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>
117 #include <linux/kmemcheck.h>
119 #include <asm/cacheflush.h>
120 #include <asm/tlbflush.h>
121 #include <asm/page.h>
124 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
125 * 0 for faster, smaller code (especially in the critical paths).
127 * STATS - 1 to collect stats for /proc/slabinfo.
128 * 0 for faster, smaller code (especially in the critical paths).
130 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
133 #ifdef CONFIG_DEBUG_SLAB
136 #define FORCED_DEBUG 1
140 #define FORCED_DEBUG 0
143 /* Shouldn't this be in a header file somewhere? */
144 #define BYTES_PER_WORD sizeof(void *)
145 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
147 #ifndef ARCH_KMALLOC_MINALIGN
149 * Enforce a minimum alignment for the kmalloc caches.
150 * Usually, the kmalloc caches are cache_line_size() aligned, except when
151 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
152 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
153 * alignment larger than the alignment of a 64-bit integer.
154 * ARCH_KMALLOC_MINALIGN allows that.
155 * Note that increasing this value may disable some debug features.
157 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
160 #ifndef ARCH_SLAB_MINALIGN
162 * Enforce a minimum alignment for all caches.
163 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
164 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
165 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
166 * some debug features.
168 #define ARCH_SLAB_MINALIGN 0
171 #ifndef ARCH_KMALLOC_FLAGS
172 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
175 /* Legal flag mask for kmem_cache_create(). */
177 # define CREATE_MASK (SLAB_RED_ZONE | \
178 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
181 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
182 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
183 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
185 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
187 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
188 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
189 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
195 * Bufctl's are used for linking objs within a slab
198 * This implementation relies on "struct page" for locating the cache &
199 * slab an object belongs to.
200 * This allows the bufctl structure to be small (one int), but limits
201 * the number of objects a slab (not a cache) can contain when off-slab
202 * bufctls are used. The limit is the size of the largest general cache
203 * that does not use off-slab slabs.
204 * For 32bit archs with 4 kB pages, is this 56.
205 * This is not serious, as it is only for large objects, when it is unwise
206 * to have too many per slab.
207 * Note: This limit can be raised by introducing a general cache whose size
208 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
211 typedef unsigned int kmem_bufctl_t
;
212 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
213 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
214 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
215 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
220 * Manages the objs in a slab. Placed either at the beginning of mem allocated
221 * for a slab, or allocated from an general cache.
222 * Slabs are chained into three list: fully used, partial, fully free slabs.
225 struct list_head list
;
226 unsigned long colouroff
;
227 void *s_mem
; /* including colour offset */
228 unsigned int inuse
; /* num of objs active in slab */
230 unsigned short nodeid
;
236 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
237 * arrange for kmem_freepages to be called via RCU. This is useful if
238 * we need to approach a kernel structure obliquely, from its address
239 * obtained without the usual locking. We can lock the structure to
240 * stabilize it and check it's still at the given address, only if we
241 * can be sure that the memory has not been meanwhile reused for some
242 * other kind of object (which our subsystem's lock might corrupt).
244 * rcu_read_lock before reading the address, then rcu_read_unlock after
245 * taking the spinlock within the structure expected at that address.
247 * We assume struct slab_rcu can overlay struct slab when destroying.
250 struct rcu_head head
;
251 struct kmem_cache
*cachep
;
259 * - LIFO ordering, to hand out cache-warm objects from _alloc
260 * - reduce the number of linked list operations
261 * - reduce spinlock operations
263 * The limit is stored in the per-cpu structure to reduce the data cache
270 unsigned int batchcount
;
271 unsigned int touched
;
274 * Must have this definition in here for the proper
275 * alignment of array_cache. Also simplifies accessing
281 * bootstrap: The caches do not work without cpuarrays anymore, but the
282 * cpuarrays are allocated from the generic caches...
284 #define BOOT_CPUCACHE_ENTRIES 1
285 struct arraycache_init
{
286 struct array_cache cache
;
287 void *entries
[BOOT_CPUCACHE_ENTRIES
];
291 * The slab lists for all objects.
294 struct list_head slabs_partial
; /* partial list first, better asm code */
295 struct list_head slabs_full
;
296 struct list_head slabs_free
;
297 unsigned long free_objects
;
298 unsigned int free_limit
;
299 unsigned int colour_next
; /* Per-node cache coloring */
300 spinlock_t list_lock
;
301 struct array_cache
*shared
; /* shared per node */
302 struct array_cache
**alien
; /* on other nodes */
303 unsigned long next_reap
; /* updated without locking */
304 int free_touched
; /* updated without locking */
308 * The slab allocator is initialized with interrupts disabled. Therefore, make
309 * sure early boot allocations don't accidentally enable interrupts.
311 static gfp_t slab_gfp_mask __read_mostly
= SLAB_GFP_BOOT_MASK
;
314 * Need this for bootstrapping a per node allocator.
316 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
317 struct kmem_list3 __initdata initkmem_list3
[NUM_INIT_LISTS
];
318 #define CACHE_CACHE 0
319 #define SIZE_AC MAX_NUMNODES
320 #define SIZE_L3 (2 * MAX_NUMNODES)
322 static int drain_freelist(struct kmem_cache
*cache
,
323 struct kmem_list3
*l3
, int tofree
);
324 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
326 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
);
327 static void cache_reap(struct work_struct
*unused
);
330 * This function must be completely optimized away if a constant is passed to
331 * it. Mostly the same as what is in linux/slab.h except it returns an index.
333 static __always_inline
int index_of(const size_t size
)
335 extern void __bad_size(void);
337 if (__builtin_constant_p(size
)) {
345 #include <linux/kmalloc_sizes.h>
353 static int slab_early_init
= 1;
355 #define INDEX_AC index_of(sizeof(struct arraycache_init))
356 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
358 static void kmem_list3_init(struct kmem_list3
*parent
)
360 INIT_LIST_HEAD(&parent
->slabs_full
);
361 INIT_LIST_HEAD(&parent
->slabs_partial
);
362 INIT_LIST_HEAD(&parent
->slabs_free
);
363 parent
->shared
= NULL
;
364 parent
->alien
= NULL
;
365 parent
->colour_next
= 0;
366 spin_lock_init(&parent
->list_lock
);
367 parent
->free_objects
= 0;
368 parent
->free_touched
= 0;
371 #define MAKE_LIST(cachep, listp, slab, nodeid) \
373 INIT_LIST_HEAD(listp); \
374 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
377 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
379 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
380 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
381 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
384 #define CFLGS_OFF_SLAB (0x80000000UL)
385 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
387 #define BATCHREFILL_LIMIT 16
389 * Optimization question: fewer reaps means less probability for unnessary
390 * cpucache drain/refill cycles.
392 * OTOH the cpuarrays can contain lots of objects,
393 * which could lock up otherwise freeable slabs.
395 #define REAPTIMEOUT_CPUC (2*HZ)
396 #define REAPTIMEOUT_LIST3 (4*HZ)
399 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
400 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
401 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
402 #define STATS_INC_GROWN(x) ((x)->grown++)
403 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
404 #define STATS_SET_HIGH(x) \
406 if ((x)->num_active > (x)->high_mark) \
407 (x)->high_mark = (x)->num_active; \
409 #define STATS_INC_ERR(x) ((x)->errors++)
410 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
411 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
412 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
413 #define STATS_SET_FREEABLE(x, i) \
415 if ((x)->max_freeable < i) \
416 (x)->max_freeable = i; \
418 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
419 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
420 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
421 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
423 #define STATS_INC_ACTIVE(x) do { } while (0)
424 #define STATS_DEC_ACTIVE(x) do { } while (0)
425 #define STATS_INC_ALLOCED(x) do { } while (0)
426 #define STATS_INC_GROWN(x) do { } while (0)
427 #define STATS_ADD_REAPED(x,y) do { } while (0)
428 #define STATS_SET_HIGH(x) do { } while (0)
429 #define STATS_INC_ERR(x) do { } while (0)
430 #define STATS_INC_NODEALLOCS(x) do { } while (0)
431 #define STATS_INC_NODEFREES(x) do { } while (0)
432 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
433 #define STATS_SET_FREEABLE(x, i) do { } while (0)
434 #define STATS_INC_ALLOCHIT(x) do { } while (0)
435 #define STATS_INC_ALLOCMISS(x) do { } while (0)
436 #define STATS_INC_FREEHIT(x) do { } while (0)
437 #define STATS_INC_FREEMISS(x) do { } while (0)
443 * memory layout of objects:
445 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
446 * the end of an object is aligned with the end of the real
447 * allocation. Catches writes behind the end of the allocation.
448 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
450 * cachep->obj_offset: The real object.
451 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
452 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
453 * [BYTES_PER_WORD long]
455 static int obj_offset(struct kmem_cache
*cachep
)
457 return cachep
->obj_offset
;
460 static int obj_size(struct kmem_cache
*cachep
)
462 return cachep
->obj_size
;
465 static unsigned long long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
467 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
468 return (unsigned long long*) (objp
+ obj_offset(cachep
) -
469 sizeof(unsigned long long));
472 static unsigned long long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
474 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
475 if (cachep
->flags
& SLAB_STORE_USER
)
476 return (unsigned long long *)(objp
+ cachep
->buffer_size
-
477 sizeof(unsigned long long) -
479 return (unsigned long long *) (objp
+ cachep
->buffer_size
-
480 sizeof(unsigned long long));
483 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
485 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
486 return (void **)(objp
+ cachep
->buffer_size
- BYTES_PER_WORD
);
491 #define obj_offset(x) 0
492 #define obj_size(cachep) (cachep->buffer_size)
493 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
494 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
495 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
499 #ifdef CONFIG_KMEMTRACE
500 size_t slab_buffer_size(struct kmem_cache
*cachep
)
502 return cachep
->buffer_size
;
504 EXPORT_SYMBOL(slab_buffer_size
);
508 * Do not go above this order unless 0 objects fit into the slab.
510 #define BREAK_GFP_ORDER_HI 1
511 #define BREAK_GFP_ORDER_LO 0
512 static int slab_break_gfp_order
= BREAK_GFP_ORDER_LO
;
515 * Functions for storing/retrieving the cachep and or slab from the page
516 * allocator. These are used to find the slab an obj belongs to. With kfree(),
517 * these are used to find the cache which an obj belongs to.
519 static inline void page_set_cache(struct page
*page
, struct kmem_cache
*cache
)
521 page
->lru
.next
= (struct list_head
*)cache
;
524 static inline struct kmem_cache
*page_get_cache(struct page
*page
)
526 page
= compound_head(page
);
527 BUG_ON(!PageSlab(page
));
528 return (struct kmem_cache
*)page
->lru
.next
;
531 static inline void page_set_slab(struct page
*page
, struct slab
*slab
)
533 page
->lru
.prev
= (struct list_head
*)slab
;
536 static inline struct slab
*page_get_slab(struct page
*page
)
538 BUG_ON(!PageSlab(page
));
539 return (struct slab
*)page
->lru
.prev
;
542 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
544 struct page
*page
= virt_to_head_page(obj
);
545 return page_get_cache(page
);
548 static inline struct slab
*virt_to_slab(const void *obj
)
550 struct page
*page
= virt_to_head_page(obj
);
551 return page_get_slab(page
);
554 static inline void *index_to_obj(struct kmem_cache
*cache
, struct slab
*slab
,
557 return slab
->s_mem
+ cache
->buffer_size
* idx
;
561 * We want to avoid an expensive divide : (offset / cache->buffer_size)
562 * Using the fact that buffer_size is a constant for a particular cache,
563 * we can replace (offset / cache->buffer_size) by
564 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
566 static inline unsigned int obj_to_index(const struct kmem_cache
*cache
,
567 const struct slab
*slab
, void *obj
)
569 u32 offset
= (obj
- slab
->s_mem
);
570 return reciprocal_divide(offset
, cache
->reciprocal_buffer_size
);
574 * These are the default caches for kmalloc. Custom caches can have other sizes.
576 struct cache_sizes malloc_sizes
[] = {
577 #define CACHE(x) { .cs_size = (x) },
578 #include <linux/kmalloc_sizes.h>
582 EXPORT_SYMBOL(malloc_sizes
);
584 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
590 static struct cache_names __initdata cache_names
[] = {
591 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
592 #include <linux/kmalloc_sizes.h>
597 static struct arraycache_init initarray_cache __initdata
=
598 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
599 static struct arraycache_init initarray_generic
=
600 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
602 /* internal cache of cache description objs */
603 static struct kmem_cache cache_cache
= {
605 .limit
= BOOT_CPUCACHE_ENTRIES
,
607 .buffer_size
= sizeof(struct kmem_cache
),
608 .name
= "kmem_cache",
611 #define BAD_ALIEN_MAGIC 0x01020304ul
613 #ifdef CONFIG_LOCKDEP
616 * Slab sometimes uses the kmalloc slabs to store the slab headers
617 * for other slabs "off slab".
618 * The locking for this is tricky in that it nests within the locks
619 * of all other slabs in a few places; to deal with this special
620 * locking we put on-slab caches into a separate lock-class.
622 * We set lock class for alien array caches which are up during init.
623 * The lock annotation will be lost if all cpus of a node goes down and
624 * then comes back up during hotplug
626 static struct lock_class_key on_slab_l3_key
;
627 static struct lock_class_key on_slab_alc_key
;
629 static inline void init_lock_keys(void)
633 struct cache_sizes
*s
= malloc_sizes
;
635 while (s
->cs_size
!= ULONG_MAX
) {
637 struct array_cache
**alc
;
639 struct kmem_list3
*l3
= s
->cs_cachep
->nodelists
[q
];
640 if (!l3
|| OFF_SLAB(s
->cs_cachep
))
642 lockdep_set_class(&l3
->list_lock
, &on_slab_l3_key
);
645 * FIXME: This check for BAD_ALIEN_MAGIC
646 * should go away when common slab code is taught to
647 * work even without alien caches.
648 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
649 * for alloc_alien_cache,
651 if (!alc
|| (unsigned long)alc
== BAD_ALIEN_MAGIC
)
655 lockdep_set_class(&alc
[r
]->lock
,
663 static inline void init_lock_keys(void)
669 * Guard access to the cache-chain.
671 static DEFINE_MUTEX(cache_chain_mutex
);
672 static struct list_head cache_chain
;
675 * chicken and egg problem: delay the per-cpu array allocation
676 * until the general caches are up.
687 * used by boot code to determine if it can use slab based allocator
689 int slab_is_available(void)
691 return g_cpucache_up
>= EARLY
;
694 static DEFINE_PER_CPU(struct delayed_work
, reap_work
);
696 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
698 return cachep
->array
[smp_processor_id()];
701 static inline struct kmem_cache
*__find_general_cachep(size_t size
,
704 struct cache_sizes
*csizep
= malloc_sizes
;
707 /* This happens if someone tries to call
708 * kmem_cache_create(), or __kmalloc(), before
709 * the generic caches are initialized.
711 BUG_ON(malloc_sizes
[INDEX_AC
].cs_cachep
== NULL
);
714 return ZERO_SIZE_PTR
;
716 while (size
> csizep
->cs_size
)
720 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
721 * has cs_{dma,}cachep==NULL. Thus no special case
722 * for large kmalloc calls required.
724 #ifdef CONFIG_ZONE_DMA
725 if (unlikely(gfpflags
& GFP_DMA
))
726 return csizep
->cs_dmacachep
;
728 return csizep
->cs_cachep
;
731 static struct kmem_cache
*kmem_find_general_cachep(size_t size
, gfp_t gfpflags
)
733 return __find_general_cachep(size
, gfpflags
);
736 static size_t slab_mgmt_size(size_t nr_objs
, size_t align
)
738 return ALIGN(sizeof(struct slab
)+nr_objs
*sizeof(kmem_bufctl_t
), align
);
742 * Calculate the number of objects and left-over bytes for a given buffer size.
744 static void cache_estimate(unsigned long gfporder
, size_t buffer_size
,
745 size_t align
, int flags
, size_t *left_over
,
750 size_t slab_size
= PAGE_SIZE
<< gfporder
;
753 * The slab management structure can be either off the slab or
754 * on it. For the latter case, the memory allocated for a
758 * - One kmem_bufctl_t for each object
759 * - Padding to respect alignment of @align
760 * - @buffer_size bytes for each object
762 * If the slab management structure is off the slab, then the
763 * alignment will already be calculated into the size. Because
764 * the slabs are all pages aligned, the objects will be at the
765 * correct alignment when allocated.
767 if (flags
& CFLGS_OFF_SLAB
) {
769 nr_objs
= slab_size
/ buffer_size
;
771 if (nr_objs
> SLAB_LIMIT
)
772 nr_objs
= SLAB_LIMIT
;
775 * Ignore padding for the initial guess. The padding
776 * is at most @align-1 bytes, and @buffer_size is at
777 * least @align. In the worst case, this result will
778 * be one greater than the number of objects that fit
779 * into the memory allocation when taking the padding
782 nr_objs
= (slab_size
- sizeof(struct slab
)) /
783 (buffer_size
+ sizeof(kmem_bufctl_t
));
786 * This calculated number will be either the right
787 * amount, or one greater than what we want.
789 if (slab_mgmt_size(nr_objs
, align
) + nr_objs
*buffer_size
793 if (nr_objs
> SLAB_LIMIT
)
794 nr_objs
= SLAB_LIMIT
;
796 mgmt_size
= slab_mgmt_size(nr_objs
, align
);
799 *left_over
= slab_size
- nr_objs
*buffer_size
- mgmt_size
;
802 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
804 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
807 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
808 function
, cachep
->name
, msg
);
813 * By default on NUMA we use alien caches to stage the freeing of
814 * objects allocated from other nodes. This causes massive memory
815 * inefficiencies when using fake NUMA setup to split memory into a
816 * large number of small nodes, so it can be disabled on the command
820 static int use_alien_caches __read_mostly
= 1;
821 static int numa_platform __read_mostly
= 1;
822 static int __init
noaliencache_setup(char *s
)
824 use_alien_caches
= 0;
827 __setup("noaliencache", noaliencache_setup
);
831 * Special reaping functions for NUMA systems called from cache_reap().
832 * These take care of doing round robin flushing of alien caches (containing
833 * objects freed on different nodes from which they were allocated) and the
834 * flushing of remote pcps by calling drain_node_pages.
836 static DEFINE_PER_CPU(unsigned long, reap_node
);
838 static void init_reap_node(int cpu
)
842 node
= next_node(cpu_to_node(cpu
), node_online_map
);
843 if (node
== MAX_NUMNODES
)
844 node
= first_node(node_online_map
);
846 per_cpu(reap_node
, cpu
) = node
;
849 static void next_reap_node(void)
851 int node
= __get_cpu_var(reap_node
);
853 node
= next_node(node
, node_online_map
);
854 if (unlikely(node
>= MAX_NUMNODES
))
855 node
= first_node(node_online_map
);
856 __get_cpu_var(reap_node
) = node
;
860 #define init_reap_node(cpu) do { } while (0)
861 #define next_reap_node(void) do { } while (0)
865 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
866 * via the workqueue/eventd.
867 * Add the CPU number into the expiration time to minimize the possibility of
868 * the CPUs getting into lockstep and contending for the global cache chain
871 static void __cpuinit
start_cpu_timer(int cpu
)
873 struct delayed_work
*reap_work
= &per_cpu(reap_work
, cpu
);
876 * When this gets called from do_initcalls via cpucache_init(),
877 * init_workqueues() has already run, so keventd will be setup
880 if (keventd_up() && reap_work
->work
.func
== NULL
) {
882 INIT_DELAYED_WORK(reap_work
, cache_reap
);
883 schedule_delayed_work_on(cpu
, reap_work
,
884 __round_jiffies_relative(HZ
, cpu
));
888 static struct array_cache
*alloc_arraycache(int node
, int entries
,
889 int batchcount
, gfp_t gfp
)
891 int memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
892 struct array_cache
*nc
= NULL
;
894 nc
= kmalloc_node(memsize
, gfp
, node
);
896 * The array_cache structures contain pointers to free object.
897 * However, when such objects are allocated or transfered to another
898 * cache the pointers are not cleared and they could be counted as
899 * valid references during a kmemleak scan. Therefore, kmemleak must
900 * not scan such objects.
902 kmemleak_no_scan(nc
);
906 nc
->batchcount
= batchcount
;
908 spin_lock_init(&nc
->lock
);
914 * Transfer objects in one arraycache to another.
915 * Locking must be handled by the caller.
917 * Return the number of entries transferred.
919 static int transfer_objects(struct array_cache
*to
,
920 struct array_cache
*from
, unsigned int max
)
922 /* Figure out how many entries to transfer */
923 int nr
= min(min(from
->avail
, max
), to
->limit
- to
->avail
);
928 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
939 #define drain_alien_cache(cachep, alien) do { } while (0)
940 #define reap_alien(cachep, l3) do { } while (0)
942 static inline struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
944 return (struct array_cache
**)BAD_ALIEN_MAGIC
;
947 static inline void free_alien_cache(struct array_cache
**ac_ptr
)
951 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
956 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
962 static inline void *____cache_alloc_node(struct kmem_cache
*cachep
,
963 gfp_t flags
, int nodeid
)
968 #else /* CONFIG_NUMA */
970 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
971 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
973 static struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
975 struct array_cache
**ac_ptr
;
976 int memsize
= sizeof(void *) * nr_node_ids
;
981 ac_ptr
= kmalloc_node(memsize
, gfp
, node
);
984 if (i
== node
|| !node_online(i
)) {
988 ac_ptr
[i
] = alloc_arraycache(node
, limit
, 0xbaadf00d, gfp
);
990 for (i
--; i
>= 0; i
--)
1000 static void free_alien_cache(struct array_cache
**ac_ptr
)
1011 static void __drain_alien_cache(struct kmem_cache
*cachep
,
1012 struct array_cache
*ac
, int node
)
1014 struct kmem_list3
*rl3
= cachep
->nodelists
[node
];
1017 spin_lock(&rl3
->list_lock
);
1019 * Stuff objects into the remote nodes shared array first.
1020 * That way we could avoid the overhead of putting the objects
1021 * into the free lists and getting them back later.
1024 transfer_objects(rl3
->shared
, ac
, ac
->limit
);
1026 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
1028 spin_unlock(&rl3
->list_lock
);
1033 * Called from cache_reap() to regularly drain alien caches round robin.
1035 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_list3
*l3
)
1037 int node
= __get_cpu_var(reap_node
);
1040 struct array_cache
*ac
= l3
->alien
[node
];
1042 if (ac
&& ac
->avail
&& spin_trylock_irq(&ac
->lock
)) {
1043 __drain_alien_cache(cachep
, ac
, node
);
1044 spin_unlock_irq(&ac
->lock
);
1049 static void drain_alien_cache(struct kmem_cache
*cachep
,
1050 struct array_cache
**alien
)
1053 struct array_cache
*ac
;
1054 unsigned long flags
;
1056 for_each_online_node(i
) {
1059 spin_lock_irqsave(&ac
->lock
, flags
);
1060 __drain_alien_cache(cachep
, ac
, i
);
1061 spin_unlock_irqrestore(&ac
->lock
, flags
);
1066 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1068 struct slab
*slabp
= virt_to_slab(objp
);
1069 int nodeid
= slabp
->nodeid
;
1070 struct kmem_list3
*l3
;
1071 struct array_cache
*alien
= NULL
;
1074 node
= numa_node_id();
1077 * Make sure we are not freeing a object from another node to the array
1078 * cache on this cpu.
1080 if (likely(slabp
->nodeid
== node
))
1083 l3
= cachep
->nodelists
[node
];
1084 STATS_INC_NODEFREES(cachep
);
1085 if (l3
->alien
&& l3
->alien
[nodeid
]) {
1086 alien
= l3
->alien
[nodeid
];
1087 spin_lock(&alien
->lock
);
1088 if (unlikely(alien
->avail
== alien
->limit
)) {
1089 STATS_INC_ACOVERFLOW(cachep
);
1090 __drain_alien_cache(cachep
, alien
, nodeid
);
1092 alien
->entry
[alien
->avail
++] = objp
;
1093 spin_unlock(&alien
->lock
);
1095 spin_lock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1096 free_block(cachep
, &objp
, 1, nodeid
);
1097 spin_unlock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1103 static void __cpuinit
cpuup_canceled(long cpu
)
1105 struct kmem_cache
*cachep
;
1106 struct kmem_list3
*l3
= NULL
;
1107 int node
= cpu_to_node(cpu
);
1108 const struct cpumask
*mask
= cpumask_of_node(node
);
1110 list_for_each_entry(cachep
, &cache_chain
, next
) {
1111 struct array_cache
*nc
;
1112 struct array_cache
*shared
;
1113 struct array_cache
**alien
;
1115 /* cpu is dead; no one can alloc from it. */
1116 nc
= cachep
->array
[cpu
];
1117 cachep
->array
[cpu
] = NULL
;
1118 l3
= cachep
->nodelists
[node
];
1121 goto free_array_cache
;
1123 spin_lock_irq(&l3
->list_lock
);
1125 /* Free limit for this kmem_list3 */
1126 l3
->free_limit
-= cachep
->batchcount
;
1128 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
1130 if (!cpus_empty(*mask
)) {
1131 spin_unlock_irq(&l3
->list_lock
);
1132 goto free_array_cache
;
1135 shared
= l3
->shared
;
1137 free_block(cachep
, shared
->entry
,
1138 shared
->avail
, node
);
1145 spin_unlock_irq(&l3
->list_lock
);
1149 drain_alien_cache(cachep
, alien
);
1150 free_alien_cache(alien
);
1156 * In the previous loop, all the objects were freed to
1157 * the respective cache's slabs, now we can go ahead and
1158 * shrink each nodelist to its limit.
1160 list_for_each_entry(cachep
, &cache_chain
, next
) {
1161 l3
= cachep
->nodelists
[node
];
1164 drain_freelist(cachep
, l3
, l3
->free_objects
);
1168 static int __cpuinit
cpuup_prepare(long cpu
)
1170 struct kmem_cache
*cachep
;
1171 struct kmem_list3
*l3
= NULL
;
1172 int node
= cpu_to_node(cpu
);
1173 const int memsize
= sizeof(struct kmem_list3
);
1176 * We need to do this right in the beginning since
1177 * alloc_arraycache's are going to use this list.
1178 * kmalloc_node allows us to add the slab to the right
1179 * kmem_list3 and not this cpu's kmem_list3
1182 list_for_each_entry(cachep
, &cache_chain
, next
) {
1184 * Set up the size64 kmemlist for cpu before we can
1185 * begin anything. Make sure some other cpu on this
1186 * node has not already allocated this
1188 if (!cachep
->nodelists
[node
]) {
1189 l3
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1192 kmem_list3_init(l3
);
1193 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
1194 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1197 * The l3s don't come and go as CPUs come and
1198 * go. cache_chain_mutex is sufficient
1201 cachep
->nodelists
[node
] = l3
;
1204 spin_lock_irq(&cachep
->nodelists
[node
]->list_lock
);
1205 cachep
->nodelists
[node
]->free_limit
=
1206 (1 + nr_cpus_node(node
)) *
1207 cachep
->batchcount
+ cachep
->num
;
1208 spin_unlock_irq(&cachep
->nodelists
[node
]->list_lock
);
1212 * Now we can go ahead with allocating the shared arrays and
1215 list_for_each_entry(cachep
, &cache_chain
, next
) {
1216 struct array_cache
*nc
;
1217 struct array_cache
*shared
= NULL
;
1218 struct array_cache
**alien
= NULL
;
1220 nc
= alloc_arraycache(node
, cachep
->limit
,
1221 cachep
->batchcount
, GFP_KERNEL
);
1224 if (cachep
->shared
) {
1225 shared
= alloc_arraycache(node
,
1226 cachep
->shared
* cachep
->batchcount
,
1227 0xbaadf00d, GFP_KERNEL
);
1233 if (use_alien_caches
) {
1234 alien
= alloc_alien_cache(node
, cachep
->limit
, GFP_KERNEL
);
1241 cachep
->array
[cpu
] = nc
;
1242 l3
= cachep
->nodelists
[node
];
1245 spin_lock_irq(&l3
->list_lock
);
1248 * We are serialised from CPU_DEAD or
1249 * CPU_UP_CANCELLED by the cpucontrol lock
1251 l3
->shared
= shared
;
1260 spin_unlock_irq(&l3
->list_lock
);
1262 free_alien_cache(alien
);
1266 cpuup_canceled(cpu
);
1270 static int __cpuinit
cpuup_callback(struct notifier_block
*nfb
,
1271 unsigned long action
, void *hcpu
)
1273 long cpu
= (long)hcpu
;
1277 case CPU_UP_PREPARE
:
1278 case CPU_UP_PREPARE_FROZEN
:
1279 mutex_lock(&cache_chain_mutex
);
1280 err
= cpuup_prepare(cpu
);
1281 mutex_unlock(&cache_chain_mutex
);
1284 case CPU_ONLINE_FROZEN
:
1285 start_cpu_timer(cpu
);
1287 #ifdef CONFIG_HOTPLUG_CPU
1288 case CPU_DOWN_PREPARE
:
1289 case CPU_DOWN_PREPARE_FROZEN
:
1291 * Shutdown cache reaper. Note that the cache_chain_mutex is
1292 * held so that if cache_reap() is invoked it cannot do
1293 * anything expensive but will only modify reap_work
1294 * and reschedule the timer.
1296 cancel_rearming_delayed_work(&per_cpu(reap_work
, cpu
));
1297 /* Now the cache_reaper is guaranteed to be not running. */
1298 per_cpu(reap_work
, cpu
).work
.func
= NULL
;
1300 case CPU_DOWN_FAILED
:
1301 case CPU_DOWN_FAILED_FROZEN
:
1302 start_cpu_timer(cpu
);
1305 case CPU_DEAD_FROZEN
:
1307 * Even if all the cpus of a node are down, we don't free the
1308 * kmem_list3 of any cache. This to avoid a race between
1309 * cpu_down, and a kmalloc allocation from another cpu for
1310 * memory from the node of the cpu going down. The list3
1311 * structure is usually allocated from kmem_cache_create() and
1312 * gets destroyed at kmem_cache_destroy().
1316 case CPU_UP_CANCELED
:
1317 case CPU_UP_CANCELED_FROZEN
:
1318 mutex_lock(&cache_chain_mutex
);
1319 cpuup_canceled(cpu
);
1320 mutex_unlock(&cache_chain_mutex
);
1323 return err
? NOTIFY_BAD
: NOTIFY_OK
;
1326 static struct notifier_block __cpuinitdata cpucache_notifier
= {
1327 &cpuup_callback
, NULL
, 0
1331 * swap the static kmem_list3 with kmalloced memory
1333 static void init_list(struct kmem_cache
*cachep
, struct kmem_list3
*list
,
1336 struct kmem_list3
*ptr
;
1338 ptr
= kmalloc_node(sizeof(struct kmem_list3
), GFP_NOWAIT
, nodeid
);
1341 memcpy(ptr
, list
, sizeof(struct kmem_list3
));
1343 * Do not assume that spinlocks can be initialized via memcpy:
1345 spin_lock_init(&ptr
->list_lock
);
1347 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1348 cachep
->nodelists
[nodeid
] = ptr
;
1352 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1353 * size of kmem_list3.
1355 static void __init
set_up_list3s(struct kmem_cache
*cachep
, int index
)
1359 for_each_online_node(node
) {
1360 cachep
->nodelists
[node
] = &initkmem_list3
[index
+ node
];
1361 cachep
->nodelists
[node
]->next_reap
= jiffies
+
1363 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1368 * Initialisation. Called after the page allocator have been initialised and
1369 * before smp_init().
1371 void __init
kmem_cache_init(void)
1374 struct cache_sizes
*sizes
;
1375 struct cache_names
*names
;
1380 if (num_possible_nodes() == 1) {
1381 use_alien_caches
= 0;
1385 for (i
= 0; i
< NUM_INIT_LISTS
; i
++) {
1386 kmem_list3_init(&initkmem_list3
[i
]);
1387 if (i
< MAX_NUMNODES
)
1388 cache_cache
.nodelists
[i
] = NULL
;
1390 set_up_list3s(&cache_cache
, CACHE_CACHE
);
1393 * Fragmentation resistance on low memory - only use bigger
1394 * page orders on machines with more than 32MB of memory.
1396 if (num_physpages
> (32 << 20) >> PAGE_SHIFT
)
1397 slab_break_gfp_order
= BREAK_GFP_ORDER_HI
;
1399 /* Bootstrap is tricky, because several objects are allocated
1400 * from caches that do not exist yet:
1401 * 1) initialize the cache_cache cache: it contains the struct
1402 * kmem_cache structures of all caches, except cache_cache itself:
1403 * cache_cache is statically allocated.
1404 * Initially an __init data area is used for the head array and the
1405 * kmem_list3 structures, it's replaced with a kmalloc allocated
1406 * array at the end of the bootstrap.
1407 * 2) Create the first kmalloc cache.
1408 * The struct kmem_cache for the new cache is allocated normally.
1409 * An __init data area is used for the head array.
1410 * 3) Create the remaining kmalloc caches, with minimally sized
1412 * 4) Replace the __init data head arrays for cache_cache and the first
1413 * kmalloc cache with kmalloc allocated arrays.
1414 * 5) Replace the __init data for kmem_list3 for cache_cache and
1415 * the other cache's with kmalloc allocated memory.
1416 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1419 node
= numa_node_id();
1421 /* 1) create the cache_cache */
1422 INIT_LIST_HEAD(&cache_chain
);
1423 list_add(&cache_cache
.next
, &cache_chain
);
1424 cache_cache
.colour_off
= cache_line_size();
1425 cache_cache
.array
[smp_processor_id()] = &initarray_cache
.cache
;
1426 cache_cache
.nodelists
[node
] = &initkmem_list3
[CACHE_CACHE
+ node
];
1429 * struct kmem_cache size depends on nr_node_ids, which
1430 * can be less than MAX_NUMNODES.
1432 cache_cache
.buffer_size
= offsetof(struct kmem_cache
, nodelists
) +
1433 nr_node_ids
* sizeof(struct kmem_list3
*);
1435 cache_cache
.obj_size
= cache_cache
.buffer_size
;
1437 cache_cache
.buffer_size
= ALIGN(cache_cache
.buffer_size
,
1439 cache_cache
.reciprocal_buffer_size
=
1440 reciprocal_value(cache_cache
.buffer_size
);
1442 for (order
= 0; order
< MAX_ORDER
; order
++) {
1443 cache_estimate(order
, cache_cache
.buffer_size
,
1444 cache_line_size(), 0, &left_over
, &cache_cache
.num
);
1445 if (cache_cache
.num
)
1448 BUG_ON(!cache_cache
.num
);
1449 cache_cache
.gfporder
= order
;
1450 cache_cache
.colour
= left_over
/ cache_cache
.colour_off
;
1451 cache_cache
.slab_size
= ALIGN(cache_cache
.num
* sizeof(kmem_bufctl_t
) +
1452 sizeof(struct slab
), cache_line_size());
1454 /* 2+3) create the kmalloc caches */
1455 sizes
= malloc_sizes
;
1456 names
= cache_names
;
1459 * Initialize the caches that provide memory for the array cache and the
1460 * kmem_list3 structures first. Without this, further allocations will
1464 sizes
[INDEX_AC
].cs_cachep
= kmem_cache_create(names
[INDEX_AC
].name
,
1465 sizes
[INDEX_AC
].cs_size
,
1466 ARCH_KMALLOC_MINALIGN
,
1467 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1470 if (INDEX_AC
!= INDEX_L3
) {
1471 sizes
[INDEX_L3
].cs_cachep
=
1472 kmem_cache_create(names
[INDEX_L3
].name
,
1473 sizes
[INDEX_L3
].cs_size
,
1474 ARCH_KMALLOC_MINALIGN
,
1475 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1479 slab_early_init
= 0;
1481 while (sizes
->cs_size
!= ULONG_MAX
) {
1483 * For performance, all the general caches are L1 aligned.
1484 * This should be particularly beneficial on SMP boxes, as it
1485 * eliminates "false sharing".
1486 * Note for systems short on memory removing the alignment will
1487 * allow tighter packing of the smaller caches.
1489 if (!sizes
->cs_cachep
) {
1490 sizes
->cs_cachep
= kmem_cache_create(names
->name
,
1492 ARCH_KMALLOC_MINALIGN
,
1493 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1496 #ifdef CONFIG_ZONE_DMA
1497 sizes
->cs_dmacachep
= kmem_cache_create(
1500 ARCH_KMALLOC_MINALIGN
,
1501 ARCH_KMALLOC_FLAGS
|SLAB_CACHE_DMA
|
1508 /* 4) Replace the bootstrap head arrays */
1510 struct array_cache
*ptr
;
1512 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1514 BUG_ON(cpu_cache_get(&cache_cache
) != &initarray_cache
.cache
);
1515 memcpy(ptr
, cpu_cache_get(&cache_cache
),
1516 sizeof(struct arraycache_init
));
1518 * Do not assume that spinlocks can be initialized via memcpy:
1520 spin_lock_init(&ptr
->lock
);
1522 cache_cache
.array
[smp_processor_id()] = ptr
;
1524 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1526 BUG_ON(cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
)
1527 != &initarray_generic
.cache
);
1528 memcpy(ptr
, cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
),
1529 sizeof(struct arraycache_init
));
1531 * Do not assume that spinlocks can be initialized via memcpy:
1533 spin_lock_init(&ptr
->lock
);
1535 malloc_sizes
[INDEX_AC
].cs_cachep
->array
[smp_processor_id()] =
1538 /* 5) Replace the bootstrap kmem_list3's */
1542 for_each_online_node(nid
) {
1543 init_list(&cache_cache
, &initkmem_list3
[CACHE_CACHE
+ nid
], nid
);
1545 init_list(malloc_sizes
[INDEX_AC
].cs_cachep
,
1546 &initkmem_list3
[SIZE_AC
+ nid
], nid
);
1548 if (INDEX_AC
!= INDEX_L3
) {
1549 init_list(malloc_sizes
[INDEX_L3
].cs_cachep
,
1550 &initkmem_list3
[SIZE_L3
+ nid
], nid
);
1555 g_cpucache_up
= EARLY
;
1557 /* Annotate slab for lockdep -- annotate the malloc caches */
1561 void __init
kmem_cache_init_late(void)
1563 struct kmem_cache
*cachep
;
1566 * Interrupts are enabled now so all GFP allocations are safe.
1568 slab_gfp_mask
= __GFP_BITS_MASK
;
1570 /* 6) resize the head arrays to their final sizes */
1571 mutex_lock(&cache_chain_mutex
);
1572 list_for_each_entry(cachep
, &cache_chain
, next
)
1573 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1575 mutex_unlock(&cache_chain_mutex
);
1578 g_cpucache_up
= FULL
;
1581 * Register a cpu startup notifier callback that initializes
1582 * cpu_cache_get for all new cpus
1584 register_cpu_notifier(&cpucache_notifier
);
1587 * The reap timers are started later, with a module init call: That part
1588 * of the kernel is not yet operational.
1592 static int __init
cpucache_init(void)
1597 * Register the timers that return unneeded pages to the page allocator
1599 for_each_online_cpu(cpu
)
1600 start_cpu_timer(cpu
);
1603 __initcall(cpucache_init
);
1606 * Interface to system's page allocator. No need to hold the cache-lock.
1608 * If we requested dmaable memory, we will get it. Even if we
1609 * did not request dmaable memory, we might get it, but that
1610 * would be relatively rare and ignorable.
1612 static void *kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
1620 * Nommu uses slab's for process anonymous memory allocations, and thus
1621 * requires __GFP_COMP to properly refcount higher order allocations
1623 flags
|= __GFP_COMP
;
1626 flags
|= cachep
->gfpflags
;
1627 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1628 flags
|= __GFP_RECLAIMABLE
;
1630 page
= alloc_pages_node(nodeid
, flags
| __GFP_NOTRACK
, cachep
->gfporder
);
1634 nr_pages
= (1 << cachep
->gfporder
);
1635 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1636 add_zone_page_state(page_zone(page
),
1637 NR_SLAB_RECLAIMABLE
, nr_pages
);
1639 add_zone_page_state(page_zone(page
),
1640 NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1641 for (i
= 0; i
< nr_pages
; i
++)
1642 __SetPageSlab(page
+ i
);
1644 if (kmemcheck_enabled
&& !(cachep
->flags
& SLAB_NOTRACK
)) {
1645 kmemcheck_alloc_shadow(page
, cachep
->gfporder
, flags
, nodeid
);
1648 kmemcheck_mark_uninitialized_pages(page
, nr_pages
);
1650 kmemcheck_mark_unallocated_pages(page
, nr_pages
);
1653 return page_address(page
);
1657 * Interface to system's page release.
1659 static void kmem_freepages(struct kmem_cache
*cachep
, void *addr
)
1661 unsigned long i
= (1 << cachep
->gfporder
);
1662 struct page
*page
= virt_to_page(addr
);
1663 const unsigned long nr_freed
= i
;
1665 kmemcheck_free_shadow(page
, cachep
->gfporder
);
1667 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1668 sub_zone_page_state(page_zone(page
),
1669 NR_SLAB_RECLAIMABLE
, nr_freed
);
1671 sub_zone_page_state(page_zone(page
),
1672 NR_SLAB_UNRECLAIMABLE
, nr_freed
);
1674 BUG_ON(!PageSlab(page
));
1675 __ClearPageSlab(page
);
1678 if (current
->reclaim_state
)
1679 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1680 free_pages((unsigned long)addr
, cachep
->gfporder
);
1683 static void kmem_rcu_free(struct rcu_head
*head
)
1685 struct slab_rcu
*slab_rcu
= (struct slab_rcu
*)head
;
1686 struct kmem_cache
*cachep
= slab_rcu
->cachep
;
1688 kmem_freepages(cachep
, slab_rcu
->addr
);
1689 if (OFF_SLAB(cachep
))
1690 kmem_cache_free(cachep
->slabp_cache
, slab_rcu
);
1695 #ifdef CONFIG_DEBUG_PAGEALLOC
1696 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1697 unsigned long caller
)
1699 int size
= obj_size(cachep
);
1701 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1703 if (size
< 5 * sizeof(unsigned long))
1706 *addr
++ = 0x12345678;
1708 *addr
++ = smp_processor_id();
1709 size
-= 3 * sizeof(unsigned long);
1711 unsigned long *sptr
= &caller
;
1712 unsigned long svalue
;
1714 while (!kstack_end(sptr
)) {
1716 if (kernel_text_address(svalue
)) {
1718 size
-= sizeof(unsigned long);
1719 if (size
<= sizeof(unsigned long))
1725 *addr
++ = 0x87654321;
1729 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1731 int size
= obj_size(cachep
);
1732 addr
= &((char *)addr
)[obj_offset(cachep
)];
1734 memset(addr
, val
, size
);
1735 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1738 static void dump_line(char *data
, int offset
, int limit
)
1741 unsigned char error
= 0;
1744 printk(KERN_ERR
"%03x:", offset
);
1745 for (i
= 0; i
< limit
; i
++) {
1746 if (data
[offset
+ i
] != POISON_FREE
) {
1747 error
= data
[offset
+ i
];
1750 printk(" %02x", (unsigned char)data
[offset
+ i
]);
1754 if (bad_count
== 1) {
1755 error
^= POISON_FREE
;
1756 if (!(error
& (error
- 1))) {
1757 printk(KERN_ERR
"Single bit error detected. Probably "
1760 printk(KERN_ERR
"Run memtest86+ or a similar memory "
1763 printk(KERN_ERR
"Run a memory test tool.\n");
1772 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1777 if (cachep
->flags
& SLAB_RED_ZONE
) {
1778 printk(KERN_ERR
"Redzone: 0x%llx/0x%llx.\n",
1779 *dbg_redzone1(cachep
, objp
),
1780 *dbg_redzone2(cachep
, objp
));
1783 if (cachep
->flags
& SLAB_STORE_USER
) {
1784 printk(KERN_ERR
"Last user: [<%p>]",
1785 *dbg_userword(cachep
, objp
));
1786 print_symbol("(%s)",
1787 (unsigned long)*dbg_userword(cachep
, objp
));
1790 realobj
= (char *)objp
+ obj_offset(cachep
);
1791 size
= obj_size(cachep
);
1792 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1795 if (i
+ limit
> size
)
1797 dump_line(realobj
, i
, limit
);
1801 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1807 realobj
= (char *)objp
+ obj_offset(cachep
);
1808 size
= obj_size(cachep
);
1810 for (i
= 0; i
< size
; i
++) {
1811 char exp
= POISON_FREE
;
1814 if (realobj
[i
] != exp
) {
1820 "Slab corruption: %s start=%p, len=%d\n",
1821 cachep
->name
, realobj
, size
);
1822 print_objinfo(cachep
, objp
, 0);
1824 /* Hexdump the affected line */
1827 if (i
+ limit
> size
)
1829 dump_line(realobj
, i
, limit
);
1832 /* Limit to 5 lines */
1838 /* Print some data about the neighboring objects, if they
1841 struct slab
*slabp
= virt_to_slab(objp
);
1844 objnr
= obj_to_index(cachep
, slabp
, objp
);
1846 objp
= index_to_obj(cachep
, slabp
, objnr
- 1);
1847 realobj
= (char *)objp
+ obj_offset(cachep
);
1848 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
1850 print_objinfo(cachep
, objp
, 2);
1852 if (objnr
+ 1 < cachep
->num
) {
1853 objp
= index_to_obj(cachep
, slabp
, objnr
+ 1);
1854 realobj
= (char *)objp
+ obj_offset(cachep
);
1855 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
1857 print_objinfo(cachep
, objp
, 2);
1864 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
, struct slab
*slabp
)
1867 for (i
= 0; i
< cachep
->num
; i
++) {
1868 void *objp
= index_to_obj(cachep
, slabp
, i
);
1870 if (cachep
->flags
& SLAB_POISON
) {
1871 #ifdef CONFIG_DEBUG_PAGEALLOC
1872 if (cachep
->buffer_size
% PAGE_SIZE
== 0 &&
1874 kernel_map_pages(virt_to_page(objp
),
1875 cachep
->buffer_size
/ PAGE_SIZE
, 1);
1877 check_poison_obj(cachep
, objp
);
1879 check_poison_obj(cachep
, objp
);
1882 if (cachep
->flags
& SLAB_RED_ZONE
) {
1883 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1884 slab_error(cachep
, "start of a freed object "
1886 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1887 slab_error(cachep
, "end of a freed object "
1893 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
, struct slab
*slabp
)
1899 * slab_destroy - destroy and release all objects in a slab
1900 * @cachep: cache pointer being destroyed
1901 * @slabp: slab pointer being destroyed
1903 * Destroy all the objs in a slab, and release the mem back to the system.
1904 * Before calling the slab must have been unlinked from the cache. The
1905 * cache-lock is not held/needed.
1907 static void slab_destroy(struct kmem_cache
*cachep
, struct slab
*slabp
)
1909 void *addr
= slabp
->s_mem
- slabp
->colouroff
;
1911 slab_destroy_debugcheck(cachep
, slabp
);
1912 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
1913 struct slab_rcu
*slab_rcu
;
1915 slab_rcu
= (struct slab_rcu
*)slabp
;
1916 slab_rcu
->cachep
= cachep
;
1917 slab_rcu
->addr
= addr
;
1918 call_rcu(&slab_rcu
->head
, kmem_rcu_free
);
1920 kmem_freepages(cachep
, addr
);
1921 if (OFF_SLAB(cachep
))
1922 kmem_cache_free(cachep
->slabp_cache
, slabp
);
1926 static void __kmem_cache_destroy(struct kmem_cache
*cachep
)
1929 struct kmem_list3
*l3
;
1931 for_each_online_cpu(i
)
1932 kfree(cachep
->array
[i
]);
1934 /* NUMA: free the list3 structures */
1935 for_each_online_node(i
) {
1936 l3
= cachep
->nodelists
[i
];
1939 free_alien_cache(l3
->alien
);
1943 kmem_cache_free(&cache_cache
, cachep
);
1948 * calculate_slab_order - calculate size (page order) of slabs
1949 * @cachep: pointer to the cache that is being created
1950 * @size: size of objects to be created in this cache.
1951 * @align: required alignment for the objects.
1952 * @flags: slab allocation flags
1954 * Also calculates the number of objects per slab.
1956 * This could be made much more intelligent. For now, try to avoid using
1957 * high order pages for slabs. When the gfp() functions are more friendly
1958 * towards high-order requests, this should be changed.
1960 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
1961 size_t size
, size_t align
, unsigned long flags
)
1963 unsigned long offslab_limit
;
1964 size_t left_over
= 0;
1967 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
1971 cache_estimate(gfporder
, size
, align
, flags
, &remainder
, &num
);
1975 if (flags
& CFLGS_OFF_SLAB
) {
1977 * Max number of objs-per-slab for caches which
1978 * use off-slab slabs. Needed to avoid a possible
1979 * looping condition in cache_grow().
1981 offslab_limit
= size
- sizeof(struct slab
);
1982 offslab_limit
/= sizeof(kmem_bufctl_t
);
1984 if (num
> offslab_limit
)
1988 /* Found something acceptable - save it away */
1990 cachep
->gfporder
= gfporder
;
1991 left_over
= remainder
;
1994 * A VFS-reclaimable slab tends to have most allocations
1995 * as GFP_NOFS and we really don't want to have to be allocating
1996 * higher-order pages when we are unable to shrink dcache.
1998 if (flags
& SLAB_RECLAIM_ACCOUNT
)
2002 * Large number of objects is good, but very large slabs are
2003 * currently bad for the gfp()s.
2005 if (gfporder
>= slab_break_gfp_order
)
2009 * Acceptable internal fragmentation?
2011 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
2017 static int __init_refok
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
2019 if (g_cpucache_up
== FULL
)
2020 return enable_cpucache(cachep
, gfp
);
2022 if (g_cpucache_up
== NONE
) {
2024 * Note: the first kmem_cache_create must create the cache
2025 * that's used by kmalloc(24), otherwise the creation of
2026 * further caches will BUG().
2028 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
2031 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2032 * the first cache, then we need to set up all its list3s,
2033 * otherwise the creation of further caches will BUG().
2035 set_up_list3s(cachep
, SIZE_AC
);
2036 if (INDEX_AC
== INDEX_L3
)
2037 g_cpucache_up
= PARTIAL_L3
;
2039 g_cpucache_up
= PARTIAL_AC
;
2041 cachep
->array
[smp_processor_id()] =
2042 kmalloc(sizeof(struct arraycache_init
), gfp
);
2044 if (g_cpucache_up
== PARTIAL_AC
) {
2045 set_up_list3s(cachep
, SIZE_L3
);
2046 g_cpucache_up
= PARTIAL_L3
;
2049 for_each_online_node(node
) {
2050 cachep
->nodelists
[node
] =
2051 kmalloc_node(sizeof(struct kmem_list3
),
2053 BUG_ON(!cachep
->nodelists
[node
]);
2054 kmem_list3_init(cachep
->nodelists
[node
]);
2058 cachep
->nodelists
[numa_node_id()]->next_reap
=
2059 jiffies
+ REAPTIMEOUT_LIST3
+
2060 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
2062 cpu_cache_get(cachep
)->avail
= 0;
2063 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
2064 cpu_cache_get(cachep
)->batchcount
= 1;
2065 cpu_cache_get(cachep
)->touched
= 0;
2066 cachep
->batchcount
= 1;
2067 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
2072 * kmem_cache_create - Create a cache.
2073 * @name: A string which is used in /proc/slabinfo to identify this cache.
2074 * @size: The size of objects to be created in this cache.
2075 * @align: The required alignment for the objects.
2076 * @flags: SLAB flags
2077 * @ctor: A constructor for the objects.
2079 * Returns a ptr to the cache on success, NULL on failure.
2080 * Cannot be called within a int, but can be interrupted.
2081 * The @ctor is run when new pages are allocated by the cache.
2083 * @name must be valid until the cache is destroyed. This implies that
2084 * the module calling this has to destroy the cache before getting unloaded.
2085 * Note that kmem_cache_name() is not guaranteed to return the same pointer,
2086 * therefore applications must manage it themselves.
2090 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2091 * to catch references to uninitialised memory.
2093 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2094 * for buffer overruns.
2096 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2097 * cacheline. This can be beneficial if you're counting cycles as closely
2101 kmem_cache_create (const char *name
, size_t size
, size_t align
,
2102 unsigned long flags
, void (*ctor
)(void *))
2104 size_t left_over
, slab_size
, ralign
;
2105 struct kmem_cache
*cachep
= NULL
, *pc
;
2109 * Sanity checks... these are all serious usage bugs.
2111 if (!name
|| in_interrupt() || (size
< BYTES_PER_WORD
) ||
2112 size
> KMALLOC_MAX_SIZE
) {
2113 printk(KERN_ERR
"%s: Early error in slab %s\n", __func__
,
2119 * We use cache_chain_mutex to ensure a consistent view of
2120 * cpu_online_mask as well. Please see cpuup_callback
2122 if (slab_is_available()) {
2124 mutex_lock(&cache_chain_mutex
);
2127 list_for_each_entry(pc
, &cache_chain
, next
) {
2132 * This happens when the module gets unloaded and doesn't
2133 * destroy its slab cache and no-one else reuses the vmalloc
2134 * area of the module. Print a warning.
2136 res
= probe_kernel_address(pc
->name
, tmp
);
2139 "SLAB: cache with size %d has lost its name\n",
2144 if (!strcmp(pc
->name
, name
)) {
2146 "kmem_cache_create: duplicate cache %s\n", name
);
2153 WARN_ON(strchr(name
, ' ')); /* It confuses parsers */
2156 * Enable redzoning and last user accounting, except for caches with
2157 * large objects, if the increased size would increase the object size
2158 * above the next power of two: caches with object sizes just above a
2159 * power of two have a significant amount of internal fragmentation.
2161 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
2162 2 * sizeof(unsigned long long)))
2163 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2164 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2165 flags
|= SLAB_POISON
;
2167 if (flags
& SLAB_DESTROY_BY_RCU
)
2168 BUG_ON(flags
& SLAB_POISON
);
2171 * Always checks flags, a caller might be expecting debug support which
2174 BUG_ON(flags
& ~CREATE_MASK
);
2177 * Check that size is in terms of words. This is needed to avoid
2178 * unaligned accesses for some archs when redzoning is used, and makes
2179 * sure any on-slab bufctl's are also correctly aligned.
2181 if (size
& (BYTES_PER_WORD
- 1)) {
2182 size
+= (BYTES_PER_WORD
- 1);
2183 size
&= ~(BYTES_PER_WORD
- 1);
2186 /* calculate the final buffer alignment: */
2188 /* 1) arch recommendation: can be overridden for debug */
2189 if (flags
& SLAB_HWCACHE_ALIGN
) {
2191 * Default alignment: as specified by the arch code. Except if
2192 * an object is really small, then squeeze multiple objects into
2195 ralign
= cache_line_size();
2196 while (size
<= ralign
/ 2)
2199 ralign
= BYTES_PER_WORD
;
2203 * Redzoning and user store require word alignment or possibly larger.
2204 * Note this will be overridden by architecture or caller mandated
2205 * alignment if either is greater than BYTES_PER_WORD.
2207 if (flags
& SLAB_STORE_USER
)
2208 ralign
= BYTES_PER_WORD
;
2210 if (flags
& SLAB_RED_ZONE
) {
2211 ralign
= REDZONE_ALIGN
;
2212 /* If redzoning, ensure that the second redzone is suitably
2213 * aligned, by adjusting the object size accordingly. */
2214 size
+= REDZONE_ALIGN
- 1;
2215 size
&= ~(REDZONE_ALIGN
- 1);
2218 /* 2) arch mandated alignment */
2219 if (ralign
< ARCH_SLAB_MINALIGN
) {
2220 ralign
= ARCH_SLAB_MINALIGN
;
2222 /* 3) caller mandated alignment */
2223 if (ralign
< align
) {
2226 /* disable debug if necessary */
2227 if (ralign
> __alignof__(unsigned long long))
2228 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2234 if (slab_is_available())
2239 /* Get cache's description obj. */
2240 cachep
= kmem_cache_zalloc(&cache_cache
, gfp
);
2245 cachep
->obj_size
= size
;
2248 * Both debugging options require word-alignment which is calculated
2251 if (flags
& SLAB_RED_ZONE
) {
2252 /* add space for red zone words */
2253 cachep
->obj_offset
+= sizeof(unsigned long long);
2254 size
+= 2 * sizeof(unsigned long long);
2256 if (flags
& SLAB_STORE_USER
) {
2257 /* user store requires one word storage behind the end of
2258 * the real object. But if the second red zone needs to be
2259 * aligned to 64 bits, we must allow that much space.
2261 if (flags
& SLAB_RED_ZONE
)
2262 size
+= REDZONE_ALIGN
;
2264 size
+= BYTES_PER_WORD
;
2266 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2267 if (size
>= malloc_sizes
[INDEX_L3
+ 1].cs_size
2268 && cachep
->obj_size
> cache_line_size() && size
< PAGE_SIZE
) {
2269 cachep
->obj_offset
+= PAGE_SIZE
- size
;
2276 * Determine if the slab management is 'on' or 'off' slab.
2277 * (bootstrapping cannot cope with offslab caches so don't do
2280 if ((size
>= (PAGE_SIZE
>> 3)) && !slab_early_init
)
2282 * Size is large, assume best to place the slab management obj
2283 * off-slab (should allow better packing of objs).
2285 flags
|= CFLGS_OFF_SLAB
;
2287 size
= ALIGN(size
, align
);
2289 left_over
= calculate_slab_order(cachep
, size
, align
, flags
);
2293 "kmem_cache_create: couldn't create cache %s.\n", name
);
2294 kmem_cache_free(&cache_cache
, cachep
);
2298 slab_size
= ALIGN(cachep
->num
* sizeof(kmem_bufctl_t
)
2299 + sizeof(struct slab
), align
);
2302 * If the slab has been placed off-slab, and we have enough space then
2303 * move it on-slab. This is at the expense of any extra colouring.
2305 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= slab_size
) {
2306 flags
&= ~CFLGS_OFF_SLAB
;
2307 left_over
-= slab_size
;
2310 if (flags
& CFLGS_OFF_SLAB
) {
2311 /* really off slab. No need for manual alignment */
2313 cachep
->num
* sizeof(kmem_bufctl_t
) + sizeof(struct slab
);
2316 cachep
->colour_off
= cache_line_size();
2317 /* Offset must be a multiple of the alignment. */
2318 if (cachep
->colour_off
< align
)
2319 cachep
->colour_off
= align
;
2320 cachep
->colour
= left_over
/ cachep
->colour_off
;
2321 cachep
->slab_size
= slab_size
;
2322 cachep
->flags
= flags
;
2323 cachep
->gfpflags
= 0;
2324 if (CONFIG_ZONE_DMA_FLAG
&& (flags
& SLAB_CACHE_DMA
))
2325 cachep
->gfpflags
|= GFP_DMA
;
2326 cachep
->buffer_size
= size
;
2327 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2329 if (flags
& CFLGS_OFF_SLAB
) {
2330 cachep
->slabp_cache
= kmem_find_general_cachep(slab_size
, 0u);
2332 * This is a possibility for one of the malloc_sizes caches.
2333 * But since we go off slab only for object size greater than
2334 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2335 * this should not happen at all.
2336 * But leave a BUG_ON for some lucky dude.
2338 BUG_ON(ZERO_OR_NULL_PTR(cachep
->slabp_cache
));
2340 cachep
->ctor
= ctor
;
2341 cachep
->name
= name
;
2343 if (setup_cpu_cache(cachep
, gfp
)) {
2344 __kmem_cache_destroy(cachep
);
2349 /* cache setup completed, link it into the list */
2350 list_add(&cachep
->next
, &cache_chain
);
2352 if (!cachep
&& (flags
& SLAB_PANIC
))
2353 panic("kmem_cache_create(): failed to create slab `%s'\n",
2355 if (slab_is_available()) {
2356 mutex_unlock(&cache_chain_mutex
);
2361 EXPORT_SYMBOL(kmem_cache_create
);
2364 static void check_irq_off(void)
2366 BUG_ON(!irqs_disabled());
2369 static void check_irq_on(void)
2371 BUG_ON(irqs_disabled());
2374 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2378 assert_spin_locked(&cachep
->nodelists
[numa_node_id()]->list_lock
);
2382 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2386 assert_spin_locked(&cachep
->nodelists
[node
]->list_lock
);
2391 #define check_irq_off() do { } while(0)
2392 #define check_irq_on() do { } while(0)
2393 #define check_spinlock_acquired(x) do { } while(0)
2394 #define check_spinlock_acquired_node(x, y) do { } while(0)
2397 static void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
2398 struct array_cache
*ac
,
2399 int force
, int node
);
2401 static void do_drain(void *arg
)
2403 struct kmem_cache
*cachep
= arg
;
2404 struct array_cache
*ac
;
2405 int node
= numa_node_id();
2408 ac
= cpu_cache_get(cachep
);
2409 spin_lock(&cachep
->nodelists
[node
]->list_lock
);
2410 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
2411 spin_unlock(&cachep
->nodelists
[node
]->list_lock
);
2415 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2417 struct kmem_list3
*l3
;
2420 on_each_cpu(do_drain
, cachep
, 1);
2422 for_each_online_node(node
) {
2423 l3
= cachep
->nodelists
[node
];
2424 if (l3
&& l3
->alien
)
2425 drain_alien_cache(cachep
, l3
->alien
);
2428 for_each_online_node(node
) {
2429 l3
= cachep
->nodelists
[node
];
2431 drain_array(cachep
, l3
, l3
->shared
, 1, node
);
2436 * Remove slabs from the list of free slabs.
2437 * Specify the number of slabs to drain in tofree.
2439 * Returns the actual number of slabs released.
2441 static int drain_freelist(struct kmem_cache
*cache
,
2442 struct kmem_list3
*l3
, int tofree
)
2444 struct list_head
*p
;
2449 while (nr_freed
< tofree
&& !list_empty(&l3
->slabs_free
)) {
2451 spin_lock_irq(&l3
->list_lock
);
2452 p
= l3
->slabs_free
.prev
;
2453 if (p
== &l3
->slabs_free
) {
2454 spin_unlock_irq(&l3
->list_lock
);
2458 slabp
= list_entry(p
, struct slab
, list
);
2460 BUG_ON(slabp
->inuse
);
2462 list_del(&slabp
->list
);
2464 * Safe to drop the lock. The slab is no longer linked
2467 l3
->free_objects
-= cache
->num
;
2468 spin_unlock_irq(&l3
->list_lock
);
2469 slab_destroy(cache
, slabp
);
2476 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2477 static int __cache_shrink(struct kmem_cache
*cachep
)
2480 struct kmem_list3
*l3
;
2482 drain_cpu_caches(cachep
);
2485 for_each_online_node(i
) {
2486 l3
= cachep
->nodelists
[i
];
2490 drain_freelist(cachep
, l3
, l3
->free_objects
);
2492 ret
+= !list_empty(&l3
->slabs_full
) ||
2493 !list_empty(&l3
->slabs_partial
);
2495 return (ret
? 1 : 0);
2499 * kmem_cache_shrink - Shrink a cache.
2500 * @cachep: The cache to shrink.
2502 * Releases as many slabs as possible for a cache.
2503 * To help debugging, a zero exit status indicates all slabs were released.
2505 int kmem_cache_shrink(struct kmem_cache
*cachep
)
2508 BUG_ON(!cachep
|| in_interrupt());
2511 mutex_lock(&cache_chain_mutex
);
2512 ret
= __cache_shrink(cachep
);
2513 mutex_unlock(&cache_chain_mutex
);
2517 EXPORT_SYMBOL(kmem_cache_shrink
);
2520 * kmem_cache_destroy - delete a cache
2521 * @cachep: the cache to destroy
2523 * Remove a &struct kmem_cache object from the slab cache.
2525 * It is expected this function will be called by a module when it is
2526 * unloaded. This will remove the cache completely, and avoid a duplicate
2527 * cache being allocated each time a module is loaded and unloaded, if the
2528 * module doesn't have persistent in-kernel storage across loads and unloads.
2530 * The cache must be empty before calling this function.
2532 * The caller must guarantee that noone will allocate memory from the cache
2533 * during the kmem_cache_destroy().
2535 void kmem_cache_destroy(struct kmem_cache
*cachep
)
2537 BUG_ON(!cachep
|| in_interrupt());
2539 /* Find the cache in the chain of caches. */
2541 mutex_lock(&cache_chain_mutex
);
2543 * the chain is never empty, cache_cache is never destroyed
2545 list_del(&cachep
->next
);
2546 if (__cache_shrink(cachep
)) {
2547 slab_error(cachep
, "Can't free all objects");
2548 list_add(&cachep
->next
, &cache_chain
);
2549 mutex_unlock(&cache_chain_mutex
);
2554 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
2557 __kmem_cache_destroy(cachep
);
2558 mutex_unlock(&cache_chain_mutex
);
2561 EXPORT_SYMBOL(kmem_cache_destroy
);
2564 * Get the memory for a slab management obj.
2565 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2566 * always come from malloc_sizes caches. The slab descriptor cannot
2567 * come from the same cache which is getting created because,
2568 * when we are searching for an appropriate cache for these
2569 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2570 * If we are creating a malloc_sizes cache here it would not be visible to
2571 * kmem_find_general_cachep till the initialization is complete.
2572 * Hence we cannot have slabp_cache same as the original cache.
2574 static struct slab
*alloc_slabmgmt(struct kmem_cache
*cachep
, void *objp
,
2575 int colour_off
, gfp_t local_flags
,
2580 if (OFF_SLAB(cachep
)) {
2581 /* Slab management obj is off-slab. */
2582 slabp
= kmem_cache_alloc_node(cachep
->slabp_cache
,
2583 local_flags
, nodeid
);
2585 * If the first object in the slab is leaked (it's allocated
2586 * but no one has a reference to it), we want to make sure
2587 * kmemleak does not treat the ->s_mem pointer as a reference
2588 * to the object. Otherwise we will not report the leak.
2590 kmemleak_scan_area(slabp
, offsetof(struct slab
, list
),
2591 sizeof(struct list_head
), local_flags
);
2595 slabp
= objp
+ colour_off
;
2596 colour_off
+= cachep
->slab_size
;
2599 slabp
->colouroff
= colour_off
;
2600 slabp
->s_mem
= objp
+ colour_off
;
2601 slabp
->nodeid
= nodeid
;
2606 static inline kmem_bufctl_t
*slab_bufctl(struct slab
*slabp
)
2608 return (kmem_bufctl_t
*) (slabp
+ 1);
2611 static void cache_init_objs(struct kmem_cache
*cachep
,
2616 for (i
= 0; i
< cachep
->num
; i
++) {
2617 void *objp
= index_to_obj(cachep
, slabp
, i
);
2619 /* need to poison the objs? */
2620 if (cachep
->flags
& SLAB_POISON
)
2621 poison_obj(cachep
, objp
, POISON_FREE
);
2622 if (cachep
->flags
& SLAB_STORE_USER
)
2623 *dbg_userword(cachep
, objp
) = NULL
;
2625 if (cachep
->flags
& SLAB_RED_ZONE
) {
2626 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2627 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2630 * Constructors are not allowed to allocate memory from the same
2631 * cache which they are a constructor for. Otherwise, deadlock.
2632 * They must also be threaded.
2634 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2635 cachep
->ctor(objp
+ obj_offset(cachep
));
2637 if (cachep
->flags
& SLAB_RED_ZONE
) {
2638 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2639 slab_error(cachep
, "constructor overwrote the"
2640 " end of an object");
2641 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2642 slab_error(cachep
, "constructor overwrote the"
2643 " start of an object");
2645 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 &&
2646 OFF_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
)
2647 kernel_map_pages(virt_to_page(objp
),
2648 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2653 slab_bufctl(slabp
)[i
] = i
+ 1;
2655 slab_bufctl(slabp
)[i
- 1] = BUFCTL_END
;
2658 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2660 if (CONFIG_ZONE_DMA_FLAG
) {
2661 if (flags
& GFP_DMA
)
2662 BUG_ON(!(cachep
->gfpflags
& GFP_DMA
));
2664 BUG_ON(cachep
->gfpflags
& GFP_DMA
);
2668 static void *slab_get_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2671 void *objp
= index_to_obj(cachep
, slabp
, slabp
->free
);
2675 next
= slab_bufctl(slabp
)[slabp
->free
];
2677 slab_bufctl(slabp
)[slabp
->free
] = BUFCTL_FREE
;
2678 WARN_ON(slabp
->nodeid
!= nodeid
);
2685 static void slab_put_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2686 void *objp
, int nodeid
)
2688 unsigned int objnr
= obj_to_index(cachep
, slabp
, objp
);
2691 /* Verify that the slab belongs to the intended node */
2692 WARN_ON(slabp
->nodeid
!= nodeid
);
2694 if (slab_bufctl(slabp
)[objnr
] + 1 <= SLAB_LIMIT
+ 1) {
2695 printk(KERN_ERR
"slab: double free detected in cache "
2696 "'%s', objp %p\n", cachep
->name
, objp
);
2700 slab_bufctl(slabp
)[objnr
] = slabp
->free
;
2701 slabp
->free
= objnr
;
2706 * Map pages beginning at addr to the given cache and slab. This is required
2707 * for the slab allocator to be able to lookup the cache and slab of a
2708 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2710 static void slab_map_pages(struct kmem_cache
*cache
, struct slab
*slab
,
2716 page
= virt_to_page(addr
);
2719 if (likely(!PageCompound(page
)))
2720 nr_pages
<<= cache
->gfporder
;
2723 page_set_cache(page
, cache
);
2724 page_set_slab(page
, slab
);
2726 } while (--nr_pages
);
2730 * Grow (by 1) the number of slabs within a cache. This is called by
2731 * kmem_cache_alloc() when there are no active objs left in a cache.
2733 static int cache_grow(struct kmem_cache
*cachep
,
2734 gfp_t flags
, int nodeid
, void *objp
)
2739 struct kmem_list3
*l3
;
2742 * Be lazy and only check for valid flags here, keeping it out of the
2743 * critical path in kmem_cache_alloc().
2745 BUG_ON(flags
& GFP_SLAB_BUG_MASK
);
2746 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2748 /* Take the l3 list lock to change the colour_next on this node */
2750 l3
= cachep
->nodelists
[nodeid
];
2751 spin_lock(&l3
->list_lock
);
2753 /* Get colour for the slab, and cal the next value. */
2754 offset
= l3
->colour_next
;
2756 if (l3
->colour_next
>= cachep
->colour
)
2757 l3
->colour_next
= 0;
2758 spin_unlock(&l3
->list_lock
);
2760 offset
*= cachep
->colour_off
;
2762 if (local_flags
& __GFP_WAIT
)
2766 * The test for missing atomic flag is performed here, rather than
2767 * the more obvious place, simply to reduce the critical path length
2768 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2769 * will eventually be caught here (where it matters).
2771 kmem_flagcheck(cachep
, flags
);
2774 * Get mem for the objs. Attempt to allocate a physical page from
2778 objp
= kmem_getpages(cachep
, local_flags
, nodeid
);
2782 /* Get slab management. */
2783 slabp
= alloc_slabmgmt(cachep
, objp
, offset
,
2784 local_flags
& ~GFP_CONSTRAINT_MASK
, nodeid
);
2788 slab_map_pages(cachep
, slabp
, objp
);
2790 cache_init_objs(cachep
, slabp
);
2792 if (local_flags
& __GFP_WAIT
)
2793 local_irq_disable();
2795 spin_lock(&l3
->list_lock
);
2797 /* Make slab active. */
2798 list_add_tail(&slabp
->list
, &(l3
->slabs_free
));
2799 STATS_INC_GROWN(cachep
);
2800 l3
->free_objects
+= cachep
->num
;
2801 spin_unlock(&l3
->list_lock
);
2804 kmem_freepages(cachep
, objp
);
2806 if (local_flags
& __GFP_WAIT
)
2807 local_irq_disable();
2814 * Perform extra freeing checks:
2815 * - detect bad pointers.
2816 * - POISON/RED_ZONE checking
2818 static void kfree_debugcheck(const void *objp
)
2820 if (!virt_addr_valid(objp
)) {
2821 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
2822 (unsigned long)objp
);
2827 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2829 unsigned long long redzone1
, redzone2
;
2831 redzone1
= *dbg_redzone1(cache
, obj
);
2832 redzone2
= *dbg_redzone2(cache
, obj
);
2837 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2840 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2841 slab_error(cache
, "double free detected");
2843 slab_error(cache
, "memory outside object was overwritten");
2845 printk(KERN_ERR
"%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2846 obj
, redzone1
, redzone2
);
2849 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2856 BUG_ON(virt_to_cache(objp
) != cachep
);
2858 objp
-= obj_offset(cachep
);
2859 kfree_debugcheck(objp
);
2860 page
= virt_to_head_page(objp
);
2862 slabp
= page_get_slab(page
);
2864 if (cachep
->flags
& SLAB_RED_ZONE
) {
2865 verify_redzone_free(cachep
, objp
);
2866 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2867 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2869 if (cachep
->flags
& SLAB_STORE_USER
)
2870 *dbg_userword(cachep
, objp
) = caller
;
2872 objnr
= obj_to_index(cachep
, slabp
, objp
);
2874 BUG_ON(objnr
>= cachep
->num
);
2875 BUG_ON(objp
!= index_to_obj(cachep
, slabp
, objnr
));
2877 #ifdef CONFIG_DEBUG_SLAB_LEAK
2878 slab_bufctl(slabp
)[objnr
] = BUFCTL_FREE
;
2880 if (cachep
->flags
& SLAB_POISON
) {
2881 #ifdef CONFIG_DEBUG_PAGEALLOC
2882 if ((cachep
->buffer_size
% PAGE_SIZE
)==0 && OFF_SLAB(cachep
)) {
2883 store_stackinfo(cachep
, objp
, (unsigned long)caller
);
2884 kernel_map_pages(virt_to_page(objp
),
2885 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2887 poison_obj(cachep
, objp
, POISON_FREE
);
2890 poison_obj(cachep
, objp
, POISON_FREE
);
2896 static void check_slabp(struct kmem_cache
*cachep
, struct slab
*slabp
)
2901 /* Check slab's freelist to see if this obj is there. */
2902 for (i
= slabp
->free
; i
!= BUFCTL_END
; i
= slab_bufctl(slabp
)[i
]) {
2904 if (entries
> cachep
->num
|| i
>= cachep
->num
)
2907 if (entries
!= cachep
->num
- slabp
->inuse
) {
2909 printk(KERN_ERR
"slab: Internal list corruption detected in "
2910 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2911 cachep
->name
, cachep
->num
, slabp
, slabp
->inuse
);
2913 i
< sizeof(*slabp
) + cachep
->num
* sizeof(kmem_bufctl_t
);
2916 printk("\n%03x:", i
);
2917 printk(" %02x", ((unsigned char *)slabp
)[i
]);
2924 #define kfree_debugcheck(x) do { } while(0)
2925 #define cache_free_debugcheck(x,objp,z) (objp)
2926 #define check_slabp(x,y) do { } while(0)
2929 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
2932 struct kmem_list3
*l3
;
2933 struct array_cache
*ac
;
2938 node
= numa_node_id();
2939 ac
= cpu_cache_get(cachep
);
2940 batchcount
= ac
->batchcount
;
2941 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2943 * If there was little recent activity on this cache, then
2944 * perform only a partial refill. Otherwise we could generate
2947 batchcount
= BATCHREFILL_LIMIT
;
2949 l3
= cachep
->nodelists
[node
];
2951 BUG_ON(ac
->avail
> 0 || !l3
);
2952 spin_lock(&l3
->list_lock
);
2954 /* See if we can refill from the shared array */
2955 if (l3
->shared
&& transfer_objects(ac
, l3
->shared
, batchcount
))
2958 while (batchcount
> 0) {
2959 struct list_head
*entry
;
2961 /* Get slab alloc is to come from. */
2962 entry
= l3
->slabs_partial
.next
;
2963 if (entry
== &l3
->slabs_partial
) {
2964 l3
->free_touched
= 1;
2965 entry
= l3
->slabs_free
.next
;
2966 if (entry
== &l3
->slabs_free
)
2970 slabp
= list_entry(entry
, struct slab
, list
);
2971 check_slabp(cachep
, slabp
);
2972 check_spinlock_acquired(cachep
);
2975 * The slab was either on partial or free list so
2976 * there must be at least one object available for
2979 BUG_ON(slabp
->inuse
>= cachep
->num
);
2981 while (slabp
->inuse
< cachep
->num
&& batchcount
--) {
2982 STATS_INC_ALLOCED(cachep
);
2983 STATS_INC_ACTIVE(cachep
);
2984 STATS_SET_HIGH(cachep
);
2986 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, slabp
,
2989 check_slabp(cachep
, slabp
);
2991 /* move slabp to correct slabp list: */
2992 list_del(&slabp
->list
);
2993 if (slabp
->free
== BUFCTL_END
)
2994 list_add(&slabp
->list
, &l3
->slabs_full
);
2996 list_add(&slabp
->list
, &l3
->slabs_partial
);
3000 l3
->free_objects
-= ac
->avail
;
3002 spin_unlock(&l3
->list_lock
);
3004 if (unlikely(!ac
->avail
)) {
3006 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, node
, NULL
);
3008 /* cache_grow can reenable interrupts, then ac could change. */
3009 ac
= cpu_cache_get(cachep
);
3010 if (!x
&& ac
->avail
== 0) /* no objects in sight? abort */
3013 if (!ac
->avail
) /* objects refilled by interrupt? */
3017 return ac
->entry
[--ac
->avail
];
3020 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
3023 might_sleep_if(flags
& __GFP_WAIT
);
3025 kmem_flagcheck(cachep
, flags
);
3030 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
3031 gfp_t flags
, void *objp
, void *caller
)
3035 if (cachep
->flags
& SLAB_POISON
) {
3036 #ifdef CONFIG_DEBUG_PAGEALLOC
3037 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
3038 kernel_map_pages(virt_to_page(objp
),
3039 cachep
->buffer_size
/ PAGE_SIZE
, 1);
3041 check_poison_obj(cachep
, objp
);
3043 check_poison_obj(cachep
, objp
);
3045 poison_obj(cachep
, objp
, POISON_INUSE
);
3047 if (cachep
->flags
& SLAB_STORE_USER
)
3048 *dbg_userword(cachep
, objp
) = caller
;
3050 if (cachep
->flags
& SLAB_RED_ZONE
) {
3051 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
3052 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
3053 slab_error(cachep
, "double free, or memory outside"
3054 " object was overwritten");
3056 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3057 objp
, *dbg_redzone1(cachep
, objp
),
3058 *dbg_redzone2(cachep
, objp
));
3060 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
3061 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
3063 #ifdef CONFIG_DEBUG_SLAB_LEAK
3068 slabp
= page_get_slab(virt_to_head_page(objp
));
3069 objnr
= (unsigned)(objp
- slabp
->s_mem
) / cachep
->buffer_size
;
3070 slab_bufctl(slabp
)[objnr
] = BUFCTL_ACTIVE
;
3073 objp
+= obj_offset(cachep
);
3074 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
3076 #if ARCH_SLAB_MINALIGN
3077 if ((u32
)objp
& (ARCH_SLAB_MINALIGN
-1)) {
3078 printk(KERN_ERR
"0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3079 objp
, ARCH_SLAB_MINALIGN
);
3085 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3088 static bool slab_should_failslab(struct kmem_cache
*cachep
, gfp_t flags
)
3090 if (cachep
== &cache_cache
)
3093 return should_failslab(obj_size(cachep
), flags
);
3096 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3099 struct array_cache
*ac
;
3103 ac
= cpu_cache_get(cachep
);
3104 if (likely(ac
->avail
)) {
3105 STATS_INC_ALLOCHIT(cachep
);
3107 objp
= ac
->entry
[--ac
->avail
];
3109 STATS_INC_ALLOCMISS(cachep
);
3110 objp
= cache_alloc_refill(cachep
, flags
);
3113 * To avoid a false negative, if an object that is in one of the
3114 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3115 * treat the array pointers as a reference to the object.
3117 kmemleak_erase(&ac
->entry
[ac
->avail
]);
3123 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3125 * If we are in_interrupt, then process context, including cpusets and
3126 * mempolicy, may not apply and should not be used for allocation policy.
3128 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3130 int nid_alloc
, nid_here
;
3132 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3134 nid_alloc
= nid_here
= numa_node_id();
3135 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3136 nid_alloc
= cpuset_mem_spread_node();
3137 else if (current
->mempolicy
)
3138 nid_alloc
= slab_node(current
->mempolicy
);
3139 if (nid_alloc
!= nid_here
)
3140 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3145 * Fallback function if there was no memory available and no objects on a
3146 * certain node and fall back is permitted. First we scan all the
3147 * available nodelists for available objects. If that fails then we
3148 * perform an allocation without specifying a node. This allows the page
3149 * allocator to do its reclaim / fallback magic. We then insert the
3150 * slab into the proper nodelist and then allocate from it.
3152 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3154 struct zonelist
*zonelist
;
3158 enum zone_type high_zoneidx
= gfp_zone(flags
);
3162 if (flags
& __GFP_THISNODE
)
3165 zonelist
= node_zonelist(slab_node(current
->mempolicy
), flags
);
3166 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
3170 * Look through allowed nodes for objects available
3171 * from existing per node queues.
3173 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
3174 nid
= zone_to_nid(zone
);
3176 if (cpuset_zone_allowed_hardwall(zone
, flags
) &&
3177 cache
->nodelists
[nid
] &&
3178 cache
->nodelists
[nid
]->free_objects
) {
3179 obj
= ____cache_alloc_node(cache
,
3180 flags
| GFP_THISNODE
, nid
);
3188 * This allocation will be performed within the constraints
3189 * of the current cpuset / memory policy requirements.
3190 * We may trigger various forms of reclaim on the allowed
3191 * set and go into memory reserves if necessary.
3193 if (local_flags
& __GFP_WAIT
)
3195 kmem_flagcheck(cache
, flags
);
3196 obj
= kmem_getpages(cache
, local_flags
, -1);
3197 if (local_flags
& __GFP_WAIT
)
3198 local_irq_disable();
3201 * Insert into the appropriate per node queues
3203 nid
= page_to_nid(virt_to_page(obj
));
3204 if (cache_grow(cache
, flags
, nid
, obj
)) {
3205 obj
= ____cache_alloc_node(cache
,
3206 flags
| GFP_THISNODE
, nid
);
3209 * Another processor may allocate the
3210 * objects in the slab since we are
3211 * not holding any locks.
3215 /* cache_grow already freed obj */
3224 * A interface to enable slab creation on nodeid
3226 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3229 struct list_head
*entry
;
3231 struct kmem_list3
*l3
;
3235 l3
= cachep
->nodelists
[nodeid
];
3240 spin_lock(&l3
->list_lock
);
3241 entry
= l3
->slabs_partial
.next
;
3242 if (entry
== &l3
->slabs_partial
) {
3243 l3
->free_touched
= 1;
3244 entry
= l3
->slabs_free
.next
;
3245 if (entry
== &l3
->slabs_free
)
3249 slabp
= list_entry(entry
, struct slab
, list
);
3250 check_spinlock_acquired_node(cachep
, nodeid
);
3251 check_slabp(cachep
, slabp
);
3253 STATS_INC_NODEALLOCS(cachep
);
3254 STATS_INC_ACTIVE(cachep
);
3255 STATS_SET_HIGH(cachep
);
3257 BUG_ON(slabp
->inuse
== cachep
->num
);
3259 obj
= slab_get_obj(cachep
, slabp
, nodeid
);
3260 check_slabp(cachep
, slabp
);
3262 /* move slabp to correct slabp list: */
3263 list_del(&slabp
->list
);
3265 if (slabp
->free
== BUFCTL_END
)
3266 list_add(&slabp
->list
, &l3
->slabs_full
);
3268 list_add(&slabp
->list
, &l3
->slabs_partial
);
3270 spin_unlock(&l3
->list_lock
);
3274 spin_unlock(&l3
->list_lock
);
3275 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, nodeid
, NULL
);
3279 return fallback_alloc(cachep
, flags
);
3286 * kmem_cache_alloc_node - Allocate an object on the specified node
3287 * @cachep: The cache to allocate from.
3288 * @flags: See kmalloc().
3289 * @nodeid: node number of the target node.
3290 * @caller: return address of caller, used for debug information
3292 * Identical to kmem_cache_alloc but it will allocate memory on the given
3293 * node, which can improve the performance for cpu bound structures.
3295 * Fallback to other node is possible if __GFP_THISNODE is not set.
3297 static __always_inline
void *
3298 __cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3301 unsigned long save_flags
;
3304 flags
&= slab_gfp_mask
;
3306 lockdep_trace_alloc(flags
);
3308 if (slab_should_failslab(cachep
, flags
))
3311 cache_alloc_debugcheck_before(cachep
, flags
);
3312 local_irq_save(save_flags
);
3314 if (unlikely(nodeid
== -1))
3315 nodeid
= numa_node_id();
3317 if (unlikely(!cachep
->nodelists
[nodeid
])) {
3318 /* Node not bootstrapped yet */
3319 ptr
= fallback_alloc(cachep
, flags
);
3323 if (nodeid
== numa_node_id()) {
3325 * Use the locally cached objects if possible.
3326 * However ____cache_alloc does not allow fallback
3327 * to other nodes. It may fail while we still have
3328 * objects on other nodes available.
3330 ptr
= ____cache_alloc(cachep
, flags
);
3334 /* ___cache_alloc_node can fall back to other nodes */
3335 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3337 local_irq_restore(save_flags
);
3338 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3339 kmemleak_alloc_recursive(ptr
, obj_size(cachep
), 1, cachep
->flags
,
3343 kmemcheck_slab_alloc(cachep
, flags
, ptr
, obj_size(cachep
));
3345 if (unlikely((flags
& __GFP_ZERO
) && ptr
))
3346 memset(ptr
, 0, obj_size(cachep
));
3351 static __always_inline
void *
3352 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3356 if (unlikely(current
->flags
& (PF_SPREAD_SLAB
| PF_MEMPOLICY
))) {
3357 objp
= alternate_node_alloc(cache
, flags
);
3361 objp
= ____cache_alloc(cache
, flags
);
3364 * We may just have run out of memory on the local node.
3365 * ____cache_alloc_node() knows how to locate memory on other nodes
3368 objp
= ____cache_alloc_node(cache
, flags
, numa_node_id());
3375 static __always_inline
void *
3376 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3378 return ____cache_alloc(cachep
, flags
);
3381 #endif /* CONFIG_NUMA */
3383 static __always_inline
void *
3384 __cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
, void *caller
)
3386 unsigned long save_flags
;
3389 flags
&= slab_gfp_mask
;
3391 lockdep_trace_alloc(flags
);
3393 if (slab_should_failslab(cachep
, flags
))
3396 cache_alloc_debugcheck_before(cachep
, flags
);
3397 local_irq_save(save_flags
);
3398 objp
= __do_cache_alloc(cachep
, flags
);
3399 local_irq_restore(save_flags
);
3400 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3401 kmemleak_alloc_recursive(objp
, obj_size(cachep
), 1, cachep
->flags
,
3406 kmemcheck_slab_alloc(cachep
, flags
, objp
, obj_size(cachep
));
3408 if (unlikely((flags
& __GFP_ZERO
) && objp
))
3409 memset(objp
, 0, obj_size(cachep
));
3415 * Caller needs to acquire correct kmem_list's list_lock
3417 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int nr_objects
,
3421 struct kmem_list3
*l3
;
3423 for (i
= 0; i
< nr_objects
; i
++) {
3424 void *objp
= objpp
[i
];
3427 slabp
= virt_to_slab(objp
);
3428 l3
= cachep
->nodelists
[node
];
3429 list_del(&slabp
->list
);
3430 check_spinlock_acquired_node(cachep
, node
);
3431 check_slabp(cachep
, slabp
);
3432 slab_put_obj(cachep
, slabp
, objp
, node
);
3433 STATS_DEC_ACTIVE(cachep
);
3435 check_slabp(cachep
, slabp
);
3437 /* fixup slab chains */
3438 if (slabp
->inuse
== 0) {
3439 if (l3
->free_objects
> l3
->free_limit
) {
3440 l3
->free_objects
-= cachep
->num
;
3441 /* No need to drop any previously held
3442 * lock here, even if we have a off-slab slab
3443 * descriptor it is guaranteed to come from
3444 * a different cache, refer to comments before
3447 slab_destroy(cachep
, slabp
);
3449 list_add(&slabp
->list
, &l3
->slabs_free
);
3452 /* Unconditionally move a slab to the end of the
3453 * partial list on free - maximum time for the
3454 * other objects to be freed, too.
3456 list_add_tail(&slabp
->list
, &l3
->slabs_partial
);
3461 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3464 struct kmem_list3
*l3
;
3465 int node
= numa_node_id();
3467 batchcount
= ac
->batchcount
;
3469 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
3472 l3
= cachep
->nodelists
[node
];
3473 spin_lock(&l3
->list_lock
);
3475 struct array_cache
*shared_array
= l3
->shared
;
3476 int max
= shared_array
->limit
- shared_array
->avail
;
3478 if (batchcount
> max
)
3480 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3481 ac
->entry
, sizeof(void *) * batchcount
);
3482 shared_array
->avail
+= batchcount
;
3487 free_block(cachep
, ac
->entry
, batchcount
, node
);
3492 struct list_head
*p
;
3494 p
= l3
->slabs_free
.next
;
3495 while (p
!= &(l3
->slabs_free
)) {
3498 slabp
= list_entry(p
, struct slab
, list
);
3499 BUG_ON(slabp
->inuse
);
3504 STATS_SET_FREEABLE(cachep
, i
);
3507 spin_unlock(&l3
->list_lock
);
3508 ac
->avail
-= batchcount
;
3509 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3513 * Release an obj back to its cache. If the obj has a constructed state, it must
3514 * be in this state _before_ it is released. Called with disabled ints.
3516 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
)
3518 struct array_cache
*ac
= cpu_cache_get(cachep
);
3521 kmemleak_free_recursive(objp
, cachep
->flags
);
3522 objp
= cache_free_debugcheck(cachep
, objp
, __builtin_return_address(0));
3524 kmemcheck_slab_free(cachep
, objp
, obj_size(cachep
));
3527 * Skip calling cache_free_alien() when the platform is not numa.
3528 * This will avoid cache misses that happen while accessing slabp (which
3529 * is per page memory reference) to get nodeid. Instead use a global
3530 * variable to skip the call, which is mostly likely to be present in
3533 if (numa_platform
&& cache_free_alien(cachep
, objp
))
3536 if (likely(ac
->avail
< ac
->limit
)) {
3537 STATS_INC_FREEHIT(cachep
);
3538 ac
->entry
[ac
->avail
++] = objp
;
3541 STATS_INC_FREEMISS(cachep
);
3542 cache_flusharray(cachep
, ac
);
3543 ac
->entry
[ac
->avail
++] = objp
;
3548 * kmem_cache_alloc - Allocate an object
3549 * @cachep: The cache to allocate from.
3550 * @flags: See kmalloc().
3552 * Allocate an object from this cache. The flags are only relevant
3553 * if the cache has no available objects.
3555 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3557 void *ret
= __cache_alloc(cachep
, flags
, __builtin_return_address(0));
3559 trace_kmem_cache_alloc(_RET_IP_
, ret
,
3560 obj_size(cachep
), cachep
->buffer_size
, flags
);
3564 EXPORT_SYMBOL(kmem_cache_alloc
);
3566 #ifdef CONFIG_KMEMTRACE
3567 void *kmem_cache_alloc_notrace(struct kmem_cache
*cachep
, gfp_t flags
)
3569 return __cache_alloc(cachep
, flags
, __builtin_return_address(0));
3571 EXPORT_SYMBOL(kmem_cache_alloc_notrace
);
3575 * kmem_ptr_validate - check if an untrusted pointer might be a slab entry.
3576 * @cachep: the cache we're checking against
3577 * @ptr: pointer to validate
3579 * This verifies that the untrusted pointer looks sane;
3580 * it is _not_ a guarantee that the pointer is actually
3581 * part of the slab cache in question, but it at least
3582 * validates that the pointer can be dereferenced and
3583 * looks half-way sane.
3585 * Currently only used for dentry validation.
3587 int kmem_ptr_validate(struct kmem_cache
*cachep
, const void *ptr
)
3589 unsigned long addr
= (unsigned long)ptr
;
3590 unsigned long min_addr
= PAGE_OFFSET
;
3591 unsigned long align_mask
= BYTES_PER_WORD
- 1;
3592 unsigned long size
= cachep
->buffer_size
;
3595 if (unlikely(addr
< min_addr
))
3597 if (unlikely(addr
> (unsigned long)high_memory
- size
))
3599 if (unlikely(addr
& align_mask
))
3601 if (unlikely(!kern_addr_valid(addr
)))
3603 if (unlikely(!kern_addr_valid(addr
+ size
- 1)))
3605 page
= virt_to_page(ptr
);
3606 if (unlikely(!PageSlab(page
)))
3608 if (unlikely(page_get_cache(page
) != cachep
))
3616 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3618 void *ret
= __cache_alloc_node(cachep
, flags
, nodeid
,
3619 __builtin_return_address(0));
3621 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
3622 obj_size(cachep
), cachep
->buffer_size
,
3627 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3629 #ifdef CONFIG_KMEMTRACE
3630 void *kmem_cache_alloc_node_notrace(struct kmem_cache
*cachep
,
3634 return __cache_alloc_node(cachep
, flags
, nodeid
,
3635 __builtin_return_address(0));
3637 EXPORT_SYMBOL(kmem_cache_alloc_node_notrace
);
3640 static __always_inline
void *
3641 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, void *caller
)
3643 struct kmem_cache
*cachep
;
3646 cachep
= kmem_find_general_cachep(size
, flags
);
3647 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3649 ret
= kmem_cache_alloc_node_notrace(cachep
, flags
, node
);
3651 trace_kmalloc_node((unsigned long) caller
, ret
,
3652 size
, cachep
->buffer_size
, flags
, node
);
3657 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_KMEMTRACE)
3658 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3660 return __do_kmalloc_node(size
, flags
, node
,
3661 __builtin_return_address(0));
3663 EXPORT_SYMBOL(__kmalloc_node
);
3665 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3666 int node
, unsigned long caller
)
3668 return __do_kmalloc_node(size
, flags
, node
, (void *)caller
);
3670 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3672 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3674 return __do_kmalloc_node(size
, flags
, node
, NULL
);
3676 EXPORT_SYMBOL(__kmalloc_node
);
3677 #endif /* CONFIG_DEBUG_SLAB */
3678 #endif /* CONFIG_NUMA */
3681 * __do_kmalloc - allocate memory
3682 * @size: how many bytes of memory are required.
3683 * @flags: the type of memory to allocate (see kmalloc).
3684 * @caller: function caller for debug tracking of the caller
3686 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3689 struct kmem_cache
*cachep
;
3692 /* If you want to save a few bytes .text space: replace
3694 * Then kmalloc uses the uninlined functions instead of the inline
3697 cachep
= __find_general_cachep(size
, flags
);
3698 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3700 ret
= __cache_alloc(cachep
, flags
, caller
);
3702 trace_kmalloc((unsigned long) caller
, ret
,
3703 size
, cachep
->buffer_size
, flags
);
3709 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_KMEMTRACE)
3710 void *__kmalloc(size_t size
, gfp_t flags
)
3712 return __do_kmalloc(size
, flags
, __builtin_return_address(0));
3714 EXPORT_SYMBOL(__kmalloc
);
3716 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, unsigned long caller
)
3718 return __do_kmalloc(size
, flags
, (void *)caller
);
3720 EXPORT_SYMBOL(__kmalloc_track_caller
);
3723 void *__kmalloc(size_t size
, gfp_t flags
)
3725 return __do_kmalloc(size
, flags
, NULL
);
3727 EXPORT_SYMBOL(__kmalloc
);
3731 * kmem_cache_free - Deallocate an object
3732 * @cachep: The cache the allocation was from.
3733 * @objp: The previously allocated object.
3735 * Free an object which was previously allocated from this
3738 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3740 unsigned long flags
;
3742 local_irq_save(flags
);
3743 debug_check_no_locks_freed(objp
, obj_size(cachep
));
3744 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3745 debug_check_no_obj_freed(objp
, obj_size(cachep
));
3746 __cache_free(cachep
, objp
);
3747 local_irq_restore(flags
);
3749 trace_kmem_cache_free(_RET_IP_
, objp
);
3751 EXPORT_SYMBOL(kmem_cache_free
);
3754 * kfree - free previously allocated memory
3755 * @objp: pointer returned by kmalloc.
3757 * If @objp is NULL, no operation is performed.
3759 * Don't free memory not originally allocated by kmalloc()
3760 * or you will run into trouble.
3762 void kfree(const void *objp
)
3764 struct kmem_cache
*c
;
3765 unsigned long flags
;
3767 trace_kfree(_RET_IP_
, objp
);
3769 if (unlikely(ZERO_OR_NULL_PTR(objp
)))
3771 local_irq_save(flags
);
3772 kfree_debugcheck(objp
);
3773 c
= virt_to_cache(objp
);
3774 debug_check_no_locks_freed(objp
, obj_size(c
));
3775 debug_check_no_obj_freed(objp
, obj_size(c
));
3776 __cache_free(c
, (void *)objp
);
3777 local_irq_restore(flags
);
3779 EXPORT_SYMBOL(kfree
);
3781 unsigned int kmem_cache_size(struct kmem_cache
*cachep
)
3783 return obj_size(cachep
);
3785 EXPORT_SYMBOL(kmem_cache_size
);
3787 const char *kmem_cache_name(struct kmem_cache
*cachep
)
3789 return cachep
->name
;
3791 EXPORT_SYMBOL_GPL(kmem_cache_name
);
3794 * This initializes kmem_list3 or resizes various caches for all nodes.
3796 static int alloc_kmemlist(struct kmem_cache
*cachep
, gfp_t gfp
)
3799 struct kmem_list3
*l3
;
3800 struct array_cache
*new_shared
;
3801 struct array_cache
**new_alien
= NULL
;
3803 for_each_online_node(node
) {
3805 if (use_alien_caches
) {
3806 new_alien
= alloc_alien_cache(node
, cachep
->limit
, gfp
);
3812 if (cachep
->shared
) {
3813 new_shared
= alloc_arraycache(node
,
3814 cachep
->shared
*cachep
->batchcount
,
3817 free_alien_cache(new_alien
);
3822 l3
= cachep
->nodelists
[node
];
3824 struct array_cache
*shared
= l3
->shared
;
3826 spin_lock_irq(&l3
->list_lock
);
3829 free_block(cachep
, shared
->entry
,
3830 shared
->avail
, node
);
3832 l3
->shared
= new_shared
;
3834 l3
->alien
= new_alien
;
3837 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3838 cachep
->batchcount
+ cachep
->num
;
3839 spin_unlock_irq(&l3
->list_lock
);
3841 free_alien_cache(new_alien
);
3844 l3
= kmalloc_node(sizeof(struct kmem_list3
), gfp
, node
);
3846 free_alien_cache(new_alien
);
3851 kmem_list3_init(l3
);
3852 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
3853 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
3854 l3
->shared
= new_shared
;
3855 l3
->alien
= new_alien
;
3856 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3857 cachep
->batchcount
+ cachep
->num
;
3858 cachep
->nodelists
[node
] = l3
;
3863 if (!cachep
->next
.next
) {
3864 /* Cache is not active yet. Roll back what we did */
3867 if (cachep
->nodelists
[node
]) {
3868 l3
= cachep
->nodelists
[node
];
3871 free_alien_cache(l3
->alien
);
3873 cachep
->nodelists
[node
] = NULL
;
3881 struct ccupdate_struct
{
3882 struct kmem_cache
*cachep
;
3883 struct array_cache
*new[NR_CPUS
];
3886 static void do_ccupdate_local(void *info
)
3888 struct ccupdate_struct
*new = info
;
3889 struct array_cache
*old
;
3892 old
= cpu_cache_get(new->cachep
);
3894 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
3895 new->new[smp_processor_id()] = old
;
3898 /* Always called with the cache_chain_mutex held */
3899 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3900 int batchcount
, int shared
, gfp_t gfp
)
3902 struct ccupdate_struct
*new;
3905 new = kzalloc(sizeof(*new), gfp
);
3909 for_each_online_cpu(i
) {
3910 new->new[i
] = alloc_arraycache(cpu_to_node(i
), limit
,
3913 for (i
--; i
>= 0; i
--)
3919 new->cachep
= cachep
;
3921 on_each_cpu(do_ccupdate_local
, (void *)new, 1);
3924 cachep
->batchcount
= batchcount
;
3925 cachep
->limit
= limit
;
3926 cachep
->shared
= shared
;
3928 for_each_online_cpu(i
) {
3929 struct array_cache
*ccold
= new->new[i
];
3932 spin_lock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3933 free_block(cachep
, ccold
->entry
, ccold
->avail
, cpu_to_node(i
));
3934 spin_unlock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3938 return alloc_kmemlist(cachep
, gfp
);
3941 /* Called with cache_chain_mutex held always */
3942 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
3948 * The head array serves three purposes:
3949 * - create a LIFO ordering, i.e. return objects that are cache-warm
3950 * - reduce the number of spinlock operations.
3951 * - reduce the number of linked list operations on the slab and
3952 * bufctl chains: array operations are cheaper.
3953 * The numbers are guessed, we should auto-tune as described by
3956 if (cachep
->buffer_size
> 131072)
3958 else if (cachep
->buffer_size
> PAGE_SIZE
)
3960 else if (cachep
->buffer_size
> 1024)
3962 else if (cachep
->buffer_size
> 256)
3968 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3969 * allocation behaviour: Most allocs on one cpu, most free operations
3970 * on another cpu. For these cases, an efficient object passing between
3971 * cpus is necessary. This is provided by a shared array. The array
3972 * replaces Bonwick's magazine layer.
3973 * On uniprocessor, it's functionally equivalent (but less efficient)
3974 * to a larger limit. Thus disabled by default.
3977 if (cachep
->buffer_size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
3982 * With debugging enabled, large batchcount lead to excessively long
3983 * periods with disabled local interrupts. Limit the batchcount
3988 err
= do_tune_cpucache(cachep
, limit
, (limit
+ 1) / 2, shared
, gfp
);
3990 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
3991 cachep
->name
, -err
);
3996 * Drain an array if it contains any elements taking the l3 lock only if
3997 * necessary. Note that the l3 listlock also protects the array_cache
3998 * if drain_array() is used on the shared array.
4000 void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
4001 struct array_cache
*ac
, int force
, int node
)
4005 if (!ac
|| !ac
->avail
)
4007 if (ac
->touched
&& !force
) {
4010 spin_lock_irq(&l3
->list_lock
);
4012 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
4013 if (tofree
> ac
->avail
)
4014 tofree
= (ac
->avail
+ 1) / 2;
4015 free_block(cachep
, ac
->entry
, tofree
, node
);
4016 ac
->avail
-= tofree
;
4017 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
4018 sizeof(void *) * ac
->avail
);
4020 spin_unlock_irq(&l3
->list_lock
);
4025 * cache_reap - Reclaim memory from caches.
4026 * @w: work descriptor
4028 * Called from workqueue/eventd every few seconds.
4030 * - clear the per-cpu caches for this CPU.
4031 * - return freeable pages to the main free memory pool.
4033 * If we cannot acquire the cache chain mutex then just give up - we'll try
4034 * again on the next iteration.
4036 static void cache_reap(struct work_struct
*w
)
4038 struct kmem_cache
*searchp
;
4039 struct kmem_list3
*l3
;
4040 int node
= numa_node_id();
4041 struct delayed_work
*work
= to_delayed_work(w
);
4043 if (!mutex_trylock(&cache_chain_mutex
))
4044 /* Give up. Setup the next iteration. */
4047 list_for_each_entry(searchp
, &cache_chain
, next
) {
4051 * We only take the l3 lock if absolutely necessary and we
4052 * have established with reasonable certainty that
4053 * we can do some work if the lock was obtained.
4055 l3
= searchp
->nodelists
[node
];
4057 reap_alien(searchp
, l3
);
4059 drain_array(searchp
, l3
, cpu_cache_get(searchp
), 0, node
);
4062 * These are racy checks but it does not matter
4063 * if we skip one check or scan twice.
4065 if (time_after(l3
->next_reap
, jiffies
))
4068 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
;
4070 drain_array(searchp
, l3
, l3
->shared
, 0, node
);
4072 if (l3
->free_touched
)
4073 l3
->free_touched
= 0;
4077 freed
= drain_freelist(searchp
, l3
, (l3
->free_limit
+
4078 5 * searchp
->num
- 1) / (5 * searchp
->num
));
4079 STATS_ADD_REAPED(searchp
, freed
);
4085 mutex_unlock(&cache_chain_mutex
);
4088 /* Set up the next iteration */
4089 schedule_delayed_work(work
, round_jiffies_relative(REAPTIMEOUT_CPUC
));
4092 #ifdef CONFIG_SLABINFO
4094 static void print_slabinfo_header(struct seq_file
*m
)
4097 * Output format version, so at least we can change it
4098 * without _too_ many complaints.
4101 seq_puts(m
, "slabinfo - version: 2.1 (statistics)\n");
4103 seq_puts(m
, "slabinfo - version: 2.1\n");
4105 seq_puts(m
, "# name <active_objs> <num_objs> <objsize> "
4106 "<objperslab> <pagesperslab>");
4107 seq_puts(m
, " : tunables <limit> <batchcount> <sharedfactor>");
4108 seq_puts(m
, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4110 seq_puts(m
, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4111 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4112 seq_puts(m
, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4117 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
4121 mutex_lock(&cache_chain_mutex
);
4123 print_slabinfo_header(m
);
4125 return seq_list_start(&cache_chain
, *pos
);
4128 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
4130 return seq_list_next(p
, &cache_chain
, pos
);
4133 static void s_stop(struct seq_file
*m
, void *p
)
4135 mutex_unlock(&cache_chain_mutex
);
4138 static int s_show(struct seq_file
*m
, void *p
)
4140 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, next
);
4142 unsigned long active_objs
;
4143 unsigned long num_objs
;
4144 unsigned long active_slabs
= 0;
4145 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
4149 struct kmem_list3
*l3
;
4153 for_each_online_node(node
) {
4154 l3
= cachep
->nodelists
[node
];
4159 spin_lock_irq(&l3
->list_lock
);
4161 list_for_each_entry(slabp
, &l3
->slabs_full
, list
) {
4162 if (slabp
->inuse
!= cachep
->num
&& !error
)
4163 error
= "slabs_full accounting error";
4164 active_objs
+= cachep
->num
;
4167 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
) {
4168 if (slabp
->inuse
== cachep
->num
&& !error
)
4169 error
= "slabs_partial inuse accounting error";
4170 if (!slabp
->inuse
&& !error
)
4171 error
= "slabs_partial/inuse accounting error";
4172 active_objs
+= slabp
->inuse
;
4175 list_for_each_entry(slabp
, &l3
->slabs_free
, list
) {
4176 if (slabp
->inuse
&& !error
)
4177 error
= "slabs_free/inuse accounting error";
4180 free_objects
+= l3
->free_objects
;
4182 shared_avail
+= l3
->shared
->avail
;
4184 spin_unlock_irq(&l3
->list_lock
);
4186 num_slabs
+= active_slabs
;
4187 num_objs
= num_slabs
* cachep
->num
;
4188 if (num_objs
- active_objs
!= free_objects
&& !error
)
4189 error
= "free_objects accounting error";
4191 name
= cachep
->name
;
4193 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
4195 seq_printf(m
, "%-17s %6lu %6lu %6u %4u %4d",
4196 name
, active_objs
, num_objs
, cachep
->buffer_size
,
4197 cachep
->num
, (1 << cachep
->gfporder
));
4198 seq_printf(m
, " : tunables %4u %4u %4u",
4199 cachep
->limit
, cachep
->batchcount
, cachep
->shared
);
4200 seq_printf(m
, " : slabdata %6lu %6lu %6lu",
4201 active_slabs
, num_slabs
, shared_avail
);
4204 unsigned long high
= cachep
->high_mark
;
4205 unsigned long allocs
= cachep
->num_allocations
;
4206 unsigned long grown
= cachep
->grown
;
4207 unsigned long reaped
= cachep
->reaped
;
4208 unsigned long errors
= cachep
->errors
;
4209 unsigned long max_freeable
= cachep
->max_freeable
;
4210 unsigned long node_allocs
= cachep
->node_allocs
;
4211 unsigned long node_frees
= cachep
->node_frees
;
4212 unsigned long overflows
= cachep
->node_overflow
;
4214 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu \
4215 %4lu %4lu %4lu %4lu %4lu", allocs
, high
, grown
,
4216 reaped
, errors
, max_freeable
, node_allocs
,
4217 node_frees
, overflows
);
4221 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4222 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4223 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4224 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4226 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4227 allochit
, allocmiss
, freehit
, freemiss
);
4235 * slabinfo_op - iterator that generates /proc/slabinfo
4244 * num-pages-per-slab
4245 * + further values on SMP and with statistics enabled
4248 static const struct seq_operations slabinfo_op
= {
4255 #define MAX_SLABINFO_WRITE 128
4257 * slabinfo_write - Tuning for the slab allocator
4259 * @buffer: user buffer
4260 * @count: data length
4263 ssize_t
slabinfo_write(struct file
*file
, const char __user
* buffer
,
4264 size_t count
, loff_t
*ppos
)
4266 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4267 int limit
, batchcount
, shared
, res
;
4268 struct kmem_cache
*cachep
;
4270 if (count
> MAX_SLABINFO_WRITE
)
4272 if (copy_from_user(&kbuf
, buffer
, count
))
4274 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4276 tmp
= strchr(kbuf
, ' ');
4281 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4284 /* Find the cache in the chain of caches. */
4285 mutex_lock(&cache_chain_mutex
);
4287 list_for_each_entry(cachep
, &cache_chain
, next
) {
4288 if (!strcmp(cachep
->name
, kbuf
)) {
4289 if (limit
< 1 || batchcount
< 1 ||
4290 batchcount
> limit
|| shared
< 0) {
4293 res
= do_tune_cpucache(cachep
, limit
,
4300 mutex_unlock(&cache_chain_mutex
);
4306 static int slabinfo_open(struct inode
*inode
, struct file
*file
)
4308 return seq_open(file
, &slabinfo_op
);
4311 static const struct file_operations proc_slabinfo_operations
= {
4312 .open
= slabinfo_open
,
4314 .write
= slabinfo_write
,
4315 .llseek
= seq_lseek
,
4316 .release
= seq_release
,
4319 #ifdef CONFIG_DEBUG_SLAB_LEAK
4321 static void *leaks_start(struct seq_file
*m
, loff_t
*pos
)
4323 mutex_lock(&cache_chain_mutex
);
4324 return seq_list_start(&cache_chain
, *pos
);
4327 static inline int add_caller(unsigned long *n
, unsigned long v
)
4337 unsigned long *q
= p
+ 2 * i
;
4351 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4357 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
, struct slab
*s
)
4363 for (i
= 0, p
= s
->s_mem
; i
< c
->num
; i
++, p
+= c
->buffer_size
) {
4364 if (slab_bufctl(s
)[i
] != BUFCTL_ACTIVE
)
4366 if (!add_caller(n
, (unsigned long)*dbg_userword(c
, p
)))
4371 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4373 #ifdef CONFIG_KALLSYMS
4374 unsigned long offset
, size
;
4375 char modname
[MODULE_NAME_LEN
], name
[KSYM_NAME_LEN
];
4377 if (lookup_symbol_attrs(address
, &size
, &offset
, modname
, name
) == 0) {
4378 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4380 seq_printf(m
, " [%s]", modname
);
4384 seq_printf(m
, "%p", (void *)address
);
4387 static int leaks_show(struct seq_file
*m
, void *p
)
4389 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, next
);
4391 struct kmem_list3
*l3
;
4393 unsigned long *n
= m
->private;
4397 if (!(cachep
->flags
& SLAB_STORE_USER
))
4399 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4402 /* OK, we can do it */
4406 for_each_online_node(node
) {
4407 l3
= cachep
->nodelists
[node
];
4412 spin_lock_irq(&l3
->list_lock
);
4414 list_for_each_entry(slabp
, &l3
->slabs_full
, list
)
4415 handle_slab(n
, cachep
, slabp
);
4416 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
)
4417 handle_slab(n
, cachep
, slabp
);
4418 spin_unlock_irq(&l3
->list_lock
);
4420 name
= cachep
->name
;
4422 /* Increase the buffer size */
4423 mutex_unlock(&cache_chain_mutex
);
4424 m
->private = kzalloc(n
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4426 /* Too bad, we are really out */
4428 mutex_lock(&cache_chain_mutex
);
4431 *(unsigned long *)m
->private = n
[0] * 2;
4433 mutex_lock(&cache_chain_mutex
);
4434 /* Now make sure this entry will be retried */
4438 for (i
= 0; i
< n
[1]; i
++) {
4439 seq_printf(m
, "%s: %lu ", name
, n
[2*i
+3]);
4440 show_symbol(m
, n
[2*i
+2]);
4447 static const struct seq_operations slabstats_op
= {
4448 .start
= leaks_start
,
4454 static int slabstats_open(struct inode
*inode
, struct file
*file
)
4456 unsigned long *n
= kzalloc(PAGE_SIZE
, GFP_KERNEL
);
4459 ret
= seq_open(file
, &slabstats_op
);
4461 struct seq_file
*m
= file
->private_data
;
4462 *n
= PAGE_SIZE
/ (2 * sizeof(unsigned long));
4471 static const struct file_operations proc_slabstats_operations
= {
4472 .open
= slabstats_open
,
4474 .llseek
= seq_lseek
,
4475 .release
= seq_release_private
,
4479 static int __init
slab_proc_init(void)
4481 proc_create("slabinfo",S_IWUSR
|S_IRUGO
,NULL
,&proc_slabinfo_operations
);
4482 #ifdef CONFIG_DEBUG_SLAB_LEAK
4483 proc_create("slab_allocators", 0, NULL
, &proc_slabstats_operations
);
4487 module_init(slab_proc_init
);
4491 * ksize - get the actual amount of memory allocated for a given object
4492 * @objp: Pointer to the object
4494 * kmalloc may internally round up allocations and return more memory
4495 * than requested. ksize() can be used to determine the actual amount of
4496 * memory allocated. The caller may use this additional memory, even though
4497 * a smaller amount of memory was initially specified with the kmalloc call.
4498 * The caller must guarantee that objp points to a valid object previously
4499 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4500 * must not be freed during the duration of the call.
4502 size_t ksize(const void *objp
)
4505 if (unlikely(objp
== ZERO_SIZE_PTR
))
4508 return obj_size(virt_to_cache(objp
));
4510 EXPORT_SYMBOL(ksize
);