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 initializations 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/proc_fs.h>
99 #include <linux/seq_file.h>
100 #include <linux/notifier.h>
101 #include <linux/kallsyms.h>
102 #include <linux/cpu.h>
103 #include <linux/sysctl.h>
104 #include <linux/module.h>
105 #include <linux/rcupdate.h>
106 #include <linux/string.h>
107 #include <linux/uaccess.h>
108 #include <linux/nodemask.h>
109 #include <linux/kmemleak.h>
110 #include <linux/mempolicy.h>
111 #include <linux/mutex.h>
112 #include <linux/fault-inject.h>
113 #include <linux/rtmutex.h>
114 #include <linux/reciprocal_div.h>
115 #include <linux/debugobjects.h>
116 #include <linux/kmemcheck.h>
117 #include <linux/memory.h>
119 #include <asm/cacheflush.h>
120 #include <asm/tlbflush.h>
121 #include <asm/page.h>
124 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
125 * 0 for faster, smaller code (especially in the critical paths).
127 * STATS - 1 to collect stats for /proc/slabinfo.
128 * 0 for faster, smaller code (especially in the critical paths).
130 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
133 #ifdef CONFIG_DEBUG_SLAB
136 #define FORCED_DEBUG 1
140 #define FORCED_DEBUG 0
143 /* Shouldn't this be in a header file somewhere? */
144 #define BYTES_PER_WORD sizeof(void *)
145 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
147 #ifndef ARCH_KMALLOC_FLAGS
148 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
151 /* Legal flag mask for kmem_cache_create(). */
153 # define CREATE_MASK (SLAB_RED_ZONE | \
154 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
157 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
158 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
159 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
161 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
163 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
164 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
165 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
171 * Bufctl's are used for linking objs within a slab
174 * This implementation relies on "struct page" for locating the cache &
175 * slab an object belongs to.
176 * This allows the bufctl structure to be small (one int), but limits
177 * the number of objects a slab (not a cache) can contain when off-slab
178 * bufctls are used. The limit is the size of the largest general cache
179 * that does not use off-slab slabs.
180 * For 32bit archs with 4 kB pages, is this 56.
181 * This is not serious, as it is only for large objects, when it is unwise
182 * to have too many per slab.
183 * Note: This limit can be raised by introducing a general cache whose size
184 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
187 typedef unsigned int kmem_bufctl_t
;
188 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
189 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
190 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
191 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
196 * Manages the objs in a slab. Placed either at the beginning of mem allocated
197 * for a slab, or allocated from an general cache.
198 * Slabs are chained into three list: fully used, partial, fully free slabs.
201 struct list_head list
;
202 unsigned long colouroff
;
203 void *s_mem
; /* including colour offset */
204 unsigned int inuse
; /* num of objs active in slab */
206 unsigned short nodeid
;
212 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
213 * arrange for kmem_freepages to be called via RCU. This is useful if
214 * we need to approach a kernel structure obliquely, from its address
215 * obtained without the usual locking. We can lock the structure to
216 * stabilize it and check it's still at the given address, only if we
217 * can be sure that the memory has not been meanwhile reused for some
218 * other kind of object (which our subsystem's lock might corrupt).
220 * rcu_read_lock before reading the address, then rcu_read_unlock after
221 * taking the spinlock within the structure expected at that address.
223 * We assume struct slab_rcu can overlay struct slab when destroying.
226 struct rcu_head head
;
227 struct kmem_cache
*cachep
;
235 * - LIFO ordering, to hand out cache-warm objects from _alloc
236 * - reduce the number of linked list operations
237 * - reduce spinlock operations
239 * The limit is stored in the per-cpu structure to reduce the data cache
246 unsigned int batchcount
;
247 unsigned int touched
;
250 * Must have this definition in here for the proper
251 * alignment of array_cache. Also simplifies accessing
257 * bootstrap: The caches do not work without cpuarrays anymore, but the
258 * cpuarrays are allocated from the generic caches...
260 #define BOOT_CPUCACHE_ENTRIES 1
261 struct arraycache_init
{
262 struct array_cache cache
;
263 void *entries
[BOOT_CPUCACHE_ENTRIES
];
267 * The slab lists for all objects.
270 struct list_head slabs_partial
; /* partial list first, better asm code */
271 struct list_head slabs_full
;
272 struct list_head slabs_free
;
273 unsigned long free_objects
;
274 unsigned int free_limit
;
275 unsigned int colour_next
; /* Per-node cache coloring */
276 spinlock_t list_lock
;
277 struct array_cache
*shared
; /* shared per node */
278 struct array_cache
**alien
; /* on other nodes */
279 unsigned long next_reap
; /* updated without locking */
280 int free_touched
; /* updated without locking */
284 * Need this for bootstrapping a per node allocator.
286 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
287 struct kmem_list3 __initdata initkmem_list3
[NUM_INIT_LISTS
];
288 #define CACHE_CACHE 0
289 #define SIZE_AC MAX_NUMNODES
290 #define SIZE_L3 (2 * MAX_NUMNODES)
292 static int drain_freelist(struct kmem_cache
*cache
,
293 struct kmem_list3
*l3
, int tofree
);
294 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
296 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
);
297 static void cache_reap(struct work_struct
*unused
);
300 * This function must be completely optimized away if a constant is passed to
301 * it. Mostly the same as what is in linux/slab.h except it returns an index.
303 static __always_inline
int index_of(const size_t size
)
305 extern void __bad_size(void);
307 if (__builtin_constant_p(size
)) {
315 #include <linux/kmalloc_sizes.h>
323 static int slab_early_init
= 1;
325 #define INDEX_AC index_of(sizeof(struct arraycache_init))
326 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
328 static void kmem_list3_init(struct kmem_list3
*parent
)
330 INIT_LIST_HEAD(&parent
->slabs_full
);
331 INIT_LIST_HEAD(&parent
->slabs_partial
);
332 INIT_LIST_HEAD(&parent
->slabs_free
);
333 parent
->shared
= NULL
;
334 parent
->alien
= NULL
;
335 parent
->colour_next
= 0;
336 spin_lock_init(&parent
->list_lock
);
337 parent
->free_objects
= 0;
338 parent
->free_touched
= 0;
341 #define MAKE_LIST(cachep, listp, slab, nodeid) \
343 INIT_LIST_HEAD(listp); \
344 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
347 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
349 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
350 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
351 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
354 #define CFLGS_OFF_SLAB (0x80000000UL)
355 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
357 #define BATCHREFILL_LIMIT 16
359 * Optimization question: fewer reaps means less probability for unnessary
360 * cpucache drain/refill cycles.
362 * OTOH the cpuarrays can contain lots of objects,
363 * which could lock up otherwise freeable slabs.
365 #define REAPTIMEOUT_CPUC (2*HZ)
366 #define REAPTIMEOUT_LIST3 (4*HZ)
369 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
370 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
371 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
372 #define STATS_INC_GROWN(x) ((x)->grown++)
373 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
374 #define STATS_SET_HIGH(x) \
376 if ((x)->num_active > (x)->high_mark) \
377 (x)->high_mark = (x)->num_active; \
379 #define STATS_INC_ERR(x) ((x)->errors++)
380 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
381 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
382 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
383 #define STATS_SET_FREEABLE(x, i) \
385 if ((x)->max_freeable < i) \
386 (x)->max_freeable = i; \
388 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
389 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
390 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
391 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
393 #define STATS_INC_ACTIVE(x) do { } while (0)
394 #define STATS_DEC_ACTIVE(x) do { } while (0)
395 #define STATS_INC_ALLOCED(x) do { } while (0)
396 #define STATS_INC_GROWN(x) do { } while (0)
397 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
398 #define STATS_SET_HIGH(x) do { } while (0)
399 #define STATS_INC_ERR(x) do { } while (0)
400 #define STATS_INC_NODEALLOCS(x) do { } while (0)
401 #define STATS_INC_NODEFREES(x) do { } while (0)
402 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
403 #define STATS_SET_FREEABLE(x, i) do { } while (0)
404 #define STATS_INC_ALLOCHIT(x) do { } while (0)
405 #define STATS_INC_ALLOCMISS(x) do { } while (0)
406 #define STATS_INC_FREEHIT(x) do { } while (0)
407 #define STATS_INC_FREEMISS(x) do { } while (0)
413 * memory layout of objects:
415 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
416 * the end of an object is aligned with the end of the real
417 * allocation. Catches writes behind the end of the allocation.
418 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
420 * cachep->obj_offset: The real object.
421 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
422 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
423 * [BYTES_PER_WORD long]
425 static int obj_offset(struct kmem_cache
*cachep
)
427 return cachep
->obj_offset
;
430 static int obj_size(struct kmem_cache
*cachep
)
432 return cachep
->obj_size
;
435 static unsigned long long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
437 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
438 return (unsigned long long*) (objp
+ obj_offset(cachep
) -
439 sizeof(unsigned long long));
442 static unsigned long long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
444 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
445 if (cachep
->flags
& SLAB_STORE_USER
)
446 return (unsigned long long *)(objp
+ cachep
->buffer_size
-
447 sizeof(unsigned long long) -
449 return (unsigned long long *) (objp
+ cachep
->buffer_size
-
450 sizeof(unsigned long long));
453 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
455 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
456 return (void **)(objp
+ cachep
->buffer_size
- BYTES_PER_WORD
);
461 #define obj_offset(x) 0
462 #define obj_size(cachep) (cachep->buffer_size)
463 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
464 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
465 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
469 #ifdef CONFIG_TRACING
470 size_t slab_buffer_size(struct kmem_cache
*cachep
)
472 return cachep
->buffer_size
;
474 EXPORT_SYMBOL(slab_buffer_size
);
478 * Do not go above this order unless 0 objects fit into the slab.
480 #define BREAK_GFP_ORDER_HI 1
481 #define BREAK_GFP_ORDER_LO 0
482 static int slab_break_gfp_order
= BREAK_GFP_ORDER_LO
;
485 * Functions for storing/retrieving the cachep and or slab from the page
486 * allocator. These are used to find the slab an obj belongs to. With kfree(),
487 * these are used to find the cache which an obj belongs to.
489 static inline void page_set_cache(struct page
*page
, struct kmem_cache
*cache
)
491 page
->lru
.next
= (struct list_head
*)cache
;
494 static inline struct kmem_cache
*page_get_cache(struct page
*page
)
496 page
= compound_head(page
);
497 BUG_ON(!PageSlab(page
));
498 return (struct kmem_cache
*)page
->lru
.next
;
501 static inline void page_set_slab(struct page
*page
, struct slab
*slab
)
503 page
->lru
.prev
= (struct list_head
*)slab
;
506 static inline struct slab
*page_get_slab(struct page
*page
)
508 BUG_ON(!PageSlab(page
));
509 return (struct slab
*)page
->lru
.prev
;
512 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
514 struct page
*page
= virt_to_head_page(obj
);
515 return page_get_cache(page
);
518 static inline struct slab
*virt_to_slab(const void *obj
)
520 struct page
*page
= virt_to_head_page(obj
);
521 return page_get_slab(page
);
524 static inline void *index_to_obj(struct kmem_cache
*cache
, struct slab
*slab
,
527 return slab
->s_mem
+ cache
->buffer_size
* idx
;
531 * We want to avoid an expensive divide : (offset / cache->buffer_size)
532 * Using the fact that buffer_size is a constant for a particular cache,
533 * we can replace (offset / cache->buffer_size) by
534 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
536 static inline unsigned int obj_to_index(const struct kmem_cache
*cache
,
537 const struct slab
*slab
, void *obj
)
539 u32 offset
= (obj
- slab
->s_mem
);
540 return reciprocal_divide(offset
, cache
->reciprocal_buffer_size
);
544 * These are the default caches for kmalloc. Custom caches can have other sizes.
546 struct cache_sizes malloc_sizes
[] = {
547 #define CACHE(x) { .cs_size = (x) },
548 #include <linux/kmalloc_sizes.h>
552 EXPORT_SYMBOL(malloc_sizes
);
554 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
560 static struct cache_names __initdata cache_names
[] = {
561 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
562 #include <linux/kmalloc_sizes.h>
567 static struct arraycache_init initarray_cache __initdata
=
568 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
569 static struct arraycache_init initarray_generic
=
570 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
572 /* internal cache of cache description objs */
573 static struct kmem_cache cache_cache
= {
575 .limit
= BOOT_CPUCACHE_ENTRIES
,
577 .buffer_size
= sizeof(struct kmem_cache
),
578 .name
= "kmem_cache",
581 #define BAD_ALIEN_MAGIC 0x01020304ul
584 * chicken and egg problem: delay the per-cpu array allocation
585 * until the general caches are up.
596 * used by boot code to determine if it can use slab based allocator
598 int slab_is_available(void)
600 return g_cpucache_up
>= EARLY
;
603 #ifdef CONFIG_LOCKDEP
606 * Slab sometimes uses the kmalloc slabs to store the slab headers
607 * for other slabs "off slab".
608 * The locking for this is tricky in that it nests within the locks
609 * of all other slabs in a few places; to deal with this special
610 * locking we put on-slab caches into a separate lock-class.
612 * We set lock class for alien array caches which are up during init.
613 * The lock annotation will be lost if all cpus of a node goes down and
614 * then comes back up during hotplug
616 static struct lock_class_key on_slab_l3_key
;
617 static struct lock_class_key on_slab_alc_key
;
619 static void init_node_lock_keys(int q
)
621 struct cache_sizes
*s
= malloc_sizes
;
623 if (g_cpucache_up
!= FULL
)
626 for (s
= malloc_sizes
; s
->cs_size
!= ULONG_MAX
; s
++) {
627 struct array_cache
**alc
;
628 struct kmem_list3
*l3
;
631 l3
= s
->cs_cachep
->nodelists
[q
];
632 if (!l3
|| OFF_SLAB(s
->cs_cachep
))
634 lockdep_set_class(&l3
->list_lock
, &on_slab_l3_key
);
636 if (!alc
|| (unsigned long)alc
== BAD_ALIEN_MAGIC
)
640 lockdep_set_class(&alc
[r
]->lock
,
646 static inline void init_lock_keys(void)
651 init_node_lock_keys(node
);
654 static void init_node_lock_keys(int q
)
658 static inline void init_lock_keys(void)
664 * Guard access to the cache-chain.
666 static DEFINE_MUTEX(cache_chain_mutex
);
667 static struct list_head cache_chain
;
669 static DEFINE_PER_CPU(struct delayed_work
, slab_reap_work
);
671 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
673 return cachep
->array
[smp_processor_id()];
676 static inline struct kmem_cache
*__find_general_cachep(size_t size
,
679 struct cache_sizes
*csizep
= malloc_sizes
;
682 /* This happens if someone tries to call
683 * kmem_cache_create(), or __kmalloc(), before
684 * the generic caches are initialized.
686 BUG_ON(malloc_sizes
[INDEX_AC
].cs_cachep
== NULL
);
689 return ZERO_SIZE_PTR
;
691 while (size
> csizep
->cs_size
)
695 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
696 * has cs_{dma,}cachep==NULL. Thus no special case
697 * for large kmalloc calls required.
699 #ifdef CONFIG_ZONE_DMA
700 if (unlikely(gfpflags
& GFP_DMA
))
701 return csizep
->cs_dmacachep
;
703 return csizep
->cs_cachep
;
706 static struct kmem_cache
*kmem_find_general_cachep(size_t size
, gfp_t gfpflags
)
708 return __find_general_cachep(size
, gfpflags
);
711 static size_t slab_mgmt_size(size_t nr_objs
, size_t align
)
713 return ALIGN(sizeof(struct slab
)+nr_objs
*sizeof(kmem_bufctl_t
), align
);
717 * Calculate the number of objects and left-over bytes for a given buffer size.
719 static void cache_estimate(unsigned long gfporder
, size_t buffer_size
,
720 size_t align
, int flags
, size_t *left_over
,
725 size_t slab_size
= PAGE_SIZE
<< gfporder
;
728 * The slab management structure can be either off the slab or
729 * on it. For the latter case, the memory allocated for a
733 * - One kmem_bufctl_t for each object
734 * - Padding to respect alignment of @align
735 * - @buffer_size bytes for each object
737 * If the slab management structure is off the slab, then the
738 * alignment will already be calculated into the size. Because
739 * the slabs are all pages aligned, the objects will be at the
740 * correct alignment when allocated.
742 if (flags
& CFLGS_OFF_SLAB
) {
744 nr_objs
= slab_size
/ buffer_size
;
746 if (nr_objs
> SLAB_LIMIT
)
747 nr_objs
= SLAB_LIMIT
;
750 * Ignore padding for the initial guess. The padding
751 * is at most @align-1 bytes, and @buffer_size is at
752 * least @align. In the worst case, this result will
753 * be one greater than the number of objects that fit
754 * into the memory allocation when taking the padding
757 nr_objs
= (slab_size
- sizeof(struct slab
)) /
758 (buffer_size
+ sizeof(kmem_bufctl_t
));
761 * This calculated number will be either the right
762 * amount, or one greater than what we want.
764 if (slab_mgmt_size(nr_objs
, align
) + nr_objs
*buffer_size
768 if (nr_objs
> SLAB_LIMIT
)
769 nr_objs
= SLAB_LIMIT
;
771 mgmt_size
= slab_mgmt_size(nr_objs
, align
);
774 *left_over
= slab_size
- nr_objs
*buffer_size
- mgmt_size
;
777 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
779 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
782 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
783 function
, cachep
->name
, msg
);
788 * By default on NUMA we use alien caches to stage the freeing of
789 * objects allocated from other nodes. This causes massive memory
790 * inefficiencies when using fake NUMA setup to split memory into a
791 * large number of small nodes, so it can be disabled on the command
795 static int use_alien_caches __read_mostly
= 1;
796 static int __init
noaliencache_setup(char *s
)
798 use_alien_caches
= 0;
801 __setup("noaliencache", noaliencache_setup
);
805 * Special reaping functions for NUMA systems called from cache_reap().
806 * These take care of doing round robin flushing of alien caches (containing
807 * objects freed on different nodes from which they were allocated) and the
808 * flushing of remote pcps by calling drain_node_pages.
810 static DEFINE_PER_CPU(unsigned long, slab_reap_node
);
812 static void init_reap_node(int cpu
)
816 node
= next_node(cpu_to_mem(cpu
), node_online_map
);
817 if (node
== MAX_NUMNODES
)
818 node
= first_node(node_online_map
);
820 per_cpu(slab_reap_node
, cpu
) = node
;
823 static void next_reap_node(void)
825 int node
= __get_cpu_var(slab_reap_node
);
827 node
= next_node(node
, node_online_map
);
828 if (unlikely(node
>= MAX_NUMNODES
))
829 node
= first_node(node_online_map
);
830 __get_cpu_var(slab_reap_node
) = node
;
834 #define init_reap_node(cpu) do { } while (0)
835 #define next_reap_node(void) do { } while (0)
839 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
840 * via the workqueue/eventd.
841 * Add the CPU number into the expiration time to minimize the possibility of
842 * the CPUs getting into lockstep and contending for the global cache chain
845 static void __cpuinit
start_cpu_timer(int cpu
)
847 struct delayed_work
*reap_work
= &per_cpu(slab_reap_work
, cpu
);
850 * When this gets called from do_initcalls via cpucache_init(),
851 * init_workqueues() has already run, so keventd will be setup
854 if (keventd_up() && reap_work
->work
.func
== NULL
) {
856 INIT_DELAYED_WORK_DEFERRABLE(reap_work
, cache_reap
);
857 schedule_delayed_work_on(cpu
, reap_work
,
858 __round_jiffies_relative(HZ
, cpu
));
862 static struct array_cache
*alloc_arraycache(int node
, int entries
,
863 int batchcount
, gfp_t gfp
)
865 int memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
866 struct array_cache
*nc
= NULL
;
868 nc
= kmalloc_node(memsize
, gfp
, node
);
870 * The array_cache structures contain pointers to free object.
871 * However, when such objects are allocated or transfered to another
872 * cache the pointers are not cleared and they could be counted as
873 * valid references during a kmemleak scan. Therefore, kmemleak must
874 * not scan such objects.
876 kmemleak_no_scan(nc
);
880 nc
->batchcount
= batchcount
;
882 spin_lock_init(&nc
->lock
);
888 * Transfer objects in one arraycache to another.
889 * Locking must be handled by the caller.
891 * Return the number of entries transferred.
893 static int transfer_objects(struct array_cache
*to
,
894 struct array_cache
*from
, unsigned int max
)
896 /* Figure out how many entries to transfer */
897 int nr
= min(min(from
->avail
, max
), to
->limit
- to
->avail
);
902 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
912 #define drain_alien_cache(cachep, alien) do { } while (0)
913 #define reap_alien(cachep, l3) do { } while (0)
915 static inline struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
917 return (struct array_cache
**)BAD_ALIEN_MAGIC
;
920 static inline void free_alien_cache(struct array_cache
**ac_ptr
)
924 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
929 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
935 static inline void *____cache_alloc_node(struct kmem_cache
*cachep
,
936 gfp_t flags
, int nodeid
)
941 #else /* CONFIG_NUMA */
943 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
944 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
946 static struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
948 struct array_cache
**ac_ptr
;
949 int memsize
= sizeof(void *) * nr_node_ids
;
954 ac_ptr
= kzalloc_node(memsize
, gfp
, node
);
957 if (i
== node
|| !node_online(i
))
959 ac_ptr
[i
] = alloc_arraycache(node
, limit
, 0xbaadf00d, gfp
);
961 for (i
--; i
>= 0; i
--)
971 static void free_alien_cache(struct array_cache
**ac_ptr
)
982 static void __drain_alien_cache(struct kmem_cache
*cachep
,
983 struct array_cache
*ac
, int node
)
985 struct kmem_list3
*rl3
= cachep
->nodelists
[node
];
988 spin_lock(&rl3
->list_lock
);
990 * Stuff objects into the remote nodes shared array first.
991 * That way we could avoid the overhead of putting the objects
992 * into the free lists and getting them back later.
995 transfer_objects(rl3
->shared
, ac
, ac
->limit
);
997 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
999 spin_unlock(&rl3
->list_lock
);
1004 * Called from cache_reap() to regularly drain alien caches round robin.
1006 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_list3
*l3
)
1008 int node
= __get_cpu_var(slab_reap_node
);
1011 struct array_cache
*ac
= l3
->alien
[node
];
1013 if (ac
&& ac
->avail
&& spin_trylock_irq(&ac
->lock
)) {
1014 __drain_alien_cache(cachep
, ac
, node
);
1015 spin_unlock_irq(&ac
->lock
);
1020 static void drain_alien_cache(struct kmem_cache
*cachep
,
1021 struct array_cache
**alien
)
1024 struct array_cache
*ac
;
1025 unsigned long flags
;
1027 for_each_online_node(i
) {
1030 spin_lock_irqsave(&ac
->lock
, flags
);
1031 __drain_alien_cache(cachep
, ac
, i
);
1032 spin_unlock_irqrestore(&ac
->lock
, flags
);
1037 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1039 struct slab
*slabp
= virt_to_slab(objp
);
1040 int nodeid
= slabp
->nodeid
;
1041 struct kmem_list3
*l3
;
1042 struct array_cache
*alien
= NULL
;
1045 node
= numa_mem_id();
1048 * Make sure we are not freeing a object from another node to the array
1049 * cache on this cpu.
1051 if (likely(slabp
->nodeid
== node
))
1054 l3
= cachep
->nodelists
[node
];
1055 STATS_INC_NODEFREES(cachep
);
1056 if (l3
->alien
&& l3
->alien
[nodeid
]) {
1057 alien
= l3
->alien
[nodeid
];
1058 spin_lock(&alien
->lock
);
1059 if (unlikely(alien
->avail
== alien
->limit
)) {
1060 STATS_INC_ACOVERFLOW(cachep
);
1061 __drain_alien_cache(cachep
, alien
, nodeid
);
1063 alien
->entry
[alien
->avail
++] = objp
;
1064 spin_unlock(&alien
->lock
);
1066 spin_lock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1067 free_block(cachep
, &objp
, 1, nodeid
);
1068 spin_unlock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1075 * Allocates and initializes nodelists for a node on each slab cache, used for
1076 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1077 * will be allocated off-node since memory is not yet online for the new node.
1078 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1081 * Must hold cache_chain_mutex.
1083 static int init_cache_nodelists_node(int node
)
1085 struct kmem_cache
*cachep
;
1086 struct kmem_list3
*l3
;
1087 const int memsize
= sizeof(struct kmem_list3
);
1089 list_for_each_entry(cachep
, &cache_chain
, next
) {
1091 * Set up the size64 kmemlist for cpu before we can
1092 * begin anything. Make sure some other cpu on this
1093 * node has not already allocated this
1095 if (!cachep
->nodelists
[node
]) {
1096 l3
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1099 kmem_list3_init(l3
);
1100 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
1101 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1104 * The l3s don't come and go as CPUs come and
1105 * go. cache_chain_mutex is sufficient
1108 cachep
->nodelists
[node
] = l3
;
1111 spin_lock_irq(&cachep
->nodelists
[node
]->list_lock
);
1112 cachep
->nodelists
[node
]->free_limit
=
1113 (1 + nr_cpus_node(node
)) *
1114 cachep
->batchcount
+ cachep
->num
;
1115 spin_unlock_irq(&cachep
->nodelists
[node
]->list_lock
);
1120 static void __cpuinit
cpuup_canceled(long cpu
)
1122 struct kmem_cache
*cachep
;
1123 struct kmem_list3
*l3
= NULL
;
1124 int node
= cpu_to_mem(cpu
);
1125 const struct cpumask
*mask
= cpumask_of_node(node
);
1127 list_for_each_entry(cachep
, &cache_chain
, next
) {
1128 struct array_cache
*nc
;
1129 struct array_cache
*shared
;
1130 struct array_cache
**alien
;
1132 /* cpu is dead; no one can alloc from it. */
1133 nc
= cachep
->array
[cpu
];
1134 cachep
->array
[cpu
] = NULL
;
1135 l3
= cachep
->nodelists
[node
];
1138 goto free_array_cache
;
1140 spin_lock_irq(&l3
->list_lock
);
1142 /* Free limit for this kmem_list3 */
1143 l3
->free_limit
-= cachep
->batchcount
;
1145 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
1147 if (!cpumask_empty(mask
)) {
1148 spin_unlock_irq(&l3
->list_lock
);
1149 goto free_array_cache
;
1152 shared
= l3
->shared
;
1154 free_block(cachep
, shared
->entry
,
1155 shared
->avail
, node
);
1162 spin_unlock_irq(&l3
->list_lock
);
1166 drain_alien_cache(cachep
, alien
);
1167 free_alien_cache(alien
);
1173 * In the previous loop, all the objects were freed to
1174 * the respective cache's slabs, now we can go ahead and
1175 * shrink each nodelist to its limit.
1177 list_for_each_entry(cachep
, &cache_chain
, next
) {
1178 l3
= cachep
->nodelists
[node
];
1181 drain_freelist(cachep
, l3
, l3
->free_objects
);
1185 static int __cpuinit
cpuup_prepare(long cpu
)
1187 struct kmem_cache
*cachep
;
1188 struct kmem_list3
*l3
= NULL
;
1189 int node
= cpu_to_mem(cpu
);
1193 * We need to do this right in the beginning since
1194 * alloc_arraycache's are going to use this list.
1195 * kmalloc_node allows us to add the slab to the right
1196 * kmem_list3 and not this cpu's kmem_list3
1198 err
= init_cache_nodelists_node(node
);
1203 * Now we can go ahead with allocating the shared arrays and
1206 list_for_each_entry(cachep
, &cache_chain
, next
) {
1207 struct array_cache
*nc
;
1208 struct array_cache
*shared
= NULL
;
1209 struct array_cache
**alien
= NULL
;
1211 nc
= alloc_arraycache(node
, cachep
->limit
,
1212 cachep
->batchcount
, GFP_KERNEL
);
1215 if (cachep
->shared
) {
1216 shared
= alloc_arraycache(node
,
1217 cachep
->shared
* cachep
->batchcount
,
1218 0xbaadf00d, GFP_KERNEL
);
1224 if (use_alien_caches
) {
1225 alien
= alloc_alien_cache(node
, cachep
->limit
, GFP_KERNEL
);
1232 cachep
->array
[cpu
] = nc
;
1233 l3
= cachep
->nodelists
[node
];
1236 spin_lock_irq(&l3
->list_lock
);
1239 * We are serialised from CPU_DEAD or
1240 * CPU_UP_CANCELLED by the cpucontrol lock
1242 l3
->shared
= shared
;
1251 spin_unlock_irq(&l3
->list_lock
);
1253 free_alien_cache(alien
);
1255 init_node_lock_keys(node
);
1259 cpuup_canceled(cpu
);
1263 static int __cpuinit
cpuup_callback(struct notifier_block
*nfb
,
1264 unsigned long action
, void *hcpu
)
1266 long cpu
= (long)hcpu
;
1270 case CPU_UP_PREPARE
:
1271 case CPU_UP_PREPARE_FROZEN
:
1272 mutex_lock(&cache_chain_mutex
);
1273 err
= cpuup_prepare(cpu
);
1274 mutex_unlock(&cache_chain_mutex
);
1277 case CPU_ONLINE_FROZEN
:
1278 start_cpu_timer(cpu
);
1280 #ifdef CONFIG_HOTPLUG_CPU
1281 case CPU_DOWN_PREPARE
:
1282 case CPU_DOWN_PREPARE_FROZEN
:
1284 * Shutdown cache reaper. Note that the cache_chain_mutex is
1285 * held so that if cache_reap() is invoked it cannot do
1286 * anything expensive but will only modify reap_work
1287 * and reschedule the timer.
1289 cancel_rearming_delayed_work(&per_cpu(slab_reap_work
, cpu
));
1290 /* Now the cache_reaper is guaranteed to be not running. */
1291 per_cpu(slab_reap_work
, cpu
).work
.func
= NULL
;
1293 case CPU_DOWN_FAILED
:
1294 case CPU_DOWN_FAILED_FROZEN
:
1295 start_cpu_timer(cpu
);
1298 case CPU_DEAD_FROZEN
:
1300 * Even if all the cpus of a node are down, we don't free the
1301 * kmem_list3 of any cache. This to avoid a race between
1302 * cpu_down, and a kmalloc allocation from another cpu for
1303 * memory from the node of the cpu going down. The list3
1304 * structure is usually allocated from kmem_cache_create() and
1305 * gets destroyed at kmem_cache_destroy().
1309 case CPU_UP_CANCELED
:
1310 case CPU_UP_CANCELED_FROZEN
:
1311 mutex_lock(&cache_chain_mutex
);
1312 cpuup_canceled(cpu
);
1313 mutex_unlock(&cache_chain_mutex
);
1316 return notifier_from_errno(err
);
1319 static struct notifier_block __cpuinitdata cpucache_notifier
= {
1320 &cpuup_callback
, NULL
, 0
1323 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1325 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1326 * Returns -EBUSY if all objects cannot be drained so that the node is not
1329 * Must hold cache_chain_mutex.
1331 static int __meminit
drain_cache_nodelists_node(int node
)
1333 struct kmem_cache
*cachep
;
1336 list_for_each_entry(cachep
, &cache_chain
, next
) {
1337 struct kmem_list3
*l3
;
1339 l3
= cachep
->nodelists
[node
];
1343 drain_freelist(cachep
, l3
, l3
->free_objects
);
1345 if (!list_empty(&l3
->slabs_full
) ||
1346 !list_empty(&l3
->slabs_partial
)) {
1354 static int __meminit
slab_memory_callback(struct notifier_block
*self
,
1355 unsigned long action
, void *arg
)
1357 struct memory_notify
*mnb
= arg
;
1361 nid
= mnb
->status_change_nid
;
1366 case MEM_GOING_ONLINE
:
1367 mutex_lock(&cache_chain_mutex
);
1368 ret
= init_cache_nodelists_node(nid
);
1369 mutex_unlock(&cache_chain_mutex
);
1371 case MEM_GOING_OFFLINE
:
1372 mutex_lock(&cache_chain_mutex
);
1373 ret
= drain_cache_nodelists_node(nid
);
1374 mutex_unlock(&cache_chain_mutex
);
1378 case MEM_CANCEL_ONLINE
:
1379 case MEM_CANCEL_OFFLINE
:
1383 return ret
? notifier_from_errno(ret
) : NOTIFY_OK
;
1385 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1388 * swap the static kmem_list3 with kmalloced memory
1390 static void __init
init_list(struct kmem_cache
*cachep
, struct kmem_list3
*list
,
1393 struct kmem_list3
*ptr
;
1395 ptr
= kmalloc_node(sizeof(struct kmem_list3
), GFP_NOWAIT
, nodeid
);
1398 memcpy(ptr
, list
, sizeof(struct kmem_list3
));
1400 * Do not assume that spinlocks can be initialized via memcpy:
1402 spin_lock_init(&ptr
->list_lock
);
1404 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1405 cachep
->nodelists
[nodeid
] = ptr
;
1409 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1410 * size of kmem_list3.
1412 static void __init
set_up_list3s(struct kmem_cache
*cachep
, int index
)
1416 for_each_online_node(node
) {
1417 cachep
->nodelists
[node
] = &initkmem_list3
[index
+ node
];
1418 cachep
->nodelists
[node
]->next_reap
= jiffies
+
1420 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1425 * Initialisation. Called after the page allocator have been initialised and
1426 * before smp_init().
1428 void __init
kmem_cache_init(void)
1431 struct cache_sizes
*sizes
;
1432 struct cache_names
*names
;
1437 if (num_possible_nodes() == 1)
1438 use_alien_caches
= 0;
1440 for (i
= 0; i
< NUM_INIT_LISTS
; i
++) {
1441 kmem_list3_init(&initkmem_list3
[i
]);
1442 if (i
< MAX_NUMNODES
)
1443 cache_cache
.nodelists
[i
] = NULL
;
1445 set_up_list3s(&cache_cache
, CACHE_CACHE
);
1448 * Fragmentation resistance on low memory - only use bigger
1449 * page orders on machines with more than 32MB of memory.
1451 if (totalram_pages
> (32 << 20) >> PAGE_SHIFT
)
1452 slab_break_gfp_order
= BREAK_GFP_ORDER_HI
;
1454 /* Bootstrap is tricky, because several objects are allocated
1455 * from caches that do not exist yet:
1456 * 1) initialize the cache_cache cache: it contains the struct
1457 * kmem_cache structures of all caches, except cache_cache itself:
1458 * cache_cache is statically allocated.
1459 * Initially an __init data area is used for the head array and the
1460 * kmem_list3 structures, it's replaced with a kmalloc allocated
1461 * array at the end of the bootstrap.
1462 * 2) Create the first kmalloc cache.
1463 * The struct kmem_cache for the new cache is allocated normally.
1464 * An __init data area is used for the head array.
1465 * 3) Create the remaining kmalloc caches, with minimally sized
1467 * 4) Replace the __init data head arrays for cache_cache and the first
1468 * kmalloc cache with kmalloc allocated arrays.
1469 * 5) Replace the __init data for kmem_list3 for cache_cache and
1470 * the other cache's with kmalloc allocated memory.
1471 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1474 node
= numa_mem_id();
1476 /* 1) create the cache_cache */
1477 INIT_LIST_HEAD(&cache_chain
);
1478 list_add(&cache_cache
.next
, &cache_chain
);
1479 cache_cache
.colour_off
= cache_line_size();
1480 cache_cache
.array
[smp_processor_id()] = &initarray_cache
.cache
;
1481 cache_cache
.nodelists
[node
] = &initkmem_list3
[CACHE_CACHE
+ node
];
1484 * struct kmem_cache size depends on nr_node_ids, which
1485 * can be less than MAX_NUMNODES.
1487 cache_cache
.buffer_size
= offsetof(struct kmem_cache
, nodelists
) +
1488 nr_node_ids
* sizeof(struct kmem_list3
*);
1490 cache_cache
.obj_size
= cache_cache
.buffer_size
;
1492 cache_cache
.buffer_size
= ALIGN(cache_cache
.buffer_size
,
1494 cache_cache
.reciprocal_buffer_size
=
1495 reciprocal_value(cache_cache
.buffer_size
);
1497 for (order
= 0; order
< MAX_ORDER
; order
++) {
1498 cache_estimate(order
, cache_cache
.buffer_size
,
1499 cache_line_size(), 0, &left_over
, &cache_cache
.num
);
1500 if (cache_cache
.num
)
1503 BUG_ON(!cache_cache
.num
);
1504 cache_cache
.gfporder
= order
;
1505 cache_cache
.colour
= left_over
/ cache_cache
.colour_off
;
1506 cache_cache
.slab_size
= ALIGN(cache_cache
.num
* sizeof(kmem_bufctl_t
) +
1507 sizeof(struct slab
), cache_line_size());
1509 /* 2+3) create the kmalloc caches */
1510 sizes
= malloc_sizes
;
1511 names
= cache_names
;
1514 * Initialize the caches that provide memory for the array cache and the
1515 * kmem_list3 structures first. Without this, further allocations will
1519 sizes
[INDEX_AC
].cs_cachep
= kmem_cache_create(names
[INDEX_AC
].name
,
1520 sizes
[INDEX_AC
].cs_size
,
1521 ARCH_KMALLOC_MINALIGN
,
1522 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1525 if (INDEX_AC
!= INDEX_L3
) {
1526 sizes
[INDEX_L3
].cs_cachep
=
1527 kmem_cache_create(names
[INDEX_L3
].name
,
1528 sizes
[INDEX_L3
].cs_size
,
1529 ARCH_KMALLOC_MINALIGN
,
1530 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1534 slab_early_init
= 0;
1536 while (sizes
->cs_size
!= ULONG_MAX
) {
1538 * For performance, all the general caches are L1 aligned.
1539 * This should be particularly beneficial on SMP boxes, as it
1540 * eliminates "false sharing".
1541 * Note for systems short on memory removing the alignment will
1542 * allow tighter packing of the smaller caches.
1544 if (!sizes
->cs_cachep
) {
1545 sizes
->cs_cachep
= kmem_cache_create(names
->name
,
1547 ARCH_KMALLOC_MINALIGN
,
1548 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1551 #ifdef CONFIG_ZONE_DMA
1552 sizes
->cs_dmacachep
= kmem_cache_create(
1555 ARCH_KMALLOC_MINALIGN
,
1556 ARCH_KMALLOC_FLAGS
|SLAB_CACHE_DMA
|
1563 /* 4) Replace the bootstrap head arrays */
1565 struct array_cache
*ptr
;
1567 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1569 BUG_ON(cpu_cache_get(&cache_cache
) != &initarray_cache
.cache
);
1570 memcpy(ptr
, cpu_cache_get(&cache_cache
),
1571 sizeof(struct arraycache_init
));
1573 * Do not assume that spinlocks can be initialized via memcpy:
1575 spin_lock_init(&ptr
->lock
);
1577 cache_cache
.array
[smp_processor_id()] = ptr
;
1579 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1581 BUG_ON(cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
)
1582 != &initarray_generic
.cache
);
1583 memcpy(ptr
, cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
),
1584 sizeof(struct arraycache_init
));
1586 * Do not assume that spinlocks can be initialized via memcpy:
1588 spin_lock_init(&ptr
->lock
);
1590 malloc_sizes
[INDEX_AC
].cs_cachep
->array
[smp_processor_id()] =
1593 /* 5) Replace the bootstrap kmem_list3's */
1597 for_each_online_node(nid
) {
1598 init_list(&cache_cache
, &initkmem_list3
[CACHE_CACHE
+ nid
], nid
);
1600 init_list(malloc_sizes
[INDEX_AC
].cs_cachep
,
1601 &initkmem_list3
[SIZE_AC
+ nid
], nid
);
1603 if (INDEX_AC
!= INDEX_L3
) {
1604 init_list(malloc_sizes
[INDEX_L3
].cs_cachep
,
1605 &initkmem_list3
[SIZE_L3
+ nid
], nid
);
1610 g_cpucache_up
= EARLY
;
1613 void __init
kmem_cache_init_late(void)
1615 struct kmem_cache
*cachep
;
1617 /* 6) resize the head arrays to their final sizes */
1618 mutex_lock(&cache_chain_mutex
);
1619 list_for_each_entry(cachep
, &cache_chain
, next
)
1620 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1622 mutex_unlock(&cache_chain_mutex
);
1625 g_cpucache_up
= FULL
;
1627 /* Annotate slab for lockdep -- annotate the malloc caches */
1631 * Register a cpu startup notifier callback that initializes
1632 * cpu_cache_get for all new cpus
1634 register_cpu_notifier(&cpucache_notifier
);
1638 * Register a memory hotplug callback that initializes and frees
1641 hotplug_memory_notifier(slab_memory_callback
, SLAB_CALLBACK_PRI
);
1645 * The reap timers are started later, with a module init call: That part
1646 * of the kernel is not yet operational.
1650 static int __init
cpucache_init(void)
1655 * Register the timers that return unneeded pages to the page allocator
1657 for_each_online_cpu(cpu
)
1658 start_cpu_timer(cpu
);
1661 __initcall(cpucache_init
);
1664 * Interface to system's page allocator. No need to hold the cache-lock.
1666 * If we requested dmaable memory, we will get it. Even if we
1667 * did not request dmaable memory, we might get it, but that
1668 * would be relatively rare and ignorable.
1670 static void *kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
1678 * Nommu uses slab's for process anonymous memory allocations, and thus
1679 * requires __GFP_COMP to properly refcount higher order allocations
1681 flags
|= __GFP_COMP
;
1684 flags
|= cachep
->gfpflags
;
1685 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1686 flags
|= __GFP_RECLAIMABLE
;
1688 page
= alloc_pages_exact_node(nodeid
, flags
| __GFP_NOTRACK
, cachep
->gfporder
);
1692 nr_pages
= (1 << cachep
->gfporder
);
1693 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1694 add_zone_page_state(page_zone(page
),
1695 NR_SLAB_RECLAIMABLE
, nr_pages
);
1697 add_zone_page_state(page_zone(page
),
1698 NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1699 for (i
= 0; i
< nr_pages
; i
++)
1700 __SetPageSlab(page
+ i
);
1702 if (kmemcheck_enabled
&& !(cachep
->flags
& SLAB_NOTRACK
)) {
1703 kmemcheck_alloc_shadow(page
, cachep
->gfporder
, flags
, nodeid
);
1706 kmemcheck_mark_uninitialized_pages(page
, nr_pages
);
1708 kmemcheck_mark_unallocated_pages(page
, nr_pages
);
1711 return page_address(page
);
1715 * Interface to system's page release.
1717 static void kmem_freepages(struct kmem_cache
*cachep
, void *addr
)
1719 unsigned long i
= (1 << cachep
->gfporder
);
1720 struct page
*page
= virt_to_page(addr
);
1721 const unsigned long nr_freed
= i
;
1723 kmemcheck_free_shadow(page
, cachep
->gfporder
);
1725 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1726 sub_zone_page_state(page_zone(page
),
1727 NR_SLAB_RECLAIMABLE
, nr_freed
);
1729 sub_zone_page_state(page_zone(page
),
1730 NR_SLAB_UNRECLAIMABLE
, nr_freed
);
1732 BUG_ON(!PageSlab(page
));
1733 __ClearPageSlab(page
);
1736 if (current
->reclaim_state
)
1737 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1738 free_pages((unsigned long)addr
, cachep
->gfporder
);
1741 static void kmem_rcu_free(struct rcu_head
*head
)
1743 struct slab_rcu
*slab_rcu
= (struct slab_rcu
*)head
;
1744 struct kmem_cache
*cachep
= slab_rcu
->cachep
;
1746 kmem_freepages(cachep
, slab_rcu
->addr
);
1747 if (OFF_SLAB(cachep
))
1748 kmem_cache_free(cachep
->slabp_cache
, slab_rcu
);
1753 #ifdef CONFIG_DEBUG_PAGEALLOC
1754 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1755 unsigned long caller
)
1757 int size
= obj_size(cachep
);
1759 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1761 if (size
< 5 * sizeof(unsigned long))
1764 *addr
++ = 0x12345678;
1766 *addr
++ = smp_processor_id();
1767 size
-= 3 * sizeof(unsigned long);
1769 unsigned long *sptr
= &caller
;
1770 unsigned long svalue
;
1772 while (!kstack_end(sptr
)) {
1774 if (kernel_text_address(svalue
)) {
1776 size
-= sizeof(unsigned long);
1777 if (size
<= sizeof(unsigned long))
1783 *addr
++ = 0x87654321;
1787 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1789 int size
= obj_size(cachep
);
1790 addr
= &((char *)addr
)[obj_offset(cachep
)];
1792 memset(addr
, val
, size
);
1793 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1796 static void dump_line(char *data
, int offset
, int limit
)
1799 unsigned char error
= 0;
1802 printk(KERN_ERR
"%03x:", offset
);
1803 for (i
= 0; i
< limit
; i
++) {
1804 if (data
[offset
+ i
] != POISON_FREE
) {
1805 error
= data
[offset
+ i
];
1808 printk(" %02x", (unsigned char)data
[offset
+ i
]);
1812 if (bad_count
== 1) {
1813 error
^= POISON_FREE
;
1814 if (!(error
& (error
- 1))) {
1815 printk(KERN_ERR
"Single bit error detected. Probably "
1818 printk(KERN_ERR
"Run memtest86+ or a similar memory "
1821 printk(KERN_ERR
"Run a memory test tool.\n");
1830 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1835 if (cachep
->flags
& SLAB_RED_ZONE
) {
1836 printk(KERN_ERR
"Redzone: 0x%llx/0x%llx.\n",
1837 *dbg_redzone1(cachep
, objp
),
1838 *dbg_redzone2(cachep
, objp
));
1841 if (cachep
->flags
& SLAB_STORE_USER
) {
1842 printk(KERN_ERR
"Last user: [<%p>]",
1843 *dbg_userword(cachep
, objp
));
1844 print_symbol("(%s)",
1845 (unsigned long)*dbg_userword(cachep
, objp
));
1848 realobj
= (char *)objp
+ obj_offset(cachep
);
1849 size
= obj_size(cachep
);
1850 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1853 if (i
+ limit
> size
)
1855 dump_line(realobj
, i
, limit
);
1859 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1865 realobj
= (char *)objp
+ obj_offset(cachep
);
1866 size
= obj_size(cachep
);
1868 for (i
= 0; i
< size
; i
++) {
1869 char exp
= POISON_FREE
;
1872 if (realobj
[i
] != exp
) {
1878 "Slab corruption: %s start=%p, len=%d\n",
1879 cachep
->name
, realobj
, size
);
1880 print_objinfo(cachep
, objp
, 0);
1882 /* Hexdump the affected line */
1885 if (i
+ limit
> size
)
1887 dump_line(realobj
, i
, limit
);
1890 /* Limit to 5 lines */
1896 /* Print some data about the neighboring objects, if they
1899 struct slab
*slabp
= virt_to_slab(objp
);
1902 objnr
= obj_to_index(cachep
, slabp
, objp
);
1904 objp
= index_to_obj(cachep
, slabp
, objnr
- 1);
1905 realobj
= (char *)objp
+ obj_offset(cachep
);
1906 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
1908 print_objinfo(cachep
, objp
, 2);
1910 if (objnr
+ 1 < cachep
->num
) {
1911 objp
= index_to_obj(cachep
, slabp
, objnr
+ 1);
1912 realobj
= (char *)objp
+ obj_offset(cachep
);
1913 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
1915 print_objinfo(cachep
, objp
, 2);
1922 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
, struct slab
*slabp
)
1925 for (i
= 0; i
< cachep
->num
; i
++) {
1926 void *objp
= index_to_obj(cachep
, slabp
, i
);
1928 if (cachep
->flags
& SLAB_POISON
) {
1929 #ifdef CONFIG_DEBUG_PAGEALLOC
1930 if (cachep
->buffer_size
% PAGE_SIZE
== 0 &&
1932 kernel_map_pages(virt_to_page(objp
),
1933 cachep
->buffer_size
/ PAGE_SIZE
, 1);
1935 check_poison_obj(cachep
, objp
);
1937 check_poison_obj(cachep
, objp
);
1940 if (cachep
->flags
& SLAB_RED_ZONE
) {
1941 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1942 slab_error(cachep
, "start of a freed object "
1944 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1945 slab_error(cachep
, "end of a freed object "
1951 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
, struct slab
*slabp
)
1957 * slab_destroy - destroy and release all objects in a slab
1958 * @cachep: cache pointer being destroyed
1959 * @slabp: slab pointer being destroyed
1961 * Destroy all the objs in a slab, and release the mem back to the system.
1962 * Before calling the slab must have been unlinked from the cache. The
1963 * cache-lock is not held/needed.
1965 static void slab_destroy(struct kmem_cache
*cachep
, struct slab
*slabp
)
1967 void *addr
= slabp
->s_mem
- slabp
->colouroff
;
1969 slab_destroy_debugcheck(cachep
, slabp
);
1970 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
1971 struct slab_rcu
*slab_rcu
;
1973 slab_rcu
= (struct slab_rcu
*)slabp
;
1974 slab_rcu
->cachep
= cachep
;
1975 slab_rcu
->addr
= addr
;
1976 call_rcu(&slab_rcu
->head
, kmem_rcu_free
);
1978 kmem_freepages(cachep
, addr
);
1979 if (OFF_SLAB(cachep
))
1980 kmem_cache_free(cachep
->slabp_cache
, slabp
);
1984 static void __kmem_cache_destroy(struct kmem_cache
*cachep
)
1987 struct kmem_list3
*l3
;
1989 for_each_online_cpu(i
)
1990 kfree(cachep
->array
[i
]);
1992 /* NUMA: free the list3 structures */
1993 for_each_online_node(i
) {
1994 l3
= cachep
->nodelists
[i
];
1997 free_alien_cache(l3
->alien
);
2001 kmem_cache_free(&cache_cache
, cachep
);
2006 * calculate_slab_order - calculate size (page order) of slabs
2007 * @cachep: pointer to the cache that is being created
2008 * @size: size of objects to be created in this cache.
2009 * @align: required alignment for the objects.
2010 * @flags: slab allocation flags
2012 * Also calculates the number of objects per slab.
2014 * This could be made much more intelligent. For now, try to avoid using
2015 * high order pages for slabs. When the gfp() functions are more friendly
2016 * towards high-order requests, this should be changed.
2018 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
2019 size_t size
, size_t align
, unsigned long flags
)
2021 unsigned long offslab_limit
;
2022 size_t left_over
= 0;
2025 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
2029 cache_estimate(gfporder
, size
, align
, flags
, &remainder
, &num
);
2033 if (flags
& CFLGS_OFF_SLAB
) {
2035 * Max number of objs-per-slab for caches which
2036 * use off-slab slabs. Needed to avoid a possible
2037 * looping condition in cache_grow().
2039 offslab_limit
= size
- sizeof(struct slab
);
2040 offslab_limit
/= sizeof(kmem_bufctl_t
);
2042 if (num
> offslab_limit
)
2046 /* Found something acceptable - save it away */
2048 cachep
->gfporder
= gfporder
;
2049 left_over
= remainder
;
2052 * A VFS-reclaimable slab tends to have most allocations
2053 * as GFP_NOFS and we really don't want to have to be allocating
2054 * higher-order pages when we are unable to shrink dcache.
2056 if (flags
& SLAB_RECLAIM_ACCOUNT
)
2060 * Large number of objects is good, but very large slabs are
2061 * currently bad for the gfp()s.
2063 if (gfporder
>= slab_break_gfp_order
)
2067 * Acceptable internal fragmentation?
2069 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
2075 static int __init_refok
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
2077 if (g_cpucache_up
== FULL
)
2078 return enable_cpucache(cachep
, gfp
);
2080 if (g_cpucache_up
== NONE
) {
2082 * Note: the first kmem_cache_create must create the cache
2083 * that's used by kmalloc(24), otherwise the creation of
2084 * further caches will BUG().
2086 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
2089 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2090 * the first cache, then we need to set up all its list3s,
2091 * otherwise the creation of further caches will BUG().
2093 set_up_list3s(cachep
, SIZE_AC
);
2094 if (INDEX_AC
== INDEX_L3
)
2095 g_cpucache_up
= PARTIAL_L3
;
2097 g_cpucache_up
= PARTIAL_AC
;
2099 cachep
->array
[smp_processor_id()] =
2100 kmalloc(sizeof(struct arraycache_init
), gfp
);
2102 if (g_cpucache_up
== PARTIAL_AC
) {
2103 set_up_list3s(cachep
, SIZE_L3
);
2104 g_cpucache_up
= PARTIAL_L3
;
2107 for_each_online_node(node
) {
2108 cachep
->nodelists
[node
] =
2109 kmalloc_node(sizeof(struct kmem_list3
),
2111 BUG_ON(!cachep
->nodelists
[node
]);
2112 kmem_list3_init(cachep
->nodelists
[node
]);
2116 cachep
->nodelists
[numa_mem_id()]->next_reap
=
2117 jiffies
+ REAPTIMEOUT_LIST3
+
2118 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
2120 cpu_cache_get(cachep
)->avail
= 0;
2121 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
2122 cpu_cache_get(cachep
)->batchcount
= 1;
2123 cpu_cache_get(cachep
)->touched
= 0;
2124 cachep
->batchcount
= 1;
2125 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
2130 * kmem_cache_create - Create a cache.
2131 * @name: A string which is used in /proc/slabinfo to identify this cache.
2132 * @size: The size of objects to be created in this cache.
2133 * @align: The required alignment for the objects.
2134 * @flags: SLAB flags
2135 * @ctor: A constructor for the objects.
2137 * Returns a ptr to the cache on success, NULL on failure.
2138 * Cannot be called within a int, but can be interrupted.
2139 * The @ctor is run when new pages are allocated by the cache.
2141 * @name must be valid until the cache is destroyed. This implies that
2142 * the module calling this has to destroy the cache before getting unloaded.
2143 * Note that kmem_cache_name() is not guaranteed to return the same pointer,
2144 * therefore applications must manage it themselves.
2148 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2149 * to catch references to uninitialised memory.
2151 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2152 * for buffer overruns.
2154 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2155 * cacheline. This can be beneficial if you're counting cycles as closely
2159 kmem_cache_create (const char *name
, size_t size
, size_t align
,
2160 unsigned long flags
, void (*ctor
)(void *))
2162 size_t left_over
, slab_size
, ralign
;
2163 struct kmem_cache
*cachep
= NULL
, *pc
;
2167 * Sanity checks... these are all serious usage bugs.
2169 if (!name
|| in_interrupt() || (size
< BYTES_PER_WORD
) ||
2170 size
> KMALLOC_MAX_SIZE
) {
2171 printk(KERN_ERR
"%s: Early error in slab %s\n", __func__
,
2177 * We use cache_chain_mutex to ensure a consistent view of
2178 * cpu_online_mask as well. Please see cpuup_callback
2180 if (slab_is_available()) {
2182 mutex_lock(&cache_chain_mutex
);
2185 list_for_each_entry(pc
, &cache_chain
, next
) {
2190 * This happens when the module gets unloaded and doesn't
2191 * destroy its slab cache and no-one else reuses the vmalloc
2192 * area of the module. Print a warning.
2194 res
= probe_kernel_address(pc
->name
, tmp
);
2197 "SLAB: cache with size %d has lost its name\n",
2202 if (!strcmp(pc
->name
, name
)) {
2204 "kmem_cache_create: duplicate cache %s\n", name
);
2211 WARN_ON(strchr(name
, ' ')); /* It confuses parsers */
2214 * Enable redzoning and last user accounting, except for caches with
2215 * large objects, if the increased size would increase the object size
2216 * above the next power of two: caches with object sizes just above a
2217 * power of two have a significant amount of internal fragmentation.
2219 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
2220 2 * sizeof(unsigned long long)))
2221 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2222 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2223 flags
|= SLAB_POISON
;
2225 if (flags
& SLAB_DESTROY_BY_RCU
)
2226 BUG_ON(flags
& SLAB_POISON
);
2229 * Always checks flags, a caller might be expecting debug support which
2232 BUG_ON(flags
& ~CREATE_MASK
);
2235 * Check that size is in terms of words. This is needed to avoid
2236 * unaligned accesses for some archs when redzoning is used, and makes
2237 * sure any on-slab bufctl's are also correctly aligned.
2239 if (size
& (BYTES_PER_WORD
- 1)) {
2240 size
+= (BYTES_PER_WORD
- 1);
2241 size
&= ~(BYTES_PER_WORD
- 1);
2244 /* calculate the final buffer alignment: */
2246 /* 1) arch recommendation: can be overridden for debug */
2247 if (flags
& SLAB_HWCACHE_ALIGN
) {
2249 * Default alignment: as specified by the arch code. Except if
2250 * an object is really small, then squeeze multiple objects into
2253 ralign
= cache_line_size();
2254 while (size
<= ralign
/ 2)
2257 ralign
= BYTES_PER_WORD
;
2261 * Redzoning and user store require word alignment or possibly larger.
2262 * Note this will be overridden by architecture or caller mandated
2263 * alignment if either is greater than BYTES_PER_WORD.
2265 if (flags
& SLAB_STORE_USER
)
2266 ralign
= BYTES_PER_WORD
;
2268 if (flags
& SLAB_RED_ZONE
) {
2269 ralign
= REDZONE_ALIGN
;
2270 /* If redzoning, ensure that the second redzone is suitably
2271 * aligned, by adjusting the object size accordingly. */
2272 size
+= REDZONE_ALIGN
- 1;
2273 size
&= ~(REDZONE_ALIGN
- 1);
2276 /* 2) arch mandated alignment */
2277 if (ralign
< ARCH_SLAB_MINALIGN
) {
2278 ralign
= ARCH_SLAB_MINALIGN
;
2280 /* 3) caller mandated alignment */
2281 if (ralign
< align
) {
2284 /* disable debug if not aligning with REDZONE_ALIGN */
2285 if (ralign
& (__alignof__(unsigned long long) - 1))
2286 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2292 if (slab_is_available())
2297 /* Get cache's description obj. */
2298 cachep
= kmem_cache_zalloc(&cache_cache
, gfp
);
2303 cachep
->obj_size
= size
;
2306 * Both debugging options require word-alignment which is calculated
2309 if (flags
& SLAB_RED_ZONE
) {
2310 /* add space for red zone words */
2311 cachep
->obj_offset
+= align
;
2312 size
+= align
+ sizeof(unsigned long long);
2314 if (flags
& SLAB_STORE_USER
) {
2315 /* user store requires one word storage behind the end of
2316 * the real object. But if the second red zone needs to be
2317 * aligned to 64 bits, we must allow that much space.
2319 if (flags
& SLAB_RED_ZONE
)
2320 size
+= REDZONE_ALIGN
;
2322 size
+= BYTES_PER_WORD
;
2324 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2325 if (size
>= malloc_sizes
[INDEX_L3
+ 1].cs_size
2326 && cachep
->obj_size
> cache_line_size() && ALIGN(size
, align
) < PAGE_SIZE
) {
2327 cachep
->obj_offset
+= PAGE_SIZE
- ALIGN(size
, align
);
2334 * Determine if the slab management is 'on' or 'off' slab.
2335 * (bootstrapping cannot cope with offslab caches so don't do
2336 * it too early on. Always use on-slab management when
2337 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2339 if ((size
>= (PAGE_SIZE
>> 3)) && !slab_early_init
&&
2340 !(flags
& SLAB_NOLEAKTRACE
))
2342 * Size is large, assume best to place the slab management obj
2343 * off-slab (should allow better packing of objs).
2345 flags
|= CFLGS_OFF_SLAB
;
2347 size
= ALIGN(size
, align
);
2349 left_over
= calculate_slab_order(cachep
, size
, align
, flags
);
2353 "kmem_cache_create: couldn't create cache %s.\n", name
);
2354 kmem_cache_free(&cache_cache
, cachep
);
2358 slab_size
= ALIGN(cachep
->num
* sizeof(kmem_bufctl_t
)
2359 + sizeof(struct slab
), align
);
2362 * If the slab has been placed off-slab, and we have enough space then
2363 * move it on-slab. This is at the expense of any extra colouring.
2365 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= slab_size
) {
2366 flags
&= ~CFLGS_OFF_SLAB
;
2367 left_over
-= slab_size
;
2370 if (flags
& CFLGS_OFF_SLAB
) {
2371 /* really off slab. No need for manual alignment */
2373 cachep
->num
* sizeof(kmem_bufctl_t
) + sizeof(struct slab
);
2375 #ifdef CONFIG_PAGE_POISONING
2376 /* If we're going to use the generic kernel_map_pages()
2377 * poisoning, then it's going to smash the contents of
2378 * the redzone and userword anyhow, so switch them off.
2380 if (size
% PAGE_SIZE
== 0 && flags
& SLAB_POISON
)
2381 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2385 cachep
->colour_off
= cache_line_size();
2386 /* Offset must be a multiple of the alignment. */
2387 if (cachep
->colour_off
< align
)
2388 cachep
->colour_off
= align
;
2389 cachep
->colour
= left_over
/ cachep
->colour_off
;
2390 cachep
->slab_size
= slab_size
;
2391 cachep
->flags
= flags
;
2392 cachep
->gfpflags
= 0;
2393 if (CONFIG_ZONE_DMA_FLAG
&& (flags
& SLAB_CACHE_DMA
))
2394 cachep
->gfpflags
|= GFP_DMA
;
2395 cachep
->buffer_size
= size
;
2396 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2398 if (flags
& CFLGS_OFF_SLAB
) {
2399 cachep
->slabp_cache
= kmem_find_general_cachep(slab_size
, 0u);
2401 * This is a possibility for one of the malloc_sizes caches.
2402 * But since we go off slab only for object size greater than
2403 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2404 * this should not happen at all.
2405 * But leave a BUG_ON for some lucky dude.
2407 BUG_ON(ZERO_OR_NULL_PTR(cachep
->slabp_cache
));
2409 cachep
->ctor
= ctor
;
2410 cachep
->name
= name
;
2412 if (setup_cpu_cache(cachep
, gfp
)) {
2413 __kmem_cache_destroy(cachep
);
2418 /* cache setup completed, link it into the list */
2419 list_add(&cachep
->next
, &cache_chain
);
2421 if (!cachep
&& (flags
& SLAB_PANIC
))
2422 panic("kmem_cache_create(): failed to create slab `%s'\n",
2424 if (slab_is_available()) {
2425 mutex_unlock(&cache_chain_mutex
);
2430 EXPORT_SYMBOL(kmem_cache_create
);
2433 static void check_irq_off(void)
2435 BUG_ON(!irqs_disabled());
2438 static void check_irq_on(void)
2440 BUG_ON(irqs_disabled());
2443 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2447 assert_spin_locked(&cachep
->nodelists
[numa_mem_id()]->list_lock
);
2451 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2455 assert_spin_locked(&cachep
->nodelists
[node
]->list_lock
);
2460 #define check_irq_off() do { } while(0)
2461 #define check_irq_on() do { } while(0)
2462 #define check_spinlock_acquired(x) do { } while(0)
2463 #define check_spinlock_acquired_node(x, y) do { } while(0)
2466 static void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
2467 struct array_cache
*ac
,
2468 int force
, int node
);
2470 static void do_drain(void *arg
)
2472 struct kmem_cache
*cachep
= arg
;
2473 struct array_cache
*ac
;
2474 int node
= numa_mem_id();
2477 ac
= cpu_cache_get(cachep
);
2478 spin_lock(&cachep
->nodelists
[node
]->list_lock
);
2479 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
2480 spin_unlock(&cachep
->nodelists
[node
]->list_lock
);
2484 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2486 struct kmem_list3
*l3
;
2489 on_each_cpu(do_drain
, cachep
, 1);
2491 for_each_online_node(node
) {
2492 l3
= cachep
->nodelists
[node
];
2493 if (l3
&& l3
->alien
)
2494 drain_alien_cache(cachep
, l3
->alien
);
2497 for_each_online_node(node
) {
2498 l3
= cachep
->nodelists
[node
];
2500 drain_array(cachep
, l3
, l3
->shared
, 1, node
);
2505 * Remove slabs from the list of free slabs.
2506 * Specify the number of slabs to drain in tofree.
2508 * Returns the actual number of slabs released.
2510 static int drain_freelist(struct kmem_cache
*cache
,
2511 struct kmem_list3
*l3
, int tofree
)
2513 struct list_head
*p
;
2518 while (nr_freed
< tofree
&& !list_empty(&l3
->slabs_free
)) {
2520 spin_lock_irq(&l3
->list_lock
);
2521 p
= l3
->slabs_free
.prev
;
2522 if (p
== &l3
->slabs_free
) {
2523 spin_unlock_irq(&l3
->list_lock
);
2527 slabp
= list_entry(p
, struct slab
, list
);
2529 BUG_ON(slabp
->inuse
);
2531 list_del(&slabp
->list
);
2533 * Safe to drop the lock. The slab is no longer linked
2536 l3
->free_objects
-= cache
->num
;
2537 spin_unlock_irq(&l3
->list_lock
);
2538 slab_destroy(cache
, slabp
);
2545 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2546 static int __cache_shrink(struct kmem_cache
*cachep
)
2549 struct kmem_list3
*l3
;
2551 drain_cpu_caches(cachep
);
2554 for_each_online_node(i
) {
2555 l3
= cachep
->nodelists
[i
];
2559 drain_freelist(cachep
, l3
, l3
->free_objects
);
2561 ret
+= !list_empty(&l3
->slabs_full
) ||
2562 !list_empty(&l3
->slabs_partial
);
2564 return (ret
? 1 : 0);
2568 * kmem_cache_shrink - Shrink a cache.
2569 * @cachep: The cache to shrink.
2571 * Releases as many slabs as possible for a cache.
2572 * To help debugging, a zero exit status indicates all slabs were released.
2574 int kmem_cache_shrink(struct kmem_cache
*cachep
)
2577 BUG_ON(!cachep
|| in_interrupt());
2580 mutex_lock(&cache_chain_mutex
);
2581 ret
= __cache_shrink(cachep
);
2582 mutex_unlock(&cache_chain_mutex
);
2586 EXPORT_SYMBOL(kmem_cache_shrink
);
2589 * kmem_cache_destroy - delete a cache
2590 * @cachep: the cache to destroy
2592 * Remove a &struct kmem_cache object from the slab cache.
2594 * It is expected this function will be called by a module when it is
2595 * unloaded. This will remove the cache completely, and avoid a duplicate
2596 * cache being allocated each time a module is loaded and unloaded, if the
2597 * module doesn't have persistent in-kernel storage across loads and unloads.
2599 * The cache must be empty before calling this function.
2601 * The caller must guarantee that noone will allocate memory from the cache
2602 * during the kmem_cache_destroy().
2604 void kmem_cache_destroy(struct kmem_cache
*cachep
)
2606 BUG_ON(!cachep
|| in_interrupt());
2608 /* Find the cache in the chain of caches. */
2610 mutex_lock(&cache_chain_mutex
);
2612 * the chain is never empty, cache_cache is never destroyed
2614 list_del(&cachep
->next
);
2615 if (__cache_shrink(cachep
)) {
2616 slab_error(cachep
, "Can't free all objects");
2617 list_add(&cachep
->next
, &cache_chain
);
2618 mutex_unlock(&cache_chain_mutex
);
2623 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
2626 __kmem_cache_destroy(cachep
);
2627 mutex_unlock(&cache_chain_mutex
);
2630 EXPORT_SYMBOL(kmem_cache_destroy
);
2633 * Get the memory for a slab management obj.
2634 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2635 * always come from malloc_sizes caches. The slab descriptor cannot
2636 * come from the same cache which is getting created because,
2637 * when we are searching for an appropriate cache for these
2638 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2639 * If we are creating a malloc_sizes cache here it would not be visible to
2640 * kmem_find_general_cachep till the initialization is complete.
2641 * Hence we cannot have slabp_cache same as the original cache.
2643 static struct slab
*alloc_slabmgmt(struct kmem_cache
*cachep
, void *objp
,
2644 int colour_off
, gfp_t local_flags
,
2649 if (OFF_SLAB(cachep
)) {
2650 /* Slab management obj is off-slab. */
2651 slabp
= kmem_cache_alloc_node(cachep
->slabp_cache
,
2652 local_flags
, nodeid
);
2654 * If the first object in the slab is leaked (it's allocated
2655 * but no one has a reference to it), we want to make sure
2656 * kmemleak does not treat the ->s_mem pointer as a reference
2657 * to the object. Otherwise we will not report the leak.
2659 kmemleak_scan_area(&slabp
->list
, sizeof(struct list_head
),
2664 slabp
= objp
+ colour_off
;
2665 colour_off
+= cachep
->slab_size
;
2668 slabp
->colouroff
= colour_off
;
2669 slabp
->s_mem
= objp
+ colour_off
;
2670 slabp
->nodeid
= nodeid
;
2675 static inline kmem_bufctl_t
*slab_bufctl(struct slab
*slabp
)
2677 return (kmem_bufctl_t
*) (slabp
+ 1);
2680 static void cache_init_objs(struct kmem_cache
*cachep
,
2685 for (i
= 0; i
< cachep
->num
; i
++) {
2686 void *objp
= index_to_obj(cachep
, slabp
, i
);
2688 /* need to poison the objs? */
2689 if (cachep
->flags
& SLAB_POISON
)
2690 poison_obj(cachep
, objp
, POISON_FREE
);
2691 if (cachep
->flags
& SLAB_STORE_USER
)
2692 *dbg_userword(cachep
, objp
) = NULL
;
2694 if (cachep
->flags
& SLAB_RED_ZONE
) {
2695 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2696 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2699 * Constructors are not allowed to allocate memory from the same
2700 * cache which they are a constructor for. Otherwise, deadlock.
2701 * They must also be threaded.
2703 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2704 cachep
->ctor(objp
+ obj_offset(cachep
));
2706 if (cachep
->flags
& SLAB_RED_ZONE
) {
2707 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2708 slab_error(cachep
, "constructor overwrote the"
2709 " end of an object");
2710 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2711 slab_error(cachep
, "constructor overwrote the"
2712 " start of an object");
2714 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 &&
2715 OFF_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
)
2716 kernel_map_pages(virt_to_page(objp
),
2717 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2722 slab_bufctl(slabp
)[i
] = i
+ 1;
2724 slab_bufctl(slabp
)[i
- 1] = BUFCTL_END
;
2727 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2729 if (CONFIG_ZONE_DMA_FLAG
) {
2730 if (flags
& GFP_DMA
)
2731 BUG_ON(!(cachep
->gfpflags
& GFP_DMA
));
2733 BUG_ON(cachep
->gfpflags
& GFP_DMA
);
2737 static void *slab_get_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2740 void *objp
= index_to_obj(cachep
, slabp
, slabp
->free
);
2744 next
= slab_bufctl(slabp
)[slabp
->free
];
2746 slab_bufctl(slabp
)[slabp
->free
] = BUFCTL_FREE
;
2747 WARN_ON(slabp
->nodeid
!= nodeid
);
2754 static void slab_put_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2755 void *objp
, int nodeid
)
2757 unsigned int objnr
= obj_to_index(cachep
, slabp
, objp
);
2760 /* Verify that the slab belongs to the intended node */
2761 WARN_ON(slabp
->nodeid
!= nodeid
);
2763 if (slab_bufctl(slabp
)[objnr
] + 1 <= SLAB_LIMIT
+ 1) {
2764 printk(KERN_ERR
"slab: double free detected in cache "
2765 "'%s', objp %p\n", cachep
->name
, objp
);
2769 slab_bufctl(slabp
)[objnr
] = slabp
->free
;
2770 slabp
->free
= objnr
;
2775 * Map pages beginning at addr to the given cache and slab. This is required
2776 * for the slab allocator to be able to lookup the cache and slab of a
2777 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2779 static void slab_map_pages(struct kmem_cache
*cache
, struct slab
*slab
,
2785 page
= virt_to_page(addr
);
2788 if (likely(!PageCompound(page
)))
2789 nr_pages
<<= cache
->gfporder
;
2792 page_set_cache(page
, cache
);
2793 page_set_slab(page
, slab
);
2795 } while (--nr_pages
);
2799 * Grow (by 1) the number of slabs within a cache. This is called by
2800 * kmem_cache_alloc() when there are no active objs left in a cache.
2802 static int cache_grow(struct kmem_cache
*cachep
,
2803 gfp_t flags
, int nodeid
, void *objp
)
2808 struct kmem_list3
*l3
;
2811 * Be lazy and only check for valid flags here, keeping it out of the
2812 * critical path in kmem_cache_alloc().
2814 BUG_ON(flags
& GFP_SLAB_BUG_MASK
);
2815 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2817 /* Take the l3 list lock to change the colour_next on this node */
2819 l3
= cachep
->nodelists
[nodeid
];
2820 spin_lock(&l3
->list_lock
);
2822 /* Get colour for the slab, and cal the next value. */
2823 offset
= l3
->colour_next
;
2825 if (l3
->colour_next
>= cachep
->colour
)
2826 l3
->colour_next
= 0;
2827 spin_unlock(&l3
->list_lock
);
2829 offset
*= cachep
->colour_off
;
2831 if (local_flags
& __GFP_WAIT
)
2835 * The test for missing atomic flag is performed here, rather than
2836 * the more obvious place, simply to reduce the critical path length
2837 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2838 * will eventually be caught here (where it matters).
2840 kmem_flagcheck(cachep
, flags
);
2843 * Get mem for the objs. Attempt to allocate a physical page from
2847 objp
= kmem_getpages(cachep
, local_flags
, nodeid
);
2851 /* Get slab management. */
2852 slabp
= alloc_slabmgmt(cachep
, objp
, offset
,
2853 local_flags
& ~GFP_CONSTRAINT_MASK
, nodeid
);
2857 slab_map_pages(cachep
, slabp
, objp
);
2859 cache_init_objs(cachep
, slabp
);
2861 if (local_flags
& __GFP_WAIT
)
2862 local_irq_disable();
2864 spin_lock(&l3
->list_lock
);
2866 /* Make slab active. */
2867 list_add_tail(&slabp
->list
, &(l3
->slabs_free
));
2868 STATS_INC_GROWN(cachep
);
2869 l3
->free_objects
+= cachep
->num
;
2870 spin_unlock(&l3
->list_lock
);
2873 kmem_freepages(cachep
, objp
);
2875 if (local_flags
& __GFP_WAIT
)
2876 local_irq_disable();
2883 * Perform extra freeing checks:
2884 * - detect bad pointers.
2885 * - POISON/RED_ZONE checking
2887 static void kfree_debugcheck(const void *objp
)
2889 if (!virt_addr_valid(objp
)) {
2890 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
2891 (unsigned long)objp
);
2896 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2898 unsigned long long redzone1
, redzone2
;
2900 redzone1
= *dbg_redzone1(cache
, obj
);
2901 redzone2
= *dbg_redzone2(cache
, obj
);
2906 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2909 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2910 slab_error(cache
, "double free detected");
2912 slab_error(cache
, "memory outside object was overwritten");
2914 printk(KERN_ERR
"%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2915 obj
, redzone1
, redzone2
);
2918 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2925 BUG_ON(virt_to_cache(objp
) != cachep
);
2927 objp
-= obj_offset(cachep
);
2928 kfree_debugcheck(objp
);
2929 page
= virt_to_head_page(objp
);
2931 slabp
= page_get_slab(page
);
2933 if (cachep
->flags
& SLAB_RED_ZONE
) {
2934 verify_redzone_free(cachep
, objp
);
2935 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2936 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2938 if (cachep
->flags
& SLAB_STORE_USER
)
2939 *dbg_userword(cachep
, objp
) = caller
;
2941 objnr
= obj_to_index(cachep
, slabp
, objp
);
2943 BUG_ON(objnr
>= cachep
->num
);
2944 BUG_ON(objp
!= index_to_obj(cachep
, slabp
, objnr
));
2946 #ifdef CONFIG_DEBUG_SLAB_LEAK
2947 slab_bufctl(slabp
)[objnr
] = BUFCTL_FREE
;
2949 if (cachep
->flags
& SLAB_POISON
) {
2950 #ifdef CONFIG_DEBUG_PAGEALLOC
2951 if ((cachep
->buffer_size
% PAGE_SIZE
)==0 && OFF_SLAB(cachep
)) {
2952 store_stackinfo(cachep
, objp
, (unsigned long)caller
);
2953 kernel_map_pages(virt_to_page(objp
),
2954 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2956 poison_obj(cachep
, objp
, POISON_FREE
);
2959 poison_obj(cachep
, objp
, POISON_FREE
);
2965 static void check_slabp(struct kmem_cache
*cachep
, struct slab
*slabp
)
2970 /* Check slab's freelist to see if this obj is there. */
2971 for (i
= slabp
->free
; i
!= BUFCTL_END
; i
= slab_bufctl(slabp
)[i
]) {
2973 if (entries
> cachep
->num
|| i
>= cachep
->num
)
2976 if (entries
!= cachep
->num
- slabp
->inuse
) {
2978 printk(KERN_ERR
"slab: Internal list corruption detected in "
2979 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2980 cachep
->name
, cachep
->num
, slabp
, slabp
->inuse
);
2982 i
< sizeof(*slabp
) + cachep
->num
* sizeof(kmem_bufctl_t
);
2985 printk("\n%03x:", i
);
2986 printk(" %02x", ((unsigned char *)slabp
)[i
]);
2993 #define kfree_debugcheck(x) do { } while(0)
2994 #define cache_free_debugcheck(x,objp,z) (objp)
2995 #define check_slabp(x,y) do { } while(0)
2998 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
3001 struct kmem_list3
*l3
;
3002 struct array_cache
*ac
;
3007 node
= numa_mem_id();
3008 ac
= cpu_cache_get(cachep
);
3009 batchcount
= ac
->batchcount
;
3010 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
3012 * If there was little recent activity on this cache, then
3013 * perform only a partial refill. Otherwise we could generate
3016 batchcount
= BATCHREFILL_LIMIT
;
3018 l3
= cachep
->nodelists
[node
];
3020 BUG_ON(ac
->avail
> 0 || !l3
);
3021 spin_lock(&l3
->list_lock
);
3023 /* See if we can refill from the shared array */
3024 if (l3
->shared
&& transfer_objects(ac
, l3
->shared
, batchcount
)) {
3025 l3
->shared
->touched
= 1;
3029 while (batchcount
> 0) {
3030 struct list_head
*entry
;
3032 /* Get slab alloc is to come from. */
3033 entry
= l3
->slabs_partial
.next
;
3034 if (entry
== &l3
->slabs_partial
) {
3035 l3
->free_touched
= 1;
3036 entry
= l3
->slabs_free
.next
;
3037 if (entry
== &l3
->slabs_free
)
3041 slabp
= list_entry(entry
, struct slab
, list
);
3042 check_slabp(cachep
, slabp
);
3043 check_spinlock_acquired(cachep
);
3046 * The slab was either on partial or free list so
3047 * there must be at least one object available for
3050 BUG_ON(slabp
->inuse
>= cachep
->num
);
3052 while (slabp
->inuse
< cachep
->num
&& batchcount
--) {
3053 STATS_INC_ALLOCED(cachep
);
3054 STATS_INC_ACTIVE(cachep
);
3055 STATS_SET_HIGH(cachep
);
3057 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, slabp
,
3060 check_slabp(cachep
, slabp
);
3062 /* move slabp to correct slabp list: */
3063 list_del(&slabp
->list
);
3064 if (slabp
->free
== BUFCTL_END
)
3065 list_add(&slabp
->list
, &l3
->slabs_full
);
3067 list_add(&slabp
->list
, &l3
->slabs_partial
);
3071 l3
->free_objects
-= ac
->avail
;
3073 spin_unlock(&l3
->list_lock
);
3075 if (unlikely(!ac
->avail
)) {
3077 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, node
, NULL
);
3079 /* cache_grow can reenable interrupts, then ac could change. */
3080 ac
= cpu_cache_get(cachep
);
3081 if (!x
&& ac
->avail
== 0) /* no objects in sight? abort */
3084 if (!ac
->avail
) /* objects refilled by interrupt? */
3088 return ac
->entry
[--ac
->avail
];
3091 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
3094 might_sleep_if(flags
& __GFP_WAIT
);
3096 kmem_flagcheck(cachep
, flags
);
3101 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
3102 gfp_t flags
, void *objp
, void *caller
)
3106 if (cachep
->flags
& SLAB_POISON
) {
3107 #ifdef CONFIG_DEBUG_PAGEALLOC
3108 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
3109 kernel_map_pages(virt_to_page(objp
),
3110 cachep
->buffer_size
/ PAGE_SIZE
, 1);
3112 check_poison_obj(cachep
, objp
);
3114 check_poison_obj(cachep
, objp
);
3116 poison_obj(cachep
, objp
, POISON_INUSE
);
3118 if (cachep
->flags
& SLAB_STORE_USER
)
3119 *dbg_userword(cachep
, objp
) = caller
;
3121 if (cachep
->flags
& SLAB_RED_ZONE
) {
3122 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
3123 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
3124 slab_error(cachep
, "double free, or memory outside"
3125 " object was overwritten");
3127 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3128 objp
, *dbg_redzone1(cachep
, objp
),
3129 *dbg_redzone2(cachep
, objp
));
3131 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
3132 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
3134 #ifdef CONFIG_DEBUG_SLAB_LEAK
3139 slabp
= page_get_slab(virt_to_head_page(objp
));
3140 objnr
= (unsigned)(objp
- slabp
->s_mem
) / cachep
->buffer_size
;
3141 slab_bufctl(slabp
)[objnr
] = BUFCTL_ACTIVE
;
3144 objp
+= obj_offset(cachep
);
3145 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
3147 #if ARCH_SLAB_MINALIGN
3148 if ((u32
)objp
& (ARCH_SLAB_MINALIGN
-1)) {
3149 printk(KERN_ERR
"0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3150 objp
, ARCH_SLAB_MINALIGN
);
3156 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3159 static bool slab_should_failslab(struct kmem_cache
*cachep
, gfp_t flags
)
3161 if (cachep
== &cache_cache
)
3164 return should_failslab(obj_size(cachep
), flags
, cachep
->flags
);
3167 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3170 struct array_cache
*ac
;
3174 ac
= cpu_cache_get(cachep
);
3175 if (likely(ac
->avail
)) {
3176 STATS_INC_ALLOCHIT(cachep
);
3178 objp
= ac
->entry
[--ac
->avail
];
3180 STATS_INC_ALLOCMISS(cachep
);
3181 objp
= cache_alloc_refill(cachep
, flags
);
3183 * the 'ac' may be updated by cache_alloc_refill(),
3184 * and kmemleak_erase() requires its correct value.
3186 ac
= cpu_cache_get(cachep
);
3189 * To avoid a false negative, if an object that is in one of the
3190 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3191 * treat the array pointers as a reference to the object.
3194 kmemleak_erase(&ac
->entry
[ac
->avail
]);
3200 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3202 * If we are in_interrupt, then process context, including cpusets and
3203 * mempolicy, may not apply and should not be used for allocation policy.
3205 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3207 int nid_alloc
, nid_here
;
3209 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3211 nid_alloc
= nid_here
= numa_mem_id();
3213 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3214 nid_alloc
= cpuset_slab_spread_node();
3215 else if (current
->mempolicy
)
3216 nid_alloc
= slab_node(current
->mempolicy
);
3218 if (nid_alloc
!= nid_here
)
3219 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3224 * Fallback function if there was no memory available and no objects on a
3225 * certain node and fall back is permitted. First we scan all the
3226 * available nodelists for available objects. If that fails then we
3227 * perform an allocation without specifying a node. This allows the page
3228 * allocator to do its reclaim / fallback magic. We then insert the
3229 * slab into the proper nodelist and then allocate from it.
3231 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3233 struct zonelist
*zonelist
;
3237 enum zone_type high_zoneidx
= gfp_zone(flags
);
3241 if (flags
& __GFP_THISNODE
)
3245 zonelist
= node_zonelist(slab_node(current
->mempolicy
), flags
);
3246 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
3250 * Look through allowed nodes for objects available
3251 * from existing per node queues.
3253 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
3254 nid
= zone_to_nid(zone
);
3256 if (cpuset_zone_allowed_hardwall(zone
, flags
) &&
3257 cache
->nodelists
[nid
] &&
3258 cache
->nodelists
[nid
]->free_objects
) {
3259 obj
= ____cache_alloc_node(cache
,
3260 flags
| GFP_THISNODE
, nid
);
3268 * This allocation will be performed within the constraints
3269 * of the current cpuset / memory policy requirements.
3270 * We may trigger various forms of reclaim on the allowed
3271 * set and go into memory reserves if necessary.
3273 if (local_flags
& __GFP_WAIT
)
3275 kmem_flagcheck(cache
, flags
);
3276 obj
= kmem_getpages(cache
, local_flags
, numa_mem_id());
3277 if (local_flags
& __GFP_WAIT
)
3278 local_irq_disable();
3281 * Insert into the appropriate per node queues
3283 nid
= page_to_nid(virt_to_page(obj
));
3284 if (cache_grow(cache
, flags
, nid
, obj
)) {
3285 obj
= ____cache_alloc_node(cache
,
3286 flags
| GFP_THISNODE
, nid
);
3289 * Another processor may allocate the
3290 * objects in the slab since we are
3291 * not holding any locks.
3295 /* cache_grow already freed obj */
3305 * A interface to enable slab creation on nodeid
3307 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3310 struct list_head
*entry
;
3312 struct kmem_list3
*l3
;
3316 l3
= cachep
->nodelists
[nodeid
];
3321 spin_lock(&l3
->list_lock
);
3322 entry
= l3
->slabs_partial
.next
;
3323 if (entry
== &l3
->slabs_partial
) {
3324 l3
->free_touched
= 1;
3325 entry
= l3
->slabs_free
.next
;
3326 if (entry
== &l3
->slabs_free
)
3330 slabp
= list_entry(entry
, struct slab
, list
);
3331 check_spinlock_acquired_node(cachep
, nodeid
);
3332 check_slabp(cachep
, slabp
);
3334 STATS_INC_NODEALLOCS(cachep
);
3335 STATS_INC_ACTIVE(cachep
);
3336 STATS_SET_HIGH(cachep
);
3338 BUG_ON(slabp
->inuse
== cachep
->num
);
3340 obj
= slab_get_obj(cachep
, slabp
, nodeid
);
3341 check_slabp(cachep
, slabp
);
3343 /* move slabp to correct slabp list: */
3344 list_del(&slabp
->list
);
3346 if (slabp
->free
== BUFCTL_END
)
3347 list_add(&slabp
->list
, &l3
->slabs_full
);
3349 list_add(&slabp
->list
, &l3
->slabs_partial
);
3351 spin_unlock(&l3
->list_lock
);
3355 spin_unlock(&l3
->list_lock
);
3356 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, nodeid
, NULL
);
3360 return fallback_alloc(cachep
, flags
);
3367 * kmem_cache_alloc_node - Allocate an object on the specified node
3368 * @cachep: The cache to allocate from.
3369 * @flags: See kmalloc().
3370 * @nodeid: node number of the target node.
3371 * @caller: return address of caller, used for debug information
3373 * Identical to kmem_cache_alloc but it will allocate memory on the given
3374 * node, which can improve the performance for cpu bound structures.
3376 * Fallback to other node is possible if __GFP_THISNODE is not set.
3378 static __always_inline
void *
3379 __cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3382 unsigned long save_flags
;
3384 int slab_node
= numa_mem_id();
3386 flags
&= gfp_allowed_mask
;
3388 lockdep_trace_alloc(flags
);
3390 if (slab_should_failslab(cachep
, flags
))
3393 cache_alloc_debugcheck_before(cachep
, flags
);
3394 local_irq_save(save_flags
);
3399 if (unlikely(!cachep
->nodelists
[nodeid
])) {
3400 /* Node not bootstrapped yet */
3401 ptr
= fallback_alloc(cachep
, flags
);
3405 if (nodeid
== slab_node
) {
3407 * Use the locally cached objects if possible.
3408 * However ____cache_alloc does not allow fallback
3409 * to other nodes. It may fail while we still have
3410 * objects on other nodes available.
3412 ptr
= ____cache_alloc(cachep
, flags
);
3416 /* ___cache_alloc_node can fall back to other nodes */
3417 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3419 local_irq_restore(save_flags
);
3420 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3421 kmemleak_alloc_recursive(ptr
, obj_size(cachep
), 1, cachep
->flags
,
3425 kmemcheck_slab_alloc(cachep
, flags
, ptr
, obj_size(cachep
));
3427 if (unlikely((flags
& __GFP_ZERO
) && ptr
))
3428 memset(ptr
, 0, obj_size(cachep
));
3433 static __always_inline
void *
3434 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3438 if (unlikely(current
->flags
& (PF_SPREAD_SLAB
| PF_MEMPOLICY
))) {
3439 objp
= alternate_node_alloc(cache
, flags
);
3443 objp
= ____cache_alloc(cache
, flags
);
3446 * We may just have run out of memory on the local node.
3447 * ____cache_alloc_node() knows how to locate memory on other nodes
3450 objp
= ____cache_alloc_node(cache
, flags
, numa_mem_id());
3457 static __always_inline
void *
3458 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3460 return ____cache_alloc(cachep
, flags
);
3463 #endif /* CONFIG_NUMA */
3465 static __always_inline
void *
3466 __cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
, void *caller
)
3468 unsigned long save_flags
;
3471 flags
&= gfp_allowed_mask
;
3473 lockdep_trace_alloc(flags
);
3475 if (slab_should_failslab(cachep
, flags
))
3478 cache_alloc_debugcheck_before(cachep
, flags
);
3479 local_irq_save(save_flags
);
3480 objp
= __do_cache_alloc(cachep
, flags
);
3481 local_irq_restore(save_flags
);
3482 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3483 kmemleak_alloc_recursive(objp
, obj_size(cachep
), 1, cachep
->flags
,
3488 kmemcheck_slab_alloc(cachep
, flags
, objp
, obj_size(cachep
));
3490 if (unlikely((flags
& __GFP_ZERO
) && objp
))
3491 memset(objp
, 0, obj_size(cachep
));
3497 * Caller needs to acquire correct kmem_list's list_lock
3499 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int nr_objects
,
3503 struct kmem_list3
*l3
;
3505 for (i
= 0; i
< nr_objects
; i
++) {
3506 void *objp
= objpp
[i
];
3509 slabp
= virt_to_slab(objp
);
3510 l3
= cachep
->nodelists
[node
];
3511 list_del(&slabp
->list
);
3512 check_spinlock_acquired_node(cachep
, node
);
3513 check_slabp(cachep
, slabp
);
3514 slab_put_obj(cachep
, slabp
, objp
, node
);
3515 STATS_DEC_ACTIVE(cachep
);
3517 check_slabp(cachep
, slabp
);
3519 /* fixup slab chains */
3520 if (slabp
->inuse
== 0) {
3521 if (l3
->free_objects
> l3
->free_limit
) {
3522 l3
->free_objects
-= cachep
->num
;
3523 /* No need to drop any previously held
3524 * lock here, even if we have a off-slab slab
3525 * descriptor it is guaranteed to come from
3526 * a different cache, refer to comments before
3529 slab_destroy(cachep
, slabp
);
3531 list_add(&slabp
->list
, &l3
->slabs_free
);
3534 /* Unconditionally move a slab to the end of the
3535 * partial list on free - maximum time for the
3536 * other objects to be freed, too.
3538 list_add_tail(&slabp
->list
, &l3
->slabs_partial
);
3543 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3546 struct kmem_list3
*l3
;
3547 int node
= numa_mem_id();
3549 batchcount
= ac
->batchcount
;
3551 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
3554 l3
= cachep
->nodelists
[node
];
3555 spin_lock(&l3
->list_lock
);
3557 struct array_cache
*shared_array
= l3
->shared
;
3558 int max
= shared_array
->limit
- shared_array
->avail
;
3560 if (batchcount
> max
)
3562 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3563 ac
->entry
, sizeof(void *) * batchcount
);
3564 shared_array
->avail
+= batchcount
;
3569 free_block(cachep
, ac
->entry
, batchcount
, node
);
3574 struct list_head
*p
;
3576 p
= l3
->slabs_free
.next
;
3577 while (p
!= &(l3
->slabs_free
)) {
3580 slabp
= list_entry(p
, struct slab
, list
);
3581 BUG_ON(slabp
->inuse
);
3586 STATS_SET_FREEABLE(cachep
, i
);
3589 spin_unlock(&l3
->list_lock
);
3590 ac
->avail
-= batchcount
;
3591 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3595 * Release an obj back to its cache. If the obj has a constructed state, it must
3596 * be in this state _before_ it is released. Called with disabled ints.
3598 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
)
3600 struct array_cache
*ac
= cpu_cache_get(cachep
);
3603 kmemleak_free_recursive(objp
, cachep
->flags
);
3604 objp
= cache_free_debugcheck(cachep
, objp
, __builtin_return_address(0));
3606 kmemcheck_slab_free(cachep
, objp
, obj_size(cachep
));
3609 * Skip calling cache_free_alien() when the platform is not numa.
3610 * This will avoid cache misses that happen while accessing slabp (which
3611 * is per page memory reference) to get nodeid. Instead use a global
3612 * variable to skip the call, which is mostly likely to be present in
3615 if (nr_online_nodes
> 1 && cache_free_alien(cachep
, objp
))
3618 if (likely(ac
->avail
< ac
->limit
)) {
3619 STATS_INC_FREEHIT(cachep
);
3620 ac
->entry
[ac
->avail
++] = objp
;
3623 STATS_INC_FREEMISS(cachep
);
3624 cache_flusharray(cachep
, ac
);
3625 ac
->entry
[ac
->avail
++] = objp
;
3630 * kmem_cache_alloc - Allocate an object
3631 * @cachep: The cache to allocate from.
3632 * @flags: See kmalloc().
3634 * Allocate an object from this cache. The flags are only relevant
3635 * if the cache has no available objects.
3637 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3639 void *ret
= __cache_alloc(cachep
, flags
, __builtin_return_address(0));
3641 trace_kmem_cache_alloc(_RET_IP_
, ret
,
3642 obj_size(cachep
), cachep
->buffer_size
, flags
);
3646 EXPORT_SYMBOL(kmem_cache_alloc
);
3648 #ifdef CONFIG_TRACING
3649 void *kmem_cache_alloc_notrace(struct kmem_cache
*cachep
, gfp_t flags
)
3651 return __cache_alloc(cachep
, flags
, __builtin_return_address(0));
3653 EXPORT_SYMBOL(kmem_cache_alloc_notrace
);
3657 * kmem_ptr_validate - check if an untrusted pointer might be a slab entry.
3658 * @cachep: the cache we're checking against
3659 * @ptr: pointer to validate
3661 * This verifies that the untrusted pointer looks sane;
3662 * it is _not_ a guarantee that the pointer is actually
3663 * part of the slab cache in question, but it at least
3664 * validates that the pointer can be dereferenced and
3665 * looks half-way sane.
3667 * Currently only used for dentry validation.
3669 int kmem_ptr_validate(struct kmem_cache
*cachep
, const void *ptr
)
3671 unsigned long size
= cachep
->buffer_size
;
3674 if (unlikely(!kern_ptr_validate(ptr
, size
)))
3676 page
= virt_to_page(ptr
);
3677 if (unlikely(!PageSlab(page
)))
3679 if (unlikely(page_get_cache(page
) != cachep
))
3687 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3689 void *ret
= __cache_alloc_node(cachep
, flags
, nodeid
,
3690 __builtin_return_address(0));
3692 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
3693 obj_size(cachep
), cachep
->buffer_size
,
3698 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3700 #ifdef CONFIG_TRACING
3701 void *kmem_cache_alloc_node_notrace(struct kmem_cache
*cachep
,
3705 return __cache_alloc_node(cachep
, flags
, nodeid
,
3706 __builtin_return_address(0));
3708 EXPORT_SYMBOL(kmem_cache_alloc_node_notrace
);
3711 static __always_inline
void *
3712 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, void *caller
)
3714 struct kmem_cache
*cachep
;
3717 cachep
= kmem_find_general_cachep(size
, flags
);
3718 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3720 ret
= kmem_cache_alloc_node_notrace(cachep
, flags
, node
);
3722 trace_kmalloc_node((unsigned long) caller
, ret
,
3723 size
, cachep
->buffer_size
, flags
, node
);
3728 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3729 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3731 return __do_kmalloc_node(size
, flags
, node
,
3732 __builtin_return_address(0));
3734 EXPORT_SYMBOL(__kmalloc_node
);
3736 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3737 int node
, unsigned long caller
)
3739 return __do_kmalloc_node(size
, flags
, node
, (void *)caller
);
3741 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3743 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3745 return __do_kmalloc_node(size
, flags
, node
, NULL
);
3747 EXPORT_SYMBOL(__kmalloc_node
);
3748 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3749 #endif /* CONFIG_NUMA */
3752 * __do_kmalloc - allocate memory
3753 * @size: how many bytes of memory are required.
3754 * @flags: the type of memory to allocate (see kmalloc).
3755 * @caller: function caller for debug tracking of the caller
3757 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3760 struct kmem_cache
*cachep
;
3763 /* If you want to save a few bytes .text space: replace
3765 * Then kmalloc uses the uninlined functions instead of the inline
3768 cachep
= __find_general_cachep(size
, flags
);
3769 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3771 ret
= __cache_alloc(cachep
, flags
, caller
);
3773 trace_kmalloc((unsigned long) caller
, ret
,
3774 size
, cachep
->buffer_size
, flags
);
3780 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3781 void *__kmalloc(size_t size
, gfp_t flags
)
3783 return __do_kmalloc(size
, flags
, __builtin_return_address(0));
3785 EXPORT_SYMBOL(__kmalloc
);
3787 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, unsigned long caller
)
3789 return __do_kmalloc(size
, flags
, (void *)caller
);
3791 EXPORT_SYMBOL(__kmalloc_track_caller
);
3794 void *__kmalloc(size_t size
, gfp_t flags
)
3796 return __do_kmalloc(size
, flags
, NULL
);
3798 EXPORT_SYMBOL(__kmalloc
);
3802 * kmem_cache_free - Deallocate an object
3803 * @cachep: The cache the allocation was from.
3804 * @objp: The previously allocated object.
3806 * Free an object which was previously allocated from this
3809 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3811 unsigned long flags
;
3813 local_irq_save(flags
);
3814 debug_check_no_locks_freed(objp
, obj_size(cachep
));
3815 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3816 debug_check_no_obj_freed(objp
, obj_size(cachep
));
3817 __cache_free(cachep
, objp
);
3818 local_irq_restore(flags
);
3820 trace_kmem_cache_free(_RET_IP_
, objp
);
3822 EXPORT_SYMBOL(kmem_cache_free
);
3825 * kfree - free previously allocated memory
3826 * @objp: pointer returned by kmalloc.
3828 * If @objp is NULL, no operation is performed.
3830 * Don't free memory not originally allocated by kmalloc()
3831 * or you will run into trouble.
3833 void kfree(const void *objp
)
3835 struct kmem_cache
*c
;
3836 unsigned long flags
;
3838 trace_kfree(_RET_IP_
, objp
);
3840 if (unlikely(ZERO_OR_NULL_PTR(objp
)))
3842 local_irq_save(flags
);
3843 kfree_debugcheck(objp
);
3844 c
= virt_to_cache(objp
);
3845 debug_check_no_locks_freed(objp
, obj_size(c
));
3846 debug_check_no_obj_freed(objp
, obj_size(c
));
3847 __cache_free(c
, (void *)objp
);
3848 local_irq_restore(flags
);
3850 EXPORT_SYMBOL(kfree
);
3852 unsigned int kmem_cache_size(struct kmem_cache
*cachep
)
3854 return obj_size(cachep
);
3856 EXPORT_SYMBOL(kmem_cache_size
);
3858 const char *kmem_cache_name(struct kmem_cache
*cachep
)
3860 return cachep
->name
;
3862 EXPORT_SYMBOL_GPL(kmem_cache_name
);
3865 * This initializes kmem_list3 or resizes various caches for all nodes.
3867 static int alloc_kmemlist(struct kmem_cache
*cachep
, gfp_t gfp
)
3870 struct kmem_list3
*l3
;
3871 struct array_cache
*new_shared
;
3872 struct array_cache
**new_alien
= NULL
;
3874 for_each_online_node(node
) {
3876 if (use_alien_caches
) {
3877 new_alien
= alloc_alien_cache(node
, cachep
->limit
, gfp
);
3883 if (cachep
->shared
) {
3884 new_shared
= alloc_arraycache(node
,
3885 cachep
->shared
*cachep
->batchcount
,
3888 free_alien_cache(new_alien
);
3893 l3
= cachep
->nodelists
[node
];
3895 struct array_cache
*shared
= l3
->shared
;
3897 spin_lock_irq(&l3
->list_lock
);
3900 free_block(cachep
, shared
->entry
,
3901 shared
->avail
, node
);
3903 l3
->shared
= new_shared
;
3905 l3
->alien
= new_alien
;
3908 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3909 cachep
->batchcount
+ cachep
->num
;
3910 spin_unlock_irq(&l3
->list_lock
);
3912 free_alien_cache(new_alien
);
3915 l3
= kmalloc_node(sizeof(struct kmem_list3
), gfp
, node
);
3917 free_alien_cache(new_alien
);
3922 kmem_list3_init(l3
);
3923 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
3924 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
3925 l3
->shared
= new_shared
;
3926 l3
->alien
= new_alien
;
3927 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3928 cachep
->batchcount
+ cachep
->num
;
3929 cachep
->nodelists
[node
] = l3
;
3934 if (!cachep
->next
.next
) {
3935 /* Cache is not active yet. Roll back what we did */
3938 if (cachep
->nodelists
[node
]) {
3939 l3
= cachep
->nodelists
[node
];
3942 free_alien_cache(l3
->alien
);
3944 cachep
->nodelists
[node
] = NULL
;
3952 struct ccupdate_struct
{
3953 struct kmem_cache
*cachep
;
3954 struct array_cache
*new[NR_CPUS
];
3957 static void do_ccupdate_local(void *info
)
3959 struct ccupdate_struct
*new = info
;
3960 struct array_cache
*old
;
3963 old
= cpu_cache_get(new->cachep
);
3965 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
3966 new->new[smp_processor_id()] = old
;
3969 /* Always called with the cache_chain_mutex held */
3970 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3971 int batchcount
, int shared
, gfp_t gfp
)
3973 struct ccupdate_struct
*new;
3976 new = kzalloc(sizeof(*new), gfp
);
3980 for_each_online_cpu(i
) {
3981 new->new[i
] = alloc_arraycache(cpu_to_mem(i
), limit
,
3984 for (i
--; i
>= 0; i
--)
3990 new->cachep
= cachep
;
3992 on_each_cpu(do_ccupdate_local
, (void *)new, 1);
3995 cachep
->batchcount
= batchcount
;
3996 cachep
->limit
= limit
;
3997 cachep
->shared
= shared
;
3999 for_each_online_cpu(i
) {
4000 struct array_cache
*ccold
= new->new[i
];
4003 spin_lock_irq(&cachep
->nodelists
[cpu_to_mem(i
)]->list_lock
);
4004 free_block(cachep
, ccold
->entry
, ccold
->avail
, cpu_to_mem(i
));
4005 spin_unlock_irq(&cachep
->nodelists
[cpu_to_mem(i
)]->list_lock
);
4009 return alloc_kmemlist(cachep
, gfp
);
4012 /* Called with cache_chain_mutex held always */
4013 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
4019 * The head array serves three purposes:
4020 * - create a LIFO ordering, i.e. return objects that are cache-warm
4021 * - reduce the number of spinlock operations.
4022 * - reduce the number of linked list operations on the slab and
4023 * bufctl chains: array operations are cheaper.
4024 * The numbers are guessed, we should auto-tune as described by
4027 if (cachep
->buffer_size
> 131072)
4029 else if (cachep
->buffer_size
> PAGE_SIZE
)
4031 else if (cachep
->buffer_size
> 1024)
4033 else if (cachep
->buffer_size
> 256)
4039 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4040 * allocation behaviour: Most allocs on one cpu, most free operations
4041 * on another cpu. For these cases, an efficient object passing between
4042 * cpus is necessary. This is provided by a shared array. The array
4043 * replaces Bonwick's magazine layer.
4044 * On uniprocessor, it's functionally equivalent (but less efficient)
4045 * to a larger limit. Thus disabled by default.
4048 if (cachep
->buffer_size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
4053 * With debugging enabled, large batchcount lead to excessively long
4054 * periods with disabled local interrupts. Limit the batchcount
4059 err
= do_tune_cpucache(cachep
, limit
, (limit
+ 1) / 2, shared
, gfp
);
4061 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
4062 cachep
->name
, -err
);
4067 * Drain an array if it contains any elements taking the l3 lock only if
4068 * necessary. Note that the l3 listlock also protects the array_cache
4069 * if drain_array() is used on the shared array.
4071 void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
4072 struct array_cache
*ac
, int force
, int node
)
4076 if (!ac
|| !ac
->avail
)
4078 if (ac
->touched
&& !force
) {
4081 spin_lock_irq(&l3
->list_lock
);
4083 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
4084 if (tofree
> ac
->avail
)
4085 tofree
= (ac
->avail
+ 1) / 2;
4086 free_block(cachep
, ac
->entry
, tofree
, node
);
4087 ac
->avail
-= tofree
;
4088 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
4089 sizeof(void *) * ac
->avail
);
4091 spin_unlock_irq(&l3
->list_lock
);
4096 * cache_reap - Reclaim memory from caches.
4097 * @w: work descriptor
4099 * Called from workqueue/eventd every few seconds.
4101 * - clear the per-cpu caches for this CPU.
4102 * - return freeable pages to the main free memory pool.
4104 * If we cannot acquire the cache chain mutex then just give up - we'll try
4105 * again on the next iteration.
4107 static void cache_reap(struct work_struct
*w
)
4109 struct kmem_cache
*searchp
;
4110 struct kmem_list3
*l3
;
4111 int node
= numa_mem_id();
4112 struct delayed_work
*work
= to_delayed_work(w
);
4114 if (!mutex_trylock(&cache_chain_mutex
))
4115 /* Give up. Setup the next iteration. */
4118 list_for_each_entry(searchp
, &cache_chain
, next
) {
4122 * We only take the l3 lock if absolutely necessary and we
4123 * have established with reasonable certainty that
4124 * we can do some work if the lock was obtained.
4126 l3
= searchp
->nodelists
[node
];
4128 reap_alien(searchp
, l3
);
4130 drain_array(searchp
, l3
, cpu_cache_get(searchp
), 0, node
);
4133 * These are racy checks but it does not matter
4134 * if we skip one check or scan twice.
4136 if (time_after(l3
->next_reap
, jiffies
))
4139 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
;
4141 drain_array(searchp
, l3
, l3
->shared
, 0, node
);
4143 if (l3
->free_touched
)
4144 l3
->free_touched
= 0;
4148 freed
= drain_freelist(searchp
, l3
, (l3
->free_limit
+
4149 5 * searchp
->num
- 1) / (5 * searchp
->num
));
4150 STATS_ADD_REAPED(searchp
, freed
);
4156 mutex_unlock(&cache_chain_mutex
);
4159 /* Set up the next iteration */
4160 schedule_delayed_work(work
, round_jiffies_relative(REAPTIMEOUT_CPUC
));
4163 #ifdef CONFIG_SLABINFO
4165 static void print_slabinfo_header(struct seq_file
*m
)
4168 * Output format version, so at least we can change it
4169 * without _too_ many complaints.
4172 seq_puts(m
, "slabinfo - version: 2.1 (statistics)\n");
4174 seq_puts(m
, "slabinfo - version: 2.1\n");
4176 seq_puts(m
, "# name <active_objs> <num_objs> <objsize> "
4177 "<objperslab> <pagesperslab>");
4178 seq_puts(m
, " : tunables <limit> <batchcount> <sharedfactor>");
4179 seq_puts(m
, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4181 seq_puts(m
, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4182 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4183 seq_puts(m
, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4188 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
4192 mutex_lock(&cache_chain_mutex
);
4194 print_slabinfo_header(m
);
4196 return seq_list_start(&cache_chain
, *pos
);
4199 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
4201 return seq_list_next(p
, &cache_chain
, pos
);
4204 static void s_stop(struct seq_file
*m
, void *p
)
4206 mutex_unlock(&cache_chain_mutex
);
4209 static int s_show(struct seq_file
*m
, void *p
)
4211 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, next
);
4213 unsigned long active_objs
;
4214 unsigned long num_objs
;
4215 unsigned long active_slabs
= 0;
4216 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
4220 struct kmem_list3
*l3
;
4224 for_each_online_node(node
) {
4225 l3
= cachep
->nodelists
[node
];
4230 spin_lock_irq(&l3
->list_lock
);
4232 list_for_each_entry(slabp
, &l3
->slabs_full
, list
) {
4233 if (slabp
->inuse
!= cachep
->num
&& !error
)
4234 error
= "slabs_full accounting error";
4235 active_objs
+= cachep
->num
;
4238 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
) {
4239 if (slabp
->inuse
== cachep
->num
&& !error
)
4240 error
= "slabs_partial inuse accounting error";
4241 if (!slabp
->inuse
&& !error
)
4242 error
= "slabs_partial/inuse accounting error";
4243 active_objs
+= slabp
->inuse
;
4246 list_for_each_entry(slabp
, &l3
->slabs_free
, list
) {
4247 if (slabp
->inuse
&& !error
)
4248 error
= "slabs_free/inuse accounting error";
4251 free_objects
+= l3
->free_objects
;
4253 shared_avail
+= l3
->shared
->avail
;
4255 spin_unlock_irq(&l3
->list_lock
);
4257 num_slabs
+= active_slabs
;
4258 num_objs
= num_slabs
* cachep
->num
;
4259 if (num_objs
- active_objs
!= free_objects
&& !error
)
4260 error
= "free_objects accounting error";
4262 name
= cachep
->name
;
4264 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
4266 seq_printf(m
, "%-17s %6lu %6lu %6u %4u %4d",
4267 name
, active_objs
, num_objs
, cachep
->buffer_size
,
4268 cachep
->num
, (1 << cachep
->gfporder
));
4269 seq_printf(m
, " : tunables %4u %4u %4u",
4270 cachep
->limit
, cachep
->batchcount
, cachep
->shared
);
4271 seq_printf(m
, " : slabdata %6lu %6lu %6lu",
4272 active_slabs
, num_slabs
, shared_avail
);
4275 unsigned long high
= cachep
->high_mark
;
4276 unsigned long allocs
= cachep
->num_allocations
;
4277 unsigned long grown
= cachep
->grown
;
4278 unsigned long reaped
= cachep
->reaped
;
4279 unsigned long errors
= cachep
->errors
;
4280 unsigned long max_freeable
= cachep
->max_freeable
;
4281 unsigned long node_allocs
= cachep
->node_allocs
;
4282 unsigned long node_frees
= cachep
->node_frees
;
4283 unsigned long overflows
= cachep
->node_overflow
;
4285 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu "
4286 "%4lu %4lu %4lu %4lu %4lu",
4287 allocs
, high
, grown
,
4288 reaped
, errors
, max_freeable
, node_allocs
,
4289 node_frees
, overflows
);
4293 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4294 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4295 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4296 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4298 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4299 allochit
, allocmiss
, freehit
, freemiss
);
4307 * slabinfo_op - iterator that generates /proc/slabinfo
4316 * num-pages-per-slab
4317 * + further values on SMP and with statistics enabled
4320 static const struct seq_operations slabinfo_op
= {
4327 #define MAX_SLABINFO_WRITE 128
4329 * slabinfo_write - Tuning for the slab allocator
4331 * @buffer: user buffer
4332 * @count: data length
4335 ssize_t
slabinfo_write(struct file
*file
, const char __user
* buffer
,
4336 size_t count
, loff_t
*ppos
)
4338 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4339 int limit
, batchcount
, shared
, res
;
4340 struct kmem_cache
*cachep
;
4342 if (count
> MAX_SLABINFO_WRITE
)
4344 if (copy_from_user(&kbuf
, buffer
, count
))
4346 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4348 tmp
= strchr(kbuf
, ' ');
4353 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4356 /* Find the cache in the chain of caches. */
4357 mutex_lock(&cache_chain_mutex
);
4359 list_for_each_entry(cachep
, &cache_chain
, next
) {
4360 if (!strcmp(cachep
->name
, kbuf
)) {
4361 if (limit
< 1 || batchcount
< 1 ||
4362 batchcount
> limit
|| shared
< 0) {
4365 res
= do_tune_cpucache(cachep
, limit
,
4372 mutex_unlock(&cache_chain_mutex
);
4378 static int slabinfo_open(struct inode
*inode
, struct file
*file
)
4380 return seq_open(file
, &slabinfo_op
);
4383 static const struct file_operations proc_slabinfo_operations
= {
4384 .open
= slabinfo_open
,
4386 .write
= slabinfo_write
,
4387 .llseek
= seq_lseek
,
4388 .release
= seq_release
,
4391 #ifdef CONFIG_DEBUG_SLAB_LEAK
4393 static void *leaks_start(struct seq_file
*m
, loff_t
*pos
)
4395 mutex_lock(&cache_chain_mutex
);
4396 return seq_list_start(&cache_chain
, *pos
);
4399 static inline int add_caller(unsigned long *n
, unsigned long v
)
4409 unsigned long *q
= p
+ 2 * i
;
4423 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4429 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
, struct slab
*s
)
4435 for (i
= 0, p
= s
->s_mem
; i
< c
->num
; i
++, p
+= c
->buffer_size
) {
4436 if (slab_bufctl(s
)[i
] != BUFCTL_ACTIVE
)
4438 if (!add_caller(n
, (unsigned long)*dbg_userword(c
, p
)))
4443 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4445 #ifdef CONFIG_KALLSYMS
4446 unsigned long offset
, size
;
4447 char modname
[MODULE_NAME_LEN
], name
[KSYM_NAME_LEN
];
4449 if (lookup_symbol_attrs(address
, &size
, &offset
, modname
, name
) == 0) {
4450 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4452 seq_printf(m
, " [%s]", modname
);
4456 seq_printf(m
, "%p", (void *)address
);
4459 static int leaks_show(struct seq_file
*m
, void *p
)
4461 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, next
);
4463 struct kmem_list3
*l3
;
4465 unsigned long *n
= m
->private;
4469 if (!(cachep
->flags
& SLAB_STORE_USER
))
4471 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4474 /* OK, we can do it */
4478 for_each_online_node(node
) {
4479 l3
= cachep
->nodelists
[node
];
4484 spin_lock_irq(&l3
->list_lock
);
4486 list_for_each_entry(slabp
, &l3
->slabs_full
, list
)
4487 handle_slab(n
, cachep
, slabp
);
4488 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
)
4489 handle_slab(n
, cachep
, slabp
);
4490 spin_unlock_irq(&l3
->list_lock
);
4492 name
= cachep
->name
;
4494 /* Increase the buffer size */
4495 mutex_unlock(&cache_chain_mutex
);
4496 m
->private = kzalloc(n
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4498 /* Too bad, we are really out */
4500 mutex_lock(&cache_chain_mutex
);
4503 *(unsigned long *)m
->private = n
[0] * 2;
4505 mutex_lock(&cache_chain_mutex
);
4506 /* Now make sure this entry will be retried */
4510 for (i
= 0; i
< n
[1]; i
++) {
4511 seq_printf(m
, "%s: %lu ", name
, n
[2*i
+3]);
4512 show_symbol(m
, n
[2*i
+2]);
4519 static const struct seq_operations slabstats_op
= {
4520 .start
= leaks_start
,
4526 static int slabstats_open(struct inode
*inode
, struct file
*file
)
4528 unsigned long *n
= kzalloc(PAGE_SIZE
, GFP_KERNEL
);
4531 ret
= seq_open(file
, &slabstats_op
);
4533 struct seq_file
*m
= file
->private_data
;
4534 *n
= PAGE_SIZE
/ (2 * sizeof(unsigned long));
4543 static const struct file_operations proc_slabstats_operations
= {
4544 .open
= slabstats_open
,
4546 .llseek
= seq_lseek
,
4547 .release
= seq_release_private
,
4551 static int __init
slab_proc_init(void)
4553 proc_create("slabinfo",S_IWUSR
|S_IRUGO
,NULL
,&proc_slabinfo_operations
);
4554 #ifdef CONFIG_DEBUG_SLAB_LEAK
4555 proc_create("slab_allocators", 0, NULL
, &proc_slabstats_operations
);
4559 module_init(slab_proc_init
);
4563 * ksize - get the actual amount of memory allocated for a given object
4564 * @objp: Pointer to the object
4566 * kmalloc may internally round up allocations and return more memory
4567 * than requested. ksize() can be used to determine the actual amount of
4568 * memory allocated. The caller may use this additional memory, even though
4569 * a smaller amount of memory was initially specified with the kmalloc call.
4570 * The caller must guarantee that objp points to a valid object previously
4571 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4572 * must not be freed during the duration of the call.
4574 size_t ksize(const void *objp
)
4577 if (unlikely(objp
== ZERO_SIZE_PTR
))
4580 return obj_size(virt_to_cache(objp
));
4582 EXPORT_SYMBOL(ksize
);