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)
213 * Manages the objs in a slab. Placed either at the beginning of mem allocated
214 * for a slab, or allocated from an general cache.
215 * Slabs are chained into three list: fully used, partial, fully free slabs.
218 struct list_head list
;
219 unsigned long colouroff
;
220 void *s_mem
; /* including colour offset */
221 unsigned int inuse
; /* num of objs active in slab */
223 unsigned short nodeid
;
229 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
230 * arrange for kmem_freepages to be called via RCU. This is useful if
231 * we need to approach a kernel structure obliquely, from its address
232 * obtained without the usual locking. We can lock the structure to
233 * stabilize it and check it's still at the given address, only if we
234 * can be sure that the memory has not been meanwhile reused for some
235 * other kind of object (which our subsystem's lock might corrupt).
237 * rcu_read_lock before reading the address, then rcu_read_unlock after
238 * taking the spinlock within the structure expected at that address.
240 * We assume struct slab_rcu can overlay struct slab when destroying.
243 struct rcu_head head
;
244 struct kmem_cache
*cachep
;
252 * - LIFO ordering, to hand out cache-warm objects from _alloc
253 * - reduce the number of linked list operations
254 * - reduce spinlock operations
256 * The limit is stored in the per-cpu structure to reduce the data cache
263 unsigned int batchcount
;
264 unsigned int touched
;
267 * Must have this definition in here for the proper
268 * alignment of array_cache. Also simplifies accessing
270 * [0] is for gcc 2.95. It should really be [].
275 * bootstrap: The caches do not work without cpuarrays anymore, but the
276 * cpuarrays are allocated from the generic caches...
278 #define BOOT_CPUCACHE_ENTRIES 1
279 struct arraycache_init
{
280 struct array_cache cache
;
281 void *entries
[BOOT_CPUCACHE_ENTRIES
];
285 * The slab lists for all objects.
288 struct list_head slabs_partial
; /* partial list first, better asm code */
289 struct list_head slabs_full
;
290 struct list_head slabs_free
;
291 unsigned long free_objects
;
292 unsigned int free_limit
;
293 unsigned int colour_next
; /* Per-node cache coloring */
294 spinlock_t list_lock
;
295 struct array_cache
*shared
; /* shared per node */
296 struct array_cache
**alien
; /* on other nodes */
297 unsigned long next_reap
; /* updated without locking */
298 int free_touched
; /* updated without locking */
302 * Need this for bootstrapping a per node allocator.
304 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
305 struct kmem_list3 __initdata initkmem_list3
[NUM_INIT_LISTS
];
306 #define CACHE_CACHE 0
308 #define SIZE_L3 (1 + MAX_NUMNODES)
311 * This function must be completely optimized away if a constant is passed to
312 * it. Mostly the same as what is in linux/slab.h except it returns an index.
314 static __always_inline
int index_of(const size_t size
)
316 extern void __bad_size(void);
318 if (__builtin_constant_p(size
)) {
326 #include "linux/kmalloc_sizes.h"
334 #define INDEX_AC index_of(sizeof(struct arraycache_init))
335 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
337 static void kmem_list3_init(struct kmem_list3
*parent
)
339 INIT_LIST_HEAD(&parent
->slabs_full
);
340 INIT_LIST_HEAD(&parent
->slabs_partial
);
341 INIT_LIST_HEAD(&parent
->slabs_free
);
342 parent
->shared
= NULL
;
343 parent
->alien
= NULL
;
344 parent
->colour_next
= 0;
345 spin_lock_init(&parent
->list_lock
);
346 parent
->free_objects
= 0;
347 parent
->free_touched
= 0;
350 #define MAKE_LIST(cachep, listp, slab, nodeid) \
352 INIT_LIST_HEAD(listp); \
353 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
356 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
358 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
359 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
360 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
370 /* 1) per-cpu data, touched during every alloc/free */
371 struct array_cache
*array
[NR_CPUS
];
372 /* 2) Cache tunables. Protected by cache_chain_mutex */
373 unsigned int batchcount
;
377 unsigned int buffer_size
;
378 /* 3) touched by every alloc & free from the backend */
379 struct kmem_list3
*nodelists
[MAX_NUMNODES
];
381 unsigned int flags
; /* constant flags */
382 unsigned int num
; /* # of objs per slab */
384 /* 4) cache_grow/shrink */
385 /* order of pgs per slab (2^n) */
386 unsigned int gfporder
;
388 /* force GFP flags, e.g. GFP_DMA */
391 size_t colour
; /* cache colouring range */
392 unsigned int colour_off
; /* colour offset */
393 struct kmem_cache
*slabp_cache
;
394 unsigned int slab_size
;
395 unsigned int dflags
; /* dynamic flags */
397 /* constructor func */
398 void (*ctor
) (void *, struct kmem_cache
*, unsigned long);
400 /* de-constructor func */
401 void (*dtor
) (void *, struct kmem_cache
*, unsigned long);
403 /* 5) cache creation/removal */
405 struct list_head next
;
409 unsigned long num_active
;
410 unsigned long num_allocations
;
411 unsigned long high_mark
;
413 unsigned long reaped
;
414 unsigned long errors
;
415 unsigned long max_freeable
;
416 unsigned long node_allocs
;
417 unsigned long node_frees
;
418 unsigned long node_overflow
;
426 * If debugging is enabled, then the allocator can add additional
427 * fields and/or padding to every object. buffer_size contains the total
428 * object size including these internal fields, the following two
429 * variables contain the offset to the user object and its size.
436 #define CFLGS_OFF_SLAB (0x80000000UL)
437 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
439 #define BATCHREFILL_LIMIT 16
441 * Optimization question: fewer reaps means less probability for unnessary
442 * cpucache drain/refill cycles.
444 * OTOH the cpuarrays can contain lots of objects,
445 * which could lock up otherwise freeable slabs.
447 #define REAPTIMEOUT_CPUC (2*HZ)
448 #define REAPTIMEOUT_LIST3 (4*HZ)
451 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
452 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
453 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
454 #define STATS_INC_GROWN(x) ((x)->grown++)
455 #define STATS_INC_REAPED(x) ((x)->reaped++)
456 #define STATS_SET_HIGH(x) \
458 if ((x)->num_active > (x)->high_mark) \
459 (x)->high_mark = (x)->num_active; \
461 #define STATS_INC_ERR(x) ((x)->errors++)
462 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
463 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
464 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
465 #define STATS_SET_FREEABLE(x, i) \
467 if ((x)->max_freeable < i) \
468 (x)->max_freeable = i; \
470 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
471 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
472 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
473 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
475 #define STATS_INC_ACTIVE(x) do { } while (0)
476 #define STATS_DEC_ACTIVE(x) do { } while (0)
477 #define STATS_INC_ALLOCED(x) do { } while (0)
478 #define STATS_INC_GROWN(x) do { } while (0)
479 #define STATS_INC_REAPED(x) do { } while (0)
480 #define STATS_SET_HIGH(x) do { } while (0)
481 #define STATS_INC_ERR(x) do { } while (0)
482 #define STATS_INC_NODEALLOCS(x) do { } while (0)
483 #define STATS_INC_NODEFREES(x) do { } while (0)
484 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
485 #define STATS_SET_FREEABLE(x, i) do { } while (0)
486 #define STATS_INC_ALLOCHIT(x) do { } while (0)
487 #define STATS_INC_ALLOCMISS(x) do { } while (0)
488 #define STATS_INC_FREEHIT(x) do { } while (0)
489 #define STATS_INC_FREEMISS(x) do { } while (0)
494 * Magic nums for obj red zoning.
495 * Placed in the first word before and the first word after an obj.
497 #define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
498 #define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
500 /* ...and for poisoning */
501 #define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
502 #define POISON_FREE 0x6b /* for use-after-free poisoning */
503 #define POISON_END 0xa5 /* end-byte of poisoning */
506 * memory layout of objects:
508 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
509 * the end of an object is aligned with the end of the real
510 * allocation. Catches writes behind the end of the allocation.
511 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
513 * cachep->obj_offset: The real object.
514 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
515 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
516 * [BYTES_PER_WORD long]
518 static int obj_offset(struct kmem_cache
*cachep
)
520 return cachep
->obj_offset
;
523 static int obj_size(struct kmem_cache
*cachep
)
525 return cachep
->obj_size
;
528 static unsigned long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
530 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
531 return (unsigned long*) (objp
+obj_offset(cachep
)-BYTES_PER_WORD
);
534 static unsigned long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
536 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
537 if (cachep
->flags
& SLAB_STORE_USER
)
538 return (unsigned long *)(objp
+ cachep
->buffer_size
-
540 return (unsigned long *)(objp
+ cachep
->buffer_size
- BYTES_PER_WORD
);
543 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
545 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
546 return (void **)(objp
+ cachep
->buffer_size
- BYTES_PER_WORD
);
551 #define obj_offset(x) 0
552 #define obj_size(cachep) (cachep->buffer_size)
553 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
554 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
555 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
560 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
563 #if defined(CONFIG_LARGE_ALLOCS)
564 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
565 #define MAX_GFP_ORDER 13 /* up to 32Mb */
566 #elif defined(CONFIG_MMU)
567 #define MAX_OBJ_ORDER 5 /* 32 pages */
568 #define MAX_GFP_ORDER 5 /* 32 pages */
570 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
571 #define MAX_GFP_ORDER 8 /* up to 1Mb */
575 * Do not go above this order unless 0 objects fit into the slab.
577 #define BREAK_GFP_ORDER_HI 1
578 #define BREAK_GFP_ORDER_LO 0
579 static int slab_break_gfp_order
= BREAK_GFP_ORDER_LO
;
582 * Functions for storing/retrieving the cachep and or slab from the page
583 * allocator. These are used to find the slab an obj belongs to. With kfree(),
584 * these are used to find the cache which an obj belongs to.
586 static inline void page_set_cache(struct page
*page
, struct kmem_cache
*cache
)
588 page
->lru
.next
= (struct list_head
*)cache
;
591 static inline struct kmem_cache
*page_get_cache(struct page
*page
)
593 if (unlikely(PageCompound(page
)))
594 page
= (struct page
*)page_private(page
);
595 return (struct kmem_cache
*)page
->lru
.next
;
598 static inline void page_set_slab(struct page
*page
, struct slab
*slab
)
600 page
->lru
.prev
= (struct list_head
*)slab
;
603 static inline struct slab
*page_get_slab(struct page
*page
)
605 if (unlikely(PageCompound(page
)))
606 page
= (struct page
*)page_private(page
);
607 return (struct slab
*)page
->lru
.prev
;
610 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
612 struct page
*page
= virt_to_page(obj
);
613 return page_get_cache(page
);
616 static inline struct slab
*virt_to_slab(const void *obj
)
618 struct page
*page
= virt_to_page(obj
);
619 return page_get_slab(page
);
622 static inline void *index_to_obj(struct kmem_cache
*cache
, struct slab
*slab
,
625 return slab
->s_mem
+ cache
->buffer_size
* idx
;
628 static inline unsigned int obj_to_index(struct kmem_cache
*cache
,
629 struct slab
*slab
, void *obj
)
631 return (unsigned)(obj
- slab
->s_mem
) / cache
->buffer_size
;
635 * These are the default caches for kmalloc. Custom caches can have other sizes.
637 struct cache_sizes malloc_sizes
[] = {
638 #define CACHE(x) { .cs_size = (x) },
639 #include <linux/kmalloc_sizes.h>
643 EXPORT_SYMBOL(malloc_sizes
);
645 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
651 static struct cache_names __initdata cache_names
[] = {
652 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
653 #include <linux/kmalloc_sizes.h>
658 static struct arraycache_init initarray_cache __initdata
=
659 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
660 static struct arraycache_init initarray_generic
=
661 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
663 /* internal cache of cache description objs */
664 static struct kmem_cache cache_cache
= {
666 .limit
= BOOT_CPUCACHE_ENTRIES
,
668 .buffer_size
= sizeof(struct kmem_cache
),
669 .name
= "kmem_cache",
671 .obj_size
= sizeof(struct kmem_cache
),
675 /* Guard access to the cache-chain. */
676 static DEFINE_MUTEX(cache_chain_mutex
);
677 static struct list_head cache_chain
;
680 * vm_enough_memory() looks at this to determine how many slab-allocated pages
681 * are possibly freeable under pressure
683 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
685 atomic_t slab_reclaim_pages
;
688 * chicken and egg problem: delay the per-cpu array allocation
689 * until the general caches are up.
699 * used by boot code to determine if it can use slab based allocator
701 int slab_is_available(void)
703 return g_cpucache_up
== FULL
;
706 static DEFINE_PER_CPU(struct work_struct
, reap_work
);
708 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
710 static void enable_cpucache(struct kmem_cache
*cachep
);
711 static void cache_reap(void *unused
);
712 static int __node_shrink(struct kmem_cache
*cachep
, int node
);
714 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
716 return cachep
->array
[smp_processor_id()];
719 static inline struct kmem_cache
*__find_general_cachep(size_t size
,
722 struct cache_sizes
*csizep
= malloc_sizes
;
725 /* This happens if someone tries to call
726 * kmem_cache_create(), or __kmalloc(), before
727 * the generic caches are initialized.
729 BUG_ON(malloc_sizes
[INDEX_AC
].cs_cachep
== NULL
);
731 while (size
> csizep
->cs_size
)
735 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
736 * has cs_{dma,}cachep==NULL. Thus no special case
737 * for large kmalloc calls required.
739 if (unlikely(gfpflags
& GFP_DMA
))
740 return csizep
->cs_dmacachep
;
741 return csizep
->cs_cachep
;
744 struct kmem_cache
*kmem_find_general_cachep(size_t size
, gfp_t gfpflags
)
746 return __find_general_cachep(size
, gfpflags
);
748 EXPORT_SYMBOL(kmem_find_general_cachep
);
750 static size_t slab_mgmt_size(size_t nr_objs
, size_t align
)
752 return ALIGN(sizeof(struct slab
)+nr_objs
*sizeof(kmem_bufctl_t
), align
);
756 * Calculate the number of objects and left-over bytes for a given buffer size.
758 static void cache_estimate(unsigned long gfporder
, size_t buffer_size
,
759 size_t align
, int flags
, size_t *left_over
,
764 size_t slab_size
= PAGE_SIZE
<< gfporder
;
767 * The slab management structure can be either off the slab or
768 * on it. For the latter case, the memory allocated for a
772 * - One kmem_bufctl_t for each object
773 * - Padding to respect alignment of @align
774 * - @buffer_size bytes for each object
776 * If the slab management structure is off the slab, then the
777 * alignment will already be calculated into the size. Because
778 * the slabs are all pages aligned, the objects will be at the
779 * correct alignment when allocated.
781 if (flags
& CFLGS_OFF_SLAB
) {
783 nr_objs
= slab_size
/ buffer_size
;
785 if (nr_objs
> SLAB_LIMIT
)
786 nr_objs
= SLAB_LIMIT
;
789 * Ignore padding for the initial guess. The padding
790 * is at most @align-1 bytes, and @buffer_size is at
791 * least @align. In the worst case, this result will
792 * be one greater than the number of objects that fit
793 * into the memory allocation when taking the padding
796 nr_objs
= (slab_size
- sizeof(struct slab
)) /
797 (buffer_size
+ sizeof(kmem_bufctl_t
));
800 * This calculated number will be either the right
801 * amount, or one greater than what we want.
803 if (slab_mgmt_size(nr_objs
, align
) + nr_objs
*buffer_size
807 if (nr_objs
> SLAB_LIMIT
)
808 nr_objs
= SLAB_LIMIT
;
810 mgmt_size
= slab_mgmt_size(nr_objs
, align
);
813 *left_over
= slab_size
- nr_objs
*buffer_size
- mgmt_size
;
816 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
818 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
821 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
822 function
, cachep
->name
, msg
);
828 * Special reaping functions for NUMA systems called from cache_reap().
829 * These take care of doing round robin flushing of alien caches (containing
830 * objects freed on different nodes from which they were allocated) and the
831 * flushing of remote pcps by calling drain_node_pages.
833 static DEFINE_PER_CPU(unsigned long, reap_node
);
835 static void init_reap_node(int cpu
)
839 node
= next_node(cpu_to_node(cpu
), node_online_map
);
840 if (node
== MAX_NUMNODES
)
841 node
= first_node(node_online_map
);
843 __get_cpu_var(reap_node
) = node
;
846 static void next_reap_node(void)
848 int node
= __get_cpu_var(reap_node
);
851 * Also drain per cpu pages on remote zones
853 if (node
!= numa_node_id())
854 drain_node_pages(node
);
856 node
= next_node(node
, node_online_map
);
857 if (unlikely(node
>= MAX_NUMNODES
))
858 node
= first_node(node_online_map
);
859 __get_cpu_var(reap_node
) = node
;
863 #define init_reap_node(cpu) do { } while (0)
864 #define next_reap_node(void) do { } while (0)
868 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
869 * via the workqueue/eventd.
870 * Add the CPU number into the expiration time to minimize the possibility of
871 * the CPUs getting into lockstep and contending for the global cache chain
874 static void __devinit
start_cpu_timer(int cpu
)
876 struct work_struct
*reap_work
= &per_cpu(reap_work
, cpu
);
879 * When this gets called from do_initcalls via cpucache_init(),
880 * init_workqueues() has already run, so keventd will be setup
883 if (keventd_up() && reap_work
->func
== NULL
) {
885 INIT_WORK(reap_work
, cache_reap
, NULL
);
886 schedule_delayed_work_on(cpu
, reap_work
, HZ
+ 3 * cpu
);
890 static struct array_cache
*alloc_arraycache(int node
, int entries
,
893 int memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
894 struct array_cache
*nc
= NULL
;
896 nc
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
900 nc
->batchcount
= batchcount
;
902 spin_lock_init(&nc
->lock
);
908 * Transfer objects in one arraycache to another.
909 * Locking must be handled by the caller.
911 * Return the number of entries transferred.
913 static int transfer_objects(struct array_cache
*to
,
914 struct array_cache
*from
, unsigned int max
)
916 /* Figure out how many entries to transfer */
917 int nr
= min(min(from
->avail
, max
), to
->limit
- to
->avail
);
922 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
932 static void *__cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
933 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
935 static struct array_cache
**alloc_alien_cache(int node
, int limit
)
937 struct array_cache
**ac_ptr
;
938 int memsize
= sizeof(void *) * MAX_NUMNODES
;
943 ac_ptr
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
946 if (i
== node
|| !node_online(i
)) {
950 ac_ptr
[i
] = alloc_arraycache(node
, limit
, 0xbaadf00d);
952 for (i
--; i
<= 0; i
--)
962 static void free_alien_cache(struct array_cache
**ac_ptr
)
973 static void __drain_alien_cache(struct kmem_cache
*cachep
,
974 struct array_cache
*ac
, int node
)
976 struct kmem_list3
*rl3
= cachep
->nodelists
[node
];
979 spin_lock(&rl3
->list_lock
);
981 * Stuff objects into the remote nodes shared array first.
982 * That way we could avoid the overhead of putting the objects
983 * into the free lists and getting them back later.
986 transfer_objects(rl3
->shared
, ac
, ac
->limit
);
988 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
990 spin_unlock(&rl3
->list_lock
);
995 * Called from cache_reap() to regularly drain alien caches round robin.
997 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_list3
*l3
)
999 int node
= __get_cpu_var(reap_node
);
1002 struct array_cache
*ac
= l3
->alien
[node
];
1004 if (ac
&& ac
->avail
&& spin_trylock_irq(&ac
->lock
)) {
1005 __drain_alien_cache(cachep
, ac
, node
);
1006 spin_unlock_irq(&ac
->lock
);
1011 static void drain_alien_cache(struct kmem_cache
*cachep
,
1012 struct array_cache
**alien
)
1015 struct array_cache
*ac
;
1016 unsigned long flags
;
1018 for_each_online_node(i
) {
1021 spin_lock_irqsave(&ac
->lock
, flags
);
1022 __drain_alien_cache(cachep
, ac
, i
);
1023 spin_unlock_irqrestore(&ac
->lock
, flags
);
1028 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1030 struct slab
*slabp
= virt_to_slab(objp
);
1031 int nodeid
= slabp
->nodeid
;
1032 struct kmem_list3
*l3
;
1033 struct array_cache
*alien
= NULL
;
1036 * Make sure we are not freeing a object from another node to the array
1037 * cache on this cpu.
1039 if (likely(slabp
->nodeid
== numa_node_id()))
1042 l3
= cachep
->nodelists
[numa_node_id()];
1043 STATS_INC_NODEFREES(cachep
);
1044 if (l3
->alien
&& l3
->alien
[nodeid
]) {
1045 alien
= l3
->alien
[nodeid
];
1046 spin_lock(&alien
->lock
);
1047 if (unlikely(alien
->avail
== alien
->limit
)) {
1048 STATS_INC_ACOVERFLOW(cachep
);
1049 __drain_alien_cache(cachep
, alien
, nodeid
);
1051 alien
->entry
[alien
->avail
++] = objp
;
1052 spin_unlock(&alien
->lock
);
1054 spin_lock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1055 free_block(cachep
, &objp
, 1, nodeid
);
1056 spin_unlock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1063 #define drain_alien_cache(cachep, alien) do { } while (0)
1064 #define reap_alien(cachep, l3) do { } while (0)
1066 static inline struct array_cache
**alloc_alien_cache(int node
, int limit
)
1068 return (struct array_cache
**) 0x01020304ul
;
1071 static inline void free_alien_cache(struct array_cache
**ac_ptr
)
1075 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1082 static int cpuup_callback(struct notifier_block
*nfb
,
1083 unsigned long action
, void *hcpu
)
1085 long cpu
= (long)hcpu
;
1086 struct kmem_cache
*cachep
;
1087 struct kmem_list3
*l3
= NULL
;
1088 int node
= cpu_to_node(cpu
);
1089 int memsize
= sizeof(struct kmem_list3
);
1092 case CPU_UP_PREPARE
:
1093 mutex_lock(&cache_chain_mutex
);
1095 * We need to do this right in the beginning since
1096 * alloc_arraycache's are going to use this list.
1097 * kmalloc_node allows us to add the slab to the right
1098 * kmem_list3 and not this cpu's kmem_list3
1101 list_for_each_entry(cachep
, &cache_chain
, next
) {
1103 * Set up the size64 kmemlist for cpu before we can
1104 * begin anything. Make sure some other cpu on this
1105 * node has not already allocated this
1107 if (!cachep
->nodelists
[node
]) {
1108 l3
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1111 kmem_list3_init(l3
);
1112 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
1113 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1116 * The l3s don't come and go as CPUs come and
1117 * go. cache_chain_mutex is sufficient
1120 cachep
->nodelists
[node
] = l3
;
1123 spin_lock_irq(&cachep
->nodelists
[node
]->list_lock
);
1124 cachep
->nodelists
[node
]->free_limit
=
1125 (1 + nr_cpus_node(node
)) *
1126 cachep
->batchcount
+ cachep
->num
;
1127 spin_unlock_irq(&cachep
->nodelists
[node
]->list_lock
);
1131 * Now we can go ahead with allocating the shared arrays and
1134 list_for_each_entry(cachep
, &cache_chain
, next
) {
1135 struct array_cache
*nc
;
1136 struct array_cache
*shared
;
1137 struct array_cache
**alien
;
1139 nc
= alloc_arraycache(node
, cachep
->limit
,
1140 cachep
->batchcount
);
1143 shared
= alloc_arraycache(node
,
1144 cachep
->shared
* cachep
->batchcount
,
1149 alien
= alloc_alien_cache(node
, cachep
->limit
);
1152 cachep
->array
[cpu
] = nc
;
1153 l3
= cachep
->nodelists
[node
];
1156 spin_lock_irq(&l3
->list_lock
);
1159 * We are serialised from CPU_DEAD or
1160 * CPU_UP_CANCELLED by the cpucontrol lock
1162 l3
->shared
= shared
;
1171 spin_unlock_irq(&l3
->list_lock
);
1173 free_alien_cache(alien
);
1175 mutex_unlock(&cache_chain_mutex
);
1178 start_cpu_timer(cpu
);
1180 #ifdef CONFIG_HOTPLUG_CPU
1183 * Even if all the cpus of a node are down, we don't free the
1184 * kmem_list3 of any cache. This to avoid a race between
1185 * cpu_down, and a kmalloc allocation from another cpu for
1186 * memory from the node of the cpu going down. The list3
1187 * structure is usually allocated from kmem_cache_create() and
1188 * gets destroyed at kmem_cache_destroy().
1191 case CPU_UP_CANCELED
:
1192 mutex_lock(&cache_chain_mutex
);
1193 list_for_each_entry(cachep
, &cache_chain
, next
) {
1194 struct array_cache
*nc
;
1195 struct array_cache
*shared
;
1196 struct array_cache
**alien
;
1199 mask
= node_to_cpumask(node
);
1200 /* cpu is dead; no one can alloc from it. */
1201 nc
= cachep
->array
[cpu
];
1202 cachep
->array
[cpu
] = NULL
;
1203 l3
= cachep
->nodelists
[node
];
1206 goto free_array_cache
;
1208 spin_lock_irq(&l3
->list_lock
);
1210 /* Free limit for this kmem_list3 */
1211 l3
->free_limit
-= cachep
->batchcount
;
1213 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
1215 if (!cpus_empty(mask
)) {
1216 spin_unlock_irq(&l3
->list_lock
);
1217 goto free_array_cache
;
1220 shared
= l3
->shared
;
1222 free_block(cachep
, l3
->shared
->entry
,
1223 l3
->shared
->avail
, node
);
1230 spin_unlock_irq(&l3
->list_lock
);
1234 drain_alien_cache(cachep
, alien
);
1235 free_alien_cache(alien
);
1241 * In the previous loop, all the objects were freed to
1242 * the respective cache's slabs, now we can go ahead and
1243 * shrink each nodelist to its limit.
1245 list_for_each_entry(cachep
, &cache_chain
, next
) {
1246 l3
= cachep
->nodelists
[node
];
1249 spin_lock_irq(&l3
->list_lock
);
1250 /* free slabs belonging to this node */
1251 __node_shrink(cachep
, node
);
1252 spin_unlock_irq(&l3
->list_lock
);
1254 mutex_unlock(&cache_chain_mutex
);
1260 mutex_unlock(&cache_chain_mutex
);
1264 static struct notifier_block cpucache_notifier
= { &cpuup_callback
, NULL
, 0 };
1267 * swap the static kmem_list3 with kmalloced memory
1269 static void init_list(struct kmem_cache
*cachep
, struct kmem_list3
*list
,
1272 struct kmem_list3
*ptr
;
1274 BUG_ON(cachep
->nodelists
[nodeid
] != list
);
1275 ptr
= kmalloc_node(sizeof(struct kmem_list3
), GFP_KERNEL
, nodeid
);
1278 local_irq_disable();
1279 memcpy(ptr
, list
, sizeof(struct kmem_list3
));
1280 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1281 cachep
->nodelists
[nodeid
] = ptr
;
1286 * Initialisation. Called after the page allocator have been initialised and
1287 * before smp_init().
1289 void __init
kmem_cache_init(void)
1292 struct cache_sizes
*sizes
;
1293 struct cache_names
*names
;
1297 for (i
= 0; i
< NUM_INIT_LISTS
; i
++) {
1298 kmem_list3_init(&initkmem_list3
[i
]);
1299 if (i
< MAX_NUMNODES
)
1300 cache_cache
.nodelists
[i
] = NULL
;
1304 * Fragmentation resistance on low memory - only use bigger
1305 * page orders on machines with more than 32MB of memory.
1307 if (num_physpages
> (32 << 20) >> PAGE_SHIFT
)
1308 slab_break_gfp_order
= BREAK_GFP_ORDER_HI
;
1310 /* Bootstrap is tricky, because several objects are allocated
1311 * from caches that do not exist yet:
1312 * 1) initialize the cache_cache cache: it contains the struct
1313 * kmem_cache structures of all caches, except cache_cache itself:
1314 * cache_cache is statically allocated.
1315 * Initially an __init data area is used for the head array and the
1316 * kmem_list3 structures, it's replaced with a kmalloc allocated
1317 * array at the end of the bootstrap.
1318 * 2) Create the first kmalloc cache.
1319 * The struct kmem_cache for the new cache is allocated normally.
1320 * An __init data area is used for the head array.
1321 * 3) Create the remaining kmalloc caches, with minimally sized
1323 * 4) Replace the __init data head arrays for cache_cache and the first
1324 * kmalloc cache with kmalloc allocated arrays.
1325 * 5) Replace the __init data for kmem_list3 for cache_cache and
1326 * the other cache's with kmalloc allocated memory.
1327 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1330 /* 1) create the cache_cache */
1331 INIT_LIST_HEAD(&cache_chain
);
1332 list_add(&cache_cache
.next
, &cache_chain
);
1333 cache_cache
.colour_off
= cache_line_size();
1334 cache_cache
.array
[smp_processor_id()] = &initarray_cache
.cache
;
1335 cache_cache
.nodelists
[numa_node_id()] = &initkmem_list3
[CACHE_CACHE
];
1337 cache_cache
.buffer_size
= ALIGN(cache_cache
.buffer_size
,
1340 for (order
= 0; order
< MAX_ORDER
; order
++) {
1341 cache_estimate(order
, cache_cache
.buffer_size
,
1342 cache_line_size(), 0, &left_over
, &cache_cache
.num
);
1343 if (cache_cache
.num
)
1346 BUG_ON(!cache_cache
.num
);
1347 cache_cache
.gfporder
= order
;
1348 cache_cache
.colour
= left_over
/ cache_cache
.colour_off
;
1349 cache_cache
.slab_size
= ALIGN(cache_cache
.num
* sizeof(kmem_bufctl_t
) +
1350 sizeof(struct slab
), cache_line_size());
1352 /* 2+3) create the kmalloc caches */
1353 sizes
= malloc_sizes
;
1354 names
= cache_names
;
1357 * Initialize the caches that provide memory for the array cache and the
1358 * kmem_list3 structures first. Without this, further allocations will
1362 sizes
[INDEX_AC
].cs_cachep
= kmem_cache_create(names
[INDEX_AC
].name
,
1363 sizes
[INDEX_AC
].cs_size
,
1364 ARCH_KMALLOC_MINALIGN
,
1365 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1368 if (INDEX_AC
!= INDEX_L3
) {
1369 sizes
[INDEX_L3
].cs_cachep
=
1370 kmem_cache_create(names
[INDEX_L3
].name
,
1371 sizes
[INDEX_L3
].cs_size
,
1372 ARCH_KMALLOC_MINALIGN
,
1373 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1377 while (sizes
->cs_size
!= ULONG_MAX
) {
1379 * For performance, all the general caches are L1 aligned.
1380 * This should be particularly beneficial on SMP boxes, as it
1381 * eliminates "false sharing".
1382 * Note for systems short on memory removing the alignment will
1383 * allow tighter packing of the smaller caches.
1385 if (!sizes
->cs_cachep
) {
1386 sizes
->cs_cachep
= kmem_cache_create(names
->name
,
1388 ARCH_KMALLOC_MINALIGN
,
1389 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1393 sizes
->cs_dmacachep
= kmem_cache_create(names
->name_dma
,
1395 ARCH_KMALLOC_MINALIGN
,
1396 ARCH_KMALLOC_FLAGS
|SLAB_CACHE_DMA
|
1402 /* 4) Replace the bootstrap head arrays */
1406 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1408 local_irq_disable();
1409 BUG_ON(cpu_cache_get(&cache_cache
) != &initarray_cache
.cache
);
1410 memcpy(ptr
, cpu_cache_get(&cache_cache
),
1411 sizeof(struct arraycache_init
));
1412 cache_cache
.array
[smp_processor_id()] = ptr
;
1415 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1417 local_irq_disable();
1418 BUG_ON(cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
)
1419 != &initarray_generic
.cache
);
1420 memcpy(ptr
, cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
),
1421 sizeof(struct arraycache_init
));
1422 malloc_sizes
[INDEX_AC
].cs_cachep
->array
[smp_processor_id()] =
1426 /* 5) Replace the bootstrap kmem_list3's */
1429 /* Replace the static kmem_list3 structures for the boot cpu */
1430 init_list(&cache_cache
, &initkmem_list3
[CACHE_CACHE
],
1433 for_each_online_node(node
) {
1434 init_list(malloc_sizes
[INDEX_AC
].cs_cachep
,
1435 &initkmem_list3
[SIZE_AC
+ node
], node
);
1437 if (INDEX_AC
!= INDEX_L3
) {
1438 init_list(malloc_sizes
[INDEX_L3
].cs_cachep
,
1439 &initkmem_list3
[SIZE_L3
+ node
],
1445 /* 6) resize the head arrays to their final sizes */
1447 struct kmem_cache
*cachep
;
1448 mutex_lock(&cache_chain_mutex
);
1449 list_for_each_entry(cachep
, &cache_chain
, next
)
1450 enable_cpucache(cachep
);
1451 mutex_unlock(&cache_chain_mutex
);
1455 g_cpucache_up
= FULL
;
1458 * Register a cpu startup notifier callback that initializes
1459 * cpu_cache_get for all new cpus
1461 register_cpu_notifier(&cpucache_notifier
);
1464 * The reap timers are started later, with a module init call: That part
1465 * of the kernel is not yet operational.
1469 static int __init
cpucache_init(void)
1474 * Register the timers that return unneeded pages to the page allocator
1476 for_each_online_cpu(cpu
)
1477 start_cpu_timer(cpu
);
1480 __initcall(cpucache_init
);
1483 * Interface to system's page allocator. No need to hold the cache-lock.
1485 * If we requested dmaable memory, we will get it. Even if we
1486 * did not request dmaable memory, we might get it, but that
1487 * would be relatively rare and ignorable.
1489 static void *kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
1497 * Nommu uses slab's for process anonymous memory allocations, and thus
1498 * requires __GFP_COMP to properly refcount higher order allocations
1500 flags
|= __GFP_COMP
;
1502 flags
|= cachep
->gfpflags
;
1504 page
= alloc_pages_node(nodeid
, flags
, cachep
->gfporder
);
1508 nr_pages
= (1 << cachep
->gfporder
);
1509 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1510 atomic_add(nr_pages
, &slab_reclaim_pages
);
1511 add_page_state(nr_slab
, nr_pages
);
1512 for (i
= 0; i
< nr_pages
; i
++)
1513 __SetPageSlab(page
+ i
);
1514 return page_address(page
);
1518 * Interface to system's page release.
1520 static void kmem_freepages(struct kmem_cache
*cachep
, void *addr
)
1522 unsigned long i
= (1 << cachep
->gfporder
);
1523 struct page
*page
= virt_to_page(addr
);
1524 const unsigned long nr_freed
= i
;
1527 BUG_ON(!PageSlab(page
));
1528 __ClearPageSlab(page
);
1531 sub_page_state(nr_slab
, nr_freed
);
1532 if (current
->reclaim_state
)
1533 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1534 free_pages((unsigned long)addr
, cachep
->gfporder
);
1535 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1536 atomic_sub(1 << cachep
->gfporder
, &slab_reclaim_pages
);
1539 static void kmem_rcu_free(struct rcu_head
*head
)
1541 struct slab_rcu
*slab_rcu
= (struct slab_rcu
*)head
;
1542 struct kmem_cache
*cachep
= slab_rcu
->cachep
;
1544 kmem_freepages(cachep
, slab_rcu
->addr
);
1545 if (OFF_SLAB(cachep
))
1546 kmem_cache_free(cachep
->slabp_cache
, slab_rcu
);
1551 #ifdef CONFIG_DEBUG_PAGEALLOC
1552 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1553 unsigned long caller
)
1555 int size
= obj_size(cachep
);
1557 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1559 if (size
< 5 * sizeof(unsigned long))
1562 *addr
++ = 0x12345678;
1564 *addr
++ = smp_processor_id();
1565 size
-= 3 * sizeof(unsigned long);
1567 unsigned long *sptr
= &caller
;
1568 unsigned long svalue
;
1570 while (!kstack_end(sptr
)) {
1572 if (kernel_text_address(svalue
)) {
1574 size
-= sizeof(unsigned long);
1575 if (size
<= sizeof(unsigned long))
1581 *addr
++ = 0x87654321;
1585 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1587 int size
= obj_size(cachep
);
1588 addr
= &((char *)addr
)[obj_offset(cachep
)];
1590 memset(addr
, val
, size
);
1591 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1594 static void dump_line(char *data
, int offset
, int limit
)
1597 printk(KERN_ERR
"%03x:", offset
);
1598 for (i
= 0; i
< limit
; i
++)
1599 printk(" %02x", (unsigned char)data
[offset
+ i
]);
1606 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1611 if (cachep
->flags
& SLAB_RED_ZONE
) {
1612 printk(KERN_ERR
"Redzone: 0x%lx/0x%lx.\n",
1613 *dbg_redzone1(cachep
, objp
),
1614 *dbg_redzone2(cachep
, objp
));
1617 if (cachep
->flags
& SLAB_STORE_USER
) {
1618 printk(KERN_ERR
"Last user: [<%p>]",
1619 *dbg_userword(cachep
, objp
));
1620 print_symbol("(%s)",
1621 (unsigned long)*dbg_userword(cachep
, objp
));
1624 realobj
= (char *)objp
+ obj_offset(cachep
);
1625 size
= obj_size(cachep
);
1626 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1629 if (i
+ limit
> size
)
1631 dump_line(realobj
, i
, limit
);
1635 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1641 realobj
= (char *)objp
+ obj_offset(cachep
);
1642 size
= obj_size(cachep
);
1644 for (i
= 0; i
< size
; i
++) {
1645 char exp
= POISON_FREE
;
1648 if (realobj
[i
] != exp
) {
1654 "Slab corruption: start=%p, len=%d\n",
1656 print_objinfo(cachep
, objp
, 0);
1658 /* Hexdump the affected line */
1661 if (i
+ limit
> size
)
1663 dump_line(realobj
, i
, limit
);
1666 /* Limit to 5 lines */
1672 /* Print some data about the neighboring objects, if they
1675 struct slab
*slabp
= virt_to_slab(objp
);
1678 objnr
= obj_to_index(cachep
, slabp
, objp
);
1680 objp
= index_to_obj(cachep
, slabp
, objnr
- 1);
1681 realobj
= (char *)objp
+ obj_offset(cachep
);
1682 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
1684 print_objinfo(cachep
, objp
, 2);
1686 if (objnr
+ 1 < cachep
->num
) {
1687 objp
= index_to_obj(cachep
, slabp
, objnr
+ 1);
1688 realobj
= (char *)objp
+ obj_offset(cachep
);
1689 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
1691 print_objinfo(cachep
, objp
, 2);
1699 * slab_destroy_objs - destroy a slab and its objects
1700 * @cachep: cache pointer being destroyed
1701 * @slabp: slab pointer being destroyed
1703 * Call the registered destructor for each object in a slab that is being
1706 static void slab_destroy_objs(struct kmem_cache
*cachep
, struct slab
*slabp
)
1709 for (i
= 0; i
< cachep
->num
; i
++) {
1710 void *objp
= index_to_obj(cachep
, slabp
, i
);
1712 if (cachep
->flags
& SLAB_POISON
) {
1713 #ifdef CONFIG_DEBUG_PAGEALLOC
1714 if (cachep
->buffer_size
% PAGE_SIZE
== 0 &&
1716 kernel_map_pages(virt_to_page(objp
),
1717 cachep
->buffer_size
/ PAGE_SIZE
, 1);
1719 check_poison_obj(cachep
, objp
);
1721 check_poison_obj(cachep
, objp
);
1724 if (cachep
->flags
& SLAB_RED_ZONE
) {
1725 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1726 slab_error(cachep
, "start of a freed object "
1728 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1729 slab_error(cachep
, "end of a freed object "
1732 if (cachep
->dtor
&& !(cachep
->flags
& SLAB_POISON
))
1733 (cachep
->dtor
) (objp
+ obj_offset(cachep
), cachep
, 0);
1737 static void slab_destroy_objs(struct kmem_cache
*cachep
, struct slab
*slabp
)
1741 for (i
= 0; i
< cachep
->num
; i
++) {
1742 void *objp
= index_to_obj(cachep
, slabp
, i
);
1743 (cachep
->dtor
) (objp
, cachep
, 0);
1750 * slab_destroy - destroy and release all objects in a slab
1751 * @cachep: cache pointer being destroyed
1752 * @slabp: slab pointer being destroyed
1754 * Destroy all the objs in a slab, and release the mem back to the system.
1755 * Before calling the slab must have been unlinked from the cache. The
1756 * cache-lock is not held/needed.
1758 static void slab_destroy(struct kmem_cache
*cachep
, struct slab
*slabp
)
1760 void *addr
= slabp
->s_mem
- slabp
->colouroff
;
1762 slab_destroy_objs(cachep
, slabp
);
1763 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
1764 struct slab_rcu
*slab_rcu
;
1766 slab_rcu
= (struct slab_rcu
*)slabp
;
1767 slab_rcu
->cachep
= cachep
;
1768 slab_rcu
->addr
= addr
;
1769 call_rcu(&slab_rcu
->head
, kmem_rcu_free
);
1771 kmem_freepages(cachep
, addr
);
1772 if (OFF_SLAB(cachep
))
1773 kmem_cache_free(cachep
->slabp_cache
, slabp
);
1778 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1779 * size of kmem_list3.
1781 static void set_up_list3s(struct kmem_cache
*cachep
, int index
)
1785 for_each_online_node(node
) {
1786 cachep
->nodelists
[node
] = &initkmem_list3
[index
+ node
];
1787 cachep
->nodelists
[node
]->next_reap
= jiffies
+
1789 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1794 * calculate_slab_order - calculate size (page order) of slabs
1795 * @cachep: pointer to the cache that is being created
1796 * @size: size of objects to be created in this cache.
1797 * @align: required alignment for the objects.
1798 * @flags: slab allocation flags
1800 * Also calculates the number of objects per slab.
1802 * This could be made much more intelligent. For now, try to avoid using
1803 * high order pages for slabs. When the gfp() functions are more friendly
1804 * towards high-order requests, this should be changed.
1806 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
1807 size_t size
, size_t align
, unsigned long flags
)
1809 unsigned long offslab_limit
;
1810 size_t left_over
= 0;
1813 for (gfporder
= 0; gfporder
<= MAX_GFP_ORDER
; gfporder
++) {
1817 cache_estimate(gfporder
, size
, align
, flags
, &remainder
, &num
);
1821 if (flags
& CFLGS_OFF_SLAB
) {
1823 * Max number of objs-per-slab for caches which
1824 * use off-slab slabs. Needed to avoid a possible
1825 * looping condition in cache_grow().
1827 offslab_limit
= size
- sizeof(struct slab
);
1828 offslab_limit
/= sizeof(kmem_bufctl_t
);
1830 if (num
> offslab_limit
)
1834 /* Found something acceptable - save it away */
1836 cachep
->gfporder
= gfporder
;
1837 left_over
= remainder
;
1840 * A VFS-reclaimable slab tends to have most allocations
1841 * as GFP_NOFS and we really don't want to have to be allocating
1842 * higher-order pages when we are unable to shrink dcache.
1844 if (flags
& SLAB_RECLAIM_ACCOUNT
)
1848 * Large number of objects is good, but very large slabs are
1849 * currently bad for the gfp()s.
1851 if (gfporder
>= slab_break_gfp_order
)
1855 * Acceptable internal fragmentation?
1857 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
1863 static void setup_cpu_cache(struct kmem_cache
*cachep
)
1865 if (g_cpucache_up
== FULL
) {
1866 enable_cpucache(cachep
);
1869 if (g_cpucache_up
== NONE
) {
1871 * Note: the first kmem_cache_create must create the cache
1872 * that's used by kmalloc(24), otherwise the creation of
1873 * further caches will BUG().
1875 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
1878 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
1879 * the first cache, then we need to set up all its list3s,
1880 * otherwise the creation of further caches will BUG().
1882 set_up_list3s(cachep
, SIZE_AC
);
1883 if (INDEX_AC
== INDEX_L3
)
1884 g_cpucache_up
= PARTIAL_L3
;
1886 g_cpucache_up
= PARTIAL_AC
;
1888 cachep
->array
[smp_processor_id()] =
1889 kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1891 if (g_cpucache_up
== PARTIAL_AC
) {
1892 set_up_list3s(cachep
, SIZE_L3
);
1893 g_cpucache_up
= PARTIAL_L3
;
1896 for_each_online_node(node
) {
1897 cachep
->nodelists
[node
] =
1898 kmalloc_node(sizeof(struct kmem_list3
),
1900 BUG_ON(!cachep
->nodelists
[node
]);
1901 kmem_list3_init(cachep
->nodelists
[node
]);
1905 cachep
->nodelists
[numa_node_id()]->next_reap
=
1906 jiffies
+ REAPTIMEOUT_LIST3
+
1907 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1909 cpu_cache_get(cachep
)->avail
= 0;
1910 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
1911 cpu_cache_get(cachep
)->batchcount
= 1;
1912 cpu_cache_get(cachep
)->touched
= 0;
1913 cachep
->batchcount
= 1;
1914 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
1918 * kmem_cache_create - Create a cache.
1919 * @name: A string which is used in /proc/slabinfo to identify this cache.
1920 * @size: The size of objects to be created in this cache.
1921 * @align: The required alignment for the objects.
1922 * @flags: SLAB flags
1923 * @ctor: A constructor for the objects.
1924 * @dtor: A destructor for the objects.
1926 * Returns a ptr to the cache on success, NULL on failure.
1927 * Cannot be called within a int, but can be interrupted.
1928 * The @ctor is run when new pages are allocated by the cache
1929 * and the @dtor is run before the pages are handed back.
1931 * @name must be valid until the cache is destroyed. This implies that
1932 * the module calling this has to destroy the cache before getting unloaded.
1936 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1937 * to catch references to uninitialised memory.
1939 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1940 * for buffer overruns.
1942 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1943 * cacheline. This can be beneficial if you're counting cycles as closely
1947 kmem_cache_create (const char *name
, size_t size
, size_t align
,
1948 unsigned long flags
,
1949 void (*ctor
)(void*, struct kmem_cache
*, unsigned long),
1950 void (*dtor
)(void*, struct kmem_cache
*, unsigned long))
1952 size_t left_over
, slab_size
, ralign
;
1953 struct kmem_cache
*cachep
= NULL
;
1954 struct list_head
*p
;
1957 * Sanity checks... these are all serious usage bugs.
1959 if (!name
|| in_interrupt() || (size
< BYTES_PER_WORD
) ||
1960 (size
> (1 << MAX_OBJ_ORDER
) * PAGE_SIZE
) || (dtor
&& !ctor
)) {
1961 printk(KERN_ERR
"%s: Early error in slab %s\n", __FUNCTION__
,
1967 * Prevent CPUs from coming and going.
1968 * lock_cpu_hotplug() nests outside cache_chain_mutex
1972 mutex_lock(&cache_chain_mutex
);
1974 list_for_each(p
, &cache_chain
) {
1975 struct kmem_cache
*pc
= list_entry(p
, struct kmem_cache
, next
);
1976 mm_segment_t old_fs
= get_fs();
1981 * This happens when the module gets unloaded and doesn't
1982 * destroy its slab cache and no-one else reuses the vmalloc
1983 * area of the module. Print a warning.
1986 res
= __get_user(tmp
, pc
->name
);
1989 printk("SLAB: cache with size %d has lost its name\n",
1994 if (!strcmp(pc
->name
, name
)) {
1995 printk("kmem_cache_create: duplicate cache %s\n", name
);
2002 WARN_ON(strchr(name
, ' ')); /* It confuses parsers */
2003 if ((flags
& SLAB_DEBUG_INITIAL
) && !ctor
) {
2004 /* No constructor, but inital state check requested */
2005 printk(KERN_ERR
"%s: No con, but init state check "
2006 "requested - %s\n", __FUNCTION__
, name
);
2007 flags
&= ~SLAB_DEBUG_INITIAL
;
2011 * Enable redzoning and last user accounting, except for caches with
2012 * large objects, if the increased size would increase the object size
2013 * above the next power of two: caches with object sizes just above a
2014 * power of two have a significant amount of internal fragmentation.
2016 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + 3 * BYTES_PER_WORD
))
2017 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2018 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2019 flags
|= SLAB_POISON
;
2021 if (flags
& SLAB_DESTROY_BY_RCU
)
2022 BUG_ON(flags
& SLAB_POISON
);
2024 if (flags
& SLAB_DESTROY_BY_RCU
)
2028 * Always checks flags, a caller might be expecting debug support which
2031 BUG_ON(flags
& ~CREATE_MASK
);
2034 * Check that size is in terms of words. This is needed to avoid
2035 * unaligned accesses for some archs when redzoning is used, and makes
2036 * sure any on-slab bufctl's are also correctly aligned.
2038 if (size
& (BYTES_PER_WORD
- 1)) {
2039 size
+= (BYTES_PER_WORD
- 1);
2040 size
&= ~(BYTES_PER_WORD
- 1);
2043 /* calculate the final buffer alignment: */
2045 /* 1) arch recommendation: can be overridden for debug */
2046 if (flags
& SLAB_HWCACHE_ALIGN
) {
2048 * Default alignment: as specified by the arch code. Except if
2049 * an object is really small, then squeeze multiple objects into
2052 ralign
= cache_line_size();
2053 while (size
<= ralign
/ 2)
2056 ralign
= BYTES_PER_WORD
;
2058 /* 2) arch mandated alignment: disables debug if necessary */
2059 if (ralign
< ARCH_SLAB_MINALIGN
) {
2060 ralign
= ARCH_SLAB_MINALIGN
;
2061 if (ralign
> BYTES_PER_WORD
)
2062 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2064 /* 3) caller mandated alignment: disables debug if necessary */
2065 if (ralign
< align
) {
2067 if (ralign
> BYTES_PER_WORD
)
2068 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2071 * 4) Store it. Note that the debug code below can reduce
2072 * the alignment to BYTES_PER_WORD.
2076 /* Get cache's description obj. */
2077 cachep
= kmem_cache_zalloc(&cache_cache
, SLAB_KERNEL
);
2082 cachep
->obj_size
= size
;
2084 if (flags
& SLAB_RED_ZONE
) {
2085 /* redzoning only works with word aligned caches */
2086 align
= BYTES_PER_WORD
;
2088 /* add space for red zone words */
2089 cachep
->obj_offset
+= BYTES_PER_WORD
;
2090 size
+= 2 * BYTES_PER_WORD
;
2092 if (flags
& SLAB_STORE_USER
) {
2093 /* user store requires word alignment and
2094 * one word storage behind the end of the real
2097 align
= BYTES_PER_WORD
;
2098 size
+= BYTES_PER_WORD
;
2100 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2101 if (size
>= malloc_sizes
[INDEX_L3
+ 1].cs_size
2102 && cachep
->obj_size
> cache_line_size() && size
< PAGE_SIZE
) {
2103 cachep
->obj_offset
+= PAGE_SIZE
- size
;
2109 /* Determine if the slab management is 'on' or 'off' slab. */
2110 if (size
>= (PAGE_SIZE
>> 3))
2112 * Size is large, assume best to place the slab management obj
2113 * off-slab (should allow better packing of objs).
2115 flags
|= CFLGS_OFF_SLAB
;
2117 size
= ALIGN(size
, align
);
2119 left_over
= calculate_slab_order(cachep
, size
, align
, flags
);
2122 printk("kmem_cache_create: couldn't create cache %s.\n", name
);
2123 kmem_cache_free(&cache_cache
, cachep
);
2127 slab_size
= ALIGN(cachep
->num
* sizeof(kmem_bufctl_t
)
2128 + sizeof(struct slab
), align
);
2131 * If the slab has been placed off-slab, and we have enough space then
2132 * move it on-slab. This is at the expense of any extra colouring.
2134 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= slab_size
) {
2135 flags
&= ~CFLGS_OFF_SLAB
;
2136 left_over
-= slab_size
;
2139 if (flags
& CFLGS_OFF_SLAB
) {
2140 /* really off slab. No need for manual alignment */
2142 cachep
->num
* sizeof(kmem_bufctl_t
) + sizeof(struct slab
);
2145 cachep
->colour_off
= cache_line_size();
2146 /* Offset must be a multiple of the alignment. */
2147 if (cachep
->colour_off
< align
)
2148 cachep
->colour_off
= align
;
2149 cachep
->colour
= left_over
/ cachep
->colour_off
;
2150 cachep
->slab_size
= slab_size
;
2151 cachep
->flags
= flags
;
2152 cachep
->gfpflags
= 0;
2153 if (flags
& SLAB_CACHE_DMA
)
2154 cachep
->gfpflags
|= GFP_DMA
;
2155 cachep
->buffer_size
= size
;
2157 if (flags
& CFLGS_OFF_SLAB
)
2158 cachep
->slabp_cache
= kmem_find_general_cachep(slab_size
, 0u);
2159 cachep
->ctor
= ctor
;
2160 cachep
->dtor
= dtor
;
2161 cachep
->name
= name
;
2164 setup_cpu_cache(cachep
);
2166 /* cache setup completed, link it into the list */
2167 list_add(&cachep
->next
, &cache_chain
);
2169 if (!cachep
&& (flags
& SLAB_PANIC
))
2170 panic("kmem_cache_create(): failed to create slab `%s'\n",
2172 mutex_unlock(&cache_chain_mutex
);
2173 unlock_cpu_hotplug();
2176 EXPORT_SYMBOL(kmem_cache_create
);
2179 static void check_irq_off(void)
2181 BUG_ON(!irqs_disabled());
2184 static void check_irq_on(void)
2186 BUG_ON(irqs_disabled());
2189 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2193 assert_spin_locked(&cachep
->nodelists
[numa_node_id()]->list_lock
);
2197 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2201 assert_spin_locked(&cachep
->nodelists
[node
]->list_lock
);
2206 #define check_irq_off() do { } while(0)
2207 #define check_irq_on() do { } while(0)
2208 #define check_spinlock_acquired(x) do { } while(0)
2209 #define check_spinlock_acquired_node(x, y) do { } while(0)
2212 static void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
2213 struct array_cache
*ac
,
2214 int force
, int node
);
2216 static void do_drain(void *arg
)
2218 struct kmem_cache
*cachep
= arg
;
2219 struct array_cache
*ac
;
2220 int node
= numa_node_id();
2223 ac
= cpu_cache_get(cachep
);
2224 spin_lock(&cachep
->nodelists
[node
]->list_lock
);
2225 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
2226 spin_unlock(&cachep
->nodelists
[node
]->list_lock
);
2230 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2232 struct kmem_list3
*l3
;
2235 on_each_cpu(do_drain
, cachep
, 1, 1);
2237 for_each_online_node(node
) {
2238 l3
= cachep
->nodelists
[node
];
2239 if (l3
&& l3
->alien
)
2240 drain_alien_cache(cachep
, l3
->alien
);
2243 for_each_online_node(node
) {
2244 l3
= cachep
->nodelists
[node
];
2246 drain_array(cachep
, l3
, l3
->shared
, 1, node
);
2250 static int __node_shrink(struct kmem_cache
*cachep
, int node
)
2253 struct kmem_list3
*l3
= cachep
->nodelists
[node
];
2257 struct list_head
*p
;
2259 p
= l3
->slabs_free
.prev
;
2260 if (p
== &l3
->slabs_free
)
2263 slabp
= list_entry(l3
->slabs_free
.prev
, struct slab
, list
);
2265 BUG_ON(slabp
->inuse
);
2267 list_del(&slabp
->list
);
2269 l3
->free_objects
-= cachep
->num
;
2270 spin_unlock_irq(&l3
->list_lock
);
2271 slab_destroy(cachep
, slabp
);
2272 spin_lock_irq(&l3
->list_lock
);
2274 ret
= !list_empty(&l3
->slabs_full
) || !list_empty(&l3
->slabs_partial
);
2278 static int __cache_shrink(struct kmem_cache
*cachep
)
2281 struct kmem_list3
*l3
;
2283 drain_cpu_caches(cachep
);
2286 for_each_online_node(i
) {
2287 l3
= cachep
->nodelists
[i
];
2289 spin_lock_irq(&l3
->list_lock
);
2290 ret
+= __node_shrink(cachep
, i
);
2291 spin_unlock_irq(&l3
->list_lock
);
2294 return (ret
? 1 : 0);
2298 * kmem_cache_shrink - Shrink a cache.
2299 * @cachep: The cache to shrink.
2301 * Releases as many slabs as possible for a cache.
2302 * To help debugging, a zero exit status indicates all slabs were released.
2304 int kmem_cache_shrink(struct kmem_cache
*cachep
)
2306 BUG_ON(!cachep
|| in_interrupt());
2308 return __cache_shrink(cachep
);
2310 EXPORT_SYMBOL(kmem_cache_shrink
);
2313 * kmem_cache_destroy - delete a cache
2314 * @cachep: the cache to destroy
2316 * Remove a struct kmem_cache object from the slab cache.
2317 * Returns 0 on success.
2319 * It is expected this function will be called by a module when it is
2320 * unloaded. This will remove the cache completely, and avoid a duplicate
2321 * cache being allocated each time a module is loaded and unloaded, if the
2322 * module doesn't have persistent in-kernel storage across loads and unloads.
2324 * The cache must be empty before calling this function.
2326 * The caller must guarantee that noone will allocate memory from the cache
2327 * during the kmem_cache_destroy().
2329 int kmem_cache_destroy(struct kmem_cache
*cachep
)
2332 struct kmem_list3
*l3
;
2334 BUG_ON(!cachep
|| in_interrupt());
2336 /* Don't let CPUs to come and go */
2339 /* Find the cache in the chain of caches. */
2340 mutex_lock(&cache_chain_mutex
);
2342 * the chain is never empty, cache_cache is never destroyed
2344 list_del(&cachep
->next
);
2345 mutex_unlock(&cache_chain_mutex
);
2347 if (__cache_shrink(cachep
)) {
2348 slab_error(cachep
, "Can't free all objects");
2349 mutex_lock(&cache_chain_mutex
);
2350 list_add(&cachep
->next
, &cache_chain
);
2351 mutex_unlock(&cache_chain_mutex
);
2352 unlock_cpu_hotplug();
2356 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
2359 for_each_online_cpu(i
)
2360 kfree(cachep
->array
[i
]);
2362 /* NUMA: free the list3 structures */
2363 for_each_online_node(i
) {
2364 l3
= cachep
->nodelists
[i
];
2367 free_alien_cache(l3
->alien
);
2371 kmem_cache_free(&cache_cache
, cachep
);
2372 unlock_cpu_hotplug();
2375 EXPORT_SYMBOL(kmem_cache_destroy
);
2377 /* Get the memory for a slab management obj. */
2378 static struct slab
*alloc_slabmgmt(struct kmem_cache
*cachep
, void *objp
,
2379 int colour_off
, gfp_t local_flags
,
2384 if (OFF_SLAB(cachep
)) {
2385 /* Slab management obj is off-slab. */
2386 slabp
= kmem_cache_alloc_node(cachep
->slabp_cache
,
2387 local_flags
, nodeid
);
2391 slabp
= objp
+ colour_off
;
2392 colour_off
+= cachep
->slab_size
;
2395 slabp
->colouroff
= colour_off
;
2396 slabp
->s_mem
= objp
+ colour_off
;
2397 slabp
->nodeid
= nodeid
;
2401 static inline kmem_bufctl_t
*slab_bufctl(struct slab
*slabp
)
2403 return (kmem_bufctl_t
*) (slabp
+ 1);
2406 static void cache_init_objs(struct kmem_cache
*cachep
,
2407 struct slab
*slabp
, unsigned long ctor_flags
)
2411 for (i
= 0; i
< cachep
->num
; i
++) {
2412 void *objp
= index_to_obj(cachep
, slabp
, i
);
2414 /* need to poison the objs? */
2415 if (cachep
->flags
& SLAB_POISON
)
2416 poison_obj(cachep
, objp
, POISON_FREE
);
2417 if (cachep
->flags
& SLAB_STORE_USER
)
2418 *dbg_userword(cachep
, objp
) = NULL
;
2420 if (cachep
->flags
& SLAB_RED_ZONE
) {
2421 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2422 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2425 * Constructors are not allowed to allocate memory from the same
2426 * cache which they are a constructor for. Otherwise, deadlock.
2427 * They must also be threaded.
2429 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2430 cachep
->ctor(objp
+ obj_offset(cachep
), cachep
,
2433 if (cachep
->flags
& SLAB_RED_ZONE
) {
2434 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2435 slab_error(cachep
, "constructor overwrote the"
2436 " end of an object");
2437 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2438 slab_error(cachep
, "constructor overwrote the"
2439 " start of an object");
2441 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 &&
2442 OFF_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
)
2443 kernel_map_pages(virt_to_page(objp
),
2444 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2447 cachep
->ctor(objp
, cachep
, ctor_flags
);
2449 slab_bufctl(slabp
)[i
] = i
+ 1;
2451 slab_bufctl(slabp
)[i
- 1] = BUFCTL_END
;
2455 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2457 if (flags
& SLAB_DMA
)
2458 BUG_ON(!(cachep
->gfpflags
& GFP_DMA
));
2460 BUG_ON(cachep
->gfpflags
& GFP_DMA
);
2463 static void *slab_get_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2466 void *objp
= index_to_obj(cachep
, slabp
, slabp
->free
);
2470 next
= slab_bufctl(slabp
)[slabp
->free
];
2472 slab_bufctl(slabp
)[slabp
->free
] = BUFCTL_FREE
;
2473 WARN_ON(slabp
->nodeid
!= nodeid
);
2480 static void slab_put_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2481 void *objp
, int nodeid
)
2483 unsigned int objnr
= obj_to_index(cachep
, slabp
, objp
);
2486 /* Verify that the slab belongs to the intended node */
2487 WARN_ON(slabp
->nodeid
!= nodeid
);
2489 if (slab_bufctl(slabp
)[objnr
] + 1 <= SLAB_LIMIT
+ 1) {
2490 printk(KERN_ERR
"slab: double free detected in cache "
2491 "'%s', objp %p\n", cachep
->name
, objp
);
2495 slab_bufctl(slabp
)[objnr
] = slabp
->free
;
2496 slabp
->free
= objnr
;
2501 * Map pages beginning at addr to the given cache and slab. This is required
2502 * for the slab allocator to be able to lookup the cache and slab of a
2503 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2505 static void slab_map_pages(struct kmem_cache
*cache
, struct slab
*slab
,
2511 page
= virt_to_page(addr
);
2514 if (likely(!PageCompound(page
)))
2515 nr_pages
<<= cache
->gfporder
;
2518 page_set_cache(page
, cache
);
2519 page_set_slab(page
, slab
);
2521 } while (--nr_pages
);
2525 * Grow (by 1) the number of slabs within a cache. This is called by
2526 * kmem_cache_alloc() when there are no active objs left in a cache.
2528 static int cache_grow(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
2534 unsigned long ctor_flags
;
2535 struct kmem_list3
*l3
;
2538 * Be lazy and only check for valid flags here, keeping it out of the
2539 * critical path in kmem_cache_alloc().
2541 BUG_ON(flags
& ~(SLAB_DMA
| SLAB_LEVEL_MASK
| SLAB_NO_GROW
));
2542 if (flags
& SLAB_NO_GROW
)
2545 ctor_flags
= SLAB_CTOR_CONSTRUCTOR
;
2546 local_flags
= (flags
& SLAB_LEVEL_MASK
);
2547 if (!(local_flags
& __GFP_WAIT
))
2549 * Not allowed to sleep. Need to tell a constructor about
2550 * this - it might need to know...
2552 ctor_flags
|= SLAB_CTOR_ATOMIC
;
2554 /* Take the l3 list lock to change the colour_next on this node */
2556 l3
= cachep
->nodelists
[nodeid
];
2557 spin_lock(&l3
->list_lock
);
2559 /* Get colour for the slab, and cal the next value. */
2560 offset
= l3
->colour_next
;
2562 if (l3
->colour_next
>= cachep
->colour
)
2563 l3
->colour_next
= 0;
2564 spin_unlock(&l3
->list_lock
);
2566 offset
*= cachep
->colour_off
;
2568 if (local_flags
& __GFP_WAIT
)
2572 * The test for missing atomic flag is performed here, rather than
2573 * the more obvious place, simply to reduce the critical path length
2574 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2575 * will eventually be caught here (where it matters).
2577 kmem_flagcheck(cachep
, flags
);
2580 * Get mem for the objs. Attempt to allocate a physical page from
2583 objp
= kmem_getpages(cachep
, flags
, nodeid
);
2587 /* Get slab management. */
2588 slabp
= alloc_slabmgmt(cachep
, objp
, offset
, local_flags
, nodeid
);
2592 slabp
->nodeid
= nodeid
;
2593 slab_map_pages(cachep
, slabp
, objp
);
2595 cache_init_objs(cachep
, slabp
, ctor_flags
);
2597 if (local_flags
& __GFP_WAIT
)
2598 local_irq_disable();
2600 spin_lock(&l3
->list_lock
);
2602 /* Make slab active. */
2603 list_add_tail(&slabp
->list
, &(l3
->slabs_free
));
2604 STATS_INC_GROWN(cachep
);
2605 l3
->free_objects
+= cachep
->num
;
2606 spin_unlock(&l3
->list_lock
);
2609 kmem_freepages(cachep
, objp
);
2611 if (local_flags
& __GFP_WAIT
)
2612 local_irq_disable();
2619 * Perform extra freeing checks:
2620 * - detect bad pointers.
2621 * - POISON/RED_ZONE checking
2622 * - destructor calls, for caches with POISON+dtor
2624 static void kfree_debugcheck(const void *objp
)
2628 if (!virt_addr_valid(objp
)) {
2629 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
2630 (unsigned long)objp
);
2633 page
= virt_to_page(objp
);
2634 if (!PageSlab(page
)) {
2635 printk(KERN_ERR
"kfree_debugcheck: bad ptr %lxh.\n",
2636 (unsigned long)objp
);
2641 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2648 objp
-= obj_offset(cachep
);
2649 kfree_debugcheck(objp
);
2650 page
= virt_to_page(objp
);
2652 if (page_get_cache(page
) != cachep
) {
2653 printk(KERN_ERR
"mismatch in kmem_cache_free: expected "
2654 "cache %p, got %p\n",
2655 page_get_cache(page
), cachep
);
2656 printk(KERN_ERR
"%p is %s.\n", cachep
, cachep
->name
);
2657 printk(KERN_ERR
"%p is %s.\n", page_get_cache(page
),
2658 page_get_cache(page
)->name
);
2661 slabp
= page_get_slab(page
);
2663 if (cachep
->flags
& SLAB_RED_ZONE
) {
2664 if (*dbg_redzone1(cachep
, objp
) != RED_ACTIVE
||
2665 *dbg_redzone2(cachep
, objp
) != RED_ACTIVE
) {
2666 slab_error(cachep
, "double free, or memory outside"
2667 " object was overwritten");
2668 printk(KERN_ERR
"%p: redzone 1:0x%lx, "
2669 "redzone 2:0x%lx.\n",
2670 objp
, *dbg_redzone1(cachep
, objp
),
2671 *dbg_redzone2(cachep
, objp
));
2673 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2674 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2676 if (cachep
->flags
& SLAB_STORE_USER
)
2677 *dbg_userword(cachep
, objp
) = caller
;
2679 objnr
= obj_to_index(cachep
, slabp
, objp
);
2681 BUG_ON(objnr
>= cachep
->num
);
2682 BUG_ON(objp
!= index_to_obj(cachep
, slabp
, objnr
));
2684 if (cachep
->flags
& SLAB_DEBUG_INITIAL
) {
2686 * Need to call the slab's constructor so the caller can
2687 * perform a verify of its state (debugging). Called without
2688 * the cache-lock held.
2690 cachep
->ctor(objp
+ obj_offset(cachep
),
2691 cachep
, SLAB_CTOR_CONSTRUCTOR
| SLAB_CTOR_VERIFY
);
2693 if (cachep
->flags
& SLAB_POISON
&& cachep
->dtor
) {
2694 /* we want to cache poison the object,
2695 * call the destruction callback
2697 cachep
->dtor(objp
+ obj_offset(cachep
), cachep
, 0);
2699 #ifdef CONFIG_DEBUG_SLAB_LEAK
2700 slab_bufctl(slabp
)[objnr
] = BUFCTL_FREE
;
2702 if (cachep
->flags
& SLAB_POISON
) {
2703 #ifdef CONFIG_DEBUG_PAGEALLOC
2704 if ((cachep
->buffer_size
% PAGE_SIZE
)==0 && OFF_SLAB(cachep
)) {
2705 store_stackinfo(cachep
, objp
, (unsigned long)caller
);
2706 kernel_map_pages(virt_to_page(objp
),
2707 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2709 poison_obj(cachep
, objp
, POISON_FREE
);
2712 poison_obj(cachep
, objp
, POISON_FREE
);
2718 static void check_slabp(struct kmem_cache
*cachep
, struct slab
*slabp
)
2723 /* Check slab's freelist to see if this obj is there. */
2724 for (i
= slabp
->free
; i
!= BUFCTL_END
; i
= slab_bufctl(slabp
)[i
]) {
2726 if (entries
> cachep
->num
|| i
>= cachep
->num
)
2729 if (entries
!= cachep
->num
- slabp
->inuse
) {
2731 printk(KERN_ERR
"slab: Internal list corruption detected in "
2732 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2733 cachep
->name
, cachep
->num
, slabp
, slabp
->inuse
);
2735 i
< sizeof(*slabp
) + cachep
->num
* sizeof(kmem_bufctl_t
);
2738 printk("\n%03x:", i
);
2739 printk(" %02x", ((unsigned char *)slabp
)[i
]);
2746 #define kfree_debugcheck(x) do { } while(0)
2747 #define cache_free_debugcheck(x,objp,z) (objp)
2748 #define check_slabp(x,y) do { } while(0)
2751 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
2754 struct kmem_list3
*l3
;
2755 struct array_cache
*ac
;
2758 ac
= cpu_cache_get(cachep
);
2760 batchcount
= ac
->batchcount
;
2761 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2763 * If there was little recent activity on this cache, then
2764 * perform only a partial refill. Otherwise we could generate
2767 batchcount
= BATCHREFILL_LIMIT
;
2769 l3
= cachep
->nodelists
[numa_node_id()];
2771 BUG_ON(ac
->avail
> 0 || !l3
);
2772 spin_lock(&l3
->list_lock
);
2774 /* See if we can refill from the shared array */
2775 if (l3
->shared
&& transfer_objects(ac
, l3
->shared
, batchcount
))
2778 while (batchcount
> 0) {
2779 struct list_head
*entry
;
2781 /* Get slab alloc is to come from. */
2782 entry
= l3
->slabs_partial
.next
;
2783 if (entry
== &l3
->slabs_partial
) {
2784 l3
->free_touched
= 1;
2785 entry
= l3
->slabs_free
.next
;
2786 if (entry
== &l3
->slabs_free
)
2790 slabp
= list_entry(entry
, struct slab
, list
);
2791 check_slabp(cachep
, slabp
);
2792 check_spinlock_acquired(cachep
);
2793 while (slabp
->inuse
< cachep
->num
&& batchcount
--) {
2794 STATS_INC_ALLOCED(cachep
);
2795 STATS_INC_ACTIVE(cachep
);
2796 STATS_SET_HIGH(cachep
);
2798 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, slabp
,
2801 check_slabp(cachep
, slabp
);
2803 /* move slabp to correct slabp list: */
2804 list_del(&slabp
->list
);
2805 if (slabp
->free
== BUFCTL_END
)
2806 list_add(&slabp
->list
, &l3
->slabs_full
);
2808 list_add(&slabp
->list
, &l3
->slabs_partial
);
2812 l3
->free_objects
-= ac
->avail
;
2814 spin_unlock(&l3
->list_lock
);
2816 if (unlikely(!ac
->avail
)) {
2818 x
= cache_grow(cachep
, flags
, numa_node_id());
2820 /* cache_grow can reenable interrupts, then ac could change. */
2821 ac
= cpu_cache_get(cachep
);
2822 if (!x
&& ac
->avail
== 0) /* no objects in sight? abort */
2825 if (!ac
->avail
) /* objects refilled by interrupt? */
2829 return ac
->entry
[--ac
->avail
];
2832 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
2835 might_sleep_if(flags
& __GFP_WAIT
);
2837 kmem_flagcheck(cachep
, flags
);
2842 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
2843 gfp_t flags
, void *objp
, void *caller
)
2847 if (cachep
->flags
& SLAB_POISON
) {
2848 #ifdef CONFIG_DEBUG_PAGEALLOC
2849 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
2850 kernel_map_pages(virt_to_page(objp
),
2851 cachep
->buffer_size
/ PAGE_SIZE
, 1);
2853 check_poison_obj(cachep
, objp
);
2855 check_poison_obj(cachep
, objp
);
2857 poison_obj(cachep
, objp
, POISON_INUSE
);
2859 if (cachep
->flags
& SLAB_STORE_USER
)
2860 *dbg_userword(cachep
, objp
) = caller
;
2862 if (cachep
->flags
& SLAB_RED_ZONE
) {
2863 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
2864 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
2865 slab_error(cachep
, "double free, or memory outside"
2866 " object was overwritten");
2868 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
2869 objp
, *dbg_redzone1(cachep
, objp
),
2870 *dbg_redzone2(cachep
, objp
));
2872 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
2873 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
2875 #ifdef CONFIG_DEBUG_SLAB_LEAK
2880 slabp
= page_get_slab(virt_to_page(objp
));
2881 objnr
= (unsigned)(objp
- slabp
->s_mem
) / cachep
->buffer_size
;
2882 slab_bufctl(slabp
)[objnr
] = BUFCTL_ACTIVE
;
2885 objp
+= obj_offset(cachep
);
2886 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
) {
2887 unsigned long ctor_flags
= SLAB_CTOR_CONSTRUCTOR
;
2889 if (!(flags
& __GFP_WAIT
))
2890 ctor_flags
|= SLAB_CTOR_ATOMIC
;
2892 cachep
->ctor(objp
, cachep
, ctor_flags
);
2897 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2900 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
2903 struct array_cache
*ac
;
2906 if (unlikely(current
->flags
& (PF_SPREAD_SLAB
| PF_MEMPOLICY
))) {
2907 objp
= alternate_node_alloc(cachep
, flags
);
2914 ac
= cpu_cache_get(cachep
);
2915 if (likely(ac
->avail
)) {
2916 STATS_INC_ALLOCHIT(cachep
);
2918 objp
= ac
->entry
[--ac
->avail
];
2920 STATS_INC_ALLOCMISS(cachep
);
2921 objp
= cache_alloc_refill(cachep
, flags
);
2926 static __always_inline
void *__cache_alloc(struct kmem_cache
*cachep
,
2927 gfp_t flags
, void *caller
)
2929 unsigned long save_flags
;
2932 cache_alloc_debugcheck_before(cachep
, flags
);
2934 local_irq_save(save_flags
);
2935 objp
= ____cache_alloc(cachep
, flags
);
2936 local_irq_restore(save_flags
);
2937 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
,
2945 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
2947 * If we are in_interrupt, then process context, including cpusets and
2948 * mempolicy, may not apply and should not be used for allocation policy.
2950 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
2952 int nid_alloc
, nid_here
;
2956 nid_alloc
= nid_here
= numa_node_id();
2957 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
2958 nid_alloc
= cpuset_mem_spread_node();
2959 else if (current
->mempolicy
)
2960 nid_alloc
= slab_node(current
->mempolicy
);
2961 if (nid_alloc
!= nid_here
)
2962 return __cache_alloc_node(cachep
, flags
, nid_alloc
);
2967 * A interface to enable slab creation on nodeid
2969 static void *__cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
2972 struct list_head
*entry
;
2974 struct kmem_list3
*l3
;
2978 l3
= cachep
->nodelists
[nodeid
];
2983 spin_lock(&l3
->list_lock
);
2984 entry
= l3
->slabs_partial
.next
;
2985 if (entry
== &l3
->slabs_partial
) {
2986 l3
->free_touched
= 1;
2987 entry
= l3
->slabs_free
.next
;
2988 if (entry
== &l3
->slabs_free
)
2992 slabp
= list_entry(entry
, struct slab
, list
);
2993 check_spinlock_acquired_node(cachep
, nodeid
);
2994 check_slabp(cachep
, slabp
);
2996 STATS_INC_NODEALLOCS(cachep
);
2997 STATS_INC_ACTIVE(cachep
);
2998 STATS_SET_HIGH(cachep
);
3000 BUG_ON(slabp
->inuse
== cachep
->num
);
3002 obj
= slab_get_obj(cachep
, slabp
, nodeid
);
3003 check_slabp(cachep
, slabp
);
3005 /* move slabp to correct slabp list: */
3006 list_del(&slabp
->list
);
3008 if (slabp
->free
== BUFCTL_END
)
3009 list_add(&slabp
->list
, &l3
->slabs_full
);
3011 list_add(&slabp
->list
, &l3
->slabs_partial
);
3013 spin_unlock(&l3
->list_lock
);
3017 spin_unlock(&l3
->list_lock
);
3018 x
= cache_grow(cachep
, flags
, nodeid
);
3030 * Caller needs to acquire correct kmem_list's list_lock
3032 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int nr_objects
,
3036 struct kmem_list3
*l3
;
3038 for (i
= 0; i
< nr_objects
; i
++) {
3039 void *objp
= objpp
[i
];
3042 slabp
= virt_to_slab(objp
);
3043 l3
= cachep
->nodelists
[node
];
3044 list_del(&slabp
->list
);
3045 check_spinlock_acquired_node(cachep
, node
);
3046 check_slabp(cachep
, slabp
);
3047 slab_put_obj(cachep
, slabp
, objp
, node
);
3048 STATS_DEC_ACTIVE(cachep
);
3050 check_slabp(cachep
, slabp
);
3052 /* fixup slab chains */
3053 if (slabp
->inuse
== 0) {
3054 if (l3
->free_objects
> l3
->free_limit
) {
3055 l3
->free_objects
-= cachep
->num
;
3056 slab_destroy(cachep
, slabp
);
3058 list_add(&slabp
->list
, &l3
->slabs_free
);
3061 /* Unconditionally move a slab to the end of the
3062 * partial list on free - maximum time for the
3063 * other objects to be freed, too.
3065 list_add_tail(&slabp
->list
, &l3
->slabs_partial
);
3070 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3073 struct kmem_list3
*l3
;
3074 int node
= numa_node_id();
3076 batchcount
= ac
->batchcount
;
3078 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
3081 l3
= cachep
->nodelists
[node
];
3082 spin_lock(&l3
->list_lock
);
3084 struct array_cache
*shared_array
= l3
->shared
;
3085 int max
= shared_array
->limit
- shared_array
->avail
;
3087 if (batchcount
> max
)
3089 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3090 ac
->entry
, sizeof(void *) * batchcount
);
3091 shared_array
->avail
+= batchcount
;
3096 free_block(cachep
, ac
->entry
, batchcount
, node
);
3101 struct list_head
*p
;
3103 p
= l3
->slabs_free
.next
;
3104 while (p
!= &(l3
->slabs_free
)) {
3107 slabp
= list_entry(p
, struct slab
, list
);
3108 BUG_ON(slabp
->inuse
);
3113 STATS_SET_FREEABLE(cachep
, i
);
3116 spin_unlock(&l3
->list_lock
);
3117 ac
->avail
-= batchcount
;
3118 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3122 * Release an obj back to its cache. If the obj has a constructed state, it must
3123 * be in this state _before_ it is released. Called with disabled ints.
3125 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
)
3127 struct array_cache
*ac
= cpu_cache_get(cachep
);
3130 objp
= cache_free_debugcheck(cachep
, objp
, __builtin_return_address(0));
3132 if (cache_free_alien(cachep
, objp
))
3135 if (likely(ac
->avail
< ac
->limit
)) {
3136 STATS_INC_FREEHIT(cachep
);
3137 ac
->entry
[ac
->avail
++] = objp
;
3140 STATS_INC_FREEMISS(cachep
);
3141 cache_flusharray(cachep
, ac
);
3142 ac
->entry
[ac
->avail
++] = objp
;
3147 * kmem_cache_alloc - Allocate an object
3148 * @cachep: The cache to allocate from.
3149 * @flags: See kmalloc().
3151 * Allocate an object from this cache. The flags are only relevant
3152 * if the cache has no available objects.
3154 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3156 return __cache_alloc(cachep
, flags
, __builtin_return_address(0));
3158 EXPORT_SYMBOL(kmem_cache_alloc
);
3161 * kmem_cache_alloc - Allocate an object. The memory is set to zero.
3162 * @cache: The cache to allocate from.
3163 * @flags: See kmalloc().
3165 * Allocate an object from this cache and set the allocated memory to zero.
3166 * The flags are only relevant if the cache has no available objects.
3168 void *kmem_cache_zalloc(struct kmem_cache
*cache
, gfp_t flags
)
3170 void *ret
= __cache_alloc(cache
, flags
, __builtin_return_address(0));
3172 memset(ret
, 0, obj_size(cache
));
3175 EXPORT_SYMBOL(kmem_cache_zalloc
);
3178 * kmem_ptr_validate - check if an untrusted pointer might
3180 * @cachep: the cache we're checking against
3181 * @ptr: pointer to validate
3183 * This verifies that the untrusted pointer looks sane:
3184 * it is _not_ a guarantee that the pointer is actually
3185 * part of the slab cache in question, but it at least
3186 * validates that the pointer can be dereferenced and
3187 * looks half-way sane.
3189 * Currently only used for dentry validation.
3191 int fastcall
kmem_ptr_validate(struct kmem_cache
*cachep
, void *ptr
)
3193 unsigned long addr
= (unsigned long)ptr
;
3194 unsigned long min_addr
= PAGE_OFFSET
;
3195 unsigned long align_mask
= BYTES_PER_WORD
- 1;
3196 unsigned long size
= cachep
->buffer_size
;
3199 if (unlikely(addr
< min_addr
))
3201 if (unlikely(addr
> (unsigned long)high_memory
- size
))
3203 if (unlikely(addr
& align_mask
))
3205 if (unlikely(!kern_addr_valid(addr
)))
3207 if (unlikely(!kern_addr_valid(addr
+ size
- 1)))
3209 page
= virt_to_page(ptr
);
3210 if (unlikely(!PageSlab(page
)))
3212 if (unlikely(page_get_cache(page
) != cachep
))
3221 * kmem_cache_alloc_node - Allocate an object on the specified node
3222 * @cachep: The cache to allocate from.
3223 * @flags: See kmalloc().
3224 * @nodeid: node number of the target node.
3226 * Identical to kmem_cache_alloc, except that this function is slow
3227 * and can sleep. And it will allocate memory on the given node, which
3228 * can improve the performance for cpu bound structures.
3229 * New and improved: it will now make sure that the object gets
3230 * put on the correct node list so that there is no false sharing.
3232 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3234 unsigned long save_flags
;
3237 cache_alloc_debugcheck_before(cachep
, flags
);
3238 local_irq_save(save_flags
);
3240 if (nodeid
== -1 || nodeid
== numa_node_id() ||
3241 !cachep
->nodelists
[nodeid
])
3242 ptr
= ____cache_alloc(cachep
, flags
);
3244 ptr
= __cache_alloc_node(cachep
, flags
, nodeid
);
3245 local_irq_restore(save_flags
);
3247 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
,
3248 __builtin_return_address(0));
3252 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3254 void *kmalloc_node(size_t size
, gfp_t flags
, int node
)
3256 struct kmem_cache
*cachep
;
3258 cachep
= kmem_find_general_cachep(size
, flags
);
3259 if (unlikely(cachep
== NULL
))
3261 return kmem_cache_alloc_node(cachep
, flags
, node
);
3263 EXPORT_SYMBOL(kmalloc_node
);
3267 * kmalloc - allocate memory
3268 * @size: how many bytes of memory are required.
3269 * @flags: the type of memory to allocate.
3270 * @caller: function caller for debug tracking of the caller
3272 * kmalloc is the normal method of allocating memory
3275 * The @flags argument may be one of:
3277 * %GFP_USER - Allocate memory on behalf of user. May sleep.
3279 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
3281 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
3283 * Additionally, the %GFP_DMA flag may be set to indicate the memory
3284 * must be suitable for DMA. This can mean different things on different
3285 * platforms. For example, on i386, it means that the memory must come
3286 * from the first 16MB.
3288 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3291 struct kmem_cache
*cachep
;
3293 /* If you want to save a few bytes .text space: replace
3295 * Then kmalloc uses the uninlined functions instead of the inline
3298 cachep
= __find_general_cachep(size
, flags
);
3299 if (unlikely(cachep
== NULL
))
3301 return __cache_alloc(cachep
, flags
, caller
);
3305 void *__kmalloc(size_t size
, gfp_t flags
)
3307 #ifndef CONFIG_DEBUG_SLAB
3308 return __do_kmalloc(size
, flags
, NULL
);
3310 return __do_kmalloc(size
, flags
, __builtin_return_address(0));
3313 EXPORT_SYMBOL(__kmalloc
);
3315 #ifdef CONFIG_DEBUG_SLAB
3316 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, void *caller
)
3318 return __do_kmalloc(size
, flags
, caller
);
3320 EXPORT_SYMBOL(__kmalloc_track_caller
);
3325 * __alloc_percpu - allocate one copy of the object for every present
3326 * cpu in the system, zeroing them.
3327 * Objects should be dereferenced using the per_cpu_ptr macro only.
3329 * @size: how many bytes of memory are required.
3331 void *__alloc_percpu(size_t size
)
3334 struct percpu_data
*pdata
= kmalloc(sizeof(*pdata
), GFP_KERNEL
);
3340 * Cannot use for_each_online_cpu since a cpu may come online
3341 * and we have no way of figuring out how to fix the array
3342 * that we have allocated then....
3344 for_each_possible_cpu(i
) {
3345 int node
= cpu_to_node(i
);
3347 if (node_online(node
))
3348 pdata
->ptrs
[i
] = kmalloc_node(size
, GFP_KERNEL
, node
);
3350 pdata
->ptrs
[i
] = kmalloc(size
, GFP_KERNEL
);
3352 if (!pdata
->ptrs
[i
])
3354 memset(pdata
->ptrs
[i
], 0, size
);
3357 /* Catch derefs w/o wrappers */
3358 return (void *)(~(unsigned long)pdata
);
3362 if (!cpu_possible(i
))
3364 kfree(pdata
->ptrs
[i
]);
3369 EXPORT_SYMBOL(__alloc_percpu
);
3373 * kmem_cache_free - Deallocate an object
3374 * @cachep: The cache the allocation was from.
3375 * @objp: The previously allocated object.
3377 * Free an object which was previously allocated from this
3380 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3382 unsigned long flags
;
3384 local_irq_save(flags
);
3385 __cache_free(cachep
, objp
);
3386 local_irq_restore(flags
);
3388 EXPORT_SYMBOL(kmem_cache_free
);
3391 * kfree - free previously allocated memory
3392 * @objp: pointer returned by kmalloc.
3394 * If @objp is NULL, no operation is performed.
3396 * Don't free memory not originally allocated by kmalloc()
3397 * or you will run into trouble.
3399 void kfree(const void *objp
)
3401 struct kmem_cache
*c
;
3402 unsigned long flags
;
3404 if (unlikely(!objp
))
3406 local_irq_save(flags
);
3407 kfree_debugcheck(objp
);
3408 c
= virt_to_cache(objp
);
3409 mutex_debug_check_no_locks_freed(objp
, obj_size(c
));
3410 __cache_free(c
, (void *)objp
);
3411 local_irq_restore(flags
);
3413 EXPORT_SYMBOL(kfree
);
3417 * free_percpu - free previously allocated percpu memory
3418 * @objp: pointer returned by alloc_percpu.
3420 * Don't free memory not originally allocated by alloc_percpu()
3421 * The complemented objp is to check for that.
3423 void free_percpu(const void *objp
)
3426 struct percpu_data
*p
= (struct percpu_data
*)(~(unsigned long)objp
);
3429 * We allocate for all cpus so we cannot use for online cpu here.
3431 for_each_possible_cpu(i
)
3435 EXPORT_SYMBOL(free_percpu
);
3438 unsigned int kmem_cache_size(struct kmem_cache
*cachep
)
3440 return obj_size(cachep
);
3442 EXPORT_SYMBOL(kmem_cache_size
);
3444 const char *kmem_cache_name(struct kmem_cache
*cachep
)
3446 return cachep
->name
;
3448 EXPORT_SYMBOL_GPL(kmem_cache_name
);
3451 * This initializes kmem_list3 or resizes varioius caches for all nodes.
3453 static int alloc_kmemlist(struct kmem_cache
*cachep
)
3456 struct kmem_list3
*l3
;
3457 struct array_cache
*new_shared
;
3458 struct array_cache
**new_alien
;
3460 for_each_online_node(node
) {
3462 new_alien
= alloc_alien_cache(node
, cachep
->limit
);
3466 new_shared
= alloc_arraycache(node
,
3467 cachep
->shared
*cachep
->batchcount
,
3470 free_alien_cache(new_alien
);
3474 l3
= cachep
->nodelists
[node
];
3476 struct array_cache
*shared
= l3
->shared
;
3478 spin_lock_irq(&l3
->list_lock
);
3481 free_block(cachep
, shared
->entry
,
3482 shared
->avail
, node
);
3484 l3
->shared
= new_shared
;
3486 l3
->alien
= new_alien
;
3489 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3490 cachep
->batchcount
+ cachep
->num
;
3491 spin_unlock_irq(&l3
->list_lock
);
3493 free_alien_cache(new_alien
);
3496 l3
= kmalloc_node(sizeof(struct kmem_list3
), GFP_KERNEL
, node
);
3498 free_alien_cache(new_alien
);
3503 kmem_list3_init(l3
);
3504 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
3505 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
3506 l3
->shared
= new_shared
;
3507 l3
->alien
= new_alien
;
3508 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3509 cachep
->batchcount
+ cachep
->num
;
3510 cachep
->nodelists
[node
] = l3
;
3515 if (!cachep
->next
.next
) {
3516 /* Cache is not active yet. Roll back what we did */
3519 if (cachep
->nodelists
[node
]) {
3520 l3
= cachep
->nodelists
[node
];
3523 free_alien_cache(l3
->alien
);
3525 cachep
->nodelists
[node
] = NULL
;
3533 struct ccupdate_struct
{
3534 struct kmem_cache
*cachep
;
3535 struct array_cache
*new[NR_CPUS
];
3538 static void do_ccupdate_local(void *info
)
3540 struct ccupdate_struct
*new = info
;
3541 struct array_cache
*old
;
3544 old
= cpu_cache_get(new->cachep
);
3546 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
3547 new->new[smp_processor_id()] = old
;
3550 /* Always called with the cache_chain_mutex held */
3551 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3552 int batchcount
, int shared
)
3554 struct ccupdate_struct
new;
3557 memset(&new.new, 0, sizeof(new.new));
3558 for_each_online_cpu(i
) {
3559 new.new[i
] = alloc_arraycache(cpu_to_node(i
), limit
,
3562 for (i
--; i
>= 0; i
--)
3567 new.cachep
= cachep
;
3569 on_each_cpu(do_ccupdate_local
, (void *)&new, 1, 1);
3572 cachep
->batchcount
= batchcount
;
3573 cachep
->limit
= limit
;
3574 cachep
->shared
= shared
;
3576 for_each_online_cpu(i
) {
3577 struct array_cache
*ccold
= new.new[i
];
3580 spin_lock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3581 free_block(cachep
, ccold
->entry
, ccold
->avail
, cpu_to_node(i
));
3582 spin_unlock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3586 err
= alloc_kmemlist(cachep
);
3588 printk(KERN_ERR
"alloc_kmemlist failed for %s, error %d.\n",
3589 cachep
->name
, -err
);
3595 /* Called with cache_chain_mutex held always */
3596 static void enable_cpucache(struct kmem_cache
*cachep
)
3602 * The head array serves three purposes:
3603 * - create a LIFO ordering, i.e. return objects that are cache-warm
3604 * - reduce the number of spinlock operations.
3605 * - reduce the number of linked list operations on the slab and
3606 * bufctl chains: array operations are cheaper.
3607 * The numbers are guessed, we should auto-tune as described by
3610 if (cachep
->buffer_size
> 131072)
3612 else if (cachep
->buffer_size
> PAGE_SIZE
)
3614 else if (cachep
->buffer_size
> 1024)
3616 else if (cachep
->buffer_size
> 256)
3622 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3623 * allocation behaviour: Most allocs on one cpu, most free operations
3624 * on another cpu. For these cases, an efficient object passing between
3625 * cpus is necessary. This is provided by a shared array. The array
3626 * replaces Bonwick's magazine layer.
3627 * On uniprocessor, it's functionally equivalent (but less efficient)
3628 * to a larger limit. Thus disabled by default.
3632 if (cachep
->buffer_size
<= PAGE_SIZE
)
3638 * With debugging enabled, large batchcount lead to excessively long
3639 * periods with disabled local interrupts. Limit the batchcount
3644 err
= do_tune_cpucache(cachep
, limit
, (limit
+ 1) / 2, shared
);
3646 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
3647 cachep
->name
, -err
);
3651 * Drain an array if it contains any elements taking the l3 lock only if
3652 * necessary. Note that the l3 listlock also protects the array_cache
3653 * if drain_array() is used on the shared array.
3655 void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
3656 struct array_cache
*ac
, int force
, int node
)
3660 if (!ac
|| !ac
->avail
)
3662 if (ac
->touched
&& !force
) {
3665 spin_lock_irq(&l3
->list_lock
);
3667 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
3668 if (tofree
> ac
->avail
)
3669 tofree
= (ac
->avail
+ 1) / 2;
3670 free_block(cachep
, ac
->entry
, tofree
, node
);
3671 ac
->avail
-= tofree
;
3672 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
3673 sizeof(void *) * ac
->avail
);
3675 spin_unlock_irq(&l3
->list_lock
);
3680 * cache_reap - Reclaim memory from caches.
3681 * @unused: unused parameter
3683 * Called from workqueue/eventd every few seconds.
3685 * - clear the per-cpu caches for this CPU.
3686 * - return freeable pages to the main free memory pool.
3688 * If we cannot acquire the cache chain mutex then just give up - we'll try
3689 * again on the next iteration.
3691 static void cache_reap(void *unused
)
3693 struct list_head
*walk
;
3694 struct kmem_list3
*l3
;
3695 int node
= numa_node_id();
3697 if (!mutex_trylock(&cache_chain_mutex
)) {
3698 /* Give up. Setup the next iteration. */
3699 schedule_delayed_work(&__get_cpu_var(reap_work
),
3704 list_for_each(walk
, &cache_chain
) {
3705 struct kmem_cache
*searchp
;
3706 struct list_head
*p
;
3710 searchp
= list_entry(walk
, struct kmem_cache
, next
);
3714 * We only take the l3 lock if absolutely necessary and we
3715 * have established with reasonable certainty that
3716 * we can do some work if the lock was obtained.
3718 l3
= searchp
->nodelists
[node
];
3720 reap_alien(searchp
, l3
);
3722 drain_array(searchp
, l3
, cpu_cache_get(searchp
), 0, node
);
3725 * These are racy checks but it does not matter
3726 * if we skip one check or scan twice.
3728 if (time_after(l3
->next_reap
, jiffies
))
3731 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
;
3733 drain_array(searchp
, l3
, l3
->shared
, 0, node
);
3735 if (l3
->free_touched
) {
3736 l3
->free_touched
= 0;
3740 tofree
= (l3
->free_limit
+ 5 * searchp
->num
- 1) /
3744 * Do not lock if there are no free blocks.
3746 if (list_empty(&l3
->slabs_free
))
3749 spin_lock_irq(&l3
->list_lock
);
3750 p
= l3
->slabs_free
.next
;
3751 if (p
== &(l3
->slabs_free
)) {
3752 spin_unlock_irq(&l3
->list_lock
);
3756 slabp
= list_entry(p
, struct slab
, list
);
3757 BUG_ON(slabp
->inuse
);
3758 list_del(&slabp
->list
);
3759 STATS_INC_REAPED(searchp
);
3762 * Safe to drop the lock. The slab is no longer linked
3763 * to the cache. searchp cannot disappear, we hold
3766 l3
->free_objects
-= searchp
->num
;
3767 spin_unlock_irq(&l3
->list_lock
);
3768 slab_destroy(searchp
, slabp
);
3769 } while (--tofree
> 0);
3774 mutex_unlock(&cache_chain_mutex
);
3776 /* Set up the next iteration */
3777 schedule_delayed_work(&__get_cpu_var(reap_work
), REAPTIMEOUT_CPUC
);
3780 #ifdef CONFIG_PROC_FS
3782 static void print_slabinfo_header(struct seq_file
*m
)
3785 * Output format version, so at least we can change it
3786 * without _too_ many complaints.
3789 seq_puts(m
, "slabinfo - version: 2.1 (statistics)\n");
3791 seq_puts(m
, "slabinfo - version: 2.1\n");
3793 seq_puts(m
, "# name <active_objs> <num_objs> <objsize> "
3794 "<objperslab> <pagesperslab>");
3795 seq_puts(m
, " : tunables <limit> <batchcount> <sharedfactor>");
3796 seq_puts(m
, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3798 seq_puts(m
, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3799 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
3800 seq_puts(m
, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3805 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
3808 struct list_head
*p
;
3810 mutex_lock(&cache_chain_mutex
);
3812 print_slabinfo_header(m
);
3813 p
= cache_chain
.next
;
3816 if (p
== &cache_chain
)
3819 return list_entry(p
, struct kmem_cache
, next
);
3822 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
3824 struct kmem_cache
*cachep
= p
;
3826 return cachep
->next
.next
== &cache_chain
?
3827 NULL
: list_entry(cachep
->next
.next
, struct kmem_cache
, next
);
3830 static void s_stop(struct seq_file
*m
, void *p
)
3832 mutex_unlock(&cache_chain_mutex
);
3835 static int s_show(struct seq_file
*m
, void *p
)
3837 struct kmem_cache
*cachep
= p
;
3838 struct list_head
*q
;
3840 unsigned long active_objs
;
3841 unsigned long num_objs
;
3842 unsigned long active_slabs
= 0;
3843 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
3847 struct kmem_list3
*l3
;
3851 for_each_online_node(node
) {
3852 l3
= cachep
->nodelists
[node
];
3857 spin_lock_irq(&l3
->list_lock
);
3859 list_for_each(q
, &l3
->slabs_full
) {
3860 slabp
= list_entry(q
, struct slab
, list
);
3861 if (slabp
->inuse
!= cachep
->num
&& !error
)
3862 error
= "slabs_full accounting error";
3863 active_objs
+= cachep
->num
;
3866 list_for_each(q
, &l3
->slabs_partial
) {
3867 slabp
= list_entry(q
, struct slab
, list
);
3868 if (slabp
->inuse
== cachep
->num
&& !error
)
3869 error
= "slabs_partial inuse accounting error";
3870 if (!slabp
->inuse
&& !error
)
3871 error
= "slabs_partial/inuse accounting error";
3872 active_objs
+= slabp
->inuse
;
3875 list_for_each(q
, &l3
->slabs_free
) {
3876 slabp
= list_entry(q
, struct slab
, list
);
3877 if (slabp
->inuse
&& !error
)
3878 error
= "slabs_free/inuse accounting error";
3881 free_objects
+= l3
->free_objects
;
3883 shared_avail
+= l3
->shared
->avail
;
3885 spin_unlock_irq(&l3
->list_lock
);
3887 num_slabs
+= active_slabs
;
3888 num_objs
= num_slabs
* cachep
->num
;
3889 if (num_objs
- active_objs
!= free_objects
&& !error
)
3890 error
= "free_objects accounting error";
3892 name
= cachep
->name
;
3894 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
3896 seq_printf(m
, "%-17s %6lu %6lu %6u %4u %4d",
3897 name
, active_objs
, num_objs
, cachep
->buffer_size
,
3898 cachep
->num
, (1 << cachep
->gfporder
));
3899 seq_printf(m
, " : tunables %4u %4u %4u",
3900 cachep
->limit
, cachep
->batchcount
, cachep
->shared
);
3901 seq_printf(m
, " : slabdata %6lu %6lu %6lu",
3902 active_slabs
, num_slabs
, shared_avail
);
3905 unsigned long high
= cachep
->high_mark
;
3906 unsigned long allocs
= cachep
->num_allocations
;
3907 unsigned long grown
= cachep
->grown
;
3908 unsigned long reaped
= cachep
->reaped
;
3909 unsigned long errors
= cachep
->errors
;
3910 unsigned long max_freeable
= cachep
->max_freeable
;
3911 unsigned long node_allocs
= cachep
->node_allocs
;
3912 unsigned long node_frees
= cachep
->node_frees
;
3913 unsigned long overflows
= cachep
->node_overflow
;
3915 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu \
3916 %4lu %4lu %4lu %4lu %4lu", allocs
, high
, grown
,
3917 reaped
, errors
, max_freeable
, node_allocs
,
3918 node_frees
, overflows
);
3922 unsigned long allochit
= atomic_read(&cachep
->allochit
);
3923 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
3924 unsigned long freehit
= atomic_read(&cachep
->freehit
);
3925 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
3927 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
3928 allochit
, allocmiss
, freehit
, freemiss
);
3936 * slabinfo_op - iterator that generates /proc/slabinfo
3945 * num-pages-per-slab
3946 * + further values on SMP and with statistics enabled
3949 struct seq_operations slabinfo_op
= {
3956 #define MAX_SLABINFO_WRITE 128
3958 * slabinfo_write - Tuning for the slab allocator
3960 * @buffer: user buffer
3961 * @count: data length
3964 ssize_t
slabinfo_write(struct file
*file
, const char __user
* buffer
,
3965 size_t count
, loff_t
*ppos
)
3967 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
3968 int limit
, batchcount
, shared
, res
;
3969 struct list_head
*p
;
3971 if (count
> MAX_SLABINFO_WRITE
)
3973 if (copy_from_user(&kbuf
, buffer
, count
))
3975 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
3977 tmp
= strchr(kbuf
, ' ');
3982 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
3985 /* Find the cache in the chain of caches. */
3986 mutex_lock(&cache_chain_mutex
);
3988 list_for_each(p
, &cache_chain
) {
3989 struct kmem_cache
*cachep
;
3991 cachep
= list_entry(p
, struct kmem_cache
, next
);
3992 if (!strcmp(cachep
->name
, kbuf
)) {
3993 if (limit
< 1 || batchcount
< 1 ||
3994 batchcount
> limit
|| shared
< 0) {
3997 res
= do_tune_cpucache(cachep
, limit
,
3998 batchcount
, shared
);
4003 mutex_unlock(&cache_chain_mutex
);
4009 #ifdef CONFIG_DEBUG_SLAB_LEAK
4011 static void *leaks_start(struct seq_file
*m
, loff_t
*pos
)
4014 struct list_head
*p
;
4016 mutex_lock(&cache_chain_mutex
);
4017 p
= cache_chain
.next
;
4020 if (p
== &cache_chain
)
4023 return list_entry(p
, struct kmem_cache
, next
);
4026 static inline int add_caller(unsigned long *n
, unsigned long v
)
4036 unsigned long *q
= p
+ 2 * i
;
4050 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4056 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
, struct slab
*s
)
4062 for (i
= 0, p
= s
->s_mem
; i
< c
->num
; i
++, p
+= c
->buffer_size
) {
4063 if (slab_bufctl(s
)[i
] != BUFCTL_ACTIVE
)
4065 if (!add_caller(n
, (unsigned long)*dbg_userword(c
, p
)))
4070 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4072 #ifdef CONFIG_KALLSYMS
4075 unsigned long offset
, size
;
4076 char namebuf
[KSYM_NAME_LEN
+1];
4078 name
= kallsyms_lookup(address
, &size
, &offset
, &modname
, namebuf
);
4081 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4083 seq_printf(m
, " [%s]", modname
);
4087 seq_printf(m
, "%p", (void *)address
);
4090 static int leaks_show(struct seq_file
*m
, void *p
)
4092 struct kmem_cache
*cachep
= p
;
4093 struct list_head
*q
;
4095 struct kmem_list3
*l3
;
4097 unsigned long *n
= m
->private;
4101 if (!(cachep
->flags
& SLAB_STORE_USER
))
4103 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4106 /* OK, we can do it */
4110 for_each_online_node(node
) {
4111 l3
= cachep
->nodelists
[node
];
4116 spin_lock_irq(&l3
->list_lock
);
4118 list_for_each(q
, &l3
->slabs_full
) {
4119 slabp
= list_entry(q
, struct slab
, list
);
4120 handle_slab(n
, cachep
, slabp
);
4122 list_for_each(q
, &l3
->slabs_partial
) {
4123 slabp
= list_entry(q
, struct slab
, list
);
4124 handle_slab(n
, cachep
, slabp
);
4126 spin_unlock_irq(&l3
->list_lock
);
4128 name
= cachep
->name
;
4130 /* Increase the buffer size */
4131 mutex_unlock(&cache_chain_mutex
);
4132 m
->private = kzalloc(n
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4134 /* Too bad, we are really out */
4136 mutex_lock(&cache_chain_mutex
);
4139 *(unsigned long *)m
->private = n
[0] * 2;
4141 mutex_lock(&cache_chain_mutex
);
4142 /* Now make sure this entry will be retried */
4146 for (i
= 0; i
< n
[1]; i
++) {
4147 seq_printf(m
, "%s: %lu ", name
, n
[2*i
+3]);
4148 show_symbol(m
, n
[2*i
+2]);
4154 struct seq_operations slabstats_op
= {
4155 .start
= leaks_start
,
4164 * ksize - get the actual amount of memory allocated for a given object
4165 * @objp: Pointer to the object
4167 * kmalloc may internally round up allocations and return more memory
4168 * than requested. ksize() can be used to determine the actual amount of
4169 * memory allocated. The caller may use this additional memory, even though
4170 * a smaller amount of memory was initially specified with the kmalloc call.
4171 * The caller must guarantee that objp points to a valid object previously
4172 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4173 * must not be freed during the duration of the call.
4175 unsigned int ksize(const void *objp
)
4177 if (unlikely(objp
== NULL
))
4180 return obj_size(virt_to_cache(objp
));