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 * Need this for bootstrapping a per node allocator.
310 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
311 struct kmem_list3 __initdata initkmem_list3
[NUM_INIT_LISTS
];
312 #define CACHE_CACHE 0
313 #define SIZE_AC MAX_NUMNODES
314 #define SIZE_L3 (2 * MAX_NUMNODES)
316 static int drain_freelist(struct kmem_cache
*cache
,
317 struct kmem_list3
*l3
, int tofree
);
318 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
320 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
);
321 static void cache_reap(struct work_struct
*unused
);
324 * This function must be completely optimized away if a constant is passed to
325 * it. Mostly the same as what is in linux/slab.h except it returns an index.
327 static __always_inline
int index_of(const size_t size
)
329 extern void __bad_size(void);
331 if (__builtin_constant_p(size
)) {
339 #include <linux/kmalloc_sizes.h>
347 static int slab_early_init
= 1;
349 #define INDEX_AC index_of(sizeof(struct arraycache_init))
350 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
352 static void kmem_list3_init(struct kmem_list3
*parent
)
354 INIT_LIST_HEAD(&parent
->slabs_full
);
355 INIT_LIST_HEAD(&parent
->slabs_partial
);
356 INIT_LIST_HEAD(&parent
->slabs_free
);
357 parent
->shared
= NULL
;
358 parent
->alien
= NULL
;
359 parent
->colour_next
= 0;
360 spin_lock_init(&parent
->list_lock
);
361 parent
->free_objects
= 0;
362 parent
->free_touched
= 0;
365 #define MAKE_LIST(cachep, listp, slab, nodeid) \
367 INIT_LIST_HEAD(listp); \
368 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
371 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
373 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
374 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
375 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
378 #define CFLGS_OFF_SLAB (0x80000000UL)
379 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
381 #define BATCHREFILL_LIMIT 16
383 * Optimization question: fewer reaps means less probability for unnessary
384 * cpucache drain/refill cycles.
386 * OTOH the cpuarrays can contain lots of objects,
387 * which could lock up otherwise freeable slabs.
389 #define REAPTIMEOUT_CPUC (2*HZ)
390 #define REAPTIMEOUT_LIST3 (4*HZ)
393 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
394 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
395 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
396 #define STATS_INC_GROWN(x) ((x)->grown++)
397 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
398 #define STATS_SET_HIGH(x) \
400 if ((x)->num_active > (x)->high_mark) \
401 (x)->high_mark = (x)->num_active; \
403 #define STATS_INC_ERR(x) ((x)->errors++)
404 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
405 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
406 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
407 #define STATS_SET_FREEABLE(x, i) \
409 if ((x)->max_freeable < i) \
410 (x)->max_freeable = i; \
412 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
413 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
414 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
415 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
417 #define STATS_INC_ACTIVE(x) do { } while (0)
418 #define STATS_DEC_ACTIVE(x) do { } while (0)
419 #define STATS_INC_ALLOCED(x) do { } while (0)
420 #define STATS_INC_GROWN(x) do { } while (0)
421 #define STATS_ADD_REAPED(x,y) do { } while (0)
422 #define STATS_SET_HIGH(x) do { } while (0)
423 #define STATS_INC_ERR(x) do { } while (0)
424 #define STATS_INC_NODEALLOCS(x) do { } while (0)
425 #define STATS_INC_NODEFREES(x) do { } while (0)
426 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
427 #define STATS_SET_FREEABLE(x, i) do { } while (0)
428 #define STATS_INC_ALLOCHIT(x) do { } while (0)
429 #define STATS_INC_ALLOCMISS(x) do { } while (0)
430 #define STATS_INC_FREEHIT(x) do { } while (0)
431 #define STATS_INC_FREEMISS(x) do { } while (0)
437 * memory layout of objects:
439 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
440 * the end of an object is aligned with the end of the real
441 * allocation. Catches writes behind the end of the allocation.
442 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
444 * cachep->obj_offset: The real object.
445 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
446 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
447 * [BYTES_PER_WORD long]
449 static int obj_offset(struct kmem_cache
*cachep
)
451 return cachep
->obj_offset
;
454 static int obj_size(struct kmem_cache
*cachep
)
456 return cachep
->obj_size
;
459 static unsigned long long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
461 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
462 return (unsigned long long*) (objp
+ obj_offset(cachep
) -
463 sizeof(unsigned long long));
466 static unsigned long long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
468 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
469 if (cachep
->flags
& SLAB_STORE_USER
)
470 return (unsigned long long *)(objp
+ cachep
->buffer_size
-
471 sizeof(unsigned long long) -
473 return (unsigned long long *) (objp
+ cachep
->buffer_size
-
474 sizeof(unsigned long long));
477 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
479 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
480 return (void **)(objp
+ cachep
->buffer_size
- BYTES_PER_WORD
);
485 #define obj_offset(x) 0
486 #define obj_size(cachep) (cachep->buffer_size)
487 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
488 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
489 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
493 #ifdef CONFIG_TRACING
494 size_t slab_buffer_size(struct kmem_cache
*cachep
)
496 return cachep
->buffer_size
;
498 EXPORT_SYMBOL(slab_buffer_size
);
502 * Do not go above this order unless 0 objects fit into the slab.
504 #define BREAK_GFP_ORDER_HI 1
505 #define BREAK_GFP_ORDER_LO 0
506 static int slab_break_gfp_order
= BREAK_GFP_ORDER_LO
;
509 * Functions for storing/retrieving the cachep and or slab from the page
510 * allocator. These are used to find the slab an obj belongs to. With kfree(),
511 * these are used to find the cache which an obj belongs to.
513 static inline void page_set_cache(struct page
*page
, struct kmem_cache
*cache
)
515 page
->lru
.next
= (struct list_head
*)cache
;
518 static inline struct kmem_cache
*page_get_cache(struct page
*page
)
520 page
= compound_head(page
);
521 BUG_ON(!PageSlab(page
));
522 return (struct kmem_cache
*)page
->lru
.next
;
525 static inline void page_set_slab(struct page
*page
, struct slab
*slab
)
527 page
->lru
.prev
= (struct list_head
*)slab
;
530 static inline struct slab
*page_get_slab(struct page
*page
)
532 BUG_ON(!PageSlab(page
));
533 return (struct slab
*)page
->lru
.prev
;
536 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
538 struct page
*page
= virt_to_head_page(obj
);
539 return page_get_cache(page
);
542 static inline struct slab
*virt_to_slab(const void *obj
)
544 struct page
*page
= virt_to_head_page(obj
);
545 return page_get_slab(page
);
548 static inline void *index_to_obj(struct kmem_cache
*cache
, struct slab
*slab
,
551 return slab
->s_mem
+ cache
->buffer_size
* idx
;
555 * We want to avoid an expensive divide : (offset / cache->buffer_size)
556 * Using the fact that buffer_size is a constant for a particular cache,
557 * we can replace (offset / cache->buffer_size) by
558 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
560 static inline unsigned int obj_to_index(const struct kmem_cache
*cache
,
561 const struct slab
*slab
, void *obj
)
563 u32 offset
= (obj
- slab
->s_mem
);
564 return reciprocal_divide(offset
, cache
->reciprocal_buffer_size
);
568 * These are the default caches for kmalloc. Custom caches can have other sizes.
570 struct cache_sizes malloc_sizes
[] = {
571 #define CACHE(x) { .cs_size = (x) },
572 #include <linux/kmalloc_sizes.h>
576 EXPORT_SYMBOL(malloc_sizes
);
578 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
584 static struct cache_names __initdata cache_names
[] = {
585 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
586 #include <linux/kmalloc_sizes.h>
591 static struct arraycache_init initarray_cache __initdata
=
592 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
593 static struct arraycache_init initarray_generic
=
594 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
596 /* internal cache of cache description objs */
597 static struct kmem_cache cache_cache
= {
599 .limit
= BOOT_CPUCACHE_ENTRIES
,
601 .buffer_size
= sizeof(struct kmem_cache
),
602 .name
= "kmem_cache",
605 #define BAD_ALIEN_MAGIC 0x01020304ul
608 * chicken and egg problem: delay the per-cpu array allocation
609 * until the general caches are up.
620 * used by boot code to determine if it can use slab based allocator
622 int slab_is_available(void)
624 return g_cpucache_up
>= EARLY
;
627 #ifdef CONFIG_LOCKDEP
630 * Slab sometimes uses the kmalloc slabs to store the slab headers
631 * for other slabs "off slab".
632 * The locking for this is tricky in that it nests within the locks
633 * of all other slabs in a few places; to deal with this special
634 * locking we put on-slab caches into a separate lock-class.
636 * We set lock class for alien array caches which are up during init.
637 * The lock annotation will be lost if all cpus of a node goes down and
638 * then comes back up during hotplug
640 static struct lock_class_key on_slab_l3_key
;
641 static struct lock_class_key on_slab_alc_key
;
643 static void init_node_lock_keys(int q
)
645 struct cache_sizes
*s
= malloc_sizes
;
647 if (g_cpucache_up
!= FULL
)
650 for (s
= malloc_sizes
; s
->cs_size
!= ULONG_MAX
; s
++) {
651 struct array_cache
**alc
;
652 struct kmem_list3
*l3
;
655 l3
= s
->cs_cachep
->nodelists
[q
];
656 if (!l3
|| OFF_SLAB(s
->cs_cachep
))
658 lockdep_set_class(&l3
->list_lock
, &on_slab_l3_key
);
661 * FIXME: This check for BAD_ALIEN_MAGIC
662 * should go away when common slab code is taught to
663 * work even without alien caches.
664 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
665 * for alloc_alien_cache,
667 if (!alc
|| (unsigned long)alc
== BAD_ALIEN_MAGIC
)
671 lockdep_set_class(&alc
[r
]->lock
,
677 static inline void init_lock_keys(void)
682 init_node_lock_keys(node
);
685 static void init_node_lock_keys(int q
)
689 static inline void init_lock_keys(void)
695 * Guard access to the cache-chain.
697 static DEFINE_MUTEX(cache_chain_mutex
);
698 static struct list_head cache_chain
;
700 static DEFINE_PER_CPU(struct delayed_work
, slab_reap_work
);
702 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
704 return cachep
->array
[smp_processor_id()];
707 static inline struct kmem_cache
*__find_general_cachep(size_t size
,
710 struct cache_sizes
*csizep
= malloc_sizes
;
713 /* This happens if someone tries to call
714 * kmem_cache_create(), or __kmalloc(), before
715 * the generic caches are initialized.
717 BUG_ON(malloc_sizes
[INDEX_AC
].cs_cachep
== NULL
);
720 return ZERO_SIZE_PTR
;
722 while (size
> csizep
->cs_size
)
726 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
727 * has cs_{dma,}cachep==NULL. Thus no special case
728 * for large kmalloc calls required.
730 #ifdef CONFIG_ZONE_DMA
731 if (unlikely(gfpflags
& GFP_DMA
))
732 return csizep
->cs_dmacachep
;
734 return csizep
->cs_cachep
;
737 static struct kmem_cache
*kmem_find_general_cachep(size_t size
, gfp_t gfpflags
)
739 return __find_general_cachep(size
, gfpflags
);
742 static size_t slab_mgmt_size(size_t nr_objs
, size_t align
)
744 return ALIGN(sizeof(struct slab
)+nr_objs
*sizeof(kmem_bufctl_t
), align
);
748 * Calculate the number of objects and left-over bytes for a given buffer size.
750 static void cache_estimate(unsigned long gfporder
, size_t buffer_size
,
751 size_t align
, int flags
, size_t *left_over
,
756 size_t slab_size
= PAGE_SIZE
<< gfporder
;
759 * The slab management structure can be either off the slab or
760 * on it. For the latter case, the memory allocated for a
764 * - One kmem_bufctl_t for each object
765 * - Padding to respect alignment of @align
766 * - @buffer_size bytes for each object
768 * If the slab management structure is off the slab, then the
769 * alignment will already be calculated into the size. Because
770 * the slabs are all pages aligned, the objects will be at the
771 * correct alignment when allocated.
773 if (flags
& CFLGS_OFF_SLAB
) {
775 nr_objs
= slab_size
/ buffer_size
;
777 if (nr_objs
> SLAB_LIMIT
)
778 nr_objs
= SLAB_LIMIT
;
781 * Ignore padding for the initial guess. The padding
782 * is at most @align-1 bytes, and @buffer_size is at
783 * least @align. In the worst case, this result will
784 * be one greater than the number of objects that fit
785 * into the memory allocation when taking the padding
788 nr_objs
= (slab_size
- sizeof(struct slab
)) /
789 (buffer_size
+ sizeof(kmem_bufctl_t
));
792 * This calculated number will be either the right
793 * amount, or one greater than what we want.
795 if (slab_mgmt_size(nr_objs
, align
) + nr_objs
*buffer_size
799 if (nr_objs
> SLAB_LIMIT
)
800 nr_objs
= SLAB_LIMIT
;
802 mgmt_size
= slab_mgmt_size(nr_objs
, align
);
805 *left_over
= slab_size
- nr_objs
*buffer_size
- mgmt_size
;
808 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
810 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
813 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
814 function
, cachep
->name
, msg
);
819 * By default on NUMA we use alien caches to stage the freeing of
820 * objects allocated from other nodes. This causes massive memory
821 * inefficiencies when using fake NUMA setup to split memory into a
822 * large number of small nodes, so it can be disabled on the command
826 static int use_alien_caches __read_mostly
= 1;
827 static int __init
noaliencache_setup(char *s
)
829 use_alien_caches
= 0;
832 __setup("noaliencache", noaliencache_setup
);
836 * Special reaping functions for NUMA systems called from cache_reap().
837 * These take care of doing round robin flushing of alien caches (containing
838 * objects freed on different nodes from which they were allocated) and the
839 * flushing of remote pcps by calling drain_node_pages.
841 static DEFINE_PER_CPU(unsigned long, slab_reap_node
);
843 static void init_reap_node(int cpu
)
847 node
= next_node(cpu_to_node(cpu
), node_online_map
);
848 if (node
== MAX_NUMNODES
)
849 node
= first_node(node_online_map
);
851 per_cpu(slab_reap_node
, cpu
) = node
;
854 static void next_reap_node(void)
856 int node
= __get_cpu_var(slab_reap_node
);
858 node
= next_node(node
, node_online_map
);
859 if (unlikely(node
>= MAX_NUMNODES
))
860 node
= first_node(node_online_map
);
861 __get_cpu_var(slab_reap_node
) = node
;
865 #define init_reap_node(cpu) do { } while (0)
866 #define next_reap_node(void) do { } while (0)
870 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
871 * via the workqueue/eventd.
872 * Add the CPU number into the expiration time to minimize the possibility of
873 * the CPUs getting into lockstep and contending for the global cache chain
876 static void __cpuinit
start_cpu_timer(int cpu
)
878 struct delayed_work
*reap_work
= &per_cpu(slab_reap_work
, cpu
);
881 * When this gets called from do_initcalls via cpucache_init(),
882 * init_workqueues() has already run, so keventd will be setup
885 if (keventd_up() && reap_work
->work
.func
== NULL
) {
887 INIT_DELAYED_WORK(reap_work
, cache_reap
);
888 schedule_delayed_work_on(cpu
, reap_work
,
889 __round_jiffies_relative(HZ
, cpu
));
893 static struct array_cache
*alloc_arraycache(int node
, int entries
,
894 int batchcount
, gfp_t gfp
)
896 int memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
897 struct array_cache
*nc
= NULL
;
899 nc
= kmalloc_node(memsize
, gfp
, node
);
901 * The array_cache structures contain pointers to free object.
902 * However, when such objects are allocated or transfered to another
903 * cache the pointers are not cleared and they could be counted as
904 * valid references during a kmemleak scan. Therefore, kmemleak must
905 * not scan such objects.
907 kmemleak_no_scan(nc
);
911 nc
->batchcount
= batchcount
;
913 spin_lock_init(&nc
->lock
);
919 * Transfer objects in one arraycache to another.
920 * Locking must be handled by the caller.
922 * Return the number of entries transferred.
924 static int transfer_objects(struct array_cache
*to
,
925 struct array_cache
*from
, unsigned int max
)
927 /* Figure out how many entries to transfer */
928 int nr
= min(min(from
->avail
, max
), to
->limit
- to
->avail
);
933 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
943 #define drain_alien_cache(cachep, alien) do { } while (0)
944 #define reap_alien(cachep, l3) do { } while (0)
946 static inline struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
948 return (struct array_cache
**)BAD_ALIEN_MAGIC
;
951 static inline void free_alien_cache(struct array_cache
**ac_ptr
)
955 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
960 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
966 static inline void *____cache_alloc_node(struct kmem_cache
*cachep
,
967 gfp_t flags
, int nodeid
)
972 #else /* CONFIG_NUMA */
974 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
975 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
977 static struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
979 struct array_cache
**ac_ptr
;
980 int memsize
= sizeof(void *) * nr_node_ids
;
985 ac_ptr
= kmalloc_node(memsize
, gfp
, node
);
988 if (i
== node
|| !node_online(i
)) {
992 ac_ptr
[i
] = alloc_arraycache(node
, limit
, 0xbaadf00d, gfp
);
994 for (i
--; i
>= 0; i
--)
1004 static void free_alien_cache(struct array_cache
**ac_ptr
)
1015 static void __drain_alien_cache(struct kmem_cache
*cachep
,
1016 struct array_cache
*ac
, int node
)
1018 struct kmem_list3
*rl3
= cachep
->nodelists
[node
];
1021 spin_lock(&rl3
->list_lock
);
1023 * Stuff objects into the remote nodes shared array first.
1024 * That way we could avoid the overhead of putting the objects
1025 * into the free lists and getting them back later.
1028 transfer_objects(rl3
->shared
, ac
, ac
->limit
);
1030 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
1032 spin_unlock(&rl3
->list_lock
);
1037 * Called from cache_reap() to regularly drain alien caches round robin.
1039 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_list3
*l3
)
1041 int node
= __get_cpu_var(slab_reap_node
);
1044 struct array_cache
*ac
= l3
->alien
[node
];
1046 if (ac
&& ac
->avail
&& spin_trylock_irq(&ac
->lock
)) {
1047 __drain_alien_cache(cachep
, ac
, node
);
1048 spin_unlock_irq(&ac
->lock
);
1053 static void drain_alien_cache(struct kmem_cache
*cachep
,
1054 struct array_cache
**alien
)
1057 struct array_cache
*ac
;
1058 unsigned long flags
;
1060 for_each_online_node(i
) {
1063 spin_lock_irqsave(&ac
->lock
, flags
);
1064 __drain_alien_cache(cachep
, ac
, i
);
1065 spin_unlock_irqrestore(&ac
->lock
, flags
);
1070 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1072 struct slab
*slabp
= virt_to_slab(objp
);
1073 int nodeid
= slabp
->nodeid
;
1074 struct kmem_list3
*l3
;
1075 struct array_cache
*alien
= NULL
;
1078 node
= numa_node_id();
1081 * Make sure we are not freeing a object from another node to the array
1082 * cache on this cpu.
1084 if (likely(slabp
->nodeid
== node
))
1087 l3
= cachep
->nodelists
[node
];
1088 STATS_INC_NODEFREES(cachep
);
1089 if (l3
->alien
&& l3
->alien
[nodeid
]) {
1090 alien
= l3
->alien
[nodeid
];
1091 spin_lock(&alien
->lock
);
1092 if (unlikely(alien
->avail
== alien
->limit
)) {
1093 STATS_INC_ACOVERFLOW(cachep
);
1094 __drain_alien_cache(cachep
, alien
, nodeid
);
1096 alien
->entry
[alien
->avail
++] = objp
;
1097 spin_unlock(&alien
->lock
);
1099 spin_lock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1100 free_block(cachep
, &objp
, 1, nodeid
);
1101 spin_unlock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1107 static void __cpuinit
cpuup_canceled(long cpu
)
1109 struct kmem_cache
*cachep
;
1110 struct kmem_list3
*l3
= NULL
;
1111 int node
= cpu_to_node(cpu
);
1112 const struct cpumask
*mask
= cpumask_of_node(node
);
1114 list_for_each_entry(cachep
, &cache_chain
, next
) {
1115 struct array_cache
*nc
;
1116 struct array_cache
*shared
;
1117 struct array_cache
**alien
;
1119 /* cpu is dead; no one can alloc from it. */
1120 nc
= cachep
->array
[cpu
];
1121 cachep
->array
[cpu
] = NULL
;
1122 l3
= cachep
->nodelists
[node
];
1125 goto free_array_cache
;
1127 spin_lock_irq(&l3
->list_lock
);
1129 /* Free limit for this kmem_list3 */
1130 l3
->free_limit
-= cachep
->batchcount
;
1132 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
1134 if (!cpumask_empty(mask
)) {
1135 spin_unlock_irq(&l3
->list_lock
);
1136 goto free_array_cache
;
1139 shared
= l3
->shared
;
1141 free_block(cachep
, shared
->entry
,
1142 shared
->avail
, node
);
1149 spin_unlock_irq(&l3
->list_lock
);
1153 drain_alien_cache(cachep
, alien
);
1154 free_alien_cache(alien
);
1160 * In the previous loop, all the objects were freed to
1161 * the respective cache's slabs, now we can go ahead and
1162 * shrink each nodelist to its limit.
1164 list_for_each_entry(cachep
, &cache_chain
, next
) {
1165 l3
= cachep
->nodelists
[node
];
1168 drain_freelist(cachep
, l3
, l3
->free_objects
);
1172 static int __cpuinit
cpuup_prepare(long cpu
)
1174 struct kmem_cache
*cachep
;
1175 struct kmem_list3
*l3
= NULL
;
1176 int node
= cpu_to_node(cpu
);
1177 const int memsize
= sizeof(struct kmem_list3
);
1180 * We need to do this right in the beginning since
1181 * alloc_arraycache's are going to use this list.
1182 * kmalloc_node allows us to add the slab to the right
1183 * kmem_list3 and not this cpu's kmem_list3
1186 list_for_each_entry(cachep
, &cache_chain
, next
) {
1188 * Set up the size64 kmemlist for cpu before we can
1189 * begin anything. Make sure some other cpu on this
1190 * node has not already allocated this
1192 if (!cachep
->nodelists
[node
]) {
1193 l3
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1196 kmem_list3_init(l3
);
1197 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
1198 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1201 * The l3s don't come and go as CPUs come and
1202 * go. cache_chain_mutex is sufficient
1205 cachep
->nodelists
[node
] = l3
;
1208 spin_lock_irq(&cachep
->nodelists
[node
]->list_lock
);
1209 cachep
->nodelists
[node
]->free_limit
=
1210 (1 + nr_cpus_node(node
)) *
1211 cachep
->batchcount
+ cachep
->num
;
1212 spin_unlock_irq(&cachep
->nodelists
[node
]->list_lock
);
1216 * Now we can go ahead with allocating the shared arrays and
1219 list_for_each_entry(cachep
, &cache_chain
, next
) {
1220 struct array_cache
*nc
;
1221 struct array_cache
*shared
= NULL
;
1222 struct array_cache
**alien
= NULL
;
1224 nc
= alloc_arraycache(node
, cachep
->limit
,
1225 cachep
->batchcount
, GFP_KERNEL
);
1228 if (cachep
->shared
) {
1229 shared
= alloc_arraycache(node
,
1230 cachep
->shared
* cachep
->batchcount
,
1231 0xbaadf00d, GFP_KERNEL
);
1237 if (use_alien_caches
) {
1238 alien
= alloc_alien_cache(node
, cachep
->limit
, GFP_KERNEL
);
1245 cachep
->array
[cpu
] = nc
;
1246 l3
= cachep
->nodelists
[node
];
1249 spin_lock_irq(&l3
->list_lock
);
1252 * We are serialised from CPU_DEAD or
1253 * CPU_UP_CANCELLED by the cpucontrol lock
1255 l3
->shared
= shared
;
1264 spin_unlock_irq(&l3
->list_lock
);
1266 free_alien_cache(alien
);
1268 init_node_lock_keys(node
);
1272 cpuup_canceled(cpu
);
1276 static int __cpuinit
cpuup_callback(struct notifier_block
*nfb
,
1277 unsigned long action
, void *hcpu
)
1279 long cpu
= (long)hcpu
;
1283 case CPU_UP_PREPARE
:
1284 case CPU_UP_PREPARE_FROZEN
:
1285 mutex_lock(&cache_chain_mutex
);
1286 err
= cpuup_prepare(cpu
);
1287 mutex_unlock(&cache_chain_mutex
);
1290 case CPU_ONLINE_FROZEN
:
1291 start_cpu_timer(cpu
);
1293 #ifdef CONFIG_HOTPLUG_CPU
1294 case CPU_DOWN_PREPARE
:
1295 case CPU_DOWN_PREPARE_FROZEN
:
1297 * Shutdown cache reaper. Note that the cache_chain_mutex is
1298 * held so that if cache_reap() is invoked it cannot do
1299 * anything expensive but will only modify reap_work
1300 * and reschedule the timer.
1302 cancel_rearming_delayed_work(&per_cpu(slab_reap_work
, cpu
));
1303 /* Now the cache_reaper is guaranteed to be not running. */
1304 per_cpu(slab_reap_work
, cpu
).work
.func
= NULL
;
1306 case CPU_DOWN_FAILED
:
1307 case CPU_DOWN_FAILED_FROZEN
:
1308 start_cpu_timer(cpu
);
1311 case CPU_DEAD_FROZEN
:
1313 * Even if all the cpus of a node are down, we don't free the
1314 * kmem_list3 of any cache. This to avoid a race between
1315 * cpu_down, and a kmalloc allocation from another cpu for
1316 * memory from the node of the cpu going down. The list3
1317 * structure is usually allocated from kmem_cache_create() and
1318 * gets destroyed at kmem_cache_destroy().
1322 case CPU_UP_CANCELED
:
1323 case CPU_UP_CANCELED_FROZEN
:
1324 mutex_lock(&cache_chain_mutex
);
1325 cpuup_canceled(cpu
);
1326 mutex_unlock(&cache_chain_mutex
);
1329 return err
? NOTIFY_BAD
: NOTIFY_OK
;
1332 static struct notifier_block __cpuinitdata cpucache_notifier
= {
1333 &cpuup_callback
, NULL
, 0
1337 * swap the static kmem_list3 with kmalloced memory
1339 static void init_list(struct kmem_cache
*cachep
, struct kmem_list3
*list
,
1342 struct kmem_list3
*ptr
;
1344 ptr
= kmalloc_node(sizeof(struct kmem_list3
), GFP_NOWAIT
, nodeid
);
1347 memcpy(ptr
, list
, sizeof(struct kmem_list3
));
1349 * Do not assume that spinlocks can be initialized via memcpy:
1351 spin_lock_init(&ptr
->list_lock
);
1353 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1354 cachep
->nodelists
[nodeid
] = ptr
;
1358 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1359 * size of kmem_list3.
1361 static void __init
set_up_list3s(struct kmem_cache
*cachep
, int index
)
1365 for_each_online_node(node
) {
1366 cachep
->nodelists
[node
] = &initkmem_list3
[index
+ node
];
1367 cachep
->nodelists
[node
]->next_reap
= jiffies
+
1369 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1374 * Initialisation. Called after the page allocator have been initialised and
1375 * before smp_init().
1377 void __init
kmem_cache_init(void)
1380 struct cache_sizes
*sizes
;
1381 struct cache_names
*names
;
1386 if (num_possible_nodes() == 1)
1387 use_alien_caches
= 0;
1389 for (i
= 0; i
< NUM_INIT_LISTS
; i
++) {
1390 kmem_list3_init(&initkmem_list3
[i
]);
1391 if (i
< MAX_NUMNODES
)
1392 cache_cache
.nodelists
[i
] = NULL
;
1394 set_up_list3s(&cache_cache
, CACHE_CACHE
);
1397 * Fragmentation resistance on low memory - only use bigger
1398 * page orders on machines with more than 32MB of memory.
1400 if (totalram_pages
> (32 << 20) >> PAGE_SHIFT
)
1401 slab_break_gfp_order
= BREAK_GFP_ORDER_HI
;
1403 /* Bootstrap is tricky, because several objects are allocated
1404 * from caches that do not exist yet:
1405 * 1) initialize the cache_cache cache: it contains the struct
1406 * kmem_cache structures of all caches, except cache_cache itself:
1407 * cache_cache is statically allocated.
1408 * Initially an __init data area is used for the head array and the
1409 * kmem_list3 structures, it's replaced with a kmalloc allocated
1410 * array at the end of the bootstrap.
1411 * 2) Create the first kmalloc cache.
1412 * The struct kmem_cache for the new cache is allocated normally.
1413 * An __init data area is used for the head array.
1414 * 3) Create the remaining kmalloc caches, with minimally sized
1416 * 4) Replace the __init data head arrays for cache_cache and the first
1417 * kmalloc cache with kmalloc allocated arrays.
1418 * 5) Replace the __init data for kmem_list3 for cache_cache and
1419 * the other cache's with kmalloc allocated memory.
1420 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1423 node
= numa_node_id();
1425 /* 1) create the cache_cache */
1426 INIT_LIST_HEAD(&cache_chain
);
1427 list_add(&cache_cache
.next
, &cache_chain
);
1428 cache_cache
.colour_off
= cache_line_size();
1429 cache_cache
.array
[smp_processor_id()] = &initarray_cache
.cache
;
1430 cache_cache
.nodelists
[node
] = &initkmem_list3
[CACHE_CACHE
+ node
];
1433 * struct kmem_cache size depends on nr_node_ids, which
1434 * can be less than MAX_NUMNODES.
1436 cache_cache
.buffer_size
= offsetof(struct kmem_cache
, nodelists
) +
1437 nr_node_ids
* sizeof(struct kmem_list3
*);
1439 cache_cache
.obj_size
= cache_cache
.buffer_size
;
1441 cache_cache
.buffer_size
= ALIGN(cache_cache
.buffer_size
,
1443 cache_cache
.reciprocal_buffer_size
=
1444 reciprocal_value(cache_cache
.buffer_size
);
1446 for (order
= 0; order
< MAX_ORDER
; order
++) {
1447 cache_estimate(order
, cache_cache
.buffer_size
,
1448 cache_line_size(), 0, &left_over
, &cache_cache
.num
);
1449 if (cache_cache
.num
)
1452 BUG_ON(!cache_cache
.num
);
1453 cache_cache
.gfporder
= order
;
1454 cache_cache
.colour
= left_over
/ cache_cache
.colour_off
;
1455 cache_cache
.slab_size
= ALIGN(cache_cache
.num
* sizeof(kmem_bufctl_t
) +
1456 sizeof(struct slab
), cache_line_size());
1458 /* 2+3) create the kmalloc caches */
1459 sizes
= malloc_sizes
;
1460 names
= cache_names
;
1463 * Initialize the caches that provide memory for the array cache and the
1464 * kmem_list3 structures first. Without this, further allocations will
1468 sizes
[INDEX_AC
].cs_cachep
= kmem_cache_create(names
[INDEX_AC
].name
,
1469 sizes
[INDEX_AC
].cs_size
,
1470 ARCH_KMALLOC_MINALIGN
,
1471 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1474 if (INDEX_AC
!= INDEX_L3
) {
1475 sizes
[INDEX_L3
].cs_cachep
=
1476 kmem_cache_create(names
[INDEX_L3
].name
,
1477 sizes
[INDEX_L3
].cs_size
,
1478 ARCH_KMALLOC_MINALIGN
,
1479 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1483 slab_early_init
= 0;
1485 while (sizes
->cs_size
!= ULONG_MAX
) {
1487 * For performance, all the general caches are L1 aligned.
1488 * This should be particularly beneficial on SMP boxes, as it
1489 * eliminates "false sharing".
1490 * Note for systems short on memory removing the alignment will
1491 * allow tighter packing of the smaller caches.
1493 if (!sizes
->cs_cachep
) {
1494 sizes
->cs_cachep
= kmem_cache_create(names
->name
,
1496 ARCH_KMALLOC_MINALIGN
,
1497 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1500 #ifdef CONFIG_ZONE_DMA
1501 sizes
->cs_dmacachep
= kmem_cache_create(
1504 ARCH_KMALLOC_MINALIGN
,
1505 ARCH_KMALLOC_FLAGS
|SLAB_CACHE_DMA
|
1512 /* 4) Replace the bootstrap head arrays */
1514 struct array_cache
*ptr
;
1516 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1518 BUG_ON(cpu_cache_get(&cache_cache
) != &initarray_cache
.cache
);
1519 memcpy(ptr
, cpu_cache_get(&cache_cache
),
1520 sizeof(struct arraycache_init
));
1522 * Do not assume that spinlocks can be initialized via memcpy:
1524 spin_lock_init(&ptr
->lock
);
1526 cache_cache
.array
[smp_processor_id()] = ptr
;
1528 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1530 BUG_ON(cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
)
1531 != &initarray_generic
.cache
);
1532 memcpy(ptr
, cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
),
1533 sizeof(struct arraycache_init
));
1535 * Do not assume that spinlocks can be initialized via memcpy:
1537 spin_lock_init(&ptr
->lock
);
1539 malloc_sizes
[INDEX_AC
].cs_cachep
->array
[smp_processor_id()] =
1542 /* 5) Replace the bootstrap kmem_list3's */
1546 for_each_online_node(nid
) {
1547 init_list(&cache_cache
, &initkmem_list3
[CACHE_CACHE
+ nid
], nid
);
1549 init_list(malloc_sizes
[INDEX_AC
].cs_cachep
,
1550 &initkmem_list3
[SIZE_AC
+ nid
], nid
);
1552 if (INDEX_AC
!= INDEX_L3
) {
1553 init_list(malloc_sizes
[INDEX_L3
].cs_cachep
,
1554 &initkmem_list3
[SIZE_L3
+ nid
], nid
);
1559 g_cpucache_up
= EARLY
;
1562 void __init
kmem_cache_init_late(void)
1564 struct kmem_cache
*cachep
;
1566 /* 6) resize the head arrays to their final sizes */
1567 mutex_lock(&cache_chain_mutex
);
1568 list_for_each_entry(cachep
, &cache_chain
, next
)
1569 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1571 mutex_unlock(&cache_chain_mutex
);
1574 g_cpucache_up
= FULL
;
1576 /* Annotate slab for lockdep -- annotate the malloc caches */
1580 * Register a cpu startup notifier callback that initializes
1581 * cpu_cache_get for all new cpus
1583 register_cpu_notifier(&cpucache_notifier
);
1586 * The reap timers are started later, with a module init call: That part
1587 * of the kernel is not yet operational.
1591 static int __init
cpucache_init(void)
1596 * Register the timers that return unneeded pages to the page allocator
1598 for_each_online_cpu(cpu
)
1599 start_cpu_timer(cpu
);
1602 __initcall(cpucache_init
);
1605 * Interface to system's page allocator. No need to hold the cache-lock.
1607 * If we requested dmaable memory, we will get it. Even if we
1608 * did not request dmaable memory, we might get it, but that
1609 * would be relatively rare and ignorable.
1611 static void *kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
1619 * Nommu uses slab's for process anonymous memory allocations, and thus
1620 * requires __GFP_COMP to properly refcount higher order allocations
1622 flags
|= __GFP_COMP
;
1625 flags
|= cachep
->gfpflags
;
1626 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1627 flags
|= __GFP_RECLAIMABLE
;
1629 page
= alloc_pages_exact_node(nodeid
, flags
| __GFP_NOTRACK
, cachep
->gfporder
);
1633 nr_pages
= (1 << cachep
->gfporder
);
1634 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1635 add_zone_page_state(page_zone(page
),
1636 NR_SLAB_RECLAIMABLE
, nr_pages
);
1638 add_zone_page_state(page_zone(page
),
1639 NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1640 for (i
= 0; i
< nr_pages
; i
++)
1641 __SetPageSlab(page
+ i
);
1643 if (kmemcheck_enabled
&& !(cachep
->flags
& SLAB_NOTRACK
)) {
1644 kmemcheck_alloc_shadow(page
, cachep
->gfporder
, flags
, nodeid
);
1647 kmemcheck_mark_uninitialized_pages(page
, nr_pages
);
1649 kmemcheck_mark_unallocated_pages(page
, nr_pages
);
1652 return page_address(page
);
1656 * Interface to system's page release.
1658 static void kmem_freepages(struct kmem_cache
*cachep
, void *addr
)
1660 unsigned long i
= (1 << cachep
->gfporder
);
1661 struct page
*page
= virt_to_page(addr
);
1662 const unsigned long nr_freed
= i
;
1664 kmemcheck_free_shadow(page
, cachep
->gfporder
);
1666 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1667 sub_zone_page_state(page_zone(page
),
1668 NR_SLAB_RECLAIMABLE
, nr_freed
);
1670 sub_zone_page_state(page_zone(page
),
1671 NR_SLAB_UNRECLAIMABLE
, nr_freed
);
1673 BUG_ON(!PageSlab(page
));
1674 __ClearPageSlab(page
);
1677 if (current
->reclaim_state
)
1678 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1679 free_pages((unsigned long)addr
, cachep
->gfporder
);
1682 static void kmem_rcu_free(struct rcu_head
*head
)
1684 struct slab_rcu
*slab_rcu
= (struct slab_rcu
*)head
;
1685 struct kmem_cache
*cachep
= slab_rcu
->cachep
;
1687 kmem_freepages(cachep
, slab_rcu
->addr
);
1688 if (OFF_SLAB(cachep
))
1689 kmem_cache_free(cachep
->slabp_cache
, slab_rcu
);
1694 #ifdef CONFIG_DEBUG_PAGEALLOC
1695 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1696 unsigned long caller
)
1698 int size
= obj_size(cachep
);
1700 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1702 if (size
< 5 * sizeof(unsigned long))
1705 *addr
++ = 0x12345678;
1707 *addr
++ = smp_processor_id();
1708 size
-= 3 * sizeof(unsigned long);
1710 unsigned long *sptr
= &caller
;
1711 unsigned long svalue
;
1713 while (!kstack_end(sptr
)) {
1715 if (kernel_text_address(svalue
)) {
1717 size
-= sizeof(unsigned long);
1718 if (size
<= sizeof(unsigned long))
1724 *addr
++ = 0x87654321;
1728 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1730 int size
= obj_size(cachep
);
1731 addr
= &((char *)addr
)[obj_offset(cachep
)];
1733 memset(addr
, val
, size
);
1734 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1737 static void dump_line(char *data
, int offset
, int limit
)
1740 unsigned char error
= 0;
1743 printk(KERN_ERR
"%03x:", offset
);
1744 for (i
= 0; i
< limit
; i
++) {
1745 if (data
[offset
+ i
] != POISON_FREE
) {
1746 error
= data
[offset
+ i
];
1749 printk(" %02x", (unsigned char)data
[offset
+ i
]);
1753 if (bad_count
== 1) {
1754 error
^= POISON_FREE
;
1755 if (!(error
& (error
- 1))) {
1756 printk(KERN_ERR
"Single bit error detected. Probably "
1759 printk(KERN_ERR
"Run memtest86+ or a similar memory "
1762 printk(KERN_ERR
"Run a memory test tool.\n");
1771 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1776 if (cachep
->flags
& SLAB_RED_ZONE
) {
1777 printk(KERN_ERR
"Redzone: 0x%llx/0x%llx.\n",
1778 *dbg_redzone1(cachep
, objp
),
1779 *dbg_redzone2(cachep
, objp
));
1782 if (cachep
->flags
& SLAB_STORE_USER
) {
1783 printk(KERN_ERR
"Last user: [<%p>]",
1784 *dbg_userword(cachep
, objp
));
1785 print_symbol("(%s)",
1786 (unsigned long)*dbg_userword(cachep
, objp
));
1789 realobj
= (char *)objp
+ obj_offset(cachep
);
1790 size
= obj_size(cachep
);
1791 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1794 if (i
+ limit
> size
)
1796 dump_line(realobj
, i
, limit
);
1800 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1806 realobj
= (char *)objp
+ obj_offset(cachep
);
1807 size
= obj_size(cachep
);
1809 for (i
= 0; i
< size
; i
++) {
1810 char exp
= POISON_FREE
;
1813 if (realobj
[i
] != exp
) {
1819 "Slab corruption: %s start=%p, len=%d\n",
1820 cachep
->name
, realobj
, size
);
1821 print_objinfo(cachep
, objp
, 0);
1823 /* Hexdump the affected line */
1826 if (i
+ limit
> size
)
1828 dump_line(realobj
, i
, limit
);
1831 /* Limit to 5 lines */
1837 /* Print some data about the neighboring objects, if they
1840 struct slab
*slabp
= virt_to_slab(objp
);
1843 objnr
= obj_to_index(cachep
, slabp
, objp
);
1845 objp
= index_to_obj(cachep
, slabp
, objnr
- 1);
1846 realobj
= (char *)objp
+ obj_offset(cachep
);
1847 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
1849 print_objinfo(cachep
, objp
, 2);
1851 if (objnr
+ 1 < cachep
->num
) {
1852 objp
= index_to_obj(cachep
, slabp
, objnr
+ 1);
1853 realobj
= (char *)objp
+ obj_offset(cachep
);
1854 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
1856 print_objinfo(cachep
, objp
, 2);
1863 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
, struct slab
*slabp
)
1866 for (i
= 0; i
< cachep
->num
; i
++) {
1867 void *objp
= index_to_obj(cachep
, slabp
, i
);
1869 if (cachep
->flags
& SLAB_POISON
) {
1870 #ifdef CONFIG_DEBUG_PAGEALLOC
1871 if (cachep
->buffer_size
% PAGE_SIZE
== 0 &&
1873 kernel_map_pages(virt_to_page(objp
),
1874 cachep
->buffer_size
/ PAGE_SIZE
, 1);
1876 check_poison_obj(cachep
, objp
);
1878 check_poison_obj(cachep
, objp
);
1881 if (cachep
->flags
& SLAB_RED_ZONE
) {
1882 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1883 slab_error(cachep
, "start of a freed object "
1885 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1886 slab_error(cachep
, "end of a freed object "
1892 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
, struct slab
*slabp
)
1898 * slab_destroy - destroy and release all objects in a slab
1899 * @cachep: cache pointer being destroyed
1900 * @slabp: slab pointer being destroyed
1902 * Destroy all the objs in a slab, and release the mem back to the system.
1903 * Before calling the slab must have been unlinked from the cache. The
1904 * cache-lock is not held/needed.
1906 static void slab_destroy(struct kmem_cache
*cachep
, struct slab
*slabp
)
1908 void *addr
= slabp
->s_mem
- slabp
->colouroff
;
1910 slab_destroy_debugcheck(cachep
, slabp
);
1911 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
1912 struct slab_rcu
*slab_rcu
;
1914 slab_rcu
= (struct slab_rcu
*)slabp
;
1915 slab_rcu
->cachep
= cachep
;
1916 slab_rcu
->addr
= addr
;
1917 call_rcu(&slab_rcu
->head
, kmem_rcu_free
);
1919 kmem_freepages(cachep
, addr
);
1920 if (OFF_SLAB(cachep
))
1921 kmem_cache_free(cachep
->slabp_cache
, slabp
);
1925 static void __kmem_cache_destroy(struct kmem_cache
*cachep
)
1928 struct kmem_list3
*l3
;
1930 for_each_online_cpu(i
)
1931 kfree(cachep
->array
[i
]);
1933 /* NUMA: free the list3 structures */
1934 for_each_online_node(i
) {
1935 l3
= cachep
->nodelists
[i
];
1938 free_alien_cache(l3
->alien
);
1942 kmem_cache_free(&cache_cache
, cachep
);
1947 * calculate_slab_order - calculate size (page order) of slabs
1948 * @cachep: pointer to the cache that is being created
1949 * @size: size of objects to be created in this cache.
1950 * @align: required alignment for the objects.
1951 * @flags: slab allocation flags
1953 * Also calculates the number of objects per slab.
1955 * This could be made much more intelligent. For now, try to avoid using
1956 * high order pages for slabs. When the gfp() functions are more friendly
1957 * towards high-order requests, this should be changed.
1959 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
1960 size_t size
, size_t align
, unsigned long flags
)
1962 unsigned long offslab_limit
;
1963 size_t left_over
= 0;
1966 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
1970 cache_estimate(gfporder
, size
, align
, flags
, &remainder
, &num
);
1974 if (flags
& CFLGS_OFF_SLAB
) {
1976 * Max number of objs-per-slab for caches which
1977 * use off-slab slabs. Needed to avoid a possible
1978 * looping condition in cache_grow().
1980 offslab_limit
= size
- sizeof(struct slab
);
1981 offslab_limit
/= sizeof(kmem_bufctl_t
);
1983 if (num
> offslab_limit
)
1987 /* Found something acceptable - save it away */
1989 cachep
->gfporder
= gfporder
;
1990 left_over
= remainder
;
1993 * A VFS-reclaimable slab tends to have most allocations
1994 * as GFP_NOFS and we really don't want to have to be allocating
1995 * higher-order pages when we are unable to shrink dcache.
1997 if (flags
& SLAB_RECLAIM_ACCOUNT
)
2001 * Large number of objects is good, but very large slabs are
2002 * currently bad for the gfp()s.
2004 if (gfporder
>= slab_break_gfp_order
)
2008 * Acceptable internal fragmentation?
2010 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
2016 static int __init_refok
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
2018 if (g_cpucache_up
== FULL
)
2019 return enable_cpucache(cachep
, gfp
);
2021 if (g_cpucache_up
== NONE
) {
2023 * Note: the first kmem_cache_create must create the cache
2024 * that's used by kmalloc(24), otherwise the creation of
2025 * further caches will BUG().
2027 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
2030 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2031 * the first cache, then we need to set up all its list3s,
2032 * otherwise the creation of further caches will BUG().
2034 set_up_list3s(cachep
, SIZE_AC
);
2035 if (INDEX_AC
== INDEX_L3
)
2036 g_cpucache_up
= PARTIAL_L3
;
2038 g_cpucache_up
= PARTIAL_AC
;
2040 cachep
->array
[smp_processor_id()] =
2041 kmalloc(sizeof(struct arraycache_init
), gfp
);
2043 if (g_cpucache_up
== PARTIAL_AC
) {
2044 set_up_list3s(cachep
, SIZE_L3
);
2045 g_cpucache_up
= PARTIAL_L3
;
2048 for_each_online_node(node
) {
2049 cachep
->nodelists
[node
] =
2050 kmalloc_node(sizeof(struct kmem_list3
),
2052 BUG_ON(!cachep
->nodelists
[node
]);
2053 kmem_list3_init(cachep
->nodelists
[node
]);
2057 cachep
->nodelists
[numa_node_id()]->next_reap
=
2058 jiffies
+ REAPTIMEOUT_LIST3
+
2059 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
2061 cpu_cache_get(cachep
)->avail
= 0;
2062 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
2063 cpu_cache_get(cachep
)->batchcount
= 1;
2064 cpu_cache_get(cachep
)->touched
= 0;
2065 cachep
->batchcount
= 1;
2066 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
2071 * kmem_cache_create - Create a cache.
2072 * @name: A string which is used in /proc/slabinfo to identify this cache.
2073 * @size: The size of objects to be created in this cache.
2074 * @align: The required alignment for the objects.
2075 * @flags: SLAB flags
2076 * @ctor: A constructor for the objects.
2078 * Returns a ptr to the cache on success, NULL on failure.
2079 * Cannot be called within a int, but can be interrupted.
2080 * The @ctor is run when new pages are allocated by the cache.
2082 * @name must be valid until the cache is destroyed. This implies that
2083 * the module calling this has to destroy the cache before getting unloaded.
2084 * Note that kmem_cache_name() is not guaranteed to return the same pointer,
2085 * therefore applications must manage it themselves.
2089 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2090 * to catch references to uninitialised memory.
2092 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2093 * for buffer overruns.
2095 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2096 * cacheline. This can be beneficial if you're counting cycles as closely
2100 kmem_cache_create (const char *name
, size_t size
, size_t align
,
2101 unsigned long flags
, void (*ctor
)(void *))
2103 size_t left_over
, slab_size
, ralign
;
2104 struct kmem_cache
*cachep
= NULL
, *pc
;
2108 * Sanity checks... these are all serious usage bugs.
2110 if (!name
|| in_interrupt() || (size
< BYTES_PER_WORD
) ||
2111 size
> KMALLOC_MAX_SIZE
) {
2112 printk(KERN_ERR
"%s: Early error in slab %s\n", __func__
,
2118 * We use cache_chain_mutex to ensure a consistent view of
2119 * cpu_online_mask as well. Please see cpuup_callback
2121 if (slab_is_available()) {
2123 mutex_lock(&cache_chain_mutex
);
2126 list_for_each_entry(pc
, &cache_chain
, next
) {
2131 * This happens when the module gets unloaded and doesn't
2132 * destroy its slab cache and no-one else reuses the vmalloc
2133 * area of the module. Print a warning.
2135 res
= probe_kernel_address(pc
->name
, tmp
);
2138 "SLAB: cache with size %d has lost its name\n",
2143 if (!strcmp(pc
->name
, name
)) {
2145 "kmem_cache_create: duplicate cache %s\n", name
);
2152 WARN_ON(strchr(name
, ' ')); /* It confuses parsers */
2155 * Enable redzoning and last user accounting, except for caches with
2156 * large objects, if the increased size would increase the object size
2157 * above the next power of two: caches with object sizes just above a
2158 * power of two have a significant amount of internal fragmentation.
2160 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
2161 2 * sizeof(unsigned long long)))
2162 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2163 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2164 flags
|= SLAB_POISON
;
2166 if (flags
& SLAB_DESTROY_BY_RCU
)
2167 BUG_ON(flags
& SLAB_POISON
);
2170 * Always checks flags, a caller might be expecting debug support which
2173 BUG_ON(flags
& ~CREATE_MASK
);
2176 * Check that size is in terms of words. This is needed to avoid
2177 * unaligned accesses for some archs when redzoning is used, and makes
2178 * sure any on-slab bufctl's are also correctly aligned.
2180 if (size
& (BYTES_PER_WORD
- 1)) {
2181 size
+= (BYTES_PER_WORD
- 1);
2182 size
&= ~(BYTES_PER_WORD
- 1);
2185 /* calculate the final buffer alignment: */
2187 /* 1) arch recommendation: can be overridden for debug */
2188 if (flags
& SLAB_HWCACHE_ALIGN
) {
2190 * Default alignment: as specified by the arch code. Except if
2191 * an object is really small, then squeeze multiple objects into
2194 ralign
= cache_line_size();
2195 while (size
<= ralign
/ 2)
2198 ralign
= BYTES_PER_WORD
;
2202 * Redzoning and user store require word alignment or possibly larger.
2203 * Note this will be overridden by architecture or caller mandated
2204 * alignment if either is greater than BYTES_PER_WORD.
2206 if (flags
& SLAB_STORE_USER
)
2207 ralign
= BYTES_PER_WORD
;
2209 if (flags
& SLAB_RED_ZONE
) {
2210 ralign
= REDZONE_ALIGN
;
2211 /* If redzoning, ensure that the second redzone is suitably
2212 * aligned, by adjusting the object size accordingly. */
2213 size
+= REDZONE_ALIGN
- 1;
2214 size
&= ~(REDZONE_ALIGN
- 1);
2217 /* 2) arch mandated alignment */
2218 if (ralign
< ARCH_SLAB_MINALIGN
) {
2219 ralign
= ARCH_SLAB_MINALIGN
;
2221 /* 3) caller mandated alignment */
2222 if (ralign
< align
) {
2225 /* disable debug if necessary */
2226 if (ralign
> __alignof__(unsigned long long))
2227 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2233 if (slab_is_available())
2238 /* Get cache's description obj. */
2239 cachep
= kmem_cache_zalloc(&cache_cache
, gfp
);
2244 cachep
->obj_size
= size
;
2247 * Both debugging options require word-alignment which is calculated
2250 if (flags
& SLAB_RED_ZONE
) {
2251 /* add space for red zone words */
2252 cachep
->obj_offset
+= sizeof(unsigned long long);
2253 size
+= 2 * sizeof(unsigned long long);
2255 if (flags
& SLAB_STORE_USER
) {
2256 /* user store requires one word storage behind the end of
2257 * the real object. But if the second red zone needs to be
2258 * aligned to 64 bits, we must allow that much space.
2260 if (flags
& SLAB_RED_ZONE
)
2261 size
+= REDZONE_ALIGN
;
2263 size
+= BYTES_PER_WORD
;
2265 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2266 if (size
>= malloc_sizes
[INDEX_L3
+ 1].cs_size
2267 && cachep
->obj_size
> cache_line_size() && size
< PAGE_SIZE
) {
2268 cachep
->obj_offset
+= PAGE_SIZE
- size
;
2275 * Determine if the slab management is 'on' or 'off' slab.
2276 * (bootstrapping cannot cope with offslab caches so don't do
2277 * it too early on. Always use on-slab management when
2278 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2280 if ((size
>= (PAGE_SIZE
>> 3)) && !slab_early_init
&&
2281 !(flags
& SLAB_NOLEAKTRACE
))
2283 * Size is large, assume best to place the slab management obj
2284 * off-slab (should allow better packing of objs).
2286 flags
|= CFLGS_OFF_SLAB
;
2288 size
= ALIGN(size
, align
);
2290 left_over
= calculate_slab_order(cachep
, size
, align
, flags
);
2294 "kmem_cache_create: couldn't create cache %s.\n", name
);
2295 kmem_cache_free(&cache_cache
, cachep
);
2299 slab_size
= ALIGN(cachep
->num
* sizeof(kmem_bufctl_t
)
2300 + sizeof(struct slab
), align
);
2303 * If the slab has been placed off-slab, and we have enough space then
2304 * move it on-slab. This is at the expense of any extra colouring.
2306 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= slab_size
) {
2307 flags
&= ~CFLGS_OFF_SLAB
;
2308 left_over
-= slab_size
;
2311 if (flags
& CFLGS_OFF_SLAB
) {
2312 /* really off slab. No need for manual alignment */
2314 cachep
->num
* sizeof(kmem_bufctl_t
) + sizeof(struct slab
);
2316 #ifdef CONFIG_PAGE_POISONING
2317 /* If we're going to use the generic kernel_map_pages()
2318 * poisoning, then it's going to smash the contents of
2319 * the redzone and userword anyhow, so switch them off.
2321 if (size
% PAGE_SIZE
== 0 && flags
& SLAB_POISON
)
2322 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2326 cachep
->colour_off
= cache_line_size();
2327 /* Offset must be a multiple of the alignment. */
2328 if (cachep
->colour_off
< align
)
2329 cachep
->colour_off
= align
;
2330 cachep
->colour
= left_over
/ cachep
->colour_off
;
2331 cachep
->slab_size
= slab_size
;
2332 cachep
->flags
= flags
;
2333 cachep
->gfpflags
= 0;
2334 if (CONFIG_ZONE_DMA_FLAG
&& (flags
& SLAB_CACHE_DMA
))
2335 cachep
->gfpflags
|= GFP_DMA
;
2336 cachep
->buffer_size
= size
;
2337 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2339 if (flags
& CFLGS_OFF_SLAB
) {
2340 cachep
->slabp_cache
= kmem_find_general_cachep(slab_size
, 0u);
2342 * This is a possibility for one of the malloc_sizes caches.
2343 * But since we go off slab only for object size greater than
2344 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2345 * this should not happen at all.
2346 * But leave a BUG_ON for some lucky dude.
2348 BUG_ON(ZERO_OR_NULL_PTR(cachep
->slabp_cache
));
2350 cachep
->ctor
= ctor
;
2351 cachep
->name
= name
;
2353 if (setup_cpu_cache(cachep
, gfp
)) {
2354 __kmem_cache_destroy(cachep
);
2359 /* cache setup completed, link it into the list */
2360 list_add(&cachep
->next
, &cache_chain
);
2362 if (!cachep
&& (flags
& SLAB_PANIC
))
2363 panic("kmem_cache_create(): failed to create slab `%s'\n",
2365 if (slab_is_available()) {
2366 mutex_unlock(&cache_chain_mutex
);
2371 EXPORT_SYMBOL(kmem_cache_create
);
2374 static void check_irq_off(void)
2376 BUG_ON(!irqs_disabled());
2379 static void check_irq_on(void)
2381 BUG_ON(irqs_disabled());
2384 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2388 assert_spin_locked(&cachep
->nodelists
[numa_node_id()]->list_lock
);
2392 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2396 assert_spin_locked(&cachep
->nodelists
[node
]->list_lock
);
2401 #define check_irq_off() do { } while(0)
2402 #define check_irq_on() do { } while(0)
2403 #define check_spinlock_acquired(x) do { } while(0)
2404 #define check_spinlock_acquired_node(x, y) do { } while(0)
2407 static void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
2408 struct array_cache
*ac
,
2409 int force
, int node
);
2411 static void do_drain(void *arg
)
2413 struct kmem_cache
*cachep
= arg
;
2414 struct array_cache
*ac
;
2415 int node
= numa_node_id();
2418 ac
= cpu_cache_get(cachep
);
2419 spin_lock(&cachep
->nodelists
[node
]->list_lock
);
2420 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
2421 spin_unlock(&cachep
->nodelists
[node
]->list_lock
);
2425 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2427 struct kmem_list3
*l3
;
2430 on_each_cpu(do_drain
, cachep
, 1);
2432 for_each_online_node(node
) {
2433 l3
= cachep
->nodelists
[node
];
2434 if (l3
&& l3
->alien
)
2435 drain_alien_cache(cachep
, l3
->alien
);
2438 for_each_online_node(node
) {
2439 l3
= cachep
->nodelists
[node
];
2441 drain_array(cachep
, l3
, l3
->shared
, 1, node
);
2446 * Remove slabs from the list of free slabs.
2447 * Specify the number of slabs to drain in tofree.
2449 * Returns the actual number of slabs released.
2451 static int drain_freelist(struct kmem_cache
*cache
,
2452 struct kmem_list3
*l3
, int tofree
)
2454 struct list_head
*p
;
2459 while (nr_freed
< tofree
&& !list_empty(&l3
->slabs_free
)) {
2461 spin_lock_irq(&l3
->list_lock
);
2462 p
= l3
->slabs_free
.prev
;
2463 if (p
== &l3
->slabs_free
) {
2464 spin_unlock_irq(&l3
->list_lock
);
2468 slabp
= list_entry(p
, struct slab
, list
);
2470 BUG_ON(slabp
->inuse
);
2472 list_del(&slabp
->list
);
2474 * Safe to drop the lock. The slab is no longer linked
2477 l3
->free_objects
-= cache
->num
;
2478 spin_unlock_irq(&l3
->list_lock
);
2479 slab_destroy(cache
, slabp
);
2486 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2487 static int __cache_shrink(struct kmem_cache
*cachep
)
2490 struct kmem_list3
*l3
;
2492 drain_cpu_caches(cachep
);
2495 for_each_online_node(i
) {
2496 l3
= cachep
->nodelists
[i
];
2500 drain_freelist(cachep
, l3
, l3
->free_objects
);
2502 ret
+= !list_empty(&l3
->slabs_full
) ||
2503 !list_empty(&l3
->slabs_partial
);
2505 return (ret
? 1 : 0);
2509 * kmem_cache_shrink - Shrink a cache.
2510 * @cachep: The cache to shrink.
2512 * Releases as many slabs as possible for a cache.
2513 * To help debugging, a zero exit status indicates all slabs were released.
2515 int kmem_cache_shrink(struct kmem_cache
*cachep
)
2518 BUG_ON(!cachep
|| in_interrupt());
2521 mutex_lock(&cache_chain_mutex
);
2522 ret
= __cache_shrink(cachep
);
2523 mutex_unlock(&cache_chain_mutex
);
2527 EXPORT_SYMBOL(kmem_cache_shrink
);
2530 * kmem_cache_destroy - delete a cache
2531 * @cachep: the cache to destroy
2533 * Remove a &struct kmem_cache object from the slab cache.
2535 * It is expected this function will be called by a module when it is
2536 * unloaded. This will remove the cache completely, and avoid a duplicate
2537 * cache being allocated each time a module is loaded and unloaded, if the
2538 * module doesn't have persistent in-kernel storage across loads and unloads.
2540 * The cache must be empty before calling this function.
2542 * The caller must guarantee that noone will allocate memory from the cache
2543 * during the kmem_cache_destroy().
2545 void kmem_cache_destroy(struct kmem_cache
*cachep
)
2547 BUG_ON(!cachep
|| in_interrupt());
2549 /* Find the cache in the chain of caches. */
2551 mutex_lock(&cache_chain_mutex
);
2553 * the chain is never empty, cache_cache is never destroyed
2555 list_del(&cachep
->next
);
2556 if (__cache_shrink(cachep
)) {
2557 slab_error(cachep
, "Can't free all objects");
2558 list_add(&cachep
->next
, &cache_chain
);
2559 mutex_unlock(&cache_chain_mutex
);
2564 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
2567 __kmem_cache_destroy(cachep
);
2568 mutex_unlock(&cache_chain_mutex
);
2571 EXPORT_SYMBOL(kmem_cache_destroy
);
2574 * Get the memory for a slab management obj.
2575 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2576 * always come from malloc_sizes caches. The slab descriptor cannot
2577 * come from the same cache which is getting created because,
2578 * when we are searching for an appropriate cache for these
2579 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2580 * If we are creating a malloc_sizes cache here it would not be visible to
2581 * kmem_find_general_cachep till the initialization is complete.
2582 * Hence we cannot have slabp_cache same as the original cache.
2584 static struct slab
*alloc_slabmgmt(struct kmem_cache
*cachep
, void *objp
,
2585 int colour_off
, gfp_t local_flags
,
2590 if (OFF_SLAB(cachep
)) {
2591 /* Slab management obj is off-slab. */
2592 slabp
= kmem_cache_alloc_node(cachep
->slabp_cache
,
2593 local_flags
, nodeid
);
2595 * If the first object in the slab is leaked (it's allocated
2596 * but no one has a reference to it), we want to make sure
2597 * kmemleak does not treat the ->s_mem pointer as a reference
2598 * to the object. Otherwise we will not report the leak.
2600 kmemleak_scan_area(&slabp
->list
, sizeof(struct list_head
),
2605 slabp
= objp
+ colour_off
;
2606 colour_off
+= cachep
->slab_size
;
2609 slabp
->colouroff
= colour_off
;
2610 slabp
->s_mem
= objp
+ colour_off
;
2611 slabp
->nodeid
= nodeid
;
2616 static inline kmem_bufctl_t
*slab_bufctl(struct slab
*slabp
)
2618 return (kmem_bufctl_t
*) (slabp
+ 1);
2621 static void cache_init_objs(struct kmem_cache
*cachep
,
2626 for (i
= 0; i
< cachep
->num
; i
++) {
2627 void *objp
= index_to_obj(cachep
, slabp
, i
);
2629 /* need to poison the objs? */
2630 if (cachep
->flags
& SLAB_POISON
)
2631 poison_obj(cachep
, objp
, POISON_FREE
);
2632 if (cachep
->flags
& SLAB_STORE_USER
)
2633 *dbg_userword(cachep
, objp
) = NULL
;
2635 if (cachep
->flags
& SLAB_RED_ZONE
) {
2636 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2637 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2640 * Constructors are not allowed to allocate memory from the same
2641 * cache which they are a constructor for. Otherwise, deadlock.
2642 * They must also be threaded.
2644 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2645 cachep
->ctor(objp
+ obj_offset(cachep
));
2647 if (cachep
->flags
& SLAB_RED_ZONE
) {
2648 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2649 slab_error(cachep
, "constructor overwrote the"
2650 " end of an object");
2651 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2652 slab_error(cachep
, "constructor overwrote the"
2653 " start of an object");
2655 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 &&
2656 OFF_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
)
2657 kernel_map_pages(virt_to_page(objp
),
2658 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2663 slab_bufctl(slabp
)[i
] = i
+ 1;
2665 slab_bufctl(slabp
)[i
- 1] = BUFCTL_END
;
2668 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2670 if (CONFIG_ZONE_DMA_FLAG
) {
2671 if (flags
& GFP_DMA
)
2672 BUG_ON(!(cachep
->gfpflags
& GFP_DMA
));
2674 BUG_ON(cachep
->gfpflags
& GFP_DMA
);
2678 static void *slab_get_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2681 void *objp
= index_to_obj(cachep
, slabp
, slabp
->free
);
2685 next
= slab_bufctl(slabp
)[slabp
->free
];
2687 slab_bufctl(slabp
)[slabp
->free
] = BUFCTL_FREE
;
2688 WARN_ON(slabp
->nodeid
!= nodeid
);
2695 static void slab_put_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2696 void *objp
, int nodeid
)
2698 unsigned int objnr
= obj_to_index(cachep
, slabp
, objp
);
2701 /* Verify that the slab belongs to the intended node */
2702 WARN_ON(slabp
->nodeid
!= nodeid
);
2704 if (slab_bufctl(slabp
)[objnr
] + 1 <= SLAB_LIMIT
+ 1) {
2705 printk(KERN_ERR
"slab: double free detected in cache "
2706 "'%s', objp %p\n", cachep
->name
, objp
);
2710 slab_bufctl(slabp
)[objnr
] = slabp
->free
;
2711 slabp
->free
= objnr
;
2716 * Map pages beginning at addr to the given cache and slab. This is required
2717 * for the slab allocator to be able to lookup the cache and slab of a
2718 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2720 static void slab_map_pages(struct kmem_cache
*cache
, struct slab
*slab
,
2726 page
= virt_to_page(addr
);
2729 if (likely(!PageCompound(page
)))
2730 nr_pages
<<= cache
->gfporder
;
2733 page_set_cache(page
, cache
);
2734 page_set_slab(page
, slab
);
2736 } while (--nr_pages
);
2740 * Grow (by 1) the number of slabs within a cache. This is called by
2741 * kmem_cache_alloc() when there are no active objs left in a cache.
2743 static int cache_grow(struct kmem_cache
*cachep
,
2744 gfp_t flags
, int nodeid
, void *objp
)
2749 struct kmem_list3
*l3
;
2752 * Be lazy and only check for valid flags here, keeping it out of the
2753 * critical path in kmem_cache_alloc().
2755 BUG_ON(flags
& GFP_SLAB_BUG_MASK
);
2756 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2758 /* Take the l3 list lock to change the colour_next on this node */
2760 l3
= cachep
->nodelists
[nodeid
];
2761 spin_lock(&l3
->list_lock
);
2763 /* Get colour for the slab, and cal the next value. */
2764 offset
= l3
->colour_next
;
2766 if (l3
->colour_next
>= cachep
->colour
)
2767 l3
->colour_next
= 0;
2768 spin_unlock(&l3
->list_lock
);
2770 offset
*= cachep
->colour_off
;
2772 if (local_flags
& __GFP_WAIT
)
2776 * The test for missing atomic flag is performed here, rather than
2777 * the more obvious place, simply to reduce the critical path length
2778 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2779 * will eventually be caught here (where it matters).
2781 kmem_flagcheck(cachep
, flags
);
2784 * Get mem for the objs. Attempt to allocate a physical page from
2788 objp
= kmem_getpages(cachep
, local_flags
, nodeid
);
2792 /* Get slab management. */
2793 slabp
= alloc_slabmgmt(cachep
, objp
, offset
,
2794 local_flags
& ~GFP_CONSTRAINT_MASK
, nodeid
);
2798 slab_map_pages(cachep
, slabp
, objp
);
2800 cache_init_objs(cachep
, slabp
);
2802 if (local_flags
& __GFP_WAIT
)
2803 local_irq_disable();
2805 spin_lock(&l3
->list_lock
);
2807 /* Make slab active. */
2808 list_add_tail(&slabp
->list
, &(l3
->slabs_free
));
2809 STATS_INC_GROWN(cachep
);
2810 l3
->free_objects
+= cachep
->num
;
2811 spin_unlock(&l3
->list_lock
);
2814 kmem_freepages(cachep
, objp
);
2816 if (local_flags
& __GFP_WAIT
)
2817 local_irq_disable();
2824 * Perform extra freeing checks:
2825 * - detect bad pointers.
2826 * - POISON/RED_ZONE checking
2828 static void kfree_debugcheck(const void *objp
)
2830 if (!virt_addr_valid(objp
)) {
2831 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
2832 (unsigned long)objp
);
2837 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2839 unsigned long long redzone1
, redzone2
;
2841 redzone1
= *dbg_redzone1(cache
, obj
);
2842 redzone2
= *dbg_redzone2(cache
, obj
);
2847 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2850 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2851 slab_error(cache
, "double free detected");
2853 slab_error(cache
, "memory outside object was overwritten");
2855 printk(KERN_ERR
"%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2856 obj
, redzone1
, redzone2
);
2859 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2866 BUG_ON(virt_to_cache(objp
) != cachep
);
2868 objp
-= obj_offset(cachep
);
2869 kfree_debugcheck(objp
);
2870 page
= virt_to_head_page(objp
);
2872 slabp
= page_get_slab(page
);
2874 if (cachep
->flags
& SLAB_RED_ZONE
) {
2875 verify_redzone_free(cachep
, objp
);
2876 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2877 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2879 if (cachep
->flags
& SLAB_STORE_USER
)
2880 *dbg_userword(cachep
, objp
) = caller
;
2882 objnr
= obj_to_index(cachep
, slabp
, objp
);
2884 BUG_ON(objnr
>= cachep
->num
);
2885 BUG_ON(objp
!= index_to_obj(cachep
, slabp
, objnr
));
2887 #ifdef CONFIG_DEBUG_SLAB_LEAK
2888 slab_bufctl(slabp
)[objnr
] = BUFCTL_FREE
;
2890 if (cachep
->flags
& SLAB_POISON
) {
2891 #ifdef CONFIG_DEBUG_PAGEALLOC
2892 if ((cachep
->buffer_size
% PAGE_SIZE
)==0 && OFF_SLAB(cachep
)) {
2893 store_stackinfo(cachep
, objp
, (unsigned long)caller
);
2894 kernel_map_pages(virt_to_page(objp
),
2895 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2897 poison_obj(cachep
, objp
, POISON_FREE
);
2900 poison_obj(cachep
, objp
, POISON_FREE
);
2906 static void check_slabp(struct kmem_cache
*cachep
, struct slab
*slabp
)
2911 /* Check slab's freelist to see if this obj is there. */
2912 for (i
= slabp
->free
; i
!= BUFCTL_END
; i
= slab_bufctl(slabp
)[i
]) {
2914 if (entries
> cachep
->num
|| i
>= cachep
->num
)
2917 if (entries
!= cachep
->num
- slabp
->inuse
) {
2919 printk(KERN_ERR
"slab: Internal list corruption detected in "
2920 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2921 cachep
->name
, cachep
->num
, slabp
, slabp
->inuse
);
2923 i
< sizeof(*slabp
) + cachep
->num
* sizeof(kmem_bufctl_t
);
2926 printk("\n%03x:", i
);
2927 printk(" %02x", ((unsigned char *)slabp
)[i
]);
2934 #define kfree_debugcheck(x) do { } while(0)
2935 #define cache_free_debugcheck(x,objp,z) (objp)
2936 #define check_slabp(x,y) do { } while(0)
2939 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
2942 struct kmem_list3
*l3
;
2943 struct array_cache
*ac
;
2948 node
= numa_node_id();
2949 ac
= cpu_cache_get(cachep
);
2950 batchcount
= ac
->batchcount
;
2951 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2953 * If there was little recent activity on this cache, then
2954 * perform only a partial refill. Otherwise we could generate
2957 batchcount
= BATCHREFILL_LIMIT
;
2959 l3
= cachep
->nodelists
[node
];
2961 BUG_ON(ac
->avail
> 0 || !l3
);
2962 spin_lock(&l3
->list_lock
);
2964 /* See if we can refill from the shared array */
2965 if (l3
->shared
&& transfer_objects(ac
, l3
->shared
, batchcount
)) {
2966 l3
->shared
->touched
= 1;
2970 while (batchcount
> 0) {
2971 struct list_head
*entry
;
2973 /* Get slab alloc is to come from. */
2974 entry
= l3
->slabs_partial
.next
;
2975 if (entry
== &l3
->slabs_partial
) {
2976 l3
->free_touched
= 1;
2977 entry
= l3
->slabs_free
.next
;
2978 if (entry
== &l3
->slabs_free
)
2982 slabp
= list_entry(entry
, struct slab
, list
);
2983 check_slabp(cachep
, slabp
);
2984 check_spinlock_acquired(cachep
);
2987 * The slab was either on partial or free list so
2988 * there must be at least one object available for
2991 BUG_ON(slabp
->inuse
>= cachep
->num
);
2993 while (slabp
->inuse
< cachep
->num
&& batchcount
--) {
2994 STATS_INC_ALLOCED(cachep
);
2995 STATS_INC_ACTIVE(cachep
);
2996 STATS_SET_HIGH(cachep
);
2998 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, slabp
,
3001 check_slabp(cachep
, slabp
);
3003 /* move slabp to correct slabp list: */
3004 list_del(&slabp
->list
);
3005 if (slabp
->free
== BUFCTL_END
)
3006 list_add(&slabp
->list
, &l3
->slabs_full
);
3008 list_add(&slabp
->list
, &l3
->slabs_partial
);
3012 l3
->free_objects
-= ac
->avail
;
3014 spin_unlock(&l3
->list_lock
);
3016 if (unlikely(!ac
->avail
)) {
3018 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, node
, NULL
);
3020 /* cache_grow can reenable interrupts, then ac could change. */
3021 ac
= cpu_cache_get(cachep
);
3022 if (!x
&& ac
->avail
== 0) /* no objects in sight? abort */
3025 if (!ac
->avail
) /* objects refilled by interrupt? */
3029 return ac
->entry
[--ac
->avail
];
3032 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
3035 might_sleep_if(flags
& __GFP_WAIT
);
3037 kmem_flagcheck(cachep
, flags
);
3042 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
3043 gfp_t flags
, void *objp
, void *caller
)
3047 if (cachep
->flags
& SLAB_POISON
) {
3048 #ifdef CONFIG_DEBUG_PAGEALLOC
3049 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
3050 kernel_map_pages(virt_to_page(objp
),
3051 cachep
->buffer_size
/ PAGE_SIZE
, 1);
3053 check_poison_obj(cachep
, objp
);
3055 check_poison_obj(cachep
, objp
);
3057 poison_obj(cachep
, objp
, POISON_INUSE
);
3059 if (cachep
->flags
& SLAB_STORE_USER
)
3060 *dbg_userword(cachep
, objp
) = caller
;
3062 if (cachep
->flags
& SLAB_RED_ZONE
) {
3063 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
3064 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
3065 slab_error(cachep
, "double free, or memory outside"
3066 " object was overwritten");
3068 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3069 objp
, *dbg_redzone1(cachep
, objp
),
3070 *dbg_redzone2(cachep
, objp
));
3072 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
3073 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
3075 #ifdef CONFIG_DEBUG_SLAB_LEAK
3080 slabp
= page_get_slab(virt_to_head_page(objp
));
3081 objnr
= (unsigned)(objp
- slabp
->s_mem
) / cachep
->buffer_size
;
3082 slab_bufctl(slabp
)[objnr
] = BUFCTL_ACTIVE
;
3085 objp
+= obj_offset(cachep
);
3086 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
3088 #if ARCH_SLAB_MINALIGN
3089 if ((u32
)objp
& (ARCH_SLAB_MINALIGN
-1)) {
3090 printk(KERN_ERR
"0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3091 objp
, ARCH_SLAB_MINALIGN
);
3097 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3100 static bool slab_should_failslab(struct kmem_cache
*cachep
, gfp_t flags
)
3102 if (cachep
== &cache_cache
)
3105 return should_failslab(obj_size(cachep
), flags
);
3108 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3111 struct array_cache
*ac
;
3115 ac
= cpu_cache_get(cachep
);
3116 if (likely(ac
->avail
)) {
3117 STATS_INC_ALLOCHIT(cachep
);
3119 objp
= ac
->entry
[--ac
->avail
];
3121 STATS_INC_ALLOCMISS(cachep
);
3122 objp
= cache_alloc_refill(cachep
, flags
);
3124 * the 'ac' may be updated by cache_alloc_refill(),
3125 * and kmemleak_erase() requires its correct value.
3127 ac
= cpu_cache_get(cachep
);
3130 * To avoid a false negative, if an object that is in one of the
3131 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3132 * treat the array pointers as a reference to the object.
3135 kmemleak_erase(&ac
->entry
[ac
->avail
]);
3141 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3143 * If we are in_interrupt, then process context, including cpusets and
3144 * mempolicy, may not apply and should not be used for allocation policy.
3146 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3148 int nid_alloc
, nid_here
;
3150 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3152 nid_alloc
= nid_here
= numa_node_id();
3153 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3154 nid_alloc
= cpuset_mem_spread_node();
3155 else if (current
->mempolicy
)
3156 nid_alloc
= slab_node(current
->mempolicy
);
3157 if (nid_alloc
!= nid_here
)
3158 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3163 * Fallback function if there was no memory available and no objects on a
3164 * certain node and fall back is permitted. First we scan all the
3165 * available nodelists for available objects. If that fails then we
3166 * perform an allocation without specifying a node. This allows the page
3167 * allocator to do its reclaim / fallback magic. We then insert the
3168 * slab into the proper nodelist and then allocate from it.
3170 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3172 struct zonelist
*zonelist
;
3176 enum zone_type high_zoneidx
= gfp_zone(flags
);
3180 if (flags
& __GFP_THISNODE
)
3183 zonelist
= node_zonelist(slab_node(current
->mempolicy
), flags
);
3184 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
3188 * Look through allowed nodes for objects available
3189 * from existing per node queues.
3191 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
3192 nid
= zone_to_nid(zone
);
3194 if (cpuset_zone_allowed_hardwall(zone
, flags
) &&
3195 cache
->nodelists
[nid
] &&
3196 cache
->nodelists
[nid
]->free_objects
) {
3197 obj
= ____cache_alloc_node(cache
,
3198 flags
| GFP_THISNODE
, nid
);
3206 * This allocation will be performed within the constraints
3207 * of the current cpuset / memory policy requirements.
3208 * We may trigger various forms of reclaim on the allowed
3209 * set and go into memory reserves if necessary.
3211 if (local_flags
& __GFP_WAIT
)
3213 kmem_flagcheck(cache
, flags
);
3214 obj
= kmem_getpages(cache
, local_flags
, numa_node_id());
3215 if (local_flags
& __GFP_WAIT
)
3216 local_irq_disable();
3219 * Insert into the appropriate per node queues
3221 nid
= page_to_nid(virt_to_page(obj
));
3222 if (cache_grow(cache
, flags
, nid
, obj
)) {
3223 obj
= ____cache_alloc_node(cache
,
3224 flags
| GFP_THISNODE
, nid
);
3227 * Another processor may allocate the
3228 * objects in the slab since we are
3229 * not holding any locks.
3233 /* cache_grow already freed obj */
3242 * A interface to enable slab creation on nodeid
3244 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3247 struct list_head
*entry
;
3249 struct kmem_list3
*l3
;
3253 l3
= cachep
->nodelists
[nodeid
];
3258 spin_lock(&l3
->list_lock
);
3259 entry
= l3
->slabs_partial
.next
;
3260 if (entry
== &l3
->slabs_partial
) {
3261 l3
->free_touched
= 1;
3262 entry
= l3
->slabs_free
.next
;
3263 if (entry
== &l3
->slabs_free
)
3267 slabp
= list_entry(entry
, struct slab
, list
);
3268 check_spinlock_acquired_node(cachep
, nodeid
);
3269 check_slabp(cachep
, slabp
);
3271 STATS_INC_NODEALLOCS(cachep
);
3272 STATS_INC_ACTIVE(cachep
);
3273 STATS_SET_HIGH(cachep
);
3275 BUG_ON(slabp
->inuse
== cachep
->num
);
3277 obj
= slab_get_obj(cachep
, slabp
, nodeid
);
3278 check_slabp(cachep
, slabp
);
3280 /* move slabp to correct slabp list: */
3281 list_del(&slabp
->list
);
3283 if (slabp
->free
== BUFCTL_END
)
3284 list_add(&slabp
->list
, &l3
->slabs_full
);
3286 list_add(&slabp
->list
, &l3
->slabs_partial
);
3288 spin_unlock(&l3
->list_lock
);
3292 spin_unlock(&l3
->list_lock
);
3293 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, nodeid
, NULL
);
3297 return fallback_alloc(cachep
, flags
);
3304 * kmem_cache_alloc_node - Allocate an object on the specified node
3305 * @cachep: The cache to allocate from.
3306 * @flags: See kmalloc().
3307 * @nodeid: node number of the target node.
3308 * @caller: return address of caller, used for debug information
3310 * Identical to kmem_cache_alloc but it will allocate memory on the given
3311 * node, which can improve the performance for cpu bound structures.
3313 * Fallback to other node is possible if __GFP_THISNODE is not set.
3315 static __always_inline
void *
3316 __cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3319 unsigned long save_flags
;
3322 flags
&= gfp_allowed_mask
;
3324 lockdep_trace_alloc(flags
);
3326 if (slab_should_failslab(cachep
, flags
))
3329 cache_alloc_debugcheck_before(cachep
, flags
);
3330 local_irq_save(save_flags
);
3333 nodeid
= numa_node_id();
3335 if (unlikely(!cachep
->nodelists
[nodeid
])) {
3336 /* Node not bootstrapped yet */
3337 ptr
= fallback_alloc(cachep
, flags
);
3341 if (nodeid
== numa_node_id()) {
3343 * Use the locally cached objects if possible.
3344 * However ____cache_alloc does not allow fallback
3345 * to other nodes. It may fail while we still have
3346 * objects on other nodes available.
3348 ptr
= ____cache_alloc(cachep
, flags
);
3352 /* ___cache_alloc_node can fall back to other nodes */
3353 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3355 local_irq_restore(save_flags
);
3356 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3357 kmemleak_alloc_recursive(ptr
, obj_size(cachep
), 1, cachep
->flags
,
3361 kmemcheck_slab_alloc(cachep
, flags
, ptr
, obj_size(cachep
));
3363 if (unlikely((flags
& __GFP_ZERO
) && ptr
))
3364 memset(ptr
, 0, obj_size(cachep
));
3369 static __always_inline
void *
3370 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3374 if (unlikely(current
->flags
& (PF_SPREAD_SLAB
| PF_MEMPOLICY
))) {
3375 objp
= alternate_node_alloc(cache
, flags
);
3379 objp
= ____cache_alloc(cache
, flags
);
3382 * We may just have run out of memory on the local node.
3383 * ____cache_alloc_node() knows how to locate memory on other nodes
3386 objp
= ____cache_alloc_node(cache
, flags
, numa_node_id());
3393 static __always_inline
void *
3394 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3396 return ____cache_alloc(cachep
, flags
);
3399 #endif /* CONFIG_NUMA */
3401 static __always_inline
void *
3402 __cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
, void *caller
)
3404 unsigned long save_flags
;
3407 flags
&= gfp_allowed_mask
;
3409 lockdep_trace_alloc(flags
);
3411 if (slab_should_failslab(cachep
, flags
))
3414 cache_alloc_debugcheck_before(cachep
, flags
);
3415 local_irq_save(save_flags
);
3416 objp
= __do_cache_alloc(cachep
, flags
);
3417 local_irq_restore(save_flags
);
3418 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3419 kmemleak_alloc_recursive(objp
, obj_size(cachep
), 1, cachep
->flags
,
3424 kmemcheck_slab_alloc(cachep
, flags
, objp
, obj_size(cachep
));
3426 if (unlikely((flags
& __GFP_ZERO
) && objp
))
3427 memset(objp
, 0, obj_size(cachep
));
3433 * Caller needs to acquire correct kmem_list's list_lock
3435 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int nr_objects
,
3439 struct kmem_list3
*l3
;
3441 for (i
= 0; i
< nr_objects
; i
++) {
3442 void *objp
= objpp
[i
];
3445 slabp
= virt_to_slab(objp
);
3446 l3
= cachep
->nodelists
[node
];
3447 list_del(&slabp
->list
);
3448 check_spinlock_acquired_node(cachep
, node
);
3449 check_slabp(cachep
, slabp
);
3450 slab_put_obj(cachep
, slabp
, objp
, node
);
3451 STATS_DEC_ACTIVE(cachep
);
3453 check_slabp(cachep
, slabp
);
3455 /* fixup slab chains */
3456 if (slabp
->inuse
== 0) {
3457 if (l3
->free_objects
> l3
->free_limit
) {
3458 l3
->free_objects
-= cachep
->num
;
3459 /* No need to drop any previously held
3460 * lock here, even if we have a off-slab slab
3461 * descriptor it is guaranteed to come from
3462 * a different cache, refer to comments before
3465 slab_destroy(cachep
, slabp
);
3467 list_add(&slabp
->list
, &l3
->slabs_free
);
3470 /* Unconditionally move a slab to the end of the
3471 * partial list on free - maximum time for the
3472 * other objects to be freed, too.
3474 list_add_tail(&slabp
->list
, &l3
->slabs_partial
);
3479 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3482 struct kmem_list3
*l3
;
3483 int node
= numa_node_id();
3485 batchcount
= ac
->batchcount
;
3487 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
3490 l3
= cachep
->nodelists
[node
];
3491 spin_lock(&l3
->list_lock
);
3493 struct array_cache
*shared_array
= l3
->shared
;
3494 int max
= shared_array
->limit
- shared_array
->avail
;
3496 if (batchcount
> max
)
3498 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3499 ac
->entry
, sizeof(void *) * batchcount
);
3500 shared_array
->avail
+= batchcount
;
3505 free_block(cachep
, ac
->entry
, batchcount
, node
);
3510 struct list_head
*p
;
3512 p
= l3
->slabs_free
.next
;
3513 while (p
!= &(l3
->slabs_free
)) {
3516 slabp
= list_entry(p
, struct slab
, list
);
3517 BUG_ON(slabp
->inuse
);
3522 STATS_SET_FREEABLE(cachep
, i
);
3525 spin_unlock(&l3
->list_lock
);
3526 ac
->avail
-= batchcount
;
3527 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3531 * Release an obj back to its cache. If the obj has a constructed state, it must
3532 * be in this state _before_ it is released. Called with disabled ints.
3534 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
)
3536 struct array_cache
*ac
= cpu_cache_get(cachep
);
3539 kmemleak_free_recursive(objp
, cachep
->flags
);
3540 objp
= cache_free_debugcheck(cachep
, objp
, __builtin_return_address(0));
3542 kmemcheck_slab_free(cachep
, objp
, obj_size(cachep
));
3545 * Skip calling cache_free_alien() when the platform is not numa.
3546 * This will avoid cache misses that happen while accessing slabp (which
3547 * is per page memory reference) to get nodeid. Instead use a global
3548 * variable to skip the call, which is mostly likely to be present in
3551 if (nr_online_nodes
> 1 && cache_free_alien(cachep
, objp
))
3554 if (likely(ac
->avail
< ac
->limit
)) {
3555 STATS_INC_FREEHIT(cachep
);
3556 ac
->entry
[ac
->avail
++] = objp
;
3559 STATS_INC_FREEMISS(cachep
);
3560 cache_flusharray(cachep
, ac
);
3561 ac
->entry
[ac
->avail
++] = objp
;
3566 * kmem_cache_alloc - Allocate an object
3567 * @cachep: The cache to allocate from.
3568 * @flags: See kmalloc().
3570 * Allocate an object from this cache. The flags are only relevant
3571 * if the cache has no available objects.
3573 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3575 void *ret
= __cache_alloc(cachep
, flags
, __builtin_return_address(0));
3577 trace_kmem_cache_alloc(_RET_IP_
, ret
,
3578 obj_size(cachep
), cachep
->buffer_size
, flags
);
3582 EXPORT_SYMBOL(kmem_cache_alloc
);
3584 #ifdef CONFIG_TRACING
3585 void *kmem_cache_alloc_notrace(struct kmem_cache
*cachep
, gfp_t flags
)
3587 return __cache_alloc(cachep
, flags
, __builtin_return_address(0));
3589 EXPORT_SYMBOL(kmem_cache_alloc_notrace
);
3593 * kmem_ptr_validate - check if an untrusted pointer might be a slab entry.
3594 * @cachep: the cache we're checking against
3595 * @ptr: pointer to validate
3597 * This verifies that the untrusted pointer looks sane;
3598 * it is _not_ a guarantee that the pointer is actually
3599 * part of the slab cache in question, but it at least
3600 * validates that the pointer can be dereferenced and
3601 * looks half-way sane.
3603 * Currently only used for dentry validation.
3605 int kmem_ptr_validate(struct kmem_cache
*cachep
, const void *ptr
)
3607 unsigned long addr
= (unsigned long)ptr
;
3608 unsigned long min_addr
= PAGE_OFFSET
;
3609 unsigned long align_mask
= BYTES_PER_WORD
- 1;
3610 unsigned long size
= cachep
->buffer_size
;
3613 if (unlikely(addr
< min_addr
))
3615 if (unlikely(addr
> (unsigned long)high_memory
- size
))
3617 if (unlikely(addr
& align_mask
))
3619 if (unlikely(!kern_addr_valid(addr
)))
3621 if (unlikely(!kern_addr_valid(addr
+ size
- 1)))
3623 page
= virt_to_page(ptr
);
3624 if (unlikely(!PageSlab(page
)))
3626 if (unlikely(page_get_cache(page
) != cachep
))
3634 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3636 void *ret
= __cache_alloc_node(cachep
, flags
, nodeid
,
3637 __builtin_return_address(0));
3639 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
3640 obj_size(cachep
), cachep
->buffer_size
,
3645 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3647 #ifdef CONFIG_TRACING
3648 void *kmem_cache_alloc_node_notrace(struct kmem_cache
*cachep
,
3652 return __cache_alloc_node(cachep
, flags
, nodeid
,
3653 __builtin_return_address(0));
3655 EXPORT_SYMBOL(kmem_cache_alloc_node_notrace
);
3658 static __always_inline
void *
3659 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, void *caller
)
3661 struct kmem_cache
*cachep
;
3664 cachep
= kmem_find_general_cachep(size
, flags
);
3665 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3667 ret
= kmem_cache_alloc_node_notrace(cachep
, flags
, node
);
3669 trace_kmalloc_node((unsigned long) caller
, ret
,
3670 size
, cachep
->buffer_size
, flags
, node
);
3675 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3676 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3678 return __do_kmalloc_node(size
, flags
, node
,
3679 __builtin_return_address(0));
3681 EXPORT_SYMBOL(__kmalloc_node
);
3683 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3684 int node
, unsigned long caller
)
3686 return __do_kmalloc_node(size
, flags
, node
, (void *)caller
);
3688 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3690 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3692 return __do_kmalloc_node(size
, flags
, node
, NULL
);
3694 EXPORT_SYMBOL(__kmalloc_node
);
3695 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3696 #endif /* CONFIG_NUMA */
3699 * __do_kmalloc - allocate memory
3700 * @size: how many bytes of memory are required.
3701 * @flags: the type of memory to allocate (see kmalloc).
3702 * @caller: function caller for debug tracking of the caller
3704 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3707 struct kmem_cache
*cachep
;
3710 /* If you want to save a few bytes .text space: replace
3712 * Then kmalloc uses the uninlined functions instead of the inline
3715 cachep
= __find_general_cachep(size
, flags
);
3716 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3718 ret
= __cache_alloc(cachep
, flags
, caller
);
3720 trace_kmalloc((unsigned long) caller
, ret
,
3721 size
, cachep
->buffer_size
, flags
);
3727 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3728 void *__kmalloc(size_t size
, gfp_t flags
)
3730 return __do_kmalloc(size
, flags
, __builtin_return_address(0));
3732 EXPORT_SYMBOL(__kmalloc
);
3734 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, unsigned long caller
)
3736 return __do_kmalloc(size
, flags
, (void *)caller
);
3738 EXPORT_SYMBOL(__kmalloc_track_caller
);
3741 void *__kmalloc(size_t size
, gfp_t flags
)
3743 return __do_kmalloc(size
, flags
, NULL
);
3745 EXPORT_SYMBOL(__kmalloc
);
3749 * kmem_cache_free - Deallocate an object
3750 * @cachep: The cache the allocation was from.
3751 * @objp: The previously allocated object.
3753 * Free an object which was previously allocated from this
3756 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3758 unsigned long flags
;
3760 local_irq_save(flags
);
3761 debug_check_no_locks_freed(objp
, obj_size(cachep
));
3762 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3763 debug_check_no_obj_freed(objp
, obj_size(cachep
));
3764 __cache_free(cachep
, objp
);
3765 local_irq_restore(flags
);
3767 trace_kmem_cache_free(_RET_IP_
, objp
);
3769 EXPORT_SYMBOL(kmem_cache_free
);
3772 * kfree - free previously allocated memory
3773 * @objp: pointer returned by kmalloc.
3775 * If @objp is NULL, no operation is performed.
3777 * Don't free memory not originally allocated by kmalloc()
3778 * or you will run into trouble.
3780 void kfree(const void *objp
)
3782 struct kmem_cache
*c
;
3783 unsigned long flags
;
3785 trace_kfree(_RET_IP_
, objp
);
3787 if (unlikely(ZERO_OR_NULL_PTR(objp
)))
3789 local_irq_save(flags
);
3790 kfree_debugcheck(objp
);
3791 c
= virt_to_cache(objp
);
3792 debug_check_no_locks_freed(objp
, obj_size(c
));
3793 debug_check_no_obj_freed(objp
, obj_size(c
));
3794 __cache_free(c
, (void *)objp
);
3795 local_irq_restore(flags
);
3797 EXPORT_SYMBOL(kfree
);
3799 unsigned int kmem_cache_size(struct kmem_cache
*cachep
)
3801 return obj_size(cachep
);
3803 EXPORT_SYMBOL(kmem_cache_size
);
3805 const char *kmem_cache_name(struct kmem_cache
*cachep
)
3807 return cachep
->name
;
3809 EXPORT_SYMBOL_GPL(kmem_cache_name
);
3812 * This initializes kmem_list3 or resizes various caches for all nodes.
3814 static int alloc_kmemlist(struct kmem_cache
*cachep
, gfp_t gfp
)
3817 struct kmem_list3
*l3
;
3818 struct array_cache
*new_shared
;
3819 struct array_cache
**new_alien
= NULL
;
3821 for_each_online_node(node
) {
3823 if (use_alien_caches
) {
3824 new_alien
= alloc_alien_cache(node
, cachep
->limit
, gfp
);
3830 if (cachep
->shared
) {
3831 new_shared
= alloc_arraycache(node
,
3832 cachep
->shared
*cachep
->batchcount
,
3835 free_alien_cache(new_alien
);
3840 l3
= cachep
->nodelists
[node
];
3842 struct array_cache
*shared
= l3
->shared
;
3844 spin_lock_irq(&l3
->list_lock
);
3847 free_block(cachep
, shared
->entry
,
3848 shared
->avail
, node
);
3850 l3
->shared
= new_shared
;
3852 l3
->alien
= new_alien
;
3855 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3856 cachep
->batchcount
+ cachep
->num
;
3857 spin_unlock_irq(&l3
->list_lock
);
3859 free_alien_cache(new_alien
);
3862 l3
= kmalloc_node(sizeof(struct kmem_list3
), gfp
, node
);
3864 free_alien_cache(new_alien
);
3869 kmem_list3_init(l3
);
3870 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
3871 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
3872 l3
->shared
= new_shared
;
3873 l3
->alien
= new_alien
;
3874 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3875 cachep
->batchcount
+ cachep
->num
;
3876 cachep
->nodelists
[node
] = l3
;
3881 if (!cachep
->next
.next
) {
3882 /* Cache is not active yet. Roll back what we did */
3885 if (cachep
->nodelists
[node
]) {
3886 l3
= cachep
->nodelists
[node
];
3889 free_alien_cache(l3
->alien
);
3891 cachep
->nodelists
[node
] = NULL
;
3899 struct ccupdate_struct
{
3900 struct kmem_cache
*cachep
;
3901 struct array_cache
*new[NR_CPUS
];
3904 static void do_ccupdate_local(void *info
)
3906 struct ccupdate_struct
*new = info
;
3907 struct array_cache
*old
;
3910 old
= cpu_cache_get(new->cachep
);
3912 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
3913 new->new[smp_processor_id()] = old
;
3916 /* Always called with the cache_chain_mutex held */
3917 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3918 int batchcount
, int shared
, gfp_t gfp
)
3920 struct ccupdate_struct
*new;
3923 new = kzalloc(sizeof(*new), gfp
);
3927 for_each_online_cpu(i
) {
3928 new->new[i
] = alloc_arraycache(cpu_to_node(i
), limit
,
3931 for (i
--; i
>= 0; i
--)
3937 new->cachep
= cachep
;
3939 on_each_cpu(do_ccupdate_local
, (void *)new, 1);
3942 cachep
->batchcount
= batchcount
;
3943 cachep
->limit
= limit
;
3944 cachep
->shared
= shared
;
3946 for_each_online_cpu(i
) {
3947 struct array_cache
*ccold
= new->new[i
];
3950 spin_lock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3951 free_block(cachep
, ccold
->entry
, ccold
->avail
, cpu_to_node(i
));
3952 spin_unlock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3956 return alloc_kmemlist(cachep
, gfp
);
3959 /* Called with cache_chain_mutex held always */
3960 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
3966 * The head array serves three purposes:
3967 * - create a LIFO ordering, i.e. return objects that are cache-warm
3968 * - reduce the number of spinlock operations.
3969 * - reduce the number of linked list operations on the slab and
3970 * bufctl chains: array operations are cheaper.
3971 * The numbers are guessed, we should auto-tune as described by
3974 if (cachep
->buffer_size
> 131072)
3976 else if (cachep
->buffer_size
> PAGE_SIZE
)
3978 else if (cachep
->buffer_size
> 1024)
3980 else if (cachep
->buffer_size
> 256)
3986 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3987 * allocation behaviour: Most allocs on one cpu, most free operations
3988 * on another cpu. For these cases, an efficient object passing between
3989 * cpus is necessary. This is provided by a shared array. The array
3990 * replaces Bonwick's magazine layer.
3991 * On uniprocessor, it's functionally equivalent (but less efficient)
3992 * to a larger limit. Thus disabled by default.
3995 if (cachep
->buffer_size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
4000 * With debugging enabled, large batchcount lead to excessively long
4001 * periods with disabled local interrupts. Limit the batchcount
4006 err
= do_tune_cpucache(cachep
, limit
, (limit
+ 1) / 2, shared
, gfp
);
4008 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
4009 cachep
->name
, -err
);
4014 * Drain an array if it contains any elements taking the l3 lock only if
4015 * necessary. Note that the l3 listlock also protects the array_cache
4016 * if drain_array() is used on the shared array.
4018 void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
4019 struct array_cache
*ac
, int force
, int node
)
4023 if (!ac
|| !ac
->avail
)
4025 if (ac
->touched
&& !force
) {
4028 spin_lock_irq(&l3
->list_lock
);
4030 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
4031 if (tofree
> ac
->avail
)
4032 tofree
= (ac
->avail
+ 1) / 2;
4033 free_block(cachep
, ac
->entry
, tofree
, node
);
4034 ac
->avail
-= tofree
;
4035 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
4036 sizeof(void *) * ac
->avail
);
4038 spin_unlock_irq(&l3
->list_lock
);
4043 * cache_reap - Reclaim memory from caches.
4044 * @w: work descriptor
4046 * Called from workqueue/eventd every few seconds.
4048 * - clear the per-cpu caches for this CPU.
4049 * - return freeable pages to the main free memory pool.
4051 * If we cannot acquire the cache chain mutex then just give up - we'll try
4052 * again on the next iteration.
4054 static void cache_reap(struct work_struct
*w
)
4056 struct kmem_cache
*searchp
;
4057 struct kmem_list3
*l3
;
4058 int node
= numa_node_id();
4059 struct delayed_work
*work
= to_delayed_work(w
);
4061 if (!mutex_trylock(&cache_chain_mutex
))
4062 /* Give up. Setup the next iteration. */
4065 list_for_each_entry(searchp
, &cache_chain
, next
) {
4069 * We only take the l3 lock if absolutely necessary and we
4070 * have established with reasonable certainty that
4071 * we can do some work if the lock was obtained.
4073 l3
= searchp
->nodelists
[node
];
4075 reap_alien(searchp
, l3
);
4077 drain_array(searchp
, l3
, cpu_cache_get(searchp
), 0, node
);
4080 * These are racy checks but it does not matter
4081 * if we skip one check or scan twice.
4083 if (time_after(l3
->next_reap
, jiffies
))
4086 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
;
4088 drain_array(searchp
, l3
, l3
->shared
, 0, node
);
4090 if (l3
->free_touched
)
4091 l3
->free_touched
= 0;
4095 freed
= drain_freelist(searchp
, l3
, (l3
->free_limit
+
4096 5 * searchp
->num
- 1) / (5 * searchp
->num
));
4097 STATS_ADD_REAPED(searchp
, freed
);
4103 mutex_unlock(&cache_chain_mutex
);
4106 /* Set up the next iteration */
4107 schedule_delayed_work(work
, round_jiffies_relative(REAPTIMEOUT_CPUC
));
4110 #ifdef CONFIG_SLABINFO
4112 static void print_slabinfo_header(struct seq_file
*m
)
4115 * Output format version, so at least we can change it
4116 * without _too_ many complaints.
4119 seq_puts(m
, "slabinfo - version: 2.1 (statistics)\n");
4121 seq_puts(m
, "slabinfo - version: 2.1\n");
4123 seq_puts(m
, "# name <active_objs> <num_objs> <objsize> "
4124 "<objperslab> <pagesperslab>");
4125 seq_puts(m
, " : tunables <limit> <batchcount> <sharedfactor>");
4126 seq_puts(m
, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4128 seq_puts(m
, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4129 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4130 seq_puts(m
, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4135 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
4139 mutex_lock(&cache_chain_mutex
);
4141 print_slabinfo_header(m
);
4143 return seq_list_start(&cache_chain
, *pos
);
4146 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
4148 return seq_list_next(p
, &cache_chain
, pos
);
4151 static void s_stop(struct seq_file
*m
, void *p
)
4153 mutex_unlock(&cache_chain_mutex
);
4156 static int s_show(struct seq_file
*m
, void *p
)
4158 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, next
);
4160 unsigned long active_objs
;
4161 unsigned long num_objs
;
4162 unsigned long active_slabs
= 0;
4163 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
4167 struct kmem_list3
*l3
;
4171 for_each_online_node(node
) {
4172 l3
= cachep
->nodelists
[node
];
4177 spin_lock_irq(&l3
->list_lock
);
4179 list_for_each_entry(slabp
, &l3
->slabs_full
, list
) {
4180 if (slabp
->inuse
!= cachep
->num
&& !error
)
4181 error
= "slabs_full accounting error";
4182 active_objs
+= cachep
->num
;
4185 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
) {
4186 if (slabp
->inuse
== cachep
->num
&& !error
)
4187 error
= "slabs_partial inuse accounting error";
4188 if (!slabp
->inuse
&& !error
)
4189 error
= "slabs_partial/inuse accounting error";
4190 active_objs
+= slabp
->inuse
;
4193 list_for_each_entry(slabp
, &l3
->slabs_free
, list
) {
4194 if (slabp
->inuse
&& !error
)
4195 error
= "slabs_free/inuse accounting error";
4198 free_objects
+= l3
->free_objects
;
4200 shared_avail
+= l3
->shared
->avail
;
4202 spin_unlock_irq(&l3
->list_lock
);
4204 num_slabs
+= active_slabs
;
4205 num_objs
= num_slabs
* cachep
->num
;
4206 if (num_objs
- active_objs
!= free_objects
&& !error
)
4207 error
= "free_objects accounting error";
4209 name
= cachep
->name
;
4211 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
4213 seq_printf(m
, "%-17s %6lu %6lu %6u %4u %4d",
4214 name
, active_objs
, num_objs
, cachep
->buffer_size
,
4215 cachep
->num
, (1 << cachep
->gfporder
));
4216 seq_printf(m
, " : tunables %4u %4u %4u",
4217 cachep
->limit
, cachep
->batchcount
, cachep
->shared
);
4218 seq_printf(m
, " : slabdata %6lu %6lu %6lu",
4219 active_slabs
, num_slabs
, shared_avail
);
4222 unsigned long high
= cachep
->high_mark
;
4223 unsigned long allocs
= cachep
->num_allocations
;
4224 unsigned long grown
= cachep
->grown
;
4225 unsigned long reaped
= cachep
->reaped
;
4226 unsigned long errors
= cachep
->errors
;
4227 unsigned long max_freeable
= cachep
->max_freeable
;
4228 unsigned long node_allocs
= cachep
->node_allocs
;
4229 unsigned long node_frees
= cachep
->node_frees
;
4230 unsigned long overflows
= cachep
->node_overflow
;
4232 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu \
4233 %4lu %4lu %4lu %4lu %4lu", allocs
, high
, grown
,
4234 reaped
, errors
, max_freeable
, node_allocs
,
4235 node_frees
, overflows
);
4239 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4240 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4241 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4242 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4244 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4245 allochit
, allocmiss
, freehit
, freemiss
);
4253 * slabinfo_op - iterator that generates /proc/slabinfo
4262 * num-pages-per-slab
4263 * + further values on SMP and with statistics enabled
4266 static const struct seq_operations slabinfo_op
= {
4273 #define MAX_SLABINFO_WRITE 128
4275 * slabinfo_write - Tuning for the slab allocator
4277 * @buffer: user buffer
4278 * @count: data length
4281 ssize_t
slabinfo_write(struct file
*file
, const char __user
* buffer
,
4282 size_t count
, loff_t
*ppos
)
4284 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4285 int limit
, batchcount
, shared
, res
;
4286 struct kmem_cache
*cachep
;
4288 if (count
> MAX_SLABINFO_WRITE
)
4290 if (copy_from_user(&kbuf
, buffer
, count
))
4292 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4294 tmp
= strchr(kbuf
, ' ');
4299 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4302 /* Find the cache in the chain of caches. */
4303 mutex_lock(&cache_chain_mutex
);
4305 list_for_each_entry(cachep
, &cache_chain
, next
) {
4306 if (!strcmp(cachep
->name
, kbuf
)) {
4307 if (limit
< 1 || batchcount
< 1 ||
4308 batchcount
> limit
|| shared
< 0) {
4311 res
= do_tune_cpucache(cachep
, limit
,
4318 mutex_unlock(&cache_chain_mutex
);
4324 static int slabinfo_open(struct inode
*inode
, struct file
*file
)
4326 return seq_open(file
, &slabinfo_op
);
4329 static const struct file_operations proc_slabinfo_operations
= {
4330 .open
= slabinfo_open
,
4332 .write
= slabinfo_write
,
4333 .llseek
= seq_lseek
,
4334 .release
= seq_release
,
4337 #ifdef CONFIG_DEBUG_SLAB_LEAK
4339 static void *leaks_start(struct seq_file
*m
, loff_t
*pos
)
4341 mutex_lock(&cache_chain_mutex
);
4342 return seq_list_start(&cache_chain
, *pos
);
4345 static inline int add_caller(unsigned long *n
, unsigned long v
)
4355 unsigned long *q
= p
+ 2 * i
;
4369 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4375 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
, struct slab
*s
)
4381 for (i
= 0, p
= s
->s_mem
; i
< c
->num
; i
++, p
+= c
->buffer_size
) {
4382 if (slab_bufctl(s
)[i
] != BUFCTL_ACTIVE
)
4384 if (!add_caller(n
, (unsigned long)*dbg_userword(c
, p
)))
4389 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4391 #ifdef CONFIG_KALLSYMS
4392 unsigned long offset
, size
;
4393 char modname
[MODULE_NAME_LEN
], name
[KSYM_NAME_LEN
];
4395 if (lookup_symbol_attrs(address
, &size
, &offset
, modname
, name
) == 0) {
4396 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4398 seq_printf(m
, " [%s]", modname
);
4402 seq_printf(m
, "%p", (void *)address
);
4405 static int leaks_show(struct seq_file
*m
, void *p
)
4407 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, next
);
4409 struct kmem_list3
*l3
;
4411 unsigned long *n
= m
->private;
4415 if (!(cachep
->flags
& SLAB_STORE_USER
))
4417 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4420 /* OK, we can do it */
4424 for_each_online_node(node
) {
4425 l3
= cachep
->nodelists
[node
];
4430 spin_lock_irq(&l3
->list_lock
);
4432 list_for_each_entry(slabp
, &l3
->slabs_full
, list
)
4433 handle_slab(n
, cachep
, slabp
);
4434 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
)
4435 handle_slab(n
, cachep
, slabp
);
4436 spin_unlock_irq(&l3
->list_lock
);
4438 name
= cachep
->name
;
4440 /* Increase the buffer size */
4441 mutex_unlock(&cache_chain_mutex
);
4442 m
->private = kzalloc(n
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4444 /* Too bad, we are really out */
4446 mutex_lock(&cache_chain_mutex
);
4449 *(unsigned long *)m
->private = n
[0] * 2;
4451 mutex_lock(&cache_chain_mutex
);
4452 /* Now make sure this entry will be retried */
4456 for (i
= 0; i
< n
[1]; i
++) {
4457 seq_printf(m
, "%s: %lu ", name
, n
[2*i
+3]);
4458 show_symbol(m
, n
[2*i
+2]);
4465 static const struct seq_operations slabstats_op
= {
4466 .start
= leaks_start
,
4472 static int slabstats_open(struct inode
*inode
, struct file
*file
)
4474 unsigned long *n
= kzalloc(PAGE_SIZE
, GFP_KERNEL
);
4477 ret
= seq_open(file
, &slabstats_op
);
4479 struct seq_file
*m
= file
->private_data
;
4480 *n
= PAGE_SIZE
/ (2 * sizeof(unsigned long));
4489 static const struct file_operations proc_slabstats_operations
= {
4490 .open
= slabstats_open
,
4492 .llseek
= seq_lseek
,
4493 .release
= seq_release_private
,
4497 static int __init
slab_proc_init(void)
4499 proc_create("slabinfo",S_IWUSR
|S_IRUGO
,NULL
,&proc_slabinfo_operations
);
4500 #ifdef CONFIG_DEBUG_SLAB_LEAK
4501 proc_create("slab_allocators", 0, NULL
, &proc_slabstats_operations
);
4505 module_init(slab_proc_init
);
4509 * ksize - get the actual amount of memory allocated for a given object
4510 * @objp: Pointer to the object
4512 * kmalloc may internally round up allocations and return more memory
4513 * than requested. ksize() can be used to determine the actual amount of
4514 * memory allocated. The caller may use this additional memory, even though
4515 * a smaller amount of memory was initially specified with the kmalloc call.
4516 * The caller must guarantee that objp points to a valid object previously
4517 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4518 * must not be freed during the duration of the call.
4520 size_t ksize(const void *objp
)
4523 if (unlikely(objp
== ZERO_SIZE_PTR
))
4526 return obj_size(virt_to_cache(objp
));
4528 EXPORT_SYMBOL(ksize
);