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/kmemtrace.h>
106 #include <linux/rcupdate.h>
107 #include <linux/string.h>
108 #include <linux/uaccess.h>
109 #include <linux/nodemask.h>
110 #include <linux/kmemleak.h>
111 #include <linux/mempolicy.h>
112 #include <linux/mutex.h>
113 #include <linux/fault-inject.h>
114 #include <linux/rtmutex.h>
115 #include <linux/reciprocal_div.h>
116 #include <linux/debugobjects.h>
118 #include <asm/cacheflush.h>
119 #include <asm/tlbflush.h>
120 #include <asm/page.h>
123 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
124 * 0 for faster, smaller code (especially in the critical paths).
126 * STATS - 1 to collect stats for /proc/slabinfo.
127 * 0 for faster, smaller code (especially in the critical paths).
129 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
132 #ifdef CONFIG_DEBUG_SLAB
135 #define FORCED_DEBUG 1
139 #define FORCED_DEBUG 0
142 /* Shouldn't this be in a header file somewhere? */
143 #define BYTES_PER_WORD sizeof(void *)
144 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
146 #ifndef ARCH_KMALLOC_MINALIGN
148 * Enforce a minimum alignment for the kmalloc caches.
149 * Usually, the kmalloc caches are cache_line_size() aligned, except when
150 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
151 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
152 * alignment larger than the alignment of a 64-bit integer.
153 * ARCH_KMALLOC_MINALIGN allows that.
154 * Note that increasing this value may disable some debug features.
156 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
159 #ifndef ARCH_SLAB_MINALIGN
161 * Enforce a minimum alignment for all caches.
162 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
163 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
164 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
165 * some debug features.
167 #define ARCH_SLAB_MINALIGN 0
170 #ifndef ARCH_KMALLOC_FLAGS
171 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
174 /* Legal flag mask for kmem_cache_create(). */
176 # define CREATE_MASK (SLAB_RED_ZONE | \
177 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
180 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
181 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
182 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE)
184 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
186 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
187 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
188 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE)
194 * Bufctl's are used for linking objs within a slab
197 * This implementation relies on "struct page" for locating the cache &
198 * slab an object belongs to.
199 * This allows the bufctl structure to be small (one int), but limits
200 * the number of objects a slab (not a cache) can contain when off-slab
201 * bufctls are used. The limit is the size of the largest general cache
202 * that does not use off-slab slabs.
203 * For 32bit archs with 4 kB pages, is this 56.
204 * This is not serious, as it is only for large objects, when it is unwise
205 * to have too many per slab.
206 * Note: This limit can be raised by introducing a general cache whose size
207 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
210 typedef unsigned int kmem_bufctl_t
;
211 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
212 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
213 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
214 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
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.
224 struct list_head list
;
225 unsigned long colouroff
;
226 void *s_mem
; /* including colour offset */
227 unsigned int inuse
; /* num of objs active in slab */
229 unsigned short nodeid
;
235 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
236 * arrange for kmem_freepages to be called via RCU. This is useful if
237 * we need to approach a kernel structure obliquely, from its address
238 * obtained without the usual locking. We can lock the structure to
239 * stabilize it and check it's still at the given address, only if we
240 * can be sure that the memory has not been meanwhile reused for some
241 * other kind of object (which our subsystem's lock might corrupt).
243 * rcu_read_lock before reading the address, then rcu_read_unlock after
244 * taking the spinlock within the structure expected at that address.
246 * We assume struct slab_rcu can overlay struct slab when destroying.
249 struct rcu_head head
;
250 struct kmem_cache
*cachep
;
258 * - LIFO ordering, to hand out cache-warm objects from _alloc
259 * - reduce the number of linked list operations
260 * - reduce spinlock operations
262 * The limit is stored in the per-cpu structure to reduce the data cache
269 unsigned int batchcount
;
270 unsigned int touched
;
273 * Must have this definition in here for the proper
274 * alignment of array_cache. Also simplifies accessing
280 * bootstrap: The caches do not work without cpuarrays anymore, but the
281 * cpuarrays are allocated from the generic caches...
283 #define BOOT_CPUCACHE_ENTRIES 1
284 struct arraycache_init
{
285 struct array_cache cache
;
286 void *entries
[BOOT_CPUCACHE_ENTRIES
];
290 * The slab lists for all objects.
293 struct list_head slabs_partial
; /* partial list first, better asm code */
294 struct list_head slabs_full
;
295 struct list_head slabs_free
;
296 unsigned long free_objects
;
297 unsigned int free_limit
;
298 unsigned int colour_next
; /* Per-node cache coloring */
299 spinlock_t list_lock
;
300 struct array_cache
*shared
; /* shared per node */
301 struct array_cache
**alien
; /* on other nodes */
302 unsigned long next_reap
; /* updated without locking */
303 int free_touched
; /* updated without locking */
307 * The slab allocator is initialized with interrupts disabled. Therefore, make
308 * sure early boot allocations don't accidentally enable interrupts.
310 static gfp_t slab_gfp_mask __read_mostly
= SLAB_GFP_BOOT_MASK
;
313 * Need this for bootstrapping a per node allocator.
315 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
316 struct kmem_list3 __initdata initkmem_list3
[NUM_INIT_LISTS
];
317 #define CACHE_CACHE 0
318 #define SIZE_AC MAX_NUMNODES
319 #define SIZE_L3 (2 * MAX_NUMNODES)
321 static int drain_freelist(struct kmem_cache
*cache
,
322 struct kmem_list3
*l3
, int tofree
);
323 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
325 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
);
326 static void cache_reap(struct work_struct
*unused
);
329 * This function must be completely optimized away if a constant is passed to
330 * it. Mostly the same as what is in linux/slab.h except it returns an index.
332 static __always_inline
int index_of(const size_t size
)
334 extern void __bad_size(void);
336 if (__builtin_constant_p(size
)) {
344 #include <linux/kmalloc_sizes.h>
352 static int slab_early_init
= 1;
354 #define INDEX_AC index_of(sizeof(struct arraycache_init))
355 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
357 static void kmem_list3_init(struct kmem_list3
*parent
)
359 INIT_LIST_HEAD(&parent
->slabs_full
);
360 INIT_LIST_HEAD(&parent
->slabs_partial
);
361 INIT_LIST_HEAD(&parent
->slabs_free
);
362 parent
->shared
= NULL
;
363 parent
->alien
= NULL
;
364 parent
->colour_next
= 0;
365 spin_lock_init(&parent
->list_lock
);
366 parent
->free_objects
= 0;
367 parent
->free_touched
= 0;
370 #define MAKE_LIST(cachep, listp, slab, nodeid) \
372 INIT_LIST_HEAD(listp); \
373 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
376 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
378 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
379 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
380 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
390 /* 1) per-cpu data, touched during every alloc/free */
391 struct array_cache
*array
[NR_CPUS
];
392 /* 2) Cache tunables. Protected by cache_chain_mutex */
393 unsigned int batchcount
;
397 unsigned int buffer_size
;
398 u32 reciprocal_buffer_size
;
399 /* 3) touched by every alloc & free from the backend */
401 unsigned int flags
; /* constant flags */
402 unsigned int num
; /* # of objs per slab */
404 /* 4) cache_grow/shrink */
405 /* order of pgs per slab (2^n) */
406 unsigned int gfporder
;
408 /* force GFP flags, e.g. GFP_DMA */
411 size_t colour
; /* cache colouring range */
412 unsigned int colour_off
; /* colour offset */
413 struct kmem_cache
*slabp_cache
;
414 unsigned int slab_size
;
415 unsigned int dflags
; /* dynamic flags */
417 /* constructor func */
418 void (*ctor
)(void *obj
);
420 /* 5) cache creation/removal */
422 struct list_head next
;
426 unsigned long num_active
;
427 unsigned long num_allocations
;
428 unsigned long high_mark
;
430 unsigned long reaped
;
431 unsigned long errors
;
432 unsigned long max_freeable
;
433 unsigned long node_allocs
;
434 unsigned long node_frees
;
435 unsigned long node_overflow
;
443 * If debugging is enabled, then the allocator can add additional
444 * fields and/or padding to every object. buffer_size contains the total
445 * object size including these internal fields, the following two
446 * variables contain the offset to the user object and its size.
452 * We put nodelists[] at the end of kmem_cache, because we want to size
453 * this array to nr_node_ids slots instead of MAX_NUMNODES
454 * (see kmem_cache_init())
455 * We still use [MAX_NUMNODES] and not [1] or [0] because cache_cache
456 * is statically defined, so we reserve the max number of nodes.
458 struct kmem_list3
*nodelists
[MAX_NUMNODES
];
460 * Do not add fields after nodelists[]
464 #define CFLGS_OFF_SLAB (0x80000000UL)
465 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
467 #define BATCHREFILL_LIMIT 16
469 * Optimization question: fewer reaps means less probability for unnessary
470 * cpucache drain/refill cycles.
472 * OTOH the cpuarrays can contain lots of objects,
473 * which could lock up otherwise freeable slabs.
475 #define REAPTIMEOUT_CPUC (2*HZ)
476 #define REAPTIMEOUT_LIST3 (4*HZ)
479 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
480 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
481 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
482 #define STATS_INC_GROWN(x) ((x)->grown++)
483 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
484 #define STATS_SET_HIGH(x) \
486 if ((x)->num_active > (x)->high_mark) \
487 (x)->high_mark = (x)->num_active; \
489 #define STATS_INC_ERR(x) ((x)->errors++)
490 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
491 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
492 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
493 #define STATS_SET_FREEABLE(x, i) \
495 if ((x)->max_freeable < i) \
496 (x)->max_freeable = i; \
498 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
499 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
500 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
501 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
503 #define STATS_INC_ACTIVE(x) do { } while (0)
504 #define STATS_DEC_ACTIVE(x) do { } while (0)
505 #define STATS_INC_ALLOCED(x) do { } while (0)
506 #define STATS_INC_GROWN(x) do { } while (0)
507 #define STATS_ADD_REAPED(x,y) do { } while (0)
508 #define STATS_SET_HIGH(x) do { } while (0)
509 #define STATS_INC_ERR(x) do { } while (0)
510 #define STATS_INC_NODEALLOCS(x) do { } while (0)
511 #define STATS_INC_NODEFREES(x) do { } while (0)
512 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
513 #define STATS_SET_FREEABLE(x, i) do { } while (0)
514 #define STATS_INC_ALLOCHIT(x) do { } while (0)
515 #define STATS_INC_ALLOCMISS(x) do { } while (0)
516 #define STATS_INC_FREEHIT(x) do { } while (0)
517 #define STATS_INC_FREEMISS(x) do { } while (0)
523 * memory layout of objects:
525 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
526 * the end of an object is aligned with the end of the real
527 * allocation. Catches writes behind the end of the allocation.
528 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
530 * cachep->obj_offset: The real object.
531 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
532 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
533 * [BYTES_PER_WORD long]
535 static int obj_offset(struct kmem_cache
*cachep
)
537 return cachep
->obj_offset
;
540 static int obj_size(struct kmem_cache
*cachep
)
542 return cachep
->obj_size
;
545 static unsigned long long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
547 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
548 return (unsigned long long*) (objp
+ obj_offset(cachep
) -
549 sizeof(unsigned long long));
552 static unsigned long long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
554 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
555 if (cachep
->flags
& SLAB_STORE_USER
)
556 return (unsigned long long *)(objp
+ cachep
->buffer_size
-
557 sizeof(unsigned long long) -
559 return (unsigned long long *) (objp
+ cachep
->buffer_size
-
560 sizeof(unsigned long long));
563 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
565 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
566 return (void **)(objp
+ cachep
->buffer_size
- BYTES_PER_WORD
);
571 #define obj_offset(x) 0
572 #define obj_size(cachep) (cachep->buffer_size)
573 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
574 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
575 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
579 #ifdef CONFIG_KMEMTRACE
580 size_t slab_buffer_size(struct kmem_cache
*cachep
)
582 return cachep
->buffer_size
;
584 EXPORT_SYMBOL(slab_buffer_size
);
588 * Do not go above this order unless 0 objects fit into the slab.
590 #define BREAK_GFP_ORDER_HI 1
591 #define BREAK_GFP_ORDER_LO 0
592 static int slab_break_gfp_order
= BREAK_GFP_ORDER_LO
;
595 * Functions for storing/retrieving the cachep and or slab from the page
596 * allocator. These are used to find the slab an obj belongs to. With kfree(),
597 * these are used to find the cache which an obj belongs to.
599 static inline void page_set_cache(struct page
*page
, struct kmem_cache
*cache
)
601 page
->lru
.next
= (struct list_head
*)cache
;
604 static inline struct kmem_cache
*page_get_cache(struct page
*page
)
606 page
= compound_head(page
);
607 BUG_ON(!PageSlab(page
));
608 return (struct kmem_cache
*)page
->lru
.next
;
611 static inline void page_set_slab(struct page
*page
, struct slab
*slab
)
613 page
->lru
.prev
= (struct list_head
*)slab
;
616 static inline struct slab
*page_get_slab(struct page
*page
)
618 BUG_ON(!PageSlab(page
));
619 return (struct slab
*)page
->lru
.prev
;
622 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
624 struct page
*page
= virt_to_head_page(obj
);
625 return page_get_cache(page
);
628 static inline struct slab
*virt_to_slab(const void *obj
)
630 struct page
*page
= virt_to_head_page(obj
);
631 return page_get_slab(page
);
634 static inline void *index_to_obj(struct kmem_cache
*cache
, struct slab
*slab
,
637 return slab
->s_mem
+ cache
->buffer_size
* idx
;
641 * We want to avoid an expensive divide : (offset / cache->buffer_size)
642 * Using the fact that buffer_size is a constant for a particular cache,
643 * we can replace (offset / cache->buffer_size) by
644 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
646 static inline unsigned int obj_to_index(const struct kmem_cache
*cache
,
647 const struct slab
*slab
, void *obj
)
649 u32 offset
= (obj
- slab
->s_mem
);
650 return reciprocal_divide(offset
, cache
->reciprocal_buffer_size
);
654 * These are the default caches for kmalloc. Custom caches can have other sizes.
656 struct cache_sizes malloc_sizes
[] = {
657 #define CACHE(x) { .cs_size = (x) },
658 #include <linux/kmalloc_sizes.h>
662 EXPORT_SYMBOL(malloc_sizes
);
664 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
670 static struct cache_names __initdata cache_names
[] = {
671 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
672 #include <linux/kmalloc_sizes.h>
677 static struct arraycache_init initarray_cache __initdata
=
678 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
679 static struct arraycache_init initarray_generic
=
680 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
682 /* internal cache of cache description objs */
683 static struct kmem_cache cache_cache
= {
685 .limit
= BOOT_CPUCACHE_ENTRIES
,
687 .buffer_size
= sizeof(struct kmem_cache
),
688 .name
= "kmem_cache",
691 #define BAD_ALIEN_MAGIC 0x01020304ul
693 #ifdef CONFIG_LOCKDEP
696 * Slab sometimes uses the kmalloc slabs to store the slab headers
697 * for other slabs "off slab".
698 * The locking for this is tricky in that it nests within the locks
699 * of all other slabs in a few places; to deal with this special
700 * locking we put on-slab caches into a separate lock-class.
702 * We set lock class for alien array caches which are up during init.
703 * The lock annotation will be lost if all cpus of a node goes down and
704 * then comes back up during hotplug
706 static struct lock_class_key on_slab_l3_key
;
707 static struct lock_class_key on_slab_alc_key
;
709 static inline void init_lock_keys(void)
713 struct cache_sizes
*s
= malloc_sizes
;
715 while (s
->cs_size
!= ULONG_MAX
) {
717 struct array_cache
**alc
;
719 struct kmem_list3
*l3
= s
->cs_cachep
->nodelists
[q
];
720 if (!l3
|| OFF_SLAB(s
->cs_cachep
))
722 lockdep_set_class(&l3
->list_lock
, &on_slab_l3_key
);
725 * FIXME: This check for BAD_ALIEN_MAGIC
726 * should go away when common slab code is taught to
727 * work even without alien caches.
728 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
729 * for alloc_alien_cache,
731 if (!alc
|| (unsigned long)alc
== BAD_ALIEN_MAGIC
)
735 lockdep_set_class(&alc
[r
]->lock
,
743 static inline void init_lock_keys(void)
749 * Guard access to the cache-chain.
751 static DEFINE_MUTEX(cache_chain_mutex
);
752 static struct list_head cache_chain
;
755 * chicken and egg problem: delay the per-cpu array allocation
756 * until the general caches are up.
767 * used by boot code to determine if it can use slab based allocator
769 int slab_is_available(void)
771 return g_cpucache_up
>= EARLY
;
774 static DEFINE_PER_CPU(struct delayed_work
, reap_work
);
776 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
778 return cachep
->array
[smp_processor_id()];
781 static inline struct kmem_cache
*__find_general_cachep(size_t size
,
784 struct cache_sizes
*csizep
= malloc_sizes
;
787 /* This happens if someone tries to call
788 * kmem_cache_create(), or __kmalloc(), before
789 * the generic caches are initialized.
791 BUG_ON(malloc_sizes
[INDEX_AC
].cs_cachep
== NULL
);
794 return ZERO_SIZE_PTR
;
796 while (size
> csizep
->cs_size
)
800 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
801 * has cs_{dma,}cachep==NULL. Thus no special case
802 * for large kmalloc calls required.
804 #ifdef CONFIG_ZONE_DMA
805 if (unlikely(gfpflags
& GFP_DMA
))
806 return csizep
->cs_dmacachep
;
808 return csizep
->cs_cachep
;
811 static struct kmem_cache
*kmem_find_general_cachep(size_t size
, gfp_t gfpflags
)
813 return __find_general_cachep(size
, gfpflags
);
816 static size_t slab_mgmt_size(size_t nr_objs
, size_t align
)
818 return ALIGN(sizeof(struct slab
)+nr_objs
*sizeof(kmem_bufctl_t
), align
);
822 * Calculate the number of objects and left-over bytes for a given buffer size.
824 static void cache_estimate(unsigned long gfporder
, size_t buffer_size
,
825 size_t align
, int flags
, size_t *left_over
,
830 size_t slab_size
= PAGE_SIZE
<< gfporder
;
833 * The slab management structure can be either off the slab or
834 * on it. For the latter case, the memory allocated for a
838 * - One kmem_bufctl_t for each object
839 * - Padding to respect alignment of @align
840 * - @buffer_size bytes for each object
842 * If the slab management structure is off the slab, then the
843 * alignment will already be calculated into the size. Because
844 * the slabs are all pages aligned, the objects will be at the
845 * correct alignment when allocated.
847 if (flags
& CFLGS_OFF_SLAB
) {
849 nr_objs
= slab_size
/ buffer_size
;
851 if (nr_objs
> SLAB_LIMIT
)
852 nr_objs
= SLAB_LIMIT
;
855 * Ignore padding for the initial guess. The padding
856 * is at most @align-1 bytes, and @buffer_size is at
857 * least @align. In the worst case, this result will
858 * be one greater than the number of objects that fit
859 * into the memory allocation when taking the padding
862 nr_objs
= (slab_size
- sizeof(struct slab
)) /
863 (buffer_size
+ sizeof(kmem_bufctl_t
));
866 * This calculated number will be either the right
867 * amount, or one greater than what we want.
869 if (slab_mgmt_size(nr_objs
, align
) + nr_objs
*buffer_size
873 if (nr_objs
> SLAB_LIMIT
)
874 nr_objs
= SLAB_LIMIT
;
876 mgmt_size
= slab_mgmt_size(nr_objs
, align
);
879 *left_over
= slab_size
- nr_objs
*buffer_size
- mgmt_size
;
882 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
884 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
887 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
888 function
, cachep
->name
, msg
);
893 * By default on NUMA we use alien caches to stage the freeing of
894 * objects allocated from other nodes. This causes massive memory
895 * inefficiencies when using fake NUMA setup to split memory into a
896 * large number of small nodes, so it can be disabled on the command
900 static int use_alien_caches __read_mostly
= 1;
901 static int __init
noaliencache_setup(char *s
)
903 use_alien_caches
= 0;
906 __setup("noaliencache", noaliencache_setup
);
910 * Special reaping functions for NUMA systems called from cache_reap().
911 * These take care of doing round robin flushing of alien caches (containing
912 * objects freed on different nodes from which they were allocated) and the
913 * flushing of remote pcps by calling drain_node_pages.
915 static DEFINE_PER_CPU(unsigned long, reap_node
);
917 static void init_reap_node(int cpu
)
921 node
= next_node(cpu_to_node(cpu
), node_online_map
);
922 if (node
== MAX_NUMNODES
)
923 node
= first_node(node_online_map
);
925 per_cpu(reap_node
, cpu
) = node
;
928 static void next_reap_node(void)
930 int node
= __get_cpu_var(reap_node
);
932 node
= next_node(node
, node_online_map
);
933 if (unlikely(node
>= MAX_NUMNODES
))
934 node
= first_node(node_online_map
);
935 __get_cpu_var(reap_node
) = node
;
939 #define init_reap_node(cpu) do { } while (0)
940 #define next_reap_node(void) do { } while (0)
944 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
945 * via the workqueue/eventd.
946 * Add the CPU number into the expiration time to minimize the possibility of
947 * the CPUs getting into lockstep and contending for the global cache chain
950 static void __cpuinit
start_cpu_timer(int cpu
)
952 struct delayed_work
*reap_work
= &per_cpu(reap_work
, cpu
);
955 * When this gets called from do_initcalls via cpucache_init(),
956 * init_workqueues() has already run, so keventd will be setup
959 if (keventd_up() && reap_work
->work
.func
== NULL
) {
961 INIT_DELAYED_WORK(reap_work
, cache_reap
);
962 schedule_delayed_work_on(cpu
, reap_work
,
963 __round_jiffies_relative(HZ
, cpu
));
967 static struct array_cache
*alloc_arraycache(int node
, int entries
,
968 int batchcount
, gfp_t gfp
)
970 int memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
971 struct array_cache
*nc
= NULL
;
973 nc
= kmalloc_node(memsize
, gfp
, node
);
975 * The array_cache structures contain pointers to free object.
976 * However, when such objects are allocated or transfered to another
977 * cache the pointers are not cleared and they could be counted as
978 * valid references during a kmemleak scan. Therefore, kmemleak must
979 * not scan such objects.
981 kmemleak_no_scan(nc
);
985 nc
->batchcount
= batchcount
;
987 spin_lock_init(&nc
->lock
);
993 * Transfer objects in one arraycache to another.
994 * Locking must be handled by the caller.
996 * Return the number of entries transferred.
998 static int transfer_objects(struct array_cache
*to
,
999 struct array_cache
*from
, unsigned int max
)
1001 /* Figure out how many entries to transfer */
1002 int nr
= min(min(from
->avail
, max
), to
->limit
- to
->avail
);
1007 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
1008 sizeof(void *) *nr
);
1018 #define drain_alien_cache(cachep, alien) do { } while (0)
1019 #define reap_alien(cachep, l3) do { } while (0)
1021 static inline struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
1023 return (struct array_cache
**)BAD_ALIEN_MAGIC
;
1026 static inline void free_alien_cache(struct array_cache
**ac_ptr
)
1030 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1035 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
1041 static inline void *____cache_alloc_node(struct kmem_cache
*cachep
,
1042 gfp_t flags
, int nodeid
)
1047 #else /* CONFIG_NUMA */
1049 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
1050 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
1052 static struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
1054 struct array_cache
**ac_ptr
;
1055 int memsize
= sizeof(void *) * nr_node_ids
;
1060 ac_ptr
= kmalloc_node(memsize
, gfp
, node
);
1063 if (i
== node
|| !node_online(i
)) {
1067 ac_ptr
[i
] = alloc_arraycache(node
, limit
, 0xbaadf00d, gfp
);
1069 for (i
--; i
>= 0; i
--)
1079 static void free_alien_cache(struct array_cache
**ac_ptr
)
1090 static void __drain_alien_cache(struct kmem_cache
*cachep
,
1091 struct array_cache
*ac
, int node
)
1093 struct kmem_list3
*rl3
= cachep
->nodelists
[node
];
1096 spin_lock(&rl3
->list_lock
);
1098 * Stuff objects into the remote nodes shared array first.
1099 * That way we could avoid the overhead of putting the objects
1100 * into the free lists and getting them back later.
1103 transfer_objects(rl3
->shared
, ac
, ac
->limit
);
1105 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
1107 spin_unlock(&rl3
->list_lock
);
1112 * Called from cache_reap() to regularly drain alien caches round robin.
1114 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_list3
*l3
)
1116 int node
= __get_cpu_var(reap_node
);
1119 struct array_cache
*ac
= l3
->alien
[node
];
1121 if (ac
&& ac
->avail
&& spin_trylock_irq(&ac
->lock
)) {
1122 __drain_alien_cache(cachep
, ac
, node
);
1123 spin_unlock_irq(&ac
->lock
);
1128 static void drain_alien_cache(struct kmem_cache
*cachep
,
1129 struct array_cache
**alien
)
1132 struct array_cache
*ac
;
1133 unsigned long flags
;
1135 for_each_online_node(i
) {
1138 spin_lock_irqsave(&ac
->lock
, flags
);
1139 __drain_alien_cache(cachep
, ac
, i
);
1140 spin_unlock_irqrestore(&ac
->lock
, flags
);
1145 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1147 struct slab
*slabp
= virt_to_slab(objp
);
1148 int nodeid
= slabp
->nodeid
;
1149 struct kmem_list3
*l3
;
1150 struct array_cache
*alien
= NULL
;
1153 node
= numa_node_id();
1156 * Make sure we are not freeing a object from another node to the array
1157 * cache on this cpu.
1159 if (likely(slabp
->nodeid
== node
))
1162 l3
= cachep
->nodelists
[node
];
1163 STATS_INC_NODEFREES(cachep
);
1164 if (l3
->alien
&& l3
->alien
[nodeid
]) {
1165 alien
= l3
->alien
[nodeid
];
1166 spin_lock(&alien
->lock
);
1167 if (unlikely(alien
->avail
== alien
->limit
)) {
1168 STATS_INC_ACOVERFLOW(cachep
);
1169 __drain_alien_cache(cachep
, alien
, nodeid
);
1171 alien
->entry
[alien
->avail
++] = objp
;
1172 spin_unlock(&alien
->lock
);
1174 spin_lock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1175 free_block(cachep
, &objp
, 1, nodeid
);
1176 spin_unlock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1182 static void __cpuinit
cpuup_canceled(long cpu
)
1184 struct kmem_cache
*cachep
;
1185 struct kmem_list3
*l3
= NULL
;
1186 int node
= cpu_to_node(cpu
);
1187 const struct cpumask
*mask
= cpumask_of_node(node
);
1189 list_for_each_entry(cachep
, &cache_chain
, next
) {
1190 struct array_cache
*nc
;
1191 struct array_cache
*shared
;
1192 struct array_cache
**alien
;
1194 /* cpu is dead; no one can alloc from it. */
1195 nc
= cachep
->array
[cpu
];
1196 cachep
->array
[cpu
] = NULL
;
1197 l3
= cachep
->nodelists
[node
];
1200 goto free_array_cache
;
1202 spin_lock_irq(&l3
->list_lock
);
1204 /* Free limit for this kmem_list3 */
1205 l3
->free_limit
-= cachep
->batchcount
;
1207 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
1209 if (!cpus_empty(*mask
)) {
1210 spin_unlock_irq(&l3
->list_lock
);
1211 goto free_array_cache
;
1214 shared
= l3
->shared
;
1216 free_block(cachep
, shared
->entry
,
1217 shared
->avail
, node
);
1224 spin_unlock_irq(&l3
->list_lock
);
1228 drain_alien_cache(cachep
, alien
);
1229 free_alien_cache(alien
);
1235 * In the previous loop, all the objects were freed to
1236 * the respective cache's slabs, now we can go ahead and
1237 * shrink each nodelist to its limit.
1239 list_for_each_entry(cachep
, &cache_chain
, next
) {
1240 l3
= cachep
->nodelists
[node
];
1243 drain_freelist(cachep
, l3
, l3
->free_objects
);
1247 static int __cpuinit
cpuup_prepare(long cpu
)
1249 struct kmem_cache
*cachep
;
1250 struct kmem_list3
*l3
= NULL
;
1251 int node
= cpu_to_node(cpu
);
1252 const int memsize
= sizeof(struct kmem_list3
);
1255 * We need to do this right in the beginning since
1256 * alloc_arraycache's are going to use this list.
1257 * kmalloc_node allows us to add the slab to the right
1258 * kmem_list3 and not this cpu's kmem_list3
1261 list_for_each_entry(cachep
, &cache_chain
, next
) {
1263 * Set up the size64 kmemlist for cpu before we can
1264 * begin anything. Make sure some other cpu on this
1265 * node has not already allocated this
1267 if (!cachep
->nodelists
[node
]) {
1268 l3
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1271 kmem_list3_init(l3
);
1272 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
1273 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1276 * The l3s don't come and go as CPUs come and
1277 * go. cache_chain_mutex is sufficient
1280 cachep
->nodelists
[node
] = l3
;
1283 spin_lock_irq(&cachep
->nodelists
[node
]->list_lock
);
1284 cachep
->nodelists
[node
]->free_limit
=
1285 (1 + nr_cpus_node(node
)) *
1286 cachep
->batchcount
+ cachep
->num
;
1287 spin_unlock_irq(&cachep
->nodelists
[node
]->list_lock
);
1291 * Now we can go ahead with allocating the shared arrays and
1294 list_for_each_entry(cachep
, &cache_chain
, next
) {
1295 struct array_cache
*nc
;
1296 struct array_cache
*shared
= NULL
;
1297 struct array_cache
**alien
= NULL
;
1299 nc
= alloc_arraycache(node
, cachep
->limit
,
1300 cachep
->batchcount
, GFP_KERNEL
);
1303 if (cachep
->shared
) {
1304 shared
= alloc_arraycache(node
,
1305 cachep
->shared
* cachep
->batchcount
,
1306 0xbaadf00d, GFP_KERNEL
);
1312 if (use_alien_caches
) {
1313 alien
= alloc_alien_cache(node
, cachep
->limit
, GFP_KERNEL
);
1320 cachep
->array
[cpu
] = nc
;
1321 l3
= cachep
->nodelists
[node
];
1324 spin_lock_irq(&l3
->list_lock
);
1327 * We are serialised from CPU_DEAD or
1328 * CPU_UP_CANCELLED by the cpucontrol lock
1330 l3
->shared
= shared
;
1339 spin_unlock_irq(&l3
->list_lock
);
1341 free_alien_cache(alien
);
1345 cpuup_canceled(cpu
);
1349 static int __cpuinit
cpuup_callback(struct notifier_block
*nfb
,
1350 unsigned long action
, void *hcpu
)
1352 long cpu
= (long)hcpu
;
1356 case CPU_UP_PREPARE
:
1357 case CPU_UP_PREPARE_FROZEN
:
1358 mutex_lock(&cache_chain_mutex
);
1359 err
= cpuup_prepare(cpu
);
1360 mutex_unlock(&cache_chain_mutex
);
1363 case CPU_ONLINE_FROZEN
:
1364 start_cpu_timer(cpu
);
1366 #ifdef CONFIG_HOTPLUG_CPU
1367 case CPU_DOWN_PREPARE
:
1368 case CPU_DOWN_PREPARE_FROZEN
:
1370 * Shutdown cache reaper. Note that the cache_chain_mutex is
1371 * held so that if cache_reap() is invoked it cannot do
1372 * anything expensive but will only modify reap_work
1373 * and reschedule the timer.
1375 cancel_rearming_delayed_work(&per_cpu(reap_work
, cpu
));
1376 /* Now the cache_reaper is guaranteed to be not running. */
1377 per_cpu(reap_work
, cpu
).work
.func
= NULL
;
1379 case CPU_DOWN_FAILED
:
1380 case CPU_DOWN_FAILED_FROZEN
:
1381 start_cpu_timer(cpu
);
1384 case CPU_DEAD_FROZEN
:
1386 * Even if all the cpus of a node are down, we don't free the
1387 * kmem_list3 of any cache. This to avoid a race between
1388 * cpu_down, and a kmalloc allocation from another cpu for
1389 * memory from the node of the cpu going down. The list3
1390 * structure is usually allocated from kmem_cache_create() and
1391 * gets destroyed at kmem_cache_destroy().
1395 case CPU_UP_CANCELED
:
1396 case CPU_UP_CANCELED_FROZEN
:
1397 mutex_lock(&cache_chain_mutex
);
1398 cpuup_canceled(cpu
);
1399 mutex_unlock(&cache_chain_mutex
);
1402 return err
? NOTIFY_BAD
: NOTIFY_OK
;
1405 static struct notifier_block __cpuinitdata cpucache_notifier
= {
1406 &cpuup_callback
, NULL
, 0
1410 * swap the static kmem_list3 with kmalloced memory
1412 static void init_list(struct kmem_cache
*cachep
, struct kmem_list3
*list
,
1415 struct kmem_list3
*ptr
;
1417 ptr
= kmalloc_node(sizeof(struct kmem_list3
), GFP_NOWAIT
, nodeid
);
1420 memcpy(ptr
, list
, sizeof(struct kmem_list3
));
1422 * Do not assume that spinlocks can be initialized via memcpy:
1424 spin_lock_init(&ptr
->list_lock
);
1426 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1427 cachep
->nodelists
[nodeid
] = ptr
;
1431 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1432 * size of kmem_list3.
1434 static void __init
set_up_list3s(struct kmem_cache
*cachep
, int index
)
1438 for_each_online_node(node
) {
1439 cachep
->nodelists
[node
] = &initkmem_list3
[index
+ node
];
1440 cachep
->nodelists
[node
]->next_reap
= jiffies
+
1442 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1447 * Initialisation. Called after the page allocator have been initialised and
1448 * before smp_init().
1450 void __init
kmem_cache_init(void)
1453 struct cache_sizes
*sizes
;
1454 struct cache_names
*names
;
1459 if (num_possible_nodes() == 1)
1460 use_alien_caches
= 0;
1462 for (i
= 0; i
< NUM_INIT_LISTS
; i
++) {
1463 kmem_list3_init(&initkmem_list3
[i
]);
1464 if (i
< MAX_NUMNODES
)
1465 cache_cache
.nodelists
[i
] = NULL
;
1467 set_up_list3s(&cache_cache
, CACHE_CACHE
);
1470 * Fragmentation resistance on low memory - only use bigger
1471 * page orders on machines with more than 32MB of memory.
1473 if (num_physpages
> (32 << 20) >> PAGE_SHIFT
)
1474 slab_break_gfp_order
= BREAK_GFP_ORDER_HI
;
1476 /* Bootstrap is tricky, because several objects are allocated
1477 * from caches that do not exist yet:
1478 * 1) initialize the cache_cache cache: it contains the struct
1479 * kmem_cache structures of all caches, except cache_cache itself:
1480 * cache_cache is statically allocated.
1481 * Initially an __init data area is used for the head array and the
1482 * kmem_list3 structures, it's replaced with a kmalloc allocated
1483 * array at the end of the bootstrap.
1484 * 2) Create the first kmalloc cache.
1485 * The struct kmem_cache for the new cache is allocated normally.
1486 * An __init data area is used for the head array.
1487 * 3) Create the remaining kmalloc caches, with minimally sized
1489 * 4) Replace the __init data head arrays for cache_cache and the first
1490 * kmalloc cache with kmalloc allocated arrays.
1491 * 5) Replace the __init data for kmem_list3 for cache_cache and
1492 * the other cache's with kmalloc allocated memory.
1493 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1496 node
= numa_node_id();
1498 /* 1) create the cache_cache */
1499 INIT_LIST_HEAD(&cache_chain
);
1500 list_add(&cache_cache
.next
, &cache_chain
);
1501 cache_cache
.colour_off
= cache_line_size();
1502 cache_cache
.array
[smp_processor_id()] = &initarray_cache
.cache
;
1503 cache_cache
.nodelists
[node
] = &initkmem_list3
[CACHE_CACHE
+ node
];
1506 * struct kmem_cache size depends on nr_node_ids, which
1507 * can be less than MAX_NUMNODES.
1509 cache_cache
.buffer_size
= offsetof(struct kmem_cache
, nodelists
) +
1510 nr_node_ids
* sizeof(struct kmem_list3
*);
1512 cache_cache
.obj_size
= cache_cache
.buffer_size
;
1514 cache_cache
.buffer_size
= ALIGN(cache_cache
.buffer_size
,
1516 cache_cache
.reciprocal_buffer_size
=
1517 reciprocal_value(cache_cache
.buffer_size
);
1519 for (order
= 0; order
< MAX_ORDER
; order
++) {
1520 cache_estimate(order
, cache_cache
.buffer_size
,
1521 cache_line_size(), 0, &left_over
, &cache_cache
.num
);
1522 if (cache_cache
.num
)
1525 BUG_ON(!cache_cache
.num
);
1526 cache_cache
.gfporder
= order
;
1527 cache_cache
.colour
= left_over
/ cache_cache
.colour_off
;
1528 cache_cache
.slab_size
= ALIGN(cache_cache
.num
* sizeof(kmem_bufctl_t
) +
1529 sizeof(struct slab
), cache_line_size());
1531 /* 2+3) create the kmalloc caches */
1532 sizes
= malloc_sizes
;
1533 names
= cache_names
;
1536 * Initialize the caches that provide memory for the array cache and the
1537 * kmem_list3 structures first. Without this, further allocations will
1541 sizes
[INDEX_AC
].cs_cachep
= kmem_cache_create(names
[INDEX_AC
].name
,
1542 sizes
[INDEX_AC
].cs_size
,
1543 ARCH_KMALLOC_MINALIGN
,
1544 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1547 if (INDEX_AC
!= INDEX_L3
) {
1548 sizes
[INDEX_L3
].cs_cachep
=
1549 kmem_cache_create(names
[INDEX_L3
].name
,
1550 sizes
[INDEX_L3
].cs_size
,
1551 ARCH_KMALLOC_MINALIGN
,
1552 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1556 slab_early_init
= 0;
1558 while (sizes
->cs_size
!= ULONG_MAX
) {
1560 * For performance, all the general caches are L1 aligned.
1561 * This should be particularly beneficial on SMP boxes, as it
1562 * eliminates "false sharing".
1563 * Note for systems short on memory removing the alignment will
1564 * allow tighter packing of the smaller caches.
1566 if (!sizes
->cs_cachep
) {
1567 sizes
->cs_cachep
= kmem_cache_create(names
->name
,
1569 ARCH_KMALLOC_MINALIGN
,
1570 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1573 #ifdef CONFIG_ZONE_DMA
1574 sizes
->cs_dmacachep
= kmem_cache_create(
1577 ARCH_KMALLOC_MINALIGN
,
1578 ARCH_KMALLOC_FLAGS
|SLAB_CACHE_DMA
|
1585 /* 4) Replace the bootstrap head arrays */
1587 struct array_cache
*ptr
;
1589 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1591 BUG_ON(cpu_cache_get(&cache_cache
) != &initarray_cache
.cache
);
1592 memcpy(ptr
, cpu_cache_get(&cache_cache
),
1593 sizeof(struct arraycache_init
));
1595 * Do not assume that spinlocks can be initialized via memcpy:
1597 spin_lock_init(&ptr
->lock
);
1599 cache_cache
.array
[smp_processor_id()] = ptr
;
1601 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1603 BUG_ON(cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
)
1604 != &initarray_generic
.cache
);
1605 memcpy(ptr
, cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
),
1606 sizeof(struct arraycache_init
));
1608 * Do not assume that spinlocks can be initialized via memcpy:
1610 spin_lock_init(&ptr
->lock
);
1612 malloc_sizes
[INDEX_AC
].cs_cachep
->array
[smp_processor_id()] =
1615 /* 5) Replace the bootstrap kmem_list3's */
1619 for_each_online_node(nid
) {
1620 init_list(&cache_cache
, &initkmem_list3
[CACHE_CACHE
+ nid
], nid
);
1622 init_list(malloc_sizes
[INDEX_AC
].cs_cachep
,
1623 &initkmem_list3
[SIZE_AC
+ nid
], nid
);
1625 if (INDEX_AC
!= INDEX_L3
) {
1626 init_list(malloc_sizes
[INDEX_L3
].cs_cachep
,
1627 &initkmem_list3
[SIZE_L3
+ nid
], nid
);
1632 g_cpucache_up
= EARLY
;
1634 /* Annotate slab for lockdep -- annotate the malloc caches */
1638 void __init
kmem_cache_init_late(void)
1640 struct kmem_cache
*cachep
;
1643 * Interrupts are enabled now so all GFP allocations are safe.
1645 slab_gfp_mask
= __GFP_BITS_MASK
;
1647 /* 6) resize the head arrays to their final sizes */
1648 mutex_lock(&cache_chain_mutex
);
1649 list_for_each_entry(cachep
, &cache_chain
, next
)
1650 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1652 mutex_unlock(&cache_chain_mutex
);
1655 g_cpucache_up
= FULL
;
1658 * Register a cpu startup notifier callback that initializes
1659 * cpu_cache_get for all new cpus
1661 register_cpu_notifier(&cpucache_notifier
);
1664 * The reap timers are started later, with a module init call: That part
1665 * of the kernel is not yet operational.
1669 static int __init
cpucache_init(void)
1674 * Register the timers that return unneeded pages to the page allocator
1676 for_each_online_cpu(cpu
)
1677 start_cpu_timer(cpu
);
1680 __initcall(cpucache_init
);
1683 * Interface to system's page allocator. No need to hold the cache-lock.
1685 * If we requested dmaable memory, we will get it. Even if we
1686 * did not request dmaable memory, we might get it, but that
1687 * would be relatively rare and ignorable.
1689 static void *kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
1697 * Nommu uses slab's for process anonymous memory allocations, and thus
1698 * requires __GFP_COMP to properly refcount higher order allocations
1700 flags
|= __GFP_COMP
;
1703 flags
|= cachep
->gfpflags
;
1704 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1705 flags
|= __GFP_RECLAIMABLE
;
1707 page
= alloc_pages_exact_node(nodeid
, flags
, cachep
->gfporder
);
1711 nr_pages
= (1 << cachep
->gfporder
);
1712 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1713 add_zone_page_state(page_zone(page
),
1714 NR_SLAB_RECLAIMABLE
, nr_pages
);
1716 add_zone_page_state(page_zone(page
),
1717 NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1718 for (i
= 0; i
< nr_pages
; i
++)
1719 __SetPageSlab(page
+ i
);
1720 return page_address(page
);
1724 * Interface to system's page release.
1726 static void kmem_freepages(struct kmem_cache
*cachep
, void *addr
)
1728 unsigned long i
= (1 << cachep
->gfporder
);
1729 struct page
*page
= virt_to_page(addr
);
1730 const unsigned long nr_freed
= i
;
1732 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1733 sub_zone_page_state(page_zone(page
),
1734 NR_SLAB_RECLAIMABLE
, nr_freed
);
1736 sub_zone_page_state(page_zone(page
),
1737 NR_SLAB_UNRECLAIMABLE
, nr_freed
);
1739 BUG_ON(!PageSlab(page
));
1740 __ClearPageSlab(page
);
1743 if (current
->reclaim_state
)
1744 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1745 free_pages((unsigned long)addr
, cachep
->gfporder
);
1748 static void kmem_rcu_free(struct rcu_head
*head
)
1750 struct slab_rcu
*slab_rcu
= (struct slab_rcu
*)head
;
1751 struct kmem_cache
*cachep
= slab_rcu
->cachep
;
1753 kmem_freepages(cachep
, slab_rcu
->addr
);
1754 if (OFF_SLAB(cachep
))
1755 kmem_cache_free(cachep
->slabp_cache
, slab_rcu
);
1760 #ifdef CONFIG_DEBUG_PAGEALLOC
1761 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1762 unsigned long caller
)
1764 int size
= obj_size(cachep
);
1766 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1768 if (size
< 5 * sizeof(unsigned long))
1771 *addr
++ = 0x12345678;
1773 *addr
++ = smp_processor_id();
1774 size
-= 3 * sizeof(unsigned long);
1776 unsigned long *sptr
= &caller
;
1777 unsigned long svalue
;
1779 while (!kstack_end(sptr
)) {
1781 if (kernel_text_address(svalue
)) {
1783 size
-= sizeof(unsigned long);
1784 if (size
<= sizeof(unsigned long))
1790 *addr
++ = 0x87654321;
1794 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1796 int size
= obj_size(cachep
);
1797 addr
= &((char *)addr
)[obj_offset(cachep
)];
1799 memset(addr
, val
, size
);
1800 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1803 static void dump_line(char *data
, int offset
, int limit
)
1806 unsigned char error
= 0;
1809 printk(KERN_ERR
"%03x:", offset
);
1810 for (i
= 0; i
< limit
; i
++) {
1811 if (data
[offset
+ i
] != POISON_FREE
) {
1812 error
= data
[offset
+ i
];
1815 printk(" %02x", (unsigned char)data
[offset
+ i
]);
1819 if (bad_count
== 1) {
1820 error
^= POISON_FREE
;
1821 if (!(error
& (error
- 1))) {
1822 printk(KERN_ERR
"Single bit error detected. Probably "
1825 printk(KERN_ERR
"Run memtest86+ or a similar memory "
1828 printk(KERN_ERR
"Run a memory test tool.\n");
1837 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1842 if (cachep
->flags
& SLAB_RED_ZONE
) {
1843 printk(KERN_ERR
"Redzone: 0x%llx/0x%llx.\n",
1844 *dbg_redzone1(cachep
, objp
),
1845 *dbg_redzone2(cachep
, objp
));
1848 if (cachep
->flags
& SLAB_STORE_USER
) {
1849 printk(KERN_ERR
"Last user: [<%p>]",
1850 *dbg_userword(cachep
, objp
));
1851 print_symbol("(%s)",
1852 (unsigned long)*dbg_userword(cachep
, objp
));
1855 realobj
= (char *)objp
+ obj_offset(cachep
);
1856 size
= obj_size(cachep
);
1857 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1860 if (i
+ limit
> size
)
1862 dump_line(realobj
, i
, limit
);
1866 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1872 realobj
= (char *)objp
+ obj_offset(cachep
);
1873 size
= obj_size(cachep
);
1875 for (i
= 0; i
< size
; i
++) {
1876 char exp
= POISON_FREE
;
1879 if (realobj
[i
] != exp
) {
1885 "Slab corruption: %s start=%p, len=%d\n",
1886 cachep
->name
, realobj
, size
);
1887 print_objinfo(cachep
, objp
, 0);
1889 /* Hexdump the affected line */
1892 if (i
+ limit
> size
)
1894 dump_line(realobj
, i
, limit
);
1897 /* Limit to 5 lines */
1903 /* Print some data about the neighboring objects, if they
1906 struct slab
*slabp
= virt_to_slab(objp
);
1909 objnr
= obj_to_index(cachep
, slabp
, objp
);
1911 objp
= index_to_obj(cachep
, slabp
, objnr
- 1);
1912 realobj
= (char *)objp
+ obj_offset(cachep
);
1913 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
1915 print_objinfo(cachep
, objp
, 2);
1917 if (objnr
+ 1 < cachep
->num
) {
1918 objp
= index_to_obj(cachep
, slabp
, objnr
+ 1);
1919 realobj
= (char *)objp
+ obj_offset(cachep
);
1920 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
1922 print_objinfo(cachep
, objp
, 2);
1929 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
, struct slab
*slabp
)
1932 for (i
= 0; i
< cachep
->num
; i
++) {
1933 void *objp
= index_to_obj(cachep
, slabp
, i
);
1935 if (cachep
->flags
& SLAB_POISON
) {
1936 #ifdef CONFIG_DEBUG_PAGEALLOC
1937 if (cachep
->buffer_size
% PAGE_SIZE
== 0 &&
1939 kernel_map_pages(virt_to_page(objp
),
1940 cachep
->buffer_size
/ PAGE_SIZE
, 1);
1942 check_poison_obj(cachep
, objp
);
1944 check_poison_obj(cachep
, objp
);
1947 if (cachep
->flags
& SLAB_RED_ZONE
) {
1948 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1949 slab_error(cachep
, "start of a freed object "
1951 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1952 slab_error(cachep
, "end of a freed object "
1958 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
, struct slab
*slabp
)
1964 * slab_destroy - destroy and release all objects in a slab
1965 * @cachep: cache pointer being destroyed
1966 * @slabp: slab pointer being destroyed
1968 * Destroy all the objs in a slab, and release the mem back to the system.
1969 * Before calling the slab must have been unlinked from the cache. The
1970 * cache-lock is not held/needed.
1972 static void slab_destroy(struct kmem_cache
*cachep
, struct slab
*slabp
)
1974 void *addr
= slabp
->s_mem
- slabp
->colouroff
;
1976 slab_destroy_debugcheck(cachep
, slabp
);
1977 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
1978 struct slab_rcu
*slab_rcu
;
1980 slab_rcu
= (struct slab_rcu
*)slabp
;
1981 slab_rcu
->cachep
= cachep
;
1982 slab_rcu
->addr
= addr
;
1983 call_rcu(&slab_rcu
->head
, kmem_rcu_free
);
1985 kmem_freepages(cachep
, addr
);
1986 if (OFF_SLAB(cachep
))
1987 kmem_cache_free(cachep
->slabp_cache
, slabp
);
1991 static void __kmem_cache_destroy(struct kmem_cache
*cachep
)
1994 struct kmem_list3
*l3
;
1996 for_each_online_cpu(i
)
1997 kfree(cachep
->array
[i
]);
1999 /* NUMA: free the list3 structures */
2000 for_each_online_node(i
) {
2001 l3
= cachep
->nodelists
[i
];
2004 free_alien_cache(l3
->alien
);
2008 kmem_cache_free(&cache_cache
, cachep
);
2013 * calculate_slab_order - calculate size (page order) of slabs
2014 * @cachep: pointer to the cache that is being created
2015 * @size: size of objects to be created in this cache.
2016 * @align: required alignment for the objects.
2017 * @flags: slab allocation flags
2019 * Also calculates the number of objects per slab.
2021 * This could be made much more intelligent. For now, try to avoid using
2022 * high order pages for slabs. When the gfp() functions are more friendly
2023 * towards high-order requests, this should be changed.
2025 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
2026 size_t size
, size_t align
, unsigned long flags
)
2028 unsigned long offslab_limit
;
2029 size_t left_over
= 0;
2032 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
2036 cache_estimate(gfporder
, size
, align
, flags
, &remainder
, &num
);
2040 if (flags
& CFLGS_OFF_SLAB
) {
2042 * Max number of objs-per-slab for caches which
2043 * use off-slab slabs. Needed to avoid a possible
2044 * looping condition in cache_grow().
2046 offslab_limit
= size
- sizeof(struct slab
);
2047 offslab_limit
/= sizeof(kmem_bufctl_t
);
2049 if (num
> offslab_limit
)
2053 /* Found something acceptable - save it away */
2055 cachep
->gfporder
= gfporder
;
2056 left_over
= remainder
;
2059 * A VFS-reclaimable slab tends to have most allocations
2060 * as GFP_NOFS and we really don't want to have to be allocating
2061 * higher-order pages when we are unable to shrink dcache.
2063 if (flags
& SLAB_RECLAIM_ACCOUNT
)
2067 * Large number of objects is good, but very large slabs are
2068 * currently bad for the gfp()s.
2070 if (gfporder
>= slab_break_gfp_order
)
2074 * Acceptable internal fragmentation?
2076 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
2082 static int __init_refok
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
2084 if (g_cpucache_up
== FULL
)
2085 return enable_cpucache(cachep
, gfp
);
2087 if (g_cpucache_up
== NONE
) {
2089 * Note: the first kmem_cache_create must create the cache
2090 * that's used by kmalloc(24), otherwise the creation of
2091 * further caches will BUG().
2093 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
2096 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2097 * the first cache, then we need to set up all its list3s,
2098 * otherwise the creation of further caches will BUG().
2100 set_up_list3s(cachep
, SIZE_AC
);
2101 if (INDEX_AC
== INDEX_L3
)
2102 g_cpucache_up
= PARTIAL_L3
;
2104 g_cpucache_up
= PARTIAL_AC
;
2106 cachep
->array
[smp_processor_id()] =
2107 kmalloc(sizeof(struct arraycache_init
), gfp
);
2109 if (g_cpucache_up
== PARTIAL_AC
) {
2110 set_up_list3s(cachep
, SIZE_L3
);
2111 g_cpucache_up
= PARTIAL_L3
;
2114 for_each_online_node(node
) {
2115 cachep
->nodelists
[node
] =
2116 kmalloc_node(sizeof(struct kmem_list3
),
2118 BUG_ON(!cachep
->nodelists
[node
]);
2119 kmem_list3_init(cachep
->nodelists
[node
]);
2123 cachep
->nodelists
[numa_node_id()]->next_reap
=
2124 jiffies
+ REAPTIMEOUT_LIST3
+
2125 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
2127 cpu_cache_get(cachep
)->avail
= 0;
2128 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
2129 cpu_cache_get(cachep
)->batchcount
= 1;
2130 cpu_cache_get(cachep
)->touched
= 0;
2131 cachep
->batchcount
= 1;
2132 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
2137 * kmem_cache_create - Create a cache.
2138 * @name: A string which is used in /proc/slabinfo to identify this cache.
2139 * @size: The size of objects to be created in this cache.
2140 * @align: The required alignment for the objects.
2141 * @flags: SLAB flags
2142 * @ctor: A constructor for the objects.
2144 * Returns a ptr to the cache on success, NULL on failure.
2145 * Cannot be called within a int, but can be interrupted.
2146 * The @ctor is run when new pages are allocated by the cache.
2148 * @name must be valid until the cache is destroyed. This implies that
2149 * the module calling this has to destroy the cache before getting unloaded.
2150 * Note that kmem_cache_name() is not guaranteed to return the same pointer,
2151 * therefore applications must manage it themselves.
2155 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2156 * to catch references to uninitialised memory.
2158 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2159 * for buffer overruns.
2161 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2162 * cacheline. This can be beneficial if you're counting cycles as closely
2166 kmem_cache_create (const char *name
, size_t size
, size_t align
,
2167 unsigned long flags
, void (*ctor
)(void *))
2169 size_t left_over
, slab_size
, ralign
;
2170 struct kmem_cache
*cachep
= NULL
, *pc
;
2174 * Sanity checks... these are all serious usage bugs.
2176 if (!name
|| in_interrupt() || (size
< BYTES_PER_WORD
) ||
2177 size
> KMALLOC_MAX_SIZE
) {
2178 printk(KERN_ERR
"%s: Early error in slab %s\n", __func__
,
2184 * We use cache_chain_mutex to ensure a consistent view of
2185 * cpu_online_mask as well. Please see cpuup_callback
2187 if (slab_is_available()) {
2189 mutex_lock(&cache_chain_mutex
);
2192 list_for_each_entry(pc
, &cache_chain
, next
) {
2197 * This happens when the module gets unloaded and doesn't
2198 * destroy its slab cache and no-one else reuses the vmalloc
2199 * area of the module. Print a warning.
2201 res
= probe_kernel_address(pc
->name
, tmp
);
2204 "SLAB: cache with size %d has lost its name\n",
2209 if (!strcmp(pc
->name
, name
)) {
2211 "kmem_cache_create: duplicate cache %s\n", name
);
2218 WARN_ON(strchr(name
, ' ')); /* It confuses parsers */
2221 * Enable redzoning and last user accounting, except for caches with
2222 * large objects, if the increased size would increase the object size
2223 * above the next power of two: caches with object sizes just above a
2224 * power of two have a significant amount of internal fragmentation.
2226 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
2227 2 * sizeof(unsigned long long)))
2228 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2229 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2230 flags
|= SLAB_POISON
;
2232 if (flags
& SLAB_DESTROY_BY_RCU
)
2233 BUG_ON(flags
& SLAB_POISON
);
2236 * Always checks flags, a caller might be expecting debug support which
2239 BUG_ON(flags
& ~CREATE_MASK
);
2242 * Check that size is in terms of words. This is needed to avoid
2243 * unaligned accesses for some archs when redzoning is used, and makes
2244 * sure any on-slab bufctl's are also correctly aligned.
2246 if (size
& (BYTES_PER_WORD
- 1)) {
2247 size
+= (BYTES_PER_WORD
- 1);
2248 size
&= ~(BYTES_PER_WORD
- 1);
2251 /* calculate the final buffer alignment: */
2253 /* 1) arch recommendation: can be overridden for debug */
2254 if (flags
& SLAB_HWCACHE_ALIGN
) {
2256 * Default alignment: as specified by the arch code. Except if
2257 * an object is really small, then squeeze multiple objects into
2260 ralign
= cache_line_size();
2261 while (size
<= ralign
/ 2)
2264 ralign
= BYTES_PER_WORD
;
2268 * Redzoning and user store require word alignment or possibly larger.
2269 * Note this will be overridden by architecture or caller mandated
2270 * alignment if either is greater than BYTES_PER_WORD.
2272 if (flags
& SLAB_STORE_USER
)
2273 ralign
= BYTES_PER_WORD
;
2275 if (flags
& SLAB_RED_ZONE
) {
2276 ralign
= REDZONE_ALIGN
;
2277 /* If redzoning, ensure that the second redzone is suitably
2278 * aligned, by adjusting the object size accordingly. */
2279 size
+= REDZONE_ALIGN
- 1;
2280 size
&= ~(REDZONE_ALIGN
- 1);
2283 /* 2) arch mandated alignment */
2284 if (ralign
< ARCH_SLAB_MINALIGN
) {
2285 ralign
= ARCH_SLAB_MINALIGN
;
2287 /* 3) caller mandated alignment */
2288 if (ralign
< align
) {
2291 /* disable debug if necessary */
2292 if (ralign
> __alignof__(unsigned long long))
2293 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2299 if (slab_is_available())
2304 /* Get cache's description obj. */
2305 cachep
= kmem_cache_zalloc(&cache_cache
, gfp
);
2310 cachep
->obj_size
= size
;
2313 * Both debugging options require word-alignment which is calculated
2316 if (flags
& SLAB_RED_ZONE
) {
2317 /* add space for red zone words */
2318 cachep
->obj_offset
+= sizeof(unsigned long long);
2319 size
+= 2 * sizeof(unsigned long long);
2321 if (flags
& SLAB_STORE_USER
) {
2322 /* user store requires one word storage behind the end of
2323 * the real object. But if the second red zone needs to be
2324 * aligned to 64 bits, we must allow that much space.
2326 if (flags
& SLAB_RED_ZONE
)
2327 size
+= REDZONE_ALIGN
;
2329 size
+= BYTES_PER_WORD
;
2331 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2332 if (size
>= malloc_sizes
[INDEX_L3
+ 1].cs_size
2333 && cachep
->obj_size
> cache_line_size() && size
< PAGE_SIZE
) {
2334 cachep
->obj_offset
+= PAGE_SIZE
- size
;
2341 * Determine if the slab management is 'on' or 'off' slab.
2342 * (bootstrapping cannot cope with offslab caches so don't do
2345 if ((size
>= (PAGE_SIZE
>> 3)) && !slab_early_init
)
2347 * Size is large, assume best to place the slab management obj
2348 * off-slab (should allow better packing of objs).
2350 flags
|= CFLGS_OFF_SLAB
;
2352 size
= ALIGN(size
, align
);
2354 left_over
= calculate_slab_order(cachep
, size
, align
, flags
);
2358 "kmem_cache_create: couldn't create cache %s.\n", name
);
2359 kmem_cache_free(&cache_cache
, cachep
);
2363 slab_size
= ALIGN(cachep
->num
* sizeof(kmem_bufctl_t
)
2364 + sizeof(struct slab
), align
);
2367 * If the slab has been placed off-slab, and we have enough space then
2368 * move it on-slab. This is at the expense of any extra colouring.
2370 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= slab_size
) {
2371 flags
&= ~CFLGS_OFF_SLAB
;
2372 left_over
-= slab_size
;
2375 if (flags
& CFLGS_OFF_SLAB
) {
2376 /* really off slab. No need for manual alignment */
2378 cachep
->num
* sizeof(kmem_bufctl_t
) + sizeof(struct slab
);
2381 cachep
->colour_off
= cache_line_size();
2382 /* Offset must be a multiple of the alignment. */
2383 if (cachep
->colour_off
< align
)
2384 cachep
->colour_off
= align
;
2385 cachep
->colour
= left_over
/ cachep
->colour_off
;
2386 cachep
->slab_size
= slab_size
;
2387 cachep
->flags
= flags
;
2388 cachep
->gfpflags
= 0;
2389 if (CONFIG_ZONE_DMA_FLAG
&& (flags
& SLAB_CACHE_DMA
))
2390 cachep
->gfpflags
|= GFP_DMA
;
2391 cachep
->buffer_size
= size
;
2392 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2394 if (flags
& CFLGS_OFF_SLAB
) {
2395 cachep
->slabp_cache
= kmem_find_general_cachep(slab_size
, 0u);
2397 * This is a possibility for one of the malloc_sizes caches.
2398 * But since we go off slab only for object size greater than
2399 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2400 * this should not happen at all.
2401 * But leave a BUG_ON for some lucky dude.
2403 BUG_ON(ZERO_OR_NULL_PTR(cachep
->slabp_cache
));
2405 cachep
->ctor
= ctor
;
2406 cachep
->name
= name
;
2408 if (setup_cpu_cache(cachep
, gfp
)) {
2409 __kmem_cache_destroy(cachep
);
2414 /* cache setup completed, link it into the list */
2415 list_add(&cachep
->next
, &cache_chain
);
2417 if (!cachep
&& (flags
& SLAB_PANIC
))
2418 panic("kmem_cache_create(): failed to create slab `%s'\n",
2420 if (slab_is_available()) {
2421 mutex_unlock(&cache_chain_mutex
);
2426 EXPORT_SYMBOL(kmem_cache_create
);
2429 static void check_irq_off(void)
2431 BUG_ON(!irqs_disabled());
2434 static void check_irq_on(void)
2436 BUG_ON(irqs_disabled());
2439 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2443 assert_spin_locked(&cachep
->nodelists
[numa_node_id()]->list_lock
);
2447 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2451 assert_spin_locked(&cachep
->nodelists
[node
]->list_lock
);
2456 #define check_irq_off() do { } while(0)
2457 #define check_irq_on() do { } while(0)
2458 #define check_spinlock_acquired(x) do { } while(0)
2459 #define check_spinlock_acquired_node(x, y) do { } while(0)
2462 static void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
2463 struct array_cache
*ac
,
2464 int force
, int node
);
2466 static void do_drain(void *arg
)
2468 struct kmem_cache
*cachep
= arg
;
2469 struct array_cache
*ac
;
2470 int node
= numa_node_id();
2473 ac
= cpu_cache_get(cachep
);
2474 spin_lock(&cachep
->nodelists
[node
]->list_lock
);
2475 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
2476 spin_unlock(&cachep
->nodelists
[node
]->list_lock
);
2480 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2482 struct kmem_list3
*l3
;
2485 on_each_cpu(do_drain
, cachep
, 1);
2487 for_each_online_node(node
) {
2488 l3
= cachep
->nodelists
[node
];
2489 if (l3
&& l3
->alien
)
2490 drain_alien_cache(cachep
, l3
->alien
);
2493 for_each_online_node(node
) {
2494 l3
= cachep
->nodelists
[node
];
2496 drain_array(cachep
, l3
, l3
->shared
, 1, node
);
2501 * Remove slabs from the list of free slabs.
2502 * Specify the number of slabs to drain in tofree.
2504 * Returns the actual number of slabs released.
2506 static int drain_freelist(struct kmem_cache
*cache
,
2507 struct kmem_list3
*l3
, int tofree
)
2509 struct list_head
*p
;
2514 while (nr_freed
< tofree
&& !list_empty(&l3
->slabs_free
)) {
2516 spin_lock_irq(&l3
->list_lock
);
2517 p
= l3
->slabs_free
.prev
;
2518 if (p
== &l3
->slabs_free
) {
2519 spin_unlock_irq(&l3
->list_lock
);
2523 slabp
= list_entry(p
, struct slab
, list
);
2525 BUG_ON(slabp
->inuse
);
2527 list_del(&slabp
->list
);
2529 * Safe to drop the lock. The slab is no longer linked
2532 l3
->free_objects
-= cache
->num
;
2533 spin_unlock_irq(&l3
->list_lock
);
2534 slab_destroy(cache
, slabp
);
2541 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2542 static int __cache_shrink(struct kmem_cache
*cachep
)
2545 struct kmem_list3
*l3
;
2547 drain_cpu_caches(cachep
);
2550 for_each_online_node(i
) {
2551 l3
= cachep
->nodelists
[i
];
2555 drain_freelist(cachep
, l3
, l3
->free_objects
);
2557 ret
+= !list_empty(&l3
->slabs_full
) ||
2558 !list_empty(&l3
->slabs_partial
);
2560 return (ret
? 1 : 0);
2564 * kmem_cache_shrink - Shrink a cache.
2565 * @cachep: The cache to shrink.
2567 * Releases as many slabs as possible for a cache.
2568 * To help debugging, a zero exit status indicates all slabs were released.
2570 int kmem_cache_shrink(struct kmem_cache
*cachep
)
2573 BUG_ON(!cachep
|| in_interrupt());
2576 mutex_lock(&cache_chain_mutex
);
2577 ret
= __cache_shrink(cachep
);
2578 mutex_unlock(&cache_chain_mutex
);
2582 EXPORT_SYMBOL(kmem_cache_shrink
);
2585 * kmem_cache_destroy - delete a cache
2586 * @cachep: the cache to destroy
2588 * Remove a &struct kmem_cache object from the slab cache.
2590 * It is expected this function will be called by a module when it is
2591 * unloaded. This will remove the cache completely, and avoid a duplicate
2592 * cache being allocated each time a module is loaded and unloaded, if the
2593 * module doesn't have persistent in-kernel storage across loads and unloads.
2595 * The cache must be empty before calling this function.
2597 * The caller must guarantee that noone will allocate memory from the cache
2598 * during the kmem_cache_destroy().
2600 void kmem_cache_destroy(struct kmem_cache
*cachep
)
2602 BUG_ON(!cachep
|| in_interrupt());
2604 /* Find the cache in the chain of caches. */
2606 mutex_lock(&cache_chain_mutex
);
2608 * the chain is never empty, cache_cache is never destroyed
2610 list_del(&cachep
->next
);
2611 if (__cache_shrink(cachep
)) {
2612 slab_error(cachep
, "Can't free all objects");
2613 list_add(&cachep
->next
, &cache_chain
);
2614 mutex_unlock(&cache_chain_mutex
);
2619 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
2622 __kmem_cache_destroy(cachep
);
2623 mutex_unlock(&cache_chain_mutex
);
2626 EXPORT_SYMBOL(kmem_cache_destroy
);
2629 * Get the memory for a slab management obj.
2630 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2631 * always come from malloc_sizes caches. The slab descriptor cannot
2632 * come from the same cache which is getting created because,
2633 * when we are searching for an appropriate cache for these
2634 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2635 * If we are creating a malloc_sizes cache here it would not be visible to
2636 * kmem_find_general_cachep till the initialization is complete.
2637 * Hence we cannot have slabp_cache same as the original cache.
2639 static struct slab
*alloc_slabmgmt(struct kmem_cache
*cachep
, void *objp
,
2640 int colour_off
, gfp_t local_flags
,
2645 if (OFF_SLAB(cachep
)) {
2646 /* Slab management obj is off-slab. */
2647 slabp
= kmem_cache_alloc_node(cachep
->slabp_cache
,
2648 local_flags
, nodeid
);
2650 * If the first object in the slab is leaked (it's allocated
2651 * but no one has a reference to it), we want to make sure
2652 * kmemleak does not treat the ->s_mem pointer as a reference
2653 * to the object. Otherwise we will not report the leak.
2655 kmemleak_scan_area(slabp
, offsetof(struct slab
, list
),
2656 sizeof(struct list_head
), local_flags
);
2660 slabp
= objp
+ colour_off
;
2661 colour_off
+= cachep
->slab_size
;
2664 slabp
->colouroff
= colour_off
;
2665 slabp
->s_mem
= objp
+ colour_off
;
2666 slabp
->nodeid
= nodeid
;
2671 static inline kmem_bufctl_t
*slab_bufctl(struct slab
*slabp
)
2673 return (kmem_bufctl_t
*) (slabp
+ 1);
2676 static void cache_init_objs(struct kmem_cache
*cachep
,
2681 for (i
= 0; i
< cachep
->num
; i
++) {
2682 void *objp
= index_to_obj(cachep
, slabp
, i
);
2684 /* need to poison the objs? */
2685 if (cachep
->flags
& SLAB_POISON
)
2686 poison_obj(cachep
, objp
, POISON_FREE
);
2687 if (cachep
->flags
& SLAB_STORE_USER
)
2688 *dbg_userword(cachep
, objp
) = NULL
;
2690 if (cachep
->flags
& SLAB_RED_ZONE
) {
2691 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2692 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2695 * Constructors are not allowed to allocate memory from the same
2696 * cache which they are a constructor for. Otherwise, deadlock.
2697 * They must also be threaded.
2699 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2700 cachep
->ctor(objp
+ obj_offset(cachep
));
2702 if (cachep
->flags
& SLAB_RED_ZONE
) {
2703 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2704 slab_error(cachep
, "constructor overwrote the"
2705 " end of an object");
2706 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2707 slab_error(cachep
, "constructor overwrote the"
2708 " start of an object");
2710 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 &&
2711 OFF_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
)
2712 kernel_map_pages(virt_to_page(objp
),
2713 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2718 slab_bufctl(slabp
)[i
] = i
+ 1;
2720 slab_bufctl(slabp
)[i
- 1] = BUFCTL_END
;
2723 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2725 if (CONFIG_ZONE_DMA_FLAG
) {
2726 if (flags
& GFP_DMA
)
2727 BUG_ON(!(cachep
->gfpflags
& GFP_DMA
));
2729 BUG_ON(cachep
->gfpflags
& GFP_DMA
);
2733 static void *slab_get_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2736 void *objp
= index_to_obj(cachep
, slabp
, slabp
->free
);
2740 next
= slab_bufctl(slabp
)[slabp
->free
];
2742 slab_bufctl(slabp
)[slabp
->free
] = BUFCTL_FREE
;
2743 WARN_ON(slabp
->nodeid
!= nodeid
);
2750 static void slab_put_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2751 void *objp
, int nodeid
)
2753 unsigned int objnr
= obj_to_index(cachep
, slabp
, objp
);
2756 /* Verify that the slab belongs to the intended node */
2757 WARN_ON(slabp
->nodeid
!= nodeid
);
2759 if (slab_bufctl(slabp
)[objnr
] + 1 <= SLAB_LIMIT
+ 1) {
2760 printk(KERN_ERR
"slab: double free detected in cache "
2761 "'%s', objp %p\n", cachep
->name
, objp
);
2765 slab_bufctl(slabp
)[objnr
] = slabp
->free
;
2766 slabp
->free
= objnr
;
2771 * Map pages beginning at addr to the given cache and slab. This is required
2772 * for the slab allocator to be able to lookup the cache and slab of a
2773 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2775 static void slab_map_pages(struct kmem_cache
*cache
, struct slab
*slab
,
2781 page
= virt_to_page(addr
);
2784 if (likely(!PageCompound(page
)))
2785 nr_pages
<<= cache
->gfporder
;
2788 page_set_cache(page
, cache
);
2789 page_set_slab(page
, slab
);
2791 } while (--nr_pages
);
2795 * Grow (by 1) the number of slabs within a cache. This is called by
2796 * kmem_cache_alloc() when there are no active objs left in a cache.
2798 static int cache_grow(struct kmem_cache
*cachep
,
2799 gfp_t flags
, int nodeid
, void *objp
)
2804 struct kmem_list3
*l3
;
2807 * Be lazy and only check for valid flags here, keeping it out of the
2808 * critical path in kmem_cache_alloc().
2810 BUG_ON(flags
& GFP_SLAB_BUG_MASK
);
2811 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2813 /* Take the l3 list lock to change the colour_next on this node */
2815 l3
= cachep
->nodelists
[nodeid
];
2816 spin_lock(&l3
->list_lock
);
2818 /* Get colour for the slab, and cal the next value. */
2819 offset
= l3
->colour_next
;
2821 if (l3
->colour_next
>= cachep
->colour
)
2822 l3
->colour_next
= 0;
2823 spin_unlock(&l3
->list_lock
);
2825 offset
*= cachep
->colour_off
;
2827 if (local_flags
& __GFP_WAIT
)
2831 * The test for missing atomic flag is performed here, rather than
2832 * the more obvious place, simply to reduce the critical path length
2833 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2834 * will eventually be caught here (where it matters).
2836 kmem_flagcheck(cachep
, flags
);
2839 * Get mem for the objs. Attempt to allocate a physical page from
2843 objp
= kmem_getpages(cachep
, local_flags
, nodeid
);
2847 /* Get slab management. */
2848 slabp
= alloc_slabmgmt(cachep
, objp
, offset
,
2849 local_flags
& ~GFP_CONSTRAINT_MASK
, nodeid
);
2853 slab_map_pages(cachep
, slabp
, objp
);
2855 cache_init_objs(cachep
, slabp
);
2857 if (local_flags
& __GFP_WAIT
)
2858 local_irq_disable();
2860 spin_lock(&l3
->list_lock
);
2862 /* Make slab active. */
2863 list_add_tail(&slabp
->list
, &(l3
->slabs_free
));
2864 STATS_INC_GROWN(cachep
);
2865 l3
->free_objects
+= cachep
->num
;
2866 spin_unlock(&l3
->list_lock
);
2869 kmem_freepages(cachep
, objp
);
2871 if (local_flags
& __GFP_WAIT
)
2872 local_irq_disable();
2879 * Perform extra freeing checks:
2880 * - detect bad pointers.
2881 * - POISON/RED_ZONE checking
2883 static void kfree_debugcheck(const void *objp
)
2885 if (!virt_addr_valid(objp
)) {
2886 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
2887 (unsigned long)objp
);
2892 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2894 unsigned long long redzone1
, redzone2
;
2896 redzone1
= *dbg_redzone1(cache
, obj
);
2897 redzone2
= *dbg_redzone2(cache
, obj
);
2902 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2905 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2906 slab_error(cache
, "double free detected");
2908 slab_error(cache
, "memory outside object was overwritten");
2910 printk(KERN_ERR
"%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2911 obj
, redzone1
, redzone2
);
2914 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2921 BUG_ON(virt_to_cache(objp
) != cachep
);
2923 objp
-= obj_offset(cachep
);
2924 kfree_debugcheck(objp
);
2925 page
= virt_to_head_page(objp
);
2927 slabp
= page_get_slab(page
);
2929 if (cachep
->flags
& SLAB_RED_ZONE
) {
2930 verify_redzone_free(cachep
, objp
);
2931 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2932 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2934 if (cachep
->flags
& SLAB_STORE_USER
)
2935 *dbg_userword(cachep
, objp
) = caller
;
2937 objnr
= obj_to_index(cachep
, slabp
, objp
);
2939 BUG_ON(objnr
>= cachep
->num
);
2940 BUG_ON(objp
!= index_to_obj(cachep
, slabp
, objnr
));
2942 #ifdef CONFIG_DEBUG_SLAB_LEAK
2943 slab_bufctl(slabp
)[objnr
] = BUFCTL_FREE
;
2945 if (cachep
->flags
& SLAB_POISON
) {
2946 #ifdef CONFIG_DEBUG_PAGEALLOC
2947 if ((cachep
->buffer_size
% PAGE_SIZE
)==0 && OFF_SLAB(cachep
)) {
2948 store_stackinfo(cachep
, objp
, (unsigned long)caller
);
2949 kernel_map_pages(virt_to_page(objp
),
2950 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2952 poison_obj(cachep
, objp
, POISON_FREE
);
2955 poison_obj(cachep
, objp
, POISON_FREE
);
2961 static void check_slabp(struct kmem_cache
*cachep
, struct slab
*slabp
)
2966 /* Check slab's freelist to see if this obj is there. */
2967 for (i
= slabp
->free
; i
!= BUFCTL_END
; i
= slab_bufctl(slabp
)[i
]) {
2969 if (entries
> cachep
->num
|| i
>= cachep
->num
)
2972 if (entries
!= cachep
->num
- slabp
->inuse
) {
2974 printk(KERN_ERR
"slab: Internal list corruption detected in "
2975 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2976 cachep
->name
, cachep
->num
, slabp
, slabp
->inuse
);
2978 i
< sizeof(*slabp
) + cachep
->num
* sizeof(kmem_bufctl_t
);
2981 printk("\n%03x:", i
);
2982 printk(" %02x", ((unsigned char *)slabp
)[i
]);
2989 #define kfree_debugcheck(x) do { } while(0)
2990 #define cache_free_debugcheck(x,objp,z) (objp)
2991 #define check_slabp(x,y) do { } while(0)
2994 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
2997 struct kmem_list3
*l3
;
2998 struct array_cache
*ac
;
3003 node
= numa_node_id();
3004 ac
= cpu_cache_get(cachep
);
3005 batchcount
= ac
->batchcount
;
3006 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
3008 * If there was little recent activity on this cache, then
3009 * perform only a partial refill. Otherwise we could generate
3012 batchcount
= BATCHREFILL_LIMIT
;
3014 l3
= cachep
->nodelists
[node
];
3016 BUG_ON(ac
->avail
> 0 || !l3
);
3017 spin_lock(&l3
->list_lock
);
3019 /* See if we can refill from the shared array */
3020 if (l3
->shared
&& transfer_objects(ac
, l3
->shared
, batchcount
))
3023 while (batchcount
> 0) {
3024 struct list_head
*entry
;
3026 /* Get slab alloc is to come from. */
3027 entry
= l3
->slabs_partial
.next
;
3028 if (entry
== &l3
->slabs_partial
) {
3029 l3
->free_touched
= 1;
3030 entry
= l3
->slabs_free
.next
;
3031 if (entry
== &l3
->slabs_free
)
3035 slabp
= list_entry(entry
, struct slab
, list
);
3036 check_slabp(cachep
, slabp
);
3037 check_spinlock_acquired(cachep
);
3040 * The slab was either on partial or free list so
3041 * there must be at least one object available for
3044 BUG_ON(slabp
->inuse
>= cachep
->num
);
3046 while (slabp
->inuse
< cachep
->num
&& batchcount
--) {
3047 STATS_INC_ALLOCED(cachep
);
3048 STATS_INC_ACTIVE(cachep
);
3049 STATS_SET_HIGH(cachep
);
3051 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, slabp
,
3054 check_slabp(cachep
, slabp
);
3056 /* move slabp to correct slabp list: */
3057 list_del(&slabp
->list
);
3058 if (slabp
->free
== BUFCTL_END
)
3059 list_add(&slabp
->list
, &l3
->slabs_full
);
3061 list_add(&slabp
->list
, &l3
->slabs_partial
);
3065 l3
->free_objects
-= ac
->avail
;
3067 spin_unlock(&l3
->list_lock
);
3069 if (unlikely(!ac
->avail
)) {
3071 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, node
, NULL
);
3073 /* cache_grow can reenable interrupts, then ac could change. */
3074 ac
= cpu_cache_get(cachep
);
3075 if (!x
&& ac
->avail
== 0) /* no objects in sight? abort */
3078 if (!ac
->avail
) /* objects refilled by interrupt? */
3082 return ac
->entry
[--ac
->avail
];
3085 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
3088 might_sleep_if(flags
& __GFP_WAIT
);
3090 kmem_flagcheck(cachep
, flags
);
3095 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
3096 gfp_t flags
, void *objp
, void *caller
)
3100 if (cachep
->flags
& SLAB_POISON
) {
3101 #ifdef CONFIG_DEBUG_PAGEALLOC
3102 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
3103 kernel_map_pages(virt_to_page(objp
),
3104 cachep
->buffer_size
/ PAGE_SIZE
, 1);
3106 check_poison_obj(cachep
, objp
);
3108 check_poison_obj(cachep
, objp
);
3110 poison_obj(cachep
, objp
, POISON_INUSE
);
3112 if (cachep
->flags
& SLAB_STORE_USER
)
3113 *dbg_userword(cachep
, objp
) = caller
;
3115 if (cachep
->flags
& SLAB_RED_ZONE
) {
3116 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
3117 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
3118 slab_error(cachep
, "double free, or memory outside"
3119 " object was overwritten");
3121 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3122 objp
, *dbg_redzone1(cachep
, objp
),
3123 *dbg_redzone2(cachep
, objp
));
3125 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
3126 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
3128 #ifdef CONFIG_DEBUG_SLAB_LEAK
3133 slabp
= page_get_slab(virt_to_head_page(objp
));
3134 objnr
= (unsigned)(objp
- slabp
->s_mem
) / cachep
->buffer_size
;
3135 slab_bufctl(slabp
)[objnr
] = BUFCTL_ACTIVE
;
3138 objp
+= obj_offset(cachep
);
3139 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
3141 #if ARCH_SLAB_MINALIGN
3142 if ((u32
)objp
& (ARCH_SLAB_MINALIGN
-1)) {
3143 printk(KERN_ERR
"0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3144 objp
, ARCH_SLAB_MINALIGN
);
3150 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3153 static bool slab_should_failslab(struct kmem_cache
*cachep
, gfp_t flags
)
3155 if (cachep
== &cache_cache
)
3158 return should_failslab(obj_size(cachep
), flags
);
3161 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3164 struct array_cache
*ac
;
3168 ac
= cpu_cache_get(cachep
);
3169 if (likely(ac
->avail
)) {
3170 STATS_INC_ALLOCHIT(cachep
);
3172 objp
= ac
->entry
[--ac
->avail
];
3174 STATS_INC_ALLOCMISS(cachep
);
3175 objp
= cache_alloc_refill(cachep
, flags
);
3178 * To avoid a false negative, if an object that is in one of the
3179 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3180 * treat the array pointers as a reference to the object.
3182 kmemleak_erase(&ac
->entry
[ac
->avail
]);
3188 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3190 * If we are in_interrupt, then process context, including cpusets and
3191 * mempolicy, may not apply and should not be used for allocation policy.
3193 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3195 int nid_alloc
, nid_here
;
3197 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3199 nid_alloc
= nid_here
= numa_node_id();
3200 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3201 nid_alloc
= cpuset_mem_spread_node();
3202 else if (current
->mempolicy
)
3203 nid_alloc
= slab_node(current
->mempolicy
);
3204 if (nid_alloc
!= nid_here
)
3205 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3210 * Fallback function if there was no memory available and no objects on a
3211 * certain node and fall back is permitted. First we scan all the
3212 * available nodelists for available objects. If that fails then we
3213 * perform an allocation without specifying a node. This allows the page
3214 * allocator to do its reclaim / fallback magic. We then insert the
3215 * slab into the proper nodelist and then allocate from it.
3217 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3219 struct zonelist
*zonelist
;
3223 enum zone_type high_zoneidx
= gfp_zone(flags
);
3227 if (flags
& __GFP_THISNODE
)
3230 zonelist
= node_zonelist(slab_node(current
->mempolicy
), flags
);
3231 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
3235 * Look through allowed nodes for objects available
3236 * from existing per node queues.
3238 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
3239 nid
= zone_to_nid(zone
);
3241 if (cpuset_zone_allowed_hardwall(zone
, flags
) &&
3242 cache
->nodelists
[nid
] &&
3243 cache
->nodelists
[nid
]->free_objects
) {
3244 obj
= ____cache_alloc_node(cache
,
3245 flags
| GFP_THISNODE
, nid
);
3253 * This allocation will be performed within the constraints
3254 * of the current cpuset / memory policy requirements.
3255 * We may trigger various forms of reclaim on the allowed
3256 * set and go into memory reserves if necessary.
3258 if (local_flags
& __GFP_WAIT
)
3260 kmem_flagcheck(cache
, flags
);
3261 obj
= kmem_getpages(cache
, local_flags
, numa_node_id());
3262 if (local_flags
& __GFP_WAIT
)
3263 local_irq_disable();
3266 * Insert into the appropriate per node queues
3268 nid
= page_to_nid(virt_to_page(obj
));
3269 if (cache_grow(cache
, flags
, nid
, obj
)) {
3270 obj
= ____cache_alloc_node(cache
,
3271 flags
| GFP_THISNODE
, nid
);
3274 * Another processor may allocate the
3275 * objects in the slab since we are
3276 * not holding any locks.
3280 /* cache_grow already freed obj */
3289 * A interface to enable slab creation on nodeid
3291 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3294 struct list_head
*entry
;
3296 struct kmem_list3
*l3
;
3300 l3
= cachep
->nodelists
[nodeid
];
3305 spin_lock(&l3
->list_lock
);
3306 entry
= l3
->slabs_partial
.next
;
3307 if (entry
== &l3
->slabs_partial
) {
3308 l3
->free_touched
= 1;
3309 entry
= l3
->slabs_free
.next
;
3310 if (entry
== &l3
->slabs_free
)
3314 slabp
= list_entry(entry
, struct slab
, list
);
3315 check_spinlock_acquired_node(cachep
, nodeid
);
3316 check_slabp(cachep
, slabp
);
3318 STATS_INC_NODEALLOCS(cachep
);
3319 STATS_INC_ACTIVE(cachep
);
3320 STATS_SET_HIGH(cachep
);
3322 BUG_ON(slabp
->inuse
== cachep
->num
);
3324 obj
= slab_get_obj(cachep
, slabp
, nodeid
);
3325 check_slabp(cachep
, slabp
);
3327 /* move slabp to correct slabp list: */
3328 list_del(&slabp
->list
);
3330 if (slabp
->free
== BUFCTL_END
)
3331 list_add(&slabp
->list
, &l3
->slabs_full
);
3333 list_add(&slabp
->list
, &l3
->slabs_partial
);
3335 spin_unlock(&l3
->list_lock
);
3339 spin_unlock(&l3
->list_lock
);
3340 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, nodeid
, NULL
);
3344 return fallback_alloc(cachep
, flags
);
3351 * kmem_cache_alloc_node - Allocate an object on the specified node
3352 * @cachep: The cache to allocate from.
3353 * @flags: See kmalloc().
3354 * @nodeid: node number of the target node.
3355 * @caller: return address of caller, used for debug information
3357 * Identical to kmem_cache_alloc but it will allocate memory on the given
3358 * node, which can improve the performance for cpu bound structures.
3360 * Fallback to other node is possible if __GFP_THISNODE is not set.
3362 static __always_inline
void *
3363 __cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3366 unsigned long save_flags
;
3369 flags
&= slab_gfp_mask
;
3371 lockdep_trace_alloc(flags
);
3373 if (slab_should_failslab(cachep
, flags
))
3376 cache_alloc_debugcheck_before(cachep
, flags
);
3377 local_irq_save(save_flags
);
3379 if (unlikely(nodeid
== -1))
3380 nodeid
= numa_node_id();
3382 if (unlikely(!cachep
->nodelists
[nodeid
])) {
3383 /* Node not bootstrapped yet */
3384 ptr
= fallback_alloc(cachep
, flags
);
3388 if (nodeid
== numa_node_id()) {
3390 * Use the locally cached objects if possible.
3391 * However ____cache_alloc does not allow fallback
3392 * to other nodes. It may fail while we still have
3393 * objects on other nodes available.
3395 ptr
= ____cache_alloc(cachep
, flags
);
3399 /* ___cache_alloc_node can fall back to other nodes */
3400 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3402 local_irq_restore(save_flags
);
3403 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3404 kmemleak_alloc_recursive(ptr
, obj_size(cachep
), 1, cachep
->flags
,
3407 if (unlikely((flags
& __GFP_ZERO
) && ptr
))
3408 memset(ptr
, 0, obj_size(cachep
));
3413 static __always_inline
void *
3414 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3418 if (unlikely(current
->flags
& (PF_SPREAD_SLAB
| PF_MEMPOLICY
))) {
3419 objp
= alternate_node_alloc(cache
, flags
);
3423 objp
= ____cache_alloc(cache
, flags
);
3426 * We may just have run out of memory on the local node.
3427 * ____cache_alloc_node() knows how to locate memory on other nodes
3430 objp
= ____cache_alloc_node(cache
, flags
, numa_node_id());
3437 static __always_inline
void *
3438 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3440 return ____cache_alloc(cachep
, flags
);
3443 #endif /* CONFIG_NUMA */
3445 static __always_inline
void *
3446 __cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
, void *caller
)
3448 unsigned long save_flags
;
3451 flags
&= slab_gfp_mask
;
3453 lockdep_trace_alloc(flags
);
3455 if (slab_should_failslab(cachep
, flags
))
3458 cache_alloc_debugcheck_before(cachep
, flags
);
3459 local_irq_save(save_flags
);
3460 objp
= __do_cache_alloc(cachep
, flags
);
3461 local_irq_restore(save_flags
);
3462 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3463 kmemleak_alloc_recursive(objp
, obj_size(cachep
), 1, cachep
->flags
,
3467 if (unlikely((flags
& __GFP_ZERO
) && objp
))
3468 memset(objp
, 0, obj_size(cachep
));
3474 * Caller needs to acquire correct kmem_list's list_lock
3476 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int nr_objects
,
3480 struct kmem_list3
*l3
;
3482 for (i
= 0; i
< nr_objects
; i
++) {
3483 void *objp
= objpp
[i
];
3486 slabp
= virt_to_slab(objp
);
3487 l3
= cachep
->nodelists
[node
];
3488 list_del(&slabp
->list
);
3489 check_spinlock_acquired_node(cachep
, node
);
3490 check_slabp(cachep
, slabp
);
3491 slab_put_obj(cachep
, slabp
, objp
, node
);
3492 STATS_DEC_ACTIVE(cachep
);
3494 check_slabp(cachep
, slabp
);
3496 /* fixup slab chains */
3497 if (slabp
->inuse
== 0) {
3498 if (l3
->free_objects
> l3
->free_limit
) {
3499 l3
->free_objects
-= cachep
->num
;
3500 /* No need to drop any previously held
3501 * lock here, even if we have a off-slab slab
3502 * descriptor it is guaranteed to come from
3503 * a different cache, refer to comments before
3506 slab_destroy(cachep
, slabp
);
3508 list_add(&slabp
->list
, &l3
->slabs_free
);
3511 /* Unconditionally move a slab to the end of the
3512 * partial list on free - maximum time for the
3513 * other objects to be freed, too.
3515 list_add_tail(&slabp
->list
, &l3
->slabs_partial
);
3520 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3523 struct kmem_list3
*l3
;
3524 int node
= numa_node_id();
3526 batchcount
= ac
->batchcount
;
3528 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
3531 l3
= cachep
->nodelists
[node
];
3532 spin_lock(&l3
->list_lock
);
3534 struct array_cache
*shared_array
= l3
->shared
;
3535 int max
= shared_array
->limit
- shared_array
->avail
;
3537 if (batchcount
> max
)
3539 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3540 ac
->entry
, sizeof(void *) * batchcount
);
3541 shared_array
->avail
+= batchcount
;
3546 free_block(cachep
, ac
->entry
, batchcount
, node
);
3551 struct list_head
*p
;
3553 p
= l3
->slabs_free
.next
;
3554 while (p
!= &(l3
->slabs_free
)) {
3557 slabp
= list_entry(p
, struct slab
, list
);
3558 BUG_ON(slabp
->inuse
);
3563 STATS_SET_FREEABLE(cachep
, i
);
3566 spin_unlock(&l3
->list_lock
);
3567 ac
->avail
-= batchcount
;
3568 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3572 * Release an obj back to its cache. If the obj has a constructed state, it must
3573 * be in this state _before_ it is released. Called with disabled ints.
3575 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
)
3577 struct array_cache
*ac
= cpu_cache_get(cachep
);
3580 kmemleak_free_recursive(objp
, cachep
->flags
);
3581 objp
= cache_free_debugcheck(cachep
, objp
, __builtin_return_address(0));
3584 * Skip calling cache_free_alien() when the platform is not numa.
3585 * This will avoid cache misses that happen while accessing slabp (which
3586 * is per page memory reference) to get nodeid. Instead use a global
3587 * variable to skip the call, which is mostly likely to be present in
3590 if (nr_online_nodes
> 1 && cache_free_alien(cachep
, objp
))
3593 if (likely(ac
->avail
< ac
->limit
)) {
3594 STATS_INC_FREEHIT(cachep
);
3595 ac
->entry
[ac
->avail
++] = objp
;
3598 STATS_INC_FREEMISS(cachep
);
3599 cache_flusharray(cachep
, ac
);
3600 ac
->entry
[ac
->avail
++] = objp
;
3605 * kmem_cache_alloc - Allocate an object
3606 * @cachep: The cache to allocate from.
3607 * @flags: See kmalloc().
3609 * Allocate an object from this cache. The flags are only relevant
3610 * if the cache has no available objects.
3612 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3614 void *ret
= __cache_alloc(cachep
, flags
, __builtin_return_address(0));
3616 trace_kmem_cache_alloc(_RET_IP_
, ret
,
3617 obj_size(cachep
), cachep
->buffer_size
, flags
);
3621 EXPORT_SYMBOL(kmem_cache_alloc
);
3623 #ifdef CONFIG_KMEMTRACE
3624 void *kmem_cache_alloc_notrace(struct kmem_cache
*cachep
, gfp_t flags
)
3626 return __cache_alloc(cachep
, flags
, __builtin_return_address(0));
3628 EXPORT_SYMBOL(kmem_cache_alloc_notrace
);
3632 * kmem_ptr_validate - check if an untrusted pointer might be a slab entry.
3633 * @cachep: the cache we're checking against
3634 * @ptr: pointer to validate
3636 * This verifies that the untrusted pointer looks sane;
3637 * it is _not_ a guarantee that the pointer is actually
3638 * part of the slab cache in question, but it at least
3639 * validates that the pointer can be dereferenced and
3640 * looks half-way sane.
3642 * Currently only used for dentry validation.
3644 int kmem_ptr_validate(struct kmem_cache
*cachep
, const void *ptr
)
3646 unsigned long addr
= (unsigned long)ptr
;
3647 unsigned long min_addr
= PAGE_OFFSET
;
3648 unsigned long align_mask
= BYTES_PER_WORD
- 1;
3649 unsigned long size
= cachep
->buffer_size
;
3652 if (unlikely(addr
< min_addr
))
3654 if (unlikely(addr
> (unsigned long)high_memory
- size
))
3656 if (unlikely(addr
& align_mask
))
3658 if (unlikely(!kern_addr_valid(addr
)))
3660 if (unlikely(!kern_addr_valid(addr
+ size
- 1)))
3662 page
= virt_to_page(ptr
);
3663 if (unlikely(!PageSlab(page
)))
3665 if (unlikely(page_get_cache(page
) != cachep
))
3673 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3675 void *ret
= __cache_alloc_node(cachep
, flags
, nodeid
,
3676 __builtin_return_address(0));
3678 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
3679 obj_size(cachep
), cachep
->buffer_size
,
3684 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3686 #ifdef CONFIG_KMEMTRACE
3687 void *kmem_cache_alloc_node_notrace(struct kmem_cache
*cachep
,
3691 return __cache_alloc_node(cachep
, flags
, nodeid
,
3692 __builtin_return_address(0));
3694 EXPORT_SYMBOL(kmem_cache_alloc_node_notrace
);
3697 static __always_inline
void *
3698 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, void *caller
)
3700 struct kmem_cache
*cachep
;
3703 cachep
= kmem_find_general_cachep(size
, flags
);
3704 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3706 ret
= kmem_cache_alloc_node_notrace(cachep
, flags
, node
);
3708 trace_kmalloc_node((unsigned long) caller
, ret
,
3709 size
, cachep
->buffer_size
, flags
, node
);
3714 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_KMEMTRACE)
3715 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3717 return __do_kmalloc_node(size
, flags
, node
,
3718 __builtin_return_address(0));
3720 EXPORT_SYMBOL(__kmalloc_node
);
3722 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3723 int node
, unsigned long caller
)
3725 return __do_kmalloc_node(size
, flags
, node
, (void *)caller
);
3727 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3729 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3731 return __do_kmalloc_node(size
, flags
, node
, NULL
);
3733 EXPORT_SYMBOL(__kmalloc_node
);
3734 #endif /* CONFIG_DEBUG_SLAB */
3735 #endif /* CONFIG_NUMA */
3738 * __do_kmalloc - allocate memory
3739 * @size: how many bytes of memory are required.
3740 * @flags: the type of memory to allocate (see kmalloc).
3741 * @caller: function caller for debug tracking of the caller
3743 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3746 struct kmem_cache
*cachep
;
3749 /* If you want to save a few bytes .text space: replace
3751 * Then kmalloc uses the uninlined functions instead of the inline
3754 cachep
= __find_general_cachep(size
, flags
);
3755 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3757 ret
= __cache_alloc(cachep
, flags
, caller
);
3759 trace_kmalloc((unsigned long) caller
, ret
,
3760 size
, cachep
->buffer_size
, flags
);
3766 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_KMEMTRACE)
3767 void *__kmalloc(size_t size
, gfp_t flags
)
3769 return __do_kmalloc(size
, flags
, __builtin_return_address(0));
3771 EXPORT_SYMBOL(__kmalloc
);
3773 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, unsigned long caller
)
3775 return __do_kmalloc(size
, flags
, (void *)caller
);
3777 EXPORT_SYMBOL(__kmalloc_track_caller
);
3780 void *__kmalloc(size_t size
, gfp_t flags
)
3782 return __do_kmalloc(size
, flags
, NULL
);
3784 EXPORT_SYMBOL(__kmalloc
);
3788 * kmem_cache_free - Deallocate an object
3789 * @cachep: The cache the allocation was from.
3790 * @objp: The previously allocated object.
3792 * Free an object which was previously allocated from this
3795 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3797 unsigned long flags
;
3799 local_irq_save(flags
);
3800 debug_check_no_locks_freed(objp
, obj_size(cachep
));
3801 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3802 debug_check_no_obj_freed(objp
, obj_size(cachep
));
3803 __cache_free(cachep
, objp
);
3804 local_irq_restore(flags
);
3806 trace_kmem_cache_free(_RET_IP_
, objp
);
3808 EXPORT_SYMBOL(kmem_cache_free
);
3811 * kfree - free previously allocated memory
3812 * @objp: pointer returned by kmalloc.
3814 * If @objp is NULL, no operation is performed.
3816 * Don't free memory not originally allocated by kmalloc()
3817 * or you will run into trouble.
3819 void kfree(const void *objp
)
3821 struct kmem_cache
*c
;
3822 unsigned long flags
;
3824 trace_kfree(_RET_IP_
, objp
);
3826 if (unlikely(ZERO_OR_NULL_PTR(objp
)))
3828 local_irq_save(flags
);
3829 kfree_debugcheck(objp
);
3830 c
= virt_to_cache(objp
);
3831 debug_check_no_locks_freed(objp
, obj_size(c
));
3832 debug_check_no_obj_freed(objp
, obj_size(c
));
3833 __cache_free(c
, (void *)objp
);
3834 local_irq_restore(flags
);
3836 EXPORT_SYMBOL(kfree
);
3838 unsigned int kmem_cache_size(struct kmem_cache
*cachep
)
3840 return obj_size(cachep
);
3842 EXPORT_SYMBOL(kmem_cache_size
);
3844 const char *kmem_cache_name(struct kmem_cache
*cachep
)
3846 return cachep
->name
;
3848 EXPORT_SYMBOL_GPL(kmem_cache_name
);
3851 * This initializes kmem_list3 or resizes various caches for all nodes.
3853 static int alloc_kmemlist(struct kmem_cache
*cachep
, gfp_t gfp
)
3856 struct kmem_list3
*l3
;
3857 struct array_cache
*new_shared
;
3858 struct array_cache
**new_alien
= NULL
;
3860 for_each_online_node(node
) {
3862 if (use_alien_caches
) {
3863 new_alien
= alloc_alien_cache(node
, cachep
->limit
, gfp
);
3869 if (cachep
->shared
) {
3870 new_shared
= alloc_arraycache(node
,
3871 cachep
->shared
*cachep
->batchcount
,
3874 free_alien_cache(new_alien
);
3879 l3
= cachep
->nodelists
[node
];
3881 struct array_cache
*shared
= l3
->shared
;
3883 spin_lock_irq(&l3
->list_lock
);
3886 free_block(cachep
, shared
->entry
,
3887 shared
->avail
, node
);
3889 l3
->shared
= new_shared
;
3891 l3
->alien
= new_alien
;
3894 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3895 cachep
->batchcount
+ cachep
->num
;
3896 spin_unlock_irq(&l3
->list_lock
);
3898 free_alien_cache(new_alien
);
3901 l3
= kmalloc_node(sizeof(struct kmem_list3
), gfp
, node
);
3903 free_alien_cache(new_alien
);
3908 kmem_list3_init(l3
);
3909 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
3910 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
3911 l3
->shared
= new_shared
;
3912 l3
->alien
= new_alien
;
3913 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3914 cachep
->batchcount
+ cachep
->num
;
3915 cachep
->nodelists
[node
] = l3
;
3920 if (!cachep
->next
.next
) {
3921 /* Cache is not active yet. Roll back what we did */
3924 if (cachep
->nodelists
[node
]) {
3925 l3
= cachep
->nodelists
[node
];
3928 free_alien_cache(l3
->alien
);
3930 cachep
->nodelists
[node
] = NULL
;
3938 struct ccupdate_struct
{
3939 struct kmem_cache
*cachep
;
3940 struct array_cache
*new[NR_CPUS
];
3943 static void do_ccupdate_local(void *info
)
3945 struct ccupdate_struct
*new = info
;
3946 struct array_cache
*old
;
3949 old
= cpu_cache_get(new->cachep
);
3951 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
3952 new->new[smp_processor_id()] = old
;
3955 /* Always called with the cache_chain_mutex held */
3956 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3957 int batchcount
, int shared
, gfp_t gfp
)
3959 struct ccupdate_struct
*new;
3962 new = kzalloc(sizeof(*new), gfp
);
3966 for_each_online_cpu(i
) {
3967 new->new[i
] = alloc_arraycache(cpu_to_node(i
), limit
,
3970 for (i
--; i
>= 0; i
--)
3976 new->cachep
= cachep
;
3978 on_each_cpu(do_ccupdate_local
, (void *)new, 1);
3981 cachep
->batchcount
= batchcount
;
3982 cachep
->limit
= limit
;
3983 cachep
->shared
= shared
;
3985 for_each_online_cpu(i
) {
3986 struct array_cache
*ccold
= new->new[i
];
3989 spin_lock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3990 free_block(cachep
, ccold
->entry
, ccold
->avail
, cpu_to_node(i
));
3991 spin_unlock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3995 return alloc_kmemlist(cachep
, gfp
);
3998 /* Called with cache_chain_mutex held always */
3999 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
4005 * The head array serves three purposes:
4006 * - create a LIFO ordering, i.e. return objects that are cache-warm
4007 * - reduce the number of spinlock operations.
4008 * - reduce the number of linked list operations on the slab and
4009 * bufctl chains: array operations are cheaper.
4010 * The numbers are guessed, we should auto-tune as described by
4013 if (cachep
->buffer_size
> 131072)
4015 else if (cachep
->buffer_size
> PAGE_SIZE
)
4017 else if (cachep
->buffer_size
> 1024)
4019 else if (cachep
->buffer_size
> 256)
4025 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4026 * allocation behaviour: Most allocs on one cpu, most free operations
4027 * on another cpu. For these cases, an efficient object passing between
4028 * cpus is necessary. This is provided by a shared array. The array
4029 * replaces Bonwick's magazine layer.
4030 * On uniprocessor, it's functionally equivalent (but less efficient)
4031 * to a larger limit. Thus disabled by default.
4034 if (cachep
->buffer_size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
4039 * With debugging enabled, large batchcount lead to excessively long
4040 * periods with disabled local interrupts. Limit the batchcount
4045 err
= do_tune_cpucache(cachep
, limit
, (limit
+ 1) / 2, shared
, gfp
);
4047 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
4048 cachep
->name
, -err
);
4053 * Drain an array if it contains any elements taking the l3 lock only if
4054 * necessary. Note that the l3 listlock also protects the array_cache
4055 * if drain_array() is used on the shared array.
4057 void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
4058 struct array_cache
*ac
, int force
, int node
)
4062 if (!ac
|| !ac
->avail
)
4064 if (ac
->touched
&& !force
) {
4067 spin_lock_irq(&l3
->list_lock
);
4069 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
4070 if (tofree
> ac
->avail
)
4071 tofree
= (ac
->avail
+ 1) / 2;
4072 free_block(cachep
, ac
->entry
, tofree
, node
);
4073 ac
->avail
-= tofree
;
4074 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
4075 sizeof(void *) * ac
->avail
);
4077 spin_unlock_irq(&l3
->list_lock
);
4082 * cache_reap - Reclaim memory from caches.
4083 * @w: work descriptor
4085 * Called from workqueue/eventd every few seconds.
4087 * - clear the per-cpu caches for this CPU.
4088 * - return freeable pages to the main free memory pool.
4090 * If we cannot acquire the cache chain mutex then just give up - we'll try
4091 * again on the next iteration.
4093 static void cache_reap(struct work_struct
*w
)
4095 struct kmem_cache
*searchp
;
4096 struct kmem_list3
*l3
;
4097 int node
= numa_node_id();
4098 struct delayed_work
*work
= to_delayed_work(w
);
4100 if (!mutex_trylock(&cache_chain_mutex
))
4101 /* Give up. Setup the next iteration. */
4104 list_for_each_entry(searchp
, &cache_chain
, next
) {
4108 * We only take the l3 lock if absolutely necessary and we
4109 * have established with reasonable certainty that
4110 * we can do some work if the lock was obtained.
4112 l3
= searchp
->nodelists
[node
];
4114 reap_alien(searchp
, l3
);
4116 drain_array(searchp
, l3
, cpu_cache_get(searchp
), 0, node
);
4119 * These are racy checks but it does not matter
4120 * if we skip one check or scan twice.
4122 if (time_after(l3
->next_reap
, jiffies
))
4125 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
;
4127 drain_array(searchp
, l3
, l3
->shared
, 0, node
);
4129 if (l3
->free_touched
)
4130 l3
->free_touched
= 0;
4134 freed
= drain_freelist(searchp
, l3
, (l3
->free_limit
+
4135 5 * searchp
->num
- 1) / (5 * searchp
->num
));
4136 STATS_ADD_REAPED(searchp
, freed
);
4142 mutex_unlock(&cache_chain_mutex
);
4145 /* Set up the next iteration */
4146 schedule_delayed_work(work
, round_jiffies_relative(REAPTIMEOUT_CPUC
));
4149 #ifdef CONFIG_SLABINFO
4151 static void print_slabinfo_header(struct seq_file
*m
)
4154 * Output format version, so at least we can change it
4155 * without _too_ many complaints.
4158 seq_puts(m
, "slabinfo - version: 2.1 (statistics)\n");
4160 seq_puts(m
, "slabinfo - version: 2.1\n");
4162 seq_puts(m
, "# name <active_objs> <num_objs> <objsize> "
4163 "<objperslab> <pagesperslab>");
4164 seq_puts(m
, " : tunables <limit> <batchcount> <sharedfactor>");
4165 seq_puts(m
, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4167 seq_puts(m
, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4168 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4169 seq_puts(m
, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4174 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
4178 mutex_lock(&cache_chain_mutex
);
4180 print_slabinfo_header(m
);
4182 return seq_list_start(&cache_chain
, *pos
);
4185 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
4187 return seq_list_next(p
, &cache_chain
, pos
);
4190 static void s_stop(struct seq_file
*m
, void *p
)
4192 mutex_unlock(&cache_chain_mutex
);
4195 static int s_show(struct seq_file
*m
, void *p
)
4197 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, next
);
4199 unsigned long active_objs
;
4200 unsigned long num_objs
;
4201 unsigned long active_slabs
= 0;
4202 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
4206 struct kmem_list3
*l3
;
4210 for_each_online_node(node
) {
4211 l3
= cachep
->nodelists
[node
];
4216 spin_lock_irq(&l3
->list_lock
);
4218 list_for_each_entry(slabp
, &l3
->slabs_full
, list
) {
4219 if (slabp
->inuse
!= cachep
->num
&& !error
)
4220 error
= "slabs_full accounting error";
4221 active_objs
+= cachep
->num
;
4224 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
) {
4225 if (slabp
->inuse
== cachep
->num
&& !error
)
4226 error
= "slabs_partial inuse accounting error";
4227 if (!slabp
->inuse
&& !error
)
4228 error
= "slabs_partial/inuse accounting error";
4229 active_objs
+= slabp
->inuse
;
4232 list_for_each_entry(slabp
, &l3
->slabs_free
, list
) {
4233 if (slabp
->inuse
&& !error
)
4234 error
= "slabs_free/inuse accounting error";
4237 free_objects
+= l3
->free_objects
;
4239 shared_avail
+= l3
->shared
->avail
;
4241 spin_unlock_irq(&l3
->list_lock
);
4243 num_slabs
+= active_slabs
;
4244 num_objs
= num_slabs
* cachep
->num
;
4245 if (num_objs
- active_objs
!= free_objects
&& !error
)
4246 error
= "free_objects accounting error";
4248 name
= cachep
->name
;
4250 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
4252 seq_printf(m
, "%-17s %6lu %6lu %6u %4u %4d",
4253 name
, active_objs
, num_objs
, cachep
->buffer_size
,
4254 cachep
->num
, (1 << cachep
->gfporder
));
4255 seq_printf(m
, " : tunables %4u %4u %4u",
4256 cachep
->limit
, cachep
->batchcount
, cachep
->shared
);
4257 seq_printf(m
, " : slabdata %6lu %6lu %6lu",
4258 active_slabs
, num_slabs
, shared_avail
);
4261 unsigned long high
= cachep
->high_mark
;
4262 unsigned long allocs
= cachep
->num_allocations
;
4263 unsigned long grown
= cachep
->grown
;
4264 unsigned long reaped
= cachep
->reaped
;
4265 unsigned long errors
= cachep
->errors
;
4266 unsigned long max_freeable
= cachep
->max_freeable
;
4267 unsigned long node_allocs
= cachep
->node_allocs
;
4268 unsigned long node_frees
= cachep
->node_frees
;
4269 unsigned long overflows
= cachep
->node_overflow
;
4271 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu \
4272 %4lu %4lu %4lu %4lu %4lu", allocs
, high
, grown
,
4273 reaped
, errors
, max_freeable
, node_allocs
,
4274 node_frees
, overflows
);
4278 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4279 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4280 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4281 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4283 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4284 allochit
, allocmiss
, freehit
, freemiss
);
4292 * slabinfo_op - iterator that generates /proc/slabinfo
4301 * num-pages-per-slab
4302 * + further values on SMP and with statistics enabled
4305 static const struct seq_operations slabinfo_op
= {
4312 #define MAX_SLABINFO_WRITE 128
4314 * slabinfo_write - Tuning for the slab allocator
4316 * @buffer: user buffer
4317 * @count: data length
4320 ssize_t
slabinfo_write(struct file
*file
, const char __user
* buffer
,
4321 size_t count
, loff_t
*ppos
)
4323 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4324 int limit
, batchcount
, shared
, res
;
4325 struct kmem_cache
*cachep
;
4327 if (count
> MAX_SLABINFO_WRITE
)
4329 if (copy_from_user(&kbuf
, buffer
, count
))
4331 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4333 tmp
= strchr(kbuf
, ' ');
4338 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4341 /* Find the cache in the chain of caches. */
4342 mutex_lock(&cache_chain_mutex
);
4344 list_for_each_entry(cachep
, &cache_chain
, next
) {
4345 if (!strcmp(cachep
->name
, kbuf
)) {
4346 if (limit
< 1 || batchcount
< 1 ||
4347 batchcount
> limit
|| shared
< 0) {
4350 res
= do_tune_cpucache(cachep
, limit
,
4357 mutex_unlock(&cache_chain_mutex
);
4363 static int slabinfo_open(struct inode
*inode
, struct file
*file
)
4365 return seq_open(file
, &slabinfo_op
);
4368 static const struct file_operations proc_slabinfo_operations
= {
4369 .open
= slabinfo_open
,
4371 .write
= slabinfo_write
,
4372 .llseek
= seq_lseek
,
4373 .release
= seq_release
,
4376 #ifdef CONFIG_DEBUG_SLAB_LEAK
4378 static void *leaks_start(struct seq_file
*m
, loff_t
*pos
)
4380 mutex_lock(&cache_chain_mutex
);
4381 return seq_list_start(&cache_chain
, *pos
);
4384 static inline int add_caller(unsigned long *n
, unsigned long v
)
4394 unsigned long *q
= p
+ 2 * i
;
4408 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4414 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
, struct slab
*s
)
4420 for (i
= 0, p
= s
->s_mem
; i
< c
->num
; i
++, p
+= c
->buffer_size
) {
4421 if (slab_bufctl(s
)[i
] != BUFCTL_ACTIVE
)
4423 if (!add_caller(n
, (unsigned long)*dbg_userword(c
, p
)))
4428 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4430 #ifdef CONFIG_KALLSYMS
4431 unsigned long offset
, size
;
4432 char modname
[MODULE_NAME_LEN
], name
[KSYM_NAME_LEN
];
4434 if (lookup_symbol_attrs(address
, &size
, &offset
, modname
, name
) == 0) {
4435 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4437 seq_printf(m
, " [%s]", modname
);
4441 seq_printf(m
, "%p", (void *)address
);
4444 static int leaks_show(struct seq_file
*m
, void *p
)
4446 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, next
);
4448 struct kmem_list3
*l3
;
4450 unsigned long *n
= m
->private;
4454 if (!(cachep
->flags
& SLAB_STORE_USER
))
4456 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4459 /* OK, we can do it */
4463 for_each_online_node(node
) {
4464 l3
= cachep
->nodelists
[node
];
4469 spin_lock_irq(&l3
->list_lock
);
4471 list_for_each_entry(slabp
, &l3
->slabs_full
, list
)
4472 handle_slab(n
, cachep
, slabp
);
4473 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
)
4474 handle_slab(n
, cachep
, slabp
);
4475 spin_unlock_irq(&l3
->list_lock
);
4477 name
= cachep
->name
;
4479 /* Increase the buffer size */
4480 mutex_unlock(&cache_chain_mutex
);
4481 m
->private = kzalloc(n
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4483 /* Too bad, we are really out */
4485 mutex_lock(&cache_chain_mutex
);
4488 *(unsigned long *)m
->private = n
[0] * 2;
4490 mutex_lock(&cache_chain_mutex
);
4491 /* Now make sure this entry will be retried */
4495 for (i
= 0; i
< n
[1]; i
++) {
4496 seq_printf(m
, "%s: %lu ", name
, n
[2*i
+3]);
4497 show_symbol(m
, n
[2*i
+2]);
4504 static const struct seq_operations slabstats_op
= {
4505 .start
= leaks_start
,
4511 static int slabstats_open(struct inode
*inode
, struct file
*file
)
4513 unsigned long *n
= kzalloc(PAGE_SIZE
, GFP_KERNEL
);
4516 ret
= seq_open(file
, &slabstats_op
);
4518 struct seq_file
*m
= file
->private_data
;
4519 *n
= PAGE_SIZE
/ (2 * sizeof(unsigned long));
4528 static const struct file_operations proc_slabstats_operations
= {
4529 .open
= slabstats_open
,
4531 .llseek
= seq_lseek
,
4532 .release
= seq_release_private
,
4536 static int __init
slab_proc_init(void)
4538 proc_create("slabinfo",S_IWUSR
|S_IRUGO
,NULL
,&proc_slabinfo_operations
);
4539 #ifdef CONFIG_DEBUG_SLAB_LEAK
4540 proc_create("slab_allocators", 0, NULL
, &proc_slabstats_operations
);
4544 module_init(slab_proc_init
);
4548 * ksize - get the actual amount of memory allocated for a given object
4549 * @objp: Pointer to the object
4551 * kmalloc may internally round up allocations and return more memory
4552 * than requested. ksize() can be used to determine the actual amount of
4553 * memory allocated. The caller may use this additional memory, even though
4554 * a smaller amount of memory was initially specified with the kmalloc call.
4555 * The caller must guarantee that objp points to a valid object previously
4556 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4557 * must not be freed during the duration of the call.
4559 size_t ksize(const void *objp
)
4562 if (unlikely(objp
== ZERO_SIZE_PTR
))
4565 return obj_size(virt_to_cache(objp
));
4567 EXPORT_SYMBOL(ksize
);