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 kmem_cache_t and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/config.h>
90 #include <linux/slab.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/seq_file.h>
98 #include <linux/notifier.h>
99 #include <linux/kallsyms.h>
100 #include <linux/cpu.h>
101 #include <linux/sysctl.h>
102 #include <linux/module.h>
103 #include <linux/rcupdate.h>
104 #include <linux/string.h>
105 #include <linux/nodemask.h>
106 #include <linux/mempolicy.h>
107 #include <linux/mutex.h>
109 #include <asm/uaccess.h>
110 #include <asm/cacheflush.h>
111 #include <asm/tlbflush.h>
112 #include <asm/page.h>
115 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
116 * SLAB_RED_ZONE & SLAB_POISON.
117 * 0 for faster, smaller code (especially in the critical paths).
119 * STATS - 1 to collect stats for /proc/slabinfo.
120 * 0 for faster, smaller code (especially in the critical paths).
122 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
125 #ifdef CONFIG_DEBUG_SLAB
128 #define FORCED_DEBUG 1
132 #define FORCED_DEBUG 0
135 /* Shouldn't this be in a header file somewhere? */
136 #define BYTES_PER_WORD sizeof(void *)
138 #ifndef cache_line_size
139 #define cache_line_size() L1_CACHE_BYTES
142 #ifndef ARCH_KMALLOC_MINALIGN
144 * Enforce a minimum alignment for the kmalloc caches.
145 * Usually, the kmalloc caches are cache_line_size() aligned, except when
146 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
147 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
148 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
149 * Note that this flag disables some debug features.
151 #define ARCH_KMALLOC_MINALIGN 0
154 #ifndef ARCH_SLAB_MINALIGN
156 * Enforce a minimum alignment for all caches.
157 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
158 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
159 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
160 * some debug features.
162 #define ARCH_SLAB_MINALIGN 0
165 #ifndef ARCH_KMALLOC_FLAGS
166 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
169 /* Legal flag mask for kmem_cache_create(). */
171 # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
172 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
173 SLAB_NO_REAP | SLAB_CACHE_DMA | \
174 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
175 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
178 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | SLAB_NO_REAP | \
179 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
180 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
187 * Bufctl's are used for linking objs within a slab
190 * This implementation relies on "struct page" for locating the cache &
191 * slab an object belongs to.
192 * This allows the bufctl structure to be small (one int), but limits
193 * the number of objects a slab (not a cache) can contain when off-slab
194 * bufctls are used. The limit is the size of the largest general cache
195 * that does not use off-slab slabs.
196 * For 32bit archs with 4 kB pages, is this 56.
197 * This is not serious, as it is only for large objects, when it is unwise
198 * to have too many per slab.
199 * Note: This limit can be raised by introducing a general cache whose size
200 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
203 typedef unsigned int kmem_bufctl_t
;
204 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
205 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
206 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-2)
208 /* Max number of objs-per-slab for caches which use off-slab slabs.
209 * Needed to avoid a possible looping condition in cache_grow().
211 static unsigned long offslab_limit
;
216 * Manages the objs in a slab. Placed either at the beginning of mem allocated
217 * for a slab, or allocated from an general cache.
218 * Slabs are chained into three list: fully used, partial, fully free slabs.
221 struct list_head list
;
222 unsigned long colouroff
;
223 void *s_mem
; /* including colour offset */
224 unsigned int inuse
; /* num of objs active in slab */
226 unsigned short nodeid
;
232 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
233 * arrange for kmem_freepages to be called via RCU. This is useful if
234 * we need to approach a kernel structure obliquely, from its address
235 * obtained without the usual locking. We can lock the structure to
236 * stabilize it and check it's still at the given address, only if we
237 * can be sure that the memory has not been meanwhile reused for some
238 * other kind of object (which our subsystem's lock might corrupt).
240 * rcu_read_lock before reading the address, then rcu_read_unlock after
241 * taking the spinlock within the structure expected at that address.
243 * We assume struct slab_rcu can overlay struct slab when destroying.
246 struct rcu_head head
;
247 kmem_cache_t
*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 [].
277 /* bootstrap: The caches do not work without cpuarrays anymore,
278 * but the cpuarrays are allocated from the generic caches...
280 #define BOOT_CPUCACHE_ENTRIES 1
281 struct arraycache_init
{
282 struct array_cache cache
;
283 void *entries
[BOOT_CPUCACHE_ENTRIES
];
287 * The slab lists for all objects.
290 struct list_head slabs_partial
; /* partial list first, better asm code */
291 struct list_head slabs_full
;
292 struct list_head slabs_free
;
293 unsigned long free_objects
;
294 unsigned long next_reap
;
296 unsigned int free_limit
;
297 spinlock_t list_lock
;
298 struct array_cache
*shared
; /* shared per node */
299 struct array_cache
**alien
; /* on other nodes */
303 * Need this for bootstrapping a per node allocator.
305 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
306 struct kmem_list3 __initdata initkmem_list3
[NUM_INIT_LISTS
];
307 #define CACHE_CACHE 0
309 #define SIZE_L3 (1 + MAX_NUMNODES)
312 * This function must be completely optimized away if
313 * a constant is passed to it. Mostly the same as
314 * what is in linux/slab.h except it returns an
317 static __always_inline
int index_of(const size_t size
)
319 if (__builtin_constant_p(size
)) {
327 #include "linux/kmalloc_sizes.h"
330 extern void __bad_size(void);
338 #define INDEX_AC index_of(sizeof(struct arraycache_init))
339 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
341 static inline void kmem_list3_init(struct kmem_list3
*parent
)
343 INIT_LIST_HEAD(&parent
->slabs_full
);
344 INIT_LIST_HEAD(&parent
->slabs_partial
);
345 INIT_LIST_HEAD(&parent
->slabs_free
);
346 parent
->shared
= NULL
;
347 parent
->alien
= NULL
;
348 spin_lock_init(&parent
->list_lock
);
349 parent
->free_objects
= 0;
350 parent
->free_touched
= 0;
353 #define MAKE_LIST(cachep, listp, slab, nodeid) \
355 INIT_LIST_HEAD(listp); \
356 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
359 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
361 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
362 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
363 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
373 /* 1) per-cpu data, touched during every alloc/free */
374 struct array_cache
*array
[NR_CPUS
];
375 unsigned int batchcount
;
378 unsigned int objsize
;
379 /* 2) touched by every alloc & free from the backend */
380 struct kmem_list3
*nodelists
[MAX_NUMNODES
];
381 unsigned int flags
; /* constant flags */
382 unsigned int num
; /* # of objs per slab */
385 /* 3) cache_grow/shrink */
386 /* order of pgs per slab (2^n) */
387 unsigned int gfporder
;
389 /* force GFP flags, e.g. GFP_DMA */
392 size_t colour
; /* cache colouring range */
393 unsigned int colour_off
; /* colour offset */
394 unsigned int colour_next
; /* cache colouring */
395 kmem_cache_t
*slabp_cache
;
396 unsigned int slab_size
;
397 unsigned int dflags
; /* dynamic flags */
399 /* constructor func */
400 void (*ctor
) (void *, kmem_cache_t
*, unsigned long);
402 /* de-constructor func */
403 void (*dtor
) (void *, kmem_cache_t
*, unsigned long);
405 /* 4) cache creation/removal */
407 struct list_head next
;
411 unsigned long num_active
;
412 unsigned long num_allocations
;
413 unsigned long high_mark
;
415 unsigned long reaped
;
416 unsigned long errors
;
417 unsigned long max_freeable
;
418 unsigned long node_allocs
;
419 unsigned long node_frees
;
431 #define CFLGS_OFF_SLAB (0x80000000UL)
432 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
434 #define BATCHREFILL_LIMIT 16
435 /* Optimization question: fewer reaps means less
436 * probability for unnessary cpucache drain/refill cycles.
438 * OTOH the cpuarrays can contain lots of objects,
439 * which could lock up otherwise freeable slabs.
441 #define REAPTIMEOUT_CPUC (2*HZ)
442 #define REAPTIMEOUT_LIST3 (4*HZ)
445 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
446 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
447 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
448 #define STATS_INC_GROWN(x) ((x)->grown++)
449 #define STATS_INC_REAPED(x) ((x)->reaped++)
450 #define STATS_SET_HIGH(x) do { if ((x)->num_active > (x)->high_mark) \
451 (x)->high_mark = (x)->num_active; \
453 #define STATS_INC_ERR(x) ((x)->errors++)
454 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
455 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
456 #define STATS_SET_FREEABLE(x, i) \
457 do { if ((x)->max_freeable < i) \
458 (x)->max_freeable = i; \
461 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
462 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
463 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
464 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
466 #define STATS_INC_ACTIVE(x) do { } while (0)
467 #define STATS_DEC_ACTIVE(x) do { } while (0)
468 #define STATS_INC_ALLOCED(x) do { } while (0)
469 #define STATS_INC_GROWN(x) do { } while (0)
470 #define STATS_INC_REAPED(x) do { } while (0)
471 #define STATS_SET_HIGH(x) do { } while (0)
472 #define STATS_INC_ERR(x) do { } while (0)
473 #define STATS_INC_NODEALLOCS(x) do { } while (0)
474 #define STATS_INC_NODEFREES(x) do { } while (0)
475 #define STATS_SET_FREEABLE(x, i) \
478 #define STATS_INC_ALLOCHIT(x) do { } while (0)
479 #define STATS_INC_ALLOCMISS(x) do { } while (0)
480 #define STATS_INC_FREEHIT(x) do { } while (0)
481 #define STATS_INC_FREEMISS(x) do { } while (0)
485 /* Magic nums for obj red zoning.
486 * Placed in the first word before and the first word after an obj.
488 #define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
489 #define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
491 /* ...and for poisoning */
492 #define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
493 #define POISON_FREE 0x6b /* for use-after-free poisoning */
494 #define POISON_END 0xa5 /* end-byte of poisoning */
496 /* memory layout of objects:
498 * 0 .. cachep->dbghead - BYTES_PER_WORD - 1: padding. This ensures that
499 * the end of an object is aligned with the end of the real
500 * allocation. Catches writes behind the end of the allocation.
501 * cachep->dbghead - BYTES_PER_WORD .. cachep->dbghead - 1:
503 * cachep->dbghead: The real object.
504 * cachep->objsize - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
505 * cachep->objsize - 1* BYTES_PER_WORD: last caller address [BYTES_PER_WORD long]
507 static int obj_dbghead(kmem_cache_t
*cachep
)
509 return cachep
->dbghead
;
512 static int obj_reallen(kmem_cache_t
*cachep
)
514 return cachep
->reallen
;
517 static unsigned long *dbg_redzone1(kmem_cache_t
*cachep
, void *objp
)
519 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
520 return (unsigned long*) (objp
+obj_dbghead(cachep
)-BYTES_PER_WORD
);
523 static unsigned long *dbg_redzone2(kmem_cache_t
*cachep
, void *objp
)
525 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
526 if (cachep
->flags
& SLAB_STORE_USER
)
527 return (unsigned long *)(objp
+ cachep
->objsize
-
529 return (unsigned long *)(objp
+ cachep
->objsize
- BYTES_PER_WORD
);
532 static void **dbg_userword(kmem_cache_t
*cachep
, void *objp
)
534 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
535 return (void **)(objp
+ cachep
->objsize
- BYTES_PER_WORD
);
540 #define obj_dbghead(x) 0
541 #define obj_reallen(cachep) (cachep->objsize)
542 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
543 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
544 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
549 * Maximum size of an obj (in 2^order pages)
550 * and absolute limit for the gfp order.
552 #if defined(CONFIG_LARGE_ALLOCS)
553 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
554 #define MAX_GFP_ORDER 13 /* up to 32Mb */
555 #elif defined(CONFIG_MMU)
556 #define MAX_OBJ_ORDER 5 /* 32 pages */
557 #define MAX_GFP_ORDER 5 /* 32 pages */
559 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
560 #define MAX_GFP_ORDER 8 /* up to 1Mb */
564 * Do not go above this order unless 0 objects fit into the slab.
566 #define BREAK_GFP_ORDER_HI 1
567 #define BREAK_GFP_ORDER_LO 0
568 static int slab_break_gfp_order
= BREAK_GFP_ORDER_LO
;
570 /* Functions for storing/retrieving the cachep and or slab from the
571 * global 'mem_map'. These are used to find the slab an obj belongs to.
572 * With kfree(), these are used to find the cache which an obj belongs to.
574 static inline void page_set_cache(struct page
*page
, struct kmem_cache
*cache
)
576 page
->lru
.next
= (struct list_head
*)cache
;
579 static inline struct kmem_cache
*page_get_cache(struct page
*page
)
581 return (struct kmem_cache
*)page
->lru
.next
;
584 static inline void page_set_slab(struct page
*page
, struct slab
*slab
)
586 page
->lru
.prev
= (struct list_head
*)slab
;
589 static inline struct slab
*page_get_slab(struct page
*page
)
591 return (struct slab
*)page
->lru
.prev
;
594 /* These are the default caches for kmalloc. Custom caches can have other sizes. */
595 struct cache_sizes malloc_sizes
[] = {
596 #define CACHE(x) { .cs_size = (x) },
597 #include <linux/kmalloc_sizes.h>
601 EXPORT_SYMBOL(malloc_sizes
);
603 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
609 static struct cache_names __initdata cache_names
[] = {
610 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
611 #include <linux/kmalloc_sizes.h>
616 static struct arraycache_init initarray_cache __initdata
=
617 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
618 static struct arraycache_init initarray_generic
=
619 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
621 /* internal cache of cache description objs */
622 static kmem_cache_t cache_cache
= {
624 .limit
= BOOT_CPUCACHE_ENTRIES
,
626 .objsize
= sizeof(kmem_cache_t
),
627 .flags
= SLAB_NO_REAP
,
628 .spinlock
= SPIN_LOCK_UNLOCKED
,
629 .name
= "kmem_cache",
631 .reallen
= sizeof(kmem_cache_t
),
635 /* Guard access to the cache-chain. */
636 static DEFINE_MUTEX(cache_chain_mutex
);
637 static struct list_head cache_chain
;
640 * vm_enough_memory() looks at this to determine how many
641 * slab-allocated pages are possibly freeable under pressure
643 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
645 atomic_t slab_reclaim_pages
;
648 * chicken and egg problem: delay the per-cpu array allocation
649 * until the general caches are up.
658 static DEFINE_PER_CPU(struct work_struct
, reap_work
);
660 static void free_block(kmem_cache_t
*cachep
, void **objpp
, int len
, int node
);
661 static void enable_cpucache(kmem_cache_t
*cachep
);
662 static void cache_reap(void *unused
);
663 static int __node_shrink(kmem_cache_t
*cachep
, int node
);
665 static inline struct array_cache
*ac_data(kmem_cache_t
*cachep
)
667 return cachep
->array
[smp_processor_id()];
670 static inline kmem_cache_t
*__find_general_cachep(size_t size
, gfp_t gfpflags
)
672 struct cache_sizes
*csizep
= malloc_sizes
;
675 /* This happens if someone tries to call
676 * kmem_cache_create(), or __kmalloc(), before
677 * the generic caches are initialized.
679 BUG_ON(malloc_sizes
[INDEX_AC
].cs_cachep
== NULL
);
681 while (size
> csizep
->cs_size
)
685 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
686 * has cs_{dma,}cachep==NULL. Thus no special case
687 * for large kmalloc calls required.
689 if (unlikely(gfpflags
& GFP_DMA
))
690 return csizep
->cs_dmacachep
;
691 return csizep
->cs_cachep
;
694 kmem_cache_t
*kmem_find_general_cachep(size_t size
, gfp_t gfpflags
)
696 return __find_general_cachep(size
, gfpflags
);
698 EXPORT_SYMBOL(kmem_find_general_cachep
);
700 /* Cal the num objs, wastage, and bytes left over for a given slab size. */
701 static void cache_estimate(unsigned long gfporder
, size_t size
, size_t align
,
702 int flags
, size_t *left_over
, unsigned int *num
)
705 size_t wastage
= PAGE_SIZE
<< gfporder
;
709 if (!(flags
& CFLGS_OFF_SLAB
)) {
710 base
= sizeof(struct slab
);
711 extra
= sizeof(kmem_bufctl_t
);
714 while (i
* size
+ ALIGN(base
+ i
* extra
, align
) <= wastage
)
724 wastage
-= ALIGN(base
+ i
* extra
, align
);
725 *left_over
= wastage
;
728 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
730 static void __slab_error(const char *function
, kmem_cache_t
*cachep
, char *msg
)
732 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
733 function
, cachep
->name
, msg
);
738 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
739 * via the workqueue/eventd.
740 * Add the CPU number into the expiration time to minimize the possibility of
741 * the CPUs getting into lockstep and contending for the global cache chain
744 static void __devinit
start_cpu_timer(int cpu
)
746 struct work_struct
*reap_work
= &per_cpu(reap_work
, cpu
);
749 * When this gets called from do_initcalls via cpucache_init(),
750 * init_workqueues() has already run, so keventd will be setup
753 if (keventd_up() && reap_work
->func
== NULL
) {
754 INIT_WORK(reap_work
, cache_reap
, NULL
);
755 schedule_delayed_work_on(cpu
, reap_work
, HZ
+ 3 * cpu
);
759 static struct array_cache
*alloc_arraycache(int node
, int entries
,
762 int memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
763 struct array_cache
*nc
= NULL
;
765 nc
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
769 nc
->batchcount
= batchcount
;
771 spin_lock_init(&nc
->lock
);
777 static void *__cache_alloc_node(kmem_cache_t
*, gfp_t
, int);
779 static inline struct array_cache
**alloc_alien_cache(int node
, int limit
)
781 struct array_cache
**ac_ptr
;
782 int memsize
= sizeof(void *) * MAX_NUMNODES
;
787 ac_ptr
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
790 if (i
== node
|| !node_online(i
)) {
794 ac_ptr
[i
] = alloc_arraycache(node
, limit
, 0xbaadf00d);
796 for (i
--; i
<= 0; i
--)
806 static inline void free_alien_cache(struct array_cache
**ac_ptr
)
819 static inline void __drain_alien_cache(kmem_cache_t
*cachep
,
820 struct array_cache
*ac
, int node
)
822 struct kmem_list3
*rl3
= cachep
->nodelists
[node
];
825 spin_lock(&rl3
->list_lock
);
826 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
828 spin_unlock(&rl3
->list_lock
);
832 static void drain_alien_cache(kmem_cache_t
*cachep
, struct kmem_list3
*l3
)
835 struct array_cache
*ac
;
838 for_each_online_node(i
) {
841 spin_lock_irqsave(&ac
->lock
, flags
);
842 __drain_alien_cache(cachep
, ac
, i
);
843 spin_unlock_irqrestore(&ac
->lock
, flags
);
848 #define alloc_alien_cache(node, limit) do { } while (0)
849 #define free_alien_cache(ac_ptr) do { } while (0)
850 #define drain_alien_cache(cachep, l3) do { } while (0)
853 static int __devinit
cpuup_callback(struct notifier_block
*nfb
,
854 unsigned long action
, void *hcpu
)
856 long cpu
= (long)hcpu
;
857 kmem_cache_t
*cachep
;
858 struct kmem_list3
*l3
= NULL
;
859 int node
= cpu_to_node(cpu
);
860 int memsize
= sizeof(struct kmem_list3
);
864 mutex_lock(&cache_chain_mutex
);
865 /* we need to do this right in the beginning since
866 * alloc_arraycache's are going to use this list.
867 * kmalloc_node allows us to add the slab to the right
868 * kmem_list3 and not this cpu's kmem_list3
871 list_for_each_entry(cachep
, &cache_chain
, next
) {
872 /* setup the size64 kmemlist for cpu before we can
873 * begin anything. Make sure some other cpu on this
874 * node has not already allocated this
876 if (!cachep
->nodelists
[node
]) {
877 if (!(l3
= kmalloc_node(memsize
,
881 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
882 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
884 cachep
->nodelists
[node
] = l3
;
887 spin_lock_irq(&cachep
->nodelists
[node
]->list_lock
);
888 cachep
->nodelists
[node
]->free_limit
=
889 (1 + nr_cpus_node(node
)) *
890 cachep
->batchcount
+ cachep
->num
;
891 spin_unlock_irq(&cachep
->nodelists
[node
]->list_lock
);
894 /* Now we can go ahead with allocating the shared array's
896 list_for_each_entry(cachep
, &cache_chain
, next
) {
897 struct array_cache
*nc
;
899 nc
= alloc_arraycache(node
, cachep
->limit
,
903 cachep
->array
[cpu
] = nc
;
905 l3
= cachep
->nodelists
[node
];
908 if (!(nc
= alloc_arraycache(node
,
914 /* we are serialised from CPU_DEAD or
915 CPU_UP_CANCELLED by the cpucontrol lock */
919 mutex_unlock(&cache_chain_mutex
);
922 start_cpu_timer(cpu
);
924 #ifdef CONFIG_HOTPLUG_CPU
927 case CPU_UP_CANCELED
:
928 mutex_lock(&cache_chain_mutex
);
930 list_for_each_entry(cachep
, &cache_chain
, next
) {
931 struct array_cache
*nc
;
934 mask
= node_to_cpumask(node
);
935 spin_lock_irq(&cachep
->spinlock
);
936 /* cpu is dead; no one can alloc from it. */
937 nc
= cachep
->array
[cpu
];
938 cachep
->array
[cpu
] = NULL
;
939 l3
= cachep
->nodelists
[node
];
944 spin_lock(&l3
->list_lock
);
946 /* Free limit for this kmem_list3 */
947 l3
->free_limit
-= cachep
->batchcount
;
949 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
951 if (!cpus_empty(mask
)) {
952 spin_unlock(&l3
->list_lock
);
957 free_block(cachep
, l3
->shared
->entry
,
958 l3
->shared
->avail
, node
);
963 drain_alien_cache(cachep
, l3
);
964 free_alien_cache(l3
->alien
);
968 /* free slabs belonging to this node */
969 if (__node_shrink(cachep
, node
)) {
970 cachep
->nodelists
[node
] = NULL
;
971 spin_unlock(&l3
->list_lock
);
974 spin_unlock(&l3
->list_lock
);
977 spin_unlock_irq(&cachep
->spinlock
);
980 mutex_unlock(&cache_chain_mutex
);
986 mutex_unlock(&cache_chain_mutex
);
990 static struct notifier_block cpucache_notifier
= { &cpuup_callback
, NULL
, 0 };
993 * swap the static kmem_list3 with kmalloced memory
995 static void init_list(kmem_cache_t
*cachep
, struct kmem_list3
*list
, int nodeid
)
997 struct kmem_list3
*ptr
;
999 BUG_ON(cachep
->nodelists
[nodeid
] != list
);
1000 ptr
= kmalloc_node(sizeof(struct kmem_list3
), GFP_KERNEL
, nodeid
);
1003 local_irq_disable();
1004 memcpy(ptr
, list
, sizeof(struct kmem_list3
));
1005 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1006 cachep
->nodelists
[nodeid
] = ptr
;
1011 * Called after the gfp() functions have been enabled, and before smp_init().
1013 void __init
kmem_cache_init(void)
1016 struct cache_sizes
*sizes
;
1017 struct cache_names
*names
;
1020 for (i
= 0; i
< NUM_INIT_LISTS
; i
++) {
1021 kmem_list3_init(&initkmem_list3
[i
]);
1022 if (i
< MAX_NUMNODES
)
1023 cache_cache
.nodelists
[i
] = NULL
;
1027 * Fragmentation resistance on low memory - only use bigger
1028 * page orders on machines with more than 32MB of memory.
1030 if (num_physpages
> (32 << 20) >> PAGE_SHIFT
)
1031 slab_break_gfp_order
= BREAK_GFP_ORDER_HI
;
1033 /* Bootstrap is tricky, because several objects are allocated
1034 * from caches that do not exist yet:
1035 * 1) initialize the cache_cache cache: it contains the kmem_cache_t
1036 * structures of all caches, except cache_cache itself: cache_cache
1037 * is statically allocated.
1038 * Initially an __init data area is used for the head array and the
1039 * kmem_list3 structures, it's replaced with a kmalloc allocated
1040 * array at the end of the bootstrap.
1041 * 2) Create the first kmalloc cache.
1042 * The kmem_cache_t for the new cache is allocated normally.
1043 * An __init data area is used for the head array.
1044 * 3) Create the remaining kmalloc caches, with minimally sized
1046 * 4) Replace the __init data head arrays for cache_cache and the first
1047 * kmalloc cache with kmalloc allocated arrays.
1048 * 5) Replace the __init data for kmem_list3 for cache_cache and
1049 * the other cache's with kmalloc allocated memory.
1050 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1053 /* 1) create the cache_cache */
1054 INIT_LIST_HEAD(&cache_chain
);
1055 list_add(&cache_cache
.next
, &cache_chain
);
1056 cache_cache
.colour_off
= cache_line_size();
1057 cache_cache
.array
[smp_processor_id()] = &initarray_cache
.cache
;
1058 cache_cache
.nodelists
[numa_node_id()] = &initkmem_list3
[CACHE_CACHE
];
1060 cache_cache
.objsize
= ALIGN(cache_cache
.objsize
, cache_line_size());
1062 cache_estimate(0, cache_cache
.objsize
, cache_line_size(), 0,
1063 &left_over
, &cache_cache
.num
);
1064 if (!cache_cache
.num
)
1067 cache_cache
.colour
= left_over
/ cache_cache
.colour_off
;
1068 cache_cache
.colour_next
= 0;
1069 cache_cache
.slab_size
= ALIGN(cache_cache
.num
* sizeof(kmem_bufctl_t
) +
1070 sizeof(struct slab
), cache_line_size());
1072 /* 2+3) create the kmalloc caches */
1073 sizes
= malloc_sizes
;
1074 names
= cache_names
;
1076 /* Initialize the caches that provide memory for the array cache
1077 * and the kmem_list3 structures first.
1078 * Without this, further allocations will bug
1081 sizes
[INDEX_AC
].cs_cachep
= kmem_cache_create(names
[INDEX_AC
].name
,
1082 sizes
[INDEX_AC
].cs_size
,
1083 ARCH_KMALLOC_MINALIGN
,
1084 (ARCH_KMALLOC_FLAGS
|
1085 SLAB_PANIC
), NULL
, NULL
);
1087 if (INDEX_AC
!= INDEX_L3
)
1088 sizes
[INDEX_L3
].cs_cachep
=
1089 kmem_cache_create(names
[INDEX_L3
].name
,
1090 sizes
[INDEX_L3
].cs_size
,
1091 ARCH_KMALLOC_MINALIGN
,
1092 (ARCH_KMALLOC_FLAGS
| SLAB_PANIC
), NULL
,
1095 while (sizes
->cs_size
!= ULONG_MAX
) {
1097 * For performance, all the general caches are L1 aligned.
1098 * This should be particularly beneficial on SMP boxes, as it
1099 * eliminates "false sharing".
1100 * Note for systems short on memory removing the alignment will
1101 * allow tighter packing of the smaller caches.
1103 if (!sizes
->cs_cachep
)
1104 sizes
->cs_cachep
= kmem_cache_create(names
->name
,
1106 ARCH_KMALLOC_MINALIGN
,
1111 /* Inc off-slab bufctl limit until the ceiling is hit. */
1112 if (!(OFF_SLAB(sizes
->cs_cachep
))) {
1113 offslab_limit
= sizes
->cs_size
- sizeof(struct slab
);
1114 offslab_limit
/= sizeof(kmem_bufctl_t
);
1117 sizes
->cs_dmacachep
= kmem_cache_create(names
->name_dma
,
1119 ARCH_KMALLOC_MINALIGN
,
1120 (ARCH_KMALLOC_FLAGS
|
1128 /* 4) Replace the bootstrap head arrays */
1132 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1134 local_irq_disable();
1135 BUG_ON(ac_data(&cache_cache
) != &initarray_cache
.cache
);
1136 memcpy(ptr
, ac_data(&cache_cache
),
1137 sizeof(struct arraycache_init
));
1138 cache_cache
.array
[smp_processor_id()] = ptr
;
1141 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1143 local_irq_disable();
1144 BUG_ON(ac_data(malloc_sizes
[INDEX_AC
].cs_cachep
)
1145 != &initarray_generic
.cache
);
1146 memcpy(ptr
, ac_data(malloc_sizes
[INDEX_AC
].cs_cachep
),
1147 sizeof(struct arraycache_init
));
1148 malloc_sizes
[INDEX_AC
].cs_cachep
->array
[smp_processor_id()] =
1152 /* 5) Replace the bootstrap kmem_list3's */
1155 /* Replace the static kmem_list3 structures for the boot cpu */
1156 init_list(&cache_cache
, &initkmem_list3
[CACHE_CACHE
],
1159 for_each_online_node(node
) {
1160 init_list(malloc_sizes
[INDEX_AC
].cs_cachep
,
1161 &initkmem_list3
[SIZE_AC
+ node
], node
);
1163 if (INDEX_AC
!= INDEX_L3
) {
1164 init_list(malloc_sizes
[INDEX_L3
].cs_cachep
,
1165 &initkmem_list3
[SIZE_L3
+ node
],
1171 /* 6) resize the head arrays to their final sizes */
1173 kmem_cache_t
*cachep
;
1174 mutex_lock(&cache_chain_mutex
);
1175 list_for_each_entry(cachep
, &cache_chain
, next
)
1176 enable_cpucache(cachep
);
1177 mutex_unlock(&cache_chain_mutex
);
1181 g_cpucache_up
= FULL
;
1183 /* Register a cpu startup notifier callback
1184 * that initializes ac_data for all new cpus
1186 register_cpu_notifier(&cpucache_notifier
);
1188 /* The reap timers are started later, with a module init call:
1189 * That part of the kernel is not yet operational.
1193 static int __init
cpucache_init(void)
1198 * Register the timers that return unneeded
1201 for_each_online_cpu(cpu
)
1202 start_cpu_timer(cpu
);
1207 __initcall(cpucache_init
);
1210 * Interface to system's page allocator. No need to hold the cache-lock.
1212 * If we requested dmaable memory, we will get it. Even if we
1213 * did not request dmaable memory, we might get it, but that
1214 * would be relatively rare and ignorable.
1216 static void *kmem_getpages(kmem_cache_t
*cachep
, gfp_t flags
, int nodeid
)
1222 flags
|= cachep
->gfpflags
;
1223 page
= alloc_pages_node(nodeid
, flags
, cachep
->gfporder
);
1226 addr
= page_address(page
);
1228 i
= (1 << cachep
->gfporder
);
1229 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1230 atomic_add(i
, &slab_reclaim_pages
);
1231 add_page_state(nr_slab
, i
);
1240 * Interface to system's page release.
1242 static void kmem_freepages(kmem_cache_t
*cachep
, void *addr
)
1244 unsigned long i
= (1 << cachep
->gfporder
);
1245 struct page
*page
= virt_to_page(addr
);
1246 const unsigned long nr_freed
= i
;
1249 if (!TestClearPageSlab(page
))
1253 sub_page_state(nr_slab
, nr_freed
);
1254 if (current
->reclaim_state
)
1255 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1256 free_pages((unsigned long)addr
, cachep
->gfporder
);
1257 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1258 atomic_sub(1 << cachep
->gfporder
, &slab_reclaim_pages
);
1261 static void kmem_rcu_free(struct rcu_head
*head
)
1263 struct slab_rcu
*slab_rcu
= (struct slab_rcu
*)head
;
1264 kmem_cache_t
*cachep
= slab_rcu
->cachep
;
1266 kmem_freepages(cachep
, slab_rcu
->addr
);
1267 if (OFF_SLAB(cachep
))
1268 kmem_cache_free(cachep
->slabp_cache
, slab_rcu
);
1273 #ifdef CONFIG_DEBUG_PAGEALLOC
1274 static void store_stackinfo(kmem_cache_t
*cachep
, unsigned long *addr
,
1275 unsigned long caller
)
1277 int size
= obj_reallen(cachep
);
1279 addr
= (unsigned long *)&((char *)addr
)[obj_dbghead(cachep
)];
1281 if (size
< 5 * sizeof(unsigned long))
1284 *addr
++ = 0x12345678;
1286 *addr
++ = smp_processor_id();
1287 size
-= 3 * sizeof(unsigned long);
1289 unsigned long *sptr
= &caller
;
1290 unsigned long svalue
;
1292 while (!kstack_end(sptr
)) {
1294 if (kernel_text_address(svalue
)) {
1296 size
-= sizeof(unsigned long);
1297 if (size
<= sizeof(unsigned long))
1303 *addr
++ = 0x87654321;
1307 static void poison_obj(kmem_cache_t
*cachep
, void *addr
, unsigned char val
)
1309 int size
= obj_reallen(cachep
);
1310 addr
= &((char *)addr
)[obj_dbghead(cachep
)];
1312 memset(addr
, val
, size
);
1313 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1316 static void dump_line(char *data
, int offset
, int limit
)
1319 printk(KERN_ERR
"%03x:", offset
);
1320 for (i
= 0; i
< limit
; i
++) {
1321 printk(" %02x", (unsigned char)data
[offset
+ i
]);
1329 static void print_objinfo(kmem_cache_t
*cachep
, void *objp
, int lines
)
1334 if (cachep
->flags
& SLAB_RED_ZONE
) {
1335 printk(KERN_ERR
"Redzone: 0x%lx/0x%lx.\n",
1336 *dbg_redzone1(cachep
, objp
),
1337 *dbg_redzone2(cachep
, objp
));
1340 if (cachep
->flags
& SLAB_STORE_USER
) {
1341 printk(KERN_ERR
"Last user: [<%p>]",
1342 *dbg_userword(cachep
, objp
));
1343 print_symbol("(%s)",
1344 (unsigned long)*dbg_userword(cachep
, objp
));
1347 realobj
= (char *)objp
+ obj_dbghead(cachep
);
1348 size
= obj_reallen(cachep
);
1349 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1352 if (i
+ limit
> size
)
1354 dump_line(realobj
, i
, limit
);
1358 static void check_poison_obj(kmem_cache_t
*cachep
, void *objp
)
1364 realobj
= (char *)objp
+ obj_dbghead(cachep
);
1365 size
= obj_reallen(cachep
);
1367 for (i
= 0; i
< size
; i
++) {
1368 char exp
= POISON_FREE
;
1371 if (realobj
[i
] != exp
) {
1377 "Slab corruption: start=%p, len=%d\n",
1379 print_objinfo(cachep
, objp
, 0);
1381 /* Hexdump the affected line */
1384 if (i
+ limit
> size
)
1386 dump_line(realobj
, i
, limit
);
1389 /* Limit to 5 lines */
1395 /* Print some data about the neighboring objects, if they
1398 struct slab
*slabp
= page_get_slab(virt_to_page(objp
));
1401 objnr
= (objp
- slabp
->s_mem
) / cachep
->objsize
;
1403 objp
= slabp
->s_mem
+ (objnr
- 1) * cachep
->objsize
;
1404 realobj
= (char *)objp
+ obj_dbghead(cachep
);
1405 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
1407 print_objinfo(cachep
, objp
, 2);
1409 if (objnr
+ 1 < cachep
->num
) {
1410 objp
= slabp
->s_mem
+ (objnr
+ 1) * cachep
->objsize
;
1411 realobj
= (char *)objp
+ obj_dbghead(cachep
);
1412 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
1414 print_objinfo(cachep
, objp
, 2);
1420 /* Destroy all the objs in a slab, and release the mem back to the system.
1421 * Before calling the slab must have been unlinked from the cache.
1422 * The cache-lock is not held/needed.
1424 static void slab_destroy(kmem_cache_t
*cachep
, struct slab
*slabp
)
1426 void *addr
= slabp
->s_mem
- slabp
->colouroff
;
1430 for (i
= 0; i
< cachep
->num
; i
++) {
1431 void *objp
= slabp
->s_mem
+ cachep
->objsize
* i
;
1433 if (cachep
->flags
& SLAB_POISON
) {
1434 #ifdef CONFIG_DEBUG_PAGEALLOC
1435 if ((cachep
->objsize
% PAGE_SIZE
) == 0
1436 && OFF_SLAB(cachep
))
1437 kernel_map_pages(virt_to_page(objp
),
1438 cachep
->objsize
/ PAGE_SIZE
,
1441 check_poison_obj(cachep
, objp
);
1443 check_poison_obj(cachep
, objp
);
1446 if (cachep
->flags
& SLAB_RED_ZONE
) {
1447 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1448 slab_error(cachep
, "start of a freed object "
1450 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1451 slab_error(cachep
, "end of a freed object "
1454 if (cachep
->dtor
&& !(cachep
->flags
& SLAB_POISON
))
1455 (cachep
->dtor
) (objp
+ obj_dbghead(cachep
), cachep
, 0);
1460 for (i
= 0; i
< cachep
->num
; i
++) {
1461 void *objp
= slabp
->s_mem
+ cachep
->objsize
* i
;
1462 (cachep
->dtor
) (objp
, cachep
, 0);
1467 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
1468 struct slab_rcu
*slab_rcu
;
1470 slab_rcu
= (struct slab_rcu
*)slabp
;
1471 slab_rcu
->cachep
= cachep
;
1472 slab_rcu
->addr
= addr
;
1473 call_rcu(&slab_rcu
->head
, kmem_rcu_free
);
1475 kmem_freepages(cachep
, addr
);
1476 if (OFF_SLAB(cachep
))
1477 kmem_cache_free(cachep
->slabp_cache
, slabp
);
1481 /* For setting up all the kmem_list3s for cache whose objsize is same
1482 as size of kmem_list3. */
1483 static inline void set_up_list3s(kmem_cache_t
*cachep
, int index
)
1487 for_each_online_node(node
) {
1488 cachep
->nodelists
[node
] = &initkmem_list3
[index
+ node
];
1489 cachep
->nodelists
[node
]->next_reap
= jiffies
+
1491 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1496 * calculate_slab_order - calculate size (page order) of slabs and the number
1497 * of objects per slab.
1499 * This could be made much more intelligent. For now, try to avoid using
1500 * high order pages for slabs. When the gfp() functions are more friendly
1501 * towards high-order requests, this should be changed.
1503 static inline size_t calculate_slab_order(kmem_cache_t
*cachep
, size_t size
,
1504 size_t align
, gfp_t flags
)
1506 size_t left_over
= 0;
1508 for (;; cachep
->gfporder
++) {
1512 if (cachep
->gfporder
> MAX_GFP_ORDER
) {
1517 cache_estimate(cachep
->gfporder
, size
, align
, flags
,
1521 /* More than offslab_limit objects will cause problems */
1522 if (flags
& CFLGS_OFF_SLAB
&& cachep
->num
> offslab_limit
)
1526 left_over
= remainder
;
1529 * Large number of objects is good, but very large slabs are
1530 * currently bad for the gfp()s.
1532 if (cachep
->gfporder
>= slab_break_gfp_order
)
1535 if ((left_over
* 8) <= (PAGE_SIZE
<< cachep
->gfporder
))
1536 /* Acceptable internal fragmentation */
1543 * kmem_cache_create - Create a cache.
1544 * @name: A string which is used in /proc/slabinfo to identify this cache.
1545 * @size: The size of objects to be created in this cache.
1546 * @align: The required alignment for the objects.
1547 * @flags: SLAB flags
1548 * @ctor: A constructor for the objects.
1549 * @dtor: A destructor for the objects.
1551 * Returns a ptr to the cache on success, NULL on failure.
1552 * Cannot be called within a int, but can be interrupted.
1553 * The @ctor is run when new pages are allocated by the cache
1554 * and the @dtor is run before the pages are handed back.
1556 * @name must be valid until the cache is destroyed. This implies that
1557 * the module calling this has to destroy the cache before getting
1562 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1563 * to catch references to uninitialised memory.
1565 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1566 * for buffer overruns.
1568 * %SLAB_NO_REAP - Don't automatically reap this cache when we're under
1571 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1572 * cacheline. This can be beneficial if you're counting cycles as closely
1576 kmem_cache_create (const char *name
, size_t size
, size_t align
,
1577 unsigned long flags
, void (*ctor
)(void*, kmem_cache_t
*, unsigned long),
1578 void (*dtor
)(void*, kmem_cache_t
*, unsigned long))
1580 size_t left_over
, slab_size
, ralign
;
1581 kmem_cache_t
*cachep
= NULL
;
1582 struct list_head
*p
;
1585 * Sanity checks... these are all serious usage bugs.
1589 (size
< BYTES_PER_WORD
) ||
1590 (size
> (1 << MAX_OBJ_ORDER
) * PAGE_SIZE
) || (dtor
&& !ctor
)) {
1591 printk(KERN_ERR
"%s: Early error in slab %s\n",
1592 __FUNCTION__
, name
);
1596 mutex_lock(&cache_chain_mutex
);
1598 list_for_each(p
, &cache_chain
) {
1599 kmem_cache_t
*pc
= list_entry(p
, kmem_cache_t
, next
);
1600 mm_segment_t old_fs
= get_fs();
1605 * This happens when the module gets unloaded and doesn't
1606 * destroy its slab cache and no-one else reuses the vmalloc
1607 * area of the module. Print a warning.
1610 res
= __get_user(tmp
, pc
->name
);
1613 printk("SLAB: cache with size %d has lost its name\n",
1618 if (!strcmp(pc
->name
, name
)) {
1619 printk("kmem_cache_create: duplicate cache %s\n", name
);
1626 WARN_ON(strchr(name
, ' ')); /* It confuses parsers */
1627 if ((flags
& SLAB_DEBUG_INITIAL
) && !ctor
) {
1628 /* No constructor, but inital state check requested */
1629 printk(KERN_ERR
"%s: No con, but init state check "
1630 "requested - %s\n", __FUNCTION__
, name
);
1631 flags
&= ~SLAB_DEBUG_INITIAL
;
1635 * Enable redzoning and last user accounting, except for caches with
1636 * large objects, if the increased size would increase the object size
1637 * above the next power of two: caches with object sizes just above a
1638 * power of two have a significant amount of internal fragmentation.
1641 || fls(size
- 1) == fls(size
- 1 + 3 * BYTES_PER_WORD
)))
1642 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
1643 if (!(flags
& SLAB_DESTROY_BY_RCU
))
1644 flags
|= SLAB_POISON
;
1646 if (flags
& SLAB_DESTROY_BY_RCU
)
1647 BUG_ON(flags
& SLAB_POISON
);
1649 if (flags
& SLAB_DESTROY_BY_RCU
)
1653 * Always checks flags, a caller might be expecting debug
1654 * support which isn't available.
1656 if (flags
& ~CREATE_MASK
)
1659 /* Check that size is in terms of words. This is needed to avoid
1660 * unaligned accesses for some archs when redzoning is used, and makes
1661 * sure any on-slab bufctl's are also correctly aligned.
1663 if (size
& (BYTES_PER_WORD
- 1)) {
1664 size
+= (BYTES_PER_WORD
- 1);
1665 size
&= ~(BYTES_PER_WORD
- 1);
1668 /* calculate out the final buffer alignment: */
1669 /* 1) arch recommendation: can be overridden for debug */
1670 if (flags
& SLAB_HWCACHE_ALIGN
) {
1671 /* Default alignment: as specified by the arch code.
1672 * Except if an object is really small, then squeeze multiple
1673 * objects into one cacheline.
1675 ralign
= cache_line_size();
1676 while (size
<= ralign
/ 2)
1679 ralign
= BYTES_PER_WORD
;
1681 /* 2) arch mandated alignment: disables debug if necessary */
1682 if (ralign
< ARCH_SLAB_MINALIGN
) {
1683 ralign
= ARCH_SLAB_MINALIGN
;
1684 if (ralign
> BYTES_PER_WORD
)
1685 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
1687 /* 3) caller mandated alignment: disables debug if necessary */
1688 if (ralign
< align
) {
1690 if (ralign
> BYTES_PER_WORD
)
1691 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
1693 /* 4) Store it. Note that the debug code below can reduce
1694 * the alignment to BYTES_PER_WORD.
1698 /* Get cache's description obj. */
1699 cachep
= (kmem_cache_t
*) kmem_cache_alloc(&cache_cache
, SLAB_KERNEL
);
1702 memset(cachep
, 0, sizeof(kmem_cache_t
));
1705 cachep
->reallen
= size
;
1707 if (flags
& SLAB_RED_ZONE
) {
1708 /* redzoning only works with word aligned caches */
1709 align
= BYTES_PER_WORD
;
1711 /* add space for red zone words */
1712 cachep
->dbghead
+= BYTES_PER_WORD
;
1713 size
+= 2 * BYTES_PER_WORD
;
1715 if (flags
& SLAB_STORE_USER
) {
1716 /* user store requires word alignment and
1717 * one word storage behind the end of the real
1720 align
= BYTES_PER_WORD
;
1721 size
+= BYTES_PER_WORD
;
1723 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
1724 if (size
>= malloc_sizes
[INDEX_L3
+ 1].cs_size
1725 && cachep
->reallen
> cache_line_size() && size
< PAGE_SIZE
) {
1726 cachep
->dbghead
+= PAGE_SIZE
- size
;
1732 /* Determine if the slab management is 'on' or 'off' slab. */
1733 if (size
>= (PAGE_SIZE
>> 3))
1735 * Size is large, assume best to place the slab management obj
1736 * off-slab (should allow better packing of objs).
1738 flags
|= CFLGS_OFF_SLAB
;
1740 size
= ALIGN(size
, align
);
1742 if ((flags
& SLAB_RECLAIM_ACCOUNT
) && size
<= PAGE_SIZE
) {
1744 * A VFS-reclaimable slab tends to have most allocations
1745 * as GFP_NOFS and we really don't want to have to be allocating
1746 * higher-order pages when we are unable to shrink dcache.
1748 cachep
->gfporder
= 0;
1749 cache_estimate(cachep
->gfporder
, size
, align
, flags
,
1750 &left_over
, &cachep
->num
);
1752 left_over
= calculate_slab_order(cachep
, size
, align
, flags
);
1755 printk("kmem_cache_create: couldn't create cache %s.\n", name
);
1756 kmem_cache_free(&cache_cache
, cachep
);
1760 slab_size
= ALIGN(cachep
->num
* sizeof(kmem_bufctl_t
)
1761 + sizeof(struct slab
), align
);
1764 * If the slab has been placed off-slab, and we have enough space then
1765 * move it on-slab. This is at the expense of any extra colouring.
1767 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= slab_size
) {
1768 flags
&= ~CFLGS_OFF_SLAB
;
1769 left_over
-= slab_size
;
1772 if (flags
& CFLGS_OFF_SLAB
) {
1773 /* really off slab. No need for manual alignment */
1775 cachep
->num
* sizeof(kmem_bufctl_t
) + sizeof(struct slab
);
1778 cachep
->colour_off
= cache_line_size();
1779 /* Offset must be a multiple of the alignment. */
1780 if (cachep
->colour_off
< align
)
1781 cachep
->colour_off
= align
;
1782 cachep
->colour
= left_over
/ cachep
->colour_off
;
1783 cachep
->slab_size
= slab_size
;
1784 cachep
->flags
= flags
;
1785 cachep
->gfpflags
= 0;
1786 if (flags
& SLAB_CACHE_DMA
)
1787 cachep
->gfpflags
|= GFP_DMA
;
1788 spin_lock_init(&cachep
->spinlock
);
1789 cachep
->objsize
= size
;
1791 if (flags
& CFLGS_OFF_SLAB
)
1792 cachep
->slabp_cache
= kmem_find_general_cachep(slab_size
, 0u);
1793 cachep
->ctor
= ctor
;
1794 cachep
->dtor
= dtor
;
1795 cachep
->name
= name
;
1797 /* Don't let CPUs to come and go */
1800 if (g_cpucache_up
== FULL
) {
1801 enable_cpucache(cachep
);
1803 if (g_cpucache_up
== NONE
) {
1804 /* Note: the first kmem_cache_create must create
1805 * the cache that's used by kmalloc(24), otherwise
1806 * the creation of further caches will BUG().
1808 cachep
->array
[smp_processor_id()] =
1809 &initarray_generic
.cache
;
1811 /* If the cache that's used by
1812 * kmalloc(sizeof(kmem_list3)) is the first cache,
1813 * then we need to set up all its list3s, otherwise
1814 * the creation of further caches will BUG().
1816 set_up_list3s(cachep
, SIZE_AC
);
1817 if (INDEX_AC
== INDEX_L3
)
1818 g_cpucache_up
= PARTIAL_L3
;
1820 g_cpucache_up
= PARTIAL_AC
;
1822 cachep
->array
[smp_processor_id()] =
1823 kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1825 if (g_cpucache_up
== PARTIAL_AC
) {
1826 set_up_list3s(cachep
, SIZE_L3
);
1827 g_cpucache_up
= PARTIAL_L3
;
1830 for_each_online_node(node
) {
1832 cachep
->nodelists
[node
] =
1834 (struct kmem_list3
),
1836 BUG_ON(!cachep
->nodelists
[node
]);
1837 kmem_list3_init(cachep
->
1842 cachep
->nodelists
[numa_node_id()]->next_reap
=
1843 jiffies
+ REAPTIMEOUT_LIST3
+
1844 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1846 BUG_ON(!ac_data(cachep
));
1847 ac_data(cachep
)->avail
= 0;
1848 ac_data(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
1849 ac_data(cachep
)->batchcount
= 1;
1850 ac_data(cachep
)->touched
= 0;
1851 cachep
->batchcount
= 1;
1852 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
1855 /* cache setup completed, link it into the list */
1856 list_add(&cachep
->next
, &cache_chain
);
1857 unlock_cpu_hotplug();
1859 if (!cachep
&& (flags
& SLAB_PANIC
))
1860 panic("kmem_cache_create(): failed to create slab `%s'\n",
1862 mutex_unlock(&cache_chain_mutex
);
1865 EXPORT_SYMBOL(kmem_cache_create
);
1868 static void check_irq_off(void)
1870 BUG_ON(!irqs_disabled());
1873 static void check_irq_on(void)
1875 BUG_ON(irqs_disabled());
1878 static void check_spinlock_acquired(kmem_cache_t
*cachep
)
1882 assert_spin_locked(&cachep
->nodelists
[numa_node_id()]->list_lock
);
1886 static inline void check_spinlock_acquired_node(kmem_cache_t
*cachep
, int node
)
1890 assert_spin_locked(&cachep
->nodelists
[node
]->list_lock
);
1895 #define check_irq_off() do { } while(0)
1896 #define check_irq_on() do { } while(0)
1897 #define check_spinlock_acquired(x) do { } while(0)
1898 #define check_spinlock_acquired_node(x, y) do { } while(0)
1902 * Waits for all CPUs to execute func().
1904 static void smp_call_function_all_cpus(void (*func
)(void *arg
), void *arg
)
1909 local_irq_disable();
1913 if (smp_call_function(func
, arg
, 1, 1))
1919 static void drain_array_locked(kmem_cache_t
*cachep
, struct array_cache
*ac
,
1920 int force
, int node
);
1922 static void do_drain(void *arg
)
1924 kmem_cache_t
*cachep
= (kmem_cache_t
*) arg
;
1925 struct array_cache
*ac
;
1926 int node
= numa_node_id();
1929 ac
= ac_data(cachep
);
1930 spin_lock(&cachep
->nodelists
[node
]->list_lock
);
1931 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
1932 spin_unlock(&cachep
->nodelists
[node
]->list_lock
);
1936 static void drain_cpu_caches(kmem_cache_t
*cachep
)
1938 struct kmem_list3
*l3
;
1941 smp_call_function_all_cpus(do_drain
, cachep
);
1943 spin_lock_irq(&cachep
->spinlock
);
1944 for_each_online_node(node
) {
1945 l3
= cachep
->nodelists
[node
];
1947 spin_lock(&l3
->list_lock
);
1948 drain_array_locked(cachep
, l3
->shared
, 1, node
);
1949 spin_unlock(&l3
->list_lock
);
1951 drain_alien_cache(cachep
, l3
);
1954 spin_unlock_irq(&cachep
->spinlock
);
1957 static int __node_shrink(kmem_cache_t
*cachep
, int node
)
1960 struct kmem_list3
*l3
= cachep
->nodelists
[node
];
1964 struct list_head
*p
;
1966 p
= l3
->slabs_free
.prev
;
1967 if (p
== &l3
->slabs_free
)
1970 slabp
= list_entry(l3
->slabs_free
.prev
, struct slab
, list
);
1975 list_del(&slabp
->list
);
1977 l3
->free_objects
-= cachep
->num
;
1978 spin_unlock_irq(&l3
->list_lock
);
1979 slab_destroy(cachep
, slabp
);
1980 spin_lock_irq(&l3
->list_lock
);
1982 ret
= !list_empty(&l3
->slabs_full
) || !list_empty(&l3
->slabs_partial
);
1986 static int __cache_shrink(kmem_cache_t
*cachep
)
1989 struct kmem_list3
*l3
;
1991 drain_cpu_caches(cachep
);
1994 for_each_online_node(i
) {
1995 l3
= cachep
->nodelists
[i
];
1997 spin_lock_irq(&l3
->list_lock
);
1998 ret
+= __node_shrink(cachep
, i
);
1999 spin_unlock_irq(&l3
->list_lock
);
2002 return (ret
? 1 : 0);
2006 * kmem_cache_shrink - Shrink a cache.
2007 * @cachep: The cache to shrink.
2009 * Releases as many slabs as possible for a cache.
2010 * To help debugging, a zero exit status indicates all slabs were released.
2012 int kmem_cache_shrink(kmem_cache_t
*cachep
)
2014 if (!cachep
|| in_interrupt())
2017 return __cache_shrink(cachep
);
2019 EXPORT_SYMBOL(kmem_cache_shrink
);
2022 * kmem_cache_destroy - delete a cache
2023 * @cachep: the cache to destroy
2025 * Remove a kmem_cache_t object from the slab cache.
2026 * Returns 0 on success.
2028 * It is expected this function will be called by a module when it is
2029 * unloaded. This will remove the cache completely, and avoid a duplicate
2030 * cache being allocated each time a module is loaded and unloaded, if the
2031 * module doesn't have persistent in-kernel storage across loads and unloads.
2033 * The cache must be empty before calling this function.
2035 * The caller must guarantee that noone will allocate memory from the cache
2036 * during the kmem_cache_destroy().
2038 int kmem_cache_destroy(kmem_cache_t
*cachep
)
2041 struct kmem_list3
*l3
;
2043 if (!cachep
|| in_interrupt())
2046 /* Don't let CPUs to come and go */
2049 /* Find the cache in the chain of caches. */
2050 mutex_lock(&cache_chain_mutex
);
2052 * the chain is never empty, cache_cache is never destroyed
2054 list_del(&cachep
->next
);
2055 mutex_unlock(&cache_chain_mutex
);
2057 if (__cache_shrink(cachep
)) {
2058 slab_error(cachep
, "Can't free all objects");
2059 mutex_lock(&cache_chain_mutex
);
2060 list_add(&cachep
->next
, &cache_chain
);
2061 mutex_unlock(&cache_chain_mutex
);
2062 unlock_cpu_hotplug();
2066 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
2069 for_each_online_cpu(i
)
2070 kfree(cachep
->array
[i
]);
2072 /* NUMA: free the list3 structures */
2073 for_each_online_node(i
) {
2074 if ((l3
= cachep
->nodelists
[i
])) {
2076 free_alien_cache(l3
->alien
);
2080 kmem_cache_free(&cache_cache
, cachep
);
2082 unlock_cpu_hotplug();
2086 EXPORT_SYMBOL(kmem_cache_destroy
);
2088 /* Get the memory for a slab management obj. */
2089 static struct slab
*alloc_slabmgmt(kmem_cache_t
*cachep
, void *objp
,
2090 int colour_off
, gfp_t local_flags
)
2094 if (OFF_SLAB(cachep
)) {
2095 /* Slab management obj is off-slab. */
2096 slabp
= kmem_cache_alloc(cachep
->slabp_cache
, local_flags
);
2100 slabp
= objp
+ colour_off
;
2101 colour_off
+= cachep
->slab_size
;
2104 slabp
->colouroff
= colour_off
;
2105 slabp
->s_mem
= objp
+ colour_off
;
2110 static inline kmem_bufctl_t
*slab_bufctl(struct slab
*slabp
)
2112 return (kmem_bufctl_t
*) (slabp
+ 1);
2115 static void cache_init_objs(kmem_cache_t
*cachep
,
2116 struct slab
*slabp
, unsigned long ctor_flags
)
2120 for (i
= 0; i
< cachep
->num
; i
++) {
2121 void *objp
= slabp
->s_mem
+ cachep
->objsize
* i
;
2123 /* need to poison the objs? */
2124 if (cachep
->flags
& SLAB_POISON
)
2125 poison_obj(cachep
, objp
, POISON_FREE
);
2126 if (cachep
->flags
& SLAB_STORE_USER
)
2127 *dbg_userword(cachep
, objp
) = NULL
;
2129 if (cachep
->flags
& SLAB_RED_ZONE
) {
2130 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2131 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2134 * Constructors are not allowed to allocate memory from
2135 * the same cache which they are a constructor for.
2136 * Otherwise, deadlock. They must also be threaded.
2138 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2139 cachep
->ctor(objp
+ obj_dbghead(cachep
), cachep
,
2142 if (cachep
->flags
& SLAB_RED_ZONE
) {
2143 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2144 slab_error(cachep
, "constructor overwrote the"
2145 " end of an object");
2146 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2147 slab_error(cachep
, "constructor overwrote the"
2148 " start of an object");
2150 if ((cachep
->objsize
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
)
2151 && cachep
->flags
& SLAB_POISON
)
2152 kernel_map_pages(virt_to_page(objp
),
2153 cachep
->objsize
/ PAGE_SIZE
, 0);
2156 cachep
->ctor(objp
, cachep
, ctor_flags
);
2158 slab_bufctl(slabp
)[i
] = i
+ 1;
2160 slab_bufctl(slabp
)[i
- 1] = BUFCTL_END
;
2164 static void kmem_flagcheck(kmem_cache_t
*cachep
, gfp_t flags
)
2166 if (flags
& SLAB_DMA
) {
2167 if (!(cachep
->gfpflags
& GFP_DMA
))
2170 if (cachep
->gfpflags
& GFP_DMA
)
2175 static void set_slab_attr(kmem_cache_t
*cachep
, struct slab
*slabp
, void *objp
)
2180 /* Nasty!!!!!! I hope this is OK. */
2181 i
= 1 << cachep
->gfporder
;
2182 page
= virt_to_page(objp
);
2184 page_set_cache(page
, cachep
);
2185 page_set_slab(page
, slabp
);
2191 * Grow (by 1) the number of slabs within a cache. This is called by
2192 * kmem_cache_alloc() when there are no active objs left in a cache.
2194 static int cache_grow(kmem_cache_t
*cachep
, gfp_t flags
, int nodeid
)
2200 unsigned long ctor_flags
;
2201 struct kmem_list3
*l3
;
2203 /* Be lazy and only check for valid flags here,
2204 * keeping it out of the critical path in kmem_cache_alloc().
2206 if (flags
& ~(SLAB_DMA
| SLAB_LEVEL_MASK
| SLAB_NO_GROW
))
2208 if (flags
& SLAB_NO_GROW
)
2211 ctor_flags
= SLAB_CTOR_CONSTRUCTOR
;
2212 local_flags
= (flags
& SLAB_LEVEL_MASK
);
2213 if (!(local_flags
& __GFP_WAIT
))
2215 * Not allowed to sleep. Need to tell a constructor about
2216 * this - it might need to know...
2218 ctor_flags
|= SLAB_CTOR_ATOMIC
;
2220 /* About to mess with non-constant members - lock. */
2222 spin_lock(&cachep
->spinlock
);
2224 /* Get colour for the slab, and cal the next value. */
2225 offset
= cachep
->colour_next
;
2226 cachep
->colour_next
++;
2227 if (cachep
->colour_next
>= cachep
->colour
)
2228 cachep
->colour_next
= 0;
2229 offset
*= cachep
->colour_off
;
2231 spin_unlock(&cachep
->spinlock
);
2234 if (local_flags
& __GFP_WAIT
)
2238 * The test for missing atomic flag is performed here, rather than
2239 * the more obvious place, simply to reduce the critical path length
2240 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2241 * will eventually be caught here (where it matters).
2243 kmem_flagcheck(cachep
, flags
);
2245 /* Get mem for the objs.
2246 * Attempt to allocate a physical page from 'nodeid',
2248 if (!(objp
= kmem_getpages(cachep
, flags
, nodeid
)))
2251 /* Get slab management. */
2252 if (!(slabp
= alloc_slabmgmt(cachep
, objp
, offset
, local_flags
)))
2255 slabp
->nodeid
= nodeid
;
2256 set_slab_attr(cachep
, slabp
, objp
);
2258 cache_init_objs(cachep
, slabp
, ctor_flags
);
2260 if (local_flags
& __GFP_WAIT
)
2261 local_irq_disable();
2263 l3
= cachep
->nodelists
[nodeid
];
2264 spin_lock(&l3
->list_lock
);
2266 /* Make slab active. */
2267 list_add_tail(&slabp
->list
, &(l3
->slabs_free
));
2268 STATS_INC_GROWN(cachep
);
2269 l3
->free_objects
+= cachep
->num
;
2270 spin_unlock(&l3
->list_lock
);
2273 kmem_freepages(cachep
, objp
);
2275 if (local_flags
& __GFP_WAIT
)
2276 local_irq_disable();
2283 * Perform extra freeing checks:
2284 * - detect bad pointers.
2285 * - POISON/RED_ZONE checking
2286 * - destructor calls, for caches with POISON+dtor
2288 static void kfree_debugcheck(const void *objp
)
2292 if (!virt_addr_valid(objp
)) {
2293 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
2294 (unsigned long)objp
);
2297 page
= virt_to_page(objp
);
2298 if (!PageSlab(page
)) {
2299 printk(KERN_ERR
"kfree_debugcheck: bad ptr %lxh.\n",
2300 (unsigned long)objp
);
2305 static void *cache_free_debugcheck(kmem_cache_t
*cachep
, void *objp
,
2312 objp
-= obj_dbghead(cachep
);
2313 kfree_debugcheck(objp
);
2314 page
= virt_to_page(objp
);
2316 if (page_get_cache(page
) != cachep
) {
2318 "mismatch in kmem_cache_free: expected cache %p, got %p\n",
2319 page_get_cache(page
), cachep
);
2320 printk(KERN_ERR
"%p is %s.\n", cachep
, cachep
->name
);
2321 printk(KERN_ERR
"%p is %s.\n", page_get_cache(page
),
2322 page_get_cache(page
)->name
);
2325 slabp
= page_get_slab(page
);
2327 if (cachep
->flags
& SLAB_RED_ZONE
) {
2328 if (*dbg_redzone1(cachep
, objp
) != RED_ACTIVE
2329 || *dbg_redzone2(cachep
, objp
) != RED_ACTIVE
) {
2331 "double free, or memory outside"
2332 " object was overwritten");
2334 "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2335 objp
, *dbg_redzone1(cachep
, objp
),
2336 *dbg_redzone2(cachep
, objp
));
2338 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2339 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2341 if (cachep
->flags
& SLAB_STORE_USER
)
2342 *dbg_userword(cachep
, objp
) = caller
;
2344 objnr
= (objp
- slabp
->s_mem
) / cachep
->objsize
;
2346 BUG_ON(objnr
>= cachep
->num
);
2347 BUG_ON(objp
!= slabp
->s_mem
+ objnr
* cachep
->objsize
);
2349 if (cachep
->flags
& SLAB_DEBUG_INITIAL
) {
2350 /* Need to call the slab's constructor so the
2351 * caller can perform a verify of its state (debugging).
2352 * Called without the cache-lock held.
2354 cachep
->ctor(objp
+ obj_dbghead(cachep
),
2355 cachep
, SLAB_CTOR_CONSTRUCTOR
| SLAB_CTOR_VERIFY
);
2357 if (cachep
->flags
& SLAB_POISON
&& cachep
->dtor
) {
2358 /* we want to cache poison the object,
2359 * call the destruction callback
2361 cachep
->dtor(objp
+ obj_dbghead(cachep
), cachep
, 0);
2363 if (cachep
->flags
& SLAB_POISON
) {
2364 #ifdef CONFIG_DEBUG_PAGEALLOC
2365 if ((cachep
->objsize
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
)) {
2366 store_stackinfo(cachep
, objp
, (unsigned long)caller
);
2367 kernel_map_pages(virt_to_page(objp
),
2368 cachep
->objsize
/ PAGE_SIZE
, 0);
2370 poison_obj(cachep
, objp
, POISON_FREE
);
2373 poison_obj(cachep
, objp
, POISON_FREE
);
2379 static void check_slabp(kmem_cache_t
*cachep
, struct slab
*slabp
)
2384 /* Check slab's freelist to see if this obj is there. */
2385 for (i
= slabp
->free
; i
!= BUFCTL_END
; i
= slab_bufctl(slabp
)[i
]) {
2387 if (entries
> cachep
->num
|| i
>= cachep
->num
)
2390 if (entries
!= cachep
->num
- slabp
->inuse
) {
2393 "slab: Internal list corruption detected in cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2394 cachep
->name
, cachep
->num
, slabp
, slabp
->inuse
);
2396 i
< sizeof(slabp
) + cachep
->num
* sizeof(kmem_bufctl_t
);
2399 printk("\n%03x:", i
);
2400 printk(" %02x", ((unsigned char *)slabp
)[i
]);
2407 #define kfree_debugcheck(x) do { } while(0)
2408 #define cache_free_debugcheck(x,objp,z) (objp)
2409 #define check_slabp(x,y) do { } while(0)
2412 static void *cache_alloc_refill(kmem_cache_t
*cachep
, gfp_t flags
)
2415 struct kmem_list3
*l3
;
2416 struct array_cache
*ac
;
2419 ac
= ac_data(cachep
);
2421 batchcount
= ac
->batchcount
;
2422 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2423 /* if there was little recent activity on this
2424 * cache, then perform only a partial refill.
2425 * Otherwise we could generate refill bouncing.
2427 batchcount
= BATCHREFILL_LIMIT
;
2429 l3
= cachep
->nodelists
[numa_node_id()];
2431 BUG_ON(ac
->avail
> 0 || !l3
);
2432 spin_lock(&l3
->list_lock
);
2435 struct array_cache
*shared_array
= l3
->shared
;
2436 if (shared_array
->avail
) {
2437 if (batchcount
> shared_array
->avail
)
2438 batchcount
= shared_array
->avail
;
2439 shared_array
->avail
-= batchcount
;
2440 ac
->avail
= batchcount
;
2442 &(shared_array
->entry
[shared_array
->avail
]),
2443 sizeof(void *) * batchcount
);
2444 shared_array
->touched
= 1;
2448 while (batchcount
> 0) {
2449 struct list_head
*entry
;
2451 /* Get slab alloc is to come from. */
2452 entry
= l3
->slabs_partial
.next
;
2453 if (entry
== &l3
->slabs_partial
) {
2454 l3
->free_touched
= 1;
2455 entry
= l3
->slabs_free
.next
;
2456 if (entry
== &l3
->slabs_free
)
2460 slabp
= list_entry(entry
, struct slab
, list
);
2461 check_slabp(cachep
, slabp
);
2462 check_spinlock_acquired(cachep
);
2463 while (slabp
->inuse
< cachep
->num
&& batchcount
--) {
2465 STATS_INC_ALLOCED(cachep
);
2466 STATS_INC_ACTIVE(cachep
);
2467 STATS_SET_HIGH(cachep
);
2469 /* get obj pointer */
2470 ac
->entry
[ac
->avail
++] = slabp
->s_mem
+
2471 slabp
->free
* cachep
->objsize
;
2474 next
= slab_bufctl(slabp
)[slabp
->free
];
2476 slab_bufctl(slabp
)[slabp
->free
] = BUFCTL_FREE
;
2477 WARN_ON(numa_node_id() != slabp
->nodeid
);
2481 check_slabp(cachep
, slabp
);
2483 /* move slabp to correct slabp list: */
2484 list_del(&slabp
->list
);
2485 if (slabp
->free
== BUFCTL_END
)
2486 list_add(&slabp
->list
, &l3
->slabs_full
);
2488 list_add(&slabp
->list
, &l3
->slabs_partial
);
2492 l3
->free_objects
-= ac
->avail
;
2494 spin_unlock(&l3
->list_lock
);
2496 if (unlikely(!ac
->avail
)) {
2498 x
= cache_grow(cachep
, flags
, numa_node_id());
2500 // cache_grow can reenable interrupts, then ac could change.
2501 ac
= ac_data(cachep
);
2502 if (!x
&& ac
->avail
== 0) // no objects in sight? abort
2505 if (!ac
->avail
) // objects refilled by interrupt?
2509 return ac
->entry
[--ac
->avail
];
2513 cache_alloc_debugcheck_before(kmem_cache_t
*cachep
, gfp_t flags
)
2515 might_sleep_if(flags
& __GFP_WAIT
);
2517 kmem_flagcheck(cachep
, flags
);
2522 static void *cache_alloc_debugcheck_after(kmem_cache_t
*cachep
, gfp_t flags
,
2523 void *objp
, void *caller
)
2527 if (cachep
->flags
& SLAB_POISON
) {
2528 #ifdef CONFIG_DEBUG_PAGEALLOC
2529 if ((cachep
->objsize
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
2530 kernel_map_pages(virt_to_page(objp
),
2531 cachep
->objsize
/ PAGE_SIZE
, 1);
2533 check_poison_obj(cachep
, objp
);
2535 check_poison_obj(cachep
, objp
);
2537 poison_obj(cachep
, objp
, POISON_INUSE
);
2539 if (cachep
->flags
& SLAB_STORE_USER
)
2540 *dbg_userword(cachep
, objp
) = caller
;
2542 if (cachep
->flags
& SLAB_RED_ZONE
) {
2543 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
2544 || *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
2546 "double free, or memory outside"
2547 " object was overwritten");
2549 "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2550 objp
, *dbg_redzone1(cachep
, objp
),
2551 *dbg_redzone2(cachep
, objp
));
2553 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
2554 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
2556 objp
+= obj_dbghead(cachep
);
2557 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
) {
2558 unsigned long ctor_flags
= SLAB_CTOR_CONSTRUCTOR
;
2560 if (!(flags
& __GFP_WAIT
))
2561 ctor_flags
|= SLAB_CTOR_ATOMIC
;
2563 cachep
->ctor(objp
, cachep
, ctor_flags
);
2568 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2571 static inline void *____cache_alloc(kmem_cache_t
*cachep
, gfp_t flags
)
2574 struct array_cache
*ac
;
2577 if (unlikely(current
->mempolicy
&& !in_interrupt())) {
2578 int nid
= slab_node(current
->mempolicy
);
2580 if (nid
!= numa_node_id())
2581 return __cache_alloc_node(cachep
, flags
, nid
);
2586 ac
= ac_data(cachep
);
2587 if (likely(ac
->avail
)) {
2588 STATS_INC_ALLOCHIT(cachep
);
2590 objp
= ac
->entry
[--ac
->avail
];
2592 STATS_INC_ALLOCMISS(cachep
);
2593 objp
= cache_alloc_refill(cachep
, flags
);
2598 static inline void *__cache_alloc(kmem_cache_t
*cachep
, gfp_t flags
)
2600 unsigned long save_flags
;
2603 cache_alloc_debugcheck_before(cachep
, flags
);
2605 local_irq_save(save_flags
);
2606 objp
= ____cache_alloc(cachep
, flags
);
2607 local_irq_restore(save_flags
);
2608 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
,
2609 __builtin_return_address(0));
2616 * A interface to enable slab creation on nodeid
2618 static void *__cache_alloc_node(kmem_cache_t
*cachep
, gfp_t flags
, int nodeid
)
2620 struct list_head
*entry
;
2622 struct kmem_list3
*l3
;
2627 l3
= cachep
->nodelists
[nodeid
];
2631 spin_lock(&l3
->list_lock
);
2632 entry
= l3
->slabs_partial
.next
;
2633 if (entry
== &l3
->slabs_partial
) {
2634 l3
->free_touched
= 1;
2635 entry
= l3
->slabs_free
.next
;
2636 if (entry
== &l3
->slabs_free
)
2640 slabp
= list_entry(entry
, struct slab
, list
);
2641 check_spinlock_acquired_node(cachep
, nodeid
);
2642 check_slabp(cachep
, slabp
);
2644 STATS_INC_NODEALLOCS(cachep
);
2645 STATS_INC_ACTIVE(cachep
);
2646 STATS_SET_HIGH(cachep
);
2648 BUG_ON(slabp
->inuse
== cachep
->num
);
2650 /* get obj pointer */
2651 obj
= slabp
->s_mem
+ slabp
->free
* cachep
->objsize
;
2653 next
= slab_bufctl(slabp
)[slabp
->free
];
2655 slab_bufctl(slabp
)[slabp
->free
] = BUFCTL_FREE
;
2658 check_slabp(cachep
, slabp
);
2660 /* move slabp to correct slabp list: */
2661 list_del(&slabp
->list
);
2663 if (slabp
->free
== BUFCTL_END
) {
2664 list_add(&slabp
->list
, &l3
->slabs_full
);
2666 list_add(&slabp
->list
, &l3
->slabs_partial
);
2669 spin_unlock(&l3
->list_lock
);
2673 spin_unlock(&l3
->list_lock
);
2674 x
= cache_grow(cachep
, flags
, nodeid
);
2686 * Caller needs to acquire correct kmem_list's list_lock
2688 static void free_block(kmem_cache_t
*cachep
, void **objpp
, int nr_objects
,
2692 struct kmem_list3
*l3
;
2694 for (i
= 0; i
< nr_objects
; i
++) {
2695 void *objp
= objpp
[i
];
2699 slabp
= page_get_slab(virt_to_page(objp
));
2700 l3
= cachep
->nodelists
[node
];
2701 list_del(&slabp
->list
);
2702 objnr
= (objp
- slabp
->s_mem
) / cachep
->objsize
;
2703 check_spinlock_acquired_node(cachep
, node
);
2704 check_slabp(cachep
, slabp
);
2707 /* Verify that the slab belongs to the intended node */
2708 WARN_ON(slabp
->nodeid
!= node
);
2710 if (slab_bufctl(slabp
)[objnr
] != BUFCTL_FREE
) {
2711 printk(KERN_ERR
"slab: double free detected in cache "
2712 "'%s', objp %p\n", cachep
->name
, objp
);
2716 slab_bufctl(slabp
)[objnr
] = slabp
->free
;
2717 slabp
->free
= objnr
;
2718 STATS_DEC_ACTIVE(cachep
);
2721 check_slabp(cachep
, slabp
);
2723 /* fixup slab chains */
2724 if (slabp
->inuse
== 0) {
2725 if (l3
->free_objects
> l3
->free_limit
) {
2726 l3
->free_objects
-= cachep
->num
;
2727 slab_destroy(cachep
, slabp
);
2729 list_add(&slabp
->list
, &l3
->slabs_free
);
2732 /* Unconditionally move a slab to the end of the
2733 * partial list on free - maximum time for the
2734 * other objects to be freed, too.
2736 list_add_tail(&slabp
->list
, &l3
->slabs_partial
);
2741 static void cache_flusharray(kmem_cache_t
*cachep
, struct array_cache
*ac
)
2744 struct kmem_list3
*l3
;
2745 int node
= numa_node_id();
2747 batchcount
= ac
->batchcount
;
2749 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
2752 l3
= cachep
->nodelists
[node
];
2753 spin_lock(&l3
->list_lock
);
2755 struct array_cache
*shared_array
= l3
->shared
;
2756 int max
= shared_array
->limit
- shared_array
->avail
;
2758 if (batchcount
> max
)
2760 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
2761 ac
->entry
, sizeof(void *) * batchcount
);
2762 shared_array
->avail
+= batchcount
;
2767 free_block(cachep
, ac
->entry
, batchcount
, node
);
2772 struct list_head
*p
;
2774 p
= l3
->slabs_free
.next
;
2775 while (p
!= &(l3
->slabs_free
)) {
2778 slabp
= list_entry(p
, struct slab
, list
);
2779 BUG_ON(slabp
->inuse
);
2784 STATS_SET_FREEABLE(cachep
, i
);
2787 spin_unlock(&l3
->list_lock
);
2788 ac
->avail
-= batchcount
;
2789 memmove(ac
->entry
, &(ac
->entry
[batchcount
]),
2790 sizeof(void *) * ac
->avail
);
2795 * Release an obj back to its cache. If the obj has a constructed
2796 * state, it must be in this state _before_ it is released.
2798 * Called with disabled ints.
2800 static inline void __cache_free(kmem_cache_t
*cachep
, void *objp
)
2802 struct array_cache
*ac
= ac_data(cachep
);
2805 objp
= cache_free_debugcheck(cachep
, objp
, __builtin_return_address(0));
2807 /* Make sure we are not freeing a object from another
2808 * node to the array cache on this cpu.
2813 slabp
= page_get_slab(virt_to_page(objp
));
2814 if (unlikely(slabp
->nodeid
!= numa_node_id())) {
2815 struct array_cache
*alien
= NULL
;
2816 int nodeid
= slabp
->nodeid
;
2817 struct kmem_list3
*l3
=
2818 cachep
->nodelists
[numa_node_id()];
2820 STATS_INC_NODEFREES(cachep
);
2821 if (l3
->alien
&& l3
->alien
[nodeid
]) {
2822 alien
= l3
->alien
[nodeid
];
2823 spin_lock(&alien
->lock
);
2824 if (unlikely(alien
->avail
== alien
->limit
))
2825 __drain_alien_cache(cachep
,
2827 alien
->entry
[alien
->avail
++] = objp
;
2828 spin_unlock(&alien
->lock
);
2830 spin_lock(&(cachep
->nodelists
[nodeid
])->
2832 free_block(cachep
, &objp
, 1, nodeid
);
2833 spin_unlock(&(cachep
->nodelists
[nodeid
])->
2840 if (likely(ac
->avail
< ac
->limit
)) {
2841 STATS_INC_FREEHIT(cachep
);
2842 ac
->entry
[ac
->avail
++] = objp
;
2845 STATS_INC_FREEMISS(cachep
);
2846 cache_flusharray(cachep
, ac
);
2847 ac
->entry
[ac
->avail
++] = objp
;
2852 * kmem_cache_alloc - Allocate an object
2853 * @cachep: The cache to allocate from.
2854 * @flags: See kmalloc().
2856 * Allocate an object from this cache. The flags are only relevant
2857 * if the cache has no available objects.
2859 void *kmem_cache_alloc(kmem_cache_t
*cachep
, gfp_t flags
)
2861 return __cache_alloc(cachep
, flags
);
2863 EXPORT_SYMBOL(kmem_cache_alloc
);
2866 * kmem_ptr_validate - check if an untrusted pointer might
2868 * @cachep: the cache we're checking against
2869 * @ptr: pointer to validate
2871 * This verifies that the untrusted pointer looks sane:
2872 * it is _not_ a guarantee that the pointer is actually
2873 * part of the slab cache in question, but it at least
2874 * validates that the pointer can be dereferenced and
2875 * looks half-way sane.
2877 * Currently only used for dentry validation.
2879 int fastcall
kmem_ptr_validate(kmem_cache_t
*cachep
, void *ptr
)
2881 unsigned long addr
= (unsigned long)ptr
;
2882 unsigned long min_addr
= PAGE_OFFSET
;
2883 unsigned long align_mask
= BYTES_PER_WORD
- 1;
2884 unsigned long size
= cachep
->objsize
;
2887 if (unlikely(addr
< min_addr
))
2889 if (unlikely(addr
> (unsigned long)high_memory
- size
))
2891 if (unlikely(addr
& align_mask
))
2893 if (unlikely(!kern_addr_valid(addr
)))
2895 if (unlikely(!kern_addr_valid(addr
+ size
- 1)))
2897 page
= virt_to_page(ptr
);
2898 if (unlikely(!PageSlab(page
)))
2900 if (unlikely(page_get_cache(page
) != cachep
))
2909 * kmem_cache_alloc_node - Allocate an object on the specified node
2910 * @cachep: The cache to allocate from.
2911 * @flags: See kmalloc().
2912 * @nodeid: node number of the target node.
2914 * Identical to kmem_cache_alloc, except that this function is slow
2915 * and can sleep. And it will allocate memory on the given node, which
2916 * can improve the performance for cpu bound structures.
2917 * New and improved: it will now make sure that the object gets
2918 * put on the correct node list so that there is no false sharing.
2920 void *kmem_cache_alloc_node(kmem_cache_t
*cachep
, gfp_t flags
, int nodeid
)
2922 unsigned long save_flags
;
2926 return __cache_alloc(cachep
, flags
);
2928 if (unlikely(!cachep
->nodelists
[nodeid
])) {
2929 /* Fall back to __cache_alloc if we run into trouble */
2931 "slab: not allocating in inactive node %d for cache %s\n",
2932 nodeid
, cachep
->name
);
2933 return __cache_alloc(cachep
, flags
);
2936 cache_alloc_debugcheck_before(cachep
, flags
);
2937 local_irq_save(save_flags
);
2938 if (nodeid
== numa_node_id())
2939 ptr
= ____cache_alloc(cachep
, flags
);
2941 ptr
= __cache_alloc_node(cachep
, flags
, nodeid
);
2942 local_irq_restore(save_flags
);
2944 cache_alloc_debugcheck_after(cachep
, flags
, ptr
,
2945 __builtin_return_address(0));
2949 EXPORT_SYMBOL(kmem_cache_alloc_node
);
2951 void *kmalloc_node(size_t size
, gfp_t flags
, int node
)
2953 kmem_cache_t
*cachep
;
2955 cachep
= kmem_find_general_cachep(size
, flags
);
2956 if (unlikely(cachep
== NULL
))
2958 return kmem_cache_alloc_node(cachep
, flags
, node
);
2960 EXPORT_SYMBOL(kmalloc_node
);
2964 * kmalloc - allocate memory
2965 * @size: how many bytes of memory are required.
2966 * @flags: the type of memory to allocate.
2968 * kmalloc is the normal method of allocating memory
2971 * The @flags argument may be one of:
2973 * %GFP_USER - Allocate memory on behalf of user. May sleep.
2975 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
2977 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
2979 * Additionally, the %GFP_DMA flag may be set to indicate the memory
2980 * must be suitable for DMA. This can mean different things on different
2981 * platforms. For example, on i386, it means that the memory must come
2982 * from the first 16MB.
2984 void *__kmalloc(size_t size
, gfp_t flags
)
2986 kmem_cache_t
*cachep
;
2988 /* If you want to save a few bytes .text space: replace
2990 * Then kmalloc uses the uninlined functions instead of the inline
2993 cachep
= __find_general_cachep(size
, flags
);
2994 if (unlikely(cachep
== NULL
))
2996 return __cache_alloc(cachep
, flags
);
2998 EXPORT_SYMBOL(__kmalloc
);
3002 * __alloc_percpu - allocate one copy of the object for every present
3003 * cpu in the system, zeroing them.
3004 * Objects should be dereferenced using the per_cpu_ptr macro only.
3006 * @size: how many bytes of memory are required.
3008 void *__alloc_percpu(size_t size
)
3011 struct percpu_data
*pdata
= kmalloc(sizeof(*pdata
), GFP_KERNEL
);
3017 * Cannot use for_each_online_cpu since a cpu may come online
3018 * and we have no way of figuring out how to fix the array
3019 * that we have allocated then....
3022 int node
= cpu_to_node(i
);
3024 if (node_online(node
))
3025 pdata
->ptrs
[i
] = kmalloc_node(size
, GFP_KERNEL
, node
);
3027 pdata
->ptrs
[i
] = kmalloc(size
, GFP_KERNEL
);
3029 if (!pdata
->ptrs
[i
])
3031 memset(pdata
->ptrs
[i
], 0, size
);
3034 /* Catch derefs w/o wrappers */
3035 return (void *)(~(unsigned long)pdata
);
3039 if (!cpu_possible(i
))
3041 kfree(pdata
->ptrs
[i
]);
3046 EXPORT_SYMBOL(__alloc_percpu
);
3050 * kmem_cache_free - Deallocate an object
3051 * @cachep: The cache the allocation was from.
3052 * @objp: The previously allocated object.
3054 * Free an object which was previously allocated from this
3057 void kmem_cache_free(kmem_cache_t
*cachep
, void *objp
)
3059 unsigned long flags
;
3061 local_irq_save(flags
);
3062 __cache_free(cachep
, objp
);
3063 local_irq_restore(flags
);
3065 EXPORT_SYMBOL(kmem_cache_free
);
3068 * kfree - free previously allocated memory
3069 * @objp: pointer returned by kmalloc.
3071 * If @objp is NULL, no operation is performed.
3073 * Don't free memory not originally allocated by kmalloc()
3074 * or you will run into trouble.
3076 void kfree(const void *objp
)
3079 unsigned long flags
;
3081 if (unlikely(!objp
))
3083 local_irq_save(flags
);
3084 kfree_debugcheck(objp
);
3085 c
= page_get_cache(virt_to_page(objp
));
3086 mutex_debug_check_no_locks_freed(objp
, obj_reallen(c
));
3087 __cache_free(c
, (void *)objp
);
3088 local_irq_restore(flags
);
3090 EXPORT_SYMBOL(kfree
);
3094 * free_percpu - free previously allocated percpu memory
3095 * @objp: pointer returned by alloc_percpu.
3097 * Don't free memory not originally allocated by alloc_percpu()
3098 * The complemented objp is to check for that.
3100 void free_percpu(const void *objp
)
3103 struct percpu_data
*p
= (struct percpu_data
*)(~(unsigned long)objp
);
3106 * We allocate for all cpus so we cannot use for online cpu here.
3112 EXPORT_SYMBOL(free_percpu
);
3115 unsigned int kmem_cache_size(kmem_cache_t
*cachep
)
3117 return obj_reallen(cachep
);
3119 EXPORT_SYMBOL(kmem_cache_size
);
3121 const char *kmem_cache_name(kmem_cache_t
*cachep
)
3123 return cachep
->name
;
3125 EXPORT_SYMBOL_GPL(kmem_cache_name
);
3128 * This initializes kmem_list3 for all nodes.
3130 static int alloc_kmemlist(kmem_cache_t
*cachep
)
3133 struct kmem_list3
*l3
;
3136 for_each_online_node(node
) {
3137 struct array_cache
*nc
= NULL
, *new;
3138 struct array_cache
**new_alien
= NULL
;
3140 if (!(new_alien
= alloc_alien_cache(node
, cachep
->limit
)))
3143 if (!(new = alloc_arraycache(node
, (cachep
->shared
*
3144 cachep
->batchcount
),
3147 if ((l3
= cachep
->nodelists
[node
])) {
3149 spin_lock_irq(&l3
->list_lock
);
3151 if ((nc
= cachep
->nodelists
[node
]->shared
))
3152 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
3155 if (!cachep
->nodelists
[node
]->alien
) {
3156 l3
->alien
= new_alien
;
3159 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3160 cachep
->batchcount
+ cachep
->num
;
3161 spin_unlock_irq(&l3
->list_lock
);
3163 free_alien_cache(new_alien
);
3166 if (!(l3
= kmalloc_node(sizeof(struct kmem_list3
),
3170 kmem_list3_init(l3
);
3171 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
3172 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
3174 l3
->alien
= new_alien
;
3175 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3176 cachep
->batchcount
+ cachep
->num
;
3177 cachep
->nodelists
[node
] = l3
;
3185 struct ccupdate_struct
{
3186 kmem_cache_t
*cachep
;
3187 struct array_cache
*new[NR_CPUS
];
3190 static void do_ccupdate_local(void *info
)
3192 struct ccupdate_struct
*new = (struct ccupdate_struct
*)info
;
3193 struct array_cache
*old
;
3196 old
= ac_data(new->cachep
);
3198 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
3199 new->new[smp_processor_id()] = old
;
3202 static int do_tune_cpucache(kmem_cache_t
*cachep
, int limit
, int batchcount
,
3205 struct ccupdate_struct
new;
3208 memset(&new.new, 0, sizeof(new.new));
3209 for_each_online_cpu(i
) {
3211 alloc_arraycache(cpu_to_node(i
), limit
, batchcount
);
3213 for (i
--; i
>= 0; i
--)
3218 new.cachep
= cachep
;
3220 smp_call_function_all_cpus(do_ccupdate_local
, (void *)&new);
3223 spin_lock_irq(&cachep
->spinlock
);
3224 cachep
->batchcount
= batchcount
;
3225 cachep
->limit
= limit
;
3226 cachep
->shared
= shared
;
3227 spin_unlock_irq(&cachep
->spinlock
);
3229 for_each_online_cpu(i
) {
3230 struct array_cache
*ccold
= new.new[i
];
3233 spin_lock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3234 free_block(cachep
, ccold
->entry
, ccold
->avail
, cpu_to_node(i
));
3235 spin_unlock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3239 err
= alloc_kmemlist(cachep
);
3241 printk(KERN_ERR
"alloc_kmemlist failed for %s, error %d.\n",
3242 cachep
->name
, -err
);
3248 static void enable_cpucache(kmem_cache_t
*cachep
)
3253 /* The head array serves three purposes:
3254 * - create a LIFO ordering, i.e. return objects that are cache-warm
3255 * - reduce the number of spinlock operations.
3256 * - reduce the number of linked list operations on the slab and
3257 * bufctl chains: array operations are cheaper.
3258 * The numbers are guessed, we should auto-tune as described by
3261 if (cachep
->objsize
> 131072)
3263 else if (cachep
->objsize
> PAGE_SIZE
)
3265 else if (cachep
->objsize
> 1024)
3267 else if (cachep
->objsize
> 256)
3272 /* Cpu bound tasks (e.g. network routing) can exhibit cpu bound
3273 * allocation behaviour: Most allocs on one cpu, most free operations
3274 * on another cpu. For these cases, an efficient object passing between
3275 * cpus is necessary. This is provided by a shared array. The array
3276 * replaces Bonwick's magazine layer.
3277 * On uniprocessor, it's functionally equivalent (but less efficient)
3278 * to a larger limit. Thus disabled by default.
3282 if (cachep
->objsize
<= PAGE_SIZE
)
3287 /* With debugging enabled, large batchcount lead to excessively
3288 * long periods with disabled local interrupts. Limit the
3294 err
= do_tune_cpucache(cachep
, limit
, (limit
+ 1) / 2, shared
);
3296 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
3297 cachep
->name
, -err
);
3300 static void drain_array_locked(kmem_cache_t
*cachep
, struct array_cache
*ac
,
3301 int force
, int node
)
3305 check_spinlock_acquired_node(cachep
, node
);
3306 if (ac
->touched
&& !force
) {
3308 } else if (ac
->avail
) {
3309 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
3310 if (tofree
> ac
->avail
) {
3311 tofree
= (ac
->avail
+ 1) / 2;
3313 free_block(cachep
, ac
->entry
, tofree
, node
);
3314 ac
->avail
-= tofree
;
3315 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
3316 sizeof(void *) * ac
->avail
);
3321 * cache_reap - Reclaim memory from caches.
3322 * @unused: unused parameter
3324 * Called from workqueue/eventd every few seconds.
3326 * - clear the per-cpu caches for this CPU.
3327 * - return freeable pages to the main free memory pool.
3329 * If we cannot acquire the cache chain mutex then just give up - we'll
3330 * try again on the next iteration.
3332 static void cache_reap(void *unused
)
3334 struct list_head
*walk
;
3335 struct kmem_list3
*l3
;
3337 if (!mutex_trylock(&cache_chain_mutex
)) {
3338 /* Give up. Setup the next iteration. */
3339 schedule_delayed_work(&__get_cpu_var(reap_work
),
3344 list_for_each(walk
, &cache_chain
) {
3345 kmem_cache_t
*searchp
;
3346 struct list_head
*p
;
3350 searchp
= list_entry(walk
, kmem_cache_t
, next
);
3352 if (searchp
->flags
& SLAB_NO_REAP
)
3357 l3
= searchp
->nodelists
[numa_node_id()];
3359 drain_alien_cache(searchp
, l3
);
3360 spin_lock_irq(&l3
->list_lock
);
3362 drain_array_locked(searchp
, ac_data(searchp
), 0,
3365 if (time_after(l3
->next_reap
, jiffies
))
3368 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
;
3371 drain_array_locked(searchp
, l3
->shared
, 0,
3374 if (l3
->free_touched
) {
3375 l3
->free_touched
= 0;
3380 (l3
->free_limit
+ 5 * searchp
->num
-
3381 1) / (5 * searchp
->num
);
3383 p
= l3
->slabs_free
.next
;
3384 if (p
== &(l3
->slabs_free
))
3387 slabp
= list_entry(p
, struct slab
, list
);
3388 BUG_ON(slabp
->inuse
);
3389 list_del(&slabp
->list
);
3390 STATS_INC_REAPED(searchp
);
3392 /* Safe to drop the lock. The slab is no longer
3393 * linked to the cache.
3394 * searchp cannot disappear, we hold
3397 l3
->free_objects
-= searchp
->num
;
3398 spin_unlock_irq(&l3
->list_lock
);
3399 slab_destroy(searchp
, slabp
);
3400 spin_lock_irq(&l3
->list_lock
);
3401 } while (--tofree
> 0);
3403 spin_unlock_irq(&l3
->list_lock
);
3408 mutex_unlock(&cache_chain_mutex
);
3409 drain_remote_pages();
3410 /* Setup the next iteration */
3411 schedule_delayed_work(&__get_cpu_var(reap_work
), REAPTIMEOUT_CPUC
);
3414 #ifdef CONFIG_PROC_FS
3416 static void print_slabinfo_header(struct seq_file
*m
)
3419 * Output format version, so at least we can change it
3420 * without _too_ many complaints.
3423 seq_puts(m
, "slabinfo - version: 2.1 (statistics)\n");
3425 seq_puts(m
, "slabinfo - version: 2.1\n");
3427 seq_puts(m
, "# name <active_objs> <num_objs> <objsize> "
3428 "<objperslab> <pagesperslab>");
3429 seq_puts(m
, " : tunables <limit> <batchcount> <sharedfactor>");
3430 seq_puts(m
, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3432 seq_puts(m
, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3433 "<error> <maxfreeable> <nodeallocs> <remotefrees>");
3434 seq_puts(m
, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3439 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
3442 struct list_head
*p
;
3444 mutex_lock(&cache_chain_mutex
);
3446 print_slabinfo_header(m
);
3447 p
= cache_chain
.next
;
3450 if (p
== &cache_chain
)
3453 return list_entry(p
, kmem_cache_t
, next
);
3456 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
3458 kmem_cache_t
*cachep
= p
;
3460 return cachep
->next
.next
== &cache_chain
? NULL
3461 : list_entry(cachep
->next
.next
, kmem_cache_t
, next
);
3464 static void s_stop(struct seq_file
*m
, void *p
)
3466 mutex_unlock(&cache_chain_mutex
);
3469 static int s_show(struct seq_file
*m
, void *p
)
3471 kmem_cache_t
*cachep
= p
;
3472 struct list_head
*q
;
3474 unsigned long active_objs
;
3475 unsigned long num_objs
;
3476 unsigned long active_slabs
= 0;
3477 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
3481 struct kmem_list3
*l3
;
3484 spin_lock_irq(&cachep
->spinlock
);
3487 for_each_online_node(node
) {
3488 l3
= cachep
->nodelists
[node
];
3492 spin_lock(&l3
->list_lock
);
3494 list_for_each(q
, &l3
->slabs_full
) {
3495 slabp
= list_entry(q
, struct slab
, list
);
3496 if (slabp
->inuse
!= cachep
->num
&& !error
)
3497 error
= "slabs_full accounting error";
3498 active_objs
+= cachep
->num
;
3501 list_for_each(q
, &l3
->slabs_partial
) {
3502 slabp
= list_entry(q
, struct slab
, list
);
3503 if (slabp
->inuse
== cachep
->num
&& !error
)
3504 error
= "slabs_partial inuse accounting error";
3505 if (!slabp
->inuse
&& !error
)
3506 error
= "slabs_partial/inuse accounting error";
3507 active_objs
+= slabp
->inuse
;
3510 list_for_each(q
, &l3
->slabs_free
) {
3511 slabp
= list_entry(q
, struct slab
, list
);
3512 if (slabp
->inuse
&& !error
)
3513 error
= "slabs_free/inuse accounting error";
3516 free_objects
+= l3
->free_objects
;
3517 shared_avail
+= l3
->shared
->avail
;
3519 spin_unlock(&l3
->list_lock
);
3521 num_slabs
+= active_slabs
;
3522 num_objs
= num_slabs
* cachep
->num
;
3523 if (num_objs
- active_objs
!= free_objects
&& !error
)
3524 error
= "free_objects accounting error";
3526 name
= cachep
->name
;
3528 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
3530 seq_printf(m
, "%-17s %6lu %6lu %6u %4u %4d",
3531 name
, active_objs
, num_objs
, cachep
->objsize
,
3532 cachep
->num
, (1 << cachep
->gfporder
));
3533 seq_printf(m
, " : tunables %4u %4u %4u",
3534 cachep
->limit
, cachep
->batchcount
, cachep
->shared
);
3535 seq_printf(m
, " : slabdata %6lu %6lu %6lu",
3536 active_slabs
, num_slabs
, shared_avail
);
3539 unsigned long high
= cachep
->high_mark
;
3540 unsigned long allocs
= cachep
->num_allocations
;
3541 unsigned long grown
= cachep
->grown
;
3542 unsigned long reaped
= cachep
->reaped
;
3543 unsigned long errors
= cachep
->errors
;
3544 unsigned long max_freeable
= cachep
->max_freeable
;
3545 unsigned long node_allocs
= cachep
->node_allocs
;
3546 unsigned long node_frees
= cachep
->node_frees
;
3548 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu \
3549 %4lu %4lu %4lu %4lu", allocs
, high
, grown
, reaped
, errors
, max_freeable
, node_allocs
, node_frees
);
3553 unsigned long allochit
= atomic_read(&cachep
->allochit
);
3554 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
3555 unsigned long freehit
= atomic_read(&cachep
->freehit
);
3556 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
3558 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
3559 allochit
, allocmiss
, freehit
, freemiss
);
3563 spin_unlock_irq(&cachep
->spinlock
);
3568 * slabinfo_op - iterator that generates /proc/slabinfo
3577 * num-pages-per-slab
3578 * + further values on SMP and with statistics enabled
3581 struct seq_operations slabinfo_op
= {
3588 #define MAX_SLABINFO_WRITE 128
3590 * slabinfo_write - Tuning for the slab allocator
3592 * @buffer: user buffer
3593 * @count: data length
3596 ssize_t
slabinfo_write(struct file
*file
, const char __user
* buffer
,
3597 size_t count
, loff_t
*ppos
)
3599 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
3600 int limit
, batchcount
, shared
, res
;
3601 struct list_head
*p
;
3603 if (count
> MAX_SLABINFO_WRITE
)
3605 if (copy_from_user(&kbuf
, buffer
, count
))
3607 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
3609 tmp
= strchr(kbuf
, ' ');
3614 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
3617 /* Find the cache in the chain of caches. */
3618 mutex_lock(&cache_chain_mutex
);
3620 list_for_each(p
, &cache_chain
) {
3621 kmem_cache_t
*cachep
= list_entry(p
, kmem_cache_t
, next
);
3623 if (!strcmp(cachep
->name
, kbuf
)) {
3626 batchcount
> limit
|| shared
< 0) {
3629 res
= do_tune_cpucache(cachep
, limit
,
3630 batchcount
, shared
);
3635 mutex_unlock(&cache_chain_mutex
);
3643 * ksize - get the actual amount of memory allocated for a given object
3644 * @objp: Pointer to the object
3646 * kmalloc may internally round up allocations and return more memory
3647 * than requested. ksize() can be used to determine the actual amount of
3648 * memory allocated. The caller may use this additional memory, even though
3649 * a smaller amount of memory was initially specified with the kmalloc call.
3650 * The caller must guarantee that objp points to a valid object previously
3651 * allocated with either kmalloc() or kmem_cache_alloc(). The object
3652 * must not be freed during the duration of the call.
3654 unsigned int ksize(const void *objp
)
3656 if (unlikely(objp
== NULL
))
3659 return obj_reallen(page_get_cache(virt_to_page(objp
)));