slabs, now we can go ahead and
1201 * shrink each nodelist to its limit.
1203 list_for_each_entry(cachep
, &cache_chain
, next
) {
1204 l3
= cachep
->nodelists
[node
];
1207 spin_lock_irq(&l3
->list_lock
);
1208 /* free slabs belonging to this node */
1209 __node_shrink(cachep
, node
);
1210 spin_unlock_irq(&l3
->list_lock
);
1212 mutex_unlock(&cache_chain_mutex
);
1218 mutex_unlock(&cache_chain_mutex
);
1222 static struct notifier_block cpucache_notifier
= { &cpuup_callback
, NULL
, 0 };
1225 * swap the static kmem_list3 with kmalloced memory
1227 static void init_list(struct kmem_cache
*cachep
, struct kmem_list3
*list
,
1230 struct kmem_list3
*ptr
;
1232 BUG_ON(cachep
->nodelists
[nodeid
] != list
);
1233 ptr
= kmalloc_node(sizeof(struct kmem_list3
), GFP_KERNEL
, nodeid
);
1236 local_irq_disable();
1237 memcpy(ptr
, list
, sizeof(struct kmem_list3
));
1238 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1239 cachep
->nodelists
[nodeid
] = ptr
;
1244 * Initialisation. Called after the page allocator have been initialised and
1245 * before smp_init().
1247 void __init
kmem_cache_init(void)
1250 struct cache_sizes
*sizes
;
1251 struct cache_names
*names
;
1255 for (i
= 0; i
< NUM_INIT_LISTS
; i
++) {
1256 kmem_list3_init(&initkmem_list3
[i
]);
1257 if (i
< MAX_NUMNODES
)
1258 cache_cache
.nodelists
[i
] = NULL
;
1262 * Fragmentation resistance on low memory - only use bigger
1263 * page orders on machines with more than 32MB of memory.
1265 if (num_physpages
> (32 << 20) >> PAGE_SHIFT
)
1266 slab_break_gfp_order
= BREAK_GFP_ORDER_HI
;
1268 /* Bootstrap is tricky, because several objects are allocated
1269 * from caches that do not exist yet:
1270 * 1) initialize the cache_cache cache: it contains the struct
1271 * kmem_cache structures of all caches, except cache_cache itself:
1272 * cache_cache is statically allocated.
1273 * Initially an __init data area is used for the head array and the
1274 * kmem_list3 structures, it's replaced with a kmalloc allocated
1275 * array at the end of the bootstrap.
1276 * 2) Create the first kmalloc cache.
1277 * The struct kmem_cache for the new cache is allocated normally.
1278 * An __init data area is used for the head array.
1279 * 3) Create the remaining kmalloc caches, with minimally sized
1281 * 4) Replace the __init data head arrays for cache_cache and the first
1282 * kmalloc cache with kmalloc allocated arrays.
1283 * 5) Replace the __init data for kmem_list3 for cache_cache and
1284 * the other cache's with kmalloc allocated memory.
1285 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1288 /* 1) create the cache_cache */
1289 INIT_LIST_HEAD(&cache_chain
);
1290 list_add(&cache_cache
.next
, &cache_chain
);
1291 cache_cache
.colour_off
= cache_line_size();
1292 cache_cache
.array
[smp_processor_id()] = &initarray_cache
.cache
;
1293 cache_cache
.nodelists
[numa_node_id()] = &initkmem_list3
[CACHE_CACHE
];
1295 cache_cache
.buffer_size
= ALIGN(cache_cache
.buffer_size
,
1298 for (order
= 0; order
< MAX_ORDER
; order
++) {
1299 cache_estimate(order
, cache_cache
.buffer_size
,
1300 cache_line_size(), 0, &left_over
, &cache_cache
.num
);
1301 if (cache_cache
.num
)
1304 BUG_ON(!cache_cache
.num
);
1305 cache_cache
.gfporder
= order
;
1306 cache_cache
.colour
= left_over
/ cache_cache
.colour_off
;
1307 cache_cache
.slab_size
= ALIGN(cache_cache
.num
* sizeof(kmem_bufctl_t
) +
1308 sizeof(struct slab
), cache_line_size());
1310 /* 2+3) create the kmalloc caches */
1311 sizes
= malloc_sizes
;
1312 names
= cache_names
;
1315 * Initialize the caches that provide memory for the array cache and the
1316 * kmem_list3 structures first. Without this, further allocations will
1320 sizes
[INDEX_AC
].cs_cachep
= kmem_cache_create(names
[INDEX_AC
].name
,
1321 sizes
[INDEX_AC
].cs_size
,
1322 ARCH_KMALLOC_MINALIGN
,
1323 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1326 if (INDEX_AC
!= INDEX_L3
) {
1327 sizes
[INDEX_L3
].cs_cachep
=
1328 kmem_cache_create(names
[INDEX_L3
].name
,
1329 sizes
[INDEX_L3
].cs_size
,
1330 ARCH_KMALLOC_MINALIGN
,
1331 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1335 while (sizes
->cs_size
!= ULONG_MAX
) {
1337 * For performance, all the general caches are L1 aligned.
1338 * This should be particularly beneficial on SMP boxes, as it
1339 * eliminates "false sharing".
1340 * Note for systems short on memory removing the alignment will
1341 * allow tighter packing of the smaller caches.
1343 if (!sizes
->cs_cachep
) {
1344 sizes
->cs_cachep
= kmem_cache_create(names
->name
,
1346 ARCH_KMALLOC_MINALIGN
,
1347 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1351 /* Inc off-slab bufctl limit until the ceiling is hit. */
1352 if (!(OFF_SLAB(sizes
->cs_cachep
))) {
1353 offslab_limit
= sizes
->cs_size
- sizeof(struct slab
);
1354 offslab_limit
/= sizeof(kmem_bufctl_t
);
1357 sizes
->cs_dmacachep
= kmem_cache_create(names
->name_dma
,
1359 ARCH_KMALLOC_MINALIGN
,
1360 ARCH_KMALLOC_FLAGS
|SLAB_CACHE_DMA
|
1366 /* 4) Replace the bootstrap head arrays */
1370 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1372 local_irq_disable();
1373 BUG_ON(cpu_cache_get(&cache_cache
) != &initarray_cache
.cache
);
1374 memcpy(ptr
, cpu_cache_get(&cache_cache
),
1375 sizeof(struct arraycache_init
));
1376 cache_cache
.array
[smp_processor_id()] = ptr
;
1379 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1381 local_irq_disable();
1382 BUG_ON(cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
)
1383 != &initarray_generic
.cache
);
1384 memcpy(ptr
, cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
),
1385 sizeof(struct arraycache_init
));
1386 malloc_sizes
[INDEX_AC
].cs_cachep
->array
[smp_processor_id()] =
1390 /* 5) Replace the bootstrap kmem_list3's */
1393 /* Replace the static kmem_list3 structures for the boot cpu */
1394 init_list(&cache_cache
, &initkmem_list3
[CACHE_CACHE
],
1397 for_each_online_node(node
) {
1398 init_list(malloc_sizes
[INDEX_AC
].cs_cachep
,
1399 &initkmem_list3
[SIZE_AC
+ node
], node
);
1401 if (INDEX_AC
!= INDEX_L3
) {
1402 init_list(malloc_sizes
[INDEX_L3
].cs_cachep
,
1403 &initkmem_list3
[SIZE_L3
+ node
],
1409 /* 6) resize the head arrays to their final sizes */
1411 struct kmem_cache
*cachep
;
1412 mutex_lock(&cache_chain_mutex
);
1413 list_for_each_entry(cachep
, &cache_chain
, next
)
1414 enable_cpucache(cachep
);
1415 mutex_unlock(&cache_chain_mutex
);
1419 g_cpucache_up
= FULL
;
1422 * Register a cpu startup notifier callback that initializes
1423 * cpu_cache_get for all new cpus
1425 register_cpu_notifier(&cpucache_notifier
);
1428 * The reap timers are started later, with a module init call: That part
1429 * of the kernel is not yet operational.
1433 static int __init
cpucache_init(void)
1438 * Register the timers that return unneeded pages to the page allocator
1440 for_each_online_cpu(cpu
)
1441 start_cpu_timer(cpu
);
1444 __initcall(cpucache_init
);
1447 * Interface to system's page allocator. No need to hold the cache-lock.
1449 * If we requested dmaable memory, we will get it. Even if we
1450 * did not request dmaable memory, we might get it, but that
1451 * would be relatively rare and ignorable.
1453 static void *kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
1459 flags
|= cachep
->gfpflags
;
1461 /* nommu uses slab's for process anonymous memory allocations, so
1462 * requires __GFP_COMP to properly refcount higher order allocations"
1464 page
= alloc_pages_node(nodeid
, (flags
| __GFP_COMP
), cachep
->gfporder
);
1466 page
= alloc_pages_node(nodeid
, flags
, cachep
->gfporder
);
1470 addr
= page_address(page
);
1472 i
= (1 << cachep
->gfporder
);
1473 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1474 atomic_add(i
, &slab_reclaim_pages
);
1475 add_page_state(nr_slab
, i
);
1477 __SetPageSlab(page
);
1484 * Interface to system's page release.
1486 static void kmem_freepages(struct kmem_cache
*cachep
, void *addr
)
1488 unsigned long i
= (1 << cachep
->gfporder
);
1489 struct page
*page
= virt_to_page(addr
);
1490 const unsigned long nr_freed
= i
;
1493 BUG_ON(!PageSlab(page
));
1494 __ClearPageSlab(page
);
1497 sub_page_state(nr_slab
, nr_freed
);
1498 if (current
->reclaim_state
)
1499 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1500 free_pages((unsigned long)addr
, cachep
->gfporder
);
1501 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1502 atomic_sub(1 << cachep
->gfporder
, &slab_reclaim_pages
);
1505 static void kmem_rcu_free(struct rcu_head
*head
)
1507 struct slab_rcu
*slab_rcu
= (struct slab_rcu
*)head
;
1508 struct kmem_cache
*cachep
= slab_rcu
->cachep
;
1510 kmem_freepages(cachep
, slab_rcu
->addr
);
1511 if (OFF_SLAB(cachep
))
1512 kmem_cache_free(cachep
->slabp_cache
, slab_rcu
);
1517 #ifdef CONFIG_DEBUG_PAGEALLOC
1518 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1519 unsigned long caller
)
1521 int size
= obj_size(cachep
);
1523 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1525 if (size
< 5 * sizeof(unsigned long))
1528 *addr
++ = 0x12345678;
1530 *addr
++ = smp_processor_id();
1531 size
-= 3 * sizeof(unsigned long);
1533 unsigned long *sptr
= &caller
;
1534 unsigned long svalue
;
1536 while (!kstack_end(sptr
)) {
1538 if (kernel_text_address(svalue
)) {
1540 size
-= sizeof(unsigned long);
1541 if (size
<= sizeof(unsigned long))
1547 *addr
++ = 0x87654321;
1551 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1553 int size
= obj_size(cachep
);
1554 addr
= &((char *)addr
)[obj_offset(cachep
)];
1556 memset(addr
, val
, size
);
1557 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1560 static void dump_line(char *data
, int offset
, int limit
)
1563 printk(KERN_ERR
"%03x:", offset
);
1564 for (i
= 0; i
< limit
; i
++)
1565 printk(" %02x", (unsigned char)data
[offset
+ i
]);
1572 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1577 if (cachep
->flags
& SLAB_RED_ZONE
) {
1578 printk(KERN_ERR
"Redzone: 0x%lx/0x%lx.\n",
1579 *dbg_redzone1(cachep
, objp
),
1580 *dbg_redzone2(cachep
, objp
));
1583 if (cachep
->flags
& SLAB_STORE_USER
) {
1584 printk(KERN_ERR
"Last user: [<%p>]",
1585 *dbg_userword(cachep
, objp
));
1586 print_symbol("(%s)",
1587 (unsigned long)*dbg_userword(cachep
, objp
));
1590 realobj
= (char *)objp
+ obj_offset(cachep
);
1591 size
= obj_size(cachep
);
1592 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1595 if (i
+ limit
> size
)
1597 dump_line(realobj
, i
, limit
);
1601 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1607 realobj
= (char *)objp
+ obj_offset(cachep
);
1608 size
= obj_size(cachep
);
1610 for (i
= 0; i
< size
; i
++) {
1611 char exp
= POISON_FREE
;
1614 if (realobj
[i
] != exp
) {
1620 "Slab corruption: start=%p, len=%d\n",
1622 print_objinfo(cachep
, objp
, 0);
1624 /* Hexdump the affected line */
1627 if (i
+ limit
> size
)
1629 dump_line(realobj
, i
, limit
);
1632 /* Limit to 5 lines */
1638 /* Print some data about the neighboring objects, if they
1641 struct slab
*slabp
= virt_to_slab(objp
);
1644 objnr
= obj_to_index(cachep
, slabp
, objp
);
1646 objp
= index_to_obj(cachep
, slabp
, objnr
- 1);
1647 realobj
= (char *)objp
+ obj_offset(cachep
);
1648 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
1650 print_objinfo(cachep
, objp
, 2);
1652 if (objnr
+ 1 < cachep
->num
) {
1653 objp
= index_to_obj(cachep
, slabp
, objnr
+ 1);
1654 realobj
= (char *)objp
+ obj_offset(cachep
);
1655 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
1657 print_objinfo(cachep
, objp
, 2);
1665 * slab_destroy_objs - destroy a slab and its objects
1666 * @cachep: cache pointer being destroyed
1667 * @slabp: slab pointer being destroyed
1669 * Call the registered destructor for each object in a slab that is being
1672 static void slab_destroy_objs(struct kmem_cache
*cachep
, struct slab
*slabp
)
1675 for (i
= 0; i
< cachep
->num
; i
++) {
1676 void *objp
= index_to_obj(cachep
, slabp
, i
);
1678 if (cachep
->flags
& SLAB_POISON
) {
1679 #ifdef CONFIG_DEBUG_PAGEALLOC
1680 if (cachep
->buffer_size
% PAGE_SIZE
== 0 &&
1682 kernel_map_pages(virt_to_page(objp
),
1683 cachep
->buffer_size
/ PAGE_SIZE
, 1);
1685 check_poison_obj(cachep
, objp
);
1687 check_poison_obj(cachep
, objp
);
1690 if (cachep
->flags
& SLAB_RED_ZONE
) {
1691 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1692 slab_error(cachep
, "start of a freed object "
1694 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1695 slab_error(cachep
, "end of a freed object "
1698 if (cachep
->dtor
&& !(cachep
->flags
& SLAB_POISON
))
1699 (cachep
->dtor
) (objp
+ obj_offset(cachep
), cachep
, 0);
1703 static void slab_destroy_objs(struct kmem_cache
*cachep
, struct slab
*slabp
)
1707 for (i
= 0; i
< cachep
->num
; i
++) {
1708 void *objp
= index_to_obj(cachep
, slabp
, i
);
1709 (cachep
->dtor
) (objp
, cachep
, 0);
1716 * slab_destroy - destroy and release all objects in a slab
1717 * @cachep: cache pointer being destroyed
1718 * @slabp: slab pointer being destroyed
1720 * Destroy all the objs in a slab, and release the mem back to the system.
1721 * Before calling the slab must have been unlinked from the cache. The
1722 * cache-lock is not held/needed.
1724 static void slab_destroy(struct kmem_cache
*cachep
, struct slab
*slabp
)
1726 void *addr
= slabp
->s_mem
- slabp
->colouroff
;
1728 slab_destroy_objs(cachep
, slabp
);
1729 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
1730 struct slab_rcu
*slab_rcu
;
1732 slab_rcu
= (struct slab_rcu
*)slabp
;
1733 slab_rcu
->cachep
= cachep
;
1734 slab_rcu
->addr
= addr
;
1735 call_rcu(&slab_rcu
->head
, kmem_rcu_free
);
1737 kmem_freepages(cachep
, addr
);
1738 if (OFF_SLAB(cachep
))
1739 kmem_cache_free(cachep
->slabp_cache
, slabp
);
1744 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1745 * size of kmem_list3.
1747 static void set_up_list3s(struct kmem_cache
*cachep
, int index
)
1751 for_each_online_node(node
) {
1752 cachep
->nodelists
[node
] = &initkmem_list3
[index
+ node
];
1753 cachep
->nodelists
[node
]->next_reap
= jiffies
+
1755 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1760 * calculate_slab_order - calculate size (page order) of slabs
1761 * @cachep: pointer to the cache that is being created
1762 * @size: size of objects to be created in this cache.
1763 * @align: required alignment for the objects.
1764 * @flags: slab allocation flags
1766 * Also calculates the number of objects per slab.
1768 * This could be made much more intelligent. For now, try to avoid using
1769 * high order pages for slabs. When the gfp() functions are more friendly
1770 * towards high-order requests, this should be changed.
1772 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
1773 size_t size
, size_t align
, unsigned long flags
)
1775 size_t left_over
= 0;
1778 for (gfporder
= 0; gfporder
<= MAX_GFP_ORDER
; gfporder
++) {
1782 cache_estimate(gfporder
, size
, align
, flags
, &remainder
, &num
);
1786 /* More than offslab_limit objects will cause problems */
1787 if ((flags
& CFLGS_OFF_SLAB
) && num
> offslab_limit
)
1790 /* Found something acceptable - save it away */
1792 cachep
->gfporder
= gfporder
;
1793 left_over
= remainder
;
1796 * A VFS-reclaimable slab tends to have most allocations
1797 * as GFP_NOFS and we really don't want to have to be allocating
1798 * higher-order pages when we are unable to shrink dcache.
1800 if (flags
& SLAB_RECLAIM_ACCOUNT
)
1804 * Large number of objects is good, but very large slabs are
1805 * currently bad for the gfp()s.
1807 if (gfporder
>= slab_break_gfp_order
)
1811 * Acceptable internal fragmentation?
1813 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
1819 static void setup_cpu_cache(struct kmem_cache
*cachep
)
1821 if (g_cpucache_up
== FULL
) {
1822 enable_cpucache(cachep
);
1825 if (g_cpucache_up
== NONE
) {
1827 * Note: the first kmem_cache_create must create the cache
1828 * that's used by kmalloc(24), otherwise the creation of
1829 * further caches will BUG().
1831 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
1834 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
1835 * the first cache, then we need to set up all its list3s,
1836 * otherwise the creation of further caches will BUG().
1838 set_up_list3s(cachep
, SIZE_AC
);
1839 if (INDEX_AC
== INDEX_L3
)
1840 g_cpucache_up
= PARTIAL_L3
;
1842 g_cpucache_up
= PARTIAL_AC
;
1844 cachep
->array
[smp_processor_id()] =
1845 kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1847 if (g_cpucache_up
== PARTIAL_AC
) {
1848 set_up_list3s(cachep
, SIZE_L3
);
1849 g_cpucache_up
= PARTIAL_L3
;
1852 for_each_online_node(node
) {
1853 cachep
->nodelists
[node
] =
1854 kmalloc_node(sizeof(struct kmem_list3
),
1856 BUG_ON(!cachep
->nodelists
[node
]);
1857 kmem_list3_init(cachep
->nodelists
[node
]);
1861 cachep
->nodelists
[numa_node_id()]->next_reap
=
1862 jiffies
+ REAPTIMEOUT_LIST3
+
1863 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1865 cpu_cache_get(cachep
)->avail
= 0;
1866 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
1867 cpu_cache_get(cachep
)->batchcount
= 1;
1868 cpu_cache_get(cachep
)->touched
= 0;
1869 cachep
->batchcount
= 1;
1870 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
1874 * kmem_cache_create - Create a cache.
1875 * @name: A string which is used in /proc/slabinfo to identify this cache.
1876 * @size: The size of objects to be created in this cache.
1877 * @align: The required alignment for the objects.
1878 * @flags: SLAB flags
1879 * @ctor: A constructor for the objects.
1880 * @dtor: A destructor for the objects.
1882 * Returns a ptr to the cache on success, NULL on failure.
1883 * Cannot be called within a int, but can be interrupted.
1884 * The @ctor is run when new pages are allocated by the cache
1885 * and the @dtor is run before the pages are handed back.
1887 * @name must be valid until the cache is destroyed. This implies that
1888 * the module calling this has to destroy the cache before getting unloaded.
1892 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1893 * to catch references to uninitialised memory.
1895 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1896 * for buffer overruns.
1898 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1899 * cacheline. This can be beneficial if you're counting cycles as closely
1903 kmem_cache_create (const char *name
, size_t size
, size_t align
,
1904 unsigned long flags
,
1905 void (*ctor
)(void*, struct kmem_cache
*, unsigned long),
1906 void (*dtor
)(void*, struct kmem_cache
*, unsigned long))
1908 size_t left_over
, slab_size
, ralign
;
1909 struct kmem_cache
*cachep
= NULL
;
1910 struct list_head
*p
;
1913 * Sanity checks... these are all serious usage bugs.
1915 if (!name
|| in_interrupt() || (size
< BYTES_PER_WORD
) ||
1916 (size
> (1 << MAX_OBJ_ORDER
) * PAGE_SIZE
) || (dtor
&& !ctor
)) {
1917 printk(KERN_ERR
"%s: Early error in slab %s\n", __FUNCTION__
,
1923 * Prevent CPUs from coming and going.
1924 * lock_cpu_hotplug() nests outside cache_chain_mutex
1928 mutex_lock(&cache_chain_mutex
);
1930 list_for_each(p
, &cache_chain
) {
1931 struct kmem_cache
*pc
= list_entry(p
, struct kmem_cache
, next
);
1932 mm_segment_t old_fs
= get_fs();
1937 * This happens when the module gets unloaded and doesn't
1938 * destroy its slab cache and no-one else reuses the vmalloc
1939 * area of the module. Print a warning.
1942 res
= __get_user(tmp
, pc
->name
);
1945 printk("SLAB: cache with size %d has lost its name\n",
1950 if (!strcmp(pc
->name
, name
)) {
1951 printk("kmem_cache_create: duplicate cache %s\n", name
);
1958 WARN_ON(strchr(name
, ' ')); /* It confuses parsers */
1959 if ((flags
& SLAB_DEBUG_INITIAL
) && !ctor
) {
1960 /* No constructor, but inital state check requested */
1961 printk(KERN_ERR
"%s: No con, but init state check "
1962 "requested - %s\n", __FUNCTION__
, name
);
1963 flags
&= ~SLAB_DEBUG_INITIAL
;
1967 * Enable redzoning and last user accounting, except for caches with
1968 * large objects, if the increased size would increase the object size
1969 * above the next power of two: caches with object sizes just above a
1970 * power of two have a significant amount of internal fragmentation.
1972 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + 3 * BYTES_PER_WORD
))
1973 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
1974 if (!(flags
& SLAB_DESTROY_BY_RCU
))
1975 flags
|= SLAB_POISON
;
1977 if (flags
& SLAB_DESTROY_BY_RCU
)
1978 BUG_ON(flags
& SLAB_POISON
);
1980 if (flags
& SLAB_DESTROY_BY_RCU
)
1984 * Always checks flags, a caller might be expecting debug support which
1987 BUG_ON(flags
& ~CREATE_MASK
);
1990 * Check that size is in terms of words. This is needed to avoid
1991 * unaligned accesses for some archs when redzoning is used, and makes
1992 * sure any on-slab bufctl's are also correctly aligned.
1994 if (size
& (BYTES_PER_WORD
- 1)) {
1995 size
+= (BYTES_PER_WORD
- 1);
1996 size
&= ~(BYTES_PER_WORD
- 1);
1999 /* calculate the final buffer alignment: */
2001 /* 1) arch recommendation: can be overridden for debug */
2002 if (flags
& SLAB_HWCACHE_ALIGN
) {
2004 * Default alignment: as specified by the arch code. Except if
2005 * an object is really small, then squeeze multiple objects into
2008 ralign
= cache_line_size();
2009 while (size
<= ralign
/ 2)
2012 ralign
= BYTES_PER_WORD
;
2014 /* 2) arch mandated alignment: disables debug if necessary */
2015 if (ralign
< ARCH_SLAB_MINALIGN
) {
2016 ralign
= ARCH_SLAB_MINALIGN
;
2017 if (ralign
> BYTES_PER_WORD
)
2018 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2020 /* 3) caller mandated alignment: disables debug if necessary */
2021 if (ralign
< align
) {
2023 if (ralign
> BYTES_PER_WORD
)
2024 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2027 * 4) Store it. Note that the debug code below can reduce
2028 * the alignment to BYTES_PER_WORD.
2032 /* Get cache's description obj. */
2033 cachep
= kmem_cache_zalloc(&cache_cache
, SLAB_KERNEL
);
2038 cachep
->obj_size
= size
;
2040 if (flags
& SLAB_RED_ZONE
) {
2041 /* redzoning only works with word aligned caches */
2042 align
= BYTES_PER_WORD
;
2044 /* add space for red zone words */
2045 cachep
->obj_offset
+= BYTES_PER_WORD
;
2046 size
+= 2 * BYTES_PER_WORD
;
2048 if (flags
& SLAB_STORE_USER
) {
2049 /* user store requires word alignment and
2050 * one word storage behind the end of the real
2053 align
= BYTES_PER_WORD
;
2054 size
+= BYTES_PER_WORD
;
2056 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2057 if (size
>= malloc_sizes
[INDEX_L3
+ 1].cs_size
2058 && cachep
->obj_size
> cache_line_size() && size
< PAGE_SIZE
) {
2059 cachep
->obj_offset
+= PAGE_SIZE
- size
;
2065 /* Determine if the slab management is 'on' or 'off' slab. */
2066 if (size
>= (PAGE_SIZE
>> 3))
2068 * Size is large, assume best to place the slab management obj
2069 * off-slab (should allow better packing of objs).
2071 flags
|= CFLGS_OFF_SLAB
;
2073 size
= ALIGN(size
, align
);
2075 left_over
= calculate_slab_order(cachep
, size
, align
, flags
);
2078 printk("kmem_cache_create: couldn't create cache %s.\n", name
);
2079 kmem_cache_free(&cache_cache
, cachep
);
2083 slab_size
= ALIGN(cachep
->num
* sizeof(kmem_bufctl_t
)
2084 + sizeof(struct slab
), align
);
2087 * If the slab has been placed off-slab, and we have enough space then
2088 * move it on-slab. This is at the expense of any extra colouring.
2090 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= slab_size
) {
2091 flags
&= ~CFLGS_OFF_SLAB
;
2092 left_over
-= slab_size
;
2095 if (flags
& CFLGS_OFF_SLAB
) {
2096 /* really off slab. No need for manual alignment */
2098 cachep
->num
* sizeof(kmem_bufctl_t
) + sizeof(struct slab
);
2101 cachep
->colour_off
= cache_line_size();
2102 /* Offset must be a multiple of the alignment. */
2103 if (cachep
->colour_off
< align
)
2104 cachep
->colour_off
= align
;
2105 cachep
->colour
= left_over
/ cachep
->colour_off
;
2106 cachep
->slab_size
= slab_size
;
2107 cachep
->flags
= flags
;
2108 cachep
->gfpflags
= 0;
2109 if (flags
& SLAB_CACHE_DMA
)
2110 cachep
->gfpflags
|= GFP_DMA
;
2111 cachep
->buffer_size
= size
;
2113 if (flags
& CFLGS_OFF_SLAB
)
2114 cachep
->slabp_cache
= kmem_find_general_cachep(slab_size
, 0u);
2115 cachep
->ctor
= ctor
;
2116 cachep
->dtor
= dtor
;
2117 cachep
->name
= name
;
2120 setup_cpu_cache(cachep
);
2122 /* cache setup completed, link it into the list */
2123 list_add(&cachep
->next
, &cache_chain
);
2125 if (!cachep
&& (flags
& SLAB_PANIC
))
2126 panic("kmem_cache_create(): failed to create slab `%s'\n",
2128 mutex_unlock(&cache_chain_mutex
);
2129 unlock_cpu_hotplug();
2132 EXPORT_SYMBOL(kmem_cache_create
);
2135 static void check_irq_off(void)
2137 BUG_ON(!irqs_disabled());
2140 static void check_irq_on(void)
2142 BUG_ON(irqs_disabled());
2145 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2149 assert_spin_locked(&cachep
->nodelists
[numa_node_id()]->list_lock
);
2153 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2157 assert_spin_locked(&cachep
->nodelists
[node
]->list_lock
);
2162 #define check_irq_off() do { } while(0)
2163 #define check_irq_on() do { } while(0)
2164 #define check_spinlock_acquired(x) do { } while(0)
2165 #define check_spinlock_acquired_node(x, y) do { } while(0)
2168 static void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
2169 struct array_cache
*ac
,
2170 int force
, int node
);
2172 static void do_drain(void *arg
)
2174 struct kmem_cache
*cachep
= arg
;
2175 struct array_cache
*ac
;
2176 int node
= numa_node_id();
2179 ac
= cpu_cache_get(cachep
);
2180 spin_lock(&cachep
->nodelists
[node
]->list_lock
);
2181 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
2182 spin_unlock(&cachep
->nodelists
[node
]->list_lock
);
2186 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2188 struct kmem_list3
*l3
;
2191 on_each_cpu(do_drain
, cachep
, 1, 1);
2193 for_each_online_node(node
) {
2194 l3
= cachep
->nodelists
[node
];
2196 drain_array(cachep
, l3
, l3
->shared
, 1, node
);
2198 drain_alien_cache(cachep
, l3
->alien
);
2203 static int __node_shrink(struct kmem_cache
*cachep
, int node
)
2206 struct kmem_list3
*l3
= cachep
->nodelists
[node
];
2210 struct list_head
*p
;
2212 p
= l3
->slabs_free
.prev
;
2213 if (p
== &l3
->slabs_free
)
2216 slabp
= list_entry(l3
->slabs_free
.prev
, struct slab
, list
);
2218 BUG_ON(slabp
->inuse
);
2220 list_del(&slabp
->list
);
2222 l3
->free_objects
-= cachep
->num
;
2223 spin_unlock_irq(&l3
->list_lock
);
2224 slab_destroy(cachep
, slabp
);
2225 spin_lock_irq(&l3
->list_lock
);
2227 ret
= !list_empty(&l3
->slabs_full
) || !list_empty(&l3
->slabs_partial
);
2231 static int __cache_shrink(struct kmem_cache
*cachep
)
2234 struct kmem_list3
*l3
;
2236 drain_cpu_caches(cachep
);
2239 for_each_online_node(i
) {
2240 l3
= cachep
->nodelists
[i
];
2242 spin_lock_irq(&l3
->list_lock
);
2243 ret
+= __node_shrink(cachep
, i
);
2244 spin_unlock_irq(&l3
->list_lock
);
2247 return (ret
? 1 : 0);
2251 * kmem_cache_shrink - Shrink a cache.
2252 * @cachep: The cache to shrink.
2254 * Releases as many slabs as possible for a cache.
2255 * To help debugging, a zero exit status indicates all slabs were released.
2257 int kmem_cache_shrink(struct kmem_cache
*cachep
)
2259 BUG_ON(!cachep
|| in_interrupt());
2261 return __cache_shrink(cachep
);
2263 EXPORT_SYMBOL(kmem_cache_shrink
);
2266 * kmem_cache_destroy - delete a cache
2267 * @cachep: the cache to destroy
2269 * Remove a struct kmem_cache object from the slab cache.
2270 * Returns 0 on success.
2272 * It is expected this function will be called by a module when it is
2273 * unloaded. This will remove the cache completely, and avoid a duplicate
2274 * cache being allocated each time a module is loaded and unloaded, if the
2275 * module doesn't have persistent in-kernel storage across loads and unloads.
2277 * The cache must be empty before calling this function.
2279 * The caller must guarantee that noone will allocate memory from the cache
2280 * during the kmem_cache_destroy().
2282 int kmem_cache_destroy(struct kmem_cache
*cachep
)
2285 struct kmem_list3
*l3
;
2287 BUG_ON(!cachep
|| in_interrupt());
2289 /* Don't let CPUs to come and go */
2292 /* Find the cache in the chain of caches. */
2293 mutex_lock(&cache_chain_mutex
);
2295 * the chain is never empty, cache_cache is never destroyed
2297 list_del(&cachep
->next
);
2298 mutex_unlock(&cache_chain_mutex
);
2300 if (__cache_shrink(cachep
)) {
2301 slab_error(cachep
, "Can't free all objects");
2302 mutex_lock(&cache_chain_mutex
);
2303 list_add(&cachep
->next
, &cache_chain
);
2304 mutex_unlock(&cache_chain_mutex
);
2305 unlock_cpu_hotplug();
2309 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
2312 for_each_online_cpu(i
)
2313 kfree(cachep
->array
[i
]);
2315 /* NUMA: free the list3 structures */
2316 for_each_online_node(i
) {
2317 l3
= cachep
->nodelists
[i
];
2320 free_alien_cache(l3
->alien
);
2324 kmem_cache_free(&cache_cache
, cachep
);
2325 unlock_cpu_hotplug();
2328 EXPORT_SYMBOL(kmem_cache_destroy
);
2330 /* Get the memory for a slab management obj. */
2331 static struct slab
*alloc_slabmgmt(struct kmem_cache
*cachep
, void *objp
,
2332 int colour_off
, gfp_t local_flags
,
2337 if (OFF_SLAB(cachep
)) {
2338 /* Slab management obj is off-slab. */
2339 slabp
= kmem_cache_alloc_node(cachep
->slabp_cache
,
2340 local_flags
, nodeid
);
2344 slabp
= objp
+ colour_off
;
2345 colour_off
+= cachep
->slab_size
;
2348 slabp
->colouroff
= colour_off
;
2349 slabp
->s_mem
= objp
+ colour_off
;
2350 slabp
->nodeid
= nodeid
;
2354 static inline kmem_bufctl_t
*slab_bufctl(struct slab
*slabp
)
2356 return (kmem_bufctl_t
*) (slabp
+ 1);
2359 static void cache_init_objs(struct kmem_cache
*cachep
,
2360 struct slab
*slabp
, unsigned long ctor_flags
)
2364 for (i
= 0; i
< cachep
->num
; i
++) {
2365 void *objp
= index_to_obj(cachep
, slabp
, i
);
2367 /* need to poison the objs? */
2368 if (cachep
->flags
& SLAB_POISON
)
2369 poison_obj(cachep
, objp
, POISON_FREE
);
2370 if (cachep
->flags
& SLAB_STORE_USER
)
2371 *dbg_userword(cachep
, objp
) = NULL
;
2373 if (cachep
->flags
& SLAB_RED_ZONE
) {
2374 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2375 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2378 * Constructors are not allowed to allocate memory from the same
2379 * cache which they are a constructor for. Otherwise, deadlock.
2380 * They must also be threaded.
2382 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2383 cachep
->ctor(objp
+ obj_offset(cachep
), cachep
,
2386 if (cachep
->flags
& SLAB_RED_ZONE
) {
2387 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2388 slab_error(cachep
, "constructor overwrote the"
2389 " end of an object");
2390 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2391 slab_error(cachep
, "constructor overwrote the"
2392 " start of an object");
2394 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 &&
2395 OFF_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
)
2396 kernel_map_pages(virt_to_page(objp
),
2397 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2400 cachep
->ctor(objp
, cachep
, ctor_flags
);
2402 slab_bufctl(slabp
)[i
] = i
+ 1;
2404 slab_bufctl(slabp
)[i
- 1] = BUFCTL_END
;
2408 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2410 if (flags
& SLAB_DMA
)
2411 BUG_ON(!(cachep
->gfpflags
& GFP_DMA
));
2413 BUG_ON(cachep
->gfpflags
& GFP_DMA
);
2416 static void *slab_get_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2419 void *objp
= index_to_obj(cachep
, slabp
, slabp
->free
);
2423 next
= slab_bufctl(slabp
)[slabp
->free
];
2425 slab_bufctl(slabp
)[slabp
->free
] = BUFCTL_FREE
;
2426 WARN_ON(slabp
->nodeid
!= nodeid
);
2433 static void slab_put_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2434 void *objp
, int nodeid
)
2436 unsigned int objnr
= obj_to_index(cachep
, slabp
, objp
);
2439 /* Verify that the slab belongs to the intended node */
2440 WARN_ON(slabp
->nodeid
!= nodeid
);
2442 if (slab_bufctl(slabp
)[objnr
] + 1 <= SLAB_LIMIT
+ 1) {
2443 printk(KERN_ERR
"slab: double free detected in cache "
2444 "'%s', objp %p\n", cachep
->name
, objp
);
2448 slab_bufctl(slabp
)[objnr
] = slabp
->free
;
2449 slabp
->free
= objnr
;
2453 static void set_slab_attr(struct kmem_cache
*cachep
, struct slab
*slabp
,
2459 /* Nasty!!!!!! I hope this is OK. */
2460 page
= virt_to_page(objp
);
2463 if (likely(!PageCompound(page
)))
2464 i
<<= cachep
->gfporder
;
2466 page_set_cache(page
, cachep
);
2467 page_set_slab(page
, slabp
);
2473 * Grow (by 1) the number of slabs within a cache. This is called by
2474 * kmem_cache_alloc() when there are no active objs left in a cache.
2476 static int cache_grow(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
2482 unsigned long ctor_flags
;
2483 struct kmem_list3
*l3
;
2486 * Be lazy and only check for valid flags here, keeping it out of the
2487 * critical path in kmem_cache_alloc().
2489 BUG_ON(flags
& ~(SLAB_DMA
| SLAB_LEVEL_MASK
| SLAB_NO_GROW
));
2490 if (flags
& SLAB_NO_GROW
)
2493 ctor_flags
= SLAB_CTOR_CONSTRUCTOR
;
2494 local_flags
= (flags
& SLAB_LEVEL_MASK
);
2495 if (!(local_flags
& __GFP_WAIT
))
2497 * Not allowed to sleep. Need to tell a constructor about
2498 * this - it might need to know...
2500 ctor_flags
|= SLAB_CTOR_ATOMIC
;
2502 /* Take the l3 list lock to change the colour_next on this node */
2504 l3
= cachep
->nodelists
[nodeid
];
2505 spin_lock(&l3
->list_lock
);
2507 /* Get colour for the slab, and cal the next value. */
2508 offset
= l3
->colour_next
;
2510 if (l3
->colour_next
>= cachep
->colour
)
2511 l3
->colour_next
= 0;
2512 spin_unlock(&l3
->list_lock
);
2514 offset
*= cachep
->colour_off
;
2516 if (local_flags
& __GFP_WAIT
)
2520 * The test for missing atomic flag is performed here, rather than
2521 * the more obvious place, simply to reduce the critical path length
2522 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2523 * will eventually be caught here (where it matters).
2525 kmem_flagcheck(cachep
, flags
);
2528 * Get mem for the objs. Attempt to allocate a physical page from
2531 objp
= kmem_getpages(cachep
, flags
, nodeid
);
2535 /* Get slab management. */
2536 slabp
= alloc_slabmgmt(cachep
, objp
, offset
, local_flags
, nodeid
);
2540 slabp
->nodeid
= nodeid
;
2541 set_slab_attr(cachep
, slabp
, objp
);
2543 cache_init_objs(cachep
, slabp
, ctor_flags
);
2545 if (local_flags
& __GFP_WAIT
)
2546 local_irq_disable();
2548 spin_lock(&l3
->list_lock
);
2550 /* Make slab active. */
2551 list_add_tail(&slabp
->list
, &(l3
->slabs_free
));
2552 STATS_INC_GROWN(cachep
);
2553 l3
->free_objects
+= cachep
->num
;
2554 spin_unlock(&l3
->list_lock
);
2557 kmem_freepages(cachep
, objp
);
2559 if (local_flags
& __GFP_WAIT
)
2560 local_irq_disable();
2567 * Perform extra freeing checks:
2568 * - detect bad pointers.
2569 * - POISON/RED_ZONE checking
2570 * - destructor calls, for caches with POISON+dtor
2572 static void kfree_debugcheck(const void *objp
)
2576 if (!virt_addr_valid(objp
)) {
2577 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
2578 (unsigned long)objp
);
2581 page
= virt_to_page(objp
);
2582 if (!PageSlab(page
)) {
2583 printk(KERN_ERR
"kfree_debugcheck: bad ptr %lxh.\n",
2584 (unsigned long)objp
);
2589 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2596 objp
-= obj_offset(cachep
);
2597 kfree_debugcheck(objp
);
2598 page
= virt_to_page(objp
);
2600 if (page_get_cache(page
) != cachep
) {
2601 printk(KERN_ERR
"mismatch in kmem_cache_free: expected "
2602 "cache %p, got %p\n",
2603 page_get_cache(page
), cachep
);
2604 printk(KERN_ERR
"%p is %s.\n", cachep
, cachep
->name
);
2605 printk(KERN_ERR
"%p is %s.\n", page_get_cache(page
),
2606 page_get_cache(page
)->name
);
2609 slabp
= page_get_slab(page
);
2611 if (cachep
->flags
& SLAB_RED_ZONE
) {
2612 if (*dbg_redzone1(cachep
, objp
) != RED_ACTIVE
||
2613 *dbg_redzone2(cachep
, objp
) != RED_ACTIVE
) {
2614 slab_error(cachep
, "double free, or memory outside"
2615 " object was overwritten");
2616 printk(KERN_ERR
"%p: redzone 1:0x%lx, "
2617 "redzone 2:0x%lx.\n",
2618 objp
, *dbg_redzone1(cachep
, objp
),
2619 *dbg_redzone2(cachep
, objp
));
2621 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2622 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2624 if (cachep
->flags
& SLAB_STORE_USER
)
2625 *dbg_userword(cachep
, objp
) = caller
;
2627 objnr
= obj_to_index(cachep
, slabp
, objp
);
2629 BUG_ON(objnr
>= cachep
->num
);
2630 BUG_ON(objp
!= index_to_obj(cachep
, slabp
, objnr
));
2632 if (cachep
->flags
& SLAB_DEBUG_INITIAL
) {
2634 * Need to call the slab's constructor so the caller can
2635 * perform a verify of its state (debugging). Called without
2636 * the cache-lock held.
2638 cachep
->ctor(objp
+ obj_offset(cachep
),
2639 cachep
, SLAB_CTOR_CONSTRUCTOR
| SLAB_CTOR_VERIFY
);
2641 if (cachep
->flags
& SLAB_POISON
&& cachep
->dtor
) {
2642 /* we want to cache poison the object,
2643 * call the destruction callback
2645 cachep
->dtor(objp
+ obj_offset(cachep
), cachep
, 0);
2647 #ifdef CONFIG_DEBUG_SLAB_LEAK
2648 slab_bufctl(slabp
)[objnr
] = BUFCTL_FREE
;
2650 if (cachep
->flags
& SLAB_POISON
) {
2651 #ifdef CONFIG_DEBUG_PAGEALLOC
2652 if ((cachep
->buffer_size
% PAGE_SIZE
)==0 && OFF_SLAB(cachep
)) {
2653 store_stackinfo(cachep
, objp
, (unsigned long)caller
);
2654 kernel_map_pages(virt_to_page(objp
),
2655 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2657 poison_obj(cachep
, objp
, POISON_FREE
);
2660 poison_obj(cachep
, objp
, POISON_FREE
);
2666 static void check_slabp(struct kmem_cache
*cachep
, struct slab
*slabp
)
2671 /* Check slab's freelist to see if this obj is there. */
2672 for (i
= slabp
->free
; i
!= BUFCTL_END
; i
= slab_bufctl(slabp
)[i
]) {
2674 if (entries
> cachep
->num
|| i
>= cachep
->num
)
2677 if (entries
!= cachep
->num
- slabp
->inuse
) {
2679 printk(KERN_ERR
"slab: Internal list corruption detected in "
2680 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2681 cachep
->name
, cachep
->num
, slabp
, slabp
->inuse
);
2683 i
< sizeof(*slabp
) + cachep
->num
* sizeof(kmem_bufctl_t
);
2686 printk("\n%03x:", i
);
2687 printk(" %02x", ((unsigned char *)slabp
)[i
]);
2694 #define kfree_debugcheck(x) do { } while(0)
2695 #define cache_free_debugcheck(x,objp,z) (objp)
2696 #define check_slabp(x,y) do { } while(0)
2699 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
2702 struct kmem_list3
*l3
;
2703 struct array_cache
*ac
;
2706 ac
= cpu_cache_get(cachep
);
2708 batchcount
= ac
->batchcount
;
2709 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2711 * If there was little recent activity on this cache, then
2712 * perform only a partial refill. Otherwise we could generate
2715 batchcount
= BATCHREFILL_LIMIT
;
2717 l3
= cachep
->nodelists
[numa_node_id()];
2719 BUG_ON(ac
->avail
> 0 || !l3
);
2720 spin_lock(&l3
->list_lock
);
2722 /* See if we can refill from the shared array */
2723 if (l3
->shared
&& transfer_objects(ac
, l3
->shared
, batchcount
))
2726 while (batchcount
> 0) {
2727 struct list_head
*entry
;
2729 /* Get slab alloc is to come from. */
2730 entry
= l3
->slabs_partial
.next
;
2731 if (entry
== &l3
->slabs_partial
) {
2732 l3
->free_touched
= 1;
2733 entry
= l3
->slabs_free
.next
;
2734 if (entry
== &l3
->slabs_free
)
2738 slabp
= list_entry(entry
, struct slab
, list
);
2739 check_slabp(cachep
, slabp
);
2740 check_spinlock_acquired(cachep
);
2741 while (slabp
->inuse
< cachep
->num
&& batchcount
--) {
2742 STATS_INC_ALLOCED(cachep
);
2743 STATS_INC_ACTIVE(cachep
);
2744 STATS_SET_HIGH(cachep
);
2746 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, slabp
,
2749 check_slabp(cachep
, slabp
);
2751 /* move slabp to correct slabp list: */
2752 list_del(&slabp
->list
);
2753 if (slabp
->free
== BUFCTL_END
)
2754 list_add(&slabp
->list
, &l3
->slabs_full
);
2756 list_add(&slabp
->list
, &l3
->slabs_partial
);
2760 l3
->free_objects
-= ac
->avail
;
2762 spin_unlock(&l3
->list_lock
);
2764 if (unlikely(!ac
->avail
)) {
2766 x
= cache_grow(cachep
, flags
, numa_node_id());
2768 /* cache_grow can reenable interrupts, then ac could change. */
2769 ac
= cpu_cache_get(cachep
);
2770 if (!x
&& ac
->avail
== 0) /* no objects in sight? abort */
2773 if (!ac
->avail
) /* objects refilled by interrupt? */
2777 return ac
->entry
[--ac
->avail
];
2780 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
2783 might_sleep_if(flags
& __GFP_WAIT
);
2785 kmem_flagcheck(cachep
, flags
);
2790 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
2791 gfp_t flags
, void *objp
, void *caller
)
2795 if (cachep
->flags
& SLAB_POISON
) {
2796 #ifdef CONFIG_DEBUG_PAGEALLOC
2797 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
2798 kernel_map_pages(virt_to_page(objp
),
2799 cachep
->buffer_size
/ PAGE_SIZE
, 1);
2801 check_poison_obj(cachep
, objp
);
2803 check_poison_obj(cachep
, objp
);
2805 poison_obj(cachep
, objp
, POISON_INUSE
);
2807 if (cachep
->flags
& SLAB_STORE_USER
)
2808 *dbg_userword(cachep
, objp
) = caller
;
2810 if (cachep
->flags
& SLAB_RED_ZONE
) {
2811 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
2812 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
2813 slab_error(cachep
, "double free, or memory outside"
2814 " object was overwritten");
2816 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
2817 objp
, *dbg_redzone1(cachep
, objp
),
2818 *dbg_redzone2(cachep
, objp
));
2820 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
2821 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
2823 #ifdef CONFIG_DEBUG_SLAB_LEAK
2828 slabp
= page_get_slab(virt_to_page(objp
));
2829 objnr
= (unsigned)(objp
- slabp
->s_mem
) / cachep
->buffer_size
;
2830 slab_bufctl(slabp
)[objnr
] = BUFCTL_ACTIVE
;
2833 objp
+= obj_offset(cachep
);
2834 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
) {
2835 unsigned long ctor_flags
= SLAB_CTOR_CONSTRUCTOR
;
2837 if (!(flags
& __GFP_WAIT
))
2838 ctor_flags
|= SLAB_CTOR_ATOMIC
;
2840 cachep
->ctor(objp
, cachep
, ctor_flags
);
2845 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2848 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
2851 struct array_cache
*ac
;
2854 if (unlikely(current
->flags
& (PF_SPREAD_SLAB
| PF_MEMPOLICY
))) {
2855 objp
= alternate_node_alloc(cachep
, flags
);
2862 ac
= cpu_cache_get(cachep
);
2863 if (likely(ac
->avail
)) {
2864 STATS_INC_ALLOCHIT(cachep
);
2866 objp
= ac
->entry
[--ac
->avail
];
2868 STATS_INC_ALLOCMISS(cachep
);
2869 objp
= cache_alloc_refill(cachep
, flags
);
2874 static __always_inline
void *__cache_alloc(struct kmem_cache
*cachep
,
2875 gfp_t flags
, void *caller
)
2877 unsigned long save_flags
;
2880 cache_alloc_debugcheck_before(cachep
, flags
);
2882 local_irq_save(save_flags
);
2883 objp
= ____cache_alloc(cachep
, flags
);
2884 local_irq_restore(save_flags
);
2885 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
,
2893 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
2895 * If we are in_interrupt, then process context, including cpusets and
2896 * mempolicy, may not apply and should not be used for allocation policy.
2898 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
2900 int nid_alloc
, nid_here
;
2904 nid_alloc
= nid_here
= numa_node_id();
2905 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
2906 nid_alloc
= cpuset_mem_spread_node();
2907 else if (current
->mempolicy
)
2908 nid_alloc
= slab_node(current
->mempolicy
);
2909 if (nid_alloc
!= nid_here
)
2910 return __cache_alloc_node(cachep
, flags
, nid_alloc
);
2915 * A interface to enable slab creation on nodeid
2917 static void *__cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
2920 struct list_head
*entry
;
2922 struct kmem_list3
*l3
;
2926 l3
= cachep
->nodelists
[nodeid
];
2931 spin_lock(&l3
->list_lock
);
2932 entry
= l3
->slabs_partial
.next
;
2933 if (entry
== &l3
->slabs_partial
) {
2934 l3
->free_touched
= 1;
2935 entry
= l3
->slabs_free
.next
;
2936 if (entry
== &l3
->slabs_free
)
2940 slabp
= list_entry(entry
, struct slab
, list
);
2941 check_spinlock_acquired_node(cachep
, nodeid
);
2942 check_slabp(cachep
, slabp
);
2944 STATS_INC_NODEALLOCS(cachep
);
2945 STATS_INC_ACTIVE(cachep
);
2946 STATS_SET_HIGH(cachep
);
2948 BUG_ON(slabp
->inuse
== cachep
->num
);
2950 obj
= slab_get_obj(cachep
, slabp
, nodeid
);
2951 check_slabp(cachep
, slabp
);
2953 /* move slabp to correct slabp list: */
2954 list_del(&slabp
->list
);
2956 if (slabp
->free
== BUFCTL_END
)
2957 list_add(&slabp
->list
, &l3
->slabs_full
);
2959 list_add(&slabp
->list
, &l3
->slabs_partial
);
2961 spin_unlock(&l3
->list_lock
);
2965 spin_unlock(&l3
->list_lock
);
2966 x
= cache_grow(cachep
, flags
, nodeid
);
2978 * Caller needs to acquire correct kmem_list's list_lock
2980 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int nr_objects
,
2984 struct kmem_list3
*l3
;
2986 for (i
= 0; i
< nr_objects
; i
++) {
2987 void *objp
= objpp
[i
];
2990 slabp
= virt_to_slab(objp
);
2991 l3
= cachep
->nodelists
[node
];
2992 list_del(&slabp
->list
);
2993 check_spinlock_acquired_node(cachep
, node
);
2994 check_slabp(cachep
, slabp
);
2995 slab_put_obj(cachep
, slabp
, objp
, node
);
2996 STATS_DEC_ACTIVE(cachep
);
2998 check_slabp(cachep
, slabp
);
3000 /* fixup slab chains */
3001 if (slabp
->inuse
== 0) {
3002 if (l3
->free_objects
> l3
->free_limit
) {
3003 l3
->free_objects
-= cachep
->num
;
3004 slab_destroy(cachep
, slabp
);
3006 list_add(&slabp
->list
, &l3
->slabs_free
);
3009 /* Unconditionally move a slab to the end of the
3010 * partial list on free - maximum time for the
3011 * other objects to be freed, too.
3013 list_add_tail(&slabp
->list
, &l3
->slabs_partial
);
3018 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3021 struct kmem_list3
*l3
;
3022 int node
= numa_node_id();
3024 batchcount
= ac
->batchcount
;
3026 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
3029 l3
= cachep
->nodelists
[node
];
3030 spin_lock(&l3
->list_lock
);
3032 struct array_cache
*shared_array
= l3
->shared
;
3033 int max
= shared_array
->limit
- shared_array
->avail
;
3035 if (batchcount
> max
)
3037 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3038 ac
->entry
, sizeof(void *) * batchcount
);
3039 shared_array
->avail
+= batchcount
;
3044 free_block(cachep
, ac
->entry
, batchcount
, node
);
3049 struct list_head
*p
;
3051 p
= l3
->slabs_free
.next
;
3052 while (p
!= &(l3
->slabs_free
)) {
3055 slabp
= list_entry(p
, struct slab
, list
);
3056 BUG_ON(slabp
->inuse
);
3061 STATS_SET_FREEABLE(cachep
, i
);
3064 spin_unlock(&l3
->list_lock
);
3065 ac
->avail
-= batchcount
;
3066 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3070 * Release an obj back to its cache. If the obj has a constructed state, it must
3071 * be in this state _before_ it is released. Called with disabled ints.
3073 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
)
3075 struct array_cache
*ac
= cpu_cache_get(cachep
);
3078 objp
= cache_free_debugcheck(cachep
, objp
, __builtin_return_address(0));
3080 /* Make sure we are not freeing a object from another
3081 * node to the array cache on this cpu.
3086 slabp
= virt_to_slab(objp
);
3087 if (unlikely(slabp
->nodeid
!= numa_node_id())) {
3088 struct array_cache
*alien
= NULL
;
3089 int nodeid
= slabp
->nodeid
;
3090 struct kmem_list3
*l3
;
3092 l3
= cachep
->nodelists
[numa_node_id()];
3093 STATS_INC_NODEFREES(cachep
);
3094 if (l3
->alien
&& l3
->alien
[nodeid
]) {
3095 alien
= l3
->alien
[nodeid
];
3096 spin_lock(&alien
->lock
);
3097 if (unlikely(alien
->avail
== alien
->limit
)) {
3098 STATS_INC_ACOVERFLOW(cachep
);
3099 __drain_alien_cache(cachep
,
3102 alien
->entry
[alien
->avail
++] = objp
;
3103 spin_unlock(&alien
->lock
);
3105 spin_lock(&(cachep
->nodelists
[nodeid
])->
3107 free_block(cachep
, &objp
, 1, nodeid
);
3108 spin_unlock(&(cachep
->nodelists
[nodeid
])->
3115 if (likely(ac
->avail
< ac
->limit
)) {
3116 STATS_INC_FREEHIT(cachep
);
3117 ac
->entry
[ac
->avail
++] = objp
;
3120 STATS_INC_FREEMISS(cachep
);
3121 cache_flusharray(cachep
, ac
);
3122 ac
->entry
[ac
->avail
++] = objp
;
3127 * kmem_cache_alloc - Allocate an object
3128 * @cachep: The cache to allocate from.
3129 * @flags: See kmalloc().
3131 * Allocate an object from this cache. The flags are only relevant
3132 * if the cache has no available objects.
3134 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3136 return __cache_alloc(cachep
, flags
, __builtin_return_address(0));
3138 EXPORT_SYMBOL(kmem_cache_alloc
);
3141 * kmem_cache_alloc - Allocate an object. The memory is set to zero.
3142 * @cache: The cache to allocate from.
3143 * @flags: See kmalloc().
3145 * Allocate an object from this cache and set the allocated memory to zero.
3146 * The flags are only relevant if the cache has no available objects.
3148 void *kmem_cache_zalloc(struct kmem_cache
*cache
, gfp_t flags
)
3150 void *ret
= __cache_alloc(cache
, flags
, __builtin_return_address(0));
3152 memset(ret
, 0, obj_size(cache
));
3155 EXPORT_SYMBOL(kmem_cache_zalloc
);
3158 * kmem_ptr_validate - check if an untrusted pointer might
3160 * @cachep: the cache we're checking against
3161 * @ptr: pointer to validate
3163 * This verifies that the untrusted pointer looks sane:
3164 * it is _not_ a guarantee that the pointer is actually
3165 * part of the slab cache in question, but it at least
3166 * validates that the pointer can be dereferenced and
3167 * looks half-way sane.
3169 * Currently only used for dentry validation.
3171 int fastcall
kmem_ptr_validate(struct kmem_cache
*cachep
, void *ptr
)
3173 unsigned long addr
= (unsigned long)ptr
;
3174 unsigned long min_addr
= PAGE_OFFSET
;
3175 unsigned long align_mask
= BYTES_PER_WORD
- 1;
3176 unsigned long size
= cachep
->buffer_size
;
3179 if (unlikely(addr
< min_addr
))
3181 if (unlikely(addr
> (unsigned long)high_memory
- size
))
3183 if (unlikely(addr
& align_mask
))
3185 if (unlikely(!kern_addr_valid(addr
)))
3187 if (unlikely(!kern_addr_valid(addr
+ size
- 1)))
3189 page
= virt_to_page(ptr
);
3190 if (unlikely(!PageSlab(page
)))
3192 if (unlikely(page_get_cache(page
) != cachep
))
3201 * kmem_cache_alloc_node - Allocate an object on the specified node
3202 * @cachep: The cache to allocate from.
3203 * @flags: See kmalloc().
3204 * @nodeid: node number of the target node.
3206 * Identical to kmem_cache_alloc, except that this function is slow
3207 * and can sleep. And it will allocate memory on the given node, which
3208 * can improve the performance for cpu bound structures.
3209 * New and improved: it will now make sure that the object gets
3210 * put on the correct node list so that there is no false sharing.
3212 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3214 unsigned long save_flags
;
3217 cache_alloc_debugcheck_before(cachep
, flags
);
3218 local_irq_save(save_flags
);
3220 if (nodeid
== -1 || nodeid
== numa_node_id() ||
3221 !cachep
->nodelists
[nodeid
])
3222 ptr
= ____cache_alloc(cachep
, flags
);
3224 ptr
= __cache_alloc_node(cachep
, flags
, nodeid
);
3225 local_irq_restore(save_flags
);
3227 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
,
3228 __builtin_return_address(0));
3232 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3234 void *kmalloc_node(size_t size
, gfp_t flags
, int node
)
3236 struct kmem_cache
*cachep
;
3238 cachep
= kmem_find_general_cachep(size
, flags
);
3239 if (unlikely(cachep
== NULL
))
3241 return kmem_cache_alloc_node(cachep
, flags
, node
);
3243 EXPORT_SYMBOL(kmalloc_node
);
3247 * kmalloc - allocate memory
3248 * @size: how many bytes of memory are required.
3249 * @flags: the type of memory to allocate.
3250 * @caller: function caller for debug tracking of the caller
3252 * kmalloc is the normal method of allocating memory
3255 * The @flags argument may be one of:
3257 * %GFP_USER - Allocate memory on behalf of user. May sleep.
3259 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
3261 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
3263 * Additionally, the %GFP_DMA flag may be set to indicate the memory
3264 * must be suitable for DMA. This can mean different things on different
3265 * platforms. For example, on i386, it means that the memory must come
3266 * from the first 16MB.
3268 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3271 struct kmem_cache
*cachep
;
3273 /* If you want to save a few bytes .text space: replace
3275 * Then kmalloc uses the uninlined functions instead of the inline
3278 cachep
= __find_general_cachep(size
, flags
);
3279 if (unlikely(cachep
== NULL
))
3281 return __cache_alloc(cachep
, flags
, caller
);
3285 void *__kmalloc(size_t size
, gfp_t flags
)
3287 #ifndef CONFIG_DEBUG_SLAB
3288 return __do_kmalloc(size
, flags
, NULL
);
3290 return __do_kmalloc(size
, flags
, __builtin_return_address(0));
3293 EXPORT_SYMBOL(__kmalloc
);
3295 #ifdef CONFIG_DEBUG_SLAB
3296 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, void *caller
)
3298 return __do_kmalloc(size
, flags
, caller
);
3300 EXPORT_SYMBOL(__kmalloc_track_caller
);
3305 * __alloc_percpu - allocate one copy of the object for every present
3306 * cpu in the system, zeroing them.
3307 * Objects should be dereferenced using the per_cpu_ptr macro only.
3309 * @size: how many bytes of memory are required.
3311 void *__alloc_percpu(size_t size
)
3314 struct percpu_data
*pdata
= kmalloc(sizeof(*pdata
), GFP_KERNEL
);
3320 * Cannot use for_each_online_cpu since a cpu may come online
3321 * and we have no way of figuring out how to fix the array
3322 * that we have allocated then....
3324 for_each_possible_cpu(i
) {
3325 int node
= cpu_to_node(i
);
3327 if (node_online(node
))
3328 pdata
->ptrs
[i
] = kmalloc_node(size
, GFP_KERNEL
, node
);
3330 pdata
->ptrs
[i
] = kmalloc(size
, GFP_KERNEL
);
3332 if (!pdata
->ptrs
[i
])
3334 memset(pdata
->ptrs
[i
], 0, size
);
3337 /* Catch derefs w/o wrappers */
3338 return (void *)(~(unsigned long)pdata
);
3342 if (!cpu_possible(i
))
3344 kfree(pdata
->ptrs
[i
]);
3349 EXPORT_SYMBOL(__alloc_percpu
);
3353 * kmem_cache_free - Deallocate an object
3354 * @cachep: The cache the allocation was from.
3355 * @objp: The previously allocated object.
3357 * Free an object which was previously allocated from this
3360 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3362 unsigned long flags
;
3364 local_irq_save(flags
);
3365 __cache_free(cachep
, objp
);
3366 local_irq_restore(flags
);
3368 EXPORT_SYMBOL(kmem_cache_free
);
3371 * kfree - free previously allocated memory
3372 * @objp: pointer returned by kmalloc.
3374 * If @objp is NULL, no operation is performed.
3376 * Don't free memory not originally allocated by kmalloc()
3377 * or you will run into trouble.
3379 void kfree(const void *objp
)
3381 struct kmem_cache
*c
;
3382 unsigned long flags
;
3384 if (unlikely(!objp
))
3386 local_irq_save(flags
);
3387 kfree_debugcheck(objp
);
3388 c
= virt_to_cache(objp
);
3389 mutex_debug_check_no_locks_freed(objp
, obj_size(c
));
3390 __cache_free(c
, (void *)objp
);
3391 local_irq_restore(flags
);
3393 EXPORT_SYMBOL(kfree
);
3397 * free_percpu - free previously allocated percpu memory
3398 * @objp: pointer returned by alloc_percpu.
3400 * Don't free memory not originally allocated by alloc_percpu()
3401 * The complemented objp is to check for that.
3403 void free_percpu(const void *objp
)
3406 struct percpu_data
*p
= (struct percpu_data
*)(~(unsigned long)objp
);
3409 * We allocate for all cpus so we cannot use for online cpu here.
3411 for_each_possible_cpu(i
)
3415 EXPORT_SYMBOL(free_percpu
);
3418 unsigned int kmem_cache_size(struct kmem_cache
*cachep
)
3420 return obj_size(cachep
);
3422 EXPORT_SYMBOL(kmem_cache_size
);
3424 const char *kmem_cache_name(struct kmem_cache
*cachep
)
3426 return cachep
->name
;
3428 EXPORT_SYMBOL_GPL(kmem_cache_name
);
3431 * This initializes kmem_list3 or resizes varioius caches for all nodes.
3433 static int alloc_kmemlist(struct kmem_cache
*cachep
)
3436 struct kmem_list3
*l3
;
3437 struct array_cache
*new_shared
;
3438 struct array_cache
**new_alien
;
3440 for_each_online_node(node
) {
3442 new_alien
= alloc_alien_cache(node
, cachep
->limit
);
3446 new_shared
= alloc_arraycache(node
,
3447 cachep
->shared
*cachep
->batchcount
,
3450 free_alien_cache(new_alien
);
3454 l3
= cachep
->nodelists
[node
];
3456 struct array_cache
*shared
= l3
->shared
;
3458 spin_lock_irq(&l3
->list_lock
);
3461 free_block(cachep
, shared
->entry
,
3462 shared
->avail
, node
);
3464 l3
->shared
= new_shared
;
3466 l3
->alien
= new_alien
;
3469 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3470 cachep
->batchcount
+ cachep
->num
;
3471 spin_unlock_irq(&l3
->list_lock
);
3473 free_alien_cache(new_alien
);
3476 l3
= kmalloc_node(sizeof(struct kmem_list3
), GFP_KERNEL
, node
);
3478 free_alien_cache(new_alien
);
3483 kmem_list3_init(l3
);
3484 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
3485 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
3486 l3
->shared
= new_shared
;
3487 l3
->alien
= new_alien
;
3488 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3489 cachep
->batchcount
+ cachep
->num
;
3490 cachep
->nodelists
[node
] = l3
;
3495 if (!cachep
->next
.next
) {
3496 /* Cache is not active yet. Roll back what we did */
3499 if (cachep
->nodelists
[node
]) {
3500 l3
= cachep
->nodelists
[node
];
3503 free_alien_cache(l3
->alien
);
3505 cachep
->nodelists
[node
] = NULL
;
3513 struct ccupdate_struct
{
3514 struct kmem_cache
*cachep
;
3515 struct array_cache
*new[NR_CPUS
];
3518 static void do_ccupdate_local(void *info
)
3520 struct ccupdate_struct
*new = info
;
3521 struct array_cache
*old
;
3524 old
= cpu_cache_get(new->cachep
);
3526 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
3527 new->new[smp_processor_id()] = old
;
3530 /* Always called with the cache_chain_mutex held */
3531 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3532 int batchcount
, int shared
)
3534 struct ccupdate_struct
new;
3537 memset(&new.new, 0, sizeof(new.new));
3538 for_each_online_cpu(i
) {
3539 new.new[i
] = alloc_arraycache(cpu_to_node(i
), limit
,
3542 for (i
--; i
>= 0; i
--)
3547 new.cachep
= cachep
;
3549 on_each_cpu(do_ccupdate_local
, (void *)&new, 1, 1);
3552 cachep
->batchcount
= batchcount
;
3553 cachep
->limit
= limit
;
3554 cachep
->shared
= shared
;
3556 for_each_online_cpu(i
) {
3557 struct array_cache
*ccold
= new.new[i
];
3560 spin_lock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3561 free_block(cachep
, ccold
->entry
, ccold
->avail
, cpu_to_node(i
));
3562 spin_unlock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3566 err
= alloc_kmemlist(cachep
);
3568 printk(KERN_ERR
"alloc_kmemlist failed for %s, error %d.\n",
3569 cachep
->name
, -err
);
3575 /* Called with cache_chain_mutex held always */
3576 static void enable_cpucache(struct kmem_cache
*cachep
)
3582 * The head array serves three purposes:
3583 * - create a LIFO ordering, i.e. return objects that are cache-warm
3584 * - reduce the number of spinlock operations.
3585 * - reduce the number of linked list operations on the slab and
3586 * bufctl chains: array operations are cheaper.
3587 * The numbers are guessed, we should auto-tune as described by
3590 if (cachep
->buffer_size
> 131072)
3592 else if (cachep
->buffer_size
> PAGE_SIZE
)
3594 else if (cachep
->buffer_size
> 1024)
3596 else if (cachep
->buffer_size
> 256)
3602 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3603 * allocation behaviour: Most allocs on one cpu, most free operations
3604 * on another cpu. For these cases, an efficient object passing between
3605 * cpus is necessary. This is provided by a shared array. The array
3606 * replaces Bonwick's magazine layer.
3607 * On uniprocessor, it's functionally equivalent (but less efficient)
3608 * to a larger limit. Thus disabled by default.
3612 if (cachep
->buffer_size
<= PAGE_SIZE
)
3618 * With debugging enabled, large batchcount lead to excessively long
3619 * periods with disabled local interrupts. Limit the batchcount
3624 err
= do_tune_cpucache(cachep
, limit
, (limit
+ 1) / 2, shared
);
3626 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
3627 cachep
->name
, -err
);
3631 * Drain an array if it contains any elements taking the l3 lock only if
3632 * necessary. Note that the l3 listlock also protects the array_cache
3633 * if drain_array() is used on the shared array.
3635 void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
3636 struct array_cache
*ac
, int force
, int node
)
3640 if (!ac
|| !ac
->avail
)
3642 if (ac
->touched
&& !force
) {
3645 spin_lock_irq(&l3
->list_lock
);
3647 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
3648 if (tofree
> ac
->avail
)
3649 tofree
= (ac
->avail
+ 1) / 2;
3650 free_block(cachep
, ac
->entry
, tofree
, node
);
3651 ac
->avail
-= tofree
;
3652 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
3653 sizeof(void *) * ac
->avail
);
3655 spin_unlock_irq(&l3
->list_lock
);
3660 * cache_reap - Reclaim memory from caches.
3661 * @unused: unused parameter
3663 * Called from workqueue/eventd every few seconds.
3665 * - clear the per-cpu caches for this CPU.
3666 * - return freeable pages to the main free memory pool.
3668 * If we cannot acquire the cache chain mutex then just give up - we'll try
3669 * again on the next iteration.
3671 static void cache_reap(void *unused
)
3673 struct list_head
*walk
;
3674 struct kmem_list3
*l3
;
3675 int node
= numa_node_id();
3677 if (!mutex_trylock(&cache_chain_mutex
)) {
3678 /* Give up. Setup the next iteration. */
3679 schedule_delayed_work(&__get_cpu_var(reap_work
),
3684 list_for_each(walk
, &cache_chain
) {
3685 struct kmem_cache
*searchp
;
3686 struct list_head
*p
;
3690 searchp
= list_entry(walk
, struct kmem_cache
, next
);
3694 * We only take the l3 lock if absolutely necessary and we
3695 * have established with reasonable certainty that
3696 * we can do some work if the lock was obtained.
3698 l3
= searchp
->nodelists
[node
];
3700 reap_alien(searchp
, l3
);
3702 drain_array(searchp
, l3
, cpu_cache_get(searchp
), 0, node
);
3705 * These are racy checks but it does not matter
3706 * if we skip one check or scan twice.
3708 if (time_after(l3
->next_reap
, jiffies
))
3711 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
;
3713 drain_array(searchp
, l3
, l3
->shared
, 0, node
);
3715 if (l3
->free_touched
) {
3716 l3
->free_touched
= 0;
3720 tofree
= (l3
->free_limit
+ 5 * searchp
->num
- 1) /
3724 * Do not lock if there are no free blocks.
3726 if (list_empty(&l3
->slabs_free
))
3729 spin_lock_irq(&l3
->list_lock
);
3730 p
= l3
->slabs_free
.next
;
3731 if (p
== &(l3
->slabs_free
)) {
3732 spin_unlock_irq(&l3
->list_lock
);
3736 slabp
= list_entry(p
, struct slab
, list
);
3737 BUG_ON(slabp
->inuse
);
3738 list_del(&slabp
->list
);
3739 STATS_INC_REAPED(searchp
);
3742 * Safe to drop the lock. The slab is no longer linked
3743 * to the cache. searchp cannot disappear, we hold
3746 l3
->free_objects
-= searchp
->num
;
3747 spin_unlock_irq(&l3
->list_lock
);
3748 slab_destroy(searchp
, slabp
);
3749 } while (--tofree
> 0);
3754 mutex_unlock(&cache_chain_mutex
);
3756 /* Set up the next iteration */
3757 schedule_delayed_work(&__get_cpu_var(reap_work
), REAPTIMEOUT_CPUC
);
3760 #ifdef CONFIG_PROC_FS
3762 static void print_slabinfo_header(struct seq_file
*m
)
3765 * Output format version, so at least we can change it
3766 * without _too_ many complaints.
3769 seq_puts(m
, "slabinfo - version: 2.1 (statistics)\n");
3771 seq_puts(m
, "slabinfo - version: 2.1\n");
3773 seq_puts(m
, "# name <active_objs> <num_objs> <objsize> "
3774 "<objperslab> <pagesperslab>");
3775 seq_puts(m
, " : tunables <limit> <batchcount> <sharedfactor>");
3776 seq_puts(m
, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3778 seq_puts(m
, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3779 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
3780 seq_puts(m
, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3785 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
3788 struct list_head
*p
;
3790 mutex_lock(&cache_chain_mutex
);
3792 print_slabinfo_header(m
);
3793 p
= cache_chain
.next
;
3796 if (p
== &cache_chain
)
3799 return list_entry(p
, struct kmem_cache
, next
);
3802 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
3804 struct kmem_cache
*cachep
= p
;
3806 return cachep
->next
.next
== &cache_chain
?
3807 NULL
: list_entry(cachep
->next
.next
, struct kmem_cache
, next
);
3810 static void s_stop(struct seq_file
*m
, void *p
)
3812 mutex_unlock(&cache_chain_mutex
);
3815 static int s_show(struct seq_file
*m
, void *p
)
3817 struct kmem_cache
*cachep
= p
;
3818 struct list_head
*q
;
3820 unsigned long active_objs
;
3821 unsigned long num_objs
;
3822 unsigned long active_slabs
= 0;
3823 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
3827 struct kmem_list3
*l3
;
3831 for_each_online_node(node
) {
3832 l3
= cachep
->nodelists
[node
];
3837 spin_lock_irq(&l3
->list_lock
);
3839 list_for_each(q
, &l3
->slabs_full
) {
3840 slabp
= list_entry(q
, struct slab
, list
);
3841 if (slabp
->inuse
!= cachep
->num
&& !error
)
3842 error
= "slabs_full accounting error";
3843 active_objs
+= cachep
->num
;
3846 list_for_each(q
, &l3
->slabs_partial
) {
3847 slabp
= list_entry(q
, struct slab
, list
);
3848 if (slabp
->inuse
== cachep
->num
&& !error
)
3849 error
= "slabs_partial inuse accounting error";
3850 if (!slabp
->inuse
&& !error
)
3851 error
= "slabs_partial/inuse accounting error";
3852 active_objs
+= slabp
->inuse
;
3855 list_for_each(q
, &l3
->slabs_free
) {
3856 slabp
= list_entry(q
, struct slab
, list
);
3857 if (slabp
->inuse
&& !error
)
3858 error
= "slabs_free/inuse accounting error";
3861 free_objects
+= l3
->free_objects
;
3863 shared_avail
+= l3
->shared
->avail
;
3865 spin_unlock_irq(&l3
->list_lock
);
3867 num_slabs
+= active_slabs
;
3868 num_objs
= num_slabs
* cachep
->num
;
3869 if (num_objs
- active_objs
!= free_objects
&& !error
)
3870 error
= "free_objects accounting error";
3872 name
= cachep
->name
;
3874 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
3876 seq_printf(m
, "%-17s %6lu %6lu %6u %4u %4d",
3877 name
, active_objs
, num_objs
, cachep
->buffer_size
,
3878 cachep
->num
, (1 << cachep
->gfporder
));
3879 seq_printf(m
, " : tunables %4u %4u %4u",
3880 cachep
->limit
, cachep
->batchcount
, cachep
->shared
);
3881 seq_printf(m
, " : slabdata %6lu %6lu %6lu",
3882 active_slabs
, num_slabs
, shared_avail
);
3885 unsigned long high
= cachep
->high_mark
;
3886 unsigned long allocs
= cachep
->num_allocations
;
3887 unsigned long grown
= cachep
->grown
;
3888 unsigned long reaped
= cachep
->reaped
;
3889 unsigned long errors
= cachep
->errors
;
3890 unsigned long max_freeable
= cachep
->max_freeable
;
3891 unsigned long node_allocs
= cachep
->node_allocs
;
3892 unsigned long node_frees
= cachep
->node_frees
;
3893 unsigned long overflows
= cachep
->node_overflow
;
3895 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu \
3896 %4lu %4lu %4lu %4lu %4lu", allocs
, high
, grown
,
3897 reaped
, errors
, max_freeable
, node_allocs
,
3898 node_frees
, overflows
);
3902 unsigned long allochit
= atomic_read(&cachep
->allochit
);
3903 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
3904 unsigned long freehit
= atomic_read(&cachep
->freehit
);
3905 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
3907 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
3908 allochit
, allocmiss
, freehit
, freemiss
);
3916 * slabinfo_op - iterator that generates /proc/slabinfo
3925 * num-pages-per-slab
3926 * + further values on SMP and with statistics enabled
3929 struct seq_operations slabinfo_op
= {
3936 #define MAX_SLABINFO_WRITE 128
3938 * slabinfo_write - Tuning for the slab allocator
3940 * @buffer: user buffer
3941 * @count: data length
3944 ssize_t
slabinfo_write(struct file
*file
, const char __user
* buffer
,
3945 size_t count
, loff_t
*ppos
)
3947 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
3948 int limit
, batchcount
, shared
, res
;
3949 struct list_head
*p
;
3951 if (count
> MAX_SLABINFO_WRITE
)
3953 if (copy_from_user(&kbuf
, buffer
, count
))
3955 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
3957 tmp
= strchr(kbuf
, ' ');
3962 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
3965 /* Find the cache in the chain of caches. */
3966 mutex_lock(&cache_chain_mutex
);
3968 list_for_each(p
, &cache_chain
) {
3969 struct kmem_cache
*cachep
;
3971 cachep
= list_entry(p
, struct kmem_cache
, next
);
3972 if (!strcmp(cachep
->name
, kbuf
)) {
3973 if (limit
< 1 || batchcount
< 1 ||
3974 batchcount
> limit
|| shared
< 0) {
3977 res
= do_tune_cpucache(cachep
, limit
,
3978 batchcount
, shared
);
3983 mutex_unlock(&cache_chain_mutex
);
3989 #ifdef CONFIG_DEBUG_SLAB_LEAK
3991 static void *leaks_start(struct seq_file
*m
, loff_t
*pos
)
3994 struct list_head
*p
;
3996 mutex_lock(&cache_chain_mutex
);
3997 p
= cache_chain
.next
;
4000 if (p
== &cache_chain
)
4003 return list_entry(p
, struct kmem_cache
, next
);
4006 static inline int add_caller(unsigned long *n
, unsigned long v
)
4016 unsigned long *q
= p
+ 2 * i
;
4030 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4036 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
, struct slab
*s
)
4042 for (i
= 0, p
= s
->s_mem
; i
< c
->num
; i
++, p
+= c
->buffer_size
) {
4043 if (slab_bufctl(s
)[i
] != BUFCTL_ACTIVE
)
4045 if (!add_caller(n
, (unsigned long)*dbg_userword(c
, p
)))
4050 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4052 #ifdef CONFIG_KALLSYMS
4055 unsigned long offset
, size
;
4056 char namebuf
[KSYM_NAME_LEN
+1];
4058 name
= kallsyms_lookup(address
, &size
, &offset
, &modname
, namebuf
);
4061 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4063 seq_printf(m
, " [%s]", modname
);
4067 seq_printf(m
, "%p", (void *)address
);
4070 static int leaks_show(struct seq_file
*m
, void *p
)
4072 struct kmem_cache
*cachep
= p
;
4073 struct list_head
*q
;
4075 struct kmem_list3
*l3
;
4077 unsigned long *n
= m
->private;
4081 if (!(cachep
->flags
& SLAB_STORE_USER
))
4083 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4086 /* OK, we can do it */
4090 for_each_online_node(node
) {
4091 l3
= cachep
->nodelists
[node
];
4096 spin_lock_irq(&l3
->list_lock
);
4098 list_for_each(q
, &l3
->slabs_full
) {
4099 slabp
= list_entry(q
, struct slab
, list
);
4100 handle_slab(n
, cachep
, slabp
);
4102 list_for_each(q
, &l3
->slabs_partial
) {
4103 slabp
= list_entry(q
, struct slab
, list
);
4104 handle_slab(n
, cachep
, slabp
);
4106 spin_unlock_irq(&l3
->list_lock
);
4108 name
= cachep
->name
;
4110 /* Increase the buffer size */
4111 mutex_unlock(&cache_chain_mutex
);
4112 m
->private = kzalloc(n
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4114 /* Too bad, we are really out */
4116 mutex_lock(&cache_chain_mutex
);
4119 *(unsigned long *)m
->private = n
[0] * 2;
4121 mutex_lock(&cache_chain_mutex
);
4122 /* Now make sure this entry will be retried */
4126 for (i
= 0; i
< n
[1]; i
++) {
4127 seq_printf(m
, "%s: %lu ", name
, n
[2*i
+3]);
4128 show_symbol(m
, n
[2*i
+2]);
4134 struct seq_operations slabstats_op
= {
4135 .start
= leaks_start
,
4144 * ksize - get the actual amount of memory allocated for a given object
4145 * @objp: Pointer to the object
4147 * kmalloc may internally round up allocations and return more memory
4148 * than requested. ksize() can be used to determine the actual amount of
4149 * memory allocated. The caller may use this additional memory, even though
4150 * a smaller amount of memory was initially specified with the kmalloc call.
4151 * The caller must guarantee that objp points to a valid object previously
4152 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4153 * must not be freed during the duration of the call.
4155 unsigned int ksize(const void *objp
)
4157 if (unlikely(objp
== NULL
))
4160 return obj_size(virt_to_cache(objp
));