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/poison.h>
93 #include <linux/swap.h>
94 #include <linux/cache.h>
95 #include <linux/interrupt.h>
96 #include <linux/init.h>
97 #include <linux/compiler.h>
98 #include <linux/cpuset.h>
99 #include <linux/seq_file.h>
100 #include <linux/notifier.h>
101 #include <linux/kallsyms.h>
102 #include <linux/cpu.h>
103 #include <linux/sysctl.h>
104 #include <linux/module.h>
105 #include <linux/rcupdate.h>
106 #include <linux/string.h>
107 #include <linux/nodemask.h>
108 #include <linux/mempolicy.h>
109 #include <linux/mutex.h>
110 #include <linux/rtmutex.h>
112 #include <asm/uaccess.h>
113 #include <asm/cacheflush.h>
114 #include <asm/tlbflush.h>
115 #include <asm/page.h>
118 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
119 * SLAB_RED_ZONE & SLAB_POISON.
120 * 0 for faster, smaller code (especially in the critical paths).
122 * STATS - 1 to collect stats for /proc/slabinfo.
123 * 0 for faster, smaller code (especially in the critical paths).
125 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
128 #ifdef CONFIG_DEBUG_SLAB
131 #define FORCED_DEBUG 1
135 #define FORCED_DEBUG 0
138 /* Shouldn't this be in a header file somewhere? */
139 #define BYTES_PER_WORD sizeof(void *)
141 #ifndef cache_line_size
142 #define cache_line_size() L1_CACHE_BYTES
145 #ifndef ARCH_KMALLOC_MINALIGN
147 * Enforce a minimum alignment for the kmalloc caches.
148 * Usually, the kmalloc caches are cache_line_size() aligned, except when
149 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
150 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
151 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
152 * Note that this flag disables some debug features.
154 #define ARCH_KMALLOC_MINALIGN 0
157 #ifndef ARCH_SLAB_MINALIGN
159 * Enforce a minimum alignment for all caches.
160 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
161 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
162 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
163 * some debug features.
165 #define ARCH_SLAB_MINALIGN 0
168 #ifndef ARCH_KMALLOC_FLAGS
169 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
172 /* Legal flag mask for kmem_cache_create(). */
174 # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
175 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
177 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
178 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
179 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
181 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
182 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
183 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
184 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
190 * Bufctl's are used for linking objs within a slab
193 * This implementation relies on "struct page" for locating the cache &
194 * slab an object belongs to.
195 * This allows the bufctl structure to be small (one int), but limits
196 * the number of objects a slab (not a cache) can contain when off-slab
197 * bufctls are used. The limit is the size of the largest general cache
198 * that does not use off-slab slabs.
199 * For 32bit archs with 4 kB pages, is this 56.
200 * This is not serious, as it is only for large objects, when it is unwise
201 * to have too many per slab.
202 * Note: This limit can be raised by introducing a general cache whose size
203 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
206 typedef unsigned int kmem_bufctl_t
;
207 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
208 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
209 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
210 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
215 * Manages the objs in a slab. Placed either at the beginning of mem allocated
216 * for a slab, or allocated from an general cache.
217 * Slabs are chained into three list: fully used, partial, fully free slabs.
220 struct list_head list
;
221 unsigned long colouroff
;
222 void *s_mem
; /* including colour offset */
223 unsigned int inuse
; /* num of objs active in slab */
225 unsigned short nodeid
;
231 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
232 * arrange for kmem_freepages to be called via RCU. This is useful if
233 * we need to approach a kernel structure obliquely, from its address
234 * obtained without the usual locking. We can lock the structure to
235 * stabilize it and check it's still at the given address, only if we
236 * can be sure that the memory has not been meanwhile reused for some
237 * other kind of object (which our subsystem's lock might corrupt).
239 * rcu_read_lock before reading the address, then rcu_read_unlock after
240 * taking the spinlock within the structure expected at that address.
242 * We assume struct slab_rcu can overlay struct slab when destroying.
245 struct rcu_head head
;
246 struct kmem_cache
*cachep
;
254 * - LIFO ordering, to hand out cache-warm objects from _alloc
255 * - reduce the number of linked list operations
256 * - reduce spinlock operations
258 * The limit is stored in the per-cpu structure to reduce the data cache
265 unsigned int batchcount
;
266 unsigned int touched
;
269 * Must have this definition in here for the proper
270 * alignment of array_cache. Also simplifies accessing
272 * [0] is for gcc 2.95. It should really be [].
277 * bootstrap: The caches do not work without cpuarrays anymore, but the
278 * cpuarrays are allocated from the generic caches...
280 #define BOOT_CPUCACHE_ENTRIES 1
281 struct arraycache_init
{
282 struct array_cache cache
;
283 void *entries
[BOOT_CPUCACHE_ENTRIES
];
287 * The slab lists for all objects.
290 struct list_head slabs_partial
; /* partial list first, better asm code */
291 struct list_head slabs_full
;
292 struct list_head slabs_free
;
293 unsigned long free_objects
;
294 unsigned int free_limit
;
295 unsigned int colour_next
; /* Per-node cache coloring */
296 spinlock_t list_lock
;
297 struct array_cache
*shared
; /* shared per node */
298 struct array_cache
**alien
; /* on other nodes */
299 unsigned long next_reap
; /* updated without locking */
300 int free_touched
; /* updated without locking */
304 * Need this for bootstrapping a per node allocator.
306 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
307 struct kmem_list3 __initdata initkmem_list3
[NUM_INIT_LISTS
];
308 #define CACHE_CACHE 0
310 #define SIZE_L3 (1 + MAX_NUMNODES)
312 static int drain_freelist(struct kmem_cache
*cache
,
313 struct kmem_list3
*l3
, int tofree
);
314 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
316 static void enable_cpucache(struct kmem_cache
*cachep
);
317 static void cache_reap(void *unused
);
320 * This function must be completely optimized away if a constant is passed to
321 * it. Mostly the same as what is in linux/slab.h except it returns an index.
323 static __always_inline
int index_of(const size_t size
)
325 extern void __bad_size(void);
327 if (__builtin_constant_p(size
)) {
335 #include "linux/kmalloc_sizes.h"
343 static int slab_early_init
= 1;
345 #define INDEX_AC index_of(sizeof(struct arraycache_init))
346 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
348 static void kmem_list3_init(struct kmem_list3
*parent
)
350 INIT_LIST_HEAD(&parent
->slabs_full
);
351 INIT_LIST_HEAD(&parent
->slabs_partial
);
352 INIT_LIST_HEAD(&parent
->slabs_free
);
353 parent
->shared
= NULL
;
354 parent
->alien
= NULL
;
355 parent
->colour_next
= 0;
356 spin_lock_init(&parent
->list_lock
);
357 parent
->free_objects
= 0;
358 parent
->free_touched
= 0;
361 #define MAKE_LIST(cachep, listp, slab, nodeid) \
363 INIT_LIST_HEAD(listp); \
364 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
367 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
369 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
370 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
371 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
381 /* 1) per-cpu data, touched during every alloc/free */
382 struct array_cache
*array
[NR_CPUS
];
383 /* 2) Cache tunables. Protected by cache_chain_mutex */
384 unsigned int batchcount
;
388 unsigned int buffer_size
;
389 /* 3) touched by every alloc & free from the backend */
390 struct kmem_list3
*nodelists
[MAX_NUMNODES
];
392 unsigned int flags
; /* constant flags */
393 unsigned int num
; /* # of objs per slab */
395 /* 4) cache_grow/shrink */
396 /* order of pgs per slab (2^n) */
397 unsigned int gfporder
;
399 /* force GFP flags, e.g. GFP_DMA */
402 size_t colour
; /* cache colouring range */
403 unsigned int colour_off
; /* colour offset */
404 struct kmem_cache
*slabp_cache
;
405 unsigned int slab_size
;
406 unsigned int dflags
; /* dynamic flags */
408 /* constructor func */
409 void (*ctor
) (void *, struct kmem_cache
*, unsigned long);
411 /* de-constructor func */
412 void (*dtor
) (void *, struct kmem_cache
*, unsigned long);
414 /* 5) cache creation/removal */
416 struct list_head next
;
420 unsigned long num_active
;
421 unsigned long num_allocations
;
422 unsigned long high_mark
;
424 unsigned long reaped
;
425 unsigned long errors
;
426 unsigned long max_freeable
;
427 unsigned long node_allocs
;
428 unsigned long node_frees
;
429 unsigned long node_overflow
;
437 * If debugging is enabled, then the allocator can add additional
438 * fields and/or padding to every object. buffer_size contains the total
439 * object size including these internal fields, the following two
440 * variables contain the offset to the user object and its size.
447 #define CFLGS_OFF_SLAB (0x80000000UL)
448 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
450 #define BATCHREFILL_LIMIT 16
452 * Optimization question: fewer reaps means less probability for unnessary
453 * cpucache drain/refill cycles.
455 * OTOH the cpuarrays can contain lots of objects,
456 * which could lock up otherwise freeable slabs.
458 #define REAPTIMEOUT_CPUC (2*HZ)
459 #define REAPTIMEOUT_LIST3 (4*HZ)
462 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
463 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
464 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
465 #define STATS_INC_GROWN(x) ((x)->grown++)
466 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
467 #define STATS_SET_HIGH(x) \
469 if ((x)->num_active > (x)->high_mark) \
470 (x)->high_mark = (x)->num_active; \
472 #define STATS_INC_ERR(x) ((x)->errors++)
473 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
474 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
475 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
476 #define STATS_SET_FREEABLE(x, i) \
478 if ((x)->max_freeable < i) \
479 (x)->max_freeable = i; \
481 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
482 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
483 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
484 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
486 #define STATS_INC_ACTIVE(x) do { } while (0)
487 #define STATS_DEC_ACTIVE(x) do { } while (0)
488 #define STATS_INC_ALLOCED(x) do { } while (0)
489 #define STATS_INC_GROWN(x) do { } while (0)
490 #define STATS_ADD_REAPED(x,y) do { } while (0)
491 #define STATS_SET_HIGH(x) do { } while (0)
492 #define STATS_INC_ERR(x) do { } while (0)
493 #define STATS_INC_NODEALLOCS(x) do { } while (0)
494 #define STATS_INC_NODEFREES(x) do { } while (0)
495 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
496 #define STATS_SET_FREEABLE(x, i) do { } while (0)
497 #define STATS_INC_ALLOCHIT(x) do { } while (0)
498 #define STATS_INC_ALLOCMISS(x) do { } while (0)
499 #define STATS_INC_FREEHIT(x) do { } while (0)
500 #define STATS_INC_FREEMISS(x) do { } while (0)
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 BUG_ON(!PageSlab(page
));
596 return (struct kmem_cache
*)page
->lru
.next
;
599 static inline void page_set_slab(struct page
*page
, struct slab
*slab
)
601 page
->lru
.prev
= (struct list_head
*)slab
;
604 static inline struct slab
*page_get_slab(struct page
*page
)
606 if (unlikely(PageCompound(page
)))
607 page
= (struct page
*)page_private(page
);
608 BUG_ON(!PageSlab(page
));
609 return (struct slab
*)page
->lru
.prev
;
612 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
614 struct page
*page
= virt_to_page(obj
);
615 return page_get_cache(page
);
618 static inline struct slab
*virt_to_slab(const void *obj
)
620 struct page
*page
= virt_to_page(obj
);
621 return page_get_slab(page
);
624 static inline void *index_to_obj(struct kmem_cache
*cache
, struct slab
*slab
,
627 return slab
->s_mem
+ cache
->buffer_size
* idx
;
630 static inline unsigned int obj_to_index(struct kmem_cache
*cache
,
631 struct slab
*slab
, void *obj
)
633 return (unsigned)(obj
- slab
->s_mem
) / cache
->buffer_size
;
637 * These are the default caches for kmalloc. Custom caches can have other sizes.
639 struct cache_sizes malloc_sizes
[] = {
640 #define CACHE(x) { .cs_size = (x) },
641 #include <linux/kmalloc_sizes.h>
645 EXPORT_SYMBOL(malloc_sizes
);
647 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
653 static struct cache_names __initdata cache_names
[] = {
654 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
655 #include <linux/kmalloc_sizes.h>
660 static struct arraycache_init initarray_cache __initdata
=
661 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
662 static struct arraycache_init initarray_generic
=
663 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
665 /* internal cache of cache description objs */
666 static struct kmem_cache cache_cache
= {
668 .limit
= BOOT_CPUCACHE_ENTRIES
,
670 .buffer_size
= sizeof(struct kmem_cache
),
671 .name
= "kmem_cache",
673 .obj_size
= sizeof(struct kmem_cache
),
677 /* Guard access to the cache-chain. */
678 static DEFINE_MUTEX(cache_chain_mutex
);
679 static struct list_head cache_chain
;
682 * vm_enough_memory() looks at this to determine how many slab-allocated pages
683 * are possibly freeable under pressure
685 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
687 atomic_t slab_reclaim_pages
;
690 * chicken and egg problem: delay the per-cpu array allocation
691 * until the general caches are up.
701 * used by boot code to determine if it can use slab based allocator
703 int slab_is_available(void)
705 return g_cpucache_up
== FULL
;
708 static DEFINE_PER_CPU(struct work_struct
, reap_work
);
710 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
712 return cachep
->array
[smp_processor_id()];
715 static inline struct kmem_cache
*__find_general_cachep(size_t size
,
718 struct cache_sizes
*csizep
= malloc_sizes
;
721 /* This happens if someone tries to call
722 * kmem_cache_create(), or __kmalloc(), before
723 * the generic caches are initialized.
725 BUG_ON(malloc_sizes
[INDEX_AC
].cs_cachep
== NULL
);
727 while (size
> csizep
->cs_size
)
731 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
732 * has cs_{dma,}cachep==NULL. Thus no special case
733 * for large kmalloc calls required.
735 if (unlikely(gfpflags
& GFP_DMA
))
736 return csizep
->cs_dmacachep
;
737 return csizep
->cs_cachep
;
740 struct kmem_cache
*kmem_find_general_cachep(size_t size
, gfp_t gfpflags
)
742 return __find_general_cachep(size
, gfpflags
);
744 EXPORT_SYMBOL(kmem_find_general_cachep
);
746 static size_t slab_mgmt_size(size_t nr_objs
, size_t align
)
748 return ALIGN(sizeof(struct slab
)+nr_objs
*sizeof(kmem_bufctl_t
), align
);
752 * Calculate the number of objects and left-over bytes for a given buffer size.
754 static void cache_estimate(unsigned long gfporder
, size_t buffer_size
,
755 size_t align
, int flags
, size_t *left_over
,
760 size_t slab_size
= PAGE_SIZE
<< gfporder
;
763 * The slab management structure can be either off the slab or
764 * on it. For the latter case, the memory allocated for a
768 * - One kmem_bufctl_t for each object
769 * - Padding to respect alignment of @align
770 * - @buffer_size bytes for each object
772 * If the slab management structure is off the slab, then the
773 * alignment will already be calculated into the size. Because
774 * the slabs are all pages aligned, the objects will be at the
775 * correct alignment when allocated.
777 if (flags
& CFLGS_OFF_SLAB
) {
779 nr_objs
= slab_size
/ buffer_size
;
781 if (nr_objs
> SLAB_LIMIT
)
782 nr_objs
= SLAB_LIMIT
;
785 * Ignore padding for the initial guess. The padding
786 * is at most @align-1 bytes, and @buffer_size is at
787 * least @align. In the worst case, this result will
788 * be one greater than the number of objects that fit
789 * into the memory allocation when taking the padding
792 nr_objs
= (slab_size
- sizeof(struct slab
)) /
793 (buffer_size
+ sizeof(kmem_bufctl_t
));
796 * This calculated number will be either the right
797 * amount, or one greater than what we want.
799 if (slab_mgmt_size(nr_objs
, align
) + nr_objs
*buffer_size
803 if (nr_objs
> SLAB_LIMIT
)
804 nr_objs
= SLAB_LIMIT
;
806 mgmt_size
= slab_mgmt_size(nr_objs
, align
);
809 *left_over
= slab_size
- nr_objs
*buffer_size
- mgmt_size
;
812 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
814 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
817 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
818 function
, cachep
->name
, msg
);
824 * Special reaping functions for NUMA systems called from cache_reap().
825 * These take care of doing round robin flushing of alien caches (containing
826 * objects freed on different nodes from which they were allocated) and the
827 * flushing of remote pcps by calling drain_node_pages.
829 static DEFINE_PER_CPU(unsigned long, reap_node
);
831 static void init_reap_node(int cpu
)
835 node
= next_node(cpu_to_node(cpu
), node_online_map
);
836 if (node
== MAX_NUMNODES
)
837 node
= first_node(node_online_map
);
839 __get_cpu_var(reap_node
) = node
;
842 static void next_reap_node(void)
844 int node
= __get_cpu_var(reap_node
);
847 * Also drain per cpu pages on remote zones
849 if (node
!= numa_node_id())
850 drain_node_pages(node
);
852 node
= next_node(node
, node_online_map
);
853 if (unlikely(node
>= MAX_NUMNODES
))
854 node
= first_node(node_online_map
);
855 __get_cpu_var(reap_node
) = node
;
859 #define init_reap_node(cpu) do { } while (0)
860 #define next_reap_node(void) do { } while (0)
864 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
865 * via the workqueue/eventd.
866 * Add the CPU number into the expiration time to minimize the possibility of
867 * the CPUs getting into lockstep and contending for the global cache chain
870 static void __devinit
start_cpu_timer(int cpu
)
872 struct work_struct
*reap_work
= &per_cpu(reap_work
, cpu
);
875 * When this gets called from do_initcalls via cpucache_init(),
876 * init_workqueues() has already run, so keventd will be setup
879 if (keventd_up() && reap_work
->func
== NULL
) {
881 INIT_WORK(reap_work
, cache_reap
, NULL
);
882 schedule_delayed_work_on(cpu
, reap_work
, HZ
+ 3 * cpu
);
886 static struct array_cache
*alloc_arraycache(int node
, int entries
,
889 int memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
890 struct array_cache
*nc
= NULL
;
892 nc
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
896 nc
->batchcount
= batchcount
;
898 spin_lock_init(&nc
->lock
);
904 * Transfer objects in one arraycache to another.
905 * Locking must be handled by the caller.
907 * Return the number of entries transferred.
909 static int transfer_objects(struct array_cache
*to
,
910 struct array_cache
*from
, unsigned int max
)
912 /* Figure out how many entries to transfer */
913 int nr
= min(min(from
->avail
, max
), to
->limit
- to
->avail
);
918 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
928 static void *__cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
929 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
931 static struct array_cache
**alloc_alien_cache(int node
, int limit
)
933 struct array_cache
**ac_ptr
;
934 int memsize
= sizeof(void *) * MAX_NUMNODES
;
939 ac_ptr
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
942 if (i
== node
|| !node_online(i
)) {
946 ac_ptr
[i
] = alloc_arraycache(node
, limit
, 0xbaadf00d);
948 for (i
--; i
<= 0; i
--)
958 static void free_alien_cache(struct array_cache
**ac_ptr
)
969 static void __drain_alien_cache(struct kmem_cache
*cachep
,
970 struct array_cache
*ac
, int node
)
972 struct kmem_list3
*rl3
= cachep
->nodelists
[node
];
975 spin_lock(&rl3
->list_lock
);
977 * Stuff objects into the remote nodes shared array first.
978 * That way we could avoid the overhead of putting the objects
979 * into the free lists and getting them back later.
982 transfer_objects(rl3
->shared
, ac
, ac
->limit
);
984 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
986 spin_unlock(&rl3
->list_lock
);
991 * Called from cache_reap() to regularly drain alien caches round robin.
993 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_list3
*l3
)
995 int node
= __get_cpu_var(reap_node
);
998 struct array_cache
*ac
= l3
->alien
[node
];
1000 if (ac
&& ac
->avail
&& spin_trylock_irq(&ac
->lock
)) {
1001 __drain_alien_cache(cachep
, ac
, node
);
1002 spin_unlock_irq(&ac
->lock
);
1007 static void drain_alien_cache(struct kmem_cache
*cachep
,
1008 struct array_cache
**alien
)
1011 struct array_cache
*ac
;
1012 unsigned long flags
;
1014 for_each_online_node(i
) {
1017 spin_lock_irqsave(&ac
->lock
, flags
);
1018 __drain_alien_cache(cachep
, ac
, i
);
1019 spin_unlock_irqrestore(&ac
->lock
, flags
);
1024 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
,
1027 struct slab
*slabp
= virt_to_slab(objp
);
1028 int nodeid
= slabp
->nodeid
;
1029 struct kmem_list3
*l3
;
1030 struct array_cache
*alien
= NULL
;
1033 * Make sure we are not freeing a object from another node to the array
1034 * cache on this cpu.
1036 if (likely(slabp
->nodeid
== numa_node_id()))
1039 l3
= cachep
->nodelists
[numa_node_id()];
1040 STATS_INC_NODEFREES(cachep
);
1041 if (l3
->alien
&& l3
->alien
[nodeid
]) {
1042 alien
= l3
->alien
[nodeid
];
1043 spin_lock_nested(&alien
->lock
, nesting
);
1044 if (unlikely(alien
->avail
== alien
->limit
)) {
1045 STATS_INC_ACOVERFLOW(cachep
);
1046 __drain_alien_cache(cachep
, alien
, nodeid
);
1048 alien
->entry
[alien
->avail
++] = objp
;
1049 spin_unlock(&alien
->lock
);
1051 spin_lock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1052 free_block(cachep
, &objp
, 1, nodeid
);
1053 spin_unlock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1060 #define drain_alien_cache(cachep, alien) do { } while (0)
1061 #define reap_alien(cachep, l3) do { } while (0)
1063 static inline struct array_cache
**alloc_alien_cache(int node
, int limit
)
1065 return (struct array_cache
**) 0x01020304ul
;
1068 static inline void free_alien_cache(struct array_cache
**ac_ptr
)
1072 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
,
1080 static int __devinit
cpuup_callback(struct notifier_block
*nfb
,
1081 unsigned long action
, void *hcpu
)
1083 long cpu
= (long)hcpu
;
1084 struct kmem_cache
*cachep
;
1085 struct kmem_list3
*l3
= NULL
;
1086 int node
= cpu_to_node(cpu
);
1087 int memsize
= sizeof(struct kmem_list3
);
1090 case CPU_UP_PREPARE
:
1091 mutex_lock(&cache_chain_mutex
);
1093 * We need to do this right in the beginning since
1094 * alloc_arraycache's are going to use this list.
1095 * kmalloc_node allows us to add the slab to the right
1096 * kmem_list3 and not this cpu's kmem_list3
1099 list_for_each_entry(cachep
, &cache_chain
, next
) {
1101 * Set up the size64 kmemlist for cpu before we can
1102 * begin anything. Make sure some other cpu on this
1103 * node has not already allocated this
1105 if (!cachep
->nodelists
[node
]) {
1106 l3
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1109 kmem_list3_init(l3
);
1110 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
1111 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1114 * The l3s don't come and go as CPUs come and
1115 * go. cache_chain_mutex is sufficient
1118 cachep
->nodelists
[node
] = l3
;
1121 spin_lock_irq(&cachep
->nodelists
[node
]->list_lock
);
1122 cachep
->nodelists
[node
]->free_limit
=
1123 (1 + nr_cpus_node(node
)) *
1124 cachep
->batchcount
+ cachep
->num
;
1125 spin_unlock_irq(&cachep
->nodelists
[node
]->list_lock
);
1129 * Now we can go ahead with allocating the shared arrays and
1132 list_for_each_entry(cachep
, &cache_chain
, next
) {
1133 struct array_cache
*nc
;
1134 struct array_cache
*shared
;
1135 struct array_cache
**alien
;
1137 nc
= alloc_arraycache(node
, cachep
->limit
,
1138 cachep
->batchcount
);
1141 shared
= alloc_arraycache(node
,
1142 cachep
->shared
* cachep
->batchcount
,
1147 alien
= alloc_alien_cache(node
, cachep
->limit
);
1150 cachep
->array
[cpu
] = nc
;
1151 l3
= cachep
->nodelists
[node
];
1154 spin_lock_irq(&l3
->list_lock
);
1157 * We are serialised from CPU_DEAD or
1158 * CPU_UP_CANCELLED by the cpucontrol lock
1160 l3
->shared
= shared
;
1169 spin_unlock_irq(&l3
->list_lock
);
1171 free_alien_cache(alien
);
1173 mutex_unlock(&cache_chain_mutex
);
1176 start_cpu_timer(cpu
);
1178 #ifdef CONFIG_HOTPLUG_CPU
1181 * Even if all the cpus of a node are down, we don't free the
1182 * kmem_list3 of any cache. This to avoid a race between
1183 * cpu_down, and a kmalloc allocation from another cpu for
1184 * memory from the node of the cpu going down. The list3
1185 * structure is usually allocated from kmem_cache_create() and
1186 * gets destroyed at kmem_cache_destroy().
1189 case CPU_UP_CANCELED
:
1190 mutex_lock(&cache_chain_mutex
);
1191 list_for_each_entry(cachep
, &cache_chain
, next
) {
1192 struct array_cache
*nc
;
1193 struct array_cache
*shared
;
1194 struct array_cache
**alien
;
1197 mask
= node_to_cpumask(node
);
1198 /* cpu is dead; no one can alloc from it. */
1199 nc
= cachep
->array
[cpu
];
1200 cachep
->array
[cpu
] = NULL
;
1201 l3
= cachep
->nodelists
[node
];
1204 goto free_array_cache
;
1206 spin_lock_irq(&l3
->list_lock
);
1208 /* Free limit for this kmem_list3 */
1209 l3
->free_limit
-= cachep
->batchcount
;
1211 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
1213 if (!cpus_empty(mask
)) {
1214 spin_unlock_irq(&l3
->list_lock
);
1215 goto free_array_cache
;
1218 shared
= l3
->shared
;
1220 free_block(cachep
, l3
->shared
->entry
,
1221 l3
->shared
->avail
, node
);
1228 spin_unlock_irq(&l3
->list_lock
);
1232 drain_alien_cache(cachep
, alien
);
1233 free_alien_cache(alien
);
1239 * In the previous loop, all the objects were freed to
1240 * the respective cache's slabs, now we can go ahead and
1241 * shrink each nodelist to its limit.
1243 list_for_each_entry(cachep
, &cache_chain
, next
) {
1244 l3
= cachep
->nodelists
[node
];
1247 drain_freelist(cachep
, l3
, l3
->free_objects
);
1249 mutex_unlock(&cache_chain_mutex
);
1255 mutex_unlock(&cache_chain_mutex
);
1259 static struct notifier_block __cpuinitdata cpucache_notifier
= {
1260 &cpuup_callback
, NULL
, 0
1264 * swap the static kmem_list3 with kmalloced memory
1266 static void init_list(struct kmem_cache
*cachep
, struct kmem_list3
*list
,
1269 struct kmem_list3
*ptr
;
1271 BUG_ON(cachep
->nodelists
[nodeid
] != list
);
1272 ptr
= kmalloc_node(sizeof(struct kmem_list3
), GFP_KERNEL
, nodeid
);
1275 local_irq_disable();
1276 memcpy(ptr
, list
, sizeof(struct kmem_list3
));
1278 * Do not assume that spinlocks can be initialized via memcpy:
1280 spin_lock_init(&ptr
->list_lock
);
1282 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1283 cachep
->nodelists
[nodeid
] = ptr
;
1288 * Initialisation. Called after the page allocator have been initialised and
1289 * before smp_init().
1291 void __init
kmem_cache_init(void)
1294 struct cache_sizes
*sizes
;
1295 struct cache_names
*names
;
1299 for (i
= 0; i
< NUM_INIT_LISTS
; i
++) {
1300 kmem_list3_init(&initkmem_list3
[i
]);
1301 if (i
< MAX_NUMNODES
)
1302 cache_cache
.nodelists
[i
] = NULL
;
1306 * Fragmentation resistance on low memory - only use bigger
1307 * page orders on machines with more than 32MB of memory.
1309 if (num_physpages
> (32 << 20) >> PAGE_SHIFT
)
1310 slab_break_gfp_order
= BREAK_GFP_ORDER_HI
;
1312 /* Bootstrap is tricky, because several objects are allocated
1313 * from caches that do not exist yet:
1314 * 1) initialize the cache_cache cache: it contains the struct
1315 * kmem_cache structures of all caches, except cache_cache itself:
1316 * cache_cache is statically allocated.
1317 * Initially an __init data area is used for the head array and the
1318 * kmem_list3 structures, it's replaced with a kmalloc allocated
1319 * array at the end of the bootstrap.
1320 * 2) Create the first kmalloc cache.
1321 * The struct kmem_cache for the new cache is allocated normally.
1322 * An __init data area is used for the head array.
1323 * 3) Create the remaining kmalloc caches, with minimally sized
1325 * 4) Replace the __init data head arrays for cache_cache and the first
1326 * kmalloc cache with kmalloc allocated arrays.
1327 * 5) Replace the __init data for kmem_list3 for cache_cache and
1328 * the other cache's with kmalloc allocated memory.
1329 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1332 /* 1) create the cache_cache */
1333 INIT_LIST_HEAD(&cache_chain
);
1334 list_add(&cache_cache
.next
, &cache_chain
);
1335 cache_cache
.colour_off
= cache_line_size();
1336 cache_cache
.array
[smp_processor_id()] = &initarray_cache
.cache
;
1337 cache_cache
.nodelists
[numa_node_id()] = &initkmem_list3
[CACHE_CACHE
];
1339 cache_cache
.buffer_size
= ALIGN(cache_cache
.buffer_size
,
1342 for (order
= 0; order
< MAX_ORDER
; order
++) {
1343 cache_estimate(order
, cache_cache
.buffer_size
,
1344 cache_line_size(), 0, &left_over
, &cache_cache
.num
);
1345 if (cache_cache
.num
)
1348 BUG_ON(!cache_cache
.num
);
1349 cache_cache
.gfporder
= order
;
1350 cache_cache
.colour
= left_over
/ cache_cache
.colour_off
;
1351 cache_cache
.slab_size
= ALIGN(cache_cache
.num
* sizeof(kmem_bufctl_t
) +
1352 sizeof(struct slab
), cache_line_size());
1354 /* 2+3) create the kmalloc caches */
1355 sizes
= malloc_sizes
;
1356 names
= cache_names
;
1359 * Initialize the caches that provide memory for the array cache and the
1360 * kmem_list3 structures first. Without this, further allocations will
1364 sizes
[INDEX_AC
].cs_cachep
= kmem_cache_create(names
[INDEX_AC
].name
,
1365 sizes
[INDEX_AC
].cs_size
,
1366 ARCH_KMALLOC_MINALIGN
,
1367 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1370 if (INDEX_AC
!= INDEX_L3
) {
1371 sizes
[INDEX_L3
].cs_cachep
=
1372 kmem_cache_create(names
[INDEX_L3
].name
,
1373 sizes
[INDEX_L3
].cs_size
,
1374 ARCH_KMALLOC_MINALIGN
,
1375 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1379 slab_early_init
= 0;
1381 while (sizes
->cs_size
!= ULONG_MAX
) {
1383 * For performance, all the general caches are L1 aligned.
1384 * This should be particularly beneficial on SMP boxes, as it
1385 * eliminates "false sharing".
1386 * Note for systems short on memory removing the alignment will
1387 * allow tighter packing of the smaller caches.
1389 if (!sizes
->cs_cachep
) {
1390 sizes
->cs_cachep
= kmem_cache_create(names
->name
,
1392 ARCH_KMALLOC_MINALIGN
,
1393 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1397 sizes
->cs_dmacachep
= kmem_cache_create(names
->name_dma
,
1399 ARCH_KMALLOC_MINALIGN
,
1400 ARCH_KMALLOC_FLAGS
|SLAB_CACHE_DMA
|
1406 /* 4) Replace the bootstrap head arrays */
1408 struct array_cache
*ptr
;
1410 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1412 local_irq_disable();
1413 BUG_ON(cpu_cache_get(&cache_cache
) != &initarray_cache
.cache
);
1414 memcpy(ptr
, cpu_cache_get(&cache_cache
),
1415 sizeof(struct arraycache_init
));
1417 * Do not assume that spinlocks can be initialized via memcpy:
1419 spin_lock_init(&ptr
->lock
);
1421 cache_cache
.array
[smp_processor_id()] = ptr
;
1424 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1426 local_irq_disable();
1427 BUG_ON(cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
)
1428 != &initarray_generic
.cache
);
1429 memcpy(ptr
, cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
),
1430 sizeof(struct arraycache_init
));
1432 * Do not assume that spinlocks can be initialized via memcpy:
1434 spin_lock_init(&ptr
->lock
);
1436 malloc_sizes
[INDEX_AC
].cs_cachep
->array
[smp_processor_id()] =
1440 /* 5) Replace the bootstrap kmem_list3's */
1443 /* Replace the static kmem_list3 structures for the boot cpu */
1444 init_list(&cache_cache
, &initkmem_list3
[CACHE_CACHE
],
1447 for_each_online_node(node
) {
1448 init_list(malloc_sizes
[INDEX_AC
].cs_cachep
,
1449 &initkmem_list3
[SIZE_AC
+ node
], node
);
1451 if (INDEX_AC
!= INDEX_L3
) {
1452 init_list(malloc_sizes
[INDEX_L3
].cs_cachep
,
1453 &initkmem_list3
[SIZE_L3
+ node
],
1459 /* 6) resize the head arrays to their final sizes */
1461 struct kmem_cache
*cachep
;
1462 mutex_lock(&cache_chain_mutex
);
1463 list_for_each_entry(cachep
, &cache_chain
, next
)
1464 enable_cpucache(cachep
);
1465 mutex_unlock(&cache_chain_mutex
);
1469 g_cpucache_up
= FULL
;
1472 * Register a cpu startup notifier callback that initializes
1473 * cpu_cache_get for all new cpus
1475 register_cpu_notifier(&cpucache_notifier
);
1478 * The reap timers are started later, with a module init call: That part
1479 * of the kernel is not yet operational.
1483 static int __init
cpucache_init(void)
1488 * Register the timers that return unneeded pages to the page allocator
1490 for_each_online_cpu(cpu
)
1491 start_cpu_timer(cpu
);
1494 __initcall(cpucache_init
);
1497 * Interface to system's page allocator. No need to hold the cache-lock.
1499 * If we requested dmaable memory, we will get it. Even if we
1500 * did not request dmaable memory, we might get it, but that
1501 * would be relatively rare and ignorable.
1503 static void *kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
1511 * Nommu uses slab's for process anonymous memory allocations, and thus
1512 * requires __GFP_COMP to properly refcount higher order allocations
1514 flags
|= __GFP_COMP
;
1516 flags
|= cachep
->gfpflags
;
1518 page
= alloc_pages_node(nodeid
, flags
, cachep
->gfporder
);
1522 nr_pages
= (1 << cachep
->gfporder
);
1523 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1524 atomic_add(nr_pages
, &slab_reclaim_pages
);
1525 add_zone_page_state(page_zone(page
), NR_SLAB
, nr_pages
);
1526 for (i
= 0; i
< nr_pages
; i
++)
1527 __SetPageSlab(page
+ i
);
1528 return page_address(page
);
1532 * Interface to system's page release.
1534 static void kmem_freepages(struct kmem_cache
*cachep
, void *addr
)
1536 unsigned long i
= (1 << cachep
->gfporder
);
1537 struct page
*page
= virt_to_page(addr
);
1538 const unsigned long nr_freed
= i
;
1540 sub_zone_page_state(page_zone(page
), NR_SLAB
, nr_freed
);
1542 BUG_ON(!PageSlab(page
));
1543 __ClearPageSlab(page
);
1546 if (current
->reclaim_state
)
1547 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1548 free_pages((unsigned long)addr
, cachep
->gfporder
);
1549 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1550 atomic_sub(1 << cachep
->gfporder
, &slab_reclaim_pages
);
1553 static void kmem_rcu_free(struct rcu_head
*head
)
1555 struct slab_rcu
*slab_rcu
= (struct slab_rcu
*)head
;
1556 struct kmem_cache
*cachep
= slab_rcu
->cachep
;
1558 kmem_freepages(cachep
, slab_rcu
->addr
);
1559 if (OFF_SLAB(cachep
))
1560 kmem_cache_free(cachep
->slabp_cache
, slab_rcu
);
1565 #ifdef CONFIG_DEBUG_PAGEALLOC
1566 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1567 unsigned long caller
)
1569 int size
= obj_size(cachep
);
1571 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1573 if (size
< 5 * sizeof(unsigned long))
1576 *addr
++ = 0x12345678;
1578 *addr
++ = smp_processor_id();
1579 size
-= 3 * sizeof(unsigned long);
1581 unsigned long *sptr
= &caller
;
1582 unsigned long svalue
;
1584 while (!kstack_end(sptr
)) {
1586 if (kernel_text_address(svalue
)) {
1588 size
-= sizeof(unsigned long);
1589 if (size
<= sizeof(unsigned long))
1595 *addr
++ = 0x87654321;
1599 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1601 int size
= obj_size(cachep
);
1602 addr
= &((char *)addr
)[obj_offset(cachep
)];
1604 memset(addr
, val
, size
);
1605 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1608 static void dump_line(char *data
, int offset
, int limit
)
1611 printk(KERN_ERR
"%03x:", offset
);
1612 for (i
= 0; i
< limit
; i
++)
1613 printk(" %02x", (unsigned char)data
[offset
+ i
]);
1620 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1625 if (cachep
->flags
& SLAB_RED_ZONE
) {
1626 printk(KERN_ERR
"Redzone: 0x%lx/0x%lx.\n",
1627 *dbg_redzone1(cachep
, objp
),
1628 *dbg_redzone2(cachep
, objp
));
1631 if (cachep
->flags
& SLAB_STORE_USER
) {
1632 printk(KERN_ERR
"Last user: [<%p>]",
1633 *dbg_userword(cachep
, objp
));
1634 print_symbol("(%s)",
1635 (unsigned long)*dbg_userword(cachep
, objp
));
1638 realobj
= (char *)objp
+ obj_offset(cachep
);
1639 size
= obj_size(cachep
);
1640 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1643 if (i
+ limit
> size
)
1645 dump_line(realobj
, i
, limit
);
1649 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1655 realobj
= (char *)objp
+ obj_offset(cachep
);
1656 size
= obj_size(cachep
);
1658 for (i
= 0; i
< size
; i
++) {
1659 char exp
= POISON_FREE
;
1662 if (realobj
[i
] != exp
) {
1668 "Slab corruption: start=%p, len=%d\n",
1670 print_objinfo(cachep
, objp
, 0);
1672 /* Hexdump the affected line */
1675 if (i
+ limit
> size
)
1677 dump_line(realobj
, i
, limit
);
1680 /* Limit to 5 lines */
1686 /* Print some data about the neighboring objects, if they
1689 struct slab
*slabp
= virt_to_slab(objp
);
1692 objnr
= obj_to_index(cachep
, slabp
, objp
);
1694 objp
= index_to_obj(cachep
, slabp
, objnr
- 1);
1695 realobj
= (char *)objp
+ obj_offset(cachep
);
1696 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
1698 print_objinfo(cachep
, objp
, 2);
1700 if (objnr
+ 1 < cachep
->num
) {
1701 objp
= index_to_obj(cachep
, slabp
, objnr
+ 1);
1702 realobj
= (char *)objp
+ obj_offset(cachep
);
1703 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
1705 print_objinfo(cachep
, objp
, 2);
1713 * slab_destroy_objs - destroy a slab and its objects
1714 * @cachep: cache pointer being destroyed
1715 * @slabp: slab pointer being destroyed
1717 * Call the registered destructor for each object in a slab that is being
1720 static void slab_destroy_objs(struct kmem_cache
*cachep
, struct slab
*slabp
)
1723 for (i
= 0; i
< cachep
->num
; i
++) {
1724 void *objp
= index_to_obj(cachep
, slabp
, i
);
1726 if (cachep
->flags
& SLAB_POISON
) {
1727 #ifdef CONFIG_DEBUG_PAGEALLOC
1728 if (cachep
->buffer_size
% PAGE_SIZE
== 0 &&
1730 kernel_map_pages(virt_to_page(objp
),
1731 cachep
->buffer_size
/ PAGE_SIZE
, 1);
1733 check_poison_obj(cachep
, objp
);
1735 check_poison_obj(cachep
, objp
);
1738 if (cachep
->flags
& SLAB_RED_ZONE
) {
1739 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1740 slab_error(cachep
, "start of a freed object "
1742 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1743 slab_error(cachep
, "end of a freed object "
1746 if (cachep
->dtor
&& !(cachep
->flags
& SLAB_POISON
))
1747 (cachep
->dtor
) (objp
+ obj_offset(cachep
), cachep
, 0);
1751 static void slab_destroy_objs(struct kmem_cache
*cachep
, struct slab
*slabp
)
1755 for (i
= 0; i
< cachep
->num
; i
++) {
1756 void *objp
= index_to_obj(cachep
, slabp
, i
);
1757 (cachep
->dtor
) (objp
, cachep
, 0);
1763 static void __cache_free(struct kmem_cache
*cachep
, void *objp
, int nesting
);
1766 * slab_destroy - destroy and release all objects in a slab
1767 * @cachep: cache pointer being destroyed
1768 * @slabp: slab pointer being destroyed
1770 * Destroy all the objs in a slab, and release the mem back to the system.
1771 * Before calling the slab must have been unlinked from the cache. The
1772 * cache-lock is not held/needed.
1774 static void slab_destroy(struct kmem_cache
*cachep
, struct slab
*slabp
)
1776 void *addr
= slabp
->s_mem
- slabp
->colouroff
;
1778 slab_destroy_objs(cachep
, slabp
);
1779 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
1780 struct slab_rcu
*slab_rcu
;
1782 slab_rcu
= (struct slab_rcu
*)slabp
;
1783 slab_rcu
->cachep
= cachep
;
1784 slab_rcu
->addr
= addr
;
1785 call_rcu(&slab_rcu
->head
, kmem_rcu_free
);
1787 kmem_freepages(cachep
, addr
);
1788 if (OFF_SLAB(cachep
)) {
1789 unsigned long flags
;
1792 * lockdep: we may nest inside an already held
1793 * ac->lock, so pass in a nesting flag:
1795 local_irq_save(flags
);
1796 __cache_free(cachep
->slabp_cache
, slabp
, 1);
1797 local_irq_restore(flags
);
1803 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1804 * size of kmem_list3.
1806 static void set_up_list3s(struct kmem_cache
*cachep
, int index
)
1810 for_each_online_node(node
) {
1811 cachep
->nodelists
[node
] = &initkmem_list3
[index
+ node
];
1812 cachep
->nodelists
[node
]->next_reap
= jiffies
+
1814 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1819 * calculate_slab_order - calculate size (page order) of slabs
1820 * @cachep: pointer to the cache that is being created
1821 * @size: size of objects to be created in this cache.
1822 * @align: required alignment for the objects.
1823 * @flags: slab allocation flags
1825 * Also calculates the number of objects per slab.
1827 * This could be made much more intelligent. For now, try to avoid using
1828 * high order pages for slabs. When the gfp() functions are more friendly
1829 * towards high-order requests, this should be changed.
1831 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
1832 size_t size
, size_t align
, unsigned long flags
)
1834 unsigned long offslab_limit
;
1835 size_t left_over
= 0;
1838 for (gfporder
= 0; gfporder
<= MAX_GFP_ORDER
; gfporder
++) {
1842 cache_estimate(gfporder
, size
, align
, flags
, &remainder
, &num
);
1846 if (flags
& CFLGS_OFF_SLAB
) {
1848 * Max number of objs-per-slab for caches which
1849 * use off-slab slabs. Needed to avoid a possible
1850 * looping condition in cache_grow().
1852 offslab_limit
= size
- sizeof(struct slab
);
1853 offslab_limit
/= sizeof(kmem_bufctl_t
);
1855 if (num
> offslab_limit
)
1859 /* Found something acceptable - save it away */
1861 cachep
->gfporder
= gfporder
;
1862 left_over
= remainder
;
1865 * A VFS-reclaimable slab tends to have most allocations
1866 * as GFP_NOFS and we really don't want to have to be allocating
1867 * higher-order pages when we are unable to shrink dcache.
1869 if (flags
& SLAB_RECLAIM_ACCOUNT
)
1873 * Large number of objects is good, but very large slabs are
1874 * currently bad for the gfp()s.
1876 if (gfporder
>= slab_break_gfp_order
)
1880 * Acceptable internal fragmentation?
1882 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
1888 static void setup_cpu_cache(struct kmem_cache
*cachep
)
1890 if (g_cpucache_up
== FULL
) {
1891 enable_cpucache(cachep
);
1894 if (g_cpucache_up
== NONE
) {
1896 * Note: the first kmem_cache_create must create the cache
1897 * that's used by kmalloc(24), otherwise the creation of
1898 * further caches will BUG().
1900 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
1903 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
1904 * the first cache, then we need to set up all its list3s,
1905 * otherwise the creation of further caches will BUG().
1907 set_up_list3s(cachep
, SIZE_AC
);
1908 if (INDEX_AC
== INDEX_L3
)
1909 g_cpucache_up
= PARTIAL_L3
;
1911 g_cpucache_up
= PARTIAL_AC
;
1913 cachep
->array
[smp_processor_id()] =
1914 kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1916 if (g_cpucache_up
== PARTIAL_AC
) {
1917 set_up_list3s(cachep
, SIZE_L3
);
1918 g_cpucache_up
= PARTIAL_L3
;
1921 for_each_online_node(node
) {
1922 cachep
->nodelists
[node
] =
1923 kmalloc_node(sizeof(struct kmem_list3
),
1925 BUG_ON(!cachep
->nodelists
[node
]);
1926 kmem_list3_init(cachep
->nodelists
[node
]);
1930 cachep
->nodelists
[numa_node_id()]->next_reap
=
1931 jiffies
+ REAPTIMEOUT_LIST3
+
1932 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1934 cpu_cache_get(cachep
)->avail
= 0;
1935 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
1936 cpu_cache_get(cachep
)->batchcount
= 1;
1937 cpu_cache_get(cachep
)->touched
= 0;
1938 cachep
->batchcount
= 1;
1939 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
1943 * kmem_cache_create - Create a cache.
1944 * @name: A string which is used in /proc/slabinfo to identify this cache.
1945 * @size: The size of objects to be created in this cache.
1946 * @align: The required alignment for the objects.
1947 * @flags: SLAB flags
1948 * @ctor: A constructor for the objects.
1949 * @dtor: A destructor for the objects.
1951 * Returns a ptr to the cache on success, NULL on failure.
1952 * Cannot be called within a int, but can be interrupted.
1953 * The @ctor is run when new pages are allocated by the cache
1954 * and the @dtor is run before the pages are handed back.
1956 * @name must be valid until the cache is destroyed. This implies that
1957 * the module calling this has to destroy the cache before getting unloaded.
1961 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1962 * to catch references to uninitialised memory.
1964 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1965 * for buffer overruns.
1967 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1968 * cacheline. This can be beneficial if you're counting cycles as closely
1972 kmem_cache_create (const char *name
, size_t size
, size_t align
,
1973 unsigned long flags
,
1974 void (*ctor
)(void*, struct kmem_cache
*, unsigned long),
1975 void (*dtor
)(void*, struct kmem_cache
*, unsigned long))
1977 size_t left_over
, slab_size
, ralign
;
1978 struct kmem_cache
*cachep
= NULL
, *pc
;
1981 * Sanity checks... these are all serious usage bugs.
1983 if (!name
|| in_interrupt() || (size
< BYTES_PER_WORD
) ||
1984 (size
> (1 << MAX_OBJ_ORDER
) * PAGE_SIZE
) || (dtor
&& !ctor
)) {
1985 printk(KERN_ERR
"%s: Early error in slab %s\n", __FUNCTION__
,
1991 * Prevent CPUs from coming and going.
1992 * lock_cpu_hotplug() nests outside cache_chain_mutex
1996 mutex_lock(&cache_chain_mutex
);
1998 list_for_each_entry(pc
, &cache_chain
, next
) {
1999 mm_segment_t old_fs
= get_fs();
2004 * This happens when the module gets unloaded and doesn't
2005 * destroy its slab cache and no-one else reuses the vmalloc
2006 * area of the module. Print a warning.
2009 res
= __get_user(tmp
, pc
->name
);
2012 printk("SLAB: cache with size %d has lost its name\n",
2017 if (!strcmp(pc
->name
, name
)) {
2018 printk("kmem_cache_create: duplicate cache %s\n", name
);
2025 WARN_ON(strchr(name
, ' ')); /* It confuses parsers */
2026 if ((flags
& SLAB_DEBUG_INITIAL
) && !ctor
) {
2027 /* No constructor, but inital state check requested */
2028 printk(KERN_ERR
"%s: No con, but init state check "
2029 "requested - %s\n", __FUNCTION__
, name
);
2030 flags
&= ~SLAB_DEBUG_INITIAL
;
2034 * Enable redzoning and last user accounting, except for caches with
2035 * large objects, if the increased size would increase the object size
2036 * above the next power of two: caches with object sizes just above a
2037 * power of two have a significant amount of internal fragmentation.
2039 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + 3 * BYTES_PER_WORD
))
2040 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2041 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2042 flags
|= SLAB_POISON
;
2044 if (flags
& SLAB_DESTROY_BY_RCU
)
2045 BUG_ON(flags
& SLAB_POISON
);
2047 if (flags
& SLAB_DESTROY_BY_RCU
)
2051 * Always checks flags, a caller might be expecting debug support which
2054 BUG_ON(flags
& ~CREATE_MASK
);
2057 * Check that size is in terms of words. This is needed to avoid
2058 * unaligned accesses for some archs when redzoning is used, and makes
2059 * sure any on-slab bufctl's are also correctly aligned.
2061 if (size
& (BYTES_PER_WORD
- 1)) {
2062 size
+= (BYTES_PER_WORD
- 1);
2063 size
&= ~(BYTES_PER_WORD
- 1);
2066 /* calculate the final buffer alignment: */
2068 /* 1) arch recommendation: can be overridden for debug */
2069 if (flags
& SLAB_HWCACHE_ALIGN
) {
2071 * Default alignment: as specified by the arch code. Except if
2072 * an object is really small, then squeeze multiple objects into
2075 ralign
= cache_line_size();
2076 while (size
<= ralign
/ 2)
2079 ralign
= BYTES_PER_WORD
;
2081 /* 2) arch mandated alignment: disables debug if necessary */
2082 if (ralign
< ARCH_SLAB_MINALIGN
) {
2083 ralign
= ARCH_SLAB_MINALIGN
;
2084 if (ralign
> BYTES_PER_WORD
)
2085 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2087 /* 3) caller mandated alignment: disables debug if necessary */
2088 if (ralign
< align
) {
2090 if (ralign
> BYTES_PER_WORD
)
2091 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2094 * 4) Store it. Note that the debug code below can reduce
2095 * the alignment to BYTES_PER_WORD.
2099 /* Get cache's description obj. */
2100 cachep
= kmem_cache_zalloc(&cache_cache
, SLAB_KERNEL
);
2105 cachep
->obj_size
= size
;
2107 if (flags
& SLAB_RED_ZONE
) {
2108 /* redzoning only works with word aligned caches */
2109 align
= BYTES_PER_WORD
;
2111 /* add space for red zone words */
2112 cachep
->obj_offset
+= BYTES_PER_WORD
;
2113 size
+= 2 * BYTES_PER_WORD
;
2115 if (flags
& SLAB_STORE_USER
) {
2116 /* user store requires word alignment and
2117 * one word storage behind the end of the real
2120 align
= BYTES_PER_WORD
;
2121 size
+= BYTES_PER_WORD
;
2123 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2124 if (size
>= malloc_sizes
[INDEX_L3
+ 1].cs_size
2125 && cachep
->obj_size
> cache_line_size() && size
< PAGE_SIZE
) {
2126 cachep
->obj_offset
+= PAGE_SIZE
- size
;
2133 * Determine if the slab management is 'on' or 'off' slab.
2134 * (bootstrapping cannot cope with offslab caches so don't do
2137 if ((size
>= (PAGE_SIZE
>> 3)) && !slab_early_init
)
2139 * Size is large, assume best to place the slab management obj
2140 * off-slab (should allow better packing of objs).
2142 flags
|= CFLGS_OFF_SLAB
;
2144 size
= ALIGN(size
, align
);
2146 left_over
= calculate_slab_order(cachep
, size
, align
, flags
);
2149 printk("kmem_cache_create: couldn't create cache %s.\n", name
);
2150 kmem_cache_free(&cache_cache
, cachep
);
2154 slab_size
= ALIGN(cachep
->num
* sizeof(kmem_bufctl_t
)
2155 + sizeof(struct slab
), align
);
2158 * If the slab has been placed off-slab, and we have enough space then
2159 * move it on-slab. This is at the expense of any extra colouring.
2161 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= slab_size
) {
2162 flags
&= ~CFLGS_OFF_SLAB
;
2163 left_over
-= slab_size
;
2166 if (flags
& CFLGS_OFF_SLAB
) {
2167 /* really off slab. No need for manual alignment */
2169 cachep
->num
* sizeof(kmem_bufctl_t
) + sizeof(struct slab
);
2172 cachep
->colour_off
= cache_line_size();
2173 /* Offset must be a multiple of the alignment. */
2174 if (cachep
->colour_off
< align
)
2175 cachep
->colour_off
= align
;
2176 cachep
->colour
= left_over
/ cachep
->colour_off
;
2177 cachep
->slab_size
= slab_size
;
2178 cachep
->flags
= flags
;
2179 cachep
->gfpflags
= 0;
2180 if (flags
& SLAB_CACHE_DMA
)
2181 cachep
->gfpflags
|= GFP_DMA
;
2182 cachep
->buffer_size
= size
;
2184 if (flags
& CFLGS_OFF_SLAB
)
2185 cachep
->slabp_cache
= kmem_find_general_cachep(slab_size
, 0u);
2186 cachep
->ctor
= ctor
;
2187 cachep
->dtor
= dtor
;
2188 cachep
->name
= name
;
2191 setup_cpu_cache(cachep
);
2193 /* cache setup completed, link it into the list */
2194 list_add(&cachep
->next
, &cache_chain
);
2196 if (!cachep
&& (flags
& SLAB_PANIC
))
2197 panic("kmem_cache_create(): failed to create slab `%s'\n",
2199 mutex_unlock(&cache_chain_mutex
);
2200 unlock_cpu_hotplug();
2203 EXPORT_SYMBOL(kmem_cache_create
);
2206 static void check_irq_off(void)
2208 BUG_ON(!irqs_disabled());
2211 static void check_irq_on(void)
2213 BUG_ON(irqs_disabled());
2216 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2220 assert_spin_locked(&cachep
->nodelists
[numa_node_id()]->list_lock
);
2224 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2228 assert_spin_locked(&cachep
->nodelists
[node
]->list_lock
);
2233 #define check_irq_off() do { } while(0)
2234 #define check_irq_on() do { } while(0)
2235 #define check_spinlock_acquired(x) do { } while(0)
2236 #define check_spinlock_acquired_node(x, y) do { } while(0)
2239 static void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
2240 struct array_cache
*ac
,
2241 int force
, int node
);
2243 static void do_drain(void *arg
)
2245 struct kmem_cache
*cachep
= arg
;
2246 struct array_cache
*ac
;
2247 int node
= numa_node_id();
2250 ac
= cpu_cache_get(cachep
);
2251 spin_lock(&cachep
->nodelists
[node
]->list_lock
);
2252 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
2253 spin_unlock(&cachep
->nodelists
[node
]->list_lock
);
2257 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2259 struct kmem_list3
*l3
;
2262 on_each_cpu(do_drain
, cachep
, 1, 1);
2264 for_each_online_node(node
) {
2265 l3
= cachep
->nodelists
[node
];
2266 if (l3
&& l3
->alien
)
2267 drain_alien_cache(cachep
, l3
->alien
);
2270 for_each_online_node(node
) {
2271 l3
= cachep
->nodelists
[node
];
2273 drain_array(cachep
, l3
, l3
->shared
, 1, node
);
2278 * Remove slabs from the list of free slabs.
2279 * Specify the number of slabs to drain in tofree.
2281 * Returns the actual number of slabs released.
2283 static int drain_freelist(struct kmem_cache
*cache
,
2284 struct kmem_list3
*l3
, int tofree
)
2286 struct list_head
*p
;
2291 while (nr_freed
< tofree
&& !list_empty(&l3
->slabs_free
)) {
2293 spin_lock_irq(&l3
->list_lock
);
2294 p
= l3
->slabs_free
.prev
;
2295 if (p
== &l3
->slabs_free
) {
2296 spin_unlock_irq(&l3
->list_lock
);
2300 slabp
= list_entry(p
, struct slab
, list
);
2302 BUG_ON(slabp
->inuse
);
2304 list_del(&slabp
->list
);
2306 * Safe to drop the lock. The slab is no longer linked
2309 l3
->free_objects
-= cache
->num
;
2310 spin_unlock_irq(&l3
->list_lock
);
2311 slab_destroy(cache
, slabp
);
2318 static int __cache_shrink(struct kmem_cache
*cachep
)
2321 struct kmem_list3
*l3
;
2323 drain_cpu_caches(cachep
);
2326 for_each_online_node(i
) {
2327 l3
= cachep
->nodelists
[i
];
2331 drain_freelist(cachep
, l3
, l3
->free_objects
);
2333 ret
+= !list_empty(&l3
->slabs_full
) ||
2334 !list_empty(&l3
->slabs_partial
);
2336 return (ret
? 1 : 0);
2340 * kmem_cache_shrink - Shrink a cache.
2341 * @cachep: The cache to shrink.
2343 * Releases as many slabs as possible for a cache.
2344 * To help debugging, a zero exit status indicates all slabs were released.
2346 int kmem_cache_shrink(struct kmem_cache
*cachep
)
2348 BUG_ON(!cachep
|| in_interrupt());
2350 return __cache_shrink(cachep
);
2352 EXPORT_SYMBOL(kmem_cache_shrink
);
2355 * kmem_cache_destroy - delete a cache
2356 * @cachep: the cache to destroy
2358 * Remove a struct kmem_cache object from the slab cache.
2359 * Returns 0 on success.
2361 * It is expected this function will be called by a module when it is
2362 * unloaded. This will remove the cache completely, and avoid a duplicate
2363 * cache being allocated each time a module is loaded and unloaded, if the
2364 * module doesn't have persistent in-kernel storage across loads and unloads.
2366 * The cache must be empty before calling this function.
2368 * The caller must guarantee that noone will allocate memory from the cache
2369 * during the kmem_cache_destroy().
2371 int kmem_cache_destroy(struct kmem_cache
*cachep
)
2374 struct kmem_list3
*l3
;
2376 BUG_ON(!cachep
|| in_interrupt());
2378 /* Don't let CPUs to come and go */
2381 /* Find the cache in the chain of caches. */
2382 mutex_lock(&cache_chain_mutex
);
2384 * the chain is never empty, cache_cache is never destroyed
2386 list_del(&cachep
->next
);
2387 mutex_unlock(&cache_chain_mutex
);
2389 if (__cache_shrink(cachep
)) {
2390 slab_error(cachep
, "Can't free all objects");
2391 mutex_lock(&cache_chain_mutex
);
2392 list_add(&cachep
->next
, &cache_chain
);
2393 mutex_unlock(&cache_chain_mutex
);
2394 unlock_cpu_hotplug();
2398 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
2401 for_each_online_cpu(i
)
2402 kfree(cachep
->array
[i
]);
2404 /* NUMA: free the list3 structures */
2405 for_each_online_node(i
) {
2406 l3
= cachep
->nodelists
[i
];
2409 free_alien_cache(l3
->alien
);
2413 kmem_cache_free(&cache_cache
, cachep
);
2414 unlock_cpu_hotplug();
2417 EXPORT_SYMBOL(kmem_cache_destroy
);
2419 /* Get the memory for a slab management obj. */
2420 static struct slab
*alloc_slabmgmt(struct kmem_cache
*cachep
, void *objp
,
2421 int colour_off
, gfp_t local_flags
,
2426 if (OFF_SLAB(cachep
)) {
2427 /* Slab management obj is off-slab. */
2428 slabp
= kmem_cache_alloc_node(cachep
->slabp_cache
,
2429 local_flags
, nodeid
);
2433 slabp
= objp
+ colour_off
;
2434 colour_off
+= cachep
->slab_size
;
2437 slabp
->colouroff
= colour_off
;
2438 slabp
->s_mem
= objp
+ colour_off
;
2439 slabp
->nodeid
= nodeid
;
2443 static inline kmem_bufctl_t
*slab_bufctl(struct slab
*slabp
)
2445 return (kmem_bufctl_t
*) (slabp
+ 1);
2448 static void cache_init_objs(struct kmem_cache
*cachep
,
2449 struct slab
*slabp
, unsigned long ctor_flags
)
2453 for (i
= 0; i
< cachep
->num
; i
++) {
2454 void *objp
= index_to_obj(cachep
, slabp
, i
);
2456 /* need to poison the objs? */
2457 if (cachep
->flags
& SLAB_POISON
)
2458 poison_obj(cachep
, objp
, POISON_FREE
);
2459 if (cachep
->flags
& SLAB_STORE_USER
)
2460 *dbg_userword(cachep
, objp
) = NULL
;
2462 if (cachep
->flags
& SLAB_RED_ZONE
) {
2463 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2464 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2467 * Constructors are not allowed to allocate memory from the same
2468 * cache which they are a constructor for. Otherwise, deadlock.
2469 * They must also be threaded.
2471 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2472 cachep
->ctor(objp
+ obj_offset(cachep
), cachep
,
2475 if (cachep
->flags
& SLAB_RED_ZONE
) {
2476 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2477 slab_error(cachep
, "constructor overwrote the"
2478 " end of an object");
2479 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2480 slab_error(cachep
, "constructor overwrote the"
2481 " start of an object");
2483 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 &&
2484 OFF_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
)
2485 kernel_map_pages(virt_to_page(objp
),
2486 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2489 cachep
->ctor(objp
, cachep
, ctor_flags
);
2491 slab_bufctl(slabp
)[i
] = i
+ 1;
2493 slab_bufctl(slabp
)[i
- 1] = BUFCTL_END
;
2497 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2499 if (flags
& SLAB_DMA
)
2500 BUG_ON(!(cachep
->gfpflags
& GFP_DMA
));
2502 BUG_ON(cachep
->gfpflags
& GFP_DMA
);
2505 static void *slab_get_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2508 void *objp
= index_to_obj(cachep
, slabp
, slabp
->free
);
2512 next
= slab_bufctl(slabp
)[slabp
->free
];
2514 slab_bufctl(slabp
)[slabp
->free
] = BUFCTL_FREE
;
2515 WARN_ON(slabp
->nodeid
!= nodeid
);
2522 static void slab_put_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2523 void *objp
, int nodeid
)
2525 unsigned int objnr
= obj_to_index(cachep
, slabp
, objp
);
2528 /* Verify that the slab belongs to the intended node */
2529 WARN_ON(slabp
->nodeid
!= nodeid
);
2531 if (slab_bufctl(slabp
)[objnr
] + 1 <= SLAB_LIMIT
+ 1) {
2532 printk(KERN_ERR
"slab: double free detected in cache "
2533 "'%s', objp %p\n", cachep
->name
, objp
);
2537 slab_bufctl(slabp
)[objnr
] = slabp
->free
;
2538 slabp
->free
= objnr
;
2543 * Map pages beginning at addr to the given cache and slab. This is required
2544 * for the slab allocator to be able to lookup the cache and slab of a
2545 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2547 static void slab_map_pages(struct kmem_cache
*cache
, struct slab
*slab
,
2553 page
= virt_to_page(addr
);
2556 if (likely(!PageCompound(page
)))
2557 nr_pages
<<= cache
->gfporder
;
2560 page_set_cache(page
, cache
);
2561 page_set_slab(page
, slab
);
2563 } while (--nr_pages
);
2567 * Grow (by 1) the number of slabs within a cache. This is called by
2568 * kmem_cache_alloc() when there are no active objs left in a cache.
2570 static int cache_grow(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
2576 unsigned long ctor_flags
;
2577 struct kmem_list3
*l3
;
2580 * Be lazy and only check for valid flags here, keeping it out of the
2581 * critical path in kmem_cache_alloc().
2583 BUG_ON(flags
& ~(SLAB_DMA
| SLAB_LEVEL_MASK
| SLAB_NO_GROW
));
2584 if (flags
& SLAB_NO_GROW
)
2587 ctor_flags
= SLAB_CTOR_CONSTRUCTOR
;
2588 local_flags
= (flags
& SLAB_LEVEL_MASK
);
2589 if (!(local_flags
& __GFP_WAIT
))
2591 * Not allowed to sleep. Need to tell a constructor about
2592 * this - it might need to know...
2594 ctor_flags
|= SLAB_CTOR_ATOMIC
;
2596 /* Take the l3 list lock to change the colour_next on this node */
2598 l3
= cachep
->nodelists
[nodeid
];
2599 spin_lock(&l3
->list_lock
);
2601 /* Get colour for the slab, and cal the next value. */
2602 offset
= l3
->colour_next
;
2604 if (l3
->colour_next
>= cachep
->colour
)
2605 l3
->colour_next
= 0;
2606 spin_unlock(&l3
->list_lock
);
2608 offset
*= cachep
->colour_off
;
2610 if (local_flags
& __GFP_WAIT
)
2614 * The test for missing atomic flag is performed here, rather than
2615 * the more obvious place, simply to reduce the critical path length
2616 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2617 * will eventually be caught here (where it matters).
2619 kmem_flagcheck(cachep
, flags
);
2622 * Get mem for the objs. Attempt to allocate a physical page from
2625 objp
= kmem_getpages(cachep
, flags
, nodeid
);
2629 /* Get slab management. */
2630 slabp
= alloc_slabmgmt(cachep
, objp
, offset
, local_flags
, nodeid
);
2634 slabp
->nodeid
= nodeid
;
2635 slab_map_pages(cachep
, slabp
, objp
);
2637 cache_init_objs(cachep
, slabp
, ctor_flags
);
2639 if (local_flags
& __GFP_WAIT
)
2640 local_irq_disable();
2642 spin_lock(&l3
->list_lock
);
2644 /* Make slab active. */
2645 list_add_tail(&slabp
->list
, &(l3
->slabs_free
));
2646 STATS_INC_GROWN(cachep
);
2647 l3
->free_objects
+= cachep
->num
;
2648 spin_unlock(&l3
->list_lock
);
2651 kmem_freepages(cachep
, objp
);
2653 if (local_flags
& __GFP_WAIT
)
2654 local_irq_disable();
2661 * Perform extra freeing checks:
2662 * - detect bad pointers.
2663 * - POISON/RED_ZONE checking
2664 * - destructor calls, for caches with POISON+dtor
2666 static void kfree_debugcheck(const void *objp
)
2670 if (!virt_addr_valid(objp
)) {
2671 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
2672 (unsigned long)objp
);
2675 page
= virt_to_page(objp
);
2676 if (!PageSlab(page
)) {
2677 printk(KERN_ERR
"kfree_debugcheck: bad ptr %lxh.\n",
2678 (unsigned long)objp
);
2683 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2685 unsigned long redzone1
, redzone2
;
2687 redzone1
= *dbg_redzone1(cache
, obj
);
2688 redzone2
= *dbg_redzone2(cache
, obj
);
2693 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2696 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2697 slab_error(cache
, "double free detected");
2699 slab_error(cache
, "memory outside object was overwritten");
2701 printk(KERN_ERR
"%p: redzone 1:0x%lx, redzone 2:0x%lx.\n",
2702 obj
, redzone1
, redzone2
);
2705 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2712 objp
-= obj_offset(cachep
);
2713 kfree_debugcheck(objp
);
2714 page
= virt_to_page(objp
);
2716 slabp
= page_get_slab(page
);
2718 if (cachep
->flags
& SLAB_RED_ZONE
) {
2719 verify_redzone_free(cachep
, objp
);
2720 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2721 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2723 if (cachep
->flags
& SLAB_STORE_USER
)
2724 *dbg_userword(cachep
, objp
) = caller
;
2726 objnr
= obj_to_index(cachep
, slabp
, objp
);
2728 BUG_ON(objnr
>= cachep
->num
);
2729 BUG_ON(objp
!= index_to_obj(cachep
, slabp
, objnr
));
2731 if (cachep
->flags
& SLAB_DEBUG_INITIAL
) {
2733 * Need to call the slab's constructor so the caller can
2734 * perform a verify of its state (debugging). Called without
2735 * the cache-lock held.
2737 cachep
->ctor(objp
+ obj_offset(cachep
),
2738 cachep
, SLAB_CTOR_CONSTRUCTOR
| SLAB_CTOR_VERIFY
);
2740 if (cachep
->flags
& SLAB_POISON
&& cachep
->dtor
) {
2741 /* we want to cache poison the object,
2742 * call the destruction callback
2744 cachep
->dtor(objp
+ obj_offset(cachep
), cachep
, 0);
2746 #ifdef CONFIG_DEBUG_SLAB_LEAK
2747 slab_bufctl(slabp
)[objnr
] = BUFCTL_FREE
;
2749 if (cachep
->flags
& SLAB_POISON
) {
2750 #ifdef CONFIG_DEBUG_PAGEALLOC
2751 if ((cachep
->buffer_size
% PAGE_SIZE
)==0 && OFF_SLAB(cachep
)) {
2752 store_stackinfo(cachep
, objp
, (unsigned long)caller
);
2753 kernel_map_pages(virt_to_page(objp
),
2754 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2756 poison_obj(cachep
, objp
, POISON_FREE
);
2759 poison_obj(cachep
, objp
, POISON_FREE
);
2765 static void check_slabp(struct kmem_cache
*cachep
, struct slab
*slabp
)
2770 /* Check slab's freelist to see if this obj is there. */
2771 for (i
= slabp
->free
; i
!= BUFCTL_END
; i
= slab_bufctl(slabp
)[i
]) {
2773 if (entries
> cachep
->num
|| i
>= cachep
->num
)
2776 if (entries
!= cachep
->num
- slabp
->inuse
) {
2778 printk(KERN_ERR
"slab: Internal list corruption detected in "
2779 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2780 cachep
->name
, cachep
->num
, slabp
, slabp
->inuse
);
2782 i
< sizeof(*slabp
) + cachep
->num
* sizeof(kmem_bufctl_t
);
2785 printk("\n%03x:", i
);
2786 printk(" %02x", ((unsigned char *)slabp
)[i
]);
2793 #define kfree_debugcheck(x) do { } while(0)
2794 #define cache_free_debugcheck(x,objp,z) (objp)
2795 #define check_slabp(x,y) do { } while(0)
2798 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
2801 struct kmem_list3
*l3
;
2802 struct array_cache
*ac
;
2805 ac
= cpu_cache_get(cachep
);
2807 batchcount
= ac
->batchcount
;
2808 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2810 * If there was little recent activity on this cache, then
2811 * perform only a partial refill. Otherwise we could generate
2814 batchcount
= BATCHREFILL_LIMIT
;
2816 l3
= cachep
->nodelists
[numa_node_id()];
2818 BUG_ON(ac
->avail
> 0 || !l3
);
2819 spin_lock(&l3
->list_lock
);
2821 /* See if we can refill from the shared array */
2822 if (l3
->shared
&& transfer_objects(ac
, l3
->shared
, batchcount
))
2825 while (batchcount
> 0) {
2826 struct list_head
*entry
;
2828 /* Get slab alloc is to come from. */
2829 entry
= l3
->slabs_partial
.next
;
2830 if (entry
== &l3
->slabs_partial
) {
2831 l3
->free_touched
= 1;
2832 entry
= l3
->slabs_free
.next
;
2833 if (entry
== &l3
->slabs_free
)
2837 slabp
= list_entry(entry
, struct slab
, list
);
2838 check_slabp(cachep
, slabp
);
2839 check_spinlock_acquired(cachep
);
2840 while (slabp
->inuse
< cachep
->num
&& batchcount
--) {
2841 STATS_INC_ALLOCED(cachep
);
2842 STATS_INC_ACTIVE(cachep
);
2843 STATS_SET_HIGH(cachep
);
2845 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, slabp
,
2848 check_slabp(cachep
, slabp
);
2850 /* move slabp to correct slabp list: */
2851 list_del(&slabp
->list
);
2852 if (slabp
->free
== BUFCTL_END
)
2853 list_add(&slabp
->list
, &l3
->slabs_full
);
2855 list_add(&slabp
->list
, &l3
->slabs_partial
);
2859 l3
->free_objects
-= ac
->avail
;
2861 spin_unlock(&l3
->list_lock
);
2863 if (unlikely(!ac
->avail
)) {
2865 x
= cache_grow(cachep
, flags
, numa_node_id());
2867 /* cache_grow can reenable interrupts, then ac could change. */
2868 ac
= cpu_cache_get(cachep
);
2869 if (!x
&& ac
->avail
== 0) /* no objects in sight? abort */
2872 if (!ac
->avail
) /* objects refilled by interrupt? */
2876 return ac
->entry
[--ac
->avail
];
2879 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
2882 might_sleep_if(flags
& __GFP_WAIT
);
2884 kmem_flagcheck(cachep
, flags
);
2889 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
2890 gfp_t flags
, void *objp
, void *caller
)
2894 if (cachep
->flags
& SLAB_POISON
) {
2895 #ifdef CONFIG_DEBUG_PAGEALLOC
2896 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
2897 kernel_map_pages(virt_to_page(objp
),
2898 cachep
->buffer_size
/ PAGE_SIZE
, 1);
2900 check_poison_obj(cachep
, objp
);
2902 check_poison_obj(cachep
, objp
);
2904 poison_obj(cachep
, objp
, POISON_INUSE
);
2906 if (cachep
->flags
& SLAB_STORE_USER
)
2907 *dbg_userword(cachep
, objp
) = caller
;
2909 if (cachep
->flags
& SLAB_RED_ZONE
) {
2910 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
2911 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
2912 slab_error(cachep
, "double free, or memory outside"
2913 " object was overwritten");
2915 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
2916 objp
, *dbg_redzone1(cachep
, objp
),
2917 *dbg_redzone2(cachep
, objp
));
2919 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
2920 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
2922 #ifdef CONFIG_DEBUG_SLAB_LEAK
2927 slabp
= page_get_slab(virt_to_page(objp
));
2928 objnr
= (unsigned)(objp
- slabp
->s_mem
) / cachep
->buffer_size
;
2929 slab_bufctl(slabp
)[objnr
] = BUFCTL_ACTIVE
;
2932 objp
+= obj_offset(cachep
);
2933 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
) {
2934 unsigned long ctor_flags
= SLAB_CTOR_CONSTRUCTOR
;
2936 if (!(flags
& __GFP_WAIT
))
2937 ctor_flags
|= SLAB_CTOR_ATOMIC
;
2939 cachep
->ctor(objp
, cachep
, ctor_flags
);
2944 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2947 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
2950 struct array_cache
*ac
;
2953 if (unlikely(current
->flags
& (PF_SPREAD_SLAB
| PF_MEMPOLICY
))) {
2954 objp
= alternate_node_alloc(cachep
, flags
);
2961 ac
= cpu_cache_get(cachep
);
2962 if (likely(ac
->avail
)) {
2963 STATS_INC_ALLOCHIT(cachep
);
2965 objp
= ac
->entry
[--ac
->avail
];
2967 STATS_INC_ALLOCMISS(cachep
);
2968 objp
= cache_alloc_refill(cachep
, flags
);
2973 static __always_inline
void *__cache_alloc(struct kmem_cache
*cachep
,
2974 gfp_t flags
, void *caller
)
2976 unsigned long save_flags
;
2979 cache_alloc_debugcheck_before(cachep
, flags
);
2981 local_irq_save(save_flags
);
2982 objp
= ____cache_alloc(cachep
, flags
);
2983 local_irq_restore(save_flags
);
2984 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
,
2992 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
2994 * If we are in_interrupt, then process context, including cpusets and
2995 * mempolicy, may not apply and should not be used for allocation policy.
2997 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
2999 int nid_alloc
, nid_here
;
3003 nid_alloc
= nid_here
= numa_node_id();
3004 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3005 nid_alloc
= cpuset_mem_spread_node();
3006 else if (current
->mempolicy
)
3007 nid_alloc
= slab_node(current
->mempolicy
);
3008 if (nid_alloc
!= nid_here
)
3009 return __cache_alloc_node(cachep
, flags
, nid_alloc
);
3014 * A interface to enable slab creation on nodeid
3016 static void *__cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3019 struct list_head
*entry
;
3021 struct kmem_list3
*l3
;
3025 l3
= cachep
->nodelists
[nodeid
];
3030 spin_lock(&l3
->list_lock
);
3031 entry
= l3
->slabs_partial
.next
;
3032 if (entry
== &l3
->slabs_partial
) {
3033 l3
->free_touched
= 1;
3034 entry
= l3
->slabs_free
.next
;
3035 if (entry
== &l3
->slabs_free
)
3039 slabp
= list_entry(entry
, struct slab
, list
);
3040 check_spinlock_acquired_node(cachep
, nodeid
);
3041 check_slabp(cachep
, slabp
);
3043 STATS_INC_NODEALLOCS(cachep
);
3044 STATS_INC_ACTIVE(cachep
);
3045 STATS_SET_HIGH(cachep
);
3047 BUG_ON(slabp
->inuse
== cachep
->num
);
3049 obj
= slab_get_obj(cachep
, slabp
, nodeid
);
3050 check_slabp(cachep
, slabp
);
3052 /* move slabp to correct slabp list: */
3053 list_del(&slabp
->list
);
3055 if (slabp
->free
== BUFCTL_END
)
3056 list_add(&slabp
->list
, &l3
->slabs_full
);
3058 list_add(&slabp
->list
, &l3
->slabs_partial
);
3060 spin_unlock(&l3
->list_lock
);
3064 spin_unlock(&l3
->list_lock
);
3065 x
= cache_grow(cachep
, flags
, nodeid
);
3077 * Caller needs to acquire correct kmem_list's list_lock
3079 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int nr_objects
,
3083 struct kmem_list3
*l3
;
3085 for (i
= 0; i
< nr_objects
; i
++) {
3086 void *objp
= objpp
[i
];
3089 slabp
= virt_to_slab(objp
);
3090 l3
= cachep
->nodelists
[node
];
3091 list_del(&slabp
->list
);
3092 check_spinlock_acquired_node(cachep
, node
);
3093 check_slabp(cachep
, slabp
);
3094 slab_put_obj(cachep
, slabp
, objp
, node
);
3095 STATS_DEC_ACTIVE(cachep
);
3097 check_slabp(cachep
, slabp
);
3099 /* fixup slab chains */
3100 if (slabp
->inuse
== 0) {
3101 if (l3
->free_objects
> l3
->free_limit
) {
3102 l3
->free_objects
-= cachep
->num
;
3104 * It is safe to drop the lock. The slab is
3105 * no longer linked to the cache. cachep
3106 * cannot disappear - we are using it and
3107 * all destruction of caches must be
3108 * serialized properly by the user.
3110 spin_unlock(&l3
->list_lock
);
3111 slab_destroy(cachep
, slabp
);
3112 spin_lock(&l3
->list_lock
);
3114 list_add(&slabp
->list
, &l3
->slabs_free
);
3117 /* Unconditionally move a slab to the end of the
3118 * partial list on free - maximum time for the
3119 * other objects to be freed, too.
3121 list_add_tail(&slabp
->list
, &l3
->slabs_partial
);
3126 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3129 struct kmem_list3
*l3
;
3130 int node
= numa_node_id();
3132 batchcount
= ac
->batchcount
;
3134 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
3137 l3
= cachep
->nodelists
[node
];
3138 spin_lock_nested(&l3
->list_lock
, SINGLE_DEPTH_NESTING
);
3140 struct array_cache
*shared_array
= l3
->shared
;
3141 int max
= shared_array
->limit
- shared_array
->avail
;
3143 if (batchcount
> max
)
3145 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3146 ac
->entry
, sizeof(void *) * batchcount
);
3147 shared_array
->avail
+= batchcount
;
3152 free_block(cachep
, ac
->entry
, batchcount
, node
);
3157 struct list_head
*p
;
3159 p
= l3
->slabs_free
.next
;
3160 while (p
!= &(l3
->slabs_free
)) {
3163 slabp
= list_entry(p
, struct slab
, list
);
3164 BUG_ON(slabp
->inuse
);
3169 STATS_SET_FREEABLE(cachep
, i
);
3172 spin_unlock(&l3
->list_lock
);
3173 ac
->avail
-= batchcount
;
3174 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3178 * Release an obj back to its cache. If the obj has a constructed state, it must
3179 * be in this state _before_ it is released. Called with disabled ints.
3181 static void __cache_free(struct kmem_cache
*cachep
, void *objp
, int nesting
)
3183 struct array_cache
*ac
= cpu_cache_get(cachep
);
3186 objp
= cache_free_debugcheck(cachep
, objp
, __builtin_return_address(0));
3188 if (cache_free_alien(cachep
, objp
, nesting
))
3191 if (likely(ac
->avail
< ac
->limit
)) {
3192 STATS_INC_FREEHIT(cachep
);
3193 ac
->entry
[ac
->avail
++] = objp
;
3196 STATS_INC_FREEMISS(cachep
);
3197 cache_flusharray(cachep
, ac
);
3198 ac
->entry
[ac
->avail
++] = objp
;
3203 * kmem_cache_alloc - Allocate an object
3204 * @cachep: The cache to allocate from.
3205 * @flags: See kmalloc().
3207 * Allocate an object from this cache. The flags are only relevant
3208 * if the cache has no available objects.
3210 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3212 return __cache_alloc(cachep
, flags
, __builtin_return_address(0));
3214 EXPORT_SYMBOL(kmem_cache_alloc
);
3217 * kmem_cache_alloc - Allocate an object. The memory is set to zero.
3218 * @cache: The cache to allocate from.
3219 * @flags: See kmalloc().
3221 * Allocate an object from this cache and set the allocated memory to zero.
3222 * The flags are only relevant if the cache has no available objects.
3224 void *kmem_cache_zalloc(struct kmem_cache
*cache
, gfp_t flags
)
3226 void *ret
= __cache_alloc(cache
, flags
, __builtin_return_address(0));
3228 memset(ret
, 0, obj_size(cache
));
3231 EXPORT_SYMBOL(kmem_cache_zalloc
);
3234 * kmem_ptr_validate - check if an untrusted pointer might
3236 * @cachep: the cache we're checking against
3237 * @ptr: pointer to validate
3239 * This verifies that the untrusted pointer looks sane:
3240 * it is _not_ a guarantee that the pointer is actually
3241 * part of the slab cache in question, but it at least
3242 * validates that the pointer can be dereferenced and
3243 * looks half-way sane.
3245 * Currently only used for dentry validation.
3247 int fastcall
kmem_ptr_validate(struct kmem_cache
*cachep
, void *ptr
)
3249 unsigned long addr
= (unsigned long)ptr
;
3250 unsigned long min_addr
= PAGE_OFFSET
;
3251 unsigned long align_mask
= BYTES_PER_WORD
- 1;
3252 unsigned long size
= cachep
->buffer_size
;
3255 if (unlikely(addr
< min_addr
))
3257 if (unlikely(addr
> (unsigned long)high_memory
- size
))
3259 if (unlikely(addr
& align_mask
))
3261 if (unlikely(!kern_addr_valid(addr
)))
3263 if (unlikely(!kern_addr_valid(addr
+ size
- 1)))
3265 page
= virt_to_page(ptr
);
3266 if (unlikely(!PageSlab(page
)))
3268 if (unlikely(page_get_cache(page
) != cachep
))
3277 * kmem_cache_alloc_node - Allocate an object on the specified node
3278 * @cachep: The cache to allocate from.
3279 * @flags: See kmalloc().
3280 * @nodeid: node number of the target node.
3282 * Identical to kmem_cache_alloc, except that this function is slow
3283 * and can sleep. And it will allocate memory on the given node, which
3284 * can improve the performance for cpu bound structures.
3285 * New and improved: it will now make sure that the object gets
3286 * put on the correct node list so that there is no false sharing.
3288 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3290 unsigned long save_flags
;
3293 cache_alloc_debugcheck_before(cachep
, flags
);
3294 local_irq_save(save_flags
);
3296 if (nodeid
== -1 || nodeid
== numa_node_id() ||
3297 !cachep
->nodelists
[nodeid
])
3298 ptr
= ____cache_alloc(cachep
, flags
);
3300 ptr
= __cache_alloc_node(cachep
, flags
, nodeid
);
3301 local_irq_restore(save_flags
);
3303 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
,
3304 __builtin_return_address(0));
3308 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3310 void *kmalloc_node(size_t size
, gfp_t flags
, int node
)
3312 struct kmem_cache
*cachep
;
3314 cachep
= kmem_find_general_cachep(size
, flags
);
3315 if (unlikely(cachep
== NULL
))
3317 return kmem_cache_alloc_node(cachep
, flags
, node
);
3319 EXPORT_SYMBOL(kmalloc_node
);
3323 * __do_kmalloc - allocate memory
3324 * @size: how many bytes of memory are required.
3325 * @flags: the type of memory to allocate (see kmalloc).
3326 * @caller: function caller for debug tracking of the caller
3328 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3331 struct kmem_cache
*cachep
;
3333 /* If you want to save a few bytes .text space: replace
3335 * Then kmalloc uses the uninlined functions instead of the inline
3338 cachep
= __find_general_cachep(size
, flags
);
3339 if (unlikely(cachep
== NULL
))
3341 return __cache_alloc(cachep
, flags
, caller
);
3345 void *__kmalloc(size_t size
, gfp_t flags
)
3347 #ifndef CONFIG_DEBUG_SLAB
3348 return __do_kmalloc(size
, flags
, NULL
);
3350 return __do_kmalloc(size
, flags
, __builtin_return_address(0));
3353 EXPORT_SYMBOL(__kmalloc
);
3355 #ifdef CONFIG_DEBUG_SLAB
3356 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, void *caller
)
3358 return __do_kmalloc(size
, flags
, caller
);
3360 EXPORT_SYMBOL(__kmalloc_track_caller
);
3365 * __alloc_percpu - allocate one copy of the object for every present
3366 * cpu in the system, zeroing them.
3367 * Objects should be dereferenced using the per_cpu_ptr macro only.
3369 * @size: how many bytes of memory are required.
3371 void *__alloc_percpu(size_t size
)
3374 struct percpu_data
*pdata
= kmalloc(sizeof(*pdata
), GFP_KERNEL
);
3380 * Cannot use for_each_online_cpu since a cpu may come online
3381 * and we have no way of figuring out how to fix the array
3382 * that we have allocated then....
3384 for_each_possible_cpu(i
) {
3385 int node
= cpu_to_node(i
);
3387 if (node_online(node
))
3388 pdata
->ptrs
[i
] = kmalloc_node(size
, GFP_KERNEL
, node
);
3390 pdata
->ptrs
[i
] = kmalloc(size
, GFP_KERNEL
);
3392 if (!pdata
->ptrs
[i
])
3394 memset(pdata
->ptrs
[i
], 0, size
);
3397 /* Catch derefs w/o wrappers */
3398 return (void *)(~(unsigned long)pdata
);
3402 if (!cpu_possible(i
))
3404 kfree(pdata
->ptrs
[i
]);
3409 EXPORT_SYMBOL(__alloc_percpu
);
3413 * kmem_cache_free - Deallocate an object
3414 * @cachep: The cache the allocation was from.
3415 * @objp: The previously allocated object.
3417 * Free an object which was previously allocated from this
3420 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3422 unsigned long flags
;
3424 BUG_ON(virt_to_cache(objp
) != cachep
);
3426 local_irq_save(flags
);
3427 __cache_free(cachep
, objp
, 0);
3428 local_irq_restore(flags
);
3430 EXPORT_SYMBOL(kmem_cache_free
);
3433 * kfree - free previously allocated memory
3434 * @objp: pointer returned by kmalloc.
3436 * If @objp is NULL, no operation is performed.
3438 * Don't free memory not originally allocated by kmalloc()
3439 * or you will run into trouble.
3441 void kfree(const void *objp
)
3443 struct kmem_cache
*c
;
3444 unsigned long flags
;
3446 if (unlikely(!objp
))
3448 local_irq_save(flags
);
3449 kfree_debugcheck(objp
);
3450 c
= virt_to_cache(objp
);
3451 debug_check_no_locks_freed(objp
, obj_size(c
));
3452 __cache_free(c
, (void *)objp
, 0);
3453 local_irq_restore(flags
);
3455 EXPORT_SYMBOL(kfree
);
3459 * free_percpu - free previously allocated percpu memory
3460 * @objp: pointer returned by alloc_percpu.
3462 * Don't free memory not originally allocated by alloc_percpu()
3463 * The complemented objp is to check for that.
3465 void free_percpu(const void *objp
)
3468 struct percpu_data
*p
= (struct percpu_data
*)(~(unsigned long)objp
);
3471 * We allocate for all cpus so we cannot use for online cpu here.
3473 for_each_possible_cpu(i
)
3477 EXPORT_SYMBOL(free_percpu
);
3480 unsigned int kmem_cache_size(struct kmem_cache
*cachep
)
3482 return obj_size(cachep
);
3484 EXPORT_SYMBOL(kmem_cache_size
);
3486 const char *kmem_cache_name(struct kmem_cache
*cachep
)
3488 return cachep
->name
;
3490 EXPORT_SYMBOL_GPL(kmem_cache_name
);
3493 * This initializes kmem_list3 or resizes varioius caches for all nodes.
3495 static int alloc_kmemlist(struct kmem_cache
*cachep
)
3498 struct kmem_list3
*l3
;
3499 struct array_cache
*new_shared
;
3500 struct array_cache
**new_alien
;
3502 for_each_online_node(node
) {
3504 new_alien
= alloc_alien_cache(node
, cachep
->limit
);
3508 new_shared
= alloc_arraycache(node
,
3509 cachep
->shared
*cachep
->batchcount
,
3512 free_alien_cache(new_alien
);
3516 l3
= cachep
->nodelists
[node
];
3518 struct array_cache
*shared
= l3
->shared
;
3520 spin_lock_irq(&l3
->list_lock
);
3523 free_block(cachep
, shared
->entry
,
3524 shared
->avail
, node
);
3526 l3
->shared
= new_shared
;
3528 l3
->alien
= new_alien
;
3531 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3532 cachep
->batchcount
+ cachep
->num
;
3533 spin_unlock_irq(&l3
->list_lock
);
3535 free_alien_cache(new_alien
);
3538 l3
= kmalloc_node(sizeof(struct kmem_list3
), GFP_KERNEL
, node
);
3540 free_alien_cache(new_alien
);
3545 kmem_list3_init(l3
);
3546 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
3547 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
3548 l3
->shared
= new_shared
;
3549 l3
->alien
= new_alien
;
3550 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3551 cachep
->batchcount
+ cachep
->num
;
3552 cachep
->nodelists
[node
] = l3
;
3557 if (!cachep
->next
.next
) {
3558 /* Cache is not active yet. Roll back what we did */
3561 if (cachep
->nodelists
[node
]) {
3562 l3
= cachep
->nodelists
[node
];
3565 free_alien_cache(l3
->alien
);
3567 cachep
->nodelists
[node
] = NULL
;
3575 struct ccupdate_struct
{
3576 struct kmem_cache
*cachep
;
3577 struct array_cache
*new[NR_CPUS
];
3580 static void do_ccupdate_local(void *info
)
3582 struct ccupdate_struct
*new = info
;
3583 struct array_cache
*old
;
3586 old
= cpu_cache_get(new->cachep
);
3588 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
3589 new->new[smp_processor_id()] = old
;
3592 /* Always called with the cache_chain_mutex held */
3593 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3594 int batchcount
, int shared
)
3596 struct ccupdate_struct
new;
3599 memset(&new.new, 0, sizeof(new.new));
3600 for_each_online_cpu(i
) {
3601 new.new[i
] = alloc_arraycache(cpu_to_node(i
), limit
,
3604 for (i
--; i
>= 0; i
--)
3609 new.cachep
= cachep
;
3611 on_each_cpu(do_ccupdate_local
, (void *)&new, 1, 1);
3614 cachep
->batchcount
= batchcount
;
3615 cachep
->limit
= limit
;
3616 cachep
->shared
= shared
;
3618 for_each_online_cpu(i
) {
3619 struct array_cache
*ccold
= new.new[i
];
3622 spin_lock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3623 free_block(cachep
, ccold
->entry
, ccold
->avail
, cpu_to_node(i
));
3624 spin_unlock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3628 err
= alloc_kmemlist(cachep
);
3630 printk(KERN_ERR
"alloc_kmemlist failed for %s, error %d.\n",
3631 cachep
->name
, -err
);
3637 /* Called with cache_chain_mutex held always */
3638 static void enable_cpucache(struct kmem_cache
*cachep
)
3644 * The head array serves three purposes:
3645 * - create a LIFO ordering, i.e. return objects that are cache-warm
3646 * - reduce the number of spinlock operations.
3647 * - reduce the number of linked list operations on the slab and
3648 * bufctl chains: array operations are cheaper.
3649 * The numbers are guessed, we should auto-tune as described by
3652 if (cachep
->buffer_size
> 131072)
3654 else if (cachep
->buffer_size
> PAGE_SIZE
)
3656 else if (cachep
->buffer_size
> 1024)
3658 else if (cachep
->buffer_size
> 256)
3664 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3665 * allocation behaviour: Most allocs on one cpu, most free operations
3666 * on another cpu. For these cases, an efficient object passing between
3667 * cpus is necessary. This is provided by a shared array. The array
3668 * replaces Bonwick's magazine layer.
3669 * On uniprocessor, it's functionally equivalent (but less efficient)
3670 * to a larger limit. Thus disabled by default.
3674 if (cachep
->buffer_size
<= PAGE_SIZE
)
3680 * With debugging enabled, large batchcount lead to excessively long
3681 * periods with disabled local interrupts. Limit the batchcount
3686 err
= do_tune_cpucache(cachep
, limit
, (limit
+ 1) / 2, shared
);
3688 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
3689 cachep
->name
, -err
);
3693 * Drain an array if it contains any elements taking the l3 lock only if
3694 * necessary. Note that the l3 listlock also protects the array_cache
3695 * if drain_array() is used on the shared array.
3697 void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
3698 struct array_cache
*ac
, int force
, int node
)
3702 if (!ac
|| !ac
->avail
)
3704 if (ac
->touched
&& !force
) {
3707 spin_lock_irq(&l3
->list_lock
);
3709 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
3710 if (tofree
> ac
->avail
)
3711 tofree
= (ac
->avail
+ 1) / 2;
3712 free_block(cachep
, ac
->entry
, tofree
, node
);
3713 ac
->avail
-= tofree
;
3714 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
3715 sizeof(void *) * ac
->avail
);
3717 spin_unlock_irq(&l3
->list_lock
);
3722 * cache_reap - Reclaim memory from caches.
3723 * @unused: unused parameter
3725 * Called from workqueue/eventd every few seconds.
3727 * - clear the per-cpu caches for this CPU.
3728 * - return freeable pages to the main free memory pool.
3730 * If we cannot acquire the cache chain mutex then just give up - we'll try
3731 * again on the next iteration.
3733 static void cache_reap(void *unused
)
3735 struct kmem_cache
*searchp
;
3736 struct kmem_list3
*l3
;
3737 int node
= numa_node_id();
3739 if (!mutex_trylock(&cache_chain_mutex
)) {
3740 /* Give up. Setup the next iteration. */
3741 schedule_delayed_work(&__get_cpu_var(reap_work
),
3746 list_for_each_entry(searchp
, &cache_chain
, next
) {
3750 * We only take the l3 lock if absolutely necessary and we
3751 * have established with reasonable certainty that
3752 * we can do some work if the lock was obtained.
3754 l3
= searchp
->nodelists
[node
];
3756 reap_alien(searchp
, l3
);
3758 drain_array(searchp
, l3
, cpu_cache_get(searchp
), 0, node
);
3761 * These are racy checks but it does not matter
3762 * if we skip one check or scan twice.
3764 if (time_after(l3
->next_reap
, jiffies
))
3767 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
;
3769 drain_array(searchp
, l3
, l3
->shared
, 0, node
);
3771 if (l3
->free_touched
)
3772 l3
->free_touched
= 0;
3776 freed
= drain_freelist(searchp
, l3
, (l3
->free_limit
+
3777 5 * searchp
->num
- 1) / (5 * searchp
->num
));
3778 STATS_ADD_REAPED(searchp
, freed
);
3784 mutex_unlock(&cache_chain_mutex
);
3786 refresh_cpu_vm_stats(smp_processor_id());
3787 /* Set up the next iteration */
3788 schedule_delayed_work(&__get_cpu_var(reap_work
), REAPTIMEOUT_CPUC
);
3791 #ifdef CONFIG_PROC_FS
3793 static void print_slabinfo_header(struct seq_file
*m
)
3796 * Output format version, so at least we can change it
3797 * without _too_ many complaints.
3800 seq_puts(m
, "slabinfo - version: 2.1 (statistics)\n");
3802 seq_puts(m
, "slabinfo - version: 2.1\n");
3804 seq_puts(m
, "# name <active_objs> <num_objs> <objsize> "
3805 "<objperslab> <pagesperslab>");
3806 seq_puts(m
, " : tunables <limit> <batchcount> <sharedfactor>");
3807 seq_puts(m
, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3809 seq_puts(m
, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3810 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
3811 seq_puts(m
, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3816 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
3819 struct list_head
*p
;
3821 mutex_lock(&cache_chain_mutex
);
3823 print_slabinfo_header(m
);
3824 p
= cache_chain
.next
;
3827 if (p
== &cache_chain
)
3830 return list_entry(p
, struct kmem_cache
, next
);
3833 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
3835 struct kmem_cache
*cachep
= p
;
3837 return cachep
->next
.next
== &cache_chain
?
3838 NULL
: list_entry(cachep
->next
.next
, struct kmem_cache
, next
);
3841 static void s_stop(struct seq_file
*m
, void *p
)
3843 mutex_unlock(&cache_chain_mutex
);
3846 static int s_show(struct seq_file
*m
, void *p
)
3848 struct kmem_cache
*cachep
= p
;
3850 unsigned long active_objs
;
3851 unsigned long num_objs
;
3852 unsigned long active_slabs
= 0;
3853 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
3857 struct kmem_list3
*l3
;
3861 for_each_online_node(node
) {
3862 l3
= cachep
->nodelists
[node
];
3867 spin_lock_irq(&l3
->list_lock
);
3869 list_for_each_entry(slabp
, &l3
->slabs_full
, list
) {
3870 if (slabp
->inuse
!= cachep
->num
&& !error
)
3871 error
= "slabs_full accounting error";
3872 active_objs
+= cachep
->num
;
3875 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
) {
3876 if (slabp
->inuse
== cachep
->num
&& !error
)
3877 error
= "slabs_partial inuse accounting error";
3878 if (!slabp
->inuse
&& !error
)
3879 error
= "slabs_partial/inuse accounting error";
3880 active_objs
+= slabp
->inuse
;
3883 list_for_each_entry(slabp
, &l3
->slabs_free
, list
) {
3884 if (slabp
->inuse
&& !error
)
3885 error
= "slabs_free/inuse accounting error";
3888 free_objects
+= l3
->free_objects
;
3890 shared_avail
+= l3
->shared
->avail
;
3892 spin_unlock_irq(&l3
->list_lock
);
3894 num_slabs
+= active_slabs
;
3895 num_objs
= num_slabs
* cachep
->num
;
3896 if (num_objs
- active_objs
!= free_objects
&& !error
)
3897 error
= "free_objects accounting error";
3899 name
= cachep
->name
;
3901 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
3903 seq_printf(m
, "%-17s %6lu %6lu %6u %4u %4d",
3904 name
, active_objs
, num_objs
, cachep
->buffer_size
,
3905 cachep
->num
, (1 << cachep
->gfporder
));
3906 seq_printf(m
, " : tunables %4u %4u %4u",
3907 cachep
->limit
, cachep
->batchcount
, cachep
->shared
);
3908 seq_printf(m
, " : slabdata %6lu %6lu %6lu",
3909 active_slabs
, num_slabs
, shared_avail
);
3912 unsigned long high
= cachep
->high_mark
;
3913 unsigned long allocs
= cachep
->num_allocations
;
3914 unsigned long grown
= cachep
->grown
;
3915 unsigned long reaped
= cachep
->reaped
;
3916 unsigned long errors
= cachep
->errors
;
3917 unsigned long max_freeable
= cachep
->max_freeable
;
3918 unsigned long node_allocs
= cachep
->node_allocs
;
3919 unsigned long node_frees
= cachep
->node_frees
;
3920 unsigned long overflows
= cachep
->node_overflow
;
3922 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu \
3923 %4lu %4lu %4lu %4lu %4lu", allocs
, high
, grown
,
3924 reaped
, errors
, max_freeable
, node_allocs
,
3925 node_frees
, overflows
);
3929 unsigned long allochit
= atomic_read(&cachep
->allochit
);
3930 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
3931 unsigned long freehit
= atomic_read(&cachep
->freehit
);
3932 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
3934 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
3935 allochit
, allocmiss
, freehit
, freemiss
);
3943 * slabinfo_op - iterator that generates /proc/slabinfo
3952 * num-pages-per-slab
3953 * + further values on SMP and with statistics enabled
3956 struct seq_operations slabinfo_op
= {
3963 #define MAX_SLABINFO_WRITE 128
3965 * slabinfo_write - Tuning for the slab allocator
3967 * @buffer: user buffer
3968 * @count: data length
3971 ssize_t
slabinfo_write(struct file
*file
, const char __user
* buffer
,
3972 size_t count
, loff_t
*ppos
)
3974 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
3975 int limit
, batchcount
, shared
, res
;
3976 struct kmem_cache
*cachep
;
3978 if (count
> MAX_SLABINFO_WRITE
)
3980 if (copy_from_user(&kbuf
, buffer
, count
))
3982 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
3984 tmp
= strchr(kbuf
, ' ');
3989 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
3992 /* Find the cache in the chain of caches. */
3993 mutex_lock(&cache_chain_mutex
);
3995 list_for_each_entry(cachep
, &cache_chain
, next
) {
3996 if (!strcmp(cachep
->name
, kbuf
)) {
3997 if (limit
< 1 || batchcount
< 1 ||
3998 batchcount
> limit
|| shared
< 0) {
4001 res
= do_tune_cpucache(cachep
, limit
,
4002 batchcount
, shared
);
4007 mutex_unlock(&cache_chain_mutex
);
4013 #ifdef CONFIG_DEBUG_SLAB_LEAK
4015 static void *leaks_start(struct seq_file
*m
, loff_t
*pos
)
4018 struct list_head
*p
;
4020 mutex_lock(&cache_chain_mutex
);
4021 p
= cache_chain
.next
;
4024 if (p
== &cache_chain
)
4027 return list_entry(p
, struct kmem_cache
, next
);
4030 static inline int add_caller(unsigned long *n
, unsigned long v
)
4040 unsigned long *q
= p
+ 2 * i
;
4054 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4060 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
, struct slab
*s
)
4066 for (i
= 0, p
= s
->s_mem
; i
< c
->num
; i
++, p
+= c
->buffer_size
) {
4067 if (slab_bufctl(s
)[i
] != BUFCTL_ACTIVE
)
4069 if (!add_caller(n
, (unsigned long)*dbg_userword(c
, p
)))
4074 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4076 #ifdef CONFIG_KALLSYMS
4079 unsigned long offset
, size
;
4080 char namebuf
[KSYM_NAME_LEN
+1];
4082 name
= kallsyms_lookup(address
, &size
, &offset
, &modname
, namebuf
);
4085 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4087 seq_printf(m
, " [%s]", modname
);
4091 seq_printf(m
, "%p", (void *)address
);
4094 static int leaks_show(struct seq_file
*m
, void *p
)
4096 struct kmem_cache
*cachep
= p
;
4098 struct kmem_list3
*l3
;
4100 unsigned long *n
= m
->private;
4104 if (!(cachep
->flags
& SLAB_STORE_USER
))
4106 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4109 /* OK, we can do it */
4113 for_each_online_node(node
) {
4114 l3
= cachep
->nodelists
[node
];
4119 spin_lock_irq(&l3
->list_lock
);
4121 list_for_each_entry(slabp
, &l3
->slabs_full
, list
)
4122 handle_slab(n
, cachep
, slabp
);
4123 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
)
4124 handle_slab(n
, cachep
, slabp
);
4125 spin_unlock_irq(&l3
->list_lock
);
4127 name
= cachep
->name
;
4129 /* Increase the buffer size */
4130 mutex_unlock(&cache_chain_mutex
);
4131 m
->private = kzalloc(n
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4133 /* Too bad, we are really out */
4135 mutex_lock(&cache_chain_mutex
);
4138 *(unsigned long *)m
->private = n
[0] * 2;
4140 mutex_lock(&cache_chain_mutex
);
4141 /* Now make sure this entry will be retried */
4145 for (i
= 0; i
< n
[1]; i
++) {
4146 seq_printf(m
, "%s: %lu ", name
, n
[2*i
+3]);
4147 show_symbol(m
, n
[2*i
+2]);
4153 struct seq_operations slabstats_op
= {
4154 .start
= leaks_start
,
4163 * ksize - get the actual amount of memory allocated for a given object
4164 * @objp: Pointer to the object
4166 * kmalloc may internally round up allocations and return more memory
4167 * than requested. ksize() can be used to determine the actual amount of
4168 * memory allocated. The caller may use this additional memory, even though
4169 * a smaller amount of memory was initially specified with the kmalloc call.
4170 * The caller must guarantee that objp points to a valid object previously
4171 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4172 * must not be freed during the duration of the call.
4174 unsigned int ksize(const void *objp
)
4176 if (unlikely(objp
== NULL
))
4179 return obj_size(virt_to_cache(objp
));