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/slab.h>
91 #include <linux/poison.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/cpuset.h>
98 #include <linux/seq_file.h>
99 #include <linux/notifier.h>
100 #include <linux/kallsyms.h>
101 #include <linux/cpu.h>
102 #include <linux/sysctl.h>
103 #include <linux/module.h>
104 #include <linux/rcupdate.h>
105 #include <linux/string.h>
106 #include <linux/uaccess.h>
107 #include <linux/nodemask.h>
108 #include <linux/mempolicy.h>
109 #include <linux/mutex.h>
110 #include <linux/fault-inject.h>
111 #include <linux/rtmutex.h>
112 #include <linux/reciprocal_div.h>
114 #include <asm/cacheflush.h>
115 #include <asm/tlbflush.h>
116 #include <asm/page.h>
119 * DEBUG - 1 for kmem_cache_create() to honour; 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 the alignment of a 64-bit integer.
152 * ARCH_KMALLOC_MINALIGN allows that.
153 * Note that increasing this value may disable some debug features.
155 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
158 #ifndef ARCH_SLAB_MINALIGN
160 * Enforce a minimum alignment for all caches.
161 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
162 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
163 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
164 * some debug features.
166 #define ARCH_SLAB_MINALIGN 0
169 #ifndef ARCH_KMALLOC_FLAGS
170 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
173 /* Legal flag mask for kmem_cache_create(). */
175 # define CREATE_MASK (SLAB_RED_ZONE | \
176 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
179 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
180 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
182 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
184 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
185 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
191 * Bufctl's are used for linking objs within a slab
194 * This implementation relies on "struct page" for locating the cache &
195 * slab an object belongs to.
196 * This allows the bufctl structure to be small (one int), but limits
197 * the number of objects a slab (not a cache) can contain when off-slab
198 * bufctls are used. The limit is the size of the largest general cache
199 * that does not use off-slab slabs.
200 * For 32bit archs with 4 kB pages, is this 56.
201 * This is not serious, as it is only for large objects, when it is unwise
202 * to have too many per slab.
203 * Note: This limit can be raised by introducing a general cache whose size
204 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
207 typedef unsigned int kmem_bufctl_t
;
208 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
209 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
210 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
211 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
216 * Manages the objs in a slab. Placed either at the beginning of mem allocated
217 * for a slab, or allocated from an general cache.
218 * Slabs are chained into three list: fully used, partial, fully free slabs.
221 struct list_head list
;
222 unsigned long colouroff
;
223 void *s_mem
; /* including colour offset */
224 unsigned int inuse
; /* num of objs active in slab */
226 unsigned short nodeid
;
232 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
233 * arrange for kmem_freepages to be called via RCU. This is useful if
234 * we need to approach a kernel structure obliquely, from its address
235 * obtained without the usual locking. We can lock the structure to
236 * stabilize it and check it's still at the given address, only if we
237 * can be sure that the memory has not been meanwhile reused for some
238 * other kind of object (which our subsystem's lock might corrupt).
240 * rcu_read_lock before reading the address, then rcu_read_unlock after
241 * taking the spinlock within the structure expected at that address.
243 * We assume struct slab_rcu can overlay struct slab when destroying.
246 struct rcu_head head
;
247 struct kmem_cache
*cachep
;
255 * - LIFO ordering, to hand out cache-warm objects from _alloc
256 * - reduce the number of linked list operations
257 * - reduce spinlock operations
259 * The limit is stored in the per-cpu structure to reduce the data cache
266 unsigned int batchcount
;
267 unsigned int touched
;
270 * Must have this definition in here for the proper
271 * alignment of array_cache. Also simplifies accessing
273 * [0] is for gcc 2.95. It should really be [].
278 * bootstrap: The caches do not work without cpuarrays anymore, but the
279 * cpuarrays are allocated from the generic caches...
281 #define BOOT_CPUCACHE_ENTRIES 1
282 struct arraycache_init
{
283 struct array_cache cache
;
284 void *entries
[BOOT_CPUCACHE_ENTRIES
];
288 * The slab lists for all objects.
291 struct list_head slabs_partial
; /* partial list first, better asm code */
292 struct list_head slabs_full
;
293 struct list_head slabs_free
;
294 unsigned long free_objects
;
295 unsigned int free_limit
;
296 unsigned int colour_next
; /* Per-node cache coloring */
297 spinlock_t list_lock
;
298 struct array_cache
*shared
; /* shared per node */
299 struct array_cache
**alien
; /* on other nodes */
300 unsigned long next_reap
; /* updated without locking */
301 int free_touched
; /* updated without locking */
305 * Need this for bootstrapping a per node allocator.
307 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
308 struct kmem_list3 __initdata initkmem_list3
[NUM_INIT_LISTS
];
309 #define CACHE_CACHE 0
311 #define SIZE_L3 (1 + MAX_NUMNODES)
313 static int drain_freelist(struct kmem_cache
*cache
,
314 struct kmem_list3
*l3
, int tofree
);
315 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
317 static int enable_cpucache(struct kmem_cache
*cachep
);
318 static void cache_reap(struct work_struct
*unused
);
321 * This function must be completely optimized away if a constant is passed to
322 * it. Mostly the same as what is in linux/slab.h except it returns an index.
324 static __always_inline
int index_of(const size_t size
)
326 extern void __bad_size(void);
328 if (__builtin_constant_p(size
)) {
336 #include "linux/kmalloc_sizes.h"
344 static int slab_early_init
= 1;
346 #define INDEX_AC index_of(sizeof(struct arraycache_init))
347 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
349 static void kmem_list3_init(struct kmem_list3
*parent
)
351 INIT_LIST_HEAD(&parent
->slabs_full
);
352 INIT_LIST_HEAD(&parent
->slabs_partial
);
353 INIT_LIST_HEAD(&parent
->slabs_free
);
354 parent
->shared
= NULL
;
355 parent
->alien
= NULL
;
356 parent
->colour_next
= 0;
357 spin_lock_init(&parent
->list_lock
);
358 parent
->free_objects
= 0;
359 parent
->free_touched
= 0;
362 #define MAKE_LIST(cachep, listp, slab, nodeid) \
364 INIT_LIST_HEAD(listp); \
365 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
368 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
370 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
371 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
372 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
382 /* 1) per-cpu data, touched during every alloc/free */
383 struct array_cache
*array
[NR_CPUS
];
384 /* 2) Cache tunables. Protected by cache_chain_mutex */
385 unsigned int batchcount
;
389 unsigned int buffer_size
;
390 u32 reciprocal_buffer_size
;
391 /* 3) touched by every alloc & free from the backend */
393 unsigned int flags
; /* constant flags */
394 unsigned int num
; /* # of objs per slab */
396 /* 4) cache_grow/shrink */
397 /* order of pgs per slab (2^n) */
398 unsigned int gfporder
;
400 /* force GFP flags, e.g. GFP_DMA */
403 size_t colour
; /* cache colouring range */
404 unsigned int colour_off
; /* colour offset */
405 struct kmem_cache
*slabp_cache
;
406 unsigned int slab_size
;
407 unsigned int dflags
; /* dynamic flags */
409 /* constructor func */
410 void (*ctor
) (void *, struct kmem_cache
*, unsigned long);
412 /* de-constructor func */
413 void (*dtor
) (void *, struct kmem_cache
*, unsigned long);
415 /* 5) cache creation/removal */
417 struct list_head next
;
421 unsigned long num_active
;
422 unsigned long num_allocations
;
423 unsigned long high_mark
;
425 unsigned long reaped
;
426 unsigned long errors
;
427 unsigned long max_freeable
;
428 unsigned long node_allocs
;
429 unsigned long node_frees
;
430 unsigned long node_overflow
;
438 * If debugging is enabled, then the allocator can add additional
439 * fields and/or padding to every object. buffer_size contains the total
440 * object size including these internal fields, the following two
441 * variables contain the offset to the user object and its size.
447 * We put nodelists[] at the end of kmem_cache, because we want to size
448 * this array to nr_node_ids slots instead of MAX_NUMNODES
449 * (see kmem_cache_init())
450 * We still use [MAX_NUMNODES] and not [1] or [0] because cache_cache
451 * is statically defined, so we reserve the max number of nodes.
453 struct kmem_list3
*nodelists
[MAX_NUMNODES
];
455 * Do not add fields after nodelists[]
459 #define CFLGS_OFF_SLAB (0x80000000UL)
460 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
462 #define BATCHREFILL_LIMIT 16
464 * Optimization question: fewer reaps means less probability for unnessary
465 * cpucache drain/refill cycles.
467 * OTOH the cpuarrays can contain lots of objects,
468 * which could lock up otherwise freeable slabs.
470 #define REAPTIMEOUT_CPUC (2*HZ)
471 #define REAPTIMEOUT_LIST3 (4*HZ)
474 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
475 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
476 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
477 #define STATS_INC_GROWN(x) ((x)->grown++)
478 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
479 #define STATS_SET_HIGH(x) \
481 if ((x)->num_active > (x)->high_mark) \
482 (x)->high_mark = (x)->num_active; \
484 #define STATS_INC_ERR(x) ((x)->errors++)
485 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
486 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
487 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
488 #define STATS_SET_FREEABLE(x, i) \
490 if ((x)->max_freeable < i) \
491 (x)->max_freeable = i; \
493 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
494 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
495 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
496 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
498 #define STATS_INC_ACTIVE(x) do { } while (0)
499 #define STATS_DEC_ACTIVE(x) do { } while (0)
500 #define STATS_INC_ALLOCED(x) do { } while (0)
501 #define STATS_INC_GROWN(x) do { } while (0)
502 #define STATS_ADD_REAPED(x,y) do { } while (0)
503 #define STATS_SET_HIGH(x) do { } while (0)
504 #define STATS_INC_ERR(x) do { } while (0)
505 #define STATS_INC_NODEALLOCS(x) do { } while (0)
506 #define STATS_INC_NODEFREES(x) do { } while (0)
507 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
508 #define STATS_SET_FREEABLE(x, i) do { } while (0)
509 #define STATS_INC_ALLOCHIT(x) do { } while (0)
510 #define STATS_INC_ALLOCMISS(x) do { } while (0)
511 #define STATS_INC_FREEHIT(x) do { } while (0)
512 #define STATS_INC_FREEMISS(x) do { } while (0)
518 * memory layout of objects:
520 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
521 * the end of an object is aligned with the end of the real
522 * allocation. Catches writes behind the end of the allocation.
523 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
525 * cachep->obj_offset: The real object.
526 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
527 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
528 * [BYTES_PER_WORD long]
530 static int obj_offset(struct kmem_cache
*cachep
)
532 return cachep
->obj_offset
;
535 static int obj_size(struct kmem_cache
*cachep
)
537 return cachep
->obj_size
;
540 static unsigned long long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
542 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
543 return (unsigned long long*) (objp
+ obj_offset(cachep
) -
544 sizeof(unsigned long long));
547 static unsigned long long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
549 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
550 if (cachep
->flags
& SLAB_STORE_USER
)
551 return (unsigned long long *)(objp
+ cachep
->buffer_size
-
552 sizeof(unsigned long long) -
554 return (unsigned long long *) (objp
+ cachep
->buffer_size
-
555 sizeof(unsigned long long));
558 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
560 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
561 return (void **)(objp
+ cachep
->buffer_size
- BYTES_PER_WORD
);
566 #define obj_offset(x) 0
567 #define obj_size(cachep) (cachep->buffer_size)
568 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
569 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
570 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
575 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
578 #if defined(CONFIG_LARGE_ALLOCS)
579 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
580 #define MAX_GFP_ORDER 13 /* up to 32Mb */
581 #elif defined(CONFIG_MMU)
582 #define MAX_OBJ_ORDER 5 /* 32 pages */
583 #define MAX_GFP_ORDER 5 /* 32 pages */
585 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
586 #define MAX_GFP_ORDER 8 /* up to 1Mb */
590 * Do not go above this order unless 0 objects fit into the slab.
592 #define BREAK_GFP_ORDER_HI 1
593 #define BREAK_GFP_ORDER_LO 0
594 static int slab_break_gfp_order
= BREAK_GFP_ORDER_LO
;
597 * Functions for storing/retrieving the cachep and or slab from the page
598 * allocator. These are used to find the slab an obj belongs to. With kfree(),
599 * these are used to find the cache which an obj belongs to.
601 static inline void page_set_cache(struct page
*page
, struct kmem_cache
*cache
)
603 page
->lru
.next
= (struct list_head
*)cache
;
606 static inline struct kmem_cache
*page_get_cache(struct page
*page
)
608 page
= compound_head(page
);
609 BUG_ON(!PageSlab(page
));
610 return (struct kmem_cache
*)page
->lru
.next
;
613 static inline void page_set_slab(struct page
*page
, struct slab
*slab
)
615 page
->lru
.prev
= (struct list_head
*)slab
;
618 static inline struct slab
*page_get_slab(struct page
*page
)
620 BUG_ON(!PageSlab(page
));
621 return (struct slab
*)page
->lru
.prev
;
624 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
626 struct page
*page
= virt_to_head_page(obj
);
627 return page_get_cache(page
);
630 static inline struct slab
*virt_to_slab(const void *obj
)
632 struct page
*page
= virt_to_head_page(obj
);
633 return page_get_slab(page
);
636 static inline void *index_to_obj(struct kmem_cache
*cache
, struct slab
*slab
,
639 return slab
->s_mem
+ cache
->buffer_size
* idx
;
643 * We want to avoid an expensive divide : (offset / cache->buffer_size)
644 * Using the fact that buffer_size is a constant for a particular cache,
645 * we can replace (offset / cache->buffer_size) by
646 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
648 static inline unsigned int obj_to_index(const struct kmem_cache
*cache
,
649 const struct slab
*slab
, void *obj
)
651 u32 offset
= (obj
- slab
->s_mem
);
652 return reciprocal_divide(offset
, cache
->reciprocal_buffer_size
);
656 * These are the default caches for kmalloc. Custom caches can have other sizes.
658 struct cache_sizes malloc_sizes
[] = {
659 #define CACHE(x) { .cs_size = (x) },
660 #include <linux/kmalloc_sizes.h>
664 EXPORT_SYMBOL(malloc_sizes
);
666 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
672 static struct cache_names __initdata cache_names
[] = {
673 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
674 #include <linux/kmalloc_sizes.h>
679 static struct arraycache_init initarray_cache __initdata
=
680 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
681 static struct arraycache_init initarray_generic
=
682 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
684 /* internal cache of cache description objs */
685 static struct kmem_cache cache_cache
= {
687 .limit
= BOOT_CPUCACHE_ENTRIES
,
689 .buffer_size
= sizeof(struct kmem_cache
),
690 .name
= "kmem_cache",
693 #define BAD_ALIEN_MAGIC 0x01020304ul
695 #ifdef CONFIG_LOCKDEP
698 * Slab sometimes uses the kmalloc slabs to store the slab headers
699 * for other slabs "off slab".
700 * The locking for this is tricky in that it nests within the locks
701 * of all other slabs in a few places; to deal with this special
702 * locking we put on-slab caches into a separate lock-class.
704 * We set lock class for alien array caches which are up during init.
705 * The lock annotation will be lost if all cpus of a node goes down and
706 * then comes back up during hotplug
708 static struct lock_class_key on_slab_l3_key
;
709 static struct lock_class_key on_slab_alc_key
;
711 static inline void init_lock_keys(void)
715 struct cache_sizes
*s
= malloc_sizes
;
717 while (s
->cs_size
!= ULONG_MAX
) {
719 struct array_cache
**alc
;
721 struct kmem_list3
*l3
= s
->cs_cachep
->nodelists
[q
];
722 if (!l3
|| OFF_SLAB(s
->cs_cachep
))
724 lockdep_set_class(&l3
->list_lock
, &on_slab_l3_key
);
727 * FIXME: This check for BAD_ALIEN_MAGIC
728 * should go away when common slab code is taught to
729 * work even without alien caches.
730 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
731 * for alloc_alien_cache,
733 if (!alc
|| (unsigned long)alc
== BAD_ALIEN_MAGIC
)
737 lockdep_set_class(&alc
[r
]->lock
,
745 static inline void init_lock_keys(void)
751 * 1. Guard access to the cache-chain.
752 * 2. Protect sanity of cpu_online_map against cpu hotplug events
754 static DEFINE_MUTEX(cache_chain_mutex
);
755 static struct list_head cache_chain
;
758 * chicken and egg problem: delay the per-cpu array allocation
759 * until the general caches are up.
769 * used by boot code to determine if it can use slab based allocator
771 int slab_is_available(void)
773 return g_cpucache_up
== FULL
;
776 static DEFINE_PER_CPU(struct delayed_work
, reap_work
);
778 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
780 return cachep
->array
[smp_processor_id()];
783 static inline struct kmem_cache
*__find_general_cachep(size_t size
,
786 struct cache_sizes
*csizep
= malloc_sizes
;
789 /* This happens if someone tries to call
790 * kmem_cache_create(), or __kmalloc(), before
791 * the generic caches are initialized.
793 BUG_ON(malloc_sizes
[INDEX_AC
].cs_cachep
== NULL
);
795 while (size
> csizep
->cs_size
)
799 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
800 * has cs_{dma,}cachep==NULL. Thus no special case
801 * for large kmalloc calls required.
803 #ifdef CONFIG_ZONE_DMA
804 if (unlikely(gfpflags
& GFP_DMA
))
805 return csizep
->cs_dmacachep
;
807 return csizep
->cs_cachep
;
810 static struct kmem_cache
*kmem_find_general_cachep(size_t size
, gfp_t gfpflags
)
812 return __find_general_cachep(size
, gfpflags
);
815 static size_t slab_mgmt_size(size_t nr_objs
, size_t align
)
817 return ALIGN(sizeof(struct slab
)+nr_objs
*sizeof(kmem_bufctl_t
), align
);
821 * Calculate the number of objects and left-over bytes for a given buffer size.
823 static void cache_estimate(unsigned long gfporder
, size_t buffer_size
,
824 size_t align
, int flags
, size_t *left_over
,
829 size_t slab_size
= PAGE_SIZE
<< gfporder
;
832 * The slab management structure can be either off the slab or
833 * on it. For the latter case, the memory allocated for a
837 * - One kmem_bufctl_t for each object
838 * - Padding to respect alignment of @align
839 * - @buffer_size bytes for each object
841 * If the slab management structure is off the slab, then the
842 * alignment will already be calculated into the size. Because
843 * the slabs are all pages aligned, the objects will be at the
844 * correct alignment when allocated.
846 if (flags
& CFLGS_OFF_SLAB
) {
848 nr_objs
= slab_size
/ buffer_size
;
850 if (nr_objs
> SLAB_LIMIT
)
851 nr_objs
= SLAB_LIMIT
;
854 * Ignore padding for the initial guess. The padding
855 * is at most @align-1 bytes, and @buffer_size is at
856 * least @align. In the worst case, this result will
857 * be one greater than the number of objects that fit
858 * into the memory allocation when taking the padding
861 nr_objs
= (slab_size
- sizeof(struct slab
)) /
862 (buffer_size
+ sizeof(kmem_bufctl_t
));
865 * This calculated number will be either the right
866 * amount, or one greater than what we want.
868 if (slab_mgmt_size(nr_objs
, align
) + nr_objs
*buffer_size
872 if (nr_objs
> SLAB_LIMIT
)
873 nr_objs
= SLAB_LIMIT
;
875 mgmt_size
= slab_mgmt_size(nr_objs
, align
);
878 *left_over
= slab_size
- nr_objs
*buffer_size
- mgmt_size
;
881 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
883 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
886 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
887 function
, cachep
->name
, msg
);
892 * By default on NUMA we use alien caches to stage the freeing of
893 * objects allocated from other nodes. This causes massive memory
894 * inefficiencies when using fake NUMA setup to split memory into a
895 * large number of small nodes, so it can be disabled on the command
899 static int use_alien_caches __read_mostly
= 1;
900 static int __init
noaliencache_setup(char *s
)
902 use_alien_caches
= 0;
905 __setup("noaliencache", noaliencache_setup
);
909 * Special reaping functions for NUMA systems called from cache_reap().
910 * These take care of doing round robin flushing of alien caches (containing
911 * objects freed on different nodes from which they were allocated) and the
912 * flushing of remote pcps by calling drain_node_pages.
914 static DEFINE_PER_CPU(unsigned long, reap_node
);
916 static void init_reap_node(int cpu
)
920 node
= next_node(cpu_to_node(cpu
), node_online_map
);
921 if (node
== MAX_NUMNODES
)
922 node
= first_node(node_online_map
);
924 per_cpu(reap_node
, cpu
) = node
;
927 static void next_reap_node(void)
929 int node
= __get_cpu_var(reap_node
);
931 node
= next_node(node
, node_online_map
);
932 if (unlikely(node
>= MAX_NUMNODES
))
933 node
= first_node(node_online_map
);
934 __get_cpu_var(reap_node
) = node
;
938 #define init_reap_node(cpu) do { } while (0)
939 #define next_reap_node(void) do { } while (0)
943 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
944 * via the workqueue/eventd.
945 * Add the CPU number into the expiration time to minimize the possibility of
946 * the CPUs getting into lockstep and contending for the global cache chain
949 static void __devinit
start_cpu_timer(int cpu
)
951 struct delayed_work
*reap_work
= &per_cpu(reap_work
, cpu
);
954 * When this gets called from do_initcalls via cpucache_init(),
955 * init_workqueues() has already run, so keventd will be setup
958 if (keventd_up() && reap_work
->work
.func
== NULL
) {
960 INIT_DELAYED_WORK(reap_work
, cache_reap
);
961 schedule_delayed_work_on(cpu
, reap_work
,
962 __round_jiffies_relative(HZ
, cpu
));
966 static struct array_cache
*alloc_arraycache(int node
, int entries
,
969 int memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
970 struct array_cache
*nc
= NULL
;
972 nc
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
976 nc
->batchcount
= batchcount
;
978 spin_lock_init(&nc
->lock
);
984 * Transfer objects in one arraycache to another.
985 * Locking must be handled by the caller.
987 * Return the number of entries transferred.
989 static int transfer_objects(struct array_cache
*to
,
990 struct array_cache
*from
, unsigned int max
)
992 /* Figure out how many entries to transfer */
993 int nr
= min(min(from
->avail
, max
), to
->limit
- to
->avail
);
998 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
1009 #define drain_alien_cache(cachep, alien) do { } while (0)
1010 #define reap_alien(cachep, l3) do { } while (0)
1012 static inline struct array_cache
**alloc_alien_cache(int node
, int limit
)
1014 return (struct array_cache
**)BAD_ALIEN_MAGIC
;
1017 static inline void free_alien_cache(struct array_cache
**ac_ptr
)
1021 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1026 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
1032 static inline void *____cache_alloc_node(struct kmem_cache
*cachep
,
1033 gfp_t flags
, int nodeid
)
1038 #else /* CONFIG_NUMA */
1040 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
1041 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
1043 static struct array_cache
**alloc_alien_cache(int node
, int limit
)
1045 struct array_cache
**ac_ptr
;
1046 int memsize
= sizeof(void *) * nr_node_ids
;
1051 ac_ptr
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1054 if (i
== node
|| !node_online(i
)) {
1058 ac_ptr
[i
] = alloc_arraycache(node
, limit
, 0xbaadf00d);
1060 for (i
--; i
<= 0; i
--)
1070 static void free_alien_cache(struct array_cache
**ac_ptr
)
1081 static void __drain_alien_cache(struct kmem_cache
*cachep
,
1082 struct array_cache
*ac
, int node
)
1084 struct kmem_list3
*rl3
= cachep
->nodelists
[node
];
1087 spin_lock(&rl3
->list_lock
);
1089 * Stuff objects into the remote nodes shared array first.
1090 * That way we could avoid the overhead of putting the objects
1091 * into the free lists and getting them back later.
1094 transfer_objects(rl3
->shared
, ac
, ac
->limit
);
1096 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
1098 spin_unlock(&rl3
->list_lock
);
1103 * Called from cache_reap() to regularly drain alien caches round robin.
1105 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_list3
*l3
)
1107 int node
= __get_cpu_var(reap_node
);
1110 struct array_cache
*ac
= l3
->alien
[node
];
1112 if (ac
&& ac
->avail
&& spin_trylock_irq(&ac
->lock
)) {
1113 __drain_alien_cache(cachep
, ac
, node
);
1114 spin_unlock_irq(&ac
->lock
);
1119 static void drain_alien_cache(struct kmem_cache
*cachep
,
1120 struct array_cache
**alien
)
1123 struct array_cache
*ac
;
1124 unsigned long flags
;
1126 for_each_online_node(i
) {
1129 spin_lock_irqsave(&ac
->lock
, flags
);
1130 __drain_alien_cache(cachep
, ac
, i
);
1131 spin_unlock_irqrestore(&ac
->lock
, flags
);
1136 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1138 struct slab
*slabp
= virt_to_slab(objp
);
1139 int nodeid
= slabp
->nodeid
;
1140 struct kmem_list3
*l3
;
1141 struct array_cache
*alien
= NULL
;
1144 node
= numa_node_id();
1147 * Make sure we are not freeing a object from another node to the array
1148 * cache on this cpu.
1150 if (likely(slabp
->nodeid
== node
))
1153 l3
= cachep
->nodelists
[node
];
1154 STATS_INC_NODEFREES(cachep
);
1155 if (l3
->alien
&& l3
->alien
[nodeid
]) {
1156 alien
= l3
->alien
[nodeid
];
1157 spin_lock(&alien
->lock
);
1158 if (unlikely(alien
->avail
== alien
->limit
)) {
1159 STATS_INC_ACOVERFLOW(cachep
);
1160 __drain_alien_cache(cachep
, alien
, nodeid
);
1162 alien
->entry
[alien
->avail
++] = objp
;
1163 spin_unlock(&alien
->lock
);
1165 spin_lock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1166 free_block(cachep
, &objp
, 1, nodeid
);
1167 spin_unlock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1173 static int __cpuinit
cpuup_callback(struct notifier_block
*nfb
,
1174 unsigned long action
, void *hcpu
)
1176 long cpu
= (long)hcpu
;
1177 struct kmem_cache
*cachep
;
1178 struct kmem_list3
*l3
= NULL
;
1179 int node
= cpu_to_node(cpu
);
1180 int memsize
= sizeof(struct kmem_list3
);
1183 case CPU_LOCK_ACQUIRE
:
1184 mutex_lock(&cache_chain_mutex
);
1186 case CPU_UP_PREPARE
:
1187 case CPU_UP_PREPARE_FROZEN
:
1189 * We need to do this right in the beginning since
1190 * alloc_arraycache's are going to use this list.
1191 * kmalloc_node allows us to add the slab to the right
1192 * kmem_list3 and not this cpu's kmem_list3
1195 list_for_each_entry(cachep
, &cache_chain
, next
) {
1197 * Set up the size64 kmemlist for cpu before we can
1198 * begin anything. Make sure some other cpu on this
1199 * node has not already allocated this
1201 if (!cachep
->nodelists
[node
]) {
1202 l3
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1205 kmem_list3_init(l3
);
1206 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
1207 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1210 * The l3s don't come and go as CPUs come and
1211 * go. cache_chain_mutex is sufficient
1214 cachep
->nodelists
[node
] = l3
;
1217 spin_lock_irq(&cachep
->nodelists
[node
]->list_lock
);
1218 cachep
->nodelists
[node
]->free_limit
=
1219 (1 + nr_cpus_node(node
)) *
1220 cachep
->batchcount
+ cachep
->num
;
1221 spin_unlock_irq(&cachep
->nodelists
[node
]->list_lock
);
1225 * Now we can go ahead with allocating the shared arrays and
1228 list_for_each_entry(cachep
, &cache_chain
, next
) {
1229 struct array_cache
*nc
;
1230 struct array_cache
*shared
= NULL
;
1231 struct array_cache
**alien
= NULL
;
1233 nc
= alloc_arraycache(node
, cachep
->limit
,
1234 cachep
->batchcount
);
1237 if (cachep
->shared
) {
1238 shared
= alloc_arraycache(node
,
1239 cachep
->shared
* cachep
->batchcount
,
1244 if (use_alien_caches
) {
1245 alien
= alloc_alien_cache(node
, cachep
->limit
);
1249 cachep
->array
[cpu
] = nc
;
1250 l3
= cachep
->nodelists
[node
];
1253 spin_lock_irq(&l3
->list_lock
);
1256 * We are serialised from CPU_DEAD or
1257 * CPU_UP_CANCELLED by the cpucontrol lock
1259 l3
->shared
= shared
;
1268 spin_unlock_irq(&l3
->list_lock
);
1270 free_alien_cache(alien
);
1274 case CPU_ONLINE_FROZEN
:
1275 start_cpu_timer(cpu
);
1277 #ifdef CONFIG_HOTPLUG_CPU
1278 case CPU_DOWN_PREPARE
:
1279 case CPU_DOWN_PREPARE_FROZEN
:
1281 * Shutdown cache reaper. Note that the cache_chain_mutex is
1282 * held so that if cache_reap() is invoked it cannot do
1283 * anything expensive but will only modify reap_work
1284 * and reschedule the timer.
1286 cancel_rearming_delayed_work(&per_cpu(reap_work
, cpu
));
1287 /* Now the cache_reaper is guaranteed to be not running. */
1288 per_cpu(reap_work
, cpu
).work
.func
= NULL
;
1290 case CPU_DOWN_FAILED
:
1291 case CPU_DOWN_FAILED_FROZEN
:
1292 start_cpu_timer(cpu
);
1295 case CPU_DEAD_FROZEN
:
1297 * Even if all the cpus of a node are down, we don't free the
1298 * kmem_list3 of any cache. This to avoid a race between
1299 * cpu_down, and a kmalloc allocation from another cpu for
1300 * memory from the node of the cpu going down. The list3
1301 * structure is usually allocated from kmem_cache_create() and
1302 * gets destroyed at kmem_cache_destroy().
1306 case CPU_UP_CANCELED
:
1307 case CPU_UP_CANCELED_FROZEN
:
1308 list_for_each_entry(cachep
, &cache_chain
, next
) {
1309 struct array_cache
*nc
;
1310 struct array_cache
*shared
;
1311 struct array_cache
**alien
;
1314 mask
= node_to_cpumask(node
);
1315 /* cpu is dead; no one can alloc from it. */
1316 nc
= cachep
->array
[cpu
];
1317 cachep
->array
[cpu
] = NULL
;
1318 l3
= cachep
->nodelists
[node
];
1321 goto free_array_cache
;
1323 spin_lock_irq(&l3
->list_lock
);
1325 /* Free limit for this kmem_list3 */
1326 l3
->free_limit
-= cachep
->batchcount
;
1328 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
1330 if (!cpus_empty(mask
)) {
1331 spin_unlock_irq(&l3
->list_lock
);
1332 goto free_array_cache
;
1335 shared
= l3
->shared
;
1337 free_block(cachep
, shared
->entry
,
1338 shared
->avail
, node
);
1345 spin_unlock_irq(&l3
->list_lock
);
1349 drain_alien_cache(cachep
, alien
);
1350 free_alien_cache(alien
);
1356 * In the previous loop, all the objects were freed to
1357 * the respective cache's slabs, now we can go ahead and
1358 * shrink each nodelist to its limit.
1360 list_for_each_entry(cachep
, &cache_chain
, next
) {
1361 l3
= cachep
->nodelists
[node
];
1364 drain_freelist(cachep
, l3
, l3
->free_objects
);
1367 case CPU_LOCK_RELEASE
:
1368 mutex_unlock(&cache_chain_mutex
);
1376 static struct notifier_block __cpuinitdata cpucache_notifier
= {
1377 &cpuup_callback
, NULL
, 0
1381 * swap the static kmem_list3 with kmalloced memory
1383 static void init_list(struct kmem_cache
*cachep
, struct kmem_list3
*list
,
1386 struct kmem_list3
*ptr
;
1388 ptr
= kmalloc_node(sizeof(struct kmem_list3
), GFP_KERNEL
, nodeid
);
1391 local_irq_disable();
1392 memcpy(ptr
, list
, sizeof(struct kmem_list3
));
1394 * Do not assume that spinlocks can be initialized via memcpy:
1396 spin_lock_init(&ptr
->list_lock
);
1398 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1399 cachep
->nodelists
[nodeid
] = ptr
;
1404 * Initialisation. Called after the page allocator have been initialised and
1405 * before smp_init().
1407 void __init
kmem_cache_init(void)
1410 struct cache_sizes
*sizes
;
1411 struct cache_names
*names
;
1416 if (num_possible_nodes() == 1)
1417 use_alien_caches
= 0;
1419 for (i
= 0; i
< NUM_INIT_LISTS
; i
++) {
1420 kmem_list3_init(&initkmem_list3
[i
]);
1421 if (i
< MAX_NUMNODES
)
1422 cache_cache
.nodelists
[i
] = NULL
;
1426 * Fragmentation resistance on low memory - only use bigger
1427 * page orders on machines with more than 32MB of memory.
1429 if (num_physpages
> (32 << 20) >> PAGE_SHIFT
)
1430 slab_break_gfp_order
= BREAK_GFP_ORDER_HI
;
1432 /* Bootstrap is tricky, because several objects are allocated
1433 * from caches that do not exist yet:
1434 * 1) initialize the cache_cache cache: it contains the struct
1435 * kmem_cache structures of all caches, except cache_cache itself:
1436 * cache_cache is statically allocated.
1437 * Initially an __init data area is used for the head array and the
1438 * kmem_list3 structures, it's replaced with a kmalloc allocated
1439 * array at the end of the bootstrap.
1440 * 2) Create the first kmalloc cache.
1441 * The struct kmem_cache for the new cache is allocated normally.
1442 * An __init data area is used for the head array.
1443 * 3) Create the remaining kmalloc caches, with minimally sized
1445 * 4) Replace the __init data head arrays for cache_cache and the first
1446 * kmalloc cache with kmalloc allocated arrays.
1447 * 5) Replace the __init data for kmem_list3 for cache_cache and
1448 * the other cache's with kmalloc allocated memory.
1449 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1452 node
= numa_node_id();
1454 /* 1) create the cache_cache */
1455 INIT_LIST_HEAD(&cache_chain
);
1456 list_add(&cache_cache
.next
, &cache_chain
);
1457 cache_cache
.colour_off
= cache_line_size();
1458 cache_cache
.array
[smp_processor_id()] = &initarray_cache
.cache
;
1459 cache_cache
.nodelists
[node
] = &initkmem_list3
[CACHE_CACHE
];
1462 * struct kmem_cache size depends on nr_node_ids, which
1463 * can be less than MAX_NUMNODES.
1465 cache_cache
.buffer_size
= offsetof(struct kmem_cache
, nodelists
) +
1466 nr_node_ids
* sizeof(struct kmem_list3
*);
1468 cache_cache
.obj_size
= cache_cache
.buffer_size
;
1470 cache_cache
.buffer_size
= ALIGN(cache_cache
.buffer_size
,
1472 cache_cache
.reciprocal_buffer_size
=
1473 reciprocal_value(cache_cache
.buffer_size
);
1475 for (order
= 0; order
< MAX_ORDER
; order
++) {
1476 cache_estimate(order
, cache_cache
.buffer_size
,
1477 cache_line_size(), 0, &left_over
, &cache_cache
.num
);
1478 if (cache_cache
.num
)
1481 BUG_ON(!cache_cache
.num
);
1482 cache_cache
.gfporder
= order
;
1483 cache_cache
.colour
= left_over
/ cache_cache
.colour_off
;
1484 cache_cache
.slab_size
= ALIGN(cache_cache
.num
* sizeof(kmem_bufctl_t
) +
1485 sizeof(struct slab
), cache_line_size());
1487 /* 2+3) create the kmalloc caches */
1488 sizes
= malloc_sizes
;
1489 names
= cache_names
;
1492 * Initialize the caches that provide memory for the array cache and the
1493 * kmem_list3 structures first. Without this, further allocations will
1497 sizes
[INDEX_AC
].cs_cachep
= kmem_cache_create(names
[INDEX_AC
].name
,
1498 sizes
[INDEX_AC
].cs_size
,
1499 ARCH_KMALLOC_MINALIGN
,
1500 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1503 if (INDEX_AC
!= INDEX_L3
) {
1504 sizes
[INDEX_L3
].cs_cachep
=
1505 kmem_cache_create(names
[INDEX_L3
].name
,
1506 sizes
[INDEX_L3
].cs_size
,
1507 ARCH_KMALLOC_MINALIGN
,
1508 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1512 slab_early_init
= 0;
1514 while (sizes
->cs_size
!= ULONG_MAX
) {
1516 * For performance, all the general caches are L1 aligned.
1517 * This should be particularly beneficial on SMP boxes, as it
1518 * eliminates "false sharing".
1519 * Note for systems short on memory removing the alignment will
1520 * allow tighter packing of the smaller caches.
1522 if (!sizes
->cs_cachep
) {
1523 sizes
->cs_cachep
= kmem_cache_create(names
->name
,
1525 ARCH_KMALLOC_MINALIGN
,
1526 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1529 #ifdef CONFIG_ZONE_DMA
1530 sizes
->cs_dmacachep
= kmem_cache_create(
1533 ARCH_KMALLOC_MINALIGN
,
1534 ARCH_KMALLOC_FLAGS
|SLAB_CACHE_DMA
|
1541 /* 4) Replace the bootstrap head arrays */
1543 struct array_cache
*ptr
;
1545 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1547 local_irq_disable();
1548 BUG_ON(cpu_cache_get(&cache_cache
) != &initarray_cache
.cache
);
1549 memcpy(ptr
, cpu_cache_get(&cache_cache
),
1550 sizeof(struct arraycache_init
));
1552 * Do not assume that spinlocks can be initialized via memcpy:
1554 spin_lock_init(&ptr
->lock
);
1556 cache_cache
.array
[smp_processor_id()] = ptr
;
1559 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1561 local_irq_disable();
1562 BUG_ON(cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
)
1563 != &initarray_generic
.cache
);
1564 memcpy(ptr
, cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
),
1565 sizeof(struct arraycache_init
));
1567 * Do not assume that spinlocks can be initialized via memcpy:
1569 spin_lock_init(&ptr
->lock
);
1571 malloc_sizes
[INDEX_AC
].cs_cachep
->array
[smp_processor_id()] =
1575 /* 5) Replace the bootstrap kmem_list3's */
1579 /* Replace the static kmem_list3 structures for the boot cpu */
1580 init_list(&cache_cache
, &initkmem_list3
[CACHE_CACHE
], node
);
1582 for_each_online_node(nid
) {
1583 init_list(malloc_sizes
[INDEX_AC
].cs_cachep
,
1584 &initkmem_list3
[SIZE_AC
+ nid
], nid
);
1586 if (INDEX_AC
!= INDEX_L3
) {
1587 init_list(malloc_sizes
[INDEX_L3
].cs_cachep
,
1588 &initkmem_list3
[SIZE_L3
+ nid
], nid
);
1593 /* 6) resize the head arrays to their final sizes */
1595 struct kmem_cache
*cachep
;
1596 mutex_lock(&cache_chain_mutex
);
1597 list_for_each_entry(cachep
, &cache_chain
, next
)
1598 if (enable_cpucache(cachep
))
1600 mutex_unlock(&cache_chain_mutex
);
1603 /* Annotate slab for lockdep -- annotate the malloc caches */
1608 g_cpucache_up
= FULL
;
1611 * Register a cpu startup notifier callback that initializes
1612 * cpu_cache_get for all new cpus
1614 register_cpu_notifier(&cpucache_notifier
);
1617 * The reap timers are started later, with a module init call: That part
1618 * of the kernel is not yet operational.
1622 static int __init
cpucache_init(void)
1627 * Register the timers that return unneeded pages to the page allocator
1629 for_each_online_cpu(cpu
)
1630 start_cpu_timer(cpu
);
1633 __initcall(cpucache_init
);
1636 * Interface to system's page allocator. No need to hold the cache-lock.
1638 * If we requested dmaable memory, we will get it. Even if we
1639 * did not request dmaable memory, we might get it, but that
1640 * would be relatively rare and ignorable.
1642 static void *kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
1650 * Nommu uses slab's for process anonymous memory allocations, and thus
1651 * requires __GFP_COMP to properly refcount higher order allocations
1653 flags
|= __GFP_COMP
;
1656 flags
|= cachep
->gfpflags
;
1658 page
= alloc_pages_node(nodeid
, flags
, cachep
->gfporder
);
1662 nr_pages
= (1 << cachep
->gfporder
);
1663 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1664 add_zone_page_state(page_zone(page
),
1665 NR_SLAB_RECLAIMABLE
, nr_pages
);
1667 add_zone_page_state(page_zone(page
),
1668 NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1669 for (i
= 0; i
< nr_pages
; i
++)
1670 __SetPageSlab(page
+ i
);
1671 return page_address(page
);
1675 * Interface to system's page release.
1677 static void kmem_freepages(struct kmem_cache
*cachep
, void *addr
)
1679 unsigned long i
= (1 << cachep
->gfporder
);
1680 struct page
*page
= virt_to_page(addr
);
1681 const unsigned long nr_freed
= i
;
1683 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1684 sub_zone_page_state(page_zone(page
),
1685 NR_SLAB_RECLAIMABLE
, nr_freed
);
1687 sub_zone_page_state(page_zone(page
),
1688 NR_SLAB_UNRECLAIMABLE
, nr_freed
);
1690 BUG_ON(!PageSlab(page
));
1691 __ClearPageSlab(page
);
1694 if (current
->reclaim_state
)
1695 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1696 free_pages((unsigned long)addr
, cachep
->gfporder
);
1699 static void kmem_rcu_free(struct rcu_head
*head
)
1701 struct slab_rcu
*slab_rcu
= (struct slab_rcu
*)head
;
1702 struct kmem_cache
*cachep
= slab_rcu
->cachep
;
1704 kmem_freepages(cachep
, slab_rcu
->addr
);
1705 if (OFF_SLAB(cachep
))
1706 kmem_cache_free(cachep
->slabp_cache
, slab_rcu
);
1711 #ifdef CONFIG_DEBUG_PAGEALLOC
1712 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1713 unsigned long caller
)
1715 int size
= obj_size(cachep
);
1717 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1719 if (size
< 5 * sizeof(unsigned long))
1722 *addr
++ = 0x12345678;
1724 *addr
++ = smp_processor_id();
1725 size
-= 3 * sizeof(unsigned long);
1727 unsigned long *sptr
= &caller
;
1728 unsigned long svalue
;
1730 while (!kstack_end(sptr
)) {
1732 if (kernel_text_address(svalue
)) {
1734 size
-= sizeof(unsigned long);
1735 if (size
<= sizeof(unsigned long))
1741 *addr
++ = 0x87654321;
1745 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1747 int size
= obj_size(cachep
);
1748 addr
= &((char *)addr
)[obj_offset(cachep
)];
1750 memset(addr
, val
, size
);
1751 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1754 static void dump_line(char *data
, int offset
, int limit
)
1757 unsigned char error
= 0;
1760 printk(KERN_ERR
"%03x:", offset
);
1761 for (i
= 0; i
< limit
; i
++) {
1762 if (data
[offset
+ i
] != POISON_FREE
) {
1763 error
= data
[offset
+ i
];
1766 printk(" %02x", (unsigned char)data
[offset
+ i
]);
1770 if (bad_count
== 1) {
1771 error
^= POISON_FREE
;
1772 if (!(error
& (error
- 1))) {
1773 printk(KERN_ERR
"Single bit error detected. Probably "
1776 printk(KERN_ERR
"Run memtest86+ or a similar memory "
1779 printk(KERN_ERR
"Run a memory test tool.\n");
1788 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1793 if (cachep
->flags
& SLAB_RED_ZONE
) {
1794 printk(KERN_ERR
"Redzone: 0x%llx/0x%llx.\n",
1795 *dbg_redzone1(cachep
, objp
),
1796 *dbg_redzone2(cachep
, objp
));
1799 if (cachep
->flags
& SLAB_STORE_USER
) {
1800 printk(KERN_ERR
"Last user: [<%p>]",
1801 *dbg_userword(cachep
, objp
));
1802 print_symbol("(%s)",
1803 (unsigned long)*dbg_userword(cachep
, objp
));
1806 realobj
= (char *)objp
+ obj_offset(cachep
);
1807 size
= obj_size(cachep
);
1808 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1811 if (i
+ limit
> size
)
1813 dump_line(realobj
, i
, limit
);
1817 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1823 realobj
= (char *)objp
+ obj_offset(cachep
);
1824 size
= obj_size(cachep
);
1826 for (i
= 0; i
< size
; i
++) {
1827 char exp
= POISON_FREE
;
1830 if (realobj
[i
] != exp
) {
1836 "Slab corruption: %s start=%p, len=%d\n",
1837 cachep
->name
, realobj
, size
);
1838 print_objinfo(cachep
, objp
, 0);
1840 /* Hexdump the affected line */
1843 if (i
+ limit
> size
)
1845 dump_line(realobj
, i
, limit
);
1848 /* Limit to 5 lines */
1854 /* Print some data about the neighboring objects, if they
1857 struct slab
*slabp
= virt_to_slab(objp
);
1860 objnr
= obj_to_index(cachep
, slabp
, objp
);
1862 objp
= index_to_obj(cachep
, slabp
, objnr
- 1);
1863 realobj
= (char *)objp
+ obj_offset(cachep
);
1864 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
1866 print_objinfo(cachep
, objp
, 2);
1868 if (objnr
+ 1 < cachep
->num
) {
1869 objp
= index_to_obj(cachep
, slabp
, objnr
+ 1);
1870 realobj
= (char *)objp
+ obj_offset(cachep
);
1871 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
1873 print_objinfo(cachep
, objp
, 2);
1881 * slab_destroy_objs - destroy a slab and its objects
1882 * @cachep: cache pointer being destroyed
1883 * @slabp: slab pointer being destroyed
1885 * Call the registered destructor for each object in a slab that is being
1888 static void slab_destroy_objs(struct kmem_cache
*cachep
, struct slab
*slabp
)
1891 for (i
= 0; i
< cachep
->num
; i
++) {
1892 void *objp
= index_to_obj(cachep
, slabp
, i
);
1894 if (cachep
->flags
& SLAB_POISON
) {
1895 #ifdef CONFIG_DEBUG_PAGEALLOC
1896 if (cachep
->buffer_size
% PAGE_SIZE
== 0 &&
1898 kernel_map_pages(virt_to_page(objp
),
1899 cachep
->buffer_size
/ PAGE_SIZE
, 1);
1901 check_poison_obj(cachep
, objp
);
1903 check_poison_obj(cachep
, objp
);
1906 if (cachep
->flags
& SLAB_RED_ZONE
) {
1907 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1908 slab_error(cachep
, "start of a freed object "
1910 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1911 slab_error(cachep
, "end of a freed object "
1914 if (cachep
->dtor
&& !(cachep
->flags
& SLAB_POISON
))
1915 (cachep
->dtor
) (objp
+ obj_offset(cachep
), cachep
, 0);
1919 static void slab_destroy_objs(struct kmem_cache
*cachep
, struct slab
*slabp
)
1923 for (i
= 0; i
< cachep
->num
; i
++) {
1924 void *objp
= index_to_obj(cachep
, slabp
, i
);
1925 (cachep
->dtor
) (objp
, cachep
, 0);
1932 * slab_destroy - destroy and release all objects in a slab
1933 * @cachep: cache pointer being destroyed
1934 * @slabp: slab pointer being destroyed
1936 * Destroy all the objs in a slab, and release the mem back to the system.
1937 * Before calling the slab must have been unlinked from the cache. The
1938 * cache-lock is not held/needed.
1940 static void slab_destroy(struct kmem_cache
*cachep
, struct slab
*slabp
)
1942 void *addr
= slabp
->s_mem
- slabp
->colouroff
;
1944 slab_destroy_objs(cachep
, slabp
);
1945 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
1946 struct slab_rcu
*slab_rcu
;
1948 slab_rcu
= (struct slab_rcu
*)slabp
;
1949 slab_rcu
->cachep
= cachep
;
1950 slab_rcu
->addr
= addr
;
1951 call_rcu(&slab_rcu
->head
, kmem_rcu_free
);
1953 kmem_freepages(cachep
, addr
);
1954 if (OFF_SLAB(cachep
))
1955 kmem_cache_free(cachep
->slabp_cache
, slabp
);
1960 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1961 * size of kmem_list3.
1963 static void __init
set_up_list3s(struct kmem_cache
*cachep
, int index
)
1967 for_each_online_node(node
) {
1968 cachep
->nodelists
[node
] = &initkmem_list3
[index
+ node
];
1969 cachep
->nodelists
[node
]->next_reap
= jiffies
+
1971 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1975 static void __kmem_cache_destroy(struct kmem_cache
*cachep
)
1978 struct kmem_list3
*l3
;
1980 for_each_online_cpu(i
)
1981 kfree(cachep
->array
[i
]);
1983 /* NUMA: free the list3 structures */
1984 for_each_online_node(i
) {
1985 l3
= cachep
->nodelists
[i
];
1988 free_alien_cache(l3
->alien
);
1992 kmem_cache_free(&cache_cache
, cachep
);
1997 * calculate_slab_order - calculate size (page order) of slabs
1998 * @cachep: pointer to the cache that is being created
1999 * @size: size of objects to be created in this cache.
2000 * @align: required alignment for the objects.
2001 * @flags: slab allocation flags
2003 * Also calculates the number of objects per slab.
2005 * This could be made much more intelligent. For now, try to avoid using
2006 * high order pages for slabs. When the gfp() functions are more friendly
2007 * towards high-order requests, this should be changed.
2009 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
2010 size_t size
, size_t align
, unsigned long flags
)
2012 unsigned long offslab_limit
;
2013 size_t left_over
= 0;
2016 for (gfporder
= 0; gfporder
<= MAX_GFP_ORDER
; gfporder
++) {
2020 cache_estimate(gfporder
, size
, align
, flags
, &remainder
, &num
);
2024 if (flags
& CFLGS_OFF_SLAB
) {
2026 * Max number of objs-per-slab for caches which
2027 * use off-slab slabs. Needed to avoid a possible
2028 * looping condition in cache_grow().
2030 offslab_limit
= size
- sizeof(struct slab
);
2031 offslab_limit
/= sizeof(kmem_bufctl_t
);
2033 if (num
> offslab_limit
)
2037 /* Found something acceptable - save it away */
2039 cachep
->gfporder
= gfporder
;
2040 left_over
= remainder
;
2043 * A VFS-reclaimable slab tends to have most allocations
2044 * as GFP_NOFS and we really don't want to have to be allocating
2045 * higher-order pages when we are unable to shrink dcache.
2047 if (flags
& SLAB_RECLAIM_ACCOUNT
)
2051 * Large number of objects is good, but very large slabs are
2052 * currently bad for the gfp()s.
2054 if (gfporder
>= slab_break_gfp_order
)
2058 * Acceptable internal fragmentation?
2060 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
2066 static int setup_cpu_cache(struct kmem_cache
*cachep
)
2068 if (g_cpucache_up
== FULL
)
2069 return enable_cpucache(cachep
);
2071 if (g_cpucache_up
== NONE
) {
2073 * Note: the first kmem_cache_create must create the cache
2074 * that's used by kmalloc(24), otherwise the creation of
2075 * further caches will BUG().
2077 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
2080 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2081 * the first cache, then we need to set up all its list3s,
2082 * otherwise the creation of further caches will BUG().
2084 set_up_list3s(cachep
, SIZE_AC
);
2085 if (INDEX_AC
== INDEX_L3
)
2086 g_cpucache_up
= PARTIAL_L3
;
2088 g_cpucache_up
= PARTIAL_AC
;
2090 cachep
->array
[smp_processor_id()] =
2091 kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
2093 if (g_cpucache_up
== PARTIAL_AC
) {
2094 set_up_list3s(cachep
, SIZE_L3
);
2095 g_cpucache_up
= PARTIAL_L3
;
2098 for_each_online_node(node
) {
2099 cachep
->nodelists
[node
] =
2100 kmalloc_node(sizeof(struct kmem_list3
),
2102 BUG_ON(!cachep
->nodelists
[node
]);
2103 kmem_list3_init(cachep
->nodelists
[node
]);
2107 cachep
->nodelists
[numa_node_id()]->next_reap
=
2108 jiffies
+ REAPTIMEOUT_LIST3
+
2109 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
2111 cpu_cache_get(cachep
)->avail
= 0;
2112 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
2113 cpu_cache_get(cachep
)->batchcount
= 1;
2114 cpu_cache_get(cachep
)->touched
= 0;
2115 cachep
->batchcount
= 1;
2116 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
2121 * kmem_cache_create - Create a cache.
2122 * @name: A string which is used in /proc/slabinfo to identify this cache.
2123 * @size: The size of objects to be created in this cache.
2124 * @align: The required alignment for the objects.
2125 * @flags: SLAB flags
2126 * @ctor: A constructor for the objects.
2127 * @dtor: A destructor for the objects.
2129 * Returns a ptr to the cache on success, NULL on failure.
2130 * Cannot be called within a int, but can be interrupted.
2131 * The @ctor is run when new pages are allocated by the cache
2132 * and the @dtor is run before the pages are handed back.
2134 * @name must be valid until the cache is destroyed. This implies that
2135 * the module calling this has to destroy the cache before getting unloaded.
2139 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2140 * to catch references to uninitialised memory.
2142 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2143 * for buffer overruns.
2145 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2146 * cacheline. This can be beneficial if you're counting cycles as closely
2150 kmem_cache_create (const char *name
, size_t size
, size_t align
,
2151 unsigned long flags
,
2152 void (*ctor
)(void*, struct kmem_cache
*, unsigned long),
2153 void (*dtor
)(void*, struct kmem_cache
*, unsigned long))
2155 size_t left_over
, slab_size
, ralign
;
2156 struct kmem_cache
*cachep
= NULL
, *pc
;
2159 * Sanity checks... these are all serious usage bugs.
2161 if (!name
|| in_interrupt() || (size
< BYTES_PER_WORD
) ||
2162 (size
> (1 << MAX_OBJ_ORDER
) * PAGE_SIZE
) || (dtor
&& !ctor
)) {
2163 printk(KERN_ERR
"%s: Early error in slab %s\n", __FUNCTION__
,
2169 * We use cache_chain_mutex to ensure a consistent view of
2170 * cpu_online_map as well. Please see cpuup_callback
2172 mutex_lock(&cache_chain_mutex
);
2174 list_for_each_entry(pc
, &cache_chain
, next
) {
2179 * This happens when the module gets unloaded and doesn't
2180 * destroy its slab cache and no-one else reuses the vmalloc
2181 * area of the module. Print a warning.
2183 res
= probe_kernel_address(pc
->name
, tmp
);
2186 "SLAB: cache with size %d has lost its name\n",
2191 if (!strcmp(pc
->name
, name
)) {
2193 "kmem_cache_create: duplicate cache %s\n", name
);
2200 WARN_ON(strchr(name
, ' ')); /* It confuses parsers */
2203 * Enable redzoning and last user accounting, except for caches with
2204 * large objects, if the increased size would increase the object size
2205 * above the next power of two: caches with object sizes just above a
2206 * power of two have a significant amount of internal fragmentation.
2208 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + 3 * BYTES_PER_WORD
))
2209 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2210 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2211 flags
|= SLAB_POISON
;
2213 if (flags
& SLAB_DESTROY_BY_RCU
)
2214 BUG_ON(flags
& SLAB_POISON
);
2216 if (flags
& SLAB_DESTROY_BY_RCU
)
2220 * Always checks flags, a caller might be expecting debug support which
2223 BUG_ON(flags
& ~CREATE_MASK
);
2226 * Check that size is in terms of words. This is needed to avoid
2227 * unaligned accesses for some archs when redzoning is used, and makes
2228 * sure any on-slab bufctl's are also correctly aligned.
2230 if (size
& (BYTES_PER_WORD
- 1)) {
2231 size
+= (BYTES_PER_WORD
- 1);
2232 size
&= ~(BYTES_PER_WORD
- 1);
2235 /* calculate the final buffer alignment: */
2237 /* 1) arch recommendation: can be overridden for debug */
2238 if (flags
& SLAB_HWCACHE_ALIGN
) {
2240 * Default alignment: as specified by the arch code. Except if
2241 * an object is really small, then squeeze multiple objects into
2244 ralign
= cache_line_size();
2245 while (size
<= ralign
/ 2)
2248 ralign
= BYTES_PER_WORD
;
2252 * Redzoning and user store require word alignment. Note this will be
2253 * overridden by architecture or caller mandated alignment if either
2254 * is greater than BYTES_PER_WORD.
2256 if (flags
& SLAB_RED_ZONE
|| flags
& SLAB_STORE_USER
)
2257 ralign
= __alignof__(unsigned long long);
2259 /* 2) arch mandated alignment */
2260 if (ralign
< ARCH_SLAB_MINALIGN
) {
2261 ralign
= ARCH_SLAB_MINALIGN
;
2263 /* 3) caller mandated alignment */
2264 if (ralign
< align
) {
2267 /* disable debug if necessary */
2268 if (ralign
> __alignof__(unsigned long long))
2269 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2275 /* Get cache's description obj. */
2276 cachep
= kmem_cache_zalloc(&cache_cache
, GFP_KERNEL
);
2281 cachep
->obj_size
= size
;
2284 * Both debugging options require word-alignment which is calculated
2287 if (flags
& SLAB_RED_ZONE
) {
2288 /* add space for red zone words */
2289 cachep
->obj_offset
+= sizeof(unsigned long long);
2290 size
+= 2 * sizeof(unsigned long long);
2292 if (flags
& SLAB_STORE_USER
) {
2293 /* user store requires one word storage behind the end of
2296 size
+= BYTES_PER_WORD
;
2298 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2299 if (size
>= malloc_sizes
[INDEX_L3
+ 1].cs_size
2300 && cachep
->obj_size
> cache_line_size() && size
< PAGE_SIZE
) {
2301 cachep
->obj_offset
+= PAGE_SIZE
- size
;
2308 * Determine if the slab management is 'on' or 'off' slab.
2309 * (bootstrapping cannot cope with offslab caches so don't do
2312 if ((size
>= (PAGE_SIZE
>> 3)) && !slab_early_init
)
2314 * Size is large, assume best to place the slab management obj
2315 * off-slab (should allow better packing of objs).
2317 flags
|= CFLGS_OFF_SLAB
;
2319 size
= ALIGN(size
, align
);
2321 left_over
= calculate_slab_order(cachep
, size
, align
, flags
);
2325 "kmem_cache_create: couldn't create cache %s.\n", name
);
2326 kmem_cache_free(&cache_cache
, cachep
);
2330 slab_size
= ALIGN(cachep
->num
* sizeof(kmem_bufctl_t
)
2331 + sizeof(struct slab
), align
);
2334 * If the slab has been placed off-slab, and we have enough space then
2335 * move it on-slab. This is at the expense of any extra colouring.
2337 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= slab_size
) {
2338 flags
&= ~CFLGS_OFF_SLAB
;
2339 left_over
-= slab_size
;
2342 if (flags
& CFLGS_OFF_SLAB
) {
2343 /* really off slab. No need for manual alignment */
2345 cachep
->num
* sizeof(kmem_bufctl_t
) + sizeof(struct slab
);
2348 cachep
->colour_off
= cache_line_size();
2349 /* Offset must be a multiple of the alignment. */
2350 if (cachep
->colour_off
< align
)
2351 cachep
->colour_off
= align
;
2352 cachep
->colour
= left_over
/ cachep
->colour_off
;
2353 cachep
->slab_size
= slab_size
;
2354 cachep
->flags
= flags
;
2355 cachep
->gfpflags
= 0;
2356 if (CONFIG_ZONE_DMA_FLAG
&& (flags
& SLAB_CACHE_DMA
))
2357 cachep
->gfpflags
|= GFP_DMA
;
2358 cachep
->buffer_size
= size
;
2359 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2361 if (flags
& CFLGS_OFF_SLAB
) {
2362 cachep
->slabp_cache
= kmem_find_general_cachep(slab_size
, 0u);
2364 * This is a possibility for one of the malloc_sizes caches.
2365 * But since we go off slab only for object size greater than
2366 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2367 * this should not happen at all.
2368 * But leave a BUG_ON for some lucky dude.
2370 BUG_ON(!cachep
->slabp_cache
);
2372 cachep
->ctor
= ctor
;
2373 cachep
->dtor
= dtor
;
2374 cachep
->name
= name
;
2376 if (setup_cpu_cache(cachep
)) {
2377 __kmem_cache_destroy(cachep
);
2382 /* cache setup completed, link it into the list */
2383 list_add(&cachep
->next
, &cache_chain
);
2385 if (!cachep
&& (flags
& SLAB_PANIC
))
2386 panic("kmem_cache_create(): failed to create slab `%s'\n",
2388 mutex_unlock(&cache_chain_mutex
);
2391 EXPORT_SYMBOL(kmem_cache_create
);
2394 static void check_irq_off(void)
2396 BUG_ON(!irqs_disabled());
2399 static void check_irq_on(void)
2401 BUG_ON(irqs_disabled());
2404 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2408 assert_spin_locked(&cachep
->nodelists
[numa_node_id()]->list_lock
);
2412 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2416 assert_spin_locked(&cachep
->nodelists
[node
]->list_lock
);
2421 #define check_irq_off() do { } while(0)
2422 #define check_irq_on() do { } while(0)
2423 #define check_spinlock_acquired(x) do { } while(0)
2424 #define check_spinlock_acquired_node(x, y) do { } while(0)
2427 static void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
2428 struct array_cache
*ac
,
2429 int force
, int node
);
2431 static void do_drain(void *arg
)
2433 struct kmem_cache
*cachep
= arg
;
2434 struct array_cache
*ac
;
2435 int node
= numa_node_id();
2438 ac
= cpu_cache_get(cachep
);
2439 spin_lock(&cachep
->nodelists
[node
]->list_lock
);
2440 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
2441 spin_unlock(&cachep
->nodelists
[node
]->list_lock
);
2445 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2447 struct kmem_list3
*l3
;
2450 on_each_cpu(do_drain
, cachep
, 1, 1);
2452 for_each_online_node(node
) {
2453 l3
= cachep
->nodelists
[node
];
2454 if (l3
&& l3
->alien
)
2455 drain_alien_cache(cachep
, l3
->alien
);
2458 for_each_online_node(node
) {
2459 l3
= cachep
->nodelists
[node
];
2461 drain_array(cachep
, l3
, l3
->shared
, 1, node
);
2466 * Remove slabs from the list of free slabs.
2467 * Specify the number of slabs to drain in tofree.
2469 * Returns the actual number of slabs released.
2471 static int drain_freelist(struct kmem_cache
*cache
,
2472 struct kmem_list3
*l3
, int tofree
)
2474 struct list_head
*p
;
2479 while (nr_freed
< tofree
&& !list_empty(&l3
->slabs_free
)) {
2481 spin_lock_irq(&l3
->list_lock
);
2482 p
= l3
->slabs_free
.prev
;
2483 if (p
== &l3
->slabs_free
) {
2484 spin_unlock_irq(&l3
->list_lock
);
2488 slabp
= list_entry(p
, struct slab
, list
);
2490 BUG_ON(slabp
->inuse
);
2492 list_del(&slabp
->list
);
2494 * Safe to drop the lock. The slab is no longer linked
2497 l3
->free_objects
-= cache
->num
;
2498 spin_unlock_irq(&l3
->list_lock
);
2499 slab_destroy(cache
, slabp
);
2506 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2507 static int __cache_shrink(struct kmem_cache
*cachep
)
2510 struct kmem_list3
*l3
;
2512 drain_cpu_caches(cachep
);
2515 for_each_online_node(i
) {
2516 l3
= cachep
->nodelists
[i
];
2520 drain_freelist(cachep
, l3
, l3
->free_objects
);
2522 ret
+= !list_empty(&l3
->slabs_full
) ||
2523 !list_empty(&l3
->slabs_partial
);
2525 return (ret
? 1 : 0);
2529 * kmem_cache_shrink - Shrink a cache.
2530 * @cachep: The cache to shrink.
2532 * Releases as many slabs as possible for a cache.
2533 * To help debugging, a zero exit status indicates all slabs were released.
2535 int kmem_cache_shrink(struct kmem_cache
*cachep
)
2538 BUG_ON(!cachep
|| in_interrupt());
2540 mutex_lock(&cache_chain_mutex
);
2541 ret
= __cache_shrink(cachep
);
2542 mutex_unlock(&cache_chain_mutex
);
2545 EXPORT_SYMBOL(kmem_cache_shrink
);
2548 * kmem_cache_destroy - delete a cache
2549 * @cachep: the cache to destroy
2551 * Remove a &struct kmem_cache object from the slab cache.
2553 * It is expected this function will be called by a module when it is
2554 * unloaded. This will remove the cache completely, and avoid a duplicate
2555 * cache being allocated each time a module is loaded and unloaded, if the
2556 * module doesn't have persistent in-kernel storage across loads and unloads.
2558 * The cache must be empty before calling this function.
2560 * The caller must guarantee that noone will allocate memory from the cache
2561 * during the kmem_cache_destroy().
2563 void kmem_cache_destroy(struct kmem_cache
*cachep
)
2565 BUG_ON(!cachep
|| in_interrupt());
2567 /* Find the cache in the chain of caches. */
2568 mutex_lock(&cache_chain_mutex
);
2570 * the chain is never empty, cache_cache is never destroyed
2572 list_del(&cachep
->next
);
2573 if (__cache_shrink(cachep
)) {
2574 slab_error(cachep
, "Can't free all objects");
2575 list_add(&cachep
->next
, &cache_chain
);
2576 mutex_unlock(&cache_chain_mutex
);
2580 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
2583 __kmem_cache_destroy(cachep
);
2584 mutex_unlock(&cache_chain_mutex
);
2586 EXPORT_SYMBOL(kmem_cache_destroy
);
2589 * Get the memory for a slab management obj.
2590 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2591 * always come from malloc_sizes caches. The slab descriptor cannot
2592 * come from the same cache which is getting created because,
2593 * when we are searching for an appropriate cache for these
2594 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2595 * If we are creating a malloc_sizes cache here it would not be visible to
2596 * kmem_find_general_cachep till the initialization is complete.
2597 * Hence we cannot have slabp_cache same as the original cache.
2599 static struct slab
*alloc_slabmgmt(struct kmem_cache
*cachep
, void *objp
,
2600 int colour_off
, gfp_t local_flags
,
2605 if (OFF_SLAB(cachep
)) {
2606 /* Slab management obj is off-slab. */
2607 slabp
= kmem_cache_alloc_node(cachep
->slabp_cache
,
2608 local_flags
& ~GFP_THISNODE
, nodeid
);
2612 slabp
= objp
+ colour_off
;
2613 colour_off
+= cachep
->slab_size
;
2616 slabp
->colouroff
= colour_off
;
2617 slabp
->s_mem
= objp
+ colour_off
;
2618 slabp
->nodeid
= nodeid
;
2622 static inline kmem_bufctl_t
*slab_bufctl(struct slab
*slabp
)
2624 return (kmem_bufctl_t
*) (slabp
+ 1);
2627 static void cache_init_objs(struct kmem_cache
*cachep
,
2628 struct slab
*slabp
, unsigned long ctor_flags
)
2632 for (i
= 0; i
< cachep
->num
; i
++) {
2633 void *objp
= index_to_obj(cachep
, slabp
, i
);
2635 /* need to poison the objs? */
2636 if (cachep
->flags
& SLAB_POISON
)
2637 poison_obj(cachep
, objp
, POISON_FREE
);
2638 if (cachep
->flags
& SLAB_STORE_USER
)
2639 *dbg_userword(cachep
, objp
) = NULL
;
2641 if (cachep
->flags
& SLAB_RED_ZONE
) {
2642 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2643 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2646 * Constructors are not allowed to allocate memory from the same
2647 * cache which they are a constructor for. Otherwise, deadlock.
2648 * They must also be threaded.
2650 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2651 cachep
->ctor(objp
+ obj_offset(cachep
), cachep
,
2654 if (cachep
->flags
& SLAB_RED_ZONE
) {
2655 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2656 slab_error(cachep
, "constructor overwrote the"
2657 " end of an object");
2658 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2659 slab_error(cachep
, "constructor overwrote the"
2660 " start of an object");
2662 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 &&
2663 OFF_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
)
2664 kernel_map_pages(virt_to_page(objp
),
2665 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2668 cachep
->ctor(objp
, cachep
, ctor_flags
);
2670 slab_bufctl(slabp
)[i
] = i
+ 1;
2672 slab_bufctl(slabp
)[i
- 1] = BUFCTL_END
;
2676 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2678 if (CONFIG_ZONE_DMA_FLAG
) {
2679 if (flags
& GFP_DMA
)
2680 BUG_ON(!(cachep
->gfpflags
& GFP_DMA
));
2682 BUG_ON(cachep
->gfpflags
& GFP_DMA
);
2686 static void *slab_get_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2689 void *objp
= index_to_obj(cachep
, slabp
, slabp
->free
);
2693 next
= slab_bufctl(slabp
)[slabp
->free
];
2695 slab_bufctl(slabp
)[slabp
->free
] = BUFCTL_FREE
;
2696 WARN_ON(slabp
->nodeid
!= nodeid
);
2703 static void slab_put_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2704 void *objp
, int nodeid
)
2706 unsigned int objnr
= obj_to_index(cachep
, slabp
, objp
);
2709 /* Verify that the slab belongs to the intended node */
2710 WARN_ON(slabp
->nodeid
!= nodeid
);
2712 if (slab_bufctl(slabp
)[objnr
] + 1 <= SLAB_LIMIT
+ 1) {
2713 printk(KERN_ERR
"slab: double free detected in cache "
2714 "'%s', objp %p\n", cachep
->name
, objp
);
2718 slab_bufctl(slabp
)[objnr
] = slabp
->free
;
2719 slabp
->free
= objnr
;
2724 * Map pages beginning at addr to the given cache and slab. This is required
2725 * for the slab allocator to be able to lookup the cache and slab of a
2726 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2728 static void slab_map_pages(struct kmem_cache
*cache
, struct slab
*slab
,
2734 page
= virt_to_page(addr
);
2737 if (likely(!PageCompound(page
)))
2738 nr_pages
<<= cache
->gfporder
;
2741 page_set_cache(page
, cache
);
2742 page_set_slab(page
, slab
);
2744 } while (--nr_pages
);
2748 * Grow (by 1) the number of slabs within a cache. This is called by
2749 * kmem_cache_alloc() when there are no active objs left in a cache.
2751 static int cache_grow(struct kmem_cache
*cachep
,
2752 gfp_t flags
, int nodeid
, void *objp
)
2757 unsigned long ctor_flags
;
2758 struct kmem_list3
*l3
;
2761 * Be lazy and only check for valid flags here, keeping it out of the
2762 * critical path in kmem_cache_alloc().
2764 BUG_ON(flags
& ~(GFP_DMA
| GFP_LEVEL_MASK
));
2766 ctor_flags
= SLAB_CTOR_CONSTRUCTOR
;
2767 local_flags
= (flags
& GFP_LEVEL_MASK
);
2768 /* Take the l3 list lock to change the colour_next on this node */
2770 l3
= cachep
->nodelists
[nodeid
];
2771 spin_lock(&l3
->list_lock
);
2773 /* Get colour for the slab, and cal the next value. */
2774 offset
= l3
->colour_next
;
2776 if (l3
->colour_next
>= cachep
->colour
)
2777 l3
->colour_next
= 0;
2778 spin_unlock(&l3
->list_lock
);
2780 offset
*= cachep
->colour_off
;
2782 if (local_flags
& __GFP_WAIT
)
2786 * The test for missing atomic flag is performed here, rather than
2787 * the more obvious place, simply to reduce the critical path length
2788 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2789 * will eventually be caught here (where it matters).
2791 kmem_flagcheck(cachep
, flags
);
2794 * Get mem for the objs. Attempt to allocate a physical page from
2798 objp
= kmem_getpages(cachep
, flags
, nodeid
);
2802 /* Get slab management. */
2803 slabp
= alloc_slabmgmt(cachep
, objp
, offset
,
2804 local_flags
& ~GFP_THISNODE
, nodeid
);
2808 slabp
->nodeid
= nodeid
;
2809 slab_map_pages(cachep
, slabp
, objp
);
2811 cache_init_objs(cachep
, slabp
, ctor_flags
);
2813 if (local_flags
& __GFP_WAIT
)
2814 local_irq_disable();
2816 spin_lock(&l3
->list_lock
);
2818 /* Make slab active. */
2819 list_add_tail(&slabp
->list
, &(l3
->slabs_free
));
2820 STATS_INC_GROWN(cachep
);
2821 l3
->free_objects
+= cachep
->num
;
2822 spin_unlock(&l3
->list_lock
);
2825 kmem_freepages(cachep
, objp
);
2827 if (local_flags
& __GFP_WAIT
)
2828 local_irq_disable();
2835 * Perform extra freeing checks:
2836 * - detect bad pointers.
2837 * - POISON/RED_ZONE checking
2838 * - destructor calls, for caches with POISON+dtor
2840 static void kfree_debugcheck(const void *objp
)
2842 if (!virt_addr_valid(objp
)) {
2843 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
2844 (unsigned long)objp
);
2849 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2851 unsigned long long redzone1
, redzone2
;
2853 redzone1
= *dbg_redzone1(cache
, obj
);
2854 redzone2
= *dbg_redzone2(cache
, obj
);
2859 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2862 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2863 slab_error(cache
, "double free detected");
2865 slab_error(cache
, "memory outside object was overwritten");
2867 printk(KERN_ERR
"%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2868 obj
, redzone1
, redzone2
);
2871 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2878 objp
-= obj_offset(cachep
);
2879 kfree_debugcheck(objp
);
2880 page
= virt_to_head_page(objp
);
2882 slabp
= page_get_slab(page
);
2884 if (cachep
->flags
& SLAB_RED_ZONE
) {
2885 verify_redzone_free(cachep
, objp
);
2886 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2887 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2889 if (cachep
->flags
& SLAB_STORE_USER
)
2890 *dbg_userword(cachep
, objp
) = caller
;
2892 objnr
= obj_to_index(cachep
, slabp
, objp
);
2894 BUG_ON(objnr
>= cachep
->num
);
2895 BUG_ON(objp
!= index_to_obj(cachep
, slabp
, objnr
));
2897 if (cachep
->flags
& SLAB_POISON
&& cachep
->dtor
) {
2898 /* we want to cache poison the object,
2899 * call the destruction callback
2901 cachep
->dtor(objp
+ obj_offset(cachep
), cachep
, 0);
2903 #ifdef CONFIG_DEBUG_SLAB_LEAK
2904 slab_bufctl(slabp
)[objnr
] = BUFCTL_FREE
;
2906 if (cachep
->flags
& SLAB_POISON
) {
2907 #ifdef CONFIG_DEBUG_PAGEALLOC
2908 if ((cachep
->buffer_size
% PAGE_SIZE
)==0 && OFF_SLAB(cachep
)) {
2909 store_stackinfo(cachep
, objp
, (unsigned long)caller
);
2910 kernel_map_pages(virt_to_page(objp
),
2911 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2913 poison_obj(cachep
, objp
, POISON_FREE
);
2916 poison_obj(cachep
, objp
, POISON_FREE
);
2922 static void check_slabp(struct kmem_cache
*cachep
, struct slab
*slabp
)
2927 /* Check slab's freelist to see if this obj is there. */
2928 for (i
= slabp
->free
; i
!= BUFCTL_END
; i
= slab_bufctl(slabp
)[i
]) {
2930 if (entries
> cachep
->num
|| i
>= cachep
->num
)
2933 if (entries
!= cachep
->num
- slabp
->inuse
) {
2935 printk(KERN_ERR
"slab: Internal list corruption detected in "
2936 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2937 cachep
->name
, cachep
->num
, slabp
, slabp
->inuse
);
2939 i
< sizeof(*slabp
) + cachep
->num
* sizeof(kmem_bufctl_t
);
2942 printk("\n%03x:", i
);
2943 printk(" %02x", ((unsigned char *)slabp
)[i
]);
2950 #define kfree_debugcheck(x) do { } while(0)
2951 #define cache_free_debugcheck(x,objp,z) (objp)
2952 #define check_slabp(x,y) do { } while(0)
2955 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
2958 struct kmem_list3
*l3
;
2959 struct array_cache
*ac
;
2962 node
= numa_node_id();
2965 ac
= cpu_cache_get(cachep
);
2967 batchcount
= ac
->batchcount
;
2968 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2970 * If there was little recent activity on this cache, then
2971 * perform only a partial refill. Otherwise we could generate
2974 batchcount
= BATCHREFILL_LIMIT
;
2976 l3
= cachep
->nodelists
[node
];
2978 BUG_ON(ac
->avail
> 0 || !l3
);
2979 spin_lock(&l3
->list_lock
);
2981 /* See if we can refill from the shared array */
2982 if (l3
->shared
&& transfer_objects(ac
, l3
->shared
, batchcount
))
2985 while (batchcount
> 0) {
2986 struct list_head
*entry
;
2988 /* Get slab alloc is to come from. */
2989 entry
= l3
->slabs_partial
.next
;
2990 if (entry
== &l3
->slabs_partial
) {
2991 l3
->free_touched
= 1;
2992 entry
= l3
->slabs_free
.next
;
2993 if (entry
== &l3
->slabs_free
)
2997 slabp
= list_entry(entry
, struct slab
, list
);
2998 check_slabp(cachep
, slabp
);
2999 check_spinlock_acquired(cachep
);
3002 * The slab was either on partial or free list so
3003 * there must be at least one object available for
3006 BUG_ON(slabp
->inuse
< 0 || slabp
->inuse
>= cachep
->num
);
3008 while (slabp
->inuse
< cachep
->num
&& batchcount
--) {
3009 STATS_INC_ALLOCED(cachep
);
3010 STATS_INC_ACTIVE(cachep
);
3011 STATS_SET_HIGH(cachep
);
3013 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, slabp
,
3016 check_slabp(cachep
, slabp
);
3018 /* move slabp to correct slabp list: */
3019 list_del(&slabp
->list
);
3020 if (slabp
->free
== BUFCTL_END
)
3021 list_add(&slabp
->list
, &l3
->slabs_full
);
3023 list_add(&slabp
->list
, &l3
->slabs_partial
);
3027 l3
->free_objects
-= ac
->avail
;
3029 spin_unlock(&l3
->list_lock
);
3031 if (unlikely(!ac
->avail
)) {
3033 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, node
, NULL
);
3035 /* cache_grow can reenable interrupts, then ac could change. */
3036 ac
= cpu_cache_get(cachep
);
3037 if (!x
&& ac
->avail
== 0) /* no objects in sight? abort */
3040 if (!ac
->avail
) /* objects refilled by interrupt? */
3044 return ac
->entry
[--ac
->avail
];
3047 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
3050 might_sleep_if(flags
& __GFP_WAIT
);
3052 kmem_flagcheck(cachep
, flags
);
3057 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
3058 gfp_t flags
, void *objp
, void *caller
)
3062 if (cachep
->flags
& SLAB_POISON
) {
3063 #ifdef CONFIG_DEBUG_PAGEALLOC
3064 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
3065 kernel_map_pages(virt_to_page(objp
),
3066 cachep
->buffer_size
/ PAGE_SIZE
, 1);
3068 check_poison_obj(cachep
, objp
);
3070 check_poison_obj(cachep
, objp
);
3072 poison_obj(cachep
, objp
, POISON_INUSE
);
3074 if (cachep
->flags
& SLAB_STORE_USER
)
3075 *dbg_userword(cachep
, objp
) = caller
;
3077 if (cachep
->flags
& SLAB_RED_ZONE
) {
3078 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
3079 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
3080 slab_error(cachep
, "double free, or memory outside"
3081 " object was overwritten");
3083 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3084 objp
, *dbg_redzone1(cachep
, objp
),
3085 *dbg_redzone2(cachep
, objp
));
3087 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
3088 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
3090 #ifdef CONFIG_DEBUG_SLAB_LEAK
3095 slabp
= page_get_slab(virt_to_head_page(objp
));
3096 objnr
= (unsigned)(objp
- slabp
->s_mem
) / cachep
->buffer_size
;
3097 slab_bufctl(slabp
)[objnr
] = BUFCTL_ACTIVE
;
3100 objp
+= obj_offset(cachep
);
3101 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
3102 cachep
->ctor(objp
, cachep
, SLAB_CTOR_CONSTRUCTOR
);
3103 #if ARCH_SLAB_MINALIGN
3104 if ((u32
)objp
& (ARCH_SLAB_MINALIGN
-1)) {
3105 printk(KERN_ERR
"0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3106 objp
, ARCH_SLAB_MINALIGN
);
3112 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3115 #ifdef CONFIG_FAILSLAB
3117 static struct failslab_attr
{
3119 struct fault_attr attr
;
3121 u32 ignore_gfp_wait
;
3122 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3123 struct dentry
*ignore_gfp_wait_file
;
3127 .attr
= FAULT_ATTR_INITIALIZER
,
3128 .ignore_gfp_wait
= 1,
3131 static int __init
setup_failslab(char *str
)
3133 return setup_fault_attr(&failslab
.attr
, str
);
3135 __setup("failslab=", setup_failslab
);
3137 static int should_failslab(struct kmem_cache
*cachep
, gfp_t flags
)
3139 if (cachep
== &cache_cache
)
3141 if (flags
& __GFP_NOFAIL
)
3143 if (failslab
.ignore_gfp_wait
&& (flags
& __GFP_WAIT
))
3146 return should_fail(&failslab
.attr
, obj_size(cachep
));
3149 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3151 static int __init
failslab_debugfs(void)
3153 mode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
3157 err
= init_fault_attr_dentries(&failslab
.attr
, "failslab");
3160 dir
= failslab
.attr
.dentries
.dir
;
3162 failslab
.ignore_gfp_wait_file
=
3163 debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
3164 &failslab
.ignore_gfp_wait
);
3166 if (!failslab
.ignore_gfp_wait_file
) {
3168 debugfs_remove(failslab
.ignore_gfp_wait_file
);
3169 cleanup_fault_attr_dentries(&failslab
.attr
);
3175 late_initcall(failslab_debugfs
);
3177 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3179 #else /* CONFIG_FAILSLAB */
3181 static inline int should_failslab(struct kmem_cache
*cachep
, gfp_t flags
)
3186 #endif /* CONFIG_FAILSLAB */
3188 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3191 struct array_cache
*ac
;
3195 ac
= cpu_cache_get(cachep
);
3196 if (likely(ac
->avail
)) {
3197 STATS_INC_ALLOCHIT(cachep
);
3199 objp
= ac
->entry
[--ac
->avail
];
3201 STATS_INC_ALLOCMISS(cachep
);
3202 objp
= cache_alloc_refill(cachep
, flags
);
3209 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3211 * If we are in_interrupt, then process context, including cpusets and
3212 * mempolicy, may not apply and should not be used for allocation policy.
3214 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3216 int nid_alloc
, nid_here
;
3218 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3220 nid_alloc
= nid_here
= numa_node_id();
3221 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3222 nid_alloc
= cpuset_mem_spread_node();
3223 else if (current
->mempolicy
)
3224 nid_alloc
= slab_node(current
->mempolicy
);
3225 if (nid_alloc
!= nid_here
)
3226 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3231 * Fallback function if there was no memory available and no objects on a
3232 * certain node and fall back is permitted. First we scan all the
3233 * available nodelists for available objects. If that fails then we
3234 * perform an allocation without specifying a node. This allows the page
3235 * allocator to do its reclaim / fallback magic. We then insert the
3236 * slab into the proper nodelist and then allocate from it.
3238 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3240 struct zonelist
*zonelist
;
3246 if (flags
& __GFP_THISNODE
)
3249 zonelist
= &NODE_DATA(slab_node(current
->mempolicy
))
3250 ->node_zonelists
[gfp_zone(flags
)];
3251 local_flags
= (flags
& GFP_LEVEL_MASK
);
3255 * Look through allowed nodes for objects available
3256 * from existing per node queues.
3258 for (z
= zonelist
->zones
; *z
&& !obj
; z
++) {
3259 nid
= zone_to_nid(*z
);
3261 if (cpuset_zone_allowed_hardwall(*z
, flags
) &&
3262 cache
->nodelists
[nid
] &&
3263 cache
->nodelists
[nid
]->free_objects
)
3264 obj
= ____cache_alloc_node(cache
,
3265 flags
| GFP_THISNODE
, nid
);
3270 * This allocation will be performed within the constraints
3271 * of the current cpuset / memory policy requirements.
3272 * We may trigger various forms of reclaim on the allowed
3273 * set and go into memory reserves if necessary.
3275 if (local_flags
& __GFP_WAIT
)
3277 kmem_flagcheck(cache
, flags
);
3278 obj
= kmem_getpages(cache
, flags
, -1);
3279 if (local_flags
& __GFP_WAIT
)
3280 local_irq_disable();
3283 * Insert into the appropriate per node queues
3285 nid
= page_to_nid(virt_to_page(obj
));
3286 if (cache_grow(cache
, flags
, nid
, obj
)) {
3287 obj
= ____cache_alloc_node(cache
,
3288 flags
| GFP_THISNODE
, nid
);
3291 * Another processor may allocate the
3292 * objects in the slab since we are
3293 * not holding any locks.
3297 /* cache_grow already freed obj */
3306 * A interface to enable slab creation on nodeid
3308 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3311 struct list_head
*entry
;
3313 struct kmem_list3
*l3
;
3317 l3
= cachep
->nodelists
[nodeid
];
3322 spin_lock(&l3
->list_lock
);
3323 entry
= l3
->slabs_partial
.next
;
3324 if (entry
== &l3
->slabs_partial
) {
3325 l3
->free_touched
= 1;
3326 entry
= l3
->slabs_free
.next
;
3327 if (entry
== &l3
->slabs_free
)
3331 slabp
= list_entry(entry
, struct slab
, list
);
3332 check_spinlock_acquired_node(cachep
, nodeid
);
3333 check_slabp(cachep
, slabp
);
3335 STATS_INC_NODEALLOCS(cachep
);
3336 STATS_INC_ACTIVE(cachep
);
3337 STATS_SET_HIGH(cachep
);
3339 BUG_ON(slabp
->inuse
== cachep
->num
);
3341 obj
= slab_get_obj(cachep
, slabp
, nodeid
);
3342 check_slabp(cachep
, slabp
);
3344 /* move slabp to correct slabp list: */
3345 list_del(&slabp
->list
);
3347 if (slabp
->free
== BUFCTL_END
)
3348 list_add(&slabp
->list
, &l3
->slabs_full
);
3350 list_add(&slabp
->list
, &l3
->slabs_partial
);
3352 spin_unlock(&l3
->list_lock
);
3356 spin_unlock(&l3
->list_lock
);
3357 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, nodeid
, NULL
);
3361 return fallback_alloc(cachep
, flags
);
3368 * kmem_cache_alloc_node - Allocate an object on the specified node
3369 * @cachep: The cache to allocate from.
3370 * @flags: See kmalloc().
3371 * @nodeid: node number of the target node.
3372 * @caller: return address of caller, used for debug information
3374 * Identical to kmem_cache_alloc but it will allocate memory on the given
3375 * node, which can improve the performance for cpu bound structures.
3377 * Fallback to other node is possible if __GFP_THISNODE is not set.
3379 static __always_inline
void *
3380 __cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3383 unsigned long save_flags
;
3386 if (should_failslab(cachep
, flags
))
3389 cache_alloc_debugcheck_before(cachep
, flags
);
3390 local_irq_save(save_flags
);
3392 if (unlikely(nodeid
== -1))
3393 nodeid
= numa_node_id();
3395 if (unlikely(!cachep
->nodelists
[nodeid
])) {
3396 /* Node not bootstrapped yet */
3397 ptr
= fallback_alloc(cachep
, flags
);
3401 if (nodeid
== numa_node_id()) {
3403 * Use the locally cached objects if possible.
3404 * However ____cache_alloc does not allow fallback
3405 * to other nodes. It may fail while we still have
3406 * objects on other nodes available.
3408 ptr
= ____cache_alloc(cachep
, flags
);
3412 /* ___cache_alloc_node can fall back to other nodes */
3413 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3415 local_irq_restore(save_flags
);
3416 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3421 static __always_inline
void *
3422 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3426 if (unlikely(current
->flags
& (PF_SPREAD_SLAB
| PF_MEMPOLICY
))) {
3427 objp
= alternate_node_alloc(cache
, flags
);
3431 objp
= ____cache_alloc(cache
, flags
);
3434 * We may just have run out of memory on the local node.
3435 * ____cache_alloc_node() knows how to locate memory on other nodes
3438 objp
= ____cache_alloc_node(cache
, flags
, numa_node_id());
3445 static __always_inline
void *
3446 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3448 return ____cache_alloc(cachep
, flags
);
3451 #endif /* CONFIG_NUMA */
3453 static __always_inline
void *
3454 __cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
, void *caller
)
3456 unsigned long save_flags
;
3459 if (should_failslab(cachep
, flags
))
3462 cache_alloc_debugcheck_before(cachep
, flags
);
3463 local_irq_save(save_flags
);
3464 objp
= __do_cache_alloc(cachep
, flags
);
3465 local_irq_restore(save_flags
);
3466 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3473 * Caller needs to acquire correct kmem_list's list_lock
3475 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int nr_objects
,
3479 struct kmem_list3
*l3
;
3481 for (i
= 0; i
< nr_objects
; i
++) {
3482 void *objp
= objpp
[i
];
3485 slabp
= virt_to_slab(objp
);
3486 l3
= cachep
->nodelists
[node
];
3487 list_del(&slabp
->list
);
3488 check_spinlock_acquired_node(cachep
, node
);
3489 check_slabp(cachep
, slabp
);
3490 slab_put_obj(cachep
, slabp
, objp
, node
);
3491 STATS_DEC_ACTIVE(cachep
);
3493 check_slabp(cachep
, slabp
);
3495 /* fixup slab chains */
3496 if (slabp
->inuse
== 0) {
3497 if (l3
->free_objects
> l3
->free_limit
) {
3498 l3
->free_objects
-= cachep
->num
;
3499 /* No need to drop any previously held
3500 * lock here, even if we have a off-slab slab
3501 * descriptor it is guaranteed to come from
3502 * a different cache, refer to comments before
3505 slab_destroy(cachep
, slabp
);
3507 list_add(&slabp
->list
, &l3
->slabs_free
);
3510 /* Unconditionally move a slab to the end of the
3511 * partial list on free - maximum time for the
3512 * other objects to be freed, too.
3514 list_add_tail(&slabp
->list
, &l3
->slabs_partial
);
3519 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3522 struct kmem_list3
*l3
;
3523 int node
= numa_node_id();
3525 batchcount
= ac
->batchcount
;
3527 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
3530 l3
= cachep
->nodelists
[node
];
3531 spin_lock(&l3
->list_lock
);
3533 struct array_cache
*shared_array
= l3
->shared
;
3534 int max
= shared_array
->limit
- shared_array
->avail
;
3536 if (batchcount
> max
)
3538 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3539 ac
->entry
, sizeof(void *) * batchcount
);
3540 shared_array
->avail
+= batchcount
;
3545 free_block(cachep
, ac
->entry
, batchcount
, node
);
3550 struct list_head
*p
;
3552 p
= l3
->slabs_free
.next
;
3553 while (p
!= &(l3
->slabs_free
)) {
3556 slabp
= list_entry(p
, struct slab
, list
);
3557 BUG_ON(slabp
->inuse
);
3562 STATS_SET_FREEABLE(cachep
, i
);
3565 spin_unlock(&l3
->list_lock
);
3566 ac
->avail
-= batchcount
;
3567 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3571 * Release an obj back to its cache. If the obj has a constructed state, it must
3572 * be in this state _before_ it is released. Called with disabled ints.
3574 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
)
3576 struct array_cache
*ac
= cpu_cache_get(cachep
);
3579 objp
= cache_free_debugcheck(cachep
, objp
, __builtin_return_address(0));
3581 if (use_alien_caches
&& cache_free_alien(cachep
, objp
))
3584 if (likely(ac
->avail
< ac
->limit
)) {
3585 STATS_INC_FREEHIT(cachep
);
3586 ac
->entry
[ac
->avail
++] = objp
;
3589 STATS_INC_FREEMISS(cachep
);
3590 cache_flusharray(cachep
, ac
);
3591 ac
->entry
[ac
->avail
++] = objp
;
3596 * kmem_cache_alloc - Allocate an object
3597 * @cachep: The cache to allocate from.
3598 * @flags: See kmalloc().
3600 * Allocate an object from this cache. The flags are only relevant
3601 * if the cache has no available objects.
3603 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3605 return __cache_alloc(cachep
, flags
, __builtin_return_address(0));
3607 EXPORT_SYMBOL(kmem_cache_alloc
);
3610 * kmem_cache_zalloc - Allocate an object. The memory is set to zero.
3611 * @cache: The cache to allocate from.
3612 * @flags: See kmalloc().
3614 * Allocate an object from this cache and set the allocated memory to zero.
3615 * The flags are only relevant if the cache has no available objects.
3617 void *kmem_cache_zalloc(struct kmem_cache
*cache
, gfp_t flags
)
3619 void *ret
= __cache_alloc(cache
, flags
, __builtin_return_address(0));
3621 memset(ret
, 0, obj_size(cache
));
3624 EXPORT_SYMBOL(kmem_cache_zalloc
);
3627 * kmem_ptr_validate - check if an untrusted pointer might
3629 * @cachep: the cache we're checking against
3630 * @ptr: pointer to validate
3632 * This verifies that the untrusted pointer looks sane:
3633 * it is _not_ a guarantee that the pointer is actually
3634 * part of the slab cache in question, but it at least
3635 * validates that the pointer can be dereferenced and
3636 * looks half-way sane.
3638 * Currently only used for dentry validation.
3640 int kmem_ptr_validate(struct kmem_cache
*cachep
, const void *ptr
)
3642 unsigned long addr
= (unsigned long)ptr
;
3643 unsigned long min_addr
= PAGE_OFFSET
;
3644 unsigned long align_mask
= BYTES_PER_WORD
- 1;
3645 unsigned long size
= cachep
->buffer_size
;
3648 if (unlikely(addr
< min_addr
))
3650 if (unlikely(addr
> (unsigned long)high_memory
- size
))
3652 if (unlikely(addr
& align_mask
))
3654 if (unlikely(!kern_addr_valid(addr
)))
3656 if (unlikely(!kern_addr_valid(addr
+ size
- 1)))
3658 page
= virt_to_page(ptr
);
3659 if (unlikely(!PageSlab(page
)))
3661 if (unlikely(page_get_cache(page
) != cachep
))
3669 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3671 return __cache_alloc_node(cachep
, flags
, nodeid
,
3672 __builtin_return_address(0));
3674 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3676 static __always_inline
void *
3677 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, void *caller
)
3679 struct kmem_cache
*cachep
;
3681 cachep
= kmem_find_general_cachep(size
, flags
);
3682 if (unlikely(cachep
== NULL
))
3684 return kmem_cache_alloc_node(cachep
, flags
, node
);
3687 #ifdef CONFIG_DEBUG_SLAB
3688 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3690 return __do_kmalloc_node(size
, flags
, node
,
3691 __builtin_return_address(0));
3693 EXPORT_SYMBOL(__kmalloc_node
);
3695 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3696 int node
, void *caller
)
3698 return __do_kmalloc_node(size
, flags
, node
, caller
);
3700 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3702 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3704 return __do_kmalloc_node(size
, flags
, node
, NULL
);
3706 EXPORT_SYMBOL(__kmalloc_node
);
3707 #endif /* CONFIG_DEBUG_SLAB */
3708 #endif /* CONFIG_NUMA */
3711 * __do_kmalloc - allocate memory
3712 * @size: how many bytes of memory are required.
3713 * @flags: the type of memory to allocate (see kmalloc).
3714 * @caller: function caller for debug tracking of the caller
3716 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3719 struct kmem_cache
*cachep
;
3721 /* If you want to save a few bytes .text space: replace
3723 * Then kmalloc uses the uninlined functions instead of the inline
3726 cachep
= __find_general_cachep(size
, flags
);
3727 if (unlikely(cachep
== NULL
))
3729 return __cache_alloc(cachep
, flags
, caller
);
3733 #ifdef CONFIG_DEBUG_SLAB
3734 void *__kmalloc(size_t size
, gfp_t flags
)
3736 return __do_kmalloc(size
, flags
, __builtin_return_address(0));
3738 EXPORT_SYMBOL(__kmalloc
);
3740 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, void *caller
)
3742 return __do_kmalloc(size
, flags
, caller
);
3744 EXPORT_SYMBOL(__kmalloc_track_caller
);
3747 void *__kmalloc(size_t size
, gfp_t flags
)
3749 return __do_kmalloc(size
, flags
, NULL
);
3751 EXPORT_SYMBOL(__kmalloc
);
3755 * krealloc - reallocate memory. The contents will remain unchanged.
3756 * @p: object to reallocate memory for.
3757 * @new_size: how many bytes of memory are required.
3758 * @flags: the type of memory to allocate.
3760 * The contents of the object pointed to are preserved up to the
3761 * lesser of the new and old sizes. If @p is %NULL, krealloc()
3762 * behaves exactly like kmalloc(). If @size is 0 and @p is not a
3763 * %NULL pointer, the object pointed to is freed.
3765 void *krealloc(const void *p
, size_t new_size
, gfp_t flags
)
3767 struct kmem_cache
*cache
, *new_cache
;
3771 return kmalloc_track_caller(new_size
, flags
);
3773 if (unlikely(!new_size
)) {
3778 cache
= virt_to_cache(p
);
3779 new_cache
= __find_general_cachep(new_size
, flags
);
3782 * If new size fits in the current cache, bail out.
3784 if (likely(cache
== new_cache
))
3788 * We are on the slow-path here so do not use __cache_alloc
3789 * because it bloats kernel text.
3791 ret
= kmalloc_track_caller(new_size
, flags
);
3793 memcpy(ret
, p
, min(new_size
, ksize(p
)));
3798 EXPORT_SYMBOL(krealloc
);
3801 * kmem_cache_free - Deallocate an object
3802 * @cachep: The cache the allocation was from.
3803 * @objp: The previously allocated object.
3805 * Free an object which was previously allocated from this
3808 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3810 unsigned long flags
;
3812 BUG_ON(virt_to_cache(objp
) != cachep
);
3814 local_irq_save(flags
);
3815 debug_check_no_locks_freed(objp
, obj_size(cachep
));
3816 __cache_free(cachep
, objp
);
3817 local_irq_restore(flags
);
3819 EXPORT_SYMBOL(kmem_cache_free
);
3822 * kfree - free previously allocated memory
3823 * @objp: pointer returned by kmalloc.
3825 * If @objp is NULL, no operation is performed.
3827 * Don't free memory not originally allocated by kmalloc()
3828 * or you will run into trouble.
3830 void kfree(const void *objp
)
3832 struct kmem_cache
*c
;
3833 unsigned long flags
;
3835 if (unlikely(!objp
))
3837 local_irq_save(flags
);
3838 kfree_debugcheck(objp
);
3839 c
= virt_to_cache(objp
);
3840 debug_check_no_locks_freed(objp
, obj_size(c
));
3841 __cache_free(c
, (void *)objp
);
3842 local_irq_restore(flags
);
3844 EXPORT_SYMBOL(kfree
);
3846 unsigned int kmem_cache_size(struct kmem_cache
*cachep
)
3848 return obj_size(cachep
);
3850 EXPORT_SYMBOL(kmem_cache_size
);
3852 const char *kmem_cache_name(struct kmem_cache
*cachep
)
3854 return cachep
->name
;
3856 EXPORT_SYMBOL_GPL(kmem_cache_name
);
3859 * This initializes kmem_list3 or resizes varioius caches for all nodes.
3861 static int alloc_kmemlist(struct kmem_cache
*cachep
)
3864 struct kmem_list3
*l3
;
3865 struct array_cache
*new_shared
;
3866 struct array_cache
**new_alien
= NULL
;
3868 for_each_online_node(node
) {
3870 if (use_alien_caches
) {
3871 new_alien
= alloc_alien_cache(node
, cachep
->limit
);
3877 if (cachep
->shared
) {
3878 new_shared
= alloc_arraycache(node
,
3879 cachep
->shared
*cachep
->batchcount
,
3882 free_alien_cache(new_alien
);
3887 l3
= cachep
->nodelists
[node
];
3889 struct array_cache
*shared
= l3
->shared
;
3891 spin_lock_irq(&l3
->list_lock
);
3894 free_block(cachep
, shared
->entry
,
3895 shared
->avail
, node
);
3897 l3
->shared
= new_shared
;
3899 l3
->alien
= new_alien
;
3902 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3903 cachep
->batchcount
+ cachep
->num
;
3904 spin_unlock_irq(&l3
->list_lock
);
3906 free_alien_cache(new_alien
);
3909 l3
= kmalloc_node(sizeof(struct kmem_list3
), GFP_KERNEL
, node
);
3911 free_alien_cache(new_alien
);
3916 kmem_list3_init(l3
);
3917 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
3918 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
3919 l3
->shared
= new_shared
;
3920 l3
->alien
= new_alien
;
3921 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3922 cachep
->batchcount
+ cachep
->num
;
3923 cachep
->nodelists
[node
] = l3
;
3928 if (!cachep
->next
.next
) {
3929 /* Cache is not active yet. Roll back what we did */
3932 if (cachep
->nodelists
[node
]) {
3933 l3
= cachep
->nodelists
[node
];
3936 free_alien_cache(l3
->alien
);
3938 cachep
->nodelists
[node
] = NULL
;
3946 struct ccupdate_struct
{
3947 struct kmem_cache
*cachep
;
3948 struct array_cache
*new[NR_CPUS
];
3951 static void do_ccupdate_local(void *info
)
3953 struct ccupdate_struct
*new = info
;
3954 struct array_cache
*old
;
3957 old
= cpu_cache_get(new->cachep
);
3959 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
3960 new->new[smp_processor_id()] = old
;
3963 /* Always called with the cache_chain_mutex held */
3964 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3965 int batchcount
, int shared
)
3967 struct ccupdate_struct
*new;
3970 new = kzalloc(sizeof(*new), GFP_KERNEL
);
3974 for_each_online_cpu(i
) {
3975 new->new[i
] = alloc_arraycache(cpu_to_node(i
), limit
,
3978 for (i
--; i
>= 0; i
--)
3984 new->cachep
= cachep
;
3986 on_each_cpu(do_ccupdate_local
, (void *)new, 1, 1);
3989 cachep
->batchcount
= batchcount
;
3990 cachep
->limit
= limit
;
3991 cachep
->shared
= shared
;
3993 for_each_online_cpu(i
) {
3994 struct array_cache
*ccold
= new->new[i
];
3997 spin_lock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3998 free_block(cachep
, ccold
->entry
, ccold
->avail
, cpu_to_node(i
));
3999 spin_unlock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
4003 return alloc_kmemlist(cachep
);
4006 /* Called with cache_chain_mutex held always */
4007 static int enable_cpucache(struct kmem_cache
*cachep
)
4013 * The head array serves three purposes:
4014 * - create a LIFO ordering, i.e. return objects that are cache-warm
4015 * - reduce the number of spinlock operations.
4016 * - reduce the number of linked list operations on the slab and
4017 * bufctl chains: array operations are cheaper.
4018 * The numbers are guessed, we should auto-tune as described by
4021 if (cachep
->buffer_size
> 131072)
4023 else if (cachep
->buffer_size
> PAGE_SIZE
)
4025 else if (cachep
->buffer_size
> 1024)
4027 else if (cachep
->buffer_size
> 256)
4033 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4034 * allocation behaviour: Most allocs on one cpu, most free operations
4035 * on another cpu. For these cases, an efficient object passing between
4036 * cpus is necessary. This is provided by a shared array. The array
4037 * replaces Bonwick's magazine layer.
4038 * On uniprocessor, it's functionally equivalent (but less efficient)
4039 * to a larger limit. Thus disabled by default.
4042 if (cachep
->buffer_size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
4047 * With debugging enabled, large batchcount lead to excessively long
4048 * periods with disabled local interrupts. Limit the batchcount
4053 err
= do_tune_cpucache(cachep
, limit
, (limit
+ 1) / 2, shared
);
4055 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
4056 cachep
->name
, -err
);
4061 * Drain an array if it contains any elements taking the l3 lock only if
4062 * necessary. Note that the l3 listlock also protects the array_cache
4063 * if drain_array() is used on the shared array.
4065 void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
4066 struct array_cache
*ac
, int force
, int node
)
4070 if (!ac
|| !ac
->avail
)
4072 if (ac
->touched
&& !force
) {
4075 spin_lock_irq(&l3
->list_lock
);
4077 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
4078 if (tofree
> ac
->avail
)
4079 tofree
= (ac
->avail
+ 1) / 2;
4080 free_block(cachep
, ac
->entry
, tofree
, node
);
4081 ac
->avail
-= tofree
;
4082 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
4083 sizeof(void *) * ac
->avail
);
4085 spin_unlock_irq(&l3
->list_lock
);
4090 * cache_reap - Reclaim memory from caches.
4091 * @w: work descriptor
4093 * Called from workqueue/eventd every few seconds.
4095 * - clear the per-cpu caches for this CPU.
4096 * - return freeable pages to the main free memory pool.
4098 * If we cannot acquire the cache chain mutex then just give up - we'll try
4099 * again on the next iteration.
4101 static void cache_reap(struct work_struct
*w
)
4103 struct kmem_cache
*searchp
;
4104 struct kmem_list3
*l3
;
4105 int node
= numa_node_id();
4106 struct delayed_work
*work
=
4107 container_of(w
, struct delayed_work
, work
);
4109 if (!mutex_trylock(&cache_chain_mutex
))
4110 /* Give up. Setup the next iteration. */
4113 list_for_each_entry(searchp
, &cache_chain
, next
) {
4117 * We only take the l3 lock if absolutely necessary and we
4118 * have established with reasonable certainty that
4119 * we can do some work if the lock was obtained.
4121 l3
= searchp
->nodelists
[node
];
4123 reap_alien(searchp
, l3
);
4125 drain_array(searchp
, l3
, cpu_cache_get(searchp
), 0, node
);
4128 * These are racy checks but it does not matter
4129 * if we skip one check or scan twice.
4131 if (time_after(l3
->next_reap
, jiffies
))
4134 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
;
4136 drain_array(searchp
, l3
, l3
->shared
, 0, node
);
4138 if (l3
->free_touched
)
4139 l3
->free_touched
= 0;
4143 freed
= drain_freelist(searchp
, l3
, (l3
->free_limit
+
4144 5 * searchp
->num
- 1) / (5 * searchp
->num
));
4145 STATS_ADD_REAPED(searchp
, freed
);
4151 mutex_unlock(&cache_chain_mutex
);
4154 /* Set up the next iteration */
4155 schedule_delayed_work(work
, round_jiffies_relative(REAPTIMEOUT_CPUC
));
4158 #ifdef CONFIG_PROC_FS
4160 static void print_slabinfo_header(struct seq_file
*m
)
4163 * Output format version, so at least we can change it
4164 * without _too_ many complaints.
4167 seq_puts(m
, "slabinfo - version: 2.1 (statistics)\n");
4169 seq_puts(m
, "slabinfo - version: 2.1\n");
4171 seq_puts(m
, "# name <active_objs> <num_objs> <objsize> "
4172 "<objperslab> <pagesperslab>");
4173 seq_puts(m
, " : tunables <limit> <batchcount> <sharedfactor>");
4174 seq_puts(m
, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4176 seq_puts(m
, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4177 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4178 seq_puts(m
, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4183 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
4186 struct list_head
*p
;
4188 mutex_lock(&cache_chain_mutex
);
4190 print_slabinfo_header(m
);
4191 p
= cache_chain
.next
;
4194 if (p
== &cache_chain
)
4197 return list_entry(p
, struct kmem_cache
, next
);
4200 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
4202 struct kmem_cache
*cachep
= p
;
4204 return cachep
->next
.next
== &cache_chain
?
4205 NULL
: list_entry(cachep
->next
.next
, struct kmem_cache
, next
);
4208 static void s_stop(struct seq_file
*m
, void *p
)
4210 mutex_unlock(&cache_chain_mutex
);
4213 static int s_show(struct seq_file
*m
, void *p
)
4215 struct kmem_cache
*cachep
= p
;
4217 unsigned long active_objs
;
4218 unsigned long num_objs
;
4219 unsigned long active_slabs
= 0;
4220 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
4224 struct kmem_list3
*l3
;
4228 for_each_online_node(node
) {
4229 l3
= cachep
->nodelists
[node
];
4234 spin_lock_irq(&l3
->list_lock
);
4236 list_for_each_entry(slabp
, &l3
->slabs_full
, list
) {
4237 if (slabp
->inuse
!= cachep
->num
&& !error
)
4238 error
= "slabs_full accounting error";
4239 active_objs
+= cachep
->num
;
4242 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
) {
4243 if (slabp
->inuse
== cachep
->num
&& !error
)
4244 error
= "slabs_partial inuse accounting error";
4245 if (!slabp
->inuse
&& !error
)
4246 error
= "slabs_partial/inuse accounting error";
4247 active_objs
+= slabp
->inuse
;
4250 list_for_each_entry(slabp
, &l3
->slabs_free
, list
) {
4251 if (slabp
->inuse
&& !error
)
4252 error
= "slabs_free/inuse accounting error";
4255 free_objects
+= l3
->free_objects
;
4257 shared_avail
+= l3
->shared
->avail
;
4259 spin_unlock_irq(&l3
->list_lock
);
4261 num_slabs
+= active_slabs
;
4262 num_objs
= num_slabs
* cachep
->num
;
4263 if (num_objs
- active_objs
!= free_objects
&& !error
)
4264 error
= "free_objects accounting error";
4266 name
= cachep
->name
;
4268 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
4270 seq_printf(m
, "%-17s %6lu %6lu %6u %4u %4d",
4271 name
, active_objs
, num_objs
, cachep
->buffer_size
,
4272 cachep
->num
, (1 << cachep
->gfporder
));
4273 seq_printf(m
, " : tunables %4u %4u %4u",
4274 cachep
->limit
, cachep
->batchcount
, cachep
->shared
);
4275 seq_printf(m
, " : slabdata %6lu %6lu %6lu",
4276 active_slabs
, num_slabs
, shared_avail
);
4279 unsigned long high
= cachep
->high_mark
;
4280 unsigned long allocs
= cachep
->num_allocations
;
4281 unsigned long grown
= cachep
->grown
;
4282 unsigned long reaped
= cachep
->reaped
;
4283 unsigned long errors
= cachep
->errors
;
4284 unsigned long max_freeable
= cachep
->max_freeable
;
4285 unsigned long node_allocs
= cachep
->node_allocs
;
4286 unsigned long node_frees
= cachep
->node_frees
;
4287 unsigned long overflows
= cachep
->node_overflow
;
4289 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu \
4290 %4lu %4lu %4lu %4lu %4lu", allocs
, high
, grown
,
4291 reaped
, errors
, max_freeable
, node_allocs
,
4292 node_frees
, overflows
);
4296 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4297 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4298 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4299 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4301 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4302 allochit
, allocmiss
, freehit
, freemiss
);
4310 * slabinfo_op - iterator that generates /proc/slabinfo
4319 * num-pages-per-slab
4320 * + further values on SMP and with statistics enabled
4323 const struct seq_operations slabinfo_op
= {
4330 #define MAX_SLABINFO_WRITE 128
4332 * slabinfo_write - Tuning for the slab allocator
4334 * @buffer: user buffer
4335 * @count: data length
4338 ssize_t
slabinfo_write(struct file
*file
, const char __user
* buffer
,
4339 size_t count
, loff_t
*ppos
)
4341 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4342 int limit
, batchcount
, shared
, res
;
4343 struct kmem_cache
*cachep
;
4345 if (count
> MAX_SLABINFO_WRITE
)
4347 if (copy_from_user(&kbuf
, buffer
, count
))
4349 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4351 tmp
= strchr(kbuf
, ' ');
4356 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4359 /* Find the cache in the chain of caches. */
4360 mutex_lock(&cache_chain_mutex
);
4362 list_for_each_entry(cachep
, &cache_chain
, next
) {
4363 if (!strcmp(cachep
->name
, kbuf
)) {
4364 if (limit
< 1 || batchcount
< 1 ||
4365 batchcount
> limit
|| shared
< 0) {
4368 res
= do_tune_cpucache(cachep
, limit
,
4369 batchcount
, shared
);
4374 mutex_unlock(&cache_chain_mutex
);
4380 #ifdef CONFIG_DEBUG_SLAB_LEAK
4382 static void *leaks_start(struct seq_file
*m
, loff_t
*pos
)
4385 struct list_head
*p
;
4387 mutex_lock(&cache_chain_mutex
);
4388 p
= cache_chain
.next
;
4391 if (p
== &cache_chain
)
4394 return list_entry(p
, struct kmem_cache
, next
);
4397 static inline int add_caller(unsigned long *n
, unsigned long v
)
4407 unsigned long *q
= p
+ 2 * i
;
4421 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4427 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
, struct slab
*s
)
4433 for (i
= 0, p
= s
->s_mem
; i
< c
->num
; i
++, p
+= c
->buffer_size
) {
4434 if (slab_bufctl(s
)[i
] != BUFCTL_ACTIVE
)
4436 if (!add_caller(n
, (unsigned long)*dbg_userword(c
, p
)))
4441 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4443 #ifdef CONFIG_KALLSYMS
4444 unsigned long offset
, size
;
4445 char modname
[MODULE_NAME_LEN
+ 1], name
[KSYM_NAME_LEN
+ 1];
4447 if (lookup_symbol_attrs(address
, &size
, &offset
, modname
, name
) == 0) {
4448 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4450 seq_printf(m
, " [%s]", modname
);
4454 seq_printf(m
, "%p", (void *)address
);
4457 static int leaks_show(struct seq_file
*m
, void *p
)
4459 struct kmem_cache
*cachep
= p
;
4461 struct kmem_list3
*l3
;
4463 unsigned long *n
= m
->private;
4467 if (!(cachep
->flags
& SLAB_STORE_USER
))
4469 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4472 /* OK, we can do it */
4476 for_each_online_node(node
) {
4477 l3
= cachep
->nodelists
[node
];
4482 spin_lock_irq(&l3
->list_lock
);
4484 list_for_each_entry(slabp
, &l3
->slabs_full
, list
)
4485 handle_slab(n
, cachep
, slabp
);
4486 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
)
4487 handle_slab(n
, cachep
, slabp
);
4488 spin_unlock_irq(&l3
->list_lock
);
4490 name
= cachep
->name
;
4492 /* Increase the buffer size */
4493 mutex_unlock(&cache_chain_mutex
);
4494 m
->private = kzalloc(n
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4496 /* Too bad, we are really out */
4498 mutex_lock(&cache_chain_mutex
);
4501 *(unsigned long *)m
->private = n
[0] * 2;
4503 mutex_lock(&cache_chain_mutex
);
4504 /* Now make sure this entry will be retried */
4508 for (i
= 0; i
< n
[1]; i
++) {
4509 seq_printf(m
, "%s: %lu ", name
, n
[2*i
+3]);
4510 show_symbol(m
, n
[2*i
+2]);
4517 const struct seq_operations slabstats_op
= {
4518 .start
= leaks_start
,
4527 * ksize - get the actual amount of memory allocated for a given object
4528 * @objp: Pointer to the object
4530 * kmalloc may internally round up allocations and return more memory
4531 * than requested. ksize() can be used to determine the actual amount of
4532 * memory allocated. The caller may use this additional memory, even though
4533 * a smaller amount of memory was initially specified with the kmalloc call.
4534 * The caller must guarantee that objp points to a valid object previously
4535 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4536 * must not be freed during the duration of the call.
4538 size_t ksize(const void *objp
)
4540 if (unlikely(objp
== NULL
))
4543 return obj_size(virt_to_cache(objp
));