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>
118 #include <linux/prefetch.h>
120 #include <asm/cacheflush.h>
121 #include <asm/tlbflush.h>
122 #include <asm/page.h>
124 #include <trace/events/kmem.h>
127 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
128 * 0 for faster, smaller code (especially in the critical paths).
130 * STATS - 1 to collect stats for /proc/slabinfo.
131 * 0 for faster, smaller code (especially in the critical paths).
133 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
136 #ifdef CONFIG_DEBUG_SLAB
139 #define FORCED_DEBUG 1
143 #define FORCED_DEBUG 0
146 /* Shouldn't this be in a header file somewhere? */
147 #define BYTES_PER_WORD sizeof(void *)
148 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
150 #ifndef ARCH_KMALLOC_FLAGS
151 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
154 /* Legal flag mask for kmem_cache_create(). */
156 # define CREATE_MASK (SLAB_RED_ZONE | \
157 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
160 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
161 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
162 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
164 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
166 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
167 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
168 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
174 * Bufctl's are used for linking objs within a slab
177 * This implementation relies on "struct page" for locating the cache &
178 * slab an object belongs to.
179 * This allows the bufctl structure to be small (one int), but limits
180 * the number of objects a slab (not a cache) can contain when off-slab
181 * bufctls are used. The limit is the size of the largest general cache
182 * that does not use off-slab slabs.
183 * For 32bit archs with 4 kB pages, is this 56.
184 * This is not serious, as it is only for large objects, when it is unwise
185 * to have too many per slab.
186 * Note: This limit can be raised by introducing a general cache whose size
187 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
190 typedef unsigned int kmem_bufctl_t
;
191 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
192 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
193 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
194 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
199 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
200 * arrange for kmem_freepages to be called via RCU. This is useful if
201 * we need to approach a kernel structure obliquely, from its address
202 * obtained without the usual locking. We can lock the structure to
203 * stabilize it and check it's still at the given address, only if we
204 * can be sure that the memory has not been meanwhile reused for some
205 * other kind of object (which our subsystem's lock might corrupt).
207 * rcu_read_lock before reading the address, then rcu_read_unlock after
208 * taking the spinlock within the structure expected at that address.
211 struct rcu_head head
;
212 struct kmem_cache
*cachep
;
219 * Manages the objs in a slab. Placed either at the beginning of mem allocated
220 * for a slab, or allocated from an general cache.
221 * Slabs are chained into three list: fully used, partial, fully free slabs.
226 struct list_head list
;
227 unsigned long colouroff
;
228 void *s_mem
; /* including colour offset */
229 unsigned int inuse
; /* num of objs active in slab */
231 unsigned short nodeid
;
233 struct slab_rcu __slab_cover_slab_rcu
;
241 * - LIFO ordering, to hand out cache-warm objects from _alloc
242 * - reduce the number of linked list operations
243 * - reduce spinlock operations
245 * The limit is stored in the per-cpu structure to reduce the data cache
252 unsigned int batchcount
;
253 unsigned int touched
;
256 * Must have this definition in here for the proper
257 * alignment of array_cache. Also simplifies accessing
263 * bootstrap: The caches do not work without cpuarrays anymore, but the
264 * cpuarrays are allocated from the generic caches...
266 #define BOOT_CPUCACHE_ENTRIES 1
267 struct arraycache_init
{
268 struct array_cache cache
;
269 void *entries
[BOOT_CPUCACHE_ENTRIES
];
273 * The slab lists for all objects.
276 struct list_head slabs_partial
; /* partial list first, better asm code */
277 struct list_head slabs_full
;
278 struct list_head slabs_free
;
279 unsigned long free_objects
;
280 unsigned int free_limit
;
281 unsigned int colour_next
; /* Per-node cache coloring */
282 spinlock_t list_lock
;
283 struct array_cache
*shared
; /* shared per node */
284 struct array_cache
**alien
; /* on other nodes */
285 unsigned long next_reap
; /* updated without locking */
286 int free_touched
; /* updated without locking */
290 * Need this for bootstrapping a per node allocator.
292 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
293 static struct kmem_list3 __initdata initkmem_list3
[NUM_INIT_LISTS
];
294 #define CACHE_CACHE 0
295 #define SIZE_AC MAX_NUMNODES
296 #define SIZE_L3 (2 * MAX_NUMNODES)
298 static int drain_freelist(struct kmem_cache
*cache
,
299 struct kmem_list3
*l3
, int tofree
);
300 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
302 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
);
303 static void cache_reap(struct work_struct
*unused
);
306 * This function must be completely optimized away if a constant is passed to
307 * it. Mostly the same as what is in linux/slab.h except it returns an index.
309 static __always_inline
int index_of(const size_t size
)
311 extern void __bad_size(void);
313 if (__builtin_constant_p(size
)) {
321 #include <linux/kmalloc_sizes.h>
329 static int slab_early_init
= 1;
331 #define INDEX_AC index_of(sizeof(struct arraycache_init))
332 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
334 static void kmem_list3_init(struct kmem_list3
*parent
)
336 INIT_LIST_HEAD(&parent
->slabs_full
);
337 INIT_LIST_HEAD(&parent
->slabs_partial
);
338 INIT_LIST_HEAD(&parent
->slabs_free
);
339 parent
->shared
= NULL
;
340 parent
->alien
= NULL
;
341 parent
->colour_next
= 0;
342 spin_lock_init(&parent
->list_lock
);
343 parent
->free_objects
= 0;
344 parent
->free_touched
= 0;
347 #define MAKE_LIST(cachep, listp, slab, nodeid) \
349 INIT_LIST_HEAD(listp); \
350 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
353 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
355 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
356 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
357 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
360 #define CFLGS_OFF_SLAB (0x80000000UL)
361 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
363 #define BATCHREFILL_LIMIT 16
365 * Optimization question: fewer reaps means less probability for unnessary
366 * cpucache drain/refill cycles.
368 * OTOH the cpuarrays can contain lots of objects,
369 * which could lock up otherwise freeable slabs.
371 #define REAPTIMEOUT_CPUC (2*HZ)
372 #define REAPTIMEOUT_LIST3 (4*HZ)
375 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
376 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
377 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
378 #define STATS_INC_GROWN(x) ((x)->grown++)
379 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
380 #define STATS_SET_HIGH(x) \
382 if ((x)->num_active > (x)->high_mark) \
383 (x)->high_mark = (x)->num_active; \
385 #define STATS_INC_ERR(x) ((x)->errors++)
386 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
387 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
388 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
389 #define STATS_SET_FREEABLE(x, i) \
391 if ((x)->max_freeable < i) \
392 (x)->max_freeable = i; \
394 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
395 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
396 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
397 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
399 #define STATS_INC_ACTIVE(x) do { } while (0)
400 #define STATS_DEC_ACTIVE(x) do { } while (0)
401 #define STATS_INC_ALLOCED(x) do { } while (0)
402 #define STATS_INC_GROWN(x) do { } while (0)
403 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
404 #define STATS_SET_HIGH(x) do { } while (0)
405 #define STATS_INC_ERR(x) do { } while (0)
406 #define STATS_INC_NODEALLOCS(x) do { } while (0)
407 #define STATS_INC_NODEFREES(x) do { } while (0)
408 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
409 #define STATS_SET_FREEABLE(x, i) do { } while (0)
410 #define STATS_INC_ALLOCHIT(x) do { } while (0)
411 #define STATS_INC_ALLOCMISS(x) do { } while (0)
412 #define STATS_INC_FREEHIT(x) do { } while (0)
413 #define STATS_INC_FREEMISS(x) do { } while (0)
419 * memory layout of objects:
421 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
422 * the end of an object is aligned with the end of the real
423 * allocation. Catches writes behind the end of the allocation.
424 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
426 * cachep->obj_offset: The real object.
427 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
428 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
429 * [BYTES_PER_WORD long]
431 static int obj_offset(struct kmem_cache
*cachep
)
433 return cachep
->obj_offset
;
436 static int obj_size(struct kmem_cache
*cachep
)
438 return cachep
->obj_size
;
441 static unsigned long long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
443 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
444 return (unsigned long long*) (objp
+ obj_offset(cachep
) -
445 sizeof(unsigned long long));
448 static unsigned long long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
450 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
451 if (cachep
->flags
& SLAB_STORE_USER
)
452 return (unsigned long long *)(objp
+ cachep
->buffer_size
-
453 sizeof(unsigned long long) -
455 return (unsigned long long *) (objp
+ cachep
->buffer_size
-
456 sizeof(unsigned long long));
459 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
461 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
462 return (void **)(objp
+ cachep
->buffer_size
- BYTES_PER_WORD
);
467 #define obj_offset(x) 0
468 #define obj_size(cachep) (cachep->buffer_size)
469 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
470 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
471 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
475 #ifdef CONFIG_TRACING
476 size_t slab_buffer_size(struct kmem_cache
*cachep
)
478 return cachep
->buffer_size
;
480 EXPORT_SYMBOL(slab_buffer_size
);
484 * Do not go above this order unless 0 objects fit into the slab.
486 #define BREAK_GFP_ORDER_HI 1
487 #define BREAK_GFP_ORDER_LO 0
488 static int slab_break_gfp_order
= BREAK_GFP_ORDER_LO
;
491 * Functions for storing/retrieving the cachep and or slab from the page
492 * allocator. These are used to find the slab an obj belongs to. With kfree(),
493 * these are used to find the cache which an obj belongs to.
495 static inline void page_set_cache(struct page
*page
, struct kmem_cache
*cache
)
497 page
->lru
.next
= (struct list_head
*)cache
;
500 static inline struct kmem_cache
*page_get_cache(struct page
*page
)
502 page
= compound_head(page
);
503 BUG_ON(!PageSlab(page
));
504 return (struct kmem_cache
*)page
->lru
.next
;
507 static inline void page_set_slab(struct page
*page
, struct slab
*slab
)
509 page
->lru
.prev
= (struct list_head
*)slab
;
512 static inline struct slab
*page_get_slab(struct page
*page
)
514 BUG_ON(!PageSlab(page
));
515 return (struct slab
*)page
->lru
.prev
;
518 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
520 struct page
*page
= virt_to_head_page(obj
);
521 return page_get_cache(page
);
524 static inline struct slab
*virt_to_slab(const void *obj
)
526 struct page
*page
= virt_to_head_page(obj
);
527 return page_get_slab(page
);
530 static inline void *index_to_obj(struct kmem_cache
*cache
, struct slab
*slab
,
533 return slab
->s_mem
+ cache
->buffer_size
* idx
;
537 * We want to avoid an expensive divide : (offset / cache->buffer_size)
538 * Using the fact that buffer_size is a constant for a particular cache,
539 * we can replace (offset / cache->buffer_size) by
540 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
542 static inline unsigned int obj_to_index(const struct kmem_cache
*cache
,
543 const struct slab
*slab
, void *obj
)
545 u32 offset
= (obj
- slab
->s_mem
);
546 return reciprocal_divide(offset
, cache
->reciprocal_buffer_size
);
550 * These are the default caches for kmalloc. Custom caches can have other sizes.
552 struct cache_sizes malloc_sizes
[] = {
553 #define CACHE(x) { .cs_size = (x) },
554 #include <linux/kmalloc_sizes.h>
558 EXPORT_SYMBOL(malloc_sizes
);
560 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
566 static struct cache_names __initdata cache_names
[] = {
567 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
568 #include <linux/kmalloc_sizes.h>
573 static struct arraycache_init initarray_cache __initdata
=
574 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
575 static struct arraycache_init initarray_generic
=
576 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
578 /* internal cache of cache description objs */
579 static struct kmem_list3
*cache_cache_nodelists
[MAX_NUMNODES
];
580 static struct kmem_cache cache_cache
= {
581 .nodelists
= cache_cache_nodelists
,
583 .limit
= BOOT_CPUCACHE_ENTRIES
,
585 .buffer_size
= sizeof(struct kmem_cache
),
586 .name
= "kmem_cache",
589 #define BAD_ALIEN_MAGIC 0x01020304ul
592 * chicken and egg problem: delay the per-cpu array allocation
593 * until the general caches are up.
605 * used by boot code to determine if it can use slab based allocator
607 int slab_is_available(void)
609 return g_cpucache_up
>= EARLY
;
612 #ifdef CONFIG_LOCKDEP
615 * Slab sometimes uses the kmalloc slabs to store the slab headers
616 * for other slabs "off slab".
617 * The locking for this is tricky in that it nests within the locks
618 * of all other slabs in a few places; to deal with this special
619 * locking we put on-slab caches into a separate lock-class.
621 * We set lock class for alien array caches which are up during init.
622 * The lock annotation will be lost if all cpus of a node goes down and
623 * then comes back up during hotplug
625 static struct lock_class_key on_slab_l3_key
;
626 static struct lock_class_key on_slab_alc_key
;
628 static struct lock_class_key debugobj_l3_key
;
629 static struct lock_class_key debugobj_alc_key
;
631 static void slab_set_lock_classes(struct kmem_cache
*cachep
,
632 struct lock_class_key
*l3_key
, struct lock_class_key
*alc_key
,
635 struct array_cache
**alc
;
636 struct kmem_list3
*l3
;
639 l3
= cachep
->nodelists
[q
];
643 lockdep_set_class(&l3
->list_lock
, l3_key
);
646 * FIXME: This check for BAD_ALIEN_MAGIC
647 * should go away when common slab code is taught to
648 * work even without alien caches.
649 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
650 * for alloc_alien_cache,
652 if (!alc
|| (unsigned long)alc
== BAD_ALIEN_MAGIC
)
656 lockdep_set_class(&alc
[r
]->lock
, alc_key
);
660 static void slab_set_debugobj_lock_classes_node(struct kmem_cache
*cachep
, int node
)
662 slab_set_lock_classes(cachep
, &debugobj_l3_key
, &debugobj_alc_key
, node
);
665 static void slab_set_debugobj_lock_classes(struct kmem_cache
*cachep
)
669 for_each_online_node(node
)
670 slab_set_debugobj_lock_classes_node(cachep
, node
);
673 static void init_node_lock_keys(int q
)
675 struct cache_sizes
*s
= malloc_sizes
;
677 if (g_cpucache_up
< LATE
)
680 for (s
= malloc_sizes
; s
->cs_size
!= ULONG_MAX
; s
++) {
681 struct kmem_list3
*l3
;
683 l3
= s
->cs_cachep
->nodelists
[q
];
684 if (!l3
|| OFF_SLAB(s
->cs_cachep
))
687 slab_set_lock_classes(s
->cs_cachep
, &on_slab_l3_key
,
688 &on_slab_alc_key
, q
);
692 static inline void init_lock_keys(void)
697 init_node_lock_keys(node
);
700 static void init_node_lock_keys(int q
)
704 static inline void init_lock_keys(void)
708 static void slab_set_debugobj_lock_classes_node(struct kmem_cache
*cachep
, int node
)
712 static void slab_set_debugobj_lock_classes(struct kmem_cache
*cachep
)
718 * Guard access to the cache-chain.
720 static DEFINE_MUTEX(cache_chain_mutex
);
721 static struct list_head cache_chain
;
723 static DEFINE_PER_CPU(struct delayed_work
, slab_reap_work
);
725 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
727 return cachep
->array
[smp_processor_id()];
730 static inline struct kmem_cache
*__find_general_cachep(size_t size
,
733 struct cache_sizes
*csizep
= malloc_sizes
;
736 /* This happens if someone tries to call
737 * kmem_cache_create(), or __kmalloc(), before
738 * the generic caches are initialized.
740 BUG_ON(malloc_sizes
[INDEX_AC
].cs_cachep
== NULL
);
743 return ZERO_SIZE_PTR
;
745 while (size
> csizep
->cs_size
)
749 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
750 * has cs_{dma,}cachep==NULL. Thus no special case
751 * for large kmalloc calls required.
753 #ifdef CONFIG_ZONE_DMA
754 if (unlikely(gfpflags
& GFP_DMA
))
755 return csizep
->cs_dmacachep
;
757 return csizep
->cs_cachep
;
760 static struct kmem_cache
*kmem_find_general_cachep(size_t size
, gfp_t gfpflags
)
762 return __find_general_cachep(size
, gfpflags
);
765 static size_t slab_mgmt_size(size_t nr_objs
, size_t align
)
767 return ALIGN(sizeof(struct slab
)+nr_objs
*sizeof(kmem_bufctl_t
), align
);
771 * Calculate the number of objects and left-over bytes for a given buffer size.
773 static void cache_estimate(unsigned long gfporder
, size_t buffer_size
,
774 size_t align
, int flags
, size_t *left_over
,
779 size_t slab_size
= PAGE_SIZE
<< gfporder
;
782 * The slab management structure can be either off the slab or
783 * on it. For the latter case, the memory allocated for a
787 * - One kmem_bufctl_t for each object
788 * - Padding to respect alignment of @align
789 * - @buffer_size bytes for each object
791 * If the slab management structure is off the slab, then the
792 * alignment will already be calculated into the size. Because
793 * the slabs are all pages aligned, the objects will be at the
794 * correct alignment when allocated.
796 if (flags
& CFLGS_OFF_SLAB
) {
798 nr_objs
= slab_size
/ buffer_size
;
800 if (nr_objs
> SLAB_LIMIT
)
801 nr_objs
= SLAB_LIMIT
;
804 * Ignore padding for the initial guess. The padding
805 * is at most @align-1 bytes, and @buffer_size is at
806 * least @align. In the worst case, this result will
807 * be one greater than the number of objects that fit
808 * into the memory allocation when taking the padding
811 nr_objs
= (slab_size
- sizeof(struct slab
)) /
812 (buffer_size
+ sizeof(kmem_bufctl_t
));
815 * This calculated number will be either the right
816 * amount, or one greater than what we want.
818 if (slab_mgmt_size(nr_objs
, align
) + nr_objs
*buffer_size
822 if (nr_objs
> SLAB_LIMIT
)
823 nr_objs
= SLAB_LIMIT
;
825 mgmt_size
= slab_mgmt_size(nr_objs
, align
);
828 *left_over
= slab_size
- nr_objs
*buffer_size
- mgmt_size
;
831 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
833 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
836 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
837 function
, cachep
->name
, msg
);
842 * By default on NUMA we use alien caches to stage the freeing of
843 * objects allocated from other nodes. This causes massive memory
844 * inefficiencies when using fake NUMA setup to split memory into a
845 * large number of small nodes, so it can be disabled on the command
849 static int use_alien_caches __read_mostly
= 1;
850 static int __init
noaliencache_setup(char *s
)
852 use_alien_caches
= 0;
855 __setup("noaliencache", noaliencache_setup
);
859 * Special reaping functions for NUMA systems called from cache_reap().
860 * These take care of doing round robin flushing of alien caches (containing
861 * objects freed on different nodes from which they were allocated) and the
862 * flushing of remote pcps by calling drain_node_pages.
864 static DEFINE_PER_CPU(unsigned long, slab_reap_node
);
866 static void init_reap_node(int cpu
)
870 node
= next_node(cpu_to_mem(cpu
), node_online_map
);
871 if (node
== MAX_NUMNODES
)
872 node
= first_node(node_online_map
);
874 per_cpu(slab_reap_node
, cpu
) = node
;
877 static void next_reap_node(void)
879 int node
= __this_cpu_read(slab_reap_node
);
881 node
= next_node(node
, node_online_map
);
882 if (unlikely(node
>= MAX_NUMNODES
))
883 node
= first_node(node_online_map
);
884 __this_cpu_write(slab_reap_node
, node
);
888 #define init_reap_node(cpu) do { } while (0)
889 #define next_reap_node(void) do { } while (0)
893 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
894 * via the workqueue/eventd.
895 * Add the CPU number into the expiration time to minimize the possibility of
896 * the CPUs getting into lockstep and contending for the global cache chain
899 static void __cpuinit
start_cpu_timer(int cpu
)
901 struct delayed_work
*reap_work
= &per_cpu(slab_reap_work
, cpu
);
904 * When this gets called from do_initcalls via cpucache_init(),
905 * init_workqueues() has already run, so keventd will be setup
908 if (keventd_up() && reap_work
->work
.func
== NULL
) {
910 INIT_DELAYED_WORK_DEFERRABLE(reap_work
, cache_reap
);
911 schedule_delayed_work_on(cpu
, reap_work
,
912 __round_jiffies_relative(HZ
, cpu
));
916 static struct array_cache
*alloc_arraycache(int node
, int entries
,
917 int batchcount
, gfp_t gfp
)
919 int memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
920 struct array_cache
*nc
= NULL
;
922 nc
= kmalloc_node(memsize
, gfp
, node
);
924 * The array_cache structures contain pointers to free object.
925 * However, when such objects are allocated or transferred to another
926 * cache the pointers are not cleared and they could be counted as
927 * valid references during a kmemleak scan. Therefore, kmemleak must
928 * not scan such objects.
930 kmemleak_no_scan(nc
);
934 nc
->batchcount
= batchcount
;
936 spin_lock_init(&nc
->lock
);
942 * Transfer objects in one arraycache to another.
943 * Locking must be handled by the caller.
945 * Return the number of entries transferred.
947 static int transfer_objects(struct array_cache
*to
,
948 struct array_cache
*from
, unsigned int max
)
950 /* Figure out how many entries to transfer */
951 int nr
= min3(from
->avail
, max
, to
->limit
- to
->avail
);
956 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
966 #define drain_alien_cache(cachep, alien) do { } while (0)
967 #define reap_alien(cachep, l3) do { } while (0)
969 static inline struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
971 return (struct array_cache
**)BAD_ALIEN_MAGIC
;
974 static inline void free_alien_cache(struct array_cache
**ac_ptr
)
978 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
983 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
989 static inline void *____cache_alloc_node(struct kmem_cache
*cachep
,
990 gfp_t flags
, int nodeid
)
995 #else /* CONFIG_NUMA */
997 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
998 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
1000 static struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
1002 struct array_cache
**ac_ptr
;
1003 int memsize
= sizeof(void *) * nr_node_ids
;
1008 ac_ptr
= kzalloc_node(memsize
, gfp
, node
);
1011 if (i
== node
|| !node_online(i
))
1013 ac_ptr
[i
] = alloc_arraycache(node
, limit
, 0xbaadf00d, gfp
);
1015 for (i
--; i
>= 0; i
--)
1025 static void free_alien_cache(struct array_cache
**ac_ptr
)
1036 static void __drain_alien_cache(struct kmem_cache
*cachep
,
1037 struct array_cache
*ac
, int node
)
1039 struct kmem_list3
*rl3
= cachep
->nodelists
[node
];
1042 spin_lock(&rl3
->list_lock
);
1044 * Stuff objects into the remote nodes shared array first.
1045 * That way we could avoid the overhead of putting the objects
1046 * into the free lists and getting them back later.
1049 transfer_objects(rl3
->shared
, ac
, ac
->limit
);
1051 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
1053 spin_unlock(&rl3
->list_lock
);
1058 * Called from cache_reap() to regularly drain alien caches round robin.
1060 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_list3
*l3
)
1062 int node
= __this_cpu_read(slab_reap_node
);
1065 struct array_cache
*ac
= l3
->alien
[node
];
1067 if (ac
&& ac
->avail
&& spin_trylock_irq(&ac
->lock
)) {
1068 __drain_alien_cache(cachep
, ac
, node
);
1069 spin_unlock_irq(&ac
->lock
);
1074 static void drain_alien_cache(struct kmem_cache
*cachep
,
1075 struct array_cache
**alien
)
1078 struct array_cache
*ac
;
1079 unsigned long flags
;
1081 for_each_online_node(i
) {
1084 spin_lock_irqsave(&ac
->lock
, flags
);
1085 __drain_alien_cache(cachep
, ac
, i
);
1086 spin_unlock_irqrestore(&ac
->lock
, flags
);
1091 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1093 struct slab
*slabp
= virt_to_slab(objp
);
1094 int nodeid
= slabp
->nodeid
;
1095 struct kmem_list3
*l3
;
1096 struct array_cache
*alien
= NULL
;
1099 node
= numa_mem_id();
1102 * Make sure we are not freeing a object from another node to the array
1103 * cache on this cpu.
1105 if (likely(slabp
->nodeid
== node
))
1108 l3
= cachep
->nodelists
[node
];
1109 STATS_INC_NODEFREES(cachep
);
1110 if (l3
->alien
&& l3
->alien
[nodeid
]) {
1111 alien
= l3
->alien
[nodeid
];
1112 spin_lock(&alien
->lock
);
1113 if (unlikely(alien
->avail
== alien
->limit
)) {
1114 STATS_INC_ACOVERFLOW(cachep
);
1115 __drain_alien_cache(cachep
, alien
, nodeid
);
1117 alien
->entry
[alien
->avail
++] = objp
;
1118 spin_unlock(&alien
->lock
);
1120 spin_lock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1121 free_block(cachep
, &objp
, 1, nodeid
);
1122 spin_unlock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1129 * Allocates and initializes nodelists for a node on each slab cache, used for
1130 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1131 * will be allocated off-node since memory is not yet online for the new node.
1132 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1135 * Must hold cache_chain_mutex.
1137 static int init_cache_nodelists_node(int node
)
1139 struct kmem_cache
*cachep
;
1140 struct kmem_list3
*l3
;
1141 const int memsize
= sizeof(struct kmem_list3
);
1143 list_for_each_entry(cachep
, &cache_chain
, next
) {
1145 * Set up the size64 kmemlist for cpu before we can
1146 * begin anything. Make sure some other cpu on this
1147 * node has not already allocated this
1149 if (!cachep
->nodelists
[node
]) {
1150 l3
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1153 kmem_list3_init(l3
);
1154 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
1155 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1158 * The l3s don't come and go as CPUs come and
1159 * go. cache_chain_mutex is sufficient
1162 cachep
->nodelists
[node
] = l3
;
1165 spin_lock_irq(&cachep
->nodelists
[node
]->list_lock
);
1166 cachep
->nodelists
[node
]->free_limit
=
1167 (1 + nr_cpus_node(node
)) *
1168 cachep
->batchcount
+ cachep
->num
;
1169 spin_unlock_irq(&cachep
->nodelists
[node
]->list_lock
);
1174 static void __cpuinit
cpuup_canceled(long cpu
)
1176 struct kmem_cache
*cachep
;
1177 struct kmem_list3
*l3
= NULL
;
1178 int node
= cpu_to_mem(cpu
);
1179 const struct cpumask
*mask
= cpumask_of_node(node
);
1181 list_for_each_entry(cachep
, &cache_chain
, next
) {
1182 struct array_cache
*nc
;
1183 struct array_cache
*shared
;
1184 struct array_cache
**alien
;
1186 /* cpu is dead; no one can alloc from it. */
1187 nc
= cachep
->array
[cpu
];
1188 cachep
->array
[cpu
] = NULL
;
1189 l3
= cachep
->nodelists
[node
];
1192 goto free_array_cache
;
1194 spin_lock_irq(&l3
->list_lock
);
1196 /* Free limit for this kmem_list3 */
1197 l3
->free_limit
-= cachep
->batchcount
;
1199 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
1201 if (!cpumask_empty(mask
)) {
1202 spin_unlock_irq(&l3
->list_lock
);
1203 goto free_array_cache
;
1206 shared
= l3
->shared
;
1208 free_block(cachep
, shared
->entry
,
1209 shared
->avail
, node
);
1216 spin_unlock_irq(&l3
->list_lock
);
1220 drain_alien_cache(cachep
, alien
);
1221 free_alien_cache(alien
);
1227 * In the previous loop, all the objects were freed to
1228 * the respective cache's slabs, now we can go ahead and
1229 * shrink each nodelist to its limit.
1231 list_for_each_entry(cachep
, &cache_chain
, next
) {
1232 l3
= cachep
->nodelists
[node
];
1235 drain_freelist(cachep
, l3
, l3
->free_objects
);
1239 static int __cpuinit
cpuup_prepare(long cpu
)
1241 struct kmem_cache
*cachep
;
1242 struct kmem_list3
*l3
= NULL
;
1243 int node
= cpu_to_mem(cpu
);
1247 * We need to do this right in the beginning since
1248 * alloc_arraycache's are going to use this list.
1249 * kmalloc_node allows us to add the slab to the right
1250 * kmem_list3 and not this cpu's kmem_list3
1252 err
= init_cache_nodelists_node(node
);
1257 * Now we can go ahead with allocating the shared arrays and
1260 list_for_each_entry(cachep
, &cache_chain
, next
) {
1261 struct array_cache
*nc
;
1262 struct array_cache
*shared
= NULL
;
1263 struct array_cache
**alien
= NULL
;
1265 nc
= alloc_arraycache(node
, cachep
->limit
,
1266 cachep
->batchcount
, GFP_KERNEL
);
1269 if (cachep
->shared
) {
1270 shared
= alloc_arraycache(node
,
1271 cachep
->shared
* cachep
->batchcount
,
1272 0xbaadf00d, GFP_KERNEL
);
1278 if (use_alien_caches
) {
1279 alien
= alloc_alien_cache(node
, cachep
->limit
, GFP_KERNEL
);
1286 cachep
->array
[cpu
] = nc
;
1287 l3
= cachep
->nodelists
[node
];
1290 spin_lock_irq(&l3
->list_lock
);
1293 * We are serialised from CPU_DEAD or
1294 * CPU_UP_CANCELLED by the cpucontrol lock
1296 l3
->shared
= shared
;
1305 spin_unlock_irq(&l3
->list_lock
);
1307 free_alien_cache(alien
);
1308 if (cachep
->flags
& SLAB_DEBUG_OBJECTS
)
1309 slab_set_debugobj_lock_classes_node(cachep
, node
);
1311 init_node_lock_keys(node
);
1315 cpuup_canceled(cpu
);
1319 static int __cpuinit
cpuup_callback(struct notifier_block
*nfb
,
1320 unsigned long action
, void *hcpu
)
1322 long cpu
= (long)hcpu
;
1326 case CPU_UP_PREPARE
:
1327 case CPU_UP_PREPARE_FROZEN
:
1328 mutex_lock(&cache_chain_mutex
);
1329 err
= cpuup_prepare(cpu
);
1330 mutex_unlock(&cache_chain_mutex
);
1333 case CPU_ONLINE_FROZEN
:
1334 start_cpu_timer(cpu
);
1336 #ifdef CONFIG_HOTPLUG_CPU
1337 case CPU_DOWN_PREPARE
:
1338 case CPU_DOWN_PREPARE_FROZEN
:
1340 * Shutdown cache reaper. Note that the cache_chain_mutex is
1341 * held so that if cache_reap() is invoked it cannot do
1342 * anything expensive but will only modify reap_work
1343 * and reschedule the timer.
1345 cancel_delayed_work_sync(&per_cpu(slab_reap_work
, cpu
));
1346 /* Now the cache_reaper is guaranteed to be not running. */
1347 per_cpu(slab_reap_work
, cpu
).work
.func
= NULL
;
1349 case CPU_DOWN_FAILED
:
1350 case CPU_DOWN_FAILED_FROZEN
:
1351 start_cpu_timer(cpu
);
1354 case CPU_DEAD_FROZEN
:
1356 * Even if all the cpus of a node are down, we don't free the
1357 * kmem_list3 of any cache. This to avoid a race between
1358 * cpu_down, and a kmalloc allocation from another cpu for
1359 * memory from the node of the cpu going down. The list3
1360 * structure is usually allocated from kmem_cache_create() and
1361 * gets destroyed at kmem_cache_destroy().
1365 case CPU_UP_CANCELED
:
1366 case CPU_UP_CANCELED_FROZEN
:
1367 mutex_lock(&cache_chain_mutex
);
1368 cpuup_canceled(cpu
);
1369 mutex_unlock(&cache_chain_mutex
);
1372 return notifier_from_errno(err
);
1375 static struct notifier_block __cpuinitdata cpucache_notifier
= {
1376 &cpuup_callback
, NULL
, 0
1379 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1381 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1382 * Returns -EBUSY if all objects cannot be drained so that the node is not
1385 * Must hold cache_chain_mutex.
1387 static int __meminit
drain_cache_nodelists_node(int node
)
1389 struct kmem_cache
*cachep
;
1392 list_for_each_entry(cachep
, &cache_chain
, next
) {
1393 struct kmem_list3
*l3
;
1395 l3
= cachep
->nodelists
[node
];
1399 drain_freelist(cachep
, l3
, l3
->free_objects
);
1401 if (!list_empty(&l3
->slabs_full
) ||
1402 !list_empty(&l3
->slabs_partial
)) {
1410 static int __meminit
slab_memory_callback(struct notifier_block
*self
,
1411 unsigned long action
, void *arg
)
1413 struct memory_notify
*mnb
= arg
;
1417 nid
= mnb
->status_change_nid
;
1422 case MEM_GOING_ONLINE
:
1423 mutex_lock(&cache_chain_mutex
);
1424 ret
= init_cache_nodelists_node(nid
);
1425 mutex_unlock(&cache_chain_mutex
);
1427 case MEM_GOING_OFFLINE
:
1428 mutex_lock(&cache_chain_mutex
);
1429 ret
= drain_cache_nodelists_node(nid
);
1430 mutex_unlock(&cache_chain_mutex
);
1434 case MEM_CANCEL_ONLINE
:
1435 case MEM_CANCEL_OFFLINE
:
1439 return notifier_from_errno(ret
);
1441 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1444 * swap the static kmem_list3 with kmalloced memory
1446 static void __init
init_list(struct kmem_cache
*cachep
, struct kmem_list3
*list
,
1449 struct kmem_list3
*ptr
;
1451 ptr
= kmalloc_node(sizeof(struct kmem_list3
), GFP_NOWAIT
, nodeid
);
1454 memcpy(ptr
, list
, sizeof(struct kmem_list3
));
1456 * Do not assume that spinlocks can be initialized via memcpy:
1458 spin_lock_init(&ptr
->list_lock
);
1460 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1461 cachep
->nodelists
[nodeid
] = ptr
;
1465 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1466 * size of kmem_list3.
1468 static void __init
set_up_list3s(struct kmem_cache
*cachep
, int index
)
1472 for_each_online_node(node
) {
1473 cachep
->nodelists
[node
] = &initkmem_list3
[index
+ node
];
1474 cachep
->nodelists
[node
]->next_reap
= jiffies
+
1476 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1481 * Initialisation. Called after the page allocator have been initialised and
1482 * before smp_init().
1484 void __init
kmem_cache_init(void)
1487 struct cache_sizes
*sizes
;
1488 struct cache_names
*names
;
1493 if (num_possible_nodes() == 1)
1494 use_alien_caches
= 0;
1496 for (i
= 0; i
< NUM_INIT_LISTS
; i
++) {
1497 kmem_list3_init(&initkmem_list3
[i
]);
1498 if (i
< MAX_NUMNODES
)
1499 cache_cache
.nodelists
[i
] = NULL
;
1501 set_up_list3s(&cache_cache
, CACHE_CACHE
);
1504 * Fragmentation resistance on low memory - only use bigger
1505 * page orders on machines with more than 32MB of memory.
1507 if (totalram_pages
> (32 << 20) >> PAGE_SHIFT
)
1508 slab_break_gfp_order
= BREAK_GFP_ORDER_HI
;
1510 /* Bootstrap is tricky, because several objects are allocated
1511 * from caches that do not exist yet:
1512 * 1) initialize the cache_cache cache: it contains the struct
1513 * kmem_cache structures of all caches, except cache_cache itself:
1514 * cache_cache is statically allocated.
1515 * Initially an __init data area is used for the head array and the
1516 * kmem_list3 structures, it's replaced with a kmalloc allocated
1517 * array at the end of the bootstrap.
1518 * 2) Create the first kmalloc cache.
1519 * The struct kmem_cache for the new cache is allocated normally.
1520 * An __init data area is used for the head array.
1521 * 3) Create the remaining kmalloc caches, with minimally sized
1523 * 4) Replace the __init data head arrays for cache_cache and the first
1524 * kmalloc cache with kmalloc allocated arrays.
1525 * 5) Replace the __init data for kmem_list3 for cache_cache and
1526 * the other cache's with kmalloc allocated memory.
1527 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1530 node
= numa_mem_id();
1532 /* 1) create the cache_cache */
1533 INIT_LIST_HEAD(&cache_chain
);
1534 list_add(&cache_cache
.next
, &cache_chain
);
1535 cache_cache
.colour_off
= cache_line_size();
1536 cache_cache
.array
[smp_processor_id()] = &initarray_cache
.cache
;
1537 cache_cache
.nodelists
[node
] = &initkmem_list3
[CACHE_CACHE
+ node
];
1540 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1542 cache_cache
.buffer_size
= offsetof(struct kmem_cache
, array
[nr_cpu_ids
]) +
1543 nr_node_ids
* sizeof(struct kmem_list3
*);
1545 cache_cache
.obj_size
= cache_cache
.buffer_size
;
1547 cache_cache
.buffer_size
= ALIGN(cache_cache
.buffer_size
,
1549 cache_cache
.reciprocal_buffer_size
=
1550 reciprocal_value(cache_cache
.buffer_size
);
1552 for (order
= 0; order
< MAX_ORDER
; order
++) {
1553 cache_estimate(order
, cache_cache
.buffer_size
,
1554 cache_line_size(), 0, &left_over
, &cache_cache
.num
);
1555 if (cache_cache
.num
)
1558 BUG_ON(!cache_cache
.num
);
1559 cache_cache
.gfporder
= order
;
1560 cache_cache
.colour
= left_over
/ cache_cache
.colour_off
;
1561 cache_cache
.slab_size
= ALIGN(cache_cache
.num
* sizeof(kmem_bufctl_t
) +
1562 sizeof(struct slab
), cache_line_size());
1564 /* 2+3) create the kmalloc caches */
1565 sizes
= malloc_sizes
;
1566 names
= cache_names
;
1569 * Initialize the caches that provide memory for the array cache and the
1570 * kmem_list3 structures first. Without this, further allocations will
1574 sizes
[INDEX_AC
].cs_cachep
= kmem_cache_create(names
[INDEX_AC
].name
,
1575 sizes
[INDEX_AC
].cs_size
,
1576 ARCH_KMALLOC_MINALIGN
,
1577 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1580 if (INDEX_AC
!= INDEX_L3
) {
1581 sizes
[INDEX_L3
].cs_cachep
=
1582 kmem_cache_create(names
[INDEX_L3
].name
,
1583 sizes
[INDEX_L3
].cs_size
,
1584 ARCH_KMALLOC_MINALIGN
,
1585 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1589 slab_early_init
= 0;
1591 while (sizes
->cs_size
!= ULONG_MAX
) {
1593 * For performance, all the general caches are L1 aligned.
1594 * This should be particularly beneficial on SMP boxes, as it
1595 * eliminates "false sharing".
1596 * Note for systems short on memory removing the alignment will
1597 * allow tighter packing of the smaller caches.
1599 if (!sizes
->cs_cachep
) {
1600 sizes
->cs_cachep
= kmem_cache_create(names
->name
,
1602 ARCH_KMALLOC_MINALIGN
,
1603 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1606 #ifdef CONFIG_ZONE_DMA
1607 sizes
->cs_dmacachep
= kmem_cache_create(
1610 ARCH_KMALLOC_MINALIGN
,
1611 ARCH_KMALLOC_FLAGS
|SLAB_CACHE_DMA
|
1618 /* 4) Replace the bootstrap head arrays */
1620 struct array_cache
*ptr
;
1622 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1624 BUG_ON(cpu_cache_get(&cache_cache
) != &initarray_cache
.cache
);
1625 memcpy(ptr
, cpu_cache_get(&cache_cache
),
1626 sizeof(struct arraycache_init
));
1628 * Do not assume that spinlocks can be initialized via memcpy:
1630 spin_lock_init(&ptr
->lock
);
1632 cache_cache
.array
[smp_processor_id()] = ptr
;
1634 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1636 BUG_ON(cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
)
1637 != &initarray_generic
.cache
);
1638 memcpy(ptr
, cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
),
1639 sizeof(struct arraycache_init
));
1641 * Do not assume that spinlocks can be initialized via memcpy:
1643 spin_lock_init(&ptr
->lock
);
1645 malloc_sizes
[INDEX_AC
].cs_cachep
->array
[smp_processor_id()] =
1648 /* 5) Replace the bootstrap kmem_list3's */
1652 for_each_online_node(nid
) {
1653 init_list(&cache_cache
, &initkmem_list3
[CACHE_CACHE
+ nid
], nid
);
1655 init_list(malloc_sizes
[INDEX_AC
].cs_cachep
,
1656 &initkmem_list3
[SIZE_AC
+ nid
], nid
);
1658 if (INDEX_AC
!= INDEX_L3
) {
1659 init_list(malloc_sizes
[INDEX_L3
].cs_cachep
,
1660 &initkmem_list3
[SIZE_L3
+ nid
], nid
);
1665 g_cpucache_up
= EARLY
;
1668 void __init
kmem_cache_init_late(void)
1670 struct kmem_cache
*cachep
;
1672 g_cpucache_up
= LATE
;
1674 /* Annotate slab for lockdep -- annotate the malloc caches */
1677 /* 6) resize the head arrays to their final sizes */
1678 mutex_lock(&cache_chain_mutex
);
1679 list_for_each_entry(cachep
, &cache_chain
, next
)
1680 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1682 mutex_unlock(&cache_chain_mutex
);
1685 g_cpucache_up
= FULL
;
1688 * Register a cpu startup notifier callback that initializes
1689 * cpu_cache_get for all new cpus
1691 register_cpu_notifier(&cpucache_notifier
);
1695 * Register a memory hotplug callback that initializes and frees
1698 hotplug_memory_notifier(slab_memory_callback
, SLAB_CALLBACK_PRI
);
1702 * The reap timers are started later, with a module init call: That part
1703 * of the kernel is not yet operational.
1707 static int __init
cpucache_init(void)
1712 * Register the timers that return unneeded pages to the page allocator
1714 for_each_online_cpu(cpu
)
1715 start_cpu_timer(cpu
);
1718 __initcall(cpucache_init
);
1721 * Interface to system's page allocator. No need to hold the cache-lock.
1723 * If we requested dmaable memory, we will get it. Even if we
1724 * did not request dmaable memory, we might get it, but that
1725 * would be relatively rare and ignorable.
1727 static void *kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
1735 * Nommu uses slab's for process anonymous memory allocations, and thus
1736 * requires __GFP_COMP to properly refcount higher order allocations
1738 flags
|= __GFP_COMP
;
1741 flags
|= cachep
->gfpflags
;
1742 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1743 flags
|= __GFP_RECLAIMABLE
;
1745 page
= alloc_pages_exact_node(nodeid
, flags
| __GFP_NOTRACK
, cachep
->gfporder
);
1749 nr_pages
= (1 << cachep
->gfporder
);
1750 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1751 add_zone_page_state(page_zone(page
),
1752 NR_SLAB_RECLAIMABLE
, nr_pages
);
1754 add_zone_page_state(page_zone(page
),
1755 NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1756 for (i
= 0; i
< nr_pages
; i
++)
1757 __SetPageSlab(page
+ i
);
1759 if (kmemcheck_enabled
&& !(cachep
->flags
& SLAB_NOTRACK
)) {
1760 kmemcheck_alloc_shadow(page
, cachep
->gfporder
, flags
, nodeid
);
1763 kmemcheck_mark_uninitialized_pages(page
, nr_pages
);
1765 kmemcheck_mark_unallocated_pages(page
, nr_pages
);
1768 return page_address(page
);
1772 * Interface to system's page release.
1774 static void kmem_freepages(struct kmem_cache
*cachep
, void *addr
)
1776 unsigned long i
= (1 << cachep
->gfporder
);
1777 struct page
*page
= virt_to_page(addr
);
1778 const unsigned long nr_freed
= i
;
1780 kmemcheck_free_shadow(page
, cachep
->gfporder
);
1782 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1783 sub_zone_page_state(page_zone(page
),
1784 NR_SLAB_RECLAIMABLE
, nr_freed
);
1786 sub_zone_page_state(page_zone(page
),
1787 NR_SLAB_UNRECLAIMABLE
, nr_freed
);
1789 BUG_ON(!PageSlab(page
));
1790 __ClearPageSlab(page
);
1793 if (current
->reclaim_state
)
1794 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1795 free_pages((unsigned long)addr
, cachep
->gfporder
);
1798 static void kmem_rcu_free(struct rcu_head
*head
)
1800 struct slab_rcu
*slab_rcu
= (struct slab_rcu
*)head
;
1801 struct kmem_cache
*cachep
= slab_rcu
->cachep
;
1803 kmem_freepages(cachep
, slab_rcu
->addr
);
1804 if (OFF_SLAB(cachep
))
1805 kmem_cache_free(cachep
->slabp_cache
, slab_rcu
);
1810 #ifdef CONFIG_DEBUG_PAGEALLOC
1811 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1812 unsigned long caller
)
1814 int size
= obj_size(cachep
);
1816 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1818 if (size
< 5 * sizeof(unsigned long))
1821 *addr
++ = 0x12345678;
1823 *addr
++ = smp_processor_id();
1824 size
-= 3 * sizeof(unsigned long);
1826 unsigned long *sptr
= &caller
;
1827 unsigned long svalue
;
1829 while (!kstack_end(sptr
)) {
1831 if (kernel_text_address(svalue
)) {
1833 size
-= sizeof(unsigned long);
1834 if (size
<= sizeof(unsigned long))
1840 *addr
++ = 0x87654321;
1844 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1846 int size
= obj_size(cachep
);
1847 addr
= &((char *)addr
)[obj_offset(cachep
)];
1849 memset(addr
, val
, size
);
1850 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1853 static void dump_line(char *data
, int offset
, int limit
)
1856 unsigned char error
= 0;
1859 printk(KERN_ERR
"%03x: ", offset
);
1860 for (i
= 0; i
< limit
; i
++) {
1861 if (data
[offset
+ i
] != POISON_FREE
) {
1862 error
= data
[offset
+ i
];
1866 print_hex_dump(KERN_CONT
, "", 0, 16, 1,
1867 &data
[offset
], limit
, 1);
1869 if (bad_count
== 1) {
1870 error
^= POISON_FREE
;
1871 if (!(error
& (error
- 1))) {
1872 printk(KERN_ERR
"Single bit error detected. Probably "
1875 printk(KERN_ERR
"Run memtest86+ or a similar memory "
1878 printk(KERN_ERR
"Run a memory test tool.\n");
1887 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1892 if (cachep
->flags
& SLAB_RED_ZONE
) {
1893 printk(KERN_ERR
"Redzone: 0x%llx/0x%llx.\n",
1894 *dbg_redzone1(cachep
, objp
),
1895 *dbg_redzone2(cachep
, objp
));
1898 if (cachep
->flags
& SLAB_STORE_USER
) {
1899 printk(KERN_ERR
"Last user: [<%p>]",
1900 *dbg_userword(cachep
, objp
));
1901 print_symbol("(%s)",
1902 (unsigned long)*dbg_userword(cachep
, objp
));
1905 realobj
= (char *)objp
+ obj_offset(cachep
);
1906 size
= obj_size(cachep
);
1907 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1910 if (i
+ limit
> size
)
1912 dump_line(realobj
, i
, limit
);
1916 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1922 realobj
= (char *)objp
+ obj_offset(cachep
);
1923 size
= obj_size(cachep
);
1925 for (i
= 0; i
< size
; i
++) {
1926 char exp
= POISON_FREE
;
1929 if (realobj
[i
] != exp
) {
1935 "Slab corruption: %s start=%p, len=%d\n",
1936 cachep
->name
, realobj
, size
);
1937 print_objinfo(cachep
, objp
, 0);
1939 /* Hexdump the affected line */
1942 if (i
+ limit
> size
)
1944 dump_line(realobj
, i
, limit
);
1947 /* Limit to 5 lines */
1953 /* Print some data about the neighboring objects, if they
1956 struct slab
*slabp
= virt_to_slab(objp
);
1959 objnr
= obj_to_index(cachep
, slabp
, objp
);
1961 objp
= index_to_obj(cachep
, slabp
, objnr
- 1);
1962 realobj
= (char *)objp
+ obj_offset(cachep
);
1963 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
1965 print_objinfo(cachep
, objp
, 2);
1967 if (objnr
+ 1 < cachep
->num
) {
1968 objp
= index_to_obj(cachep
, slabp
, objnr
+ 1);
1969 realobj
= (char *)objp
+ obj_offset(cachep
);
1970 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
1972 print_objinfo(cachep
, objp
, 2);
1979 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
, struct slab
*slabp
)
1982 for (i
= 0; i
< cachep
->num
; i
++) {
1983 void *objp
= index_to_obj(cachep
, slabp
, i
);
1985 if (cachep
->flags
& SLAB_POISON
) {
1986 #ifdef CONFIG_DEBUG_PAGEALLOC
1987 if (cachep
->buffer_size
% PAGE_SIZE
== 0 &&
1989 kernel_map_pages(virt_to_page(objp
),
1990 cachep
->buffer_size
/ PAGE_SIZE
, 1);
1992 check_poison_obj(cachep
, objp
);
1994 check_poison_obj(cachep
, objp
);
1997 if (cachep
->flags
& SLAB_RED_ZONE
) {
1998 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1999 slab_error(cachep
, "start of a freed object "
2001 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2002 slab_error(cachep
, "end of a freed object "
2008 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
, struct slab
*slabp
)
2014 * slab_destroy - destroy and release all objects in a slab
2015 * @cachep: cache pointer being destroyed
2016 * @slabp: slab pointer being destroyed
2018 * Destroy all the objs in a slab, and release the mem back to the system.
2019 * Before calling the slab must have been unlinked from the cache. The
2020 * cache-lock is not held/needed.
2022 static void slab_destroy(struct kmem_cache
*cachep
, struct slab
*slabp
)
2024 void *addr
= slabp
->s_mem
- slabp
->colouroff
;
2026 slab_destroy_debugcheck(cachep
, slabp
);
2027 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
2028 struct slab_rcu
*slab_rcu
;
2030 slab_rcu
= (struct slab_rcu
*)slabp
;
2031 slab_rcu
->cachep
= cachep
;
2032 slab_rcu
->addr
= addr
;
2033 call_rcu(&slab_rcu
->head
, kmem_rcu_free
);
2035 kmem_freepages(cachep
, addr
);
2036 if (OFF_SLAB(cachep
))
2037 kmem_cache_free(cachep
->slabp_cache
, slabp
);
2041 static void __kmem_cache_destroy(struct kmem_cache
*cachep
)
2044 struct kmem_list3
*l3
;
2046 for_each_online_cpu(i
)
2047 kfree(cachep
->array
[i
]);
2049 /* NUMA: free the list3 structures */
2050 for_each_online_node(i
) {
2051 l3
= cachep
->nodelists
[i
];
2054 free_alien_cache(l3
->alien
);
2058 kmem_cache_free(&cache_cache
, cachep
);
2063 * calculate_slab_order - calculate size (page order) of slabs
2064 * @cachep: pointer to the cache that is being created
2065 * @size: size of objects to be created in this cache.
2066 * @align: required alignment for the objects.
2067 * @flags: slab allocation flags
2069 * Also calculates the number of objects per slab.
2071 * This could be made much more intelligent. For now, try to avoid using
2072 * high order pages for slabs. When the gfp() functions are more friendly
2073 * towards high-order requests, this should be changed.
2075 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
2076 size_t size
, size_t align
, unsigned long flags
)
2078 unsigned long offslab_limit
;
2079 size_t left_over
= 0;
2082 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
2086 cache_estimate(gfporder
, size
, align
, flags
, &remainder
, &num
);
2090 if (flags
& CFLGS_OFF_SLAB
) {
2092 * Max number of objs-per-slab for caches which
2093 * use off-slab slabs. Needed to avoid a possible
2094 * looping condition in cache_grow().
2096 offslab_limit
= size
- sizeof(struct slab
);
2097 offslab_limit
/= sizeof(kmem_bufctl_t
);
2099 if (num
> offslab_limit
)
2103 /* Found something acceptable - save it away */
2105 cachep
->gfporder
= gfporder
;
2106 left_over
= remainder
;
2109 * A VFS-reclaimable slab tends to have most allocations
2110 * as GFP_NOFS and we really don't want to have to be allocating
2111 * higher-order pages when we are unable to shrink dcache.
2113 if (flags
& SLAB_RECLAIM_ACCOUNT
)
2117 * Large number of objects is good, but very large slabs are
2118 * currently bad for the gfp()s.
2120 if (gfporder
>= slab_break_gfp_order
)
2124 * Acceptable internal fragmentation?
2126 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
2132 static int __init_refok
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
2134 if (g_cpucache_up
== FULL
)
2135 return enable_cpucache(cachep
, gfp
);
2137 if (g_cpucache_up
== NONE
) {
2139 * Note: the first kmem_cache_create must create the cache
2140 * that's used by kmalloc(24), otherwise the creation of
2141 * further caches will BUG().
2143 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
2146 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2147 * the first cache, then we need to set up all its list3s,
2148 * otherwise the creation of further caches will BUG().
2150 set_up_list3s(cachep
, SIZE_AC
);
2151 if (INDEX_AC
== INDEX_L3
)
2152 g_cpucache_up
= PARTIAL_L3
;
2154 g_cpucache_up
= PARTIAL_AC
;
2156 cachep
->array
[smp_processor_id()] =
2157 kmalloc(sizeof(struct arraycache_init
), gfp
);
2159 if (g_cpucache_up
== PARTIAL_AC
) {
2160 set_up_list3s(cachep
, SIZE_L3
);
2161 g_cpucache_up
= PARTIAL_L3
;
2164 for_each_online_node(node
) {
2165 cachep
->nodelists
[node
] =
2166 kmalloc_node(sizeof(struct kmem_list3
),
2168 BUG_ON(!cachep
->nodelists
[node
]);
2169 kmem_list3_init(cachep
->nodelists
[node
]);
2173 cachep
->nodelists
[numa_mem_id()]->next_reap
=
2174 jiffies
+ REAPTIMEOUT_LIST3
+
2175 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
2177 cpu_cache_get(cachep
)->avail
= 0;
2178 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
2179 cpu_cache_get(cachep
)->batchcount
= 1;
2180 cpu_cache_get(cachep
)->touched
= 0;
2181 cachep
->batchcount
= 1;
2182 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
2187 * kmem_cache_create - Create a cache.
2188 * @name: A string which is used in /proc/slabinfo to identify this cache.
2189 * @size: The size of objects to be created in this cache.
2190 * @align: The required alignment for the objects.
2191 * @flags: SLAB flags
2192 * @ctor: A constructor for the objects.
2194 * Returns a ptr to the cache on success, NULL on failure.
2195 * Cannot be called within a int, but can be interrupted.
2196 * The @ctor is run when new pages are allocated by the cache.
2198 * @name must be valid until the cache is destroyed. This implies that
2199 * the module calling this has to destroy the cache before getting unloaded.
2203 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2204 * to catch references to uninitialised memory.
2206 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2207 * for buffer overruns.
2209 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2210 * cacheline. This can be beneficial if you're counting cycles as closely
2214 kmem_cache_create (const char *name
, size_t size
, size_t align
,
2215 unsigned long flags
, void (*ctor
)(void *))
2217 size_t left_over
, slab_size
, ralign
;
2218 struct kmem_cache
*cachep
= NULL
, *pc
;
2222 * Sanity checks... these are all serious usage bugs.
2224 if (!name
|| in_interrupt() || (size
< BYTES_PER_WORD
) ||
2225 size
> KMALLOC_MAX_SIZE
) {
2226 printk(KERN_ERR
"%s: Early error in slab %s\n", __func__
,
2232 * We use cache_chain_mutex to ensure a consistent view of
2233 * cpu_online_mask as well. Please see cpuup_callback
2235 if (slab_is_available()) {
2237 mutex_lock(&cache_chain_mutex
);
2240 list_for_each_entry(pc
, &cache_chain
, next
) {
2245 * This happens when the module gets unloaded and doesn't
2246 * destroy its slab cache and no-one else reuses the vmalloc
2247 * area of the module. Print a warning.
2249 res
= probe_kernel_address(pc
->name
, tmp
);
2252 "SLAB: cache with size %d has lost its name\n",
2257 if (!strcmp(pc
->name
, name
)) {
2259 "kmem_cache_create: duplicate cache %s\n", name
);
2266 WARN_ON(strchr(name
, ' ')); /* It confuses parsers */
2269 * Enable redzoning and last user accounting, except for caches with
2270 * large objects, if the increased size would increase the object size
2271 * above the next power of two: caches with object sizes just above a
2272 * power of two have a significant amount of internal fragmentation.
2274 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
2275 2 * sizeof(unsigned long long)))
2276 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2277 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2278 flags
|= SLAB_POISON
;
2280 if (flags
& SLAB_DESTROY_BY_RCU
)
2281 BUG_ON(flags
& SLAB_POISON
);
2284 * Always checks flags, a caller might be expecting debug support which
2287 BUG_ON(flags
& ~CREATE_MASK
);
2290 * Check that size is in terms of words. This is needed to avoid
2291 * unaligned accesses for some archs when redzoning is used, and makes
2292 * sure any on-slab bufctl's are also correctly aligned.
2294 if (size
& (BYTES_PER_WORD
- 1)) {
2295 size
+= (BYTES_PER_WORD
- 1);
2296 size
&= ~(BYTES_PER_WORD
- 1);
2299 /* calculate the final buffer alignment: */
2301 /* 1) arch recommendation: can be overridden for debug */
2302 if (flags
& SLAB_HWCACHE_ALIGN
) {
2304 * Default alignment: as specified by the arch code. Except if
2305 * an object is really small, then squeeze multiple objects into
2308 ralign
= cache_line_size();
2309 while (size
<= ralign
/ 2)
2312 ralign
= BYTES_PER_WORD
;
2316 * Redzoning and user store require word alignment or possibly larger.
2317 * Note this will be overridden by architecture or caller mandated
2318 * alignment if either is greater than BYTES_PER_WORD.
2320 if (flags
& SLAB_STORE_USER
)
2321 ralign
= BYTES_PER_WORD
;
2323 if (flags
& SLAB_RED_ZONE
) {
2324 ralign
= REDZONE_ALIGN
;
2325 /* If redzoning, ensure that the second redzone is suitably
2326 * aligned, by adjusting the object size accordingly. */
2327 size
+= REDZONE_ALIGN
- 1;
2328 size
&= ~(REDZONE_ALIGN
- 1);
2331 /* 2) arch mandated alignment */
2332 if (ralign
< ARCH_SLAB_MINALIGN
) {
2333 ralign
= ARCH_SLAB_MINALIGN
;
2335 /* 3) caller mandated alignment */
2336 if (ralign
< align
) {
2339 /* disable debug if necessary */
2340 if (ralign
> __alignof__(unsigned long long))
2341 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2347 if (slab_is_available())
2352 /* Get cache's description obj. */
2353 cachep
= kmem_cache_zalloc(&cache_cache
, gfp
);
2357 cachep
->nodelists
= (struct kmem_list3
**)&cachep
->array
[nr_cpu_ids
];
2359 cachep
->obj_size
= size
;
2362 * Both debugging options require word-alignment which is calculated
2365 if (flags
& SLAB_RED_ZONE
) {
2366 /* add space for red zone words */
2367 cachep
->obj_offset
+= sizeof(unsigned long long);
2368 size
+= 2 * sizeof(unsigned long long);
2370 if (flags
& SLAB_STORE_USER
) {
2371 /* user store requires one word storage behind the end of
2372 * the real object. But if the second red zone needs to be
2373 * aligned to 64 bits, we must allow that much space.
2375 if (flags
& SLAB_RED_ZONE
)
2376 size
+= REDZONE_ALIGN
;
2378 size
+= BYTES_PER_WORD
;
2380 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2381 if (size
>= malloc_sizes
[INDEX_L3
+ 1].cs_size
2382 && cachep
->obj_size
> cache_line_size() && ALIGN(size
, align
) < PAGE_SIZE
) {
2383 cachep
->obj_offset
+= PAGE_SIZE
- ALIGN(size
, align
);
2390 * Determine if the slab management is 'on' or 'off' slab.
2391 * (bootstrapping cannot cope with offslab caches so don't do
2392 * it too early on. Always use on-slab management when
2393 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2395 if ((size
>= (PAGE_SIZE
>> 3)) && !slab_early_init
&&
2396 !(flags
& SLAB_NOLEAKTRACE
))
2398 * Size is large, assume best to place the slab management obj
2399 * off-slab (should allow better packing of objs).
2401 flags
|= CFLGS_OFF_SLAB
;
2403 size
= ALIGN(size
, align
);
2405 left_over
= calculate_slab_order(cachep
, size
, align
, flags
);
2409 "kmem_cache_create: couldn't create cache %s.\n", name
);
2410 kmem_cache_free(&cache_cache
, cachep
);
2414 slab_size
= ALIGN(cachep
->num
* sizeof(kmem_bufctl_t
)
2415 + sizeof(struct slab
), align
);
2418 * If the slab has been placed off-slab, and we have enough space then
2419 * move it on-slab. This is at the expense of any extra colouring.
2421 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= slab_size
) {
2422 flags
&= ~CFLGS_OFF_SLAB
;
2423 left_over
-= slab_size
;
2426 if (flags
& CFLGS_OFF_SLAB
) {
2427 /* really off slab. No need for manual alignment */
2429 cachep
->num
* sizeof(kmem_bufctl_t
) + sizeof(struct slab
);
2431 #ifdef CONFIG_PAGE_POISONING
2432 /* If we're going to use the generic kernel_map_pages()
2433 * poisoning, then it's going to smash the contents of
2434 * the redzone and userword anyhow, so switch them off.
2436 if (size
% PAGE_SIZE
== 0 && flags
& SLAB_POISON
)
2437 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2441 cachep
->colour_off
= cache_line_size();
2442 /* Offset must be a multiple of the alignment. */
2443 if (cachep
->colour_off
< align
)
2444 cachep
->colour_off
= align
;
2445 cachep
->colour
= left_over
/ cachep
->colour_off
;
2446 cachep
->slab_size
= slab_size
;
2447 cachep
->flags
= flags
;
2448 cachep
->gfpflags
= 0;
2449 if (CONFIG_ZONE_DMA_FLAG
&& (flags
& SLAB_CACHE_DMA
))
2450 cachep
->gfpflags
|= GFP_DMA
;
2451 cachep
->buffer_size
= size
;
2452 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2454 if (flags
& CFLGS_OFF_SLAB
) {
2455 cachep
->slabp_cache
= kmem_find_general_cachep(slab_size
, 0u);
2457 * This is a possibility for one of the malloc_sizes caches.
2458 * But since we go off slab only for object size greater than
2459 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2460 * this should not happen at all.
2461 * But leave a BUG_ON for some lucky dude.
2463 BUG_ON(ZERO_OR_NULL_PTR(cachep
->slabp_cache
));
2465 cachep
->ctor
= ctor
;
2466 cachep
->name
= name
;
2468 if (setup_cpu_cache(cachep
, gfp
)) {
2469 __kmem_cache_destroy(cachep
);
2474 if (flags
& SLAB_DEBUG_OBJECTS
) {
2476 * Would deadlock through slab_destroy()->call_rcu()->
2477 * debug_object_activate()->kmem_cache_alloc().
2479 WARN_ON_ONCE(flags
& SLAB_DESTROY_BY_RCU
);
2481 slab_set_debugobj_lock_classes(cachep
);
2484 /* cache setup completed, link it into the list */
2485 list_add(&cachep
->next
, &cache_chain
);
2487 if (!cachep
&& (flags
& SLAB_PANIC
))
2488 panic("kmem_cache_create(): failed to create slab `%s'\n",
2490 if (slab_is_available()) {
2491 mutex_unlock(&cache_chain_mutex
);
2496 EXPORT_SYMBOL(kmem_cache_create
);
2499 static void check_irq_off(void)
2501 BUG_ON(!irqs_disabled());
2504 static void check_irq_on(void)
2506 BUG_ON(irqs_disabled());
2509 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2513 assert_spin_locked(&cachep
->nodelists
[numa_mem_id()]->list_lock
);
2517 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2521 assert_spin_locked(&cachep
->nodelists
[node
]->list_lock
);
2526 #define check_irq_off() do { } while(0)
2527 #define check_irq_on() do { } while(0)
2528 #define check_spinlock_acquired(x) do { } while(0)
2529 #define check_spinlock_acquired_node(x, y) do { } while(0)
2532 static void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
2533 struct array_cache
*ac
,
2534 int force
, int node
);
2536 static void do_drain(void *arg
)
2538 struct kmem_cache
*cachep
= arg
;
2539 struct array_cache
*ac
;
2540 int node
= numa_mem_id();
2543 ac
= cpu_cache_get(cachep
);
2544 spin_lock(&cachep
->nodelists
[node
]->list_lock
);
2545 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
2546 spin_unlock(&cachep
->nodelists
[node
]->list_lock
);
2550 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2552 struct kmem_list3
*l3
;
2555 on_each_cpu(do_drain
, cachep
, 1);
2557 for_each_online_node(node
) {
2558 l3
= cachep
->nodelists
[node
];
2559 if (l3
&& l3
->alien
)
2560 drain_alien_cache(cachep
, l3
->alien
);
2563 for_each_online_node(node
) {
2564 l3
= cachep
->nodelists
[node
];
2566 drain_array(cachep
, l3
, l3
->shared
, 1, node
);
2571 * Remove slabs from the list of free slabs.
2572 * Specify the number of slabs to drain in tofree.
2574 * Returns the actual number of slabs released.
2576 static int drain_freelist(struct kmem_cache
*cache
,
2577 struct kmem_list3
*l3
, int tofree
)
2579 struct list_head
*p
;
2584 while (nr_freed
< tofree
&& !list_empty(&l3
->slabs_free
)) {
2586 spin_lock_irq(&l3
->list_lock
);
2587 p
= l3
->slabs_free
.prev
;
2588 if (p
== &l3
->slabs_free
) {
2589 spin_unlock_irq(&l3
->list_lock
);
2593 slabp
= list_entry(p
, struct slab
, list
);
2595 BUG_ON(slabp
->inuse
);
2597 list_del(&slabp
->list
);
2599 * Safe to drop the lock. The slab is no longer linked
2602 l3
->free_objects
-= cache
->num
;
2603 spin_unlock_irq(&l3
->list_lock
);
2604 slab_destroy(cache
, slabp
);
2611 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2612 static int __cache_shrink(struct kmem_cache
*cachep
)
2615 struct kmem_list3
*l3
;
2617 drain_cpu_caches(cachep
);
2620 for_each_online_node(i
) {
2621 l3
= cachep
->nodelists
[i
];
2625 drain_freelist(cachep
, l3
, l3
->free_objects
);
2627 ret
+= !list_empty(&l3
->slabs_full
) ||
2628 !list_empty(&l3
->slabs_partial
);
2630 return (ret
? 1 : 0);
2634 * kmem_cache_shrink - Shrink a cache.
2635 * @cachep: The cache to shrink.
2637 * Releases as many slabs as possible for a cache.
2638 * To help debugging, a zero exit status indicates all slabs were released.
2640 int kmem_cache_shrink(struct kmem_cache
*cachep
)
2643 BUG_ON(!cachep
|| in_interrupt());
2646 mutex_lock(&cache_chain_mutex
);
2647 ret
= __cache_shrink(cachep
);
2648 mutex_unlock(&cache_chain_mutex
);
2652 EXPORT_SYMBOL(kmem_cache_shrink
);
2655 * kmem_cache_destroy - delete a cache
2656 * @cachep: the cache to destroy
2658 * Remove a &struct kmem_cache object from the slab cache.
2660 * It is expected this function will be called by a module when it is
2661 * unloaded. This will remove the cache completely, and avoid a duplicate
2662 * cache being allocated each time a module is loaded and unloaded, if the
2663 * module doesn't have persistent in-kernel storage across loads and unloads.
2665 * The cache must be empty before calling this function.
2667 * The caller must guarantee that no one will allocate memory from the cache
2668 * during the kmem_cache_destroy().
2670 void kmem_cache_destroy(struct kmem_cache
*cachep
)
2672 BUG_ON(!cachep
|| in_interrupt());
2674 /* Find the cache in the chain of caches. */
2676 mutex_lock(&cache_chain_mutex
);
2678 * the chain is never empty, cache_cache is never destroyed
2680 list_del(&cachep
->next
);
2681 if (__cache_shrink(cachep
)) {
2682 slab_error(cachep
, "Can't free all objects");
2683 list_add(&cachep
->next
, &cache_chain
);
2684 mutex_unlock(&cache_chain_mutex
);
2689 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
2692 __kmem_cache_destroy(cachep
);
2693 mutex_unlock(&cache_chain_mutex
);
2696 EXPORT_SYMBOL(kmem_cache_destroy
);
2699 * Get the memory for a slab management obj.
2700 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2701 * always come from malloc_sizes caches. The slab descriptor cannot
2702 * come from the same cache which is getting created because,
2703 * when we are searching for an appropriate cache for these
2704 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2705 * If we are creating a malloc_sizes cache here it would not be visible to
2706 * kmem_find_general_cachep till the initialization is complete.
2707 * Hence we cannot have slabp_cache same as the original cache.
2709 static struct slab
*alloc_slabmgmt(struct kmem_cache
*cachep
, void *objp
,
2710 int colour_off
, gfp_t local_flags
,
2715 if (OFF_SLAB(cachep
)) {
2716 /* Slab management obj is off-slab. */
2717 slabp
= kmem_cache_alloc_node(cachep
->slabp_cache
,
2718 local_flags
, nodeid
);
2720 * If the first object in the slab is leaked (it's allocated
2721 * but no one has a reference to it), we want to make sure
2722 * kmemleak does not treat the ->s_mem pointer as a reference
2723 * to the object. Otherwise we will not report the leak.
2725 kmemleak_scan_area(&slabp
->list
, sizeof(struct list_head
),
2730 slabp
= objp
+ colour_off
;
2731 colour_off
+= cachep
->slab_size
;
2734 slabp
->colouroff
= colour_off
;
2735 slabp
->s_mem
= objp
+ colour_off
;
2736 slabp
->nodeid
= nodeid
;
2741 static inline kmem_bufctl_t
*slab_bufctl(struct slab
*slabp
)
2743 return (kmem_bufctl_t
*) (slabp
+ 1);
2746 static void cache_init_objs(struct kmem_cache
*cachep
,
2751 for (i
= 0; i
< cachep
->num
; i
++) {
2752 void *objp
= index_to_obj(cachep
, slabp
, i
);
2754 /* need to poison the objs? */
2755 if (cachep
->flags
& SLAB_POISON
)
2756 poison_obj(cachep
, objp
, POISON_FREE
);
2757 if (cachep
->flags
& SLAB_STORE_USER
)
2758 *dbg_userword(cachep
, objp
) = NULL
;
2760 if (cachep
->flags
& SLAB_RED_ZONE
) {
2761 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2762 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2765 * Constructors are not allowed to allocate memory from the same
2766 * cache which they are a constructor for. Otherwise, deadlock.
2767 * They must also be threaded.
2769 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2770 cachep
->ctor(objp
+ obj_offset(cachep
));
2772 if (cachep
->flags
& SLAB_RED_ZONE
) {
2773 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2774 slab_error(cachep
, "constructor overwrote the"
2775 " end of an object");
2776 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2777 slab_error(cachep
, "constructor overwrote the"
2778 " start of an object");
2780 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 &&
2781 OFF_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
)
2782 kernel_map_pages(virt_to_page(objp
),
2783 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2788 slab_bufctl(slabp
)[i
] = i
+ 1;
2790 slab_bufctl(slabp
)[i
- 1] = BUFCTL_END
;
2793 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2795 if (CONFIG_ZONE_DMA_FLAG
) {
2796 if (flags
& GFP_DMA
)
2797 BUG_ON(!(cachep
->gfpflags
& GFP_DMA
));
2799 BUG_ON(cachep
->gfpflags
& GFP_DMA
);
2803 static void *slab_get_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2806 void *objp
= index_to_obj(cachep
, slabp
, slabp
->free
);
2810 next
= slab_bufctl(slabp
)[slabp
->free
];
2812 slab_bufctl(slabp
)[slabp
->free
] = BUFCTL_FREE
;
2813 WARN_ON(slabp
->nodeid
!= nodeid
);
2820 static void slab_put_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2821 void *objp
, int nodeid
)
2823 unsigned int objnr
= obj_to_index(cachep
, slabp
, objp
);
2826 /* Verify that the slab belongs to the intended node */
2827 WARN_ON(slabp
->nodeid
!= nodeid
);
2829 if (slab_bufctl(slabp
)[objnr
] + 1 <= SLAB_LIMIT
+ 1) {
2830 printk(KERN_ERR
"slab: double free detected in cache "
2831 "'%s', objp %p\n", cachep
->name
, objp
);
2835 slab_bufctl(slabp
)[objnr
] = slabp
->free
;
2836 slabp
->free
= objnr
;
2841 * Map pages beginning at addr to the given cache and slab. This is required
2842 * for the slab allocator to be able to lookup the cache and slab of a
2843 * virtual address for kfree, ksize, and slab debugging.
2845 static void slab_map_pages(struct kmem_cache
*cache
, struct slab
*slab
,
2851 page
= virt_to_page(addr
);
2854 if (likely(!PageCompound(page
)))
2855 nr_pages
<<= cache
->gfporder
;
2858 page_set_cache(page
, cache
);
2859 page_set_slab(page
, slab
);
2861 } while (--nr_pages
);
2865 * Grow (by 1) the number of slabs within a cache. This is called by
2866 * kmem_cache_alloc() when there are no active objs left in a cache.
2868 static int cache_grow(struct kmem_cache
*cachep
,
2869 gfp_t flags
, int nodeid
, void *objp
)
2874 struct kmem_list3
*l3
;
2877 * Be lazy and only check for valid flags here, keeping it out of the
2878 * critical path in kmem_cache_alloc().
2880 BUG_ON(flags
& GFP_SLAB_BUG_MASK
);
2881 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2883 /* Take the l3 list lock to change the colour_next on this node */
2885 l3
= cachep
->nodelists
[nodeid
];
2886 spin_lock(&l3
->list_lock
);
2888 /* Get colour for the slab, and cal the next value. */
2889 offset
= l3
->colour_next
;
2891 if (l3
->colour_next
>= cachep
->colour
)
2892 l3
->colour_next
= 0;
2893 spin_unlock(&l3
->list_lock
);
2895 offset
*= cachep
->colour_off
;
2897 if (local_flags
& __GFP_WAIT
)
2901 * The test for missing atomic flag is performed here, rather than
2902 * the more obvious place, simply to reduce the critical path length
2903 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2904 * will eventually be caught here (where it matters).
2906 kmem_flagcheck(cachep
, flags
);
2909 * Get mem for the objs. Attempt to allocate a physical page from
2913 objp
= kmem_getpages(cachep
, local_flags
, nodeid
);
2917 /* Get slab management. */
2918 slabp
= alloc_slabmgmt(cachep
, objp
, offset
,
2919 local_flags
& ~GFP_CONSTRAINT_MASK
, nodeid
);
2923 slab_map_pages(cachep
, slabp
, objp
);
2925 cache_init_objs(cachep
, slabp
);
2927 if (local_flags
& __GFP_WAIT
)
2928 local_irq_disable();
2930 spin_lock(&l3
->list_lock
);
2932 /* Make slab active. */
2933 list_add_tail(&slabp
->list
, &(l3
->slabs_free
));
2934 STATS_INC_GROWN(cachep
);
2935 l3
->free_objects
+= cachep
->num
;
2936 spin_unlock(&l3
->list_lock
);
2939 kmem_freepages(cachep
, objp
);
2941 if (local_flags
& __GFP_WAIT
)
2942 local_irq_disable();
2949 * Perform extra freeing checks:
2950 * - detect bad pointers.
2951 * - POISON/RED_ZONE checking
2953 static void kfree_debugcheck(const void *objp
)
2955 if (!virt_addr_valid(objp
)) {
2956 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
2957 (unsigned long)objp
);
2962 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2964 unsigned long long redzone1
, redzone2
;
2966 redzone1
= *dbg_redzone1(cache
, obj
);
2967 redzone2
= *dbg_redzone2(cache
, obj
);
2972 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2975 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2976 slab_error(cache
, "double free detected");
2978 slab_error(cache
, "memory outside object was overwritten");
2980 printk(KERN_ERR
"%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2981 obj
, redzone1
, redzone2
);
2984 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2991 BUG_ON(virt_to_cache(objp
) != cachep
);
2993 objp
-= obj_offset(cachep
);
2994 kfree_debugcheck(objp
);
2995 page
= virt_to_head_page(objp
);
2997 slabp
= page_get_slab(page
);
2999 if (cachep
->flags
& SLAB_RED_ZONE
) {
3000 verify_redzone_free(cachep
, objp
);
3001 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
3002 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
3004 if (cachep
->flags
& SLAB_STORE_USER
)
3005 *dbg_userword(cachep
, objp
) = caller
;
3007 objnr
= obj_to_index(cachep
, slabp
, objp
);
3009 BUG_ON(objnr
>= cachep
->num
);
3010 BUG_ON(objp
!= index_to_obj(cachep
, slabp
, objnr
));
3012 #ifdef CONFIG_DEBUG_SLAB_LEAK
3013 slab_bufctl(slabp
)[objnr
] = BUFCTL_FREE
;
3015 if (cachep
->flags
& SLAB_POISON
) {
3016 #ifdef CONFIG_DEBUG_PAGEALLOC
3017 if ((cachep
->buffer_size
% PAGE_SIZE
)==0 && OFF_SLAB(cachep
)) {
3018 store_stackinfo(cachep
, objp
, (unsigned long)caller
);
3019 kernel_map_pages(virt_to_page(objp
),
3020 cachep
->buffer_size
/ PAGE_SIZE
, 0);
3022 poison_obj(cachep
, objp
, POISON_FREE
);
3025 poison_obj(cachep
, objp
, POISON_FREE
);
3031 static void check_slabp(struct kmem_cache
*cachep
, struct slab
*slabp
)
3036 /* Check slab's freelist to see if this obj is there. */
3037 for (i
= slabp
->free
; i
!= BUFCTL_END
; i
= slab_bufctl(slabp
)[i
]) {
3039 if (entries
> cachep
->num
|| i
>= cachep
->num
)
3042 if (entries
!= cachep
->num
- slabp
->inuse
) {
3044 printk(KERN_ERR
"slab: Internal list corruption detected in "
3045 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
3046 cachep
->name
, cachep
->num
, slabp
, slabp
->inuse
);
3047 print_hex_dump(KERN_ERR
, "", DUMP_PREFIX_OFFSET
, 16, 1, slabp
,
3048 sizeof(*slabp
) + cachep
->num
* sizeof(kmem_bufctl_t
),
3054 #define kfree_debugcheck(x) do { } while(0)
3055 #define cache_free_debugcheck(x,objp,z) (objp)
3056 #define check_slabp(x,y) do { } while(0)
3059 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
3062 struct kmem_list3
*l3
;
3063 struct array_cache
*ac
;
3068 node
= numa_mem_id();
3069 ac
= cpu_cache_get(cachep
);
3070 batchcount
= ac
->batchcount
;
3071 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
3073 * If there was little recent activity on this cache, then
3074 * perform only a partial refill. Otherwise we could generate
3077 batchcount
= BATCHREFILL_LIMIT
;
3079 l3
= cachep
->nodelists
[node
];
3081 BUG_ON(ac
->avail
> 0 || !l3
);
3082 spin_lock(&l3
->list_lock
);
3084 /* See if we can refill from the shared array */
3085 if (l3
->shared
&& transfer_objects(ac
, l3
->shared
, batchcount
)) {
3086 l3
->shared
->touched
= 1;
3090 while (batchcount
> 0) {
3091 struct list_head
*entry
;
3093 /* Get slab alloc is to come from. */
3094 entry
= l3
->slabs_partial
.next
;
3095 if (entry
== &l3
->slabs_partial
) {
3096 l3
->free_touched
= 1;
3097 entry
= l3
->slabs_free
.next
;
3098 if (entry
== &l3
->slabs_free
)
3102 slabp
= list_entry(entry
, struct slab
, list
);
3103 check_slabp(cachep
, slabp
);
3104 check_spinlock_acquired(cachep
);
3107 * The slab was either on partial or free list so
3108 * there must be at least one object available for
3111 BUG_ON(slabp
->inuse
>= cachep
->num
);
3113 while (slabp
->inuse
< cachep
->num
&& batchcount
--) {
3114 STATS_INC_ALLOCED(cachep
);
3115 STATS_INC_ACTIVE(cachep
);
3116 STATS_SET_HIGH(cachep
);
3118 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, slabp
,
3121 check_slabp(cachep
, slabp
);
3123 /* move slabp to correct slabp list: */
3124 list_del(&slabp
->list
);
3125 if (slabp
->free
== BUFCTL_END
)
3126 list_add(&slabp
->list
, &l3
->slabs_full
);
3128 list_add(&slabp
->list
, &l3
->slabs_partial
);
3132 l3
->free_objects
-= ac
->avail
;
3134 spin_unlock(&l3
->list_lock
);
3136 if (unlikely(!ac
->avail
)) {
3138 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, node
, NULL
);
3140 /* cache_grow can reenable interrupts, then ac could change. */
3141 ac
= cpu_cache_get(cachep
);
3142 if (!x
&& ac
->avail
== 0) /* no objects in sight? abort */
3145 if (!ac
->avail
) /* objects refilled by interrupt? */
3149 return ac
->entry
[--ac
->avail
];
3152 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
3155 might_sleep_if(flags
& __GFP_WAIT
);
3157 kmem_flagcheck(cachep
, flags
);
3162 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
3163 gfp_t flags
, void *objp
, void *caller
)
3167 if (cachep
->flags
& SLAB_POISON
) {
3168 #ifdef CONFIG_DEBUG_PAGEALLOC
3169 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
3170 kernel_map_pages(virt_to_page(objp
),
3171 cachep
->buffer_size
/ PAGE_SIZE
, 1);
3173 check_poison_obj(cachep
, objp
);
3175 check_poison_obj(cachep
, objp
);
3177 poison_obj(cachep
, objp
, POISON_INUSE
);
3179 if (cachep
->flags
& SLAB_STORE_USER
)
3180 *dbg_userword(cachep
, objp
) = caller
;
3182 if (cachep
->flags
& SLAB_RED_ZONE
) {
3183 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
3184 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
3185 slab_error(cachep
, "double free, or memory outside"
3186 " object was overwritten");
3188 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3189 objp
, *dbg_redzone1(cachep
, objp
),
3190 *dbg_redzone2(cachep
, objp
));
3192 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
3193 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
3195 #ifdef CONFIG_DEBUG_SLAB_LEAK
3200 slabp
= page_get_slab(virt_to_head_page(objp
));
3201 objnr
= (unsigned)(objp
- slabp
->s_mem
) / cachep
->buffer_size
;
3202 slab_bufctl(slabp
)[objnr
] = BUFCTL_ACTIVE
;
3205 objp
+= obj_offset(cachep
);
3206 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
3208 if (ARCH_SLAB_MINALIGN
&&
3209 ((unsigned long)objp
& (ARCH_SLAB_MINALIGN
-1))) {
3210 printk(KERN_ERR
"0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3211 objp
, (int)ARCH_SLAB_MINALIGN
);
3216 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3219 static bool slab_should_failslab(struct kmem_cache
*cachep
, gfp_t flags
)
3221 if (cachep
== &cache_cache
)
3224 return should_failslab(obj_size(cachep
), flags
, cachep
->flags
);
3227 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3230 struct array_cache
*ac
;
3234 ac
= cpu_cache_get(cachep
);
3235 if (likely(ac
->avail
)) {
3236 STATS_INC_ALLOCHIT(cachep
);
3238 objp
= ac
->entry
[--ac
->avail
];
3240 STATS_INC_ALLOCMISS(cachep
);
3241 objp
= cache_alloc_refill(cachep
, flags
);
3243 * the 'ac' may be updated by cache_alloc_refill(),
3244 * and kmemleak_erase() requires its correct value.
3246 ac
= cpu_cache_get(cachep
);
3249 * To avoid a false negative, if an object that is in one of the
3250 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3251 * treat the array pointers as a reference to the object.
3254 kmemleak_erase(&ac
->entry
[ac
->avail
]);
3260 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3262 * If we are in_interrupt, then process context, including cpusets and
3263 * mempolicy, may not apply and should not be used for allocation policy.
3265 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3267 int nid_alloc
, nid_here
;
3269 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3271 nid_alloc
= nid_here
= numa_mem_id();
3273 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3274 nid_alloc
= cpuset_slab_spread_node();
3275 else if (current
->mempolicy
)
3276 nid_alloc
= slab_node(current
->mempolicy
);
3278 if (nid_alloc
!= nid_here
)
3279 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3284 * Fallback function if there was no memory available and no objects on a
3285 * certain node and fall back is permitted. First we scan all the
3286 * available nodelists for available objects. If that fails then we
3287 * perform an allocation without specifying a node. This allows the page
3288 * allocator to do its reclaim / fallback magic. We then insert the
3289 * slab into the proper nodelist and then allocate from it.
3291 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3293 struct zonelist
*zonelist
;
3297 enum zone_type high_zoneidx
= gfp_zone(flags
);
3301 if (flags
& __GFP_THISNODE
)
3305 zonelist
= node_zonelist(slab_node(current
->mempolicy
), flags
);
3306 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
3310 * Look through allowed nodes for objects available
3311 * from existing per node queues.
3313 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
3314 nid
= zone_to_nid(zone
);
3316 if (cpuset_zone_allowed_hardwall(zone
, flags
) &&
3317 cache
->nodelists
[nid
] &&
3318 cache
->nodelists
[nid
]->free_objects
) {
3319 obj
= ____cache_alloc_node(cache
,
3320 flags
| GFP_THISNODE
, nid
);
3328 * This allocation will be performed within the constraints
3329 * of the current cpuset / memory policy requirements.
3330 * We may trigger various forms of reclaim on the allowed
3331 * set and go into memory reserves if necessary.
3333 if (local_flags
& __GFP_WAIT
)
3335 kmem_flagcheck(cache
, flags
);
3336 obj
= kmem_getpages(cache
, local_flags
, numa_mem_id());
3337 if (local_flags
& __GFP_WAIT
)
3338 local_irq_disable();
3341 * Insert into the appropriate per node queues
3343 nid
= page_to_nid(virt_to_page(obj
));
3344 if (cache_grow(cache
, flags
, nid
, obj
)) {
3345 obj
= ____cache_alloc_node(cache
,
3346 flags
| GFP_THISNODE
, nid
);
3349 * Another processor may allocate the
3350 * objects in the slab since we are
3351 * not holding any locks.
3355 /* cache_grow already freed obj */
3365 * A interface to enable slab creation on nodeid
3367 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3370 struct list_head
*entry
;
3372 struct kmem_list3
*l3
;
3376 l3
= cachep
->nodelists
[nodeid
];
3381 spin_lock(&l3
->list_lock
);
3382 entry
= l3
->slabs_partial
.next
;
3383 if (entry
== &l3
->slabs_partial
) {
3384 l3
->free_touched
= 1;
3385 entry
= l3
->slabs_free
.next
;
3386 if (entry
== &l3
->slabs_free
)
3390 slabp
= list_entry(entry
, struct slab
, list
);
3391 check_spinlock_acquired_node(cachep
, nodeid
);
3392 check_slabp(cachep
, slabp
);
3394 STATS_INC_NODEALLOCS(cachep
);
3395 STATS_INC_ACTIVE(cachep
);
3396 STATS_SET_HIGH(cachep
);
3398 BUG_ON(slabp
->inuse
== cachep
->num
);
3400 obj
= slab_get_obj(cachep
, slabp
, nodeid
);
3401 check_slabp(cachep
, slabp
);
3403 /* move slabp to correct slabp list: */
3404 list_del(&slabp
->list
);
3406 if (slabp
->free
== BUFCTL_END
)
3407 list_add(&slabp
->list
, &l3
->slabs_full
);
3409 list_add(&slabp
->list
, &l3
->slabs_partial
);
3411 spin_unlock(&l3
->list_lock
);
3415 spin_unlock(&l3
->list_lock
);
3416 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, nodeid
, NULL
);
3420 return fallback_alloc(cachep
, flags
);
3427 * kmem_cache_alloc_node - Allocate an object on the specified node
3428 * @cachep: The cache to allocate from.
3429 * @flags: See kmalloc().
3430 * @nodeid: node number of the target node.
3431 * @caller: return address of caller, used for debug information
3433 * Identical to kmem_cache_alloc but it will allocate memory on the given
3434 * node, which can improve the performance for cpu bound structures.
3436 * Fallback to other node is possible if __GFP_THISNODE is not set.
3438 static __always_inline
void *
3439 __cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3442 unsigned long save_flags
;
3444 int slab_node
= numa_mem_id();
3446 flags
&= gfp_allowed_mask
;
3448 lockdep_trace_alloc(flags
);
3450 if (slab_should_failslab(cachep
, flags
))
3453 cache_alloc_debugcheck_before(cachep
, flags
);
3454 local_irq_save(save_flags
);
3456 if (nodeid
== NUMA_NO_NODE
)
3459 if (unlikely(!cachep
->nodelists
[nodeid
])) {
3460 /* Node not bootstrapped yet */
3461 ptr
= fallback_alloc(cachep
, flags
);
3465 if (nodeid
== slab_node
) {
3467 * Use the locally cached objects if possible.
3468 * However ____cache_alloc does not allow fallback
3469 * to other nodes. It may fail while we still have
3470 * objects on other nodes available.
3472 ptr
= ____cache_alloc(cachep
, flags
);
3476 /* ___cache_alloc_node can fall back to other nodes */
3477 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3479 local_irq_restore(save_flags
);
3480 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3481 kmemleak_alloc_recursive(ptr
, obj_size(cachep
), 1, cachep
->flags
,
3485 kmemcheck_slab_alloc(cachep
, flags
, ptr
, obj_size(cachep
));
3487 if (unlikely((flags
& __GFP_ZERO
) && ptr
))
3488 memset(ptr
, 0, obj_size(cachep
));
3493 static __always_inline
void *
3494 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3498 if (unlikely(current
->flags
& (PF_SPREAD_SLAB
| PF_MEMPOLICY
))) {
3499 objp
= alternate_node_alloc(cache
, flags
);
3503 objp
= ____cache_alloc(cache
, flags
);
3506 * We may just have run out of memory on the local node.
3507 * ____cache_alloc_node() knows how to locate memory on other nodes
3510 objp
= ____cache_alloc_node(cache
, flags
, numa_mem_id());
3517 static __always_inline
void *
3518 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3520 return ____cache_alloc(cachep
, flags
);
3523 #endif /* CONFIG_NUMA */
3525 static __always_inline
void *
3526 __cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
, void *caller
)
3528 unsigned long save_flags
;
3531 flags
&= gfp_allowed_mask
;
3533 lockdep_trace_alloc(flags
);
3535 if (slab_should_failslab(cachep
, flags
))
3538 cache_alloc_debugcheck_before(cachep
, flags
);
3539 local_irq_save(save_flags
);
3540 objp
= __do_cache_alloc(cachep
, flags
);
3541 local_irq_restore(save_flags
);
3542 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3543 kmemleak_alloc_recursive(objp
, obj_size(cachep
), 1, cachep
->flags
,
3548 kmemcheck_slab_alloc(cachep
, flags
, objp
, obj_size(cachep
));
3550 if (unlikely((flags
& __GFP_ZERO
) && objp
))
3551 memset(objp
, 0, obj_size(cachep
));
3557 * Caller needs to acquire correct kmem_list's list_lock
3559 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int nr_objects
,
3563 struct kmem_list3
*l3
;
3565 for (i
= 0; i
< nr_objects
; i
++) {
3566 void *objp
= objpp
[i
];
3569 slabp
= virt_to_slab(objp
);
3570 l3
= cachep
->nodelists
[node
];
3571 list_del(&slabp
->list
);
3572 check_spinlock_acquired_node(cachep
, node
);
3573 check_slabp(cachep
, slabp
);
3574 slab_put_obj(cachep
, slabp
, objp
, node
);
3575 STATS_DEC_ACTIVE(cachep
);
3577 check_slabp(cachep
, slabp
);
3579 /* fixup slab chains */
3580 if (slabp
->inuse
== 0) {
3581 if (l3
->free_objects
> l3
->free_limit
) {
3582 l3
->free_objects
-= cachep
->num
;
3583 /* No need to drop any previously held
3584 * lock here, even if we have a off-slab slab
3585 * descriptor it is guaranteed to come from
3586 * a different cache, refer to comments before
3589 slab_destroy(cachep
, slabp
);
3591 list_add(&slabp
->list
, &l3
->slabs_free
);
3594 /* Unconditionally move a slab to the end of the
3595 * partial list on free - maximum time for the
3596 * other objects to be freed, too.
3598 list_add_tail(&slabp
->list
, &l3
->slabs_partial
);
3603 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3606 struct kmem_list3
*l3
;
3607 int node
= numa_mem_id();
3609 batchcount
= ac
->batchcount
;
3611 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
3614 l3
= cachep
->nodelists
[node
];
3615 spin_lock(&l3
->list_lock
);
3617 struct array_cache
*shared_array
= l3
->shared
;
3618 int max
= shared_array
->limit
- shared_array
->avail
;
3620 if (batchcount
> max
)
3622 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3623 ac
->entry
, sizeof(void *) * batchcount
);
3624 shared_array
->avail
+= batchcount
;
3629 free_block(cachep
, ac
->entry
, batchcount
, node
);
3634 struct list_head
*p
;
3636 p
= l3
->slabs_free
.next
;
3637 while (p
!= &(l3
->slabs_free
)) {
3640 slabp
= list_entry(p
, struct slab
, list
);
3641 BUG_ON(slabp
->inuse
);
3646 STATS_SET_FREEABLE(cachep
, i
);
3649 spin_unlock(&l3
->list_lock
);
3650 ac
->avail
-= batchcount
;
3651 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3655 * Release an obj back to its cache. If the obj has a constructed state, it must
3656 * be in this state _before_ it is released. Called with disabled ints.
3658 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
,
3661 struct array_cache
*ac
= cpu_cache_get(cachep
);
3664 kmemleak_free_recursive(objp
, cachep
->flags
);
3665 objp
= cache_free_debugcheck(cachep
, objp
, caller
);
3667 kmemcheck_slab_free(cachep
, objp
, obj_size(cachep
));
3670 * Skip calling cache_free_alien() when the platform is not numa.
3671 * This will avoid cache misses that happen while accessing slabp (which
3672 * is per page memory reference) to get nodeid. Instead use a global
3673 * variable to skip the call, which is mostly likely to be present in
3676 if (nr_online_nodes
> 1 && cache_free_alien(cachep
, objp
))
3679 if (likely(ac
->avail
< ac
->limit
)) {
3680 STATS_INC_FREEHIT(cachep
);
3681 ac
->entry
[ac
->avail
++] = objp
;
3684 STATS_INC_FREEMISS(cachep
);
3685 cache_flusharray(cachep
, ac
);
3686 ac
->entry
[ac
->avail
++] = objp
;
3691 * kmem_cache_alloc - Allocate an object
3692 * @cachep: The cache to allocate from.
3693 * @flags: See kmalloc().
3695 * Allocate an object from this cache. The flags are only relevant
3696 * if the cache has no available objects.
3698 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3700 void *ret
= __cache_alloc(cachep
, flags
, __builtin_return_address(0));
3702 trace_kmem_cache_alloc(_RET_IP_
, ret
,
3703 obj_size(cachep
), cachep
->buffer_size
, flags
);
3707 EXPORT_SYMBOL(kmem_cache_alloc
);
3709 #ifdef CONFIG_TRACING
3711 kmem_cache_alloc_trace(size_t size
, struct kmem_cache
*cachep
, gfp_t flags
)
3715 ret
= __cache_alloc(cachep
, flags
, __builtin_return_address(0));
3717 trace_kmalloc(_RET_IP_
, ret
,
3718 size
, slab_buffer_size(cachep
), flags
);
3721 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
3725 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3727 void *ret
= __cache_alloc_node(cachep
, flags
, nodeid
,
3728 __builtin_return_address(0));
3730 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
3731 obj_size(cachep
), cachep
->buffer_size
,
3736 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3738 #ifdef CONFIG_TRACING
3739 void *kmem_cache_alloc_node_trace(size_t size
,
3740 struct kmem_cache
*cachep
,
3746 ret
= __cache_alloc_node(cachep
, flags
, nodeid
,
3747 __builtin_return_address(0));
3748 trace_kmalloc_node(_RET_IP_
, ret
,
3749 size
, slab_buffer_size(cachep
),
3753 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
3756 static __always_inline
void *
3757 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, void *caller
)
3759 struct kmem_cache
*cachep
;
3761 cachep
= kmem_find_general_cachep(size
, flags
);
3762 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3764 return kmem_cache_alloc_node_trace(size
, cachep
, flags
, node
);
3767 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3768 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3770 return __do_kmalloc_node(size
, flags
, node
,
3771 __builtin_return_address(0));
3773 EXPORT_SYMBOL(__kmalloc_node
);
3775 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3776 int node
, unsigned long caller
)
3778 return __do_kmalloc_node(size
, flags
, node
, (void *)caller
);
3780 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3782 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3784 return __do_kmalloc_node(size
, flags
, node
, NULL
);
3786 EXPORT_SYMBOL(__kmalloc_node
);
3787 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3788 #endif /* CONFIG_NUMA */
3791 * __do_kmalloc - allocate memory
3792 * @size: how many bytes of memory are required.
3793 * @flags: the type of memory to allocate (see kmalloc).
3794 * @caller: function caller for debug tracking of the caller
3796 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3799 struct kmem_cache
*cachep
;
3802 /* If you want to save a few bytes .text space: replace
3804 * Then kmalloc uses the uninlined functions instead of the inline
3807 cachep
= __find_general_cachep(size
, flags
);
3808 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3810 ret
= __cache_alloc(cachep
, flags
, caller
);
3812 trace_kmalloc((unsigned long) caller
, ret
,
3813 size
, cachep
->buffer_size
, flags
);
3819 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3820 void *__kmalloc(size_t size
, gfp_t flags
)
3822 return __do_kmalloc(size
, flags
, __builtin_return_address(0));
3824 EXPORT_SYMBOL(__kmalloc
);
3826 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, unsigned long caller
)
3828 return __do_kmalloc(size
, flags
, (void *)caller
);
3830 EXPORT_SYMBOL(__kmalloc_track_caller
);
3833 void *__kmalloc(size_t size
, gfp_t flags
)
3835 return __do_kmalloc(size
, flags
, NULL
);
3837 EXPORT_SYMBOL(__kmalloc
);
3841 * kmem_cache_free - Deallocate an object
3842 * @cachep: The cache the allocation was from.
3843 * @objp: The previously allocated object.
3845 * Free an object which was previously allocated from this
3848 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3850 unsigned long flags
;
3852 local_irq_save(flags
);
3853 debug_check_no_locks_freed(objp
, obj_size(cachep
));
3854 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3855 debug_check_no_obj_freed(objp
, obj_size(cachep
));
3856 __cache_free(cachep
, objp
, __builtin_return_address(0));
3857 local_irq_restore(flags
);
3859 trace_kmem_cache_free(_RET_IP_
, objp
);
3861 EXPORT_SYMBOL(kmem_cache_free
);
3864 * kfree - free previously allocated memory
3865 * @objp: pointer returned by kmalloc.
3867 * If @objp is NULL, no operation is performed.
3869 * Don't free memory not originally allocated by kmalloc()
3870 * or you will run into trouble.
3872 void kfree(const void *objp
)
3874 struct kmem_cache
*c
;
3875 unsigned long flags
;
3877 trace_kfree(_RET_IP_
, objp
);
3879 if (unlikely(ZERO_OR_NULL_PTR(objp
)))
3881 local_irq_save(flags
);
3882 kfree_debugcheck(objp
);
3883 c
= virt_to_cache(objp
);
3884 debug_check_no_locks_freed(objp
, obj_size(c
));
3885 debug_check_no_obj_freed(objp
, obj_size(c
));
3886 __cache_free(c
, (void *)objp
, __builtin_return_address(0));
3887 local_irq_restore(flags
);
3889 EXPORT_SYMBOL(kfree
);
3891 unsigned int kmem_cache_size(struct kmem_cache
*cachep
)
3893 return obj_size(cachep
);
3895 EXPORT_SYMBOL(kmem_cache_size
);
3898 * This initializes kmem_list3 or resizes various caches for all nodes.
3900 static int alloc_kmemlist(struct kmem_cache
*cachep
, gfp_t gfp
)
3903 struct kmem_list3
*l3
;
3904 struct array_cache
*new_shared
;
3905 struct array_cache
**new_alien
= NULL
;
3907 for_each_online_node(node
) {
3909 if (use_alien_caches
) {
3910 new_alien
= alloc_alien_cache(node
, cachep
->limit
, gfp
);
3916 if (cachep
->shared
) {
3917 new_shared
= alloc_arraycache(node
,
3918 cachep
->shared
*cachep
->batchcount
,
3921 free_alien_cache(new_alien
);
3926 l3
= cachep
->nodelists
[node
];
3928 struct array_cache
*shared
= l3
->shared
;
3930 spin_lock_irq(&l3
->list_lock
);
3933 free_block(cachep
, shared
->entry
,
3934 shared
->avail
, node
);
3936 l3
->shared
= new_shared
;
3938 l3
->alien
= new_alien
;
3941 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3942 cachep
->batchcount
+ cachep
->num
;
3943 spin_unlock_irq(&l3
->list_lock
);
3945 free_alien_cache(new_alien
);
3948 l3
= kmalloc_node(sizeof(struct kmem_list3
), gfp
, node
);
3950 free_alien_cache(new_alien
);
3955 kmem_list3_init(l3
);
3956 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
3957 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
3958 l3
->shared
= new_shared
;
3959 l3
->alien
= new_alien
;
3960 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3961 cachep
->batchcount
+ cachep
->num
;
3962 cachep
->nodelists
[node
] = l3
;
3967 if (!cachep
->next
.next
) {
3968 /* Cache is not active yet. Roll back what we did */
3971 if (cachep
->nodelists
[node
]) {
3972 l3
= cachep
->nodelists
[node
];
3975 free_alien_cache(l3
->alien
);
3977 cachep
->nodelists
[node
] = NULL
;
3985 struct ccupdate_struct
{
3986 struct kmem_cache
*cachep
;
3987 struct array_cache
*new[0];
3990 static void do_ccupdate_local(void *info
)
3992 struct ccupdate_struct
*new = info
;
3993 struct array_cache
*old
;
3996 old
= cpu_cache_get(new->cachep
);
3998 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
3999 new->new[smp_processor_id()] = old
;
4002 /* Always called with the cache_chain_mutex held */
4003 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
4004 int batchcount
, int shared
, gfp_t gfp
)
4006 struct ccupdate_struct
*new;
4009 new = kzalloc(sizeof(*new) + nr_cpu_ids
* sizeof(struct array_cache
*),
4014 for_each_online_cpu(i
) {
4015 new->new[i
] = alloc_arraycache(cpu_to_mem(i
), limit
,
4018 for (i
--; i
>= 0; i
--)
4024 new->cachep
= cachep
;
4026 on_each_cpu(do_ccupdate_local
, (void *)new, 1);
4029 cachep
->batchcount
= batchcount
;
4030 cachep
->limit
= limit
;
4031 cachep
->shared
= shared
;
4033 for_each_online_cpu(i
) {
4034 struct array_cache
*ccold
= new->new[i
];
4037 spin_lock_irq(&cachep
->nodelists
[cpu_to_mem(i
)]->list_lock
);
4038 free_block(cachep
, ccold
->entry
, ccold
->avail
, cpu_to_mem(i
));
4039 spin_unlock_irq(&cachep
->nodelists
[cpu_to_mem(i
)]->list_lock
);
4043 return alloc_kmemlist(cachep
, gfp
);
4046 /* Called with cache_chain_mutex held always */
4047 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
4053 * The head array serves three purposes:
4054 * - create a LIFO ordering, i.e. return objects that are cache-warm
4055 * - reduce the number of spinlock operations.
4056 * - reduce the number of linked list operations on the slab and
4057 * bufctl chains: array operations are cheaper.
4058 * The numbers are guessed, we should auto-tune as described by
4061 if (cachep
->buffer_size
> 131072)
4063 else if (cachep
->buffer_size
> PAGE_SIZE
)
4065 else if (cachep
->buffer_size
> 1024)
4067 else if (cachep
->buffer_size
> 256)
4073 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4074 * allocation behaviour: Most allocs on one cpu, most free operations
4075 * on another cpu. For these cases, an efficient object passing between
4076 * cpus is necessary. This is provided by a shared array. The array
4077 * replaces Bonwick's magazine layer.
4078 * On uniprocessor, it's functionally equivalent (but less efficient)
4079 * to a larger limit. Thus disabled by default.
4082 if (cachep
->buffer_size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
4087 * With debugging enabled, large batchcount lead to excessively long
4088 * periods with disabled local interrupts. Limit the batchcount
4093 err
= do_tune_cpucache(cachep
, limit
, (limit
+ 1) / 2, shared
, gfp
);
4095 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
4096 cachep
->name
, -err
);
4101 * Drain an array if it contains any elements taking the l3 lock only if
4102 * necessary. Note that the l3 listlock also protects the array_cache
4103 * if drain_array() is used on the shared array.
4105 static void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
4106 struct array_cache
*ac
, int force
, int node
)
4110 if (!ac
|| !ac
->avail
)
4112 if (ac
->touched
&& !force
) {
4115 spin_lock_irq(&l3
->list_lock
);
4117 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
4118 if (tofree
> ac
->avail
)
4119 tofree
= (ac
->avail
+ 1) / 2;
4120 free_block(cachep
, ac
->entry
, tofree
, node
);
4121 ac
->avail
-= tofree
;
4122 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
4123 sizeof(void *) * ac
->avail
);
4125 spin_unlock_irq(&l3
->list_lock
);
4130 * cache_reap - Reclaim memory from caches.
4131 * @w: work descriptor
4133 * Called from workqueue/eventd every few seconds.
4135 * - clear the per-cpu caches for this CPU.
4136 * - return freeable pages to the main free memory pool.
4138 * If we cannot acquire the cache chain mutex then just give up - we'll try
4139 * again on the next iteration.
4141 static void cache_reap(struct work_struct
*w
)
4143 struct kmem_cache
*searchp
;
4144 struct kmem_list3
*l3
;
4145 int node
= numa_mem_id();
4146 struct delayed_work
*work
= to_delayed_work(w
);
4148 if (!mutex_trylock(&cache_chain_mutex
))
4149 /* Give up. Setup the next iteration. */
4152 list_for_each_entry(searchp
, &cache_chain
, next
) {
4156 * We only take the l3 lock if absolutely necessary and we
4157 * have established with reasonable certainty that
4158 * we can do some work if the lock was obtained.
4160 l3
= searchp
->nodelists
[node
];
4162 reap_alien(searchp
, l3
);
4164 drain_array(searchp
, l3
, cpu_cache_get(searchp
), 0, node
);
4167 * These are racy checks but it does not matter
4168 * if we skip one check or scan twice.
4170 if (time_after(l3
->next_reap
, jiffies
))
4173 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
;
4175 drain_array(searchp
, l3
, l3
->shared
, 0, node
);
4177 if (l3
->free_touched
)
4178 l3
->free_touched
= 0;
4182 freed
= drain_freelist(searchp
, l3
, (l3
->free_limit
+
4183 5 * searchp
->num
- 1) / (5 * searchp
->num
));
4184 STATS_ADD_REAPED(searchp
, freed
);
4190 mutex_unlock(&cache_chain_mutex
);
4193 /* Set up the next iteration */
4194 schedule_delayed_work(work
, round_jiffies_relative(REAPTIMEOUT_CPUC
));
4197 #ifdef CONFIG_SLABINFO
4199 static void print_slabinfo_header(struct seq_file
*m
)
4202 * Output format version, so at least we can change it
4203 * without _too_ many complaints.
4206 seq_puts(m
, "slabinfo - version: 2.1 (statistics)\n");
4208 seq_puts(m
, "slabinfo - version: 2.1\n");
4210 seq_puts(m
, "# name <active_objs> <num_objs> <objsize> "
4211 "<objperslab> <pagesperslab>");
4212 seq_puts(m
, " : tunables <limit> <batchcount> <sharedfactor>");
4213 seq_puts(m
, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4215 seq_puts(m
, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4216 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4217 seq_puts(m
, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4222 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
4226 mutex_lock(&cache_chain_mutex
);
4228 print_slabinfo_header(m
);
4230 return seq_list_start(&cache_chain
, *pos
);
4233 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
4235 return seq_list_next(p
, &cache_chain
, pos
);
4238 static void s_stop(struct seq_file
*m
, void *p
)
4240 mutex_unlock(&cache_chain_mutex
);
4243 static int s_show(struct seq_file
*m
, void *p
)
4245 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, next
);
4247 unsigned long active_objs
;
4248 unsigned long num_objs
;
4249 unsigned long active_slabs
= 0;
4250 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
4254 struct kmem_list3
*l3
;
4258 for_each_online_node(node
) {
4259 l3
= cachep
->nodelists
[node
];
4264 spin_lock_irq(&l3
->list_lock
);
4266 list_for_each_entry(slabp
, &l3
->slabs_full
, list
) {
4267 if (slabp
->inuse
!= cachep
->num
&& !error
)
4268 error
= "slabs_full accounting error";
4269 active_objs
+= cachep
->num
;
4272 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
) {
4273 if (slabp
->inuse
== cachep
->num
&& !error
)
4274 error
= "slabs_partial inuse accounting error";
4275 if (!slabp
->inuse
&& !error
)
4276 error
= "slabs_partial/inuse accounting error";
4277 active_objs
+= slabp
->inuse
;
4280 list_for_each_entry(slabp
, &l3
->slabs_free
, list
) {
4281 if (slabp
->inuse
&& !error
)
4282 error
= "slabs_free/inuse accounting error";
4285 free_objects
+= l3
->free_objects
;
4287 shared_avail
+= l3
->shared
->avail
;
4289 spin_unlock_irq(&l3
->list_lock
);
4291 num_slabs
+= active_slabs
;
4292 num_objs
= num_slabs
* cachep
->num
;
4293 if (num_objs
- active_objs
!= free_objects
&& !error
)
4294 error
= "free_objects accounting error";
4296 name
= cachep
->name
;
4298 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
4300 seq_printf(m
, "%-17s %6lu %6lu %6u %4u %4d",
4301 name
, active_objs
, num_objs
, cachep
->buffer_size
,
4302 cachep
->num
, (1 << cachep
->gfporder
));
4303 seq_printf(m
, " : tunables %4u %4u %4u",
4304 cachep
->limit
, cachep
->batchcount
, cachep
->shared
);
4305 seq_printf(m
, " : slabdata %6lu %6lu %6lu",
4306 active_slabs
, num_slabs
, shared_avail
);
4309 unsigned long high
= cachep
->high_mark
;
4310 unsigned long allocs
= cachep
->num_allocations
;
4311 unsigned long grown
= cachep
->grown
;
4312 unsigned long reaped
= cachep
->reaped
;
4313 unsigned long errors
= cachep
->errors
;
4314 unsigned long max_freeable
= cachep
->max_freeable
;
4315 unsigned long node_allocs
= cachep
->node_allocs
;
4316 unsigned long node_frees
= cachep
->node_frees
;
4317 unsigned long overflows
= cachep
->node_overflow
;
4319 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu "
4320 "%4lu %4lu %4lu %4lu %4lu",
4321 allocs
, high
, grown
,
4322 reaped
, errors
, max_freeable
, node_allocs
,
4323 node_frees
, overflows
);
4327 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4328 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4329 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4330 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4332 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4333 allochit
, allocmiss
, freehit
, freemiss
);
4341 * slabinfo_op - iterator that generates /proc/slabinfo
4350 * num-pages-per-slab
4351 * + further values on SMP and with statistics enabled
4354 static const struct seq_operations slabinfo_op
= {
4361 #define MAX_SLABINFO_WRITE 128
4363 * slabinfo_write - Tuning for the slab allocator
4365 * @buffer: user buffer
4366 * @count: data length
4369 static ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
4370 size_t count
, loff_t
*ppos
)
4372 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4373 int limit
, batchcount
, shared
, res
;
4374 struct kmem_cache
*cachep
;
4376 if (count
> MAX_SLABINFO_WRITE
)
4378 if (copy_from_user(&kbuf
, buffer
, count
))
4380 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4382 tmp
= strchr(kbuf
, ' ');
4387 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4390 /* Find the cache in the chain of caches. */
4391 mutex_lock(&cache_chain_mutex
);
4393 list_for_each_entry(cachep
, &cache_chain
, next
) {
4394 if (!strcmp(cachep
->name
, kbuf
)) {
4395 if (limit
< 1 || batchcount
< 1 ||
4396 batchcount
> limit
|| shared
< 0) {
4399 res
= do_tune_cpucache(cachep
, limit
,
4406 mutex_unlock(&cache_chain_mutex
);
4412 static int slabinfo_open(struct inode
*inode
, struct file
*file
)
4414 return seq_open(file
, &slabinfo_op
);
4417 static const struct file_operations proc_slabinfo_operations
= {
4418 .open
= slabinfo_open
,
4420 .write
= slabinfo_write
,
4421 .llseek
= seq_lseek
,
4422 .release
= seq_release
,
4425 #ifdef CONFIG_DEBUG_SLAB_LEAK
4427 static void *leaks_start(struct seq_file
*m
, loff_t
*pos
)
4429 mutex_lock(&cache_chain_mutex
);
4430 return seq_list_start(&cache_chain
, *pos
);
4433 static inline int add_caller(unsigned long *n
, unsigned long v
)
4443 unsigned long *q
= p
+ 2 * i
;
4457 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4463 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
, struct slab
*s
)
4469 for (i
= 0, p
= s
->s_mem
; i
< c
->num
; i
++, p
+= c
->buffer_size
) {
4470 if (slab_bufctl(s
)[i
] != BUFCTL_ACTIVE
)
4472 if (!add_caller(n
, (unsigned long)*dbg_userword(c
, p
)))
4477 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4479 #ifdef CONFIG_KALLSYMS
4480 unsigned long offset
, size
;
4481 char modname
[MODULE_NAME_LEN
], name
[KSYM_NAME_LEN
];
4483 if (lookup_symbol_attrs(address
, &size
, &offset
, modname
, name
) == 0) {
4484 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4486 seq_printf(m
, " [%s]", modname
);
4490 seq_printf(m
, "%p", (void *)address
);
4493 static int leaks_show(struct seq_file
*m
, void *p
)
4495 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, next
);
4497 struct kmem_list3
*l3
;
4499 unsigned long *n
= m
->private;
4503 if (!(cachep
->flags
& SLAB_STORE_USER
))
4505 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4508 /* OK, we can do it */
4512 for_each_online_node(node
) {
4513 l3
= cachep
->nodelists
[node
];
4518 spin_lock_irq(&l3
->list_lock
);
4520 list_for_each_entry(slabp
, &l3
->slabs_full
, list
)
4521 handle_slab(n
, cachep
, slabp
);
4522 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
)
4523 handle_slab(n
, cachep
, slabp
);
4524 spin_unlock_irq(&l3
->list_lock
);
4526 name
= cachep
->name
;
4528 /* Increase the buffer size */
4529 mutex_unlock(&cache_chain_mutex
);
4530 m
->private = kzalloc(n
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4532 /* Too bad, we are really out */
4534 mutex_lock(&cache_chain_mutex
);
4537 *(unsigned long *)m
->private = n
[0] * 2;
4539 mutex_lock(&cache_chain_mutex
);
4540 /* Now make sure this entry will be retried */
4544 for (i
= 0; i
< n
[1]; i
++) {
4545 seq_printf(m
, "%s: %lu ", name
, n
[2*i
+3]);
4546 show_symbol(m
, n
[2*i
+2]);
4553 static const struct seq_operations slabstats_op
= {
4554 .start
= leaks_start
,
4560 static int slabstats_open(struct inode
*inode
, struct file
*file
)
4562 unsigned long *n
= kzalloc(PAGE_SIZE
, GFP_KERNEL
);
4565 ret
= seq_open(file
, &slabstats_op
);
4567 struct seq_file
*m
= file
->private_data
;
4568 *n
= PAGE_SIZE
/ (2 * sizeof(unsigned long));
4577 static const struct file_operations proc_slabstats_operations
= {
4578 .open
= slabstats_open
,
4580 .llseek
= seq_lseek
,
4581 .release
= seq_release_private
,
4585 static int __init
slab_proc_init(void)
4587 proc_create("slabinfo",S_IWUSR
|S_IRUSR
,NULL
,&proc_slabinfo_operations
);
4588 #ifdef CONFIG_DEBUG_SLAB_LEAK
4589 proc_create("slab_allocators", 0, NULL
, &proc_slabstats_operations
);
4593 module_init(slab_proc_init
);
4597 * ksize - get the actual amount of memory allocated for a given object
4598 * @objp: Pointer to the object
4600 * kmalloc may internally round up allocations and return more memory
4601 * than requested. ksize() can be used to determine the actual amount of
4602 * memory allocated. The caller may use this additional memory, even though
4603 * a smaller amount of memory was initially specified with the kmalloc call.
4604 * The caller must guarantee that objp points to a valid object previously
4605 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4606 * must not be freed during the duration of the call.
4608 size_t ksize(const void *objp
)
4611 if (unlikely(objp
== ZERO_SIZE_PTR
))
4614 return obj_size(virt_to_cache(objp
));
4616 EXPORT_SYMBOL(ksize
);